universidade técnica de lisboa instituto superior técnico · i universidade técnica de lisboa...
TRANSCRIPT
I
Universidade Técnica de Lisboa
Instituto Superior Técnico
Coupled and extended dynamical systems. Generating complexity from simple dynamics
Rui Pedro Ribeiro Ferreira de Carvalho Mestre
Dissertação para obtenção do Grau de Doutor em Física
Outubro de 2001
II
Title: Coupled and extended dynamical systems. Generating complexity from simple
dynamics
Abstract
The systems considered in this thesis are built upon simple units, which display emergent
behaviour when coupled. We show that this happens on extended systems, but also in
systems as simple as two coupled one-dimensional maps.
The systems studied are regarded as applied dynamical systems, in the sense that they can
be used in data processing (chapter 1); for modelling real world economic phenomena
(chapter 2); or as a framework to gain a deeper understanding of higher dimensional
coupled dynamical systems, which are currently models for biological and chemical
processes (chapters 3 and 4).
Key-words: coupled maps, multistability, cellular automata, econophysics, symbolic
dynamics
1
Contents
List of Figures 30.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50.2 Some definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1 Extended Dynamical Systems for Pattern Recognition 15
1.1 Cellular Automata . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151.1.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161.1.2 Phenomenology . . . . . . . . . . . . . . . . . . . . . . . . . . 181.1.3 Transients and cycles on small lattices . . . . . . . . . . . . . 211.1.4 Algebraic methods . . . . . . . . . . . . . . . . . . . . . . . . 21
1.2 Parallel computer architectures . . . . . . . . . . . . . . . . . . . . . 241.3 A Cellular Automata algorithm for pattern recognition in HMPID . . 25
1.3.1 The ALICE case study . . . . . . . . . . . . . . . . . . . . . . 261.3.2 The Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 281.3.3 The SPOT as a generalized CA . . . . . . . . . . . . . . . . . 33
1.4 Tests on the algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 341.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2 Financial markets as extended dynamical systems with interactionsthrough the aggregate pattern 422.1 Introduction to complexity in economics and finance . . . . . . . . . . 42
2.1.1 Low dimensional chaos in economic and financial data . . . . . 432.1.2 Complex systems in Economics . . . . . . . . . . . . . . . . . 452.1.3 Stylized facts of finance . . . . . . . . . . . . . . . . . . . . . . 45
2.2 A physicist’s approach to modeling market phenomena . . . . . . . . 462.2.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 462.2.2 Lévy distributions and paretian tails . . . . . . . . . . . . . . 522.2.3 Empirical statistical regularities in prices . . . . . . . . . . . . 542.2.4 Agent-based models . . . . . . . . . . . . . . . . . . . . . . . . 62
2.3 Farmer’s model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2
2.3.1 Derivation of the market impact function . . . . . . . . . . . . 682.3.2 Price dynamics caused by common trading strategies . . . . . 72
2.4 The random linear Farmer’s model . . . . . . . . . . . . . . . . . . . 752.4.1 The strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . 762.4.2 Value investors . . . . . . . . . . . . . . . . . . . . . . . . . . 772.4.3 Trend followers and value investors trading on the critical point 82
2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3 Multistability and Synchronization in Coupled MapsIntroduction 89
3.1 Multistability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 893.2 Coupled unimodal maps . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.2.1 Applications of two coupled maps . . . . . . . . . . . . . . . . 933.2.2 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.3 Two quadratic coupled maps: from synchronization to multistability . 953.3.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 953.3.2 Periodic orbits with small periods . . . . . . . . . . . . . . . . 973.3.3 Higher period phenomenology . . . . . . . . . . . . . . . . . . 1033.3.4 The non-symmetric orbits . . . . . . . . . . . . . . . . . . . . 1063.3.5 Multistability . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4 Symbolic Dynamics for Tent-Maps 115
4.1 Two Diffusively Coupled Modified Tent Maps . . . . . . . . . . . . . 1154.1.1 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1154.1.2 Rigorous expression for the orbits as a function of the codes . 1164.1.3 Numerical studies in parameter space . . . . . . . . . . . . . . 1214.1.4 A period doubling route of the coupled modified tent map . . 1234.1.5 Border collision bifurcations . . . . . . . . . . . . . . . . . . . 123
4.2 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
A Essential Definitions and Background Material 128
A.1 The Tent Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128A.2 Two Diffusively Coupled Tent Maps . . . . . . . . . . . . . . . . . . . 132
B Notation and Tools 152
Bibliography 162
3
List of Figures
1.1 Examples of Wolfram’s four qualitative classes. Top to bottom, left toright, they are: (a) Class 1: ECA 32; (b) Class 2: ECA 44; (c) Class3: ECA 90; (d) Class 4: Binary range-2 (k − 2, r − 2) CA 1771476584. 20
1.2 Space-time plot of ECA 90 evolving from an initial condition consistingof a single nonzero site. The space-time pattern is just Pascal’s trianglemodulo 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.3 The reference system and geometrical quantities. . . . . . . . . . . . . 291.4 SPOT apllied to two gaussians with a transfer coefficient κ = 0.1. The
Gaussian with larger mass is identified by a higher peak even after 20steps of the algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.5 SPOT apllied to two gaussians with a transfer coefficient κ = 0.2. TheGaussian with larger mass is not identified (the transfer coefficient istoo large). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
1.6 SPOT apllied to two gaussians with a transfer coefficient κ = 0.25.The Gaussian with lower mass is identified (the transfer coefficient istoo large). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.7 15 tracks, 100 photons p/track, 10% noise. . . . . . . . . . . . . . . . 381.8 100% recognized tracks . . . . . . . . . . . . . . . . . . . . . . . . . . 381.9 50 tracks, 100 photons p/track, No noise. . . . . . . . . . . . . . . . . 391.10 50 tracks, 30 photons p/track, 10% noise. . . . . . . . . . . . . . . . . 40
2.1 x (t) versus t for the random map 2.1 with at defined by 2.2. . . . . . 492.2 Densities of the extreme value distributions. . . . . . . . . . . . . . . 512.3 Hill plot of 5000 Pareto observations, γ = 1. . . . . . . . . . . . . . . 532.4 Hill estimator for the tails of rn when N = 103, p = 9.5 · 10−3 and
α = 0.1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 822.5 Simulation with Nv = Ntf = 5 · 103, λ = 1, η = 0.1, θmin = 1 and
θmax = 100, c = 8 · 10−2 and p = 10−2. . . . . . . . . . . . . . . . . . 86
4
3.1 On the left we have plotted the phase-opposition 2-periodic and 4-periodic orbits at µ∞. The colours denote stability of the orbits: redfor unstable, blue for saddle-node and green for stable. On the right,the corresponding eigenvalues, Re (λ) in blue and red, respectively, andIm (λ) in green. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
3.2 On the left we have plotted the non-symmetric 2-periodic and 4-periodicorbits at µ∞ which are composed or different period orbits for ε = 0.The colours denote stability of the orbits: red for unstable, blue forsaddle-node and green for stable. On the right, the correspondingeigenvalues, Re (λ) in blue and red, respectively, and Im (λ) in green. 105
3.3 Bifurcation values of ε for the phase opposition orbits. From top tobottom, birth, pitchfork and Hopf bifurcations. . . . . . . . . . . . . . 107
3.4 Periodic orbits for the unidimensional map at µ = µ∞. . . . . . . . . 1073.5 An example of 4-periodic orbits for which δ = 2 at µ∞. . . . . . . . . 1083.6 ε-evolution of a period-4 distance-2 orbit and its bifurcating orbits at
µ = µ∞. The axes are the space variables, x and y. . . . . . . . . . . 1103.7 Stability lines of orbits from distance 1 (top) to 4 (bottom). . . . . . 1113.8 Scaling properties of D(p, δ). . . . . . . . . . . . . . . . . . . . . . . . 113
4.1 The modified tent map fa,± (x) = 1− a¯x± 2−a
a
¯is plotted for a = 3
2
(dashed line) together with the tent map, f2 (x) (solid line). . . . . . 1174.2 Attractors in parameter space (a, ε) for the coupled modified tent map.
Initial condition is (0.2, 0.7) and the color map is in Fig . . . . . . . . 1224.3 Colourmap for Figure 4.2. . . . . . . . . . . . . . . . . . . . . . . . . 1234.4 The fixed point is admissible and stable for 0.481 ≤ ε ≤ 0.5. . . . . . 1254.5 A period-2 orbit which is stable for 0.288 ≤ ε ≤ 0.48. . . . . . . . . . 1264.6 A period-4 orbit which is stable for 0.268 ≤ ε ≤ 0.286. . . . . . . . . . 1264.7 The orbits from Figures 4.4, 4.5 and 4.6 superimposed and only the
stability region is plotted. It is apparent that period doubling bifurca-tions occur upon collision with the admissibility axes. . . . . . . . . . 127
A.1 The limit set of Ta, Λa, plotted against 1/a. . . . . . . . . . . . . . . 129A.2 Period-2 orbits for a = 2. . . . . . . . . . . . . . . . . . . . . . . . . . 136A.3 Period-3 orbits for a = 2. . . . . . . . . . . . . . . . . . . . . . . . . . 137A.4 Period-4 orbits for a = 2. . . . . . . . . . . . . . . . . . . . . . . . . . 138A.5 Period-4 orbits for a = 2 (cont.). . . . . . . . . . . . . . . . . . . . . . 139A.6 A mesh of 100 by 100 initial conditions is iterated for 103 steps and
the state of the coupled tent map with a = 1/2 − 10−15 (a = 1/2 is asource of numeric problems) is then plotted. . . . . . . . . . . . . . . 141
5
0.1 Introduction
The systems considered in this thesis are built upon simple units, which display
emergent behaviour when coupled. We show that this happens on extended systems,
but also in systems as simple as two coupled one-dimensional maps.
The systems presented in the following chapters are regarded as applied dynam-
ical systems, in the sense that they can be used in data processing (chapter 1); for
modelling real world economic phenomena (chapter 2); or as a framework to gain
a deeper understanding of higher dimensional coupled dynamical systems, which are
currently models for biological and chemical processes (e.g. ecosystem dynamics[113],
evolution of genetic sequences[24] and reaction diffusion systems[72]) (chapters 3 and
4).
In chapter 1 a generalized cellular automata algorithm is proposed for Cherenkov
ring detection at ALICE (CERN), showing that spatially extended systems can be
successfully implemented for technological applications. The SPOT algorithm is one
of the two proposed competing solutions proposed for particle identification at ALICE
and is currently being implemented at CERN by José Barbosa. This chapter is the
result of a collaboration with João Seixas and José Barbosa at and has been published
in [12].
Chapter 2 comprises a modification of a model proposed by J. Doyne Farmer for
financial markets. Each trader is coupled to other traders trough a collective variable
and the market is seen as a high-dimensional affine random dynamical system. At
the critical point, the model is characterized by on-off intermittency and displays a
statistics compatible with the stylized facts of finance (viz. heavy tailed distributions
(exponent ' 3) and volatility clustering). This work has resulted in a paper acceptedfor publication at the proceedings of ”Complex Behavior in Economics —Modelling,
6
Computing and Mastering Complexity”, Aix-en-Provence 2000[18].
Chapter 3 is the outcome of a collaboration with Rui Vilela Mendes on the study
of two coupled quadratic maps that resulted in the publication [20]. Bastien Fer-
nandez later questioned our proof, which lead to another work, this time studying
the same system in greater depth[19]. Through this two publications it is apparent
that phenomena reported for spatially extended systems, such as multistability [64],
are already present for a system of two coupled one-dimensional maps. Indeed, my
interest on two coupled maps comes from the fact that phenomena reminiscent of
those observed for higher dimensional systems can be rigorously studied on a low
dimensional setting.
Chapter 4 results from work done in collaboration with Ricardo Coutinho on two
coupled tent-like maps. We elaborate a rigorous framework to follow orbits from the
codes for two coupled tent-like maps and show that two coupled modified tent maps
have a period-doubling bifurcation route (which is absent from the uncoupled maps).
By using symbolic dynamics to follow the orbits, our formalism permits a rigorous
determination of an orbit, given its code and the map’s parameters, independently of
its stability. Results are also given in Appendix A for two coupled tent maps. This
is a much simpler situation for which proofs are extensively presented. Appendix B
states the notation and tools used. The outcome of the work from this chapter will
be submitted for publication shortly.
0.2 Some definitions
Let f be a map depending on parameter µ, let x0 be an initial condition, dimx = p.
A number n, n = 1, 2, . . . of successive applications of f generates a sequence of points
xn = x(n, x0, µ) called iterated sequence, or discrete phase trajectory, or orbit. The
7
map f can be considered as an implicit definition of the function xn, which plays
the role of the recurrence solution. Although theoretically quite satisfactory, such a
definition practically is almost useless, because in general the function xn is unknown,
except for the linear case, and for very few examples in the non-linear case. In all
non-contrived cases xn cannot be expressed explicitly in terms of known elementary
and transcendental functions.
In order to characterize the non-classical transcendental function xn, it is possible
to use the analogue of the characterization of a function of the complex variable
via the knowledge of its singularities (poles, zeros and essential singularities). So
a meaningful characterization of xn consists of the identification of its singularities,
and the behaviour of the latter as the parameter µ varies. Any change in the nature
of singularities so obtained, or any change of their qualitative properties, is called
a bifurcation. In the parameter space, the boundary separating behaviours of xn
which are qualitatively different, is called a set of bifurcation values of the system
parameters. This approach, which will be used in this thesis, constitutes what is
called a qualitative method of the theory of dynamical systems.
The simplest singularities are zero-dimensional: cycles of order k (or periodic
points of period k), denoted also k-periodic orbits. A k-periodic orbit is a set of
k consecutive points x∗i , i = 1, 2, . . . , k, such that: x∗i = fk (x∗i ), x∗i 6= f l (x∗i ) for
1 < l < k (a cycle of order k = 1 is called a fixed point). A k-periodic orbit is attracting
if all the eigenvalues of the Jacobian matrix of fk (when the map is differentiable at
the periodic points), written in one of the cycle point, have their modulus less than
one. A k-periodic orbit is repulsive if at least one of the eigenvalues is larger than one
in modulus. These eigenvalues are called the multipliers of the cycle, and are denoted
λi. Then a cycle is attracting (or asymptotically stable) if all |λi| < 1, i = 1, 2, . . . , p.It is repulsive if one of the multipliers is such that |λ| > 1.
8
With dimx = p, the pmultipliers λi characterize completely the iterated sequences
behaviour in a sufficiently small neighbourhood of the cycle points, if none of them
has a modulus equal to one. When f is defined by differentiable functions, in pres-
ence of parameter variations, the crossing through and eigenvalue having a modulus
equal to one, gives rise to a bifurcation leading at least to local qualitative change of
iterated sequences. The corresponding bifurcation set corresponds to a critical case
in the sense of Lyapunov, i.e. the eigenvalues do not permit to identify the dynamic
behaviour of iterated sequences, whatever small is the considered neighbourhood of
a cycle point. This behaviour depends on the nonlinear part of fk at that point.
Manifolds of dimension d = 1, 2, . . . , p− 1 (dimx = p), invariant, or mapped intoitself by f or f−1 (resp. fk or f−k) may constitute singularities of higher degree of
complexity with respect to fixed points and cycles, the dimension of which is zero. For
example, singularities of dimension p − 1, invariant by f−1, of f−k, separating openregions (basins) of initial conditions giving rise to different qualitative dynamics in
the p-dimensional space, play a fundamental role. A map may also generate singular-
ities whose dimension is not an integer. These singularities constitute what is called
”fractal sets”, either attracting, or repulsive for the points located in a sufficiently
small neighbourhood of such a set. So a basin boundary may be a repulsive fractal
set, sometimes called chaotic basin boundary.
The Schwarzian derivative of f at a point x, denoted f, x, is given by:
f, x = f000(x)
f 0 (x)− 32
µf00(x)
f 0 (x)
¶The derivative of a map with negative Schwarzian derivative has no positive local
minima or negative local maxima[111]. Julia proved that for the quadratic polynomi-
als f (x) = λx (1− x) there is at most one periodic sink and it has seemed plausibleto conjecture that this is always the case for smooth endomorphisms satisfying
9
f (0) = f (1) = 0
f has a single critical point in [0, 1]
f00(x) < 0 everywhere
Singer has proved that this conjecture is true for functions f with negative Schwarzian
derivative[111]. Indeed, he gives an example of a degree four polynomial having
positive Schwarzian derivative1 with only one critical point and non-positive second
derivative, but a fixed point and a 2-periodic orbit.
A map f : X → X is said to be topologically semi-conjugate to a map g : Y → Y
if there is an endomorphism h of X onto Y (i.e., continuous map of X onto Y )
such that h f = g h. The map f is said to be topological conjugate to g if h is ahomeomorphism (one-to-one, onto, continuous map with continuous inverse).
Semi-conjugacies preserve orbits and invariant sets. A semi-conjugacy may col-
lapse some intervals to points or blow-up some points to intervals (because it is not
invertible). If g is semi-conjugate to f and f is transitive, then g is transitive, but
the reverse implication does not necessarily hold unless f and g are conjugate [11]. If
f and g are conjugate, then so are fn and gn for any n ∈ N.
A map f : X → X is said to be topological transitive if for every pair of non-empty
open sets U, V ⊂ X, there exists n ∈ N such that fn(U) ∩ V 6= 0. If f is transitive,fn (X) is dense in X for n ∈ N, whence inverse images of non-empty open sets arenon-empty . The topological transitivity of f means that its domain cannot be broken
up into topologically non-trivial pieces which do not interact dynamically with each
1Condition f, x < 0 is satisfied everywhere by polynomials whose first derivative decomposesinto linear factors over the reals, but not by those whose first derivatives have complex roots. In thelatter case there are stable orbits which attract critical points in C−R. These critical points aredifficult to detect by ”real” techniques.
10
other [11, 31]. Topological transitivity forms part of a popular definition of chaos in
discrete dynamical systems :
Definition 1 ([31]) Let V be a set. f : V → V is said to be chaotic on V if
1. f has sensitive dependence on initial conditions;
2. f is topologically transitive;
3. periodic points are dense in V .
One of the most fundamental concepts in the study of dynamical systems is that of
the rate of growth of a quantity with time. Perhaps the most familiar example of this
is the rate of expansion or contraction of infinitesimal perturbations (i.e. of tangent
vectors). Recall that if f : M → M is a diffeomorphism of a compact manifold M ,
x ∈M and v ∈ TxM , then this is given by2
λ (x, v) = limn→∞
1
nlog kDxfnvk (0.1)
whenever this limit exists. It is well known that given an invariant measure µ, the
limit does exist for µ-almost all x by the Oseledec multiplicative ergodic theorem.
Furthermore, if µ is ergodic, then λ (x, v) takes on only a finite number of possible
values, called Lyapunov exponents[35] (LEs):
Theorem 1 Let µ be a probability measure on a spaceM , and f :M →M a measure
preserving map such that ρ is ergodic. Let also T : M →the m × m matrices be a
measurable map such that Zρ (dx) log+ kT (x)k <∞
2The reader should be aware that the same variable, λ, will be used both for the eigenvalues ofthe jacobian matrix of f and for its Lyapunov exponents. Confusion should not arise, though, as itshould be obvious from the context to which case we are refering.
11
where log+ u = max (0, log u). Define T nx = T (fn−1x) . . . T (f). Then, for µ-almost
all x, the following limit exists:
limn→∞
(T n∗x Tnx )1/2n = Λx
The logarithms of the eigenvalues of Λx are called the Lyapunov (characteristic) ex-
ponents.
Let us denote the logarithms of the eigenvalues of Λx by λ1 ≥ λ2 ≥ . . . . Orderthem, without repetition from multiplicity, as λ(1) ≥ λ(2) ≥ . . . and let m(i) be the
multiplicity of λ(i). Let E(i)x be the subspace of Rm corresponding to the eigenvalues
≤ exp ¡λ(i)¢ of Λx. Then Rm = E(1)x ⊃ E(2)x ⊃ . . . and the following holdsTheorem 2 For µ-almost all x,
limn→∞
1
nlog kTnx uk = λ(i)
if u ∈ E(i)x \E(i+1)x . In particular, for all vectors u that are not in the subspace E(2)x
(viz., almost all u), the limit is the largest Lyapunov exponent.
A positive Lyapunov exponent is the operational definition of chaos and is the
primary dynamical invariant used to characterize a chaotic process.
There are dynamical systems where the randomness is not simply given by an
additive noise[8, 13]. This kind of systems has been the subject of interest in the last
few years in relation to the problems involving disorder, such as the characterization of
the so-called on-off intermittency[100] and to model transport problems in turbulent
flows[122]. In these systems, in general, the random part represents an ensemble of
hidden variables believed to be implicated in the dynamics.
Suppose that ψ is a measure-preserving transformation on (X,B,m) and ϕx is
a measure-preserving transformation on (Y,A, µ) for each x ∈ X. Assume that the
12
mapping (x, y) → ϕx (y) is measurable with respect to the σ−algebras B × A and
A. The transformation θ defined on the product space (X × Y,B ×A,m× µ) byθ (x, y) = (ψ (x) ,ϕx (y)) is measure-preserving and is called the skew product of ψ
and ϕx[62].A random dynamical system is a skew product where the evolution of the system
after a time t is given by a deterministic mapping
ω(t) = θtω
y (t) = φ (t,ω) y
where θt is a flow and φ (t,ω) is a cocycle (i.e. is such that φ(t + s,ω) = φ(t, θsω) φ (s,ω) for all t, s > 0). ω represents the state of a dynamical system that models the
noise process and y represents the dynamical system forced by the noise. We write
θtω (resp. φ (t,ω) y) to mean a nonlinear map θt (resp. φ (t,ω)) applied to the point ω
(resp. y). In particular, the action of θ and φ need not be linear. In the case that the
evolution is chaotic we can see the above system as random forcing of a deterministic
system φ where the evolution of ω is ‘hidden’. By looking at such systems one can
get a more detailed picture of the dependence of a dynamical system on noise than
is possible by, for example, a Fokker-Planck approach.
We assume that the evolution θ has an ergodic invariant probability measure Pwith respect to which an initial condition for ω is chosen from a set of full measure.
Random dynamical systems studies the dynamics of the full system relative to the
dynamics of the measure-preserving transformation (θ, P ) ( µ is defined on subsets
in some σ-algebra).
Symbolic dynamics is the study of abstract symbolic systems characterized by a
topological space, Σ, and a shift map, σ : Σ→ Σ.
13
We begin by considering a map F : X→ X. A set Λ is invariant under F if
Λ = F (Λ). The long time behaviour of the system is captured in the restriction of
the dynamics to the invariant set, F |Λ. When Λ is attractive, F |Λ captures the rangeof system motions (the stable motions) that are physically observable. Sometimes the
dynamics on the invariant set is quite simple —periodic motion for example. Other
times computational experiments of nonlinear systems have revealed a wealth of com-
plicated ”chaotic” behavior. The famous attractors of the Lorenz equations and the
Hénon map are two such examples. The beauty of symbolic dynamics is that the
abstract symbolic description is immune to the practical difficulties associated with
iteration of unstable orbits. We associate an abstract symbolic system with F by
constructing a partitioning of Λ so that to each orbit x ∈ Λ, we associate a unique
symbolic itinerary. The itinerary identifies the partition elements visited by x under
the action of the dynamics.
Suppose we partition Λ with a finite collection of disjoint open sets,H = H0,H1, . . . ,Hk,whose closures cover Λ (i.e. Λ ⊂ ∪kj=1cl [Hj ]). For our purpose we think of the par-tition as a ”mesh” superimposed on Λ. This partition is a Markov partition when
the intersections of successive iterates of the original grid produce a nested set of
successively finer and finer grids until in the limit, each point in Λ falls into its own
infinitesimally small partition element.
Let Σk be the space of bi-infinite sequences on a k letter alphabet 1, 2, . . . , k,
Σk ≡ s = (. . . s−3s−2s−1 · s0s1s2 . . . ) : sj ∈ 1, 2, . . . , k
The correspondence between the partition elements and points in the phase space is
given by the itinerary map h : Λ→ Σk, where ∀x ∈ Λ
h (x)j = sj if Fj (x) ∈ Hsj
14
The shift map is the continuous map σ : Σ −→ Σ, such that σ (. . . , θ−1.θ0, θ1, . . . )
= (. . . , θ0.θ1, θ2, . . . ). For every element θ = (. . . , θ−2, θ−1.θ0, θ1, . . . ) ∈ Σ we call
θu ≡ (. . . , θ−2, θ−1) the tail of θ, and θs ≡ (θ0, θ1, . . . ) the head of θ. Let Cu (resp.Cs) be the set of all tails (resp. heads) of the elements of Σ, so that Σ = Cu × Cs.Obviously, an arbitarily given symbolic sequence may not correspond to any actual
orbit in the dynamics. We must formulate the conditions for a symbolic sequence to
be allowed by the dynamics. A symbolic sequence is called an admissible sequence at
a given parameter, if one can choose an initial point to produce a numerical orbit of
the map which leads to the symbolic sequence, using a given partition of the interval.
15
Chapter 1
Extended Dynamical Systems for
Pattern Recognition
1.1 Cellular Automata
Cellular Automata (CA) are spatially-extended dynamical systems with the par-
ticular feature of being as discrete as possible: in space, in time, and in site value.
Space and time are indexed by integers, and the value of each spatial site is taken
from some finite alphabet of formal symbols. The rule governing a CAs evolution is
extremely simple, being a spatially and temporally invariant, locally coupled, parallel
update of the lattice.
Despite their extreme simplicity, CA exhibit an extremely wide range of behaviour,
from spatially uniform stationary states to spatially and temporally periodic evolu-
tion to spatial disorder to fully-developed, spatio-temporally turbulent behaviour.
Many CA evolve to patterns which can be decomposed into two components: (i) a
spatio-temporally uniform, periodic, or turbulent background, and (ii) particles or
dislocations moving against the background. The particles are defined in terms of the
16
background pattern: they are points at which the pattern breaks down.
The common methodology in analyzing nonlinear spatially-extended systems is
the following: first the characteristic pattern(s) into which the system evolves are
identified, then a representation that fits these patterns is formulated, and finally the
characteristic patterns are used as the underlying structure in terms of which the
system is understood. Another way of saying this is that the characteristic patterns
emerging in a system’s evolution comprise a pattern basis that governs the system’s
behaviour and guides its analysis. This methodology breaks down in the face of
spatio-temporal disorder, by failing at the very first step: no characteristic patterns
can be identified.
1.1.1 Definitions
A one-dimensional cellular automaton consists of a countable array of discrete
sites or cells i and a discrete-time local update rule φ operating in parallel on local
neighborhoods of a given radius r. At each time t each site takes on a value σit in
a finite alphabet A of k primitive symbols: σit ∈ 0, 1, . . . , k − 1 ≡ A. The localsite-update operation is written
σit+1 = φ¡σi−rt , . . . , σi+rt
¢The space index i and the time index t are omitted when they are unimportant.
The string ηit = σi−rt . . .σi+rt of 2r+1 symbols is the parent neighborhood of σit+1,
its child symbol. The local update rule is therefore also written σit+1 = φ (ηit). The
set of distinct ordered pairs (η, σ) with σ = φ (η) is the rule table, of which the η’s
are the input portion and the σ’s are the output portion.
The state (or configuration) st of the CA at time t is the configuration of the finite
or infinite spatial array: st ∈ AN , where AN is the set of all possible site values onan array of N cells. The array is also called the CA’s lattice.
17
The extended state space, denoted by A∗, is the union of all states of any N :
A∗ =[N≥0
AN
with A0 = ∅.The CA global update rule Φ : AN → AN applies φ in parallel to all sites in the
lattice: st = Φst−1. For finite N it is also necessary to specify a boundary condition
(this is usually periodic).
The simplest class of cellular automata are the elementary CA (ECA) for which
(k, r) = (2, 1); that is, binary cell values and nearest-neighbour interactions, giving
a total of k2r+1 = 8 possible parent neighbourhoods. Since for each such parent
neighbourhood there are k possible values of the child cell, there is a total of kk2r+1
different local update rules φ. Thus there are 28 = 256 different elementary CA rules.
For the class of binary CA with the next-larger radius, binary range-2 CA in which
(k, r) = (2, 2), there are 32 different parent neighbourhoods, hence 232 different rules.
A numerical index #(φ) ∈ Z+, the rule number, is assigned to each function φ
(and therefore to each Φ), using a conventional numbering scheme as follows[119]. The
different parent neighbourhoods η are regarded as numbers in base k, and arranged in
decreasing numerical order, say from left to right. Immediately beneath each parent
neighbourhood, its child symbol σ = φ(η) is written. The sequence of 2r + 1 child
symbols is then regarded as a number, again in base k. The ECA having rule index
18, for example, has the following rule table:
NeighbourhoodNext site value
1110
1100
1010
1001
0110
0100
0011
0000
σi−1t σitσi+1t
σit+1
Table 1.1: ECA having rule 18.
In some CA applications, it is useful to designate a particular symbol bσ ∈ A(usually 0) as a quiescent symbol, and consider other members of A as somehow
18
denoting excited cell values. It is usual to require that the CA rule leave blocks of
quiescent cells invariant; i.e.,
φ (bσ, bσ, . . . , bσ) = bσAlternatively, we can look at how a CA processes entire sets of spatial configurations.
This is analogous in dynamical systems theory to the study of the evolution of state
ensembles, rather than individual orbits. To this end we define the ensemble evolution
operator Φ by
Lt = ΦLt−1 = ΦtL0
where L0 ⊆ A∗ is any set of spatial configurations used as initial conditions, and theensemble evolves according to
Lt = st : st = φst−1, ∀st−1 ∈ Lt−1
Note that no restrictions are made on the size of the lattice; a single ensemble may
contain states with different lattice sizes.
Many CA are irreversible or dissipative in the sense that the global update rule is
many-to-one, in which case Lt ⊂ Lt−1.
1.1.2 Phenomenology
CA exhibit a remarkable variety in behaviour. Space-time plots of different CA
range from being entirely ordered to spatio-temporally disordered. At least one CA
has been shown to be usable as a random number generator[121].
Nevertheless, after an extensive survey, Wolfram[120] proposed that CA rules can
be classified in four qualitative classes, based on a visual examination of space-time
patterns exhibited by the CA at long times:
19
1. (Spatio-temporally uniform state). Almost all initial configurations relax
after a transient period to the same fixed configuration (e.g. all 1s);
2. (Separated simple or periodic structures). Almost all initial configurations
relax after a transient period to some fixed point or some temporally periodic
cycle of configurations, but which one depends on the initial configuration. (It
should be pointed out that on finite lattices, there is only a finite number (2N)
of possible configurations, so all rules ultimately lead to periodic behaviour.
Class 2 refers not to this type of periodic behaviour but rather to cycles with
periods much shorter than 2N);
3. (Chaotic space-time pattern). Almost all initial configurations relax after a
transient period to chaotic behaviour;
4. (Complex localized structures). Some initial configurations result in com-
plex localized structures, sometimes long-lived.
An example of each of these classes is shown in Fig 1.1. It is important to em-
phasize that this classification scheme is phenomenological at root. The scheme was
originally motivated by the hierarchy of behaviour in dynamical systems theory —i.e.,
the sequence: fixed point, limit cycle, quasiperiodicity, chaos.
Classes 1 and 2 are more or less self-explanatory. Note, however, that tem-
porally periodic evolution of a spatially disordered configuration (spatial chaos) falls
into class 2. In class 3, chaotic is taken to mean spatio-temporal disorder, in the
sense that the number of distinct space-time patches grows exponentially with the
linear size of the patch along both the spatial and temporal axes. Class 4 has been
the subject of speculation about its possible computational capacity. It has been
suggested that CA in this class may be capable of universal computation, that is, of
implementing a universal Turing machine[120].
20
Figure 1.1: Examples of Wolfram’s four qualitative classes. Top to bottom, left toright, they are: (a) Class 1: ECA 32; (b) Class 2: ECA 44; (c) Class 3: ECA 90; (d)Class 4: Binary range-2 (k − 2, r − 2) CA 1771476584.
21
1.1.3 Transients and cycles on small lattices
Because the state space is discrete, on a finite lattice of any given size the number
of possible configurations is necessarily finite. This means that, in principle at least,
all possible configurations of the system can be enumerated (say, as points on a sheet of
paper), and the step-by-step evolution can be represented by connecting those points
with directed segments. Once this has been done for each state in the state space,
the resulting graph is a state transition diagram[82]. The state transition diagram
is a directed graph, often consisting of multiple disconnected components, with each
node represents a unique configuration. Because the CA rule is deterministic, there
is exactly one edge leaving each node. The majority of CA rules are dissipative,
which shows up as multiple edges entering nodes. The branching in the trees is a
measure of the amount of dissipation in the system. The so-called Garden-of-Eden
configurations, which have no predecessors, have in-degree 0.
1.1.4 Algebraic methods
The algebraic analysis of CA considers finite CA with periodic boundary condi-
tions. The possible site values are taken to be elements of a finite commutative ring
Rk with k elements. Typically, Rk is chosen to be Zk, so that site values are treated
as integers modulo k.
On a lattice of size N , each configuration st ∈ AN is represented by a characteristicpolynomial
A(t) (x) =N−1Xi=0
a(t)i x
i
where x ∈ R is a formal parameter and a(t)i is the numeric value of cell σit.
22
The local update rule φ¡σi−rt , . . . , σi+rt
¢is then expressed as a function
a(t)i = F
³a(t−1)i−r , . . . , a
(t)i+r
´Additive (or linear) CA are those in which the function F has the form of a simple
linear combination of its arguments, i.e.,
a(t)i = α−ra
(t−1)i−r + α−r+1a
(t−1)i−r+1 + . . .+ α+ra
(t−1)i+r
where each coefficient αj is a fixed element of Rk and all arithmetic is performed in
Rk. For such CA, the global evolution Φ (st) can be represented by the multiplication
of the characteristic polynomial A(t) (x) by the fixed dipolynomial (i.e., a polynomial
with both positive and negative powers of x)
T (x) = α−rx−r + α−r+1x−r+1 + . . .+ α+rxr
according to the relation
A(t) (x) = T (x)A(t−1) (x) mod¡xN − 1¢
In this way, the evolution of an additive CA is expressed in the language of lin-
ear algebra. Additive CA obey a linear superposition principle in the sense that the
site-by-site sum (mod k) of two configurations evolving under the CA rule is itself a
configuration evolving according to the same rule. Thus, for any two states A(0) (x)
and B(0) (x) (expressed by their characteristic polynomials), the sum C(0) (x) =
A(0) (x) + B(0) (x) evolves such that C(t) (x) = A(t) (x) + B(t) (x) for each time t.
In particular, this means that the evolution of an arbitrary initial condition s0 can be
expressed as a superposition of the evolution from single-seed initial conditions, each
consisting of a single nonzero cell.
One of the most extensively studied additive CA is ECA 90, for which the evolution
is governed by the dipolynomial
T (x) = x+ x−1
23
For example, an initial condition s0 = . . . 00100 . . . containing a single nonzero cell
evolves after t time steps to a state with characteristic polynomial
T (x) 1 =¡x+ x−1
¢t=
tXi=0
µt
i
¶x2i−t
Note that this is the formula for Pascal’s triangle modulo 2. Figure 1.2 shows a space-
time plot of the evolution of ECA 90 on a periodic lattice of size N = 1000, starting
from a single-seed initial condition with a single nonzero cell at site i = 500. Only
the central portion of the lattice is shown. Because ECA 90 is linear, the evolution
of an arbitrary initial condition is given by the site-by-site sum (mod 2) of copies
of the evolution of this initial condition, shifted to the left or right as appropriate.
Random initial conditions evolving under ECA 90 exhibit spatio-temporal disorder
or chaos, despite the linearity of the rule. The disorder arises from superimposing
many shifted copies of the space-time pattern of Figure 1.2. Reference [82] contains
extensive calculations of the number and lengths of transients and cycles for ECA 90
on finite periodic lattices of size N .
Figure 1.2: Space-time plot of ECA 90 evolving from an initial condition consistingof a single nonzero site. The space-time pattern is just Pascal’s triangle modulo 2.
24
1.2 Parallel computer architectures
Since the development of the electronic computer in the 1940s, the serial processing
computational paradigm has successfully held sway. It has developed to the point
where it is now ubiquitous. However, there are many tasks which are yet to be
successfully tackled computationally. A case in point is the multifarious activities
that the human brain performs regularly, including pattern recognition, associative
recall, etc. which is often extremely difficult to do using traditional computation.
This problem has led to the development of non-standard techniques to tackle
situations at which biological information processing systems excel. One of the more
successful of such developments aims at reverse-engineering the biological appara-
tus itself to find out why and how it works. Parallel computation models (cellular
automata and neural networks) have grown up on the premise that the massively par-
allel distributed processing and connectionist structure observed in the brain is the
key behind its superior performance. By implementing these features in the design of
a new class of architectures and algorithms, it is hoped that machines will approach
human-like ability in handling real-world situations.
There are two major types of parallel computers. The Single-Instruction-Multiple-
Data (SIMD) architecture is characterized by the fact that each processor executes
the same instruction simultaneously. Examples of this type of architecture are the
Connection Machine CM-2, the ASTRAmachine1 and the Maspar MP-1. In Multiple-
Instruction-Multiple-Data (MIMD) computers, each processor operates autonomously
and executes instructions independently of the other processors. Examples of this
type of architecture are the Connection Machine CM-5, Intel Paragon, and Fujitsu
VPP500[117].
In the last ten years interesting developments have taken place: the spectacu-
1The ASTRA machine was developed at CERN.
25
lar success of Intel machines, the diminishing interest in parallel machines, and the
simultaneous disappearance of mainframes. The relative cheapness of PCs having
workstation performance together with extensive networking created a situation that
can be regarded as the MIMD version of parallel computing. Owing to the expected
increase of the bandwidth of network communication lines, these loose MIMD systems
will become more and more tightly connected. Therefore, if one wants to speak about
‘real’ parallel computing one should define it in a restricted sense, close to the SIMD
architecture. Here we concentrate on parallel computing based on simple processing
elements but in enormous quantities, which are so cheap. CA algorithms for pattern
recognition have the advantage that parallel hardware implementations are presently
available, thus allowing for the fulfillment of very complex tasks on-line.
We define the goal of computer based pattern recognition as the classification of
inputs according to pre-set categories, by deciding whether a given configuration is
an example of a given pattern2.
1.3 ACellular Automata algorithm for pattern recog-
nition in HMPID
Adaptable Neural Networks (ANNs) are known to give quite good answers to the
problem of pattern recognition whenever the desired input/output connection is very
hard to obtain.
In some situations in high energy physics, however, we need only to recognize a
single pattern scaled to different sizes. This is the case, e.g., for the situation we are
going to analyze in this chapter, viz. Cherenkov circle recognition.
When a charged particle with velocity v passes through a medium of index n such
2Intuitively, a pattern is a structural feature or collection of features.
26
that v > c/n, it emits light around a cone whose angle θ is given by cos θ = 1/βn
where β = v/c is the particle velocity relative to the speed of light. This is called
the Cherenkov effect. From the front view, the cone appears as a circle, whose
radius decreases as the particle mass increases. RICH detectors take advantage of the
Cherenkov radiation produced by high-momentum particles transversing a radiating
layer. The cone of Cherenkov light is read out, by an array of photomultipliers placed
parallel to the quartz, as a ring of photons enclosing a MIP (Minimum Ionizing
Particle - the signal that the particle itself generates when crosses the array). The size
of the ring, i.e. the Cherenkov angle, yields the particle momentum. As the intensity
of the Cherenkov light is usually low, only a few photons are emitted, which has
the result that the Cherenkov ring appears not as a full ring but only as several dots
lying on a circle. Hence, the challenge is to reconstruct the circle which passes through
these points knowing that the circle can be deformed as the light originating from
the particle usually goes through different materials and therefore can be diffracted
or even reflected. Furthermore, one has to deal with some noise inherent to any high
sensitivity detection.
As we will show, a very simple and efficient method can be devised in which one
makes no use of ANNs. Instead, one can use massive parallelism with a convenient
architecture and obtain excellent results. Since ANNs always pose the problems of
high connectivity and other hardware implementation problems, the simplicity of the
present approach may reveal itself to be quite useful and rewarding.
1.3.1 The ALICE case study
In the ALICE (A Large Ion Collider Experiment) case at CERN the cathode is
segmented into 162 × 162 pads per module, each pad having a size of 8 × 8 mm[2],with a 2 mm anode-to-cathode gap. In the acceptance region covered by the RICH,
27
the maximum particle density reaches 100 m−2 (including the expected background)
and the predicted densities of charged pions and kaons with momenta above 1 GeV/c
(signal particles) are v 5m−2 at the position of the RICH, with incident angle 0 <
θin < 15 [29]. In the low magnetic field of ALICE, tracks with these momenta are
almost straight. This means that most of the patterns will be closed ellipses, a feature
that eases the pattern recognition[2].
The Pattern Recognition algorithm will have to take into account not only the
data structure as detected by the High Momentum Particle ID (HMPID), but the
geometry of ALICE, so as to build a reliable Cherenkov angle reconstruction method.
We know that the data will be a large-sized map of hits with a very complex
pattern where individual RICH patterns can hardly be identified by eye and that
approximately 9.3% of the pads have a signal above threshold [83].
In order to extract the Cherenkov signal for a chosen track from the described
cluster distribution and track-impact information, we define the detection zone (cor-
responding to the area that would encompass the biggest possible ellipses) on the
detector plane where Cherenkov photon clusters emitted by the chosen track may be
observed. Then we create an activation mesh for the three angles to be detected: the
particle’s polar angle, θ, the particle’s azimuthal angle, ϕ and the Cherenkov angle,
Ω. We reconstruct the corresponding Cherenkov angle of emission, Ωi, for each mesh
point (θi,ϕi) and each cluster (or pad) in the fiducial area and we increment the ac-
tivation of the mesh point (θi,ϕi,Ωi) by one. We must apply a correction algorithm
that accounts for the presence of background noise and the SPOT algorithm to reduce
clusters in space (θ,ϕ,Ω) to one point. Finally the cell with the highest activation
on the (θ,ϕ,Ω) mesh contains the particles’ incidence and Cherenkov angles.
28
1.3.2 The Algorithm
Assumptions
The algorithm is designed so that from a ring-shaped distribution of photons
we reconstruct the angles under which the incident particle entered the detector
and the Cherenkov cone of emission on the radiator. We have made the follow-
ing assumptions[2] :
• the origin of photons is chosen to be one point on the track path through theradiator. The coordinates of this point vary with the particle’s incidence angle
(because of the absorption in the radiator), and are close to the middle of the
radiator. It is a good approximation to admit that the particle emits exactly
on the middle of the radiator.
• all the photons are assumed to be of the same energy (6.85 eV in the ALICEHMPID case), corresponding to the mean energy of the photons producing
photoelectrons . Therefore the only information we have is whether a photon
was present or not.
• the coordinates of the MIPs are known. It is fairly easy to recognize a MIP as,on one hand, it is a much bigger cluster than the ones produced by photons
and, on the other hand, their activation is much higher.
The reference system has the origin at the hitting point of the Minimum Ionizing
Particle (MIP) on the photocathode as illustrated in Figure 1.3. In this figure we also
define all the geometrical quantities as shown.
29
Figure 1.3: The reference system and geometrical quantities.
30
Determination of the Cherenkov angle, Ω.
If the incidence is orthogonal to the (x, y) plane, the emitted Cherenkov cone is
given by:
x2
R2+y2
R2=(z − l)2l2
(1.1)
here R is the radius of the circle resulting from the intersection of the cone with the
photocathode plane.
When the particles’ incidence angles, θp and ϕp are not zero, (1.1) is transformed
by means of two rotations. The first, around the x axis by an angle θp, transforms it
into:
x2
R2+(y cos θp + z sin θp)
2
R2=(−y sin θp + z cos θp − l)2
l2(1.2)
Instead of a circle, we now observe an ellipse on the photocathode plane which is,
minus a rotation, determined by making z = 0 on (1.2),
x2
R2+(y cos θp)
2
R2=(−y sin θp − l)2
l2(1.3)
The second rotation, around the z axis, by the angle ϕp, is now made, transforming
(1.3) into:
(−y sinϕp + x cosϕp)2R2
+cos θ2p (y cosϕp + x sinϕp)
2
R2=
=(− sin θp (y cosϕp + x sinϕp)− l)2
l2(1.4)
As tanΩ = Rl,
31
tan2Ωp =(−y sinϕp + x cosϕp)2 + cos θ2p (y cosϕp + x sinϕp)2
(sin θp (y cosϕp + x sinϕp) + l)2 (1.5)
Knowing θp and ϕp, for each activated pad with coordinates (x, y), we detect a
Cherenkov angle, Ωp.
The maximum zone where photons from the chosen track may be observed is a
circle with radius:
RMax = h tan(ΩMax + θMax) (1.6)
where h is the height of the radiator (to the pad plane) and ΩMaxthe maximum
Cherenkov angle.
The angles θp and ϕp are known from the TPC chamber data ??. The goal of our
algorithm is ito identify the Cherenkov angle Ωp, given a set of activated pads with
coordinates (x, y) inside the maximum zone. Equation (1.5) permits the reconstruc-
tion of Ωp for each pad. Therefore, to each Cherenkov cone, there will correspond a
”cloud” of reconstructed points on space (θ,ϕ,Ω). The SPOT algorithm will reduce
this cloud to its center of mass and extract the Cherenkov angle corresponding to the
MIP.
Background Correction
The background has been parameterized with the analytical form [2]:
Fbkg(ηc) = [tan ηc(1 + tan2 ηc)]
α +A+B tan ηc (1.7)
where α = 5.52, A = −7.80 and B = 22.02. In our case we assumed the backgroundnoise would generate an occupancy of about 10% i.e. the noise level will be 10%,
which is a good approximation.
32
The photons detected on the photocathode from the same track do not fall, how-
ever, on a perfect ellipse. Each photon’s trajectory is slightly altered so that the
resulting figure on the photocathode plane is an elliptical ring with average thickness
2σ. This implies that the ring will encompass background (noise) photons as well as
the true clusters from Cherenkov radiation. Thus a correction for background noise
is compulsory3.
We determine the average number of background photons inside the ring defined
by the two ellipses¡θp [i] ,ϕp [j] ,Ωp [k] +
σNΩ
2π
¢and (θp [i] ,ϕp [j] ,Ωp [k]) as
Npho = NoiseLevel × (Area(θp[i],ϕp[j],Ωp[k]+
σNΩ2π )− Area
(θp[i],ϕp[j],Ωp[k])) (1.8)
Thus Npho is calculated for each (θp [i] ,ϕp [j] ,Ωp [k]) and the matrix element decre-
ments by Npho:
(θp [i] ,ϕp [j] ,Ωp [k]) = (θp [i] ,ϕp [j] ,Ωp [k])−Npho (1.9)
The calculation of (1.8) requires the determination of the area of the ellipse defined
by the intersection of the Cherenkov cone (θp [i] ,ϕp [j] ,Ωp [k]) with the photocathode
plane.
The ellipse defined by the intersection of the Cherenkov cone (θp [i] ,ϕp [j] ,Ωp [k])
with the photocathode plane is
cot2Ω− tan2 θpl2
x2 + (cos θp cotΩ− sin θp tan θp tanΩ
l)2 ×
× (y − 2l sin θp sin2Ω
cos(2θp) + cos(2Ω))2 = 1 (1.10)
3The activation of cell (thetap [i] ,phip [j] ,Omegap[Ω(x,y)]) increases with Ω(x,y) independentlyof the number of Cherenkov photons inside the cell due to an increase with Ω(x,y) of the number ofbackground points inside the same cell.
33
with area:
A =2√2πl2 cos2 θp sin
2Ω cosΩ
(cos(2θp) + cos(2Ω))3/2(1.11)
Eq. (1.11) is used to calculate Npho and thus the noise correction.
1.3.3 The SPOT as a generalized CA
As one has to deal with some noise inherent to any high sensitivity detection,
the Cherenkov photons are spread along cells on the same cluster on the (θp,ϕp,Ωp)
space, creating the suitable conditions for applying the SPOT algorithm[4].
Consider a discretized bidimensional space, S ⊂ Z2, whose ’nodes’ are denoted by©¡Xi,Yj
¢ª, i = 1, . . . , N and j = 1, . . . ,M . Further, associate to every node the
function F t : S×N→ R. Prior to the application of the SPOT algorithm, nodes have
an ’initial condition’ activation F0. The SPOT updates nodes in parallel as follows:
1. At step t, consider the ’cubic’ neighbourhood of cell©¡Xi,Yj
¢ª:
N ti,j=
©¡Xtk,Y
tl
¢ªk ∈ i− 1, i, i+ 1 and l ∈ j − 1, j, j + 1
2. Determine k, l such that max ¡F t ¡N ti,j
¢¢= F t
¡¡Xtk,Y
tl
¢¢3. Let
F t+1¡¡Xtk,Y
tl
¢¢:= F t
¡¡Xtk,Y
tl
¢¢+
1Xm=−1m6=k
1Xn=−1n6=l
κF t ¡¡Xt2m−1,Y
t2n−1
¢¢F t+1
¡Ni,j\ ¡Xtk,Y
tl
¢¢:= (1− κ)F t ¡Ni,j\
¡Xtk,Y
tl
¢¢where 0 < κ < 1.
34
At each application of the SPOT algorithm, the highest activation cell in each
neighbourhood is reinforced with the sum of the activation of the neighbouring cells
(multiplied by a factor κ, the transfer coefficient). Conversely, the cells with activation
lower than the maximum have their activation decreased by a factor κ.
The SPOT algorithm ”transfers the weights” in S. If the transfer coefficient, κ, isproperly configured, the algorithm reduces a ”cloud” of points to its center of mass
with an activation proportional to the cloud’s mass. As an illustration we show the
action of the SPOT algorithm on two gaussians, one with a lower maximum but a
higher mass than a second one. The algorithm is able to extract the dominant mass
6contribution if the activation coefficient is κ = 0.1 —this is the value we shall use for
Cherenkov angle recognition. For higher activation coefficients, it fails to perfom the
correct identification. This is apparent on Figures 3.3.2 ,1.5, and 1.6. On each case,
we consider a mesh of 50 × 50 nodes. The transfer coefficient, κ, is varied from 0.1
to 0.25 and the evolution of the system is observed for each picture at t = 0, t = 20,
t = 40 and t = 60.
Each node has a real value, therefore there is an infinite alphabet of primitive
symbols for each site (as opposed to a finite alphabet for CA as defined in Section
1.1.1). This leads to the classification of the SPOT as a generalized CA.
1.4 Tests on the algorithm
35
Figure 1.4: SPOT apllied to two gaussians with a transfer coefficient κ = 0.1. TheGaussian with larger mass is identified by a higher peak even after 20 steps of thealgorithm.
36
Figure 1.5: SPOT apllied to two gaussians with a transfer coefficient κ = 0.2. TheGaussian with larger mass is not identified (the transfer coefficient is too large).
37
Figure 1.6: SPOT apllied to two gaussians with a transfer coefficient κ = 0.25. TheGaussian with lower mass is identified (the transfer coefficient is too large).
38
020406080100
120
140 0
2040
6080
100
120
140
Figure 1.7: 15 tracks, 100 photons p/track, 10% noise.
Figure 1.8: 100% recognized tracks
39
020406080100
120
140 0
2040
6080
100
120
140
Figure 1.9: 50 tracks, 100 photons p/track, No noise.
As we stated above, our case study is the ALICE HMPID detector. We must,
therefore test our algorithm in a corresponding environment. We thus make the
following assumptions:
• The height from radiator to pad plane is 7.5 cm and the pad area is 0.8 × 0.8cm2;
• The maximum incidence angle, θp, is 10;
• The Cherenkov angle is in the continuous range 25;
• Each photon track can have a shift up to 2 from the original course;
• We considered as successfully detected tracks only the ones in which the differ-ence between the detected and real Cherenkov angles was under 1.
40
020406080100
120
140 0
2040
6080
100
120
140
Figure 1.10: 50 tracks, 30 photons p/track, 10% noise.
Next we present three generated events as shown in Figures 1.7-1.10. In these
Figures we also show the pattern recognition performed by our algorithm.
The algorithm performs well for multiplicities up to 30 tracks, detecting the to-
tality of the tracks. On higher multiplicities environments the probability of misiden-
tification becomes important, reaching 66% for a full event.
1.5 Conclusions
Our proposed algorithm performs reasonably well, considering the complexity of
the problem. The algorithm is geometry independent since it detects ellipses as well
as circles.
Noise is an important factor. The algorithm’s performance degrades fast with
increasing noise level and this is the problem most difficult to overcome. For occu-
41
pations up to 5% the reconstruction is perfect, for occupations above 15% it is less
than half. For a perfect reconstruction each track should, at least, have 50 photons.
This seems to be the only way to eliminate the problems posed by high noise levels.
The SPOT introduces only a small correction. It is essential, though, when the mesh
has a fine granularity, because this implies that reconstructed points have a higher
probability of not being clustered in the same cell.
We consider this algorithm to be very promising but it still relies on a number of
approximations. Some of them are more stringent than the ones found in real cases.
A final assessment of the quality of the algorithm can thus only be obtained when
applying it to real data in a definite experimental environment.
42
Chapter 2
Financial markets as extended
dynamical systems with
interactions through the aggregate
pattern
2.1 Introduction to complexity in economics and
finance
There are two contrasting viewpoints concerning the explanation of observed fluc-
tuations in economic and financial markets. According to the first (Neoclassical)
view the main source of fluctuations is to be found in exogenous, random shocks to
fundamentals1. In the absence of shocks, prices and other variables would converge
to a steady state (growth) path, completely determined by fundamentals. According
1Economic fundamentals are structural characteristics of an economy (e.g. preferences, endow-ments, production sets, technological changes, transport improvements, population growth, andchanges in tastes) which may change over time due to endogenous reasons.
43
to the second (Keynesian) view a significant part of observed fluctuations is caused
by nonlinear economic laws. Even in the absence of any external shocks, nonlinear
market laws can generate endogenous business fluctuations. The neoclassical view is
intimately related to the concept of rational expectations, whereas animal spirits or
market psychology have been an important Keynesian theme.
In finance the two different viewpoints lead to opposite views concerning the ef-
ficiency of financial markets. In the efficient market hypothesis (EMH ) the current
price already contains all information and past prices cannot help in predicting future
prices. The modern view of the EMH admits that the market may overreact in some
circumstances, but underreact in others according to pure chance (in other words,
the expected value of abnormal returns is zero)[38]. Parametric stochastic processes
have been used in the empirical literature that are consistent with the EMH. Ex-
amples include random walk processes, GARCH-processes and the like. In contrast,
Keynes already argued that stock prices are not only determined by fundamentals,
but in addition market psychology and investors animal spirits influence financial
markets significantly. In the Keynesian view, simple technical trading rules, such as
extrapolation of a trend, may help predict future price changes.
We shall discuss a modification of Doyne Farmer’s model[40] for financial markets
that conforms to the stylized facts of finance. Farmer’s model is Keynesian in the
sense that it assumes the market to be a consequence of traders’ psychology, but it is
also compatible with the EMH on large time scales[40]. Let us start by putting into
context the frame upon which the model is built (and to which it reacts).-
2.1.1 Low dimensional chaos in economic and financial data
Based on a numerical computation of fractal dimension, several researchers claimed
to observe low-dimensional chaos in price series. Such computations are done by mea-
44
suring the coarse-grained size of a set, in this case a possible attractor of returns in
a state space whose variables are lagged returns, as a function of the scale of the
coarse-graining. If this behaves as a power law in the limit where the scale is small, it
implies low-dimensional chaos. But it is very easy to be fooled when performing such
calculations. It is critical to test against a carefully formulated null hypothesis[118].
Scheinkman and LeBaron showed in their classical paper[108, p.334] that the
claims of low-dimensional chaos in price series were not well justified. The authors
state that a substantial part of the variation on weekly returns is coming from non-
linearities as opposed to randomness. However, they are very careful in their claim:
”...the data are not incompatible with a theory where some of the variation in weekly
returns could come from nonlinearities as opposed to randomness and are not com-
patible with a theory that predicts that the returns are generated by IID random vari-
ables”. While nonlinearity is clearly present, there is no convincing evidence of low-
dimensionality. The power-law scaling that people thought they saw was apparently
just an artifact of the finite size of their data sets.
New light has been shed on this debate by the discovery of chaotic, seemingly ran-
dom looking dynamical behaviour in simple deterministic (macro) economic models
(see e.g. review [89]). It has been shown that simple nonlinear general equilibrium
models, satisfying the current dominating assumptions in economic theory (i.e. util-
ity or profit maximization agents, rational, self-fulfilling expectations and market
clearing), can generate chaotic equilibrium dynamics, but in most of these cases the
dynamics can be reduced to a one- or two-dimensional nonlinear difference equation.
The (chaotic) time series generated by these models all seem to be clearly different
from actual (macro) economic data however. Apparently, the models are still ”too
simple to be true”.
In summary, it seems to be fair to say that there is no convincing evidence for
45
a low dimensional chaotic explanation of economic or financial data. Still, there
seems to be a ”chaos model-data paradox” in economics: chaos is hard to find in
economic and financial data, but easily generated by economic equilibrium models
under increasingly plausible conditions.
2.1.2 Complex systems in Economics
During the late 1980’s a parallel research agenda had priority in economic research.
Its main concern was no longer to seek analytical solutions to economic problems,
but to study the unfolding of the patterns its agents (co)create. Low dimensional
dynamical systems were no longer the chosen tool and economic structures were seen
as high dimensional.
Complex economic environments typically take as their microeconomic founda-
tions direct interactions between economic agents. What this means is that the
individual agents are conceptualized as decision makers that adapt and react with
strategy and foresight directly to the decisions of others, or indirectly to the aggre-
gate pattern they generate.
Complex economic environments are typically characterized by positive feedback
effects (a behavioural choice by one agent makes similar choices on the part of other
agents more likely) and increasing returns (if a firm duplicates it’s resources, it’s pro-
duction is more than duplicated). These systems are typically evolving processes,
endemically nonlinear when positive feedback phenomena are present, and path de-
pendent (i.e. nonergodic).
2.1.3 Stylized facts of finance
Share prices and foreign exchange rates (or their logarithms) obey three stylized
facts[77] which every candidate model should display: price variations are uncorre-
46
lated (thus, one is unable to reject the hypothesis that financial prices follow a random
walk or a martingale); on the other hand, the amplitude of price variation (volatil-
ity) is non-homogeneous and its correlations are very long-ranged (these periods of
quiescence and turbulence are known in the finance literature as volatility clusters);
finally, the returns unconditional distributions are fat tailed (they decay slower than
a Gaussian).
2.2 A physicist’s approach to modelingmarket phe-
nomena
Since the 1980s, the deterioration of the academic job market in physics has been
attracting a large number of physicists to investment banks: many of them now
working as ”quants”, designing sophisticated new derivative products or developing
numerically intensive data analysis techniques for price and volatility forecasting.
More recently, several teams of physicists have launched their own firms, offering
services in the fields of financial software design and forecasting.
There exists however another set of motivations – scientific ones – which have
also been prompting theoretical physicists to become interested in economics and
finance. Although this phenomenon may seem a bit mysterious to the outsider, we
will attempt to convince the reader that it is not: economic and financial systems
may well be considered as objects of high interest for researchers with a background
in theoretical physics.
2.2.1 Methodology
The modeling of financial markets has multiple facets, depending on the level
of description and the objectives one has in mind. At an empirical level, financial
47
econometrics analyzes the statistical properties of various time series obtained from
market data. Once some statistical regularities (if any) have been identified, one
would like to build theoretical models of markets that are able to explain their origin
in terms of economic behaviour of market participants and their interactions (one
tries to explain the behaviour of macrovariables – here: prices, trading volumes,
etc. – starting from microvariables, i.e. individual agents behaviour. If the number
of agents is large, then we obtain a setting similar to that of the thermodynamic
limit, which has prompted several authors to draw analogies with interacting particle
systems. However, contrarily to the case of physical systems where homogeneity of the
individual components often simplifies the analysis, in economic systems heterogeneity
of behaviour and beliefs seems to be the rule rather than the exception.
Last but not least, an important difference between physical and socioeconomic
systems is the role of anticipations in the later: the agents beliefs determine their
behaviour and therefore the state of the market variables which in turn influences
these beliefs. This important point indicates the necessity of taking into account
various feedback effects when modeling market behaviour.
The approach we propose is not grounded on statistical mechanics (as most of
the work on ’econophysics’ has been), but is built upon high dimensional dynamical
systems. The later provides a toolbox of concepts and methods which may be applied
to analyze the behaviour of large systems given a rudimentary description of their
individual components.
High dimensional dynamical systems shall thus be combined with statistical tools
in a complementary way.
48
Random Dynamical Systems and On-off Intermittency
Random maps exhibit very interesting features ranging from stable or quasi-stable
behaviours, to chaotic behaviours and intermittency. In particular on-off intermit-
tency is an aperiodic switching between static, or laminar, behaviour and chaotic
bursts of oscillation. It can be generated by systems having an unstable invariant
manifold, within which it is possible to find a suitable attractor (i.e. a fixed point).
On-off intermittency in dynamical systems is triggered by the repeated variation
of one dynamical variable through the bifurcation point of another dynamical vari-
able. The first variable acts as a time-dependent parameter, while the response of
the second variable comprises the intermittent signal[60]. This mechanism can be
easily recreated in dynamical systems that are skew products[100], including random
dynamical systems.
A random map can be defined in the following way. Denoting with x(t) the state
of the system at discrete time t, the evolution law is given by
x (t+ 1) = f (x (t) , J (t))
where J (t) is a random variable. Let us discuss a one-dimensional randommap which,
in spite of its simplicity, captures some basic features of this kind of systems[13]:
x (t+ 1) = atx (t) (1− x (t)) (2.1)
where at is a random dichotomous variable given by
at =
4 w.p. p
1/2 w.p. 1− p(2.2)
For x (t) close to zero, we can neglect the non-linear terms to obtain
x (t) =t−1Yj=0
ajx (0)
49
From the law of large numbers one has that the typical behaviour is
x (t) ∼ x (0) ehln ait
Since hln ai = p ln 4 + (1− p) ln 1/2 = (3p− 1) ln 2, one has that for p < pc = 1/3,
x (t) → 0 for t →∞. On the contrary, for p > pc, after a certain time, x (t) escapesfrom the fixed point zero and the non-linear term becomes relevant. Figure 2.1 shows
a typical on-off intermittency behaviour for p slightly larger than pc. Note that, in
spite of this irregular behaviour, numerical computations show that the Lyapunov
Exponent λJ is negative for p < epc ' 0.5: this is essentially due to the non-linear
terms.
There seems to be growing accumulating reason to believe that chaotic transitions
through on-off intermittency could be one of the routes to chaos of random dynamical
systems as period doubling, intermittency, quasiperiodicity, and crises are those of
deterministic dynamical systems[104].
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 500 1000 1500 2000 2500 3000 3500 4000
x t
t
Figure 2.1: x (t) versus t for the random map 2.1 with at defined by 2.2.
50
Extreme value theory
Starting in the early nineties, a new type of analysis has appeared in empirical
finance, that highlights typical features by abstracting from a specific distributional
shape and concentrating on the behaviour of the tails instead of trying to fit the
entire distribution. The theoretical background to this research program is provided
by statistical extreme value theory.
Statistical extreme value theory asserts that under very mild conditions the dis-
tribution of (appropriately scaled) extreme realizations (sample maxima or minima)
of a time series of IID observations converges to one out of three types of so-called
extreme value distributions.
For concreteness, consider a stationary sequence of IID variables X1,X2, . . . , Xn,
e.g. Xi = log(Pi), where Pi is the price of a single asset at time i. Define the
maximum:
M = max (X1,X2, . . . , Xn)
Under widely applicable conditions, the distribution function (df) of M , F , can be
approximated by a member of the following class of extreme value distributions[37]:
Hξ,µ,ψ = exp
(−µ1 + ξ
x− µψ
¶−1/ξ+
)
where (·)+ = max(·, 0). This three—parameter family of distributions has a locationparameter µ ∈ R, a scale parameter ψ > 0 and (most importantly) a shape parameterξ ∈ R . The case ξ = 0 is to be interpreted as
H0,µ,ψ = exp
½− exp
µ−x− µ
ψ
¶¾, x ∈ R
and is referred to as the double exponential or Gumbel distribution. For ξ > 0, Hξ,µ,ψ
is called the Fréchet distribution, for ξ < 0 the Weibull. An important distinction is
51
that the Fréchet has unbounded support to the right and theWeibull has unbounded
support to the left. Figure 2.2 contains the density functions for the standard cases
Hξ,0,1 for ξ = 0, ξ = 2/3, ξ = −2/3.
0 10
0.00.10.20.30.40.50.
Weibull
Gumbel
Frechet
Figure 2.2: Densities of the extreme value distributions.
From Figure 2.2, we see the typical skew behaviour of extreme value distributions.
Moreover, in the case ξ > 0 which is most important for finance, the tail 1−Hξ,0,1 (x)
behaves like x−1/ξ, i.e. is fat—tailed. In order to be clear about the significance of the
extreme value pdfs and their link to the normal pdf, observe that for X1, . . . , Xn IID,
N (µ,σ2), and Mn = max (X1, X2, . . . ,Xn)
p (Mn > x) ≈ pµx− bnan
¶= H0,0,1
µx− bnan
¶for suitable sequences (an) and (bn) which can be calculated explicitly as functions of
n, µ and σ. Hence the two—sided, skew Gumbel pdf approximates the law governing
the largest observation in a normal sample. A similar result, with different (an) and
(bn)’s holds for instance for exponential and lognormal data.
A popular estimator of ξ is the Hill estimator obtained as follows. Suppose one
observes X1, . . . ,Xn and orders these observations as
X(1) ≥ . . . ≥ X(n)
52
The Hill estimator based on k + 1 upper order statistics is[33]
Hk,n :=1
k
kXi=1
logX(i)X(k+1)
for k = 1, . . . , n− 1.The Hill estimator is consistent for ξ in the following sense: if (kn)n∈N is an
intermediate sequence, that is,
kn →∞, kn/n→ 0 (2.3)
and Xn is IID, then
Hkn,nP−→ ξ (2.4)
Because of condition (2.3) on the number of order statistics, it is not clear how to
apply the consistency result (2.4). What is done in practice is to construct a Hill
plot, defined as ©¡k,H−1
k,n
¢, 1 ≤ k ≤ n− 1ª
and then try to infer the value of ξ from the stable region in the graph.
2.2.2 Lévy distributions and paretian tails
Lévy distributions (noted Lµ(x) below) appear naturally in the context of the
central limit theorem, because of their stability property under addition (a property
shared by Gaussians). The tails of the Lévy distributions are however much ’fatter’
than those of Gaussians, and are thus useful to describe multiscale phenomena (i.e.
when both very large and very small values of a quantity can commonly be observed).
An important constitutive property of these Lévy distributions is their power-law
behaviour for large arguments, often called ’Pareto tails’:
Lµ(x) ∼ µAµ±|x|1+µ for x→ ±∞
53
1.1
1.2
1.4
1.8
0 1000 2000 3000 4000 5000number of order statistics
Hill
Estimator
Figure 2.3: Hill plot of 5000 Pareto observations, γ = 1.
where 0 < µ < 2 is a certain exponent (often called α), and Aµ± two constants which
we call tail amplitudes, or scale parameters: indeed, Aµ± gives the order of magnitude
of the large (positive or negative) fluctuations of x. One can of course observe Pareto
tails with µ ≥ 2, however, those tails do not correspond to the asymptotic behaviourof a Lévy distribution.
If one adds random variables distributed according to an arbitrary law P1 (x1), one
constructs a random variable which has in general a different probability distribution
(P (x,N) = [P1 (x1)]∗N). However, for certain special distributions, the law of the sum
has exactly the same shape2 as the elementary distribution —these are called stable
laws. The distribution of increments on a certain time scale is thus scale invariant,
provided the variable X is properly rescaled. The family of all possible stable laws
coincide (for continuous variables) with the Lévy distributions defined above3, which
2The fact that two distributions have the ’same shape’ means that one can find a (N dependent)translation and dilation of x such that the two laws coincide:
P (x,N) dx = P1 (x1) dx1 where x = aNx1 + bN
3For discrete variables, one should also add the Poisson distribution.
54
include Gaussians as the special case4 µ = 2.
The power-law distribution of wealth discovered by Vilfredo Pareto (1848—1923)
in the 19th century predates any power laws in physics. And indeed, since Pareto, the
existence of power laws has been controversial. One underlying reason is that power-
law probability distributions are necessarily approximations. An inverse power-law
cumulative distribution f (r) ∼ |r|−α with an exponent α > 0 is not integrable at
zero, and similarly, with an exponent α ≤ 0, it is not integrable at infinity. Thus,a power-law probability distribution cannot be exactly true for a variable with an
unbounded range. When they apply at all, power-law distributions are necessarily
only part of a more complete description, valid within certain limits.
2.2.3 Empirical statistical regularities in prices
The distribution of price fluctuations is one of the most basic properties of mar-
kets. For some markets the historical data spans a century at a daily timescale,
and for at least the last decade every transaction is recorded. Nonetheless, the price
distributions functional form is still a topic of active debate. Naively, central-limit
theorem arguments suggest a Gaussian (normal) distribution.
If p(t) is the price at time t, the log-return rτ (t) is defined as rτ (t) = log p(t+ τ)−log p(t). Dividing τ into N subintervals, the total log-return rτ (t) is by definition the
sum of the log-returns in each subinterval. If the price changes in each subinterval
4The characteristic function of a symmetric Lévy distribution is
bLµ (z) = exp (−aµ |z|µ)Thus in the limit µ = 2 one recovers the definition of a Gaussian with zero mean. Recall thatcharacteristic function of a Gaussian with mean m and variance σ2 is
bPG (z) = expµ−σ2z22
+ imz
¶
55
are independent and identically distributed (IID) with a well-defined second moment,
under the central limit theorem the cumulative distribution function f(rτ ) should
converge to a normal distribution for large τ . This approach was inaugurated by
Louis Bachelier who first introduced the idea that stock market prices behave as a
random walk and who considered Brownian motion as a candidate for modeling price
fluctuations.
For real financial data, however, convergence is very slow. While the normal
distribution provides a good approximation for the center of the distribution for large
τ , for smaller values of τ there are strong deviations from normality. This is surprising,
given that the autocorrelation of log-returns is typically very close to zero for times
longer than about 15 to 30 minutes[81, 14]. What is the nature of these deviations
from normality and what is their cause?
The Efficient Market Hypothesis
Although it turned out later that the Normal distribution can usually be over-
whelmingly rejected by explicit statistical tests, due to its familiarity it is still often
used in theoretical work as well as by market practitioners. But why should we as-
sume that single price changes are governed by random motion? Shouldn’t there
be systematic economic factors behind price variations in financial markets? Thus,
before looking at empirical results, we will shortly outline the economic reasoning
behind the IID assumption for price changes.
For a long time, the only available theoretical background to the statistical be-
haviour of asset prices has been the so-called Efficient Market Hypothesis (EMH)[39].
It states that, at any point in time, asset prices should reflect the discounted expected
stream of earnings from holding the underlying asset. Thereby, the expectations are
to be understood in a mathematical sense as the expected future values affected by
56
uncertainty which are computed using all presently known information about future
circumstances.
Although mutatis mutandis the efficient market hypothesis can be applied to other
financial markets as well (for example, foreign exchange, precious metal or future
markets), we will concentrate on share markets for ease of exposition. In this case,
future earning streams are most easily identified as dividend payments. The price
predicted by the EMH should, therefore, be given by:
pt =∞Xi=1
δiE [dt+i|It] (2.5)
In (2.5) pt denotes the price at time t, dt+i denotes the dividend paid in a future
period t + i, and δ < 1 is a discount factor reflecting the fact, that the expected
payment of one dollar in the future has less value today than immediate availability
of the same amount. E[.] finally, is the mathematical expectation operator which is
computed conditional on the current information set It.
We can also express the asset pricing equation in a somewhat different manner if
we do not assume that the investor will hold the asset forever, but rather considers
to sell it at time t+ 1 in order to consume his receipts or make another investment.
Under this perspective, future earnings are composed of the dividend plus the selling
price at t+ 1 which are both uncertain in t. Hence, the price should equal:
pt =N−1Xi=1
δiE [dt+i|It] + δNE [pt+N + dt+N |It] (2.6)
It is, in fact, easy to show that both variants are identical. However, there is
a slight difference between equations (2.5) and (2.6): in order to make sure that
the sum on the RHS of (2.5) converges, we have to assume that the mathematical
’transversality condition’ limi→∞ δiE [pt+i] → 0 holds. Although this condition may
appear innocuous, its violation allows for price paths which give rise to so-called
rationally expected speculative bubbles.
57
Let us assume that the transversality condition holds and prices can be expressed
by (2.5). Then, price changes come about by arrival of some new items of informa-
tion about future dividends (i.e. an increase of the information set) which changes
the rationally computed expectations of future prices and dividends. Of course, such
information continuously hits the market: one can, for example, think of such ele-
mentary items as weather or climatic events affecting output in certain industries,
political changes (like introduction of new taxes on emissions, changes of monetary
policy) as well as any kind of firm-specific elements (e.g. all types of news about
management policies). According to the EMH, the entirety of all these factors would
explain the history of price changes: the news arrival process is crucial for and, in
fact, governs the characteristics of the distribution of price increments. As ’news’
should be events that are independent of the preceding ones, the ’independence’ part
of the IID assumption can be immediately justified by the above considerations. If
we are prepared to accept the second part as well (identical distribution), we end up
with the central limit law for low-frequency price changes as the aggregate of high-
frequency ones. Note that under the EMH, the distribution of price changes is a
mere reflection of the distribution of arriving news (which, however, is an aggregate
of diverse individual events that cannot be quantified). Scaling laws for price changes
would, therefore, have to be explained by similar scaling laws for the news arrival
process.
Scaling properties of financial data
The actual distribution of log-returns has fat tails. That is, there is a higher prob-
ability for extreme values than for a normal distribution. As one symptom of this, the
fourth moment is larger than expected for a Gaussian. We can measure this deviation
58
in a scale-independent manner by using the kurtosis5 κ =(r − hri)4® / (r − hri)2®2
(hi indicates a time average). In the early 1960s, Benoit Mandelbrot (now famousas the grandfather of fractals) and Eugene Fama (now famous as the high priest of
efficient market theory) presented empirical evidence that f was a stable Lévy distri-
bution. Based on daily prices in different markets, Mandelbrot and Fama measured
µ ' 1.7, a result that suggested that short-term price changes were ill-behaved: if thevariance doesn’t exist, most statistical properties are ill defined. Furthermore, at the
time of publication of Mandelbrot’s and Fama’s papers, almost no statistical method
was known for estimating the parameters and testing goodness-of-fit of the Lévy dis-
tributions. Nevertheless, the argument of convergence of aggregate returns towards
one of the stable laws appeared so convincing, that a number of researchers accepted
indication of α < 2 as evidence in favour of Lévy-stable distributions without further
testing of fit.
Subsequent studies demonstrated, however, that the behaviour is more compli-
cated than this[93, 80, 75, 78, 53, 101]. Indeed, the intermittency of the process (i.e.
a measure of how heavy the tails are) is difficult to measure quantitatively, since by
definition one is observing rare events. First, for larger values of τ , the distribution
becomes progressively closer to normal. Second, investigations of larger data sets
make it clear that large returns asymptotically follow a power law f(r) ∼ |r|−α, withα > 2. This finding is incompatible with the Lévy distribution. The difference in the
value of α is very important: with α > 2, the second moment (the variance) is well
defined. A value 2 < α < ∞ is incompatible with the stable Lévy distribution and
indicates that simply generalizing the central limit theorem with long tails is not the
correct explanation.
5The kurtosis is defined such that κ = 0 for a Gaussian distribution, a positive value of κindicating a slow asymptotic decay of the PDF.
59
Physicists have contributed by studying really large data sets and looking at the
scalings in close detail. A group at Olsen and Associates, led by Michel Dacorogna,
studied intraday price movements in foreign exchange markets[30]. Another group
at Boston University, led by Rosario Mantegna and Eugene Stanley, studied the
intraday movements of the S&P index[80, 53]. More recently, they studied the five-
minute returns of 1,000 individual stocks traded on the AMEX, NASDAQ, and NYSE
exchanges, over a two-year period involving roughly 40 million records [12]. In this
case, they observed the power-law scaling over about 90 standard deviations. For
larger values of |r|, these results dramatically illustrate that the pdf is approximatelya power law with α ' 3. Thus, the mean and variance are well-defined, the kurtosisclearly diverges, and the behaviour of the skewness is not so clear.
Most of the tools used by physicists (scaling laws[28], wavelets and multiresolu-
tion analysis[7], neural networks and nonlinear prediction methods[44]) are beyond
the scope of the statistical framework of econometrics and have not been used by
mainstream economists but offer interesting potentialities for the analysis of financial
time series. However with a few exceptions they mostly deal with ”unconditional”
objects (unconditional probability distributions, correlation functions, etc.) which are
the quantities often used in the stochastic description of physical systems, while the
econometric approach has mostly focused on conditional probabilities and forecasts.
Also, whereas many stochastic models in physics have a theoretical underpinning in
’first principles’ or physical models, this is not the case in finance, implying that one
should be cautious when drawing conclusions based on purely statistical observations.
Indeed, sloppy statistical analysis has led to mistakes in the past. In the 1980s,
there was considerable interest in the possibility that price changes might be described
by a low-dimensional chaotic attractor. More careful statistical analysis by José
60
Scheinkman and Blake LeBaron showed that the claims of low-dimensional chaos in
price series were not well-justified [108].
The power law for large price moves is a very different story. To detect a chaotic
attractor based on its fractal dimension in state space requires a test of the distri-
butions fine-grained, microscopic properties. Low-dimensional chaos is a very strong
hypothesis, because it would imply deep structure and short-term predictability in
prices. A power law in the tails of the returns, in contrast, is just a statement about
the frequency of large events and is a much weaker hypothesis. As we have discussed
above, this becomes clear in the context of extreme value theory, the data clearly
pointing to a fat tail decaying as a power law with α > 2[80, 53, 101, 81]. Sur-
prisingly, this implies that the price-formation process cannot be fully understood
in terms of central limit theorem arguments, even in a generalized form. Power-law
tails do obey a sort of partial central limit theorem: For a random variable with
tail exponent α, the sum of N variables will also have the same tail exponent α[45].
This does not mean that the full distribution is stable, however, because the distri-
butions central part, as well as the power laws cutoff, will generally vary. The fact
that the distributions shape changes with the time window considered makes it clear
that the random process underlying prices must have nontrivial temporal structure.
This complicates statistical analysis of prices, both for theoretical and practical pur-
poses, and gives an important clue about the behaviour of economic agents and the
price-formation process. But unlike low-dimensional chaos, it does not imply that the
direction of short-time price movements is predictable.
Volatility and intermittency
Another feature of asset returns is that, although they are serially uncorrelated
(i.e. the two point correlation is not significantly different from zero), their absolute
61
values or squares show positive two-point correlation functions and in particular they
decrease slowly to zero, pointing to a possible long-range dependence. This means
that the amplitudes of successive price movements are persistent but not necessarily
their signs. This phenomenon, called ”volatility clustering”, is a universal feature
observed in many different markets.
While the autocorrelation of log-returns, ρ (τ) ∼ hrτ (t+ τ ) rτ (t)i, is generallyvery small on timescales longer than a day, this is not true for the volatility (which can
be defined, for example, as r2 or |r| ). The volatility on successive days is positivelycorrelated, and these correlations remain positive for weeks or months. Clustered
volatility can cause fat tails in the pdf of the returns. For example, the sum of
normally distributed variables with different variances has a high kurtosis (although
it does not have power-law tails) [14]. To understand the statistical properties of
price changes, we need a more sophisticated model that accounts for the probability
distributions temporal variation.
Clustered volatility is traditionally described by simple ad hoc time-series models
with names that include the letters ARCH (for AutoRegressive Conditional Hetero-
scedasticity)[16]. Such models involve linear relationships between the square or ab-
solute value of current and past log-returns. Volatility at one time influences volatility
at subsequent times. ARCH-type models can be effective for forecasting volatility,
and there is a large body of work devoted to problems of parameter estimation, vari-
ations on the basic model, and so forth. ARCH models are not compatible with all
of the empirical properties of price fluctuations, however.
A good price-fluctuations model should connect the behaviour onmultiple timescales.
A natural test is the behaviour of moments, in this case h|rτ |qi as a function of q andτ . Early research reports approximate power-law scaling with τ , with different slopes
for each value of q [51]. In the jargon of dynamical systems theory, this suggests a
62
fractal random process. A slope that is a linear function of q implies a simple fractal
process, and a slope that is a nonlinear function of q implies a multifractal or mul-
tiscaling process. Indeed, several different calculations seem to show that the slope
varies nonlinearly with q, suggesting that the price process is multifractal.
One thing that does seem clear is that conventional ARCH-type models are incom-
patible with the scaling properties of price fluctuations [81, 14]. While ARCH-type
models can indeed give rise to fat-tailed probability distributions with α > 2, they
cannot explain other properties of the price fluctuations. ARCH-type models fit at
a given timescale τ do not appear to do a good job of explaining the volatility at a
different timescale. Furthermore, conventional ARCH models do not have asymptotic
power-law decay in the volatility autocorrelation function. The most likely explana-
tion is that ARCH models are misspecified – their simple linear structure is not
general enough to fully capture the real temporal structure of volatility. Given that
they are completely ad hoc models, this is not surprising.
There are still missing pieces and several open questions to be answered before
we will have a good random-process model linking the behaviour of prices across a
range of different time scales. Physicists have contributed to the theory and data
analysis leading to the current understanding. But to have a good theory of how
prices behave, we will need to explain the behaviour of agents on whom they depend.
2.2.4 Agent-based models
The elementary building blocks of financial markets are human agents, each buying
and selling based on his or her own point of view. To have a full and fundamental
description of markets requires models that take intelligence, learning, and complex
decision making into account, not to mention emotion, irrationality, and intuition.
The traditional view in economics is that financial agents are completely rational
63
with perfect foresight. Markets are always in equilibrium, which in economics means
that trading always occurs at a price that conforms to everyone’s expectations of
the future. Markets are efficient, which means that there are no patterns in prices
that can be forecast based on a given information set. The only possible changes in
price are random, driven by unforecastable external information. Profits occur only
by chance. Traditional economic theory is furthermore based on the concept of the
”representative agent”, a (hypothetical) individual whose profit- or utility-maximizing
behaviour reproduces the main features of aggregate (macro) variables such as price,
supply and demand.
The behavioural economists have presented evidence of irrational behaviour and
market anomalies that historically would have allowed excess profits. Anecdotal evi-
dence suggests that some individuals might indeed make statistically significant excess
profits.
It is fair to say that the physicists studying these problems tend toward the more
radical end of the spectrum. While bounded rationality is a nice idea, it is only part
of the story. People are not identical finite-capacity calculating machines differing
only in their utility functions. Equally important is the diversity of viewpoints in-
duced by nature and nurture. Formulating successful predictive models is extremely
difficult and requires both hard work and intelligence. To make a good model, it is
necessary to specialize, which stimulates diversification of financial strategies. As a
result, financial agents are very heterogeneous. Some agents are more skilled than
others, and the excess profits of such agents are not necessarily reasonable. The
behavioural economists are clearly right that people are not fully rational and that
this can play an important role in setting prices. But where do we go from there?
Despite the idiosyncrasies of human psychology, is there a statistical mechanics that
can explain some of the statistical properties of the market, and perhaps take such
64
idiosyncrasies into account? Agent-based modeling offers one approach to addressing
these problems[70] and its approach bears implications very different from the rep-
resentative agent methodology. While the analysis and the results are qualitatively
identical at both the micro- and the macro-level in representative agent models, in
an agent-based model the properties of macro variables are not necessarily identi-
cal to those of the corresponding micro-variables nor does the mere aggregation of
micro-components always yield sensible macroeconomic relationships.
Efforts in this direction range from simple, metaphorical models, such as those of
evolutionary game theory, to complicated simulations, such as the Santa Fe Institute
stock market model [71, 10]. The SFI model, which was a collaboration between two
economists, a physicist, and a computer scientist, was a significant accomplishment.
It demonstrated that many of the dynamical properties of real markets, such as
clustered volatility and fat tails, emerge automatically when a market simulation
allows the views of the participants to be dynamic. It was a good start, but in
part because of the complexity of the numerical simulations, it left many unanswered
questions.
Many dynamic trading models attempt to use strategies patterned after strate-
gies actually used in real markets (as opposed to arbitrary, abstract strategies that
have often been used in other game-theoretic models). The decisions are based on
information internal to the market, such as prices and their history, and possibly in-
formation external to the market, and can be public (such as prices) or private (such
as conversations between traders).
Despite the wide variation in financial-trading strategies, we can classify many of
them into broad groups. Strategies that depend only on the price history are called
technical trading or chartist strategies. Trend-following strategies are a commonly
used special case of technical strategies in which the holdings of the asset positively
65
correlate with past price movements. Value or fundamental strategies are strategies
based on perceived value (a model for what something ought to be worth), as opposed
to its current price. Value strategies tend to buy when an asset is undervalued and
sell when it is overvalued.
Many authors have studied trend and value strategies, with a variety of differences
in the details of the implementations. Certain conclusions seem to emerge that are
independent of the details. For example, trend strategies tend to induce trends and
therefore positive autocorrelations in the price [40], which was also evident in earlier
work by economists [32]. Several new features, however, are apparent with a simple
price formation rule that were not recognized in earlier studies by economists because
of the more cumbersome framework they used to formulate the problem. For exam-
ple, trend-following strategies also induce oscillations because, to keep risk bounded,
strategies are formulated in terms of positions (holdings), whereas changes in price
are caused by orders, which are changes in positions. Thus, the price dynamics have
second-order, oscillatory terms[40]. Earlier studies also showed that trends are self-
reinforcing – trend strategies tend to induce trends in the market [32]. Some have
mistaken this to mean ”the more the merrier” that the profits of trend strategies are
enhanced by other identical trend strategies. A more careful analysis disproves this.
While trend strategies indeed create profit opportunities, these opportunities are for
other trend strategies, not for the same trend strategy.
A study of value-investing strategies also shows some interesting results. Not sur-
prisingly, most sensible value strategies induce negative autocorrelations in the price.
Some value strategies cause prices and perceived values to track each other. But
surprisingly, many sensible value strategies do not have this property[40]. Another
interesting set of questions concerns the case where the perceived values are heteroge-
neous (that is, people have different opinions about what something is worth). If all
66
the strategies are linear, the market behaves just as though there were a single agent
whose perceived value is the mean of the all the diverse values. If the strategies are
nonlinear, however, the diversity of views results in excess volatility.
Preliminary results suggest that it might be possible to provide a quantitative
explanation for some of the statistical properties of prices. For example, simulations
by Thomas Lux and Michele Marchesi [78] (an economist and a physicist) use trend
following and value-investing strategies. They let agents switch from one group to
the other. They observe power-law scaling in the tails of the log-returns,with a tail
exponent α ≈ 3, similar to that observed in real data. Other simulations also findpower-law scaling in the tails in models that allow diffusion of private information[61].
The traditional approach in mathematical finance models is to start by specifying
a stochastic process which represents the evolution of market prices. The investor
or portfolio manager is then faced with an optimization problem in this random
environment, which one proceeds to solve in order to obtain the strategy that the
agent should optimally adopt. In this point of view, the price process is exogenous,
i.e. given a priori and not affected by the action undertaken by the agents: the market
price is said to be perfect elastic on demand.
In a real market where agents are not only price takers but also price makers,
market supply and demand depend on the strategies adopted by the agents which
will in turn affect the evolution of prices. Given that the agents use as an input their
observations of past price movements, we are in the presence of a feedback mechanism
where prices and strategies mutually influence each other. The properties of the price
process must therefore be endogenously determined as the result of this feedback
mechanism instead of being postulated a priori.
All of these models show variations in the price, reminiscent of boom-bust cycles.
One source of these irregular cycles is the interaction between trend and value strate-
67
gies. We can describe one such scenario roughly as follows: Suppose the market is
strongly undervalued. The value investors buy, thereby raising the price. As the price
rises, it creates a trend, causing trend followers to buy, raising the price even further.
As the price becomes overvalued, the value investors sell, the trend is damped, and
the trend followers eventually sell, and so on. In practice, there are many other effects
and the resulting oscillations are highly irregular.
An economist would criticize the results in the above cited studies for several
reasons. A minor point concerns questions of whether the price-formation process
some of these models use is sufficiently realistic. Use of an entirely ad hoc price-
formation rule might cause inefficiencies or spurious dynamical effects.
Perhaps more persuasively, all of these dynamic trading models have qualita-
tive features such as clustered volatility and fat tails in common, even though the
strategies and in some cases the price formation rules are quite different. So far,
no careful studies have compared different methods of price formation or determined
which market properties depend on the method of price formation. Perhaps the main
contribution of physicists here has been the use of really simple methods for price
formation, within which the dynamics are obvious and simple examples are easily
solved analytically.
As a stronger criticism, many of these models contain only limited mechanisms
for individual agents to learn, or perhaps more important, selection mechanisms that
allow the agent population as a whole to learn. With appropriate selection mecha-
nisms under standard dogma, the market should become efficient and the price should
be random (or at least random enough that it is impossible for any agent to make
profits).
This study of dynamic agent-based trading models is still in its infancy, and many
interesting problems remain to be addressed. To make these models more convincing,
68
more work is needed to make them more realistic and better grounded in economic
theory.
2.3 Farmer’s model
The model builds upon a vision of the market based on interaction between micro
and macro dynamics. The former is the dynamics of individual agents, who invest
according to fundamental or technical strategies. The later is the dynamics of fun-
damental value, a random walk for the scope of this chapter, and prices. Traders do
not communicate with each other; instead they interact with the market maker and
base their decisions upon present and past values of the price or fundamental value.
Therefore, traders’ decisions (the finegrained variables) are influenced by prices and
fundamental values (the coarse grained variables), and the following price is deter-
mined by traders’ investments. The market is born out of this feedback loop between
finegrained and coarsegrained variables.
2.3.1 Derivation of the market impact function
Let there be N directional traders, labeled by the superscript i, holding x(i)t shares
at time t. Although this is not necessary, Farmer[40] assumes synchronous trading at
times . . . , t−1, t, t+1, . . . . Let the position of the ith directional trader be a functionx(i)t+1 = x(i)
³Pt, Pt−1, . . . , I
(i)t
´, where Pt is the price of a single asset at time t and
I(i)t represents any additional external information. The function x(i) can be thought
of as the strategy or decision rule of agent i. The order ω(i)t is determined from the
position through the relation
ω(i)t = x
(i)t − x(i)t−1 (2.7)
A single timestep in the trading process can be decomposed into two parts:
69
1. The directional traders observe the most recent prices and information at time
t and submit orders ω(i)t+1.
2. The market maker fills all the orders at the new price Pt+1.
To keep things simple, we will assume that the price Pt is a positive real num-
ber, and that positions, orders, and strategies are anonymous. This motivates the
assumption that the market maker bases price formation only on the net order
ω =NXi=1
ω(i)
The algorithm the market maker uses to compute the fill price for the net order is an
increasing function of ω
Pt+1 = f (Pt,ω)
Note that because orders are anonymous, with more than one trader the fill price is
unknown to individual agents when orders are placed.
An approximation of the market impact function can be derived by assuming that
it is of the form
f (Pt,ω) = Ptφ (ω) (2.8)
where φ is an increasing function with φ (0) = 1. Taking logarithms and expanding
in a Taylor’s series, providing the derivative φ0(0) exists, to leading order
logPt+1 − logPt ≈ ω
λ
This functional form for φ will be called the log-linear market impact function. λ is
a scale factor that normalizes the order size, and will be called the liquidity. It is the
order size that will cause the price to change by a factor of e, and can be measured
in either units of shares or money.
70
For an equilibrium model the clearing price depends only on the current demand
functions. For a general nonlinear price formation rule, in contrast, the price at
any time depends on the full sequence of previous net orders. The log-linear rule is
somewhere in between: the price change over any given period of time depends only
on the net imbalance during that time. In fact, we can make this a requirement,
and use it to derive the log-linear rule: suppose we require that two orders placed in
succession result in the same price as a single order equal to their sum, i.e.
f (f (P,ω1) ,ω2) = f (P,ω1 + ω2) (2.9)
By grouping orders pairwise, repeated application of (2.9) makes it clear that the
price change in any time interval only depends on the sum of the net orders in this
interval. Substituting (2.8) into (2.9) yields
φ (ω1 + ω2) = φ (ω1)φ (ω2)
This functional equation has the non-constant6 solution
φ(ω) = akω (2.10)
Putting equation (2.10) into (2.8) and taking logarithms gives
logPt+1 = logPt +ω
λ(2.11)
where λ = 1k log a
is the market marker liquidity7. Notice that the uncertainty about
a in equation (2.10) is absorbed in the expression of λ as a scale factor.
6φ(ω) = 0 and φ(ω) = 1 are also solutions. These are not relevant as they do not satisfy thecondition that φ should be increasing.
7It is out of the scope of this text to consider a non-constant market maker liquidity.
71
The log-linear price formation rule is only an approximation. It is, however,
perhaps the simplest one that gives reasonable results. In addition to the path-
independence derived above, it has several other special properties. It simplifies the
attribution of profits by making it possible to decompose them pairwise based on
correlations of position. It also implies that total realized wealth is conserved in
closed systems. The demonstration of these properties is interesting in part because
it makes it clear how nonlinearities lead to path dependence, non-decomposability,
and non-conservation of realized wealth. The derivation above implicitly assumes
that the market impact is permanent. That is, the market impact at a given time
changes the price, and without any other changes in demand, this price change persists
indefinitely.
The assumption that the market impact function φ depends only on the excess
demand is equivalent to saying that the market is risk neutral. Real market makers
try to keep their positions as small as possible. This makes the price formation
process path dependent and time dependent. The assumption that φ depends only in
the excess demand also neglects the possibility that other forms of information, not
reflected by demand, might influence the price. This is modeled as a noise term in
(2.12).
There is an implicit assumption in this derivation that the market is symmetric,
i.e. that there is no a priori difference in the market impact function of buy and sell
orders. This should be a good assumption for currency markets, and many derivative
markets, but it is probably not a good assumption for stock markets, where there are
often asymmetries in the liquidity of buying and selling.
The most relevant related work in the market making literature[92] is the model
of continuous actions and insider trading of Kyle[69]. He derives a price formation
rule in which the prices (rather than the logarithm of prices) depend linearly on the
72
orders. This fails to satisfy the boundary condition that prices remain positive.
We can now write down a dynamical system describing the interaction between
trading decisions and prices. Letting pt = logPt, and adding noise, equation (2.11)
becomes
pt+1 = pt +1
λ
NXi=1
ω(i) (pt, pt−1, . . . , It) + ξt+1 (2.12)
The random term is added to account for possible external perturbations to the price
that are not driven by trading, such as news announcements or perceived arbitrage
possibilities in related markets. The dynamics of (2.12) are completely general. De-
pending on the collection of functions ω(i), defined in terms of the positions in (2.7),
equation (2.12) can have stable fixed points, limit cycles, or chaotic attractors, or
they can be globally stable.
The choice of a discrete time, synchronized trading process is a matter of conve-
nience. One could have used an asynchronous process with random updating, or a
continuous time Weiner process (which has advantages for obtaining analytic results).
The time ∆t corresponding to a single iteration should be thought of as the timescale
on which the fastest traders observe and react to the price, e.g. a minute to a day.
2.3.2 Price dynamics caused by common trading strategies
Because the market maker is built into the model, we can begin by studying
the price dynamics of each strategy by itself, and use this insight to understand the
heterogeneous situation where many strategies trade at once.
We will assume that (the logarithm of ) the perceived value is given by an exoge-
nous random process
νt+1 = νt + ηt+1
73
where ηt is a normal, IID noise process with standard deviation ση and mean µη. We
will also assume that all agents perceive the same value.
The position is the net holding of the asset, i.e. if the initial holding is zero, since
all orders are filled, the position is the sum of all previous orders.
Value investors
For the simplest class of value strategies the position of agent i is of the form
x(i)t+1 = x
(i) (νt, pt) = V(i) (pt − νt)
where V (i) is a generally decreasing function8 with V (i) (0) = 0. This class of strategies
only depends on the difference between (log) price and (log) fundamental value,
mt = pt − νt
Such a strategy takes a long position when the asset is underpriced (mt < 0) and a
short position if the asset is overpriced (mt > 0).
To first order the position can be approximated by
x(i)t+1 = −c(i)mt
Thus the order size can be approximated by
ω(i)n+1 ' −c(i)∆mn = −c(i)(rn − ηn) (2.13)
where ∆mn = mn −mn−1 (∆ is the difference operator), rn = pn − pn−1 is the one-period return and ηn = ∆νn is the random increment to value. The scale parameter
c(i) is proportional to the capital of trader i, and controls the size of his orders and
positions.8Generally decreasing means that V either decreases or remains constant, and is not constant
everywhere.
74
It can be shown[40] that the first autocorrelation of the log-returns, rt = pt−pt−1,is generally negative for the simple state-independent value strategy above. However,
the price and value are not cointegrated. That is, the price and value can wander
arbitrarily far apart, and the mispricing is a random walk[40]. This is a surprising
result, since it implies that, even if everyone agrees about value, the price does not
track value at all.
Trend strategies
Trend followers, also sometimes called positive-feedback investors, invest based
on the belief that price changes have inertia. A trend strategy has a positive (long)
position if prices have recently been going up, and a negative (short) position if they
have recently been going down. More precisely, a trading strategy is trend following
on timescale θ if the position has a positive correlation with past price movements on
timescale θ, i.e.
correlation (xt+1, (pt − pt−θ)) > 0
A strategy can be trend following on some timescales but not on others.
An example of a trend following strategy on timescale θ is
ω(j)n+1 = c
(j)(rn − rn−θ) (2.14)
It can be shown that trend followers induce positive short term autocorrelations
in the log-returns. Perhaps more surprising, they also induce negative terms in the
autocorrelation function on timescales θ. Such negative autocorrelations cause the
prices to oscillate[40].
The possibility for complicated behaviour of the autocorrelation function arises
because the strategy is formulated in terms of the position, but the market impact is
based on the order, so that there are second order, oscillatory terms in the dynamics.
75
This is true for any strategy, though this effect is particularly pronounced for trend
strategies.
2.4 The random linear Farmer’s model
The main proposed contribution of this chapter is to state conditions under which,
on the random linear Farmer’s model, the market dynamics generates a fat tailed
unconditional distribution for the returns with clustered volatility. Lux and Marchesi
[78] have answered this question for a nonlinear model. We would like to address
the question of how much of the observed financial time series can be explained
within the framework of a simple affine random dynamical system. As we will see,
although under reasonable conditions the returns’ statistics seems considerably well
adjusted with the above stylized facts (displaying ’realistic’[52] tails in the range 2 ∼ 4and volatility clustering), the price series displays extremely wild oscillations not in
agreement with a model were price and fundamental value cointegrate. The extent
up to which this is relevant may be questionable, as stock prices and dividends series
do not seem to cointegrate universally[107].
Farmer characterizes each agent by her capital and strategy. We propose an
additional parameter, the unconditional probability for the agent to be active at each
time step, p.
The introduction of an unconditional probability for the agents’ activity is jus-
tified on the following possible sources of uncertainty: inherent unpredictability of
agents’ behaviour (random preferences) and lack of knowledge on the part of the
econometrician (random characteristics)[15]. Therefore, we define the random linear
Farmer’s model as the model presented on[40] where agents i = 1 . . . N follow linear
position based value investor or trend follower strategies, ω(i), which are active with
76
probability (w.p.) p(i). Nevertheless, it is outside of the scope of this chapter to con-
sider dynamical effects caused by evolution or capital reinvestment. Indeed, traders
will have constant capital and strategies.
2.4.1 The strategies
We consider a system with Nv agents holding position based value strategies and
Nt agents with position based trend following strategies, so that N = Nv +Nt is the
total number of agents. Value investors are indexed 1, . . . , Nv and trend followers are
indexed Nv + 1, . . . ,N .
Replacing (2.13) and (2.14) on (2.12), defining the sums of the agents normalized
capital for each class of strategies as αvi = 1λ
PNvi=1 c
(i) for value investors and αtf =
1λ
PNj=Nv+1
c(j) for trend followers9, and the difference between the value investors’
and trend followers’ total normalized capital as ∆α = αvi − αtf , yields:
rn+1 = −∆α rn − α tfrn−θ + αvi ηn
pn+1 = pn + rn+1(2.15)
We follow Farmer and allow the trend follower’s θ to be agent dependent, that is
θ ≡ θ(i).
To the above framework, we add to each agent i a probability for her to trade
(either buy or sell) at each time step, p(i). Therefore, the market has a non-constant
(in time) number of participants. It will be useful to define the following
δθ(i),n = 1 if trend follower i has time lag θ(i) and
is active at time step n and 0 otherwise;
δi,n = 1 if trader i is active at time step n and 0 otherwise
(2.16)
9We will drop the subscripts vi and tf , whenever they are superfulous.
77
2.4.2 Value investors
On this section we assume that the only strategy present on the market is value
investing. Under the assumptions that the traders’ normalized capital and trading
probability are constants, we show that unconditional distribution of the returns can
display Pareto tails and determine an expression for the tail exponent. A simple
physical explanation is given for this effect.
Define the total value investors’ and trend followers normalized capital respectively
as αvin =PNv
i=1 α(i)δi,n and αtfn =
PNi=Nv+1
α(i)δi,n.
Consider αvin À αtfn ∀n ∈ N. This is the situation when value investors dominatethe market (for the scope of this section, we will consider that only value investors
are present, that is, N = Nv and define αn ≡ αvin ). Then equation (2.15) becomes
rn+1 = −αnrn + αnηn (2.17)
Let the agents’ normalized capital and trading probability be constants, respec-
tively α(i)n = α and p(i) = p. Then αn has a binomial distribution,
αn =
0 w.p. (1− p)N
α w.p. p(1− p)N−1N...
...
(N − 1)α w.p. pN−1(1− p)¡ NN−1
¢Nα w.p. pN
Define
Ik,k+1 =
#1
k+1pM (k+1)(0)
,1
kpMk(0)
"(2.18)
Where M (k)(0) is the kth moment of the distribution of αn. We will show that if
one chooses N, k ∈ N2, p ∈ ]0, 1[ and α ∈ Ik,k+1, then the tails of the limiting
78
distribution of rn are asymptotic to a power law with exponent −µ for a certainµ ∈ [k, k + 1] and p small enough.The proof will follow three steps. First we show (Lemma 1) that the Ik,k+1 are
ordered intervals. Second we show (Lemma 2) that for α ∈ Ik,k+1and k ∈ N, rnconverges in distribution to a unique limiting distribution. Finally, we state Kesten’s
Theorem (Theorem 3) which gives conditions for the presence of fat tails and permits
the determination of the tail exponent on one-dimensional affine random dynamical
systems. Although we do not determine a simple closed form expression for the tail
exponent, we show that it can be bounded by two adjacent integers (Proposition 4).
Lemma 1 ∃p0 ∈]0, 1[ ∀p < p0
1k+1pM (k+1)(0)
<1
kpM (k)(0)
(2.19)
Proof: The kth moment of αµn is given by
M (k)(0) =NXi=0
ikpi(1− p)N−i N
i
(2.20)
A Taylor series expansion of (2.20) in p to second order about the origin yields10
M (k)(0) =
Np k = 1
Np£1 + (2k−1 − 1)(N − 1)p¤+O3(p) k ≥ 2
So that ³M(k+1)(0)
M(k)(0)
´k=
³1+(2k−1)(N−1)p1+(2k−1−1)(N−1)p
´k=
³1 + 2k−1(N−1)p
1+(2k−1−1)(N−1)p´k (2.21)
We want to prove that for p small enoughµM (k+1)(0)
M (k)(0)
¶k> M (k)(0)
10We observe that only terms i < 3 contribute to a second order approximation in p.
79
Which is equivalent toµ1 +
2k−1(N − 1)p1 + (2k−1 − 1)(N − 1)p
¶k>¡1 + (2k−1 − 1)(N − 1)p¢Np (2.22)
Expanding both sides in a Taylor series in p about the origin, we obtain the relation
1 + 2k−1k(N − 1)p+O2(p) > Np+O2(p)
which, to first order in p, is equivalent to
N >2k−1k
2k−1k − 1 −1
(2k−1k − 1) p
which is valid for k > 1 and N > 1.
¤
Lemma 2 Choose N, k ∈ N2, p ∈ ]0, 1[ and α ∈ Ik,k+1. Then rn converges indistribution to a unique limiting distribution.
Proof:
The solution of (2.17) is given by[56]
rn =nXk=1
ÃnY
j=k+1
−αj!αkηk +
ÃnYj=1
−αj!r0
with the definition thatQnj=n+1−αj = 1. The Lyapunov number of (2.17) is given
by
hαni = αM (1)(0) = αNp (2.23)
If hαni < 1 (negative Lyapunov exponent)nYj=1
−αvij →n→∞
0
80
exponentially fast and under very weak conditions on the product αkηk [56], the
distribution of rn will converge independently of r0 to that of the series
nXk=1
ÃnY
j=k+1
−αj!αkηk (2.24)
Thus, if α < 1M(1)(0)
, rn converges in distribution to a unique limiting distribution,
that of 2.24. As the binomial moments M (k)(0) are increasing functions of k this
condition is satisfied for α ∈ Ik,k+1 ∀k ∈ N.¤
Equation (2.17) is well known to display Pareto tails. In fact, Kesten[65] states
the following
Theorem 3 (theorem Kesten) Consider a stochastic difference equation:
xt+1 = mt xt−1 + qt t = 1, 2, . . .
where the pairs mt, qt are i.i.d. real valued random variables.
If xt converges in distribution to a unique limiting distribution, qt/ (1−mt) is non-
degenerate and there exists some µ > 0 such that:
i) 0 < h|qt|µi <∞ii)h|mt|µi = 1iii)
|mt|µ log+ |mt|®<∞
then the tails of the limiting distribution are asymptotic to a power law, i.e. they obey
a law of the type
Prob (|xt| > y) ≈ c · y−µ (2.25)
The derivation of (2.25) uses results from renewal theory of large positive excur-
sions of a random walk biased towards −∞ (see [45], sections V I6 − 8, XI1, XI6,XII4b and XII5 for an outline of the proof when qt is positive). As can be observed
81
from (2.24), the additive term in (2.17), αnηn, provides a reinjection mechanism,
allowing rn to fluctuate without converging to zero, as it would if αnηn vanished[114].
Proposition 4 Choose N, k ∈ N2, p ∈ ]0, 1[ and α ∈ Ik,k+1. Then the tails
of the limiting distribution of rn are asymptotic to a power law with exponent −µfor a certain µ ∈ [k, k + 1] and p small enough, i.e. they obey a law of the typeP (|rn| > x) ≈ c · x−µ ∃µ ∈ [k, k + 1] such that
¡αvin¢µ®
= 1 (2.26)
Proof : If the agents’ normalized capitals and trading probabilities are constants,
respectively α(i)n = α and p(i) = p,
¡αvin¢µ®
= αµNXi=0
iµ pi(1− p)N−iµN
i
¶(2.27)
Theorem 3 requires the existence of a µ > 0 such that (2.26). Although the estimation
of µ from (2.26) is in general not possible, one can estimate an interval for µ. Choose
p ∈ ]0, 1[ and N, k ∈ N2. As the binomial moments M (k)(0) are continuous and
increasing functions of k, equation (2.26) is satisfied if there exists an α such thatD(αvin )
kE= αkM (k)(0) < 1D
(αvin )k+1E= αk+1M (k+1)(0) > 1
(2.28)
From (2.27) and (2.28), one deduces that α ∈ Ik,k+1.¤
Intuitively, heavy tails appear on the system as the expected value of the traders
capital is less than one ((2.17) has a negative Lyapunov exponent), but it attains
very high values as a massive number of traders simultaneously enter the market
(this happens intermittently), causing explosions in price.
82
Numerical calculations show that for N = 103, p = 9.5 · 10−3 and α = 0.1,
hαni = .95,(αn)
2® = 0.996598 and (αn)3® = 1.13478. Thus, one should expect a µvery close to 2.
Figure 2.4 is the modified Hill plot[33] for rn which numerically confirms these
calculations.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.5
1
1.5
2
2.5
3
Theta
1/H
Hill Estimator
Figure 2.4: Hill estimator for the tails of rn when N = 103, p = 9.5 ·10−3 and α = 0.1.
2.4.3 Trend followers and value investors trading on the crit-
ical point
In this section we will consider the case where trend followers and value investors
are both present on the market simultaneously.
83
To better understand (2.15), we notice that
rn+1 = 4mn+1 + ηn+1
so that the dynamics of the mispricing is
4mn+1 = −4α 4mn − αtf 4mn−θ − ηn+1 + αtf ηn − αtfηn−θ (2.29)
Equation (2.29) is a discrete dynamical system with delay structure. By making
the substitution
(X1n, . . . , X
θn,X
θ+1n )T ≡ (4mn−θ, . . . ,4mn−1,4mn)
T (2.30)
the dynamics of 4mn becomes
Xn+1 = Jm Xn + ηn (2.31)
where Jm is a (θ + 1)× (θ + 1) companion matrix,
Jm =
0 1
1
...
1
−αtf 0 . . . 0 −∆α
(2.32)
and
ηn,θ =
0
...
−ηn+1 + αtf ηn − αtf ηn−θ
(2.33)
84
The sum of the capital of active trend followers with time lag θ(i) at time step n
is
αtfn (θ(i)) =
NtXk=1
α(k)n δθ(i),n
and the difference in capital between (active) value investors and trend followers total
capital at time step n is
∆αn =NvXi=1
α(i)n δi,n −NX
j=Nv+1
α(j)n δj,n
Thus, matrix Jm(n) can be written as:
Jm(n) =
0 1 0 . . . 0
0 0 1 . . . 0
...
0 0 . . . 0 1
−αtfn (θmax) −αtfn (θmax−1) . . . −− αtfn (θ1) ∆αn
(2.34)
where θmax = maxθ(i) and θmax−j = maxθ(i)\θmax, . . . , θmax−j+1 Equation 2.31finally becomes:
Xn+1 = Jm (n)Xn + ηn (2.35)
where
ηn =
0
...
−ηn+1 + αtf ηn −Pθmax
θ=1 αtfn (θi) ηn−θ
(2.36)
Equation (2.35), together with equation (2.34) and (2.36), define an Affine Random
Dynamical System[8].
85
Affine Random Dynamical Systems verify Oseledets’s multiplicative ergodic the-
orem [8] and thus the Lyapunov numbers are the eigenvalues of
limn→∞
[(J∗m)n (Jm)
n]1/2n
If the largest Lyapunov number of (2.35) is less than one and agents trade w.p. 1, then
(2.35) is contractive and ηn is a stationary process. This means that the mispricing
is a random walk and the returns follow a stationary process.
Notice that if p(i) 6= 1, ηn is not stationary and the mispricing will not follow a
random walk, as αtfn (θi) in (2.36) will be time dependent. An interesting question is
’What is the asymptotic statistics of rn if the traders’ capital and activity probability
are chosen so that (2.15) displays intermittent bursts?’.
For correlations of the returns to be small, trend followers and value investors
should have the same parameters [40]. To simplify the analysis, we will consider that
these are trader independent. Thus each trader is characterized by her strategy (value
investor/trend follower), her capital, c, and her activity probability, p. From equation
(2.34), we note that h∆αni = 0, so that the trend followers’ lag in time provides
the oscillations observed on the system. As equation (2.15) approaches the critical
point (zero Lyapunov exponent), the system is characterized by on-off intermittent
behaviour (see section 2.2.1). Indeed, although price oscillations are frequently too
wild, statistics of the returns shows two of the stylized facts of financial time series:
fat tails (indexes between 2 and 4) and volatility clustering.
The system was simulated for Nv = Ntf = 5 ·103, λ = 1, θ(i) uniformly distributedbetween θmin = 1 and θmax = 100, η = 0.1, c = 8 · 10−2 and p = 10−2 (see Figure
2.5). For these parameters, the system is on the critical (non hyperbolic) point
of zero Lyapunov exponent and the statistics shown are highly non-trivial. The
returns distribution is fat tailed with an index of approximately 3. Although the
Autocorrelation Function (ACF) of the returns shows no long range correlations, the
86
volatility ACF is highly correlated. These results were consistent for a large range of
parameter values.
2 4 6 8 10x 104
-100
-50
0
50
100
150
200
time
p t and
ν t
← p
ν ↑
Log Price and Log Value
0 1000 2000 3000 4000 50000
0.2
0.4
0.6
0.8
1
eλ
time
Lyapunov Number
2 4 6 8 10x 104
-30
-20
-10
0
10
20
30
time
r t
Returns
0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
θ
1/H
Hill Estimator for the Returns
Figure 2.5: Simulation with Nv = Ntf = 5 · 103, λ = 1, η = 0.1, θmin = 1 andθmax = 100, c = 8 · 10−2 and p = 10−2.
The reader should notice that on Section 2.4.2, Pareto tails are generated by a
linear process which is contractive on average, but intermittently explosive. This is
not so simply generated on a multi-dimensional system when value investors and trend
followers are together as the trend followers cannot be reduced to one representative
agent (while the value investors can). Nevertheless, at the non-hyperbolic point, if
the traders’ capital is sufficiently high and their trading probability sufficiently low,
the returns have a heavy tailed asymptotic statistics and the volatility is clustered. A
87
possible explanation for this phenomena are nonlinear effects generated by a varying
number of active trend followers.
2.5 Conclusions
We have shown that when the dominant strategy is value investing, if there exist
N,α, p and µ ∈ [k, k + 1] that satisfy (2.28) then
Prob (|rn| > x) ≈ c · x−µ (2.37)
and the returns display a probability distribution asymptotic to a power tail of expo-
nent µ. We have also shown that, if one fixes N and p, one can determine heuristically
α from (2.27) such that the probability distribution of rn is given by (2.37) with the
desired µ.
The association of an unconditional trading probability to each agent, permits
intermittent bursts on market price and a fat tailed distribution for the returns. This
is, to our knowledge, the first model that discusses fat tail distributions caused by
value investor strategies only.
We have shown that when trend followers and value investors are together on the
market trading at each time step, the mispricing is a random walk, independently
of the maximum lag in time, θmax of the trend followers. A mixture of negative
short-term dependence (the value investors’) with long-term positive dependence (the
trend followers) obscures the underlying dependence structure leading to apparent
insignificant correlations[40].
Numerical simulations show that when the system is near the critical point (zero
Lyapunov exponent) and each trader has a relatively high capital and a low trading
probability, distribution of the returns is fat tailed and volatility clustering is observed.
88
The random drive on the system leads to on-off intermittent behaviour at the critical
point. We propose this mechanism as an explanation of the heavy tailed distributions
observed in finance.
Whether the present work is relevant and lessons for the market can be drawn
from it depends on the reliability of simulated price and return time series (where it is
noted that no mechanism provides any cointegration between price and fundamental
value) and on the acceptance of the underlying model [40].
89
Chapter 3
Multistability and Synchronization
in Coupled MapsIntroduction
From a theoretical point of view, it is often useful to study structure formation
with the assistance of models that yield the same dynamical behaviour as the systems
under study, but for which the analysis is simpler than the starting equations. A deep
understanding of theoretical models can, on the other hand, allow predictions on the
behaviour of experimental systems.
The interest of Coupled Map Lattices comes from their ability to reproduce ex-
perimental situations qualitatively, while remaining amenable to analysis.
3.1 Multistability
Many processes in nature do not possess only one long-term asymptotic state or
attractor, but are rather characterized by a large number of coexisting attractors
for a fixed set of parameters. This implies that the attractor eventually reached
by a trajectory of the system depends strongly on the initial condition. This phe-
90
nomenon, called multistability, is found commonly in many fields of science such as
mechanics[17], neuroscience[36, 50], optics[6, 3, 67, 63, 105], fluid dynamics [110],
condensed matter physics[103]. Which behaviour is observed at a given time is highly
sensitive to minor perturbations. The two attributes, accessibility to many states
and sensitivity, allows us to influence, manipulate, and control the dynamics of such
a system.
A well studied case of multistable behaviour is the bistability associated with a
subcritical Hopf bifurcation[109]. Singer has proved that an endomorphism of [0, 1],
F , with a negative Schwarzian derivative has only finitely many periodic sinks, each
corresponding to a critical point of F [111].
It has been known that several bidimensional maps are multistable. Among them,
the Duffing oscillator[5] and the Hénon map[54]. Multistability within the phase-
locked regions of the Arnol’d circle map can be induced by forcing. Indeed, regions
of multistability appear due to the emergence of additional pairs of invariant curves
as a result of smooth and nonsmooth saddle-node or pitchfork bifurcations under the
variation of the forcing amplitude. As a result, these regions look like overlaps of
phase-locked regions with the same rotation number[94].
Recent studies show that multistability is often present in coupled maps. For
small values of the coupling parameter, such as those corresponding to the so-called
”anti-integrable” limit which tries to extend zero-coupling behaviour to small, but
finite coupling strengths, CMLs are often characterized by ”blocked” states, or by
the presence of ”defects”, i.e. spatial structures with regular dynamics but located
arbitrarily in space[21]. An interesting kind of mean-field multistability is found for
CML’s, where the collective attractor depends on the initial density for the Perron-
Frobenius equation[74]. Kaneko has observed attractors with a variety of clusterings
(on the partially ordered phase) with an increase of a or a decrease of ε (the coupling
91
parameter) on globally coupled quadratic maps (f = 1 − ax2) [64]. Further, multi-stability is present in the anti-integrable limit for a large class of network interaction,
not necessarily linear, in networks of weakly coupled bistable units[79].
It has been found that increasing time delays induces a route leading to multi-
stability in networks of oscillators[98]. Also, coupled oscillators with time-delayed
interaction under a pinning force (phenomenological model of oscillatory neural net-
work) display multistability for large coupling strengths[66, 23]. Arrays of diffusively
coupled Van der Pol oscillators (models for e.g. the mammalian small intestine and of
the vortex shedding in a flow past cone-shaped body) also display multistability[95].
Of the above discussed routes to multistability, only [64] can, under a tuning of
the parameters, lead to an arbitrarily large number of coexisting attractors (with an
infinite number of units). Our interest lies in routes to multistability that are not
dependent on the dimensions of the system and that lead to an arbitrary number of
sinks.
We are concerned with multistability in noninvertible maps, and thus will not re-
fer to the Newhouse phenomena (see e.g. [84] and references therein for a review and
a discussion of a generalization to higher dimensions). Nevertheless, it is worthwhile
mentioning that a possible dynamical mechanism leading to multistability is the addi-
tion of small dissipative perturbations to conservative systems. Conservative systems
have a large number of coexisting invariant sets, namely periodic orbits, invariant
tori and cantori. By adding a small amount of dissipation to a conservative system
some of the invariant sets become attractors. Not all invariant sets of the conserva-
tive system will survive when the dissipation is added. However, if the dissipation
is sufficiently small, many attractors (mainly periodic orbits) have been observed in
typical systems[102, 49, 48].
Finally, it is the aim of this chapter to show that for parameter values near the
92
Feigenbaum period-doubling accumulation point, quadratic maps coupled by con-
vex coupling may have a large number of stable periodic orbits. The emphasis on
quadratic maps near the Feigenbaum accumulation point has a motivation close to
the idea of control of chaos.
Control of chaos[97, 1] deals with the local control of unstable orbits, either to
achieve their stabilization or, alternatively, to target a dynamical system to some
desired final state. Reliable stabilization of unstable periodic orbit requires either
the knowledge of a good model of the system or an accurate local reconstruction of
the dynamics. This is feasible in some low-dimensional systems, but it seems rather
problematic for high-dimensional ones. The typical situation in control of chaos, is
that of a strange attractor with an infinite number of embedded periodic orbits, all
of them unstable. If, instead of an infinite number of unstable periodic orbits, one
would have, for example, a large number of sinks, the controlling situation would
seem more promising. The sinks would of course have very small basins of attraction.
Nevertheless, the control need not be so accurate, because it suffices to keep the
system inside the desired basin of attraction. This, in principle, makes for a more
robust control.
3.2 Coupled unimodal maps
The system we consider is a coupled map lattice of two sites. This means the
dynamics is given by the composition of the direct product of two identical one-
dimensional maps with convex coupling.
There are several reasons for our interest in only two sites[47]:
1. We observed a large number of coexisting attractors (multistability) when the
sites were coupled on the Feigenbaum accumulation point and the coupling
93
strength was small enough;
2. The two-dimensional coupled map is a paradigm for the Master dynamics in
secure communications, thus the technological implications of a knowledge of
the regimes visited by the map might be of technological relevance;
3. By coupling only two one-dimensional maps, we expect to be able to solve the bi-
dimensional dynamics rigorously, thus reaching results that could be generalized
for an infinite chain of coupled maps (Coupled Map Lattice), which have been
proposed as models for the time evolution of reaction-diffusion systems (lattice
of ODEs). Understanding the dynamics of the former may then be useful in
predicting the asymptotic behaviour of the latter, in particular, in distinguishing
between chaotic and stable regimes and describing the various patterns in the
stable case;
4. Mathematically, we want to describe the topological changes in the non-wandering
set of two-dimensional non-hyperbolic systems. Just like the Hénon map, our
dynamical system is derived from a simple one-dimensional map through cou-
pling and demonstrates rich dynamics.
3.2.1 Applications of two coupled maps
Multistability in Information Processing
Searching with chaos for a stored memory (as an attractor) has been discussed
in relation to the observation of chaos in neural systems[112]. With the use of the
stochastic nature of chaos, searching for different regions is possible, while deter-
ministic dynamics restricts the space of search, which may eliminate unnecessary
wandering.
94
Two coupled maps in secure communications
Systems with two diffusively coupled maps arise in connection with the develop-
ment of new principles for secure communication based, for instance, on the so called
master—slave configuration (see e.g. [106, 59]).
Master :
xn+1 = f (xn + ε (zn − xn))zn+1 = f (zn + ε (xn − zn))
Slave :
yn+1 = f (yn + ε (zn − yn))un+1 = f (zn + ε (yn − zn))
where f (x) is a one-dimensional (chaotic) map. When transmitting the signal zn from
the master system we want to be able to read the message xn from the slave system.
This will be possible if the parameters of the coupled system can be adjusted such
that yn synchronizes with xn. On the other hand, to mask xn during the transmission
process, synchronization between xn and zn should be avoided. Whenever xn and yn
synchronize, un and zn will also synchronize, and by comparing the variables of the
slave system, one can immediately detect whether synchronization has occurred or
not.
3.2.2 Related work
Motivated by the development of new principles for secure communication based
on the so-called master-slave configuration, Taborov and Maistrenko[116] study the
transverse destabilization of periodic orbits on two diffusively coupled logistic maps.
Although they are interested in the region a > a∞, the process of ’transverse desta-
bilization of low periodic orbits’ reported agrees quantitatively with the bifurcation
route of what we will call phase opposition orbits.
95
3.3 Two quadratic coupled maps: from synchro-
nization to multistability
3.3.1 Definitions
We consider the discrete dynamical system generated by the composition of the
direct product of two identical one dimensional maps, with a diffusive coupling.
The phase space is the squareM = [−1, 1]2. The dynamics is obtained by iteratingthe map Fε defined as follows. Given a point (x, y) ∈M, its image by Fε, denoted
as (x, y) is given by x = (1− ε)f(x) + εf(y)
y = (1− ε)f(y) + εf(x)(3.1)
where 0 6 ε 6 12and f(z) = 1− µz2, 0 < µ ≤ 2 and z ∈ [−1, 1].
The map Fε can be viewed as the composition LεF0 where F0(x, y) = (f(x), f(y))and Lε(x, y) = ((1− ε)x + εy, (1− ε)y + εx). In particular, for ε = 0, the dynamics
is uncoupled and each coordinate evolves independently of the other under f .
Moreover, since f maps the interval [−1, 1] into itself, the convex combination in(3.1) ensures that (x, y) ∈M whenever (x, y) ∈M and the dynamics is well-defined
in this set.
The orbit issued from the initial condition (x, y) is the sequence (xt, yt)t∈N de-fined by (x0, y0) = (x, y) and (xt+1, yt+1) = Fε(x
t, yt) for all t > 0.
The analysis of the dynamics turns out to be more convenient in the variables
(s, d) defined by s = x+y2and d = x−y
2. The dynamics in these variables is given by
s = 1− µ2(s2 + d2)d = −β s d
(3.2)
96
where β = 2µ(1− 2ε).The condition dt = 0 implies dt+1 = 0 and st+1 = f(st) and the corresponding
orbit is said to be synchronized (from t on). A part of the analysis will be devoted
to the orbits which are asymptotically synchronized. An orbit is said to synchronize
if limt→∞ dt = 0 and if all orbits synchronize we say we have synchronization. The
dynamics of the synchronized orbit is the one of f which is rather well known[25].
The map Fε commutes with the reflection over the diagonal
σFε = Fεσ (3.3)
where σ is an exchange operator acting on (x, y) such that σ (x, y) = (y, x) (in (s, d)
coordinates, the symmetry is θ(s, d) = (s,−d)). Hence, for every orbit, (st, dt)t∈Nthere is a symmetric one, (st,−dt)t∈N. Therefore, if W s
(s,d) is the stable manifold
of (s, d), then W s(s,−d) = θ(W s
(s,d)). Moreover, the symmetry also commutes with
the pre-image. Consequently, if W u(s,d) is the unstable manifold of (s, d), we have
W u(s,−d) = θ(W u
(s,d)). Since these manifolds are tangent to the eigen-direction of the
linear dynamics at (s, d), if u is an eigen-vector of the Jacobian at a periodic point
(s, d), then θu is an eigen-vector of the periodic point (s,−d) with the same eigen-value. The set of all fixed points of θ forms a synchronization line d = 0 (x = y) in
the state space. It follows from (3.3) that the synchronization line is invariant under
Fε1.
In addition to θ, one has another symmetry for initial conditions. From (3.2), it
is clear that if (st, dt)t∈N is an orbit, then (−s0, d0), (st,−dt)t≥1 is also an orbit.(Whereas θ holds for any local map f , the latter symmetry holds only for even local
maps).
1The symmetry σ is a property of the coupling and, thus, independent of the local map f . Wewill make use of this result again in Appendix A.
97
3.3.2 Periodic orbits with small periods
Our analysis of the phenomenology starts by determining conditions on the param-
eters for which the system has the ”simplest” dynamical regime, i.e. synchronization.
Indeed, if ε = 12, then β = 0 and for any point (s, d) ∈M we have synchronization
(after the first iteration of Fε). However, synchronization occurs under milder con-
ditions. Indeed, for any orbit, one has |βst| ≤ |β| |st| ≤ |β| as |st| ≤ 1 for all t ∈ N,so that dt will decay exponentially to zero with t and we have synchronization. The
synchronization condition can be relaxed to β < 1 as β is always non-negative. This
condition is (see Figure 3.3.2)
µ < µ1(ε) :=1
2(1− 2ε) (3.4)
Condition (3.4) is not optimal. For instance, if µ < 34, then f has an attracting fixed
point in the interior of [−1, 1] and one can show that synchronization occurs. When ε
is sufficiently small, this global attraction of the synchronized fixed point holds even
though µ ≥ µ1(ε).The local map, f , has a fixed point in [−1, 1] and thus Fε has a synchronized fixed
point inM whose s-component is
s∗ =−1 +√1 + 4µ
2µ
Non-synchronized orbits exist only if µ ≥ µ1(ε). In particular, they are created frombifurcations of the synchronized fixed point, (s∗, 0).(Fε has also two non-synchronized
fixed points (i.e. with d 6= 0), which are always outsideM for 0 < µ < 2 and on the
boundary ofM for µ = 22. Thus, they do not play any role in the complexification
of the dynamics inM for 0 < µ < 2.)
2This is no longer true when f = µx(1− x) (see [116]).
98
0 1/4 1/21/2
3/4
5/4
2
ε
µµ∞
µ1µ(2)PDµ(2)PFµ(2)Hµ(4)PD
The multipliers, i.e. the eigenvalues of the Jacobian of Fε, at the fixed point (s∗, 0)
are f 0(s∗) and (1− 2ε)f 0(s∗), respectively. The first one corresponds to the directionof the diagonal and the second one to the orthogonal direction, which will be referred
as the anti-diagonal.
Computing s∗ from the equation f(s∗) = s∗, one checks that the derivative f0(s∗)
is negative for any µ. If ε > 0, the condition f0(s∗) = −1 determines a period-
doubling bifurcation along the diagonal, the one creating the 2-periodic synchronized
orbit.
On the other hand, the conditions ε > 0 and (1 − 2ε)f 0(s∗) = −1 determine aperiod-doubling bifurcation along the anti-diagonal. Expressed in the parameters µ
and ε, these conditions are:
ε > 0 and µ = µ(2)PD(ε) :=3− 4ε
4(1− 2ε)2 (3.5)
At this bifurcation, the fixed point (s∗, 0) becomes unstable and a 2-periodic orbit
with components, say (s1, d1) and (s2, d2), is created (the transversality condition is
99
satisfied provided the curve a = µ(2)PD(ε) is crossed). When created, this orbit is
locally stable along the anti-diagonal. Moreover, since f 0(s∗) < (1 − 2ε)f 0(s∗) < 0,
before and sufficiently close to the bifurcation, the point (s∗, 0) is unstable along the
diagonal, so is this periodic orbit when created.
Since the bifurcating direction is transverse to the diagonal, we have d1d2 < 0
sufficiently close to the bifurcation. Because of the symmetry θ, the system has also
a 2-periodic orbit with components (s1,−d1) and (s2,−d2). But then the system hastwo periodic orbits created by a co-dimension 1 bifurcation. This is impossible and,
therefore, sufficiently close to the bifurcation, we have d1 = −d2 and s1 = s2.Since the only fixed point insideM is the synchronized one, we shall now consider
2-periodic orbits. The 2-periodic orbits of Fε are determined by (s1, d1) (s2, d2),
where
s1 = s2 =1
β(3.6)
or
s22 =γ − 1±p(γ + 1)(γ − 3)
2µγ(3.7)
for γ = β2
µ= 4µ(1− 2ε)2.
Solution (3.6) is the s-component of the symmetric 2-periodic orbit, the orbit
bifurcating from the synchronous fixed point:
s1 = s2 =1β
d1 = −d2 = 1β
p4µ(1− 2ε)2 + 4ε− 3
(3.8)
100
This orbit exists when d1 = −d2 is real, that is for µ > µ(2)PD(ε), in agreement withthe previously determined condition (3.5).
Equation (3.7) determines four distinct non-symmetric 2-periodic orbits, whose
components are:
s22,± =γ−1±√(γ+1)(γ−3)2µγ
s1,± = 1β2s2,±
d22,± =1µ− s22,± − 1
µβ2s2,±
d1,±± = −β s2,± d2,±
(3.9)
As a simplification of notation for these orbits, it is convenient to define
s± = |s2,±|d± = ±
q1µ− s2± − 1
µβ2s±
(3.10)
The orbits with s-component s2,± = −s± are not insideM and will be discharged
from our analysis3.
The two remaining orbits are determined by (s−, d−) ( 1β2s− ,−β s− d−) and
(s+, d+) ( 1β2s+
,−β s+ d+). These orbits are θ-symmetric, that is (s±, d±) = θ Fε(s∓, d∓).
From (3.9), it can be seen4 that these orbits exist for γ > 3, that is for
µ > µ(2)PF (ε) :=3
4(1− 2ε)2 (3.11)
For µ = µ(2)PF (ε), that is, for γ = 3, s+ = s− = 1βand this last equality implies
that the symmetric and non-symmetric orbits have the same coordinates. Hence, the3They are created through a codimension 1 period-doubling bifurcation of the non-synchronized
fixed points.4It is straightforward to verify that the radicand in the expression of d± is never negative for
γ > 3.
101
non-symmetric orbits appear from a pitchfork bifurcation at µ = µ(2)PF (ε). We have
checked that for ε = 0, these orbits’ components are the combinations of the fixed
point and the components of the period-2 orbit of f .
These results can be confirmed and extended by proceeding to a spectral analysis
of the eigen-values of the Jacobian matrix for the phase-opposition 2-periodic orbit.
Assuming ε > 0 and µ > µ(2)PD(ε), the eigenvalue equation of the Jacobian associated
to this orbit is:
λ2 − 2 £(1− 2ε) (S2 −D2) + 2ε2S2¤λ+ (1− 2ε)2 (S2 −D2)2 = 0 (3.12)
where
S = −2µs1 = 11−2ε
D = −2µd1 =√4µ(1−2ε)2−(3−4ε)
1−2ε(3.13)
For ε = 0, λ0 = S2−D2 = 4(1−µ), which are the eigen-values of the local map. Whenµ > 1, λ0 will be negative. For ε > 0, if the eigenvalues are real, they are non-negative,
as proved by the following. We start with the remark that for the phase-opposition
orbit, S 6= 0 for µ > 0, as can be deduced from (3.8). The discriminant of the solutionto equation (3.12),
∆ = 4ε2S2¡(1− 2ε)(S2 −D2) + ε2S2
¢(3.14)
is non-negative by hypothesis, which implies (1− 2ε)(S2−D2)+2ε2S2 ≥ 0 for ε ≥ 0,so that λ+ ≥ 0. Furthermore, the determinant of the Jacobian matrix is non-negative,
|J| = λ+λ− = (1− 2ε)2(S2 −D2)2 ≥ 0 (3.15)
102
so that λ− ≥ 0.Supposing 0 < ε < 1
2, the condition ∆ = 0 implies S2 − D2 < 0 (see (3.14)).
Furthermore, the conditions ∆ > 0 and D2 ∼ 0 imply S2 −D2 > 0. Hence, by the
Intermediate Value Theorem, there exists ∆ > 0 for which S2 − D2 = 0. This last
equality is equivalent to S = ±D, a crossing of the axes x = 0∨ y = 0 that occurs atµ(2)λ−=0(ε) :=
1−ε(1−2ε)2 and yields λ− = 0 and λ+ = 4ε
2S.
We conclude that ∆ ≥ 0 implies λ+ > 0 and λ− ≥ 0, where λ− = 0 for µ =
µ(2)λ−=0(ε) > µ(2)PF (ε) > µ(2)PD(ε).
An analysis of the spectra of λ± shows that when
µ(2)PD(ε) < µ ≤ µ(2)PF (ε)
we have λ+ ≥ 1 and 0 < λ− < 1. For µ = µ(2)PF (ε) we have λ+ = 1 and there is a
pitch-fork bifurcation on the eigen-direction corresponding to λ+.
Further analysis of the eigenvalues shows that when
µ(2)PF (ε) < µ ≤ µ(2)complex(ε) := (2− 3ε)24(1− 2ε)3
we have 0 < λ+ < 1 and 0 ≤ λ− < 1 and the phase-opposition orbit is stable. For
µ = µ(2)λ−=0(ε) we have λ− = 0. By the arguments developed on the previous section,
µ(2)λ−=0(ε) < µ(2)complex(ε), thus λ− has a local minimum at µ = µ(2)λ−=0.
For µ > µ(2)complex(ε) and ε 6= 0, the eigen-values are complex conjugate, λ+ = λ−
and |λ±| < 1. When the curve µ = µ(2)H(ε) := 5−6ε4(1−2ε)2 is crossed, a Hopf bifurcation
destabilizes this orbit and creates a stable close invariant ”circle”. Indeed, one checks
that the transversality and non-resonance conditions are satisfied (see, for instance,
[55]). A numerical calculation shows that the stable invariant ”circle” is destroyed
when µ is sufficiently large or ε sufficiently small. The presence and the normal form
of a stable invariant circle in coupled quadratic maps was studied in [68].
103
3.3.3 Higher period phenomenology
This orbit exists with the same symmetry for ε = 0 and µ > 34. In this case, it ap-
pears from a co-dimension 2 bifurcation at µ = 34which also creates the synchronized
2-periodic.
The previous argument developed at the synchronized fixed point extends to the
iterates of Fε and describes the appearance of the following orbits.
Definition 2 An orbit with the property st+2n = st, dt 6= 0 and dt+2n = −dt, t ≥ 0,n ≥ 0 is called a phase opposition (2n+1-periodic) orbit.
Let n ∈ N and let (si, 0)1≤i≤2n be the components of a 2n-periodic synchronizedorbit, the existence of which depends on a. By the chain rule and since the Jacobian
is diagonal in (s, d)-coordinates, the corresponding multiplier along the anti-diagonal
is
(1− 2ε)2n2nYi=1
f 0(si)
The condition that this multiplier is equal to −1 determines a co-dimension 1 period-doubling bifurcation. Since 0 < ε < 1
2, (1−2ε)2n < 1 ∀n ∈ N and this bifurcation can
only occur after the corresponding one along the diagonal has occurred. Applying
the previous arguments to each point (si, 0), we conclude that the bifurcation leads
to a phase-opposition 2n+1-periodic orbit. This orbit is locally stable along the anti-
diagonal (with eigen-value, say, 0 < Λ− < 1 along this direction) and unstable along
the diagonal (with eigen-value, say, Λ+ > 1).
Not only the 2n+1-periodic phase opposition orbit exits if the 2n-periodic syn-
chronized orbit is unstable along the diagonal, but it seems that it exists only if the
2n-periodic phase opposition periodic orbit exists.
104
-1 0 1-1
0
1
s
d
0 0.1 0.2 0.3 0.4-2
-1
0
1
2
ελ
-1 0 1-1
0
1
s
d
0 0.1 0.2 0.3 0.4-2
-1
0
1
2
3
ε
λ
Figure 3.1: On the left we have plotted the phase-opposition 2-periodic and 4-periodicorbits at µ∞. The colours denote stability of the orbits: red for unstable, blue forsaddle-node and green for stable. On the right, the corresponding eigenvalues, Re (λ)in blue and red, respectively, and Im (λ) in green.
105
This is indeed the case for the 4-periodic orbit whose existence condition is µ >
µ(4)symm(ε) :=4(1−2ε)2+14(1−2ε)2 ≥ µ(2)PD(ε) if 0 < ε ≤ 1
4and µ(4)symm(ε) < 2.
Figure 3.1 shows the phase opposition 2-periodic and 4-periodic orbits and their
stability. Figure 3.2 shows the equivalent diagrams for non-symmetric 2-periodic and
4-periodic orbits, which are compositions of 2-periodic and 1-periodic orbits of f , and
4-periodic and 2-periodic orbits of fµ, respectively, for ε = 0.
-1 0 1-1
0
1
s
d
0 0.05 0.1-2
-1
0
1
2
ε
λ
-1 0 1-1
0
1
s
d
0 0.02 0.04 0.06-2
-1
0
1
2
3
ε
λ
Figure 3.2: On the left we have plotted the non-symmetric 2-periodic and 4-periodicorbits at µ∞ which are composed or different period orbits for ε = 0. The coloursdenote stability of the orbits: red for unstable, blue for saddle-node and green for sta-ble. On the right, the corresponding eigenvalues, Re (λ) in blue and red, respectively,and Im (λ) in green.
Furthermore, a numerical calculation at µ = µ∞ , reported in Figure 3.3, shows
that the succession of bifurcations of a phase opposition orbit does not depend on the
106
period. On this picture, we have plotted the values of ε for the Hopf bifurcation, the
pitchfork bifurcation and the period-doubling bifurcation creating the orbit, versus
the power of the period. For each period, the phenomenology is identical to that
described in the previous section, with an adequate change of scale in ε. In addition,
the picture shows that several phase opposition orbits may be stable for ε > 0 fixed.
This stabilization is an effect of the coupling that will be discussed below.
Finally, since the phase opposition orbits are the first orbits to appear when the pa-
rameters are varied from synchronization and since the first such orbit that is created
is of period 2, it follows that a necessary and sufficient condition for synchronization
is µ ≤ µ(2)symm(ε), the condition for the existence of the latter.
3.3.4 The non-symmetric orbits
We now analyze the existence and the stability of other period-2p orbits for µ =
µ∞. Our interest in µ∞ comes from the scaling properties of f which are reflected
on the scaling laws for the periods and values of ε at which the bifurcations of Fε
occur (see Figures 3.3 and 3.7). We only consider the orbits which for ε = 0 have the
same period on projection to both axis x and y. These orbits are followed numerically
when ε increases and are referred using the phase shift of their components at ε = 0.
For µ = µ∞, the map f has a period-2p orbit for each p ∈ N, whose componentsfor p up to 5 are shown in Figure 3.4. In this picture, the numbers reflect the order in
which the components are visited and the tree structure represents the origin of each
component in the bifurcation cascade. An important notion is the dyadic distance δ
between the components of an orbit. δ is the number of steps one has to go back in
the bifurcation tree to meet a common component.
The dyadic distance is used to characterize the families of periodic orbits that
we are considering. For instance, the coordinates of each component of a synchro-
107
1 2 3 4 5 6 7 8 9 10 11
10-5
10-4
10-3
10-2
10-1
100
p
epsi
lon
Figure 3.3: Bifurcation values of ε for the phase opposition orbits. From top tobottom, birth, pitchfork and Hopf bifurcations.
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
0
1
2
3
4
5
1
1 2
1 23 4
1 23 45 67 8
1 23 45 67 89 1011 1213 1415 161 23 45 67 89 1011 1213 1415 1617 1819 2021 2223 2425 2627 2829 3031 32
x
p
Figure 3.4: Periodic orbits for the unidimensional map at µ = µ∞.
108
nized orbit are at distance 0, those of a phase-opposition orbit are at distance 1.
Accordingly, when we speak of distance−k orbit we refer to the dyadic distance ofthe coordinates of its components.
For any δ ≥ 1, there are 2δ−1 different orbits with distance δ which have co-
ordinates out of phase by 2p−δ + α12p−δ+1 + α22
p−δ+2 + · · · + αd−12p−1 steps, with
αi ∈ 0, 1. The distance of a period-2p orbit is at most p (p ∈ N).
-1 0 1-1
0
1
s
d
0 0.01 0.02-2
-1.5
-1
-0.5
0
0.5
1
ε
λ
-1 0 1-1
0
1
s
d
0 0.005 0.01-2
-1.5
-1
-0.5
0
0.5
1
ε
λ
Figure 3.5: An example of 4-periodic orbits for which δ = 2 at µ∞.
For ε = 0, the only symmetric orbits are those at distance 0 and 1. This property
is preserved for ε > 0 as shows Figure 3.5 for δ = 2. The succession of bifurcations
of orbits with distance δ ≥ 2 should then differ from those with distance 1 (see the
evolution of the eigenvalues of the orbits on Figure 3.5). Indeed, for ε = 0, the orbits
109
are unstable. When ε increases, they suffer two collisions with orbits of twice the
period when the eigenvalues cross −1, after which the orbits become stable. (Whendecreasing ε, these collisions would be period-doubling bifurcations). If ε increases
further, the orbit collides with an unstable one of the same period in a saddle-node
bifurcation when the larger eigenvalue reaches 1. For larger values of ε, the orbit does
not exist. The unstable orbit with which it collides is one that at ε = 0 has period
2p in one projection and 2p−1 in the other. For higher dyadic distances, the overall
variation of the eigenvalues is similar to the δ = 2 case.
Figure 3.6 shows a typical example of these phenomena for the case δ = 2. Between
ε = 0 and the point labelled 1 in the figure, the orbit is unstable. The point 1
corresponds to the smaller eigenvalue crossing -1 (see Figure 3.5). Therefore, between
the point 1 and 2, the orbit is stable. It disappears at the point 2 when it collides
with an unstable orbit of the same period.
3.3.5 Multistability
We have seen that the coupling stabilizes the orbits with distance larger than 0
at µ = µ∞. There are indeed two mechanisms responsible for this stabilization. The
determinant of the Jacobian of a period-2p orbit is
(1− 2ε)2p(−2µ∞)2p+12pYi=1
xiyi.
The term (1− 2ε)2p, consequence of the coupling, decreases as p increases. However,there is yet a second stabilizing mechanism. Denote by Γ(ε) the remaining factor in
the determinant
Γ(ε) = (−2µ∞)2p+12pYi=1
xiyi.
Without coupling, Γ(0) is simply the square of the multiplier of f for the periodic
orbit. From the properties of the Feigenbaum—Cvitanovic functional equation, follows
110
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
21
Figure 3.6: ε-evolution of a period-4 distance-2 orbit and its bifurcating orbits atµ = µ∞. The axes are the space variables, x and y.
111
that this factor converges to a fixed value around −1.6 when p increases[20]. However,the coupling changes the position of the orbit components in such a way that this
factor also decreases. It is the combined action of this decrease with the contraction
of the coupling that brings the eigenvalues into the interior of the unit circle and
stabilizes the orbits.
For small ε there is a simple geometrical interpretation for the variation of Γ(ε).
On the one dimensional map, the productQ2p
i=1 xi remains constant, when p grows, as
each time the period doubles, the doubling in the number of factors greater than one
is compensated by the fact that the component of the orbit closest to zero approaches
zero a little more.
1 2 3 4 5 6 7 8 9 10 1110-6
10-5
10-4
10-3
10-2
10-1
100
p
epsi
lon
Figure 3.7: Stability lines of orbits from distance 1 (top) to 4 (bottom).
For the unstable orbits along the period-doubling chain, the orbit components
closest to zero alternate on each side of the origin. The contracting effect of the convex
112
coupling tends to bring the orbits back in the period-doubling hierarchy. Therefore,
because the component closest to zero has to move across the origin for the orbit to
approach the one with half the period, this implies that the product of the coordinates
is going to decrease. The greater the dyadic distance between the orbit projections
on the axis, the greater will be the perturbation that the original (one-dimensional)
orbits suffer. Therefore one expects the contracting effect in Γ(ε) to increase with
the dyadic distance. This effect is quite apparent on Figure 3.7 which shows the
stabilizing and destabilizing lines for orbits with distance from 1 to 4. The shift
downwards of the stable regions for successively larger dyadic distances implies that
the smaller ε is, the larger the number of distinct stable orbits that are obtained. An
accurate numerical estimate of the number of distinct orbits is obtained by computing
the derivative D (p, δ) = ∂³Q2p
i=1 xiyi´/∂ε at ε = 0 for each p and dyadic distance
δ. Actually this derivative provides an accurate estimate of Γ (ε) itself, because this
one varies almost linearly with ε for most of the stable range of the orbits. On Figure
3.8, the scaling properties, when p grows, of this derivative are shown. From these
results one computes
log (−D (p, δ)) = g (δ)− 2p+1
with
g (δ) ' 0.907δ + 3.475
Notice that in Figure 3.8 there is more than one data point for each pair (p, δ) which
correspond to non-equivalent orbits with the same dyadic distance.
Two other useful estimates are:
• the value of the smallest ε parameter that stabilizes an orbit of dyadic distanceδ equal to the power p
log εminδ=p = −1.535p− 1.547
113
0 1 2 3 4 53.5
4
4.5
5
5.5
6
6.5
7
7.5
8
8.5
d
log|
D(p
,d)|
+2N
. p=1x p=2* p=3o p=4+ p=5
Figure 3.8: Scaling properties of D(p, δ).
• the value of the largest ε parameter for which a δ = 1 orbit is stable
log cmaxδ=1 = −0.679p− 1.235
From this, one obtains the result that at least k distinct stable orbits are obtained
if
0 < ε . exp (−0.99− 1.22k)
k is only a lower bound on the number of distinct stable periodic orbits, because here
we have studied only orbits with the same period under projection in the two axis.
3.4 Conclusions
For sufficiently small ε an arbitrarily large number of distinct stable periodic orbits
is obtained. However, for any fixed ε, it is an arbitrarily large number that is obtained,
114
not an infinite number. Most orbits either synchronize (and are then unstable) or
disappear as ε grows. As a result, a reasoning based on the implicit function theorem,
as used in [20] is misleading. Given a sequence of orbits of different periods, even
if they remain as orbits for a small perturbation, that does not mean that their
(smallest) periods remain distinct.
115
Chapter 4
Symbolic Dynamics for Tent-Maps
4.1 Two Diffusively Coupled Modified Tent Maps
We assume the reader goes through Appendices A and B, and is familiar with the
notation therein before entering this chapter. Most of the concepts we shall use are
developed in detail in these Appendices in a much simpler setting. Furthermore, we
will omit proofs whenever similar reasonings can be found in the former Appendices.
4.1.1 Main results
We show that periodic orbits can be rigorously followed as the coupling strength
is varied.
We apply the method to the coupled tent map (see Appendix A) and obtain the
following results:
1. we state rigorous expressions for the Lyapunov numbers as a function of local
map’s parameter, a, the coupling strength, ε and the orbit’s codes, θi and ξi;
116
2. we derive rigorous expressions for the orbit’s points as a function of a, ε and
the orbit’s codes;
3. we prove that all periodic orbits are unstable for ε < 1−a2;
4. we prove a sufficient condition for synchronization and that for ε = 1−a2the
coupled map undergoes a blowout bifurcation.
The method is also applied to the coupled modified tent map, with the following
results:
1. we state rigorous expressions for the Lyapunov numbers as a function of local
map’s parameter, a, the coupling strength, ε and the orbit’s codes, θi and ξi;
2. we derive rigorous expressions for the orbit’s points as a function of a, ε and
the orbit’s codes;
3. we study a period-doubling bifurcation route with the help of the analytic tools
developed in 1 and . For orbits of period up to 4:
(a) we extract the orbit’s codes and (rigorously) follow the orbit numerically
as the coupling parameter, ε, is varied;
(b) we show that the period-doubling bifurcations occur as a period-2p orbit
collides with the admissibility axes and that a period-2p+1 orbit is born out
of this collision on a border-collision bifurcation. Further this new orbit
differs from the previous one on only one code;
4.1.2 Rigorous expression for the orbits as a function of the
codes
Consider two diffusively coupled unidimensional maps,
117
xi+1 = (1− ε)f(xi) + εf(yi)
yi+1 = (1− ε)f(yi) + εf(xi)(4.1)
Equation (4.1) can be written in vector form as
Xi+1 =
xi+1yi+1
= 1− ε ε
ε 1− ε
fa,+ (xi)fa,− (yi)
(4.2)
Consider the modified tent map (see Fig 4.1),
fa,± (x) = 1− a¯x± 2− a
a
¯(4.3)
For a = 2, fa,± (x) ≡ Ta (xi) and (4.2) is reduced to the coupled map in Appendix A.
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
xn
x n+1
← f3/2,-(x)f3/2,+(x) →
Figure 4.1: The modified tent map fa,± (x) = 1 − a ¯x± 2−aa
¯is plotted for a = 3
2
(dashed line) together with the tent map, f2 (x) (solid line).
The following expression will be useful in what follows: fa,+ (xi)fa,− (yi)
= U−a ¯xi + 2−a
a
¯¯yi − 2−a
a
¯ (4.4)
118
Where U =
11
. For ε ≤ 1/2,r := − log ¡√1− 2ε¢ (4.5)
and
1− ε ε
ε 1− ε
= √1− 2ε [r; 1] (4.6)
Thus, from (4.6) and (4.4)
Xi+1 =√1− 2ε [r; 1]
U−a ¯xi + 2−a
a
¯¯yi − 2−a
a
¯ (4.7)
Let the codes, θi, ξi ∈ −1, 12, be defined by¯xi +
2− aa
¯= θi
µxi +
2− aa
¶(4.8)¯
yi − 2− aa
¯= ξi
µyi − 2− a
a
¶Further, define ηi ∈ −1, 1 by
ηi = θiξi
Thus (4.7) can be written as
Xi+1 =√1− 2ε [r; 1]U− (2− a)√1− 2εθi [r; ηi]V−
√1− 2εaθi [r; ηi]Xi (4.9)
where V =
1
−1
.
119
Proposition 5 For ε < 1/2,
Xi =nXk=0
µ −1√1− 2ε
¶k1
ak+1θi,i+k
"−r
k−1Xj=0
ηi,i+j; ηi,i+k
#U
− 2− aa
nXk=0
µ −1√1− 2εa
¶kθi,i+k−1
"−r
k−1Xj=0
ηi,i+j; ηi,i+k−1
#V
+
µ −1√1− 2ε
¶n+1θi,i+n
"−r
nXj=0
ηi,i+j; ηi,i+n
#Xi+n+1 (4.10)
Proof: By induction on n.¤
Proposition 6 For a period n orbit, the coupled map’s Lyapunov numbers are
e±λ =
a¡√1− 2ε¢1∓ 1
n
Pn−1j=0 η1,1+j if η1,n = 1
a√1− 2ε if η1,n = −1
Proof: Similar to the proof of Proposition 10.
Proposition 7
Xi =
q−1Xk=0
µ −1√1− 2ε
¶k1
ak+1θi,i+kχU
ηi,i+kΓ−ηi,i+k
Pk−1j=0 ηi,i+j −
³−1√1−2εa
´qθi,i+q−1ηi+k+1,i+q−1Γ
Γ−ηi,i+kPk−1j=0 ηi,i+j −
³−1√1−2εa
´qθi,i+q−1Γ
ηi,i+kPq−1j=k ηi,i+j
− 2− aa
q−1Xk=0
µ −1√1− 2εa
¶k θi,i+k−1χV
ηi,i+k−1Γ−ηi,i+k−1Pk−1j=0 ηi,i+j −
³−1√1−2εa
´qθi,i+q−1ηi+k,
−Γ−ηi,i+k−1Pk−1j=0 ηi,i+j +
³−1√1−2εa
´qθi,i+q−1Γ
−ηi,i+k−1P
where
χU = ηi,i+k +
µ1
(1− 2ε) a2¶q
ηi+k+1,i+q−1 −
−µ −1√
1− 2εa¶q
θi,i+q−12
(ηi,i+k + ηi+k+1,i+q−1)³Γηi,i+k
Pq−1j=0 ηi,i+j + Γ−ηi,i+k
χV = ηi,i+k−1 +µ
1
(1− 2ε) a2¶q
ηi+k,i+q−1 −
−µ −1√
1− 2εa¶q
θi,i+q−12
(ηi,i+k−1 + ηi+k,i+q−1)³Γηi,i+k−1
Pq−1j=0 ηi,i+j + Γ−ηi,i+k
121
Proof: Similar to the proofs of Propositions 11 and 12.
Expression (4.12) permits a rigorous determination of a periodic orbit as a function
of the maps’ parameters, a and ε, and the codes, θi and ξi for i = 1, . . . , q . This allows
a rigorous study of the orbit’s points, independently of its stability and provides us
with a method to rigorously follow by continuation stable or unstable periodic orbits.
4.1.3 Numerical studies in parameter space
A numeric study in parameter space (a, ε) of the attractors of (4.2) is presented
in Figures 4.2 and 4.3. The colours in Figure 4.2 have the following meaning (see
Figure 4.3 for the colour map):
1. Colour n is plotted when the trajectory of (4.2) converges to a periodic orbit of
period n, where n < 13;
2. If the trajectory diverges, then colour 15 is plotted (this has not been used in
Figure 4.2);
3. If the maximum Lyapunov exponent is positive, then colour 14 is plotted (see
large light yellow area in Figure 4.2);
122
4. If the |max λ1,λ2| < χ, where λ1,λ2 are the Lyapunov exponents of themap and χ ≥ 0 is chosen, colour 13 is plotted. In Figure 4.2, χ = 0 and colour13 is left unused;
5. If none of the above cases occurs (e.g. the trajectory of (4.2) converges to a
high period attractor), colour 0 is plotted.
A careful examination of Figure 4.2 reveals a period-doubling route: period-1
(blue), period-2 (green), period-4 (dark red) and period-8 (dark brown). Higher
periods are not present as not enough colours are available to plot them.
Figure 4.2: Attractors in parameter space (a, ε) for the coupled modified tent map.Initial condition is (0.2, 0.7) and the color map is in Fig .
123
Figure 4.3: Colourmap for Figure 4.2.
4.1.4 A period doubling route of the coupled modified tent
map
Our study proceeds by extracting the codes and determining numerically the ad-
missibility condition of these orbits.
Our analysis supposes that ε is being increased from 0 to 1/2 and the orbits are
being followed according to (4.12) for a given code. The fixed point, period-2 and
period-4 orbits on the period-doubling bifurcation route are plotted for a = 1.3877
and ε ∈ [0, 0.5] on Figures 4.4, 4.5 and 4.6, respectively. Orbits are plotted with acolour that represents their stability —blue for an unstable orbit and red for a stable
one. Figure 4.7 shows the same orbits together, but only the stable region is plotted.
Periodic orbits bifurcate at the admissibility axes (which are plotted in magenta);
this coincides with the values of ε for which the orbit is no longer admissible. One
should be aware that orbits in Figures 4.4-4.6 are not admissible for all ε ∈ [0, 1/2].Indeed, this is apparent as the plotted orbits are discontinuous in ε: they stop being
admissible when they ”collide” with the admissibility axes, but if ε is raised further,
they become admissible again.
4.1.5 Border collision bifurcations
The modified coupled tent map, as given by (4.2), is a piecewise differentiable
two dimensional map. By a piecewise differentiable two dimensional map we mean
124
that the map is continuous and there is one (or more) curve (which we call the
border) separating the phase space into two (or more) regions, such that the map
is differentiable on both sides of the border, but not on it. In particular we are
concerned with the case where the map’s derivative changes discontinuously across
the border. This circumstance leads to a rich class of bifurcation phenomena called
border-collision bifurcations[34].
Table 4.1: Codes extracted from the period-doubling route orbits up to period 8. Thecode that changes with each bifurcation is underlined.Period Stability Region1 0 0.481 ≤ ε ≤ 1/22 0 1 0.288 ≤ ε < 0.4814 0 1 -1 1 0.2663 ≤ ε < 0.2888 0 1 -1 1 1 1 -1 1 0.2632 ≤ ε < 0.2663
A border collision bifurcation can be defined as follows. Let J be the Jacobian
matrix of a map F evaluated at a periodic orbit of period p. Let m be the number
of real eigenvalues of J smaller than −1, and let n be the number of real eigenvaluesof J greater than +1. The orbit index I of the periodic orbit is
I =
0 if m is odd
−1 if m is even and n is odd
1 if both m and n are even
We define a border-collision bifurcation as a bifurcation at a periodic orbit on the
border of two regions when the orbit index of the periodic orbit before the crossing of
the border is different from the orbit index of the fixed point after the collision[90].
The orbits on the period-doubling bifurcation route are plotted in Figures 4.4-4.6
and have the following characteristics:
1. The only admissible codes are θi, ξi ∈ −1,−1 , 1,−1 , 1, 1 . Therefore,the two codes can be replaced by one only, e.g. their semi-sum ψi =
θi+ξi2. As a
125
consequence, the orbits do not have points on the second quadrant (this is the
”smallest quadrant”)- see Figure 4.7;
2. At each period-doubling bifurcation only one of the codes changes. Indeed, this
can be observed from Table 4.1 which contains the extracted codes. We have
verified this phenomena for periods up to 256 in the bifurcation cascade.
-1 0 1-1
0
1 theta:1 ksi:-1
Figure 4.4: The fixed point is admissible and stable for 0.481 ≤ ε ≤ 0.5.
Finally, the regions of ε for which the orbits are stable are presented on Table 4.2.
Period Stability Region1 0.481 ≤ ε ≤ 1/22 0.288 ≤ ε < 0.4814 0.2663 ≤ ε < 0.2888 0.2632 ≤ ε < 0.266316 0.2622 ≤ ε < 0.263232 0.26109 ≤ ε < 0.262264 0.26103145 ≤ ε < 0.26109128 0.26103097 ≤ ε < 0.26103145
Table 4.2: Stability regions for a = 1.3877 of the period-doubling orbits.
126
-1 0 1-1
0
1 theta:1 1 ksi:1 -1
Figure 4.5: A period-2 orbit which is stable for 0.288 ≤ ε ≤ 0.48.
-1 0 1-1
0
1 theta:1 1 1 -1 ksi:1 -1 1 -1
Figure 4.6: A period-4 orbit which is stable for 0.268 ≤ ε ≤ 0.286.
127
-1 0 1-1
0
1
Figure 4.7: The orbits from Figures 4.4, 4.5 and 4.6 superimposed and only thestability region is plotted. It is apparent that period doubling bifurcations occurupon collision with the admissibility axes.
4.2 Conclusions
Symbolic dynamics applied to two diffusively coupled tent-like maps is a powerful
framework for the study of periodic orbit phenomena. Namely, we have developed
a framework to study bi-dimensional coupled tent-like maps. Surprisingly, coupled
modified tent maps can display a bifurcation route absent from the local (uncoupled)
maps: a period-doubling bifurcation route. We have studied in detail the orbits in
this bifurcation route and have extracted their code. We have concluded that at each
border-collision bifurcation only one code changes.
The method can be applied if the local maps have the same absolute derivative
on both sides of the critical point and is, therefore, not valid for skew tent maps.
Work still in progress involves a rigorous proof of the existence of the period
doubling bifurcation route to chaos.
128
Appendix A
Essential Definitions and
Background Material
This Appendix assembles some background material on the dynamics of tent maps
and coupled tent maps. Together with Appendix B, it allows an understanding of the
dynamics of the diffusively coupled tent map. Most of the proofs are similar (but less
extensive), mutatis mutandis, to the proofs of Chapter 4 and will be omitted from
the later.
A.1 The Tent Map
The tent map Ta : R −→ R is composed of two straight line segments of slope 1a
and −1a:
xi+1 ≡ Ta (xi) = 1− a− |xi|a
, a ∈·1
2, 1
·(A.1)
For a given tent-map Ta, Lebesgue-almost every point in [0, 1] is attracted to the sameset Λa, a finite union of intervals. Figure A.1 plots Λa against 1/a.
129
1-1 -0.6 -0.2 0.2 0.6
1/a
log2/16
log2/8
log2/4
log2/2
log2
Figure A.1: The limit set of Ta, Λa, plotted against 1/a.
130
As characterized by its Lyapunov exponent, log 1a, the map is chaotic for a ∈ £1
2, 1£.
The dynamics under iteration of the tent map has been extensively studied as one
of the simplest examples of discrete dynamical systems.
Similar to the well-known universality of the quadratic map, the tent map also
governs the low-dimensional behavior of a wide class of nonlinear phenomena. In
Lorenz’s 1963 paper on deterministic nonperiodic flow, which introduced the ’Lorenz
attractor’, Lorenz computed a statistic of orbits on the attractor that appeared to
be described by iterating a tent-like map [73]. In 1979, Parry exhibited a semi-
conjugacy between the symmetric tent map and a simplified model of the Lorenz
attractor (which is a piecewise linear rotation on S1) [99]. Moon[88] has recently
shown that the dynamics of the Ginzburg-Landau equation, in its description of the
modulational instability of a wave train, is reducible to the tent map (under this
reduction, varying the height of the tent map corresponds to varying the wavelength
of the initial modulational instability).
Milnor and Thurston [115, 86] showed that the logistic map, gλ (x) := λx (1− x),is topological semi-conjugate to some tent map with the same topological entropy
Th(λ)(x). Here the conjugacy endomorphism is generally not invertible; it may have
’flat spots’. The tent map essentially captures the dynamics of the logistic map on
the subset [−1, 1] where it is ’expanding’.We use Σ to denote the symbol space −1, 1Z equipped with the standard product
topology.
Definition 3 The itinerary of x ∈ R under Ta is
ıa (x) ≡©θ ∈ Cs | θiT ia (x) ≥ 0, ∀i ≥ 0
ª
131
Alternatively, ıa : R−→Cs is the multivalued function given by
ıa (x)i =
+1 if T ia (x) > 0∗ if T ia (x) = 0−1 if T ia (x) < 0
Where ∗ plays the role of a ”joker”, i.e. is both +1 and −1.Then (A.1) can be written as
xi = (1− a) θi − aθixi+1 (A.2)
Replacing i by i+ 1 and i+ 2,
xi = (1− a) θi − aθixi+1 (A.3)
xi+1 = (1− a) θi+1 − aθi+1xi+2 (A.4)
xi+2 = (1− a) θi+2 − aθi+2xi+3 (A.5)
Inserting (A.4) into (A.3) yields
xi = (1− a) θi − aθi,i+1+ (−a)2θi,i+1xi+2 (A.6)
Inserting (A.5) into (A.6),
xi = (1− a)©θi − aθi,i+1 + (−a)2θi,i+2
ª+ (−a)3xi+3θi,i+2
Generalizing (see Appendix B for notation),
Proposition 8
xi = (1− a)nXk=0
(−a)k θi,i+k + (−a)n+1 xi+n+1θi,i+n ∀n ∈ N
Proof: By induction on n.¤
132
A.2 Two Diffusively Coupled Tent Maps
Consider two diffusively coupled tent maps,
xi+1 = (1− ε)Ta (xi) + εTa (yi)yi+1 = (1− ε)Ta (yi) + εTa (xi)
(A.7)
Equation (A.7) can be written in vector form as
Xi+1 =
xi+1yi+1
= 1− ε ε
ε 1− ε
Ta (xi)Ta (yi)
(A.8)
For 1−ε√1−2ε ≥ 1⇐⇒ ε ≤ 1/2,
r := arg cosh
µ1− ε√1− 2ε
¶= − log ¡√1− 2ε¢ (A.9)
The connection matrix will be denoted by
[r; 1] :=
cosh r sinh r
sinh r cosh r
=
1−ε√1−2ε
ε√1−2ε
ε√1−2ε
1−ε√1−2ε
=
1√1− 2ε
1− ε ε
ε 1− ε
(A.10)
and
[r; 1]
1 0
0 η
= [r; η] (A.11)
The tent maps can be written in the form Ta (xi)Ta (yi)
= 1− aaU−1
a
|xi||yi|
(A.12)
133
Where U : =
11
. From (A.10) and (A.12), we can now write (A.8) as
Xi+1 =√1− 2ε [r; 1]
µ1− aaU−1
aXi
¶(A.13)
Let the maps’ codes, θi, ξi ∈ −1, 12, be defined by
|xi| = θixi
|yi| = ξiyi
Further, define ηi ∈ −1, 1 by
ηi = θiξi
Thus (A.13) can be written as
Xi+1 =√1− 2ε1− a
a[r; 1]U−
√1− 2εa
θi [r; ηi]Xi (A.14)
So that
Xi =a√1− 2εθi [r; ηi]
−1µ√
1− 2ε1− aa
[r; 1]U−Xi+1¶
(A.15)
From Proposition 34 (see Appendix B),
[r; ηi]−1 [r; 1] = [−ηir; ηi] [r; 1]
= [0; ηi]
And (A.15) can be simplified to
Xi = (1− a) θi [0; ηi]U− a√1− 2εθi [−rηi; ηi]Xi+1 (A.16)
134
Let us generate expressions for Xi+1 and Xi+2 from (A.16) and substitute them into
the rhs of (A.16). We obtain
Xi = (1− a)½θi [0; ηi]− a√
1− 2εθi,i+1 [−rηi; ηi,i+1] +
+
µ −a√1− 2ε
¶2θi,i+2 [−rηi − rηi,i+1; ηi,i+2]
)U+
+
µ −a√1− 2ε
¶3θi,i+2 [−rηi − rηi,i+1 − rηi,i+2; ηi,i+2]Xi+3 (A.17)
Generalizing (A.17), we obtain the following
Proposition 9 For ε < 1/2,
Xi = (1− a)nXk=0
µ −a√1− 2ε
¶kθi,i+k
"−r
k−1Xj=0
ηi,i+j ; ηi,i+k
#U+ (A.18)
µ −a√1− 2ε
¶n+1θi,i+n
"−r
nXj=0
ηi,i+j; ηi,i+n
#Xi+n+1
The results that follow are only valid for periodic orbits, for which (A.18) contains
only a finite number of terms. Indeed, for a periodic orbit the Lyapunov numbers can
be rigorously determined from the codes:
Proposition 10 For a period q orbit, the coupled tent’s Lyapunov numbers are
e±λ =
(√1−2ε)1∓
1qPq−1j=0 η1,1+j
aif η1,q = 1
√1−2εa
if η1,q = −1(A.19)
Expression (A.18) leads to a rigorous determination of the points of the orbit in
Proposition 12 as a function of the codes, the parameter of the local map (a) and the
coupling (ε).
Definition 4
αi,k = rηi,i+k
k−1Xj=0
ηi,i+j
βi,k = −rq−1Xj=k
ηi+k+1,i+j
Proposition 11 For a q—periodic orbit,
Xi = (1− a)q−1Xk=0
³−a√1−2ε
´ke−αi,k
ξi,i+k −³
−a√1−2ε
´qξi+k+1,i+q−1eαi,k−β
θi,i+k −³
−a√1−2ε
´qθi+k+1,i+q−1eαi,k−β
ηi,i+k +¡a2
1−2ε¢q
ηi+k+1,i+q−1 −³
−a√1−2ε
´qθi,i+q−1 (ηi,i+k + ηi+k+1,i+q−1
Proposition 12
Xi = (1− a)q−1Xk=0
(−a)k ξi,i+kΓ
k−ηi,i+kPk−1j=0 ηi,i+j − (−a)q ξi+k+1,i+q−1Γq+k+ηi,i+k
θi,i+kΓk−ηi,i+k
Pk−1j=0 ηi,i+j − (−a)q θi+k+1,i+q−1Γq+k+ηi,i+
Φ
where
Φ = ηi,i+k + (−a)2q Γ2qηi+k+1,i+q−1 − (−a)q
2θi,i+q−1 (ηi,i+k + ηi+k+1,i+q−1)Γq
³Γηi,i+k
Pq−1j=0 ηi,
and
Γ = er =1√1− 2ε
136
From (A.20), we plot the non-trivial1 and non-synchronized period-2 (see Figure
A.2), period-3 (see FigureA.3 ) and period-4 orbits (see Figures A.4 and A.5). We
omit the symmetric orbits (orbits that are symmetric relatively to the diagonal x = y
—see Section 3.3.1). A green circle is used to denote the points of the orbit for ε = 0
and the colour denotes the stability of the orbit2 (blue for an unstable orbit).
-1 0 1-1
0
1
θ = -1+1 ξ = +1-1
-1 0 1-1
0
1
θ = -1+1 ξ = +1+1
Figure A.2: Period-2 orbits for a = 2.
1Some orbits are not admissible for ε 6= 0 (they are only in [−1, 1]2 for ε = 0) and are thus notplotted.
2The only colour present on the Figures is blue, as the orbits are unstable when admissible.
137
θ = -1-1+1 ξ = -1+1-1
θ = -1-1+1 ξ = -1+1+1
θ = -1-1+1 ξ = +1+1-1
θ = -1-1+1 ξ = +1-1+1
θ = -1-1+1 ξ = +1+1+1
θ = -1+1+1 ξ = +1+1-1
θ = -1+1+1 ξ = +1+1+1
Figure A.3: Period-3 orbits for a = 2.
138
θ = -1-1-1+1 ξ = -1-1+1-1
θ = -1-1-1+1 ξ = -1+1-1-1
θ = -1-1-1+1 ξ = -1-1+1+1
θ = -1-1-1+1 ξ = +1+1-1-1
θ = -1-1-1+1 ξ = +1-1-1+1
θ = -1-1-1+1 ξ = -1+1-1+1
θ = -1-1-1+1 ξ = +1-1+1-1
θ = -1-1-1+1 ξ = -1+1+1+1
θ = -1-1-1+1 ξ = +1+1+1-1
θ = -1-1-1+1 ξ = +1+1-1+1
θ = -1-1-1+1 ξ = +1-1+1+1
θ = -1-1-1+1 ξ = +1+1+1+1
θ = -1-1+1+1 ξ = -1+1+1-1
θ = -1-1+1+1 ξ = +1+1-1-1
θ = -1-1+1+1 ξ = -1+1-1+1
θ = -1-1+1+1 ξ = +1-1+1-1
Figure A.4: Period-4 orbits for a = 2.
139
θ = -1-1+1+1 ξ = -1+1+1+1
θ = -1-1+1+1 ξ = +1+1+1-1
θ = -1-1+1+1 ξ = +1+1-1+1
θ = -1-1+1+1 ξ = +1-1+1+1
θ = -1-1+1+1 ξ = +1+1+1+1
θ = -1+1-1+1 ξ = -1+1+1+1
θ = -1+1-1+1 ξ = +1+1+1-1
θ = -1+1+1+1 ξ = +1+1+1-1
θ = -1+1+1+1 ξ = +1+1-1+1
θ = -1+1+1+1 ξ = +1+1+1+1
Figure A.5: Period-4 orbits for a = 2 (cont.).
140
Proposition 13 All periodic orbits with period q ≥ 2 are unstable for ε < 1−a2.
Proposition 14 A sufficient condition for synchronization of the coupled tent map
is
1− a2
< ε ≤ 12
Proposition 15 A sufficient condition for synchronization of a period q orbit of the
coupled tent map is
θi = ξi, ∀i ∈ Z
Proposition 16 The periodic orbits’ transversal Lyapunov exponent is zero for ε =
1−a2.
From Proposition 16, for ε = 1−a2the transversal Lyapunov Exponent, λ⊥, crosses
zero and the coupled tent map undergoes a blowout bifurcation. For a brief discussion,
we state the following
Definition 5 (Blowout Bifurcation [96]) Suppose that f : M → M , where M is
an m-dimensional manifold, and that under the action of f there exists an invariant
n-dimensional submanifold, N . Further, suppose that A ⊂ N is a chaotic attractor
for the restricted mapping f |N : N → N , with natural measure l (.). Then we say
that the system undergoes a blowout bifurcation, if the normal Lyapunov exponent
with respect to l passes through zero.
In Figure A.6 a mesh of 104 initial conditions for the coupled tent map (a =
12− 10−15) is iterated for 103 steps and plotted for several different values of ε. For ε
below, but arbitrarily close to ε = 1−a2, a non-synchronized attractor is born (observe
that this attractor intercepts the even quadrants).
141
-1 1-1
1
ε=1e-015
-1 1-1
1
ε=0.05
-1 1-1
1
ε=0.1
-1 1
-1
1
ε=0.2
-1 1
-1
1
ε=0.24
-1 1
-1
1
ε=0.247
-1 1-1
1
ε=0.25
-1 1-1
1
ε=0.3
-1 1-1
1
ε=0.5
Figure A.6: A mesh of 100 by 100 initial conditions is iterated for 103 steps and thestate of the coupled tent map with a = 1/2− 10−15 (a = 1/2 is a source of numericproblems) is then plotted.
142
Proof of Proposition 9: DefinePk
j=0Aj by
−1Xj=0
Aj = 0
kXj=0
Aj =k−1Xj=0
Aj +Ak
We proceed the proof by induction on n.
For n = 0,
Xi = (1− a) θi [0; ηi]U+µ −a√
1− 2ε¶θi [−rηi; ηi]Xi+1
Thus we recover (A.14).
The n → n + 1 step is proved as follows. From the hypothesis, (A.16) and
Proposition 30 (see Appendix B),
Xi = (1− a)nXk=0
µ −a√1− 2ε
¶kθi,i+k
"−r
k−1Xj=0
ηi,i+j ; ηi,i+k
#U+
µ −a√1− 2ε
¶n+1θi,i+n
"−r
nXj=0
ηi,i+j ; ηi,i+n
#·½
(1− a) θi+n+1 [0; ηi+n+1]U− a√1− 2εθi+n+1 [−rηi+n+1; ηi+n+1]Xi+n+2
¾So that
Xi = (1− a)nXk=0
µ −a√1− 2ε
¶kθi,i+k
"−r
k−1Xj=0
ηi,i+j ; ηi,i+k
#U+
(1− a)µ −a√
1− 2ε¶n+1
θi,i+n+1
"−r
nXj=0
ηi,i+j; ηi,+n+1
#U+
µ −a√1− 2ε
¶n+2θi,i+n+1
"−r
nXj=0
ηi,i+j; ηi,i+n
#[−rηi+n+1; ηi+n+1]Xi+n+2
= (1− a)n+1Xk=0
µ −a√1− 2ε
¶kθi,i+k
"−r
k−1Xj=0
ηi,i+j ; ηi,i+k
#U+
+
µ −a√1− 2ε
¶n+2θi,i+n+1
"−r
n+1Xj=0
ηi,i+j; ηi,i+n+1
#Xi+n+2 (A.23)
143
¤
Proof of Proposition 10: As n is a dumb index, we replace n+1 by q in Proposition
9, yielding
Xi+q =
µ−√1− 2εa
¶qθi,i+q−1
"−r
q−1Xj=0
ηi,i+j; ηi,i+q−1
#−1(Xi − (1− a)
q−1Xk=0
µ −a√1− 2ε
¶kθi,i+k
"−r
k−1Xj=0
ηi,i+j; ηi,i+k
#U
)(A.24)
Thus, we reach an expression of the type Xi+q = T q (Xi). Derivating both members
of (A.24), we obtain
Dq =∂Xi+q∂Xi
=
µ−√1− 2εa
¶qθi,i+q−1
"−r
q−1Xj=0
ηi,i+j ; ηi,i+q−1
#−1
=
µ−√1− 2εa
¶qθi,i+q−1
"rη1,q
q−1Xj=0
ηi,i+j; ηi,i+q−1
#(A.25)
From Proposition 32 (see Appendix B) and noting that i and q are dumb indexes on
(A.25), one concludes that the eigenvalues of Dq are
λq =
³−√1−2ε
a
´qθ1,qe
±rPq−1j=0 ηi,i+j η1,q = 1
±³−√1−2ε
a
´qθ1,q η1,q = −1
(A.26)
From (A.9), (A.26) can be simplified to
λq =
¡−1a
¢q ¡√1− 2ε¢q∓Pq−1
j=0 ηi,i+j θ1,q η1,q = 1
±³−√1−2ε
a
´qθ1,q η1,q = −1
(A.27)
From Proposition 33 (see Appendix B) and (A.9), the eigenvalues of DT , ϑq, are also
given by (A.27), that is, ϑq = λq. Therefore, the Lyapunov numbers are given by
eλ = (λqϑq)1/2q =
(√1−2ε)1∓
1qPq−1j=0 ηi,i+j
aη1,q = 1
√1−2εa
η1,q = −1Finally, note that the Lyapunov numbers do not depend on i, as
Pq−1j=0 ηi,i+j is inde-
pendent of i.¤
Proof of Proposition 11: Consider an orbit of period q. Proposition 9 with n = q − 1 (n
Xi = (1− a)q−1Xk=0
µ −a√1− 2ε
¶kθi,i+k
"−r
k−1Xj=0
ηi,i+j; ηi,i+k
#U+
µ −a√1− 2ε
¶qθi,i+q−1
"−r
q−1Xj=0
ηi,i+j ; ηi,i+q−1
#Xi
so that
Xi = (1− a)q−1Xk=0
µ −a√1− 2ε
¶kθi,i+k
Ã1−
µ −a√1− 2ε
¶qθi,i+q−1
"−r
q−1Xj=0
ηi,i+j ; ηi,i+q−1
#!−1
Observe that3
Ξ :=
Ã1−
µ −a√1− 2ε
¶qθi,i+q−1
"−r
q−1Xj=0
ηi,i+j ; ηi,i+q−1
#!−1 "−r
k−1Xj=0
ηi,i+j; ηi,i+k
#U
=
Ã"rηi,i+k
k−1Xj=0
ηi,i+j; ηi,i+k
#−µ −a√
1− 2ε¶q
θi,i+q−1
"rηi,i+k
k−1Xj=0
ηi,i+j; ηi,i+k
#"−r
q−1Xj=0
=
Ã"rηi,i+k
k−1Xj=0
ηi,i+j; ηi,i+k
#−µ −a√
1− 2ε¶q
θi,i+q−1
"−r
q−1Xj=k
ηi+k+1,i+j; ηi+k+1,i+q−1
#!
3Note that a−1b = a−1¡b−1¢−1
=¡b−1a
¢−1and
h−rPk−1
j=0 ηi,i+j ; ηi,i+ki−1
=hrηi,i+k
Pk−1j=0 ηi,i+j ; ηi,i+k
i
To simplify (A.29), we note that
(g [α; η]− h [β; ξ])−1U
=
g coshα− h cosh β gη sinhα− hξ sinh βg sinhα− h sinhβ gη coshα− hξ cosh β
−1U=
1
g2η + h2ξ − gh (η + ξ) cosh (α− β)
gη (coshα− sinhα)− hξ (cosh β −g (coshα− sinhα)− h (cosh β − si
=1
g2η + h2ξ − gh (η + ξ) cosh (α− β)
gηe−α − hξe−βge−α − he−β
We note that (A.29) can be simplified using (A.30) where
g ≡ 1
η ≡ ηi,i+k
ξ ≡ ηi+k+1,i+q−1
α ≡ rηi,i+kk−1Xj=0
ηi,i+j
β ≡ −rq−1Xj=k
ηi+k+1,i+j
h ≡µ −a√
1− 2ε¶q
θi,i+q−1
leading to:
Ξ =e−αi,k
ηi,i+k +¡a2
1−2ε¢q
ηi+k+1,i+q−1 −³
−a√1−2ε
´qθi,i+q−1 (ηi,i+k + ηi+k+1,i+q−1) cosh ηi,i+k −
³−a√1−2ε
´qθi,i+q−1ηi+k+1,i+q−1e−βi,k+αi,k
1−³
−a√1−2ε
´qθi,i+q−1e−βi,k+αi,k
Thus, (A.28) is simplified to
Xi = (1− a)q−1Xk=0
³−a√1−2ε
´ke−αi,k
ξi,i+k −³
−a√1−2ε
´qξi+k+1,i+q−1eαi,k−β
θi,i+k −³
−a√1−2ε
´qθi+k+1,i+q−1eαi,k−β
ηi,i+k +¡a2
1−2ε¢q
ηi+k+1,i+q−1 −³
−a√1−2ε
´qθi,i+q−1 (ηi,i+k + ηi+k+1,i+q−1
¤
Proof of Proposition 12: From Propositions 18 and 21 (see Appendix B),
αi,k = rηi,i+k
k−1Xj=0
ηi,i+j
= rk−1Xj=0
ηi+k+1,i+kηi+k+1,i+j
= rk−1Xj=0
ηi+k+1,i+j
From Definition 4 and Propositions 18 and 21 (see Appendix B),
αi,k − βi,k = r
Ãk−1Xj=0
ηi+k+1,i+j +
q−1Xj=k
ηi+k+1,i+j
!
= r
q−1Xj=0
ηi+k+1,i+j
= rηi+k+1,i+k
q−1Xj=0
ηi+k+1,i+j
= rηi,i+k
q−1Xj=0
ηi,i+j
= log
µ1√1− 2ε
¶ηi,i+kPq−1
j=0 ηi,i+j
Therefore,
cosh (αi,k − βi,k) =eαi,k−βi,k + e−(αi,k−βi,k)
2
=
¡√1− 2ε¢−ηi,i+kPq−1
j=0 ηi,i+j +¡√1− 2ε¢ηi,i+kPq−1
j=0 ηi
2
=1
2
³Γηi,i+k
Pq−1j=0 ηi,i+j + Γ−ηi,i+k
Pq−1j=0 ηi,i+j
´Finally,
eαi,k−βi,k = Γηi,i+kPq−1j=0 ηi,i+j
and
e−αi,k = e−rηi,i+kPk−1j=0 ηi,i+j
= Γ−ηi,i+kPk−1j=0 ηi,i+j
Using (A.33), (A.34) and (A.35) in Proposition 11
Xi = (1− a)q−1Xk=0
(−a)k ξi,i+kΓ
k−ηi,i+kPk−1j=0 ηi,i+j − (−a)q ξi+k+1,i+q−1Γq+k+ηi,i+k(
Pq−1j=k ηi,i+j
θi,i+kΓk−ηi,i+k
Pk−1j=0 ηi,i+j − (−a)q θi+k+1,i+q−1Γq+k+ηi,i+k(
Pq−1j=k ηi,i+j
Φ
Φ = ηi,i+k + (−a)2q Γ2qηi+k+1,i+q−1 − (−a)2
θi,i+q−1 (ηi,i+k + ηi+k+1,i+q−1)Γq³Γηi,i+k
Pq−1j=0 ηi
¤
Proof of Proposition 13: We show that the minimum Lyapunov number for ε < 1−a2is alwa
of period greater or equal to two. Thus, let us consider the two cases in (A.19), noting that
i) η1,q = −1. Then for 0 ≤ ε < 12,
e±λ =
√1− 2εa
>1√1− 2ε ≥ 1
ii) η1,q = 1. For 0 ≤ ε < 12, the smallest Lyapunov number is4
min(e±λ) = min(
¡√1− 2ε¢1∓ 1
q
Pq−1j=0 η1,1+j
a)
> min(
¡√1− 2ε¢1∓ 1
q
Pq−1j=0 η1,1+j
1− 2ε )
= min
Ã(1− 2ε)∓ 1
2q
Pq−1j=0 η1,1+j
√1− 2ε
!
=(1− 2ε) 12q
Pq−1j=0 η1,1+j
√1− 2ε
=¡√1− 2ε¢ 12(−1+ 1
q
Pq−1j=0 η1,1+j)
> 1
as −1 + 1q
Pq−1j=0 η1,1+j < 0.¤
4The positive synchronized fixed point Lyapunov numbers are 1 for a = 1− 2ε. What follows does not t
150
Proof of Proposition 14: (the reader is referred to Chapter ?? for background on
synchronization).
Under the variables si =xi+yi2
and di =xi−yi2, the dynamics is given by si+1 =
1−aa− |xi|+|yi|
a
di+1 =1−aa− |xi|−|yi|
a
Therefore,
di+1 =
−1−2εadi if si − di > 0 and si + di > 0
1−2εasi if si − di > 0 and si + di < 0
−1−2εasi if si − di < 0 and si + di > 0
1−2εadi if si − di < 0 and si + di < 0
We remember the reader that the map is said to synchronize if limi→∞ di = 0. Noting
that |di| < 1 and |si| < 1, we conclude that a sufficient condition for synchronizationis
1− 2εa
< 1⇔ ε >1− a2
Note that we are only interested in ε ≤ 12.
¤
Proof of Proposition 15: Immediate from the symmetry of (A.20).
151
¤
Proof of Proposition 16: We shall show that the transversal Lyapunov exponent
is zero for ε = 1−a2.
From Proposition ??, for 1−a2< ε ≤ 1
2all orbits are on the diagonal x = y.
Therefore, for 1−a2< ε ≤ 1
2, ηi,j = 1 ∀i, j ∈ 1, . . . q. Thus, (A.25) is simplified to
Dq =µ−√1− 2ε
a
¶qθi,i+q−1 [rq; 1]
=
µ−√1− 2εa
¶qθi,i+q−1
·−q2log (1− 2ε) ; 1
¸(A.39)
Where expression (A.9) has been used. From Proposition 32 (see Appendix B), the
eigenvalues and eigenvectors of (A.39) are given by
λ± (a, ε) =³−√1−2ε
a
´qθi,i+q−1 exp(∓ q
2log (1− 2ε))
−→v± = ±11
Therefore, the transversal eigenvector is −→v− = −11
, with corresponding eigenvalueλ− =
³−√1−2ε
a
´qθi,i+q−1 exp(
q2log (1− 2ε)). But λ− crosses the circle of radius 1 for
ε = 1−a2,
λ− (a, ε) = (−a)−q2 θi,i+q−1 exp(
q
2log a)
= (−1)− q2 θi,i+q−1
Finally, from Proposition 33, the transversal Lyapunov exponent is
λ⊥ = limq→∞
1
2qlog°°(−1)−q°° = 0
¤
152
Appendix B
Notation and Tools
In this Appendix we will summarize the main tools that we use in the study of
coupled tent maps. For this Appendix we always assume i, j, k, l, q ∈ Z.Let θii∈Z ∈ −1, 1Z. Define
a) θi,i = θi;
b) θi,i+n+1 = θi,i+nθi+n+1,i+n+1 .
Noting that θi,i ∈ −1, 1, b) can be written as
b’) θi,i+n = θi,i+n+1θi+n+1,i+n+1.
Thus, the sequence θi,ji,j∈Z ∈ −1, 1Z2
is well defined.
Proposition 17 θ2i,j = 1
Proposition 18 θi,j is a product with |j − i+ 1| terms defined by
θi,j =
jQs=i
θs if j ≥ i i)
1 if j = i− 1 ii)i−1Qs=j+1
θs if j ≤ i− 2 iii)
153
Proof: Replacing i+ n by j on b) gives
θi,j+1 = θi,jθj+1,j+1 = θi,jθj+1 (B.1)
Replacing i by j + 1 yields
θj+1,j+1 = θj+1,jθj+1
As θj+1,j+1 = θj+1, we conclude that θj+1,j = 1, which is ii).
From B.1 and ii) θi,j+1 = θi,jθj+1
θi,i−1 = 1(B.2)
which proves i).
Multiplying (B.2) by θj+1 and replacing j by j − 1, we obtain
θi,j−1 = θi,jθj (B.3)
for j = i− 1, i− 2, . . . . Together with ii)
θi,i−1 = 1
this proves iii).
Note, in particular, that relations (B.2) and (B.3) are equivalent to b) and b0), so
that together with a) they completely define the sequence θi,ji,j∈Z2 .¤
Proposition 19 θi,kθj,k does not depend on k.
Proof:
θi,k+1θj,k+1 = θi,kθj,kθ2k+1 = θi,kθj,k
θi,k−1θj,k−1 = θi,kθj,kθ2k = θi,kθj,k
¤
154
Proposition 20
θi,kθj,kθi,lθj,l = 1 (B.4)
Proof: Multiplying both members by θi,lθj,l yields
θi,kθj,k = θi,lθj,l
which is the result of Proposition 19.
¤
Proposition 21 θi,kθi,l does not depend on i.
Proof: Multiplying both members of (B.4) by θj,kθj,l,
θi,kθi,l = θj,kθj,l
¤
Proposition 22
θi,kθj,l = θj,kθi,l
Proof: Directly from Proposition 20.
Proposition 23
θi,j = θj+1,i−1
Proof: From Proposition 22 and 18,
θi,jθj+1,i−1 = θj+1,jθi,i−1 = 1
¤
155
Proposition 24
θi+1,j = θi,jθi
θi−1,j = θi,jθi−1
Proof: From Proposition 23,
θi,jθi = θj+1,i−1θi
= θj+1,i
= θi+1,j
and
θi,jθi−1 = θj+1,i−1θi−1
= θj+1,i−2
= θi−1,j
¤
An orbit is said to be (eventually) periodic if
∃i0 ∈ Z ∀i ≥ i0 ∃q ∈ Z : θi+q = θi (B.5)
Proposition 25 (If (B.5)).111
θi+q,j+q = θi,j
Proof: The proof follows by induction on j. For j = i,
θi+q,i+q = θi,i = θi
Assume the Proposition for j. Then
θi+q,j+1+q = θi+q,j+qθj+1+q = θi,jθj+1+q = θi,j+1
156
and
θi+q,j−1+q = θi+q,j+qθj+q = θi,jθj = θi,j−1
¤
Proposition 26 (If (B.5)).
θi+q,j = θi,j+q
Proof: From Propositions 25 and 22,
1 = θi+q,j+qθi,j = θi+q,jθi,j+q
¤
Proposition 27 (If (B.5)).
θi,i+q−1 = θj,j+q−1
Proof: From Propositions 23, 22, 25 and 23 again,
θj,j+q−1θi,i+q−1 = θj,j+q−1θi+q,i−1
= θj,i−1θi+q,j+q−1
= θj,i−1θi,j−1
= θi,j−1θi,j−1
= 1
¤
Proposition 28 ¯¯n−1Xl=0
θi,k+l
¯¯ =
¯¯n−1Xl=0
θj,k+l
¯¯
157
Proof: Remembering that θi,j ∈ −1, 1 and using Propositions 21and 18¯¯n−1Xl=0
θi,k+l
¯¯ =
¯¯θi,j−1
n−1Xl=0
θi,k+l
¯¯
=
¯¯n−1Xl=0
θi,j−1θi,k+l
¯¯
=
¯¯n−1Xl=0
θj,j−1θj,k+l
¯¯ =
¯¯n−1Xl=0
θj,k+l
¯¯
¤
Proposition 29 (If (B.5)). If θi,i+q−1 = 1, then¯¯q−1Xl=0
θi,i+l
¯¯ =
¯¯q−1Xl=0
θj,j+k
¯¯
Proof: Noting that θi,j ∈ −1, 1, from Propositions 28 and 27,¯¯q−1Xl=0
θi,i+l
¯¯ =
¯¯q−1Xl=0
θi+1,i+l
¯¯
=
¯¯q−1Xl=0
θi+1,i+1+l + θi+1,i − θi+1,i+q
¯¯
=
¯¯q−1Xl=0
θi+1,i+1+l
¯¯
¤
Define
[α; η] =
coshα η sinhα
sinhα η coshα
for α ∈ R, θ, η ∈ −1, 12.
Proposition 30
[α; η] [β; ξ] = [α+ ηβ; ηξ]
159
Proof: As η ∈ −1, 1, η2 = 1 and sinh(ηα) = η sinh(α) and cosh(ηα) = cosh(α),
thus we have
[α; η] [β; ξ] =
coshα η sinhα
sinhα η coshα
coshβ ξ sinh β
sinhβ ξ coshβ
=
cosh(α) cosh (β) + η sinh (α) sinh (β) ξ (η cosh (β) sinh (α) + cosh (α) sinh (β))
cosh(β) sinh (α) + η cosh (α) sinh (β) ξ (η cosh (α) cosh (β) + sinh (α) sinh (β))
=
cosh(α) cosh (ηβ) + sinh (α) sinh (ηβ) ηξ (cosh (ηβ) sinh (α) + cosh (α) sinh (ηβ)
cosh(ηβ) sinh (α) + cosh (α) sinh (ηβ) ηξ (cosh (α) cosh (ηβ) + sinh (α) sinh (ηβ)
=
cosh (α+ ηβ) ηξ sinh (α+ ηβ)
sinh (α+ ηβ) ηξ cosh (α+ ηβ)
= [α+ ηβ, ηξ]
¤
Proposition 31
[α; η]T = [ηα; η]
Proof: As η ∈ −1, 1, η2 = 1 and sinh(ηα) = η sinh(α) and cosh(ηα) = cosh(α).
160
Thus, coshα η sinhα
sinhα η coshα
T = coshα sinhα
η sinhα η coshα
=
cosh ηα η sinh ηα
sinh ηα η cosh ηα
¤
Proposition 32 The eigenvalues and eigenvectors of [α; η] are, respectively,λ1 = exp(±α)
−→v1 = ±11
if η = 1
λ−1 = ±1
−→v−1 = ¡tanh α
2
¢∓11
if η = −1
Proposition 33 The eigenvalues of [α; η]T are λ1 = exp(±α) if η = 1
λ−1 = ±1 if η = −1Proof: Follows from Propositions 31 and 32.
¤
Proposition 34
[α; η]−1 = [−ηα; η]
Proof: From Proposition 30,
[α; η] [−ηα; η] = £α+ η(−ηα); η2¤= [0; 1]
161
and
[−ηα; η] [α; η] = £−ηα + ηα; η2¤
= [0; 1]
¤
Proposition 35
N−1Yj=0
[αj ; ηj] =
"N−1Xj=0
Ãj−1Yk=0
ηk
!αj;
N−1Yk=0
ηk
#
Proof: The proof follows by induction on N . For N = 2, we recover Proposition 30.
The N → N + 1 step proceeds as follows,
NYj=0
[αj ; ηj ] =
ÃN−1Yj=0
[αj ; ηj]
![αN ; ηN ]
=
"N−1Xj=0
Ãj−1Yk=0
ηk
!αj;
N−1Yk=0
ηk
#[αN ; ηN ]
=
"N−1Xj=0
Ãj−1Yk=0
ηk
!αj +
ÃN−1Yk=0
ηk
!αN ;
ÃN−1Yk=0
ηk
!ηN
#
162
Bibliography
[1] Focus Issue on Control and Synchronization of Chaos, Chaos, 7 (4) (1997).
[2] ALICE Technical Design Report of the High Momentum Particle Identification
Detector . Technical Report, 18 August 1998.
[3] Aguado, Ramón and Gloria Platero. “Photoinduced Multistable Phenom-
ena in the Tunneling Current through Doped Superlattices,” Phys. Rev. Lett.,
81 (22):4971—4974 (1998).
[4] Altherr, T. and J. Seixas. “Cherenkov ring recognition using a non-adaptable
network,” Nuclear Instruments and Methods in Physics Research, 335—338
(1992).
[5] Arecchi, F. T., et al. “Generalized multistability and noise-induced jumps in a
nonlinear dynamical system,” Phys. Rev. A, 32 :402—408 (1985).
[6] Arecchi, F.T., et al. “Experimental Evidence of Subharmonic Bifurcations,
Multistability, and Turbulence in a Q-Switched Gas Laser,” Phys. Rev. Lett.,
49 (17):1217—1220 (1982).
[7] Arneodo, Alain, et al. “Causal cascade in the stock market from the infrared
to the ultraviolet,” Eur. Phys. J. B , 2 :277—282 (1998).
[8] Arnold, Ludwig. Random Dynamical Systems. Springer-Verlag, 1998.
163
[9] Arthur, W. Brian. “Complexity and the Economy,” Science, 284 (2):107—109
(1999).
[10] Arthur, W. Brian, et al. “Asset Pricing Under Endogenous Expectations in
an Artificial Stock Market.” The Economy as an Evolving, Complex System II
Addison-Wesley, 1997.
[11] Banks, John. “Regular periodic decompositions for topologically transitive
maps,” Ergodic Theory Dynam. Systems, 17 :505—529 (1997).
[12] Barbosa, Jose, et al. “A Cellular Automata Algorithm for Pattern Recog-
nition in HMPID.” New Computing Techniques in Physics Research edited
by G. Athanasiu and D. Perret-Gallix, 8—17, Parisianou, 2000. Available at
ftp://ftp.econodynamics.com/pub/hmpid.zip.
[13] Boffetta, G., et al., “Predictability: a way to characterize Complexity,” 2001.
available at http://xxx.lanl.gov/abs/nlin.CD/0101029.
[14] Bouchaud, Jean-Philippe and Marc Potters. Theory of Financial Risk: from
data analysis to risk management . Science and Finance, 2000.
[15] Brock, William A. and Patrick de Fontnouvelle. “Expectational diversity in
monetary economies,” J. Econ. Dynam. Control , 24 :725—759 (2000).
[16] Campbell, John, et al. The Econometrics of Financial Markets. Princeton
University Press, 1997.
[17] Campbell, Sue Ann, et al. “Complex dynamics and multistability in a damped
harmonic oscillator with delayed negative feedback,” Chaos, 5 (4):640—645
(1995).
164
[18] Carvalho, Rui. “The Dynamics of the Linear Random Farmer Model.” Complex
Behavior in Economics- Modeling, Computing and Mastering Complexity 2000.
(to appear) Available at ftp://ftp.econodynamics.com/pub/apfa.pdf.
[19] Carvalho, Rui, et al. “From Synchronization to Multistability in Two Cou-
pled QuadraticMaps,” (2000). submitted to Phys. Lett. A. Available at
http://xxx.lanl.gov/abs/nlin/0005053.
[20] Carvalho, Rui, et al. “Feigenbaum Networks,” Phys. D, 126 :27—37 (1999).
[21] Chate, Hugues and Jerome Losson. “Non-trivial Collective behavior in coupled
map lattices: A transfer operator perspective,” Phys. D, 103 :51—72 (1997).
[22] Chiarella, Carl. “The Dynamics of Speculative Behaviour,” Ann. Oper. Res.,
37 :101—123 (1992).
[23] Choi, M. Y., et al. “Synchronization in a system of globally coupled oscillators
with time delay,” Phys. Rev. E , 61 (1):371—381 (2000).
[24] Cocho, G. and G. Martinez-Mekler. “On a coupled map lattice formulation of
genetic sequences,” Phys. D, 51 :119—130 (1991).
[25] Collet, Pierre and Jean-Pierre Eckmann. Iterated maps on the interval as dy-
namical systems (second Edition), 1 . Progess in Physics. Boston: Birkhauser,
1983.
[26] Cont, Rama. “Modeling economic randomness: statistical mechanics of market
phenomena.” Statistical Physics on the Eve of the 2st Century 14 . Series on
Advances in Statistical Mechanics, edited by Murray T Batchelor and Luc T
Wille, 47—64, World Scientific, 1999.
165
[27] Cont, Rama and Jean-Philippe Bouchaud. “Herd Behaviour and aggregate
fluctuations in financial markets,” cond-mat/9712318v2 (1998).
[28] Cont, Rama, et al. “Scaling in stock market data: stable laws and beyond.”
Scale Invariance and Beyond edited by B. Dubrulle, et al., Berlin: Springer,
1997.
[29] Cozza, D., et al. Development of a pattern recognition method for the HMPID
in ALICE . ALICE Internal Note, September 1998.
[30] Dacorogna, M., et al. The Distribution of Extremal Foreign Exchange Rate
Returns in Extremely Large Data Sets. Technical Report, 1995.
[31] Devaney. Chaotic Dynamical Systems (Second Edition). Addison-Wesley, 1989.
[32] Ding, Z., et al. “A Long Memory Property of Stock Returns and a New Model,”
J. Empirical Finance, 1 :83 (1993).
[33] Dress, Holger, et al., “How to make a Hill plot,” 1999. Available at
http://www.orie. cornell.edu/trlist/trlist.html.
[34] Dutta, Mitrajit, et al. “Multiple Attractor Bifurcations: A Source of Unpre-
dictability in Piecewise Smooth Systems,” Phys. Rev. Lett., 83 (21):4281—4284
(1999).
[35] Eckmann, J.-P. and D. Ruelle. “Ergodic theory of chaos and strange attractors,”
Rev. Modern Phys., 57 (3):617—656 (1985).
[36] Eguia, M. C., et al. “Information transmission and recovery in neural commu-
nications channels,” Phys. Rev. E , 62 (5):7111—7122 (2000).
[37] Embrechts, Paul, et al. “Living on the Edge,” RISK , 11 (1):96—100 (1998).
166
[38] Fama, E.F. “Market efficiency, long-term returns and behavioral finance,” J.
Finan. Econ., 49 :283—306 (1998).
[39] Fama, Eugene F. “Efficient Capital Markets: A Review of Theory and Empirical
Work,” J. Finance, 25 :383—417 (1970).
[40] Farmer, J. Doyne. “Market force, ecology, and evolution,” Santa Fe Institute
Working Paper 98-12-117 (1998).
[41] Farmer, J. Doyne. “Physicists attempt to scale the ivory towers of finance,”
Computing in Science and Engineering, 1 :26—39 (1999).
[42] Farmer, J. Doyne. “A simple model for the nonequilibrium dynamics and evo-
lution of a financial market,” Int. J. Theor. Appl. Fin., 3 (3):425—441 (2000).
[43] Farmer, J. Doyne and Shareen Joshi. “The price dynamics of common trading
strategies,” J. Econ. Behav. Organ. (2001). To appear.
[44] Farmer, J. Doyne and J. J. Sidorowich. “Predicting chaotic time series,” Phys.
Rev. Lett., 62 :845—848 (1987).
[45] Feller, W. An Introduction to Probability Theory and Its Applications , 2 . New
York: John Wiley & Sons, 1971.
[46] Fernandez, Bastien. Instabilités et fronts dans les réseaux d’applications cou-
plées. PhD dissertation, Université de la Méditerranée, 1197.
[47] Fernandez, Bastien and Miaohua Jiang, “Dynamics of two diffusively coupled
unimodal maps with zero topological entropy,” 2000. Personal Communication.
[48] Feudel, Ulrike and Celso Grebogi. “Multistability and the control of complex-
ity,” Chaos, 7 (4):597—604 (1997).
167
[49] Feudel, Ulrike, et al. “Map with more than 100 coexisting low-period attrac-
tors,” Phys. Rev. E , 54 (1):71—81 (1996).
[50] Foss, Jennifer, et al. “Multistability and Delayed Recurrent Loops,” Phys. Rev.
Lett., 76 (4):708—711 (1996).
[51] Ghashghaie, S., et al. “Turbulent Cascades in Foreign Exchange Markets,”
Nature, 381 :767—770 (1996).
[52] Gopikrishnan, P., et al. “Inverse Cubic Law for the Distribution of Stock Price
Variations,” Eur. Phys. J. B , 3 :139—140 (1998).
[53] Gopikrishnan, Parameswaran, et al. “Scaling of the distribution of fluctuations
of financial market indices,” Phys. Rev. E , 60 :5305—5316 (1999).
[54] Grebogi, C., et al. “Crises, sudden changes in chaotic attractors, and transient
chaos,” Phys. D, 7 :181—200 (1983).
[55] Guckenheimer, J. and P.J. Holmes. Nonlinear Oscillations, Dynamical Systems,
and Bifurcations of Vector Fields (fifth Edition). Berlin: Springer, 1997.
[56] Haan, Laurens de, et al. “Extremal Behaviour of Solutions to a Stochastic
Difference Equation with Applications to ARCH Processes,” Stochastic Process.
Appl., 32 :213—224 (1989).
[57] Hanson, James E. Computational Mechanics of Cellular Automata. PhD dis-
sertation, University of California, 1993.
[58] Hao, Bai-Lin and Wei-Mou Zhang. Applied Symbolic Dynamics and Chaos.
Singapore: World Scientific, 1998.
[59] Hasler, M., et al. “Simple example of partial synchronization of chaotic sys-
tems,” Phys. Rev. E , 58 :6843—6846 (1998).
168
[60] Heagy, J.F., et al. “Characterization of on-off intermittency,” Phys. Rev. E ,
49 (2):1140—1150 (1994).
[61] Iori, Giulia. “A microsimulation of traders activity in the stock market: the
role of heterogeneity, agents’ interactions and trade frictions,” J. Econ. Behav.
Organ. (2001). in press.
[62] Ito, Kiyosi, editor. Encyclopedic Dictionary of Mathematics. MIT Press, 1996.
[63] Jonsson, Fredrik and Christos Flytzanis. “Polarization State Controlled Mul-
tistability of a Nonlinear Magneto-optic Cavity,” Phys. Rev. Lett., 82 (7):1426—
1429 (1999).
[64] Kaneko, Kunihiko. “Dominance of Milnor Attractors and Noise-Induced Selec-
tion in a Multiattractor System,” Phys. Rev. Lett., 78 (14):2736—2739 (1997).
[65] Kesten, Harry. “Random Difference Equations and Renewal Theory for Prod-
ucts of Random Matrices,” Acta Math., (131):207—248 (1973).
[66] Kim, Seunghwan, et al. “Multistability in Coupled Oscillator Systems with
Time Delay,” Phys. Rev. Lett., 79 (15):2911—2914 (1997).
[67] Komarov, A. K. and K. P. Komarov. “Multistability and hysteresis phenemena
in passive mode-locked lasers,” Phys. Rev. E , 62 (6):R7607—7610 (2000).
[68] Kurten, Karl E. and Grégoire Nicolis. “Bifurcation scenarios and quasiperiod-
icity in coupled maps,” Phys. A, 245 :446—452 (1997).
[69] Kyle, Albert S. “Continuous Auctions and Insider Trading,” Econometrica,
53 (6):1315—1335 (1985).
[70] LeBaron, Blake. “Agent-based computational finance: Suggested readings and
early research,” J. Econ. Dynam. Control , 24 :679—702 (2000).
169
[71] LeBaron, Blake, et al. “Time series properties of an artificial stock market,” J.
Econ. Dynam. Control , 23 :1487—1516 (1999).
[72] Levine, Herbert and William N. Reynolds. “Coupled map lattice techniques for
simulating interfacial phenomena in reaction-diffusion systems,” Chaos, 2 :337—
342 (1992).
[73] Lorenz, Edward N. “Deterministic Nonperiodic Flow,” J. Atmospheric Sci.,
20 :130—141 (1963).
[74] Losson, Jérôme, et al. “Aperiodic mean-field evolutions in coupled map lat-
tices,” Phys. Rev. E , 57 (5):4921—4932 (1998).
[75] Lux, Thomas. “The socio-economic dynamics of speculative markets: interact-
ing agents, chaos, and the fat tails of return distributions,” J. Econ. Behav.
Organ., 33 :143—165 (1998).
[76] Lux, Thomas and Marcel Ausloos. “Market Fluctuations I: Scaling, Multi-
Scaling and Their Possible Origins.” Theories of Disasters: Scaling Laws Gov-
erning Weather, Body and Stock Market Dynamics edited by A. Bunde and H.-
J. Schellnhuber, Berlin: Springer. (in press) Available at http://www.bwl.uni-
kiel.de/vwlinstitute/gwrp/german/publ/publ.html.
[77] Lux, Thomas and Michele Marchesi. Volatility Clustering in Financial Mar-
kets: A Microsimulation of Interacting Agents. Technical Report B-437, Bonn
University, July 1998.
[78] Lux, Thomas and Michele Marchesi. “Scaling and Criticality in a Stochastic
Multi-Agent Model of a Financial Market,” Nature, 397 :498 — 500 (1999).
170
[79] MacKay, R. S. and J.-A. Sepulchre. “Multistability in networks of weakly
coupled bistable units,” Phys. D, 82 :243—254 (1995).
[80] Mantegna, Rosario and H. Eugene Stanley. “Scaling Behavior in the Dynamics
of an Economic Index,” Nature, 376 :46—49 (1995).
[81] Mantegna, Rosario and H. Eugene Stanley. An Introduction to Econophysics:
correlation and complexity in finance. Cambridge University Press, 1999.
[82] Martin, Olivier, et al. “Algebraic Properties of Cellular Automata,” Comm.
Math. Phys., 93 :219—258 (1984).
[83] Mauro, A. di and A. Morsch. A Full Simulation of the RICH Response in the
ALICE Radiation Environment. ALICE Internal Note, September 1998.
[84] Mendes, Rui Vilela. “Multistability in Dynamical Systems.” 1999. Available at
http://xxx.lanl.gov/abs/nlin.CD/9904004.
[85] Mendes, Rui Vilela, et al. A process-reconstruction analysis of market fluc-
tuations. Technical Report, 2001. available at http://xxx.lanl.gov/abs/cond-
mat/0102301.
[86] Milnor, J. and W. Thurston. “On Iterated Maps of the Interval.” Lecture Notes
in Mathematics 1342 , edited by J. Alexander, 465—563, Springer, 1988.
[87] Mira, Chistian, et al. Chaotic Dynamics in Two-Dimensional Nonivertible
Maps, 20 . Nonlinear Science. Singapore: World Scientific, 1996.
[88] Moon, Hie Tae. “Strange attractor of the modulated Stokes wave: A universal
form,” Phys. Rev. E , 47 (2):R772—R775 (1993).
[89] Nishimura, Kazuo and Gerhard Sorger. “Optimal Cycles and Chaos: A Survey,”
Stud. Nonlinear Dynam. Econometrics, 1 (1):11—28 (1996).
171
[90] Nusse, Helena E., et al. “Border collision bifurcations: an explanation for
observed bifurcation phenomena,” Phys. Rev. E , 49 :1073—1076 (1994).
[91] Ódor, G., et al., “Parallel Algorithms for the ASTRA simd machine,” 1995.
[92] O’Hara, Maureen. Market Microstructure Theory. Blackwell, 1995.
[93] Olsen, Richard B., et al. Going Back to the Basics -Rethinking Market Effi-
ciency. Technical Report, 1992.
[94] Osinga, Hinke, et al. “Multistability and nonsmooth bifurcations in the
quasiperiodically forced circle map,” To appear in International Journal of Bi-
furcations and Chaos. Available at http://xxx.lanl.gov/abs/nlin.CD/0005032.
[95] Osipov, G. V. and M. M. Sushchik. “Synchronized clusters and multistabil-
ity in arrays of oscillators with different natural frequencies,” Phys. Rev. E ,
58 (6):7198—7207 (1998).
[96] Ott, E. and J. C. Sommerer. “Blowout bifurcations: the occurence of riddled
basins and on-off intermittency,” Phys. Lett. A, 188 :39—47 (1994).
[97] Ott, Edward, et al. “Controlling Chaos,” Phys. Rev. Lett., 64 (11):1196—1199
(1990).
[98] Park, Seon Hee, et al. “Effects of time-delayed interaction on dynamic patterns
in a coupled phase oscillator system,” Phys. Rev. E , 60 (4):4962—4965 (1999).
[99] Parry, William. “The Lorenz attractor and a related population model.” Er-
godic Theory, Proceedings, Oberwolfach, Germany 1978 729 . Lecture Notes in
Mathematics, edited by M. Denker and K. Jacobs, New York: Springer, 1979.
[100] Platt, N., et al. “On-Off Intermittency: A Mechanism for Bursting,” Phys. Rev.
Lett., 70 (3):279—282 (1993).
172
[101] Plerou, Vasiliki, et al. “Scaling of the distribution of price fluctuations of indi-
vidual companies,” Phys. Rev. E , 60 :6519—6529 (1999).
[102] Poon, Leon and Celso Grebogi. “Controlling Complexity,” Phys. Rev. Lett.,
75 (22):4023—4026 (1995).
[103] Prengel, F., et al. “Simple model for multistability and domain formation in
semiconductor superlattices,” Phys. Rev. B , 50 (3):1705—1712 (1994).
[104] Rim, Sunghwan, et al. “Chaotic Transition of Random Dynamical Systems and
Chaos Synchronization by Common Noises,” Phys. Rev. Lett., 85 (11):2304—
2307 (2000).
[105] Rossi, Alfredo De, et al. “Stability, Multistability, and Wobbling of Optical
Gap Solitons,” Phys. Rev. Lett., 81 (1):85—88 (1998).
[106] Rulkov, Nikolai F. “Images of synchronized chaos: Experiments with Circuits,”
Chaos, 6 :262—279 (1996).
[107] Sarno, Lucio and Mark P. Taylor. “Moral hazard, asset price bubbles, capital
flows, and the East Asian crisis: the first tests,” J. Int. Money Finance, 18 :637
(1999).
[108] Scheinkman, José and Blake LeBaron. “Nonlinear Dynamics and Stock Re-
turns,” J. Bus., 62 (3):311—337 (1989).
[109] Seydel, Rudiger. Practical Bifurcation and Stability Analysis. New York:
Springer-Verlag, 1994.
[110] Shiau, Yuo-Hsien, et al. “Multistability and symmetry breaking in the two-
dimensional flow around a square cylinder,” Phys. Rev. E , 60 (5):6188—6191
(1999).
173
[111] Singer, D. “Stable orbits and bifurcation of maps of the interval,” SIAM J.
Appl. Math., 35 :260—267 (1978).
[112] Skarda, C. A. and W. J. Freeman. “How brains make chaos in order to make
sense of the world.,” Behavioral and Brain Sciences, 10 :161—195 (1987).
[113] Sole, Ricardo V., et al. “Nonequilibrium dynamics in lattice ecosystems:
Chaotic stability and dissipative structures,” Chaos, 2 (3):387—395 (1992).
[114] Sornette, Didier and Rama Cont. “Convergent Multiplicative Processes Re-
pelled from Zero: Power Laws and Truncated Power Laws,” J. Phys. I , 7 :431—
444 (1997).
[115] Strien, Sebastien van. “Smooth Dynamics on the Interval.” New Directions in
Dynamical Systems edited by T. Bedford and J. Swift, Cambridge University
Press, 1988.
[116] Taborov, A.V. and Yu. L. Maistrenko. “Partial synchronization in a system of
coupled logistic maps,” Int. J. Bif. Chaos, 10 (5):1051—1066 (2000).
[117] Thearling, Kurt. “Massively Parallel Architectures and Algorithms for Time
Series Analysis,” Santa Fe Institute Working Paper 92-11-055 (1996).
[118] Theiler, James, et al. “Testing for Nonlinearity in Time Series: The Method of
Surrogate Data,” Phys. D, 58 :77—94 (1992).
[119] Wolfram, Stephen. “Statistical Mechanics of Cellular Automata,” Rev. Modern
Phys., 55 :601—644 (1983).
[120] Wolfram, Stephen. “Universality and Complexity in Cellular Automata,” Phys.
D, 10 :1—35 (1984).
174
[121] Wolfram, Stephen. “Random Sequence Generation by Cellular Automata,”
Adv. in Appl. Math., 7 :123—169 (1986).
[122] Yu, Lei, et al. “Transitions to Chaos for Random Dynamical Systems,” Phys.
Rev. Lett., 65 (24):2935—2938 (1990).