You are on page 1of 169

POPULATION DYNAMICS

STUDY SEGREGATION AS A NONLINEAR


PHENOMENON

By
Hezi Yizhaq

SUBMITTED IN PARTIAL FULFILLMENT OF THE


REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
AT
BEN-GURION UNIVERSITY
OF THE NEGEV, BEER-SHEVA
DECEMBER 2001

c Copyright by Hezi Yizhaq, 2001


°
Abstract

key-words:
attractors, bifurcation, bistable systems, cross diffusion, Eckhaus instability, fronts,
nonlinear dynamics, linear stability, Lotka Volterra competition model, mathemati-
cal modeling, neighborhood change, nonlocal migration, numerical simulations, NWS
equation, pitchfork bifurcation, self-organization, stable and unstable manifolds, pat-
tern formation, reaction diffusion equations, residential segregation, tipping point phe-
nomenon, transition zone, Turing instability, transverse instability, zigzag instability

Residential segregation phenomena of mixed populations in urban environments


have extensively been studied by geographers and social scientists. In many cases
socio-economic considerations have been reported as the dominant driving forces for
segregation. This research attempts modeling segregation phenomena in mathemati-
cal terms using observations made by geographers and social scientists.
A new mathematical model of residential segregation in urban neighborhoods is
suggested. The model consists of evolution equations of three continuous variables,
population densities of immigrants (v) and of veterans (u), and a socio-economic
status variable (s). The equations describe instant spatial averages over distances
large in comparison to a single house and small in comparison to the city size. The two
populations affect one another in two possible ways: A direct one when the response of
one population to changes in the other is immediate due to prejudice preferences such
as the black-white preferences in American cities. Such considerations are suitable to
cases where aversion forces between the two population groups are the driving forces
for segregation. The second way that one population affects the other is in an indirect
way, when the response is mediated by status changes. The immigration to Israeli
cities from the former Soviet Union during the 90’s is an example where economic
considerations were dominant. In the basic model only status considerations have
been taken into account. Mathematically this means no direct coupling between the
population densities.
The main assumptions of the model are:
• All three variables can change in space (x) and time (t).

v
vi

• We associate the u-population with high socio-economic status and the v-


population with low status.

• The rate of change of each population is due to growth contributions (logistic


growth and status driven growth) and migration contributions (density and
status driven migration).

• The migration can be divided to short range migration which is modeled by


diffusion terms, and long range migration, which is modeled mathematically by
nonlocal terms.

• The status variable is affected by the population densities; u-residents raise the
status while v-residents lower it.

• To saturate the status growth, balancing terms are included in the status equa-
tion.

Mathematical analysis and computer simulations of the model reproduce a vari-


ety of socio-spatial phenomena. The model associates segregation with a pitchfork
instability of a symmetric uniform mixed state to a pair of asymmetric uniform mixed
states where one state has a majority of u-residents and the other has a majority of v-
residents. Beyond this instability the system becomes bistable and patterns involving
domains of one mixed population state in a neighborhood of the other are possi-
ble. Such patterns describe segregation; regions over-represented with v-residents are
surrounded by regions under-represented with v-residents. An opposite patterns are
hold for u-residents. The segregation becomes stronger and stronger as the status gap
between the two populations is increased. Variable segregation of this kind has been
reported for various ethnic groups in European and American cities. The existence
of variable segregation states is a new prediction of the model compared to other
continuous mathematical models of intra-urban segregation.
A second instability designates a transition to a stronger form of segregation
involving enclaves of pure population (strong segregation). Such strong segregation
has been reported mostly with black-white segregation in American cities. In an
asymmetric version of the model, enclaves of pure population in a neighborhood of
mixed population may also exist.
The model suggests a third type of mechanism for segregation which is new in the
context of urban geography research. This pattern of segregation which we denote as
weak segregation is related to a finite wavenumber instability (Turing instability) of
a homogeneous mixed population state. The asymptotic state beyond the instability
is a stationary periodic pattern representing a weakly segregated state, consisting of
alternating domains. Regions with increased values of u have lower values of v and vice
versa. The driving force of segregation is the positive feedback effect of migration due
vii

to status considerations: economically strong residents migrate to regions with higher


status and thus raise the status further. The migration of economically weak residents
away from such regions adds to this effect. Similar considerations hold for regions
with low status. Non-local migrations can increase the segregation between the two
populations and can also increase the wavelength of weak segregation patterns.
Neighborhood change processes such as displacement of transition zones (zones
between segregated domains) and tipping point phenomena (population inversion) are
associated with fronts propagation and phase space structures. Front solutions de-
scribing transition zones between different neighborhoods appear in parameter ranges
pertaining to bistability of two population states. They are confined to narrow but
finite spatial domains where the population densities and the status gradually change.
The width of the transition zones is controlled mainly by the different diffusion coef-
ficients and by the status gap between the two populations. Numerical simulations
show that the width increases as the status gap between the two populations is de-
creased and vice versa. If the symmetry between the two populations is broken the
transition zone starts moving otherwise it remains stationary. Its velocity is affected
by a number of factors including the growth rate of the two populations, housing
capacities, and non-local migration and is computed by numeric simulations. The
moving interface can be halted by non-local migration terms.
The sudden reversal of a population distribution (e.g. natives majority to immi-
grants majority) once a critical population threshold is exceeded has been termed
tipping point phenomenon. The tipping point phenomenon in the model is associated
with the stable manifold of the unstable uniform mixed population state that went
through the pitchfork bifurcation. This stable manifold divides the three-dimensional
phase space into two distinct parts each containing one of the two asymmetric mixed
population states. The stable manifold defines a threshold: perturbing one of these
asymmetric states beyond the stable manifold leads to a convergence to the other
asymmetric state or to tipping.
Analysis of time-evolution trajectories in phase space during the tipping process
reveals five stages: initial state, fast response, slow growth, rapid growth and con-
vergence. From a dynamical systems standpoint, these stages can be explained by
considering the distances of points along these trajectories to the stable and unstable
population states. In the vicinities of these states the dynamics are slow and approx-
imately linear whereas far away from them the dynamics are fast and nonlinear.
Another interesting phenomenon, that can be investigated with the proposed
model is the critical size of a minority enclave, below which the enclave shrinks and
disappears. The critical enclave size is a function of the parameters that control the
status gap between the two populations, and of the populations growth rates. As we
increase the status gap, the critical size of the enclave increases which reflects the fact
that it is much harder for a low status minority group to penetrate to high status
viii

neighborhood. For a higher growth rate of the minority population, the critical size
gets smaller. From the geographic point this phenomenon of critical size is related to
the critical mass of the minority population that may trigger the institutionalization
process which will enhance the migration to the enclave (chain migration). On the
majority side, the concentration of minority families should be sufficiently large in
order to trigger an irreversible process (in the black-white segregation context it is
called a white flight). During this stage the majority households leave the area as
they realize that their way of life is threatened and they may be better-off by moving
elsewhere in the city.
In addition, the model suggests possible processes that have not been reported
yet, such as morphological instability leading to a finger-like invasion of one domain
into the another. This instability, which exists in our model leads to a corrugated
enclave shape. These uneven shapes are usually explained by anisotropy of the urban
space, but from the point of view of dynamical systems it may appear even in a
homogeneous environment.
Adding prejudice considerations between the two populations reduces the domain
of coexistence of stable mixed states and increases the coexistence range of stable
pure segregation. This extension of the model sheds light on the long debate about
the reasons for ethnic segregation.
The results of this research have been compared with results of discrete cellular
automata (CA) type models proposed by urban geographers. Good agreement has
been obtained with City-1, a model proposed by Prof. Portugali and coworkers. The
model deals with two population groups, new immigrants and veterans. It was found
that when the veterans are antagonistic to new immigrants, a spatial segregation be-
tween the two groups occurs all over the city. A similar result has been obtained with
our continuum urban model that includes a direct coupling between the two residen-
tial groups. In contrast to the City-1 model our model predicts spatial segregation
even when only economic considerations are taken into account.
The model is an over simplification of a complex urban residential reality, yet it
captures some basic urban processes and sheds light on their mechanisms and their
inter-relations. The main significance of the model lies in the conceptual framework
it introduces by relating socio-spatial phenomena to dynamical system and pattern
formation theories. Here also lies the advantage of the present model over CA models,
which do not lend themselves to mathematical analysis as continuum models do, and
thus are not helpful in revealing the mathematical rationale underlying socio-spatial
phenomena.
Besides the urban context, the model equations have interesting pattern formation
aspects. The uniform mixed population state of the symmetric version of the model
may become unstable in two distinct ways: (a) by a pitchfork bifurcation to a pair
of asymmetric uniform states, and (b) by a Turing instability characterized by an
ix

intrinsic wavelength that depends on diffusion coefficients and kinetic constants. Both
instabilities lead to pattern formation. The two asymmetric states created by the
pitchfork bifurcation allows for domain patterns involving islands of one state in
neighborhoods of the other state. These patterns are affected by possible instabilities
of the interface separating domains of different states, and by nonlocal couplings.
The Turing instability leads to small amplitude stationary periodic patterns. These
patterns may be affected by secondary instabilities such as Eckhaus and zigzag. The
two mechanisms of pattern formation may also act in a coupled way, leading to large
amplitude periodic patterns.
The Turing instability of the uniform mixed population state was first studied by
performing a linear stability analysis. Conditions for the instability have been found
and tested by numerically solving the model equations. The new states the system
evolves to as a result of the Turing instability were studied by deriving an amplitude
equation. The amplitude equation, which took the form of a Newell-Whithead-Segel
(NWS) equation, has stationary periodic solutions, representing stripe patterns in
two space dimensions. These solutions have been compared with stationary periodic
solutions of the original model equations and a good agreement has been found. The
amplitude equation has been used to study secondary instabilities, that is, instabilities
of the periodic solutions. Criteria, expressed in terms of the model parameters, have
been found for the Eckhaus and the zigzag instabilities. Other instabilities such as
cross-roll instability that changes the orientation of the stripes and oblique instability
also have been found numerically by solving the model with appropriate choice of
parameters and initial conditions.
The zigzag patterns also arise spontaneously in textures that form in a weakly
confined system from a uniform state. Starting from a uniform state with random
noise, the concentration slowly evolves to form a system of well organized domains,
separated by lines of defects. The stripes are not parallel, but clearly form local zigzag
patterns. These results show good agreement with numerical simulations reported for
the Schnackenberg model, a simplified variant of the well known “Brussellator” model.
In the parameter range where bistability of two asymmetric uniform states exists,
front solutions appear. We have analysed the stability of these solutions to modula-
tions of a linear front line and found conditions for a linear front instability that leads
to labyrinthine patterns. The analysis was carried out semi-analytically by linear sta-
bility analysis of the front in 2D and computing the eigenvalues of Jacobian matrix.
Carrying out this computation involves a calculation of eigenvalues of a huge sparse
matrix and the method is very sensitive to the choice of parameters that control the
resolution. Analysis has been done for the FHN model where the analytic result is
known and this helped to determine the optimal choice of parameters for this compu-
tation. The results for the proposed model stand in good agreement with numerical
solutions of the full model.
x

It is shown by numerical simulations that labyrinthine patterns can evolve even


when the front is linear stable due a nonlinear transverse instabilty. This new mech-
anism results from the Turing instability of the unstable branch (UHSS) i.e where
a spatial mode different from zero is selected. This scenario is a generic mechanism
and was observed in other models such as Swift-Hohenberg model (SH) and forced
complex Ginzburg Landau model (FCGL).
When the Turing instability of the uniform mixed population state and the pitch-
fork bifurcation of this state occur close to one another they can no longer be treated
independently. In this parameter range we find that the Turing instability (finite
k instability) is amplified by the coupling to the pitchfork bifurcation (zero mode
instability) and large amplitude patterns form. Numerical studies of this coupling
between the two instabilities have been done in one and two dimension. The asymp-
totic patterns depend on the ratio between the growth rate of the of the mode with
maximum growth rate to the growth rate of the zero mode. For lower values of this
ratio large domain patterns dominant and as we increase the value of this ratio strip
patterns becomes dominant.
Numerical simulations with global coupling terms show an increasing of the Tur-
ing space in the monostable regime and an increasing of the amplitude of the Turing
patterns. In the bistable regime these terms lead to a coarsening of the Turing pat-
terns. During this coarsening process, small domains fuses to create larger alternating
domains.
The implication of Turing Instability to ecological systems is included in appendix
B of this dissertation. It is shown that two competing species modeled by the classical
symmetric Lotka-Volterra (LV) equations with long range diffusion and cross diffusion
interaction may lead to a spatial segregation due to a Turing kind instability of the
mixed state. It is a well known result that the LV model with self diffusion terms,
that stands for migration of species to less crowded areas, does not support Turing
patterns. It is shown also that adding only cross diffusion terms that describe repulsive
force between the two species is not sufficient to reproduce spatial patterns.
Adding higher order diffusion terms (i.e forth order spatial derivatives) to the LV
model may induce a diffusion instability. These higher order diffusion terms represent
a long range diffusion. The importance of the long range diffusion lies in situations
where the density of the species is not small and the Fickian diffusion is not sufficiently
the accurate description.
Even the pure states (i.e states that consist from one species only) can undergo a
diffusion instability which leads to a coexistence of species in spite of the high value of
interspecific competition coefficient in contrast to the competitive exclusion principle.
The competitive exclusion principle, is one of the fundamental ideas in population
ecology, asserts the impossibility of the stable coexistence of two or more species
competing for the same resources in the same restricted area.
xi

Although heterogeneity has long been recorded in ecology, it has traditionally been
recorded as a consequence of environmental variability. The extended LV competition
model suggests that the distribution of populations in space may thus result from
intrinsic mechanisms, that is, the coupling between species interaction and different
kind of short range and long range diffusion terms. The addition of long range
diffusion terms to population dynamics models can show symmetry-breaking in a
wide range of models. Such phenomena are of general importance in understanding
diversity in ecological systems.
Acknowledgements

I would like to thank Ehud Meron, my supervisor, for his many suggestions and
constant support during this long and tedious research, and for his great help during
all last five years.
To Dr.Boris Portnov for his help and discussing the urban aspects of this work.
To Prof. Yair Zarmi for his help in mathematical analysis.
To Aric Yoechelis with whom I discussed the transverse instability analysis , and
to Jost von Hardenberg that helped me with the numerical codes.
To Lilian Naman the secretary of the department of Energy and Environmental
Physics and Helen from the Jane Schapiro library of the Jacob Blaustein Institute.
To Hadassa Soreq in helping me getting the grants on time.
Of course, I am grateful to my wife Ofra and my charming children Noa and Gal
for their patience and love. For being so many hours without me. Without them this
work would never have come into existence (literally).
Finally, I wish to thank my mountain bike that let me relax during the hard
work and the magnificent desert around Sde-Boker area.

Sde Boker, Israel Hezi Yizhaq


January , 2002

xii
Table of Contents

Abstract v

Acknowledgements xii

Table of Contents xiii

1 Introduction 1
1.1 Introduction to population dynamics in urban environments . . . . . 4
1.1.1 Ethnic segregation in cities . . . . . . . . . . . . . . . . . . . . 4
1.1.2 The human ecology approach . . . . . . . . . . . . . . . . . . 6
1.1.3 Massey’s model of ethnic segregation . . . . . . . . . . . . . . 7
1.1.4 Self-Organization and the City . . . . . . . . . . . . . . . . . . 8
1.2 Mathematical models of intra-urban population dynamics . . . . . . . 9
1.2.1 Discrete models . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2.2 Continuum models . . . . . . . . . . . . . . . . . . . . . . . . 11

2 Proposed model 21
2.1 Model equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.1.1 Rescaling equations . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Stationary and homogeneous solutions . . . . . . . . . . . . . . . . . 25
2.3 Linear stability analysis to uniform perturbations . . . . . . . . . . . 26
2.4 Turing instability analysis . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4.1 Turing instability in population models . . . . . . . . . . . . . 30
2.4.2 Turing instability of the M state . . . . . . . . . . . . . . . . . 31
2.4.3 Turing instability analysis of the non symmetric mixed states . 32
2.5 Reducing the the model to Lotka-Volterra competition model . . . . . 37
2.5.1 Liapunov function to the reduced model . . . . . . . . . . . . 39
2.5.2 Turing instability analysis of Eqs. 2.5.1 . . . . . . . . . . . . . 42

xiii
CONTENTS xiv

3 Urban Dynamics Aspects of the Model 43


3.1 Segregation mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2 Neighborhood change . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2.1 Transition zones . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2.2 Tipping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2.3 Expansion of a minority enclave . . . . . . . . . . . . . . . . . 52
3.2.4 Uneven expansion of a minority enclave . . . . . . . . . . . . . 57
3.3 Patterns with nonlocal migration . . . . . . . . . . . . . . . . . . . . 60
3.4 Weak segregation as a result of Turing instability of the M state . . . 64
3.5 Including discrimination in the model . . . . . . . . . . . . . . . . . . 66
3.6 Making the control parameter µ a dynamic variable . . . . . . . . . . 72

4 Pattern Formation Aspects 74


4.1 NWS amplitude equation for the Turing instability of M state . . . . 75
4.1.1 Derivation of the NWS equation . . . . . . . . . . . . . . . . . 77
4.1.2 Numerical testing of the amplitude equation . . . . . . . . . . 82
4.1.3 Amplitude instabilities . . . . . . . . . . . . . . . . . . . . . . 85
4.2 Patterns from ”Turing instability” of the unstable M state . . . . . . 91
4.3 Transverse instability . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.3.1 Numerical linear stability analysis of interface . . . . . . . . . 100
4.3.2 Testing the method with the analytic result of the FHN model 101
4.3.3 Numerical results for the urban model . . . . . . . . . . . . . 103

5 Numerical Methods 109


5.1 Spatial Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.2 Time Integration Method . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.3 Numerical Simulation of Global Coupling . . . . . . . . . . . . . . . . 113
5.4 Front Transverse Instability Analysis . . . . . . . . . . . . . . . . . . 114

6 Discussion 116
6.1 Urban Dynamics Modeling Aspects . . . . . . . . . . . . . . . . . . . 116
6.2 Pattern Formation Aspects . . . . . . . . . . . . . . . . . . . . . . . . 121

Bibliography 124

A Urban Hierarchy 140

B Spatial patterns in LV competition model 141


B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
B.2 The LV system with long range diffusion terms . . . . . . . . . . . . . 144
CONTENTS xv

B.3 Turing instability analysis . . . . . . . . . . . . . . . . . . . . . . . . 146


B.4 Turing patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
B.5 Morphological instabilities . . . . . . . . . . . . . . . . . . . . . . . . 150
B.6 Ecological implications . . . . . . . . . . . . . . . . . . . . . . . . . . 151
List of Figures

1.1 Bifurcation diagram for Beckmann & Puu model . . . . . . . . . . . . 19

2.1 Bifurcation diagrams for stationary uniform solutions . . . . . . . . . 28


2.2 Bifurcation diagram for the asymmetric model . . . . . . . . . . . . . 29
2.3 Finite wave instability of M state. . . . . . . . . . . . . . . . . . . . . 32
2.4 The polynom p(k 2 ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.5 Bifurcation diagram with Turing instability domains . . . . . . . . . . 36
2.6 Growth curves for the asymmetric model . . . . . . . . . . . . . . . . 38
2.7 Bifurcation diagram of the Lotka-Volterra competition model . . . . . 40

3.1 Segregation forms in the symmetric mode . . . . . . . . . . . . . . . . 45


3.2 Two index of segregation for the symmetric model as a function of
the parameter µ. Decreases the value of µ increases the segregation as
depicted by the dissimilarity and the exposure indexes. . . . . . . . . 46
3.3 Stationary fronts for the symmetric model . . . . . . . . . . . . . . . 48
3.4 The transition zone between P− and P+ . . . . . . . . . . . . . . . . 49
3.5 Trajectories near the saddle point . . . . . . . . . . . . . . . . . . . . 52
3.6 Time dependence of v population . . . . . . . . . . . . . . . . . . . . 53
3.7 Time derivative of v population . . . . . . . . . . . . . . . . . . . . . 54
3.8 Temporal evolution of the interface zone . . . . . . . . . . . . . . . . 55
3.9 Time evolution of u and v at a point that is initially at N+ state . . . 56
3.10 Critical radius rc of a minority enclave as a function of µ . . . . . . . 57
3.11 Undulating transition zone . . . . . . . . . . . . . . . . . . . . . . . . 59
3.12 Enclave expansion in a heterogeneous environment . . . . . . . . . . 60
3.13 Effect of nonlocal migration on weak segregation patterns . . . . . . . 62
3.14 Development of a stationary front due to nonlocal migration . . . . . 63
3.15 Segregation as a result of nonlocal migration . . . . . . . . . . . . . . 64
3.16 Turing patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.17 The effect of nonlaocal migration on weak segregation patterns- two
dimensional simulation. . . . . . . . . . . . . . . . . . . . . . . . . . . 66

xvi
LIST OF FIGURES xvii

3.18 Bifurcation diagram for the extended model with direct coupling be-
tween u and v . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4.1 Behavior of the natural stability curve, η(k) near the critical wavenum-
ber kc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.2 Amplitude of u vs η . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.3 Stability boundaries of the amplitude equation . . . . . . . . . . . . . 84
4.4 Numerical testing of the amplitude equation . . . . . . . . . . . . . . 86
4.5 Initial condition of Figure 4.4 . . . . . . . . . . . . . . . . . . . . . . 87
4.6 Formation of zigzag patterns in a weakly confined system . . . . . . . 88
4.7 Formation of zigzag patterns for η = 0.04 . . . . . . . . . . . . . . . . 88
4.8 Cross roll instability . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.9 Oblique instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.10 Development of Eckhaus instability . . . . . . . . . . . . . . . . . . . 90
4.11 Defect creation due the Eckhaus instability . . . . . . . . . . . . . . . 91
4.12 Zigzag stripes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.13 Zigzag instability for η = 0.02, q = −0.4 . . . . . . . . . . . . . . . . 92
4.14 Growth curves showing the finite wavenumber instability of the unsta-
ble M . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.15 Boundaries for Turing instability of the unstable M state . . . . . . . 94
4.16 One dimensional stationary patterns that evolve from random initial
perturbation of the unstable state . . . . . . . . . . . . . . . . . . . . 95
4.17 Large amplitude patterns for Ω ' 1.01 . . . . . . . . . . . . . . . . . 96
4.18 Labyrinthine patterns for Ω = 1.467 . . . . . . . . . . . . . . . . . . . 97
4.19 Ω and kmax as a function of δ . . . . . . . . . . . . . . . . . . . . . . 97
4.20 Labyrinthine pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.21 The NIB front bifurcation and transverse instability boundaries for the
FHN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.22 Transverse instability boundaries . . . . . . . . . . . . . . . . . . . . 104
4.23 Dynamics in the different regions . . . . . . . . . . . . . . . . . . . . 106
4.24 The evolution of an initially perturbed front that it is linear stable to
transverse perturbations . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.25 The interaction of Turing instability of the stable state with the trans-
verse instability. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

B.1 Dispersion curves for the M state . . . . . . . . . . . . . . . . . . . . 147


B.2 Dispersion curves for the bistable states . . . . . . . . . . . . . . . . . 148
B.3 Turing patterns beginning from the M state . . . . . . . . . . . . . . 149
B.4 Turing patterns starting from the P+ . . . . . . . . . . . . . . . . . . 150
B.5 Coexistence between Hπ and H0 hexagons . . . . . . . . . . . . . . . 151
B.6 Linear stable front . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
LIST OF FIGURES xviii

B.7 Nonlinear front transverse instability . . . . . . . . . . . . . . . . . . 152


List of Tables

1.1 Mathematical modelling methods of intra-urban population patterns . 20

2.1 Stationary uniform solutions of the symmetric model . . . . . . . . . 26

3.1 Urban interpretation of the parameters . . . . . . . . . . . . . . . . . 48


3.2 Representative papers on social class and ethnic class models for ethnic
segregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.3 Representative papers on social and ethnic class models for ethnic seg-
regation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.4 Stationary uniform solutions of Eqs. 3.6.1 . . . . . . . . . . . . . . . 73

6.1 A mapping between urban dynamics phenomena and dynamical system


concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

xix
Chapter 1

Introduction

A dynamical model of a human society begins with the realization that,


in addition to its internal structure, the system is firmly embedded in an
environment with which it exchanges matter, energy, and information.
Think, for instance, of a town in which a raw materials and agricultural
products arrive continuously, finished goods are exported, while mass me-
dia and professional communication keep the various groups aware of the
present situation of the immediate trends.

Gregoire Nicolis [1]


The urban environment resembles in many respects a nonlinear dissipative system
maintained out of equilibrium by influx and outflux of energy and matter [2]. The
driving force that keeps the urban system far from equilibrium is a continuous inflow
of information, raw materials or residents. Instabilities often break the symmetry of
the system and lead to self-organizing phenomena, such as the appearance of spatial
patterns. These phenomena involve collective behavior where the various modes of
the system follow the evolution of a few critical modes.
Dissipative systems have generic properties common to complex systems arising in
different disciplines such as: chemical systems [3], fluids [4] and granular materials
[5, 6]. They commonly evolve to a finite asymptotic states, called attractors, from
a broad range of initial conditions. The existence of attractors can often simplify
the dynamical description of the systems and reduce the number of variables. The
nonlinear nature of the system allows for multiplicity of attractors (e.g bistability,
tristability etc. ), which means that the time evolution of the system depends on the

1
CHAPTER 1. INTRODUCTION 2

initial state and that threshold behaviors may occur.


The nonlinear nature of the system may show up as a saturated growth process, as
a driving force of a decay process, or as a qualitative change in the system behavior as
a control parameter is varied, a phenomenon which called bifurcation or a stratigraphic
change in the urban context [2].
We construct continuous evolution equations for population densities regarded as
averaged quantities over space which are large relative to a single house but small with
respect to the system size. In the context of these equations segregation phenomena
are tightly related to multiplicity of uniform population states and to instabilities
thereof. The evolution equations take the form of reaction diffusion equations which
have been studied in the context of chemical reactions [4], biological systems [7] and
ecological systems [8, 9]. The thesis is divided into six chapters:

Chapter 1 presents the residential segregation background and mathematical


models of intra-urban population dynamics.

Chapter 2 explains the proposed model, and investigates the homogeneous


solutions and their stability to uniform and non uniform perturbations.

Chapter 3 relates the model to urban dynamics phenomenon, mainly to segre-


gation mechanisms and to interface dynamics between two adjacent neigh-
borhoods. This chapter also investigates an extended version of the model
with a direct coupling between the two populations groups.

Chapter 4 presents a mathematical analysis of the model’s instabilities: Tur-


ing instability of the mixed state by deriving an amplitude equation, and
analyzes the transverse instability that leads to labyrinthine patterns.

Chapter 5 describes the numerical techniques that are used in the numerical
simulations.

Chapter 6 discusses the model and its relevance to population dynamics in


an urban environment, and its contributions to pattern formation theory.

Appendix A describes shortly the urban hierarchies.


CHAPTER 1. INTRODUCTION 3

Appendix B presents a modification of the classical Lotka-Volterra two species com-


petition model by adding long range diffusion which may induce Turing instability
even in the symmetric case.
CHAPTER 1. INTRODUCTION 4

1.1 Introduction to population dynamics in urban


environments
The spatial segregation and concentration of population groups are as old
as the hills. They go back at least to 2000BC or thereabouts, when the city
of Babylon was described as being composed of distinct quarters.

Ronald van Kempen and A. Sule Özüekern [10]

Residential segregation in cities has been studied extensively by social scientists


from different fields: economy, sociology, and geography [11, 12, 13, 14, 10, 15, 16, 2].
The interest in this issue is due in part to a common perception of residential segrega-
tion as an undesirable, and even dangerous process. As commonly acknowledged, resi-
dential segregation could restrict the participation of disadvantaged population groups
(specifically immigrants) in various aspects of civil society, limits socio-economic op-
portunities of weak population strata, and has negative effects on various aspects of
local development, such as provision of commercial facilities and social services in
urban neighborhoods with economically weak populations [10].

1.1.1 Ethnic segregation in cities

Spatial segregation can be seen as the residential separation of groups within a broader
population [10]. A group is said to be completely mixed in a spatial sense when
its members are distributed uniformly throughout the city. Spatial segregation exists
when some areas show an over-representation and other areas an under-representation
of members of a group. Spatial segregation can appear on different scales, from the
city scale to the house scale. Spatial segregation implies spatial concentration and
both are often considered negative or, at best temporary problems as they curtail
the opportunities for the segregated members to participate in the civil life. Spatial
segregation often relates to segregation in school and it has been shown that children
with a foreign background have less chance of receiving a good education if they live
in a concentration area.
The urban ghetto or ethnic enclave has no official definition. As it is typically used
in discussion of poverty, the terms refer to inner-city neighborhoods with very high
CHAPTER 1. INTRODUCTION 5

levels of poverty [17]. In the U.S.A, these neighborhoods are usually predominantly
black or Hispanic. Segregation or unevenness appears in many aspects of life such as,
socio-economic status, life cycle and ethnicity.

There are many dimensions to the spatial segregation of different racial or ethnic
groups and no one can capture them all [18]. One which is often used is the index
of dissimilarity. Indexing areas by i, it is defined as:

1 X ¯¯ vi ui ¯¯
N
diss = ¯ − ¯ (1.1.1)
2 i=1 V U

where vi is the number of v residents in area i, V is the total number of v-residents


in the city and similar definitions hold for u-residents. If the two populations are
distributed evenly throughout the city diss will equal zero. If the two groups reside
in completely different areas, diss will equal one. This index can be shown to answer
the question: what share of v population would need to change areas in order to get
even distribution ? [19] Typically, a dissimilarity index of less than 0.3 is considered
low, an index between 0.3 and 0.6 is considered moderate, and an index above 0.6 is
considered high [20]. Applied to racial segregation by level of SES (socio-economic
status), diss measures the extent to which two populations of the same SES live in
the same neighborhood [21].

Another index of segregation that is relevant to this work is the exposure index
x Py *. It measures the racial residential segregation in terms of the average proportion
of the population that is of one group(y) in the neighborhoods occupied by another
group(x) [21]. The exposure index is defined as:
X
x Py ∗ = (vi /V )(ui /ti ) (1.1.2)
i

where ti is the total population of neighborhood i. Applied to residential segregation


by level of SES, x Py * measures the extent to which v-population live in neighbor-
hoods with u-residents of the same SES, that is exposed to them. x Py *=0.5 means
that the two groups are evenly distributed throughout the city, and x Py *=0 means a
strong segregation. These two indexes are more relevant in analyzing real urban data
CHAPTER 1. INTRODUCTION 6

for a city that consists of many different neighborhoods or tracts, in which each of
them is perfectly homogeneous (see Appendix A for more details about geographic
hierarchy of the city and about census tracts). For this research which concerns with
the dynamics of two different neighborhoods, their use will be limited.

1.1.2 The human ecology approach


One of the classic models of neighborhood change is the invasion-succession model
developed by the Chicago School [22]. The terms ”invasion” and “succession”, taken
from plant and animal ecology, were used to describe the process of neighborhood
population alteration. The driving force of the succession process is competition for
housing between the newcomers and the local population [23]. Invasion and succes-
sion involve a chain reaction, with each preceding immigrant wave moving outwards
and being succeeded by more recent, poorer immigrants [10]. The final pattern of
segregation was a “mosaic of social worlds” and seen as a natural equilibrium.
Through the years this model has been used to describe neighborhood racial
change and social-status transitions. Duncan & Duncan [24] identified four basic
stages: penetration, invasion, consolidation, and piling up. The Duncans used the
term “succession” in a more general sense:

An area undergoes “succession” when one type of land use replaces an-
other. The term ”residential succession” means, more specifically, the
replacement of one population group by another.

The ecological succession model contends that new immigrants concentrate in eth-
nic ghettos or low-cost housing areas and will move to better neighborhoods only
after they improve their socio-economic status. The human ecology approach was
criticized by others who claimed it paid little attention to the actual factors that in-
fluence succession and neglected intentional and political factors. It was argued that
the invasion-succession model was relevant only to American cities or to black-white
segregation. The applicability of these ideas to other places was questioned [25]. An-
other criticism concerns the use of ecological models to human systems as the former
is based on cultural and social processes which the ecological models do not take into
account directly.
CHAPTER 1. INTRODUCTION 7

1.1.3 Massey’s model of ethnic segregation


In an important paper Massey [26] revised the Chicago ecology theory and proposed a
modern theory for segregation which emphasized the interplay of two opposing spatial
forces: concentration and dispersion. The concentration of ethnic groups is rooted in
the spatial differentiation of the urban economy (e.g. low cost housing ringing city
industrial centers, close to jobs and not requiring payment for public transport), and
reinforced by the nature of immigrants and immigration. Concentration is driven by
succession which is the process of neighborhood change that occurs when a new ethnic
group enters a residential area and displaces the original inhabitants. Succession in
turn is affected by institutionalization and chain migration.
Institutionalization occurs in a neighborhood once it achieves a critical mass of im-
migrants, ethnic stores spring up, and specialized services are established etc., giving
rise to a clearly identifiable enclave of ethnic institutions. This increases the area’s at-
tractiveness to immigrants and plays as a positive feedback mechanism, contributing
to the maintenance of ethnic solidarity and community involvement, so that neigh-
borhoods are more likely to retain their ethnic residents and remain segregated. The
institutionalization process reinforces the chain migration which guides immigrants
to the neighborhood by social networks or family relations. If migration continues
after an area has become predominantly migrant, the ethnic enclave can expand to
adjacent areas.
Contracting concentration or succession is the process of dispersion or spatial as-
similation driven by acculturation which is the gradual acquisition of language, values
and manners of the host society, and by social mobility. Immigrants are encouraged
to move in order to improve their socioeconomic status and position in society. Thus,
dispersion acts to reduce segregation over time and to increase the mixing between
immigrants and veterans. This process is more effective when the social, economic,
and cultural differences between the two groups are small.
Massey examines data from the US, Canada, Australia, Western Europe and Is-
rael to evaluate his model. He excludes Black-White segregation from the analysis,
since research suggests [27] that black segregation stems more from white prejudice
than from the socioeconomic processes specified by the ecological model. Basically, he
founds out that, segregation patterns are consistent with this model. Massey’s model
will be an explanatory model for this thesis and in the following chapters concepts
CHAPTER 1. INTRODUCTION 8

such as dispersion, concentration and institutionalization will be formulated mathe-


matically. In view of the next subsection we can consider segregation or ethnicity as
an emergent phenomenon that results when people with similar characteristics share
a common position which defined by their socioeconomic status and their common
cultural identity.

1.1.4 Self-Organization and the City


Self organization is a property of complex nonlinear systems involving many degrees
of freedom. Under suitable conditions these systems can generate collective behaviors
were the many degrees of freedom are “enslaved” to a smaller set of leading degrees
of freedom or “order parameters”. Collective behaviors of this kind often lead to
pattern formation phenomena where ordered spatio-temporal spontaneously emerged
[4]. One of the famous example for this behavior is the Bènard cell: a layer of
fluid that is heated from below between two plates. As the temperature difference
is above some critical value, the fluid begins to exhibit series of convective cells.
The initial symmetry of the system has broken and a new pattern spontaneously
emerged. A typical Bènard cell in usual laboratory conditions is in the millimeter
range [28], whereas the intermolecular forces acts in Angstrom range, so this system
exhibits long-range correlations. The Bènard cell is a system that maintained far for
equilibrium (in this case equilibrium means no temperature difference between the
plates) by continuous supply of heat (energy) to the system.
Bènard cell exhibits the essential properties of self organizing systems, the system
is open and far from equilibrium and is driven by nonlinear mechanism which is
responsible for the threshold phenomena or the instabilities. By changing a control
parameter the system can undergo a series of bifurcations where the system moves
from one state to another. To quote Allen [29]:

The symbol of such a situation is that of a tree, for which there is neither
a single deterministic path, nor total freedom of action for the future from
wherever one starts.

If the system is near a bifurcation point, fluctuations about the average behavior can
kick the system to another branch, which means a new dissipative structure. This
new dissipative pattern persists until the next bifurcation point is reached [30]. Thus,
CHAPTER 1. INTRODUCTION 9

even in systems which their time evolution is governed by a deterministic set of partial
differential equations, the final state of the system is not predictable, because near
the bifurcation points the system is very sensitive to small perturbations, so it can
“selects” one of the many spatial structures.
In equilibrium the system always seeks to adopt the configuration that has the
lowest free energy. In self-organizing systems there is no such minimization principle
that would determine the choice of pattern adopted near the instability point. It was
even proven by Rolf Landauer [31] (p. 261) that such a principle can not be applied
to pattern selection out of equilibrium. Powerful mathematical techniques have been
developed over the years such as amplitude and phase equations, in order to study
the patterns selection problem [4].
The city can be regarded as an open system that is maintained far equilibrium by
a constant supply of goods and raw materials and by immigration and emigrations
of residents of different groups. The individuals in the city can exhibit a collective
behavior due to mass media and social interactions. Regarding the city as a self orga-
nizing system is consistent with the social aspects of individuals (homo socialis) [32].
The social person behave like others and thus, exhibit a kind of collective behavior.
In contrast, the economic person (homo economic) acts in a rational behavior on the
basis of the utility maximization principle. We can make the following interpretation
concerning these two types of behaviors, the economic person drives the urban system
towards an equilibrium point, whereas the social person drives the system far from
equilibrium and emphasizes the nonlinear nature of the city.

1.2 Mathematical models of intra-urban popula-


tion dynamics
Existing mathematical models of residential segregation can be divided into two main
categories: (i) discrete models and (ii) continuum models. In their turn, discrete
models are subdivided into regional models and cellular automata models and multi-
agent models. In the following subsections, each of these models will be discussed in
some detail.
CHAPTER 1. INTRODUCTION 10

1.2.1 Discrete models


In the framework of a regional model (RM), the city is divide to N discrete zones, each
of which is usually characterized by two sets of variables: the first variable represents
the properties of the city’s physical environment, and the second variable describes the
distinctive characteristics of the city’s population, such as the proportion of white and
blue collar residents [33, 34]. Each zone is described by a set of non-linear ordinary
differential equations which determined the time evolution of the number of jobs or
residents of a particular type at a given zone [35].
In contrast, the models of cellular automata (CA) represent urban space as two
dimensional grid. Each cell in the grid is found in some state that belongs to a
finite set. The cells evolve in discrete time steps according to local transition rules
which apply simultaneously to all sites on the grid. In most of the CA models the
transition rules of a site depend only on a local neighborhood of cells around it
[36, 37, 38, 39, 40, 41]. In the context of residential segregation in a city Portugali
and his coworkers [42] proposed a CA model called “City”. In “City” each cell
might be seen as a house or a place and occupied by individuals which belonged to
two groups blue and green, each of which subdivided according to the preferences
of its members into segregatives and neutrals. Individuals can occupy or leave their
sites according to probabilistic transition rules which are computed at each time step.
“City” is used as a heuristic tool for the study the sociospatial segregation and to
demonstrate some generic self-organization principles such as slaving, captivity and
emerging of global patterns from local interactions.
In order to account for the human aspects of populations in the city i.e the ability
of individuals to change their characteristics during the time evolution of the city and
to consider individuals desires and decisions in a direct manner, a more elaborate
approach was developed called Multi-Agent (MA) models [38]. MA models consists
of two layers, the first is the infrastructure layer which is simulated as a CA model,
and the second layer represents the residents (free agents) who can immigrate into
the city, occupy and migrate according to their decisions. These decisions are based
on regarding information from three levels of urban hierarchies: the individual, the
neighborhood and the whole city (which can be regarded as some kind of global
coupling interaction). The free agents are characterized by their economic status and
cultural identity which can also be changed during the time evolution, leading to the
CHAPTER 1. INTRODUCTION 11

emergence of new socio-cultural groups. Thus, we can regard the MA models as two
coupled CA models that simulate the interaction between the residents of different
groups and the city’s housing infrastructure. MA models can take into account the
cognitive gap or dissonance between individual’s wants and his actual behavior [32].
One example for this phenomenon within a city would be an individual living in a
neighborhood which does not fit his wants. This discrepancy between his intention
and behavior can be relaxed by a change of his wants or migrations to other site in
the city. We will consider this phenomenon and relates it to our proposed model in
Chapter 3 of this dissertation.

1.2.2 Continuum models


Continuum models (CM), described by partial differential equations represent space
with a continuous coordinate system whose smallest (infinitesimal) scales are still
large in comparison with a single house. Along this continuous space populations
interact and disperse. According to the underlying assumptions of such an approach,
urban systems evolve over space and time and are characterized by non-linear interac-
tions among the variables and geographical diffusion processes [43]. The continuum
modelling approach has the advantage of exploiting advanced methods of the theory
of pattern formation that are not available for the analysis of discrete models. These
methods include linear stability analysis, deriving amplitude and phase equations,
velocity-curvature relations and linear stability analysis of planar fronts. The im-
portance of these analytical methods is due to the fact that the city is an extended
system where spatial variations of the state variables may occur on length scales much
smaller than the system size.

Gurtin’s Model

One of the earliest studies of residential segregation using a continuum modelling


approach is that of Gurtin [44]. In his study, he addressed the tipping point phe-
nomenon, by which a new minority entering a neighborhood causes the evacuation
of the veteran population. Gurtin used the following simple linear model for two
cultural groups:
dn1
= a11 n1 + a12 n2
∂t
dn2 (1.2.1)
∂t
= a21 n1 + a22 n2 .
CHAPTER 1. INTRODUCTION 12

To account for situations in which the two groups repel each other, he assumed that
a12 < 0 and a21 < 0. The solution to Eq. 1.2.1 can be written as:
© ª
n1 (t) = eλt © ω1 (a12 N2 + βN1 ) sinh ωt + N1 cosh ωtª
(1.2.2)
n2 (t) = eλt ω1 (a21 N1 − βN2 ) sinh ωt + N2 cosh ωt

where β = (a11 − a22 )/2, λ = (a11 + a22 )/2, ω = (β2 + a12 a21 )1/2 , and N1 = n1 (0),
N2 = n2 (0). Gurtin introduced a tipping ratio A defined as

ω−β −a12
A= = (1.2.3)
−a21 ω+β

and concluded that if N1 /N2 < A the region will eventually be populated only by
group 2; if N1 /N2 > A the region will eventually be populated only by group 1.
Thus, the region will become segregated in a finite time for all choices of initial
conditions except those satisfying N1 /N2 = A which is rather a very speciel case.
Gurtin also defined the segregation time which is the time T at which, say, n1 (T ) = 0
(when N1 /N2 < A) and is given by
µ ¶
1 −1 ω
T = tanh (1.2.4)
ω αω + (α − 1)β

where α = AN2 /N1 > 1. The closer N1 /N2 is to the tipping ratio A, the longer it
will take for N1 to vanish. According to Eq. 1.2.1 the segregation between the two
groups is strong and coexistence is rare. Because the model is linear no bistability
of states can occur. It can be applied to black-white segregation and according to
Grodzin [45], Gurtin estimated the tipping ratio between blacks and white as 1/4.
Gurtin tried also to propose a PDE model which allows local population transport
mechanism i.e.diffusion. He considered situations in which the population flux of each
group depends only on the other group, and it is in a direction of decreasing density
of the other group. The model takes the form
∂n1 2
∂t
= c ∂∂xn22
(1.2.5)
∂n2 2
∂t
= c ∂∂xn21
CHAPTER 1. INTRODUCTION 13

According to this model any differences in density grow without bound and for every
x with n1 −n2 6= 0 either n1 or n2 will eventually vanish. This model has the drawback
that after the first vanishing of one of the populations, it will become negative. Eqs.
1.2.5 are too simple and the total population remains constant in time. No population
growth is included, but we will use some of his ideas as starting point for the proposed
model.
In another version, Gurtin added long-range local transport i.e. a model in which
people are allowed to migrate between any two points instantaneously. His idea is
based on a function

G1 (x, y, t) = ϕ1 (x, y) [n2 (y, t) − n2 (x, t)] (1.2.6)

where G1 (x, y, t) denotes the number of people of group 1, per unit lengths and per
unit time, flowing from y to x at time t. According to Eq. 1.2.6, G1 (x, y, t) is
proportional to the density difference n2 (y, t) − n2 (x, t) and the function ϕ1 (x, y) is
assumed to be known a priori. This choice for G1 (x, y, t) states that population 1
moves to locations with lower density of population 2. A similar equation holds for
the population 2 which means that the two populations repel each other. To account
for the total number of people of group 1, per unit time, entering x from all other
point in [0, L] we have to integrate over y, and the full model can be written as:

∂n1 (x,t) RL
∂t
= G1 (x, y, t)dy + a11 n1 (x, t) + a12 n2 (x, t)
0
(1.2.7)
∂n2 (x,t) RL
∂t
= G2 (x, y, t)dy + a21 n1 (x, t) + a22 n2 (x, t)
0

Eq. 1.2.7 is a system of integral-differential equations and Gurtin considers only a


simple case where a12 = a21 = 0, a11 = a22 = a and ϕi (x, t) = k/L > 0 are constant.
Eqs. 1.2.7 then take the simpler form
∂n1 (x,t)
= an1 − kn2 + Lk N2
∂t
∂n2 (x,t) (1.2.8)
∂t
= an2 − kn1 + Lk N1
CHAPTER 1. INTRODUCTION 14

where N1 is the total population 1 and is given by


ZL
N1 = n1 (x, t)dx . (1.2.9)
0

A similar equation holds for N2 . The system 1.2.8 has the following solution:
½ 1£ ¤ ¾
n1 (x, 0) + n2 (x, 0) − L1 (N1 (0) + N2 (0)) e−kt +
n1 (x, t) = e at 2 £ ¤ (1.2.10)
1
2
n1 (x, 0) − n2 (x, 0) − L1 (N1 (0) + N2 (0)) ekt + N1L(0) ,

with a similar form for n2 . From this solution the conditions for segregation can be
derived. Thus at points for which the initial density difference happens to be the
same as the average initial density difference, i.e.

n1 (x, 0) − n2 (x, 0) = hN1 (0)i − hN2 (0)i (1.2.11)

the density will remain positive for all time. If 1.2.11 is not satisfied, the population
of one of the groups will tend to zero at that point and this locality will become
segregated. Thus, for example, if at a point n1 (x, 0) − n2 (x, 0) > hN1 (0)i − hN2 (0)i is
valid then n2 (x, t) → 0 in a finite time. We will apply this idea of non-local migration
and the dependence on the average density of a population in the whole system, in
our proposed model in a slightly different manner.

O’Neill’s model

In another study, O’Neill [47] introduced a model that incorporates logistic growth
of a minority population with spatial diffusion:
∂n
= an(1 − n) + D∇2 n (1.2.12)
∂t
.
This equation is a classical model of ecological importance known as the Fisher equa-
tion, and is used to model the spatial spread of a favored gene. This model produces
waves of invaders that travel at a velocity c(t) which approaches an asymptotic veloc-

ity 4aD [8]. If the Fisher model is analyzed in two dimensions, circular waves form
CHAPTER 1. INTRODUCTION 15

and spread outward in the same asymptotic velocity as calculated in the one dimen-
sional case. The Fisher model has been used to make predictions of range expansion
using microscale data on individual movement for a variety of animals. The Fisher
equation has been extended to heterogeneous environments by allowing the growth
coefficient and the diffusion coefficient to vary in space [48]. By selecting appropriate
values of a(x) and D(x) it is possible to set up environments consisting of a mosaic
of patches.
O’Neill checked the validity of the equation in its radial formulation for data
collected from a seventy-seven block area in Chicago during 1968-72. According
to this model, the black ghetto’s growth is in the form of travelling wave, that is,
one of the stable solutions of Fisher’s equation [49]. It is important to note here
that the Fisher equation describes the dynamics of one population only, so in the
context of black-white dynamics it is assumed that the system is governed by the
black population parameters and the white population dynamics is enslaved or less
relevant to the specific case of minority group invasion.

Zhang’s model

More recently, Zhang [50] proposed a one-dimensional predator-prey model to de-


scribe the relation among residential density, land rent, and spatial diffusion effects.
He showed mathematically the non-existence of periodic solutions in time, and the
non-existence of periodic patterns in space. In another study [51], Zhang studied a
Lotka-Volterra competition model in order to simulate the dynamics of two popula-
tion groups competing for space:
ut = (²1 − µ11 u − µ12 v)u + D1 ∇2 u
(1.2.13)
vt = (²2 − µ21 u − µ22 v)v + D2 ∇2 v .

Here u(x, t) and v(x, t) are the group densities and x denote the distance from the
CBD (Central Business District) and µ12 and µ21 are the measure of the strength of
action of one group upon the other.1 The diffusion terms describe the tendency of
individuals to move to less populated areas in the city. Zhang studied the stationary
solutions of this model in one dimension with Neumann boundary conditions. Eq.
1.2.13 has four stationary and uniform solutions:
1
In ecology these terms are referred to interspecific competition coefficients and the µ11 and µ22
are called intraspecific competition coefficients.
CHAPTER 1. INTRODUCTION 16

1. The trivial solution (0, 0) which is unstable for all values of parameters.

2. A pure u-population solution: (0, ²1 /µ11 ) which is stable for µ²11 1


> µ²21
2
, µ²22
2
< µ²12
1
.
³ ´
²1 µ22 −²2 µ12 ²2 µ11 −²1 µ21
3. Both groups coexist: µ11 ,
µ22 −µ12 µ21 µ11 µ22 −µ12 µ21
, which is stable for µ²11 1
<
²2 ²2 ²1
,
µ21 µ22
< µ12 .
²1 ²2
4. A pure v-population solution: (0, ²/µ22 ) which is stable for µ11
< , ²2
µ21 µ22
> ²1
µ12
.
The system is bistable for µ²11
1
> µ²21
2
, µ²22
2
> µ²12
1
, either u-population or v-population
wins depending on the initial condition. For the symmetric case of Eq. 1.2.13 the
system can be bistable or only the mixed solution is stable (A bifurcation diagram
for this version is shown in Figure 2.7).
The system 1.2.13 has travelling wave front solutions of constant shape and speed
that connect the two pure states. Hosono [52] employed the matched asymptotic
expansion technique to prove the existence of this travelling wave front solution. The
speed of the travelling wave is heuristically derived in the following way (assume that
the v-population is the invading population and that the pure state (0, ²1 /µ11 ) is
unstable): At the wave front the densities of the two populations are approximately
given by (²1 /µ10 ) and so the the second equation of 1.2.13 can be approximately
expressed as µ ¶
2 ²1
vt = D2 ∇ v + ²2 − µ21 v (1.2.14)
µ11

which has the form of a known equation called the Skellam equation and its solution
for the front velocity is known. So we can write the minimum speed of the travelling
wave as:
p
c2 = 2 ²2 D2 (1 − ²1 µ21 /²2 µ11 (1.2.15)

If v population invades an area that is absent of any competing population i.e. the
trivial unstable state (0, 0) the second equation of 1.2.13 can be expressed as

vt = D2 ∇2 v + (²2 − µ22 v) v (1.2.16)

which has the form as Fisher equation 1.2.12 and the speed of the travelling wave is:
p
c̄2 = 2 ²2 D2 (1.2.17)
CHAPTER 1. INTRODUCTION 17

In this case the travelling wave front connects the pure state (0, ²/µ22 ) with the
unstable state (0, 0) state. It is obviously clear that the presence of a competing
population will lower the propagation speed of the invading population and also that
this speed will be lower as this competition is stronger (higher value of µ21 in 1.2.15).
The Lotka-Volterra competition model was used by Okubo et. al to model the spread
of the grey squirrel in Britain after it was introduced from North America [48].
Okubo and his colleagues performed numerical computations of the model in one and
two dimensions when expanding species (grey squirrel) meets uniformly distributed
competitors (red squirrel).
Jorné and Carmi [53] showed that Lotka-Volterra competition equations do not
have time oscillations (i.e. Hopf bifurcation, see the next chapter for details of this
proof). The Zhang model has only bistability of pure population states solutions
and the social interaction between the two ethnic groups is aversion. More moderate
interactions based on socioeconomic considerations were not taken into account. We
will show in the next chapter that the proposed model can be reduced to the Zhang
model.

Beckmann and Puu model

The last model we include in this introduction describes the time evolution of one pop-
ulation density so it can not account for residential segregation between two groups.
We choose to describe it here because this model predicts spatial periodic patterns
that emerge from an instability of the uniform states solutions to nonhomogeneous
perturbations. This mechanism is new in the context of population dynamics in urban
environment. The model was proposed by Beckmann and Puu [43]. In their model
they changed the simple logistic growth and added a nonlinear term that for small
densities increases and for high densities decreases. They also changed the diffusion
term so it describes migration of population to the direction of the steepest increase
of population density:
∂p 1
= p(1 + α(βp2 − p3 ) − γp) + ∇2 α(3βp2 − 2p3 ) (1.2.18)
∂t 6

where p(x, t) is the population density in a location x and in a given time t. p(x, t)
CHAPTER 1. INTRODUCTION 18

stands for the total population density without further division to subgroups. They
studied the stationary solutions and their stability. One solution is the trivial solution
p = 0 and it is always unstable. The number of non-trivial stationary and uniform
solutions depends on the discriminant D = Q3 + R2 where
γ
Q= 3α
− β9
3 (1.2.19)
R= 1

+ β27 − βγ

If D < 0 there are three different nontrivial stationary states and if D > 0 there is
only one. When D = 0 at least two roots are coincident. The stability of a state P
depends on the sign of:

µ = 1 + α(3βP 2 − 4P 3 ) − 2γP (1.2.20)

Stability obtains when µ < 0. A bifurcation diagram for α = 1.2 and β = 3 is


shown in Figure 1.1. The bifurcation parameter is γ. For 3.358 < γ < 3.426 the
system is bistable and exhibit hysteresis loop formed by two back-to-back saddle-
node bifurcations.
Beckmann and Puu showed that by adding the non-linear diffusion term each of
the bistable states can undergo a kind of diffusion instability which drives the system
to a pattern that consists of periodic domains with a wavelength that depends on the
amplitude. This model exhibits some of the essential features of non-linear system
such as bifurcations, multistability and diffusion instability and in their own words
[43]

we may expect economically sensible spatial population patterns with al-


ternating concentrations and rarefactions.

Continuum models can be analyzed with powerful analytical methods such as


linear stability, bifurcation analysis, order parameter equations for front motion [54,
55], amplitude equation [56] and phase equation. Table 1.1 summarizes in brief the
aforementioned and some other mathematical models of urban segregation. All the
models in question have at least two drawbacks. First, most of them predict that only
strong segregation (under which a particular urban area is populated predominantly
by one population group) may exist as an asymptotic state. This assumption does
CHAPTER 1. INTRODUCTION 19

Figure 1.1: Bifurcation diagram for stationary uniform solutions of Eqs. 1.2.18 show-
ing the density p. Solid (dashed) lines stand for stabe (unstable) solutions. The
trivial state p = 0 exists for all values of γ but is always unstable. Other values of
parameters: α = 1.2, β = 3.

not always hold since populations in urban areas tend to be mixed, with residents
of different ethnicity and income groups reside next to each other, specifically in
marginal areas between “pure” ethnic enclaves [14]. Second, these models do not
account for socio-economic status considerations which take place in the population
dynamics. They consider only direct interaction between the two populations as in
the ecological models. As we show in this dissertation adding the status consideration
allows for mixed population states.
CHAPTER 1. INTRODUCTION 20

Model Type Representation Papers Advantages Disadvantages


CA & Portugali (2000) [2] Can account for Not easily
Multi-agent Benenson (1999) [38] the behavior of amenable to
Morrill (1965, 1970) [57, 58] individuals components mathematical
Rose (1972) [59] of the system. techniques.
Woods (1981) [60] Can simulate
Hansell and Clark (1970) [61] changes in individuals
Gaylord & D’Andria properties in time. city
(1998) [40]
Regional Allen (1999) [35] Can simulate Very complex
Models Allen & Sanglier (1981) [62] the entire and have many
Allen et al. (1985) [33] urban system parameters that
Sanglier (1989) [63] and even an have to
Straussfogel 1991 [30] entire country. be defined.
Pumain (1987) [64] Not easily to
amenable to
mathematical
analysis.
Ecological Gurtin (1974) [44] Can be analyzed Difficulties
Models Hotelling (1978) [65] with powerful in parameters
Rosser (1980) [66] analytic methods estimate.
Ishikawa (1980) [67] of pattern
O’Neill (1981) [47] formation theories.
Zhang (1988) [50] Models with few
Zhang (1989) [51] variables that
Zhang (1990) [69] can lead to
Beckmann & Puu (1990) [43] a qualitative
Meron (1999) [68] understanding.
Yizhaq & Meron (2002) [70]

Table 1.1: Mathematical modelling methods of intra-urban population patterns


Chapter 2

Proposed model

. . . physico-chemical systems giving rise to transition phenomena, long


range order, and symmetry breaking far from equilibrium can serve as
an archetype for understanding other types of systems that show complex
behaviour-systems for which the evolution laws of the variables involved are
not known to any comparable degree of detail. More important, in many
of these systems the very choice of what should be a pertinent variable may
well be part of the problem we try to solve.

Grégoire Nicolis and Ilya Prigogine [28]

The approach in this dissertation to modeling segregation, is based on the following


assumptions:

• There are circumstances where migration within and into the city is mostly
affected by a few independent factors and all other aspects of complex urban
environment are (i) instantaneously determined (enslaved) by this factors, or
(ii) regardable as a small random force, or (iii) negligible. The immigration to
Israeli cities from the former Soviet Union during the 90’s is an example where
economic considerations were dominant.

• Segregation patterns can be characterized by a set of continuous state variable


(population densities, socio-economic status etc.) that depend on space and
time and describe instant averages over distances large in comparison to a single
house and small in comparison to the neighborhood or the city size [51].

21
CHAPTER 2. PROPOSED MODEL 22

• Population can be modeled in two ways: (i) each population is assigned a density
variable, (ii) a single population density is defined along with other variables
that encode the differences among the populations. The former case is suitable
in situations where a sharp distinction between two or more population can be
made, e.g. on the basis of national or ethnic identity. The latter case applies to
situations where the discerning property varies continuously between extreme
values, e.g. a Jewish population consisting of secular and orthodox residents
with various degrees of Jewish tradition awareness, or cultural identity [2] (p.
178)

2.1 Model equations


We consider two populations in a residential area whose densities are u(x, t) and
v(x, t). We will refer to these populations as to u-residents and v-residents, respec-
tively. The two populations represent different ethnic or socio-economic groups. We
parametrize this difference by a status variable s(x, t) and associate the u-residents
with high status and the v-residents with low status. All three variables may change
in space (x) and time (t). The rates of change of u and v have two contributions
pertaining to growth and migration:
∂u
= Gu (u, s) + Mu (u, s) ,
∂t
∂v
= Gv (v, s) + Mv (v, s) , (2.1.1)
∂t
The growth contributions are modeled by

Gu (u, s) = α1 u − α2 u2 + α3 (s − sR )u ,
Gv (v, s) = β1 v − β2 v 2 − β3 (s − sR )v , (2.1.2)

where all coefficients, αi and βi ; i = 1, 2, 3, are positive. The terms α1 u and β1 v de-
scribe natural growth while the terms −α2 u2 and −β2 v 2 account for growth saturation
due to limited housing capacity. The larger α2 (β2 ) the smaller the housing capacity
for u-residents (v-residents). The contributions α3 (s − sR )u and −β3 (s − sR )v, where
sR is a constant reference status, describe immigration into the city and emigration
out of the city driven by status considerations. A neighborhood with a high status
CHAPTER 2. PROPOSED MODEL 23

(s > sR ) attracts u-residents and expel v-residents while a low status neighborhood
(s < sR ) attracts v-residents and expel u-residents [71]. We chose these contribu-
tions to be proportional to the densities u and v in order to account for the effects
of institutionalization. A higher density of v-residents, for example, in a low status
neighborhood strengthens the local network of social institutions which facilitate the
absorption of new v-residents.
The migration contributions are modeled by
Z Z
Mu (u, s) = [u(x , t) − u(x, t)]pu (x − x)dx − [s(x0 , t) − s(x, t)]ps (x0 − x)dx0 ,
0 0 0

Z Z
Mv (v, s) = [v(x , t) − v(x, t)]qv (x − x)dx + [s(x0 , t) − s(x, t)]qs (x0 − x)dx0 ,
0 0 0

The first terms on the right hand sides represent migration from densely populated
regions to less populated ones. The second terms represent migration driven by status
considerations. In these expressions dx0 is an infinitesimal area element representing
a small part of the residential area that is still large in comparison to a single house,
pu , ps , qv , qs are positive weight functions and the integration extends over the whole
residential area. The integral forms of these terms reflect the fact that migration
within the city is generally nonlocal; a resident may move to a distant neighborhood
if this neighborhood meets his needs. Thus, the density of u-residents at location
x will increase due to migration from highly populated places x0 (u(x0 ) − u(x) > 0)
and decrease due to migration to less populated areas (u(x0 ) − u(x) < 0). Similarly,
the density of u-residents at location x will increase due to migration from places x0
with lower status (s(x0 ) − s(x) < 0), and decrease due to migration to places x0 with
higher status (s(x0 ) − s(x) > 0). Similar considerations hold for v-residents except
that status gradients act in the opposite sense.
While the migration within the city is generally nonlocal there are also consider-
ations which favor local migration [72]. Families, for example, will generally prefer
to move within their neighborhoods to avoid pulling out children from local schools
or to retain networks of social and organizational relations. These considerations
determine the shapes of the weight functions pu , ps , qv , and qs . We assume that
each of these functions has two contributions: a constant term representing nonlocal
migration where all migration distances are given equal weights, and a term peeked
at x0 = x representing local migration. A possible choice for pu is
pu (x0 − x) = au + bu exp(−|x0 − x|2 /d2 ) . (2.1.3)
CHAPTER 2. PROPOSED MODEL 24

In this equation, au is the constant contribution term and bu exp(− |x0 − x|2 /d2 ) is
the term that peaks at x’=(0,0) and decays as we go away from the origin. The
meaning of this term is that most migrants move over short distances (see distance-
decay curves in [72]. This behavior is more typical for poor people [17] (p.91).
Similar forms can be chosen for ps , qv and qs . Expanding u(x0 , t) around u(x, t) in
Taylor series we find
Z
[u(x0 , t) − u(x, t)]pu (x0 − x)dx0 = α4 [< u > −u] + D1 ∇2 u , (2.1.4)

where α4 and D1 are constant parameters, ∇2 is the Laplacian operator in the plane
and Z
1
< u >= u(x0 , t)dx0 (2.1.5)
A A
is the average of u over the whole residential area, A. Similar forms are obtained for
the other integral terms. Summing up all migration contributions we find
Mu (u, s) = D1 ∇2 u − D2 ∇2 s + α4 [< u > −u] − α5 [< s > −s] ,
Mv (v, s) = D3 ∇2 v + D4 ∇2 s + β4 [< v > −v] + β5 [< s > −s] . (2.1.6)
Equations (2.1.1), (2.1.2) and (2.1.6) describe how the socio-economic status af-
fects the population densities. The population densities affect the status as well. New
u-residents in a neighborhood raise the status whereas new v-residents lower it. This
is a positive feedback effect which leads to unbounded status values: high (low) status
attracts u-residents (v-residents) which raise (lower) the status further up (down). In
practice status values and status gradients can never grow or decrease without bound.
We therefore model the rate equation for s as
∂s
= γ1 u − γ2 v − γ3 (s − sR ) − γ4 (s − sR )3 + D5 ∇2 s , (2.1.7)
∂t
where the first two terms on the right hand side describe the positive feedback and
the rest are balancing terms. 1
Eqs. (2.1.1), (2.1.2), (2.1.6) and (2.1.7) constitute the model:
∂u
∂t
= α1 u − α2 u2 + α3 (s − sR )u + D1 ∇2 u − D2 ∇2 s + α4 [< u > −u] − α5 [< s > −s]
∂v
∂t
= β1 v − β2 v 2 − β3 (s − sR )v + D3 ∇2 v + D4 ∇2 s + β4 [< v > −v] + β5 [< s > −s]
∂s
∂t
= γ1 u − γ2 v − γ3 (s − sR ) − γ4 (s − sR )3 + D5 ∇2 s
(2.1.8)
1
We choose to introduce the power of the nonlinear term of the balancing term to be cubic in s,
in order to prevent divergence of the status in case that s − sR < 0, because then −x(s − sR )3 > 0,
while for example −x(s − sR )2 < 0.
CHAPTER 2. PROPOSED MODEL 25

2.1.1 Rescaling equations


Some of the constants in these equations are redundant and can be eliminated by
rescaling space, time and the dynamical variables u, v and s. Define:
q
t̃ = α1 t, x̃ = Dα1
1
x, ũ = αα21 u, ṽ = ββ12 v, s̃ = αα31 (s − sR ) (2.1.9)

With Eqs. (2.1.9) and dropping the tilde notation for convenience, we can rewrite
(2.1.8) in the following form:
∂u
∂t
= u − u2 + us + ∇2 u − δ1 ∇2 s + ρ1 [< u > −u] − ρ2 [< s > −s >]
∂v
∂t
= αv − v 2 − βvs + δ2 ∇2 v + δ3 ∇2 s + ρ3 [< v > −v] + ρ4 [< s > −s] (2.1.10)
∂s
∂t
= ²(u − γv − µs) − ξs3 + δ4 ∇2 s

The new parameters in (2.1.10), expressed in terms of the old parameters in (2.1.8)
are given by:

α = αβ11 , β = αβ33 , ² = αα13αγ12 , γ = αα2 γ3 3 , µ = αα2 γ3 3


ξ = γα4 α2 1 , δ1 = DD2 α2
1 α3
, δ2 = D 3
D1
, δ3 = αβ23D 4
D1 (2.1.11)
3
δ4 = αα13αγ12DD5
1
, ρ1 = αα4 α2 1 , ρ2 = αα5 α3 1 , ρ3 = β4βα2 1 , ρ4 = β5 α1
α3

A few more constants can be viewed as unfolding parameters of an invariant form


of the model that reflects a symmetry between the two populations. Setting α = 1,
δ1 = δ2 = δ, β = 1, δ2 = 1, γ = 1 we obtain (hereafter the “symmetric model”):

∂u
= u − u2 + su + ∇2 u − δ∇2 s + ρ1 [< u > −u] − ρ2 [< s > −s] ,
∂t
∂v
= v − v 2 − sv + ∇2 v + δ∇2 s + ρ1 [< v > −v] + ρ2 [< s > −s] ,
∂t
∂s
= ²(u − v − µs) − ξs3 + δ4 ∇2 s . (2.1.12)
∂t
The symmetry between the two populations follows from the invariance of Eqs.
(2.1.12) under the transformation u → v, v → u, s → −s.

2.2 Stationary and homogeneous solutions


The simplest solutions of Eqs. (2.1.12) are stationary uniform solutions. The popu-
lation states they represent do not describe segregation, but the number of different
CHAPTER 2. PROPOSED MODEL 26

States Symbol Solution (u, v, s)


No-Population O (0, 0, 0)
Symmetric mixed population M (1, 1, 0)
Non-Symmetric mixed population N− (1 − η, 1 + η, −η)
N+ (1 + η, 1 − η, η)
Pure population P+ (p, 0, sp )
P− (0, p, −sp )

Table 2.1: Stationary uniform solutions of the symmetric model

solutions and their stability do contain important information about possible forms of
segregation as we discuss in the next chapter. To find stationary uniform solutions we
set all time and space derivatives to zero. We consider first the symmetric model Eqs.
(2.1.12). Solving the resulting equations , we obtain the population states shown in
Table 2.1. Note that the non-local terms UN L and VN L vanish for uniform solutions.
Thus, the solutions displayed in Table 1 solve also the non-local symmetric model
(Eqs. (2.1.12) with the terms UN L and VN L included).

The parameters in Table 2.1 were calculated using Mathematica and are given by
p
η = ²(2 − µ)
1/3 2/3
p = 1 + 16 [ 62Ω ² (1 − µ) + 2 ξ Ω ]
Ω 21/3 ²(µ−1) (2.2.1)
sp = 321/3 −
ξ pΩ
Ω ≡ (27²ξ 2 + 108²3 (µ − 1)3 ξ 3 + 729²2 ξ 4 )1/3

2.3 Linear stability analysis to uniform perturba-


tions
The uniform states that can be realized in practice are those which are linearly stable.
Roughly speaking, a given state is stable if small perturbation of this state decay to
zero as time evolve. In order to study the stability of the stationary and uniform
solutions (u0 , v0 , s0 ) of the symmetric model Eqs. 2.1.12 we write:

u(t) = u0 + u1 (t), v(t) = v0 + v1 (t), s(t) = s0 + s1 (t), (2.3.1)


CHAPTER 2. PROPOSED MODEL 27

where u1 (t), v1 (t), s1 (t) are infinitesimal uniform perturbations. The stability to non-
uniform perturbations will be studied in Sec. 2.5. Insert 2.3.1 in 2.1.12 and neglecting
nonlinear terms in u1 (t), v1 (t), s1 (t) we obtain the following linear system:
 ∂u1    
∂t
1 − 2u0 + s0 0 u0 u1
 ∂v1  =  0 1 − 2v0 − s0 −v0   v1  (2.3.2)
∂t
∂s1 2
∂t
² −² −²µ − 3ξs 0 s1

In order to get stability, all the eigenvalues of the matrix in the right side of 2.3.2,
must have negative real parts. This matrix is known as the Jacobian or community
matrix [73, 49]. Figure 2.1 shows the results of this analysis for the symmetric model.
The no-population state O exists for all parameters, but is always unstable. This
reflects our assumption that the area under consideration has all the infrastructure
and facilities to serve as a residential area and therefore residents immediately occupy
vacant regions. The other stationary uniform solutions of 2.1.12 and their stability are
described by the bifurcation diagrams shown in Figure 2.1. The bifurcation parameter
is µ. The significance of this parameter will be discussed in the next chapter. At all
values of µ there exists a symmetric mixed state M: (u, v, s) = (1, 1, 0), representing
a uniform and equal distribution of two population u = 1 and v = 1 at any location.
This state is stable for relatively large µ. As µ is decreased past the critical value,
µ = 2, the mixed state M loses stability in a pitchfork bifurcation to a pair of
asymmetric mixed population states N± : (u, v, s) = (1 ± η, 1 ∓ η, ±η) where η =
p
²(2 − µ). The mixed population states represents uniform but unequal distributions
of the two populations. In N+ there is a majority to u-residents whereas in N− the
majority is of v-residents. As µ is further decreased the N± states become more and
more asymmetric until a second bifurcation point, µ = 1, is reached where they lose
stability in a pair of transcritical bifurcations to a pair of pure population states: P+
: (u, v, s) = (p, 0, sp ) and P− : (u, v, s) = (0, p, −sp ). 2
The actual urban reality is asymmetric with respect to the two populations. People
of lower socio-economic status normally live in denser residential areas, their natural
growth rate is often higher, and their mobility is lower [72, 17]. The symmetric
model 2.1.12 does not strictly apply to this situation. Its significance however lies
in unraveling the main bifurcations of the system. The behaviors of the asymmetric
2
For ² = 1 the transcritical bifurcation occurs at µ = 2 − ξ.
CHAPTER 2. PROPOSED MODEL 28

2 P+ (a)
N+
u1
N− M
P−
0

0 1
µ 2 3

2 P− (b)
N−
v1
M
P+ N+
0

0 1 µ 2 3

1 P+ (c)
N+
s 0
M
P− N−
−1

0 1 µ 2 3

Figure 2.1: Bifurcation diagrams for stationary uniform solutions of Eqs. (2.1.12)
showing (a) the density u, (b) the density v, and (c) the status s. Solid (dashed) lines
stand for stable (unstable) solutions. The symmetric state M exists for all values
of µ and is stable for µ > 2. The Pure states, P± , exist for all values of µ and are
stable for µ < 1. The non-symmetric states, N± , exist for µ < 2 and are stable for
1 < µ < 2. They appear in a pitchfork bifurcation of the symmetric state at µ = 2
and lose stability to the pure states in a pair of transcritical bifurcation at µ = 1.
Parameters: ξ = 1, ² = 1, ρ1 = ρ2 = 0.
CHAPTER 2. PROPOSED MODEL 29

2.5
u
2 P+
N+
1.5

1 N−

0.5

P−
0

−0.5
0 0.5 1 1.5
µ 2 2.5 3 3.5

Figure 2.2: Bifurcation diagram for the asymmetric model (α = 1.1). Solid (dashed)
lines stand for stable (unstable) solutions. The M state is no more a solution of the
system and there is a coexistence of states N+ and P− for 0.699 < µ < 1.1. The N−
state is stable for µ > α.

system 2.1.10 near the bifurcation point are captured by the universal unfolding of
these bifurcations [74]. Figure 2.2 shows how the the bifurcation diagram is modified
when we consider the asymmetric equations with α = 1.1. In this case the N±
states are no longer symmetric. For ² = 1 the N− state is stable for µ > 2 − α.
The asymmetric model allows for additional state coexistence ranges, for instance
coexistence the N+ state with the P− state as Fig. 2.2 shows.

2.4 Turing instability analysis


In the previous section we considered the stability of the M state to uniform pertur-
bations. Stability of the M state requires the decay of nonuniform perturbations as
well. In this section we show that in addition to the pitchfork bifurcation to uniform
CHAPTER 2. PROPOSED MODEL 30

states, the M state may also go through a Turing type instability (see below). We
also study in this section the possible Turing instabilities of the non-symmetric mixed
population states N± .

2.4.1 Turing instability in population models


A Turing bifurcation involves the destabilization of a homogeneous steady state to
form an inhomogeneous state whose wavelength depends on the kinetic parameters
and diffusion coefficients of the system [75].
The importance of Turing instability for biological pattern formation has been a
lively topic of debate [7] since the publication of Turing’s seminal paper The Chemi-
cal basis of morphogenesis in 1952 [76]. The key insight of Turing was that diffusion,
which is usually thought of as stabilizing mechanism, can actually act as a destabiliz-
ing mechanism of the spatially uniform state. This is the case, for example, in a class
of two-component activator-inhibitor systems, provided the inhibitor diffuses more
rapidly than the activator. For the mathematical conditions for Turing instability
for activator-inhibitor systems see [77]. Experimental evidence of Turing patterns
was discovered only recently, and a number of experiments on the Turing instability
have now been carried out on the chloride-iodide-malonic acid (CIMA) reaction in
a gel reactor [78, 79]. These experiments have demonstrated the existence and the
stability of a variety of regular patterns, including stripes, hexagons and rhombs.
Many theoretical biologists have proposed various models of biological pattern
formation based on Turing idea. For example Murray [49] showed that such reaction
diffusion models are able to reproduce animal coat patterns like the Tiger’s spots or
Zebra’s stripes. In the context of ecology heterogeneity has traditionally been inter-
preted as a consequence of environmental variability, but it may also results from
intrinsic mechanisms, such as the Turing instability [80]. This instability has sub-
sequently been explored for two-species interactions in ecological context by Segel &
and Jackson [81], Levin & Segel [82], and White & Gilligan [83] who proposed a
three species system involving hosts, parasites and hyperparasites in homogeneous
envirionment. The possibility of diffusive instabilities can be shown for a wide variety
of ecological systems and as P. Kareiva wrote [84] ”Perhaps much of the pattern
CHAPTER 2. PROPOSED MODEL 31

attributed to underlying environmental mosaics could be more parsimoniously ex-


plained by diffusive instabilities”. 3
Turing patterns are well studied for two species systems. The following charac-
teristic properties of these patterns have been identified [85]:
• The patterns appear as a result of reaction and diffusion only, no convective
transport is involved.

• They grow form uniform steady states, which are stable to homogeneous per-
turbation but unstable to inhomogeneous perturbations of a well defined wave
lengths.

• They are spatially inhomogeneous, steady in time and have intrinsic wave length
that it is independent on the dimension and geometry of the system, provided,
the system extension is larger than at least a few wavelengths.

2.4.2 Turing instability of the M state


We explore now the possible destabilization of the M state, (u, v, s) = (1, 1, 0), by
a Turing instability in the range µ > 2. To analyze the stability of the M state to
nonhomogenous perturbations we subsistute in 2.1.12 the following perturbed form
of this state:       
u 1 ũ
 v  =  1  +  ṽ  eikx+σt + c.c. . (2.4.1)
s 0 s̃

Linearization around the M state leads to the characteristic determinant:


¯ ¯
¯ −(1 + k 2 ) − σ 0 1 + δk 2 ¯
¯ ¯
¯ 0 −(1 + k ) − σ −(1 + δk ) ¯¯ = 0
2 2
(2.4.2)
¯
¯ ε −² −(²µ + δ4 )k 2 ¯

which have the following solutions for σ:


σ1 = −1 ¡− k 2 √ ¢
σ 2,3 = 21 −A ∓ A2 − 4B
(2.4.3)
A ≡ 1 + k 2 + δ4 k 2 + ²µ
B ≡ −2² + ²µ − k 2 (2δ² − δ4 − ²µ) + δ4 k 4
3
In Appendix B we show how the classical Lotka-Volterra two species competition model can be
modified in order to support the existence of Turing instability.
CHAPTER 2. PROPOSED MODEL 32

0.01

σ3 µ<µc
0
µ=µc
−0.01
µ>µc

−0.02

−0.03
0 0.2 0.4 0.6 0.8 1 1.2 1.4
k
Figure 2.3: Growth curves showing a finite wavenumber instability of the uniform
mixed state. Parameters:² = 1, ξ = 1, δ = 1.12, δ4 = 0.1.

The instability occurs for σ3 > 0 or when


µ < µc √
2 2²δ4 (δ−1)−δ (2.4.4)
µc ≡ 2δ − ²

Figure 2.3 shows the growth curves near the onset of the instability for different values
of the parameter µ. Below the instability (µ > µc ) the growth rate is negative for all
wave-numbers implying the decay of all perturbation and the stability of the uniform
p
state. At the instability point (µ = µc ) one mode, k = kc ((kc2 = 2²δ4 (δ − 1))
becomes marginal while all other still decay. Beyond the instability (µ < µc ) a band
of modes centered about kc grow and evolve into a stationary periodic pattern. The
patterns that evolve from the Turing instability of the M state and its relation to
residential segregation phenomena will be discuss in the next two chapters.

2.4.3 Turing instability analysis of the non symmetric mixed


states
The N± states are stable to uniform perturbations in the range 1 < µ < 2, but may
undergo Turing instabilities. To figure that out we perform stability analysis of these
CHAPTER 2. PROPOSED MODEL 33

states to nonuniform perturbations, similar to the stability analysis of the M state


presented in the previous section. A perturbed form similar to 2.4.1 is written for
the N± states: (u, v, s) = (1 ± η, 1 ∓ η, ±η), and is used in the symmetric model
2.1.12. To simplify the calculation we will restrict the analysis to ² = 1 and ξ = 1.
Linearizing around the N± we find the characteristic determinant
¯ ¯
¯ f1 − k 2 − σ 0 f + δk 2 ¯
¯ 3 ¯
¯ 0 2
g2 − k − σ g3 − δk 2 ¯=0 (2.4.5)
¯ ¯
¯ 1 −1 2
h3 − δ4 k − σ ¯

where f1 = 1 − 2u − s, f3 = u, g2 = 1 − 2v − s, g3 = −v, h3 = −µ − 3s2 and


√ √ √
u = 1 + 2 − µ, v = 1 − 2 − µ, s = 2 − µ (because of symmetry the same
results are obtained The characteristic equation for the eigenvalues σ can be written
in the schematic form as

a0 σ 3 + a1 σ 2 + a2 σ + a3 = 0 (2.4.6)

where the ai ’s defined as


a0 = 1
a1 = 8 + (2 + δ4 )k 2 − 2µ
a2 = (1 + 2δ4 )k 4 + (14 − 2δ − 4µ + 2δ4 )k 2 + 9 − 3µ (2.4.7)
a3 = δ4 k 6 + (6 − 2δ − 2µ + 2δ4 )k 4 + (10 − 2δ − 4µ − δ4 + µδ4 )k 2
+(−4 + 6µ − 2µ2 )

The Ruth-Hurowitz conditions [77] for 2.4.7 imply that in order to have Re(σ) < 0
we need that a1 > 0, a3 > 0, a1 a2 − a3 > 0. The last term a3 is the interesting one
[86], since by changing sign from positive to negative it may destabilize the system.
We need to solve the following equation

a3 ≡ p(k 2 ) = b0 k 6 − b1 k 4 + b2 k 2 + b3 = 0 (2.4.8)

where the bi ’s are defined as


b0 = δ4
b1 = −6 + 2δ + 2µ − 2δ4
(2.4.9)
b2 = 10 − 2δ − 4µ − δ4 + µδ4
b3 = −4 + 6µ − 2µ2
CHAPTER 2. PROPOSED MODEL 34

A feasible minimum point of p(k 2 ) has to occur in the positive quadrant and this point
is also the global minimum in this quadrant. A necessary and sufficient condition for
the Turing instability is that this minimum will be equal to zero as shown in Fig.
2.4 . ( p(k 2 ) intersect the vertical axis at b3 = 2(µ − 2)(µ − 1) which is positive
for 1 < µ < 2). We will find a “weaker” condition by finding the extremum points
of p(k 2 ). These points can be found by solving the equation dp(k 2 )/dk 2 = 0. The
solution gives the expression for kc2 :
p
2 b1 ± b21 − 3b0 b2
kc = (2.4.10)
3b0

In order that kc2 will be positive we have to add the conditions b21 − 3b0 b2 > 0 and
b1 > 0 or b2 < 0. Substitute 2.4.9 in 2.4.10 we get these conditions written with the
model’s parameters: µ > µ2 or µ < µ1 where
µ p q ¶
1 2 2
µ1,2 = 24 − 4δ4 + 3δ4 − δ ∓ 3δ4 16 − 8δ4 + δ4 − 16δ + 32(δ − 1) (2.4.11)
8

and µ > 3 + δ4 − δ or µ > (10 − 2δ − δ4 )/(4 − δ4 ), so we have four cases according to


the different values of µ1 and µ2 :

1. µ > µ2 and µ > 3 + δ4 − δ

2. µ > µ2 and µ > (10 − 2δ − δ4 )/(4 − δ4 )

3. µ < µ1 and µ > 3 + δ4 − δ

4. µ < µ1 and µ > (10 − 2δ − δ4 )/(4 − δ4 )

For δ = 1.12 and δ4 = 0.1 the above conditions are:

1. µ > 1.96313 and µ > 1.98

2. µ > 1.96313 and µ > 1.9641

3. µ < 1.70437 and µ > 1.98

4. µ < 1.70437 and µ > 1.98641


CHAPTER 2. PROPOSED MODEL 35

0.1
p(K)

0.05 µ=1.98
µ=1.983

µ=1.9845
0

µ=1.985

−0.05
−1.5 −1 −0.5 0 0.5 1 1.5
K
Figure 2.4: The function p(k 2 ) for different values of µ. Increases the value of µ lowers
the minimum and for µ = 1.9845 the minimum is equal to zero and this is the critical
value of µ (k 2 = K).

Only the first two cases are possible and because that these conditions are necessary
and not sufficient we conclude that 1.98 < µnc < 2 is a necessary condition δ = 1.12
and δ4 = 0.1. Exact numeric calculation shows that the instability occurs at µ = 1.985
as shown in Fig. 2.4.
We can find the relation between δ and δ4 according the condition b1 > 0 so that
the instability will be restricted to 1 < µ < 2. We get that 1 + δ4 < δ < 2 + δ4 . Figure
2.5 shows the bifurcation diagram of u in a parameter range that includes finite wave
number instabilities of the mixed states we denote the onset of the M state instability
by µc and the onset of the N± states by µnc .

So far we considered Turing instabilities in the symmetric model. The mixed


population states may become Turing unstable in the asymmetric model too, as Fig.
2.6a shows. Unlike the symmetric model a parameter range can be found where one
CHAPTER 2. PROPOSED MODEL 36

2.5

u
2 P+

1.5 N+

1
M
0.5 N−
P−
0

−0.5
0 0.5 1 1.5 2
µ nc µ c
2.5 3
µ 3.5

Figure 2.5: Bifurcation diagram for stationary uniform solution showing in a pa-
rameter range that includes Turing wave instability of the mixed population states.
Solid (dashed) lines denote stable (unstable) states. Dotted lines denote solutions
that lose their stability to finite wave instability. The M state is Turing unstable
for 2 < µ < µc . The N± states are Turing unstable for µnc < µ < 2. Parameters:
δ = 0.3, δ4 = 0.1. For this choice of parameters µc = 2.2101, µnc = 1.8916.
CHAPTER 2. PROPOSED MODEL 37

of the mixed population states becomes Turing unstable, while the other one remains
stable. A case like this is shown in Fig.2.6b. The pure states may also go through
Turing instabilities but the patterns that developed may involve negative values of
the population densities and therefore are not realistic in the urban context.

2.5 Reducing the the model to Lotka-Volterra com-


petition model
The lotka-Volterra competition model is one of the classic models in ecology. The
model describes an interspecific competition between two population in homogeneous
environment. The model able to generate a range of possible outcomes: the pre-
dictable exclusion of one species by another, exclusion dependent on initial densities,
and stable coexistence. In the context of residential segregation Zhang [51] used
this model to describe the interaction between two population groups. Although the
lotka-Volterra competition model is too simple to describe the complex phenomena
of competition in ecology and it serve as a starting point for more elaborated models.
From the mathematical point of view, the lotka-Volterra competition model has
been investigated thoroughly. One important result is, that this system has a Li-
apunuv function which is an energy-like function that decreases along trajectories
[87]. Liapunov function is a continuously differentiable, real valued V (x) with the
following properties:

1. V (x) is positive definite i.e. it equal zero only at the fixed points.

2. All trajectories flow “downhill” toward the fixed point (dV /dt < 0 for all x 6= x∗
where x∗ is the fixed point.)

System that have Liapunov function are belonged to a more general class called
gradient systems. In a gradient system no closed orbit in the phase space is allowed
which exclude time oscillations or the existence of Hopf bifurcation. Unfortunately,
there is no systematic way to construct Liapunov functions and the task is even
more difficult for systems with three variables. As the following subsection shows the
proposed model can be reduced to the Lotka-Volterra model which has a Liapunov
function. So only with this restrictions the urban model is a gradient system. 4
4
Oscillations in populations exist in predator-prey models in population ecology, and the basic
CHAPTER 2. PROPOSED MODEL 38

Figure 2.6: Growth curves for the asymmetric model. (a) Both N± states are Turing
unstable for δ = 1.4. (b) Only the N+ state is Turing unstable for a lower value of δ:
δ = 1.2. Other parameters values: α = 1.1, ² = 1, ξ = 1, δ4 = 0.1.
CHAPTER 2. PROPOSED MODEL 39

2.5.1 Liapunov function to the reduced model


The model equations 2.1.12 can be reduced to a system of two species Lotka-Volterra
competition model by applying the following assumption. First we neglect the non-
local migration terms so ρi = 0 for i = 1, 2, 3, 4. For ξ = 0, δ4 = 0 and for ² À 1,
then we eliminate s from the equation for the status in 2.1.12 and get s = (u − v)/µ.
Substitute this expression for s in the equations for u and v gives:
ut = u − (1 − µ−1 )u2 − µ−1 uv + (1 − λ)∇2 u + λ∇2 v
(2.5.1)
vt = v − (1 − µ−1 )v 2 − µ−1 uv + (1 − λ)∇2 v + λ∇2 u

where ut and vt stand for partial time derivative and λ = δ/µ. Eqs. 2.5.1 describe a
competition interaction between two populations and for δ = 0 it similar to Zhang’s
model (1.2.13). The migrational cross diffusion terms describe local repulsion interac-
tion between the two population. The Lotka-Volterra competition model is a central
model in ecology to describe the invasion of competing species [48, 77].
System 2.5.1 has four uniform homogeneous solutions, which their stability are
shown in Figure 2.7:
1. (u, v) = (0, 0) is always unstable.

2. (u, v) = (µ/(µ − 1)) and (u, v) = (0, µ/(µ − 1)) are stable for 1 < mu < 2.

3. (u, v) = (1, 1) hereafter will be termed the emphnontrivial solution [88] is stable
for µ > 2.
Murray [89] showed that the effect of diffusion on the Lotka-Volterra equations
with equal diffusion coefficients is to damp out all spatial variations in a system with
zero flux boundary conditions. Jorné [53] showed that for Eqs. 2.5.1 in the limit
δ = 0 takes as a Liapunuv function the spatial integral V which is positive definite,
Z ½ µ ¶ µ ¶¾
u(x, t) v(x, t)
V = u(x, t) + v(x, t) − u0 ln − v0 ln dx > 0 (2.5.2)
u0 v0
where u0 and v0 is the nontrivial solution 5 and whose time derivative is
Z ½ ¾
dV u0 v0
= ut + vt − ut − vt dx (2.5.3)
dt u(x, t) v(x, t)
interaction between u and v in our model is competition as we can conclude from the fact that their
spatial distributions are out of phase.
5
This definition of Liapunuv function is not valid for the bistbale case as it diverges.
CHAPTER 2. PROPOSED MODEL 40

2.5

1.5
u
1

0.5

−0.5
1.5 2 2.5 3
µ 3.5 4 4.5 5 5.5

Figure 2.7: Bifurcation diagram of the Lotka-Volterra competition model eqs. 2.5.1.
The symmetric state (u, v) = (1, 1) is stable for µ > 2. For 1 < µ < 2 the system is
bistable and only pure population states exists.

Using the fact that (u0 , v0 ) = (1, 1), we obtain


dV
R
dt R
= − {(u − 1) [(1 − µ−1 )(u − 1) + µ−1 (v − 1)]}dx
− R £{(v − 1) [(1 − µ−1 )(v − 1) + −1
¤ µ (u − 1)]}dx (2.5.4)
1 2 1 2
+ (1 − u )∇ u + (1 − v )∇ v dx

The first two integrals can be written as


Z Z
© ª
f (u, v)dx = −(u − 1)2 − (v − 1)2 + µ−1 (u − v)2 dx (2.5.5)

In order to prove that 2.5.5 is negative for all initial conditions with u > 0 and v > 0,
we will look for the maximum point of f (u, v). f (u, v) is a function of two variables
and has a maximum at (u0 , v0 ) = (1, 1) for all µ > 2 and the value of this maximum
is 0. We can prove the existence of this maximum by looking at the different partial
derivatives of f (u, v):
∂2f
∂f
∂u
= −2(u − 1) + 2(u−v)
µ ∂u2
= −2 + µ2
∂f 2(u−v) ∂2f
∂v
= −2(v − 1) + µ ∂u2
= −2 + µ2 (2.5.6)
∂2f 2
∂u∂v
= − µ
CHAPTER 2. PROPOSED MODEL 41

The conditions for a maximum point of two variable function f (u, v) are:

1. fu = fv = 0 at (1, 1).

2. fuu < 0 and fvv < 0

3. fuu fvv > (fuv )2

The two first conditions are clearly valid and third condition become 4/µ < 2 which
is true for µ > 2. Thus, we conclude that 2.5.5 is negative and equal to zero only at
the fixed point.
Define the third integral in 2.5.4 as I. I can be evaluated by using the divergence
theorem and the no flux boundary condition n · ∇u = n · ∇v = 0 where n is the outer
normal vector and the identity:

1 2 1 (∇u)2
∇ u = ∇ · ( ∇u) + (2.5.7)
u u u2

we obtain R£ ¤
IR = (1 − u1 )∇2 u R+ (1 − vv0 )∇2 v dx =
(1 − uu0 )n · ∇u + (1 − vv0 )n · ∇v+ (2.5.8)
∂ R ∂ R
2 2
− (1 − u0 ) (∇u) u2
dx − (1 − v0 ) (∇v)
v2
dx

Using the no-flux boundary conditions makes the surface integrals vanish yielding
I = 0 which ensures that dV /dt is negative and is equal to zero only at (u0 , v0 ) = (1, 1).
This result means that Lotka-Volterra competition equations are a gradient system
so no Hopf bifurcation can take place. Hastings [88] proved global stability of the
nontrivial state of a more general system in m dimensional space and n interacting
species, which Zhang’s model is a private case of it:
à n
! m µ ¶
∂Ni X X ∂ ∂Ni
= Ni ri + aij Nj + Di i = 1, n (2.5.9)
∂t j=1 k=1
∂x k ∂x k

Hastings concluded that as dispersal of each species is from areas of higher population
density to areas of lower population density no spatial depended solution can exists.
For the symmetric model eqs 2.1.12 it remains an open question if the system is
CHAPTER 2. PROPOSED MODEL 42

gradient. Extensive numerical search for the Hopf instabilityy has been carried for
different values of parameter but gave negative result.

2.5.2 Turing instability analysis of Eqs. 2.5.1


We can analyze directly the stability of the mixed state (u, v) = (1, 1) of eqs. 2.5.1
to non-homogeneous perturbations by introducing small perturbations,
µ ¶ µ ¶ µµ ¶ ¶
u 1 ũ ikx
= + e + c.c. (2.5.10)
v 1 ṽ

where ũ ¿ 1 and ṽ ¿ 1 and c.c stands for complex conjugation. Substitute 2.5.10
in 2.5.1 and after linearization around the mixed state, we get the characteristic
determinant
¯ −1 ¯
¯ µ − 1 − (1 − λ)k 2 − σ −(µ−1 + λk 2 ) ¯
¯ ¯=0 (2.5.11)
¯ −(µ−1 + λk 2 ) µ−1 − 1 − (1 − λ)k 2 − σ ¯

obtaining the characteristic equation


£ ¤2
µ−1 − 1 − (1 − λ)k 2 − σ = (µ−1 + λk 2 )2 (2.5.12)

which gives the solutions for σ

σ1 = 2µ−1 − 1 + (2λ − 1)k 2


(2.5.13)
σ2 = −1 − k 2

In order to have a finite wavenumber instability we need that at least that one of the
two eigenvalues will be positive for finite range of wavenumber k. σ2 is negative for
all values of k and σ1 is negative for µ > 2 at k = 0 but it it is positive for larger
values of k, which precludes the existing of a finite wavenumber instability. Appendix
B shows what kind of diffusing terms we have to add to 2.5.1 to get a finite wave
instability and discuss it from an ecological point of view.
Chapter 3

Urban Dynamics Aspects of the


Model

In the approach to urban and regional modeling derived from self-organizing


system in physics and chemistry, key events are the occurrence of spatial
instabilities. During these, unpredictable events may play a vital role,
nudging the system on to one or another branch of organization.
Peter Allen, [34]

In this chapter we apply the mathematical model to two types urban phenomena,
residential segregation and neighborhood change processes. We present mechanisms
of weak and strong segregation and modes of neighborhood changes involving invasion
and tipping processes.

3.1 Segregation mechanisms


The phenomenon of segregation refers to situations where population groups are
nonuniformly distributed in space. Segregation exists when some areas show an over-
representation and other areas an under-representation of a given population group.
Segregation in our model is achieved by decreasing the parameter µ. The smaller
µ the stronger the response of the status s to population differences u − v and the
stronger the drive for segregation. An inspection of the bifurcation diagrams of Fig.
2.1 suggests three parameter ranges of different behaviors: (i) µ > 2, no segregation,1
1
later in this chapter that for this range segregation can occur due to finite wavenumber instability.

43
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 44

(ii) 1 < µ < 2, variable segregation, and (iii) µ < 1 strong segregation. In case (i) only
one stable uniform state, M, exists, implying uniform distribution of mixed popu-
lation. In case (ii) two stable asymmetric mixed population states coexist, N− and
N+ . The system is bistable and patterns involving domains of one mixed population
state in a neighborhood of the other are possible as Fig. 3.1a shows. Such patterns
describe segregation; regions over-represented with v-residents are surrounded by re-
gions under-represented with v-residents. An opposite pattern holds for u-residents.
The segregation is weak close to µ = 2 because the population distribution of the
two asymmetric states N− and N+ , are close to one another. The segregation be-
comes stronger and stronger as µ is decreased towards µ = 1. In case (iii), the mixed
population states have lost stability to the pair of pure states, P− and P+ . In this
parameter range enclaves of pure v-population in neighborhoods of pure u-population
and vice versa may appear as demonstrated in Fig. 3.1b. In the asymmetric model
enclaves of pure population in a neighborhood of mixed population may also exist.
The parameter µ measures the the status gap that develops between distinct
population states, e.g. N± . According to 2.1, the status gap, ∆s, between N+
p
and N− is 2η = ²(2 − µ). The smaller µ the larger the socioeconomic inequality
between the two populations. This interpretation is consistent with Darroch and
Marston’s finding [90] that group differences in social status are clearly related to
group differences in residential segregation.
The quantity µp − µ can be associated with the concept of cognitive dissonance
[2]. To see this we consider the more generic non-symmetric model Eqs. 2.1.10
and the bifurcation diagram shown in Figure 2.2. Negative values of this quantity
amount to a single stable mixed state M. In this parameter range the proportions
of the two populations are such that social conflicts are weak and no segregation
occurs. However, as µp − µ increases the representation of v-residents increases at
the expense of the u residents and cognitive dissonance develops: the high status
u-residents find it less satisfying to live in neighborhoods where v-residents form a
majority. A critical point is then reached µp − µ = 0 (µp = 2 in the symmetric model)
where u individuals have to decide either to follow their intentions and wants and
migrate to neighborhoods that fit their high socioeconomic status (thus giving rise
to segregation), or to change their intentions (giving more weight to values such as
social integration) and stay at their present neighborhoods. This dilemma and the
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 45

2
1.5
u
1
v
0.5
s
0
−0.5
(a)
−1
0 20 40 60 80 100
x

2
u
1.5
1
s
0.5
0
v
−0.5
−1
(b)
0 20 40 60 80 100
x

Figure 3.1: Segregation forms in the symmetric model 2.1.12. (a) Variable segregation
(µ = 1.8): A domain of N− state in a N+ neighborhood, representing a region with
over representation of v-residents and lower status. (b) Strong segregation (µ = 0.9):
A domain of P− in a P+ neighborhood, representing an enclave of pure v-residents
in a neighborhood of pure u-population. Values of other parameters: ξ = 1, ² = 1,
δ = 0.6, δ4 = 0.1.
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 46

1
0.9
0.8
0.7 dissimilarity
0.6 index
0.5
0.4
0.3 exposure
0.2 index
0.1
0
1 1.2 1.4 µ 1.6 1.8 2

Figure 3.2: Two index of segregation for the symmetric model as a function of the
parameter µ. Decreases the value of µ increases the segregation as depicted by the
dissimilarity and the exposure indexes.

consequence actions taken by individuals are reflected in the model by the appearance
of the stable N+ state along with the stable N− state at µp − µ = 0. For now islands
representing neighborhoods with majority of u-resident (the N+ state), can be form
in areas occupied predominantly by v-residents (N− state). The islands contain
u-residents that migrated from lower status surrounding areas whereas these areas
contain u-residents that seek integration or those that do not have the economic
ability to move to the higher socioeconomic status areas.
Figure 3.2 shows the two measures of segregation: index of dissimilarity and
exposure index (see chapter 1) as a function of the parameter µ for two adjacency
neighborhoods, one in the N+ state and the second in N− state. The exposure index
x Py * equal to 0.5 at µ = 2 which means no segregation and x Py *=0 at µ = 1 which
means strong segregation. The index of dissimilarity is 0 at µ = 2 and equal to 1 at
µ = 1 which means strong segregation. x Py * is a linear function of µ where as the
index of dissimilarity is a nonlinear function of µ. We conclude that this two measures
of segregation are consistent with our statement that as µ decreases the segregation
becomes stronger.
In the asymmetric model the asymmetry of the N+ and N− states is already
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 47

pronounced. Thus, in systems with significant asymmetry between the two popu-
lations, segregation phenomena involving domains of asymmetric states can not be
weak. Variable segregation has been reported for various ethnic groups in European
and American cities. Black African in London [91] and Turks and Moroccans in
Amsterdam [10] are two examples. Strong segregation has been reported mostly in
the context of black-white segregation in American cities [59, 92, 19].

3.2 Neighborhood change


3.2.1 Transition zones
An interface in the context of dynamical system is a localized structure that forms
a transitions zone between two different stable uniform states of the system [68].
The interface represents a stable balance between two competing processes, i.e. the
tendency of the system to converge locally to one of the two states because, their
stability, and diffusion. The effect of diffusion is to smooth out the interface and give
it characteristic width; without diffusion an infinitely sharp interface would form. In
the proposed model, this smoothing effect is archived by the self-diffusion terms. The
cross-diffusion terms are governed by the value of the parameter δ (in the symmetric
model Eqs. 2.1.12); this parameter represents local repulsion interaction between the
two population. Table 3.1 summarizes the urban dynamic interpretation of the main
parameters in the model (without the nonlocal migration parameters).
Increasing the value of δ will narrowing the width of the interface and vice versa.
Another factor that influences the width of the interface is the value of the parameter
µ, which controls the status gap between the two populations. As we increase the
value of µ, the segregation becomes stronger and interface becomes narrower. For the
symmetric model the interface is part of the stationary front solution that connects
the two state like N− and N+ or P− and P+ .
Figure 3.3 shows these fronts for different value of µ and and δ. These fronts appear
to be similar in shape to those found in various empirical studies (see for example
[47] and [48] for invasion models in ecology).
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 48

δ=0 δ=0.3 δ=0.6 δ=0.9

µ=0.6

µ=1.8

µ=1.95

Figure 3.3: Stationary fronts for the symmetric model Eqs. 2.1.12 for different values
of µ and δ. Decrease in the value of δ makes the interface shallower as a a consequence
of the decrease in the repulsion between the two populations in the transition zone
locality. Increase in the value of µ has the same effect, as the status gap between the
two populations gets smaller. Values of other parameters: α = 1, ξ = 1, δ4 = 0.1

Parameters Urban interpretation of the parameters


µ status gap between the two populations;
the smaller µ the larger the gap
δ local repulsive force between the two populations;
based on socioeconomic considerations. δ stands for δ1 = δ2
δ4 diffusion of the status without population change;
its value is relative small.
² strength of indirect coupling (via the status variable)
between the two populations.
α the growth rate of the v-populations.
α = 1 represents the symmetric case
δ1 the tendency of u-residents to move to areas with higher status value.
δ2 the tendency of v-residents to move to areas with lower status value

Table 3.1: Urban interpretation of the parameters


CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 49

2.5
Transition Zone
2
u v
1.5 Lower status Higher status
u−residents v−residents
1
s
0.5

−0.5
only u−residents only v−residents
−1
Neighborhood Change
−1.5
0 5 10 15 x 20 25 30

Figure 3.4: The transition zone between P− and P+ . If the rate growth of the v-
residents is higher than that of the u-residents the interface will start to move leftward
and the v-population will invade the u-population neighborhood. We can interpret
the status values in the transition zone and the mixed population as that the higher
status (compared to the status of the P− ) v residents are located near the interface
and the same for the lower status (compared to the status of the P− ) u-residents.
The transition zone has a finite wide because of the diffusion of the two populations.
Parameters: ² = 1, ξ = 1, µ = 0.7, δ = 0.7, δ4 = 0.01

From the geographic standpoint, the transition zone is the area with a mixed
(u, v) population. The transition zone remains stable if no substantial changes in
its population makeup occur. Figure 3.4 illustrates the transition zone between P−
and P+ . If the symmetry between the two groups is broken (for example the growth
rate of the v-residents is higher than that of the u-residents due to which the former
population expands more rapidly), the transition zone starts to move into the P+ .
Neighborhood change by interface motion may thus stem from competition for hosing
in the transition zone [94, 93, 17]. This competition, and thus the interface between
the two populations, are affected by a number of factors including the growth rate of
the two population, housing capacities, and non-local migration, etc. (see [70]).
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 50

Transition zones can be stationary or propagating depending on parameter values.


A stationary transition zone is normally conceived as being in a stable equilibrium,
while a propagating transition zone is regarded as being unstable due to unbalanced
processes which lead one neighborhood to grow at the expense of the other [93].
This view of stability differs from the stability concept in dynamical systems theory.
Roughly speaking, a state of a dynamical system, stationary or time depended, is
stable if small perturbations of the state decay to zero in the course of time. The
model equations have interface solutions propagating at constant speeds which are
stable in this sense. They describe propagating transition zones which are stable to
various perturbations, e.g. temporal initiatives to slow down their motion.

3.2.2 Tipping
The sudden reversal of a population distribution (e.g. natives rich to immigrant
rich) once a critical population threshold is exceeded [22, 60] has been termed the
tipping point phenomenon. Supporting evidence has been provided in several studies
[45, 95]. Tipping point models focus on temporal changes of populations in a given
neighborhood with the implicit assumption of uniform population distributions. To
account for the tipping point phenomenon in the model we restrict the analysis to
uniform solutions of Eqs. 2.1.10 (dropping space derivatives). Calculation of the
eigenvalues of the unstable state M in a wide parameter range of bistability always
found two eigenvalues with negative real parts. This implies a two dimensional stable
manifold of the M state which divides the three-dimensional phase space (spanned
by u, v and s) into two distinct parts or basin of attraction. The initial condition of
the system determined the long time behavior of the system and lead to threshold
phenomenon. If the initial state is located in the same side of the stable manifold as
the the N+ , the system will converge in time to the N+ . If on the other hand the
initial state is located in the phase space in the other side of the stable manifold, the
time evolution will take the system to the N− state. The stable manifold is also called
basin boundary [87] and the two trajectories that comprise the stable manifold are
traditionally called separtics as they partition the phase space into regions of different
long-term behavior.
Mathematical investigating of phase-space trajectories near the saddle point M
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 51

suggests four kinds of transient behaviors. Figure 3.5 shows such trajectories in the
u − v plane near the saddle point M. As the phase portrait shows, a saddle point
combines a stabilizing action in one direction and a destabilizing action along another
direction [28]. The dynamics far from the fixed point is fast and slow near the fixed
point as the trajectories in the phase space are almost linear.
The transients are affected by two main factors, (i) the direction of the stable
manifold at the transient point in phase space, and (ii) the proximity to the stable
manifold. Consider first the case where tipping fails, that is, at long times the system
converges to the N+ state. Starting at any point along the A trajectory in Fig. 3.5,
or at any point along the D trajectory the system will eventually evolve to the same
state, N+ . The two trajectories, however, represent different transients, as Fig. 3.6
shows, because they correspond to different directions of the stable manifold. In A
the v-population first sharply decreases while in D it first sharply increases (stage
II). Following this fast initial stage where the v-population in both cases decreases
(stage III). The difference in the time scale follows from the proximity to the stable
manifold. The closer the initial point to the stable manifold, the closer the system
approaches the unstable M state, and the sharper the difference in time scales. The
next stage (stage IV) is faster as the system departs from the unstable M, but as
it approaches the asymptotic state, N+ , the time evolution is slow again (stage V).
Stage I corresponds in all cases to the perturbation (e.g. invasion) that posits the
system in its initial state.
Consider now the case where tipping succeeds, that is, at long time the system
converges to the N− state. Starting at any point along the B trajectory, or at any
point along the C trajectory the system will eventually evolve toward the same state,
N− . The different initial states (pertaining to different directions of the stable man-
ifold of M) lead to different responses in stages II: in trajectory C the v-population
continues the perturbation trend and keep increasing, while in trajectory B the v-
population first decreases. The subsequent stages are similar in both cases: slow
(stage III), fast (stage IV) and slow (stage V) increase of the v-population toward its
asymptotic value at the N− state.
To farther verify the stages in the different trajectories we can plot the time deriva-
tive of v which describes the growth rate of v population. These graphs are shown
in Figure 3.7. We can identify the maximum (minimum) with the passage from the
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 52

0.9
N− B
v
0.86
A

0.82

M
0.78 C

D
0.74

N+
0.7
1.2 1.24 1.28
u 1.32 1.36 1.4

Figure 3.5: Trajectories near the saddle point (u, v, s) = (1.288, 0.812, 0.288) in the
(u−v) phase plane. The figure shows the projection of the three dimension trajectories
on the plane. Trajectories B and C describe an invasion that shifts the system to N−
state where trajectories A and D describe a weaker invasion that do not have any
effect on long time and the system remain in N+ state. Parameters: µ = 1.57,
α = 1.1, ² = 1, ξ = 1.

rapid growth (decline) stage to the convergence stage. Deskins [92] found the same
maximum in his research of the morphology of Detroit black ghetto population and he
related it to a rapid growth of the black population with relatively little expansion of
the ghetto taking place. In A and D we can recognize a local minimum which is not so
prominent which define the passage to the stage of slow decline. The existence of the
tipping point phenomenon has been questioned by several authors [96, 94, 22, 60].
The mathematical considerations described above suggest the phenomenon should
exist.

3.2.3 Expansion of a minority enclave


The process of neighborhood change, which is described by a moving front, is termed
invasion-succession process. This term is taken from animal/plant ecology [22] and
is used to describe the process of the expansion of various ethnic enclaves in the
U.S [47]. Omer [97] concluded that a diffusion process can describe the expansion
of the Arab minority in Jaffa. Travelling waves are common solutions for reaction
diffusion equations. Figure 3.8 shows how the proposed model predicts the moving
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 53

I 2

v 1 A
1.5
B
0.8
I 1

II III IV V II III IV V
0.6 0.5
0 100 200 0 100 200

C 0.8 D
1.5
I
II
0.7
1

I II
III IV V III IV V
0.5 0.6
0 100 200 0 100 200
t
Figure 3.6: Time dependence of v population pertain to the four trajectories of Figure
3.5.In trajectories A and D the v population converge to the N+ whereas trajectories
B and C show a tipping process which shift the v population toward the N− state.
We can identify five stages : (I) initial condition, (II) transient (III) slow growth in
cases B and C (decline for A and D), (IV) rapid growth (decline), (V) convergence.
(see the above text for more explanation).
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 54

−3
x 10
5 0.1

v’(t) A B
0.05
0
0
−5
−0.05
II II
III IV V III IV V
−10 −0.1
0 100 200 0 100 200
−3
x 10
0.1 10

0.05
C D
5
0
0
−0.05
II III IV V II III IV V
−0.1 −5
0 100 200 0 100 200
t
Figure 3.7: Time derivative of v for the different cases of Figure 3.6 showing the
different stages of the dynamics (the first stage is not show in this figure).

interface caused by a higher growth rate (α > 1) of the v-residents. The left frame in
this figure shows the initial condition; a sharp interface separating a v − population
(white domain) from a u-population (black domain), with an enclave of u-population
inside the domain of v-population (the black square). The parameters chosen are
such that pure populations states P− and P+ are unstable but the mixed population
states N− and N+ are stable. The successive frames (from left to right) show the
time evolution of the initial population distribution. The white and black domains
first converge to the stable state N− and N+ states, respectively, and the initial sharp
interface widens up to its characteristic width. Along with these processes the black
domain (N− ) invade the white domain (N+ ) and merge into a single black domain
that continues invading into the white domain. Such merging processes have been
reported by Deskins [92] in Detroit black ghetto and by Hansell & Clark [61] in the
Milwaukee black ghetto expansion during the years (1950-1960).
Figure 3.9 shows the time evolution of an enclave at a point, which is initially
at an area with a majority of u-residents (N+ ) and as the moving front advances, it
shifts the system to a state with v-residents majority (N− ). The behavior is similar
to C type trajectory in Fig. 3.5 which after an initial growth the tipping process
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 55

Figure 3.8: Temporal evolution of the interface zone. The interface moves due to
higher growth of the weak population. Black areas correspond to regions occupied by
v-population and shrinking white areas correspond to areas occupied by u-population.
Simulation parameters: µ = 1.9, α = 1.006, ² = 1, δ1 = 0.95, δ2 = 0.6, δ4 = 0.001

takes place and shift the system towards N− state. Compare to the non-spatial case,
here the time evolution is more gradual.
The discussion of the tipping point phenomenon was restricted so far to uniform
solutions of the model equations. In this context, the initial invasion that induces
the tipping was assumed to be uniform in space. Very often the invasion process
occurs locally. A “successful” tipping results in a nucleus of the inversed population
that grows or approaches a fixed size. Numerical studies of the non-symmetric model
indicate, however that the growth of the nucleus is not guaranteed, even if locally
population inversion took place; the nucleus must have a minimal size or a
minimal radius otherwise it will diminish in size and disappear.
Another interesting phenomenon, that can be investigated with the proposed
model, is the critical size of a minority enclave, which allows the enclave either to
remain stable or to expand, if, for example, the growth rate of the v-population is
greater than the growth rate of the u-population. We shall refer to such a critical size
of the enclave as the critical nucleus, Rc . If R < Rc then the enclave will shrink and
we call this radius as sub-critical nucleus while if R > Rc then the enclave will grow
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 56

70

60 IV
v III
50

I II
40

30
0 50 100 150 200 250 300 350 400 450

70

60
u
50

40

30
0 50 100 150 200 250 300 350 400 450
t

Figure 3.9: Time evolution of u and v at a point that is initially at N+ state. Can
be identified five basic stages in time evolution of v-population. (a) The percentage
of v-population from the total population. (b) The percentage of u-population. (see
the above text for more explanation. The values of parameters are as in Figure 3.8.

and take over the whole domain and we call it super-critical nucleus. In ecology, a
similar problem was investigated by Gandhi et al. [98, 99], using discrete models and
reaction diffusion equations. For the interface motion in the thin front limit we have
the following equation [100]
cn = c0 − Dκ (3.2.1)

where cn is the normal velocity of the front, c0 is the planar front velocity, κ is
the local curvature of the front and D is a constant that depend on the specific
equations. For the symmetric model Eqs. 2.1.12, the planar front velocity is zero,
so the enclave shrink in time and for circular front we have κ = 1/r and smaller
enclave will extinct faster. Figure 3.10 shows the dependance of the critical radius on
the status gap for two values of asymmetry between the two populations. It shows
that the critical radius increases as the status gap increases (or µ decreases). It
farther shows that the critical radius decreases as the asymmetry increases. Why
should a minority enclave reach a certain critical mass in order to expand
more rapidly in a majority-populated area? The answer to this question is
rather straightforward: If the relationship between the majority and minority is that
of repulsion, the concentration of minority families in the majority populated area
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 57

16

14

12

10
rc 8 α=1.005

6
α=1.01
4

2
1.7 1.75 µ 1.8 1.85 1.9

Figure 3.10: Critical radius rc of a minority enclave as a function of µ for two dif-
ferent values of α. Decreases the value off µ increases the status gap between the
two populations as a consequence the critical radius is larger, as more minority pop-
ulation needs to establish a stable enclave. for higher values of α the growth rate
of the v-residents is greater and this results in a lower value of the critical radius.
Enclaves with an initial radius that is less than rc disappear in the course of time and
institutionalization process takes place.

should be sufficiently large in order for the majority households to realize that their
way of life in the area is irreversibly threatened and that they may be better-off by
moving else-where. On the minority side, reaching a certain level of concentration
may trigger the ’economies of scale’ that may help to support more specialized, ethnic
stores and institutions (the process of institutionalisation, [26]). According to the
proposed model, the critical mass of the minority enclave increases in line with the
status gap, as indicated by lower values of µ. In part, this reflects the fact that for a
low status minority group is much harder to penetrate to a high status neighborhood.

3.2.4 Uneven expansion of a minority enclave


Interface in two-dimensional systems may become unstable also to transverse insta-
bility, that is, perturbations leading to a curved interface [68]. Imagine a straight
interface perturbed locally so as to have a bulge. In the course of time this bulge
may smooth out and disappear if the interface is stable, or grow into a finger if it
is sufficiently unstable. Transverse instabilities are often develop when the interface
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 58

motion triggers a process that acts to slow down the speed of propagation. If this
process is diffusing fast enough to the two sides of the bulge it may slow down the
propagation of those parts of the interface and let the bulge grow.
In non-linear systems we can distinguish among three different growth modes. (1)
Even growth, where the boundary retains its circular shape. (2) Domain splitting,
where the growth is accompanied by splitting into disjoint domains. (3) Fingering
and tip splitting, which are caused by the transverse instability.
This instability, which exists in our model leads to a corrugated enclave shapes.
In the proposed the transverse instability mechanism can also stems from a Turing
instability of the unstable branch (see the next chapter for details about this analysis).
These uneven shapes were explained by an anisotropy of the urban space, but from the
point of view of dynamical systems this phenomenon can occurs even in homogeneous
environment. The transverse instability was addressed as a Stefan moving boundary
problem by Rosser [66], to explain the wedge shaped ghettos. He concluded that
if a portion of a ghetto’s boundary begins to move more rapidly then the others, it
will continue to do so and the outcome will be a wedge-shaped ghetto. Downs [93]
proposed that a finger like spread of a minority can be explained, if only economic
consideration dominate. Downs argued that if minority household in the transition
zone have higher income than the majority household, they will try to move into the
majority area and this process will continue till a new equilibrium will occur.
In the proposed model increasing the value of δ and keeping the value of the other
parameters fixed imposes the transition from an even growth mode to uneven growth
and to labyrinthine pattern, which is hardly seen in reality due to streets and other
barriers (highways, streets and etc.) that are common in the urban space. These
mode of growth are driven by the cross diffusion terms in the model. Figure 3.11
shows numerical simulation of the symmetric model with a set of parameters beyond
the transverse instability of the front that connects N− with N+ . For each value of
µ there is a critical value δt that for δ > δt the front becomes unstable to transverse
perturbations (see Figure 4.22 and the next chapter for mathematical treatment of
the phenomenon.
Evidences or clues for this instability in a black ghetto expansion patterns was
mentioned by Morrill [57], who argued that a ghetto does not progress evenly and
smoothly in all directions but exhibits an uneven edge which moves at different rates
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 59

(a) (b) (c)

Figure 3.11: Undulating transition zone. The evolution of an initially perturbed


front which unstable to transverse perturbations. The light and the dark regions
correspond to the N+ and N− states. The frames a,b,c pertain to times t=0, 4000,
10000. The computational parameters are: µ = 1.88, δ = 1.12, δ4 = 0.1.

at different direction.
We can explain this uneven growth of the transition zone by the behaviour of the
u residents at the interface which make an adjustment to a perceived threat from
the v residents [59]. This mechanism which is called filtering [60] responsible for
the migration of u-residents out of the initial enclave, but now as in contrast to
the planar front case, they can move not only forward but also left and right. A
result of this process their density in the front of the initial bulge is smaller than
that of the stationary front and higher in the edges of the bulge, so that locally the
equilibrium condition of the stationary front is violated. Because of this redistribution
of u-residents the initial v bulge start to move forward as v-residents invade the u-
population area. In the light of this interpretation this mode of growth is possible
only with a mixed population states N± and impossible if the front connect pure
population states.
This preliminary results which argue, that corrugated fronts (transition zone) may
develop due to a transverse front instability even in homogeneous systems, is quite
new in the urban patterns research and it may be hard to prove it because the real
urban space is heterogenous.
The proposed model can simulate an uneven expansion of a minority enclave that
results from heterogenous space, by considering space dependence of the parameters.
Figure 3.12 shows such an enclave expansion in a space that consists of four neigh-
borhoods with a different value of µ. The v-enclave expand into the neighborhoods
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 60

Figure 3.12: Enclave expansion in a heterogeneous environment. The initial enclave


grows into the quarters with the lower status (µ = 1.8) where the growth in the
quarters with a higher status gap (µ = 1.2) is depressed. Values of parameters: ² = 1
,ξ = 1, α = 1.02, δ = 0.9, δ4 = 0.001

with a higher value of µ which the u-residents there are with a lower status, thus
they follow the path with the lowest resistance. Although the growth rate of the
v-population is higher they don’t invade into the high status u neighborhoods.

3.3 Patterns with nonlocal migration


In recent years it becomes clear that pattern formation in several experimental sys-
tems is affected by nonlocal affects [101]. Specifically, the dynamic behavior is
affected by a weighted spatial average value of one or several variables. Systems with
such nonlocal effects are described in general by a model of the form
½ Z ¾
∂u 2
= D∇ u + f u, p(u(x))dx (3.3.1)
∂t

where in many cases, Z Z


p(u(x))dx = u(x)dx = hui (3.3.2)

The nonlocal effect introduces a new mode of communication among the various
elements of the distributed system. This can generates a very rich set of patterns,
some of which are not seen in the absence of the coupling [101]. Examples of such
systems cases ferromagnetic systems, catalytic ribbons kept at a constant temperature
by electrical heating, catalytic ring, and reactive gel surrounded by active fluid in a
well mixed vessel (more physical systems can be found in [102, 103, 3]). It was found
that global coupling interaction introduces novel types of bifurcation, affects pattern
selection, and increases the sensitivity of the attained pattern to the initial conditions.
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 61

In Eqs. 2.1.12 the nonlocal migrations terms are (with ρ1 6= ρ3 and ρ2 6= ρ4 )

UN L = ρ1 [hui − u] − ρ2 [hsi − s]
(3.3.3)
VN L = ρ3 [hvi − v] + ρ4 [hsi − s]

where hui is the average of u over the whole residential area A and defined as
Z
1
hui = u(x0 , t)dx0 (3.3.4)
A
A

In the geographic literature the term residential mobility is generally used to refer
to short distance local moves while migration is used for moves of greater distance,
from city to city and state to state [104, 72]. In most cases the movements of
residents are of short distance and moves with increasing distance decline. This local
or short migration described by the diffusion terms in the model. There are other
circumstances, particularly when the segregation is weak, where nonlocal migration
might be important. The first terms in the expression for UN L and Vnl describe, that
at any site residents are leaving for less populated sites and new residents are coming
from denser regions (“global crowding effect”). The second terms are generalizations
of the cross-diffusion of the status in 2.1.12. At any site the u-residents are leaving
for higher status sites and new u-residents are coming from lower status regions.
The opposite is true for v-residents. In most cases ρ2 > ρ4 as the v-residents have
more economic ability and the necessary information in order to perform intra-urban
migration.
In the following simulations two assumption were used to simplify the study:

1. We set ρ1 = ρ3 = 0, because the nonlocal migration terms that are controlled


by these parameters are less likely to occur.

2. For simplicity we will assume that ρ2 = ρ4 = ρ ¿ 1 so we left only with one


parameter.

We have seen in the last chapter that for 2 < µ < µc the M state is Turing unstable.
Adding nonlocal migration to the dynamics affects the patterns in two ways: the first
it increase the segregation i.e the difference in population densities between the two
groups, as shown in Figure 3.13a and Figure 3.13b. Second it enables Turing patterns
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 62

(a)
u
1

0 10 20 30 40 50 60 70 80 90 100
1.5
(b)
1

0.5
0 10 20 30 40 50 60 70 80 90 100
1.2

1.1 (c)
1

0.9
0 10 20 30 40 50 60 70 80 90 100
x

Figure 3.13: Effect of nonlocal migration on weak segregation patterns. (a) Regular
Turing patterns with no global coupling. Parameters: µ = 2.01, ρ = 0. (b) Nonlocal
migration increases the amplitude of the patterns. Parameters: µ = 2.01, ρ = 0.05.
(c) Turing Patterns for µ > µc Parameters: µ = 2.1, ρ = 0.05. In all the figures
the initial condition is random noise from the M state. Values of other parameters:
² = 1, ξ = 1, δ = 1.12, δ4 = 0.1.

for µ > µc as shown in Figure 3.13c. The M state is unstable for 2 < µ < µN L and
µN L is a function of ρ. We can see also that the wavelength of the final pattern
increases compare to the ’regular Turing pattern’ (the one with ρ = 0). The same
result have been observed [3] in a well mixed vessel, if global interaction between
the gel and the vessel is taken into control. Spatial patterns can result from global
control, even when reaction-diffusion mechanism cannot create patterns on their on.
Nonlocal migration can lead to develop of segregation in the bistable region. Fig-
ure 3.14 shows how a stationary front develops from initial random perturbations of
the N+ . Without the global coupling the N+ is stable to these perturbations, but the
nonlocal migration amplify the initial perturbations and creates a front that connect
the N− with the N+ . The parameter ρ enhances the segregation between the two
residential groups in the model as it introduces a positive feedback mechanism that
coupled the whole space. We assume that residents in a specific location have global
information about the population density and status distribution in the entire city.
Nonlocal migration can lead to a creation of an enclave that begins from localized
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 63

1.5
(a)
u 1

0.5
0 20 40 60 80 100
3

2 (b)
1

0
0 20 40 60 80 100
1.5

(c)
1

0.5
0 20 40 60 80 100
x

Figure 3.14: Development of a stationary front due to nonlocal migration. (a) The
initial state is random noise around the N+ . (b) Without nonlocal migrations (ρ = 0)
the final state is the uniform N+ . (c) The random noise is amplified due to nonlocal
migration and a stationary front that connects the two new stable states is developed.
Other parameters values: ² = 1, ξ = 1, µ = 1.95, δ = 1, δ4 = 0.01

finite perturbations from the stable N+ state as shown in Figure 3.15. Without these
terms the N+ state is stable. This effect of non-local migration can enable the tip-
ping phenomena of a neighborhood even if the initial state is belongs to the basin of
attraction of the N+ (see [70]).
The nonlocal migration or the global coupling interaction in the model finds sup-
port with Anderson’s description of segregation. Anderson [105] proposed a simple
dynamic model which view the segregation process from the perspective of selective
migrations to describe segregation in Sweden. Anderson identify three principal stages
in the segregation process. The first is the proto-segregation process that may vary
over time and space. The second stage is called the segregation-generating migration
and it could be said to start when a neighborhood reaches a level of immigrants higher
than the city average. In the last stage which Anderson termed segregation-generated
migration the native residents start to leave the area mostly because of the decrease
in the status of the neighborhood. This indirect effect can be relate to the nonlo-
cal migration status terms in 3.3.3 and in Anderson’s words “As the residential area
starts getting a ’doubtful reputation’, a process that often takes place outside the
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 64

0.9
initial perturbations (a)
v 0.8

0.7
0 20 40 60 80 100
3

2
(b)
u 1

0
0 20 40 60 80 100
1.5

(c)
u 1

0.5
0 20 40 60 80 100
x

Figure 3.15: Segregation as a result of nonlocal migration. (a) Initial state (b) Only
local migration (ρ = 0.05), the perturbations decay and to N+ state. (c) Nonlocal
migration drives the system into segregated pattern with two v enclaves. parameters
values: ² = 1, ξ = 1, µ = 1.95, δ = 1, δ4 = 0.01

area itself, people within the area will become affected.” Residents with extensive
social networks will be affected more than those who have more local. In the view of
our mathematical model Anderson describes segregation that start to develop from
the N+ state and enhanced due to nonlocal migration with ρ1 > ρ2 . Ellen [106] in
his proposed race-based neighborhood projection model also supports the nonlocal
migration term for u. In the context of white-black segregation he suggests that rela-
tive status matters and that if the status income of white homeowner is high relative
to the tract mean, he is more likely to move away than those which their status is
less than the average.

3.4 Weak segregation as a result of Turing insta-


bility of the M state
Segregation in the context of the model may appear as an instability of a uniform
mixed state to a finite wavenumber mode at a critical µ value (see previous chapter).
Figure 3.16 shows a typical segregation pattern. The over and under-presentation
of a given population group at neighborhoods grow as µ decreases but remain small
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 65

1.2
u
1

v
0.8

0.6

0.4

0.2
s
0

−0.2
0 20 40 60 80 100
x

Figure 3.16: A stationary periodic solution of Eqs. 2.1.12 beyond the primary insta-
bility at µ = µc , representing weak segregation. Parameters: ² = 1, ξ = 1, µ = 2.001,
δ4 = 0.2, δ = 1.12,

as compared with the absolute population densities. 2 We refer to this form of


segregation as to weak segregation. The driving force for weak segregation is the
positive feedback effect of migration due to status considerations: u-residents migrate
to regions with higher status and thus raise the status farther. The migrations of v-
residents away from such regions adds to this effect. Similar considerations hold for
regions with low status. The spatial modulations of the M state are very weak close
to µc and increase as µ departs from the crucial value. The periodic solutions may
undergo secondary instabilities, e.g. Eckhaus instability or zigzag instability 3 , that
destroy their periodicity either transiently or permanently (see the next chapter).
The N± states also Turing unstable and so patterns of weak segregation allow for
these state. It is very hard to identify the weak segregation patterns in a real city
because, the the urban space is almost never homogeneous and it is hard to find a
neighborhood which the initial state is M state or N± . The mechanism of pattern
formation due to finite wavenumber instability in the context of population dynamics
have been suggested by ecologists [82, 107, 83] as one mechanism that can produce
patchy structure in predator-prey systems.
From the urban dynamics standpoint this hypothesis needs farther research al-
though Puu and Beckmann [43] suggested that it may exists. Guest and Weed in
2
One of the effects of nonlocal migration as shown in the former section is to increase the disparity
between the densities of the two groups.
3
See the next chapter for mathematical analysis of these instabilities
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 66

(a) (b)
250 250 1.4
1.15
200 1.1 200 1.2
1.05
150 150 1
1

100 0.95 100 0.8


0.9
50 0.85 50 0.6
0.8
50 100 150 200 250 50 100 150 200 250

Figure 3.17: The effect of nonlaocal migration on weak segregation patterns- two
dimensional simulation. (a) Local migration only. (b) With nonlocal migration ρ =
0.08. Parameters: µ = 2.01, ² = 1, ξ = 1, δ1.12, δ4 = 0.1

their study of residential segregation Cleveland, Boston and Seattle [90] concluded
that some segregation among ethnic groups can be realized on the basis of simple
chance and that the minimum degree of this segregation is unknown. This descrip-
tion of segregation between two equal ethnic groups resembles the mechanism of the
week segregation in the proposed model and may be a preliminary clue to its existence
in cities.
Adding non-local migration terms to the model can affect the patterns of the weak
segregation in a dramatic way as shown in Figure 3.17. The segregation becomes
stronger and the labyrinthine pattern changes to a patchy environment, that is more
similar to real segregation patterns. The nonlocal migrations terms in in the model,
delineate an ideal situation in which the dynamics in a certain site, is coupled to
whole space. In reality this assumption can be relaxed so, only a limited environment
around the site will contribute to the dynamics.

3.5 Including discrimination in the model


To include discrimination between the two residential groups, we need to add a direct
coupling between u and v. The simplest way to do it, is by adding competitions
terms like in the Lotka-Volterra model. For convenience we will add these terms to
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 67

the symmetric model Eqs. 2.1.12 without the diffusion terms:


∂u
∂t
= u − u2 + us − η1 uv
∂v
∂t
= v − v 2 − vs − η2 uv (3.5.1)
∂s
∂t
= ε(u − v − µs) − ξs3

The parameters η1 and η2 control the strength of the repulsion interaction and are
called competition coefficients [108] (p. 379) Foe example η1 measures the competitive
effect of v-residents on the u-residents. The interpretation of this term is that u-
residents do not like to live with v-residents mainly because they believe that property
values will fall down [109]. By changing the signs of η1 and η2 in 3.5.1 we can account
for other kinds of relations between the two groups [51].4
In human ecology the new terms in 3.5.1 describe mutual avoidance between the
two groups and can be explained by the principle of homophily which says that people
are attracted to those like themselves and repelled by others who are racially different
[110]. For simplicity we will take η1 = η2 ≡ η and analyze the linear stability of the
stationary uniform states. Figure 3.18 shows the bifurcation diagram for η = 0.3.
The significance of the new terms is that they reduce the domain of bistability of the
mixed states N± . The states N± are stable for µ1 < µ < µ2 where
−1−7η+η 2 −η 3
µ1 = (η−1)(1+η)2
(3.5.2)
2
µ2 = 1−η

It stems from 3.5.1 that for η = 1 only the pure states p+ and p− are stable, which
means that only strong segregation exists. We conclude that if the racial preferences
considerations are strong compare to the status ones, the mixed states will be stable
for a small range of the control parameter or even missing.
Adding the diffusion terms and including local direct avoidance interaction be-
tween the two groups will results in the following model
∂u
∂t
= u − u2 + us − η1 uv + ∇2 u − δ∇2 s + σ1 ∇2 v
∂v
∂t
= v − v 2 − vs − η2 uv + ∇2 v + δ∇2 s + σ2 ∇2 u (3.5.3)
∂s
∂t
= ε(u − v − µs) − ξs3 + δ4 ∇2 s
4
For example η1 6= 0 and η2 = 0 means that the v residents are natural to u-residents whereas
the u-residents are averse to the v-residents.
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 68

2
P+
u1 N+ M
P− N−
0

−1

−2
0 0.5 1 1.5 2 µ 2.5 3 3.5 4

Figure 3.18: Bifurcation diagram for the extended model with direct coupling between
u and v (η = 0.3). Solid (dashed) lines stand for stable (unstable) solutions. The
prejudice terms decrease the domain of bistability. The pitchfork bifurcation occurs
at µ2 = 2/(1 − η) and for η = 1 only strong segregation states exist. The symmetric
mixed states exists for all values of µ and it is stable for µ > 2.857. The pure
states N+ and N− exist for all values of µ but are stable for 0 < µ < 2.567. The
Npm exist for µ < 2.857 and are stable for 2.567 < µ < 2.8567. They appear in a
pitchfork bifurcation of symmetric state at µ = 2.857, and loose stability in a pair of
transcritical bifurcations at µ = 2.567
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 69

The extended model can shed light on the old time debate about the different
factors of residential segregation. The vast literature on this subject suggests two
social models [111]. The first is the social class which is strongly implied that dif-
ferences in socioeconomic status largely determine the dissimilarities in residential
segregation. The second model is the ethnic model, which gives more significance
to non-socioeconomic differences between the groups, such as prejudice, various sorts
of market forces and illegal discriminatory acts. The ethnic model explains the high
and the persistent segregation of Negroes in American cities. Table 3.2 and table 3.3
summarize part of the voluminous papers in this subject, according to their explana-
tion to the residential segregation in the view of this two models. The racial and the
socioeconomic factors interact in a complex way to shape the segregation patterns in
the city, and the extent of each factor differs between the ethnic groups.
In the context of white-black segregation in American cities it is clear that racial
preferences of the white population reinforce the patterns of separation between the
two groups and is one of the main reasons of persistence of segregation-“if segregation
was a matter of income, rich blacks would live with rich whites and poor blacks with
poor whites. This does not happen” [109]. Because of the interaction of social and
racial factors, integrated neighborhoods are very difficult to maintain and too often
it is only a temporary condition that exists between the arrival of the first black
household and withdrawal of the last white one [112].
In the extended model 3.5.1, it is clear that the domain of stability of the mixed
states has been reduced compare to the pure socioeconomic model 2.1.12. The signif-
icance of the mixed states that they allow rich (poor) minority residents to live with
rich (poor) majority residents. If the prejudice and the discrimination are strong
enough, the mixed states disappear and the segregation is strong. For example in
Australia where the discrimination is small and the social class model can apply
more confidently, the Asia born immigrants live in houses that at least as good as
Australian born residents [25]. The immigration to Israeli cities from the former
Soviet Union during the 90’s is another example where economic considerations were
dominant.
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 70

Ethnic class Subject


Grodzins 1957 [45] white-black segregation in 207 American cities
Taeuber 1965 [46] white black segregation in American large cities
Schelling 1978 [113] theoretical preferences model
Clark 1988 [114] black, white and Hispanic segregation in Los Angles
Galster 1989 [115] black-white segregation in Cuyahoga County,
Cleveland Ohio
Farely 1994 [109] black-white segregation in Detroit area
Sims 1999 [116] black, white and Asian segregation in
five metropolitan cities in U.S
Massey et al. 1999 [117] black, Hispanic and Asian segregation in U.S 1980 cens.
White et al. 2000 [118] white, black, Hispanic and Asian segregation in
50 cities in U.S
Social class Subject
Steinnes 1977 [119] black-white segregation in Chicago
Guest and Weed 1976 [90] Multi-ethnic segregation in Boston,
Seattle and Cleveland
Darroch and Marston 1971 [111] multi-ethnic segregation in Toronto, Canada
Massey 1979 [27] black, white and Spanish in 29 cities in U.S
Anas 1980 [120] theoretical model
Massey 1981 [121] black, white and Hispanic segregation in 6 southwestern
cities in U.S
Massey and Denton 1985 [122] Hispanic-black and Hispanic-PuretoRican segregation in
1970 census data
Hwang and Murdock 1998 [110] Angolo, black, Asian and Hispanic segregation in 1672
U.S. suburban areas
Zang 2000 [25] black-white segregation in metropolitan areas with over
40000 blacks in 1990
st Jhon and Clymer 2000 [21] segregation of Asian immigrants in Melbourne
and Sidney, Australia

Table 3.2: Representative papers on social class and ethnic class models for ethnic
segregation
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 71

Social class and ethnic class Subject


Goring 1978 [96] theoretical paper
Frey 1979 [123] white-black segregation in 39 American cities
Clark 1986 [11] white-black segregation in 38 metropo. areas in U.S
Galster 1987 [124] black-white segregation in 40 SMSAs in U.S
Denton and Massey 1989 [125] black, Hispanic and Asian segregation in 60
U.S metropolitan areas
Galster 1988 [126] black-white segregation in 1980 SMSAs in U.S
Clark 88 [114] theoretical paper
Galster 1989 [115] theoretical paper
Clark 1989 [127] theoretical paper
Massey and Fong 1990 [128] black, Hispanic and Angolo in San-Francisco
metropolitan area
Massey and Eggers 1990 [129] black, Hispanic, white and Asian segregation in 60
U.S metropolitan areas
Denton and Massey 1991 [130] multi-ethnic segregation in U.S metropolitan areas
Massey and Eggers 1993 [20] segregation between income groups in 60
U.S metropolitan areas
Morrill 1995 [14] white, black and Asian segregation in Seattle
Boswell and Cruz-Báes 1997 [131] white, Hispanic, black and Asian segregation in Miami
Harris 1999 [71] black-white segregation in Census 1980 in U.S
Massey et al. 2000 [13] black, white and Hispanic segregation in 60
U.S metropolitan areas

Table 3.3: Representative papers on social and ethnic class models for ethnic segre-
gation
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 72

3.6 Making the control parameter µ a dynamic


variable
This section suggests another extension to the proposed mode by adding a forth
dynamic equation and briefly discusses the uniform and stationary solutions associate
with this new version. This section can be viewed as one of the many possibilities in
which one can extend the model and apply it to more complicate situations.
In their City-1 model Portugali et al. [2] characterized each individuals by two
economic properties status and tendency and also by a socio-cultural origin. Sta-
tus refers to the individual’s current economic ability which is similar to the status
variable in our model, and tendency which is the potential to improve status. The
tendency influence the decline/gain of status, for instance the status may be low in
the first period after arrival; with time some immigrants will improve their status and
others might deteriorate their previous status.
The status gap in our model is determined by the value of the parameter µ which
its value is fixed in time and space. One possible extension of the model is to redefine
µ as a forth dynamic variable of the model and let it change in time and space. A
preliminary model can be 5

ut = u − u2 + us
vt = v − v 2 − vs
(3.6.1)
st = ²(u − γv − µs) − ξs3
µt = µ − µ2 + ζµs2

where µ stands for the the status gap between the two population or as “acclimation”
parameter that measures the capability of residents to mix with different population
group. Low values of µ imply high status gap and vice versa. According to the forth
equation of 3.6.1 the growth rate of µ has a logistic part and a term that introduces
a non-linear positive coupling with the status s. The parameter γ can be interpreted
as a tendency parameter. In this dissertation all simulations were done with γ = 1,
but assign a lower value to γ means that v-residents lower less the status, so it defines
an immigrant group with a higher economic potential.
Uniform solutions of Eqs. 3.6.1 with ζ = 1 and γ = 1 reveal four new more
solutions to those of the symmetric model Eqs. 2.1.12. Table 3.4 shows the Uniform
5
For simplicity only the kinetic terms are shown.
CHAPTER 3. URBAN DYNAMICS ASPECTS OF THE MODEL 73

States Symbol Solution (u, v, s, µ)


No-Population O− (0, 0, 0, 0)
O+ (0, 0, 0, 1)
Symmetric mixed population M+ (1, 1, 0, 1)
M− (1, 1, 0, 0)
Non-Symmetric mixed population N− (um , vm , sm , µm )
N+ (vm , um , −sm , µm )
Pure population P++ (p1, 0, sp1 , µ1 )
P+− (p2 , 0, sp2 , 0)
P−− (0, p2 , −sp2 , 0)
P−+ (0, p1 , −sp1 , µ1 )

Table 3.4: Stationary uniform solutions of Eqs. 3.6.1

and stationary solutions of the new model.


These new uniform and stationary solutions enable the existence of pure states with
different value of status as sp1 < sp2 which correspond to different values of µ. The
state P++ has a higher value of µ and a lower value of status than the state P+− . The
interpretation of these two states is that the forth equation subdivides the u-residents
group farther to higher and lower status sub-groups. Similar arguments hold for P−+
and P− states. The mixed states population remain solutions of the new model. It is
important to note that in Eqs. 3.6.1 µ can vary in time, for example an initial v-group
with a lower value of µ which means strong segregation may be less segregated if µ
will grow in time which means better a “acclimation” in the city. More studies are
need to be done in order to develop this new model but this initial attempt is seen
plausible.
Chapter 4

Pattern Formation Aspects

A complex system is exactly that; there are many things going on simulta-
neously. If you search carefully, you can find your favorite toy: fractals,
chaos, self-organized critically, phase transitions analogies, Lotka-Volterra
predator-prey oscillations, etc., in some corner, in a relatively well devel-
oped and isolated way. But do not expect any single insight to explain it
all.

Rolf Landauer [31] (p. 252)

Unlike the previous chapter where the focus was on transient behavior such as tran-
sition zone, displacement or tipping point phenomenon. Most of the patterns to be
considered here are asymptotic, that is, patterns that emerge as time goes to infinity.
These patterns are quite different from the transient ones, but do contain information
about the long time evolution of urban patterns with their transient character.
This chapter is devoted to pattern formation aspects of the proposed model, and
analyze it as a reaction diffusion model without reference to urban contexts. This
chapter explore in a mathematical detail some generic aspects of pattern formation
phenomenon such as amplitude instabilities, (Eckhaus and zigzag instability), front
transverse instability and patterns that emerge from the unstable branch which may
have a maximum growth rate for a finite wavenumber.
In section 1 we derive an amplitude equation (NWS equation) for the stripes
that develop beyond the Turing instability of the M state and then we show by
numeric simulation the different instabilities associated with this amplitude equation.

74
CHAPTER 4. PATTERN FORMATION ASPECTS 75

In section 2 we show that in the bistable regime perturbations around the unstable
M state can grow to to irregular patterns which their structure depend on the growth
rate curve of the unstable M state. In section 3 we analyze the transverse instability
of the front and show that two mechanisms can lead to labyrinth forms: one is a
linear instability of the front to transverse perturbations and the second is nonlinear
instability front instability. The driving force behind this instability is the “Turing
instability” of the unstable M state i.e the growth rate curve has a maximum for
k > 0. For the sake of simplicity, most of the analysis is done for the symmetric
model eqs. 2.1.12.

4.1 NWS amplitude equation for the Turing insta-


bility of M state
A common analytical method for studying periodic patterns near their onset is the
derivation of evolution equations for the amplitude A of the fastest growing modes
kc . The amplitude is introduced by the relation

u(x, y, t) ≈ u0 + u10 A(x, y, t) exp(ikc x) + c.c , (4.1.1)

where u is the set of dynamic variables of the model (u, v and s in our case), u0
represents the uniform stationary state that undergoes the instability, u10 is a set of
constants and c.c. stands for the complex conjugate. The space and time dependence
of the amplitude A are assumed to be weak according with the behavior of the system
near instability points (the weak spatial dependence accounts for the dynamics of the
narrow band of modes that grow together with the central kc mode).
The advantage of this method lies the in the universal form of the amplitude
equation. Similar instabilities (finite wavenumber, stationary) lead to the same form
of equation irrespective of the particular system under consideration. The details
of the particular system come through the coefficients of the universal terms in the
equation. Thus, once the amplitude equation is derived all of its known properties
(solutions, instabilities, etc.) can readily be translated to the particular system.
For stationary finite wavelength instabilities the amplitude equation takes the form
CHAPTER 4. PATTERN FORMATION ASPECTS 76

[4, 85, 132]


µ ¶2
∂A ∂ 1 ∂2
τ = ηA + ξ02 + A − g1 |A|2 A, (4.1.2)
∂t ∂x 2ik c ∂y 2

where in the context of 2.1.12 η = (µc − µ)/µc , and the other coefficients need to be
evaluated. This equation, referred to as the Newell-Whitehead -Segel equation, can be
regarded as the normal form of a symmetry breaking bifurcation leading to roll stripe
patterns and allowing for modulation in both x and y directions. Equation 4.1.2 has
a family of exact solutions
A = Aq q exp(iqx)
η−ξ02 q 2 (4.1.3)
Aq = g1
,

where q = k − kc , representing stationary, spatially periodic solutions of the original



system with wavenumbers in a η neighborhood of kc . Stability analysis of these so-
lutions give a well known criteria for longitudinal (Eckhaus) and transverse (zigzag)
instabilities (hereafter “secondary instabilities”) [4]: η < 3ξ02 q 2 for Eckhaus instabil-
ity and q < 0 for zigzag instability.
Experimental evidences for the existence of Eckhaus instability has been reported
using a convecting liquid crystal layer [134]. The zigzag instability has been ob-
served experimentally in connection with the Bénard problem [133] and was studied
numerically with the Schnackenberg model [135]. 1
The Eckhaus and the zigzag instabilities provide a partial solution to the selection
of patterns problem beyond the critical point, since they imply the collapse of certain
types of structure that would appear to be allowed on the basis of linear instability
analysis.
1
The Schnackenberg model is a variant of the well known Brusselator model and after rescaling
the evolution equations take the form:

ut = γ(a − u + u2 v) + ∆u
(4.1.4)
vt = γ(b − u2 v) + d∆v

This model belongs to the class of the so-called “activator-substrate depleted” models i.e. when the
pattern emerges, the densities of u and v are out of phase by π. The same relation holds in our
model between u and v.
CHAPTER 4. PATTERN FORMATION ASPECTS 77

4.1.1 Derivation of the NWS equation


We derive an amplitude equation for the finite wavenumber instability of the M state:
(u, v, s) = (1, 1, 0) to spatially periodic solutions. First define new variables (α, β, γ)
which are deviations from the M state
α=u−1
β =v−1 (4.1.5)
γ=s

Substitution of 4.1.5 in 2.1.12 yields the following equations for (α, β, γ)


αt = −α − α2 + γα + γ + ∇2 α − δ∇2 γ
βt = −β − β 2 − γβ − γ + ∇2 β + δ∇2 γ (4.1.6)
γt = ²(α − β − µγ) − ξγ 3 + δ4 ∇2 γ .

A solution to 4.1.6 will be sought as an asymptotic expansion


       
α α1 α2 α3
 β  = √η  β1  + η  β2  + η 3/2  β3  + . . . (4.1.7)
γ γ1 γ2 γ3

where η := (µc − µ)/µc and µc > 2 is the critical value of µ where the M state
becomes. At µ = 2 the M goes through a pitchfork bifurcation (an instability to a
zero wavenumber mode). We assume here that µ > 2. Substitution of µ = µc (1 − η)
in 4.1.6 yields the following equations which contain the small parameter η
αt = −α − α2 + γα + γ + ∇2 α − δ∇2 γ
βt = −β − β 2 − γβ − γ + ∇2 β + δ∇2 γ (4.1.8)
γt = ² (α − β − µc (1 − η)γ) − ξγ 3 + δ4 ∇2 γ

We study 4.1.8 by means of multiples scale analysis. We assume that


   
α1 α0
 β1  =  β0  A(X, Y, T ) exp(ikc x) + c.c. (4.1.9)
γ1 γ0

where 4.1.9 X and Y stand for the slow space coordinates and are given by
X = η 1/2 x
Y = η 1/4 y (4.1.10)
T = ηt
CHAPTER 4. PATTERN FORMATION ASPECTS 78

.
The scaling of X and Y are motivated by the shape of the natural stability curve.
The scaling of T is in accordance with the expression for the growth rate.
From the chain rule of differentiation we therefore must take the replacements

∂x → ∂x + η 1/2 ∂X , ∂y → ∂y + η 1/4 y, ∂t → ∂t + η∂T , (4.1.11)

where on the right-hand side ∂x , ∂y and ∂t now act only on the fast variables x, y and
t . The Laplacian operator becomes

2
∇2 → ∂x2 + η 1/2 (2∂x ∂X + ∂Y2 ) + η∂X (4.1.12)


Insertion of 4.1.12, 4.1.11 and 4.1.9 in 4.1.8, we find to order η
 
α1
L  β1  = 0 , (4.1.13)
γ1

where L is the operator


 
−1 + ∂x2 0 1 − δ∂x2
L= 0 −1 + ∂x2 −1 + δ∂x2  (4.1.14)
² −² −²µc + δ4 ∂x2 .

We seek a solution in the form of 4.1.9 and by choosing γ0 = 1 we can find expressions
for α0 and β0
2
α0 = 1+δk
1+k2c
c
(4.1.15)
β0 = −α0

To the order of η we find


       2 
α2 α1 2α02 − 2α0 α0 − α0
L  β2  = L1  β1  +  2β02 + 2β 0  |A|2 +  β02 + β0  A2 exp(2ikc x) + c.c.
γ2 γ1 0 0
(4.1.16)
CHAPTER 4. PATTERN FORMATION ASPECTS 79

where the operator L1 is defined as


 
−1 0 δ
L1 =  0 −1 −δ  (2∂x ∂X + ∂Y2 ) . (4.1.17)
0 0 −δ4

We attempt a solution of 4.1.16 in the form as


        
α2 A1 A2 A0
 β2  =  B1  exp(ikc x) +  B2  exp(2ikc x) + c.c. +  B0  . (4.1.18)
γ2 C1 C2 C0

In order to find the coefficients A1 , B1 and C1 we have to solve the following system
of equations
−(1 + kc2 )A1 + (1 + δkc2 )C1 = Ψ(−α0 + δ)
−(1 + kc2 )A1 − (1 + δkc2 )C1 = Ψ(−β0 − δ) (4.1.19)
2
²A1 − ²B1 − (²µc − δ4 kc )C1 = −δ4 Ψ ,

where Ψ = (2∂x ∂X + ∂Y2 )A. The solution is C1 = 0, A1 = −δ4 Ψ/2², B1 = −A1 . The
coefficients A0 , B0 and C0 can be found by solving the system
   
−A0 + C0 2α02 − 2α0
 −B0 − C0  =  2β02 + 2β 0  |A|2 . (4.1.20)
²A0 − ²B0 − ²µc C0 0

The solution of 4.1.20 is

C0 = 0, A0 = B0 = 2α0 (1 − α0 ) |A|2 . (4.1.21)

The coefficients A2 , B2 and C2 can be found in the same way, which yields
α0 (1 − α0 ) 2
C2 = 0, A 2 = B2 = A . (4.1.22)
4kc2 + 1

At the order η 3/2 we get


       
α3 α2 α1 −2α1 α2 + γ1 α2 + α1 γ2
L  β3  = −L2  β2  + L1  β1  −  −2β1 β2 − γ1 β2 − β1 γ2  , (4.1.23)
γ3 γ2 γ1 ²µc γ1 − ξγ13
CHAPTER 4. PATTERN FORMATION ASPECTS 80

where the operator L2 is defined as


 2 2

−∂T + ∂X 0 −δ∂X
L2 =  0 2
−∂T + ∂X 2
δ∂X . (4.1.24)
2
0 0 −∂T + δ4 ∂X

A solvability condition at order η 3/2 will give the amplitude equation [4]. The solv-
ability condition can be written in a formal form as

(D0 , G) = 0 (4.1.25)

where D0 is an eigenvector of the adjoint operator L+ with zero eigenvalue and G


are terms in the rhs of 4.1.23 that multiply ∓exp(ikc x). The adjoint operator L+ is
given by  
−1 + ∂x2 0 ²
L+ =  0 −1 + ∂x2 −² . (4.1.26)
2 2 2
1 − δ∂x −1 + δ∂x −²µc + δ4 ∂x

The zero eigenvector D0 which satisfies L+ D0 = 0, is found to be


 
z
D0 =  −z  exp(ikc x) , (4.1.27)
1

where s
²δ4
z≡ . (4.1.28)
2(δ − 1)

Tedious algebraic manipulations and returning to the fast variables we finally obtain
to the NWS amplitude equation for A(x, y, t)
1 2 2
τ At = ηA + ξ02 (∂x + ∂ ) A − g1 |A|2 A , (4.1.29)
2ikc y

with the explicit expressions of the coefficient τ, kc , ξ0 and g1 in terms of the symmetric
CHAPTER 4. PATTERN FORMATION ASPECTS 81

model’s parameters
2
q
τ= 2zα0 +1
²µc³
, ξ02 = 4zk c δ4
²2 µc
, kc2 = 2²(δ−1)
−1
δ4 ´
g1 = ²µ1c 2zα0 (2α0 − 1)(1 − α0 )(2 + 4k21+1 ) + 3ξ . (4.1.30)
√ c

1+δkc2 2 2δ4 (²−1)−δ4


α0 = 1+kc2
, µc = 2δ − ²
.

Equation 4.1.29 has a family of exact solutions

A = Aq
q exp(iqx)
η−ξ02 q 2 (4.1.31)
Aq = g1
,

where q = k − kc , representing stationary, spatially periodic solutions of the original



system with wavenumbers in a η neighborhood of kc . These solutions which describe
stripes parallel to the y direction are unstable to longitudinal perturbation (“Eckhaus
instability”) for η < 3ξ02 q 2 where ξ0 is given by 4.1.30, and to transverse perturbations
(“zigzag instability”) for q < 0. Numerical testing of these instabilities are shown in
the next subsection.
The natural stability curve for the Turing instability of the M state for the sym-
metric model eqs. 2.1.12 can be obtained by solving the equation σ(k) = 0 for µ (σ is
the largest eigenvalue of Jaccobian at M state and is defined as σ3 (k) in 2.4.3.) and
substitution of the solution in η = (µc − µ(k))/µc gives the function η(k)
p
(1 + k 2 )δ4 + 2(δ − 1)² − 23/2 (1 + k 2 ) (δ − 1)δ 4 ²
η(k) = p (4.1.32)
(1 + k 2 )(δ4 + 2δ² − 23/2 (δ − 1)δ 4 ²)

Figure 4.1a shows the natural stability curve, η(k) for different values of ². The
behavior of η(k) near the the critical wavenumber kc can be studied by Taylor ex-
pansion of η(k) in powers of k − kc . Figure 4.1b shows the quadratic coefficient a2
which determines the shape of the parabola near kc , as a function of the parameter
². Decreasing the value of ² increases the band width for a given η value.

Figure 4.2 shows the amplitude of the stripes solution (u variable is shown) of the

full model and those obtained from the amplitude equation 4.1.29 (amplitude=4α0 η)
as a function of η. Increases the values of η increases the amplitude of u. As Figure
CHAPTER 4. PATTERN FORMATION ASPECTS 82

0.06 0.08
η (a) a2 (b)
0.05
0.06

0.04
0.04

0.03

0.02
0.02 ε=1.2

ε=1 0
0.01
ε=0.6
0 −0.02
0 0.5
k 1 0.4 0.6
ε 0.8 1

Figure 4.1: Behavior of natural stability curve, η(k) near the critical wavenumber kc .
(a) Graphs of η(k) for different values of ² show that for lower values of ² the natural
curves are flatter near the instability threshold which means the dynamics beyond
the instability involve a wider band of wave numbers. (b) The quadratic coefficient
in the Taylor expansion of η(k) vs. ² which determines the shape of the curve near
the threshold. For lower values of ² the parabola is flatter. Parameters: δ = 1.12,
δ4 = 0.1, ξ = 1.

4.2 that the amplitude equation solution is more accurate near the onset of the in-
stability (lower values of η and as we move away from the onset of instability (higher
values of η the difference between the two curves is larger. As discussed before, the
stripes pattern undergo two secondary instabilities: Eckhaus and zigzag. Figure 4.3,
shows the natural stability curve and the Eckhaus stability curve. The area between
the q = 0 axis and ηE curve corresponds to the domain of stability of roles patterns.2

4.1.2 Numerical testing of the amplitude equation


In order to compare the amplitude equation solution 4.1.29 with the numeric solution
of the symmetric model eqs. 2.1.12, we have to write the complex amplitude A as
A = p + iq where p and q are real functions of space and time. Inserting A = p + iq
2
Note that because µc ≥ 2, there exists a maximum value of η, ηmax = (µc − 2)/µc . For example
for ² = 1, δ = 1.2 and δ4 = 0.1, ηmax = 0.047619.
CHAPTER 4. PATTERN FORMATION ASPECTS 83

Amplitude equation
0.8
Amplitude

0.6
Full model
0.4

0.2

0
−0.01 0 0.01 0.02
η 0.03 0.04 0.05

Figure 4.2: Bifurcation diagram: amplitude of u pattern vs η = (µ − µc )/µc as


computed from numerical simulation of the full model and from the amplitude equa-

tion 4.1.29 (amplitude=4α0 η). Increases the value of η increases the amplitude of
u and enlarge the discrepancy between the full model and the amplitude equation.
Parameters: ² = 1, ξ = 1, δ = 1.2, δ4 = 0.1.
CHAPTER 4. PATTERN FORMATION ASPECTS 84

0.05

η
η
E A
0.04

C B
0.03

D E
0.02

η
L
0.01

0
−0.8 −0.5 −0.2 0.1 0.4 0.7

q
Figure 4.3: Stability boundaries of the amplitude equation 4.1.29 showing the reduce
control parameter η vs. q = k − kc . ηL is the linear stability curve and ηE is
the Eckhaus stability curve i.e ηE (q) < 3ξ02 q 2 . The strips are unstable to transverse
perturbations for q < 0. Points A, B, C, D define conditions for numerical experiments
in the next subsection. Values of parameters: ² = 1, ξ = 1, δ = 1.2, δ4 = 0.1
CHAPTER 4. PATTERN FORMATION ASPECTS 85

in 4.1.29 yields a coupled PDE equations for p and q


³ ´
1 1
2
τ pt = ηp + ξ0 pxx + 2kc (qxyy + qyyx ) − 4k2 pyyyy − g1 (p2 + q 2 )p
³ c ´ (4.1.33)
1 1
τ qt = ηq + ξ0 qxx − 2kc (pxyy + pyyx ) − 4k2 qyyyy − g1 (p2 + q 2 )q
2
c

3
The variables u, v and s relate to the p and q amplitude’s variables through

u = 1 + α = 1 + ηα1 + ηα2 + . . .

v = 1 + β = 1 + ηβ1 + ηβ2 + . . . (4.1.34)

s = γ = ηγ1 + ηγ2 + . . .

and for η ¿ 1 we can neglect the higher order terms in η. Using the expressions for
α1 , β1 and γ1 , eqs. 4.1.9 and 4.1.15 give

u(x, y, t) ≈ 1 + 2α0 (p(x, y, t) cos kc x − q(x, y, t) sin kc x)


v(x, y, t) ≈ 1 − 2α0 (p(x, y, t) cos kc x − q(x, y, t) sin kc x) (4.1.35)
s(x, y, t) ≈ 2 (p(x, y, t) cos kc x − q(x, y, t) sin kc x)

A comparison of the u component between the amplitude equation solution and the
full model equations is shown in Figure 4.4. The initial condition is identical and is
shown in the same scale, in Figure 4.5. We see that the amplitude equation reproduce
the solution to a very high accurate as far as the amplitude of the solution and its
wavelength are conserved. The error in the phase becomes significance at times
t À η1 = 103 , because of the truncation of the amplitude equation at the cubic order.

4.1.3 Amplitude instabilities


According to the theory of pattern formation, beyond the bifurcation point, an in-
creasing number of modes can become unstable. Nevertheless, only a limited range
of modes can be selected through a nonlinear coupling. The Eckhaus instability and
the zigzag instability are typical phase instabilities which may occur for stationary
periodic patterns. In this subsection we study numerically these instabilities by nu-
merical simulations of the full model eqs. 2.1.12 with values of parameters which
3
The finite difference formulas for the mixed derivatives pxxy and pyyx are given in chapter 5:
Numerical Methods
CHAPTER 4. PATTERN FORMATION ASPECTS 86

1.2

u Amplitude equation
Full model
1.15

1.1

1.05

0.95

0.9

0.85
0 5 10 15 20 25 30 35 40 45
x
Figure 4.4: Numerical testing of the amplitude equation in 2D. The initial condition
is: p0(x, y) = q0(x, y) = 0.01cos(kc x). The figure shows a cross section of u as
computed from the amplitude equation (solid line), and a u as computed from the
full model (dashed line). Parameters: η = 0.001, δ = 1.12, δ4 = 0.1, ξ = 0.2, ² = 1,
time=20000, (grid:128x128).
CHAPTER 4. PATTERN FORMATION ASPECTS 87

u0
1.15

1.1

1.05

0.95

0.9

0.85
0 5 10 15 20
x 25 30 35 40 45

Figure 4.5: The initial amplitude of u component is identical. The wavenumber is


twice the critical wavenumber kc and this mode decays while the critical mode grows.

are suggested by the amplitude equation 4.1.29. These simulations show up how the
different instabilities are expressed in the urban model. The following examples were
computed with no-flux boundary conditions. In fluid dynamics, there are good ar-
guments to predict that the stripes tend to form along a direction perpendicular to
the walls [135]. The different orientation of the walls imposes that, ultimately, some
blocks are parallel to the walls or that topological defects form and move slowly in
the structure. Figure 4.6 shows this tendency of stripes to reorient parallel to the
walls 4 . For higher value of η the spontaneous stripes are more wavy and irregular as
shown in Figure 4.7. The spatial modulations are large as µ departs from the critical
value µc .
For wavenumbers with k > kc (i.e q > 0) the amplitude can become unstable
through cross-roll instability and oblique instability. These instabilities correspond
to large wavenumbers. Since for a given η, the oscillation amplitudes are larger for
modes with k ∼ kc than for marginal modes (i.e modes that are Eckhaus unstable),
the system tends to destroy the latter in order to develop the former. The cross-
roll instability develops from stripes with a small amplitude perturbations as shown
4
The following two figures represents patterns that emerged from random noise around the uni-
form M state their locations in Figure 4.3 are not shown.
CHAPTER 4. PATTERN FORMATION ASPECTS 88

(a) (b)

Figure 4.6: Formation of zigzag patterns in a weakly confined system. The zigzag
patterns arise spontaneously in textures that form in weakly confined system from a
uniform state. Starting from a uniform state with random noise, the concentrations
slowly evolve to form a system of well organized domains, separated by lines of defects.
The frame a pertains to time t=2000, and b pertains to time t=6000. Values of
parameters: η = 0.01, δ = 1.12, δ4 = 0.01, grid: 256x256.

Figure 4.7: Formation of zigzag patterns starting from a uniform state with random
noise for η = 0.04. The pattern is more irregular compare to Figure 4.6 and resembles
a labyrinthine pattern; grid: 256x256 and t = 12000.
CHAPTER 4. PATTERN FORMATION ASPECTS 89

(a) (b)
Figure 4.8: Cross roll instability (point B on Figure 4.3). η = 0.03, q = 0.5 a: initial
pattern (vertical stripes with small amplitude horizontal stripes added 0.1cos(kc y)).
b: final pattern (t=6000).

in Figure 4.8. Adding perturbations in different orientation to stripes leads to the


oblique instability. The perturbations grow and impose their own wavenumber and
orientation to the eventual pattern as shown in Figure 4.9.
Figure 4.10 shows a development of the Eckhaus instability from a spatial periodic
pattern for η = 0.04 and q = −0.6. New roll pairs nucleate, leading to a higher wave
number close to the critical wave number kc . In systems with large aspect ration,
topological defects play an important role in wavenumber selection mechanism or the
transition to turbulence [137]. Figure 4.11 show such defects. For parameters close
ηE curve. The system reduce its wavenumber by a defects creation.
Figure 4.12 shows a stationary zigzag stripes. Stable zigzag patterns arise from
the destabilization of the parallel stripes by a small perturbation made by oblique
stripes, with a wavelength large with respect to the critical wave length kc [136].
Deeper in the regime of zigzag instability no saturation of the zigzag mode occurs
and the zigs and the zags of successive strips reconnect to yield a new set of straight
stripes [138]. When the control parameter is large (higher value of η in our model),
the zigzag pattern is stable. On the other hand for lower values, the zigzag structure
does not saturate and it leads to an oblique roll pattern via a reconnection process of
the roll parameter. Figure 4.13 show such pattern for η = 0.02
CHAPTER 4. PATTERN FORMATION ASPECTS 90

(a) (b)

Figure 4.9: Oblique instability (point A on Figure 4.3). η = 0.04, q = 0.6, a: initial
pattern (vertical stripes with small amplitude at different orientation). For example
the initial condition for u is: u(x, y, 0) = 1 + 0.2cos(1.6x) + 0.05cos(kc y − x). b: final
pattern (t=3000). Stripes reorientation occur in the interior of the grid, while near
the walls the no-flux boundary conditions impose that some stripes remain parallel.

(a)

(b)

(c)

Figure 4.10: Development of Eckhaus instability from a spatially periodic pattern for
η = 0.04 and q = 0.6 (correspond to point A Figure 4.3) at successive times (a) t=0,
(b) t=1200, (c) t=2400 . New roll pairs nucleate, leading to a wave higher number.
Values of other parameters :² = 1, ξ = 1, δ = 1.12, δ4 = 0.1 and dx = 0.306, dt = 0.01
CHAPTER 4. PATTERN FORMATION ASPECTS 91

Figure 4.11: Defect creation by the Eckhaus instability. The initial pattern is stripes
with small random perturbations, η = 0.02 and q = 0.35 (correspond to point E
Figure 4.3) grid:256x256, t=3000.

4.2 Patterns from ”Turing instability” of the un-


stable M state
In the bistable region of the symmetric model 2.1.12 i.e. for 1 < µ < 2 the M state is
linearly unstable to homogeneous perturbations, but it may be also unstable to non-
uniform perturbations i.e it may have a maximum growth rate for finite wavenumber
kc which is different from zero. This instability is different from the ’regular Turing
instability’ because it starts from an unstable state.5
When the Turing instability of the uniform mixed population state and the pitch-
fork bifurcation of this state occur close to one another they can no longer be treated
independently. In this parameter range we find that the Turing instability (finite k
instability) is amplified by the coupling to the pitchfork bifurcation (zero mode insta-
bility) and large amplitude patterns form. The asymptotic patterns depend on the
ratio between the growth rate of the of the mode with maximum growth rate to the
growth rate of the zero mode. For lower values of this ratio large domain patterns
5
Engelhardt [86] called this instability as secondary diffusion instability.
CHAPTER 4. PATTERN FORMATION ASPECTS 92

Figure 4.12: Zigzag stripes (correspond to point D on Figure 4.3) on 256x256 grid
with no-flux boundary conditions with the following set of parameters: η = 0.03,
q = −0.5 (grid:256x256). The stripes at the edges of the grid are affected from the
boundary condition.

(a) (b)

Figure 4.13: Zigzag instability for η = 0.02 and q = −0.4 (correspond to point D
on Figure 4.3). a: initial pattern as in Figure 4.9. b: final pattern (t=6000) the
amplitude of the zigzag structure does not saturate and a reconnection process of
the stripe pattern leads finally to oblique roll pattern. For this values of parameters
the initial pattern is also Eckhaus unstable and it is clear from Figure 4.13 that the
number of the oblique stripes is greater that the initial stripes number.
CHAPTER 4. PATTERN FORMATION ASPECTS 93

0.25
σ
0.2
µ=1.5

0.15
µ=1.6

0.1 µ=1.7

µ=1.8
0.05

µ=1.9
0

−0.05
0 0.5 1 1.5 2 2.5 3 3.5
k

Figure 4.14: Growth curves showing the finite wavenumber instability of the unstable
M state for δ4 = 0.05 and δ = 1.0. For µ = 1.9 the growth curve has maximum at
k = 0. Other values of parameters: ² = 1, ξ = 1

dominant and as we increase the value of this ratio strip patterns becomes dominant.
This is a generic mechanism of pattern formation for bistable systems like the SH
model [139], the reduced non-oscillatory EOE model [86] and FHN model [140]. In
this section we study this mechanism numerically in one and two dimension 6 .
Finite amplitude perturbations that emanate from one of the bistable states can
also evolve to patterns due to the interaction with the diffusive instability of the
unstable state. The situation can be more complex if this instability of the unstable
branch interact with the regular Turing instability of the stable states for µnc < µ < 2
(see the bifurcation diagram in 2.5). Figure 4.14 shows forms of growth curves for
different values of µ and δ. As µ is decreased below µ = 2 the maximal growth
rate at k = 0 (due to the pitchfork bifurcation of the M state) disappears and a
new maximum at a finite wavenumber k > 0 appears. Similar behavior is found by
increasing δ at constant µ
Decreases the value of δ for a fixed value of µ decreases the wave bandwidth which
is unstable and for δ < δc the curve does not have a local maximum for k > 0 and
the asymptotic solution is a front that connects the two stable states. Hereafter we
refer to the Turing instability of the M state only in case that the growth rate curve
6
In all the simulations in this section the initial state is the M state with random perturbations,
in contrast to the simulations in the next section which start from a perturbed front that connect
the two stable asymmetric states N+ and N− .
CHAPTER 4. PATTERN FORMATION ASPECTS 94

1.1

δc
periodic pattrens
1

0.9

0.8
large domains

0.7
1 1.2 1.4 µ 1.6 1.8 2

Figure 4.15: Boundaries for Turing instability of the unstable M state for different
values of δ4 (δ4 = {0.15, 0.1, 0.05}, dotted, solid and dashed lines). For δ > δc the
growth curve has maximum for k > 0.

has maximum for k > 0. A similar description is hold for increasing the value of µ
for a fixed value of δ. Close to δc the growth rate is small and the pattern is consists
of several modes. Increases farther the value of δ select one mode and the pattern is
similar to regular Turing pattern. For a fixed value of δ4 we can plot the values of δc
as shown in Figure 4.15 for different values of µ and δ. An analytic expression for δc
can be obtain by solving the equation σ 0 (k) = 0 and demanding that the maximum
will be at k > 0. This analysis gives
p
−(1 − δ4 − ²µ + δ4 ²µ) + (1 + δ4 ) 1 + 8² − 2²µ + ²2 µ2
δc = (4.2.1)

Figure 4.16 shows 1D patterns that evolve from random perturbations of the M for
different values of δ and at µ = 1.9. For δ < δc the asymptotic state is a stationary
front that connects N+ domain with N− domain. For higher values of δ two or more
modes are selected. Farther increasing the value of δ results in a periodic pattern
which pertains to a selection of one dominant mode. The main point here is that
the shape of the growth curve of the unstable state determine the final pattern. The
formation of large domains of the two asymmetric stable N+ and N− separated by
CHAPTER 4. PATTERN FORMATION ASPECTS 95

δ=1.04 δ=1.02
1.4 1.4

u u
1 1

0.6 0.6
0 50 100 0 50 100

δ=1 δ=0.98
1.4 1.4

u u
1 1

0.6 0.6
0 50 100 0 50 100

Figure 4.16: One dimension patterns that evolve from random initial perturbation of
the unstable state M, for different values of δ at µ = 1.9. Other values parameters:
² = 1, δ4 = 0.05, ξ = 1. For these parameters values δc = 1.0054.

smooth interfaces or fronts (in the absence of a transverse instability. See section
4.3).
A maximum at finite k favors the formation of large amplitude patterns with
a characteristic wave length-scale of order k −1 . Figure 4.16 shows a transition from
large domains at δ < δc to periodic patterns at δ > δc . Note that unlike regular Turing
patterns that appear at δ > δc , these patterns have large amplitudes, comparable to
the amplitudes of a front that connects the two states N+ and N− . This property is
a result of the coupling between a finite wavenumber mode and a zero-wavenumber
mode. Figure 4.17 shows the time evolution of the large amplitude two dimensional
patterns that evolve from random perturbations around the unstable M state. The
resulting patterns consists of large domains of the bistable N+ and N− states. We
can attribute these large domain patterns to the competition between the growth rate
of the zero wavenumber mode σ0 and the growth rate of the finite wavenumber σkmax .
It seems that the ratio Ω = σkmax /σ0 determines the long time patterns. The
CHAPTER 4. PATTERN FORMATION ASPECTS 96

(a) (b) (c) (d)

(e) (f) (g) (h)

Figure 4.17: Time evolution of large amplitude patterns which result of the coupling
between a finite wavenumber mode and a zero-wavenumber mode. The initial state is
random perturbations around the unstable M state which evolve to a large domains
patterns. The frames pertain to times t=250, 500, 1000, 1250, 1500, 1750, 2000. The
computational parameters are: ξ = 1, ² = 1, µ = 1.9, δ = 1, δ4 = 0.05 on a 256 × 256
grid size.

onset of the Turing instability of the unstable M state is for Ω = 1. As we increase


the value of Ω the finite wavenumber mode with the maximum growth rate (kmax )
becomes more dominant and the patterns are of labyrinthine type, consist of narrower
stripes as shown in Figure 4.18. In addition, in two dimensions the fronts that connect
the two stable states can be unstable to transverse instabilities as discussed in the
next section. This instability may causes to undulating fronts. Figure 4.19 shows Ω
and kmax as a function of the δ. Increases the value of δ results in higher values of Ω
which favors the selection of kmax wavenumber. For higher values of δ this value of
kmax is larger which corresponds to small domains pattern.
CHAPTER 4. PATTERN FORMATION ASPECTS 97

(a) (b) (c) (d)

Figure 4.18: Time evolution of labyrinthine patterns for Ω À 1. The initial state is
the same as in Figure 4.17 but for a larger value of δ (δ = 1.05). The large value of Ω
enable the selection of the mode with the maximum growth rate. The frames pertain
to times t=250, 750, 1250, 1500, 2000.

1.5 1
Ω kmax
1.4 0.8

1.3 0.6

1.2 0.4

1.1 0.2

1 0
0.99 1 1.01 1.02 1.03 1.04 1.05 1.06 1.07
δ
Figure 4.19: The thick line shows Ω as a function of δ for ξ = 1, ² = 1, µ = 1.9,
δ4 = 0.05 fixed. The thin line shows kmax as a function of δ. Increases the value of δ
and the keeping the other parameters fixed, results in a passage from large domains
pattern to a labyrinthine patterns.
CHAPTER 4. PATTERN FORMATION ASPECTS 98

4.3 Transverse instability


In bistable situations front structure and dynamics of front are affected by front in-
stabilities including in particular an instability to curvature perturbations (transverse
instability) and a pitchfork bifurcation of a stationary front to counter-propagating
fronts (this instability is called Ising-Bloch bifurcation). Transverse fronts instabil-
ities can be found analytically using linear stability analysis of planar fronts or by
deriving velocity curvature relations [141, 142].
Imagine a straight interface perturbed locally so as to have a bulge. In the course
of time this bulge may smooth out and disappear if the interface is stable, or grow into
a finger if it is sufficiently unstable. This instability is numerically demonstrated in
Figure 4.20. Shown in this figure is a simulation of equations 2.1.12, starting from an
almost flat interface. The interface initially develops bulges which grow into fingers
and tip split. the tip split grows into pairs of fingers and the process continues until
a stationary labyrinth fills the whole system.
The final labyrinthine pattern is an attractor of the dynamics, one out of many
other attractors, all having the same qualitative appearance but differ in qualitative
details. Transverse instabilities often develop when the interface motion triggers a
process that acts to slow down the speed of propagation. If this process is diffusing
fast enough to the two sides of an initial bulge, it may slow down the propagation
of those parts of the interface and let the bulge grow [68]. Recent investigations
of transverse instability in bistable systems reveal a new mechanism for labyrinthine
patten formation which we may call a nonlinear transverse instability. 7 The main
point is that the front can be unstable to finite amplitude perturbation, even if the
front is linearly stable. The driving force of this nonlinear instability is the Turing
instability of the unstable middle branch. The Turing instability of the unstable
branch favors the formation of finite wavenumber patterns and perturbations of flat
front that shift the system away from the basin of attraction of a flat front will not
decay but rather grow as the simulation in Figure 4.24 demonstrates.
This new mechanism reminisces the results of Lee and his coworkers [143] who
claimed that they had discovered a new kind of pattern forming mechanism gener-
ated by interacting chemical fronts in a spatially extended chemical system. Their
experiment consists of a continuously fed jel layer which supports a chemical reaction
7
The mathematical analysis of the SH model and FHN model were conducted with Arik Yochelis
CHAPTER 4. PATTERN FORMATION ASPECTS 99

a b

c d

e f

Figure 4.20: The evolution of an initially perturbed stationary domain above the
transverse instability line. In this case the front is linearly unstable to transverse
perturbations (see Figure 4.22 for more details). The frames pertain to times t=0,
5000, 10000, 15000, 20000, 25000. The computational parameters are: ξ = 1, ² = 1,
µ = 1.95, δ = 1.12, δ4 = 0.1 on a 128 × 128 grid size and with ∆x = 0.39, ∆t = 0.01.
CHAPTER 4. PATTERN FORMATION ASPECTS 100

which gives two states of Ph concentration, high and low. They found that as the flow
rate of chemicals is increased, the uniform and linearly stable stable low Ph state,
can be destabilized by a finite amplitude perturbation. The perturbation then prop-
agates throughout the low Ph state region, giving rise to lamellar-like pattern. The
fronts that develop approach each other until they reach a certain distance (0.4mm)
and then they stopped. Increasing the flow rate beyond a critical value results in
disappearing of this pattern.

4.3.1 Numerical linear stability analysis of interface


The associated eigenvalues problem to the front solution for the symmetric model
2.1.12 can be formally written as
   
ũ ũ

L ṽ  = λ ṽ 
 (4.3.1)
s̃ s̃

where L is the linear operator (for simplicity of computations we choose ² = ξ = 1)


 
1 − 2I1 + I3 + ∇2 0 I 1 − δ∇2
L= 0 1 − 2I2 − I3 + ∇2 −I2 + δ∇2  (4.3.2)
2 2
1 −1 −µ − 3I3 + δ4 ∇ ,

and (I1 , I2 , I3 ) is the front solution in two dimensions, and (ũ, ṽ, s̃) are the eigenmodes.
A positive eigenvalue λ corresponds to linear instability of the front.
In order to perform the numerical computation of the eigenvalues, a discrete ver-
sion of the Laplacian operator ∇2 is used
1
∇2 u = (ui+1,j − 4ui,j + ui−1,j + ui,j+1 + ui,j−1 ) (4.3.3)
h2

Including the no flux boundary conditions leads to a sparse matrix whose dimension
is [(n + 1)2 ]2 where n is the number of grid points used in the computation. To get
CHAPTER 4. PATTERN FORMATION ASPECTS 101

an idea how this matrix looks like, we write it for n = 2:


 
−4 2 0 2 0 0 0 0 0
 1 −4 1 0 2 0 0 0 0 
 
 0 2 −4 0 0 2 0 0 0 
 
 1 0 0 −4 2 0 1 0 0 
1  0


1 0 1 −4 1 0 1 0 (4.3.4)
h2 
 0


 0 1 0 2 −4 0 0 1 
 0 0 0 2 0 0 −4 2 0 
 
 0 0 0 0 2 0 1 −4 1 
0 0 0 0 0 2 0 2 −4 .

The matrix 4.3.4 belongs to a broad class of differentiation matrices (see [144] for more
details and spectral representation of these matrices). Because that the symmetric
model consists of three equations the dimension of the matrix is [3(n+1)2 ]2 . Since the
dimension grows like 9n2 it becomes very large as n grows. This make the computation
time consuming. The analysis has three stages:

1. Numerical solution of the stationary front by solving the symmetric model.

2. Construction of the matrix representation of L

3. Finding the eigenvalues of this matrix with a Matlab code

The numerical technique which is used in this analysis is explained in the next chapter.

4.3.2 Testing the method with the analytic result of the FHN
model
The FitzHugh-Nagumo model is [55]
∂u
= ²−1 (u − u3 − v) + δ −1 ∇2 u
∂t
∂v (4.3.5)
∂t
= u − a1 v − a0 + ∇2 v .

We consider a parameter range where 4.3.5 has two stable uniform states: an ”up”
state (u+ , v+ ) and a ”down” state (u− , v− ). For this model there is an analytic result
for the stationary front (Ising front) that connects the two stable states [55, 141].
Assuming ²/δ ¿ 1, for δ > δI where δI = 89 ²(a1 + 1/2)3 (for a0 = 0) planar fronts are
CHAPTER 4. PATTERN FORMATION ASPECTS 102

unstable to transverse perturbations. Figure 4.21 shows the NIB bifurcation and the
transverse instability for a1 = 4.
The growth rate of perturbations of the trivial unstable state (u, v) = (0, 0) is
calculated from the linearization of Eq. 4.3.5 giving the stability matrix
µ −1 ¶
² − δ −1 k 2 −²−1
A= (4.3.6)
1 −a1 − k 2 ,

The growth rate is found by solving det|A − δI| = 0 which gives


1 ³ √ ´
σ= T + T 2 − 4δ²D , (4.3.7)
2²δ

where
T = δ + k 2 ² + k 2 δ² − ²δa1
(4.3.8)
D = (δ + k 2 ²)(k 2 − a1 ) + δ .

We find δT the value of δ corresponding to the Turing instability of the unstable


solution, as a function of ² by demanding that σ(k) = 0 for some finite wavenumber
k > 0, by solving the equation σ 0 (kc ) = 0 where kc is the wavenumber with a maximum
growth rate. Using the fact that k is real gives

B + (1 + a1 ²) B − 2²
δT (²) = (4.3.9)

where
B = 1 + ²(2a1 + a21 ² − 2) (4.3.10)

Figure 4.21 shows δI (²) and δT (²) for a1 = 4. Labyrinthine patterns develop from
the linear transverse instability at δ > δI values which are far below δT . This implies
that the mechanism that stands behind the formation of the labyrinthine patterns
is not related to the Turing instability of the unstable uniform state. The δ = δI (²)
analytical result for the transverse front in the FHN model 4.3.5 allows testing the
validity of the numerical code for linear stability of the interface. The crosses in
Figure 4.21 show parameter values where the maximal eigenvalue, computed by the
numerical stability analysis code, crosses the zero. As the figure indicates these
parameters values agree very well with the analytic results.
CHAPTER 4. PATTERN FORMATION ASPECTS 103

1.8
(a) 85 (b)
δI
1.4 75
δT
1 65
δF
0.6 55
0.012 0.016 0.02 0.012 0.016 0.02
ε ε

Figure 4.21: The NIB front bifurcation and transverse instability boundaries for the
FHN model eqs. 4.3.5 for a1 = 4. (a) δF (²) and δI (²) are the boundaries for the NIB
front bifurcation and the transverse instability of the stationary front (Ising front).
The marked points in the figure are results from the numerical analysis described in
this section.(b) The curve δT (²) stands for the finite wave number instability of the
trivial state (u, v) = (0, 0). For δ < δT the maximal growth rate occurs at a zero
wavenumber while for δ > δT it occurs at a finite wavenumber.

4.3.3 Numerical results for the urban model


The semi-analytical method described in 4.3.1 was used to analyze the linear stability
of the front to transverse perturbations for the symmetric model eqs. 2.1.12. For each
value of µ the value of the parameter δ which first becomes positive was obtained
numerically (∆δI = 0.01). These values are described by stars in Figure 4.22 and
the straight line δI is the regression curve which fits these values. For values of δ
above the line δI the front is linearly unstable to transverse perturbations, whereas
for values of δ below δI the front is linearly stable to transverse perturbations but
large enough perturbations can grow and develop in time. This analysis shows good
agreement with the numerical simulations of the full model equations 2.1.12.
The numerical results are shown in Figure 4.22 for δ4 = 0.1. According to Figure
4.22 there are four regions in the (µ − δ) plane

1. Region I: Fronts are linearly stable to transverse perturbations and the growth
rate curve σ(k) for the unstable state M has a maximum for k = 0. Patterns in
this region consists of large domains of the two states N+ and N− .
CHAPTER 4. PATTERN FORMATION ASPECTS 104

1.6
δ
1.5 δ TN Stripes and Hexagons

1.4
III IV
1.3
Labyrinthine patterns

1.2 δI
1.1
Coexistence of Labyrinthine Patterns
and Large Domains Patterns
1
II δT I
Large Domains Patterns
0.9
1.5 1.6 1.7
µ 1.8 1.9 2

Figure 4.22: Phase diagram for the symmetric model eqs. 2.1.12 for δ4 = 0.1. δT N
define the Turing instability curve for the N± states (δT N = 3 + δ4 − µ). δI is the
boundary for the transverse instability of the front and was computed numerically (the
stars represent the computed values). δT is the boundary for the Turing instability
of the unstable M state. Above this line the growth curve has maximum for k > 0.
These line divide the phase space to four regions: (I) which in only large domain
patterns can exists, (II) where the non-linear transverse instability can drive the front
into labyrinthine pattern depend on the size of the initial amplitude perturbation, (III)
where the front is linear unstable and transverse perturbations grow to labyrinthine
patterns, in domain (IV) the two stables states N± are also Turing unstable and
strips and hexagons patterns may develop.
CHAPTER 4. PATTERN FORMATION ASPECTS 105

2. Region II: Fronts are linearly unstable to transverse perturbations, but the
growth rate has a maximum growth rate at a finite k. Small perturbations
decay to a planar front and larger enough can develop to labyrinthine patterns.

3. Region III: Fronts are linearly unstable and transverse perturbations develop to
labyrinthine patterns.

4. Region IV: The N± states are also Turing unstable and regular Turing patterns
can develop which include stripes and hexagons.

Figure 4.23 shows the different pattern formation mechanisms in the above mentioned
regions for µ = 1.97 close to the pitchfork bifurcation.
The conditions in region II suggests the existence of a nonlinear front instability.
The nonlinear transverse instability is depicted in Figure 4.24. The small pertur-
bations are damped because the front is linear stable, while the larger amplitude
perturbations grow and finally drive the system to a labyrinthine pattern.
Recent theoretical studies analyze the consequences of the interaction between
the Turing instability and the pitchfork bifurcation of uniform states [145, 3]. This
can modify the usual sequence of Turing planform bifurcation. It is important to
note that pure stripes or hexagons are generally not the solutions to the problem. In
particular the two inverted types of hexagons structures can coexist [146] as shown
in Figure 4.25. Hexagonal patterns are superpositions of stripes disposed at angels 0,
π/3, 2π/3. According to the value of the sum of their phases (0 or π), the maxima of
concentration u form a triangular lattice or a honeycomb, the opposite situation holds
for v. These two types are called respectively Hπ and H0 [135]. Fronts between the
two kinds of hexagons have been observed experimentally with SF6 near its critical
point and in oscillated granular layers [3]. Because of the Turing instability of the
unstable branch no small amplitude (e.g. periodic patterns around one of the stable
states N± ) have been found by numerical simulations eqs. 2.1.12.
CHAPTER 4. PATTERN FORMATION ASPECTS 106

δ=1.17
Region IV

δ=1.14
Region IV

δ=1.12
Region III

δ=1.05
Region II

δ=1
Region I

Figure 4.23: Numerical simulations showing the dynamics in the different regions
described in 4.22 for µ = 1.97 and δ4 = 0.1. In region IV for δ = 1.17 the final
pattern consists of spots and stripes. For a lower value of δ (δ = 1.14 but still in
region IV, the final pattern is of two kinds of hexagons. In region III beyond the
linear transverse instability (δI ) the initial perturbed front develops to a labyrinth
pattern. In region II and region I the perturbed front is linearly stable and the long
time pattern consists of two domains which are separating by an interface which is
width increases as the value of δ is decreased.
CHAPTER 4. PATTERN FORMATION ASPECTS 107

(a) (b) (c)

(d) (e) (f)

Figure 4.24: The evolution of an initially perturbed front that it is linear stable to
transverse perturbations (see Figure 4.22). The light and the dark regions correspond
to the N+ and N− states. The small amplitude perturbations decay to a planar front
due to the linear stability of the front, where the larger amplitude perturbations grow
and finally drive the system into a lamellar pattern. The frames a,b,c,d,e,f pertain
to times t=0, 2000, 4000, 6000, 8000, 10000. The computational parameters are:
µ = 1.82, δ = 1.12, δ4 = 0.1.
CHAPTER 4. PATTERN FORMATION ASPECTS 108

(a) (b)

Figure 4.25: The interaction of Turing instability of the stable state with the trans-
verse instability in a larger domain (100x100) than Figure ?? showing the same pat-
terns. The patterns in this figure are large amplitude patterns in the sense that their
amplitude connects the N± states. (a) Coexistence between Hπ and H0 hexagons
and stripes for conditions such that the Turing instability occurs in the vicinity of the
pitchfork bifurcation for µ = 1.985. (b) For µ = 1.99 the pattern consist of stripes.
Other parameters values: ξ = 1, ² = 1, δ = 1.12, δ4 = 0.1. White (Black) color stands
for maximum (minimum) value of u.
Chapter 5

Numerical Methods

Numerical simulations of the model’s equations were crucial to the development and
testing of the ideas presented in this dissertation. This chapter provides a description
of the methods and algorithms that were used for a numerical computing of solutions
in one and two dimensions and the algorithms for computing the eigenvalues associate
with the front stability analysis (see the previous chapter).
The numerical solutions were computed using the method of lines approach for
partial differential equations that decouples the discretization of the spatial and tem-
poral operators into independent problems. The spatial derivatives were represented
on uniform grids with finite difference approximations, and the resulting set of coupled
ordinary differential equations solved by a suitable explicit method.

5.1 Spatial Discretization


The first step in the method of lines approach for PDE equations is the discretization
of the spatial operators. The model equation can be formally written as [141]
Ut = f (U) + DL(U) (5.1.1)

with the L operator given by L = ∂ 2 /∂x2 in one dimension or L = ∇2 in two


dimensions. The vector U(x, y, t) represents the system variables (u, v, s), and f (U)
for the 2.1.12 is the kinetic part
 
u − u2 + us
f(U) =  v − v 2 − vs  (5.1.2)
3
ε(u − v − µs) − ξs

109
CHAPTER 5. NUMERICAL METHODS 110

The boundary conditions are chosen to be Neumann (no-flux). For most physical
cases, the Neumann boundary conditions are the most appropriate. For example, at
the boundaries of a stirred chemical reactor, the chemical species are not fixed to have
a specific concentration. However, the chemicals cannot have any flow through the
wall of the container. For an urban system the Neumann boundary condition means
that even though the residents of the urban area migrate into and out of the urban
system, the outside world is almost unaffected [51].
In one dimension the no-flux boundary conditions are implemented as symmetric
boundary conditions where all off the odd derivatives are zero at the boundaries,
x = 0 and x = Lx ,
Ux (0, t) = Uxxx (0, t) = . . . = 0
(5.1.3)
Ux (Lx , t) = Uxxx (Lx , t) = . . . = 0

For two dimensions the simulations were carried out on a uniform square grid which
adds the following conditions,

Uy (x, 0, t) = Uyyy (x, 0, t) = . . . = 0


(5.1.4)
Uy (x, Ly , t) = Uyyy (x, Ly , t) = . . . = 0

where Ly is the size of the domain in the y direction. In one dimension the continuous
operators were discredited on a uniform mesh of N + 1 points on varying domain size.
The grid points are labelled by xi , where i = 1, 2, . . . , N + 1. The number of grid
points were chosen to satisfy the condition ∆x = Lx /N ≤ 1. 1 This choice ensured,
that there were at least 6 to 8 points across the the front region. For higher values of
δ the front is very narrow and a greater number of grid points has to be used. The
simulations were checked against higher grid resolutions, to ensure that the patterns
were qualitatively unchanged.
The function f was represented on the grid points, xi , by its value at each location
f(Ui ), where Ui = U(xi ). The operator L was replaced in the interior regions of the
domain, with a discrete operator using 2nd and 4th order central finite difference
approximations [147],
Ui+1 −2Ui +Ui−1
LUi = ∆x2
+ O(∆x2 )
−Ui+2 +16Ui+1 −30Ui +16Ui−1 −Ui−2 (5.1.5)
LUi = 12∆x4
+ O(∆x4 )
1
Most of the simulations were done with 256 x 256 grid points and with ∆x = ∆y ≈ 0.39
CHAPTER 5. NUMERICAL METHODS 111

At the boundaries the same formulas were used to compute the derivatives by
adding ghost points outside the computational domain, one beyond each end in the
case of 2nd order approximation and two for 4th order finite difference scheme. The
solution is not integrated forward in time on those ghost points. They are updated
between each time step by filling their values by reflecting the solution about the
two boundary points U0 and UN . At the right end of the domain we obtain UN +1 =
UN −1 , UN +2 = UN −2 , and at the left end, U−1 = U1 , U−2 = U2 .
In two dimensions the function f was represented in the same way as in one
dimension, and the dimensional operator L = ∇2 was discretized using the five or the
nine points cross, depending on the order of the finite difference approximation.

LU = L̂xx U + L̂yy U (5.1.6)

where L̂xx and L̂yy are the one dimensional finite difference given in 5.1.5. Thus for
the 2nd order approximation we get,
Ui+1,j − 2Ui,j + Ui−1,j Ui,j+1 − 2Ui,j + Ui,j−1
LUi,j = + + O(∆x2 , ∆y 2 ) (5.1.7)
∆x2 ∆x2

In solving the amplitude equation 4.1.33 we have to use finite approximations for
mixed derivatives operators L̂xyy and L̂xxy . This can be done by writing

L̂xyy = L̂x L̂yy


(5.1.8)
L̂yyx = L̂y L̂yx

where L̂x and L̂y are given by [147]


Ui+1,j −Ui−1,j
Lx Ui,j = 2∆x
+ O(∆x2 )

Ui,j+1 −Ui,j−1
Ly Ui,j = 2∆y
+ O(∆x2 ) (5.1.9)

Ui+1,j+1 −Ui+1,j−1 −Ui−1,j+1 +Ui−1,j−1


Lyx Ui,j = 4∆x2
+ O(∆x2 , ∆y 2 )
CHAPTER 5. NUMERICAL METHODS 112

yielding the expressions (for ∆x = ∆y)


Ui+1,j+1 −2Ui+1,j +Ui+1,j−1 −Ui−1,j+1 +2Ui−1,j −Ui−1,j−1
Lxyy Ui,j = 2∆x3
+ O(∆x2 )
(5.1.10)
Ui+1,j+2 −2Ui+1,j +2Ui−1,j −Ui−1,j+2 +Ui+1,j−2 −Ui−1,j−2 2
Lyyx Ui,j = 8∆x3
+ O(∆x )

In the numerical solution of the amplitude equation we need also to use an expression
for ∂yyyy [147]

Ui,j+2 − 4Ui,j+1 + 6Ui,j − 4Ui,j−1 + Ui−2,j


Lyyyy Ui,j = + O(∆x2 ) (5.1.11)
h4
The Laplacian operator ∇4 U = Uxxxx + Uxxyy + Uyyyy was used in the numeric simu-
lations in Appendix B which can be written as [150]
 
Ui−2,j + 2Ui−1,j+1 − 8Ui,j + 2Ui−1,j−1
1
∇4 Ui,j = 4  +Ui,j+2 − 8Ui,j+1 + 20Ui,j − 8Ui,j−1 + Ui,j+2  (5.1.12)
h +2Ui+1,j+1 − 8Ui+1,j + 2U i+1,j−1 + Ui+2,j

5.2 Time Integration Method


After the spatial operators are discretized, we are left with a set of coupled differential
equations of the form
U = F(U, x, t) (5.2.1)

where F represents the spatially discretized right hand of Eqns. 5.1.1. We may then
use numerical methods for ordinary differential equations to step the solution forward
in time. In one dimension we used the one step Runge-Kutta method and the second
order Adams-Bashforth.
The Runge-Kutta method [148] uses in each time step, four auxiliary quantaties
(k1 , k2 , k3 , k4 ) which can be written as

k1 = ∆xF (Un , xi , tn )
k2 = ∆xF (Un + 21 k1 , xi + 12 ∆x, tn )
(5.2.2)
k3 = ∆xF (Un + 12 k2 , xi + 12 ∆x, tn )
k4 = ∆xF (Un + k3 , xi + ∆x, tn )
CHAPTER 5. NUMERICAL METHODS 113

and the value for Un+1 at location (i, j) is


1
Un+1 = Un + (k1 + 2k2 + 2k3 + k4 ) (5.2.3)
6

It can be shown that the numerical truncation error per step is of order ∆x4 [149].
In two dimensions the explicit second order Adams-Bashforth method [150] with a
fixed time step was used. For a time step of size h = ∆t the solution of 5.2.1 at
time-step n + 1 is in term of the solution at the previous two time steps, n and n − 1,
3h h
Un+1 = Un + Fn − Fn−1 + O(h3 ) (5.2.4)
2 2

where Fn = F (Un , x, tn ). To start the integration from the initial conditions we set
F−1 = F0 and take one step with the local second order method
U1 = U0 + hF0 + O(h2 ) (5.2.5)
Generally, for explicit methods the step size in t should be much smaller than the
step size in x and we must have h/∆x2 ≤ 0.5 in order for this method to work [151].
In most simulations we used the values h = 0.01 and ∆x = 0.39 which fulfills this
condition.

5.3 Numerical Simulation of Global Coupling


The numerical approximation for the global coupling in 2.1.12 was done in one di-
mension with Simpsons’s rule of integration [148]. In each time step the terms hui,
hvi and hsi were
Zl
1
hui = u(x)dx (5.3.1)
l
0
Rl
and similar expressions for v and s. The approximate value I˜ for u(x)dx can be
0
computed with the following algorithm
s0 = u0 + u2n
s1 = u1 + u3 + . . . + u2n−1
(5.3.2)
s2 = u2 + u4 + . . . + u2n
I˜ = ∆l
3
(s0 + 4s1 + 2s2 )
CHAPTER 5. NUMERICAL METHODS 114

where ∆l = l/2n. In two dimensions the hui was computed directly from the definition
of arithmetic average
n n
1 XX
hui = 2 ui,j (5.3.3)
n i=1 j=1

5.4 Front Transverse Instability Analysis


In order to compute the largest eigenvalue in the transverse instability analysis (see
in the previous chapter), M atlabr eigs routine has been used [152]. If A is a real
sparse matrix eigs(A) proceeds as follows [153]. First the matrix A is reduced to
Hassenberg form using Householder’s transformation method thus, A = PHP0 where
P is the unitary transformation matrix and H is Hessenberg matrix. 2
If A is a symmetric matrix the transform creates a tridiagonal matrix (which is not
the case for our model). Then the eigenvalues and the eigenvalues of the real upper
Hessenberg matrix are found by the iterative QR procedure. The QR procedure
involves decomposing the Hessenberg matrix into an upper triangular matrix and
unitary matrix. The method is as follows:

1. k=0 (k is the index of iteration)

2. Decompose Hk into Qk and Rk such Hk = Qk Rk where Hk is a Hessenberg


matrix or a triadiagonal matrix.

3. Compute Hk+1 = Rk Qk . The estimate of the eigenvalues equal to the values


of the leading diagonal of Hk

4. Check the accuracy of the eigenvalues. If the process has not converged, k =
k + 1; repeat from (2).
2
The Hessenberg matrix is similar to the triangular matrix except that it has non-zero elements
on the diagonals adjacent to the leading diagonal, for example the following matrix is 5x5 Hessenberg
matrix  
x x x x x
 x x x x x 
 
 0 x x x x  (5.4.1)
 
 0 0 x x x 
0 0 0 x x
CHAPTER 5. NUMERICAL METHODS 115

A one dimensional code with n = 1000 and dx = 0.06 has been used to check if the
zero eigenvalue that belongs to the translation mode is obtained and it was proved to
the order of 10−7 . This method of analysis can be applied in one dimension to check
the existence of front bifurcation in which a stationary front loses its stability to a
pair of counter-propagating fronts. So far this front bifurcation has not been found
in our model.
Chapter 6

Discussion

This dissertation as an interdisciplinary work has two main aspects. The first deals
with urban dynamics issues, mainly with residential segregation between two groups.
The significance of the of the model lies in the conceptual framework it introduces by
relating socio-spatial phenomena to dynamical system and pattern formation theories.
The second aspect is the contributions the model make to generic pattern formation
phenomena especially with bistable systems such as FHN model(Fits-Hugh Nagumo),
SH (Swift-Hohenberg) model, FCGL (Forced Complex Ginzburg-Landaw) model and
Lotka Volterra competition model with long-range diffusion terms (see Appendix B).
We will discuss these two aspects separately.

6.1 Urban Dynamics Modeling Aspects


The main goal of this research is to introduce a mathematical model based on non-
linear dynamics theory and demonstrate that it produces forms that qualitatively
are resembling urban internal patterns, and apply tentative interpretations to these
patterns. The justification for this approach is grounded in seeing the city as a self-
organizing system. Such system exhibits nonlinearity, instabilities, fractal structures
and chaos [2]. As in physicochemical systems and biological systems, the starting
point consists in choosing a model, that if does not describe in detail the mechanisms
of the systems, at least summarize its essential characteristics [3]. In chemistry and
biology and even in ecology and urban population dynamics, evolution equations are
often simply not known, and the use of reaction-diffusion models can often be justified
a fortiori.

116
CHAPTER 6. DISCUSSION 117

Urban Dynamics Phenomena Dynamical System Concept


Short range migration Diffusion terms
Non local migration Nonlocal migration terms
Stability of the socio-spatial structure Linear stability
Residential segregation Nonhomogenous solutions
Transition zones Front solutions
Uneven growth of a minority enclave Front transverse instability
Tipping point model Basins of attraction
Succession Moving front solution

Table 6.1: A mapping between urban dynamics phenomena and dynamical system
concepts

The model successfully reproduces residential segregation patterns such as strong,


variable and weak segregation between two distinct population. The model also repro-
duces neighborhood change phenomena such as tipping, transition zones and critical
minority enclave size. Table 6.1 sums up the mapping between urban dynamics phe-
nomena and dynamical system concepts as suggested by the model.

Cities do not develop on homogeneous planes but, rather in locations which have
rivers and valleys, and along coasts and around lakes [11]. There are areas of industry
and commerce in the city which act as natural barriers. Although this infrastructure
is certainly not determinant of the urban structure it shape how cities change and
grow over time. In some instances the location of minority areas near the downtown
may have been directly influenced by elements of urban structure. Thus, the fact that
a minority enclave has grown in a consistent, coherent, and particular direction may
well be related to significant geographic factors such as a major river or the associated
industrial land use around major railroads, which in turn influenced the direction of
residential expansion.
The proposed model can account for these effects of infrastructure by letting
the different parameters in the model to be spatial dependence. For instance we
can model a rich neighborhood with a small value of µ only in the vicinity of this
specific neighborhood. A more difficult task is modeling of streets or main highways
which define borders between neighborhoods and tracts in the city. One possibility
CHAPTER 6. DISCUSSION 118

is to use the White and Engelen’s idea [155] in their cellular automata simulation
model of Cincinnati. Their model was developed with three active land-use types
(housing, industry, and commerce) which can change, and three land use features
(railway, road and river). The main point is that cells occupied with fixed features
can affect the dynamics of other cells but cannot themselves be changed. We can
define a transportation system on the grid with a fixed value for zero population of
both groups. This transportation network will act as a barrier to diffusion so we
must include non-local migration terms in the model in order to have influence on
the dynamics.
Adding an infrastructure to the model is analogous to two-layer discrete models
considered recently by Benenson [38]. The first layer is the city’s infrastructure-
representing the urban housing properties; the second later represents individual cit-
izens and reflects their migratory movements. In this approach one has to be careful
not include too much details and mimic some regional models. This overcrowding of
details will mask the advantage of the continuum approach. Preliminary attempts to
include the infrastructure layer have been done but much work is still needed in or-
der to successfully integrate the two layers without breaking the continuum modeling
approach.
The practical-minded geographer might wonder why anyone would attempt to
explain spatial pattering without environmental heterogeneity when all environments
are so obviously heterogeneous. This is the same as asking why anyone would want to
explain population fluctuations in terms of interactions between different population
types when the urban environment so obviously fluctuates in time. The point is to rec-
ognize that heterogeneous spatial environments are not necessarily needed to produce
striking spatial irregularities in population densities. The proposed model suggests
the existing of Turing patterns for a specific set of parameters. These patterns exhibit
another kind of mechanism for segregation. As shown in this dissertation, nonlocal
migration can even make these segregation patterns stronger. This is one example
where the pattern formation theory ideas can initiate a new insight into socio-spatial
phenomenon.
The extended version of the model, which includes a direct interaction between
the two populations can shed light on the long time debate on the causes of residential
segregation. Adding a repulsion interaction between the two populations reduces the
CHAPTER 6. DISCUSSION 119

domain of stability of the mixed population states and increasing the intensity of
this repulsion interaction results with only strong segregation states. However as
the model predicts, taking into account only status considerations can also produce
strong segregation. Thus even without racial preferences, segregation will continue
to be part of the social landscape of the modern city. It is that complex interplay
between the two motives that makes various segregation patterns.
These new terms can simulate different kind of interaction between two groups by
changing their signs. With this recognition we can relate the model results with the
cellular automata models of Portugali and his coworkers [2]. Their heuristic Model
City simulates the interaction between two groups of individuals, Green and Blues,
each of which subdivided according to the preferences of its members into segregatives
and neutrals. Simulations with only segregative members of the two groups lead to a
segregated city as in our model. Increasing the proportions of neutrals in both groups
decreases the level of segregation. This is analogous to decreasing the strength of
direct interaction between u and v in our model. The difference is that in our model
we can still get segregation with neutrals only by indirect interaction through the
status variable.
The view of the city as a complex, dissipative dynamic system [2] implies the
evolution of the system towards asymptotic states or attractors. As a consequence the
actual long-term state of the system may differ significantly from the initial setting.
Thus planning process that fails to appreciate correctly the attractors of the system
may lead to undesired outcomes such as the case of Brasìlia.
The urban design, or the Pilot plan, of Brasìlia was conceived by Lucio Costa,
a modernist architect and a student of famous modernist Le Curbusier. The basic
idea of the Modernism is that unplanned development leads to chaos, whereas totally
centralized city planning could build an ideal city, which in turn leads an ideal society.
The planing of an ideal city was the most basic goal of Costa’s Pilot plan and one
of the most important goals was to create an egalitarian society, so that people of all
income levels could live together and would interact on a personal and classless level,
which means no residential segregation.
As Brasìlia was being built, vast numbers of workers came to the city. The workers
were to stay during the three years of construction in temporary campsites, which
would later disappear. But these workers ended up getting permission from the
CHAPTER 6. DISCUSSION 120

government to bring their families and stayed on [156]. The residents of the con-
struction camps were encouraged to move farther from the city into shanty towns,
which evolved into satellite towns of Brasìlia. These satellite towns grew more and
more self-sufficient of the Pilot Plan. Modernists predicted that these unplanned
developments would result in chaos; in reality they were surprised by the resultant
order which evolved. The satellite towns of Brasìlia exhibit the self-organizing nature
of the modern city. The modernist view that an ideal city would produce an ideal
society is clearly objectional as it did not take into account the human aspect of the
city and therefore failed.
Can we predict or control the urban behavior? Ideally, we would like a planning
process to determine the initial system in a unique way, so that the system will
evolve in time towards the desired state assuming that the system is deterministic.
In reality the limited information urban planners have does not allow a complete
determination of the initial state. In a simple state having few attractors, there is
still a good chance that the system will evolve towards the desired state. However
in a dissipative nonlinear system the insufficient knowledge of the initial state makes
the final state completely uncertain. One way to overcome this inherent difficulty
of urban planning is to make it flexible enough so that during its evolution, the
system can be “kicked out” of trajectories that lead to states. Brasìlia might be an
example of a rigid planning that did not leave enough room for such corrections. To
perform such corrections, planners have to be aware of the nonlinear nature of the
urban dynamics, and that there are more plausible for these corrections acts such near
bifurcation points where a small difference in the value of a control parameter can
result in different attractors or near these points, a small perturbation can shift the
system from one base of attraction to another. Mathematical modeling that focuses
on restricted aspects of the urban environment and captures the essential instabilities
and bifurcations, may help achieving this goal.
Another direction of future study is to introduce in the model exogenous force.
There are many exogenous events and forces that influence behavior; these may be
a consequence of economic and political decisions (relevant, or irrelevant, local or
remote, and of a capricious nature). For instance the government may develop special
policies to guarantee coexistence of residence [51]. This can be modeled by terms like
Ei (x, t) in the model equations, where Ei (x, t) are exogenous “shocks”. For example
CHAPTER 6. DISCUSSION 121

we can rewrite the equation for v

vt = αv − v 2 − vs + E2 (x, t) + ∇2 v − δ∇2 s (6.1.1)

where E2 (x, t) is in-migration of v-residents to the urban area which do not depend
on the population density or on status, but on space and time only.
The proposed model is deterministic. Such models do not account for the effects
of noise despite the fact that it is always present in actual population dynamics
and arises from different sources, such as intrinsic stochasticity associated with the
dynamics of individuals and the random variability of the environment. In a recent
paper [157], it was shown that consideration of noise in a classical Lotka-Volerra
model of two-species competition
xt = µ(1 − x − βy) + D∇2 x + fx (x, y) + ξx (t)
(6.1.2)
yt = µ(1 − y − βx) + D∇2 y + fy (x, y) + ξy (t)

where x and y are the population densities and fi (x, y)ξi (t) (i = x, y) model the the
contribution of random forces. They also account for a non-stationary environment by
allowing β to be β = 1 + ² + αsin(ω0 t). The significance of this model is that system’s
response to a periodic force may be enhanced by the presence of the noise (stochastic
resonance). They also showed the appearance of spatial patterns which do not arise
in the deterministic version of 6.1.2. As our urban model can be approximated to
the classical Lotka-Volterra competition model (see chapter 2 for more details of this
approximation), this constructive role of noise can be taken into account in a similar
way. This will enable a better comparison with Portugali’s CA models which can be
considered a stochastic CA model, as the transition rules are probabilistic [42].

6.2 Pattern Formation Aspects


The proposed model shows a generic mechanism of large amplitude labyrinthine pat-
terns formation in bistable systems. Several other pattern-forming systems, such as
gas discharges, liquid crystals, and nonlinear optics, also exhibit bistable regimes
[3, 145]. Two situations leading to bistability are a pitchfork bifurcation (generally
imperfect) like in the proposed urban model or the SH-model, from one HSS (homoge-
neous stable state) toward two different HSS. The second is an hysteresis loop formed
CHAPTER 6. DISCUSSION 122

by two back-to-back saddle-node bifurcation, as encountered in the FHN-model and


in the FIS system. It was shown that the linear front instability is a sufficient ingre-
dient in obtaining labyrinths, however labyrinths can also be formed when the front
is linear stable to transverse perturbations due to a finite wavenumber instability of
the unstable branch. We refer to this instability as nonlinear front instability. The
nonlinear front instability is induced for perturbations with an amplitude larger than
a critical value This new mechanism has been found in several models such as Swift-
Hohenberg (SH), and Forced Complex Ginzburg-Landau model (FCGL) (see [158]).
The existence of this threshold suggests that a competition mechanism is in place.
The competition is between the linear stability of the front, which acts to smooth the
boundaries, and the growth mode (k > 0) of the UHSS state that destabilizes the
front. If the amplitude of the initial perturbations is below the threshold, the linear
stability “wins” and the front remains planar. In the other case the perturbations
will grow and the front will not remain anymore planar.
One still open question is why in the FHN model only linear stability of the front is
the only mechanism acting whereas in the other models both mechanisms are acting.
Linear stability of the front to transverse perturbations has been performed nu-
merically and the results were confronted with numeric simulations of the full model
equations. This method is a computer time demanding task as one needs to find
eigenvalues of a huge sparse matrix and is sensitive to a choice of parameters.
In recent papers [3, 140, 139, 146], the coupling of the active spatial modes with a
quasi-neutral homogeneous mode generated by the bifurcations inducing the bistabil-
ity, has been studied analytically in 1D and 2D. The analysis relies on the fact that
the homogeneous perturbations about the uniform states quasi-natural, so the zero
mode is active and must be taken into account, since it will affect the amplitude and
the stability of the patterns. The concentration field C can be approximated by a
linear superposition of these active modes:
" m
#
X
c = e1 A0 + e2 Ai exp[iqi · r] + c.c (6.2.1)
i=1

where (|qi | = qc , e1 and e2 are the eigenvectores of the Jacobian matrix of the
homogeneous and the spatial systems. In 1D, it can give rise to a rich variety of
subharmonic and superharmonic patterns. In 2D one can get coexistence of H0 and
CHAPTER 6. DISCUSSION 123

Hπ hexagons and stripes. These patterns are an outcome of the interaction of the
diffusive instability of one of the HSS and the homogeneous bifurcations. As the
urban model shows, a different mechanism can work in bistable systems when the
unstable branch is also Turing unstable. This new mechanism can lead to a similar
patterns even if the the two HSS are Turing stable. The amplitude of the mixed
Turing homogeneous patterns is on the order of the difference in amplitudes between
the two HSS and is thus usually quite large.
The significance of this new mechanism is that it takes into consideration the
influence of an unstable state on the dynamics. This makes the whole picture more
complicated because in bistable systems the patterns can include different spatial
modes from stable and unstable branches solutions. In principle we think that the
mathematical description of this interaction has to be based on an equation like 6.2.1
and as was shown in chapter 4, the shape of the growth curve determines how many
modes interact with the zero mode.
To conclude, this study explores the implications of the instabilities and non-
linearity in urban pattern formation processes. The main concern is with internal
urban structures and processes such as residential segregation, transition zones and
neighborhood change. We try to give to these geographic concepts a mathemati-
cal interpretation. The importance of the study lies in the conceptual framework
it introduces by relating socio-spatial phenomena to dynamical systems and pattern
formation theory. Thus, this study can be a good starting point for more elaborate
models that will try to deep our understanding in the variety of urban forms.
Bibliography

[1] Nicolis, G. (1989). Physics of far from equilibrium systems and self-organization,
in The New Physics, Ed. Paul Davis, 316-347. Cambridge University Press.

[2] Portugali, J. (2000). Self-Organization and the city. Berlin: Springer-Verlag.

[3] De Wit, A. (1999). Spatial Patterns and spatiotemporal dynamics in chemical


systems. Advances in Chemical Physics, 109, 435-513.

[4] Cross, M. C. and Hohenberg, P. C. (1993). Pattern Formation outside of equi-


librium. Review of Modern Physics, 65, 3, 851-1112.

[5] Anderson, R. S. (1990). Eolian ripples as examples of self-organization in geo-


morphological systems. Earth-Sciences Reviews, 29, 77-96.

[6] Werner, B. T. (1995). Eolian dunes: Computer simulations and attractor inter-
pretation. Geology, 23, 12, 1107-1110.

[7] Harrison, L. G. (1987). Mini Review, What is the status of reaction-diffusion


thirty-four years after Turing? Journal of Therotical Biology, 125, 369-384.

[8] Holmes, E. E., Lewis, M. A., Banks, J. E., & Veit, R. R. (1994). Partial differen-
tial equations in ecology: Spatial interactions and population dynamics. Ecolgy,
75, 1, 17-29.

[9] Okubo, A. (1980). Diffusion and Ecological Problems: Mathematical Models..


Biomathematics Volume 10. Springe-Verlag: Berlin Heidelberg.

124
BIBLIOGRAPHY 125

[10] Kempen, R., & Weesep, J. (1998). Ethnic residential patterns in Dutch cities:
Backgrounds, shifts and consequences. Urban Studies, 10, 1813-1833.

[11] Clark, W. A. V. (1986). Residential segregation in American cities: a review and


interpretation. Population Research and Policy Review, 5, 95-127.

[12] Taylor, C., Gorard, S., & Fitz, J. (2000). A reexamination of segregation indices
in terms of compositional invariance. Social Research Update, 30, 1-9.

[13] Massey, D. S., & Fischer, M. J. (2000). How segregation concentrates poverty.
Ethnic and Racial Studies, 13, 4, 670-691.

[14] Morrill, R. (1995). Racial segregation and class in a liberal metropolis. Geograph-
ical Analysis, 27, 1, 22-41.

[15] White, P. (1998). The settlement patterns of developed world migrants in Lon-
don. Urban Studies, 35, 10, 1725-1744.

[16] Lever, W. F., & Paddison, R. (1998). Special issue: Ethnic segregation in cities:
New forms and explanations in a dynamic world. Urban Studies, 35, 10.

[17] Lynn, L. E., Jr., & McGeary, M. G. H., Ed. (1990). Inner-City Poverty in the
United States. National Academy Press. Washington D.C .

[18] Massey, D., & Denton, N. (1988). The dimensions of residential segregation.
Social Forces, 76, 281-315.

[19] Cutler, D. M., Glaeser, E. L., & Vigdor, J. L. (1997). The rise and decline of the
American ghetto. NBER Working Paper Series, working paper 5881.

[20] Massey, D., & Denton, N. (1993). American Apartheid: Segregation and Making
of the Underclass. Harward University Press.

[21] St.John, C. & Clymer, R. (2000). Racial residential segregation by level of so-
cioeconomic status. Social Science Quarterly, 81, 3, 701-715.
BIBLIOGRAPHY 126

[22] Schwirian, K. P. (1983). Models of neighborhood change. Annual Review of


Sociology, 9, 83-102.

[23] Park, R. E. (1925/1974). The city: suggestions for the investigation of human
behavior in the urban environment in: R. E. Park, E. W. Burgess and R. D.
McKenzie (eds.) The City, pp. 47-62. Chicago/London: University of Chicago
Press.

[24] Duncan,O. D., & Duncan, B. (1957). The Negro Popualtion of Chicago. Chicago:
University Chicago Press.

[25] Zang, X. (2000). Ecological succession and Asian immigrants in Australia. In-
ternational Migration, 38, 1, 109-117.

[26] Massey, D. S. (1985). Ethnic residential segregation: A theoretical synthesis and


empirical review. Sociology and Social Research, 69, 315-350.

[27] Massey, D. (1979). Effects of socioeconomic factors on the residential segregation


of blacks and Spanish American in U.S. urbanized areas. American Sociological
Review, 44, 1015-1022.

[28] Nicolis, G. & Prigogine, I. (1989). Exploring Complexity.New York: W. H. Free-


man and Company.

[29] Allen, P. M. (1982). Evolution, modelling, and design in a complex world. Envi-
ronmental and Planning B, 9, 95-111.

[30] Straussfogel, D. (1991). Modeling Suburbanization as a evolutionary system dy-


namic. Geographical Analysis, 23, 1, 1- 24.

[31] Ball, P. (1999). The Self-Made Tapestry: Pattern Formation in Nature. Oxford:
Oxford University Press.

[32] Portugali, J., Benenson, I.,& Omer, I. (1997). Spatial cognitive dissonance and
socio-spatial emergence in a self-organizing city. Environment and Planning B,
24, 263-285.
BIBLIOGRAPHY 127

[33] Allen, P. M. (1985). Towards a new synthesis in the modelling of evolving complex
systems. Environmental and Planning B: Planning and Design, 12, 65-84.

[34] Allen, P. M. (1997). Cities and regions as self-organizing systems: Models of


complexity. Gordon and Breach Science Publications.

[35] Allen, P. M (1999). Popualation growth and Envirionment as a self-organizing


system. Discrete Dynamics in Nature and society, 3, 81-108.

[36] Portugali, J. & Benenson, I. (1995). Artificial planning experience by means


of heuristic sell-space model: Simulating international migration in the urban
process. Environment and Planning A, 27, 1647- 1665.

[37] Clarke, K.C, Hoppen, S., & Gaydos, L. (1996). A self modifying cellular automa-
ton model of historical urbanization in the San-Francisco Bay area. Environmen-
tal and Planning B, 24, 247-261.

[38] Benenson, I. (1999). Modelling population dynamics in the city: from a regional
to a multi-agent Approach. Discrete Dynamics in Nature and Society, 3, 149-170.

[39] Gilbert, N., & Troitzsch, K. G. (1999). Simulation for the Social Scientist. Open
University Press, Buckingham, Philadelphia.

[40] Gaylord, R. G., & D’Andria, L. J. (1998). Simulating Society. A Mathematica


Toolkit for Modelling socio-economic behaviour. Springer-Verlag, New York.

[41] Itami, R. A. (1994). Simulation spatial dynamics: cellular automata theory.


Landscape and Urban Planning, 30, 27-44.

[42] Portugali, J., Benenson, I., & Omer, I. (1994). Sociospatial residential dynamics:
Instability within a self-organizing city. Geographical Analysis, 26, 321-340.

[43] Beckmann, M. J., & Puu, T. (1990). Spatial structures. Springer-Verlag, Berlin.
BIBLIOGRAPHY 128

[44] Gurtin, M. E. (1974). Some mathematical models for population dynamics that
lead to segregation. Quarterly of Applied Mathematics, XXXII, 1-9.

[45] Grodzins, M. (1957). Metropolitan segregation. Scientific American, 197, 4, 33-


41.

[46] Taeuber, A. F. (1965). Negrows in Cities. Chicago: Aldine.

[47] O’Neill, W. D. (1981). Estimation of logistic growth model describing neighbor-


hood change. Geographical Analysis, 13, 391-397.

[48] Shigesada, N., & Kawasaki, K. (1997). Biological Invasions: Theory and Practice.
Oxford University Press.

[49] Murray, J. D. (1989). Mathematical Biology. Springer-Verlag, Berlin.

[50] Zhang, W. B. (1988). Pattern formation of an urban system. Geographical Anal-


ysis, 20, 75- 84.

[51] Zhang, W. D. (1989). Coexistence and separation of two residential groups - An


interactional spatial dynamics approach. Geographical Analysis, 21, 91-102.

[52] Hosono, Y. (1989). Singular perturbation analysis of traveling waves for diffu-
sive Lotka-Volterra competition models. Numerical and Applied Mathematics, C.
Brezinski (editor) 687-692.

[53] Jorné, J. & Carmi, S. (1977). Liapunov stability of the diffusive Lotka-Volterra
Equations. Mathematical Biosciences, 37, 51-61.

[54] Hagberg, A., Meron, E., Rubinstein, I., & Zaltzman, B. (1996). Controlling
domain patterns far from equilibrium. Physical Review Letters, 76, 427-430.

[55] Hagberg, A. & Meron, E. (1997). The dynamics of curved fronts: Beyond geom-
etry. Physical Review Letters, 78, 6, 1166-1169.
BIBLIOGRAPHY 129

[56] Newell, A. C. (1974). Envelope Equations. In: Nonlinear Wave Motion. American
Mathematical Society, Providence.

[57] Morrill, R. L. (1965). The Negro ghetto: Problems and alternatives. Geographical
Review, 55, 339-361.

[58] Morrill, R. L. (1970). The shape of diffusion in space and time. Economic Geog-
raphy, 259-268.

[59] Rose, H. M. (1972). The spatial development of black residential subsystems.


Economic Geography, 48, 43-66.

[60] Woods, R. I. (1981). Spatiotemporal models of ethnic segregation and their im-
plications for housing policy. Envirionment and Planning A, 13, 1415-1433.

[61] Hansell, CH. R., & Clark, W. A. V. (1970). The expansion of the negro ghetto in
Milwaukee: A description and simulation model. Tijdschrift voor Economische
en Sociale Geografie, 61, 267-277.

[62] Allen, P. M., & Sanglier, M. (1981). Urban evolution, self-organization, and
decision-making. Environmental and Planning A, 13, 167-183.

[63] Sanglier, M., & Allen, P. M. (1989). Evolutionary models of urban systems: an
application to the Belgian provinces. Environmental and Planning A, 21, 477-
498.

[64] Pumain, D., Saint-Julien, Th., & Sanders, L. (1987). Application of a dynamic
urban model. Geographical Analysis, 19, 2, 152-166.

[65] Hotelling, H. (1978). A mathematical theory of migration. Environmental and


Planning A, 10, 1223-1239.

[66] Rosser, J. B. (1980). The dynamics of ghetto boundary movement and ghetto
shape. Urban Studies, 17, 231-235.
BIBLIOGRAPHY 130

[67] Ishikawa, H. (1980). A new model for the population density distribution in an
isolated city. Geographical Analysis, 12, 3, 223- 225.

[68] Meron, E. (1999). Self organization in interface dynamics and urban develop-
ment, Discrete Dynamics in Nature and Society, 3, 125-136.

[69] Zhang, W. D. (1990). Stability versus instability in urban pattern formation.


Occasional Paper Series on Socio-Spatial Dynamics, 1, 1, 41-56.

[70] Yizhaq, H., & Meron, E. (2002). Urban segregation as a nonlinear phenomenon.
Non-linear Dynamics, Psychology, and Life Sciences, 6, 3, 269-283.

[71] Harris, D. R. (1999). ’Property values drop when blacks move in, because . . . ”:
Racial and socioeconomic determination of neighborhood desirability. American
Sociological Review, 64, 461-479.

[72] Clark, W. A. V. (1980). Residential mobility and neighborhoods change: Some


implications for racial residential segregation. Urban Geography, 1, 2, 95-107.

[73] May. R. M. (1974). Stability and complexity in model ecosystems. 2nd Ed. Prince-
ton Univer-sity Press, Princeton, New-Jersey.

[74] Golubitsky, M. & Schaeffer, D. G. (1985). Singularities and groups in bifurcation


theory, Vol. 1.. New York: Springer-Verlag.

[75] Kapral, R. (1995). Pattren formation in chemical systems. Physica D, 86, 149-
157.

[76] Turing, A. (1952). The chemical basis of morphogensis. Philosophical Transac-


tions of the Royal Society London B, 237, 37.

[77] Edelstein-Keshet, L. (1988). Mathematical Models in Biology. New York: Ran-


dom House.
BIBLIOGRAPHY 131

[78] Ouyang, Q., & Swinney, H. L. (1991). Transition from a unifrom state to hexag-
onal and striped Turing patterns. Nature, 352, 610-611.

[79] Epstein, I. R., & Showalter, K. (1996). Nonlinear chemical dynamics: Oscilla-
tions, patterns, and chaos. Journal of Physical Chemistry, 100, 13132-13147.

[80] Bascompte, J., & Solé, R. V. (1995). Rethinking complexity: modeling spa-
tiotemporal dynamics in ecolgy. Trends in Ecological Evolution, 10, 361-366.

[81] Segel, L. A., & Jackson, J. L. (1972). Dissipative structure: An expalnation and
ecological example. Journal of Theortecial Biology, 37, 545-559.

[82] Levin, S. A., & Segel, L. A. (1985). Pattern generation in space and aspect.
SIAM Review, 27, 45-67.

[83] White, K. A. J., & Gilligan, C. A. (1998). Spatial heterogeneity in three species,
plant parasite hyperparasite systems. Phil. Trans. R. Soc. Lond. B, 353, 543-557.

[84] Kareiva, P. (1990). Population dynamics in spatially complex environment: The-


ory and data. Philosophical Transactions of the Royal Society of London B, 330,
175-190.

[85] Walgraef, D. (1996). Spatio-Temporal Pattren Formation. New York: Springer-


Verlag.

[86] Engelhardt, R. (1994). Modeling Pattern formation in Reaction-Diffusion Sys-


tems. Master dissertation: Department of Chemistry, University of Copenhagen,
Denmark.

[87] Strogatz, S. H. (1994). Nonlinear Dynamics and Chaos. Perseus Publishing:


Cambridge, Massachusetts.
BIBLIOGRAPHY 132

[88] Hastings, A. (1978). Global Stability in Lotka-Volterra Systems with Diffusion.


Jornal of Mathematical Biology, 6, 163-168.

[89] Murray, J. D. (1975). Nonexistence of wave solutions for the class of reaction-
diffusion equations given by the Volterra interacting popualtion equations with
diffusion. Journal of Theoretical Biology, 52, 459-469.

[90] Guest, A. M., & Weed, J. A. (1976). Ethnic residential segregation: patterns of
change, American Journal of Physics, 81, 1088-1112.

[91] Daley, P. O. (1998). Black Africans in Great Britain: Spatial concentration and
segregation. Urban Studies, 10, 1703-1724.

[92] Deskins, D. R. (1981). Morphogensis of a black ghetto. Urban Geography, 2, 2,


95-114.

[93] Downs, A. (1981). Neighborhoods and urban development. Washington, D. C. :


The Brookings Insititution.

[94] Schwab, W. A., & Marsh, E. (1980). The tipping-point model: Prediction of
change in the composition of Cleveland, Ohio, neighborhoods 1940 -1970. En-
vironment and Planning A, 12, 385-398.

[95] Wolf, E. P. (1963). The tipping point in racially changing neighborhoods. Journal
of American Insititute of Planners, XXIX, 217-222.

[96] Goring, J. M. (1978). Neighborhood tipping and racial transition: A review of


social science evidence. Journal of American Planners Association, 45, 506-514.

[97] Omer, I. (1995). Ethnic residential segregation as a structuration process: case


study the Arab in Jaffa. Ph.D. Dissertation, Tel-Aviv University. (In Hebrew)
BIBLIOGRAPHY 133

[98] Gandhi, A., et. al (1998). ’Critical slowing down’ in time-to-extinction: an ex-
ample of critical phenomenon in ecology. Journal of Theoretical Biology, 192,
363-376.

[99] Gandhi, A., et. al (1999). Nucleation and relaxation from meta-stability in spatial
ecological models. Journal of Theoretical Biology, 200, 121-146.

[100] Fife, P., & Mcleod, J. B. (1977). The approach of solutions of non-linear diffu-
sion equations to travelling front solutions. Arc. Rat. Mech. Anal. 65, 335-361.

[101] Middya, U., & Luss, D. (1995). Impact of global interaction on pattern forma-
tion on a disk. Journal of Chemical Physics, 102, 12, 5029-5036.

[102] Middya, U. et al. (1993). Pattern selection in controlled reaction-diffusion sys-


tems. Journal of Chemical Physics, 98, 4, 2823-2836.

[103] Middya, U., Luss, D., & Sheintuch, M. (1994). Impact of global interaction and
symmetry on pattern selection and bifurcation. Journal of Chemical Physics,
101, 6, 4688-4696.

[104] Clark W. A. V. (1970). Measurement and expalnation in intra-urban residential


mobility. Tijdschrift voor Economische en Sociale Geografie, 61, 49-57.

[105] Anderson, R. (1998). Socio-spatial dynamics: Ethnic divisions of mobility and


housing in post-palme Sweden. Urban Studies, 35, 3, 397-428.

[106] Ellen, I. G. (1997). Race-Based Neighborhood Projection: A proposed frame-


work for understanding new data on racial segregation. Taub Urban Research
Center.

[107] Rothschild, B. J., & Ault, J. S. (1996). Popualtion-dynamic instability as a


cause of patch structure. Ecological Modeling, 93, 237-249.
BIBLIOGRAPHY 134

[108] Brown, D., & Rothery, P. (1993). Models in Biology: Mathematics, Statistics
and Computing. John Wiley & Sons.

[109] Farely, R., Jackson, T., & Reeves, K. (1994). Stereotypes and segeragtion:
Neighborhoods in the Detroit area. American Journal of Sociology, 100, 750-
780.

[110] Hwang, S., & Murdock, S. H. (1998). Racial attraction or racial avoidence in
American suburbs. Social Forces, 77, 2, 541-566.

[111] Darroch, A. G., & Marston, W. G. (1971). The social class basis of ethnic
segeragtion : The Canadian case. American Journal of Sociology, 77, 491-510.

[112] Smith, R. A. (1993). Creating stable racially integrated communities: A review.


Journal of Urban Affairs, 15, 2, 115-140.

[113] Schelling, T. C. (1978). Micromotives and macrobehavior. New York: W .W.


Norton & Company, Inc.

[114] Clark, W. A. V. (1988). Understanding residential segregation in American


cities: Interpreting the evidence. Population Research and Policy Review, 7,
113-121.

[115] Galster, G. (1989). Residential segregation in American cities: A further re-


sponse to Clark. Population Research and Policy Review, 8, 181-192.

[116] Sims, M. (1999). High-status residential segregation among racial and ethnic
groups in five metro areas, 1980-1990. Social Science Quarterly, 80, 3, 556-573.

[117] Massey, S. D., & Fischer, M. J. (1999). Does rising income bring integration?
New results for Blacks, Hispanics, and Asian in 1990. Social Science Research,
28, 316-326.
BIBLIOGRAPHY 135

[118] White, M. J., & Sassler, S. (2000). Judging not only by colour: ethnicity, na-
tivity, and neighborhood attainment. Social Science Quarterly, 81, 4, 997-1013.

[119] Steiness, D. N. (1977). Alternatives models of neighborhood change. Social


Forces, 55, 4, 1043-1057.

[120] Anas, A. (1980). A model of residential change and neighborhood tipping. Jour-
nal of urban Economics, 7, 358-370.

[121] Massey, D. S. 1981. Social class and ethnic segregation: A reconsideration of


methods and conclusions. American Sociological Review, 46, 641-650.

[122] Massey, D. S., & Denton, N. A. (1985). Spatial assimilation as a socioeconomic


outcome. American Sociological Review, 50, 94-106.

[123] Frey, W. H. (1979). Central city white flight: Racial and nonracial causes.
American Social Review, 44, 425-448.

[124] Galster, G. C. (1987). Residential segregation and interracial economic dispar-


ities: A simulta-neous-equations approach. Journal of Urban Economics, 21,
22-24.

[125] Denton, N. A., & Massey, D.S. (1989). Residential segregation of Blacks, His-
panics, and Asians by socioeconomic status and generations, Social science Quar-
terly, 69, 797-817.

[126] Galster, G. (1988). Residntial segregation in American cities: A contrary review.


Population Research and Policy Review, 7, 93-112.

[127] Clark, W. A. V. (1989). Residential segregation in American cities: Common


ground and differences in interpretation. Population Research and Policy Review,
8, 193-197.
BIBLIOGRAPHY 136

[128] Massey, D. s, & Fong, E. (1990). Segregation and neighborhood quality: Blacks,
Hispanics, and Asians in the San Francisco metropolitan area. Social Forces, 69,
1, 15-32.

[129] Massey, D. S., & Eggers, M. L. (1990). The ecolog of inequality: Minorities
and concentration of poverty, 1970-1980. American Journal of Sociology, 95,
1153-1188.

[130] Denton, N. A., & Massey, D.S. (1991). Patterns of neigborhood transition in a
multiethnic world: U.S. metropolitan areas. Demography, 28, 41-63.

[131] Boswell, T. D., & Cruz-Báez, A. D. (1997). Residential segregation by socioe-


conomic class in metropolitan Miami: 1990. Urban Geography, 18, 6, 474-496.

[132] Nicolis, G. (1995). Introduction to Nonlinear Science. Cambridge: Cambridge


University Press.

[133] Busse, F. (1978). Non linear properties of thermal convections. Pep.Progr.Phys.,


41, 1921-1967.

[134] Lowe, M., & Gollub, J. P. (1985). Pattern selection near the onset of convection:
The Eckhaus insatbility. Physical Review Letters, 55, 23, 2575-2578.

[135] Dufiet, V., & Boissonade, J. (1992). Numerical studies of Turing patterns se-
lection in a two-dimensional systems. Physica A, 188, 158-171.

[136] Dufiet, V., & Boissonade, J. (1991). Conventional and unconventional Turing
patterns. Journal of Chemical Physics, 96, 1, 664-673.

[137] Sakaguchi, H. (1991). Zigzag instability and reorientation of roll pattern in the
Newell-Whitehead equation. Progress in Theoretical Physics, 86, 4, 759-763.

[138] Borckmans, P., De Wit, A., & Dewel, G. (1992). Competition in ramped Turing
structures. Physica A, 188, 137-157.
BIBLIOGRAPHY 137

[139] Hilali, M’F., Dewel, P., & Borckmans, P. (1996). Subharmonic and strong res-
onances through coupling with zero mode. Physics Letters A, 217, 263-267.

[140] Métens, S., et al. (1997). Pattern selection in bistable systems. Europhysics
Letters, 37, 2, 109-114.

[141] Hagberg, A. (1994). Fronts and Pattrens in Reaction Diffusion Equations. A


disertation submitted to the faculty of the graduate interdisciplinary, program
in applied mathematics. The University of Arizona.

[142] Aranson, I. S., Tsimring, L. S., & Vinokur, V. M. (1999). Hexagons and interface
in a vibrated granular layer. Physical Review E, 59, 1327-1330.

[143] Lee, K., et al. (1993). Experimental observation of self-replicating spots in a


reaction-diffusion system. Nature, 369, 215-218.

[144] Trefethen, L. N., (2000). Spectral Methods in Matlab. Siam, Philadelphia, PA.

[145] Davis, P. W., et al. (1998). Dividing blobs, Chemical flowers, and pattrened
islands in reaction-diffusion systems. Journal of Physical Chemistry A, 102, 43,
8236-8247.

[146] Dewel, G., et al. (1995). Resonant pattrens through coupling with zero mode.
Physical Review Letters, 74, 4667-4650.

[147] Lam, C. (1994). Applied Numerical Methods for Differential Equations. Singa-
pure: Prentice Hall.

[148] Kreyszig, E. (1988). Advanced Engineering Mathmatics. Sixth edition. New


York: John Wiley & Sons.

[149] Collatz, L. (1966). The Numerical Treatment of Differential Equations. 3rd ed.
New York: Springer.
BIBLIOGRAPHY 138

[150] Fausett, L. V. (1999). Applied Numerical Analysis Using Matlab. New jersy:
Prentice Hall.

[151] Farlow, S. J. (1982). Partial Diffential Equations for Scientists & Engineers.
New York: John Wiley & Sons

[152] Pärt-Enander, E., & Sjöberg, A. (1999). The Matlab Handbook. London: Pren-
tice Hall.

[153] Lindfield, G., & Penny, J. (1995). Numerical Methods Using Matlab. New York:
Ellis Horwood.

[154] White, R., & Engelen, G. (1997). Cellular automata as the basis of integrated
dynamic regional modelling. Environment and planning B: Planning and Design.
24, 235-246.

[155] White, R., & Engelen, G. (1997). Cellular automata as the basis of integrated
dynamic regional modelling. Environment and planning B: Planning and Design.
24, 235-246.

[156] Wright, C. W., & Turkienicz, B. (1988). Brasília and the ageing of modernism.
Cities, November, 347-364.

[157] Vilar, J. M. G., & Solé, R. V. (1998). Effects of Noise in Symmetric two species
competitions. Physical Review Letters, 18, 4099-4102.

[158] Yochelis, A., et al. , in final etiding.

[159] White, M. J. (1987). American Neighborhoods and Residential Differentiation.


New York: Russell sage Foundation.

[160] Levin, S. A., & Pacala, W. (1997). Theories of simplification and scaling of
spatially distributed processes. In Spatial Ecolgy: The role of space in population
dynamics and interspecific interactions. Edited by D. Tilman and P. Kareiva.
Princton University Press: Prinston, New Jersi.
BIBLIOGRAPHY 139

[161] Shigesada, N., Kawasaki, k., & Teramoto, E. (1979). Spatial segregation of
interacting species. Jounal of Theoretical Biology, 79, 83-99.

[162] Levin, S. A. (1974). Dispersion and population interacions. American Naturalist,


108, 207-228.

[163] Mimura, M., & Kawasaki, K. (1980). Spatial segregation in competitive


interaction-diffusion equations. Jounal of Mathematical Biology, 9, 49-64.

[164] Solé, R. V., Bascompte, J., & Valls, J. (1992). Stability and complexity of
spatially extended two-speices competition. Jounal of Theoretical Biology, 159,
469-480.

[165] Bertsch, M., Gurtin, M. E., Hilhorst, D., & Peletier, L. A. (1985). On interacting
populations that disperse to avoid crowding: preservation of segeragtion. Journal
of Mathematical Biology, 23, 1-13.

[166] Cohen, D., & Murray, J. D. (1981). A generalized diffusion model for growth
and dispersal in a population. Journal of Mathematical Biology, 12, 237-249.

[167] Travis, J. M. J., & French, R. F. (2000). Dispersal functions and spatial models:
expanding our dispersal toolbox. Ecology Letters, 3, 163-165.
Appendix A

Urban Hierarchy

This appendix presents a brief overview of geographic concepts related to the internal
structures of a metropolitan area which are frequently used in geographic studies.
The outline of geographical hierarchy of the census is: blocks that nest to block
groups which nest to tracts, which in turn nest into counties, and several counties
comprise the SMSA. This structure indicates the heterogeneous character of an urban
environment and relates to the infrastructure of the city like main roads, rivers etc.
The Standard Metropolitan Area (SMSA) classification is one of a large population
nucleus, together with adjacent communities which have a high degree of economic
and social integration with that nucleus [159]. The SMSA is consists of several cities
and counties. For example in 1980 census inn U.S.A there were 318 SMASs in fifty
counties.
A city can be divided farther to tracts which are small, and designed to be homoge-
neous with respect to population characterizes, economic status, and living condition.
Tracts generally have between 2500 and 8000 residents. Boundaries of tracts should
follow permanent and easily recognized lines in order to facilitate the allocation of
house number ranges as well as enumeration.
Blocks are the small geographic unit for which information is available, but they
suffer from widely varying sizes, and need not exhaustively covered all the SMSA.
So the tracts are the most useful geographic unit for the study of small area social
phenomena within urban areas.

140
Appendix B

Spatial patterns in LV competition


model

It is obvious that the entire nature of movement and the scale of interac-
tions play fundamental roles in the persistence of species and the mainte-
nance of biodiversity.

Levin and Pacala 1997 (p.276) [160]

In this appendix we will study the effect of adding long-range diffusion terms to the
classic symmetric Lotka-Volterra (LV) competition model on the existence of Turing
instabilities of the different stationary uniform solutions. The ecological implications
of these terms will be also shortly discussed. In B.1 we give a brief description about
different approaches in ecology to model spatial heterogenous environment. In B.2
we introduce the the LV model with long-range diffusion terms. In B.3 we analyze
the Turing instability of the model and in B.4 we show results of numeric simulations
of the Turing patterns. In B.5 we show that fronts connected the bistable states may
be unstable to transverse perturbations. This instability may lead to labyrinthine
patterns. The last section B.6 discusses the ecological aspects of these instabilities.

B.1 Introduction
Spatial heterogeneity (commonly called patchiness, patch structure or patch distribu-
tion) is a characteristic feature of many ecological systems and has a profound effects
on the dynamics of invasion, segregation and persistence of populations [83, 107]. The

141
APPENDIX B. SPATIAL PATTERNS IN LV COMPETITION MODEL 142

most fundamental distinction between different models of heterogenous environments


is the manner in which spatial dimension is represented. Three main approaches ex-
ist: island models, stepping-stone models and continuum models [84]. The island
models imagine the world subdivided into large collection of patches and the coupling
is via a common pool of dispersers. Such models do not have an explicit spatial di-
mension, and dispersal rates refer to the fraction of individuals that move and not to
the distances individuals move.
The stepping-stone model divide the world into patches that have fixed coordi-
nates, and can be used to contrast the consequence of long-range versus short-range
dispersal. The continuum approach which is used here typically takes the form of
PDE or more specially of reaction diffusion equations. In most reaction-diffusion
models the environment is assumed homogeneous, and the questions of interest con-
cern what types of spatiotemporal patterns in population density emerge as a result
of coupling dispersal with local dynamics. To address this issue, reaction diffusion
models have used standard continuous time Lotka-Volterra equations with diffusing
occurring at a constant rate which corresponds to an assumption of random motion
(see [84] for more details about these three approaches of modeling).
The evolution of spatial heterogeneity may be environmentally driven, when local
or regional variability in driving variables imposes a spatial structure on population
density. Spatial heterogeneity may also arise, however, in homogeneous environment
where the evolution and persistence of the pattern is driven by the growth and migra-
tion of the interacting populations [81, 9]; mechanism which is called Turing insta-
bility or in the ecology context diffusion driven instability. The parameters domain
where a critical diffusion rate can produce stationary heterogeneous spatial patterns
from the model is called Turing space.
Hastings [88] proved a general result for Lotka-Volterra equations
à n
! m µ ¶
∂Ni X X ∂ ∂Ni
= Ni ri + aij Nj + Di i = 1, n (B.1.1)
∂t j=1 k=1
∂x k ∂x k

to describe the population dynamics of n species, with population densities Ni . The


ri ’s are measure of the intrinsic growth rate of each species, and the aij measure
interspecific and intraspecific interaction. The Di are measures of dispersal rate and
k is the dimension of space. For B.1.1 no Spatial heterogeneity will occur, as long as
dispersal of each species is from areas of higher population density to areas of lower
APPENDIX B. SPATIAL PATTERNS IN LV COMPETITION MODEL 143

density. This result is true even if the dispersal terms may include a response to other
species, or to time, as long as the net flow is toward lower population densities. It
follows that Turing patterns can develop in a predator-prey system of two interacting
species. The key ingredients for these pattern formation are [8] (1) the predator
disperses faster than the prey (2) at low densities, an increase in prey density tends
to increase the net rate of prey population growth, and (3) an increase in predator
densities decreases both prey and predator population growth.
In contrast to predator-prey systems, in LV competition models [53, 88] no Turing
patterns can develop. During last three decades several attempts have been made
to modify the simple LV competition models in order to get spatial heterogeneity
solutions. Mimura et al. [163] considered two competing species that migrate under
self and cross diffusion. Their model for two species u and v can be written as:
ut = ∂x2 [(α + β1 v)u] + (R1 − au − bv)u
(B.1.2)
vt = ∂x2 [(α + β2 u)v] + (R2 − bu − av)v

subject to no-flux boundary condition. This means that there isn’t any local migra-
tion of population through the boundaries of the system. Change in the population
density can occur inside the system. The nonlinear diffusion terms are introduced
to describe a situation that individuals are randomly walking and disperse repul-
sively and that the dispersive force will be increased by repulsive interference with
the increase of population density. These terms can be deduced from Morisita’s
phenomenological theory of “environmental density” [161]. It was shown that the
mixed population state (u, v) = (ū, v̄) can develop to spatial patterns which exhibit
a segregation phenomenon between the species for appropriate choice of diffusion
coefficients.
Levin [162] used a two patches version of the LV symmetric model (island models
approach) and concluded that, when there are at least two patches in the environment,
coexistence of two species that otherwise exclude each other is possible. Each species
establish itself in one patch sufficiently to withstand invasion of the other species,
and each is found in the other patch due to sustained migration form the favorite
territory.
Sol’e et al. [164] used another approach which was based on the Coupled Map
Lattice (CML) formalism that describes the spatiotemporal organization of a con-
tinuous variable of state through a discrete time and space. They found that CML
APPENDIX B. SPATIAL PATTERNS IN LV COMPETITION MODEL 144

version of the symmetric LV competition model showed global coexistence of the two
species even in a parameter range which allows existence of one species only. This
global persistence of both competing species in spite of local exclusion is equivalent
to Turing patterns. This result is in contrast to the ecological idea that persistence
needs a reduction of of the interspecific competition pressure. In the following section
it will be shown in a different way how to get Turing instability in the symmetric LV
model for the bistable states and for the mixed population states.

B.2 The LV system with long range diffusion terms


The symmetric two species LV model that will be analyzed is:

ut = au − bu2 − cuv + d∇2 u + δ∇2 v − ²∇4 u


(B.2.1)
vt = av − bv 2 − cuv + d∇2 v + δ∇2 u − ²∇4 v

with no flux boundary conditions, ∇2 is the two dimensional Laplacian operator and
all the coefficients are non-negative. The symmetrical case have been selected because
if coexistence is possible here, it will be easily found in some ecological interaction
involving real species, where the asymmetries are inevitable [164]. The cross diffusion
terms (δ∇2 v and δ∇2 u) describe avoidance interaction between the two populations
[163, 165] and have an aggregation effect on each of the population groups. Self
diffusion terms (d∇2 u and d∇2 v) describe migration from higher densities locations
to lower densities population locations.
The higher order diffusion terms defined by the forth order spatial derivatives
represents a long range diffusion effect [82]. Cohen and Murray [166] derived a
general non-Fickian diffusion theory from an approach based on Landau-Ginzburg
free energy model. The generalized diffusion approch takes into account the diffusive
gradient necessary to maintain a pattern even in the experimentally observed situation
of a single diffusing species. 1 They argued that with the complexity of ecological
systems it seems restrictive to consider spatial effects to simply Fickian diffusion.
1
Their diffusion part of their model include also nonlinear diffusion term

nt = −Dk∇4 n + DA∇2 n + DB∇2 n3 + G(n) (B.2.2)

with A < 0 the equation can exhibit Turing instability .


APPENDIX B. SPATIAL PATTERNS IN LV COMPETITION MODEL 145

Murray [49] discussed another approach which supports using long range diffusion
terms. He considered a generic form of flux:

J~ = G ∇n(~x + ~r, t) (B.2.3)


~
r∈N (x)

where N (x) is some neighborhood around the point x over which effects are noticed
at x, and G is some functional of the gradient. Murray showed that the first order
correction to the flux is ∇(∇2 n) and this term leads to a higher order diffusion term.
The higher order diffusion term can be interpreted as a contribution from the average
of nearest averages. 2 The importance of the long range diffusion is emphasized
in situations where the density of the species is not small and the Fickian diffusion
description is not sufficiently accurate.
In a recent work, Levin and Pacala [160] (chap. 12) used a moment expan-
sion method to “fix” the diffusion approximation. In this approach, the number of
individuals at a point x at time t + δt is given by the integral representation
Z
N (x, t + δt) = N (x, t)(1 − kδt) + kδt N (x − y, t)P (y)dy (B.2.4)

here the integral is taken over all space. This reflects the assumption that the distance
of moving is not limited to the nearest neighbors; rather it is governed by the prob-
ability distribution P (x) for moving distance x. Expansion of N (x − y, t) in Taylor
series about x, will gives at the lowest order term in the expansion a diffusion term,
but the higher order term in the expansion give the forth order derivative.
The stationary and uniform solutions of Eqs. B.2.1 and their stability are well
known [48]:

1. O: (u, v) = (0, 0) is always unstable.

2. P+ : (u, v) = (a/b, 0) and P− : (u, v) = (0, a/b) are stable for b > c i.e. when
the intraspecific coefficient is greater than the interspecific coefficient.
2
An intuitive explanation for the long range influence of the higher order diffusion term can be
realized from the finite difference approximations (in 1D). For example for “ordinary” diffusion the
discrete second order for site i involves sites which are located at i ± 1, whereas the higher order
diffusion approximation involves sites which are located at i ± 2.
APPENDIX B. SPATIAL PATTERNS IN LV COMPETITION MODEL 146

3. M: a/(b + c), a/(b + c) is stable for b < c i.e. when the interspecific competition
is weaker than the intraspecific competition model.

In the next section we will perform a linear stability analysis of these uniform solutions
to nonhomogeneous perturbations.

B.3 Turing instability analysis


To carry on the linear stability analysis to nonhomogeneous perturbations, we write
u(x, t) and v(x, t) as
u(x, t) = u0 + ũ(x, t)
(B.3.1)
v(x, t) = v0 + ṽ(x, t)

where the small perturbations are in the form

ũ(x, t) = ũ0 eσt−ikx + c.c


(B.3.2)
ṽ(x, t) = ṽ0 eσt−ikx + c.c

where k is the wavenumber and c.c stands for complex conjugation. Instability will
occur if Re(σ) > 0 for some values of the parameters. Define the kinetic terms in
B.2.1 as
f (u, v) = au − bu2 − cuv
(B.3.3)
g(u, v) = av − bv 2 − cuv

and substitute B.3.2 in B.2.1 gives after linearization the characteristic determinant
¯ ∂f ¯
¯ − dk 2
− ²k 4
− σ ∂f
− δk 2 ¯
¯ ∂u ∂g ∂v ¯=0 (B.3.4)
¯ − δk 2 ∂g
− dk 2
− ²k 4
− σ ¯
∂u ∂v

We have to compute B.3.4 at the uniform states M and P± . For the state M, we get
the following expression for the largest eigenvalue:
a(c − b) − (b + c)(d − δ + ²k 2 )k 2
σ1 (k) = (B.3.5)
b+c

In order to have σ1 > 0 we must require that δ > d, which means that the avoidance
interaction between the two species is stronger than the self diffusion. It is also
APPENDIX B. SPATIAL PATTERNS IN LV COMPETITION MODEL 147

0.2
σ1
0.1

ε=0.2
0
ε=0.28
ε=0.4
−0.1

−0.2

−0.3
0 0.5
k 1 1.5

Figure B.1: Dispersion curves for different values of ² showing the onset of the finite
wavenumber instability of the M state. For the set of parameters: a = 1, b = 0.8,
c = 0.6, d = 0.4 and δ = 0.8, ²c = 0.28

clear that without the long range diffusion i.e. ² = 0 the model will have a serious
drawback that the growth rate diverges at k → ∞, indicating that the it is not well
posed. Thus, the long-range diffusion terms exclude this short wave length and have
a stabilizing effect.
For given values of the other parameters in the model we can find the critical
value ²c that for ² < ²c the M state will be Turing unstable. Solving σ1 = 0 for k and
require that k has to be a real number gives

(b + c)(d − δ)
²c = (B.3.6)
4a(b − c)

To summarize, the M is Turing unstable for δ > d and for 0 < ² < ²c . The maximum
unstable wavenumber is r
δ−d
k0 = (B.3.7)

Figure B.1 shows the dispersion curves for different values of ².


The bistable states P± can be also Turing unstable. The same analysis as for the
APPENDIX B. SPATIAL PATTERNS IN LV COMPETITION MODEL 148

σ2
0.05

−0.05

ε=0.4
ε=0.5
−0.15
ε=0.55

−0.25
0 0.4
k 0.8 1.2

Figure B.2: Growth curves showing a finite wave instability of the uniform states P± .
Parameters: a = 1, b = 0.8, c = 1, d = 0.3 and δ = 0.8.

M state, gives the following expression for the largest eigenvalue σ2 (k)
p
−ac − 2bk 2 (d + ²k 2 ) + a2 (c − 2b)2 + 4acbδk 2 + 4b2 δ 2 k 4
σ2 (k) = (B.3.8)
2b

Figure B.2 shows the growth curves for different values of ²

B.4 Turing patterns


We will show here the results of numerical simulations of the Turing patterns for
the two cases. The initial conditions are random perturbations around the uniform
state. Figure B.3 shows the Turing patterns that evolve from the finite wavenumber
instability of the M state. Zero flux boundary conditions are used in all the following
simulations. The segregation between the two species is strong, the pattern consists
of patches with only one species. Figure B.3a shows the dispersion curves σ1 (k) and
σ2 (k) for the values of parameters used in the simulation. The large amplitude pattern
can be understood by the shape of the growth rate curve σ2 (k) which has a maximum
for k > 0. The perturbations that start to grow from the M state interact with the
unstable modes of the bistable states and this interaction drives the pattern to large
APPENDIX B. SPATIAL PATTERNS IN LV COMPETITION MODEL 149

(a) (b)
0.2

0.1 σ2
0

−0.1
σ1

−0.2

0 0.5 1 1.5 k 2

Figure B.3: Turing patterns beginning from M state. (a) Growth curves: The M
state has a finite wavenumber instability and the states P± are linear unstable but
σ2 (k) has a maximum growth for k > 0. (b) Stationary Turing patterns, represent
the density of u; the corresponding patterns of v are similar with inversion of density.
White (dark) areas correspond to maximum (u = 1.25) (minimum, u = 0) values of
u. Parameters: a = 1, b = 0.8, c = 0.6, d = 0.4, δ = 0.8, ² = 0.2; grid:128x128,
t=1000.
APPENDIX B. SPATIAL PATTERNS IN LV COMPETITION MODEL 150

(a) (b)
0.5

σ1

σ2
0

−0.5
0 0.5 1
k 1.5 2

Figure B.4: Turing patterns starting from the P+ . (a) Dispersion curves for the
parameters used in this simulation . The curve σ2 (k) shows a finite wavenumber
instability of the uniform bistable states, whereas the curve σ1 (k) shows a maximum
growth rate at k > kc for the M state. (b) The initial pattern is the uniform P+ with
small random perturbations. Parameters: a = 1, b = 0.8, c = 1, d = 0.3, δ = 0.8,
² = 0.2; grid: 128x128, dt = 0.001.

amplitude modulations. Figure B.4b shows the Turing patterns that emerge from the
P+ and correspondence growth rate are shown in Figure B.4a. For this numerical
simulation an additional rule is used in order to ensure non-negative populations i.e
u(i, j) = 0 if u(i, j) < 0 and a similar rule for v-population [164].
The Turing patterns in both cases exhibit strong segregation between the two
species because of the avoidance interaction between them. No small amplitude pat-
terns around the mixed population state have been obtained.

B.5 Morphological instabilities


The stationary front solution that connect the two stable states P+ and P− , can be
unstable to transverse perturbations. The long time pattern depends on the Turing
instability of the stable state P± and the unstable state M. If the two stable states
have a finite wavenumber instability and the M state has a maximum growth rate for
k > 0 the perturbations grows and we can get coexistence of two kind of hexagons
as shown in Figure B.5. A different behavior occurs if σ2 (k) < 0 for all k i.e. the P±
APPENDIX B. SPATIAL PATTERNS IN LV COMPETITION MODEL 151

(a) (b) (c)


0.5
σ1

σ2
0

−0.5
0 1 k 2

Figure B.5: Coexistence between Hπ and H0 hexagons. (a) The initial condition is
a front with a small perturbation. (b) The coupling of the transverse instability of
the front with the finite wave instability of the P± leads to the hexagons patterns
(t = 500). (c) Growth rate curves: σ1 (k) shows a maximum growth rate for k > 0
and σ2 (k) shows a finite wavenumber instability. Parameters: a = 1, b = 0.8, c = 1,
d = 0.3, δ = 0.8, ² = 0.2.

states are Turing stable. The perturbations decay and the species remain segregated
as shown in Figure B.6 Transverse perturbations can grow even if the front is linearly
stable due to nonlinear instability of the front. The driving force for this mechanism is
the Turing instability of the unstable M state which has a maximum growth rate for
k > 0. Figure B.7 shows this instability. Small perturbation decay but perturbations
with larger amplitude grow. Also shown in this Figure the growth rate of the uniform
state. We see that that the extended LV competition model exhibits the same pattern
formations scenarios as the urban model (see chapter 4) and the long range diffusion
terms are equivalent (in pattern formation sense) to the third dynamic equation for
the status variable of the urban model.

B.6 Ecological implications


From the ecological point of view, the main significance of these patterns is that they
allow coexistence of the competing species despite of the local exclusion principle
that will drive one of the species to extinction. The number of patches resulting
from the finite wavenumber instability depends on the size of the spatial domain. If
space is reduced beyond a minimum threshold i.e 2π/k > L where L is the length of
the system, no spatial patterns will exists, and there is an abrupt reduction in the
APPENDIX B. SPATIAL PATTERNS IN LV COMPETITION MODEL 152

(a) (b) 2 σ1 (c)

−2 σ2
0 1 k 2

Figure B.6: Linear stable front. (a) Initial condition is the same as in Figure B.5. (b)
The perturbation decays and the front is almost planar (t=500). (c) Growth curves
shows that the P± are Turing stable. Parameters: a = 4, b = 0.7, c = 3, d = 0.3,
δ = 0.7, ² = 0.2.

(a) (b) (c)

4
(d) (e) (f)
2 σ1
0

−2
σ2

−4
0 1 k 2 3

Figure B.7: Nonlinear front transverse instability. Small amplitude perturbations


decay in the course of time, while larger amplitude grow. Frames (a-e) show snapshot
at different times: 1000,2000, 3000, 4000, 5000. (f) Growth rate curves for the M
state (σ1 ) and for the P± states (σ2 ). Other values of parameters: a = 2, b = 0.7,
c = 2, d = 0.3, δ = 0.8, ² = 0.1, grid: 256x256
APPENDIX B. SPATIAL PATTERNS IN LV COMPETITION MODEL 153

number of competitors. A practical consequence for the management and protection


of natural resources is that an ecosystem cannot be reserved without protecting a
large enough spatial domain [164].
As we showed, adding long-rage diffusion terms to the classic LV competition
model with cross diffusion terms results in rich spatial patterns. These terms might be
important in situations where the density of the species is not small. We can interpret
these terms as a first order approximation to general nonlocal migration terms. In a
recent paper Travis and French [167] pointed out the essential of dispersal in spatially
models, and often is oversimplified. They suggested a greater communication between
empiricists and theoreticians in consider the exact form of dispersal incorporated. The
model takes into account three kinds of diffusion terms, two local (“nearest neighbor”)
and one of longer range. The cross diffusion terms destabilize the system were the
second order diffusion terms have a stabilizing effect. This interplay between the
different dispersion terms in the model leads to stationary spatial patterns. This
emphasizes that more complicate forms of diffusion can alter the behavior of the
ecological system and it follows therefore that a more through consideration of this
dispersal process is to be encouraged by both theoreticians and empiricists.
The ecological conventional wisdom concerning coexistence between competitors
is that, it is only possible if there is some kind of ecological differences between the
opponents. As the proposed model suggests such coexistence by compartmentation
to patches or niches is possible even for two identical species when we add long range
diffusion. From a physical point of view these spatial patterns may be regarded as
symmetry breaking phenomena typical of systems far from thermodynamic equilib-
rium [4].

You might also like