Lecture Notes in Physics 915

Charalampos (Haris) Skokos

GeorgA.Gottwald
JacquesLaskar Editors

Chaos
Detection
and
Predictability
Lecture Notes in Physics

Volume 915

Founding Editors
W. Beiglbck
J. Ehlers
K. Hepp
H. Weidenmller

Editorial Board
M. Bartelmann, Heidelberg, Germany
B.-G. Englert, Singapore, Singapore
P. Hnggi, Augsburg, Germany
M. Hjorth-Jensen, Oslo, Norway
R.A.L. Jones, Sheffield, UK
M. Lewenstein, Barcelona, Spain
H. von Lhneysen, Karlsruhe, Germany
J.-M. Raimond, Paris, France
A. Rubio, Donostia, San Sebastian, Spain
M. Salmhofer, Heidelberg, Germany
S. Theisen, Golm, Germany
D. Vollhardt, Augsburg, Germany
J. Wells, Michigan, USA
G.P. Zank, Huntsville, USA
The Lecture Notes in Physics

The series Lecture Notes in Physics (LNP), founded in 1969, reports new devel-
opments in physics research and teaching-quickly and informally, but with a high
quality and the explicit aim to summarize and communicate current knowledge in
an accessible way. Books published in this series are conceived as bridging material
between advanced graduate textbooks and the forefront of research and to serve
three purposes:
to be a compact and modern up-to-date source of reference on a well-defined
topic
to serve as an accessible introduction to the field to postgraduate students and
nonspecialist researchers from related areas
to be a source of advanced teaching material for specialized seminars, courses
and schools
Both monographs and multi-author volumes will be considered for publication.
Edited volumes should, however, consist of a very limited number of contributions
only. Proceedings will not be considered for LNP.
Volumes published in LNP are disseminated both in print and in electronic for-
mats, the electronic archive being available at springerlink.com. The series content
is indexed, abstracted and referenced by many abstracting and information services,
bibliographic networks, subscription agencies, library networks, and consortia.
Proposals should be sent to a member of the Editorial Board, or directly to the
managing editor at Springer:

Christian Caron
Springer Heidelberg
Physics Editorial Department I
Tiergartenstrasse 17
69121 Heidelberg/Germany
[email protected]

More information about this series at http://www.springer.com/series/5304

Charalampos (Haris) Skokos Georg A. Gottwald
Jacques Laskar
Editors

Chaos Detection
and Predictability

123
Editors
Charalampos (Haris) Skokos Georg A. Gottwald
Department of Mathematics and Applied School of Mathematics and Statistics
Mathematics University of Sydney
University of Cape Town Sydney, Australia
Rondebosch, South Africa

Jacques Laskar
Observatoire de Paris
IMCCE
Paris, France

ISSN 0075-8450 ISSN 1616-6361 (electronic)

Lecture Notes in Physics
ISBN 978-3-662-48408-1 ISBN 978-3-662-48410-4 (eBook)
DOI 10.1007/978-3-662-48410-4
Library of Congress Control Number: 2015959798

Springer Heidelberg New York Dordrecht London

Springer-Verlag Berlin Heidelberg 2016
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made.

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
Preface

Being able to distinguish chaoticity from regularity in deterministic dynamical

systems, as well as to specify the subspace of the phase space in which instabilities
are expected to occur, is of utmost importance in as disparate areas as astronomy,
particle physics, and climate dynamics. The presence of chaos introduces limitations
in our ability to accurately predict the evolution of a dynamical system at scales of
different sizes. In many practical applications it is of great importance to determine
the significance of this effect to the overall dynamics of the system. For this reason,
the development of precise and efficient numerical tools for distinguishing between
order and chaos, both locally and globally, becomes imperative, especially in the
case of multidimensional systems, whose phase space is not easily visualized.
Nowadays there exists a plethora of such methods.
The workshop Methods of Chaos Detection and Predictability: Theory and
Applications which was held in June 2013 at the Max Planck Institute for the
Physics of Complex Systems, in Dresden, Germany, brought together specialists
who have developed such methods, as well as researchers applying those techniques
to a variety of problems in the natural sciences. This book reviews the theory and
numerical implementation of several of the existing methods of chaos detection and
predictability and presents the current state of the art. Its chapters are written by
the creators of these methods and/or by well-established experts included in the
workshops list of invited speakers.
The most commonly employed method for investigating chaotic dynamics is the
computation of the Lyapunov Exponents (LEs). These are asymptotic measures
characterizing the average rate of growth (or shrinking) of small perturbations to
the orbits of a dynamical system, with the positivity of the maximum LE (mLE)
indicating chaoticity. The basic concepts of LEs are presented in the first chapter of
the book written by U. Parlitz, where the particular case of the LEs estimation for
time series is discussed and analyzed in depth.

v
vi Preface

As successful and illuminating LEs have been to characterize chaoticity in

deterministic dynamical systems, they suffer in certain situations from serious
drawbacks: For example, their computed values can vary significantly in time
and may only be used in the long time limit when the exponents have converged
with satisfactory accuracy. Furthermore, in the case of (noisy) experimental data,
they rely on phase space reconstruction methods, whose inattentive implementation
might produce unreliable results.
In the last two decades, several methods have been developed for the fast and
reliable determination of the regular or chaotic nature of orbits which were aimed
to surmount the shortcomings of the traditional methods involving LEs and phase
space reconstruction. These methods can be divided in two broad categories: those
which are based on the study of the evolution of deviation vectors from a given
orbit, like the computation of the mLE, and those which rely on the analysis of the
particular orbit itself.
A technique closely related to the computation of the mLE, which exploits the
information provided by the short time evolution of a deviation vector, is the Fast
Lyapunov Indicator (FLI) discussed in the second chapter of the book by E. Lega,
M. Guzzo, and C. Froeschl. The next chapter by R. Barrio deals with some variants
of the FLI method, namely, the Orthogonal Fast Lyapunov Indicator (OFLI and
OFLI2). The method of the Mean Exponential Growth factor of Nearby Orbits
(MEGNO), which again is based on the evolution of one deviation vector from the
reference orbit, is presented in the next chapter by P. Cincotta and M. Giordano.
The utilization of more than one deviation vector for the characterization of chaos
is considered in the next chapter by Ch. Skokos and Th. Manos where the methods
of the Smaller (SALI) and the Generalized Alignment Index (GALI) are presented.
The method of the Relative Lyapunov Indicator (RLI) where the differences of the
finite-time estimators of the mLE of two nearby orbits are used to characterize chaos
is the content of the next chapter by Z. Sndor and N. Maffione.
In the following chapter by G. Gottwald and I. Melbourne, the 0-1 test for
chaos is discussed in detail. Contrary to the five previous chapters, the analysis in
the 0-1 test for chaos is performed directly on the actual orbit (or time series).
The presence of chaos and a positive mLE is often seen as a limitation to the
predictability time of the underlying system, which is crudely estimated to be
inversely proportional to the mLE (the so-called Lyapunov time). The situation in
complex systems evolving on several temporal scales, like for example, in weather
forecasting models, can be, however, much more intricate as is shown in the last
chapter of the book by S. Siegert and H. Kantz: reliable predictions can be made
for times much longer than suggested by the predictability horizon implied by the
Lyapunov time.
We hope that this book will be useful both for young scholars, like graduate
students, Ph.D. candidates, and postdocs, and for specialists aiming at an up-to-
date review of some of the most widely used techniques of chaos detection and
predictability.
Preface vii

We thank the Springer Editorial Board, the authors, as well as the reviewers of
all chapters, for their work and effort which made possible the publication of this
volume.

Cape Town, South Africa Ch. Skokos

Sydney, Australia G. Gottwald
June 2015
The original version of the book was revised: The first editors name was corrected.

The publisher apologizes for having published an early draft edition of the preface.
This has been updated to the final version. The Erratum to the book is available at
DOI 10.1007/978-3-662-48410-4_9
Contents

1 Estimating Lyapunov Exponents from Time Series . .. . . . . . . . . . . . . . . . . . . . 1

Ulrich Parlitz
2 Theory and Applications of the Fast Lyapunov Indicator
(FLI) Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 35
Elena Lega, Massimiliano Guzzo, and Claude Froeschl
3 Theory and Applications of the Orthogonal Fast Lyapunov
Indicator (OFLI and OFLI2) Methods . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 55
Roberto Barrio
4 Theory and Applications of the Mean Exponential Growth
Factor of Nearby Orbits (MEGNO) Method. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 93
Pablo M. Cincotta and Claudia M. Giordano
5 The Smaller (SALI) and the Generalized (GALI)
Alignment Indices: Efficient Methods of Chaos Detection . . . . . . . . . . . . . . 129
Charalampos (Haris) Skokos and Thanos Manos
6 The Relative Lyapunov Indicators: Theory and Application
to Dynamical Astronomy .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 183
Zsolt Sndor and Nicols Maffione
7 The 0-1 Test for Chaos: A Review .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 221
Georg A. Gottwald and Ian Melbourne
8 Prediction of Complex Dynamics: Who Cares About Chaos? . . . . . . . . . . 249
Stefan Siegert and Holger Kantz
Erratum to: Chaos Detection and Predictability . . . . . . . . .. . . . . . . . . . . . . . . . . . . . E1
Charalampos (Haris) Skokos, Georg A. Gottwald, Jacques Laskar

xi
Chapter 1
Estimating Lyapunov Exponents from Time
Series

Ulrich Parlitz

Abstract Lyapunov exponents are important statistics for quantifying stability

and deterministic chaos in dynamical systems. In this review article, we first
revisit the computation of the Lyapunov spectrum using model equations. Then,
employing state space reconstruction (delay coordinates), two approaches for
estimating Lyapunov exponents from time series are presented: methods based on
approximations of Jacobian matrices of the reconstructed flow and so-called direct
methods evaluating the evolution of the distances of neighbouring orbits. Most
direct methods estimate the largest Lyapunov exponent, only, but as an advantage
they give graphical feedback to the user to confirm exponential divergence. This
feedback provides valuable information concerning the validity and accuracy of the
estimation results. Therefore, we focus on this type of algorithms for estimating
Lyapunov exponents from time series and illustrate its features by the (iterated)
Hnon map, the hyper chaotic folded-towel map, the well known chaotic Lorenz-
63 system, and a time continuous 6-dimensional Lorenz-96 model. These examples
show that the largest Lyapunov exponent from a time series of a low-dimensional
chaotic system can be successfully estimated using direct methods. With increasing
attractor dimension, however, much longer time series are required and it turns out
to be crucial to take into account only those neighbouring trajectory segments in
delay coordinates space which are located sufficiently close together.

1.1 Introduction

Lyapunov exponents are a fundamental concept of nonlinear dynamics. They

quantify local stability features of attractors and other invariant sets in state space.
Positive Lyapunov exponents indicate exponential divergence of neighbouring
trajectories and are the most important attribute of chaotic attractors. While the
computation of Lyapunov exponents for given dynamical equations is straight

U. Parlitz ()
Max Planck Institute for Dynamics and Self-Organization, Am Faberg 17, 37077 Gttingen,
Germany
e-mail: [email protected]

Springer-Verlag Berlin Heidelberg 2016 1

Ch. Skokos et al. (eds.), Chaos Detection and Predictability, Lecture Notes
in Physics 915, DOI 10.1007/978-3-662-48410-4_1
2 U. Parlitz

forward, their estimation from time series remains a delicate task. Given a univariate
(scalar) time series the first step is to use delay coordinates to reconstruct the state
space dynamics. Using the reconstructed states there are basically two approaches
to solve the estimation problem: With Jacobian matrix based methods a (local)
mathematical model is fitted to the temporal evolution of the states that can then
be used like any other dynamical equation. Using this approach in principle all
Lyapunov exponents can be estimated if the chosen black-box model is in very good
agreement with the underlying dynamics. In practical applications such a high level
of fidelity is often difficult to achieve, in particular since the time series typically
contain only limited information about contracting directions in state space. With
the second approach for estimating (at least the largest) Lyapunov exponents the
local divergence of trajectory segments in reconstructed state space is assessed
directly. Advantage of this kind of direct methods is their low number of estimation
parameters, easy implementation, and last but not least, direct graphical feedback
about the (non-) existence of exponential divergence in the given time series.
The following presentation is organised as follows: In Sect. 1.2 the standard
algorithm for computing Lyapunov exponents using dynamical model equations
is revisited. Methods for computing Lyapunov exponents from time series are
presented in Sect. 1.3. In Sect. 1.4 four dynamical systems are introduced to generate
time series which are then in Sect. 1.5 used as examples for illustrating and
evaluating features of direct estimation of the largest Lyapunov exponent. The
examples are: the Hnon map, the hyper chaotic folded-towel map, the Lorenz-
63 system, and a 6-dimensional Lorenz-96 model. These time discrete and time
continuous models exhibit deterministic chaos of different dimensionality and
complexity. In Sect. 1.6 a summary is given and the Appendix contains some
information for those readers who are interested in implementing Jacobian based
estimation algorithms.

1.2 Computing Lyapunov Exponents Using Model Equations

Lyapunov exponents characterize and quantify the dynamics of (infinitesimally)

small perturbations of a state or trajectory in state space. Let the dynamical model
be a M-dimensional discrete

x.n C 1/ D g.x.n// (1.1)

or a continuous
dx
xP D D f.x/ (1.2)
dt
1 Estimating Lyapunov Exponents from Time Series 3

dynamical system generating a flow

 t W RM ! RM (1.3)

with discrete t D n 2 Z or continuous t 2 R time. The temporal evolution of an

infinitesimally small perturbation y of the state x

D t .x/  y (1.4)

is governed by the linearized dynamics where D t .x/ denotes the Jacobian matrix of
the flow  t . For discrete systems this Jacobian can be computed using the recursion
scheme

Dx  nC1 .x/ D Dx g. n .x//  Dx  n .x/ (1.5)

with initial value Dx  0 .x/ D IM where IM denotes the M  M identity matrix.

For continuous systems (1.2) additional linearized ordinary differential equations
(ODEs)

d
Y D Dx f. t .x//  Y (1.6)
dt

have to be solved where  t .x/ is a solution of Eq. (1.2) with initial value x and Y
is a M  M matrix that is initialized as Y.0/ D IM . The solution Y.t/ provides the
Jacobian of the flow D t .x/ that describes the local dynamics along the trajectory
given by the temporal evolution  t .x/ of the initial state x. Since Eq. (1.6) is a linear
ODE its solutions consist of exponential functions and the Jacobian of the flow
D t .x/ maps a sphere of initial values close to x to an ellipsoid centered at  t .x/ as
illustrated in Fig. 1.1. This evolution of the tangent space dynamics can be analyzed
using a singular value decomposition (SVD) of the Jacobian of the flow D t .x/

D t .x/ D U  S  V tr (1.7)

1 u(1)
v(1) v (2) D (x) = U S V tr
2 u(2)
x (x)

Fig. 1.1 Temporal evolution of an infinitesimally small sphere in state space

4 U. Parlitz

where S D diag.1 ; : : : ; M / is a M  M diagonal matrix containing the singular

values 1  2  : : :  M  0 and U D .u.1/ ; : : : ; u.M/ / and V D .v.1/ ; : : : ; v.M/ /
are orthogonal matrices, represented by orthonormal column vectors u.i/ 2 RM and
v.i/ 2 RM , respectively. V tr is the transposed of V coinciding with the inverse V 1 D
V tr , because V is orthogonal. For the same reason U tr D U 1 and by multiplying
by V from the right we obtain D t .x/  V D U  S or

D t .x/v.m/ D m u.m/ .m D 1; : : : ; M/: (1.8)

The column vectors of the matrices V and U span the initial sphere and the ellipsoid,
as illustrated in Fig. 1.1, where the singular values m .t/ give the lengths of the
principal axes of the ellipsoid at time t. On average m .t/ increases or decreases
exponentially during the temporal evolution and the Lyapunov exponents m are the
mean logarithmic growth rates of the lengths of the principal axes

1
m D lim ln m .t/: (1.9)
t!1 t

The existence of the limit in Eq. (1.9) is guaranteed by the Theorem of Oseledec
[33] stating that the Oseledec matrix
 1
.x/ D lim D t .x/tr  D t .x/ 2t (1.10)
t!1

exists. For dissipative systems one set of exponents is associated with each attractor
and for almost all initial states x from each attractor .x/ takes the same value.
The logarithms of the eigenvalues m of this symmetric positive definite M  M
matrix are the Lyapunov exponents of the attractor or invariant set the initial state x
belongs to

m D ln m .m D 1; : : : ; M/: (1.11)

Using the SVD of the Jacobian matrix of the flow the Oseledec matrix for finite time
t can be written
  2t1  1
V  S  U tr  U  S  V tr D V  S2  V tr 2t (1.12)

1=t
with eigenvalues m . Taking the logarithm 1t ln m and performing the limit t !
1 we obtain the Lyapunov exponents (1.11). Unfortunately, this definition and
illustration of the Lyapunov exponents cannot be used directly for their numerical
computation, because the Jacobian matrix D t .x/ consists of elements that are expo-
nentially increasing or decreasing in time resulting in values beyond the numerical
resolution and representation of variables. To avoid these severe numerical problems
in 1979 Shimada and Nagashima [45] and in 1980 and Benettin et al. [3] suggested
algorithms that exploit the fact that the growth rate of k-dimensional volumes .k/
1 Estimating Lyapunov Exponents from Time Series 5

(in the M-dimensional state space) is given by the sum of the largest k Lyapunov
exponents

X
k
.k/ D m (1.13)
mD1

and the Lyapunov exponents can by computed from the volume growth rates as
1 D .1/ , 2 D .2/  1 , 3 D .3/  2  1 , etc. The volume growth
rates .k/ can be computed using a QR decomposition of the Jacobian of the flow
D t .x/. Let O.k/ D .o.1/ ; : : : ; o.k/ / be an orthogonal matrix whose column vectors
oj span a k-dimensional infinitesimal volume with k ranging from 1 to M. After
time t this volume is transformed by the Jacobian matrix into a parallelepiped
P.k/ .t/ D D t .x/  O.k/ . To computed the volume spanned by the column vectors
of P.k/ .t/ we perform a QR-decomposition of P.k/ .t/

P.k/ .t/ D D t .x/  O.k/ D Q.k/ .t/  R.k/ .t/ (1.14)

where Q.k/ .t/ is a matrix with k orthonormal columns and R.k/ .t/ is an upper
triangular matrix with non-negative diagonal elements. The volume V .k/ .t/ of P.k/ .t/
.k/
at time t is given by the product of the diagonal elements Rii .t/ of R.k/ .t/

.k/ .k/
Y
k
.k/
V .k/ .t/ D R11 .t/  : : :  Rkk .t/ D Rii .t/: (1.15)
iD1

The mean logarithmic growth rate of the k-dimensional volume is thus given by

1X
k
1 .k/
.k/ D lim ln V .k/ .t/ D lim ln Rii .t/: (1.16)
t!1 t t!1 t
iD1

Using this relation and Eq. (1.13) we can conclude that the first k Lyapunov
exponents 1 ; : : : ; k are given by

1 .k/
i D lim ln Rii .t/: (1.17)
t!1 t

If one would perform the QR-decomposition (1.14) of the Jacobian D t .x/ after a
very long period of time (to approximate the limit t ! 1) then one would be faced
with the same numerical problems that were mentioned above in the context of
the Oseledec matrix. The advantage of the volume approach via QR-decomposition
is, however, that this decomposition can be computed recursively for small time
intervals avoiding any numerical over or underflow. To exploit this feature the period
of time 0; t is divided into N time intervals of length T D t=N and the Jacobian
matrices D T . tn .x// are computed at times tn D nT (n D 0; : : : ; N  1) along
6 U. Parlitz

the orbit. Employing the chain rule the Jacobian matrix D t .x/ can be written as a
product of Jacobian matrices D T . tn .x//

Y
N1
D t .x/ D D T . tN1 .x//  : : :  D T . t0 .x// D D T . tn .x//: (1.18)
nD0

Using QR-decompositions

QO .k/ .tnC1 /  RO .k/ .tnC1 / D D T . tn .x//  QO .k/ .tn / (1.19)

the full period of time 0; t used for averaging the local expansion rates can
be decomposed into a sequence of relatively short intervals 0; T with Jacobian
matrices D T . tn .x// that are not suffering from numerical difficulties. Applying
the QR-decompositions (1.19) recursively we obtain a scheme for computing the
QR-decomposition of the Jacobian matrix of the full time step

D t .x/  O.k/ D D T . tN1 .x//  : : :  D T . t0 .x//  O.k/

D D T . tN1 .x//  : : :  D T . t1 .x//  QO .k/ .t1 /  RO .k/ .t1 /
D D T . tN1 .x//  : : :  D T . t2 .x//  QO .k/ .t2 /  RO .k/ .t2 /  RO .k/ .t1 /
::
:
D D T . tN1 .x//  QO .k/ .tN1 /  RO .k/ .tN1 /  : : :  RO .k/ .t1 /
D QO .k/ .tN /  RO .k/ .tN /  : : :  RO .k/ .t1 /

which provides Q.k/ .t/ D QO .k/ .tN / and the required matrix R.k/ .t/ as a product

R.k/ .t/ D RO .k/ .tN /  : : :  RO .k/ .t1 /: (1.20)

For the diagonal elements of the upper triangular matrices holds the relation

.k/
Y
N
.k/
Rii .t/ D RO ii .tn / (1.21)
nD1

.k/
and substituting Rii .t/ in Eq. (1.17) (with t D NT) we obtain the following
expression for the ith Lyapunov exponent (with i  k  M)

1 X O .k/
N
i D lim ln Rii .tn /: (1.22)
N!1 NT
nD1

Using this approach the computation of all Lyapunov exponents of a given

dynamical system became a standard procedure [12, 19, 46, 51] providing the full
1 Estimating Lyapunov Exponents from Time Series 7

set of Lyapunov exponents constituting the Lyapunov spectrum which is an ordered

set of real numbers f1 ; 2 ; : : : ; m g. If the system undergoes aperiodic oscillations
(after transients decayed) and if the largest Lyapunov exponent 1 is positive, then
the corresponding attractor is said to be chaotic and to show sensitive dependence
on initial conditions. If more than one Lyapunov exponent is positive the underlying
dynamics is called hyper chaotic.
The (ordered) spectrum 1  2  : : :  m can be used to compute the
KaplanYorke dimension (also called Lyapunov dimension) [34]
Pk
iD1 i
DKY D k C (1.23)
jkC1 j

where k is the maximum integer such that the sum of the k largest exponents is
still non-negative. DKY is an upper bound for the information dimension of the
underlying attractor.

1.3 Estimating Lyapunov Exponents from Time Series

All methods for computing Lyapunov exponents are based on state space reconstruc-
tion from some observed (univariate) time series [11, 42, 43, 49]. For reconstructing
the dynamics most often delay coordinates are used due to their efficacy and
robustness.
To reconstruct the multi-dimensional dynamics from an observed (univariate)
time series fsn g sampled at times tn D nt we use delay coordinates providing the
N  D trajectory matrix

X D .x1 ; x2 ; : : : ; xN /tr (1.24)

where each row is a reconstructed state vector1

xn D .sn ; snCL ; : : : ; snC.D1/L / (1.25)

at time n (with lag L and dimension D). From a time series fsn g of length Nd a
total number of N D Nd  .D  1/L states can be reconstructed. To achieve useful
(nondistorted) reconstructions the time window length .D  1/L of the delay vector
should cover typical time scales of the dynamics like natural periods or the first zero
or minimum of the autocorrelation function or the (auto) mutual information [1, 26].
Since Lyapunov exponents are invariant with respect to diffeomorphic changes of
the coordinate system the Lyapunov exponents estimated for the reconstructed flow

1
We use forward delay coordinates here. Delay reconstruction backward in time provides
equivalent results.
8 U. Parlitz

will coincide with those of the original system. Technically, different approaches
exist for computing Lyapunov exponents from embedded time series. Jacobian-
based methods employ the standard algorithm outlined in Sect. 1.2 except for the
computation of the Jacobian matrix D t .x/ which is now based on approximations
of the flow in the reconstructed state space. This class of methods will be briefly
presented in Sect. 1.3.1. In particular with noisy data reliable estimation of the
Jacobian matrix may be a delicate task. This is one of the reasons why several
authors proposed methods for estimating the largest Lyapunov exponent directly
from diverging trajectories in reconstructed state space. Such direct methods will be
discussed in detail in Sect. 1.3.2 and will be illustrated and evaluated in Sect. 1.5.
They do not require Jacobian matrices but are mostly used to compute the largest
Lyapunov exponent, only. A major advantage of direct methods, however, is the fact
that they provide direct visual feedback to the user whether the available time series
really exhibits exponential divergence on small scales. Therefore, we shall focus on
this class of methods in the following.

1.3.1 Jacobian-Based Methods

With Jacobian methods, first a model is fitted to the data and then the Jacobian
matrices of the model equations are used to compute the Lyapunov exponents using
standard algorithms (see Sect. 1.2) which have been developed for the case when the
equations of the dynamical system are known [3, 12, 19, 45]. In this context usually
local linear approximations are used for modeling the flow in reconstructed state
space [14, 22, 29, 36, 40, 47, 48, 53, 54]. An investigation of the data requirements
for Jacobian-based methods may be found in [13, 15]. Technical details and more
information about the implementation of Jacobian-based methods are given in the
Appendix.
To employ the standard algorithm for computing Lyapunov exponents (Sect. 1.2)
also for time series analysis the Jacobian matrices along the orbit in reconstruction
space are required and have to be estimated from the temporal evolution of
reconstructed states. Here two major challenges occur:
(a) The Jacobian matrices (derivatives) have to be estimated using reconstructed
states that are scattered along the unstable direction(s) of the attractor but not
in transversal directions (governed by contracting dynamics). This may result
in ill-posed estimation problems and is a major obstacle for estimating negative
Lyapunov exponents. Furthermore, the estimation problem is often even more
delicate because we aim at approximating (partial) derivatives (the elements of
the Jacobian matrix) from typically noisy data where estimating derivatives is a
notoriously difficult problem.
(b) To properly unfold the attractor and the dynamics in reconstruction space
the embedding dimension D has in general to be larger than the dimension
of the original state space M (see Sect. 1.3). Therefore, a straightforward
1 Estimating Lyapunov Exponents from Time Series 9

computation of Lyapunov exponents using (estimated) D  D Jacobian matrices

in reconstruction space and QR-decomposition (see Sect. 1.2) will provide D
Lyapunov exponents, although the underlying M-dimensional system possesses
M < D exponents, only. The additional D  M Lyapunov exponents are called
parasitic or spurious exponents and they have to be identified (or avoided),
because their values are not related to the dynamics to be characterized.
Spurious Lyapunov exponents can take any values [8], depending on details
of the approximation scheme used to estimate the Jacobian matrices, the local
curvature of the reconstructed attractor, and perturbations of the time series (e.g.,
noise). Therefore, without taking precautions spurious Lyapunov exponents can
occur between true exponents and may spoil in this sense the observed spectrum
(resulting in false conclusion about the number of positive exponents or the Kaplan
Yorke dimension, for example). To cope with this problem many authors presented
different approaches for avoiding spurious Lyapunov exponents or for reliably
detecting them [44].
To identify spurious Lyapunov exponents one can estimate the local thickness of
the attractor along the directions associated with the different Lyapunov exponents
[6, 7] or compare the exponents obtained with those computed for the time reversed
series [35, 36], because spurious exponents correspond to directions where the
attractor is very thin and because in general they do not change their signs upon
time reversal (in contrast to the true exponents). The latter method, however, works
only for data of very high quality that enable also a correct estimation of negative
Lyapunov exponents which in most practical situations is not the case. Furthermore,
in some cases also spurious Lyapunov exponents may change signs and can then
not be distinguished from true exponents. Another method for identifying spurious
Lyapunov exponents employing covariant Lyapunov vectors been suggested in
[27, 52].
Spurious Lyapunov exponents can be avoided by globally unfolding the dynam-
ics in a D-dimensional reconstruction space and locally approximating the (tangent
space) dynamics in a lower dimensional d-dimensional space (with d  M). This
can be done using two different delay coordinates where the set of indices of
neighbouring points of a reference point is identified using a D-dimensional delay
reconstruction and then these indices are used to reconstruct states representing
proper neighbours in a d-dimensional delay reconstruction (with d < D) which
is used for subsequent modeling of the dynamics (flow and its Jacobian matrices)
[6, 7, 14]. To cover relevant time scales it is recommended [14] to use for both delay
reconstructions different lags LD and Ld so that the delay vectors span the same or
similar windows in time (i.e., .D  1/LD  .d  1/Ld ). An alternative approach
for evaluating the dynamics in a lower dimensional space employs local projections
into d-dimensional subspaces of the D-dimensional delay embedding space given
by singular value decompositions of local trajectory matrices [10, 48].
For evaluating the uncertainty in Lyapunov exponent computations from time
series employing Jacobian based algorithms bootstrapping methods have been
suggested [28].
10 U. Parlitz

1.3.2 Direct Methods

There are (slightly) different ways to implement a direct method for estimating
the largest Lyapunov exponent and they all rely on the fact that almost all tangent
vectors (or perturbations) converge to the subspace spanned by the first Lyapunov
vector(s) with an asymptotic growth rate given by the largest Lyapunov exponent
1 (see Sect. 1.5.1). In practice, however, from a time series of finite length only a
finite number of reconstructed states is available with a finite lower bound for their
mutual distances. If the nearest neighbour xm.n/ of a reference point xn is chosen
from the set of reconstructed states the trajectory segments emerging from both
states will (on average) diverge exponentially until the distance kxm.n/Ck  xnCk k
exceeds a certain threshold and ceases to grow but oscillates bounded by the size
of the attractor. For direct methods it is crucial that the reorientation towards the
most expanding direction takes place and is finished before the distance between the
states saturates. Then the period of exponential growth characterised by the largest
Lyapunov exponent can be detected and estimated for some period of time as a linear
segment in a suitable semi-logarithmic plot. This feature is illustrated in Fig. 1.2a
showing the average of the logarithms of distances of neighbouring trajectories vs.
time on a semi-logarithmic scale. In phase I the difference vector between states
from both trajectories converges towards the most expanding direction. Then in
phase II exponential divergence results in a linear segment until in phase III states
from both trajectory segments are so far away from each other that nonlinear folding
occurs and the mean distance converges to a constant value (which is related to the
diameter of the attractor).
Different implementations of the direct approach have been suggested in the past
25 years [18, 25, 30, 38, 41] that are based on the following considerations.
Let x.m.n// be a neighbour of the reference state x.n/ (with respect to the
Euclidean norm or any other norm) and let both states be temporally separated

jm.n/  nj > w (1.26)

Fig. 1.2 (a) Sketch showing the mean logarithmic distance of neighbouring states on different
trajectory segments vs. time. (b) Illustration motivating the exclusion of temporal neighbours
(Theiler window)
1 Estimating Lyapunov Exponents from Time Series 11

where w is a characteristic time scale (e.g., a mean period) of the time series.
The temporal separation (also called Theiler window w [50]) is necessary to make
sure that this pair of neighbouring states can be considered as initial conditions of
different trajectory segments and the largest Lyapunov exponent can be estimated
from the mean rate of separation of states along these two orbits. Figure 1.2b shows
an illustration of a case where a Theiler window of w D 3 would be necessary to
exclude neighbours of the state marked by a big dot to avoid temporal neighbours
on the same trajectory segment (small dots preceding and succeeding the reference
state within the circle, indicating the search radius).
We shall quantify the separation of states by the distance

d.m.n/; n; k/ D kx.m.n/ C k/  x.n C k/k (1.27)

of the neighbouring states after k time steps (i.e. a period of time T D kt). Most
often [18, 30, 38, 41] the Euclidean norm

dE .m.n/; n; k/ D kx.m.n/ C k/  x.n C k/k2 (1.28)

is used to define this distance, although Kantz [25] pointed out that it is sufficient to
consider the difference

dL .m.n/; n; k/ D jx.m.n/ C .D  1/L C k/  x.n C .D  1/L C k/j (1.29)

of the last components of both reconstructed states, because these projections also
grow exponentially with the largest Lyapunov exponent. Here L denotes again
the time lag used for delay reconstruction. With the same argument, one can also
consider the difference of the first component

dF .m.n/; n; k/ D jx.m.n/ C k/  x.n C k/j (1.30)

and in the following we shall compare all three choices. Within the linear
approximation (very small d.m.n/; n; k/) the temporal evolution of the distance
d.m.n/; n; k/ is given by

O
d.m.n/; n; k/  d.m.n/; n; 0/ e1 .n/kt (1.31)

where d.m.n/; n; 0/ stands for the initial separation of both orbits and O 1 .n/ denotes
the (largest) local expansion rate of orbits starting at x.n/. Taking the logarithm we
obtain
1
O 1 .n/  ln.d.m.n/; n; k//  ln.d.m.n/; n; 0// (1.32)
kt
 
1 d.m.n/; n; k/
D ln :
kt d.m.n/; n; 0/
12 U. Parlitz

Here and in the following expansion rates and Lyapunov exponents are computed
using the natural logarithm ln./.
Since expansion rates vary on the attractor we have to average along the available
trajectory by choosing for each reference state x.n/ some neighbouring states
fx.m.n// W m.n/ 2 Un g where Un defines the chosen neighbourhood of x.n/ that
can be of fixed mass (a fixed number K of nearest neighbours of x.n/) or of fixed
size (all points with distance smaller than a given bound ). Often a fixed mass with
K D 1 (i.e., using only the nearest neighbour) is used [18, 38, 41], but a fixed size
may in some cases be more appropriate to avoid mixing of scales [25, 30]. In the
following jUn j denotes the number of neighbours of x.n/.
Furthermore, it may be appropriate to use not all available reconstructed states
x.n/ (n D 1; : : : ; N) as reference points but only a subset R consisting of Nr D
jRj points. This speeds up computations and may even result in better results if R
contains only those reconstructed states that possess very close neighbours (where
d.m.n/; n; 0/ is very small). This issue will be discussed and demonstrated in the
results section.
With averaged logarithmic distances

1 X 1 X
E.k/ D ln.d.m; n; k// (1.33)
Nr jUn j
n2R m2Un

and
 
1 X 1 X d.m; n; k/
S.k/ D ln D E.k/  E.0/ (1.34)
Nr jUn j d.m; n; 0/
n2R m2Un

and the local expansion rates (1.32) the averaged growth rate can be expressed as
 
1 XO 1 1 X 1 X d.m; n; k/
N 1 D 1 .n/  ln (1.35)
Nr Nr kt jUn j d.m; n; 0/
n2R n2R m2Un

1 1
D S.k/ D E.k/  E.0/ (1.36)
kt kt
providing the relations

S.k/  ktN 1 (1.37)

and

E.k/  ktN 1 C E.0/: (1.38)

Here E.k/ stands for EE .k/, EF .k/, or EL .k/ depending on the distance measure dE ,
dF , or dL used when computing E in Eq. (1.33).
1 Estimating Lyapunov Exponents from Time Series 13

In 1987 Sato et al. [41] suggested to estimated the largest Lyapunov exponent
by the slope of a linear segment of the graph obtained when plotting S.k/ vs. kt.
The same approach was suggested later in 1993 by Gao and Zheng [18]. The same
year Rosenstein et al. [38] recommended to avoid the normalization by the initial
distance d.m.n/; n; 0/ in Eqs. (1.32) and (1.34) and to plot E.k/ vs. kt. As can
be seen from Eqs. (1.37) and (1.38) both procedures are equivalent, because both
graphs differ only be a constant shift E.0/. Instead of estimating the slope in S.k/
vs. kt Sato et al. [41] and Kurths and Herzel [30] independently suggested in 1987
to consider

E.k C l/  E.k/ S.k C l/  S.k/

N 1  D (1.39)
lt lt

and to identify a plateau in the graph E.k Cl/E.k/ vs. k (that should occur for the
same range of k values where the linear segment occurs with the previous methods).
This is basically a finite differences approximation of the slope of the graph E.k/ vs.
kt. To obtain best results the time interval lt should be large but .l C k/t must
not exceed the linear scaling region(s) where distances grow exponentially (and this
range is in general not known a priori).
In 1985 Wolf et al. [51] suggested a method to estimate the largest Lyapunov
exponent(s) which avoids the saturation of mutual distances of reference states
and local neighbours due to nonlinear folding. The main idea is to monitor the
distance between the reference orbit and the neighbouring orbit and to replace
(once a threshold is exceeded) the neighbouring state by another neighbouring
state that is closer to the reference orbit and which lies on or near the line from
the current reference state to the last point of the previous neighbouring orbit
in order to preserve the (local) direction corresponding to the largest Lyapunov
exponent. Criteria for the replacement threshold and other details of the algorithm
are given in [51], including a FORTRAN program. In principle, it is possible to
use this strategy also for computing the second largest Lyapunov exponent [51],
but this turns out to be quite difficult. When applied to stochastic time series
the Wolf algorithm yields inconclusive results and may provide any value for
the Lyapunov exponent depending on computational parameters and pre-filtering
[9]. Due to its robustness the Wolf-algorithm is often used for the analysis of
experimental data (see, for example, [16, 17]). A drawback of this method (similar
to Jacobian based algorithms) is the fact that the user has no possibility to check
whether exponential growth underlies the estimated values or not. Even if the
amount of data available or the type and quality of the time series would not
be sufficient to quantify exponential divergence the algorithm would provide a
number that might be misinterpreted as the largest Lyapunov exponent of the
underlying process. Therefore, we do not consider this method in more detail in
the following.
14 U. Parlitz

1.4 Example Time Series

To illustrate and evaluate the direct method for estimating the largest Lyapunov
exponent, time series generated by four different chaotic dynamical systems are
used that will be introduced in the following subsections.

1.4.1 The Hnon Map

The first system is the Hnon map [21]

x1 .n C 1/ D 1  ax21 .n/ C bx2 .n/ (1.40a)

x2 .n C 1/ D x1 .n/ (1.40b)

with parameters a D 1:4 and b D 0:3. The Lyapunov exponents of this system
are 1 D 0:420 and 2 D 1:624 (computed with the natural logarithm ln./,
note that for the Hnon map 1 C 2 D ln.b/ D 1:204). In the following
we shall assume that a x1 time series of length Nd D 4096 is given.2 A special
feature of the Hnon map is that its original coordinates .x1 .n/; x2 .n// coincide with
2-dimensional delay coordinates .x1 .n/; x1 .n  1// D .x1 .n/; x2 .n//. Figure 1.3a
shows the Hnon attractor reconstructed from a clean fx1 .n/g time series and in
Fig. 1.3b a reconstruction is given based on a time series with additive measurement
noise of signal-to-noise ration (SNR) of 30 dB (generated by adding normally
distributed random numbers).

(a) 1.5 (b) 1.5

1 1
x1(n1) = x2(n)

x1(n1) = x2(n)

0.5 0.5

0 0

0.5 0.5

1 1

1.5 1.5
2 1 0 1 2 2 1 0 1 2
x (n) x (n)
1 1

Fig. 1.3 Attractor of the Hnon map (1.40) (a) without noise and (b) with noise (SNR D 30 dB)

2
Since x2 .n C 1/ D x1 .n/ any x2 time series will give the same results.
1 Estimating Lyapunov Exponents from Time Series 15

1.4.2 The Folded-Towel Map

The second system is the folded-towel map introduced in 1979 by Rssler [39]

x.n C 1/ D 3:8x.n/.1  x.n//  0:05.y.n/ C 0:35/.1  2z.n// (1.41a)

y.n C 1/ D 0:1.y.n/ C 0:35/.1  2z.n//  1.1  1:9x.n// (1.41b)
z.n C 1/ D 3:78z.n/.1  z.n// C 0:2y.n/ (1.41c)

which generates the chaotic attractor shown in Fig. 1.4. The folded-towel map has
two positive Lyapunov exponents 1 D 0:427, 2 D 0:378, and a negative exponent
3 D 3:30. The KaplanYorke dimension of this attractor equals DKY D 2:24.
Figure 1.5 shows delay reconstructions based on x, y, and z time series of length
Nd D 65;536 (i.e. 64k). In the first row (Fig. 1.5ac) clean data are used while
for the reconstructions shown in the second row (Fig. 1.5df) noisy data (64k)
with signal-to-noise ratio (SNR) of 30 dB are used that were obtained by adding
normally distributed random numbers to the clean data shown in the first row. These
noisy time series will be used to evaluate the robustness of methods for estimating
Lyapunov exponents.

0.8

0.6
z(n)

0.4

0.2

0
0.1
1
0
0.5
0.1
y(n) 0 x(n)
Fig. 1.4 Hyperchaotic attractor of the folded-towel map (1.41), original coordinates .x; y; z/
(length Nd D 65;536)
16 U. Parlitz

(a) (b) (c)

1 1
0.1
x(n+2)

y(n+2)

z(n+2)
0.5 0 0.5

0.1
0 0
1 1
1 0.1 1
0.5 0 0.1 0.5
0.5 0 0.5
0.1 0.1
x(n+1) 0 0 y(n+1) z(n+1) 0 0
x(n) y(n) z(n)

(d) (e) (f)

1 1
0.1
x(n+2)

y(n+2)

z(n+2)
0.5 0 0.5

0.1
0 0
1 1
1 0.1 1
0.5 0 0.1 0.5
0.5 0 0.5
0.1 0.1
x(n+1) 0 0 y(n+1) z(n+1) 0 0
x(n) y(n) z(n)

Fig. 1.5 Delay reconstructions of the attractor of the folded-towel map (1.41). (a)(c) Without
and (d)(f) with additive (measurement) noise of SNR 30 dB. Reconstruction from (a), (d) fx.n/g
time series, (b), (e) fy.n/g time series, and (c), (f) fz.n/g time series

1.4.3 Lorenz-63 System

As an example of a low dimensional continuous time system we use the Lorenz-63

system [31] given by the following set of ordinary differential equations (ODEs)

xP 1 D .x2  x1 / (1.42a)
xP 2 D x1 .R  x3 /  x2 (1.42b)
xP 3 D x1 x2  bx3 : (1.42c)

With parameter values  D 16, R D 45:92, and b D 4 this systems generates a

chaotic attractor with Lyapunov exponents 1 D 1:51, 2 D 0, and 3 D 22:5.

1.4.4 Lorenz-96 System

As an example of a continuous time system exhibiting complex dynamics we shall

employ a 6-dimensional Lorenz-96 system [32] describing a ring of 1-dimensional
dynamical elements. The differential equations for the model read

dxi .t/
D xi1 .t/.xiC1 .t/  xi2 .t//  xi .t/ C f (1.43)
dt
1 Estimating Lyapunov Exponents from Time Series 17

2 15

1
10
0

1 5

1
x
2
0
3

4 5

5
10
0 2000 4000 6000 8000 10000 0 5 10 15 20
t t

Fig. 1.6 (a) Convergence of Lyapunov exponents of the 6-dimensional Lorenz-96 system (1.43)
generated with parameter value f D 10. (b) Typical oscillation

with i D 1; 2; : : : ; 6, x1 .t/ D x5 .t/; x0 .t/ D x6 .t/, and x7 .t/ D x1 .t/. With a
forcing parameter f D 10 the system generates a chaotic attractor characterized by
a Lyapunov spectrum f1:249; 0:000; 0:098; 0:853; 1:629; 4:670g and a result-
ing KaplanYorke dimension of DKY D 4:18. Figure 1.6a shows the convergence of
the six Lyapunov exponents upon their computation using the full model equations
(1.43) and in Fig. 1.6b, a typical time series of the Lorenz-96 system is plotted.

1.5 Estimation of Largest Lyapunov Exponents Using Direct

Methods

1.5.1 Convergence of Small Perturbations

As illustrated in Fig. 1.2a any (random) perturbation first undergoes a transient phase
I and converges to the direction of the Lyapunov vector(s) corresponding to the
largest Lyapunov exponent. Then, in phase II, it grows linearly (on a logarithmic
scale) until the perturbation exceeds the linear range (phase III). In the following
we shall study this convergence process for the four example systems which were
introduced in the previous section. The asymptotic average stretching of almost any
initial perturbation (tangent vector) z.0/ is given by the largest Lyapunov exponent
1 . Using the tangent space basis fv.1/ ; : : : ; v.m/ g provided by the SVD (1.7) the
18 U. Parlitz

initial tangent vector can be written as

X
M
z.0/ D ci v.m/ D V  c (1.44)
mD1

where c D .c1 ; : : : ; cM / denotes the vector of projection coefficients. Its temporal

evolution is thus given by

X
M
z.t/ D D t .x/  z.0/ D U  S  V tr  z.0/ D U  S  c D cm m u.m/ : (1.45)
mD1

If we approximate the singular values m by em t we obtain for (the square of) the
Euclidean norm of z.t/

X
M
Z 2 .t/ D kz.t/k2 D c2m e2m t (1.46)
mD1

and this yields

!1=t
1 1 1 XM
lim ln .Z.t// D lim ln .kz.t/k/ D lim ln c2m e2m t D 1
t!1 t t!1 t t!1 2
mD1
(1.47)

because the term em t with the largest m dominates the sum as time t goes to infinity.
The speed of convergence depends on the full Lyapunov spectrum. Figure 1.6 shows
ln.Z.t// D ln.kz.t/k/ (as defined in Eq. (1.47)) vs. t for the Hnon map (1.40), the
folded towel map (1.41), the Lorenz-63 system (1.42), and the Lorenz-96 system
(1.43). While the local slopes of the Hnon map and the folded towel map reach the
value of the largest Lyapunov exponent after a period of time of about t  1 the
random initial tangent vectors z.0/ of the Lorenz-96 system need about twice the
time and converge to 1 only after t  2 (Fig. 1.7).

1.5.2 Hnon Map

Figure 1.8 shows an application of the direct estimation method to a fx1 .n/g time
series of the Hnon map (1.40). The time series has a length of N D 4096 samples,
the lag equals L D 1, and different reconstruction dimensions D D 2, D D 4,
and D D 6 are used. Figure 1.8a shows EE .k/ vs. kt where t D 1 denotes the
1 Estimating Lyapunov Exponents from Time Series 19

(a) 3.5 (b) 2

Henon map
3 folded towel map 1.5
Lorenz63
2.5
Lorenz96
1

ln(Z)/ t
2
ln(Z)

1.5 0.5

1
0
Henon map
0.5
folded towel map
0.5 Lorenz63
0
Lorenz96
0.5 1
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
t t

Fig. 1.7 (a) Average values of ln.Z.t// D ln.kz.t/k/ vs. t (see Eq. (1.47)) for the Hnon map
(red), the folded towel map (blue, dotted line), The Lorenz-63 system (green, dashed-dotted line)
and the Lorenz-96 system (dashed line). The curves are computed by averaging 2000 realizations
with randomly chosen initial vectors z.0/ with kz.0/k D 1. (b) Local slopes of curves shown in
(a) indicating the convergence to the value of the corresponding largest Lyapunov exponent (given
by horizontal dashed lines)

(a) (c) (e)

0 0 0

2 2 2
E

L
E
E
E

4 4 4
D=2 D=2 D=2
D=4 D=4 D=4
6 6 6
D=6 D=6 D=6

0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
kt kt kt

(b) 0.5 (d) 0.5 (f) 0.5

0.4 0.4 0.4

EE/ t

E / t

EL/ t

0.3 0.3 0.3

F

0.2 0.2 0.2

D=2 D=2 D=2
0.1 D=4 0.1 D=4 0.1 D=4
D=6 D=6 D=6
0 0 0
5 10 15 5 10 15 5 10 15
kt kt kt

Fig. 1.8 Direct estimation of the largest Lyapunov exponent from a Hnon time series for different
reconstruction dimensions D D 2, D D 4, D D 6 using a lag of L D 1. The diagrams (a), (c), and
(e) show EE , EL and EF vs. kt with t D 1 for different measures of distance (1.28), (1.29), and
(1.30). In (b), (d), and (f) the corresponding slopes E=t vs. kt (Eq. (1.48)) are shown. The
dashed lines indicate the true result 1 D 0:42
20 U. Parlitz

sampling time and EE (1.33) is computed with the Euclidean distance dE (1.28). In
Fig. 1.8b the slope

dEE EE .k C 1/  EE .k/ EE

..k C 0:5/t/  D (1.48)
dt t t
is plotted. Figure 1.8c, d and e, f show the corresponding diagrams obtained with
the distance measures dF (1.30) and dL (1.29), respectively. In all diagrams three
phases occur (see also Fig. 1.2):
First the difference vector x.m.n/ C k/  x.n C k/ converges for increasing k to
the subspace spanned by the first Lyapunov vector(s). The slope increases.
Then the difference vector experiences the expansion rate given by the largest
Lyapunov exponent. The slope is constant indicating exponential divergence.
Finally, the lengths of the difference vector exceeds the range of the linearized
dynamics and its length saturates due to nonlinear folding in the (reconstructed)
state space. The slope decreases.
The lengths of the linear scaling regions in Fig. 1.8a, c, e and of the plateaus
in Fig. 1.8b, d, f shrink with increasing embedding dimension. They also shrink, if
the length of the time series is reduced or the number of nearest neighbours K is
increased.
The results shown in Fig. 1.8 are computed by using each reconstructed state
as a reference point. Reliable estimates may be obtained, however, already with a
subset of reference points which reduces computation time almost linearly. This
subset can be randomly selected from all reconstructed states or it can be chosen
to include only those reconstructed states that possess the nearest neighbours. The
latter choice has the advantage that more steps of the diverging neighbouring
trajectory segments are governed by the linearised flow and exhibit exponential
growths resulting in longer scaling regions. Figure 1.9 shows results based on
those 25 % of the total number N of reference points that possess the closest
neighbours (i.e., where the chosen distance measure dE , dF , or dL (see Eqs. (1.28)
(1.30)) takes the smallest values). The scaling regions are extended compared
to Fig. 1.8 but the local slopes plotted in Fig. 1.9c, d, f show more statistical
fluctuations due to the smaller number of reference points (for D D 6 we
have Nr D 1018 reference points in Fig. 1.9e, f compared to Nr D 4070 in
Fig. 1.8e, f).
To illustrate the impact of (additive) measurement noise Fig. 1.10 shows results
obtained with a noisy Hnon time series (compare Fig. 1.2b). As can be seen in
all diagrams noise leads to shorter scaling intervals and a bias towards smaller
values underestimating the largest Lyapunov exponent. Decreasing the number Nr of
reference points (with nearest neighbours) reduces the bias but increases statistical
fluctuations.
1 Estimating Lyapunov Exponents from Time Series 21

(a) 0 (c) 0 (e) 0

2 2 2
EE

L
4 4 4

E
E
6 D=2 6 D=2 6 D=2
D=4 D=4 D=4
D=6 D=6 D=6
8 8 8
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
kt kt kt

(b) 0.5 (d) 0.5 (f) 0.5

0.4 0.4 0.4

EE/ t

E / t

EL/ t
0.3 0.3 0.3
F

0.2 0.2 0.2

D=2 D=2 D=2
0.1 D=4 0.1 D=4 0.1 D=4
D=6 D=6 D=6
0 0 0
5 10 15 5 10 15 5 10 15
kt kt kt

Fig. 1.9 Direct estimation of the largest Lyapunov exponent from a Hnon time series for different
reconstruction dimensions D D 2, D D 4, D D 6 using a lag of L D 1. The diagrams (a), (c),
and (e) show EE , EL and EF vs. kt with t D 1 for different measures of distance (1.28), (1.29),
and (1.30). In (b), (d), and (f) the corresponding slopes E=t vs. kt (Eq. (1.48)) are shown.
The dashed lines indicate the true result 1 D 0:42. In contrast to Fig. 1.8 only those 25 % of the
reconstructed states with closest neighbours have been used as reference points

1.5.3 Folded Towel Map

To address the question whether the direct methods also work with hyper-chaotic
dynamics we shall now analyze time series generated by the folded-towel map
(1.41). Figure 1.11 shows results obtained from a fx.n/g time series of length
Nd D 65;536 using all N reconstructed states as reference points. As can be seen no
linear scaling region exists, because this time series provides poor reconstructions
of the underlying attractor (compare the reconstruction shown in Fig. 1.4a). Results
can be improved by using a longer time series and only those reference points
with very close neighbours. Alternatively, one may consider reconstructions based
on a fy.n/g time series which provide better unfolding of the chaotic attractor
(compare Fig. 1.4b). Figure 1.12 shows results computed using a fy.n/g time
series from the folded-towel map (1.41) with length Nd D 65;536, where only
10 % of the reconstructed states (with closest neighbours) are used for estimating
exponential divergence. As can be seen the fy.n/g time series is more suited for
estimating the largest Lyapunov exponent of the folded towel map and exhibits for
reconstruction dimensions D D 4 and D D 6 the expected scaling behaviour.
For D D 2 no clear scaling occurs and results differ significantly from those
obtained with D D 4 and D D 6, because 2-dimensional delay coordinates are not
22 U. Parlitz

(a) 1 (c) 1 (e) 1

0 0 0
1 1 1
2 2 2
EE

L
E
E
3 3 3
4 D=2 4 D=2 4 D=2
D=4 D=4 D=4
5 5 5
D=6 D=6 D=6
6 6 6
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
kt kt kt

(b) 0.5 (d) 0.5 (f) 0.5

0.4 0.4 0.4

EE/ t

E / t

EL/ t
0.3 0.3 0.3
F

0.2 0.2 0.2

D=2 D=2 D=2
0.1 D=4 0.1 D=4 0.1 D=4
D=6 D=6 D=6
0 0 0
5 10 15 5 10 15 5 10 15
kt kt kt

Fig. 1.10 Direct estimation of the largest Lyapunov exponent from a noisy Hnon time series
(SNR 30 dB) for different reconstruction dimensions D D 2, D D 4, D D 6 using a lag of L D 1.
The diagrams (a), (c), and (e) show EE , EL and EF vs. kt with t D 1 for different measures
of distance (1.28), (1.29), and (1.30). In (b), (d), and (f) the corresponding slopes E=t vs. kt
(Eq. (1.48)) are shown. All reconstructed states are used as reference points and the dashed lines
indicate the true result 1 D 0:42

(a) 0 (c) 0 (e) 0

2 2 2
E

4 4 4
E
E
E

D=2 D=2 D=2

6 6 6
D=4 D=4 D=4
D=6 D=6 D=6
8 8 8
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
kt kt kt

(b) (d) (f)

0.6 0.6 0.6
EE/ t

E / t

EL/ t

0.4 0.4 0.4

F

0.2 0.2 0.2

D=2 D=2 D=2
D=4 D=4 D=4
0 D=6 0 D=6 0 D=6

5 10 15 5 10 15 5 10 15
kt kt kt

Fig. 1.11 Direct estimation of the largest Lyapunov exponent from a fx.n/g time series of the
folded-towel map of length Nd D 65;536 for different embedding dimensions D D 2, D D 4,
D D 6 using a lag of L D 1. The diagrams (a), (c), and (e) show E vs. kt with t D 1 for the
Euclidean norm. In (b), (d), and (f) the corresponding slopes E=t vs. k (Eq. (1.48)) are shown.
The dashed lines indicate the true result 1 D 0:43
1 Estimating Lyapunov Exponents from Time Series 23

(a) 2 (c) 2 (e) 2

4 4 4
EE

L
E
E
6 6 6
D=2 D=2 D=2
8 D=4 8 D=4 8 D=4
D=6 D=6 D=6

0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
kt kt kt
(b) (d) (f)
0.6 0.6 0.6
EE/ t

E / t

EL/ t
0.4 0.4 0.4
F

0.2 0.2 0.2

D=2 D=2 D=2
D=4 D=4 D=4
0 D=6 0 D=6 0 D=6

5 10 15 5 10 15 5 10 15
kt kt kt

Fig. 1.12 Direct estimation of the largest Lyapunov exponent from a fy.n/g time series of the
folded-towel map of length Nd D 65;536 for different embedding dimensions D D 2, D D 4,
D D 6 using a lag of L D 1. The diagrams (a), (c), and (e) show E vs. kt with t D 1 for the
Euclidean norm. In (b), (d), and (f) the corresponding slopes E=t vs. k (Eq. (1.48)) are shown.
The dashed lines indicate the true result 1 D 0:43. Only those 10 % of the reconstructed states
possessing the most nearest neighbours are used as reference points for estimating exponential
divergence (Nr D 6551 for D D 6)

sufficient for reconstructing this chaotic attractor (with KaplanYorke dimension

DKY D 2:24).
Figure 1.13 shows results obtained with a noisy fy.n/g time series (SNR
30 dB) generated by the folded-towel map (compare Fig. 1.4e) with reconstruction
dimension D D 4, D D 6, and D D 8 and 10 % reference points. Scaling intervals
are barely visible due to the added measurement noise.

1.5.4 Lorenz-63

We shall now use as data source the Lorenz-63 system which is an example of a low
dimensional continuous system exhibiting deterministic chaos. Figure 1.14 shows
results for a x1 time series of length Nd D 65;536 sampled with t D 0:025 for
reconstruction dimensions D D 4, D D 12, and D D 21 using a delay of L D 1.
The resulting time windows .D  1/L covered by the delay vectors are 3, 11, and
20, respectively, where the latter corresponds to a typical oscillation period of the
Lorenz-63 system. Here the sampling time t D 0:025 is much smaller compared to
24 U. Parlitz

(a) (c) 3 (e) 3

2
4 4
3
5 5
EE

L
4

E
E
6 6
5 D=4 D=4 D=4
D=6 7 D=6 7 D=6
6 D=8 D=8 D=8
8 8
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
kt kt kt

(b) (d) (f)

0.6 0.6 0.6
EE/ t

E / t

EL/ t
0.4 0.4 0.4
F

0.2 0.2 0.2

D=4 D=4 D=4
D=6 D=6 D=6
0 D=8 0 D=8 0 D=8

5 10 15 5 10 15 5 10 15
kt kt kt

Fig. 1.13 Direct estimation of the largest Lyapunov exponent from a noisy fy.n/g time series
(SNR D 30 dB) of the folded-towel map of length Nd D 65;536 for different embedding
dimensions D D 4, D D 6, D D 8 using a lag of L D 1. The diagrams (a), (c), and (e) show E
vs. kt with t D 1 for the Euclidean norm. In (b), (d), and (f) the corresponding slopes E=t
vs. k (Eq. (1.48)) are shown. The dashed lines indicate the true result 1 D 0:43. Only 10 % of
the reconstructed states with the smallest distances to their neighbours are used for estimating
(exponential) growth rates

the iterated maps considered so far. To avoid strong fluctuations of the slope values
the derivative E=t is estimated by

E E.t C 3t/  E.t  3t/

.t/  (1.49)
t 6t

where E-values at t 3 are used when estimating E=t at time t. Note that the
oscillations are less pronounced for higher reconstruction dimensions. Only 20 % of
the reconstructed states are used as reference points (those which possess the closest
neighbours). The linear scaling regions are clearly visible in the semi-logarithmic
diagrams.
Figure 1.15 shows diagrams with reconstruction dimensions D D 6, D D 11,
and D D 21 and corresponding lags L D 4, L D 2, and L D 1, respectively. In
this case all reconstructed states represent the same windows in time with a length
of .D  1/L D 5  4 D 10  2 D 20  1 D 20 time steps of size t D 0:025,
i.e. a period of time of length 20  0:025 D 0:5 which is close to the period of the
natural oscillations of the Lorenz-63 system. The results for all three state space
1 Estimating Lyapunov Exponents from Time Series 25

(a) 4 (c) 2 (e) 2

3
1 1
2
0 0
1
EE

L
E
E
0 1 1

1 D=4 2 D=4 2 D=4

D=12 D=12 D=12
2 3 3
D=21 D=21 D=21
3
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
kt kt kt
(b) 2.5 (d) 2.5 (f) 2.5

2 2 2
E / t

E / t

EL/ t
1.5 1.5 1.5
E

1 1 1
D=4 D=4 D=4
0.5 D=12 0.5 D=12 0.5 D=12
D=21 D=21 D=21
0 0 0
1 2 3 4 1 2 3 4 1 2 3 4
kt kt kt

Fig. 1.14 Direct estimation of the largest Lyapunov exponent from a fx.n//g time series of the
Lorenz-63 system of length Nd D 65;536 for different reconstruction dimensions D D 4, D D 12,
and D D 21, all with a lag of L D 1. As reference points only those 20 % of the reconstructed
states are used that possess the nearest neighbours. The diagrams (a), (c), and (e) show E vs. kt
with t D 0:025 for the Euclidean norm. In (b), (d), and (f) the corresponding slopes E=t vs.
k (Eq. (1.48)) are shown. The dashed lines indicate the true result 1 D 1:51

reconstruction coincide very well and the amplitude of oscillations of the slope is
much smaller compared to the results shown in Fig. 1.14.

1.5.5 Lorenz-96

Although it possesses only a single positive Lyapunov exponent the 6-dimensional

Lorenz-96 systems turns out to be a surprisingly challenging case for estimating the
largest Lyapunov exponent from time series. Figure 1.16 shows estimation results
for time series of different lengths (first column: Nd D 10;000, second column
Nd D 100;000, third column Nd D 1;000;000) and a different number of reference
points given by those reconstructed states with closest neighbours (first row: 1 %,
second row: 10 %). All examples employing 10 % of the reconstructed states as
reference points provide diagrams where no suitable scaling region exists (even
with Nd D 1;000;000 data points, see Fig. 1.16F, f). If only 1 % of the reconstructed
states is used, the diagram based on Nd D 100;000 samples (Fig. 1.16B, b) gives
a rough estimate of 1 and with Nd D 1;000;000 data points a linear scaling
26 U. Parlitz

(a) 4 (c) 2
(e) 2
3
1 1
2
0 0
1
EE

L
E
E
0 1 1

1 D=6 2 D=6 2 D=6

D=11 D=11 D=11
2 3 3
D=21 D=21 D=21
3
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
kt kt kt
(b) 2.5 (d) 2.5 (f) 2.5

2 2 2
E / t

EF/ t

EL/ t
1.5 1.5 1.5
E

1 1 1
D=6 D=6 D=6
0.5 D=11 0.5 D=11 0.5 D=11
D=21 D=21 D=21
0 0 0
1 2 3 4 1 2 3 4 1 2 3 4
kt kt kt

Fig. 1.15 Direct estimation of the largest Lyapunov exponent from a fx.n//g time series of the
Lorenz-63 system of length Nd D 65;536 for different reconstruction dimensions D D 6, D D 11,
and D D 21, with lags L D 4, L D 2, and L D 1, respectively. As reference points only those 20 %
of the reconstructed states are use that possess the nearest neighbours. The diagrams (a), (c), and
(e) show E vs. kt with t D 0:025 for the Euclidean norm. In (b), (d), and (f) the corresponding
slopes E=t vs. k (Eq. (1.48)) are shown. The dashed lines indicate the true result 1 D 1:51

regime (with the correct slope) is clearly visible in Fig. 1.16C, c. The reconstruction
dimensions used here are D D 9; 18, and 36 with lags L D 4, 2, and 1, respectively,
resulting in window lengths 8  4 D 32, 17  2 D 34, and 35  1 D 35. The
slopes given in Fig. 1.16 were computed with Eq. (1.48) and only the case of
the Euclidean norm EE is shown here, because EF and EL show very similar
results. The observation that a time series of length Nd D 1;000;000 (at least)
is required to obtain reliable and correct results is consistent with the results of
Eckmann and Ruelle [13] who estimated that the amount of required data points
increases as a power of the attractor dimension. For comparison, the Kaplan
Yorke dimension of the Lorenz-96 attractor (DKY D 4:18) is more than twice as
large as the dimension of the Lorenz-63 model and so instead of 64k data a time
series of length longer than 642 k D 4M would be necessary to obtain comparable
results.3

3
This is just a rough estimate, because the choice of the sampling time t and the resulting
distribution of reconstructed states on the attractor have also to be taken into account when
estimating the required length of the time series.
1 Estimating Lyapunov Exponents from Time Series 27

Fig. 1.16 Direct estimation of the largest Lyapunov exponent from a fx1 .n//g time series of the
Lorenz-96 system for different reconstruction dimensions D D 9, D D 18, and D D 36 with
corresponding lags L D 4, L D 2, and L D 1, respectively. Diagrams (A)(F) show EE vs.
kt with t D 0:025 and diagrams (a)(f) give the corresponding local slopes E=t vs. k
(Eq. (1.48)) (EE is the error with respect to the Euclidean norm). The dashed lines indicate the
true result 1 D 1:249. In diagrams (A)(C), (a)(c) only 1 % of the reconstructed states with the
smallest distances to their neighbours are selected for estimating (exponential) growth rates, while
in (D)(F), (d)(f) 10 % are used. Diagrams (A), (a) and (D), (d) are generated using Nd D 10;000
samples, figures (B), (b) and (E), (e) are computed from Nd D 100;000 data points, and diagrams
(C), (c) and (F), (f) show results obtained from a time series of length Nd D 1;000;000

1.6 Conclusion

Estimating Lyapunov exponents from time series is a challenging task and since any
algorithm provides results (i.e., numbers) some error control is very important
to avoid misleading interpretation of the values obtained. From the variety of
28 U. Parlitz

estimation methods, currently only the direct methods provide some feedback to
the user whether local exponential divergence is properly identified or not. The
presented examples included cases where this was not the case, due to:
(a) a too short time series (compared to the dimension of the underlying attractor),
resulting in neighbouring reconstructed states whose distances exceed the range
of validity of locally linearized dynamics, see for example Fig. 1.16A, B
(b) (measurement) noise, see for example Fig. 1.13, or
(c) an observable which is not suitable (to faithfully unfold the dynamics in
reconstruction space), see for example Fig. 1.11.
This failure was in all cases directly visible in the semi-logarithmic diagrams
showing the average growth of mutual distances of neighbouring states vs. time,
where no linear scaling region could be identified. If, on the contrary, such a linear
scaling region exists then it provides strong evidence for deterministic chaos and
the estimated slope can be trusted to be a good estimate of the largest Lyapunov
exponent. The choice of the norm for quantifying the divergence of trajectories
turned out to be noncritical because all three norms used (EE , EF , and EL , see
Sect. 1.3.2) used exhibited equivalent performance.
A particular challenge are time series from high dimensional chaotic attractors.
Eckmann and Ruelle [13] estimated that the number of required data points Nd
exponentially grows with the attractor dimension Da as Nd  constDa . The
results obtained for the folded-towel map (Sect. 1.5.3) and the 6-dimensional
Lorenz-96 model (Sect. 1.5.5) confirmed this (pessimistic) prediction. Although
the 6-dimensional Lorenz-96 model possesses a chaotic attractor with a single
positive Lyapunov exponent it possesses a KaplanYorke dimension of DKY D 4:18.
Due to this relatively high attractor dimension, satisfying estimates of the largest
Lyapunov exponent were obtained only from very long time series (Fig. 1.16C) and
if only those trajectory segments are used for estimating local divergence which
started from very closely neighbouring reconstructed states (1 % in Fig. 1.16C).
This selection of suitable reference points is very similar to a fixed size approach
(see Sect. 1.3.2) using a relatively small radius and the results obtained for the
folded-towel map and the Lorenz-96 model indicate its importance for coping with
high dimensional chaos. On the other hand, these examples clearly show that data
requirements (and practical difficulties) increase exponentially with the dimension
of the underlying attractor (at least for the direct estimation methods employed here)
and this fact imposes fundamental bounds for estimating Lyapunov exponents from
time series generated by processes of medium or even high complexity.

Acknowledgements Inspiring scientific discussions with S. Luther and all members of the
Biomedical Physics Research Group and financial support from the German Federal Ministry
of Education and Research (BMBF) (project FKZ 031A147, GO-Bio) and the German Research
Foundation (DFG) (Collaborative Research Centre SFB 937 Project A18) are gratefully acknowl-
edged.
1 Estimating Lyapunov Exponents from Time Series 29

Appendix

Let ' k W RD ! RD be the induced flow in reconstruction space mapping

reconstructed states xn D .sn ; snCL ; : : : ; snC.D1/L / to their future values ' k .xn / D
xnCk . To estimate the D  D Jacobian matrices D' k .x/ from the temporal evolution
of the reconstructed states fxn gNnD1 the flow ' k has to be approximated by a general
ansatz (black-box model) like a neural network [20] or a superposition of I basis
functions bi W RD ! R providing an approximating function
!
  X I X
I
k
.x/ D 1k .x/; : : : ; Dk .x/ D ci1 bi .x/; : : : ; ciD bi .x/ D b.x/  C
iD1 iD1
(1.50)
where C D .cij / denotes a I  D matrix of coefficients with columns c. j/ ( j D
1; : : : ; D) that have to be estimated and b.x/ D .b1 .x/; : : : ; bI .x// is a row vector
consisting of the values of all basis functions evaluated at the state x.4
For the special choice k D L (evolution time step equals the lag of the delay
coordinates) the first D  1 components of the map k .xn / are known (due to the
delay reconstruction) and only for the last component an approximation is required
!
X
I
L
.x/ D snCL ; snC2L ; : : : ; snC.D1/L ; ci bi .x/ (1.51)
iD1

D .xn2 ; : : : ; xnD ; b.x/  c/: (1.52)

With this notation the approximation D k .x/ of the desired Jacobian matrix
D' k .x/ of the (induced) flow '.x/ in embedding space can be written as
0 1
@b1 @bI
: : :
B @x: 1 : @x: 1 C
D k .x/ D B C
@ :: : : :: A  C D G  C (1.53)
@b1 @bI
@xD
: : : @xD

where G will be called derivative matrix in the following.

For k D L the Jacobian matrix of the approximating function k
is given as
0 1
0 1 0 ::: 0
B 0 0 1 ::: 0 C
B C
B :: :: :: : : :: C
D .x/ D B
L
B : : : : :
C
C (1.54)
B C
@ 0 0 0 ::: 1 A
PI @bi P @bi
iD1 ci @x1 : : : : : : : : : IiD1 ci @x D

4
The matrix C and its column vectors c.j/ depend on the time step k. To avoid clumsy notation this
dependance is not explicitly indicated.
30 U. Parlitz

Linear basis functions bi .x/ can be used to model the (linearized) flow (very) close
to the reference points xn along the orbits. To approximate the flow in a larger
neighbourhood of xn or even globally, nonlinear basis functions are required, like
multidimensional polynomials [2, 47], or radial basis functions [23, 24, 35].
To estimate the coefficient matrix C in Eq. (1.50) or the coefficient vector c in
Eq. (1.51) we select a set of representative states fzj g whose temporal evolution
' k .zj / is known. For local modeling this set of states consists of nearest neighbours
fxm.n/ W m.n/ 2 Un g of the reference point xn where Un defines the chosen
neighbourhood that can be of fixed mass (a fixed number K of nearest neighbours
of xn ) or of fixed size (all points with distance smaller than a given bound ). For
global modeling of the flow the set fzj g is usually a (randomly sampled) subset of
all reconstructed states. Let
0 k 1 1
'1 .z / : : : 'Dk .z1 /
B :: C
Y D @ ::: ::
: : A (1.55)
'1k .zJ / : : : 'Dk .zJ /

be a J  D matrix whose rows are components the (known) future values ' k .zj / of
the J states fzj g and let
0 1
b1 .z1 / : : : bI .z1 /
B :: :: :: C
BD@ : : : A (1.56)
b1 .z / : : : bI .zJ /
J

be the J  I (design) matrix [37] whose rows are the basis functions bi ./ evaluated
at the selected states fzj g. Using this notation the approximation task can be stated
as a minimization problem with a cost function

g.c.j/ / D kB  c.j/  y.j/ k2 (1.57)

where y.j/ denotes the j-th column of the matrix Y (given in Eq. (1.55)), or

g.C/ D kB  C  Yk2F (1.58)

where k  kF D denotes the Frobenius matrix norm (also called Schur norm).
The solution of this optimization problem may suffer from the fact that typically
the states fzj g cover only some subspace of the reconstructed state space. Therefore,
in particular for local modeling ill-posed optimization problems may occur with
many almost equivalent solutions. For estimating Lyapunov exponents we prefer
to select solutions for the coefficient matrix C that provide partial derivatives
(elements of the Jacobian matrix) with small magnitudes, because in this way
spurious Lyapunov exponents are shifted towards 1. This goal can be achieved
by TikhonovPhilips regularization where the cost function of the optimization
1 Estimating Lyapunov Exponents from Time Series 31

problem (1.58) is extended by a term

kA  Ck resulting in

g.c.j/ / D kB  c.j/  y.j/ k2 C

2 kA  c.j/ k2 (1.59)

where A denotes a so-called stabilizer matrix and

2 R is the regularization
parameter that is used to control the impact of the regularization term on the solution
of the minimization problem. If the identity matrix is used as stabilizer A D I
then kc.j/k is minimized and the solution with the smallest coefficients is selected
(also called Tikhonov stabilization). Another possible choice is the derivative matrix
(1.53) A D G. In this case we minimize the sum of all squared singular values i of
D k .x/ D U  S  V tr , because
 
kG  Ck2F D kD k
.x/k2F D trace D k
.x/tr  D k
.x/ (1.60)

  X
D
D trace V  S2  V tr D trace.S2 / D i2 (1.61)
iD1

and so we minimize Lyapunov exponents by maximizing contraction rates.

To solve the optimization problem (1.59) we rewrite it as an augmented least
squares problem with a cost function
   
B y.j/
.j/
g.c / D k .j/
c  k2 D kBO  c.j/  yO.j/ k2 (1.62)

A 0

that can be minimized by a solution of the corresponding normal equations

 
Btr  B C
2 Atr  A  c.j/ D Btr  y.j/ (1.63)

using a sequence of Householder transformations [35] or by employing the singular

value decomposition of the matrix BO D UBO  SBO  VBtrO providing the minimal solution
[37]

c.j/ D VBO  SB1 .j/

O  UBO  yO :
tr
(1.64)

for each column cO.j/ or

tr O
C D VBO  SB1
O  UBO  Y (1.65)
 
Y
for the full coefficient matrix C where YO D .
0
32 U. Parlitz

For the stabilizer A D I the elements of the diagonal matrix S1

O
are given by
B

O i
(1.66)
O 2i C
2

O
where O i are the diagonal elements of SBO (i.e., the singular values of B).

References

1. Abarbanel, H.D.I.: Analysis of Observed Chaotic Data. Springer, New York (1996)
2. Abarbanel, H.D.I., Brown, R., Kennel, M.B.: Lyapunov exponents in chaotic systems: their
importance and their evaluation using observed data. Int. J. Mod. Phys. B 5, 13471375 (1991)
3. Benettin, G., Galgani, L., Giorgilli, A., Strelcyn, J.M.: Lyapunov characteristic exponents for
smooth dynamical systems and for hamiltonian systems; a method for computing all of them.
Part II: Numerical application. Meccanica 15, 2130 (1980)
4. Briggs, K.: An improved method for estimating Liapunov exponents of chaotic time series.
Phys. Lett. A 151, 2732 (1990)
5. Brown, R.: Calculating Lyapunov exponents for short and/or noisy data sets. Phys. Rev. E
47(6), 39623969 (1993)
6. Brown, R., Bryant, P., Abarbanel, H.D.I.: Computing the Lyapunov spectrum of a dynamical
system from an observed time series. Phys. Rev. A 43, 27872806 (1991)
7. Bryant, P., Brown, R., Abarbanel, H.D.I.: Lyapunov exponents from observed time series. Phys.
Rev. Lett. 65, 15231526 (1990)
8. Cenys, A.: Lyapunov spectrum of the maps generating identical attractors. Europhys. Lett.
21(4), 407411 (1993)
9. Dmmig, M., Mitschke, F.: Estimation of Lyapunov exponents from time series: the stochastic
case. Phys. Lett. A 178, 385394 (1993)
10. Darbyshire, A.G., Broomhead, D.S.: Robust estimation of tangent maps and Liapunov spectra.
Physica D 89(34), 287305 (1996)
11. Dechert, W.D., Genay, R.: The topological invariance of Lyapunov exponents in embedded
dynamics. Physica D 90, 4055 (1996)
12. Eckmann, J.-P., Ruelle, D.: Ergodic theory of chaos and strange attractors. Rev. Mod. Phys. 57,
617656 (1985)
13. Eckmann, J.-P., Ruelle, D.: Fundamental limitations for estimating dimensions and Lyapunov
exponents in dynamical systems. Physica D 56, 185187 (1992)
14. Eckmann, J.-P., Kamphorst, S.O., Ruelle, D., Ciliberto, S.: Lyapunov exponents from time
series. Phys. Rev. A 34, 49714979 (1986)
15. Ellner, S., Gallant, A.R., McCaffrey, D., Nychka, D.: Convergence rates and data requirements
for Jacobian-based estimates of Lyapunov exponents from data. Phys. Lett. A 153, 357363
(1991)
16. Fell, J., Beckmann, P.: Resonance-like phenomena in Lyapunov calculations from data
reconstructed by the time-delay method. Phys. Lett. A 190, 172176 (1994)
17. Fell, J., Rschke, j., Beckmann, P.: Deterministic chaos and the first positive Lyapunov
exponent: a nonlinear analysis of the human electroencephalogram during sleep. Biol. Cybern.
69, 139146 (1993)
18. Gao, J., Zheng, Z.: Local exponential divergence plot and optimal embedding of a chaotic time
series. Phys. Lett. A 181, 153158 (1993)
19. Geist, K., Parlitz, U., Lauterborn, W.: Comparison of different methods for computing
Lyapunov exponents. Prog. Theor. Phys. 83, 875893 (1980)
1 Estimating Lyapunov Exponents from Time Series 33

20. Gencay, R., Dechert, W.D.: An algorithm for the n Lyapunov exponents of an n-dimensional
unknown dynamical system. Physica D 59, 142157 (1992)
21. Hnon, M.: A two-dimensional mapping with a strange attractor. Commun. Math. Phys. 50(1),
6977 (1976)
22. Holzfuss, J., Lauterborn, W.: Liapunov exponents from a time series of acoustic chaos. Phys.
Rev. A 39, 21462152 (1989)
23. Holzfuss, J., Parlitz, U.: Lyapunov exponents from time series. In: Arnold, L., Crauel, H.,
Eckmann, J.-P. (eds.) Proceedings of the Conference Lyapunov Exponents, Oberwolfach 1990.
Lecture Notes in Mathematics, vol. 1486, pp. 263270. Springer, Berlin (1991)
24. Kadtke, J.B., Brush, J., Holzfuss, J.: Global dynamical equations and Lyapunov exponents
from noisy chaotic time series. Int. J. Bifurcat. Chaos 3, 607616 (1993)
25. Kantz, H.: A robust method to estimate the maximal Lyapunov exponent of a time series. Phys.
Lett. A 185, 7787 (1994)
26. Kantz, H., Schreiber, T.: Nonlinear Time Series Analysis. Cambridge University Press,
Cambridge (2004)
27. Kantz, H., Radons, G., Yang, H.: The problem of spurious Lyapunov exponents in time series
analysis and its solution by covariant Lyapunov vectors. J. Phys. A: Math. Theor. 46, 254009
(2013)
28. Kostelich, E.: Bootstrap estimates of chaotic dynamics. Phys. Rev. E 64, 016213 (2001)
29. Kruel, Th.M., Eiswirth, M., Schneider, F.W.: Computation of Lyapunov spectra: effect of
interactive noise and application to a chemical oscillator. Physica D 63, 117137 (1993)
30. Kurths, J., Herzel, H.: An attractor in solar time series. Physica D 25, 165172 (1987)
31. Lorenz, E.N.: Deterministic nonperiodic flow. J. Atmos. Sci. 20(2), 130141 (1963)
32. Lorenz, E.N.: Predictability a problem partly solved. In: Proceedings of the Seminar on
Predictability, vol. 1, pp. 118. ECMWF, Reading (1996)
33. Oseledec, V.I.: A multiplicative ergodic theorem. Lyapunov characteristic numbers for dynam-
ical systems. Trans. Moscow Math. Soc. 19, 197231 (1968)
34. Ott, E.: Chaos in Dynamical Systems. Cambridge University Press, Cambridge (1993)
35. Parlitz, U.: Identification of true and spurious Lyapunov exponents from time series. Int. J.
Bifurcat. Chaos 2, 155165 (1992)
36. Parlitz, U.: Lyapunov exponents from Chuas circuit. J. Circuits Syst. Comput. 3, 507523
(1993)
37. Press, W.H., Teukolsky, S.A., Vetterling, W.T., Flannery, B.P.: Numerical Recipes: The Art of
Scientific Computing, 3rd edn. Cambridge University Press, Cambridge (2007)
38. Rosenstein, M.T., Collins, J.J., de Luca, C.J.: A practical method for calculating largest
Lyapunov exponents from small data sets. Physica D 65, 117134 (1993)
39. Rssler, O.E.: An equation for hyperchaos. Phys. Lett. A 71, 155157 (1979)
40. Sano, M., Sawada, Y.: Measurement of the Lyapunov spectrum from a chaotic time series.
Phys. Rev. Lett. 55, 10821085 (1985)
41. Sato, S., Sano, M., Sawada Y.: Practical methods of measuring the generalized dimension and
largest Lyapunov exponent in high dimensional chaotic systems. Prog. Theor. Phys. 77, 15
(1987)
42. Sauer, T., Yorke, J.A.: How many delay coordinates do you need? Int. J. Bifurcat. Chaos 3,
737744 (1993)
43. Sauer, T., Yorke, J., Casdagli, M.: Embedology. J. Stat. Phys. 65, 579616 (1991)
44. Sauer, T.D., Tempkin, J.A., Yorke, J.A.: Spurious Lyapunov exponents in attractor reconstruc-
tion. Phys. Rev. Lett. 81, 43414344 (1998)
45. Shimada, I., Nagashima, T.: A numerical approach to ergodic problems of dissipative dynami-
cal systems. Prog. Theor. Phys. 61, 16051616 (1979)
46. Skokos, Ch.: The Lyapunov characteristic exponents and their computation. In: Lecture Notes
in Physics, vol. 790, pp. 63135. Springer, Berlin (2010)
47. Stoop, R., Meier, P.F.: Evaluation of Lyapunov exponents and scaling functions from time
series. J. Opt. Soc. Am. B 5, 10371045 (1988)
34 U. Parlitz

48. Stoop, R., Parisi, J.: Calculation of Lyapunov exponents avoiding spurious elements. Physica
D 50, 8994 (1991)
49. Takens, F.: Detecting strange attractors in turbulence. In: Lecture Notes in Mathematics,
vol. 898, pp. 366381. Springer, Berlin (1981)
50. Theiler, J.: Estimating fractal dimension. J. Opt. Soc. Am. A 7, 10551073 (1990)
51. Wolf, A., Swift, J.B., Swinney, L., Vastano, J.A.: Determining Lyapunov exponents from a
time series. Physica D 16, 285317 (1985)
52. Yang, H.-L., Radons, G., Kantz, H.: Covariant Lyapunov vectors from reconstructed dynamics:
the geometry behind true and spurious Lyapunov exponents. Phys. Rev. Lett. 109(24), 244101
(2012)
53. Zeng, X., Eykholt, R., Pielke, R.A.: Estimating the Lyapunov-exponent spectrum from short
time series of low precision. Phys. Rev. Lett. 66, 32293232 (1991)
54. Zeng, X., Pielke, R.A., Eykholt, R.: Extracting Lyapunov exponents from short time series of
low precision. Mod. Phys. Lett. B 6, 5575 (1992)
Chapter 2
Theory and Applications of the Fast Lyapunov
Indicator (FLI) Method

Elena Lega, Massimiliano Guzzo, and Claude Froeschl

Abstract In the last 20 years numerical experiments have allowed to study

dynamical systems in a new way providing interesting results. The development
of tools for the detection of regular and chaotic orbits has been one of the key
points to access the global properties of dynamical systems. In many cases the
visualization of suitably chosen sections of the phase space has been determinant
for the comprehension of the fascinating and complex interplay between order and
chaos. The Fast Lyapunov Indicator introduced in Froeschl et al. (Celest Mech
Dyn Astron 67:4162, 1997) and further developed in Guzzo et al. (Physica D
163(12):125, 2002), is an easy to implement and sensitive tool for the detection
of order and chaos in dynamical systems. Closely related to the computation of
the Largest Lyapunov Exponent, the Fast Lyapunov Indicator relies on the idea
that the computation of tangent vectors contains a lot of information even on short
integration times, while for the Largest Lyapunov Indicator large integration times
are required in order to accurately approximate a limit value. The aim of this Chapter
is to provide the definition of the Fast Lyapunov indicator and some simple examples
of applications for readers that would like to implement and use the indicator for
the first time. We associate to each example of application the references to more
specific papers that we have published during these years.

E. Lega ()
Universit de Nice Sophia Antipolis, CNRS UMR 7293, Observatoire de la Cte dAzur, Bv. de
lObservatoire, CS 34229, 06304 Nice cedex 4, France
e-mail: [email protected]
M. Guzzo
Dipartimento di Matematica, Universit degli studi di Padova, Via Trieste, 63 - 35121 Padova,
Italy
e-mail: [email protected]
C. Froeschl
Universit de Nice Sophia Antipolis, CNRS UMR 7293, Observatoire de la Cte dAzur, Bv. de
lObservatoire, B.P. 4229, 06304 Nice cedex 4, France
e-mail: [email protected]

Springer-Verlag Berlin Heidelberg 2016 35

Ch. Skokos et al. (eds.), Chaos Detection and Predictability, Lecture Notes
in Physics 915, DOI 10.1007/978-3-662-48410-4_2
36 E. Lega et al.

2.1 Introduction

The study of the interplay between order and chaos is one of the keys for under-
standing the behaviour of complex systems. Since the pioneering work of Hnon
and Heiles [27] the use of numerical simulations together with the development
of different tools for the detection of chaos has provided interesting results in
different domains of physics (celestial mechanics, particle accelerators, dynamical
astronomy, statistical physics, plasma physics).
In their study, Hnon and Heiles, searching for the existence of a third integral
of motion in a galactic potential, were surprised by finding that ordered and
chaotic motions co-existed for some values of the total energy of the system.
As usual in numerical experiments the authors searched for eventual numerical
errors. Listening to a seminar by Arnold about new theoretical results on stability
of quasi-integrable Hamiltonian systems (the nowadays celebrated KAM theorem
[1, 28, 43]), M. Hnon understood that order and chaos are complementary rather
than antagonist dynamical behaviours and got convinced on the numerical results
he had obtained with C. Heiles in their study of the galactic potential. Since then,
numerical experiments have become a sort of laboratory, very often used to extend
the domain of validity of theorems and therefore showing their interest for physical
problems.
As an example, we consider the problem of the long term stability properties
of a dynamical system, problem of particular interest in the domain of celestial
mechanics. During the last decade the numerical detection of the resonances of a
system using dynamical indicators has been one of the major tools for studying the
long-term stability in the specific case of celestial mechanics (for recent examples,
see [15, 16, 29, 42, 4951, 55]). The reason is that many problems of interest for
celestial mechanics can be studied with KAM [1, 28, 43] and Nekhoroshev theorems
[48]. For small values of the perturbation parameters the KAM theorem leaves the
possibility of large instabilities only on a peculiar subset of the phase space, the so-
called Arnold web. According to the Nekhoroshev theorem, on the Arnold web the
diffusion times are expected to increase at least exponentially with an inverse power
of the norm of the perturbation. This phenomenon of extremely slow diffusion
was introduced by Arnold [2] on an ad-hoc model well suited to the mathematical
demonstration rather than for numerical experiments.
We recall that, for many years, researchers were convinced that Arnolds
diffusion could not be detected numerically, and therefore, in some sense, the
phenomenon was not interesting for the study of physical systems.
The Fast Lyapunov Indicator (FLI hereafter), introduced in [12] and further
developed in [21], is an easy (to implement) and sensitive tool for the detection of
the Arnold web of a system. The FLI method was first tested by comparing results
obtained with other chaos indicators. A detailed comparison with the frequency
analysis application on two and four dimensional mapping [30, 31] can be found in
[10, 32]. The comparison with other chaos indicators was presented in [33]. Without
entering in the details, we can say that the FLI belongs to the class of the so called
2 Theory and Applications of the Fast Lyapunov Indicator (FLI) Method 37

finite time chaos indicators (such as the Finite Time Lyapunov Exponent [52], the
MEGNO [8, 9] as well as OFLI and OFLI2 [3]) which are able to discriminate
between regular orbits and chaotic orbits on times significantly smaller than the time
required for a reliable estimation of the largest characteristic Lyapunov exponent or
of the frequency.
The detailed detection of the resonances obtained with the FLI on models which
satisfy the hypothesis of both KAM and Nekhoroshev theorems allowed us to
measure directly the quantitative features of the Arnolds diffusion [14, 18, 22
26, 35, 36, 53] showing its interest for physical systems. Later in [17, 24, 3740]
we have used the FLI for the detection of the stable and unstable manifolds.
More recently the FLI has been applied to the planar circular restricted three body
problem for the detection and characterization of close encounters and resonances
[19, 41]; more precisely, we have formulated the FLI method using the LeviCivita
regularization in order to handle the singularity of the gravitational potential.
The majority of our studies concern conservative systems, however, we have used
the FLI for studying the dynamics of dissipative systems in [57]; more recently
we have provided an application of the FLI to track the diffusion of orbits of a
quasi integrable Hamiltonian system perturbed with a very small non-Hamiltonian
perturbation [18].
In this Chapter, rather than providing a review of the results obtained with the
FLI, we present the indicator for readers that would like to implement it for the
first time. At this purpose we provide in Sect. 2.2 the definition and use of the FLI
on a simple 2-dimensional discrete model: the standard map. On this model we
try to answer to some frequently asked questions about the implementation and
use of the method. In Sect. 2.3 we show the use of the FLI for the computation
of the stable and unstable manifolds. In Sect. 2.4 we provide an application on a
generic Hamiltonian model. In Sect. 2.5 we show an application of the FLI for the
detection of the resonances of a quasi-integrable Hamiltonian system with 3 degrees
of freedom and we show how to use the FLI to follow the diffusion of orbits along
resonant lines. Conclusions are provided in Sect. 2.6.

2.2 Definition of the Fast Lyapunov Indicator

Given a set of differential equations:

dx
dt
D F.x/ ; x D .x1 ; x2 ; ::::xn / (2.1)

for any solution x.t/ with initial condition x.0/ the evolution v.t/ of any tangent
vector with initial value v.0/ is obtained by integrating the variational equations:
( dx
dt
D F.x/
(2.2)
@F
dv
dt
D @x
v:
38 E. Lega et al.

If instead one considers the discrete-time dynamics defined by the map

x.t C 1/ D .x.t//; (2.3)

the evolution of the tangent vector is defined by:


x.t C 1/ D .x.t//
@ (2.4)
v.t C 1/ D @x .x.t//v.t/:

With this setting, for both systems (2.1) and (2.3), the simplest definition of the fast
Lyapunov indicator of a point x.0/ and of a tangent vector v.0/, at time t, is:

kv.t/k
FLIt .x.0/; v.0// D log : (2.5)
kv.0/k

The FLI is defined in such a way that, unless v.0/ belongs to some lower
dimensional linear spaces, the quantity FLIt .x.0/; v.0//=t tends to the largest
Lyapunov exponent as t goes to infinity. If Eq. (2.1) is Hamiltonian and if the motion
is regular (except for some peculiar hyperbolic structures, such as whiskered tori)
then the largest Lyapunov exponent is zero, otherwise it is positive. This property
has been largely used to discriminate between chaotic and ordered motions.
However, among regular motions the Lyapunov exponent does not distinguish
between circulation and libration orbits. In contrast, the FLI distinguishes between
them ([13, 34], see Sect. 2.1).
Therefore, the computation of FLIt .x; v/ on grids of initial conditions x and for
the same fixed tangent vector v allows one to detect the distribution of invariant tori
and resonances (i.e. circulation and libration orbits) in relatively short CPU times
[11, 13].
We remark that the FLI depends parametrically on the initial vector v.0/ and on
the integration time t. A frequently asked question concerns the choice of v.0/ for
the practical implementation of the method. As for the computation of the largest
Lyapunov exponent, one has in principle to avoid special choices of v.0/. In order
to reduce the dependence of the computation on the choice of the initial tangent
vector we suggested in [18] to compute the average (or alternatively the maximum)
of the FLIs obtained for an orthonormal basis of tangent vectors. It happens that
any orthonormal basis is suitable to detect the dynamics of the system. A second
frequently asked question concerns the choice of the integration time t. We answer
to both questions in the following using the standard map as a model problem.

2.2.1 The Standard Map as a Model Problem

We consider as a model problem the two-dimensional standard map, whose phase

space variables are denoted by .I; '/ 2 R  S1 , and whose dynamics is defined by

.I.t C 1/; '.t C 1// D .I.t/; '.t// (2.6)

2 Theory and Applications of the Fast Lyapunov Indicator (FLI) Method 39

with

.I; '/ D .I C sin.' C I/; ' C I/;

and is a parameter. As it is well known, the map has interesting dynamics

for 0. As an example, in Fig. 2.1, we report the phase-portrait of the map
for D 0:3: we can appreciate the presence of invariant curves, as well as of
a small chaotic zone around the hyperbolic fixed point .0; 0/. We select three
initial conditions on the phase-portrait of Fig. 2.1, according to different dynamical
features: we select an initial condition in the small chaotic region around the
origin, a second one corresponding to a resonant libration and finally a third one
corresponding to a circulation curve (see Fig. 2.1). In Fig. 2.2 we report the time
evolution of the FLI for the three orbits. We remark that about 10 iterations of the
map are enough to differentiate the chaotic orbit, whose tangent vector growths
approximately exponentially with time, from the regular libration and circulation,
whose tangent vectors increase almost linearly with time (correspondingly, the
FLI increases almost linearly with time for the chaotic orbit, and approximately
logarithmically for the regular motions).
In order to reduce the fluctuations that appear on Fig. 2.2 one can conveniently
compute, instead of the indicator defined in (2.5), the indicator:

FLI.x.0/; v.0/; t/ D sup log jjv.k/jj (2.7)

0kt

Fig. 2.1 On the left: a set of orbits of the standard mapping for D 0:3. The black points
correspond to a resonant libration orbit of initial conditions .I.0/; '.0// D .0; 1:5/ and to a
circulation orbit of initial conditions .I.0/; '.0// D .1:5; 0/. On the right: enlargement around
the hyperbolic fixed point at I.0/ D 0, '.0/ D 0
40 E. Lega et al.

20

18
Libration Orbit
16 Circulation Orbit
Chaotic Orbit
14

12
FLI

10

0
1 10 100 1000
t

Fig. 2.2 Evolution with time of the FLI for the standard map of Eq. (2.6) for D 0:3 for 3 orbits
of initial conditions .105 ; 0/ corresponding to the small chaotic region around the hyperbolic
fixed point at the origin (right panel of Fig. 2.1), .0; 1:5/ and .1:5; 0/ corresponding respectively to
the libration orbit and to the circulation curve marked with points in Fig. 2.1

Figure 2.3 shows that computing the FLI as in (2.7) the fluctuations become
negligible. We remark that, while the Largest Lyapunov exponent is zero for both
libration and circulation orbits, their corresponding FLI are different. In fact, using
a refined perturbation theory we have shown in [21] that the value of the FLI differs
at order 0 in , between libration and circulation motions even for more general
systems. In Fig. 2.4 we show the FLI value at t D 1000, obtained for a set of 900 
900 orbits of the standard map with D 0:3 and with I.0/ and '.0/ regularly
spaced in the interval  W . We have considered 2 orthogonal initial vectors
v.0/ D .1; 0/ and w.0/ D .0; 1/ and we have computed the FLI value using Eq. (2.7)
on both vectors; we plot the largest between the two FLI values. When compared
with Fig. 2.1 we clearly see that the three different dynamics are well distinguished:
the largest FLI values corresponding to chaotic motions, the intermediate values
to circulation orbits and the lower values to libration orbits. In [4] it was shown
on a pendulum problem that some spurious pattern appear when using FLI (their
Fig. 3). We notice that considering the largest between the two FLI values obtained
on orthogonal initial vectors, there are no spurious structures in the FLI computation
shown in Fig. 2.4.

2.2.2 The Choice of the Integration Time

A second frequently asked question concerns the choice of the integration time.
A practical way to choose a suitable integration time is to compute the FLI for
different time values and see for which time the picture gets stable. For example,
2 Theory and Applications of the Fast Lyapunov Indicator (FLI) Method 41

3.5

3
Libration orbit
Circulation orbit
2.5

2
FLI

1.5

0.5

0
1 10 100 1000
t

Fig. 2.3 Time evolution of the FLI for the circulation orbit and the libration orbit of Fig. 2.2. The
FLI is computed as in (2.7). When considering the supremum of log jjv.t/jj the fluctuations, which
are due to the geometry of the orbits, become negligible and libration motion is well distinguished
from circulation

Fig. 2.4 Computation of the FLI for t D 1000 using Eq. (2.7) on a grid of 900  900 initial
conditions regularly spaced in the interval  W . Precisely, two FLIs have been computed on 2
orthogonal initial vectors v.0/ D .1; 0/ and w.0/ D .0; 1/, the largest FLI value is plotted
42 E. Lega et al.

Fig. 2.5 Computation of the FLI as in Fig. 2.4 at t D 10 (top left), t D 100 (top right), t D 1000
(bottom left), t D 10;000 (bottom right)

lets consider the FLI computation shown in Fig. 2.4 for different times, say t D
10; 100; 1000; 10;000. We see clearly on Fig. 2.5 that t D 10 is a too short time to
distinguish the dynamics while already at t D 100 we clearly distinguish between
the different motions. Few more weakly chaotic orbits are detected at t D 1000 and
no difference appears between t D 1000 and t D 10;000. For this case, t D 1000
has to be considered a suitable integration time.
We remark that, the FLI chart obtained in [14] on the relatively short time t '
1000 provided a representation of the geometry of the resonances which allowed to
follow the diffusion of orbits up to the very long times t ' 1011 .
2 Theory and Applications of the Fast Lyapunov Indicator (FLI) Method 43

For further details about the sensitivity of the method in detecting high order
resonances we refer to [21].

2.3 The FLI for the Computation of the Stable and Unstable
Manifolds

Since the work of Poincar it is well known that the complexity of chaotic motions
in deterministic systems can be appreciated from the analysis of the stable and
unstable manifolds associated to hyperbolic orbits. Different methods can be found
in the literature for the detection of hyperbolic manifolds. The FLI method for the
computation of the hyperbolic manifolds has been introduced in [13, 17, 54] and
used in [24, 25, 39, 40] to investigate the relation between the topology of hyperbolic
manifolds and diffusion. Recently, the method has been used for the detection of the
tube manifolds related to the Lyapunov periodic orbits [41] and for the detection of
multiple close encounters [19] in the case of the restricted planar three body prob-
lem. We do not enter in the details here, we just recall that we have recently provided
[20] an analytic description of the growth of tangent vectors for orbits with initial
conditions which are close to the stable-unstable manifolds of a hyperbolic saddle
point; as a matter of fact, we explain why the Fast Lyapunov Indicator detects the
stable-unstable manifolds of all fixed points which satisfy a certain condition and we
provide a suitably modified Fast Lyapunov Indicator if the condition is not satisfied.
Here we illustrate the use of the FLI in detecting hyperbolic manifolds associated
to the hyperbolic point .0; 0/ on the standard map of Eq. (2.6). Let us recall that
the for a two-dimensional standard map the unstable manifold Wu .Ih ; 'h / of an
hyperbolic point .Ih ; 'h / 2 R  S1 is the set of .I; '/ such that:

Wu .Ih ; 'h / D f.I.0/; '.0// W lim d..Ih ; 'h /; .I.t/; '.t/// D 0g;
t!1

the stable manifold Ws .Ih ; 'h /, is the set of .I; '/ such that:

Ws .Ih ; 'h / D f.I.0/; '.0// W lim d..Ih ; 'h /; .I.t/; '.t/// D 0g:
t!1

The numerical localization of the unstable manifold of an hyperbolic fixed point

can be obtained by propagating a small neighborhood of initial conditions up
to a time T of the order of some Lyapunov times of the fixed point (see [40]
and references therein). In such a way, one directly constructs a neighborhood of
a finite piece of the unstable manifold (for the stable manifold one repeats the
construction for the inverse flow). This method gives very good results for fixed
points of two dimensional maps, because the neighborhoods of the fixed points are
two dimensional and can be propagated with reasonable CPU times. Figure 2.6, left
panel shows the detection of a piece of the stable and of the unstable manifolds of
the standard map of Eq. (2.6) for D 0:3. Figure 2.6, right panel shows the value
of the FLI obtained on a two dimensional grid of regularly spaced initial conditions
44 E. Lega et al.

-0,02 0 0,02
I

Fig. 2.6 Left panel: Detection of a piece of the stable and unstable manifold of the standard map
computed with the usual method of set propagation (see [40] and references therein). Right panel:
Representation of the FLI for the standard map (2.6) for t D 50. Precisely, we plot the average of
two FLIs obtained on the direct and on the inverse map. We can appreciate the details of the lobes
associated to the hyperbolic manifolds and the agreement with the results of the usual method of
set propagation

for a short integration time t D 50. Precisely, two FLIs have been computed one on
the direct and one on the inverse map, and the average of the two FLIs is plotted.
We can appreciate the details of the lobes associated to the hyperbolic manifolds.
The comparison with Fig. 2.6, left panel shows the quality of the detection of pieces
of hyperbolic manifolds as obtained with the FLI computation. The advantage is
that the use of the FLI method easily extends to higher dimensional systems and
moreover one does not need to know in advance the local approximations of the
hyperbolic manifolds.

2.4 Application to a Continuous System

The FLI is easily implemented also for generic continuous dynamical systems.
We consider here, as an example, the computation of the FLI for a particle in
an accelerated logarithmic potential which models the mean motion of stars in
a flat rotation curve galaxy that sustains an asymmetric jet, whose dynamics has
been previously studied in [47] using the traditional method of Poincar surface
of sections. The problem of the influence of stellar jets on the dynamics of
protoplanetary discs was studied in [4446]. In [47] the motion of stars in a flat
rotation curve galaxy that sustains wind episodes was modeled by:

1 2 h2 1
H.
; z; p
; pz / D .p
C p2z / C z2 C log.
2 C z2 /  z (2.8)
2 2
2
2 Theory and Applications of the Fast Lyapunov Indicator (FLI) Method 45

where
and z are the cylindrical coordinates of the star in a reference frame with
origin on the galactic center, p
and pz are the corresponding conjugate momenta,
hz is the projection of the angular momentum along the direction of acceleration
(the z-axis). H and hz are constants of motion of the system. In order to numerically
compute the FLI, we integrate the Hamilton equations of (2.8) with a symplectic
integrator and we write the variational equations of the map representing the
numerical integrator. When dealing with multi-dimensional systems it is evident
that we cant visualize the whole phase space as we did for the 2-dimensional
standard map. However, we can still provide a global view of the dynamics on
suitably chosen 2-dimensional sections. Precisely, the surface of constant energy
H is three dimensional, and we can further reduce the study to a two dimensional
space by fixing the value of one of the three independent variables.
Figure 2.7 shows the FLI computed on a bidimensional grid of 1000  1000
initial conditions regularly spaced in
and z for E D 0:75 and h2z D 0:05831
with integration time t D 200. The other initial conditions are pz D 0 and p
is

Fig. 2.7 FLI computation of 1000  1000 initial conditions regularly spaced in
and z for H D
0:75 and h2z D 0:05831 with integration time t D 200. The other initial conditions are pz D 0
and p
is obtained from the energy equation. For all initial conditions we have chosen v.0/ D
p
.1; 1; 0:5. 5  1/; 1/
46 E. Lega et al.

Fig. 2.8 FLI computation as in Fig. 2.7. To compare the results obtained with the FLI to those of
the traditional method of surface of section we have drawn with black points the intersections of a
set of orbits with the plane
; z obtained setting pz D 0

obtained
p from (2.8). For all initial conditions we have chosen v.0/ D .1; 1; 0:5
. 5  1/; 1/.
We now compare the result of the FLI computation with the results obtained
with the traditional method of surface of section. In Fig. 2.8 we plot with black
points the intersections of a set of orbits with the plane
; z obtained setting pz D 0.
We can observe that, as usual, the larger FLI values correspond to chaotic orbits
(dispersed points on the surface of section) while intermediate and lower FLI values
provide regular motions (closed curves on the surface of section). For the specific
case of H D 0:75 and h2z D 0:05831 the system has a large chaotic region for
smaller and larger elevations z. Moreover, using the FLI, we do not only recover the
results in [47] concerning the integrability, but we easily obtain much more details
in the dynamics. This appears clearly in Fig. 2.9 where the FLI is computed zooming
out Fig. 2.7 and we can see the complexity of the chaotic structures. In the bottom
part of the figure (for
close to 0.25) we recognize the typical lobes related to the
hyperbolic manifold of hyperbolic periodic orbits.
2 Theory and Applications of the Fast Lyapunov Indicator (FLI) Method 47

Fig. 2.9 Zoom of Fig. 2.7. The FLI is computed for 10001000 initial conditions regularly spaced
in
and z for H D 0:75 and h2z D 0:05831 with integration time t D 100

2.5 The FLI for Detecting the Geography of Resonances

The problem of long term stability of an Hamiltonian system is strongly related

to the famous KAM [1, 28, 43] and Nekhoroshev [48] theorems which leave the
possibility of a slow drift of the orbits on a peculiar subset of the phase space, the
so called Arnolds web. The detection of the Arnolds web is therefore the first step
to achieve if we are interested in studying the long term stability properties of a
system.
We consider the following quasi-integrable hamiltonian (that we studied in
several papers, see for example [22, 25, 35]):

I12 I22 1
H D C CI3 C f ; f D ; (2.9)
2 2 cos.'1 / C cos.'2 / C cos.'3 / C 3 C c

where is a small parameter, c > 0, and .I; '/ are action angle variables. In the
integrable case (defined by D 0) the actions are constants of motion while the
angles '1 .t/ D '1 .0/ C I1 t, '2 .t/ D '2 .0/ C I2 t, '3 .t/ D '3 .0/ C t rotate with
frequencies !1 D I1 , !2 D I2 , !3 D 1. Therefore, a couple of actions I1 ; I2
48 E. Lega et al.

characterizes an invariant torus T3 . For any small different from zero, H is not
expected to be integrable. However, if is sufficiently small, the KAM theorem
applies1 : for any invariant torus of the original system with Diophantine non-
resonant frequencies there exists an invariant torus in the perturbed system which is
a small deformation of the unperturbed one.
The phase space has dimension 6, therefore, as for the system discussed
in Sect. 2.4, we need to properly choose sections in order to provide a visual
representation of the dynamics. Since the action I3 does not enter the equations of
motion of all the other variables, we can consider the time evolution in the reduced
phase-space .I1 ; I2 ; '1 ; '2 ; '3 /, and then in this space we consider various sections
by fixing the values of some of the variables, for example, we fix the angles and
consider the section:

S0 D f.I1 ; I2 ; '1 ; '2 ; '3 / 2 R2  T3 W '1 ; '2 ; '3 D 0g: (2.10)

or, alternatively, we fix one action and two angles and consider the section:

S1 D f.I1 ; I2 ; '1 ; '2 ; '3 / 2 R2  T3 W I2 D I2 .0/; '2 ; '3 D 0g (2.11)

or:

S2 D f.I1 ; I2 ; '1 ; '2 ; '3 / 2 R2  T3 W I1 D I1 .0/; '2 ; '3 D 0g: (2.12)

In Fig. 2.10 we show the FLI computed for the three different sections S0 ; S1 ; S2
represented in the three dimensional space .I1 ; I2 ; '1 /. Precisely, in the horizontal
plane we have represented the FLI computed on section S0 using a grid of 500  500
initial conditions regularly spaced in .I1 ; I2 / in the interval 0:5; 1:5; on the
vertical plane on the left (right) we have represented the FLI computed on section
S1 (S2 ) using a grid of 500  500 initial conditions regularly spaced in .I1 ; '1 /
(respectively .I2 ; '1 /) in the intervals 0:5; 1:5 and 0; 2, with respectively
I2 .0/ D 1:5, I1 .0/ D 1:5.
In the horizontal plane we clearly see a web of resonance, located near the
straight lines defined by: k1 !1 C k2 !2 C k3  k1 I1 C k2 I2 C k3 D 0, with
k1 ; k2 ; k3 2 Zn0. For examples the resonances I1 D 0 and I2 D 0 appear as
large lines in the horizontal plane. They both have a chaotic boundary (shown in
yellow). When looking at the vertical panels it appears clearly that the amplitude of
the resonances change with values of the angles.
We now study the evolution of a set of N D 100 chaotic orbits with initial
conditions I1 .i/ D 1:5, I2 .i/ D 0, '2 .i/ D '3 .i/ D 0, '1 D  C 106 i, i D 1; ::::N.
The orbits evolve in a multi dimensional space, therefore it is useful to consider
the points of the orbits intersecting two dimensional sections. On the FLI map of
Fig. 2.11 we plot as black dots the points of the orbits which have returned after

1
H is real analytic and H0 is isoenergetically non-degenerate.
2 Theory and Applications of the Fast Lyapunov Indicator (FLI) Method 49

I2
I1

2 2

1
1

0
0

1.5 1.5

I1
I2

0.5 0.5

Fig. 2.10 FLI computation for the Hamiltonian of Eq. (2.9) for a value of the perturbing parameter
D 0:04. The integration time is t D 100. The initial conditions are regularly spaced on grids of
500  500 points on three different sections of the phase space defined in the text as S0 (horizontal
plane), S1 (vertical left plane) and S2 (vertical right plane)

some time on the sections S2 and S0 . Of course, since computed orbits are discrete,
we represent the points that return in a small neighbourhood of S0 defined by

SQ0 D f.I1 ; I2 ; '1 ; '2 ; '3 / W j'1 j  0:05; j'2 j  0:05; '3 D 0g

and those that return in a small neighbourhood of S2 defined by

SQ2 D f.I1 ; I2 ; '1 ; '2 ; '3 / W j'1 j  0:05; jI1  1:5j < 0:005; '3 D 0g:

Reducing the size of the neighborhood of the sections S0 and S2 reduce the number
of points but doesnt change the results. In Fig. 2.11 we consider three different
integration times: t D 2 105 (left panel), t D 2 107 (middle panel), t D 2 108 (right
panel).
On the vertical plane the black points cover the whole chaotic region associated
to the chaotic border of the resonance I2 D 0 already on short times. As we can
observe in Fig. 2.11 the number of points returning in SQ2 increases with time but the
region covered by the orbits doesnt change.
On the horizontal plane, it is interesting to observe that the orbits slowly diffuse
along the resonance I2 D 0, precisely the points that return in SQ0 cover a portion
50 E. Lega et al.

I2
0.5

0.5

1.5

I1
1

Fig. 2.11 Zoom of Fig. 2.10 around the resonance I2 D 0. We plot as black dots the points
of a set of N D 100 chaotic orbits which have returned after some time on the sections S2
and S0 . Precisely, we represent the points that return in a small neighbourhood of S0 defined by
SQ0 D f.I1 ; I2 ; '1 ; '2 ; '3 / W j'1 j  0:05; j'2 j  0:05; '3 D 0g and those that return in a small
neighbourhood of S2 defined by SQ2 D f.I1 ; I2 ; '1 ; '2 ; '3 / W j'1 j  0:05; jI1  1:5j < 0:005; '3 D
0g. The three panels correspond to different integrations times: t D 2 105 (left panel), t D 2 107
(middle panel) and t D 2 108 (right panel). The initial conditions are chosen in the neighbourhood
of the hyperbolic point: I1 .i/ D 1:5, I2 .i/ D 0, '2 .i/ D '3 .i/ D 0, '1 D  C 106 i, i D 1; ::::N

of the chaotic border of the resonance that increases for increasing values of the
integration time. Let us remark that the FLI, computed on a short total time t D 100,
allows to properly follow orbits on a much larger integration time (t D 2 108 on the
right panel of Fig. 2.11).
Let us remark that the section S0 is particularly suited to the detection of the
slow diffusion since the large oscillations of the action I2 that we observe on section
S2 are filtered when considering the points that return in SQ0 . Moreover, the FLI
chart allows us to check that the orbits really diffuse along the chaotic border of a
resonance.
We have quantified the diffusion by measuring a diffusion coefficient using the
points of sets of orbits returning in the section SQ0 using various Hamiltonian or
discrete mapping models in [14, 18, 22, 2426, 35, 36]. The computation of the
diffusion coefficient has many technical difficulties that we do not recall here. The
interested reader can refer to [25] for a numerical characterization of the statistical
2 Theory and Applications of the Fast Lyapunov Indicator (FLI) Method 51

properties of the diffusion and to [18] for the technical aspects of the computation
of the diffusion coefficient.
In [22] we have shown that the diffusion of orbits occurring along the peculiar
set of resonances has a global character and in [26] we measured a diffusion
coefficient decreasing exponentially through 40 orders of magnitude thus showing
that Arnolds diffusion concerns and can be measured on systems of physical
interest.

2.6 Conclusions

Since the pioneering work of Hnon and Heiles [27] it appeared clearly that
understanding the dynamical behaviours of a system required global studies of the
phase space. Different tools for the detection of chaotic and ordered motions have
been developed since then, providing interesting results in different domains of
physics (celestial mechanics, particle accelerators, dynamical astronomy, statistical
physics, plasma physics). The Fast Lyapunov Indicator was introduced in [12] as an
easy (to implement) and sensitive tool for distinguish between ordered and chaotic
motion. The method was used in [13] for the detection of the Arnold web of a system
and further developed in [21] using a refined perturbation theory which provided the
behaviour of the FLI for different orbits.
In the last 10 years we have used the FLI for studies for which a global
visualization of the phase space was one of the key ingredients for understanding
the problem. For example, the Arnold web computed with short integration time
allowed us to choose possibly diffusing initial conditions and to follow their
evolution on much longer integration time. With a rather technical method we have
measured diffusion coefficients decreasing faster than a power low and possibly
exponentially through many orders of magnitude [14, 18, 22, 2426, 35, 36] showing
the interest of Arnolds diffusion for physical systems.
Later in [17, 24, 3840] we have used the FLI for the detection of the stable
and unstable manifolds. This is in general a difficult task that requires sophisticated
methods. We have recently provided [20] an analytic description of the growth
of tangent vectors for orbits with initial conditions which are close to the stable-
unstable manifolds of a hyperbolic saddle point and we have explained the condition
for the detection the stable-unstable manifolds with the FLI.
Moreover, the FLI has been applied to the planar circular restricted three body
problem for the detection and characterization of close encounters and resonances
[19, 41].
In this chapter we have we presented the indicator for readers that would like
to implement it for the first time. At this purpose we have chosen simple discrete
and continuous model problems giving the element to reproduce this cases. As
examples of applications we have shown (1) the use of the FLI for the detection
of the stable/unstable manifold of a two dimensional model, (2) the FLI for the
52 E. Lega et al.

detection of the resonances of continuous systems and (3) we have explained how
to use the indicator to follow the diffusion of orbits along resonant lines.

References

1. Arnold, V.I.: Proof of a theorem by A.N. Kolmogorov on the invariance of quasi-periodic

motions under small perturbations of the Hamiltonian. Russ. Math. Surv. 18, 9 (1963)
2. Arnold, V.I.: Instability of dynamical systems with several degrees of freedom. Sov. Math.
Dokl. 6, 581585 (1964)
3. Barrio, R.: Sensitivity tools vs. Poincar sections. Chaos Solitons Fractals 25, 711726 (2005)
4. Barrio, R., Borczyk, W., Breiter, S.: Spurious structures in chaos indicators maps. Chaos
Solitons Fractals 40, 16971714 (2009)
5. Celletti, A., Froeschl, C., Lega, E.: Dissipative and weakly-dissipative regimes in nearly-
integrable mappings. Discrete Contin. Dyn. Syst. Ser. A 16(4), 757781 (2006)
6. Celletti, A., Froeschl, C., Lega, E.: Dynamics of the conservative and dissipative spinorbit
problem. Planet. Space Sci. 55, 889899 (2007)
7. Celletti, A., Stefanelli, L., Lega, E., Froeschl, C.: Global dynamics of the regularized restricted
three-body problem with dissipation. Celest. Mech. Dyn. Astron. 109, 265284 (2011)
8. Cincotta, P.M., Sim, C.: Simple tools to study global dynamics in non-axisymmetric galactic
potentials - I. Astron. Astrophys. Suppl. Ser. 147, 205 (2000)
9. Cincotta, P.M., Giordano, C.M., Sim, C.: Phase space structure of multi-dimensional systems
by means of the mean exponential growth factor of nearby orbits. Physica D 182(34), 151178
(2003)
10. Froeschl, C., Lega, E.: Weak chaos and diffusion in Hamiltonian systems. From Nekhoroshev
to Kirkwood. In: Roy, A.E. (eds.) The Dynamics of Small Bodies in the Solar System: A Major
Key to Solar System Studies. NATO/ASI. Kluwer Academic, Boston (1998)
11. Froeschl, C., Lega, E.: On the structure of symplectic mappings. The fast Lyapunov indicator:
a very sensitive tool. Celest. Mech. Dyn. Astron. 78(1/4), 167195 (2000)
12. Froeschl, C., Lega, E., Gonczi, R.: Fast Lyapunov indicators. Application to asteroidal motion.
Celest. Mech. Dyn. Astron. 67, 4162 (1997)
13. Froeschl, C., Guzzo, M., Lega, E.: Graphical evolution of the Arnold web: from order to
chaos. Science 289(5487), 21082110 (2000)
14. Froeschl, C., Guzzo, M., Lega, E.: Local and global diffusion along resonant lines in discrete
quasi-integrable dynamical systems. Celest. Mech. Dyn. Astron. 92(13), 243255 (2005)
15. Guzzo, M.: The web of three-planets resonances in the outer solar system. Icarus 174(1), 273
284 (2005)
16. Guzzo, M.: The web of three-planet resonances in the outer solar system II: a source of orbital
instability for Uranus and Neptune. Icarus 181, 475485 (2006)
17. Guzzo, M.: Chaos and diffusion in dynamical systems through stableunstable manifolds.
In: Perozzi, Mello, F. (eds.) Space Manifolds Dynamics: Novel Spaceways for Science and
Exploration. Novel Spaceways for scientific and exploration missions, a dynamical systems
approach to affordable and sustainable space applications held in Fucino Space Centre
(Avezzano), 1517 October 2007. Springer, New York/Dordrecht/Heidelberg/London (2010)
18. Guzzo, M., Lega, E.: The numerical detection of the Arnold web and its use for long-term
diffusion studies in conservative and weakly dissipative systems. Chaos 23, 23124 (2013)
19. Guzzo, M., Lega, E.: On the identification of multiple close encounters in the planar circular
restricted three-body problem. Mon. Not. R. Astron. Soc. Lett. 428, 26882694 (2013)
20. Guzzo, M., Lega, E.: Evolution of the tangent vectors and localization of the stable and unstable
manifolds of hyperbolic orbits by Fast Lyapunov Indicators. SIAM J. Appl. Math. 74(4), 1058
1086 (2014)
2 Theory and Applications of the Fast Lyapunov Indicator (FLI) Method 53

21. Guzzo, M., Lega, E., Froeschl, C.: On the numerical detection of the effective stability of
chaotic motions in quasi-integrable systems. Physica D 163(12), 125 (2002)
22. Guzzo, M., Lega, E., Froeschl, C.: First numerical evidence of Arnold diffusion in quasi-
integrable systems. Discrete Continuous Dyn. Syst. Ser. B 5(3), 687698 (2005)
23. Guzzo, M., Lega, E., Froeschl, C.: Diffusion and stability in perturbed non-convex integrable
systems. Nonlinearity 19, 10491067 (2006)
24. Guzzo, M., Lega, E., Froeschl, C.: A numerical study of the topology of normally hyperbolic
invariant manifolds supporting Arnold diffusion in quasi-integrable systems. Physica D 238,
17971807 (2009)
25. Guzzo, M., Lega, E., Froeschl, C.: A numerical study of Arnold diffusion in a priori unstable
systems. Commun. Math. Phys. 290, 557576 (2009)
26. Guzzo, M., Lega, E., Froeschl, C.: First numerical investigation of a conjecture by N.N.
Nekhoroshev about stability in quasi-integrable systems. Chaos 21(3), 033101-1033101-12
(2011)
27. Hnon, M., Heiles, C.: The applicability of the third integral of motion: some numerical
experiments. Astron. J. 69, 7379 (1964)
28. Kolmogorov, A.N.: On the conservation of conditionally periodic motions under small
perturbation of the Hamiltonian. Dokl. Akad. Nauk. SSSR 98, 527530 (1954)
29. Laskar, J.: The chaotic motion of the Solar system. A numerical estimate of the size of the
chaotic zones. Icarus 88, 266291 (1990)
30. Laskar, J.: Frequency analysis for multi-dimensional systems. Global dynamics and diffusion.
Physica D 67, 257281 (1993)
31. Laskar, J., Froeschl, C., Celletti, A.: The measure of chaos by the numerical analysis of the
fundamental frequencies. Application to the standard mapping. Physica D 56, 253 (1992)
32. Lega, E., Froeschl, C.: Fast Lyapunov Indicators. Comparison with other chaos indicators.
Application to two and four dimensional maps. In: Henrard, J., Dvorak, R. (eds.) The
Dynamical Behaviour of our Planetary System. Springer, The Netherlands (1997)
33. Lega, E., Froeschl, C.: Comparison of convergence towards invariant distributions for rotation
angles, twist angles and local Lyapunov characteristic numbers. Planet. Space Sci. 46, 1525
1534 (1998)
34. Lega, E., Froeschl, C.: On the relationship between fast Lyapunov indicator and periodic
orbits for symplectic mappings. Celest. Mech. Dyn. Astron. 81, 129147 (2001)
35. Lega, E., Guzzo, M., Froeschl, C.: Physica D 182, 179187 (2003)
36. Lega, E., Froeschl, C., Guzzo, M.: Diffusion in Hamiltonian quasi-integrable systems. In:
Benest, D., Froeschl, C., Lega, E. (eds.) Topics in Gravitational Dynamics. Lecture Notes in
Physics, vol. 729. Springer, Berlin (2007)
37. Lega, E., Guzzo, M., Froeschl, C.: Measure of the exponential splitting of the homoclinic
tangle in four dimensional symplectic mappings. Celest. Mech. Dyn. Astron. 104, 191204
(2009)
38. Lega, E., Guzzo, M., Froeschl, C.: A numerical study of the size of the homoclinic tangle of
hyperbolic tori and its correlation with Arnold diffusion in Hamiltonian systems. Celest. Mech.
Dyn. Astron. 107, 129144 (2010)
39. Lega, E., Guzzo, M., Froeschl, C.: A numerical study of the hyperbolic manifolds in a priori
unstable systems. A comparison with Melnikov approximations. Celest. Mech. Dyn. Astron.
107, 115127 (2010)
40. Lega, E., Guzzo, M., Froeschl, C.: Numerical Studies of hyperbolic manifolds supporting
diffusion in symplectic mappings. Eur. Phys. J. Spec. Top. 186, 331 (2010)
41. Lega, E., Guzzo, M., Froeschl, C.: Detection of Close encounters and resonances in three
body problems through Levi-Civita regularization. Mon. Not. R. Astron. Soc. Lett. 418, 107
113 (2011)
42. Mitchenko, T.A., FerrazMello, S.: Astron. J. 122, 474481 (2001)
43. Moser, J.: On invariant curves of area-preserving maps of an annulus. Commun. Pure Appl.
Math. 11, 81114 (1958)
44. Namouni, F.: Astron. J. 130, 280 (2005)
54 E. Lega et al.

45. Namouni, F.: LNP 729, 233 (2007)

46. Namouni, F., Guzzo, M.: Celest. Mech. Dyn. Astron. 99, 31 (2007)
47. Namouni, F., Guzzo, M., Lega, E.: On the integrability of stellar motion in an accelerated
logarithmic potential. Astron. Astrophys. 489, 1363 (2008)
48. Nekhoroshev, N.N.: Exponential estimates of the stability time of near-integrable Hamiltonian
systems. Russ. Math. Surv. 32, 165 (1977)
49. Robutel, P.: Frequency map analysis and quasiperiodic decompositions. In: Benest et al. (eds.)
Hamiltonian Systems and Fourier Analysis, pp. 179198. Taylor and Francis. Adv. Astron.
Astrophys., Cambridge Sci. Publ., Cambridge (2005)
50. Robutel, P., Galern, F.: The resonant structure of Jupiters Trojan asteroids I. Long term stability
and diffusion. Mon. Not. R. Astron. Soc. 372, 14631482 (2006)
51. Robutel, P., Laskar, J.: Frequency map and global dynamics in the Solar System I. Icarus 52(1),
428 (2001)
52. Tang, X.Z., Boozer, A.H.: Finite time Lyapunov exponent and advection-diffusion equation.
Physica D 95(34), 283305 (1996)
53. Todorovic, N., Guzzo, M., Lega, E., Froeschl, C.: A numerical study of the stabilization effect
of steepness. Celest. Mech. Dyn. Astron. 110(4), 389398 (2011)
54. Villac, B.F.: Using FLI maps for preliminary spacecraft trajectory design in multi-body
environments. Celest. Mech. Dyn. Astron. 102, 2948 (2008)
55. Wayne, B.H., Malykh, A.V., Danforth, C.M.: The interplay of chaos between the terrestrial and
giant planets. Mon. Not. R. Astron. Soc. 407(3), 18591865 (2010)
Chapter 3
Theory and Applications of the Orthogonal Fast
Lyapunov Indicator (OFLI and OFLI2) Methods

Roberto Barrio

Dedicated to the Memory of Eugenio Barrio (19342014)

Abstract During the last decades the Nonlinear Dynamics field has produced a
large number of numerical techniques oriented to the analysis of the behavior of the
orbits in different systems. These methods are mainly focused to distinguish chaotic
from regular behavior. Among the variational methods, based into the variational
equations, we discuss in this paper the so-called Orthogonal Fast Lyapunov
Indicator (OFLI and OFLI2) methods that are variants of the FLI method but
designed to obtain also some information about the periodic orbits of the systems.
We review the OFLI and OFLI2 methods and we show several computational
aspects related with avoiding the appearance of spurious structures, with their use
in the analysis of regular/chaotic behaviors, but also with the analysis of periodic
orbits and regular regions, and with the efficient computation of the solution of the
variational equations by means of Taylor series methods. Finally, the methods are
shown in several Hamiltonian problems, as well as in several classical dissipative
systems, as the Lorenz and Rssler models.

3.1 Orthogonal Fast Lyapunov Indicators

When we intend to analyze the behavior of a dynamical system one of the most
interesting questions is if it is possible to know if a given initial condition generates
a chaotic orbit or not. In fact, this question cannot be answered rigorously without a
carefully theoretical study of the particular problem. Therefore, this has been done
only for some important problems (note that nowadays the computer assisted proof
of chaos is an active research field [1, 2]). Thus, a numerical evidence of the behavior
of a dynamical system has become an invaluable tool in the analysis of a problem.
One of the most popular techniques is the computation of Poincar sections,

R. Barrio ()
Computational Dynamics Group (CODY), IUMA and Departamento de Matemtica Aplicada,
Universidad de Zaragoza, E50009 Zaragoza, Spain
e-mail: [email protected]; http://cody.unizar.es

Springer-Verlag Berlin Heidelberg 2016 55

Ch. Skokos et al. (eds.), Chaos Detection and Predictability, Lecture Notes
in Physics 915, DOI 10.1007/978-3-662-48410-4_3
56 R. Barrio

which allow us to distinguish regular from chaotic orbits. However, the Poincar
sections have several drawbacks: they are useful only for two-degrees of freedom
Hamiltonian systems, we have to select carefully the two-dimensional surface that is
transverse to most of the trajectories for a fixed value of the Hamiltonian and, finally,
in the case of chaotic regions it is quite difficult to distinguish among different
structures. Another important technique is the Maximum Lyapunov Exponent
(MLE) that study the divergence among trajectories of close initial conditions. The
definition of the MLE for an initial value problem

dy
D f.t; y/; y.t0 / D y0 (3.1)
dt
is given by
1 ky.t/k
MLE D lim ln
t!C1 t ky.t0 /k
being y.t/ the solution of the first order variational equations

dy @f.t; y/
D y; y.t0 / D y0 : (3.2)
dt @y

The value of the MLE gives a way of measuring the degree of sensitivity to initial
conditions [3], and so it has been used as an indicator of chaos. The problem of
the MLE is that its practical computation is not so simple because as its definition
is a limit we have to integrate the system up to a long time and so the computer
time is large, being therefore useful just for the analysis of a short number of
orbits. It is interesting to note that the equations (3.2) are the first order sensitivity
equations with respect to the initial conditions if we take y0 D I, being I the
identity matrix (note that in (3.2) we have just selected one particular directional
derivative specified by the initial conditions y0 ). The last few decades several fast
chaos indicators have been designed to overcome the drawbacks of the MLE, among
others the frequency map analysis [4, 5], the Mean Exponential Growth factor
of Nearby Orbits (MEGNO) [6, 7], the Fast Lyapunov Indicator (FLI) [8, 9], the
Smaller ALigment Index (SALI) [10], the 0-1 test [11, 12], spike-counting diagrams
[13], kneading invariants [14], and so on (see this volume for more indicators).
Among all of the above the FLI seems to be one of the fastest and more efficient
ways of studying numerically the dynamical behavior of a set of orbits [15], but
it maintains some difficulties in locating periodic orbits and it can generate some
spurious structures.
On the present paper we intend to analyze briefly the use of two modifications of
the FLI chaos indicator, the OFLI [16] and the OFLI2 [1719]. The Fast Lyapunov
Indicator (FLI) [8] was introduced as the initial part (up to a stopping time tf ) of the
computation of the Maximum Lyapunov Exponent (MLE [3]):

FLI.y.t0 /; y.t0 /; tf / WD sup log ky.t/k;

t0 <t<tf
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 57

where y.t/ and y.t/ are the solutions of the system (3.1) and the first order
variational equations (3.2).
In order to detect easily periodic orbits, a variation called OFLI [16] (Orthogonal
Fast Lyapunov Indicator) was introduced by Fouchard, Lega, Froeschl C. and
Froeschl Ch. in 2002 and it is defined by

OFLI.y.t0 /; y.t0 /; tf / WD sup log ky? .t/k

t0 <t<tf

where y? is the component of y orthogonal to the flow at that point. The problem
is that these indicators (and most of the methods that are on the literature) may
exhibit spurious structures [19]. To minimize the appearance of these spurious
artifacts and to accelerate the detection of chaos, Barrio in 2005 developed the
OFLI2 method [1719] (also denominated OFLI2TT ). The method is based on the
use of the second order variational equations. We use as numerical ODE integrator
a specially developed Taylor method [20, 21] that gives a fast and accurate numerical
integration as we will explain below in the Appendix. The OFLI2 looks for detecting
the set of initial conditions where we may expect sensitive dependence on initial
conditions. The OFLI2 indicator at the final time tf is given by

1 2
OFLI2.y.t0 /; tf / WD sup log kfy.t/ C y.t/g? k; (3.3)
t0 <t<tf 2

where y and 2 y are the first and second order sensitivities with respect to carefully
chosen initial vectors and x? stands for the orthogonal component to the flow of the
vector x. Note that in many practical applications the computation of sensitivities is
an important task as it gives the dependence of a system with respect to the initial
conditions or parameters of the problem, that is defined by the corresponding partial
derivatives and directional derivatives. In our case we need the variational equations
up to second order and we fix the initial conditions:

dy
D f.t; y/; y.t0 / D y0 ;
dt
d y @f.t; y/ f.t0 ; y0 /
D y; y.t0 / D ; (3.4)
dt @y kf.t0 ; y0 /k
d 2 yj @fj 2 @2 fj
D y C y> 2 y; 2 y.t0 / D 0:
dt @y @y

Note that the last line of Eq. (3.4) is written for a single jth component 2 yj of the
second order variational equations to simplify the notation.
58 R. Barrio

The evolution of the FLI and OFLI indicators is explained in [9] for quasi-
integrable systems:
Proposition Given the Hamiltonian function

H .I; / D h.I/ C f .I; /; (3.5)

with action-angle variables I1 ; : : : ; In 2 R and 1 ; : : : ; n 2 S and a small

parameter, and suppose that the functions h and f satisfy the hypotheses of both
KAM and Nekhoroshev theorems then the following estimates were given:
1. If the initial conditions are on the KAM torus
 2 
@ h 
k.I .t/;  .t//k D 
 @I 2 .I.t0 // I.t 
0  t C O. t/ C O.1/;
/

for some > 0.

2. If the initial conditions are on a regular resonant motion

k.I .t/;  .t//k D kC ort I.t0 /k t C O. t/ C t O.

2 /
p 
CO. t/ C O p1 ;

with some > 0,  Zn a d-dimensional lattice that defines a resonance

through the relation  .@h=@I/ D 0 where  denotes the Euclidean projection
of a vector onto the linear space spanned by  and C a linear operator
depending on the resonant lattice  and the initial action I.t0 /.
For the OFLI2 it is possible to obtain similar estimates [17]. Again, if we assume
that the Hamiltonian (3.5) satisfies the hypotheses of the KAM theorem then, we
obtain the following estimate for some > 0 and for initial conditions on the KAM
torus

k. 2 I .t/; 2  .t/k D k. 2 I 0 .0/; 2  0 .0/k t C O. t/:

Then, we expect the OFLI and OFLI2 behave as log t [17] for initial conditions on
a KAM tori and on a regular resonant motion but with different rate of growing
(and so they grow linearly in a logarithmic time scale), tend to a constant value
for periodic orbits and grow exponentially (in a logarithmic time scale) for chaotic
orbits.
As exemplary problem, we provide some tests on the very classical Hnon
Heiles Hamiltonian system [22] given by the Hamiltonian

1 2 1
H .x; y; X; Y/ D .X C Y 2 C x2 C y2 / C x2 y  y3 (3.6)
2 3
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 59

OFLI2
8
d

6 c

b
4

a
0
0 2 4
10 10 10
time

Fig. 3.1 Evolution of the OFLI2 values of four orbits of the HnonHeiles Hamiltonian with
energy E D 1=12: a (a periodic orbit), b and c (orbits on a KAM torus) and d (a chaotic
orbit close to the hyperbolic point)

and so the differential system is

xP D X; XP D x  2x y;
yP D Y; YP D y  x2 C y2 :

The reason of using this problem is understandable: it was extensively used as a

benchmark for chaos indicators.
In Fig. 3.1 we show the time evolution of the OFLI2 values in the time interval
[1, 10,000] for four particular orbits of the HnonHeiles problem. The orbits are
indicated with the letters a (a periodic orbit), b and c (orbits on a KAM torus)
and d (a chaotic orbit close to the hyperbolic point). From the different behaviour
of the curves we observe how the indicator works and how it is easy to classify the
orbits comparing one each other.
From now on, all the biparametric figures with the OFLI2 results use the red
(white) color to point the chaotic regions and blue (black) for the most regular ones
in the color (B&W) plots, and being the intermediate colors the transition from one
to the another situation.
60 R. Barrio

3.2 Locating Periodic Orbits and Avoiding Spurious

Structures

In this section we study two of the points where the OFLI2 and OFLI techniques
give a better performance than other chaos indicators (in fact two of the reasons for
their design): in avoiding the appearance of spurious structures inside the regular
regions and in some analysis of periodic orbits and regular regions, besides the
identification of regular/chaotic behavior.

3.2.1 Avoiding Spurious Structures

One of the problems of the FLI and OFLI indicators is the dependence of the
results on the initial conditions of the variational equations. This fact was already
detected in [8], where a set of different initial tangent vectors were considered for
a symplectic mapping showing how the magnitude of the FLI indicator changes.
In any case, for symplectic mappings the most important fact is the choice of the
same tangent vector for the whole set of orbits (see [8]). In the continuous case the
dependency on the initial tangent vector is higher that in the symplectic mappings
and we have to proceed with more care, especially for short integration times. We
have to remark that these indicators give in all cases good results for a long time
analysis of an orbit but, depending on the initial conditions, the short time analysis
of a region (that is, as a global picture) may give us wrong pictures (although
the individual analysis of each orbit is correct due to the local character of the
indicators).
We have performed [17] several numerical tests with the HnonHeiles system
(3.6) to show the temporary dependence on the initial conditions of the variational
equations and we have computed the OFLI values for four sets of initial vectors,
three of them forming an orthonormal basis of the orthogonal subspace to the flow
at the initial time and the unitary initial tangent vector given by

yT .t0 / D f.t0 /=kf.t0 /k D .t1 ; t2 ; t3 ; t4 /;

yO1 .t0 / D . t2 ; t1 ; t4 ; t3 /;
yO2 .t0 / D .t4 ; t3 ; t2 ; t1 /;
yO3 .t0 / D .t3 ; t4 ; t1 ; t2 /:

We use t0 D 0 and f.t0 / D .x0 ; X0 ; y0 ; Y0 / for the HnonHeiles system.

On each picture of Fig. 3.2 we show 400  400 values of the OFLI depending
on the initial conditions of the sensitivity vector for the HnonHeiles system
with energyD 1=12 and x D 0. The stopping time in the numerical integration
is tf D 100. From the pictures we may appreciate as the whole figure presents
different configurations depending on the set of initial conditions of the sensitivity
vector. The first set, O1 , gives an acceptable figure but the other two orthogonal
vectors, O2 and O3 , generate some spurious structures (although the analysis of
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 61

each particular orbit is correct, the problem is the different time scales to reach a
global picture). The tangent initial conditions, T, gives a correct analysis but the
evolution is too slow (the values of the OFLI are of order 1017 ) and so it is difficult
to distinguish the dynamical structures. Note that with this orthonormal basis one
vector gives a good result, another one outlines a good result and two of them give
some spurious structures. We have chosen them on purpose, just to show that among
the infinity possibilities we may find good and bad ones. On the middle plots of
Fig. 3.2 we show on the pictures the evolution of the OFLI on the line x D Y D 0
(the discontinuous line on the top figures) for different initial conditions for the
sensitivity values (we plot the OFLI value at tf D 100). On the bottom pictures
we show the evolution of the OFLI values in the time interval 1; 10000 for four
particular orbits (remarked with a dotted vertical line on the pictures on the middle)
indicated with the letters a (a periodic orbit), b and c (orbits on a KAM torus) and
d (a chaotic orbit close to the hyperbolic point). We observe that the OFLI with
tangent initial conditions T gives very small values and therefore, the evolution is
much slower than in the other cases (for example the periodic orbit has not reach
yet its constant value), the OFLIO2 gives constant values also for non-periodic orbits
although it points out clearly the hyperbolic point, the OFLIO3 does not clearly
identify the periodic orbit for a low time integration and OFLIO1 seems to give
a correct information. Therefore, from these pictures it seems necessary a more
detailed study (see [19] for more details) on how to select the initial conditions of
the variational equations.
Recalling that variational methods tend to estimate the maximum Lyapunov
characteristic exponent, let us first review some results of the Lyapunov exponents
theory. In the continuous case, we have a dynamical system [23] on the state space
M defined by a diffeomorphic flow map

 t W M ! M;
y 7!  t .y/

given by an ordinary differential equation yP D f.y/ with formal solution y D

y.tI y0 / D  t .y0 / 2 M. The Lyapunov exponents are based on the solution on
the standard orthonormal basis of the linearized flow map Y that maps the tangent
space Ty0 M into T t .y0 / M and is given by the resolvent (or stability matrix) of the
matrix linear system

P D Df.t; y.t// Y.t/ WD @f.t; y.t// Y.t/;

Y.t/ Y.t0 / D Y;
@y

where Y denotes any orthonormal basis (usually we take Y D I, that is, the identity
matrix). Once we have the resolvent we have all the solutions of the variational
equations on the form y.t/ D Y.t/ y0 . The resolvent matrix may have complex
eigenvalues, soin order to simplify the stability analysisit is common to use
the singular value decomposition (SVD) of the resolvent [24]. That is, to put
Y D UDV > with U and V orthogonal matrices and D diagonal. The diagonal
 62

T O1 O2 O
3

OFLIT OFLIO OFLIO OFLIO

1 2 3
-16.5
d c b 4 d c b a 4 d c b a 4 d c b a
-17 a
-17.5
2 2 2
-18
-18.5 0 0 0
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4
coordinate y coordinate y coordinate y coordinate y
8 8 8
-10 d c
c c
6 6 6
b d d d
-15 4 b 4 4 b
a 2 2 a,b 2
-20 0 a 0 c 0 a
0 2 4 0 2 4 0 2 4 0 2 4
10 10 10 10 10 10 10 10 10 10 10 10
time time time time
Fig. 3.2 Top: OFLI values for different initial conditions for the sensitivity values (T, O1 , O2 or O3 ) for the HnonHeiles system
R. Barrio
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 63

elements of D are the square roots of the eigenvalues of the matrix Y > Y, which
is now a symmetric positive definite matrix, hence its eigenvalues are real and its
eigenvectors form an orthogonal basis. We denote by i .tI Y/ the eigenvalues of
Y > Y at time t and using the initial matrix Y. We call local or finite-time Lyapunov
exponents the real numbers

1
i .tI Y/ WD ln i .tI Y/
2t
and local or finite-time Lyapunov vectors the eigenvectors of

Y > .tI Y/ Y.tI Y/:

The matrix Y > Y may be also interpreted in a more geometrical setting as the flat
metric tensor of Eulerian space transformed to Lagrangian coordinates (for details
see [25]).
The asymptotic behavior of the finite-time Lyapunov exponents and vectors is
governed by the multiplicative ergodic theorem of Osedelec [26] that states that for
any ergodic probability measure p on the state space M D Rn and for any solution
y.t/ of the differential equation we have [27]:
1. For p-almost all v 2 Rn , there exists a finite exponent

1 kY.t/vk
 D lim ln ;
t!1 t kvk

that does not depend on the initial time and takes at most n values 1  2 
: : :  n (the Lyapunov exponents).
2. There exists the limit matrix,

1=2t
L.t0 / D lim Y > .tI Y/Y.tI Y/ :
t!1

The non-integer power of the matrix Y > Y is defined by diagonalization. Note

that this fact asses that the finite-time Lyapunov exponents i .tI Y/ will have a
limit, the Lyapunov exponents i of the orbit.
3. There exists a sequence of embedded subspaces

Sn .t0 /
Sn1 .t0 /
: : :
S1 .t0 / D Rn ;

such that on the complement Si .t0 /nSiC1 .t0 / of SiC1 .t0 / in Si .t0 / the exponential
growth (or decay) rate is i .
The eigenvectors li of the limit matrix L.t0 / are called Lyapunov vectors and
the ith one belongs to Si .t0 /nSiC1 .t0 /. The convergence of the Lyapunov vectors
is exponential, so its behaviour at finite time describes well its asymptotic limit,
64 R. Barrio

but the convergence of the Lyapunov exponents is very slow. This fact is usually
employed to obtain the MLE (the kernel of all the variational indicators). However,
problems appear when all i D 0. Now, following the ergodic theorem it is not
easy to compute for all the orbits the same Lyapunov exponent (in fact the same
finite-time Lyapunov exponent).
The key point to study some spurious patterns is to study the directions associated
with zero Lyapunov exponents.
Proposition The function V D f.t; y/ is the solution of the variational equation
(3.2) with initial conditions y0 D f.t0 ; y0 /. Moreover, if the support of the ergodic
measure p does not reduce to a fixed point then these initial conditions in the
variational equations generate a zero Lyapunov exponent.
The above Proposition [28] establishes that for any orbit at least one Lyapunov
exponent vanishes. We may enforce the above result just pointing that the solution
of the variational equations using any vector tangent to the flow will generate a
solution tangent to the flow with the same proportionality constant.
But, what happens if we work with Hamiltonian systems? Now the differential
system and the Lyapunov spectra possess a specific structure [29]. Given a 2n
degrees of freedom Hamiltonian function H the differential system is given by

yP D J rH ;

with J the skew-symmetric matrix

 
0 I
JD :
I 0

In this case the stability matrix Y is symplectic (that is, Y J Y > D J) and, for
conservative Hamiltonians, if i is a Lyapunov exponent then also i is another
one: the exponents are grouped in pairs. Therefore, as at least one Lyapunov
exponent is zero, automatically two of them are zero. Moreover [30], if H is
constant then for any solution y.t/ of (3.4) one has

d
hy.t/; rH i D 0:
dt
An important consequence of the above result is that if a solution of the variational
equation is orthogonal to rH at any time, then it will always remain orthogonal.
Also, the projection of y.t/ onto such a vector is constant. Besides, the vector rH
is a Lyapunov vector associated with a zero Lyapunov exponent [30]. In fact, any
given conserved quantity gives two zero Lyapunov exponents. Also, by Noethers
theorem, a symmetry in the dynamics implies a zero Lyapunov exponent. A com-
pletely different behavior is associated to the projection onto the tangential direction
of the flow: no answer can be given for hy.t/; f.t; y/i D hy.t/; J rH i. So, even
the first Lyapunov vector cannot be made orthogonal to the flow at every instant.
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 65

These two special directions J rH and rH are usually considered as marginal.

The displacement in the direction of J rH gives just a displacement in the
reference trajectory and the displacement in the direction of rH will give rise
to a transfer to a nearby orbit with a Hamiltonian value different from that of
the reference one. Therefore, several strategies have been designed to separate
these Lyapunov vectors from the rest [30] (in [31] a special algorithm is designed
to maintain these vectors for dissipative systems). But what happens if all the
Lyapunov exponents vanish as it happens for regular orbits? This is precisely the
situation where the Chaos Indicators based on first order variational equations seem
to have problems and it may generate spurious patterns. So, a priori, methods based
on second (or higher) order variational equations (OFLI2) are more suitable when all
the Lyapunov exponents vanish. The main reason of this better performance is just a
probabilistic approach. As it is shown below, the appearance of spurious structures
is related with the points where the initial conditions of the variational equations are
aligned close to the directions with some minimal Lyapunov exponents, and in the
case of the OFLI2, as we have two terms there is a low chance that our fixed initial
conditions point at the same time to these special directions for the first and second
order terms.
In conservative Hamiltonians with one degree of freedom the situation is quite
simple: for each orbit both Lyapunov exponents vanish. The direction tangent to
the flow generates a very low value of the variational Chaos Indicators because
for periodic orbits the ratio kf.t/k=kf.t0 /k has only small variations. Thus we may
expect that if the initial conditions of the variational equations follow that direction
only for some orbits, then we will obtain spurious patterns. The methods that use
first order variational equations (MEGNO and FLI) exhibit a spurious line, as shown
in Fig. 3.3. In order to have an initial vector y0 D .x0 ; y0 /> for the variational
equations tangent to the flow in the pendulum equations

xP D y; yP D  sin x;

we need

y0 y0 D  sin.x0 / x0 :

And so, for y0 0,

x0
y0 D  sin.x0 /:
y0

In Fig. 3.4 we show (in red online) the curve of points where the initial variation
vector .1; 1/> is tangent to the flow. This figure coincides quite well with the
observed spurious patterns in Fig. 3.3.
According to the above discussion, it seems reasonable to avoid the tangent
direction when we use first order variational equations. In Hamiltonian systems with
just one degree of freedom there is actually a single choice to avoid the tangent: the
66 R. Barrio

MEGNO FLI

velocity X
velocity X

coordinate x coordinate x
OFLI2
velocity X

coordinate x

Fig. 3.3 MEGNO and FLI plots for the pendulum problem using the vector .1; 1/> as initial
conditions of the variational equations, and OFLI2 plot

vector orthogonal to the flow, that in this case coincides with the gradient of the
Hamiltonian rH . Then the vector y evolving in time always has an orthogonal
component although a tangent one will also appear. This fact prevent us from
approaching the Lyapunov vector as fast as we would like.
Let us see how this strategy works in the HnonHeiles Hamiltonian. We
compute the FLI indicators using initial vectors rH =krH k. The results shown
in Fig. 3.5 suggest that now the FLI indicator works quite well, as OFLI2 does
automatically. Therefore, as proposed in [19], a good set of initial conditions for
the variational equations for any variational chaos indicator is given by y0 D
rH =krH k.
To summarize this problem of spurious structures we remark that, obviously,
this problem has been tackled by many other researchers using other approaches.
One option is to use random initial conditions for the variational equations (as used
in some articles for the SALI method [10]) and another option is to compute the
chaos indicator values for two different initial orthogonal vectors for the variational
equations and to use their mean value (as done for the FLI method in [32]). Both
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 67

velocity X 1

3
2 0 2
coordinate x

Fig. 3.4 FLI plot of Fig. 3.3, its contour plot and (in red) the theoretical predicted spurious pattern
for the chosen initial conditions of the variational equations .1; 1/>

options provide better results than the standard application, and in most cases they
eliminate the spurious structures. The only problem is that we may miss some of
the information inside the regular region, and in order to obtain a result that fits
well with, for instance, the skeleton of periodic orbits (with the OFLI or OFLI2
indicators), we have to use just one initial condition because we compare one value
with the ones close to it. In that situation the above approaches, OFLI2 or y0 D
rH =krH k for OFLI, provide a quite suitable option.

3.2.2 Locating Periodic Orbits

A quite important problem in studying Dynamical Systems is to locate the position

of the periodic orbits, what Poincar called the skeleton of the system. There are
several methods in literature, but a good point of the OFLI and OFLI2 methods
is that they can be used to locate approximate values of periodic orbits. Note that
for a continuous flow there always exists a differential rotation along any trajectory
that produces an increase of the tangential component to the flow of the variational
equations. Therefore the FLI and similar techniques cannot detect periodic orbits,
whereas the orthogonal versions, OFLI and OFLI2, as they cancel that component
when computing values for a periodic orbit their value reaches, after a transient
behavior, a constant value [16]. In our case, to compute the skeleton of periodic
orbits after a first scanning with the OFLI2, we have used a systematic search
approach that takes advantage of the symmetries of the system (see [33] for a
complete description). Another standard method is the use of modified Newton
68 R. Barrio

BAD FLI GOOD FLI

velocity Y

velocity Y
coordinate y coordinate y
OFLI2
velocity Y

coordinate y

Fig. 3.5 FLI plots for the HnonHeiles Hamiltonian using the gradient to the Hamiltonian as
initial conditions of the variational equations (good FLI) and another initial vector (bad FLI), and
OFLI2 plot

methods [34], that permits to obtain the initial conditions of the periodic orbits with
any desired precision.
On Fig. 3.6 we show the Poincar sections and the OFLI2 pictures at tf D 300
for the HnonHeiles problem for different values of the energy E on the surface
x D 0. For E D 1=12 most of the orbits are regular as shows the Poincar sections
and as reflects the OFLI2 picture. Note that the OFLI2 locates the separatrices
and no spurious structures are present. For E D 1=8 the OFLI2 gives much more
information than the Poincar sections and locates, without a selection of the orbits,
the periodic orbits and the chain of islands inside the chaotic sea.
In order to study with more detail the use of the OFLI and OFLI2 indicators
in the location of periodic orbits we show on the top of Fig. 3.7 the skeleton of
symmetric periodic orbits for the HnonHeiles system up to multiplicity 12 [35, 36]
on the line x D Y D 0. Note that in such a plot any point corresponds to the
initial conditions of one orbit. The plot correspond to values of the energy below the
escape region. The forbidden region is located outside the thick black line. We note
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 69

0.5
velocity Y
E: 1/12 E: 1/12

-0.5
-0.4 -0.2 0 0.2 0.4 0.6
coordinate y

E: 1/8

0.5
E: 1/8
velocity Y

-0.5
-0.4 -0.2 0 0.2 0.4 0.6
coordinate y

Fig. 3.6 Poincar surfaces of section and OFLI2 plots for the HnonHeiles system on the section
x D 0 with energy E D 1=12 and E D 1=8

the presence of the families of the normal modes 4;7;8 (the black lines originating
at E D 0) that configure the behavior for large E as the other families of periodic
orbits accumulate around them and define the boundaries of the exit basins [35].
On the bottom picture we present the values of the OFLI2 indicator showing how
the minimum values are related with the location of the periodic orbits. The three
normal modes are pointed out with solid red dots, whereas several higher order
multiplicities are pointed out with red circles. Note, as said in [16], that the OFLI and
OFLI2 indicators just can tell us that in the neighborhood of a minimum value there
should be a periodic orbit, but to locate the exact values of the periodic orbits one has
to use, as commented before, another method using these values as approximations
(for instance by applying a Newton method). In any case, to locate periodic orbits
is not easy as for resonant orbits close to a periodic orbit the OFLI and OFLI2
indicators behave temporarily as a constant value, and only after a transient time
their value increase.
70 R. Barrio

0.1667
0.16

0.14

0.12

0.1
Energy E

ion
0.08
E=1/12

reg
Forb

n
0.06

de
idde

bid
0.04
n re

For
gion

0.02

0
-0.5 -0.25 0 0.25 0.5 0.75 1
coordinate y
OFLI2

E=1/12
0
-0.4 -0.2 0 0.2 0.4
coordinate y

Fig. 3.7 Top: Skeleton of symmetric periodic orbits for the HnonHeiles system up to multiplic-
ity 12 on the line x D Y D 0. Bottom: OFLI2 for the value of the energyD 1=12 pointing out
the minimum values at the location of several periodic orbits. The three solid red dots point the
location of three normal modes

3.3 Applications: Hamiltonian Systems

Most of the chaos indicators are designed for Hamiltonian systems. In this section
we show several examples of applications of the OFLI2 indicator in different studies
[18, 3541].
As it has been noted in the previous section the OFLI2 indicator permits to
locate periodic orbits. Therefore, it is interesting to study the skeleton of periodic
orbits of a system and to compare with the OFLI2 plot in order to complete the
analysis. We present an example of this application in the Copenhagen problem
[39].
The three-body problem is one of the oldest problems in dynamical systems.
The Restricted Three-Body Problem (RTBP) supposes that the mass of one of the
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 71

three bodies is negligible. It was considered by [42] and [43], and it can serve as
an example of classical chaos. For the remaining two bodies, the case of equal
masses was first investigated by Strmgren and his colleagues of the Copenhagen
group. Its name is derived from the series of papers published by them starting in
1913.
Defining the distances to the respective primaries as:

r12 D .x C /2 C y2 ;
r22 D .x  .1  //2 C y2 ;

where  D m1 =.m1 C m2 / with m1 and m2 the masses of the two bodies, the
equations of motion of the Restricted Three-Body Problem are

xC x1C
xR  2Py D x  .1  /  ;
r13 r23
y y
yR C 2Px D y  .1  / 3   3 :
r1 r2

The only known integral of motion of this problem is the Jacobian constant

1 
C D .Px2 C yP 2 / C 2 C 2 C x2 C y2
r1 r2

that can be used to define the effective energy as EJ D C=2. Since the Copenhagen
problem is the particular case of m1 D m2 we have to take the value  D 1=2. That
is, it is the restricted planar three-body problem where the two big spherical and
homogeneous primaries have equal masses and rotate with constant angular velocity
in circular orbits around their centre of mass, while a small massless particle moves
under the resultant Newtonian action of the two primaries. On the top of Fig. 3.8 we
draw a simple picture of the problem with the two primaries in blue color. Note that
this problem has been revisited during the last 20 years (>1990) after the discovery
of many extrasolar planetary systems which are two-body systems where the two
primaries have almost equal masses.
In Fig. 3.8 we show the evolution with the energy of the OFLI2 and the
skeleton of symmetric periodic orbits for the Copenhagen problem [39], studying
the variation in the x-axis by fixing y D xP D 0 and depending on the energy EJ . We
note that these pictures give us a clear idea of the evolution of the system. When
EJ is low the system has large forbidden regions and the motion is highly regular.
Increasing EJ the system is more and more complex. The OFLI2 plot shows that the
chaotic behavior appears mainly in the range EJ 2 1:75; 0. When EJ is very high
the behavior is again more regular. The position of the two primaries are marked
with discontinuous vertical lines. The skeleton of symmetric periodic orbits is done
up to global multiplicity m D 4. On the figure we have used a color code for the
different multiplicities. The region with a great number of periodic orbits denotes
72 R. Barrio

Copenhagen r1
r2
problem

2
OFLI2
1.5

0.5
Energy EJ

0. 5

1. 5

2. 5
4 3 2 1 0 1 2
2
PO
1.5 limit
m=1
1 m=2
m=3
0.5 m=4
Energy EJ

0. 5

1. 5

2. 5
4 3 2 1 0 1 2
coordinate x

Fig. 3.8 OFLI2 plot and the skeleton of symmetric periodic orbits (PO) up to multiplicity m D
4 showing the evolution with energy (coordinate x versus energy EJ plots) for the Copenhagen
problem

also the regions with chaotic behavior (see the OFLI2 plot). Note that each point on
the curves stands for the initial conditions of a symmetric periodic orbit and each
curve is a family of periodic orbits.
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 73

Thanks to the OFLI2 indicator one is able to not only compare with the skeleton
of periodic orbits, but also to join both methods of analysis giving a much more
detailed study of how the system changes with a parameter. To that goal we show
another example [40], with an interesting generic family of Hamiltonian systems,
which is given by H D 12 .Px2 C yP 2 / C V.x; y/ with the quartic potential

1 2 1
V.x; y/ D n.x C y2 / C x2 y2 C .x4 C y4 /; (3.7)
2 4
that was proposed by Andrle [44] for a stellar system with an axis and a plane
of symmetry, and later used in many applications. This potential depends on
parameters n; ; 2 R, and it is known to be integrable for some values of the
parameters. This Hamiltonian presents the D4 symmetry, that is, it is invariant
under a rotation by =2. It has also the time-reversal symmetry. One particular and
interesting case is the dihedral potential [45] (n D 2; D 1=4; D 1)
1 2 1
V.x; y/ D .x C y2 /2  .x2 C y2 /  x2 y2 : (3.8)
4 4
This problem also appears when studying the BogdanovTakens bifurcation at
the origin. This bifurcation has interest in fluid dynamics related to convection
problems in a container such as in a magnetoconvention [46, 47] model with
a vertical magnetic field, but it can also appear in models with a salt gradient
(thermohaline convection), a Coriolis force or other stabilizing effects (see [46] for
more details). Due to this stabilizing vertical gradient, the otherwise O.2/ symmetry
of the convection problem is broken and a D4 symmetric system (assuming a square
container; other symmetries are possible in other containers) remains that depends
on a parameter. The variation of this parameter allows to study the Bogdanov
Takens codimension two bifurcation on those systems. In a limit case the system
is Hamiltonian [45] with the potential of Eq. (3.8).
Figure 3.9 shows the evolution of the dihedral potential as the energy grows. We
have combined the skeleton of periodic orbits together with several OFLI2 plots
for some values of the energy. This plot shows not only the periodic orbits, but
also the KAM tori around the periodic orbits and the chaotic regions. We can see
how the system evolves as the energy grows. For very low values of the energy
(E D 0:7; E D 0:5), the system appears to be very chaotic and it is divided in
two disconnected regions. When the energy increases to E D 0:0, the two regions
touch and they are connected at x D 0. In the OFLI2 plots we can see several
complex structures with islands due to periodic orbits (blue), separated by chaotic
regions (red). If we further increase the energy, we see at E D 2:0, E D 5:0 and E D
10:0 that the two regions are now completely connected and new structures appear
74 R. Barrio

10

6
Energy E

5
2
y
e
at

2 0
in
rd

0
oo

Coo
C

rdin 2 5
ate
x

Fig. 3.9 Several OFLI2 plots on the plane .x; y/ for different values f0:7; 0:5; 0; 2; 5; 10g of the
energy together with the skeleton of symmetric periodic orbits for the dihedral squared symmetric
Hamiltonian

and merge. As observed, there is a correspondence with the skeleton of periodic

orbits. At E D 5:0 new structures appear that evolve as the energy grows, and at
E D 10:0 those islands have merged into a bigger structure. A similar evolution
happens around the edges, where some families appear, giving rise to some new
islands. These plots show us how the system evolves when the energy changes, but
to complement this study we have to analyze the bifurcations of the system (for
more details see [40]).
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 75

3.3.1 Detecting New Dynamics: Safe Regions

Another interesting application of the fast chaos indicators, and in particular of

OFLI2 is to detect new phenomena. The option of obtaining a large number of
biparametric plots in a reasonable CPU time permits to explore large regions to
find interesting behaviours. For instance, for the HnonHeiles system it was more
or less established that just a few after the escape energy all the stable regions
disappear. In [48], after a systematic study of biparametric plots it was shown that
what disappear are just the stable regions due to KAM tori and several stable regions
appear inside the escape region. These regions are denominated safe regions in
[35, 48] as they permit to have bounded regular regions in large areas of fast escape
dynamics.
In Fig. 3.10 we show on the top an OFLI2 plot (1000  1000 points) of the
HnonHeiles system on the .y; E/ plane. Up to the escape energy Ee D 1=6 most
of the orbits are escape orbits. Thanks to a large number of OFLI2 plots it has been
possible to locate several safe regions far from the escape energy. On the bottom
plot we present a magnification of one of such regions with some key bifurcations.
In the plot we show the OFLI2 plots of just the bounded regular region on color
scale, and on the right we present schematic Poincar sections computed from the
normal form of the different bifurcations. We suppose that the bifurcation occurs at
the value of the parameter PB (in our case the energy E) and we write P D PB C".
Note that we only show one direction, from " < 0 to " > 0, but it could be the
opposite depending on the particular bifurcation. The main family of multiplicity
m D 1 is the one in black. The other families bifurcate from this on the points
marked with a circle. The main family, and in fact all the safe region, appears with a
saddle-node bifurcation (SN) and corresponds to the case where two periodic orbits
are created (or destroyed), one stable and another one unstable. This is the only way
of creating new families of periodic orbits, apart from the boundaries of the domain
of definition of the Poincar map. Another bifurcation that appear in this region
is an isochronous pitchfork bifurcation (P). It is a symmetric pitchfork bifurcation:
from a symmetric periodic orbit two new isochronous periodic orbits are created but
with fewer symmetries (the symmetry in this case is on the y-axis, not D3 ) and the
main symmetric periodic orbit changes its stability character after the bifurcation.
Besides, we plot the generic touch-and-go bifurcation, as an example of a known
generic bifurcation where an unstable periodic orbit of multiplicity m D 3 touches
[33] the center m D 1 periodic orbit and bounces while the main orbit remains
stable.
Therefore, we have shown that the use of a fast chaos indicator like OFLI/OFLI2
provides a quite useful tool in the numerical study of Hamiltonian systems.
76 R. Barrio

Ee=1/6

safe region
0.2535 < 0 = 0 > 0
SN 1

0.2534
m=1
2

m=3
m=1
Saddle-node bifurcation
0.2533 P generic

P
Energy E

0.2532
2

TG m=1
Pitchfork bifurcation
0.2531 symmetric
TG 1 1

0.253 1 1

1 1

m=3
0.2529
SN Touch-and-go bifurcation
0.11 0.105 0.1 0.095 0.09 0.085 0.08 0.075 0.07 generic
Coordinate y

Fig. 3.10 Top: OFLI2 plot of the HnonHeiles system on the .y; E/ plane (dark blue color
denotes escape orbits). Bottom: magnification of a bounded region (safe region) inside the escape
region (only the OFLI2 values of the bounded regular orbits are shown). On the left, both OFLI2
and families of periodic orbits (the main bifurcation parameters are PB .P/ D EP ' 0:25331,
PB .TG/ D ETG ' 0:25314 and PB .SN/ D ESN ' 0:25292). On the right, schematic Poincar
surface sections of some of the bifurcations involved in this region
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 77

3.4 Applications: Dissipative Systems

The fast chaos indicators may be also used in the study of dissipative systems, and
not only on Hamiltonian dynamics. As now the dynamics is quite different it is
advisable to use the fast techniques as preliminary studies to select the correct region
to later perform a more detailed analysis. In our studies [4955] we have called
this combined use BPD (Biparametric Phase Diagrams), as the global procedure is
designed mainly for biparametric plots, first using OFLI2 and later refined using
MLE to obtain the final plots. We illustrate this procedure with two paradigmatic
dissipative systems: the Lorenz and the Rssler model.
The Lorenz model [56] is one of the most classical low-dimensional chaotic
problems because it is one of the first models with the presence of chaotic behavior
and chaotic attractors.
The Lorenz model has attracted the attention of a large number of researchers
and a great number of papers continue to appear covering the model [2, 4951, 57
61]. The Lorenz model is a simplification of a more complicated model, presented
by Saltzman [62], to describe buoyancy-driven convection patterns in the classical
rectangular Rayleigh-Bnard problem applied to the thermal convection between
two plates perpendicular to the direction of the Earths gravitational force. After
several simplifications, Lorenz arrived at his famous equations:

dx dy dz
D  x C  y; D xz C r x  y; D xy  b z; (3.9)
dt dt dt
where t is a dimensionless time, and  (the Prandtl number), r and b are three
dimensionless control parameters.
In [49], the authors made an extensive numerical study of the Lorenz model
based of the OFLI2 chaos indicator. From the numerical tests, they conjectured
that the region of parameters where the Lorenz model is chaotic is bounded for
fixed r. We have determined the complete parametric chaotic region for the classical
Lorenz system [49, 50]. In Fig. 3.11 we show some of the findings from [4951].
The three plots Fig. 3.11a1, a2 and a3 are done first with the OFLI2 Chaos Indicator
(to locate the interesting values of parameters) and later with the Maximum
Lyapunov Exponent (to provide a more standard output) using .x.0/; y.0/; z.0// D
.60; 60; 10/ as the initial conditions and fixing the third parameter (the one that is not
on the biparametric plot) at the Saltzman values .r; b; / D .28; 8=3; 10/. The white
color is associated with chaotic behavior, whereas the black color is associated with
regular behavior. The biparametric plot -a1.a- is a magnification of a1, and on the top
we present a MLE curve at .b; / D .8=3; 10/. One of the remarkable properties of
Fig. 3.11 is that in the .; b/ plane (plot -a2-) the chaotic region seems to be bounded.
In [50] such a conjecture was proved, establishing that the Lorenz system (3.9) has,
for any given fixed parameter value r > 1, a chaotic region bounded in b, and if
b  > 0 then the region is bounded in  too (in fact, outside a bounded region
every positive semiorbit of the Lorenz system converges to an equilibrium).
78 R. Barrio

a1

3
2 b=8/3
MLE

1
0

-1

a1.a

b=8/3
parameter b

parameter r

a2
parameter

a3

parameter r

Fig. 3.11 OFLI2 and MLE (BPD) plots for the Lorenz model depending on the parameters of the
system (black color denotes regular behavior and white chaotic behavior). Panel -a1- presents the
BPD plot for the .r; b/ parametric plane (the panel -a1.a- depicts a magnification and the top plot
shows the MLE in the straight line b D 8=3), -a2- the .; b/ plane and -a3- the .r;  / plane
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 79

100

b 50

800
0 400
100 200 300 400 500 0
r

Fig. 3.12 Three-dimensional parametric evolution of the Lorenz model giving a complete study
of the chaotic region

Moreover, joining several BPD (MLE and OFLI2 plots) we have been able to
provide a complete three-dimensional parametric evolution of the Lorenz model
giving a detailed study of the complete chaotic region [49, 50]. This is shown in
Fig. 3.12.
Another low dimensional paradigmatic problem that have been frequently
studied is the Rssler model [63]. The Rssler equations are given by

xP D .y C z/;
yP D x C ay; (3.10)
zP D b C z.x  c/;

with a; b; c 2 R, and they are assumed to be positive and dimensionless. This model
is a famous prototype of a continuous dynamical system exhibiting chaotic behavior
with minimum ingredients.
The Rssler system is not always dissipative as the divergence is given by
divff.x; y; z/g D a  c C x. Thus, in a large region of parameters (especially when a
grows) and for large values of the variable x, we will have a positive divergence and
so escape orbits. Therefore, apart from regular and chaotic orbits, the Rssler system
also has escape orbits with transient chaos or regular behavior before escaping.
This fact makes more difficult a theoretical and numerical analysis of this problem.
Note that there may exist escape trajectories when the divergence of the system is
negative: the initial volume eventually shrinks along the target trajectory, but the
trajectory goes to infinity without bounds. But in the dissipative case escape orbits
will not occur in large sets in the phase space as it happens in the positive divergence
case, where this is the normal behavior.
In Fig 3.13 we present some BPD diagrams on the .c; b/ plane for different values
of the parameter a. The blue color is associated with regular behavior, the red color
with chaotic one and white color denotes escape orbits. The plots show us a pattern
structure that is repeated when the parameter c grows, that consist on interlacing
bands in groups of two. Also different bands of regular motion appear inside the
chaotic region [52]. When the parameter a grows, more and more of such pair of
80 R. Barrio

3.5 40
a=0.1 a=0.2
3
30
2.5

b 2 b 20
1.5
1
10
0.5
0 0
0 50 100 150 200 250 0 50 100 150 200 250
c c
120 200
a=0.3
a=0.35
160
80
120
b b
80
40
40

0 0
0 50 100 150 200 250 0 50 100 150 200 250
c c

Fig. 3.13 BPD diagrams of the Rssler system on the .c; b/ plane for different values of the
parameter a

structures appear and the escape region grows in size, becoming the dominating
behavior.
In the Rssler model, when one plots a BPD in the parametric plane .a; c/ it
is possible to observe several spiral structures [52, 64, 65] formed by branches
of regular and chaotic orbits. These structures configure the global parametric
organization of the Rssler system, and in fact of any dissipative system with
Shilnikov saddle-foci. Therefore, thanks to the fast chaos indicator OFLI2 and the
MLE (BPD) we can easily detect these interesting spirals. The remaining question
is: how are they formed? To answer that question in [53, 54] an extensive use of
BPD, combined with bifurcation analysis, has been done.
Figure 3.14 outlines the theoretical skeleton [53, 54] of the bifurcation unfold-
ing around the spiral structure. The top picture sketches phenomenologically a
caricature of the bifurcation structure of the spiral structure along with shrimps
(also called swallows, periodic connecting windows, crossroads, : : : in literature).
In it, the saddle-node bifurcation curves originating from the B point (a Belyakov
point) demarcate the boundaries of shrimps near the spiral. Indeed, the spiral can
generate an infinite chain of shrimps. A BPD zoom of the Rssler bifurcation
diagram in the bottom picture depicts a few such shrimps, S2j and S2j1 ,
which are singled out by the saddle-node curves (solid thick red curves). The
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 81

regular
chaotic
S2j-
1
B-point
S2j+
1

theoretical Homoclinic bif.

scenario S 2j

BPD
+
bifurcations

25

S2j-
20 1

15 S2j+1
c
10
S2j

0.1 0.12 0.14 0.16 0.18 0.2 0.22

a
Fig. 3.14 Outline of the spiral structures in dissipative systems with Shilnikov saddle-foci (gray
stands for regular motion, whereas light gray stands for chaotic motion). Top: Phenomenological
sketch of the spiral structure formed by the shrimps. Bottom: magnification of the bifurcation
portrait of the spiral hub, overlaid with the homoclinic bifurcation curve (black) and the principal
folded (thick red) and cusp-shaped (thin blue) bifurcation curves setting the boundaries for largest
shrimps in the Rssler model

cusp-shaped saddle-node bifurcation curves (light thin blue curves) join the succes-
sive shrimps. Thus, both fold- and cusp-shaped bifurcation curves of saddle-node
periodic orbits determine the local structure of the hub and the shrimps. The latter
serve as connection centers between hubs that contribute toward the formation of
characteristic spiral structures in the bifurcation diagram of the system.
This generic scenario explains the formation of the spiral structures and
shrimps in the biparameter space of a system with a Shilnikov saddle-focus
[53, 54]. The skeleton of the structure is due to fold- and cusp-shaped bifurcation
82 R. Barrio

curves of saddle-node periodic orbits that accompany the homoclinics of the saddle-
focus. These bifurcation curves distinctively shape the shrimps zones in the
vicinity of the spiral hub. Again, a massive use of the OFLI2 indicator has permitted
to study in detail such a new phenomena.

3.5 Conclusions

In the literature there are a large plethora of fast chaos indicators. Among them,
the OFLI and OFLI2 (Orthogonal Fast Lyapunov Indicators) provide with some
of the fastest methods, but they also permit to locate periodic orbits in Hamiltonian
systems as shown in this paper. Besides, we have given a detailed analysis of the way
to minimize the appearance of spurious structures in generic variational indicator
methods (choosing as initial conditions of the variational equations the direction of
the gradient of the Hamiltonian function). This problem is automatically avoided
by the OFLI2 method. The combined used of OFLI/OFLI2 with other techniques,
as skeletons of periodic orbits, bifurcation analysis and so on, gives rise to a quite
powerful numerical study of Hamiltonian and dissipative dynamical systems. This
fact has been shown by presenting several new phenomena detected in Hamiltonian
dynamics (new insights in Copenhagen and dihedral problems, and location of
safe regions in open systems) and dissipative systems (statement of boundness
of chaotic region in Lorenz system and the location and study of spiral structures in
systems with a Shilnikov saddle-focus). Finally, the efficient numerical integration
of the flow and the automatic determination of the partial derivatives of the solution
is performed by using the Taylor series method (TIDES software). This numerical
ODE solver provides with the most efficient numerical method for high-precision
requirements.

Appendix: Adaptive ODE IntegratorTaylor Series Method

A basic feature in using any chaos indicator is to obtain numerically solutions of

ODE systems and variational equations. In Dynamical Systems there are also more
requirements, for example, in the process of determination of periodic orbits we
obviously have to integrate the differential system, normally for a short time, with
very high precision, especially for highly unstable periodic orbits. Moreover, in the
study of the bifurcations and stability of periodic orbits we also have to integrate
the first order variational equations using as initial conditions the identity matrix as
also occurs in the use of chaos indicators. To reach this goal we may, obviously, use
any numerical ODE method like, for example, a RungeKutta method. The last few
years, in the computational dynamics community [66] one of the preferred methods
is the Taylor series method.
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 83

The Taylor method is one of the oldest numerical methods for solving ordinary
differential equations but it is scarcely used in the numerical analysis community.
Its formulation is quite simple [67]. Let us consider the initial value problem yP D
f.t; y/. Now, the value of the solution at ti (that is, y.ti /) is approximated by yi from
the nth degree Taylor series of y.t/ at t D ti (the function f has to be a smooth
function). So, denoting hi D ti  ti1 ,

y.t0 / DW y0 ;
1 df.ti1 ; yi1 / 2
y.ti / ' yi1 C f.ti1 ; yi1 / hi C hi C : : :
2 dt
1 dn1 f.ti1 ; yi1 / n
::: C hi DW yi :
n dtn1
Therefore, the problem is reduced to the determination of the Taylor coefficients
f1=. j C 1/ dj f=dtj g. This may be done quite efficiently by means of the automatic
differentiation (AD) techniques (for more details see [20]). Note that the Taylor
method has several good features; one of them is that it gives directly a dense
output in the form of a power series being therefore quite useful when an event
location criteria may be used (as in the computation of Poincar sections), it can be
formulated as an interval method giving guaranteed integration methods (used, by
instance, in the computer assisted proof of chaos [1] and skeletons of periodic orbits
[68]), Taylor methods may manage directly high order differential equations just
taking into account that the Taylor coefficients for the solution and its derivatives
are evidently related, Taylor methods of degree n are also of order n and so Taylor
methods of high degree give us numerical methods of high order (therefore, they are
very useful for high-precision solution of ODEs, as needed, for example, in some
fine studies in dynamical systems [69] and in the computation of unstable periodic
orbits [34, 70]).
Just as a short look at the practical implementation of the Taylor series method
we remark that in the literature there are efficient variable-stepsize variable-order
(VSVO) formulations. For example, in [20, 71, 72] the variable-stepsize formulation
is based on the error estimator using the last two coefficients and gives the following
stepsize prediction
8 ! n1
1 ! 1n 9
< Tol Tol =
hiC1 D fac  min ;
: kf .n1/ f.n2/ .ti /k1
1 1 .n1/
k n f .ti /k1 ;

where fac is a safety factor and Tol the user error tolerance. A very simple order
selection that only depends on the user error tolerance is given [73] by the formula
n.Tol/ D  12 ln Tol. See [20, 71] for a more extensive analysis and comparison
with variable-stepsize variable-order formulations of the Taylor method. In Fig. 3.15
we present some comparisons on the HnonHeiles problem with initial conditions
.x0 ; y0 ; X0 ; Y0 / D .0; 0:52; 0:371956090598519; 0/ and E D 0:157494996 in
the time interval 0; 200 using the Taylor method (software TIDES [72]) and
84 R. Barrio

102
DP dop853 QP dop853
odex odex
102 TIDES TIDES
101
CPU time

CPU time
100

103
101

1010 105 1030 1020 1010

Relative error Relative error
100
MP
80
CPU time

60

40

20
TIDES
0
100 200 300 400 500
Log10(Relative error)

Fig. 3.15 Comparison of the CPU time in seconds vs. relative error in the numerical integration
of a KAM orbit of the HnonHeiles problem using two well established codes, dop853 (explicit
RungeKutta method) and odex (extrapolation method), and the Taylor series method (TIDES
code) with variable-order and variable-stepsize in double-precision (DP), quadruple-precision (QP)
and multiple-precision (MP)

the well established codes dop853 and odex developed by Hairer and Wanner
[74]. These codes are based on an explicit RungeKutta of order 8(5,3) given by
Dormand and Prince with stepsize control and dense output and the extrapolation
method, respectively. All the methods are compared only in double and quadruple
precision using the Lahey LF 95 compiler (fortran) because the dop853 cannot
be directly used in multiple precision. The multiple-precision tests are done using
C++ and the GMP and MPFR [75] multiple precision packages. From Fig. 3.15 we
note that for low precision the dop853 code is a bit faster but when the precision
demands are increased the Taylor method is by far the fastest, being for very high
precision the only reliable method. Moreover, we can appreciate the different slope
of the variable order method (Taylor method) and the fixed order one (dop853),
being clear that for high precision the variable order schemes become the more
competitive because they are more versatile.
For the computation of the OFLI and OFLI2 we are interested not only in the
differential equations but also in the variational equations. In order to avoid their
explicit generation we have devised [21] an alternative that permits us to obtain the
solution of the variational equations without computing them explicitly. Therefore,
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 85

we have to obtain a numerical solution of y.t/ and Ly.t0 / y.t/, being Ly.t0 / y.t/ the
Lie derivative of the solution y.t/ with respect to the vector y.t0 / (that is, in this
case the directional derivative). Note that the partial derivatives of the solution with
respect to the initial conditions are given by
 
D Le1 y.t/ j Le2 y.t/ j : : : j Len y.t/

with .e1 ; e2 ; : : : ; en / the canonical base of Rn .

The Taylor series method computes the Taylor series of the solution of the
differential equation and the Taylor series of the partial derivatives of the solution
@y.ti /
y.ti / D  y.t0 / D Ly.t0 / y.ti /
@y.t0 /
1
' Ly.t0 / y.ti1 / C Ly.t0 / f.ti1 / hi C Ly.t0 / f.2/ .ti1 / h2i
2
1
C::: C Ly.t0 / f.n1/ .ti1 / hni :
n
We now may compute the coefficients 1=. j C 1/ Ly.t0 / f. j/ .ti1 / by rules of auto-
matic differentiation of the elementary functions (, , =, ln, sin, : : :) obtained in
[21]. Automatic differentiation gives a recursive procedure to obtain the numerical
value of the reiterated derivatives of the elementary functions at a given point. We
present here, as example, the rules for the sum, product by a constant, product,
division and real power of functions (see [21] for the complete list of rules of any
elementary operation):
Proposition If f .t; y.t//; g.t; y.t// W .t; y/ 2 RsC1 7! R are functions of class C n
and given a vector v 2 Rs , we denote

1 dj f .t/ 1
f j; 0 WD ; f j; 1 WD Lv f . j/ ;
j dtj j

that is, the jth Taylor coefficient of the function f .t; y.t// and of its Lie derivative
with respect to v, respectively. Then, we have
(i) If h.t/ D f .t/ g.t/ then hn; i D f n; i gn; i .
(ii) If h.t/ D f .t/ with 2 R then hn; i D f n; i .
(iii) If h.t/ D f .t/  g.t/ then

X
n
h n; 0
D f nj; 0  g j; 0 ;
jD0
X
n
 
h n; 1
D f nj; 0  g j; 1 C f nj; 1  g j; 0 :
jD0
86 R. Barrio

(iv) If h.t/ D f .t/=g.t/ then

 X
n1 
n; 0 1 n; 0 j; 0 nj; 0
h D f  h f ;
g0;0
 jD0
1
hn; 1 D 0;0 f n; 1  hn; 0  f 0; 1
g
X 
n1
 j; 0 nj; 1 j; 1 nj; 0

 h f Ch f :
jD0

(v) If h.t/ D f .t/ with 2 R and f 0;0 6D 0, then

h0; 0 D . f 0;0 .t// ;

1 X 
n1
h n; 0
D 0;0 n  j. C 1/ h j; 0  f nj; 0 ;
nf jD0
1
h 0; 1
D 0;0 h0; 0  f 0; 1 ;
f 
1
hn; 1 D 0;0  nhn; 0  f 0; 1
nf
X 
n1
  
C n  j. C 1/ h j; 0  f nj; 1 C h j; 1  f nj; 0 :
jD0

The use of high-precision numerical integrators in the determination of periodic

orbits is justified, for instance, by the search of highly unstable periodic orbits [34].
In Fig. 3.16 we show some comparisons for the Lorenz model (3.9) in double
precision, all obtained with the code TIDES using the traditional way to compute
the solution of the variational equations (VAR), that is writing them explicitly, and
with the use of the extended Taylor series method (ETS) and using TIDES with this
capability (using the extended Automatic Differentiation rules of Proposition 3.5).
In the pictures we present computational relative error vs. CPU time diagrams
in seconds. The extended Taylor series method is the fastest option with a low
difference, but the most important thing is that the difference in the formulation
is very high. Everyone knows how cumbersome is to write variational equations
of order one, two and higher!! The picture is done for computing the complete
order two and just the partial derivative @2 x=@x20 . Note that TIDES can compute
also sensitivities with respect to parameters of a system, not only with respect to the
initial conditions.
To end this section we remark that the use of the Taylor series method is currently
helped by the free available new state-of-the-art numerical library TIDES (Taylor
Integrator of Differential EquationS) that has just been developed by Profs. Abad,
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 87

2
x/ x20: ETS
2
10
2
x/ x2: VAR
0
Full Order 2: ETS
CPU time Full Order 2: VAR

3
10

14 10 5 2
10 10 10 10
Relative error

Fig. 3.16 Computational relative error in the computation of sensitivities vs. CPU time diagrams
in seconds for the Lorenz model in double-precision using TIDES code using the extended Taylor
series method for the solution of the variational equations (ETS) or just the standard Taylor series
method with explicit formulation of the variational equations (VAR)

Barrio, Blesa and Rodrguez [72, 76]. The reader can contact the authors to obtain
the software.1
Nowadays it is quite standard to preserve several geometric properties of the
differential systems by means of geometric integrators. This kind of methods are
specially useful when we want to solve a problem with not very high precision
but with a constant value of the energy, for instance. The problem for very long
numerical integrations is that it doesnt matter how you perform the integration,
finally the rounding errors of the computer will affect the integration, giving an
increment of the error in the geometric object [77]. The optimal error in these
quantities was studied first by Brouwer [78], who established that the error in
energy grows at least as O.t1=2 /. This error is obtained for long integrations
of careful used symplectic integrators [77] or when one is able to suppress the
truncation error in any numerical integrator (and also with a careful use, of
course). In other circumstances we may observe a typical linear growing O.t/.
In the case of the positions, we will have a root-mean-squared (RMS) error
O.t3=2 / in the best case, and a typical error O.t2 / (as in any non-symplectic RK
code).

1
TIDES: a Taylor series Integrator for Differential EquationS (GNU free software). Webpage:
http://cody.unizar.es/software.html and http://sourceforge.net/projects/tidesodes/.
88 R. Barrio

Now, we just show how easy is to eliminate the truncation error in the Taylor
series method, and so, in TIDES. The advantage is that using the error estimator
of the Taylor method, as they are based in just studying some Taylor coefficients,
we may use any tolerance level. A completely different situation occurs when our
error estimator is based on the substraction of two similar expressions (like in some
formulations of embedded RungeKutta pairs where the local error estimator is
given by the substraction of the solution of two methods of different order) that
makes impossible to use them for tolerances lower than the rounding error due to the
catastrophic digit cancelation. So, if we fix the tolerance far below the roundoff
unit of the computer we, in theory, can control the truncation error. This technique
has been used previously by the group of Carles Sim [79] and by others [8082].
We have to combine this technique with a compensated sum formulation [83] of
the time increment as we use variable stepsize strategies (in contrast with symplectic
integrators that have to use fixed stepsize implementations). So, the truncation-free
formulation can be described as:
(use TOL  u, with u D the roundoff unit) C (compensated sum)
TIDES uses compensated sum in some stages of the method, so, if we want to
preserve some geometric properties of the systems we just have to fix a low enough
tolerance level. Obviously, this approach is computationally more expensive than
other approaches and it is valid only if you also look for high precision numerical
results.
In Fig. 3.17 we present the evolution of the error using TIDES with the truncated-
free formulation. It is clear that this approach permits to achieve the optimal
Brouwers law (see Fig. 3.17), like well-programmed symplectic integrators [77],
but it can be used in variable-stepsize formulations being therefore a quite flexible
approach.

TIDES: Truncation-free formulation

O(t3/2 )
10 -10 O(t1/2 ) error in position

error in Energy
10 -15

10 4 10 5 10 6
time

Fig. 3.17 Evolution of the error in the position and in the Energy of one orbit of the Hnon
Heiles system using a tolerance lower than the unit roundoff of the computer (truncation-free
formulation). For the position we show the evolution of the error for the x and X variables of the
Hamiltonian (3.6)
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 89

Acknowledgements The author thanks his colleagues and friends Dr. Fernando Blesa and Dr.
Sergio Serrano for many interesting discussions and common work on this subject. The author
thanks the referees for their help in improving the paper. The author has been supported during this
research by the Spanish Research Grant MTM2012-31883, MTM2015-64095-P and by Gobierno
de Aragon and Fondo Social Europeo.

References

1. Galias, Z., Zgliczynski, P.: Computer assisted proof of chaos in the Lorenz equations. Physica
D 115, 165188 (1998)
2. Tucker, W.: A rigorous ODE solver and Smales 14th problem. Found. Comput. Math. 2, 53
117 (2002)
3. Skokos, C.: The Lyapunov characteristic exponents and their computation. In: Souchay, J.J.,
Dvorak, R. (eds.) Dynamics of Small Solar System Bodies and Exoplanets. Lecture Notes in
Physics, vol. 790, pp. 63135. Springer, Berlin, Heidelberg (2010)
4. Laskar, J.: Frequency analysis for multi-dimensional systems. Global dynamics and diffusion.
Physica D 67, 257281 (1993)
5. Laskar, J.: Frequency analysis of a dynamical system. Celest. Mech. Dyn. Astron. 56, 191196
(1993)
6. Cincotta, P.M., Sim, C.: Simple tools to study global dynamics in non-axisymmetric galactic
potentials - I. Astron. Astrophys. Suppl. 147, 205228 (2000)
7. Cincotta, P.M., Giordano, C.M., Sim, C.: Phase space structure of multi-dimensional systems
by means of the mean exponential growth factor of nearby orbits. Physica D 182, 151178
(2003)
8. Froeschl, C., Lega, E.: On the structure of symplectic mappings. The fast Lyapunov indicator:
a very sensitivity tool. Celest. Mech. Dyn. Astron. 78, 167195 (2000)
9. Guzzo, M., Lega, E., Froeschl, C.: On the numerical detection of the effective stability of
chaotic motions in quasi-integrable systems. Physica D 163, 125 (2002)
10. Skokos, C., Antonopoulos, C., Bountis, T.C., Vrahatis, M.N.: Detecting order and chaos in
Hamiltonian systems by the SALI method. J. Phys. A 37, 62696284 (2004)
11. Gottwald, G.A., Melbourne, I.: A new test for chaos in deterministic systems. Proc. R. Soc.
Lond. Ser. A Math. Phys. Eng. Sci. 460, 603611 (2004)
12. Gottwald, G.A., Melbourne, I.: Testing for chaos in deterministic systems with noise. Physica
D 212, 100110 (2005)
13. Barrio, R., Shilnikov, A.: Parameter-sweeping techniques for temporal dynamics of neuronal
systems: case study of Hindmarsh-Rose model. J. Math. Neurosci. 1(6), 20 (2011)
14. Barrio, R., Shilnikov, A., Shilnikov, L.: Kneadings, symbolic dynamics and painting Lorenz
chaos. Int. J. Bifurcat. Chaos 22, 1230016, 24 (2012)
15. Darriba, L.A., Maffione, N.P., Cincotta, P.M., Giordano, C.M.: Comparative study of varia-
tional chaos indicators and ODEs numerical integrators. Int. J. Bifurcat. Chaos 22, 1230033,
33 (2012)
16. Fouchard, M., Lega, E., Froeschl, C., Froeschl, C.: On the relationship between fast
Lyapunov indicator and periodic orbits for continuous flows. Celest. Mech. Dyn. Astron. 83,
205222 (2002)
17. Barrio, R.: Sensitivity tools vs. Poincar sections. Chaos Solitons Fractals 25, 711726 (2005)
18. Barrio, R.: Painting chaos: a gallery of sensitivity plots of classical problems. Int. J. Bifurcat.
Chaos 16, 27772798 (2006)
19. Barrio, R., Borczyk, W., Breiter, S.: Spurious structures and chaos indicators. Chaos Solitons
Fractals 40(4), 16971714 (2007)
20. Barrio, R., Blesa, F., Lara, M.: VSVO formulation of the Taylor method for the numerical
solution of ODEs. Comput. Math. Appl. 50, 93111 (2005)
90 R. Barrio

21. Barrio, R.: Sensitivity analysis of ODEs/DAEs using the Taylor series method. SIAM J. Sci.
Comput. 27, 19291947 (2006)
22. Hnon, M., Heiles, C.: The applicability of the third integral of motion: some numerical
experiments. Astron. J. 69, 7379 (1964)
23. Geist, K., Parlitz, U., Lauterborn, W.: Comparison of different methods for computing
Lyapunov exponents. Prog. Theor. Phys. 83, 875893 (1990)
24. Dieci, L., Van Vleck, E.S.: Lyapunov and other spectra: a survey. In: Collected Lectures on the
Preservation of Stability Under Discretization (Fort Collins, CO, 2001), pp. 197218. SIAM,
Philadelphia, PA (2002)
25. Thiffeault, J.L., Boozer, A.H.: Geometrical constraints on finite-time Lyapunov exponents in
two and three dimensions. Chaos 11, 1628 (2001)
26. Osedelec, V.I.: A multiplicative ergodic theorem. Lyapunov characteristic numbers for dynam-
ical systems. Trans. Moscow Math. Soc. 19, 197231 (1968)
27. Legras, B., Vautard, R.: A guide to Liapunov vectors. In: Palmer, T. (ed.) Predictability. ECWF
Seminar. vol. I, pp. 135146. ECMWF, Reading (1996)
28. Haken, H.: At least one Lyapunov exponent vanishes if the trajectory of an attractor does not
contain a fixed point. Phys. Lett. A 94, 7172 (1983)
29. Meyer, H.D.: Theory of the Liapunov exponents of Hamiltonian systems and a numerical study
on the transition from regular to irregular classical motion. J. Chem. Phys. 84, 31473161
(1986)
30. Yamaguchi, Y.Y., Iwai, T.: Geometric approach to Lyapunov analysis in Hamiltonian dynamics.
Phys. Rev. E (3) 64, 066206, 16 (2001)
31. Grond, F., Diebner, H.H., Sahle, S., Mathias, A., Fischer, S., Rossler, O.E.: A robust, locally
interpretable algorithm for Lyapunov exponents. Chaos Solitons Fractals 16, 841852 (2003)
32. Guzzo, M., Lega, E.: The numerical detection of the Arnold web and its use for long-term
diffusion studies in conservative and weakly dissipative systems. Chaos 23, 023124 (2013)
33. Barrio, R., Blesa, F.: Systematic search of symmetric periodic orbits in 2DOF Hamiltonian
systems. Chaos Solitons Fractals 41, 560582 (2009)
34. Abad, A., Barrio, R., Dena, A.: Computing periodic orbits with arbitrary precision. Phys. Rev.
E 84, 016701 (2011)
35. Barrio, R., Blesa, F., Serrano, S.: Fractal structures in the Hnon-Heiles Hamiltonian.
Europhys. Lett. 82, 10003 (2008)
36. Barrio, R., Blesa, F., Serrano, S.: Bifurcations and chaos in Hamiltonian systems. Int. J.
Bifurcat. Chaos 20, 12931319 (2010)
37. Barrio, R., Blesa, F., Serrano, S.: Qualitative analysis of the .N C 1/-body ring problem. Chaos
Solitons Fractals 36, 10671088 (2008)
38. Barrio, R., Blesa, F., Elipe, A.: On the use of chaos indicators in rigid-body motion. J.
Astronaut. Sci. 54, 359368 (2006)
39. Barrio, R., Blesa, F., Serrano, S.: Periodic, escape and chaotic orbits in the Copenhagen and the
.n C 1/-body ring problems. Commun. Nonlinear Sci. Numer. Simul. 14, 22292238 (2009)
40. Blesa, F., Piasecki, S., Dena, ., Barrio, R.: Connecting symmetric and asymmetric families
of periodic orbits in squared symmetric Hamiltonians. Int. J. Mod. Phys. C 23, 1250014, 22
(2012)
41. Blesa, F., Seoane, J.M., Barrio, R., Sanjun, M.A.F.: To escape or not to escape, that is the
questionperturbing the Hnon-Heiles Hamiltonian. Int. J. Bifurcat. Chaos 22, 1230010, 9
(2012)
42. Euler, L.: Theoria motuum lunae (E418) Petrop. Parisin. Lond. (1772)
43. Jacobi, C.G.J.: Comp. Rend. 3, 5961 (1836)
44. Andrle, P.: A third integral of motion in a system with a potential of the fourth degree. Bull.
Astron. Inst. Czechoslov. 17, 169175 (1966)
45. Armbruster, D., Guckenheimer, J., Kim, S.h.: Chaotic dynamics in systems with square
symmetry. Phys. Lett. A 140, 416420 (1989)
46. Rucklidge, A.: Global bifurcations in the Takens-Bogdanov normal form with D4 symmetry
near the O(2) limit. Phys. Lett. A 284, 99111 (2001)
3 Orthogonal Fast Lyapunov Indicators (OFLI and OFLI2) 91

47. Proctor, M.R.E., Weiss, N.O.: Magnetoconvection. Rep. Prog. Phys. 45, 1317 (1982)
48. Barrio, R., Blesa, F., Serrano, S.: Bifurcations and safe regions in open Hamiltonians. New J.
Phys. 11, 053004 (2009)
49. Barrio, R., Serrano, S.: A three-parametric study of the Lorenz model. Physica D 229, 4351
(2007)
50. Barrio, R., Serrano, S.: Bounds for the chaotic region in the Lorenz model. Physica D 238,
16151624 (2009)
51. Barrio, R., Blesa, F., Serrano, S.: Behavior patterns in multiparametric dynamical systems:
Lorenz model. Int. J. Bifurcat. Chaos 22, 1230019, 14 (2012)
52. Barrio, R., Serrano, S.: Qualitative analysis of the Rssler equations: bifurcations of limit
cycles and chaotic attractors. Physica D 238, 10871100 (2009)
53. Barrio, R., Blesa, F., Serrano, S., Shilnikov, A.: Global organization of spiral structures in
biparameter space of dissipative systems with Shilnikov saddle-foci. Phys. Rev. E 84, 035201
(2011)
54. Barrio, R., Blesa, F., Serrano, S.: Topological changes in periodicity hubs of dissipative
systems. Phys. Rev. Lett. 108, 214102 (2012)
55. Serrano, S., Barrio, R., Dena, A., Rodrguez, M.: Crisis curves in nonlinear business cycles.
Commun. Nonlinear Sci. Numer. Simul. 17, 788794 (2012)
56. Lorenz, E.: Deterministic nonperiodic flow. J. Atmos. Sci. 20, 130141 (1963)
57. Doedel, E.J., Krauskopf, B., Osinga, H.M.: Global bifurcations of the Lorenz manifold.
Nonlinearity 19, 29472972 (2006)
58. Dullin, H.R., Schmidt, S., Richter, P.H., Grossmann, S.K.: Extended phase diagram of the
Lorenz model. Int. J. Bifurcat. Chaos 17, 30133033 (2007)
59. Guckenheimer, J., Holmes, P.: Nonlinear Oscillations, Dynamical Systems, and Bifurcations
of Vector Fields. Applied Mathematical Sciences, vol. 42. Springer, New York (1990)
60. Shil0 nikov, A.L., Shil0 nikov, L.P., Turaev, D.V.: Normal forms and Lorenz attractors. Int. J.
Bifurcat. Chaos 3, 11231139 (1993)
61. Viana, M.: Whats new on Lorenz strange attractors? Math. Intelligencer 22, 619 (2000)
62. Saltzman, B.: Finite amplitude free convection as an initial value problem - 1. J. Atmos. Sci.
19, 329341 (1962)
63. Rssler, O.E.: An equation for continuous chaos. Phys. Lett. A 57, 397398 (1976)
64. Gaspard, P., Kapral, R., Nicolis, G.: Bifurcation phenomena near homoclinic systems: a two-
parameter analysis. J. Stat. Phys. 35, 697727 (1984)
65. Gallas, J.A.C.: The structure of infinite periodic and chaotic hub cascades in phase diagrams
of simple autonomous flows. Int. J. Bifurcat. Chaos 20, 197211 (2010)
66. Guckenheimer, J., Meloon, B.: Computing periodic orbits and their bifurcations with automatic
differentiation. SIAM J. Sci. Comput. 22, 951985 (2000)
67. Corliss, G., Chang, Y.F.: Solving ordinary differential equations using Taylor series. ACM
Trans. Math. Softw. 8, 114144 (1982)
68. Barrio, R., Rodrguez, M., Blesa, F.: Computer-assisted proof of skeletons of periodic orbits.
Comput. Phys. Commun. 183, 8085 (2012)
69. Sim, C.: Dynamical systems, numerical experiments and super-computing. Mem. Real Acad.
Cienc. Artes Barcelona 61, 336 (2003)
70. Dena, ., Rodrguez, M., Serrano, S., Barrio, R.: High-precision continuation of periodic
orbits. Abstr. Appl. Anal. 12 (2012). Art. ID 716024
71. Barrio, R.: Performance of the Taylor series method for ODEs/DAEs. Appl. Math. Comput.
163, 525545 (2005)
72. Abad, A., Barrio, R., Blesa, F., Rodrguez, M.: Algorithm 924: TIDES, a Taylor series
integrator for differential equationS. ACM Trans. Math. Softw. 39, 5:15:28 (2012)
73. Jorba, ., Zou, M.: A software package for the numerical integration of ODEs by means of
high-order Taylor methods. Exp. Math. 14, 99117 (2005)
74. Hairer, E., Nrsett, S.P., Wanner, G.: Solving Ordinary Differential Equations. I. Springer
Series in Computational Mathematics, vol. 8. Springer, Berlin (1993)
92 R. Barrio

75. Fousse, L., Hanrot, G., Lefvre, V., Plissier, P., Zimmermann, P.: MPFR: a multiple-precision
binary floating-point library with correct rounding. ACM Trans. Math. Softw. 33, 13:113:15
(2007)
76. Abad, A., Barrio, R., Blesa, F., Rodrguez, M.: TIDES tutorial: integrating ODEs by using the
Taylor series method. Monografas de la Academia de Ciencias de la Universidad de Zaragoza
36, 1116 (2011)
77. Hairer, E., McLachlan, R.I., Razakarivony, A.: Achieving Brouwers law with implicit Runge-
Kutta methods. BIT 48, 231243 (2008)
78. Brouwer, D.: On the accumulation of errors in numerical integration. Astron. J. 30, 149153
(1937)
79. Sim, C.: Global dynamics and fast indicators. In: Global Analysis of Dynamical Systems,
pp. 373389. Institute of Physics, Bristol (2001)
80. Sharp, P.W.: N-body simulations: the performance of some integrators. ACM Trans. Math.
Softw. 32, 375395 (2006)
81. Grazier, K.R., Newman, W.I., Hyman, J.M., Sharp, P.W.: Long simulations of the Solar System:
Brouwers law and chaos. ANZIAM J. 46, C1086C1103 (2004/05)
82. Grazier, K.R., Newman, W.I., Hyman, J.M., Sharp, P.W., Goldstein, D.J.: Achieving Brouwers
law with high-order Strmer multistep methods. ANZIAM J. 46, C786C804 (2004/05)
83. Higham, N.J.: Accuracy and Stability of Numerical Algorithms, 2nd edn. SIAM, Philadelphia,
PA (2002)
Chapter 4
Theory and Applications of the Mean
Exponential Growth Factor of Nearby Orbits
(MEGNO) Method

Pablo M. Cincotta and Claudia M. Giordano

Abstract In this chapter we discuss in a pedagogical way and from the very
beginning the Mean Exponential Growth factor of Nearby Orbits (MEGNO)
method, that has proven, in the last ten years, to be efficient to investigate both
regular and chaotic components of phase space of a Hamiltonian system. It is a fast
indicator that provides a clear picture of the resonance structure, the location of
stable and unstable periodic orbits as well as a measure of hyperbolicity in chaotic
domains which coincides with that given by the maximum Lyapunov characteristic
exponent but in a shorter evolution time. Applications of the MEGNO to simple
discrete and continuous dynamical systems are discussed and an overview of the
stability studies present in the literature encompassing quite different dynamical
systems is provided.

4.1 Introduction

One of the most challenging aspects of dynamical systems, particularly of those

that present a divided phase space, is the understanding of global properties in
phase space. Unfortunately, for instance, global instabilities of near-integrable
multidimensional Hamiltonian systems are far from being well understood, so in
this chapter we should focus on local features, that is, the dynamical behavior in a
small domain around a given point of the phase space of the system.
An example of the study of the local dynamics in every point of phase space
concerns the so-called chaos detection tools. This implies the characterization of
the dynamical flow around a given initial condition, that is for instance, how two
orbits starting very close to each other evolve with time t. A well known result
is that for ordered or regular motion, the separation between these two initially

P.M. Cincotta () C.M. Giordano

Grupo de Caos en Sistemas Hamiltonianos, Facultad de Ciencias Astronmicas y Geofsicas,
Universidad Nacional de La Plata and Instituto de Astrofsica de La Plata (CONICET), La Plata,
Argentina
e-mail: [email protected]; [email protected]

Springer-Verlag Berlin Heidelberg 2016 93

Ch. Skokos et al. (eds.), Chaos Detection and Predictability, Lecture Notes
in Physics 915, DOI 10.1007/978-3-662-48410-4_4
94 P.M. Cincotta and C.M. Giordano

nearby orbits grows linearly with time (or in some particular cases at some power
of t); while in those domains where the motion is unstable, chaotic, this separation
grows exponentially with t. The rate of this exponential divergence, defined as the
limit when t ! 1, is given by the so-called maximum Lyapunov Characteristic
Exponent (mLCE). Therefore if we know how to compute efficiently this separation
for large times we can obtain a picture of the local dynamics at any given point
of phase space. Indeed, in case of regular motion the mLCE vanishes and it has a
positive value for chaotic motion (and for unstable periodic orbits).
Another way to characterize the local dynamics is through a spectral analysis.
In fact, regular motion proceeds on invariant tori with a constant frequency vector
while, when the dynamics is chaotic, the frequencies are no longer local integrals
of motion but change with time. Therefore if we managed to develop an accurate
technique to measure the frequency of the motion, we could be able to separate
the dynamics in regular and chaotic. Moreover, in the regular regime it would be
possible to compute the full set of local integrals of motion (that is, the components
of the frequency vector).
Since the eighties and mid of the nineties two well known techniques have been
available in the literature, an algorithm to compute the mLCE, see for instance [5],
and the so called Frequency Map Analysis (FMA [43]). The first one obviously
provides the rate of divergence of nearby orbits while the second one is a very
precise method to obtain the frequencies of the motion. Both approaches were
widely used in many physical and astronomical applications; in particular the FMA
was the natural technique to investigate the dynamics of planets and, by means of
this tool it was shown that the Solar System as a whole dynamical system is not
stable and in fact it is chaotic or marginal unstable [42, 44, 45].
Actually, the mLCE and the FMA (besides the well known Poincar surface of
section for systems with two degrees of freedom), were popular chaos detection
tools in dynamical astronomy at those times.
However, computers were not fast enough to deal with large samples of orbits
and quite long integration times. For a set of M & 106 orbits, 12  M > 107
nonlinear coupled differential equations should be numerically integrated over long
time intervals and with high accuracy in order to get numerical values of the
mLCEs close to the expected theoretical ones. For instance, for regular motion the
theoretical mLCE ! 0 when t ! 1 as  ln t=t, so for an evolution time t  104 ,
a null mLCE numerically means  103 . Thus, it was not possible to distinguish
a regular orbit from a chaotic one with a mLCE  103 : Therefore, much larger
evolution times would be necessary to discriminate the nature of the motion by the
numerical asymptotic value. Thus it becomes clear that 20 years ago, this was a
severe restriction to derive precise values of the mLCE.
Since in the end of the nineties several fast dynamical indicators appeared in
the literature, some of the most popular ones in dynamical astronomy are largely
discussed in the present volume. All of them rest on the same theoretical arguments
4 Theory and Applications of the MEGNO 95

behind the mLCE, by following the evolution of the flow in a small neighborhood of
a given initial condition. Besides, a few new techniques, based on spectral analysis
have also been developed which are in fact, slight variations of the FMA. At the end
of this chapter we will briefly refer to several of such chaos detection tools.

4.1.1 The MEGNO: Brief History

The MEGNO belongs to the class of the so-called fast dynamical indicators. It is,
in fact, a byproduct of a former fast indicator, the Conditional Entropy of Nearby
Orbits, first proposed in [57] and improved in [10] and [11].
The MEGNO was announced in [11], but neither a description of the method
nor a name was provided. In [12], the MEGNO was introduced, but that work was
not devoted exclusively to the MEGNO, but to discuss analytical and numerical
methods for describing global dynamics in non-axisymmetric galactic potentials
in both regimes, regular and chaotic. The MEGNO was addressed there just as an
additional and simple tool, and its name (MEGNO) was proposed, following the
strong suggestion of one of the reviewers of the paper. The MEGNO was introduced
as an efficient way to derive accurately the mLCE. Indeed, in the Introduction of
that paper, the authors wrote : : : Alternative techniques were proposed to separate
ordered and stochastic motion, to classify orbits in families, to describe the global
structure of phase space, but not to get the LCN in shorter times. In Sect. 3 we
shall resume this point together with some comparisons with the new technique
here presented (MEGNO).: : : This new tool has proven to be useful for studying
global dynamics and succeeds in revealing the hyperbolic structure of phase-space,
the source of chaotic motion. The MEGNO provides a measure of chaos that is
proportional to the LCN, so that it allows to derive the actual LCN but in realistic
physical times: : :
It was in [13] that the MEGNO was discussed in detail and a generalization
of the original method was presented with applications to both multidimensional
Hamiltonian flows and maps.
This chapter is organized as follows. In Sect. 4.2 we address the theory of the
MEGNO in a simple fashion, without any intention to enunciate theorems and
their concomitant proofs: just several numerical examples would serve to show the
expected theoretical behavior of this dynamical indicator. In Sect. 4.3 we present
some applications to Hamiltonian flows and symplectic maps. We also yield the
results of an exhaustive comparative study of different indicators of chaos in
Sect. 4.4. We discuss further applications of the MEGNO that can be found along
the literature, from realistic planetary models to bifurcation analysis, in Sect. 4.5. A
thorough discussion is provided in the last section.
96 P.M. Cincotta and C.M. Giordano

4.2 The Mean Exponential Growth Factor of Nearby

Orbits (MEGNO)

Herein we address the MEGNOs theory following the original presentation given
in [13] but in a more pedagogical way.
To that aim, let us consider the phase space state vector

x D .p; q/ 2 B
R2N ; (4.1)

and introduce the vector field, also defined in B

 
@H @H
v.x/ D  ; ; (4.2)
@q @p

where H.p; q/ refers to an N-dimensional Hamiltonian, assumed to be autonomous

just for the sake of simplicity. The formulation given below however is completely
independent of the system being Hamiltonian as well as of the phase space
coordinates adopted to express the state vector x. In case of a Hamiltonian system,
since the motion in general takes place on a compact energy surface Mh D fx D
.p; q/ 2 B W H.p; q/ D hg, thus x 2 B0 Mh , where dim.Mh / D 2N  1.
Therefore, the equations of motion in B0 have the simple form

xP D v.x/: (4.3)

Let '.t/ denote a given solution of the flow (4.3), for a given initial condition x0 ,

'.t/ D x.tI x0 /; x0 2 B0 : (4.4)

For any such an orbit ' the mLCE, .'/, is defined as

1 k.'.t//k
.'/ D lim 1 .'.t//; 1 .'.t// D ln ; (4.5)
t!1 t k 0 k

where .'.t// and 0 are infinitesimal displacements from ' at times t and 0,
respectively, and k  k denotes the usual Euclidean norm.1
In fact, .'.t// is the time evolution of the difference ' 0 .t/  '.t/, being ' 0 .t/
a nearby orbit to '.t/ whose initial condition is x00 D x0 C x0 , for kx0 k small
enough. The evolution of ' 0 .t/  '.t/ after linearizing the flow around '.t/ is then
computed. Therefore we are evaluating the flow (4.3) and its first variation on a
single orbit, instead of computing the evolution of ' 0 .t/ and '.t/ and performing
their difference. This is the very same way in which the algorithm to compute the

1
Let us note that any other norm could be used all the same.
4 Theory and Applications of the MEGNO 97

mLCE was developed. Though it would be possible to integrate the flow to get ' 0
and ' starting at x00 and x0 respectively, when performing the difference ' 0 .t/  '.t/,
both in Mh , then after a large but finite time t, the separation between the two orbits
would reach, in a chaotic domain, an upper bound k' 0 .t/  '.t/k  d, where d is
the maximum size of the accessible region in Mh . The limiting case when k' 0 .t/ 
'.t/k D d corresponds to the completely ergodic case, in which any orbit, and also
the difference of nearby ones, could fill densely the energy surface Mh .
In any case, the computation of ' 0 .t/  '.t/ would provide the right physical
insight about the nature of the dynamics in a small neighborhood of x0 , but
computationally this is not the best option, since the mLCE measures the divergence
of .t/ D k' 0 .t/  '.t/k when t ! 1 and 0 D k' 0 .0/  '.0/k ! 0.
It is well known that  provides relevant information about the flow in a small
domain around '. Indeed, recasting (4.5) in the form

Z !
1 t P
.'.s// P
.'/ D lim ds D ; (4.6)
t!1 t 0 .'.s//

where  kk is the Euclidean norm, P  d=dt D P  =kk, and ./ denotes

time-average, thus it is explicit that the mLCE measures the mean exponential rate
of divergence of nearby orbits.
Thus defined, the so-called tangent vector satisfies the first variational equation
of the flow (4.3):

P D .'.t//; (4.7)

where .'.t//  Dx v.'.t// is the Jacobian matrix of the vector field v evaluated on
'.t/.
Let us now introduce a slightly different sensitive function on the orbit '.t/ which
is closely related to the integral in (4.6); the Mean Exponential Growth factor of
Nearby Orbits (MEGNO), Y.'.t//, through
Z P
2 t
.'.s//
Y.'.t// D sds: (4.8)
t 0 .'.s//

Recall that in case of an exponential increase of , as it occurs for an unstable

periodic orbit or a chaotic one, .'.t// D 0 exp.t/;  > 0, Y.'.t// can be
considered as a weighted variant of the integral in (4.6). Indeed, instead of the
instantaneous rate of growth, , we average the logarithm of the growth factor,
ln..'.t//=0 / D t. Further variants will be considered in Sect. 4.2.2 where the
generalization of the MEGNO is addressed.
In what follows we consider some, though quite special, very representative
solutions of (4.7) in order to show how Y.'.t// serves to provide clear indication on
the character of the motion in each case.
98 P.M. Cincotta and C.M. Giordano

Thus, let us first consider any orbit 'q .t/ on a N-dimensional irrational torus in a
non-isochronous or nonlinear system. Therefore we can locally define action-angle
variables .I; / such that .t/ D .I/t C 0 ; I  I0 , being I0 a constant and, for any
set of generalized coordinates .p; q/ the solution of (4.3) can be expanded in Fourier
series in with coefficients that depend on I. Therefore for any such quasiperiodic
orbit, 'q , the solution of (4.7) in generalized coordinates has the form
    
'q .t/  0 1 C wq .t/ C t q C uq .t/ ; (4.9)

where q > 0 is the absolute value of the linear rate of divergence around 'q , wq .t/
and uq .t/ are oscillating functions (in general quasiperiodic and with zero average)
of bounded amplitude, that satisfy juq .t/j  bq < q , for some positive constant bq .2
The quantity q is a measure of the lack of isochronicity around the orbit and it is
related to the absolute value of the maximum eigenvalue of the nonlinearity matrix

@!i @2 H
D :
@Ij @Ii @Ij

Recall that for a linear or quasi-linear system, such as the harmonic oscillator,  D 0
for all '. Indeed, the linear divergence of two nearby quasiperiodic orbits reflects the
fact that they move on nearby N-dimensional tori. Since we assume that ! depends
on I, two nearby tori have a small differentaction vector,
 say I and I C I, and thus
.I C I/ D .I/ C . However if det @!i =@Ij D 0, the system behaves as a
linear one and no divergence between two nearby orbits is expected.
From (4.8) and (4.9), keeping in mind that juq j is bounded by bq , it is
straightforward to see that Y.'q .t// oscillates around 2 with bounded amplitude,
verifying that

  q C b q bq
jY 'q .t/  2j  4 ln 8 ; t ! 1; (4.10)
q  b q q

where the last approximation holds if bq  q . The time evolution of Y.'q .t// is
given by

  2 ln.1 C q t/  
Y 'q .t/  2  C O 'q .t/ ; (4.11)
q t

where O denotes an oscillating term (with zero average) due to the quasiperiodic
character of both wq .t/ and uq .t/. Though
 
lim Y 'q .t/ (4.12)
t!1

2
Anyway (4.9) could be empirically derived by numerical means.
4 Theory and Applications of the MEGNO 99

 
does not exist due to the oscillatory term O 'q .t/ in (4.11), introducing the time-
average
Z
1 t
Y.'q .t//  Y.'q .s//ds; (4.13)
t 0

it can readily be seen from (4.10), (4.11) and (4.13) that

   
Y 'q  lim Y 'q .t/ D 2: (4.14)
t!1

Therefore, for quasiperiodic motion, Y.'/ converges to a constant value, which is

independent of 'q .t/.
The above results still hold in case of a regular orbit '.t/ that is not purely stable
quasiperiodic. Let us restrict ourselves to 2-dimensional (2D) Hamiltonian systems,
though the arguments given below could be straightforwardly extended to higher
dimensions and let '.t/ be close to a stable periodic orbit, 's .t/. Since O.'.t// in
(4.11) involves nearly periodic terms, and both  and b= are small, it follows from
(4.10) and (4.11) that Y.'.t// oscillates around 2 with a small amplitude and that
Y.'.t// converges to 2 slower the smaller is . When '.t/ ! 's .t/; both u.t/;  !
0, and Y ! 0 as t ! 1. In this limiting case, the oscillations of Y.'.t// about 0
are due to the presence of the term w.t/ in (4.9).
Meanwhile, whenever '.t/ is close to an unstable periodic orbit, 'u .t/, Y.'.t//
behaves in a different fashion since in such a case, the motion in any small
neighborhood of 'u .t/, U, is mainly determined by its associated stable and unstable
manifolds. For a sufficiently large motion time, '.t/ will pass close to 'u .t/
several times. Suppose that between two successive close approaches with 'u .t/,
'.t/ spends a time 1 within U and a time 2 outside U. During the interval
1 , .'.t//  .'u .t//  0 exp .t/ with  > 0, while, during 2 , .'.t//
approximately obeys (4.9). The interaction time between '.t/ and 'u .t/, 1 ,
is larger the closer the orbits are to each other. Thus, Y.'.t// should exhibit
quasiperiodic oscillations modulated by periodic pulses, of period  2 , width
 1 and similar amplitude. Analogous considerations apply to Y.'.t// but, due
to the averaging, the amplitude of the pulses should decrease as  1=t. In general,
Y.'.t// will approach 2 from above and, after a total evolution time t, Y.'.t// will
be larger the smaller is the distance k'.t/  'u .t/k. In the limit, when '.t/ ! 'u .t/,
1 ! t and .'.t// grows exponentially with time, so that Y.'.t// 2 (see
the forthcoming Eq. (4.17) and the so-called right-stop criterion discussed in
Sect. 4.2.4 that applies for maps).
In case of an irregular orbit, 'i .t/, within any chaotic component, the solution of
(4.7), besides oscillation terms which are irrelevant in this case, is

.'i .t//  0 ei t ; (4.15)

100 P.M. Cincotta and C.M. Giordano

i being the 'i .t/s mLCE. Thus,

Q i .t//;
Y .'i .t//  i t C O.' (4.16)

with OQ some oscillating term of bounded amplitude which is in general neither

periodic nor quasiperiodic, but it has zero average.3 Note that in a chaotic domain
the orbits proceed on a D-dimensional manifold where N < D < 2N  1. In these
domains, tori are in general destroyed and the dynamics is said to be hyperbolic
since a chaotic orbit could be thought as a slight distortion of an unstable P-periodic
orbit with P 1.
On averaging (4.16) over a large time interval, we obtain
i
Y .'i .t//  t; t ! 1: (4.17)
2
Therefore, for a chaotic orbit, Y .'i .t// and Y .'i .t// grow linearly with time, at
a rate equal to the mLCE of the orbit or one half of it, respectively (see below).
Only when the phase space has an hyperbolic structure, does Y grow with time.
Otherwise, it saturates to a constant value, even in the degenerated cases in which
grows with some power of t, say n, and therefore Y ! 2n as t ! 1.
The MEGNOs temporal evolution allows for being summed up as a single
expression valid for any kind of motion, which is not the case for 1 or any other
chaos indicator. In fact, the asymptotic behavior of Y.'.t// may be written in the
fashion

Y.'.t//  a' t C b' (4.18)

where a' D ' =2 and b'  0 for chaotic motion, while a' D 0 and b'  2 for
stable quasiperiodic motion. Departures from the value b'  2 indicate that ' is
close to some periodic orbit, being b' . 2 and b' & 2 for stable or near-unstable
periodic orbits, respectively.
Notice that O 1  Y.'.t//=t verifies
2
O 1 .'q .t//  ; O 1 .'i .t//  i ; t ! 1; (4.19)
t
which show that, for regular motion O 1 converges to 0 faster than 1 , which it does
as ln t=t, while for chaotic motion both magnitudes approach the positive mLCE at
a similar rate.
As it turns out from (4.18) and perhaps the key point of the MEGNO method
(but not widespread used) is that, since for chaotic motion Y grows linearly with
time with a rate =2, a very accurate estimate of the mLCE can be obtained in rather

3
Since the motion is bounded in phase space, any orbit '.t/ should be an oscillating function of
time of bounded amplitude, despite if it is regular or chaotic. For unstable or chaotic orbits the
main secular growth is given by the exponential term and therefore it is always possible to separate
it from a purely oscillating term with zero average.
4 Theory and Applications of the MEGNO 101

short times by means of a linear least squares fit on Y .'.t//. The main feature of
this procedure is that it takes advantage of all the dynamical information contained
in Y .'.t// regarding the whole interval .t0 ; t/; t t0 and on the fact that Y has
a smooth behavior. Since for purely quasiperiodic orbits Y .'.t// approaches the
constant value 2 quite faster than for nearly stable and near-unstable periodic orbits,
the mLCE derived from a linear least squares fit of the MEGNO would also yield
information on elliptic and hyperbolic points as well.

4.2.1 Comparison of Theoretical and Numerical Results

In order to illustrate the predicted MEGNOs behavior, we regard the well known
2D HnonHeiles model [35],

1 2 1 y3
H. px ; py ; x; y/ D . px C p2y / C .x2 C y2 / C x2 y  ; (4.20)
2 2 3
where x; y; px ; py 2 R: This Hamiltonian was proposed in the sixties to investigate
the existence of the so-called third integral of motion in the Galaxy. We consider
the energy level h D 0:118. The phase space at this energy level displays at least
two main unconnected chaotic domains having different mLCEs as shown by the
Poincar surfaces of section presented in Fig. 4.1 (see, for instance, [11]).

0.5

0.4

0.3

0.2

0.1
py

-0.1

-0.2

-0.3

-0.4

-0.5
-0.4 -0.2 0 0.2 0.4 0.6
y

Fig. 4.1 .y; py /-Surfaces of section for the HnonHeiles Hamiltonian for h D 0:118; x D 0;
px > 0. The arrows indicate the location of the five initial conditions, from left to right (sp), (up),
(qp), (c1), (c2). See text
102 P.M. Cincotta and C.M. Giordano

We picked up initial conditions for five representative orbits from the surface x D
0: one close to the stable 1-periodic orbit at .y; py / D .0:295456; 0/ (sp); another
one looking like stable quasiperiodic at .0:483; 0/ (qp); a third one at .0:46912; 0/
also quasiperiodic but close to an unstable 4-periodic orbit (up); and two irregular
orbits, one in the stochastic layer surrounding a 5-periodic island chain (or at a 5 W m
resonance for m 2 Z0 ) (c1) at .0:509; 0/, and the other one lying in a large chaotic
sea (c2) at .0:56; 0:112/.
We computed Y and Y by means of (4.8) and (4.13) respectively; note that
the renormalization of , if necessary, proceeds naturally from (4.8). Along this
work all the numerical integrations were carried out by recourse to a RungeKutta
7/8th order integrator (the Dopri8 routine, see [58] and [34]), the accuracy in the
conservation of the energy in this case being 1013 . The initial tangent vector is
chosen at random and with unit norm.4
In Fig. 4.2 we show that both Y and Y evolve with time as predicted. Indeed, in
Fig. 4.2a we observe that, for the stable quasiperiodic orbit (qp), Y oscillates around
the value 2 with an amplitude . 1, while Y shows a very fast convergence to the
actual average (see below).
Figure 4.2b displays the typical behavior of a trajectory close to an unstable
periodic orbit. While the (up) orbit is far away from the hyperbolic point, both
Y and Y evolve as in the quasiperiodic case. However, when this quasiperiodic orbit
passes close to the unstable one, the mutual interaction causes the oscillations of Y
to exhibit a strong modulation, which is damped in Y as t increases. Thus, after the
first close approach at t  2000; Y > 2 (mainly due to the cumulative effect on the
average) but, for t large enough, it asymptotically approaches 2.
Also for the irregular orbits (c1) and (c2) we compute the time-evolution of Y
and Y. The results are given in Fig. 4.2c, where both Y and 2Y are plotted together
to show that, as follows from (4.16) and (4.17), both quantities have the same time-
rate. Since the trajectories belong to unconnected chaotic domains, the time-rate
(i.e. the mLCE) is different for the two orbits.
In Fig. 4.2d, the temporal evolution of Y for all the three regular orbits are
compared. For the stable quasiperiodic orbit (qp), Y reaches 2 much faster than
for the orbit (sp), which is close to a stable periodic one. In fact, Y.'sp / . 2 over the
full time interval. The time evolution of both, Y.'sp / and Y.'qp /, fit very well (4.11),
on neglecting oscillations and being sp < qp . We note again just for Y, that the
orbits (qp) and (up) evolve in a rather similar way, as long as the interaction between
(up) and its nearby unstable periodic orbit is weak. Therefore, a least squares fit on
Y could distinguish clearly quasiperiodic orbits from stable and unstable periodic
orbits.
In order to show that O 1 ! mLCE when t ! 1, in Fig. 4.2e we display its
time evolution together with that of 1 for three of the orbits, namely, (sp), (c1) and

4
One should verify that the tangent vector has a non-vanishing component normal to the flow,
particularly in the regular component, in order to ensure the linear divergence of nearby orbits
(see [16]).
4 Theory and Applications of the MEGNO 103

a b
Y(t) Y(t)
< Y > (t) 5 < Y > (t)
3

2.5 4

2
3

1.5
2
1

1
0.5

0 0
0 1000 2000 3000 4000 5000 6000 7000 0 1000 2000 3000 4000 5000 6000 7000
t t
c d
Y(t) 3.5 < Y > (t)
400
2 < Y > (t)
up
350 3

300 2.5
c2 qp
250
2
200 sp
1.5
150
1
100
c1 0.5
50

0 0
0 1000 2000 3000 4000 5000 6000 7000 0 1000 2000 3000 4000 5000 6000 7000
t t
e
0.1 mLCE
c1 Y/t

0.01
c2

0.001
sp

0.0001
1000 2000 3000 4000 5000 6000 7000
t

Fig. 4.2 Time evolution of Y and Y (< Y > in the figure) for the orbits: (a) (qp) stable
quasiperiodic; (b) (up) quasiperiodic but close to an unstable 4-periodic orbit; (c) (c1) and
(c2) irregular, embedded in two different chaotic domains. (d) Y (< Y > in the figure) for three
regular orbits: (sp) close to a stable periodic orbit, (qp), and (up); (e) time evolution of O 1 (Y=t in
the figure) and the mLCE, 1 for (sp), (c1) and (c2) computed using the algorithm given in [5]

(c2). We observe that for the chaotic orbits, both magnitudes converge to the same
positive mLCE at the same rate. For the regular orbit (sp) instead, we note that O 1
decreases faster than 1 , the expected final values (see (4.19) and discussion below),
0:0013 and 0:00028 respectively, being the latter close to the computed one.
In the case of chaotic motion, both Y and Y evolve almost linearly with time
over the whole time interval, as shown in Fig. 4.2c. The deviations from the linear
trend, for instance in (c2), are presumably caused by stickiness. Indeed, during those
104 P.M. Cincotta and C.M. Giordano

2.5 -1

0.04 0.04 -2
2
-3

0.02 0.02 -4
1.5
-5
py

py
0 1 0 -6

-7
0.5
-0.02 -0.02 -8

-9
0
-0.04 -0.04 -10

-0.5 -11
-0.1 -0.05 0 0.05 0.1 -0.1 -0.05 0 0.05 0.1
y y

Fig. 4.3 .y; py /-Surface of section of the HnonHeiles Hamiltonian for h D 0:118; x D 0; px >
0. The left panel corresponds to MEGNO contour plot in logarithmic scale .log 2  0:3/ and in
the right panel the mLCE, also in logarithmic scale, derived from a linear least squares fit on Y.
See text for details

time intervals, ts , in which Y is almost flat, the orbit remains close to some small
stability domain embedded in the chaotic sea. In this particular example, stickiness
does not significantly reduce the linear trend but, whenever it is strong, it does
influence the mean time-rate of both Y and Y and consequently, the derived mLCE.
However, the same effect would be present in the numerical computation of the
mLCE, since the stickiness phenomena affects the evolution of .t/ and therefore
if .t/ does not increase exponentially within ts , the evolution of 1 .t/ would
decrease with time as  ln t=t while t 2 ts .
Finally, in Fig. 4.3 we present a small domain of the .y; py /-surface of section of
the HnonHeiles Hamiltonian for h D 0:118; x D 0; px > 0 given in Fig. 4.1, the
contour plots providing, in logarithmic scale, the MEGNO and the mLCE computed
by a least squares fit on the time evolution of Y over the time interval .t0 ; t/ with t D
104 ; t0 D 2  103 . A given t0 > 0 is adopted in order to avoid the initial transient;
thus, the least squares fit is performed over the 80 % of the full time interval. From
these two plots we observe that the MEGNO provides a clear picture of the dynamics
but the accurate value of the mLCE obtained following this alternative procedure
furnishes more information than the MEGNO itself. Indeed, both plots show up
the very same information in the chaotic domain, however, the MEGNO does not
separate clearly the thin unstable domain inside the stability island as the mLCE
computed by a least squares fit does. Note that using a simple least squares fit on the
time evolution on Y over the 80 % of the whole time interval, we reach values of the
mLCE for regular motion of the order of 1010 or lower considering motion times
t  104 , when the expected lower value of the mLCE by recourse to the classical
algorithm would be 103 . This is, in our opinion, one of the main results provided
by the MEGNO: its very accurate determination of the positive and null mLCE, for
chaotic and regular motion respectively.
4 Theory and Applications of the MEGNO 105

Further details on the MEGNOs performance when applied to the study of the
dynamics of 2D Hamiltonians, as well as other advantages of deriving the mLCE
from a least squares fit on Y are given in [12].

4.2.2 Generalization of the MEGNO

Let us generalize the MEGNO by introducing the exponents .m; n/ such that
Z P
t
.'.s//
Ym;n .'.t/ / D .m C 1/ t n
.s/m ds; (4.21)
0 .'.s//

now defining
Z
1 t  
Y m;n .'.t/ / D Ym;n 'q .s/ ds; (4.22)
tmCnC1 0

and analyze whether any benefit would turn out when taking values for the
exponents .m; n/; m  0 other than the natural choice .1; 1/ which yielded (4.8)
and (4.13). Note that in the limit when t ! 1; Y0;1 !  as defined in (4.6).
The time evolution of Ym;n for regular, quasiperiodic motion, is given by
!
  X
m1
.1/k tmCnk
Ym;n 'q .t/  .m C 1/
kD0
.m  k/kq
!
mt
n
ln.1 C q t/  
C.m C 1/ .1/ C O 'q .t/ ; (4.23)
mq

which naturally reduces to (4.11)

 for .m; n/ D .1; 1/. This expression is obtained
by replacing the value of 'q .t/ given by (4.9) in (4.21). Notice that for t large
enough we get
 
Ym;n 'q .t/ .m C 1/
mCn
 ; (4.24)
t m

so the ratio Ym;n =tmCn saturates to a constant as t ! 1.

Moreover, from both (4.22) and (4.23) it follows that

  .m C 1/
Y m;n 'q .t/  ; t ! 1; (4.25)
m .m C n C 1/

which is also a fixed constant not depending on the orbit.

106 P.M. Cincotta and C.M. Giordano

For an irregular orbit, 'i , with mLCE i , we have

Ym;n .'i .t/ /

 i t C OQ .'i .t// ; (4.26)
tmCn
while, on considering a sufficiently large time, we obtain

i t
Y m;n .'i .t/ /  : (4.27)
.m C n C 2/

For a chaotic orbit then, both Ym;n =tmCn and Y m;n , thus defined, grow linearly
with time, at a rate that is proportional to the mLCE of the orbit.
Therefore, the asymptotic behavior of Y m;n can be recast as

Y m;n .'.t//  a' t C b' ; (4.28)

where now a' D i =.m C n C 2/ and b'  0 for irregular, chaotic motion, while
a' D 0 and b'  .mC1/=m.mCn C1/ for stable, quasiperiodic motion. As it turns
out from (4.28), the mLCE can also be recovered by a simple linear least squares fit
on Y m;n .'.t//.
Notice that O 1;m;n D Ym;n =tmCnC1 satisfies

.m C 1/
O 1;m;n .'q .t//  ; O 1;m;n .'i .t//  i ; t ! 1; (4.29)
mt

so that, for regular motion, O 1;m;n also converges to 0 faster than 1  ln t=t, while
for chaotic motion, both magnitudes approach the positive mLCE at a similar rate.
An exhaustive comparison of the generalized MEGNOs performance for differ-
ent exponents .m; n/ revealed that, besides the natural choice .1; 1/, the values
.2; 0/ serve to distinguish regular from chaotic behavior in a quite efficient manner
(see below).
Just for the sake of illustration, let us turn back to the 2D HnonHeiles example
given in Sect. 4.2.1. For the same three regular orbits labeled as (sp), (qp) and (up),
we computed both Ym;n and Y m;n , by means of (4.21) and (4.22) respectively, for
three different choices of .m; n/, namely, .1; 1/, .2; 0/ and .3; 1/.
In Fig. 4.4 we show that for regular motion, Y m;n evolves with time as predicted
by (4.25). Indeed, the temporal evolution of Y m;n for the three regular orbits is seen
to tend to the asymptotic values 2, 1=2 and 4=15, when the exponents are .1; 1/,
.2; 0/ and .3; 1/, respectively. We note that, for the stable quasiperiodic orbit (qp),
Y m;n converges to the value given in (4.25), a faster convergence being observed the
larger is m. Also for the orbit close to a stable periodic one (sp), does Y m;n reach the
constant value (4.25) faster as a larger exponent m is considered. Notice however
that for m D 2 much smaller oscillations around the asymptotic value (4.25) are
observed in the case of the trajectory close to an unstable periodic orbit (up). Let us
note that the exponent n is dummy in the present discussion.
4 Theory and Applications of the MEGNO 107

3.5

2.5 up

qp (1,-1)
2
<Ym,n>

sp
1.5

(2,0)
0.5

(3,1)
0

0 2000 4000 6000 8000 10000

t

Fig. 4.4 Time evolution of Y m;n (< Y >m;n in the plot) for the regular orbits (sp), (qp) and (up) in
the HnonHeiles model, for different values of the exponents .m; n/. Figure taken from [13]

From this comparison we conclude that the choice of exponents .2; 0/ allows for
clearly separating regular and chaotic regime even in rather short evolution times.
Furthermore, if we use as a dynamical indicator the quantity 4Y 2;0 , we see that for
regular orbits it tends to 2, as Y 1;1 does, while for orbits with exponential instability
it tends to i t. Then, either a linear fit or simply 4Y 2;0 .'i .t//=t provides an estimate
of the mLCE. However, the choice .1; 1/ for the exponents offers the additional
benefit of more clearly identifying stable and unstable periodic motion as well.
Anyway, though all the eventual advantages of the generalized MEGNO showed
above, the use of the classical MEGNO, Y 1;1 , is widespread. Therefore, we will
show the results for Y 2;0 when dealing with discrete applications and discuss below
an interesting connection between the classical MEGNO and another well known
chaos indicator.

4.2.3 The Connection Between the MEGNO and the FLI

The standard MEGNO, defined adopting the value of the exponents .1; 1/, exhibits
an intrinsic relation with the classical Fast Lyapunov Indicator (FLI) [53], as we will
see in the sequel. For that sake we recall that in [18] the authors define the FLI for
a given solution of the flow (4.3), '.t/, in terms of the norm of the tangent vector
108 P.M. Cincotta and C.M. Giordano

 kk as

FLI.'.t// D ln .'.t//; (4.30)

expression that has been used to obtain analytical results in both [18] and [33].
Thus, the time average of the FLI in the interval .0; t/ is given by
Z
1 t
FLI.'.t// D ln .'.s//ds: (4.31)
t 0

The MEGNO is twice a time weighted average of the relative divergence of orbits
as it can be seen from (4.8). In order to show the relation between the MEGNO and
the FLI, let us rewrite (4.8) in the fashion:
Z
2 t
d
Y.'.t// D .ln .'.s/// ds: (4.32)
t 0 ds

After a simple manipulation we obtain

( Z )
1 t
Y.'.t// D 2 ln .'.t//  ln .'.s//ds ; (4.33)
t 0

where the value .0/ D 1 has been taken. From (4.30), (4.31) and (4.33) we
conclude that the MEGNO is twice the difference between the FLI and its time
average over the interval .0; t/,

Y.'.t// D 2 FLI.'.t//  FLI.'.t// : (4.34)

This result serves to understand two facts that have been recently mentioned in the
literature. One point is that the MEGNO criterion takes advantage of the dynamical
information of the evolution of the tangent vector along the complete orbit, as stated
in [70] and [37]. Equation (4.34) tells us exactly in which way it encompasses this
information: at every time the MEGNO subtracts from the FLI its average value.
The other point worth discussing, which is explicitly mentioned in [7] and [3],
is the reason by which the MEGNO gives account of the degree of chaoticity of an
orbit in an absolute scale while the FLI just gives relative values; i.e. in the case of
regular orbits the MEGNO tends asymptotically towards a constant value (2), while
the FLI behaves logarithmically, not allowing to count with a time independent
criterion to establish the threshold that separates chaotic from regular motion.
Just to illustrate this situation let us consider the case of an ideal KAM regular
orbit. Therefore the norm of the tangent vectors behaves as (4.9) and besides
oscillations .'q .t//  1 C t ( > 0 and 0 D 1). In this case it is

FLI.'q .t//  ln.1 C t/ (4.35)

4 Theory and Applications of the MEGNO 109

and

ln.1 C t/
FLI..'q .t///  ln.1 C t/ C  1: (4.36)
t
Therefore, on regarding (4.34) there results
 
ln.1 C t/
Y.'q .t// D 2 1  ; (4.37)
t

and we rediscover the already mentioned asymptotic limit of the MEGNO for
regular orbits.
On the other hand, in the case of an ideal chaotic orbit, with .'i .t//  e t (being
 the mLCE), the MEGNO-FLI relation allows to prove that both indicators behave
similarly, that is linearly with time with a slope equal to .
In order to show the MEGNO-FLI relation we consider again the HnonHeiles
model for the same energy level, h D 0:118, and two orbits one quasiperiodic at
y D 0:2; py D 0 inside the largest island, and a chaotic one at y D 0:18; py D 0 in
thechaotic sea. Just to eliminate

oscillations, we compute Y.'.t// and the average
of FLI.'.t//  FLI.'.t// for these two orbits. The results presented in Fig. 4.5
show an excellent agreement between both magnitudes.

18
<Y>
2 < FLI - <FLI> >
16

14
c

12

10

2
qp

0
0 100 200 300 400 500 600 700 800
t

Fig. 4.5 Illustration of the relation (4.34)in averagefor a quasiperiodic orbit (qp) and
a chaotic (c) one for the HnonHeiles system at h D 0:118. Note that both curves are
indistinguishable
110 P.M. Cincotta and C.M. Giordano

Therefore in view of the close relation between the MEGNO and the FLI, any
improvement concerning the FLI, as for instance the alternative version of the
FLI, the so-called Orthogonal Fast Lyapunov Indicator (OFLI)see [16] and the
corresponding chapter in this volume, applies naturally to improve the MEGNO
itself.

4.2.4 The MEGNO for Maps

In this subsection we briefly show how the MEGNO should be implemented to

discrete dynamical systems. For dealing with maps, this numerical tool is defined
essentially as before, but summing over the iterates of the map instead of integrating
with respect to t, and taking the differential map in place of the first variational
equations.
For a given initial point P0 , iterates under a given map T are computed yielding
points Pk D T k .P0 /. An initial random and unitary tangent vector v0 , is transported
under the differential map DT, to obtain vectors vk D DT k .P0 /v0 . Then, after N
iterates, the (generalized) MEGNO is computed by means of

X
N  
kvk k
Ym;n .N/ D .m C 1/ N n
ln km ; (4.38)
kD1
kvk1 k

and

1 X
N
Y m;n .N/ D Ym;n .k/ : (4.39)
N mCnC1 kD1

We have considered different values for the exponents m and n. Again, it turned
out that the larger m, the faster Y m;n converges to a constant value for regular
motion, but, for m rather large, small oscillations show up. However, the bumpy
late evolution of Y m;n (which is also present in the continuous case, as Fig. 4.4
shows, in the case of (up) orbits) is diminished if the iteration is stopped when the
distance between the initial and final points is minimum (right-stop condition).
On returning close to the initial point, the effect of the periodic or quasiperiodic
oscillations added to a regular behavior is minimized. This sort of refinement in
regards to the stop time in the case of maps has proven rather efficient in smoothing
such oscillations.
The choice .2; 0/ for the exponents, together with the right-stop condition,
have shown to provide a fairly good fast dynamical indicator for maps. A minor
additional modification is also convenient with the choice .m; n/ D .2; 0/. Let us
define the parameter

4Y 2;0 .N/  2
YO 2;0 .N/ D ; (4.40)
N
4 Theory and Applications of the MEGNO 111

which when N ! 1, YO 2;0 ! 0 for orbits lying on tori, while YO 2;0 ! i in the
case of chaotic orbits that lie in a higher dimensional domain. So, negative values
(close to 0) of YO 2;0 .N/ arise for regular orbits (provided N is taken not too small),
while small positive values would identify mild chaos.

4.3 Applications

4.3.1 A System of Continuous Time: The Arnold Model

Let us consider the well known classical Arnold Hamiltonian [1], which is the
paradigmatic model that leads to the so-called (and perhaps controversial) Arnold
diffusion. We will address this simple but very representative nonlinear model
because, in our opinion, it has not been discussed in a plain manner for non-
mathematical readers yet. In fact, though Sect. 7 in [8] is devoted to present Arnold
diffusion in an heuristic way by recourse to this model, unfortunately that section of
the outstanding review by B. Chirikov seems not to be widespread in the nonlinear
community. The Arnold model is also well discussed in the lectures of Giorgilli [24],
though in a more mathematical fashion.
The Arnold Hamiltonian has the form
1 2
H.I1 ; I2 ; 1 ; 2 ; tI "; / D.I C I22 / C ".cos 1  1/.1 C B.2 ; t//
2 1
B.2 ; t/ D sin 2 C cos t; (4.41)

with I1 ; I2 2 R; 1 ; 2 ; t 2 S1 ; where  should be exponentially small with respect

to ", so that "  "  1 ( just in Arnold formulation, however see below).
For " D 0 we have two integrals of motion, namely I1 and I2 which determine the
invariant tori supporting the quasiperiodic motion with frequencies !1 D I1 ; !2 D
I2 . Therefore we have a very simple dynamical system consisting of two uncoupled
free rotators, so that, 1 .t/ D I1 t C 10 ; 2 .t/ D I2 t C 20 .
For " 0,  D 0 we still have two integrals,

1 2
H1 .I1 ; 1 I "/ D I C ".cos 1  1/; I2 ; (4.42)
2 1
and the unperturbed Hamiltonian could be written as

1
H0 .I1 ; I2 ; 1 I "/ D H1 .I1 ; 1 I "/ C I22 : (4.43)
2
Notice that H1 is the pendulum model for the resonance !1 D 0; H1  h1 D 2"
corresponds to the exact resonance or stable equilibrium point at .I1 ; 1 / D .0; /
112 P.M. Cincotta and C.M. Giordano

while h1 D 0 to the separatrix and thus .I1 ; 1 / D .0; 0/ is the unstable point or
whiskered torus.5
The associated frequencies are now !1 D !p .h1 ; "/; !2 D I2 ; where !p .h1 ; "/ is
the pendulum frequency,
!0 ."/
!p .h1 ; "/ D ; 2"  h1 < 0;
2K .kh1 /
(4.44)
!r .h1 ; "/
!p .h1 ; "/ D   ; h1 > 0I
2K kh11
p
where kh21 D .h1 C 2"/=2"; !0 ."/  " is the small oscillation frequency,
!r .h1 ; "/ D !0 ."/kh1 is the half-rotation frequency and K. / is the complete
elliptical integral of the first kind. For rotations, the second in (4.44) provides
the half-rotation frequency, in order to avoid the jump of a factor 2 between
the frequency at both sides of the separatrix. Therefore in the oscillation regime
!p .h1 ; "/  !0 ."/ and close to the separatrix for both oscillations and rotations,
!p .jh1 j  1; "/  !sx .h1 ; "/ takes the asymptotic form

!0 ."/
!sx .h1 ; "/ D  ; !sx .h1 ; "/ ! 0 as jh1 j ! 0: (4.45)
32"
ln jh 1j

p
In the rotation regime, for h1 large enough 2!p .h1 ; "/  2h1  I1 . Figure 4.6
shows the dependence of !p on h1 , for " D 0:15. p
The resonance !1 D 0 has a half-width p .I1 / D 2 " in action space, so the
r

variation of I1 is bounded by jI1 j  2 " whilep I2 remains constant. Therefore in

.I1 ; I2 / plane, !1 ! !sx .h1 ; "/ ! 0 when I1 ! 2 ".
For " 0;  0 the original system (4.41) can be written as

H.I1 ; I2 ; 1 ; 2 ; tI "; / D H0 .I1 ; I2 ; 1 I "/ C V.1 ; 2 ; tI "/; (4.46)

V.1 ; 2 ; tI "/ D ".sin 2 C cos t/.cos 1  1/;

where H0 is given by (4.43) and 2 .t/ D !2 t C 20 . Therefore the full Hamiltonian

is a simple pendulum and a free rotator coupled by V.1 ; 2 ; tI "/.
Since the perturbation depends on 2 and t, it affects the phase oscillations at
the resonance !1 D 0 and leads to the formation of the stochastic layer around its
separatrix. Moreover, due to the dependence of V on 2 , the perturbation changes
not only I1 but I2 as well, and then motion along the stochastic layer should proceed.
Due to the stochasticity of the motion inside the layer, the variation of I2 should be

5
The whiskered torus is a generalization of a saddle equilibrium point and it is defined as the
connected intersection of the stable and unstable manifolds or, in Arnold language, arriving and
departing whiskers, W  and W C respectively (see [1, 24] for further details).
4 Theory and Applications of the MEGNO 113

0.9
=0.15 p
sx

0.6

0.3

0
-0.3 0 0.3 0.6 0.9
h1

Fig. 4.6 Frequency of the pendulum H1 given by (4.44), !p , and the approximation (4.45), !sx ,
for " D 0:15. The separatrix corresponds to h1 D 0 and the small oscillation frequency, !0 D
p
" p 0:39. Within the oscillation domain h1  0 and !.h1 /  !0 , while for h1 large enough,
2!  2h1

I2

I1

I2
0
1 6

Fig. 4.7 Sketch of diffusion along I2 . The arrow indicates that I2 lies on the stochastic layer of
the resonance !1 D 0

also stochastic, giving rise to diffusion in I2 , as sketched in Fig. 4.7. In consequence,

as I2 would change unboundedly, a gross instability could set up. This is the way
in which Arnold diffusion is described in an heuristic way in [8]. However, in this
model, since the perturbation V vanishes at I1 D 0; 1 D 0, it is possible to build
up a transition chain [1, 24] such that if !2 is irrational, then all tori defined by
114 P.M. Cincotta and C.M. Giordano

I1 D 0; I2 D !2 > 0 are transition tori,6 and when t ! 1, jI2 .t/  I2 .0/j D O.1/,
independently of " and also of . Therefore a large variation of I2 could take
place. Let us state that by large variation we mean that I2 could vary over a finite
domain, which does not imply that it can be proved that I2 changes without any
bound. In fact, this is an open subject of research from a theoretical point of view.
Therefore, any demonstration that diffusion might spread along the resonance web
is quite far to be obtained, as pointed out in [46] and [9].
In the full Hamiltonian (4.46) however, !1 D 0 is just one of the six first order
resonances. Indeed, multiplying the different harmonics and using trigonometric
relationships in V.1 ; 2 ; tI "/ we obtain the following primary resonances at order
" and ":

!1 D 0; !2 D 0; !1 !2 D 0 !1 1 D 0; (4.47)

which are depicted in Fig. 4.8 in frequency space, illustrating their respective widths.
In (4.47) but in the action or energy space, we should use either the approximations

2
1 = 0

1 = 1
1 = 2
I2 = 2

2 = 0
0
0 1 2
I1 = 1

Fig. 4.8 Primary resonances in Arnold model (4.46) in the domain I1 ; I2  0 and considering
.!1 ; !2 / D .I1 ; I2 /. The resonance !1 D 0 has an amplitude V10 D " while for the rest, Vmn D
"  V10

6
Roughly, a transition torus is a whiskered torus satisfying that points belonging to its arriving
whisker W  , intersect any manifold which is transverse to its departing whisker W C . Therefore
a transition chain is a set of k transition tori satisfying that WlC of the l-transition torus intersects

transversally WlC1 of the .l C 1/-transition torus.
4 Theory and Applications of the MEGNO 115

p p
!1  I1 in case I1 2 ", p while !1 D !p .h1 ; "/ for I1 < 2 " .h1 < 0/ or
!1 D 2!p .h1 ; "/pin case I1 & 2 " .h1 > 0/.
For I1 2 " the resonance lines intersect at seven fixed different points
namely, .I1 ; I2 / D .0; 0/; .0; 1/; .1; 1/.7 Hence, as pointed out by Chirikov [8],
the diffusion would spread over all this resonance set. Notice however, that for " 
"  1 the diffusion rate should be negligible along all resonances except for !1 D
0, since this resonance is the one that has the main strength, its amplitude being
", while all the remaining resonances have amplitudes "  ". Indeed, it can be
shownp (see for instance [8] and [9]) that the diffusion rate depends exponentially on
1= Vmn , where Vmn stands for the amplitude of the above considered resonances.
Considering the fully perturbed motion, besides the ones given in (4.47), the full
set of resonances is an integer linear combination of the form

m1 !1 C m2 !2 C m3 D 0; m1 ; m2 ; m3 2 Z; (4.48)
p
where again, !1  I1 or !p .h1 ; "/ depending on the value of I1 =2 ". Therefore, the
true picture of the Arnold web in action space8 should be much more complex than
the one presented in Fig. 4.8, since in thatpcase it is assumed that "  "  1 and
away from the origin it holds that I1 2 " so that !1 D 2!p  I1 . In this case we
expect vertical resonances for m2 D 0, horizontal ones for m1 D 0 and an infinite
but countable set of curves for m1 ; m2 0 (see below).
For the sake of illustration, we present first the result of a numerical experiment
adopting " D 0:05 and  D 0:0001, such that the condition "  "  1 is fulfilled.
Figure 4.9 shows the actual resonances while plotting just the MEGNO values larger
than 2:05 for 106 initial conditions in the I1 ; I2 space with 1 D ; 2 D t D 0 after
a total motion time 104 . This plot should be compared to Fig. 4.8 where the main
resonances, !1 D 0; !2 D 0; !1 D !2 , and !1  1 are clearly distinguished.
The expected width of the main resonance !1 D 0, .I1 /r  0:45 is fully
consistent with the computed one, and regarding the rest of the resonances, their
width should be rather small, close to 4103 , and thus they show up approximately
as a single curve. Some other resonances do not appear as lines, while the one at
!1 D 1 do not arise exactly at I1 D 1. Indeed, if we take the resonance condition
given by (4.48) for m2 0, we can rewrite it using the right value for !1 ,
m1 m3
!2 D  !p .h1 ; "/  ; (4.49)
m2 m2

p
7
Note that for I1  2 "; !1 D 2!p .h1 ; "/ and the resonances should not intersect in the same
set of points, since for instance the resonance !1 D !2 leads to a curve in the .I1 ; I2 / plane that
changes with ".
8
The web of all resonances such as (4.48) for all m1 ; m2 ; m3 2 Z.
116 P.M. Cincotta and C.M. Giordano

= 0.05, = 0.0001

1.4

1.2

0.8
I2

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
I1

Fig. 4.9 Actual resonances in the Arnold model according to a MEGNO mapping on the .I1 ; I2 /-
plane for 1 D ; 2 D t D 0 and "    . Region in black corresponds to chaotic domains,
while those in white correspond to periodic or quasiperiodic motion (see text)

and several of the observed curves follow the very same pattern of !p .h1 ; "/ given
in Fig. 4.6.9 In order to compare both figures recall that in Fig. 4.9, 1 D  so h1
and I1 are simply related by I12 D 2h1 C 4".
Many other resonances are obtained by means of the MEGNO for two sets of
larger values of " and , the results being displayed in Fig. 4.10. These are somewhat
closer to a more realistic case since in a generic Hamiltonian, it is not possible to
reduce the perturbation in such a way that it becomes exponentially small with
respect to the integrable part. The assumption " &  represents a typical situation
in a system involving an integrable Hamiltonian plus a perturbation, which in fact
is an artificial separation in a real problem (see for instance [66]).
In Fig. 4.10 we use the .h1 ; I2 /-plane to display the resonances just to simplify the
comparison of the pattern shown by high order resonances with the plot in Fig. 4.6.
Several resonances of the form (4.49) can be observed, namely those of very low
order, like !1 D 0 of width 2" (measured in h1 ) where the separatrix appears at
h1 D 0. Many other high order ones show up exhibiting a similar pattern as that
of !p .h1 ; "/. Close to the separatrix all resonances accumulate at .h1 ; I1 / D .0; 0/
following the very same behavior as !p .

9
See next page for the estimation of the right position in the .I1 ; I2 /-plane of the !1 D 1 resonance.
4 Theory and Applications of the MEGNO 117

= 0.15, = 0.003 = 0.25, = 0.01

2 2

1.5 1.5

1 1
I2

I2
0.5 0.5

0 0
-0.2 0 0.2 0.4 0.6 -0.4 -0.2 0 0.2 0.4
h1 h1

Fig. 4.10 True pictures through a MEGNO map on the .h1 ; I2 /-plane for 106 initial values of
.h1 ; I2 / for a total motion time 104 . White corresponds to regions of regular motion where Y < 2:05,
while those in black correspond to chaotic motion (Y  2:05). Section at 1 D ; 2 D t D 0 for
h1  2"; I2 D !2 > 0

Let us take for instance the MEGNOs plot for " D 0:15. The resonance !1 D 0
should have a half-width 2" D 0:3, which is fully consistent with the computed
one, and the separatrix appears at h1 D 0 as expected. For the resonance !1 D 1,
the approximate value of I1r D 1 in fact corresponds to hr1 D 0:2. However if we
use the approximation (4.45) for the resonance condition 2!p .h1 ; "/ D 1, it leads
to hr1  0:4 .I1r  1:2/ very close to the computed one. The obtained picture for
" D 0:25 shows a similar structure but, as expected, resonances are wider and many
other high order resonances appear, particularly in the region close to the separatrix.
In both MEGNO contour plots the center of any resonance channel corre-
sponds to 2D elliptic tori while the borders (the stochastic layer or homoclinic
tangle) to 2D hyperbolic tori. At the intersection of two or more resonances a
periodic orbit appears, which could be stable or unstable. In general, the intersection
of two elliptic 2D tori leads to a stable periodic orbit and to a small domain of stable
motion. From Fig. 4.10 we see that the MEGNO plots reveal the stability character
of all the periodic orbits as well as a clear picture of the dynamics on the whole
domain. However, from these plots nothing could be inferred concerning diffusion
in action space, since as we have already pointed out, the MEGNO, as most chaos
indicators, only provides information about the local dynamics of the Hamiltonian
flow. Therefore we only have at hand just the behavior of the flow in any rather small
open domain of every selected point in the grid. Nothing could be said about if it is
possible that a chaotic orbit could explore a finite domain.
118 P.M. Cincotta and C.M. Giordano

4.3.2 Models of Discrete Time

4.3.2.1 The Rational Shifted Standard Map

Let us consider the so-called Rational Shifted Standard Map (RSSMsee [13] for
some additional details). This is a 2D area-preserving discrete dynamical system
given by

y0 D y C " f .x/; x0 D x C " y0 ; (4.50)

with x 2 0; 2/, y 2 0; 2="/, and where

sin .x C '/  sin '

f .x/ D  ; D p : (4.51)
1   cos x 1  2 C 1  2

Notice that (4.50) and (4.51) define some sort of Standard Map (SM) modified in
order to have a no longer symmetric nor entire function f . Indeed, symmetry is lost
through the introduction of the phase ', while the insertion of the denominator, with
the parameter  2 0; 1/, breaks the entire character of f . The quantity  is fixed so
that f has zero average, in order the RSSM be area-preserving.
After rescaling the y-variable by means of y ! "y such that both x; y 2 0; 2/,
the RSSM reads

y0 D y C "2 f .x/; x0 D x C y0 ; (4.52)

and adopts an even closer form to that of the SM.

On expanding

1
D 1 C  cos x C 2 cos2 x C 3 cos3 x C : : : ; (4.53)
1   cos x

and adopting ' D 0 in order to emphasize the comparison with the SM, after taking
into account some trivial trigonometric identities, there results
 
sin x 2  2
f .x/ D D 1C sin x C sin 2x C sin 3x C : : : : (4.54)
1   cos x 4 2 4

To analyze the effect of changing ', we perform the shift x ! x C ' after which
the equation for x in (4.52) remains invariant. On fixing ' D 
 
sin x 2  2
f .x/ D D 1C sin x  sin 2x C sin 3x C : : : ; (4.55)
1 C  cos x 4 2 4

the map for x; y 2 0; 2/ is seen to have a different dependence on the parameters

than in the case in which ' D 0. Therefore a strong dependence of the dynamics on
4 Theory and Applications of the MEGNO 119

' is expected. Herein we will consider the two limiting values, ' D 0; , in order
to reduce the number of free parameters and to clearly show the differences with the
SM.
Thus, from (4.54) and (4.55) it becomes clear that the RSSM shows up all the
harmonics, instead of the solely term in sin x present in the SM. Furthermore, the
resonances width depends not only on "2 , as it is the case in the SM, but on 
as well, and the resonance structure of both maps is similar when  ! 0. In the
RSSM, for  0, all resonances (like y=2 D 0; 1=3; 1=2; 2=3) appear at order "2
and at different orders in , while in the SM, for instance the semi-integer resonance
as y D 1=2 appears at "4 and those at y D 1=3; 2=3 show up at order "6 , so as 
increases the resonances interaction in the RSSM is stronger than in the SM.
The potential function for f  V 0 is

1 n o
V.x/ D ln 1   cos x ;  0: (4.56)


Expanding V.x/ in powers of  and using the 2-periodic in its Fourier form, the
potential U.x/ of the corresponding Hamiltonian has the form
(  1
"2 2 X
U.x/ D 1C cos.x C nt/ C
4 2 4 nD1
1 1
)
 X 2 X
C cos.2x C nt/ C cos.3x C nt/ C : : : ; (4.57)
4 nD1 12 nD1

while the kinetic energy is given by yO 2 =2, being yO D y=2. Thus we can easily see
how resonances appear at different orders in " and .
The MEGNO has been applied to (4.52) in an equispaced grid of 1000  1000
pixels in the domain .x=2; y=2/ 2 0; 1/0; 1/, to obtain YO 2;0 .N/ for N D 11;000
(see (4.40) and discussion below). The results for ' D 0 and ' D  are presented in
Fig. 4.11, for " D 0:8 and two different values of . There the pixels corresponding
to initial conditions of regular behavior are plotted in white and those of chaotic
behavior in black.10 While for ' D 0 the regular regime prevails (plots on the
left), the dynamics for ' D  displays several chaotic domains (plots on the right)
surrounding stochastic layers of resonances or as it seems, a connected chaotic open
domain, but rotational invariant curves (joining the vertical boundaries) still exist.
The variation of ' from 0 to  has a quite notorious effect on the dynamics as
already mentioned in the theoretical discussion. The figures on the top corresponds
to  D 0:1 while those on the bottom to  D 0:2. We can notice that increasing
the value of  changes the stability of the periodic orbit at .0:5; 0:5/ in the case

10
We take slightly different threshold values in the figures just to display the global behavior, since
for ' D 0 the map is mostly regular while for ' D  it is strongly chaotic.
120 P.M. Cincotta and C.M. Giordano

= 0.8, =0.10, =0 = 0.8, =0.10, =

1 1

0.8 0.8

0.6 0.6
y/2

y/2
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/2 x/2
= 0.8, =0.20, =0 = 0.8, =0.20, =
1 1

0.8 0.8

0.6 0.6
y/2

y/2

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/2 x/2

Fig. 4.11 YO 2;0 -levels for the RSSM corresponding to " D 0:8, for ' D 0 (on the left) and ' D 
(on the right). The figures on the top correspond to  D 0:1 and those on the bottom to  D 0:2
Regions of regular behavior are depicted in white and those of chaotic behavior in black. The
threshold values are 2  104 for ' D 0 and 2  102 for ' D 

of ' D 0. Meanwhile, for ' D  the chaotic regime increases as larger values of
 are adopted. Notice that the MEGNO also succeeds in unveiling the high order
resonance structure of this map.

4.3.2.2 The Coupled Rational Shifted Standard Map

Let us now turn to the Coupled Rational Shifted Standard Map (CRSSM), consisting
of two coupled RSSM, defined by

y01 D y1 C "1 f1 .x1 / C C f3 .x1 C x2 / C  f3 .x1  x2 /;

y02 D y2 C "2 f2 .x2 / C C f3 .x1 C x2 /   f3 .x1  x2 /;
4 Theory and Applications of the MEGNO 121

x01 D x1 C "1 y01 ;

x02 D x2 C "2 y02 ; (4.58)

with xi 2 0; 2/, yi 2 0; 2="i /, i D 1; 2, and where

sin .x C 'i / i sin 'i

fi .x/ D  i ; i D q ; i D 1; 3;
1  i cos x 1  2i C 1  2i
(4.59)

with i 2 0; 1/ and again i fixed so that the fi have zero average. Notice that two
coupling terms in .x1 C x2 / and .x1  x2 / have been added, C and  being the
coupling parameters. This map provides a more realistic representation of nonlinear
resonance interactions than two coupled Standard Maps, so its dynamics would well
serve as an improved simple model for many dynamical scenarios.
Again as in the RSSM, rescaling the y-variables, the CRSSM can be recast as

y01 D y1 C "21 f1 .x1 / C "1 C f3 .x1 C x2 / C "1  f3 .x1  x2 /;

y02 D y2 C "22 f2 .x2 / C "2 C f3 .x1 C x2 /  "2  f3 .x1  x2 /;
x01 D x1 C y01 ;
x02 D x2 C y02 ; (4.60)

where .xi ; yi / 2 0; 2/  0; 2/.
The full set of primary resonances is determined by

k1 y1 C k2 y2 C 2k3 D 0; k1 ; k2 ; k3 2 Z: (4.61)

Therefore, in the action plane, horizontal resonances correspond to the uncoupled

.x2 ; y2 / map and appear for k1 D 0, the vertical ones correspond to the uncoupled
.x1 ; y1 / map obtained by setting k2 D 0, while the coupling resonances given by
y2 D .k1 y1 C 2k3 /=k2 are dense (but countable) in the .y1 ; y2 /-space.
The MEGNO has been computed for an equispaced grid of 1000  1000 pixels
in the domain .y1 =2; y2 =2/ 2 0; 1/  0; 1/. The initial values for the remaining
variables are x1 D 0; x2 D 0. The right-stop condition described in Sect. 4.2.4 has
been applied so that for each initial condition the iteration is stopped after N iterates,
with 10000 < N < 11000, when the distance between the N-th iteration of the map
and the initial condition is minimum.
A difference should be remarked with the action space of Arnold model discussed
in the previous section. Indeed, by making the cross product .x1 ; y1 /  .x2 ; y2 / at
x1 D x2 D 0, depending on the adopted value of ' and considering Fig. 4.11, we
should expect that in the .y1 ; y2 /-plane, only do the hyperbolic 2D tori (or to be
precise, the homoclinic tangle) show up for ' D 0 while for ' D , the picture
should be similar to that of Fig. 4.10, since both, the elliptic and hyperbolic 2D
tori would be present, as well as the nearly resonant 3D tori that are trapped in
resonances.
122 P.M. Cincotta and C.M. Giordano

1= 2 = 0.3, + = = 0.05, 1 = 2 = 3 = 0.25, 1= 2= 3 = 0 1= 2 = 0.3, + = = 0.05, 1 = 2 = 3 = 0.25, 1= 2= 3 =

1 1

0.8 0.8

0.6 0.6
y2 / 2

y2 / 2
0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
y1 / 2 y1 / 2

Fig. 4.12 YO 2;0 -levels for the CRSSM for different values of the parameters 'i , 0 in the plot on the
left and  in that on the right. The contour plots correspond to YO 2;0 binned in three intervals; pixels
corresponding to initial conditions of regular behavior are plotted in white, and those of chaotic
behavior in black (red). With gray (green) we identify mild chaotic or even quasiperiodic motion

In order to illustrate the efficiency of the MEGNO to display the full dynamics
of this 4D map, we show the results for i D 0, i.e, 'i D 0 and 'i D ; i D
1; 2; 3, "1 D "2 D 0:3, 1 D 2 D 3 D 0:25, and C D  D 0:05, given
in Fig. 4.12. The contour-like plots exhibit the obtained values for log.YO 2;0 / given
by (4.40) scaled in order to range from 5  log.YO 2;0 / < 3, to log.YO 2;0 /  3.
Recall that YO 2;0 ! 0 for quasiperiodic motion while YO 2;0 > 0 indicates chaotic
dynamics. The initial conditions corresponding to regular orbits have been depicted
in white, while those in black (red) are chaotic. The orbits holding intermediate
values of YO 2;0 are plotted in gray (green) and considered as, possibly, quasiperiodic
or mildly chaotic.
Though the Arnold web could be obtained also by means of other chaos
indicators, let us mention that since the MEGNO and its generalized version have a
clear threshold value, both of them allow for separating regular and chaotic orbits,
providing for the latter a measure of the mLCE. Therefore, instead of performing
an automatic contour plot it is possible to select the MEGNO ranges to be depicted.
This has several benefits when we are interested in separating regular motion and
chaotic motion with different degrees of hyperbolicity.
The resonances defined by (4.61) can be clearly distinguished. The wider ones
are the integer resonances of the uncoupled maps, however for ' D 0 almost all
resonances have a rather small width due to the fact that in such a case we see
only the hyperbolic part, except for a few high order resonances which show up
as narrow channels. The opposite picture corresponds to ' D , where most
resonances reveal their 2D elliptic and hyperbolic tori. Note that the periodic orbit
at each resonance intersection is, as expected, unstable for ' D 0, while it is stable
4 Theory and Applications of the MEGNO 123

1= 2 = 0.3, + = = 0.05, 1 = 2 = 3 = 0.25, 1= 2= 3 = 0 1= 2 = 0.3, + = = 0.05, 1 = 2 = 3 = 0.25, 1= 2= 3 =

0.4 0.4

0.35 0.35

0.3 0.3

y2 / 2
y2 / 2

0.25 0.25

0.2 0.2

0.15 0.15
0.15 0.2 0.25 0.3 0.35 0.4 0.15 0.2 0.25 0.3 0.35 0.4
y1 / 2 y1 / 2

Fig. 4.13 Zoom of Fig. 4.12 in the domain 0:15  y1 ; y2  0:4

for ' D . The complement of the set of 2D elliptic tori, 2D hyperbolic tori and
periodic orbits, corresponds to 3D tori, where the motion is quasiperiodic.
In Fig. 4.13 we present a zoom of Fig. 4.12 corresponding to the region 0:15 
y1 ; y2  0:4. In these YO 2;0 -contour plots we can distinguish many resonances of very
high order as well as the dynamics in the resonance crossings. We can also notice
how the stable and unstable manifolds bend to lead to either a regular or a chaotic
domain. These manifolds are very important since they are the objects able to carry
the motion arriving along one of the resonances either to the other parts of the
resonance or to a different resonance. Besides, from the numerical results provided
by the MEGNO, we could infer the true effect of the intersection of resonances of
different order.

4.4 Comparison of Different Chaos Indicators

Some comparisons between particular indicators in the framework of specific

studies were carried out, as those given for instance in [3, 47] and [37]. However,
no systematic comparisons of the performance of several chaos indicators had been
accomplished up to our comparative studies given in [48] and [49], which we briefly
describe in the forthcoming subsections.

4.4.1 Comparative Studies for a Hamiltonian Flow

As already mentioned, the standard MEGNO has become a widespread technique

for the study of Hamiltonian systems, particularly in the field of dynamical
astronomy and astrodynamics, then, a comparison with other dynamical indicators
124 P.M. Cincotta and C.M. Giordano

was in order. Therefore, in [49] a rather complex nonlinear system was addressed
that reproduces many characteristics of real elliptical galaxies, namely, the self-
consistent model introduced in [56]. Such a model was used as the scenario for a
comprehensive comparison between the MEGNO and the mLCE, and even with the
FLI. A detailed numerical and statistical study of a sample of orbits in the triaxial
galactic system showed that the MEGNO is a suitable fast indicator to separate
regular from chaotic motion and that it is particularly useful to investigate the nature
of orbits that have a small but positive mLCE.
A rather good correlation was obtained between the MEGNO and the mLCE
values for short, moderate and large integration times when considering just chaotic
orbits, while the MEGNO provided much better results for regular motion. The
FLI also looked like a reliable fast indicator, but since it has no reference value
for regular motion, it might be useful to explore the phase space rather than to
investigate the nature of a given orbit, unless of course the time evolution of such
indicator was followed.
In [50] the same self-consistent triaxial stellar dynamical model was studied
for different energy levels by means of some selected variational indicators and
spectral analysis methods. Therein, the comparison of several variational indicators
on different scenarios was addressed. Indeed, the Average Power Law Exponent
(APLE) [47] and the MEGNOs estimation of the mLCE by a least squares fit of
its time evolution were compared. The spectral analysis method selected for that
investigation was the Frequency Modified Fourier Transform (FMFT) [63], which
is just a slight variation of the FMA. Besides, a comparative study of the APLE,
the FLI, the Orthogonal Fast Lyapunov Indicator (OFLI) [16] and the estimation
of the mLCE obtained from the MEGNOs slope yielded as a result that the latter
could be an appropriate alternative to the MEGNO when studying large samples
of initial conditions. In fact, it succeeded in separating the chaotic and the regular
components and in identifying the different levels of hyperbolicity (or exponential
rate of divergence of nearby orbits) as well. Further, it turned out to be more reliable
than the FMFT while describing chaotic domains.

4.4.2 Comparative Studies for Maps

In [48] the efficiency of several variational indicators of chaos when applied to

mappings was compared. We considered the mLCE, the MEGNO, the Smaller
Alignment Index (SALI) [67], its generalized version, the Generalized Alignment
Index (GALI) [69], the FLI [19], the Dynamical Spectra of stretching numbers
(SSN) [71] and the corresponding Spectral Distance (D) and the Relative Lyapunov
Indicator (RLI) [62], which is based on the evolution of the difference between
two close orbits. As a result of several experiments presented therein concerning
two different 4D mappings, namely, a variant of Froeschls symplectic mapping
[14, 17, 67, 68] and a system comprising two coupled Standard Maps, it was shown
that a package composed of the FLI and the RLI (when a global analysis of the phase
4 Theory and Applications of the MEGNO 125

portrait is pursued) and of the MEGNO and the SALI (if the objective is the analysis
of individual orbits) turned out to be the best choices to yield a good description of
the dynamics of the systems under study.

4.5 Further Applications of the MEGNO

In recent years the MEGNO has been widely used mainly in the field of dynamics
of multi-planet extrasolar systems to address stability and habitability studies as
well as in the solar system, galactic dynamics, astrobiology and chemistry. Herein,
we include some references that would serve as illustration of this issue. We refer
the reader to the original papers for details regarding the concomitant physical
problems.
For extrasolar dynamical studies see for instance [20, 2632, 52, 55, 61]. For
research concerning Solar System dynamics we refer for example to [22] and [37],
where the MEGNO technique is applied to the investigation of the dynamics of
Jovian irregular satellites, to [60] which is devoted to the resonant structure of
Jupiters Trojan asteroids, its long-term stability and diffusion or to [21] where the
evection resonance is considered. Interesting results in astrobiology are obtained
while studying the dynamical habitability of exoplanetary systems (see [15, 36, 54]).
Further applications of the MEGNO can be found in the study of space debris
motion as in [38, 39, 70] among others, and of the chaotic motion of geosynchronous
satellites as in [6, 40, 41].
As far as galactic studies are concerned, we can refer for instance to [7, 51, 72].
The use of this chaos indicator in rigid-body motion can be found in [2], and
in the realm of chemistry in the analysis of intramolecular dynamics [65], or while
revisiting the problem of driven coupled Morse oscillators [64]. Finally, bifurcations
and chaos in different scenarios are studied by means of the MEGNO for instance
in [4, 23, 25, 59], among many others.

4.6 Discussion

In this review we have described a rather simple technique, the Mean Exponential
Growth factor of Nearby Orbits (MEGNO), which succeeds in providing detailed
indications on the dynamics of continuous dynamical systems and maps. The
intrinsic connection of this technique with the FLI and the mLCE is also presented.
The MEGNO furnishes an efficient algorithm that allows not only to clearly
identify regular and irregular motion as well as stable and unstable periodic orbits,
but also to obtain a quite good estimate of the mLCE in comparatively very short
evolution times, for both ordered and chaotic components of phase space. This is
a particular feature of this indicator that is not shared with many other techniques.
In fact, we could deem that the derivation of the mLCE by a least squares fit of the
126 P.M. Cincotta and C.M. Giordano

time evolution of the MEGNO is an alternative algorithm to get the time-scale for
exponential divergence of nearby orbits but in rather short times in comparison with
the classical approach.
Thus, by the application of this single tool it is possible to grasp the dynamics of
the system over the whole phase space, and this procedure is a first attempt to get
dynamical information about the motion using the whole orbit.
Moreover, there exists profuse numerical evidence of the MEGNO being a fast
indicator capable of unveiling the hyperbolic structure of the phase space, as well
as yielding a clear picture of the resonance structure in any dimensional systems.
Besides, the MEGNO is shown to provide the actual size of a resonance of very
high order as well as to reveal its internal structure.
Let us mention that the application of this technique to many different dynamical
systems along the literature shows that it could be useful to investigate stability
domains in exoplanetary models, chemical dynamics, space debris as well as to
discuss purely theoretical features like bifurcation analysis.
Finally, regarding which is the more suitable chaos detection tool (based on the
evolution of the tangent vector) we claim from our experience and in view of the
nowadays available computational resources that, it is just a matter of the gained
expertise on the adopted technique. However, let us say that the MEGNO is the one
with a theoretical threshold value that allows to clearly separate regular from chaotic
motion as well as it provides an accurate estimate of the mLCE by means of a very
simple algorithm.
Therefore, a combination of any such indicator together with an accurate spectral
technique, like the FMA, would be the best option to display the full dynamics of
nonlinear systems which in general present a divided phase space.

Acknowledgements We should state clear that also Prof. C. Sim worked very hard developing
this technique. Further, we are grateful to him for his illuminating discussions, comments and
suggestions that helped us to write this chapter in a comprehensive way and any mistake is nothing
but our own responsibility.
We also deeply appreciate the recommendations of the two thoughtful reviewers of this effort
that serve to improve the final version.
This work was supported by grants from CONICET, UNLP and Instituto de Astrofsica de La
Plata.

References

1. Arnold, V.I.: Soviet Math. Dokl. 5, 581585 (1964)

2. Barrio, R., Blesa, F., Elipe, A.: J. Astronaut. Sci. 54(34), 359368 (2006)
3. Barrio, R., Borczyk, W., Breiter, S.: Chaos Solitons Fractals 40, 16971714 (2009)
4. Barrio, R., Blesa, F., Serrano, S.: Int. J. Bifurcat. Chaos 20(5), 12931319 (2010)
5. Benettin, G., Galgani, L., Giorgilli, A., Strelcyn, J.M.: Meccanica 15(1) Part I: Theory, 920;
Part II: Numerical applications, 2130 (1980)
6. Breiter, S., Wytrzyszczak, I., Melendo, B.: Adv. Space Res. 35(7), 13131317 (2006)
7. Breiter, S., Fouchard, M., Ratajczak, R.: Mon. Not. R. Astron. Soc. 383(1), 200208 (2008)
4 Theory and Applications of the MEGNO 127

8. Chirikov, B.V.: Phys. Rep. 52, 263379 (1979)

9. Cincotta, P.M.: New Astron. Rev. 46, 1339 (2002)
10. Cincotta, P.M., Sim, C.: Celest. Mech. Dyn. Astron. 73(14), 195209 (1999)
11. Cincotta, P.M., Sim, C.: In: Gurzadyan, V.G., Ruffini, R. (eds.) The Chaotic Universe.
Advanced Series in Astrophysics and Cosmology, vol. 10, pp. 247258. World Scientific
(2000). ISBN: 978-981-279-362-1
12. Cincotta, P.M., Sim, C.: Astron. Astrophys. Suppl. 147, 205228 (2000)
13. Cincotta, P.M., Giordano, C.M., Sim, C.: Physica D 182, 11178 (2003)
14. Contopoulos, G., Giorgilli, A.: Meccanica 23, 1928 (1988)
15. Dvorak, R., Pilat-Lohinger, E., Bois, E., Schwarz, R., Funk, B., Beichman, C., Danchi, W.,
et al.: Astrobiology 10(1), 3343 (2010)
16. Fouchard, M., Lega, E., Froeschl, Ch., Froeschl, Cl.: Celest. Mech. Dyn. Astron. 83, 205
222 (2002)
17. Froeschl, C.: Astron. Astrophys. 16, 172189 (1972)
18. Froeschl, C., Lega, E.: Celest. Mech. Dyn. Astron. 78, 167195 (2000)
19. Froeschl, C., Gonczi, R., Lega, E.: Planet. Space Sci. 45, 881886 (1997)
20. Frouard, J., Compre, A.: Icarus 220(1), 149161 (2012)
21. Frouard, J., Fouchard, M., Vienne, A.: Astron. Astrophys. 515(7) (2010). doi: 10.1051/0004-
6361/200913048
22. Frouard, J., Vienne, A., Fouchard, M.: Astron. Astrophys. 532 (2011). doi: 10.1051/0004-
6361/201015873
23. Gelfreich, V., Sim, C., Vieiro, A.: Physica D 243(1), 92110 (2013)
24. Giorgilli, A.: In: Benest, D., Froeschl, C. (eds.) Les Methodes Modernes de la Mecanique
Celeste, pp. 249284. Frontires (1990). ISBN: 2-8633209-2
25. Gonchenko, S.V., Sim, C., Vieiro, A.: Nonlinearity 26(3), 621678 (2013)
26. Gozdziewski, K., Migaszewski, C.: Mon. Not. R. Astron. Soc. Lett. 397(1), L16L20 (2009)
27. Gozdziewski, K., Konacki, M., Maciejewski, A.J.: Astrophys. J. Lett. 622(2I), 11361148
(2005)
28. Gozdziewski, K., Konacki, M., Wolszczan, A.: Astrophys. J. Lett. 619(2I), 10841097 (2005)
29. Gozdziewski, K., Breiter, S., Borczyk, W.: Mon. Not. R. Astron. Soc. 383(3), 989999 (2006)
30. Gozdziewski, K., Konacki, M., Maciejewski, A.J.: Astrophys. J. Lett. 645(1I), 688703 (2006)
31. Gozdziewski, K., Migaszewski, C., Musielinski, A.: In: Proceedings of the International
Astronomical Union 3, Symposium 249, pp. 447460 (2007)
32. Gozdziewski, K., Slonina, M., Migaszewski, C., Rozenkiewicz, A.: Mon. Not. R. Astron. Soc.
430(1), 533545 (2013)
33. Guzzo, M., Lega, E., Froeschl, C.: Physica D 163, 125 (2002)
34. Hairer, E., Nrsett, S., Wanner, G.: Solving Ordinary Differential Equations I: Nonstiff
Problems. Springer, New York (1987)
35. Hnon, M., Heiles, C.: Astron. J. 69, 73 (1964)
36. Hinse, T.C., Michelsen, R., Jorgensen, U.G., Gozdziewski, K., Mikkola, S.: Astron. Astrophys.
488(3), 11331147 (2008)
37. Hinse, T.C., Christou, A.A., Alvarellos, J.L.A., Gozdziewski, K.: Mon. Not. R. Astron. Soc.
404(2), 837857 (2010)
38. Hubaux, C., Lemitre, A., Delsate, N., Carletti, T.: Adv. Space Res. 49(10), 14721486 (2012)
39. Hubaux, Ch., Libert, A.S., Delsate, N., Carletti, T.: Adv. Space Res. 51(1), 2538 (2013)
40. Kuznetsov, E.D., Kaiser, G.T.: Cosmic Res. 45(4), 359367 (2007)
41. Kuznetsov, E.D., Kudryavtsev, A.O.: Russ. Phys. J. 52(8), 841849 (2009)
42. Laskar, J.: Icarus 88, 266291 (1990)
43. Laskar, J.: Physica D 67, 257281 (1993)
44. Laskar, J.: Astron. Astrophys. 287(1), L9L12 (1994)
45. Laskar, J., Froeschl, C., Celletti, A.: Physica D 56(23), 253269 (1992)
46. Lochak, P.: In: Sim, C. (ed.) Hamiltonian Systems with Three or More Degrees of Freedom.
NATO ASI Series, pp. 168183. Kluwer Academic Publishers, The Netherlands (1999)
47. Lukes-Gerakopoulos, G., Voglis, N., Efthymiopoulos, C.: Physica A 387, 19071925 (2008)
128 P.M. Cincotta and C.M. Giordano

48. Maffione, N.P., Giordano, C.M., Cincotta, P.M.: Celest. Mech. Dyn. Astron. 111, 285307
(2011)
49. Maffione, N.P., Giordano, C.M., Cincotta, P.M.: Int. J. Non Linear Mech. 46, 2334 (2011)
50. Maffione, N.P., Darriba, L.A., Cincotta, P.M., Giordano, C.M.: Mon. Not. R. Astron. Soc.
429(3), 27002717 (2013)
51. Manos, T., Bountis, T., Skokos, C.: J. Phys. A Math. Theor. 46(25), 254017 (2013)
52. Mart, J.G., Giuppone, C.A., Beaug, C.: Mon. Not. R. Astron. Soc. 433(2), 928934 (2013)
53. Mestre, M., Cincotta, P.M., Giordano, C.M.: Mon. Not. R. Astron. Soc. Lett. 404, L100L103
(2011)
54. Migaszewski, C., Gozdziewski, K., Hinse, T.C.: Mon. Not. R. Astron. Soc. 395(3), 12041212
(2008)
55. Migaszewski, C., Slonina, M., Gozdziewski, K.: Mon. Not. R. Astron. Soc. 427(1), 770789
(2012)
56. Muzzio, J.C., Carpintero, D.D., Wachlin, F.C.: Celest. Mech. Dyn. Astron. 91(12), 173 (2005)
57. Nez, J.A., Cincotta, P.M., Wachlin, F.C.: Celest. Mech. Dyn. Astron. 54(12), 4353 (1996)
58. Prince, P., Dormand, J.: J. Comput. Appl. Math. 35, 67 (1981)
59. Puig, J., Sim, C.: Regul. Chaotic Dyn. 16(1), 6178 (2011)
60. Robutel, P., Gabern, F.: Mon. Not. R. Astron. Soc. 372(4), 14631482 (2006)
61. Saito, M.M., Tanikawa, K., Orlov, V.V.: Celest. Mech. Dyn. Astron. 116(1), 110 (2013)
62. Sndor, Z., rdi, B., Efthymiopoulos, C.: Celest. Mech. Dyn. Astron. 78, 113123 (2000)
63. Sidlichovsk, M., Nesvorn, D.: Celest. Mech. Dyn. Astron. 65, 137148 (1997)
64. Sethi, A., Keshavamurthy, S.: Mol. Phys. 110(910), 717727 (2012)
65. Shchekinova, E., Chandre, C., Lan, Y., Uzer, T.: J. Chem. Phys. 121(8), 34713477 (2004)
66. Sim, C.: In: Broer, H.W., Krauskopf, B., Vegter, G. (eds.) Global Analysis of Dynamical
Systems, pp. 373390. IOP Publishing, Bristol (2001)
67. Skokos, Ch.: J. Phys. A Math. Gen. 34, 1002910043 (2001)
68. Skokos, Ch., Contopoulos, G., Polymilis, C.: Celest. Mech. Dyn. Astron. 65, 223251 (1997)
69. Skokos, Ch., Bountis, T., Antonopoulus, Ch.: Physica D 231, 3054 (2007)
70. Valk, S., Delsate, N., Lemaitre, A., Carletti, T.: Adv. Space Res. 43(10), 15091526 (2009)
71. Voglis, N., Contopoulos, G., Efthymiopoulos, C.: Celest. Mech. Dyn. Astron. 73, 211220
(1999)
72. Zorzi, A.F., Muzzio, J.C.: Mon. Not. R. Astron. Soc. 423(2), 19551963 (2012)
Chapter 5
The Smaller (SALI) and the Generalized (GALI)
Alignment Indices: Efficient Methods of Chaos
Detection

Charalampos (Haris) Skokos and Thanos Manos

Abstract We provide a concise presentation of the Smaller (SALI) and the

Generalized Alignment Index (GALI) methods of chaos detection. These are
efficient chaos indicators based on the evolution of two or more, initially distinct,
deviation vectors from the studied orbit. After explaining the motivation behind the
introduction of these indices, we sum up the behaviors they exhibit for regular and
chaotic motion, as well as for stable and unstable periodic orbits, focusing mainly
on finite-dimensional conservative systems: autonomous Hamiltonian models and
symplectic maps. We emphasize the advantages of these methods in studying the
global dynamics of a system, as well as their ability to identify regular motion
on low dimensional tori. Finally we discuss several applications of these indices
to problems originating from different scientific fields like celestial mechanics,
galactic dynamics, accelerator physics and condensed matter physics.

The original version of this chapter was revised.

An erratum to this chapter can be found at DOI 10.1007/978-3-662-48410-4_9
Ch. (Haris) Skokos ()
Department of Mathematics and Applied Mathematics, University of Cape Town, Rondebosch
7701, South Africa
e-mail: [email protected]
T. Manos
Center for Applied Mathematics and Theoretical Physics (CAMTP), University of Maribor,
Krekova 2, SI-2000 Maribor, Slovenia
School of Applied Sciences, University of Nova Gorica - Vipavska 11c, SI-5270 Ajdovcina,
Slovenia
Institute of Neuroscience and Medicine Neuromodulation (INM-7), Research Center Jlich,
52425 Jlich, Germany
e-mail: [email protected]

Springer-Verlag Berlin Heidelberg 2016 129

Ch. Skokos et al. (eds.), Chaos Detection and Predictability, Lecture Notes
in Physics 915, DOI 10.1007/978-3-662-48410-4_5
130 Ch. (Haris) Skokos and T. Manos

5.1 Introduction and Basic Concepts

A fundamental aspect in studies of dynamical systems is the identification of chaotic

behavior, both locally, i.e. in the neighborhood of individual orbits, and globally,
i.e. for large samples of initial conditions. The most commonly used method to
characterize chaos is the computation of the maximum Lyapunov exponent (mLE)
1 . In general, Lyapunov exponents (LEs) are asymptotic measures characterizing
the average rate of growth or shrinking of small perturbations to orbits of dynamical
systems. They were introduced by Lyapunov [55] and they were applied to
characterize chaotic motion by Oseledec in [70], where the Multiplicative Ergodic
Theorem (which provided the theoretical basis for the numerical computation of the
LEs) was stated and proved. For a recent review of the theory and the numerical
evaluation of LEs the reader is referred to [81]. The numerical evaluation of the
mLE was achieved in the late 1970s [11, 32, 69] and allowed the discrimination
between regular and chaotic motion. This evaluation is performed through the time
evolution of an infinitesimal perturbation of the orbits initial condition, which is
described by a deviation vector from the orbit itself. The evolution of the deviation
vector is governed by the so-called variational equations [32].
In practice, 1 is evaluated as the limit for t ! 1 of the finite time maximum
Lyapunov exponent

1 kw.t/k
1 .t/ D ln ; (5.1)
t kw.0/k

where t denotes the time and kw.0/k, kw.t/k are the Euclidean norms1 of the
deviation vector w at times t D 0 and t > 0 respectively. Thus

1 D lim 1 .t/: (5.2)

t!1

The computation of the mLE was extensively used for studying chaos and it is
still implemented nowadays for this purpose. Nevertheless, one of its major practical
disadvantages is the slow convergence of the finite time Lyapunov exponent (5.1)
to its limit value (5.2). Since 1 .t/ is influenced by the whole evolution of the
deviation vector, the time needed for it to converge to 1 is not known a priori,
and in many cases it may become extremely long. This delay can result in CPU-
time expensive computations, especially when the study of many orbits is required
for the global investigation of a system. In order to overcome this problem several
other fast chaos detection techniques have been developed over the years; some of
which are presented in this volume.
Throughout this chapter we consider finite-dimensional conservative dynamical
systems and in particular, autonomous Hamiltonian models and symplectic maps
(except from Sect. 5.4.3 where a time dependent Hamiltonian system is studied).

1
We note that the value of 1 is independent of the used norm.
5 The SALI and the GALI Methods of Chaos Detection 131

In these systems regular motion occurs on the surface of a torus in the systems
phase space and is characterized by 1 D 0. Any deviation vector w.0/ from a
regular orbit eventually falls on the tangent space of this torus and its norm will
approximately grow linearly in time, i.e. eventually becoming proportional to t,
kw.t/k / t. Consequently, 1 .t/ / ln t=t, which practically means that 1 .t/ tends
asymptotically to zero following the power law t1 because the values of ln t change
much slower than t as time grows (see for example [11, 26] and Sect. 5.3 of [81]). On
the other hand, in the case of chaotic orbits the use of any initial deviation vector in
(5.1) and (5.2) practically leads to the computation of the mLE 1 > 0 because this
vector eventually is stretched towards the direction associated to the mLE, assuming
of course that 1 > 2 , with 2 being the second largest LE. We note here that,
from the first numerical attempts to evaluate the mLE [11, 32] it became apparent
that a random choice of the initial deviation vector w.0/ leads with probability one
to the computation of 1 . This means that, the choice of w.0/ does not affect the
limiting value of 1 .t/, but only the initial phases of its evolution. This behavior
introduces some difficulties when we want to evaluate the whole spectrum of LEs
of chaotic orbits because any set of initially distinct deviation vectors eventually
end up to vectors aligned along the direction defined by the mLE. It is worth-noting
that even in cases where we could theoretically know the initial choice of deviation
vectors which would lead to the evaluation of LEs other than the maximum one, the
unavoidable numerical errors in the computational procedure will lead again to the
computation of the mLE [15]. This problem was bypassed by the development of a
procedure based on repeated orthonormalizations of the evolved deviation vectors
[10, 1215, 78, 95].
Although the eventual coincidence of distinct initial deviation vectors for chaotic
orbits with 1 > 2 was well-known from the early 1980s, this property was
not directly used to identify chaos for about two decades until the introduction of
the Smaller Alignment Index (SALI) method in [79]. In the 1990s some indirect
consequences of the fact that two initially distinct deviation vectors eventually
coincide for chaotic motion, while they will have different directions on the tangent
space of the torus for regular ones, were used to determine the nature of orbits, but
not the fact itself. In particular, in [91] the spectra of what was named the stretching
number, i.e. the quantity

kw.tCt/k
ln kw.t/k
D ; (5.3)
t
where t is a small time step, were considered. The main outcome of that paper was
that the spectra for two different initial deviations are the same for chaotic orbits,
but different for ordered orbits, as was stated in the abstract of Voglis et al. [91].
This feature was later quantified in [92] by the introduction of a quantity measuring
the difference of two spectra, the so-called spectral distance. In [92] it was
shown that this quantity attains constant, positive values for regular orbits, while it
becomes zero for chaotic ones. It is worth noting that in [91] it was explained that the
observed behavior of the two spectra was due to the fact that the deviation vectors
132 Ch. (Haris) Skokos and T. Manos

eventually coincide for chaotic orbits, producing the same sequences of stretching
numbers, while they remain different for regular ones resulting in different spectra of
stretching numbers. Nevertheless, instead of directly checking the matching (or not)
of the two deviation vectors the method developed in [91, 92] requires unnecessary,
additional computations as it goes through the construction of the two spectra and
the evaluation of their distance. Naturally, this procedure is influenced by the
whole time evolution of the deviation vectors, which in turn results in the delay
of the matching of the two spectra with respect to the matching of the two deviation
vectors.
Apparently, the direct determination of the possible coincidence (or not) of the
deviation vectors is a much faster and more efficient approach to reveal the regular
or chaotic nature of orbits than the evaluation of the spectral distance, as it requires
less computations (see [79] for a comparison between the two approaches). This
observation led to the introduction in [79] of the SALI method which actually checks
the possible coincidence of deviation vectors, while the later introduced Generalized
Alignment Index (GALI) [84] extends this criterion to more deviation vectors. As
we see in Sect. 5.3 this extension allows the correct characterization of chaotic orbits
also in the case where the spectrum of the LEs is degenerate and the second, or even
more, largest LEs are equal to 1 .
In order to illustrate the behaviors of both the SALI and the GALI methods
for regular and chaotic motion we use in this chapter some simple models of
Hamiltonian systems and symplectic maps.
In particular, as a two degrees of freedom (2D) Hamiltonian model we consider
the well-known Hnon-Heiles system [43], described by the Hamiltonian

1 2 1 1
H2 D .p1 C p22 / C .q21 C q22 / C q21 q2  q32 : (5.4)
2 2 3
We also consider the 3D Hamiltonian system

3
X !i
H3 D .q2i C p2i / C q21 q2 C q21 q3 ; (5.5)
iD1
2

initially studied in [15, 32]. Note that !i in (5.5) are some constant coefficients. As
a model of higher dimensions we use the ND Hamiltonian

N 
1X 2 X 1
N
2 1 4
HN D p C .qiC1  qi / C .qiC1  qi / ; (5.6)
2 iD1 i iD0
2 4

which describes a chain of N particles with quadratic and quartic nearest neighbor
interactions, known as the FermiPastaUlam model (FPU-) [36], where q0 D
qNC1 D 0. In all the above-mentioned ND Hamiltonian models, qi , pi , i D 1; 2; : : : N
are respectively the generalized coordinates and the conjugate momenta defining the
2N-dimensional (2Nd) phase space of the system.
5 The SALI and the GALI Methods of Chaos Detection 133

As a symplectic map model we consider in our presentation the 2M-dimensional

(2Md) system of coupled standard maps studied in [51]

x0j D xj C y0j

Kj         

y0j D yj C sin 2xj  sin 2 xjC1  xj C sin 2 xj1  xj ;

2 2
(5.7)

where j D 1; 2; : : : ; M is the index of each standard map, Kj and  are the models
parameters and the prime (0 ) denotes the new values of the variables after one
iteration of the map. We note that each variable is given modulo 1, i.e. 0  xj < 1,
0  yj < 1 and also that the conventions x0 D xM and xMC1 D x1 hold.
In order to make this chapter more focused and easier to read we decided not
to present any analytical proofs for the various mathematical statements given in
the text; we prefer to direct the reader to the publications where these proofs can
be found. Nevertheless, we want to emphasize here that all the laws describing the
behavior of the SALI and the GALI have been obtained theoretically and they are
not numerical estimations or fits to numerical data. Indeed, these laws succeed to
accurately reproduce the evolution of the indices in actual numerical simulations,
some of which are presented in the following sections.
The chapter is organized as follows. In Sect. 5.2 the SALI method is presented
and the behavior of the index for regular and chaotic orbits is discussed. Section 5.3
is devoted to the GALI method. After explaining the motivation that led to the
introduction of the GALI, the definition of the index is given and its practical
computation is discussed in Sect. 5.3.1. Then, in Sect. 5.3.2 the behavior of the
index for regular and chaotic motion is presented and several example orbits
of Hamiltonian systems and symplectic maps of various dimensions are used to
illustrate these behaviors. The ability of the GALI to identify motion on low
dimensional tori is presented in Sect. 5.3.3, while Sect. 5.3.4 is devoted to the
behavior of the index for stable and unstable periodic orbits. In Sect. 5.4 several
applications of the SALI and the GALI methods are presented. In particular, in
Sect. 5.4.1 we explain how the SALI and the GALI can be used for understanding
the global dynamics of a system, while specific applications of the indices to
various dynamical models are briefly discussed in Sect. 5.4.2. The particular case
of time dependent Hamiltonians is considered in Sect. 5.4.3. Finally, in Sect. 5.5 we
summarize the advantages of the SALI and the GALI methods and briefly discuss
some recent comparative studies of different chaos indicators.

5.2 The Smaller Alignment Index (SALI)

The idea behind the SALIs introduction was the need for a simple, easily
computed quantity which could clearly identify the possible alignment of two
multidimensional vectors. As has been already explained, it was well-known that
134 Ch. (Haris) Skokos and T. Manos

any two deviation vectors from a chaotic orbit with 1 > 2 are stretched towards
the direction defined by the mLE, eventually becoming aligned having the same or
opposite directions. Thus, it would be quite helpful to devise a quantity which could
clearly indicate this alignment.
Since we are only interested in the direction of the two deviation vectors and not
in their actual size, we can normalize them before checking their alignment. This
process also eliminates the problem of potential numerical overflow due to vectors
growth in size, which appears especially in the case of chaotic orbits. So in practice,
we let the two deviation vectors evolve under the systems dynamics (according
to the variational equations for Hamiltonian models, or the so-called tangent map
for symplectic maps) normalizing them after a fixed number of evolution steps to
a predefined norm value. For simplicity in our presentation we consider the usual
Euclidean norm (denoted by k  k) and renormalize the evolved vectors to unity.
In the case of chaotic orbits this procedure is schematically shown in Fig. 5.1
O 1 .0/, w
where the two initially distinct unit deviation vectors2 w O 2 .0/ converge to the
same direction. We emphasize that Fig. 5.1 is just a schematic representation on the
plane of the real deviation vectors which are objects evolving in multidimensional
spaces. Since the mLE 1 > 0 denotes the mean exponential rate of each vectors
stretching, they are elongated at some later time t > 0,3 becoming w1 .t/, w2 .t/,

Fig. 5.1 Schematic representation of the evolution of two deviation vectors and of the correspond-
ing SALI for a chaotic orbit. Two initially distinct unit deviation vectors w O 1 .0/, w
O 2 .0/ from point
P.0/ of a chaotic orbit become w1 .t/, w2 .t/ after some time t > 0 when the orbit reaches point P.t/,
with wO 1 .t/, w
O 2 .t/ being the unit vectors along these directions. The length of the shortest diagonals
of the grey-shaded parallelograms defined by w O 1 .0/, w
O 2 .0/ and w
O 1 .t/, w
O 2 .t/ are the values of the
SALI.0/ and the SALI.t/ respectively

2
We note that throughout this chapter we use the hat symbol (O) to denote a unit vector.
3
For Hamiltonian systems the time is a continuous variable, while for maps it is a discrete one
counting the maps iterations.
5 The SALI and the GALI Methods of Chaos Detection 135

while the corresponding unit vectors are w O 1 .t/, w O 2 .t/. Then the diagonals of the
parallelograms defined by w O 1 .t/, w
O 2 .t/, both for t D 0 and t > 0, depict the sum
and the difference of the two unit vectors.
In the particular case shown in Fig. 5.1 the two unit vectors tend to align by
becoming equal. This means that kw O 1 .t/  w
O 2 .t/k ! 0 and kw O 1 .t/ C wO 2 .t/k ! 2.
Of course the dynamics could have led the vectors to become opposite. In that case
we get kw O 1 .t/  w
O 2 .t/k ! 2 and kw O 1 .t/ C wO 2 .t/k ! 0. Since we are not interested
in the particular orientation of the deviation vectors, i.e. whether they become equal
or opposite to each other, when we check their possible alignment, a rather natural
choice is to define the minimum of norms kw O 1 .t/ C w O 2 .t/k, kwO 1 .t/  w
O 2 .t/k as an
indicator of the vectors alignment. This is the reason of the appellation, as well as
of the definition of the SALI in [79] as

SALI.t/ D min fkw O 2 .t/k; kw

O 1 .t/ C w O 1 .t/  w
O 2 .t/kg ; (5.8)
wi .t/
with wO i .t/ D kw i .t/k
, i D 1; 2 being unit vectors.
Naturally, in order for the SALI to be efficiently used as a chaos indicator it
should exhibit distinct behaviors for chaotic and regular orbits. As explained before
the SALI becomes zero for chaotic orbits. On the other hand, in the case of regular
orbits deviation vectors fall on the tangent space of the torus on which the motion
occurs, having in general different directions as there is no reason for them to be
aligned [82, 91]. This behavior is shown schematically in Fig. 5.2. Thus, in this case
the index should be always different from zero. In practice, the values of the SALI
exhibit bounded fluctuations around some constant, positive number.

Fig. 5.2 Schematic representation of the evolution of two deviation vectors for a regular orbit.
The motion takes place on a torus. We consider two initially distinct unit deviation vectors w O 1 .0/,
wO 2 .0/ from point P.0/, which are not necessarily on the tangent space of the torus (this space is
depicted as a shaded parallelogram passing through P.0/). As time evolves the deviation vectors
tend to fall on the torus tangent space and the corresponding unit vectors w O 1 .t/, w
O 2 .t/ at time
t > 0 are closer to the current tangent space (i.e. the grey-shaded parallelogram passing through
P.t/), as the shortening of the perpendicular to the tangent spaces dotted lines from the edges of the
deviation vectors indicate. Since there is no reason for the alignment of the two deviation vectors,
the SALI will not become zero
136 Ch. (Haris) Skokos and T. Manos

Thus, in order to compute the SALI we follow the evolution of two initially
distinct, random, unit deviation vectors wO 1 .0/, w O 2 .0/. Choosing these vectors to pbe
also orthogonal sets the initial SALI to its highest possible value (SALI.0/ D 2)
and ensures that they are considerably different from each other, which has proved
to be a very good computational practice. Then, every t D  time units we normalize
the evolved vectors w1 .i/, w2 .i/, i D 1; 2; : : :, to wO 1 .i/, w
O 2 .i/ and evaluate the
SALI.i/ from (5.8). This algorithm is described in pseudo-code in Table 5.1 of
the Appendix. A MAPLE code for this algorithm, developed specifically for the
Hnon-Heiles system (5.4) can be found in Chap. 5 of [20].
The completely different behaviors of the SALI for regular and chaotic orbits are
clearly seen in Fig. 5.3,4 where some representative results are shown for the 2D
Hamiltonian system (5.4) and the 6d symplectic map

x01 D x1 C y01
K 
y01 D y1 C sin .2x1 /  fsin 2 .x2  x1 / C sin 2 .x3  x1 /g
2 2
0 0
x2 D x2 C y2
K 
y02 D y2 C sin .2x2 /  fsin 2 .x3  x2 / C sin 2 .x1  x2 /g
2 2
0 0
x3 D x3 C y3
K 
y03 D y3 C sin .2x3 /  fsin 2 .x1  x3 / C sin 2 .x2  x3 /g ;
2 2
(5.9)

obtained by considering M D 3 coupled standard maps with K1 D K2 D K3 D K in

(5.7). From the results of Fig. 5.3 we see that for both systems the SALI of regular
orbits (black, solid curves) remains practically constant and positive, i.e.

SALI / constant: (5.10)

On the other hand, the SALI of chaotic orbits (black, dashed curve in Fig. 5.3a and
grey, solid curve in Fig. 5.3b) exhibits a fast decrease to zero after an initial transient
time interval, reaching very small values around the computers accuracy (1016 ).
Actually, it was shown in [83] that the SALI tends to zero exponentially fast in such
cases, following the law

SALI.t/ / exp .1  2 /t; (5.11)

where 1 , 2 (1  2 ) are the first (i.e. the mLE) and the second largest LEs
respectively. As an example demonstrating the validity of this exponential-decay
law we plot in Fig. 5.4 the evolution of the SALI (solid curve) of the chaotic orbit of

4
We note that throughout this chapter the logarithm to base 10 is denoted by log.
5 The SALI and the GALI Methods of Chaos Detection 137

(a) (b)
100

102

104

106
SALI

108

1010

12
10

14
10

16
10
10 100 1000
t

Fig. 5.3 The time evolution of the SALI for a regular and a chaotic orbit of (a) the 2D Hamiltonian
system (5.4) for H2 D 0:125 (after [83]) and (b) the 6d map (5.9) for K D 3 and  D 0:1 (after
[79]). In (a) the time t is continuous, while in (b) it is discrete and counts the maps iterations n.
The initial conditions of the orbits are: (a) q1 D 0, q2 D 0:1, p1 D 0:49058, p2 D 0 (regular orbit;
solid curve) and q1 D 0, q2 D 0:25, p1 D 0:42081, p2 D 0 (chaotic orbit; dashed curve), and
(b) x1 D 0:55, y1 D 0:05, x2 D 0:55, y2 D 0:01, x3 D 0:55, y3 D 0 (regular orbit; black curve)
and x1 D 0:55, y1 D 0:05, x2 D 0:55, y2 D 0:21, x3 D 0:55, y3 D 0 (chaotic orbit; grey curve)

Fig. 5.4 The evolution of the SALI (solid curve) for the chaotic orbit of Fig. 5.3a as a function of
time t. The dashed line corresponds to a function proportional to exp .1 t/ for 1 D 0:047. Note
that the t-axis is linear (after [83])
138 Ch. (Haris) Skokos and T. Manos

Fig. 5.3a using a linear horizontal axis for time t. Since for 2D Hamiltonian systems
2 D 0, (5.11) becomes

SALI.t/ / exp .1 t/; (5.12)

For this particular orbit the mLE was found to be 1  0:047 in [83]. From Fig. 5.4
we see that (5.12) with 1 D 0:047 (dashed line) reproduces correctly the evolution
of the SALI.5
Thus, the completely different behavior of the SALI for regular (5.10) and
chaotic (5.11) orbits permits the clear and efficient distinction between the two
cases. In [79, 83] a comparison of the SALIs performance with respect to other
chaos detection techniques was presented and the efficiency of the index was
discussed. A main advantage of the SALI method is its ability to detect chaotic
motion faster than other techniques which depend on the whole time evolution
of deviation vectors, like the mLE and the spectral distance, because the SALI
is determined by the current state of these vectors and is not influenced by their
evolution history. Hence, the moment the two vectors are close enough to each
other the SALI becomes practically zero and guarantees the chaotic nature of the
orbit beyond any doubt. In addition, the evaluation of the SALI is simpler and
more straightforward with respect to other methods that require more complicated
computations. Such aspects were discussed in [83] where a comparison of the index
with the Relative Lyapunov Indicator (RLI) [77] and the so-called 01 test [39]
was presented. Another crucial characteristic
p of the SALI is that it attains values
in a given interval, namely SALI.t/ 2 0; 2, which does not change in time as
is for example the case for the Fast Lyapunov Indicator (FLI) [37]. Thus, setting
a realistic threshold value below which the SALI is considered to be practically
zero (and the corresponding orbit is characterized as chaotic), allows the fast
and accurate discrimination between regular and chaotic motion. Due to all these
features the SALI became a reliable and widely used chaos indicator as its numerous
applications to a variety of dynamical systems over the years prove. Some of these
applications are discussed in Sect. 5.4.

5.3 The Generalized Alignment Index (GALI)

A fundamental difference between the SALI and other, commonly applied chaos
indicators, is that it uses information from the evolution of two deviation vectors
instead of just one. A consequence of this feature is the appearance of the two

5
We note that here, as well as in several, forthcoming figures in this chapter, the evaluation of
the LEs is done only for confirming the theoretical predictions for the time evolution of the SALI
(Eq. (5.12) in the current case) and later on of the GALIs, and it is not needed for the computation
of the SALI and the GALIs.
5 The SALI and the GALI Methods of Chaos Detection 139

largest LEs in (5.11). After performing this first leap from using only one deviation
vector, the question of going even further arises naturally. To formulate this in other
words: why should we stop in using only two deviation vectors? Can we extend
the definition of the SALI to include more deviation vectors? Assuming that this
extension is possible, what will we gain from it? Will the use of more than two
deviation vectors lead to the introduction of a new chaoticity index which will permit
the acquisition of a deeper understanding of the systems dynamics, exhibiting at the
same time a better numerical performance than the SALI? For instance, from (5.11)
we realize that in the case of a chaotic orbit with 1  2 the convergence of the
SALI to zero will be extremely slow. As a result long integrations would be required
in order for the index to distinguish this orbit from a regular one for which the SALI
remains practically constant. Although the existence of such chaotic orbits is not
very probable the drawback of the SALI remains. An alternative way to state this
problem is the following: can we construct a new index whose behavior in the case
of chaotic orbits will depend on more LEs than the two largest ones so that it can
overcome the discrimination problem for 1  2 ?
Indeed, such an index can be constructed. The key point to its development is
the observation that the SALI is closely related to the area of the parallelogram
defined by the two deviation vectors.6 From the schematic representation of the
deviation vectors evolution in Fig. 5.1 we see that when the SALI vanishes one of
the diagonals of the parallelogram also vanishes, and consequently its area becomes
zero. The area A2 of a usual 2d parallelogram is equal to the norm of the exterior
product of its two sides v1 , v2 , and also equal to the half of the product of its
diagonals lengths

kv1 C v2 k  kv1  v2 k
A2 D kv1  v2 k D : (5.13)
2
In a similar way, the area A of the parallelogram of Fig. 5.1 is given by the
generalization of the exterior product of vectors to higher dimensions, i.e. the so-
called wedge product denoted by .^/,7 so that

O1 Cw
kw O 2 k  kw
O1 w
O 2k
A D kw O 2k D
O1 ^w : (5.14)
2

Note the analogy of this equation to (5.13).8

Based on the fact that the SALI is related to the area of the parallelogram defined
by two unit deviation vectors, the extension of the index to include more vectors

6
Note that this parallelogram is not the usual 2d parallelogram on the plane because its sides (the
deviation vectors) are not 2d vectors.
7
For a brief introduction to the notion of the wedge product the reader is referred to the Appendix A
of [84] and the Appendix of [81].
8
A proof of the second equality of (5.14) can be found in the Appendix B of [84].
140 Ch. (Haris) Skokos and T. Manos

is straightforward: the new quantity is defined as the volume of the parallelepiped

formed by more than two deviation vectors. This volume is computed as the norm
of the wedge product of these vectors. These arguments led to the introduction in
[84] of the Generalized Alignment Index of order k (GALIk ) as

O 1 .t/ ^ w
GALIk .t/ D kw O 2 .t/ ^ : : : ^ w
O k .t/k; (5.15)

where wO i are unit vectors as in (5.8). In this definition the number of used deviation
vectors should not exceed the dimension of the systems phase space, because in this
case the k vectors will become linearly dependent and the corresponding volume
will be by definition zero, as is for example the area defined by two vectors having
the same direction. Thus, for an ND Hamiltonian system with N  2 or a 2Nd
symplectic map with N  1, we consider only GALIs with 2  k  2N.
By its definition the GALIk is a quantity clearly indicating the linear dependence
(GALIk D 0) or independence (GALIk > 0) of k deviation vectors. The SALI
has the same discriminating ability as SALI D 0 indicates that the two vectors
are aligned, i.e. they are linearly dependent, while SALI > 0 implies that the
vectors are not aligned, which means that they are linearly independent. Actually,
the connection between the two indices can be quantified explicitly. Indeed, it was
proved in the Appendix B of [84] that

max fkw O 2 .t/k; kw

O 1 .t/ C w O 1 .t/  w
O 2 .t/kg
GALI2 D SALI  : (5.16)
2

Since
p the max fkwO 1 .t/ C w
O 2 .t/k; kw
O 1 .t/  w
O 2 .t/kg is a number in the interval
2; 2 we conclude that

GALI2 / SALI; (5.17)

which means that the GALI2 is practically equivalent to the SALI. This is another
evidence that the GALI definition (5.15) is a natural extension of the SALI for more
than two deviation vectors.

5.3.1 Computation of the GALI

Let us discuss now how one can actually calculate the value of the GALIk for an ND
Hamiltonian system (N  2) or a 2Nd symplectic map (N  1). For this purpose
we consider the k  2N matrix
2 3
w11 .t/ w12 .t/    w1 2N .t/
6 w21 .t/ w22 .t/    w2 2N .t/ 7
6 7
A.t/ D 6 : :: :: 7 (5.18)
4 :: : : 5
wk1 .t/ wk2 .t/    wk 2N .t/
5 The SALI and the GALI Methods of Chaos Detection 141

having as rows the 2N coordinates of the k unit deviation vectors w O i .t/ with respect
to the usual orthonormal basis eO 1 D .1; 0; 0; : : : ; 0/, eO 2 D .0; 1; 0; : : : ; 0/, : : :,
eO 2N D .0; 0; 0; : : : ; 1/. We note that the elements of A.t/ satisfy the condition
P 2N 2
jD1 wij .t/ D 1 for i D 1; 2; : : : ; k as each deviation vector has unit norm.
We can now follow two routes for evaluating the GALIk .t/. According to the
first one we compute the GALIk by evaluating the norm of the wedge product of k
vectors as
8 0 2 312 91=2

w1i1 .t/ w1i2 .t/    w1ik .t/ >
>

>
>
< X B 6 w .t/ w .t/    w .t/ 7C =
B 6 2i1 2i 2 2i k 7C
GALIk .t/ D Bdet 6 : :: :: 7C ;

@ 4 :: : : 5A > >
1i1 <i2 < <ik 2N >
>
: wki1 .t/ wki2 .t/    wkik .t/ ;
(5.19)

where the sum is performed over all the possible combinations of k indices out of
2N (a proof of this equation can be found in [84]). In practice this means that in
our calculation we consider all the k  k determinants of A.t/. Equation (5.19) is
particularly useful for the theoretical description of the GALIs behavior (actually
expressions (5.22) and (5.23) below were obtained by using this equation), but
not very efficient from a practical point of view. The reason is that the number of
determinants appearing in (5.19) can increase enormously when N grows, leading
to unfeasible numerical computations.
A simpler, straightforward and computationally more efficient approach to
evaluate the GALIk was developed in [85], where it was proved that the index is
equal to the product of the singular values zi , i D 1; 2; : : : ; k of AT .t/ (the transpose
of matrix A.t/), i.e.

Y
k
GALIk .t/ D zi .t/: (5.20)
iD1

We note that the singular values of AT .t/ are obtained by performing the Singular
Value Decomposition (SVD) procedure to AT .t/. According to the SVD method (see
for instance Sect. 2.6 of [74]) the 2N  k matrix AT is written as the product of a
2N k column-orthogonal matrix U (UT U D Ik , with Ik being the kk unit matrix),
a k  k diagonal matrix Z having as elements the positive or zero singular values zi ,
i D 1; : : : ; k, and the transpose of a k  k orthogonal matrix V (VT  V D Ik ), i.e.

AT D U  Z  VT : (5.21)

In practice, in order to compute the GALI of order k we follow the evolution of

k initially distinct, random, orthonormal deviation vectors w O 2 .0/, : : :, w
O 1 .0/, w O k .0/.
Similarly to the computation of the SALI, choosing orthonormal vectors ensures
that all of them are sufficiently far from linear dependence and gives to the GALIk
142 Ch. (Haris) Skokos and T. Manos

its largest possible initial value GALIk D 1. Afterwards, every t D  time units we
O 1 .i/,
normalize the evolved vectors w1 .i/, w2 .i/, : : :, wk .i/, i D 1; 2; : : :, to w
wO 2 .i/, : : :, w
O k .i/ and set them as rows of a matrix A.i/ (5.18). Then, according
to (5.20) the GALIk .i/ is computed as the product of the singular values of matrix
AT .i/. This algorithm is described in pseudo-code in Table 5.2 of the Appendix.
A MAPLE code computing all the possible GALIs (i.e. GALI2 , GALI3 and GALI4 )
for the 2D Hamiltonian (5.4) can be found in Chap. 5 of [20].

5.3.2 Behavior of the GALI for Chaotic and Regular Orbits

After defining the new index and explaining a practical way to evaluate it, let us
discuss its ability to discriminate between chaotic and regular motion. As we have
already mentioned, in the case of a chaotic orbit all deviation vectors eventually
become aligned to the direction defined by the largest LE. Thus, they become
linearly dependent and consequently the volume they define vanishes, meaning
that the GALIk , 2  k  2N, will become zero. Actually, in [84] it was shown
analytically that in this case the GALIk .t/ decreases to zero exponentially fast with
an exponent which depends on the k largest LEs as

GALIk .t/ / exp f .1  2 / C .1  3 / C    C .1  k / tg : (5.22)

Note that for k D 2 we get the exponential law (5.11) in agreement with the
equivalence between the GALI2 and the SALI (5.17).
Let us now consider the case of regular motion in a ND Hamiltonian system or a
2Nd symplectic map with N  2. In general, this motion occurs on an Nd torus in the
systems 2Nd phase space. As we discussed in Sect. 5.2, in this case any deviation
vector eventually falls on the Nd tangent space of the torus (Fig. 5.2). Consequently,
the k initially distinct, linearly independent deviation vectors we follow in order to
compute the evolution of the GALIk eventually fall on the Nd tangent space of the
torus, without necessarily having the same directions. Thus, if we do not consider
more deviation vectors than the dimension of the tangent space (k  N) we end up
with k linearly independent vectors on the torus tangent space and consequently the
volume of the parallelepiped they define (i.e. the GALIk ) will be different from zero.
As we see later on, numerical simulations show that in this case the GALIk exhibits
small fluctuations around some positive value. If, on the other hand, we consider
more deviation vectors than the dimension of the tangent space (N < k  2N) the
deviation vectors eventually become linearly dependent, as we end up with more
vectors in the torus tangent space than the spaces dimension. Thus, the volume
that these vectors define will vanish and the GALIk will become zero. Specifically,
in [84] it was shown analytically that in this case the GALIk tends to zero following
a power law whose exponent depends on the torus dimension and on the number k
5 The SALI and the GALI Methods of Chaos Detection 143

of deviation vectors considered, i.e. GALIk / t2.kN/ . In summary the behavior of

the GALIk for regular orbits is
(
constant if 2  k  N
GALIk .t/ / 1 (5.23)
t2.kN/
if N < k  2N:

From this equation we see that SALI / GALI2 / constant, in accordance to (5.10).

5.3.2.1 Some Illustrative Paradigms

In what follows we illustrate the different behaviors of the GALIk by computing

its evolution for some representative chaotic and regular orbits of various ND
autonomous Hamiltonians and 2Nd symplectic maps. Before doing so let us note
that for these systems the LEs comes in pairs of values having opposite signs

i D 2NiC1 ; i D 1; 2; : : : ; N; (5.24)

while, moreover

N D NC1 D 0 (5.25)

for Hamiltonian systems [14, 41, 81].

Hamiltonian Systems

Initially, we consider the 2D Hamiltonian (5.4) which has a 4d phase space. For this
system we can define the GALIk for k D 2, 3 and 4. Then, according to (5.24) and
(5.25), the LEs satisfy the conditions 1 D 4 , 2 D 3 D 0. Thus, according to
(5.22) the evolution of the GALIs for a chaotic orbit is given by

GALI2 .t/ / e1 t ; GALI3 .t/ / e21 t ; GALI4 .t/ / e41 t : (5.26)

On the other hand, for a regular orbit (5.23) indicates that

1 1
GALI2 .t/ / constant; GALI3 .t/ / ; GALI4 .t/ / 4 : (5.27)
t2 t
From the results of Fig. 5.5, where the time evolution of the GALI2 , the GALI3 and
the GALI4 for a chaotic orbit (actually the one considered in Figs. 5.3a and 5.4)
and a regular orbit are plotted, we see that the laws (5.26) and (5.27) describe quite
accurately the obtained numerical data.
144 Ch. (Haris) Skokos and T. Manos

Fig. 5.5 The time evolution of the GALI2 , the GALI3 and the GALI4 for (a) a chaotic and (b) a
regular orbit of the 2D Hamiltonian (5.4) for H2 D 0:125. The chaotic orbit is the one considered
in Fig. 5.3a, while the initial conditions of the regular orbit are q1 D 0, q2 D 0, p1 D 0:5,
p2 D 0. The straight lines correspond in (a) to functions proportional to exp.1 t/, exp.21 t/
and exp.41 t/, for 1 D 0:047 and in (b) to functions proportional to t2 and t4 . The slope of
each line is mentioned in the legend. Note that the horizontal, time axis in (a) is linear, while in (b)
is logarithmic (after [84])

For a 3D Hamiltonian like (5.5) the theoretical prediction (5.22) gives

GALI2 .t/ / e.1 2 /t ; GALI3 .t/ / e.21 2 /t ; GALI4 .t/ / e.31 2 /t ;
GALI5 .t/ / e41 t ; GALI6 .t/ / e61 t ;
(5.28)
for a chaotic orbit, because, according to (5.24) and (5.25), 1 D 6 , 2 D 5
and 3 D 4 D 0. On the other hand, a regular orbit lies on a 3d torus and according
to (5.23) the GALIs should behave as

1
GALI2 .t/ / constant; GALI3 .t/ / constant; GALI4 .t/ / 2 ;
t (5.29)
1 1
GALI5 .t/ / 4 ; GALI6 .t/ / 6 :
t t
In Fig. 5.6 we plot the time evolution of the various GALIs for a chaotic (Fig. 5.6a)
and a regular (Fig. 5.6b) orbit of the 3D Hamiltonian (5.5). From the plotted results
we see that the behaviors of the GALIs are very well approximated by (5.28)
and (5.29). We note here that the constant values that the GALI2 and the GALI3
eventually attain in Fig. 5.6b are not the same. Actually, the limiting value of GALI3
is smaller than the one of GALI2 .
As an example of evaluating the GALIs for multidimensional Hamiltonians we
consider model (5.6) for N D 8 particles. This corresponds to an 8D Hamiltonian
system H8 , having a 16d phase space, which allows the definition of several GALIs:
5 The SALI and the GALI Methods of Chaos Detection 145

Fig. 5.6 The time evolution of the GALIk , k D 2; 3; : : : ; 6 for (a)p

a chaotic and (b)pa regular orbit
of the 3D Hamiltonian (5.5) with H3 D 0:09, !1 D 1, !2 D 2 and !3 D 3. The initial
conditions of the orbits are: (a) q1 D 0, q2 D 0, q3 D 0, E1 D 0:03, E2 D 0:03, E3 D 0:03,
and (b) q1 D 0, q2 D 0, q3 D 0, E1 D 0:005, E2 D 0:085, E3 D 0, where the quantities
E1 , E2 , E3 (usually
p referred as the harmonic energies) are related to the momenta p1 , p2 , p3
through pi D 2Ei =!i , i D 1; 2; 3. The straight lines in (a) correspond to functions proportional
to exp.1  2 /t, exp.21  2 /t, exp.31  2 /t, exp.41 t/ and exp.61 t/ for
1 D 0:03, 2 D 0:008, which are accurate numerical estimations of the orbits two largest LEs
(see [84] for more details). The straight lines in (b) correspond to functions proportional to t2 ,
t4 and t6 . The slope of each line is mentioned in the legend. The horizontal, time axis is linear
in (a) and logarithmic in (b) (after [84])

starting from GALI2 up to GALI16 . In Fig. 5.7 the time evolution of several of these
indices are shown for a chaotic (Fig. 5.7a, b) and a regular (Fig. 5.7c, d) orbit. From
these results we again conclude that the laws (5.22) and (5.23) are quite accurate in
describing the time evolution of the GALIs.
The first seven indices, GALI2 up to GALI8 , exhibit completely different
behaviors for chaotic and regular motion: they tend exponentially fast to zero
for a chaotic orbit (Fig. 5.7a, b), while they attain constant, positive values for a
regular one (Fig. 5.7c). This characteristic makes them ideal numerical tools for
discriminating between the two cases, as we see in Sect. 5.4.1.1 where some specific
numerical examples are discussed in detail.
Although the constancy of the GALIk , k D 1; : : : ; 8 for regular orbits is predicted
from (5.23), nothing is yet said about the actual values of these constants. It is
evident from Fig. 5.7c that these values decrease as the order k of the GALIk
increases, something which was also observed in Fig. 5.6b for the 3D Hamiltonian
(5.5). For the regular orbit of Fig. 5.7c we see that GALI8  107 . One might
argue that this very small value could be considered to be practically zero and that
the orbit might be (wrongly) classified as chaotic. The flaw in this argumentation
is that the possible smallness of GALI8  107 is of relative nature as this
value should be compared with the values that the index reaches for actual chaotic
146 Ch. (Haris) Skokos and T. Manos

Fig. 5.7 The time evolution of the GALIk , k D 2; : : : ; 8; 10; 12; 14; 16 for a chaotic (panels (a)
and (b)) and a regular orbit (panels (c) and (d)) of the ND Hamiltonian (5.6) with N D 8 and D
1:5. The initial conditions of the chaotic orbit are Q1 D Q4 D 2, Q2 D Q5 D 1, Q3 D Q6 D 0:5,
8   8  
2X ij 2X ij
Q7 D Q8 D 0:1, Pi D 0 where Qi D qj sin , Pi D pj sin , i D 1; : : : ; 8
3 jD1 9 3 jD1 9
(see [85] for more details). The initial conditions of the regular orbit are q1 D q2 D q3 D q8 D
0:05, q4 D q5 D q6 D q7 D 0:1, pi D 0, i D 1; : : : ; 8. The straight lines in (a) and (b) correspond
to exponential functions of the form (5.22) for 1 D 0:170, 2 D 0:141, 3 D 0:114, 4 D 0:089,
5 D 0:064, 6 D 0:042, 7 D 0:020, which are estimations (obtained in [85]) of the orbits
seven largest LEs. The straight lines in (d) correspond to functions proportional to t4 , t8 , t12
and t16 . The slope of each line is mentioned in the legend. Note the huge range differences in the
horizontal, time axes between panels (a) and (b), where the axes are linear, and panels (c) and (d)
where the axes are logarithmic (after [85])

orbits. For instance, the chaotic orbit of Fig. 5.7b has GALI8  1040 , after only
t  160 time units! At the same time we get GALI8  101 for the regular orbit
(Fig. 5.7c). In addition, extrapolating the results of GALI8 for the chaotic orbit in
5 The SALI and the GALI Methods of Chaos Detection 147

Fig. 5.7b to e.g. t  105 we would obtain values extremely smaller than the value
GALI8  107 archived for the regular orbit in Fig. 5.7c.
The necessity to determine an appropriate threshold value for the GALIk , 2 
k  N, below which orbits will be securely classified as chaotic, becomes evident
from the above analysis. Since a theoretical, or even an empirical (numerical)
relation between the order k of the GALIk and the constant value it reaches for
regular orbits is still lacking, one efficient way to determine this threshold value is
by computing the GALIk for some representative chaotic and regular orbits of each
studied system. Then, a safe policy is to define this threshold to be a few orders of
magnitude smaller than the minimum value obtained by the GALIk for the tested
regular orbits. For example, based on the results of Fig. 5.6 for the 3D Hamiltonian
(5.5) this threshold value for the GALI3 could be set to be 108 , while for the
system of Fig. 5.7 a reliable threshold value for the GALI8 could be 1016 .
The results of Fig. 5.7 verify the predictions of (5.22) and (5.23) that the GALIs
of order 8 < k  16 tend to zero both for chaotic and regular orbits. Nevertheless,
the completely different way they do so, i.e. they decay exponentially fast for chaotic
orbits, while they follow a power law decay for regular ones, allows us again to
develop a well-tailored strategy to discriminate between the two cases. The different
decay laws result in enormous differences in the time the indices need to reach any
predefined low value. Thus, the measurement of this time can be used to characterize
the nature of the orbits, as we see in Sect. 5.4.1.2. For example, for the chaotic orbit
of Fig. 5.7b GALI16  1030 after about t  25 time units, while it reaches the
same small value after about t  105 time units for the regular orbit of Fig. 5.7d; a
time interval which is larger by a factor  4000 with respect to the chaotic orbit!

Symplectic Maps

Although up to now our discussion concerned the implementation of the GALIs to

Hamiltonian systems, the indices follow laws (5.22) and (5.23) also for symplectic
maps (with the obvious substitution of the continuous time t by a discrete one which
counts the maps iterations n) as the representative results of Figs. 5.8 and 5.9 clearly
verify. In particular, in Fig. 5.8 we see the behavior of the GALIs for a chaotic
(Fig. 5.8a) and a regular (Fig. 5.8b) orbit of the 4d map

x01 D x1 C y01
K 
y01 D y1 C sin .2x1 /  sin 2 .x2  x1 /
2 2 (5.30)
x02 D x2 C y02
K 
y02 D y2 C sin .2x2 /  sin 2 .x1  x2 / ;
2 2
148 Ch. (Haris) Skokos and T. Manos

Fig. 5.8 The evolution of the GALI2 , the GALI3 and the GALI4 with respect to the number of
iterations n for (a) a chaotic and (b) a regular orbit of the 4d map (5.30) with K D 0:5 and  D
0:05. The initial conditions of the orbits are: (a) x1 D 0:55, y1 D 0:1, x2 D 0:005, y2 D 0:01, and
(b) x1 D 0:55, y1 D 0:1, x2 D 0:54, y2 D 0:01. The straight lines in (a) correspond to functions
proportional to exp.1 2 /n, exp.21 n/ and exp.41 n/ for 1 D 0:07, 2 D 0:008, which
are the orbits LEs obtained in [66]. The straight lines in (b) represent functions proportional to
n2 and n4 . The slope of each line is mentioned in the legend. Note that the horizontal axis is
linear in (a) and logarithmic in (b) (after [66])

Fig. 5.9 The evolution of the GALIk , k D 2; 3; : : : ; 6 with respect to the number of iterations n
for (a) a chaotic (after [64]) and (b) a regular orbit of the 6d map (5.9) with K D 3 and  D 0:1.
The initial conditions of the orbits are: (a) x1 D x2 D x3 D 0:8, y1 D 0:05, y2 D 0:21, y3 D 0:01,
and (b) x1 D x2 D x3 D 0:55, y1 D 0:05, y2 D 0:21, y3 D 0. The straight lines in (a) correspond
to functions proportional to exp.1  2 /n, exp.21  2  3 /n, exp.31  2 /n,
exp.41 n/ and exp.61 n/ for 1 D 0:70, 2 D 0:57, 3 D 0:32, which are the orbits LEs
obtained in [64]. The straight lines in (b) represent functions proportional to n2 , n4 and n6 .
The slope of each line is mentioned in the legend. Note that the horizontal axis is linear in (a) and
logarithmic in (b)
5 The SALI and the GALI Methods of Chaos Detection 149

obtained from (5.7) for M D 2 and K1 D K2 D K, while in Fig. 5.9 a chaotic

(Fig. 5.9a) and a regular (Fig. 5.9b) orbit of the 6d map (5.9) are considered.
These results illustrate the fact that the GALIk has the same behavior for
Hamiltonian flows and symplectic maps. For instance, even by simple inspection
we conclude that the GALIs behave similarly in Figs. 5.5 and 5.8, which refer to a
2D Hamiltonian and a 4d map respectively, as well as in Figs. 5.6 and 5.9, which
refer to a 3D Hamiltonian and a 4d map respectively.

5.3.2.2 The Case of 2d Maps

Equations (5.22) and (5.23) describe the behavior of the GALIs for ND Hamiltonian
systems and 2Nd symplectic maps with N  2. What happens if N D 1? The case of
an 1D, time independent Hamiltonian is not very interesting because such systems
are integrable and chaos does not appear. But, this is not the case for 2d maps, which
can exhibit chaotic behavior.
In 2d maps only the GALI2 (which, according to (5.17) is equivalent to the SALI)
is defined. For chaotic orbits the GALI2 decreases exponentially to zero according
to (5.22), which becomes

GALI2 .n/ / SALI.n/ / exp .21 n/ ; (5.31)

in this particular case, since, according to (5.24) 1 D 2 > 0. Note that in
(5.31) we have substituted the continuous time t of (5.22) by the number n of maps
iterations. The agreement between the prediction (5.31) and actual, numerical data
can be seen for example in Fig. 5.10a where the evolution of the SALI (/ GALI2 )
is plotted for a chaotic orbit of the 2d standard map

x01 D x1 C y01
K (5.32)
y01 D y1 C sin .2x1 / ;
2
obtained from (5.7) for M D 1. Thus, we conclude that (5.22) is also valid for 2d
maps.
But what happens in the case of regular orbits? Is (5.23) still valid for k D 2
and N D 1? First of all let us note that for these particular values of k and N only
the second branch of (5.23) is meaningful, and it provides the prediction that the
GALI2 tends to zero as n2 . This result is interesting, as this is the first case of
regular motion for which no GALI remains constant. But actually the vanishing
of the GALI2 in this case is not surprising. Regular motion in 2d maps occurs on
1d invariant curves. So, any deviation vector from a regular orbit eventually falls
on the tangent space of this curve, which of course has dimension 1. Thus, the
two deviation vectors needed for the computation of the GALI2 eventually becomes
150 Ch. (Haris) Skokos and T. Manos

Fig. 5.10 The evolution of the SALI (which in practice is the GALI2 ) with respect to the number
of iterations n for (a) a chaotic and (b) a regular orbit of the 2d map (5.32) with K D 2. The initial
conditions of the orbits are: (a) x1 D y1 D 0:2, and (b) x1 D 0:4, y1 D 0:8. The straight line in (a)
corresponds to a function proportional to exp.21 n/ for 1 D 0:438, which is the orbits mLE
obtained in [65], while the line in (b) represents a function proportional to n2 . The slope of each
line is mentioned in the legend. Note that the horizontal axis is linear in (a) and logarithmic in (b)
(after [65])

collinear and consequently GALI2 ! 0. Actually the prediction obtained by (5.23),

that for regular orbits of 2d maps

1
GALI2 .n/ / SALI.n/ / ; (5.33)
n2
is correct, as for example the results of Fig. 5.10b show.
In conclusion we note that the behavior of the SALI/GALI2 for chaotic and
regular orbits in 2d maps is respectively given by (5.31) and (5.33), which are
obtained from (5.22) and (5.23) for k D 2 and N D 1. The different behaviors
of the index for chaotic (exponential decay) and regular motion (power law decay)
were initially observed in [79], although the exact functional laws (5.31) and (5.33)
were derived later [83, 84]. As was pointed out even from the first paper on the
SALI [79], these differences allow us to use the SALI/GALI2 to distinguish between
chaotic and regular motion also in 2d maps (see for instance [65, 79]).

5.3.3 Regular Motion on Low Dimensional Tori

An important feature of the GALIs is their ability to identify regular motion on low
dimensional tori. In order to explain this capability let us assume that a regular orbit
lies on an sd torus, 2  s  N, in the 2Nd phase space on an ND Hamiltonian
5 The SALI and the GALI Methods of Chaos Detection 151

system or a 2Nd map with N  2. Then, following similar arguments to the ones
made in Sect. 5.3.2 for regular motion on an Nd torus, we conclude that the GALIk
eventually remains constant for 2  k  s, because in this case the k deviation
vectors will remain linearly independent when they eventually fall on the sd tangent
space of the torus. On the other hand, any s < k  2N deviation vectors eventually
become linearly dependent as there will be more vectors on the torus tangent space
than the spaces dimension, and consequently the GALIk will vanish. In this case,
the way the GALIk tends to zero depends not only on k and N, as in (5.23), but also
on the dimension s of the torus. Actually, it was shown analytically in [27, 85] that
for regular orbits on an sd torus the GALIk behaves as
8

< constant if 2  k  s
1
GALIk .t/ / tks if s < k  2N  s (5.34)
: 1 if 2N  s < k  2N:
t2.kN/

It is worth noting that for s D N we retrieve (5.23) as the second branch of (5.34)
becomes meaningless, while by setting k D 2, s D 1 and N D 1 we get (5.33).
The validity of (5.34) is supported by the results of Fig. 5.11 where two
representative regular orbits of the H8 Hamiltonian, obtained by setting N D 8 in
(5.6), are considered (we note that Fig. 5.7 refers to the same model). The first orbit
(Fig. 5.11a, b) lies on a 2d torus as the constancy of only GALI2 indicates. The decay
of the remaining GALIs is well reproduced by the power laws (5.34) for N D 8 and
s D 2. The second orbit (Fig. 5.11c, d) lies on a 4d torus and consequently the
GALI2 , the GALI3 and the GALI4 remain constant, while all other indices follow
power law decays according to (5.34) for N D 8 and s D 4.
In Fig. 5.12 we see the evolution of some GALIs for regular motion on low
dimensional tori of the 40d map obtained by (5.7) for M D 20. The results of
Fig. 5.12a denote that the orbit lies on a 3d torus in the 40d phase space of the map,
while in the case of Fig. 5.12b the motion takes place on a 6d torus. The plotted
straight lines help us verify that for both orbits the behaviors of the decaying GALIs
are accurately reproduced by (5.34) for N D 20, s D 3 (Fig. 5.12a) and N D 20,
s D 6 (Fig. 5.12b).

5.3.3.1 Searching for Regular Motion on Low Dimensional Tori

Equation (5.34), as well as the results of Figs. 5.11 and 5.12 imply that the GALIs
can be also used for identifying regular motion on low dimensional tori. From
(5.34) we deduce that the dimension of the torus on which the regular motion
occurs coincides with the largest order k of the GALIs for which the GALIk
remains constant. Based on this remark we can develop a strategy for locating low
dimensional tori in the phase space of a dynamical system. The GALIk of initial
conditions resulting in motion on an sd torus eventually will remain constant for
2  k  s, while it will decay to zero following the power law (5.34) for k > s. So,
152 Ch. (Haris) Skokos and T. Manos

Fig. 5.11 The time evolution of the GALIk , k D 2; : : : ; 9; 11; 13; 14; 16 for a regular orbit lying
on a 2d torus (panels (a) and (b)) and for another one lying on a 4d torus (panels (c) and (d)) of
the 8D Hamiltonian H8 considered in Fig. 5.7. The initial conditions of the first orbit are Q1 D 2,
P1 D 0, Qi D Pi D 0, i D 2; : : : ; 8 (the definition of these variables is given in the caption of
Fig. 5.7). The initial conditions of the second orbit are qi D 0:1, pi D 0, i D 1; : : : ; 8. The plotted
straight lines correspond to the power law predictions (5.34) for N D 8, s D 2 (panels (a) and (b))
and for N D 8, s D 4 (panels (c) and (d)). The slope of each line is mentioned in the legend (after
[85])

after some relatively long time interval, all the GALIs of order k > s will have much
smaller values than the ones of order k  s. Thus, in order to identify the location
of sd tori, 2  s  N, in the 2Nd phase space of a dynamical system we evaluate at
first various GALIs for several initial conditions and then find the initial conditions
which result in large GALIk values for k  s and small values for k > s.
As was mentioned in Sect. 5.3.2.1, the constant, final values of the GALIs for
regular motion decrease with the order of the GALI (see Figs. 5.6b, 5.7c, 5.9b, 5.11c
5 The SALI and the GALI Methods of Chaos Detection 153

Fig. 5.12 The evolution of several GALIs for a regular orbit lying (a) on a 3d torus and (b) on a
6d torus of the 40d map obtained by setting M D 20 in (5.7). In (a) the initial conditions of the
orbit are x11 D 0:65, x12 D 0:55, xi D 0:5 8i 11; 12, and yi D 0, i D 1; : : : ; 20, while the
parameters of the map are set to  D 0:001 and Ki D K D 2, i D 1; : : : ; 20. In (b)  D 0:00001
and Ki are set in triplets of 1:35, 1:45, 1:55 (i.e. K1 D 1:35, K2 D 1:45, K3 D 1:55,
K4 D 1:35, : : :, K20 D 1:45), while the orbits exact initial conditions can be found in [21].
The plotted straight lines correspond to the power law predictions (5.34) for (a) N D 20, s D 3
and (b) N D 20, s D 6. The slope of each line is mentioned in the legend (after [21])

and 5.12). Since this decrease has not been quantified yet, a good computational
approach in the quest for low dimensional tori is to normalize the values of the
GALIs for each individual orbit by dividing them by the largest GALIk value,
max .GALIk /, obtained by all orbits in the studied ensemble at the end time t D te
of the integration. In this way we define the normalized GALIk

GALIk .t/
gk .t/ D : (5.35)
max GALIk .te /

Then, by coloring each initial condition according to its gk .te / value we can
construct phase space charts where the position of low dimensional tori is easily
located.
To illustrate this method we present (following [38]) the search for low dimen-
sional tori in a subspace of the 8d phase space of the 4D Hamiltonian system
H4 obtained by setting N D 4 and D 1:5 in (5.6). In order to facilitate the
visualization of the whole procedure we restrict our search in the subspace .q3 ; q4 /
by setting the other initial conditions of the studied orbits to q1 D q2 D 0:1,
p1 D p2 D p3 D 0, while p4 > 0 is evaluated so that H4 D 0:010075. In Fig. 5.13
we color each permitted initial condition in the .q3 ; q4 / plane according to its g2 , g3
and g4 value at t D te D 106 time units (panels (a), (b) and (c) respectively).
For this particular Hamiltonian we can have regular motion on 2d, 3d and 4d
tori. Let us see now how we can exploit the results of Fig. 5.13 to locate such tori.
154 Ch. (Haris) Skokos and T. Manos

g2 g3 g4
100
0.10
(a) (b) (c) -1
10
0.05
-2
q4

10 gk

0.00 10-3

-4
10
0.00 0.05 0.10 0.00 0.05 0.10 0.00 0.05 0.10
q3 q3 q3

Fig. 5.13 Regions of different gk (5.35) values for (a) k D 2, (b) k D 3, (c) k D 4, in the subspace
.q3 ; q4 / of the 4D Hamiltonian H4 obtained from (5.6) for N D 4 and D 1:5. The remaining
coordinates of the considered initial conditions are set to q1 D q2 D 0:1, p1 D p2 D p3 D 0,
while p4 > 0 is evaluated so that H4 D 0:010075. White regions correspond to forbidden initial
conditions. The color scales shown at the right of the panels are used to color each point according
to the orbits gk value at t D 106 . The points with coordinates q3 D 0:106, q4 D 0:0996 (marked
by a triangle), q3 D 0:085109, q4 D 0:054 (marked by a square) and q3 D 0:025, q4 D 0 (marked
by a circle) correspond to regular orbits on a 2d, a 3d and a 4d torus respectively (after [38])

Fig. 5.14 The time evolution of the GALI2 , the GALI3 and the GALI4 of regular orbits lying on
a (a) 2d, (b) 3d, (c) 4d torus of the 4D Hamiltonian considered in Fig. 5.13. The initial conditions
of these orbits are respectively marked by a triangle, a square and a circle in Fig. 5.13 (after [38])

Motion on 2d tori results in large final g2 values and to small g3 and g4 . So, such
tori should be located in regions colored in yellow or light red in Fig. 5.13a and
in black in Fig. 5.13b, c. A region which satisfies these requirements is located at
the upper border of the colored areas in Fig. 5.13. The evolution of the GALIs of
an orbit with initial conditions in that region (denoted by a triangle in Fig. 5.13) is
shown in Fig. 5.14a and it verifies that the motion takes place on a 2d torus, as only
the GALI2 remains constant.
5 The SALI and the GALI Methods of Chaos Detection 155

Extending the same argumentation to higher dimensions we see that motion on

a 3d torus can occur in regions colored in yellow or light red in both Fig. 5.13a, b
and in black in Fig. 5.13c. The initial condition of an orbit of this kind is marked by
a small square in Fig. 5.13. The evolution of this orbits GALIs (Fig. 5.14b) verifies
that the orbit lies on a 3d torus, because only the GALI2 and the GALI3 remain
constant. Orbits on 4d tori is the most common situation of regular motion for this
4D Hamiltonian system. This is evident from the results of Fig. 5.13 because most
of the permitted area of initial conditions correspond to high g2 , g3 and g4 values.
A randomly chosen initial condition in this region (marked by a circle in Fig. 5.13)
results indeed to regular motion on a 4d torus as the constancy of its GALIk , k D
2; 3; 4 in Fig. 5.14c clearly indicates.
We note that initial conditions leading to chaotic motion in this system would
correspond to very small g2 , g3 and g4 values (due to the exponential decay of
the associated GALIs) and consequently would be colored in black in all panels
of Fig. 5.13. The lack of such regions in Fig. 5.13 signifies that all considered
initial conditions lead to regular motion. This happens because regions of chaotic
motion occupy a tiny fraction of the systems phase space, because its nonlinearity
strength is very small. Therefore, chaotic motion is not captured by the grid of initial
conditions of Fig. 5.13.

5.3.4 Behavior of the GALI for Periodic Orbits

Let us now discuss the behavior of the GALIs for periodic orbits of period T;
i.e. orbits satisfying the condition x.t C T/ D x.t/, with x.t/ being the coordinate
vector in the systems phase space. In the presentation of this topic we mainly follow
the analysis performed in [67]. The linear stability of periodic orbits is defined by the
eigenvalues of the so-called monodromy matrix, which is obtained by the solution
of the variational equations (for Hamiltonian systems) or by the evolution of the
tangent map (for symplectic maps) for one period T (see for example [22, 80] and
Sect. 3.3 of [53]). When all eigenvalues lie on the unit circle in the complex plane the
orbit is characterized as elliptic, while otherwise it is called hyperbolic (unstable).
For a detailed presentation of the various stability types of periodic orbits the reader
is referred for example to [22, 40, 44, 45, 80].
The presence of periodic orbits influence significantly the dynamics. In most
systems we observe that the majority of non-periodic orbits in the vicinity of an
elliptic one are regular. So, although initial conditions near an elliptic orbit can
lead to chaos, regular orbits exhibiting a time evolution similar to the elliptic orbit
itself prevail. If one assumes that the elliptic orbit is integrable and in its vicinity
the KolmogorovArnoldMoser (KAM) theorem (see for example Sect. 3.2 of
[53] and references therein) can be applied (for which one needs to check a non-
degeneracy condition which is typically satisfied), then there is large measure of
orbits on KAM tori nearby. In Hamiltonian systems of dimension larger than 2 the
phenomenon of Arnold diffusion (see for example Chap. 6 of [53] and references
156 Ch. (Haris) Skokos and T. Manos

therein) typically would lead to an escape of orbits from the neighborhood of the
elliptic orbit. However, it is generally believed that Arnold diffusion occurs on a
slow time scale, and we do not expect interference with the GALI method. Of
course, regular behavior on nearby KAM tori does not imply that the elliptic orbit
itself is stable (e.g. Appendix of [34]). On the other hand, in chaotic Hamiltonian
systems and symplectic maps orbits in the vicinity of an unstable periodic orbit
typically behave chaotically and diverge from the periodic one exponentially fast.
This divergence is characterized by LEs (with at least one of them being positive)
which are determined by the eigenvalues of the monodromy matrix (e.g. [13, 84]
and Sect. 5.2b of [53]). Thus, following arguments similar to the ones developed
in Sect. 5.3.2 for chaotic orbits, we easily see that the GALIk of unstable periodic
orbits decreases to zero following the exponential law (5.22), i. e.

GALIk .t/ / exp f .1  2 / C .1  3 / C    C .1  k / tg ; (5.36)

where i , i D 1; : : : ; k are the periodic orbits k largest LEs.

In Fig. 5.15a we see that the evolution of the GALIs for an unstable periodic
orbit of the 2D Hamiltonian (5.4) is well approximated by (5.36) for 1 D 0:084.
This value is the orbits mLE determined by the eigenvalues of the corresponding
monodromy matrix (see [67] for more details). We also note that according to (5.24)
and (5.25) we set 1 D 4 , and 2 D 3 D 0 in (5.36). The agreement between the
numerical data and the theoretical prediction (5.36) is lost after about t  350 time
units. This happens because the numerically computed orbit eventually deviates

Fig. 5.15 The time evolution of (a) the GALI2 , the GALI3 , the GALI4 and (b) the finite time
mLE 1 of an unstable periodic orbit of the 2D Hamiltonian (5.4) for H2 D 0:125. The initial
conditions of the orbit are q1 D 0, q2 D 0:2083772012, p1 D 0:4453146996, p2 D 0:1196065752.
The straight lines in (a) correspond to functions proportional to exp.1 t/, exp.21 t/ and
exp.41 t/, for 1 D 0:084, which is the mLE of the periodic orbit. The slope of each line is
mentioned in the legend. The horizontal dotted line in (b) indicates the value 1 D 0:084 (after
[67])
5 The SALI and the GALI Methods of Chaos Detection 157

from the unstable periodic one due to unavoidable computational inaccuracies

and enters the chaotic region around the periodic orbit. In general, this region
is characterized by different LEs with respect to the ones of the periodic orbit.
The effect of this behavior on the orbits finite time mLE 1 (5.1) is seen in
Fig. 5.15b. The computed 1 deviates from the value 1 D 0:084 (marked by a
horizontal dotted line) at about the same time the GALI2 changes its decreasing rate
in Fig. 5.15a. Eventually, 1 stabilizes at another positive value, which characterizes
the chaoticity of the region around the periodic orbit.
On the other hand, the case of stable periodic orbits is a bit more complicated,
because the GALIs behave differently for Hamiltonian flows and symplectic maps.
In [67] it was shown analytically that for stable periodic orbits of ND Hamiltonian
systems, with N  2, the GALIs decay to zero following the following power laws
(
1
if 2  k  2N  1
GALIk .t/ / tk1
1 (5.37)
t2N
if k D 2N:

We observe that this equation can be derived from (5.34), which describes the
behavior of the GALIs for motion on an sd tori, by setting s D 1. We note that
the first branch of (5.34) is meaningless for s D 1, while the other two branches
take the forms appearing in (5.37). The connection between (5.34) and (5.37) is not
surprising if we notice that a periodic orbit is nothing more than an 1d closed curve
in the systems phase space, having the some dimension with an 1d torus.
Small, random perturbations from the stable periodic orbit generally results in
regular motion on an Nd torus. So, the GALIs of the perturbed orbit will follow
(5.23). Thus, in general, the GALIs of regular orbits in the vicinity of a stable
periodic orbit behave differently with respect to the indices of the periodic orbit
itself (except from the GALI2N and the GALI2N1 , which respectively follow the
laws / t2N and / t.2N2/ in both cases). The most profound change happens for
the GALIs of order 2  k  N because, according to (5.23), they remain constant
in the neighborhood of the periodic orbit, while they decay to zero following the
power law (5.37) for the periodic orbit.
The correctness of (5.37) becomes evident from the results of Fig. 5.16a, where
the time evolution of the GALIs of a stable periodic orbit of the 2D Hamiltonian
(5.4) is shown. In particular, we see that the indices decay to zero following the
power laws GALI2 / t1 , GALI3 / t2 , GALI4 / t4 predicted from (5.37).
According to (5.23) the GALIs of regular orbits in the neighborhood of the stable
periodic orbit should behave as GALI2 / constant, GALI3 / t2 and GALI4 / t4 .
Thus, only the GALI2 is expected to behave differently for regular orbits in the
vicinity of the periodic orbit of Fig. 5.16a. The results of Fig. 5.16b show that this is
actually true. The GALI2 of the neighboring regular orbits initially follows the same
power law decay of the periodic orbit (GALI2 / t1 ), but later on it stabilizes to a
constant positive value. We see that the further the orbit is located from the periodic
one the sooner the GALI2 deviates from the power law decay.
158 Ch. (Haris) Skokos and T. Manos

Fig. 5.16 (a) The time evolution of the GALI2 , the GALI3 and the GALI4 for a stable periodic
orbit of the 2D Hamiltonian (5.4) for H2 D 0:125. The orbits initial conditions are q1 D 0 D q10 ,
q2 D 0:35207 D q20 , p1 D 0:36427 D p10 , p2 D 0:14979 D p20 . The straight lines correspond
to functions proportional to t1 , t2 and t4 . The slope of each line is mentioned in the legend. (b)
The same plot as in (a) where apart from the GALIs of the stable periodic orbit (red curves) the
indices of two neighboring, regular orbits are also plotted. Their initial conditions are q1 D q10 ,
p2 D p20 for both of them, while q2 D q20 C 0:00793 (green curves), and q2 D q20 C 0:02793
(blue curves). In both cases the p1 > 0 initial condition is set so that H2 D 0:125. Note that the
curves of the GALI3 and the GALI4 for all three orbits overlap each other (after [67])

These differences of the GALI2 values can be used to identify the location of
stable periodic orbits in the systems phase space, although the index was not
developed for this particular purpose.9 This becomes evident from the result of
Fig. 5.17 where the values of the GALI2 at t D 105 for several orbits of the Hnon-
Heiles system (5.4) are plotted as a function of the q2 coordinate of the orbits
initial conditions. The remaining coordinates are q1 D p2 D 0, while p1 > 0 is set
so that H2 D 0:125. Actually these initial conditions lie on the symmetry line of the
subspace defined by q1 D 0, p1 > 0, i.e. the horizontal line p2 D 0 in Figs. 5.19 and
5.20 below. This line passes through the initial condition of some periodic orbits of
the system. For the construction of Fig. 5.17 we considered an ensemble of 7 000
orbits whose q2 coordinates are equally distributed in the interval 0:1  q2  0:6.
The data points are line connected, so that the changes of the GALI2 values become
easily visible.
In Fig. 5.17 regions of relatively large GALI2 values (&104 ) correspond to
regular (periodic or quasiperiodic) motion. Chaotic orbits and unstable periodic
orbits have very small GALI2 values (.1012 ), while domains with intermediate
values (1012 . GALI2 . 104 ) correspond to sticky chaotic orbits. An interesting

9
It is worth mentioning here that other chaos indicators, like the Orthogonal Fast Lyapunov
Indicator (OFLI) and its variations [7, 8], are quite successful in performing this task as they were
actually designed for this purpose.
5 The SALI and the GALI Methods of Chaos Detection 159

Fig. 5.17 The values of the GALI2 at t D 105 for several orbits of the 2D Hamiltonian (5.4)
as a function of the q2 coordinate of the orbits initial conditions. The remaining coordinates are
q1 D p2 D 0, while p1 > 0 is set so that H2 D 0:125. Actually these initial conditions lie on the
p2 D 0 line of Figs. 5.19 and 5.20. The numerical data (black points) are line connected (grey line)
in order to facilitate the visualization of the value changes (after [67])

feature of Fig. 5.17 is the appearance of some relatively narrow regions where the
GALI2 decreases abruptly obtaining values 104 . GALI2 . 101 ; the most
profound one being in the vicinity of q2  0:3. These regions correspond to the
immediate neighborhoods of stable periodic orbits, with the periodic orbit itself
been located at the point with the smallest GALI2 value.
The creation of these characteristic pointy shapes is due to the behavior
depicted in Fig. 5.16b: the GALI2 has relatively small values on the stable periodic
orbit, for which it decreases as / t1 , while it attains constant, positive values for
regular orbits in the vicinity of the periodic orbit. These constant values increase as
the orbits initial conditions depart further away from the periodic orbit. So, more
generally, the appearance of such pointy formations in GALIk plots (2  k  N)
provide good indications for the location of stable periodic orbits.
Let us now turn our attention to maps. In 2Nd symplectic maps stable periodic
orbits of period l correspond to l distinct points (the so-called stable fixed points
of order l). Any deviation vector from the periodic orbit rotates around each fixed
160 Ch. (Haris) Skokos and T. Manos

point. This behavior can be easily seen in the case of 2d maps where the tori around
a stable fixed point correspond to closed invariant curves which can be represented,
through linearization, by ellipses (see for example Sect. 3.3b of [53]). Thus, any
k initially distinct deviation vectors needed for the computation of the GALIk will
rotate around the fixed point keeping on average the angles between them constant.
Consequently the volume of the parallelepiped they define, i.e. the value of the
GALIk , will remain practically constant. Thus, in the case of stable periodic orbits
of 2Nd maps, with N  1 we have

GALIk .t/ / constant; for 2  k  2N: (5.38)

This behavior is clearly seen in Fig. 5.18a where the evolution of the GALI2 , the
GALI3 and the GALI4 for a stable periodic orbit of period 7 of the 4d map (5.30) is
plotted.
Again small perturbations of the periodic orbits initial conditions generally
result in motion on an Nd tori. Then, the evolution of the corresponding GALIs
is provided by (5.23) for N  2, while the GALI2 will decrease to zero according
to (5.33) for 2d maps. So, the most striking difference between the behavior of the
GALIk of a stable periodic orbit and of a neighboring, regular orbit appears for
k > N, because in this case the GALIk remains constant for the periodic orbit, while
it decays to zero for the neighboring one. Differences of this kind can be observed
in Fig. 5.18b.

Fig. 5.18 The evolution of the GALI2 , the GALI3 and the GALI4 with respect to the number of
iterations n for (a) a stable periodic orbit and (b) a nearby regular orbit, of the 4d map (5.30) with
K D 0:9 and  D 0:05. The initial conditions of the orbits are: (a) x1 D 0:23666, y1 D 0:0,
x2 D 0:23666, y2 D 0:0, and (b) x1 D 0:23, y1 D 0:0, x2 D 0:236, y2 D 0:0
5 The SALI and the GALI Methods of Chaos Detection 161

5.4 Applications

The ability of the SALI and the GALI methods to efficiently discriminate between
chaotic and regular motion was described in detail in the previous sections, where
some exemplary Hamiltonian systems and symplectic maps were considered. In
what follows we present applications of this ability to various dynamical systems
originating from different research fields.

5.4.1 Global Dynamics

In Sect. 5.3.2 we discussed how one can use the various GALIs to reveal the chaotic
or regular nature of individual orbits in the 2Nd phase space of a dynamical system.
Additionally, in Sect. 5.3.4 we saw how the measurement of the GALI2 values for
an ensemble of orbits can facilitate the uncovering of some dynamical properties of
the studied system, in particular the pinpointing of stable periodic orbits (Fig. 5.17),
while in Sect. 5.3.3.1 we described how a more general search can help us locate
motion on low dimensional tori.
Now we see how one can use the GALIs in order to study the global dynamics of
a system. For simplicity we use in our analysis the 2D Hamiltonian system (5.4), but
the methods presented below can be (and actually have already been) implemented
to higher-dimensional systems.

5.4.1.1 Investigating Global Dynamics by the GALIk with 2 k N

According to (5.22) and (5.23) the GALIk , with 2  k  N, behaves in a completely

different way for chaotic (exponential decay) and regular (remains practically
constant) orbits. Thus, by coloring each initial condition of an ensemble of orbits
according to its GALIk value at the end of a fixed integration time we can produce
color plots where regions of chaotic and regular motion are easily seen. In addition,
by choosing an appropriate threshold value for the GALIk , below which the orbit
is characterized as chaotic (see Sect. 5.3.2.1 on how to set up this threshold), we
can efficiently determine the strength of chaos by calculating the percentage of
chaotic orbits in the studied ensemble. Then, by performing the same analysis for
different parameter values of the system we can determine its physical mechanisms
that increase or suppress chaotic behavior.
A practical question arises though: which index should one use for this kind
of analysis? The obvious advantage of the GALI2 /SALI is its easy computation
according to (5.8), which requires the evolution of only two deviation vectors.
On the other hand, evaluating the GALIs of order up to k D N is more CPU-
time consuming as the computation of the index from (5.20) requires the evolution
of more deviation vectors, as well as the implementation of the SVD algorithm.
An advantage of these higher order indices is that they tend to zero faster than
162 Ch. (Haris) Skokos and T. Manos

the GALI2 /SALI for chaotic orbits. So, reaching their threshold value which
characterizes an orbit as chaotic, requires in general, less computational effort.
This feature is particularly useful when we want to estimate the percentage of
chaotic orbits, as there is no need to continue integrating orbits which have been
characterized as chaotic (see Sect. 5.2 of [84] for an example of this kind). Thus, we
conclude that the reasonable choices for such global studies are the GALI2 /SALI
and the GALIN .
In order to illustrate this process, let us consider the 2D Hnon-Heiles system
(5.4), for which GALIN  GALI2 , since N D 2. In Fig. 5.19 we see color plots
of its Poincar surface of section defined by q1 D 0 (a concise description of the
construction of a surface of section can be found for instance in Sect. 1.2b of [53]).
The remaining initial conditions of each orbit are its coordinates on the .q2 ; p2 /
plane of Fig. 5.19, while p1 > 0 is set so that H2 D 0:125. For each panel of
Fig. 5.19 a 2d grid of approximately 350;000 equally distributed initial conditions is
considered. Each point on the .q2 ; p2 / plane is colored according to its log.GALI2 /
value at t D 2000, while white regions denote not permitted initial conditions.
Regions colored in yellow or light red correspond to regular orbits, while dark blue
and black domains contain chaotic ones. Intermediate colors at the borders between
these two regions indicate sticky chaotic orbits.
This kind of color plots can reveal fine details of the underlying dynamics, like for
example the small yellow islands of regular motion inside the large, black chaotic
sea, as well as allow the accurate estimation of the percentage of chaotic or regular
orbits in the studied ensemble. Naturally the denser the used grid is, the finer the
uncovered details become, but unfortunately the higher the needed computational
effort gets. In an attempt to speed up the whole process the following procedure
was followed in [3] where the dynamics of the Hnon-Heiles system (5.4) was
studied. The final GALI2 /SALI value and the corresponding color was assigned
not only to the initial condition of the studied orbit, but also to all intersection
points of the orbit with the surface of section. This assignment can be extended
even further by additionally taking into account the symmetry of Hamiltonian (5.4)
with respect to the q2 variable, which results in structures symmetric with respect
to the p2 D 0 axis in Fig. 5.19. Consequently, points symmetric to this axis should
have the same GALI2 /SALI value. So, orbits with initial conditions on grid points
to which a color has already been assigned, as they were intersection points with
the surface of section of previously computed orbits, are not computed again and so
the construction of color plots like the ones of Fig. 5.19 is speeded up significantly.
In [3] it was shown that this approach achieves very accurate estimations of the
percentages of chaotic orbits with respect to the ones obtaining by coloring each
and every initial condition according to the indexs value at the end of the integration
time (this is actually how Fig. 5.19 was produced).
Let us now discuss the differences between panels (a) and (b) of Fig. 5.19. In both
figures the chaotic regions are practically the same. Nevertheless, in the yellow and
light red colored domains, where regular motion occurs, some spurious structures
appear in Fig. 5.19a, which are not present in Fig. 5.19b. For example, inside the
large stability island with 0 . q2 . 0:5 at the right side of Fig. 5.19a we observe an
almost horizontal formation colored in light red, while similar colored arcs appear
5 The SALI and the GALI Methods of Chaos Detection 163

Fig. 5.19 Regions of different values of the GALI2 on the Poincar surface of section defined
by q1 D 0 of the 2D Hamiltonian (5.4) for H2 =0.125. A set of approximately 350;000 equally
spaced initial conditions on the grid .q2 ; p2 / 2 0:5; 0:7  0:5; 0:5 is used. White regions
correspond to forbidden initial conditions. The color scales shown at the right of the panels are
used to color each point according to the orbits log.GALI2 / value at t D 2000. In (a) the same set
of initial orthonormal deviation vectors was used for the computation of the GALI2 of each initial
condition, while in (b) a different, randomly produced set of vectors was used for each orbit

inside many other islands of regular motion. These artificial features emerge when
one uses exactly the same set of orthonormal, initial deviation vectors for every
studied orbit, as we did in Fig. 5.19a. The appearance of such features in color plots
164 Ch. (Haris) Skokos and T. Manos

of other chaos detection methods has already been reported in the literature [9]. A
simple way to avoid them is to use a different, random set of initial, orthonormal
vectors for the computation of the GALI2 , as we did in Fig. 5.19b. By doing so,
these spurious features disappear and only structures related to the actual dynamics
of the system remain, like for instance the cyclical chain of the light red colored,
elongated regions inside the big stability island at the right side of Fig. 5.19b. This
structure indicates the existence of some higher order stability islands, which are
surrounded by an extremely thin chaotic layer. This layer is not visible for the
resolution used in Fig. 5.19b. A magnification, and a much finer grid would reveal
this tiny chaotic region.

5.4.1.2 Investigating Global Dynamics by the GALIk with N < k 2N

As was clearly explained in Sect. 5.3.2.1 the GALIs of order N < k  2N tend to
zero both for chaotic and regular orbits, but with very different time rates as (5.22)
and (5.23) state. This deference can be also used to investigate global dynamics, but
following an alternative approach to the one developed in Sect. 5.4.1.1. Since these
GALIs decay to zero exponentially fast for chaotic orbits, but follow a much slower
power law decay for regular ones, the time tth they need to reach an appropriately
chosen, small threshold value will be significantly different for the two kinds of
orbits. We note that both the exponential and the power law decays become faster
with increasing order k of GALIk . Consequently, the creation of huge differences
in the GALIk values, which allow the discrimination between chaotic and regular
motion, will appear earlier for larger k values. So, in general, the overall required
computational time decreases significantly by using a higher order GALIk , despite
the integration of more deviation vectors, since this integration will be terminated
earlier. Thus, the best choice in investigations of this kind is to use the GALI2N .
Let us illustrate this approach by computing the GALI4 for the 2D Hnon-Heiles
system (5.4), at a grid in its q1 D 0 surface of section. The outcome of this procedure
is seen in Fig. 5.20, where each initial condition is colored according to the time tth
needed for its GALI4 to become  1012 . Each orbit is integrated up to t D 500 time
units and if its GALI4 value at the end of the integration is larger than the threshold
value 1012 the corresponding tth value is set to tth D 500 and the initial condition
is colored in blue according to the color scales seen below the panel of Fig. 5.20.
Regions of regular motion correspond to large tth values and are colored in blue,
while all the remaining colored domains contain chaotic orbits. Again, white regions
correspond to forbidden initial conditions. This approach yields a very detailed chart
of the dynamics, analogous to the one seen in Fig. 5.19.
An advantage of the current approach is its ability to clearly reveal various
degrees of chaotic behavior in regions not colored in blue. Strongly chaotic orbits
are colored in red and yellow as their GALI4 becomes 1012 quite fast. Orbits with
larger tth values correspond to chaotic orbits which need more time in order to show
their chaotic nature, while the sticky chaotic regions are characterized by even
higher tth values and are colored in light blue. We note that for every initial condition
5 The SALI and the GALI Methods of Chaos Detection 165

Fig. 5.20 Regions of different values of the time tth needed for the GALI4 to become less than
1012 on the q1 D 0 surface of section of the 2D Hnon-Heiles system (5.4). Each orbit is
integrated up to t D 500 time units. White regions correspond to forbidden initial conditions.
The color scales shown below the panel are used to color each point according to the orbits tth
value (after [84])

we used a different, random set of orthonormal deviation vectors in order to avoid

the appearance of possible spurious structures, like the ones seen in Fig. 5.19a.

5.4.2 Studies of Various Dynamical Systems

The SALI and the GALI methods have been used broadly for the study of the phase
space dynamics of several models originating from different scientific fields. These
studies include the characterization of individual orbits as chaotic or regular, as well
as the consideration of large ensembles of initial conditions along the lines presented
in Sect. 5.4.1, whenever a more global understanding of the underlying dynamics
was needed.
In this section we present a brief, qualitative overview of such investigations. For
this purpose we focus mainly on the outcomes of these studies avoiding a detailed
presentation of mathematical formulas and equations for each studied model.

5.4.2.1 An Accelerator Map Model

Initially, let us discuss two representative applications of the SALI. The first one
concerns the study of a 4d symplectic map which describes the evolution of a
charged particle in an accelerator ring having a localized thin sextupole magnet.
166 Ch. (Haris) Skokos and T. Manos

Fig. 5.21 Regions of different SALI values of (a) the 4d uncontrolled accelerator map studied in
[19] and (b) the controlled map constructed in [16]. The coordinates x1 , x3 respectively describe
horizontal and vertical deflections of a charged particle from the ideal circular orbit passing from
x1 D x3 D 0 in some appropriate units (see [19] for more details). 16;000 uniformly distributed
initial conditions on the grid .x1 ; x3 / 2 1; 1  1; 1 were evolved for 105 iterations of each
map and colored according to the orbits log.SALI/ value, using the color scales shown at the
right of the panels. The white colored regions correspond to orbits that escape in less than 105
iterations. Red points denote chaotic orbits, while regular ones are colored in blue. The increase of
the stability region around the point x1 D x3 D 0 is evident (after [17])

The specific form of this map can be found in [19] where the SALI method was
used for the construction of phase space color charts where regions of chaotic and
regular motion were clearly identified, as well as for evaluating the percentage of
chaotic orbits.
Later on, in [16, 17] this map was used to test the efficiency of chaos control
techniques for increasing the stability domain (the so-called dynamic aperture)
around the ideal circular orbit of this simplified accelerator model. These techniques
turned out to be quite successful, as the addition of a rather simple control term,
which potentially could be approximated by real multipole magnets, increased the
stability region of the map as can be seen in Fig. 5.21.

5.4.2.2 A Hamiltonian Model of a Bose-Einstein Condensate

Let us now turn our attention to a 2D Hamiltonian system describing the interaction
of three vortices in an atomic Bose-Einstein condensate, which was studied in [52].
By means of SALI color plots the extent of chaos in this model was accurately
measured and its dependence on physically important parameters, like the energy
and the angular momentum of the vortices, were determined.
In real experiments, from which the study of this model was motivated, the
life time of Bose-Einstein condensates is limited. For this reason the time in
which the chaotic nature of orbits is uncovered played a significant role in the
analysis presented in [52]. Actually, different degrees of chaoticity are revealed by
5 The SALI and the GALI Methods of Chaos Detection 167

(a) 0,25 3000

2500
0,20
2000
0,15
J1 1500
0,10
1000

0,05
500

0,00 0
0,0 0,2 0,4 0,6 0,8 1,0 1,2 1,4 1,6 1,8 2,0
f1/
(b) 0,25 (c) 0,25

0,20 0,20

0,15 0,15
J1 J1
0,10 0,10

0,05 0,05

0,00 0,00
0,0 0,2 0,4 0,6 0,8 1,0 1,2 1,4 1,6 1,8 2,0 0,0 0,2 0,4 0,6 0,8 1,0 1,2 1,4 1,6 1,8 2,0
f1/ f1/
(d) 0,25 (e) 0,25

0,20 0,20

0,15 0,15
J1 J1
0,10 0,10

0,05 0,05

0,00 0,00
0,0 0,2 0,4 0,6 0,8 1,0 1,2 1,4 1,6 1,8 2,0 0,0 0,2 0,4 0,6 0,8 1,0 1,2 1,4 1,6 1,8 2,0

f1/ f1/

Fig. 5.22 (a) Regions of different values of the time tth needed for the SALI to become less than
1012 for a 2D Hamiltonian describing the interaction of three vortices in an atomic Bose-Einstein
condensate. The explicit definition of the coordinates J1 and 1 = can be found in [52] where this
model was studied in detail. Each orbit is integrated up to t D 3000 time units. White regions
correspond to regular orbits, while black areas at the upper two corners, as well as in the middle
of the vertical axes at both sides of the plot, denote not permitted initial conditions. The color
scales shown at the right of the panel are used to color each point according to the orbits tth value.
The initial conditions of (a) are decomposed in four different sets according to their tth value: (b)
140  tth  500, (c) 500 < tth  1000, (d) 1000 < tth  1500 and (e) 1500 < tth  2000 (after
[52])

registering the time tth that the SALI of a chaotic orbit requires in order to become
1012 (Fig. 5.22). This approach is similar to the one presented in Sect. 5.4.1.2,
and allows the identification of regions with different strengths of chaos.
168 Ch. (Haris) Skokos and T. Manos

The chaotic orbits of Fig. 5.22a are decomposed in Figs. 5.22be in four different
sets according to their tth value: tth 2 140; 500 (Fig. 5.22b), tth 2 .500; 1000
(Fig. 5.22c), tth 2 .1000; 1500 (Fig. 5.22d) and tth 2 .1500; 2000 (Fig. 5.22e),
where time is measured in some appropriate units (see [52] for more details). From
these results we see that, as the initial conditions move further away from the center
of the x-shaped region of Fig. 5.22a the orbits need more time to show their chaotic
nature and consequently, some of them can be considered as regular from a practical
(experimental) point of view. For instance, in real experiments one would expect to
detect chaotic motion in regions shown in Fig. 5.22b where orbits have relatively
small tth values. Thus, an analysis of this kind can provide practical information
about where one should look for chaotic behavior in actual experimental set ups.

5.4.2.3 Further Applications of the SALI and the GALI Methods

The SALI and the GALI methods have been successfully employed in studies
of various physical problems and mathematical toy models, as well as for the
investigation of fundamental aspects of nonlinear dynamics (e.g. see [30]). In what
follows we briefly present some of these studies
In [65] the SALI/GALI2 method was used for the global study of the standard
map (5.32). By considering large ensembles of initial conditions the percentage
of chaotic motion was accurately computed as a function of the maps parameter
K. This work revealed the periodic re-appearance of small (even tiny) islands
of stability in the systems phase space for increasing values of K. Subsequent
investigations of the regular motion of the standard map in [62] led to the clear
distinction between typical islands of stability and the so-called accelerator modes,
i.e. motion resulting in an anomalous enhancement of the linear in time orbits
diffusion. Typically, this motion is highly superdiffusive and is characterized by
a diffusion exponent  2.
In [21] the GALI was used for the detection of chaotic orbits in many dimensions,
the prediction of slow diffusion, as well as the determination of quasiperiodic
motion on low dimensional tori in the system (5.7) of many coupled standard maps.
Additional applications of the SALI in studying maps can be found in [73], where
the index was used for shedding some light in the properties of accelerator models,
while in [76] a coupled logistic type predator-prey model describing population
growths in biological systems was considered. Further studies of 2d and 4d maps
based on the SALI method were performed in [35].
Models of dynamical astronomy and galactic dynamics are considered to be the
spearhead of the chaos detection methods [31]. Actually, many of these methods
have been used, or often even constructed, to investigate the properties of such
systems. Several applications of the SALI to systems of this kind can be found
in the literature. In [18, 88, 89] the stability properties of orbits in a particular
few-body problem, the so-called the Sitnikov problem, were studied, while in [94]
the long term stability of two-planet extrasolar systems initially trapped in the 3:1
mean motion resonance was investigated. The SALI was also used to study the
5 The SALI and the GALI Methods of Chaos Detection 169

dynamics of the Caledonian symmetric four-body problem [90], as well as the

circular restricted three-body problem [75].
In systems modeling the dynamics of galaxies special care should be taken with
respect to the determination of the star motions nature, because this has to be done
as fast as possible and in physically relevant time intervals (e.g. smaller than the
age of the universe). Hence, in order to check the adequacy of a proposed galactic
model, in terms of being able to sustain structures resembling the ones seen in
observations of real galaxies, the detection of chaotic and regular motion for rather
small integration times is imperative. The SALI and the GALI methods have proved
to be quite efficient tools for such studies, as they allow the fast characterization
of orbits. This ability reduces significantly the required computational burden, as in
many cases the determination of the orbits nature is achieved before the predefined,
final integration time.
In particular, the SALI method has been used successfully in studying the chaotic
motion and spiral structure in self-consistent models of rotating galaxies [93], the
dynamics of self-consistent models of cuspy triaxial galaxies with dark matter
haloes [23], the orbital structure in N body models of barred-spiral galaxies [42],
the secular evolution of elliptical galaxies with central masses [50], the chaotic
component of cuspy triaxial stellar systems [25], as well as the chaoticity of non-
axially symmetric galactic models [97] and of models with different types of dark
matter halo components [96].
The SALI was used in [65] for investigating the dynamics of 2D and 3D
Hamiltonian models of rotating bared galaxies. This work was extended in [60]
by using the GALI for studying the global dynamics of different galactic models
of this type. In particular, the effects of several parameters related to the shape and
the mass of the disk, the bulge and the bar components of the models, as well as
the rotation speed of the bar, on the amount of chaos appearing in the system were
determined. Moreover, the implementation of the GALI3 in the 3D Hamiltonians
allowed the detection of regular motion on low (2d) dimensional tori, although these
systems support, in general, 3d orbits. The astronomical significance of these orbits
was discussed in detail in [60].
Implementations of the SALI to nuclear physics systems can be found in [56
58, 86, 87] where the chaotic behavior of boson models is investigated, as well as in
[5] where the dynamics of a Hamiltonian model describing a confined microplasma
was studied. Recently the SALI and the GALI methods, together with other chaos
indicators, were reformulated in the framework of general relativity, in order to
become invariant under coordinate transformation [54].
The SALI and the GALI have been also used to study the dynamics of nonlinear
lattice models. Applications of these indices to the FermiPastaUlam model can
be found in [1, 2, 4, 2729, 71, 85] where the properties of regular motion on low
dimensional tori, the long term stability of orbits, as well as the interpretation of
FermiPastaUlam recurrences were studied. In [63] the GALI method managed to
capture the appearance of a second order phase transition that the Hamiltonian Mean
Field model exhibits at a certain energy density. The index successfully verified also
other characteristics of the system, like the sharp transition from weak to strong
170 Ch. (Haris) Skokos and T. Manos

chaos. Further applications of the SALI method to other models of nonlinear lattices
can be found in [4, 72].
In addition, the SALI was further used in studying the chaotic and regular nature
of orbits in non-Hamiltonian dynamical systems [6, 47], some of which model
chaotic electronic circuits [46, 48, 49].

5.4.3 Time Dependent Hamiltonians

The applications presented so far concerned autonomous dynamical systems. How-

ever, there are several phenomena in nature whose modeling requires the invocation
of parameters that vary in time. Whenever these phenomena are described according
to the Hamiltonian formalism, the corresponding Hamiltonian function is not an
integral of motion as its value does not remain constant as time evolves.
The SALI and the GALI methods can be also used to determine the chaotic or
regular nature of orbits in time dependent systems as long as, their phase space
does not shrink ceaseless or expand unlimited, with respect to its initial volume,
during the considered times. This property allows us to utilize the time evolution of
the volume defined by the deviation vectors, as in the case of the time independent
models, and estimate accurately its possible decay for time intervals where the total
phase space volume has not changed significantly.
In conservative time independent Hamiltonians orbits can be periodic (stable or
unstable), regular (quasiperiodic) or chaotic and their nature does not change in
time. Sticky chaotic orbits may exhibit a change in their orbital morphologies from
almost quasiperiodic to completely chaotic behaviors, but in reality their nature
does not change as they are weakly chaotic orbits. On the other hand, in time
dependent models, individual orbits can display abrupt transitions from regular to
chaotic behavior, and vice versa, during their time evolution. This is an intriguing
characteristic of these systems which should be captured by the used chaos indicator.
Such transitions between chaotic and regular behaviors can be seen for example in
N body simulations of galactic models. For this reason, time dependent analytic
potentials trying to mimic the evolution of N body galactic systems, are expected to
exhibit similar transitions.
An analytic time dependent bared galaxy model consisting of a bar, a disk
and a bulge component, whose masses vary linearly in time was studied in [68].
The time dependent nature of the model influences drastically the location and
the size of stability islands in the systems phase space, leading to a continuous
interplay between chaotic and regular behaviors. The GALI was able to capture
subtle changes in the nature of individual orbits (or ensemble of orbits) even
for relatively small time intervals, verifying that it is an ideal diagnostic tool for
detecting dynamical transitions in time dependent systems.
Although both 2D and 3D time dependent Hamiltonian models were studied in
[68], we further discuss here only the 3D model in order to illustrate the procedure
followed for detecting the various dynamical epochs in the evolution of an orbit. The
5 The SALI and the GALI Methods of Chaos Detection 171

main idea for doing that is the re-initialization of the computation of the GALIk , with
2  k  N, whenever the index reaches a predefined low value (which signifies
chaotic behavior) by considering k new, orthonormal deviation vectors resetting
GALIk D 1.
Let us see this procedure in more detail. In [68] the evolution of the GALI3
was followed for each studied orbit. The three randomly chosen, initial orthonormal
deviation vectors set GALI3 D 1 in the beginning of the numerical simulation
.t D 0/. These vectors were evolved according to the dynamics induced by the 3D,
time dependent Hamiltonian up to the time t D td that the GALI3 became smaller
than 108 for the fist time. At that point the time t D td was registered and three new,
random, orthonormal vectors were considered resetting GALI3 D 1. Afterwards,
the evolution of these vectors was followed until the next, possible occurrence of
GALI3 < 108 . Then the same process was repeated.
Why was this procedure implemented? What is the reason behind this strategy?
In order to reveal this reason let us assume that an orbit initially behaves in a chaotic
way and later on it drifts to a regular behavior. The volume formed by the deviation
vectors will shrink exponentially fast, becoming very small during the initial chaotic
epoch and will remain small throughout the whole evolution in the regular epoch,
unless one re-initializes the deviation vectors and the volume they define. In this
way the deviation vectors will be able to feel the new, current dynamics.
An example case of this kind is shown in Fig. 5.23. In particular, in Fig. 5.23a
we see that the evolution of the finite time mLE 1 is not able to provide valid
information about the different dynamical epochs that the studied orbit experiences.
This is due to the indexs averaging nature which takes into account the whole
history of the evolution. On the other hand, the re-initialized GALI3 (whose time
evolution is shown in Fig. 5.23b) clearly succeeds in depicting the transitions
between regular epochs, where it oscillates around positive values (such time
intervals are denoted by I and III in Fig. 5.23a, b), and chaotic ones, where it exhibits
repeated exponential decays to very small values (epoch II). From the results of
Fig. 5.23a it becomes evident that the computation of the mLE cannot be used as a
reliable criterion for determining the chaotic or regular nature of the orbit in these
three time intervals.
Another way to visualize the results of Fig. 5.23b is through the measurement of
the time td needed for the repeated re-initializations of the GALI3 , or in other words,
of the time needed for the GALI3 to decrease from GALI3 D 1 to GALI3  108 .
In Fig. 5.23c we present td as a function of the evolution time of the orbit. From
the results of this figure we see that during the time interval 7500 . t . 14;000
the value of td is rather small, indicating strong chaotic motion. For smaller times,
t . 7500, the GALI3 takes a long time to become small, suggesting the presence
of regular motion or of (relatively) weaker chaotic motion. The upwardly pointing
arrow, after t & 15;000, shows that the GALI3 no longer falls to zero, which again
indicates the appearance of a regular epoch.
After the first, successful application of the GALIs to time dependent Hamilto-
nians in [68], the same approach was followed for the study of a more sophisticated
time dependent galactic model in [61]. This analytic Hamiltonian model succeeded
172 Ch. (Haris) Skokos and T. Manos

-1.4
(a)
-1.6

-1.8 I II III
-2
log(1)
-2.2

-2.4

-2.6

-2.8

-3
0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000

t
1
(b)
0

-1

-2
log(GALI 3)

-3

-4

-5

-6

-7 I II III
-8
0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000

t
8000
(c)
7000

6000

5000

4000
td

3000

2000

1000

0
0 2500 5000 7500 10000 12500 15000

t
Fig. 5.23 Time evolution of (a) the finite time mLE 1 , (b) the re-initialized GALI3 , and (c)
the time td needed for the re-initialized GALI3 to decrease from GALI3 D 1 to GALI3  108
for a particular orbit of the 3D time dependent galactic model studied in [68]. The orbit changes
its dynamical nature from regular to chaotic and again to regular. Three characteristic epochs are
located between the vertical dashed gray lines in (a) and (b) and are denoted by I (regular), II
(chaotic) and III (regular). The arrow at the right end of (c) indicates that after t & 15;000 the
GALI3 in (b) does not fall back to zero (until of course, the final integration time t D 20;000),
which is a clear indication that in this time interval the orbit is regular (after [68])
5 The SALI and the GALI Methods of Chaos Detection 173

to incorporate the evolution of the basic morphological features of an actual N body

simulation, by allowing all the relevant parameters of its dynamical components to
vary in time.

5.5 Summary

In this chapter we presented how the SALI and the various GALIs can be used to
study the chaotic behavior of dynamical systems.
Following the history of the evolution of these indices, we initially presented in
Sect. 5.2 the underlying idea behind the introduction of the SALI: the index actually
quantifies the possible alignment of two initially distinct deviation vectors. The
natural generalization of this idea, by considering more than two deviation vectors
and checking if they become linearly dependent, led later on, to the introduction of
the GALI, as we explained in Sect. 5.3. The close relation between the two indices
was also pointed out, as according to (5.17) the GALI2 and the SALI practically
coincide

GALI2 / SALI:

Avoiding the presentation of mathematical proofs (which the interested reader

can find in the related references), we formulated in Sect. 5.3 the laws that the
indices follow for chaotic and regular orbits, providing also several numerical results
which demonstrate their validity.
In particular, for ND Hamiltonian systems (N  2) and 2Nd symplectic maps
(N  1) the GALIk tends exponentially to zero for chaotic orbits and unstable
periodic orbits following (5.22)

GALIk .t/ / exp f .1  2 / C .1  3 / C    C .1  k / tg ;

while for regular motion on as sd torus, with 2  s  N, the evolution of the GALIk
is given by (5.34)
8

< constant if 2  k  s
1
GALIk .t/ / tks if s < k  2N  s
: 1 if 2N  s < k  2N:
t2.kN/

The latter formula is quite general as (a) for s D N it provides (5.23), which
describes the behavior of the GALIk for motion on an Nd torus, i.e. the most
common situation of regular motion in the 2Nd phase space of the system, (b) for
k D 2, s D 1 and N D 1 it gives (5.33), which describes the power law decay of
the GALI2 in the case of a 2d map (the GALI2 is the only possible GALI in this
case), and (c) for s D 1 it becomes (5.37), which provides the power law decay of
174 Ch. (Haris) Skokos and T. Manos

the GALIk for stable periodic orbits of Hamiltonian systems (we remind that in the
case of stable periodic in maps all the GALIs remain constant (5.38)).
In our presentation, we paid much attention to issues concerning the actual
computation of the indices. In Sect. 5.3.1 we explained in detail an efficient way
to evaluate the GALIk , which is based on the SVD procedure (5.20), while in the
Appendix we provide pseudo-codes for the computation of the SALI and the GALI.
In Sect. 5.3.3.1 we discussed a numerical strategy for the detection of regular motion
on low dimensional tori (see Figs. 5.13 and 5.14), while in Sect. 5.3.4 we showed
how the evaluation of the GALI for an ensemble of orbits can lead to the location of
stable periodic orbits (see Figs. 5.17 and 5.18). In addition, the effect of the choice
of the initial deviation vectors on the color plots depicting the global dynamics
of a system, was discussed in Sect. 5.4.1.1, where specific strategies to avoid the
appearance of spurious structures in these plots were presented (see Fig. 5.19).
One of the main advantages of the SALI and the GALI methods is their ability
to discriminate between chaotic and regular motion very efficiently. The GALIk
with 2  k  N tends exponentially fast to zero for chaotic orbits, while it attains
positive values for regular ones. Due to these different behaviors these indices,
and in particular the GALI2 /SALI and the GALIN , can reveal even tiny details of
the underlying dynamics, if one follows the procedure presented in Sect. 5.4.1.1.
Implementing the numerical strategies developed in Sect. 5.4.1.2 we can also use
the completely different time rates with which the GALIk with N < k  2N,
tends to zero (exponentially fast for chaotic orbits and power law decay for regular
ones) in order to study the dynamics globally. Finally, in Sect. 5.4.3 a particular
numerical method, the re-initialization of the GALIk , proved to be the suitable
approach to reveal even brief changes in the dynamical nature of orbits in time
dependent Hamiltonians.
The SALI and the GALI have already proven their usefulness in chaos studies
as their many applications to a variety of dynamical systems show (see Sect. 5.4.2).
Nevertheless, several other chaos indicators have been developed over the years. A
few, sporadic comparisons between some of these methods have been performed
in studies of particular dynamical systems (e.g. [9, 75, 79, 83]). Recently, detailed
and systematic comparisons between many chaos indicators based on the evolution
of deviation vectors were conducted [33, 59], and the SALI method was added in
the software package LP-VIcode [24], which includes several of these indicators.
The main outcome of these comparative studies was that the use of more than one
chaos indicators is useful, if not imperative, for revealing the dynamics of a system.

Acknowledgements Many of the results described in this chapter were obtained in close
collaboration with Prof. T. Bountis, Dr. Ch. Antonopoulos and Dr. E. Gerlach. This work was
partially supported by the European Union (European Social Fund - ESF) and Greek national funds
through the Operational Program Education and Lifelong Learning of the National Strategic
Reference Framework (NSRF)Research Funding Program: THALES. Ch. S. would like to
thank the Research Office of the University of Cape Town for the Research Development Grant
which funded part of this study, as well as the Max Planck Institute for the Physics of Complex
Systems in Dresden for its hospitality during his visit in December 2014January 2015, when part
of this work was carried out. In addition, Ch. S. thanks T. van Heerden for the careful reading of the
5 The SALI and the GALI Methods of Chaos Detection 175

manuscript and for his valuable comments. We are also grateful to the three anonymous referees
whose constructive remarks helped us improve the content and the clarity of the chapter.

Appendix: Pseudo-Codes for the Computation of the SALI

and the GALIk

We present here pseudocodes for the numerical computation of the SALI

(Table 5.1) and the GALIk (Table 5.2) methods, according to the algorithms
presented in Sects. 5.2 and 5.3.1 respectively.

Table 5.1 Numerical computation of the SALI

Input: 1. Hamilton equations of motion and variational equations, or
equations of the map and of the tangent map.
2. Initial condition for the orbit x.0/.
3. Initial orthonormal deviation vectors w1 .0/, w2 .0/.
4. Renormalization time  .
5. Maximum time: TM and small threshold value of the SALI: Sm .
Step 1 Set the stopping flag, SF 0, the counter, i 1, and the orbit
characterization variable, OC regular.
Step 2 While .SF D 0/ Do
Evolve the orbit and the deviation vectors from time t D .i  1/
to t D i , i.e. Compute x.i / and w1 .i /, w2 .i /.
Step 3 Normalize the two vectors, i.e.
Set w1 .i / w1 .i /=kw1 .i /k and w2 .i / w2 .i /=kw2 .i /k.
Step 4 Compute and Store the current value of the SALI:
SALI.i / D min fkw1 .i / C w2 .i /k; kw1 .i /  w2 .i /kg.
Step 5 Set the counter i i C 1.
Step 6 If SALI..i  1/ / < Sm  Then
Set SF 1 and OC chaotic.
End If
Step 7 If .i > TM / Then
Set SF 1.
End If
End While
Step 8 Report the time evolution of the SALI and the nature of the orbit.
The algorithm for the computation of the SALI according to Eq. (5.8). The program computes the
evolution of the SALI with respect to time t up to a given upper value of time t D TM or until
the index becomes smaller than a low threshold value Sm . In the latter case the studied orbit is
considered to be chaotic
176 Ch. (Haris) Skokos and T. Manos

Table 5.2 Numerical computation of the GALIk

Input: 1. Hamilton equations of motion and variational equations, or
equations of the map and of the tangent map.
2. Order k of the desired GALI.
3. Initial condition for the orbit x.0/.
4. Initial orthonormal deviation vectors w1 .0/, w2 .0/, : : :, wk .0/.
5. Renormalization time  .
6. Maximum time: TM and small threshold value of the GALI: Gm .
Step 1 Set the stopping flag, SF 0, the counter, i 1, and the orbit
characterization variable, OC regular.
Step 2 While .SF D 0/ Do
Evolve the orbit and the deviation vectors from time t D .i  1/
to t D i , i.e. Compute x.i / and w1 .i /, w2 .i /, : : :, wk .i /.
Step 3 Normalize the vectors:
Do for j D 1 to k
Set wj .i / wj .i /=kwj .i /k.
End Do
Step 4 Compute and Store the current value of the GALIk :
Create matrix A.i / having as rows the deviation vectors w1 .i /,
w2 .i /, : : :, wk .i /.
Compute the singular values z1 .i /, z2 .i /, : : :, zk .i / of
matrix AT .i / by applying the SVD algorithm.
Qk
GALIk .i / D jD1 zj .i /.
Step 5 Set the counter i i C 1.
Step 6 If GALIk ..i  1/ / < Gm  Then
Set SF 1 and OC chaotic.
End If
Step 7 If .i > TM / Then
Set SF 1.
End If
End While
Step 8 Report the time evolution of the GALIk and the nature of the orbit.
The algorithm for the computation of the GALIk according to Eq. (5.20). The program computes
the evolution of the GALIk with respect to time t up to a given upper value of time t D TM or
until the index becomes smaller than a low threshold value Gm . In the latter case the studied orbit
is considered to be chaotic
5 The SALI and the GALI Methods of Chaos Detection 177

References

1. Antonopoulos, Ch., Bountis, T.: Detecting order and chaos by the Linear Dependence Index
(LDI) method. ROMAI J. 2(2), 113 (2006)
2. Antonopoulos, Ch., Christodoulidi, H.: Weak chaos detection in the Fermi-Pasta-Ulam-
system using q-Gaussian statistics. Int. J. Bifurcat. Chaos 21, 2285 (2011)
3. Antonopoulos, Ch., Manos, A., Skokos, Ch.: SALI: an efficient indicator of chaos with
application to 2 and 3 degrees of freedom Hamiltonian systems. In: Tsahalis, D.T. (ed.)
From Scientific Computing to Computational Engineering. Proceedings of the 1st International
Conference, vol. III, pp. 10821088. Patras University Press, Patras (2005)
4. Antonopoulos, Ch., Bountis, T., Skokos, Ch.: Chaotic dynamics of N-degree of freedom
Hamiltonian systems. Int. J. Bifurcat. Chaos 16, 17771793 (2006)
5. Antonopoulos, Ch., Basios, V., Bountis, T.: Weak chaos and the Melting Transition in a
confined microplasma system. Phys. Rev. E 81, 016211 (2010)
6. Antonopoulos, C., Basios, V., Demongeot, J, Nardone, P., Thomas, R.: Linear and nonlinear
arabesques: a study of closed chains of negative 2-element circuits. Int. J. Bifurcat. Chaos 23,
1330033 (2013)
7. Bario, R.: Sensitivity tools vs. Poincar sections. Chaos Solitons Fractals 25, 711726 (2005)
8. Bario, R.: Painting chaos: a gallery of sensitivity plots of classical problems. Int. J. Bifurcat.
Chaos 16, 27772798 (2006)
9. Barrio, R., Borczyk, W., Breiter, S.: Spurious structures in chaos indicators maps. Chaos
Solitons Fractals 40, 16971714 (2009)
10. Benettin, G., Galgani, L.: Lyapunov characteristic exponents and stochasticity. In: Laval, G.,
Grsillon, D. (eds.) Intrinsic Stochasticity in Plasmas, pp. 93114. Edit. Phys., Orsay (1979)
11. Benettin, G., Galgani, L., Strelcyn, J.M.: Kolmogorov entropy and numerical experiments.
Phys. Rev. A 14, 23382344 (1976)
12. Benettin, G., Galgani, L., Giorgilli, A., Strelcyn, J.M.: Tous les nombres caractristiques sont
effectivement calculables. C. R. Acad. Sc. Paris Sr. A 286, 431433 (1978)
13. Benettin, G., Froeschl, C., Scheidecker, J.P.: Kolmogorov entropy of a dynamical system with
an increasing number of degrees of freedom. Phys. Rev. A 19, 24542460 (1979)
14. Benettin, G., Galgani, L., Giorgilli, A., Strelcyn, J.M.: Lyapunov characteristic exponents for
smooth dynamical systems and for Hamiltonian systems; A method for computing all of them.
Part 1: theory. Meccanica (March) 920 (1980)
15. Benettin, G., Galgani, L., Giorgilli, A., Strelcyn, J.M.: Lyapunov characteristic exponents for
smooth dynamical systems and for Hamiltonian systems; A method for computing all of them.
Part 2: Numerical application. Meccanica (March) 2130 (1980)
16. Boreux, J., Carletti, T., Skokos, Ch., Vittot, M.: Hamiltonian control used to improve the beam
stability in particle accelerator models. Commun. Nonlinear Sci. Numer. Simul. 17, 17251738
(2012)
17. Boreux, J., Carletti, T., Skokos, Ch., Papaphilippou, Y., Vittot, M.: Efficient control of
accelerator maps. Int. J. Bifurcat. Chaos 22(9), 1250219 (2012)
18. Bountis, T., Papadakis, K.E.: The stability of vertical motion in the N-body circular Sitnikov
problem. Celest. Mech. Dyn. Astron. 104, 205225 (2009)
19. Bountis, T., Skokos, Ch.: Application of the SALI chaos detection method to accelerator
mappings. Nucl. Instrum. Methods Phys. Res. A 561, 173179 (2006)
20. Bountis, T.C., Skokos, Ch.: Complex Hamiltonian Dynamics. Springer, Berlin (2012)
21. Bountis, T., Manos, T., Christodoulidi, H.: Application of the GALI Method to localization
dynamics in nonlinear systems. J. Comp. Appl. Math. 227, 1726 (2009)
22. Broucke, R.A.: Periodic orbits in the elliptic restricted threebody problem. NASA, Jet
Propulsion Laboratory, Tech. Rep. 32-1360 (1969)
23. Capuzzo-Dolcetta, R., Leccese, L., Merritt, D., Vicari, A.: Self-consistent models of cuspy
triaxial galaxies with dark matter haloes. Astrophys. J. 666, 165180 (2007)
178 Ch. (Haris) Skokos and T. Manos

24. Carpintero, D.D., Maffione, N., Darriba, L.: LPVIcode: a program to compute a suite of
variational chaos indicators. Astron. Comput. 5, 1927 (2014)
25. Carpintero, D.D., Muzzio, J.C., Navone, H.D.: Models of cuspy triaxial stellar systems III.
The effect of velocity anisotropy on chaoticity. Mon. Not. R. Astron. Soc. 438, 28712881
(2014)
26. Casati, G., Chirikov, B.V., Ford, J.: Marginal local instability of quasi-periodic motion. Phys.
Lett. A 77, 9194 (1980)
27. Christodoulidi, H., Bountis, T.: Low-dimensional quasiperiodic motion in Hamiltonian sys-
tems. ROMAI J. 2(2), 3744 (2006)
28. Christodoulidi, H., Efthymiopoulos, Ch.: Low-dimensional q-tori in FPU lattices: dynamics
and localization properties. Physica D 261, 92 (2013)
29. Christodoulidi, H., Efthymiopoulos, Ch., Bountis, T.: Energy localization on q-tori, long-term
stability, and the interpretation of Fermi-Pasta-Ulam recurrences. Phys. Rev. E 81, 016210
(2010)
30. Cincotta, P.M., Efthymiopoulos, C., Giordano, C.M., Mestre, M.F.: Chirikov and Nekhoroshev
diffusion estimates: bridging the two sides of the river. Physica D 266, 4964 (2014)
31. Contopoulos, G.: Order and Chaos in Dynamical Astronomy. Springer, Berlin, Heidelberg
(2002)
32. Contopoulos, G., Galgani, L., Giorgilli, A.: On the number of isolating integrals in Hamiltonian
systems. Phys. Rev. A 18, 11831189 (1978)
33. Darriba, L.A., Maffione, N.P., Cincotta, P.M., Giordano, C.M.: Comparative study of varia-
tional chaos indicators and ODEs numerical integrators. Int. J. Bifurcat. Chaos 22, 1230033
(2012)
34. Dullin, H.R., Meiss, J.D., Sterling, D.: Generic twistless bifurcations. Nonlinearity 13, 203
224 (2000)
35. Faranda, D., Mestre, M.F., Turchetti, G.: Analysis of round off errors with reversibility test as
a dynamical indicator. Int. J. Bifurcat. Chaos 22, 1250215 (2012)
36. Fermi, E., Pasta, J., Ulam, S.: Studies of nonlinear problems. I. Los Alamos Rep LA-1940
(1955)
37. Froeschl, C., Gonczi, R., Lega, E.: The fast Lyapunov indicator: a simple tool to detect weak
chaos. Application to the structure of the main asteroidal belt. Planet. Space Sci. 45, 881886
(1997)
38. Gerlach, E., Eggl, S., Skokos, Ch.: Efficient integration of the variational equations of multi
dimensional Hamiltonian systems: application to the FermiPastaUlam lattice. Int. J. Bifurcat.
Chaos 22, 1250216 (2012)
39. Gottwald, G.A., Melbourne, I.: A new test for chaos in deterministic systems. Proc. R. Soc.
Lond. A 460, 603611 (2004)
40. Hadjidemetriou, J.: The stability of periodic orbits in the three-body problem. Celest. Mech.
12, 255276 (1975)
41. Haken, H.: At least one Lyapunov exponent vanishes if the trajectory of an attractor does not
contain a fixed point. Phys. Lett. A 94, 7172 (1983)
42. Harsoula, M., Kalapotharakos, C.: Orbital structure in N-body models of barred-spiral galaxies.
Mon. Not. R. Astron. Soc. 394, 16051619 (2009)
43. Hnon, M., Heiles, C.: The applicability of the third integral of motion: Some numerical
experiments. Astron. J. 69, 7379 (1964)
44. Howard, J.E., Dullin, H.R.: Linear stability of natural symplectic maps. Phys. Lett. A 246,
273283 (1998)
45. Howard, J.E., MacKay, R.S.: Linear stability of symplectic maps. J. Math. Phys. 28, 1036
1051 (1987)
46. Huang, G., Cao, Z.: Numerical analysis and circuit realization of the modified L chaotic
system. Syst. Sci. Control Eng. 2, 7479 (2014)
47. Huang, G-Q, Wu, X.: Analysis of Permanent-Magnet Synchronous Motor Chaos System.
Lecture Notes in Computer Science, vol. 7002, pp. 257263. Springer, Berlin, Heidelberg
(2011)
5 The SALI and the GALI Methods of Chaos Detection 179

48. Huang, G.Q., Wu, X.: Analysis of new four-dimensional chaotic circuits with experimental and
numerical methods. Int. J. Bifurcat. Chaos 22, 1250042 (2012)
49. Huang, G., Zhou, Y.: Circuit simulation of the modified Lorenz system. J. Inf. Comput. Sci.
10, 47634772 (2013)
50. Kalapotharakos, C.: The rate of secular evolution in elliptical galaxies with central masses.
Mon. Not. R. Astron. Soc. 389, 17091721 (2008)
51. Kantz, H., Grassberger, P.: Internal Arnold diffusion and chaos thresholds in coupled symplec-
tic maps. J. Phys. A 21, L127133 (1988)
52. Kyriakopoulos, N., Koukouloyannis, V., Skokos, Ch., Kevrekidis, P.: Chaotic behavior of three
interacting vortices in a confined Bose-Einstein condensate. Chaos 24, 024410 (2014)
53. Lichtenberg, A.J., Lieberman, M.A.: Regular and Chaotic Dynamics, 2nd edn. Springer, Berlin
(1992)
54. Lukes-Gerakopoulos, G.: Adjusting chaotic indicators to curved spacetimes. Phys. Rev. D 89,
043002 (2014)
55. Lyapunov, A.M.: The general problem of the stability of motion. Taylor and Francis, London
(1992) (English translation from the French: Liapounoff, A.: Problme gnral de la stabilit
du mouvement. Annal. Fac. Sci. Toulouse 9, 203474 (1907). The French text was reprinted in
Annals Math. Studies vol. 17. Princeton University Press (1947). The original was published
in Russian by the Mathematical Society of Kharkov in 1892)
56. Macek, M., Strnsk, P., Cejnar, P., Heinze, S., Jolie, J., Dobe, J.: Classical and quantum
properties of the semiregular arc inside the Casten triangle. Phys. Rev. C 75, 064318 (2007)
57. Macek, M, Dobe, J., Cejnar, P.: Occurrence of high-lying rotational bands in the interacting
boson model. Phys. Rev. C 82, 014308 (2010)
58. Macek, M, Dobe, J., Strnsk, P., Cejnar, P.: Regularity-induced separation of intrinsic and
collective dynamics. Phys. Rev. Lett. 105, 072503 (2010)
59. Maffione, N.P., Darriba, L.A., Cincotta, P.M., Giordano, C.M.: A comparison of different
indicators of chaos based on the deviation vectors: Application to symplectic mappings. Celest.
Mech. Dyn. Astron. 111, 285307 (2011)
60. Manos, T., Athanassoula, E.: Regular and chaotic orbits in barred galaxies - I. Applying the
SALI/GALI method to explore their distribution in several models. Mon. Not. R. Astron. Soc.
415, 629642 (2011)
61. Manos, T., Machado, R.E.G.: Chaos and dynamical trends in barred galaxies: bridging the gap
between N-body simulations and time-dependent analytical models. Mon. Not. R. Astron. Soc.
438, 22012217 (2014)
62. Manos, T., Robnik, M.: Survey on the role of accelerator modes for the anomalous diffusion:
The case of the standard map. Phys. Rev. E 89, 022905 (2014)
63. Manos, T., Ruffo, S.: Scaling with system size of the Lyapunov exponents for the Hamiltonian
mean field model. Transp. Theory Stat. Phys. 40, 360381 (2011)
64. Manos, T., Skokos, Ch., Bountis, T.: Application of the Generalized Alignment Index (GALI)
method to the dynamics of multi-dimensional symplectic maps. In: Chandre C., Leoncini, X.,
Zaslavsky, G. (eds.) Chaos, Complexity and Transport: Theory and Applications. Proceedings
of the CCT 07, pp. 356364. World Scientific, Singapore (2008)
65. Manos, T., Skokos, Ch., Athanassoula, E., Bountis, T.: Studying the global dynamics of
conservative dynamical systems using the SALI chaos detection method. Nonlinear Phenomen.
Complex Syst. 11(2), 171176 (2008)
66. Manos, T., Skokos, Ch., Bountis, T.: Global dynamics of coupled standard maps. In: Contopou-
los, G., Patsis, P.A. (eds.) Chaos in Astronomy, Astrophysics and Space Science Proceedings,
pp. 367371. Springer, Berlin, Heidelberg (2009)
67. Manos, T., Skokos, Ch., Antonopoulos, Ch.: Probing the local dynamics of periodic orbits by
the generalized alignment index (GALI) method. Int. J. Bifurcat. Chaos 22, 1250218 (2012)
68. Manos, T., Bountis, T., Skokos, Ch.: Interplay between chaotic and regular motion in a time-
dependent barred galaxy model. J. Phys. A Math. Theor. 46, 254017 (2013)
69. Nagashima, T., Shimada, I.: On the C-system-like property of the Lorenz system. Prog. Theor.
Phys. 58, 13181320 (1977)
180 Ch. (Haris) Skokos and T. Manos

70. Oseledec, V.I.: A multiplicative ergodic theorem. Ljapunov characteristic numbers for dynam-
ical systems. Trans. Moscow Math. Soc. 19, 197231 (1968)
71. Paleari, S., Penati, T.: Numerical Methods and Results in the FPU Problem. Lecture Notes in
Physics, vol. 728, pp. 239282. Springer, Berlin (2008)
72. Panagopoulos, P., Bountis, T.C., Skokos, Ch.: Existence and stability of localized oscillations
in 1-dimensional lattices with soft spring and hard spring potentials. J. Vib. Acoust. 126, 520
527 (2004)
73. Petalas Y.G., Antonopoulos, C.G., Bountis, T.C., Vrahatis, M.N.: Evolutionary methods for
the approximation of the stability domain and frequency optimization of conservative maps.
Int. J. Bifurcat. Chaos 18, 22492264 (2008)
74. Press, W.H., Teukolsky, S.A., Vetterling, W.T., Flannery, B.P.: Numerical Recipes in Fortran
77, 2nd edn. The Art of Scientific Computing. Cambridge University Press, Cambridge (1992)
75. Racoveanu, O.: Comparison of chaos detection methods in the circular restricted three-body
problem. Astron. Nachr. 335, 877885 (2014)
76. Saha, L.M., Sahni, N.: Chaotic evaluations in a modified coupled logistic type predator-prey
model. Appl. Math. Sci. 6(139), 69276942 (2012)
77. Sndor, Zs., rdi, B., Szll, A., Funk, B.: The relative Lyapunov indicator: an efficient method
of chaos detection. Celest. Mech. Dyn. Astron. 90, 127138 (2004)
78. Shimada, I., Nagashima, T.: A numerical approach to ergodic problem of dissipative dynamical
systems. Prog. Theor. Phys. 61, 16051615 (1979)
79. Skokos, Ch.: Alignment indices: a new, simple method for determining the ordered or chaotic
nature of orbits. J. Phys. A 34, 1002910043 (2001)
80. Skokos, Ch.: On the stability of periodic orbits of high dimensional autonomous Hamiltonian
systems. Physica D 159, 155179 (2001)
81. Skokos, Ch.: The Lyapunov Characteristic Exponents and their Computation. Lecture Notes in
Physics, vol. 790, pp. 63135. Springer, Berlin, Heidelberg (2010)
82. Skokos, Ch., Antonopoulos, Ch., Bountis, T.C., Vrahatis, M.N.: How does the smaller
alignment index (SALI) distinguish order from chaos? Prog. Theor. Phys. Suppl. 150, 439
443 (2003)
83. Skokos, Ch., Antonopoulos, Ch., Bountis, T.C., Vrahatis, M.N.: Detecting order and chaos in
Hamiltonian systems by the SALI method. J. Phys. A 37, 62696284 (2004)
84. Skokos, Ch., Bountis, T.C., Antonopoulos, Ch.: Geometrical properties of local dynamics in
Hamiltonian systems: the generalized alignment index (GALI) method. Physica D 231, 3054
(2007)
85. Skokos, Ch., Bountis, T.C., Antonopoulos, Ch.: Detecting chaos, determining the dimensions
of tori and predicting slow diffusion in Fermi-Pasta-Ulam lattices by the generalized alignment
index method. Eur. Phys. J. Spec. Top. 165, 514 (2008)
86. Strnsk, P., Cejnar, P., Macek, M.: Order and chaos in the Geometric Collective Model. Phys.
At. Nucl. 70(9), 15721576 (2007)
87. Strnsk, P., Hruka, P., Cejnar, P.: Quantum chaos in the nuclear collective model: classical-
quantum correspondence. Phys. Rev. E 79, 046202 (2009)
88. Soulis, P., Bountis, T., Dvorak, R.: Stability of motion in the Sitnikov 3-body problem. Celest.
Mech. Dyn. Astron. 99, 129148 (2007)
89. Soulis, P.S., Papadakis, K.E., Bountis, T.: Periodic orbits and bifurcations in the Sitnikov four-
body problem. Celest. Mech. Dyn. Astron. 100, 251266 (2008)
90. Szll, A., rdi, B., Sndor, Z., Steves, B.: Chaotic and stable behavior in the Caledonian
Symmetric Four-Body problem. Mon. Not. R. Astron. Soc. 347, 380388 (2004)
91. Voglis, N., Contopoulos, G., Efthymiopoulos, C.: Method for distinguishing between ordered
and chaotic orbits in four-dimensional maps. Phys. Rev. E 57, 372377 (1998)
92. Voglis, N., Contopoulos, G., Efthymiopoulos, C.: Detection of ordered and chaotic motion
using the dynamical spectra. Celest. Mech. Dyn. Astron. 73, 211220 (1999)
93. Voglis, N., Harsoula, M., Contopoulos, G.: Orbital structure in barred galaxies. Mon. Not. R.
Astron. Soc. 381, 757770 (2007)
5 The SALI and the GALI Methods of Chaos Detection 181

94. Voyatzis, G.: Chaos, order, and periodic orbits in 3:1 resonant planetary dynamics. Astrophys.
J. 675, 802816 (2008)
95. Wolf, A., Swift, J.B., Swinney, H.L., Vastano, J.A.: Determining Lyapunov exponents from a
time series. Physica D 16, 285317 (1985)
96. Zotos, E.E.: Classifying orbits in galaxy models with a prolate or an oblate dark matter halo
component. Astron. Astrophys. 563, A19 (2014)
97. Zotos, E.E., Caranicolas, N.D.: Order and chaos in a new 3D dynamical model describing
motion in non-axially symmetric galaxies. Nonlinear Dyn. 74, 12031221 (2013)
Chapter 6
The Relative Lyapunov Indicators: Theory
and Application to Dynamical Astronomy

Zsolt Sndor and Nicols Maffione

Abstract A recently introduced chaos detection method, the Relative Lyapunov

Indicator (RLI) is investigated in the cases of symplectic mappings and continuous
Hamiltonian systems. It is shown that the RLI is an efficient numerical tool in
determining the true nature of individual orbits, and in separating ordered and
chaotic regions of the phase space of dynamical systems. A comparison between
the RLI and some other variational indicators are presented, as well as the recent
applications of the RLI to various problems of dynamical astronomy.

6.1 Introduction

One of the most important questions of investigating a dynamical system with

i > 1 degrees of freedom is to identify the ordered or chaotic behaviour of its
orbits. If the dynamical system is governed by ordered orbits its time evolution
is predictable. On the contrary, if the dynamical system evolves through chaotic
orbits, its long-term behaviour cannot be predicted. In this paper we are considering
a special class of dynamical systems called Hamiltonian systems. The phase space
of a Hamiltonian-system usually contains both regions for ordered (predictable) and
chaotic (unpredictable) motion, therefore the informations about the locations and
extent of these regions are of high interest in investigating the evolution of such
systems. A typical class of Hamiltonian systems are the planetary systems such as
the Solar System, or extrasolar planetary systems.

Z. Sndor ()
Computational Astrophysics Group, Konkoly Observatory of the Hungarian Academy
of Sciences, Budapest, Hungary
e-mail: [email protected]
N. Maffione
Grupo de Caos en Sistemas Hamiltonianos, Facultad de Ciencias Astronmicas y Geofsicas,
Universidad Nacional de la Plata, Buenos Aires, Argentina
e-mail: [email protected]

Springer-Verlag Berlin Heidelberg 2016 183

Ch. Skokos et al. (eds.), Chaos Detection and Predictability, Lecture Notes
in Physics 915, DOI 10.1007/978-3-662-48410-4_6
184 Z. Sndor and N. Maffione

The ordered behaviour of an orbit or trajectory1 is strongly related to its stability.

By the term stability we mean that the trajectories are located in a bounded
region of the phase space. If the region of chaotic motion is not bounded in the
phase space, the trajectories could leave that domain through chaotic diffusion.
In this case the chaotic trajectories become unstable. Thus one way to perform
stability investigations of dynamical systems is the detection of ordered or chaotic
behaviour of their orbits. The stability of the planets or asteroids in the Solar
System is an outstanding question of dynamical astronomy. The ongoing discovery
of exoplanetary systems made the stability investigations of planetary systems even
more important.
In recent years there has been a growing interest in development and applica-
tion of different chaos detection methods. Beside the traditional tools such as
the largest Lyapunov Characteristic Exponent (LCE) or Lyapunov Characteristic
Number (LCN; [5]) and the frequency analysis [26], several new methods have been
developed, which can be used to detect the ordered and chaotic nature of individual
orbits, or to separate regions of ordered and chaotic motions in the phase space of
a dynamical system. These methods are the Fast Lyapunov Indicator (FLI; [17],
[20]), the Orthogonal Fast Lyapunov Indicator (OFLI; [16]), the Mean Exponential
Growth factor of Nearby Orbits (MEGNO; [9, 10]), the Spectral Distance (SD; [49]),
the Smaller ALignment Index (SALI; [40, 42]), the Generalized ALignment Index
(GALI; [43, 44]), and finally, the Relative Lyapunov Indicator (RLI; [36, 37]), whose
analysis is the main scope of this paper. We note that all of these methods are based
on the time evolution of the infinitesimally small tangent vector to the orbit, which
is provided by the variational equations. Thus these chaos detection methods can be
classified as variational indicators.
In what follows, after recalling the definition of the RLI, we present its behaviour
in symplectic mappings and continuous Hamiltonian systems. The efficiency of the
RLI is presented by a comparative study with the already mentioned variational
indicators. This paper closes with a chapter presenting the recent applications of the
RLI.

6.1.1 Definition of the RLI

The ordered or the chaotic nature of a trajectory can be characterized most precisely
by the calculation of the LCE:
1 jj.t/jj
L1 .x
/ D lim log ;
t!1 t jj 0 jj

1
In the case of continuous dynamical systems the trajectory is a continuous curve in the phase
space given by the points representing the time evolution of an initial state. In discrete dynamical
systems the set of the discrete points representing the time evolution of the system is called as
orbit.
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 185

where x
2 Rn is the initial state of the system, or in other words, the starting point
of the trajectory, and .t/ is the image of an initial infinitesimally small deviation
vector 0 between two nearby trajectories after time t. The time-evolution of  is
given by the equations of motions and their linearized equations:

dx.t/ d.t/
D f x.t/ ; D Dfx.t/ ;
dt dt

where Dfx.t/ is the Jacobian matrix of the function f W Rn ! Rn evaluated at

x.t/ 2 Rn , and x W R ! Rn ,  W R ! Rn are vector-valued functions, too. In
Hamiltonian systems if L1 .x
/ D 0, the orbit emanating from the initial state x

is ordered, if L1 .x
/ > 0 it is chaotic. A serious disadvantage of the calculation of
the LCE is that it can be obtained as a limit, thus its value can only be extrapolated,
which makes the identification of weakly chaotic orbits unreliable.
In practice, one calculates only the finite-time approximation of the LCE, called
the finite-time Lyapunov Indicator (LI):

1 jj.t/jj
L.x; t/ D log :
t jj 0 jj

It is obvious that by calculating the LI for short time, the true nature of individual
orbits cannot be identified. However, the basic features of the phase space of a
system (the existence and approximate position of regular regions and extended
chaotic domains) can be discovered very quickly by calculating a large number of
LIs for short time. Let x be a vector variable taken along a line, which is going
through both regular and chaotic regions of the phase space of a dynamical system.
Then by fixing the integration time tint , one can calculate the curve L.x; tint /. In the
case of regular regions (KAM tori, islands of stability) the curve L.x; tint / varies
smoothly, while in the case of an extended chaotic region it shows large fluctuations
[11]. However, in the case of weak chaos the fluctuations of the curve L.x; tint / are
not large enough to decide the true nature of the investigated region. In order to
measure the fluctuations of the curve of the finite-time LI at x
, we introduce the
quantity:

L.x
; t/ D jL.x
C x
; t/  L.x
; t/j ; (6.1)

which is the difference between the finite-time LI of two neighbouring orbits, and
x
is the distance between the two initial condition vectors. This quantity has
been introduced and called RLI in [36, 37]. Definition (6.1) contains x
as a free
parameter, which should be chosen small enough to reflect the local properties of the
phase space. In our numerical investigations we have experienced that the arbitrary
choice of x
in a quite large interval jjx
jj 2 1014 ; 107  does not modify
essentially the behaviour of the RLI as a function of the time. For ordered orbits the
RLI shows linear dependence on jjx
jj, while for chaotic orbits the RLI practically
is invariant with respect to the choice of jjx
jj.
186 Z. Sndor and N. Maffione

Although there is not developed a strict mathematical theory describing the time
behaviour of the RLI so far, the results of numerical simulations clearly show its
power in separating ordered and chaotic orbits. An intuitive explanation could be
that in the case of ordered orbits the time evolution of the two LI curves (L.x
; t/
and L.x
C x
; t/) practically cannot be distinguished meaning that they converge
with the same (or very similar) rate to the LCE D 0 limit. On the other hand, the
convergence rate of the LI of two close chaotic orbits (separated in the phase space
by the vector x
) could be very different, which is reflected in the time evolution
of the RLI. In the next sections of the paper we shall investigate through extensive
numerical experiments the completely different behaviour of the RLI as a function
of time in the cases of ordered and chaotic orbits, which makes it a suitable tool of
chaos detection.

6.1.2 Properties of the RLI in Chaos Detection

In order to eliminate the high frequency fluctuations of the curve L.x; t/ for a fixed
x 2 Rn , we suggested the following smoothing

t=t
1 X
hL.x/i.t/ D L.x; i  t/ ;
t iD1

where t is the stepsize. In numerical experiments we always use the above

smoothed value of the RLI.
The different behaviour of the RLI for ordered and chaotic orbits are first
presented for discrete Hamiltonian systems, such as the following 2D:

x01 D x1 C x2
(6.2)
x02 D x2    sin.x1 C x2 / mod .2/ ;

and 4D symplectic mapping:

8 0

x D x1 C x2
< 10
x2 D x2    sin.x1 C x2 /    1  cos.x1 C x2 C x3 C x4 /
x03 D x3 C x4
: 0
x4 D x4   sin.x3 C x4 /    1  cos.x1 C x2 C x3 C x4 / mod .2/ ;
(6.3)
where  and are the non-linearity parameters, and  is the coupling parameter of
the 4D mapping.
In the case of the 2D symplectic mapping (6.2) the initial conditions of the
ordered orbit are x1 D 2, x2 D 0, while the initial conditions of the chaotic orbit
are x1 D 3, x2 D 0. In both cases  D 0:5. The phase plots of these orbits are
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 187

a 1.5 b 0.01

0.0001
1
1e06
0.5

RLI
1e08
0
x2

1e10
0.5
1e12
1
1e14
1.5
3 2 1 0 1 2 3 0 5000 10000 15000 20000
x1 Number of iterations

Fig. 6.1 (a) (left) The phase plot of an ordered and a chaotic orbit in the mapping (6.2); (b) (right)
the behaviour of the RLI as the function of time for a chaotic orbit (upper curve) and for an ordered
orbit (lower curve)

a 3 b 0.0001
2
1e06
1
1e08
RLI
x2

0
1e10
1
1e12
2
1e14
3
3 2 1 0 1 2 3 0 5000 10000 15000 20000
x1 Number of iterations

Fig. 6.2 (a) (left) The phase plot of an ordered and a chaotic orbit in the mapping (6.3); (b) (right)
the behaviour of the RLI as the function of time for a chaotic orbit (upper curve) and for an ordered
orbit (lower curve)

shown in Fig. 6.1a and the corresponding time behaviour of the RLI is displayed in
Fig. 6.1b. In the case of the 4D mapping (6.3) the following initial conditions are
used: x1 D 0:5, x2 D 0, x3 D 0:5, x4 D 0 for the ordered, and x1 D 3, x2 D 0,
x3 D 0:5, x4 D 0 for the chaotic orbit. In both cases the parameters are  D 0:5,
D 0:1 and  D 0:001. The projections of these orbits onto the x1  x2 plane are
shown in Fig. 6.2a. The behaviour of the RLI as the functions of time of the ordered
and chaotic orbits are plotted in Fig. 6.2b. Studying Figs. 6.1b and 6.2b one can see
that the RLI for an ordered orbit is almost constant. The RLI of a chaotic orbit grows
very rapidly, and after reaching a maximum value it decreases very slowly.
The maximum value of the RLI of a chaotic orbit is much higher (in the examples
shown by 910 orders of magnitude) than the almost constant value of the RLI of an
ordered orbit. It can be seen that by using the RLI, the ordered or chaotic nature of
orbits can be identified after a few hundred iterations of the investigated mapping.
A crucial test for a chaos detection method is whether it separates the weakly
chaotic orbits (also called sticky orbits) from the ordered orbits. In order to
188 Z. Sndor and N. Maffione

demonstrate this property of the RLI we used the 4D symplectic mapping and initial
conditions for an ordered and a weakly chaotic orbit as [40, 49]:
8 0

x D x1 C x02
< 10
x2 D x2 C .K=2/ sin.2x1 /  .=/ sin2.x3  x1 /
(6.4)
x03 D x3 C x04
: 0
x4 D x4 C .K=2/ sin.2x3 /  .=/ sin2.x1  x3 / mod .1/ ;

where K is the non-linearity and is the coupling parameter. According to [49] the
orbit with the parameters K D 3, D 0:1 and with initial conditions x1 D 0:55,
x2 D 0:1, x3 D 0:62, x4 D 0:2 is ordered, while the orbit with the same initial
conditions and parameter K, but with D 0:3051 is slightly chaotic tending to a
very small LCE. The projection of the weakly chaotic orbit on the x1  x2 plane is
shown in Fig. 6.3a, and it seems to be an ordered orbit on a torus. Using the RLI
(Fig. 6.3b) one can see that the chaotic nature of this orbit can be detected after
about N  5  106 iterations.
Finally, we should discuss the role of the initial separation between the two orbits,
which is one of the free parameters of the RLI (the other one is the length of the
time needed to calculate the RLI, as will be mentioned later on). In what follows,
we give an evidence that jjx
jj can be chosen arbitrarily from a quite large interval
1014 ; 107 . The smallest value of this interval is due to the finite representation
of numbers by computers. On the other hand, the largest value of the above interval
should also be small enough in order to the RLI reflect the local property of the
phase space around the orbit under study.
Figure 6.4 shows the dependence of the RLI (obtained after a fixed number of
iterations, which in this particular case was 2  104 ), on jjx
jj for the 4D ordered
orbit of mapping (6.3), shown in Fig. 6.2a. In Fig. 6.4 both the horizontal and vertical
axes are scaled logarithmically. Studying Fig. 6.4a one can see that for the ordered
orbit the RLI changes linearly with respect to jjx
jj. In Fig. 6.4b we display the

a 1 b 1e07

1e08
0.8

1e09
0.6
RLI

1e10
x2

0.4
1e11

0.2
1e12

0 1e13
0 0.2 0.4 0.6 0.8 1 10000 100000 1e+06 1e+07
x1 Number of iterations

Fig. 6.3 (a) (left) The phase plot of a weakly chaotic orbit in the mapping (6.4); (b) (right) the
behaviour of the RLI as the function of time for the weakly chaotic orbit seen in the left panel
(upper curve) and for an ordered orbit (lower curve)
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 189

a 1e08 b 0.0001
1e09
8e05
1e10
1e11 6e05

RLI
RLI

1e12
1e13 4e05

1e14
2e05
1e15
1e16 0
1e14 1e13 1e12 1e11 1e10 1e09 1e08 1e07 1e14 1e13 1e12 1e11 1e10 1e09 1e08 1e07
Initial separation Initial separation

Fig. 6.4 (a) (left) The linear dependence of the final value of the RLI on the initial separation for
an ordered orbit; (b) (right) the final value of the RLI versus the initial separation for a chaotic
orbit

dependence of the RLI (after 2  104 iterations) on jjx

jj for the chaotic 4D orbit
shown in Fig. 6.2a. One can see that in the chaotic case the final value of the RLI
practically does not depend on the choice of jjx
jj. Thus in the above sense the
RLI is invariant with respect to the choice of the initial separation. This invariance
property can be explained by the fact that the RLI of a chaotic orbit practically
measures the average width of the oscillation of the LI curve (as a function of time),
which does not depend heavily on the choice of the initial separation.

6.2 A Short Comparison of the RLI to Other Methods

of Chaos Detection

In this section we compare the performances of the RLI with the variational
indicators mentioned in Sect. 6.1: the LI, the FLI and the OFLI, the MEGNO, the
SD, the SALI and the GALI.
In Sect. 6.2.1 we analyze the dependence of the RLI on its free parameters and
in the following one we compare the typical behaviours of the RLI with other
techniques in the well-known HnonHeiles model [21] (hereinafter HH model).
In Sect. 6.2.3 we apply the RLI and other indicators to study dynamical systems
of different complexity: the 4D symplectic mapping (6.3) presented in Sect. 6.1.2
and a rather complex 3D potential resembling a Navarro, Frenk and White triaxial
halo (hereinafter NFW model; [48]). We compare the phase space portraits given
by the RLI and the other methods to decide whether the results are comparable. In
Sect. 6.2.4 we briefly discuss the dependence of the RLI on the computing times.
The BulirschStoer integrator is used throughout this section.
190 Z. Sndor and N. Maffione

6.2.1 The Dependence of the RLI on the Free Parameters

The RLI has two free parameters: (a) the initial separation .x
/ between the basis
orbit and its shadow (Sect. 6.1.1) and (b) its threshold (the threshold is a value that
separates chaotic from regular motion and it is related with the length of the time
needed to calculate the RLI). These free parameters are user-choice quantities.
Thus, it is of interest to study the dependence of the RLI on both of them.
The following experiments are undertaken on the HH model:

1 2  1 2  1
H D px C p2y C x C y2 C x2  y  y3 ;
2 2 3
where H is the Hamiltonian and x; y; px ; py are the usual phase space variables.

6.2.1.1 The Initial Separation Parameter

The dependence of the RLI on the initial separation parameter is strongly related
to the type of orbit under study (see Sect. 6.1.1). Therefore, we take four different
types of orbits with initial conditions located on the line defined by x D py D 0 and
y 2 0:1 W 0:1 and the energy surface E D 0:118, namely, a regular orbit close
to a stable periodic orbit (r-sp); a quasiperiodic orbit (r-qp); a regular orbit close to
an unstable periodic orbit (r-up); and a chaotic orbit inside a stochastic layer (c-sl).
The initial conditions are taken from [10]. The integration time is 104 units of time
(hereinafter u.t.), which is enough time to provide a reliable characterization of the
orbits and the stepsize of the numerical integration is 0:01. We note that these values
have been used in the following numerical experiments, too.
In Fig. 6.5a, we present the final values (i.e. the values of the indicator at the end
of the integration time) of the RLI as a function of the initial separation parameter2
for the orbits introduced earlier. We show that the initial separation parameter does
not significantly affect the RLI when we apply the indicator to the chaotic orbit c-
sl, but it does when we apply it to the regular orbits r-sp, r-qp and r-up (and
confirming the results shown in Sect. 6.1.2). This dependence of the RLI on its free
parameter has severe implications in the selection of the threshold. For instance, if
we start the computation with an initial separation of 1014 , the relation shown in
Fig. 6.5a will indicate that a good candidate for the threshold to distinguish between
the chaotic orbit c-sl (RLI 0:1) and the regular orbit r-sp (RLI 1013:5 ) can
be 1012 . Then, the orbits with values of the RLI higher than 1012 will be classified
as chaotic orbits. However, this choice of the threshold leads to a misclassification
of the regular orbits r-up (RLI 1010 ) and r-qp (RLI 1012 ). Furthermore,
since the correspondence between the RLI and the initial separation parameter for

2
The values for the parameter have been taken from the interval suggested in Sect. 6.1.2.
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 191

1 2

1.8
0.01
1.6
0.0001 1.4
r-sp
r-qp 1.2
1e-06 r-up
RLI

c-sl 1
1e-08
0.8 RLI
MEGNO
1e-10 0.6 FLI/OFLI
SALI
0.4 GALI2
1e-12 GALI3
0.2 GALI4
1e-14 0
1e-14 1e-13 1e-12 1e-11 1e-10 1e-09 1e-08 1e-07 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Initial Separation Time

Fig. 6.5 (a) (left) The RLI values as a function of the initial separation parameter for the orbits
r-sp, r-qp, r-up and c-sl; (b) (right) the normalized approximation rates for several chaos
indicators, including the RLI (see text for further details)

the regular orbits of the sample tends to be linear (see Sect. 6.1.2), the threshold
1012 does not work at all for an initial separation greater than 1011 .
In order to determine a reliable threshold for the RLI, the relationship with the
initial separation parameter should be done by computing the indicator for a group
of orbits known to be regular but with some level of instability (e.g. regular orbits
close to a hyperbolic object such as an unstable periodic orbit, see [27]).

6.2.1.2 The Threshold

In this section we investigate how the RLI and other indicators (listed above)
depend on their thresholds. For the following experiment in the HH model we have
adopted a sample of 125751 initial conditions in the region defined by x D 0,
y 2 0:1 W 0:1, py 2 0:05 W 0:05 and on the energy surface defined by
E D 0:118. The thresholds of the LI, the RLI, the MEGNO, the SALI, the FLI/OFLI
and the GALIs are shown in Table 6.1, where t is the time (see [13]). From here,
the initial separation parameter will be 1012 , unless stated otherwise. The threshold
used for the RLI has been computed following the remarks discussed in the previous
section.
To proceed with the experiment we define the approximation rate as the rate
of convergence with a final percentage of chaotic orbits. This rate will show a
combination of the reliability of the indicator and the accuracy of the selected
threshold if the final percentage of chaotic orbits approaches the true percentage
of chaotic orbits in the system. Therefore, as we require a reliable final percentage
of chaotic orbits, we consider the percentage of chaotic orbits given by the LI by
104 u.t.: 39:92 %, the true percentage of chaotic orbits in the system. Both
the overwhelming number of papers claiming the reliability of the LI as a chaos
indicator and the experimental evidence showing that 104 u.t. seems to be a reliable
convergent time for all the indicators in the experiment (see, for instance, Sect. 6.2.2,
192 Z. Sndor and N. Maffione

Table 6.1 Thresholds for Indicator Threshold

several indicators, including
the RLI LI ln.t/=t
RLI 1010
MEGNO 2
SALI 104
FLI/OFLI t
GALI2 t1
GALI3 t2
GALI4 t4

CIs vs RLI (HH model)

(a) 30 (d) 1
25 0.01
0.0001
20 I II
MEGNO

1e-06

SALI
15 1e-08 I II
10 1e-10
1e-12
5 1e-14
2
0 1e-16
10 100 300 1000 10000 10 100 400 800 3000 10000
(b) 1e+16
(e) 1
1e+14 0.01
1e+12 I II 0.0001
1e+10 1e-06
GALI3
OFLI

1e+08 1e-08
1e+06 1e-10
10000 1e-12 I II
100 1e-14
1 1e-16
10 100 300 1000 10000 10 100 300 600 3000 10000

(c) 1 (f) 1
I II 0.01
0.1 0.0001
1e-06
GALI4
LI

0.01 1e-08
1e-10
0.001 1e-12
1e-14 I II
0.0001 1e-16
10 100 300 600 3000 10000 10 100 200 550 3000 10000

(g) 1
0.01
I II
r-sp 0.0001
r-qp 1e-06
RLI

r-up
c-sl 1e-08
c-cs 1e-10
Thresholds
1e-12
1e-14
10 80 200 850 3000 10000

Time [u.t.]

Fig. 6.6 Behaviours of (a) the MEGNO, (b) the OFLI, (c) the LI, (d) the SALI, (e) the GALI3 ,
(f) the GALI4 and (g) the RLI for orbits r-sp, r-qp, r-up, c-sl and c-cs. The thresholds as
well as the lines I and II are included (see text for further details)

Fig. 6.6) supporting this statement. Thus, we normalize the time evolution of the
percentage of chaotic orbits given by the methods with this true percentage.
Hence, the values of the normalized approximation rates higher than 1 show
percentages of chaotic orbits higher than the true percentage.
We test the reliability of the thresholds given in Table 6.1 according to the above
mentioned rates. The results are indicated in Fig. 6.5b. The convergence towards a
constant rate of the RLI and the FLI/OFLI is faster than that of the other indicators of
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 193

the sample. Despite this rapid tendency towards a constant value for both indicators,
the final percentages of chaotic orbits given by the RLI and the FLI/OFLI are
higher than that of the LI, which means that the values of the rates are above 1.
Nevertheless, this slight difference in the final percentages of chaotic orbits can
always be fixed with a small adjustment of the corresponding thresholds. Since the
results for the MEGNO show substantial disagreement between the percentages
of the chaotic orbits given by the indicator and the LI, a significant empirical
adjustment of the MEGNOs threshold should be made to avoid an overestimation
in the number of chaotic orbits. The final percentages given by the SALI and the
GALIs are in perfect agreement with the true percentage. However, their tendency
towards a stable percentage of chaotic orbits is slower than the one showed by the
RLI or the FLI/OFLI.
Thus, among the above mentioned CIs, the RLI and the FLI/OFLI show the best
approximation rates, i.e. the best combination of the reliability of the indicators and
the accuracy of their thresholds. For further details on the experiment, refer to [13].
By 104 u.t. the threshold taken for the GALI4 reaches the computers precision
(1016 ) and thus, every chaotic orbit lies beyond such precision. Therefore, the last
point for GALI4 in Fig. 6.5b falls apart from the tendency.
Now that we have finished studying the importance of a wise selection of the
free parameters of the RLI, we calibrated the indicator following the suggestions
mentioned here and continue comparing its performance with other indicators.

6.2.2 Expected Behaviour of the Indicators in the HH Model

In this experiment, done in the HH model, our goal is to compare the typical
behaviours of the RLI to the other techniques and show its advantages and
disadvantages.
We take the orbits of Sect. 6.2.1.1 and a chaotic orbit inside the chaotic sea (c-cs)
with initial conditions on the line defined by x D py D 0 and y 2 0:1 W 0:1 and
the energy surface E D 0:118 (the initial conditions are taken from [10]). The final
integration time is 104 u.t. and the stepsize is 0:01.
Figure 6.6 shows the time evolution curves of several indicators for the five
different types of orbits introduced at the beginning of the section. Some of the
main features of a chaos indicator are the speed of convergence and the resolving
power. The former is the time it takes to distinguish between chaotic and regular
motion. In order to visualize this quantity, in Fig. 6.6 we introduce the vertical lines
I and II. The first one shows the time after which the orbit c-cs is clearly
identified as a chaotic orbit with the indicator and the second one plays the same
role as I for the orbit c-sl.
On panel (a) we present the typical behaviours of the MEGNO (see [9]). The
values of the indicator for the three regular orbits tend towards the theoretical
asymptotic threshold, 2, in different ways (see the right bottom of panel a). The
values for the chaotic orbits increase linearly with time. At 300 u.t. (see line I),
194 Z. Sndor and N. Maffione

the separation between orbit c-cs and the threshold line is already significant.
Hence, orbit c-cs is clearly identified as a chaotic orbit then. Only 100 u.t. later
(line II) the same happens with orbit c-sl.
On panel (b) we show the time evolution curves of the OFLI [16]. The values
of the indicator for two of the regular orbits increase linearly with time (see the
threshold and its expression in Table 6.1) while the values for the chaotic orbits
increase exponentially fast with time. This distinction between both tendencies can
be made at the same times that have been shown by the MEGNO. Besides, orbit r-
sp has an almost constant value because it is very close to a stable periodic orbit.
On panel (c) we present the LI (see e.g. [41]). The values of the indicator for
the three regular orbits decrease with time (see the threshold and its expression in
Table 6.1) while the values for the chaotic orbits tend towards a constant value which
depends on the chaoticity of the orbit. The distinction between the regular orbits and
the chaotic orbit c-sl is made by the LI later than by the previous indicators: the
orbit leaves the linear tendency of the threshold around 600 u.t. (line II).
On panel (d) we present the time evolution curves of the SALI [40]. The values
of the indicator for the regular orbits r-qp and r-up oscillate within the interval
(0,2), while the orbit r-sp tends towards 0 following a power law behaviour. The
chaotic orbits decrease exponentially fast with time. The time needed for the SALI
to clearly identify the chaotic orbits is the time used by the chaotic orbits to reach
the threshold (see its value in Table 6.1 and lines I and II in the figure to locate
the times). The indicator delays making this distinction (in fact, it does so later than
the LI) because for smaller values of the integration time, the chaotic orbits decrease
with a power law as the regular orbit r-sp does.
On panels (e) and (f) we show the time evolution curves of the GALI3 and the
GALI4 , respectively (the GALI2 and the SALI have almost identical behaviours and
the former is not included). Their theoretical thresholds (Table 6.1; see [8, 29, 43, 44]
for further details) yield good estimations of the time needed for the indicators
to distinguish the chaotic orbits from the regular orbits. The GALI3 makes this
distinction in the same time as the LI did (see lines I and II). Once again, the
reason for this delay is that the GALI3 decreases with a power law for regular orbits
as well as for chaotic orbits at the beginning of the integration interval. Nevertheless,
the higher the order of the GALI, the faster its tendency towards 0 for the chaotic
orbits. Then, the GALI4 has registered the best time so far to distinguish the chaotic
orbit c-cs: 200 u.t.
On panel (g) we present the time evolution curves of the RLI (Sect. 6.1.2). The
values of the indicator for the three regular orbits are in the interval .1012 ; 1010 /
according to the initial separation of 1012 (see Fig. 6.5a, in Sect. 6.2.1.1). Thus, as
the value 1010 (depicted with a dotted line in the figure) that have been selected
in Sect. 6.2.1.2 is in the limit of the interval, it is not reliable as a threshold any
more. Therefore, we selected the value 108 for the threshold. The characterization
of the regular orbits does not clearly differentiate among them as the MEGNO, the
OFLI or the SALI. The values for the chaotic orbits increase with time until they
reach a constant value. On the one hand, orbit c-cs is clearly identified as a chaotic
orbit by 80 u.t. (line I) when the orbit reaches the threshold. This is the fastest
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 195

characterization of the chaotic orbit c-cs. On the other hand, the RLI identifies
orbit c-sl as a chaotic orbit around 850 u.t. (line II), which is the slowest
characterization.
All the indicators delay in making a reliable characterization of the chaotic orbit
c-sl, which shows that the chaotic orbit c-cs has a larger LI than orbit c-sl.
Finally, the characterization of the five representative orbits made by the RLI as
well as its speed of convergence is similar to the other techniques. Thus, the RLI is
most welcome to the group of fast variational indicators.

6.2.3 Performances of the Indicators Under Different

Scenarios

We have seen in the previous section how similar are the performances of the RLI
and the other fast indicators in the rather simple HH model. Here we will focus on
experiments in scenarios that are different from the HH model to determine whether
the RLI is a reliable technique for studying different or more complex systems than
the HH model.

6.2.3.1 The 4D Symplectic Mapping

The time evolution curves of the indicators (used in Sect. 6.2.2) are not efficient
to analyze a large number of orbits. The appropriate way to gather information in
these cases is in terms of the final values of the methods. Thus, let us now turn
our attention to the study of the resolving power of the techniques using their final
values.
The following study will be conducted in the 4D mapping (6.3) presented in
Sect. 6.1.2 by adopting different samples of initial conditions and 105 iterations.
The version of the MEGNO considered here is the MEGNO(2,0), whose threshold
value is 0:5 (see [10]).
The large number of iterations used in the experiments deserves a further
explanation. In Fig. 6.7, we present the RLI mappings for 103 (left panel), 5  103
(middle panel) and 104 (right panel) iterations. The RLI mapping corresponding
to 103 iterations presents a very noisy phase space portrait probably due to a
combination of a poor election of the initial deviation vectors (see for instance [1, 2])
and the short number of iterations. It is also clear from the figure that the phase space
portrait presents a stable picture after 5  103 iterations. Furthermore, increasing the
number of iterations helps to resolve very sticky orbits but no further advantage is
observed. Thus, we iterate the map 105 times in order to distinguish the most sticky
regions.
196 Z. Sndor and N. Maffione

3
RLI mapping, 10 iterations RLI mapping, 5x103 iterations 4
RLI mapping, 10 iterations
0 0 0 0 0 0
-2 -0.5 -2 -2
-0.5 -0.5
-4 -4 -4
-1 -1 -1
-6 -6 -6
-1.5 -1.5 -1.5

x2

x2
x2

-8 -8 -8
-2 -10 -2 -10 -2 -10
-12 -12 -12
-2.5 -2.5 -2.5
-14 -14 -14
-3 -3 -3
-16 -16 -16
-3 -2.5 -2 -1.5 -1 -0.5 0 -3 -2.5 -2 -1.5 -1 -0.5 0 -3 -2.5 -2 -1.5 -1 -0.5 0
x1 x1 x1

Fig. 6.7 The RLI mapping on gray-scale plots composed of 106 initial conditions, for 103 (left),
5  103 (middle) and 104 (right) iterations

RLI mapping 0.54

0 0
MEGNO(2,0) 0.52
-2
-0.5 0.5
-4 0.48
-1
0.46
-6
-3 -2.5 -2 -1.5 -1 -0.5 0
-1.5
x2

-8
0
-2 LI
-2 -10 RLI
-4
-6
LI, RLI

-12
-2.5 -8
-10
-14
-12
-3 -14
-16 -16
-3 -2.5 -2 -1.5 -1 -0.5 0
-3 -2.5 -2 -1.5 -1 -0.5 0
x1

Fig. 6.8 (a) (left) The RLI mapping on gray-scale plot composed of 106 initial conditions, for 105
iterations (in logarithmic scale); (b) (right) the MEGNO(2,0), the LI and the RLI final values for
103 initial conditions along the line x2 D 3, for 105 iterations (the LI and the RLI final values
are in logarithmic scale)

Initial Conditions Inside High-Order Resonances

In Fig. 6.8a we present the RLI mapping for a region of the phase space corre-
sponding to the 4D mapping and is composed of 106 initial conditions. The main
resonance as well as the high-order resonances are clearly depicted in dark gray (i.e.
small values of the RLI) while the stochastic layers inside the main stability island
and the chaotic sea are depicted in light gray (large values of the RLI). We show an
horizontal line of initial conditions .x1 2 ; 0, x2 D 3, x3 D 0:5 and x4 D 0/
used in the following experiment to compare the performances of the RLI with the
mostly used variational indicator, the LI, and with the MEGNO(2,0), which is faster
than the LI and which is also a reliable indicator. With a diagonal segment we depict
the initial conditions .x1 D x2 2 1:03; 0:8, x3 D 0:5 and x4 D 0/ used in the
experiment of next section.
In Fig. 6.8b, we compare the performances of the RLI, the LI and the
MEGNO(2,0) on 103 equidistant initial conditions lying on the horizontal line
that crosses the high-order resonances in Fig. 6.8a. This figure clearly shows that
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 197

the RLI unzips the hidden structure inside the high-order resonances better than
the LI. Furthermore, the RLI and the MEGNO(2,0) reveal similar structures (the
Y-range of the MEGNO(2,0) has been centered on the threshold and shortened
to amplify the details of the revealed structure). For further discussions on the
experiment, refer to [27].
On behalf of the previous experiments the RLI is not only more reliable than the
LI to reveal small scale structures, but also an accurate indicator to describe a large
array of initial conditions.

Sticky Orbits

Sticky orbits are the most difficult type of orbit to characterize by a variational indi-
cator. Thus, we further analyze the identification of this type of orbits (Sect. 6.1.2)
to study the performance of the RLI.
In Fig. 6.9, we show the sticky region enclosed in the interval .1:03; 0:8/ (and
depicted earlier in Fig. 6.8a with a diagonal segment) in terms of the final values of
the same indicators previously used. We also point out the final values of three
representative orbits, two chaotic orbits (one of them which is sticky chaotic) and a
regular orbit. The thresholds are also included.

(a) -1.5 (b) -2

Threshold Threshold
-2 Chaotic orbit -4 Chaotic orbit
Sticky chaotic orbit Sticky chaotic orbit
-2.5 -6
Regular orbit Regular orbit
LCN

RLI

-3 -8

-3.5 -10

-4 -12

-4.5 -14
-1 -0.95 -0.9 -0.85 -0.8 -1 -0.95 -0.9 -0.85 -0.8
x1=x2 x1=x2

(c) 350
Threshold
300
Chaotic orbit
250 Sticky chaotic orbit
MEGNO(2,0)

200 Regular orbit

150

100

50

-50
-1 -0.95 -0.9 -0.85 -0.8
x1=x2

Fig. 6.9 Sticky region inside the interval .1:03; 0:8/ for 105 iterations and for three indicators:
(a) the LI, (b) the RLI and (c) the MEGNO(2,0). The representative orbits are depicted with
points of different types. The thresholds are depicted with a dashed line (the LI and the RLI final
values are in logarithmic scale). See text for details. The figures were taken from [27] and slightly
modified
198 Z. Sndor and N. Maffione

In Fig. 6.9a, the LI perfectly identifies the three representative orbits and their
domains. In Fig. 6.9b, some sticky and chaotic orbits share the same RLI final
values (103:5 ) which hide the different levels of chaoticity (such is the case
of the representative chaotic and sticky chaotic orbits, both of them have similar
RLI values). This is not the case for the other indicator shown in the figure: the
MEGNO(2,0) on panel (c). The MEGNO(2,0) has completely different final values
for the sticky and the chaotic orbits.
The RLI shows a reliable performance revealing the global characteristics of the
region, such as the regular domain (where we find the representative regular orbit)
and some small high-order resonances (e.g. x1 D x2  0:85). Nevertheless, it
does not distinguish the sticky from the chaotic orbit in the experiment (see [27] for
further details).
In the following section, we follow with the experiments in a scenario of
completely different nature and much more complex than the HH model or the 4D
mapping.

6.2.3.2 The 3D NFW Model

According to the current paradigm of a hierarchically clustering universe, large

galaxies formed through the accretion and mergers of smaller objects. The imprints
of such events should be well preserved in the outer stellar haloes where dynamical
mixing processes are not significant in relatively short times (for instance, many
stellar streams have been identified in the outer regions of the Milky Way [4, 22,
28]). Furthermore, this galaxy formation paradigm predicts that the centres of the
accreted component of stellar haloes should contain the oldest products of accretion
events [12] in the formation history of the galaxy (such as the Milky Way), and
therefore, this substructure might be hidden in its inner regions (e.g. close to the
Solar neighbourhood). However, in order to study the phase space portraits of stars
in small volumes in the inner regions of the stellar haloes to quantify and classify the
substructure, we need a model of the Dark Matter (DM) halo that hosts the galaxy.
In [33, 34] the authors introduced a universal density profile for DM haloes (i.e.
haloes with masses ranging from dwarf galaxy haloes to those of major clusters):
the NFW profile. However, in Cold Dark Matter (CDM) cosmologies DM haloes
are not spherical. Furthermore, numerical simulations suggest that their shape vary
with radius. In [48] the authors have built a triaxial extension of the NFW profile,
resembling a triaxial DM halo; the corresponding potential is the so-called NFW
model. Therefore, the NFW model is a triaxial potential used in galactic dynamics
associated with equilibrium density profiles of DM haloes in CDM cosmologies.
In a forthcoming paper we study the phase space portraits of stellar particles
inside solar neighbourhood-like volumes to gain insights about the formation history
of Milky Way-like galaxies. Hereinafter, we use some results from that investigation
to demonstrate the reliability of the RLI on a rather complex model.
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 199

The solar neighbourhood volume is a sphere of 2.5 Kpc of radius located at 8 Kpc
from the center of the NFW model (denoted by N in the following equation):
 
A rp
N D  ln 1 C ;
rp rs

where A is the constant:

G M200
AD ;
ln .1 C CNFW /  CNFW = .1 C CNFW /

with G, the gravitational constant, M200 , the virial mass of the DM halo and
CNFW , the concentration parameter used to describe the shape of the density profile.
Besides, rp follows the relation:

.rs C r/re
rp D ;
rs C re

where rs is the scale radius defined by dividing the virial radius of the DM halo by
CNFW . The scale radius represents a transition scale between an ellipsoidal and a
near spherical shape of the N . The re is an ellipsoidal radius:
r  y 2  z 2
x 2
re D C C ;
a b c

where b=a and c=a are the principal axial ratios with a the major axis and where
we require a2 C b2 C c2 D 3 (see [48]). The values of the constants used in the
following experiment are taken at redshift z D 0 and listed in Table 6.2 (see [19]
and references therein for further details on the model).
In order to begin the study of the (6 dimensional) regions of interest in the
phase space of the NFW model, we needed to restrain some of the variables that
defined the original sample of 22,500 initial conditions. We fixed the positions of
the particles to the centre of the solar neighbourhood. Then, the stellar particles
had the following positions at the beginning of the simulation: x D 8, y D 0 and
z D 0 (in [Kpc]). The initial velocity in the polar axis (vz ) was restrained to the
value 250 in [km s1 ]. The energy (E) was restrained to the interval .Emb ; Elb /
with Emb  195;433 the energy of the most bound particle, and Elb  59;293
(in [M Kpc2 Gyr2 ] with [M ] the mass in solar mass units) the energy of the

Table 6.2 Constants used for A 4158670.1856267899

the N potential
rs 19.044494521343964
a 1.3258820840000000
b 0.86264540200000000
c 0.70560584600000000
200 Z. Sndor and N. Maffione

10 [Gyrs] 150 periods

1 0 1 0

-2 -2
0.8 0.8

-4 -4

0.6 0.6

log10(RLI)

log10(RLI)
Enorm

Enorm
-6 -6

0.4 -8 0.4 -8

-10 -10
0.2 0.2

-12 -12
0 0
-2000 0 2000 -2000 0 2000
Lz Lz

Fig. 6.10 Gray-scale plots of the RLI mapping for the velocity surface vz D 250. (left) For a
fixed integration time of 10 Gyrs. (right) For a fixed number (150) of radial periods. The values of
the indicator are in logarithmic scale

least bound particle of the sample. The angular momentum (Lz ) was restrained to
the interval .2000; 2000/ in [Kpc km s1 ]. We integrated the initial conditions for
different integration times (1 u.t. corresponds to 1 Gyr).
In Fig. 6.10, we present the phase space portraits given by the RLI for two
different choices of the integration times. On the left panel, we integrate the orbits
for a fixed time interval of the order of the Hubble time, i.e. 10 Gyrs. On the right
panel, we integrate the orbits for a fixed number of radial periods. The radial period
of the stellar orbits in galactic potentials such as the NFW model scales as E3=2 .
Then, we integrate the orbits for 150 radial periods in order to have a stable portrait
of the phase space. It is evident that 10 Gyrs (left panel) is not enough to classify
properly the orbits with the RLI (or any other indicator). Then, on the right panel
the time interval used was 57;750 Gyrs where 750 Gyrs is enough time to set
reliable values of the RLI for the least bound particle of the sample. However,
the most sticky regions are not clearly depicted yet. Therefore, in the following
experiment we choose to scale the integration time linearly with the energy of the
orbit. The linear relation between the computed integration times and the energy
overestimates the former for the most bound particles. Indeed, the time interval used
for the experiment was 204;750 Gyrs where the integration times are clearly
larger than those applied with the E3=2 scale. The larger integration times given
by the linear scale improve the identification of the most sticky regions which helps
to evaluate the performance of the indicators.
On the left panels of Figs. 6.11 and 6.12 we present the gray-scale plots of the
final values of four different chaos indicators in the .Enorm ; Lz / plane3 : the RLI
and the MEGNO; panels (a) and (b1) of Fig. 6.11, respectively; the OFLI and the
1/GALI3 : panels (c1) and (d1) of Fig. 6.12, respectively. The GALI3 is inverted in

3
The Enorm is the normalized energy: .E  Emb / = .Elb  Emb /.
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 201

(a) 1 0

-2
0.8

-4

0.6

log10(RLI)
Enorm
-6

0.4 -8

-10
0.2

-12
0
-2000 0 2000
Lz

(b1) 1 30 (b2) 1
700

25
0.8 0.8 600

20 500

MEGNOsat
0.6 0.6
MEGNO
Enorm

Enorm

400
15
0.4 0.4 300

10
200
0.2 0.2
5 100

0 0 0
-2000 0 2000 -2000 0 2000
Lz Lz

Fig. 6.11 Gray-scale plots for the velocity surface vz D 250. (a) The RLI mapping (the values
of the indicator are in logarithmic scale), (b1) the MEGNO mapping and (b2) the MEGNOsat
mapping

the gray-scale plots to make the comparison of the portraits with the other indicators
easier.
In many situations it is useful to define a saturation value by which the chaos
indicator saturates. For instance, the OFLI and the GALI3 for chaotic motion
increases or decreases exponentially fast, respectively. Then, if the chaotic nature
of an orbit is well characterized by the OFLI when the indicator reaches 1016 or
by the GALI3 when it reaches the computers precision (1016 ), the computation
should be stopped. Hence, the values 1016 and 1016 can be used as saturation values
for the OFLI and the GALI3 , respectively. Another example is the MEGNO: the
MEGNO has an asymptotic value for regular orbits, 2 (see Table 6.1), and increases
linearly for chaotic orbits. Then, if the MEGNO reaches the value 30, the orbit is
undoubtfully chaotic and it is worthless to continue the computation of the indicator.
Then, the value 30 can be used as a saturation value for the MEGNO. The time of
saturation, that is the time by which the indicator saturates, it is a quantity useful in
recovering the chaoticity levels in the chaotic domain: the smaller the value of the
time of saturation, the more chaotic the orbit (see [27, 43]). Finally, if the indicator
202 Z. Sndor and N. Maffione

(c1) 1 16 (c2) 1
700
14
0.8 0.8 600
12
500

log10(OFLI)
0.6 10 0.6

OFLIsat
Enorm

Enorm
400
8
0.4 0.4 300
6
200
0.2 4 0.2
100
2
0 0 0
-2000 0 2000 -2000 0 2000
Lz Lz
1
(d1) 1 16 (d2)
700
14
0.8 600
0.8
12
500
log10(GALI3-1)

10 0.6
0.6

GALI3sat
Enorm

400
Enorm

8
0.4 300
0.4
6
200
4 0.2
0.2
100
2

0 0
0 0
-2000 0 2000
-2000 0 2000
Lz
Lz

Fig. 6.12 Gray-scale plots for the velocity surface vz D 250. (c1) The OFLI mapping and (c2)
the OFLIsat mapping, (d1) the GALI3 mapping and (d2) the GALIsat 3 mapping (the final values of
the indicators are in logarithmic scale)

saturates (i.e. the indicator reaches its saturation value), the integration times will be
the times of saturation, but, if the indicator does not saturate, the integration times
will be the final integration times.
On the right panels of Figs. 6.11 and 6.12 we present gray-scale plots of the
integration times used for three of the four indicators above mentioned. On panel
(b2) in Fig. 6.11 we present the integration times for the MEGNO or MEGNOsat
and in Fig. 6.12: panels (c2) and (d2), the integration times for the OFLI and the
GALI3 , or OFLIsat and GALIsat 3 , respectively.
The discussion below is not intended to analyze the dynamics of the system,
which is the aim of a future work. Our goal here is to demonstrate that the
performance of the RLI in this complex system is as good as the performances
of the other three widespread indicators.
On the left panels of Figs. 6.11 and 6.12, we can clearly see the great level of
coincidence among the phase space portraits of the four indicators. The regular
component is composed of symmetrical structures around the Lz axis and the four
indicators represent these structures with very similar shapes, sizes and shades of
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 203

gray. This shows that the indicators do not only agree in the location, extension and
shape of the domains of regular motion, but also in the description of these domains.
The chaotic domain is described equivalently by the four indicators and their
corresponding times of saturation. For instance, the RLI (Fig. 6.11a) shows that
the chaoticity (light gray) is inversely proportional to Enorm and does not depend
on the Lz . That is, more bound orbits are more chaotic orbits. We arrive at the
same (trivial) conclusion with the information given by the integration times of
the other three indicators (Figs. 6.11b2 and 6.12c2, d2). The MEGNO (Fig. 6.11b1)
shows an uniform and almost white color for the chaotic domain. This implies
that the saturation value (30) has been reached by these chaotic orbits and thus,
no further structure is revealed. However, the MEGNOsat (Fig. 6.11b2) shows such
structure (surrounded by orbits that did not saturate).4 This structure revealed by
the MEGNOsat shows that the times of saturation are directly proportional to
Enorm or, once again, that chaoticity is inversely proportional to Enorm and also
does not depend on the Lz . Similar conclusions can be drawn from Fig. 6.12 for
the OFLI (panel c1) and the OFLIsat (panel c2) and the GALI3 (panel d1) and
GALIsat3 (panel d2). However, the region composed of orbits that have reached
the associated saturation values (1016 and 1016 for the OFLI and the GALI3 ,
respectively) within the interval of integration is now much extended. On the one
hand, this region in Fig. 6.12c1,d1 is depicted in an uniform and almost white color
and thus, the structure cannot be revealed. On the other hand, the times of saturation
in Fig. 6.12c2,d2 fulfill the missing information.
In the next experiment, we follow two orbits in the NFW model for a timespan
of 1000 Gyrs (i.e. 77 Hubble times) in order to have convergent final values of
all the indicators in the study. We compute the time evolution curves of several
indicators, including the RLI, and present the results for a chaotic and a regular
orbit (cha and reg, respectively) in Fig. 6.13. In order to distinguish efficiently
the chaotic orbit from the regular one, we proceed as in Sect. 6.2.2 and use the same
thresholds used there for the MEGNO, the OFLI, the LI, the SALI and the RLI.
The threshold for the GALI3 in the NFW model will be the same constant used for
the SALI (see [8]) because the model is a 3 degree of freedom (hereinafter d.o.f.)
system. The threshold for the GALI5 will be t4 , with t the time (see [13] for further
details). In Fig. 6.13, we mark with the vertical line I the time after which the orbit
cha is clearly identified as a chaotic orbit.
Figure 6.13 shows that the accurate identification of the orbit cha as a chaotic
orbit by the MEGNO (panel a), the OFLI (panel b), the GALI5 (panel f) and the RLI
(panel g) is made within a Hubble time (13 Gyrs). The above mentioned indicators
show that the orbit will behave as a chaotic orbit within a physical meaningful time
span (i.e. the age of the Universe) which is important to understand the dynamics of
a real system like a galaxy.

4
Remember that the integration time varies with the Enorm , which explains the transition from dark
to light gray in the background of the plots on the left side of Figs. 6.11 and 6.12. Also the time of
integration is fixed to 0 where there are not initial conditions.
204 Z. Sndor and N. Maffione

CIs vs RLI (NFW model)

(a) (d)
30 1
25 0.01
0.0001
20 I
MEGNO

1e-06

SALI
15 1e-08
10 1e-10 I
1e-12
5 1e-14
2
0 1e-16
3 7 10 100 1000 3 10 37 100 1000
(b) (e)
1e+16 1
1e+14 0.01
1e+12 I 0.0001
1e+10 1e-06

GALI3
OFLI

1e+08 1e-08
1e+06 1e-10
10000 1e-12 I
100 1e-14
1 1e-16
3 4.5 10 100 1000 3 10 25 100 1000
(c) (f)
10 1
0.01
1 0.0001
GALI5 1e-06
LI

0.1 1e-08
1e-10 I
0.01 I 1e-12
1e-14
0.001 1e-16
3 10 25 100 1000 3 4.5 10 100 200 1000
(g)
1

cha 0.01
reg
Thresholds 0.0001 I
RLI

1e-06

1e-08

1e-10
3 7.5 20 100 1000

Time [u.t.]

Fig. 6.13 Behaviours of (a) the MEGNO, (b) the OFLI, (c) the LI, (d) the SALI, (e) the GALI3 ,
(f) the GALI5 and (g) the RLI for orbits cha and reg. The thresholds as well as the line I are
included (see text for further details)

In this section, we present results that support the RLI as a reliable indicator. This
technique shows a phase space portrait very similar to those shown by the MEGNO,
the OFLI or the GALI3 . Furthermore, the RLI identified the chaotic nature of the
orbit cha very fast in the second experiment (the fastest being the OFLI and the
GALI5 ), within a Hubble time. These results put the RLI on an equal footing with
the other fast variational indicators.

6.2.4 A Short Discussion on the Computing Times

The computing times of the indicators are specially crucial for time-consuming
simulations and their estimation helps for an efficient usage of the computational
resources. However, such estimation is not an easy task. The computing times
depend on a wide variety of factors such as the complexity of the model and
that of the indicators algorithm, its numerical implementation, the hardware, etc.
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 205

Although making a detailed study of the computing times is not our concern (see
[13] for further information on the subject), we would like to point out the fact that
the easy algorithm of the RLI helps for a rather fast computation of the indicator.
Furthermore, we are not dealing with the computing times themselves. We are going
to register the ratios between the computing times of the different techniques and
the computing time of the LIs5 for orbits in two of the systems previously studied in
the chapter, the HH and the NFW models. If the ratio is above 1, then the computing
time of the indicator is larger than the computing time of the LIs.
In this experiment, we used the following hardware/software configuration: an
Intel Core i5 with four cores, CPU at 2.67 GHz, 3 GB of RAM, an OS of 32 bits, and
the gfortran compiler of gcc version 4.4.4, without any optimizations. The code
to compute the indicators is the LP-VIcode, the acronym for La Plata Variational
Indicators code. The alpha version of the LP-VIcode was introduced in [14]
and currently, it is a fully operational code that computes a suite of ten variational
indicators (see [6]).6 The initial setup of the LP-VIcode is the following: the
integration time is 1000 u.t. (or 1000 Gyrs in the NFW model), the step of integration
is 0.001 u.t. (or 1 Myr in the NFW model).
We registered the ratios for two pairs of orbits and for every indicator. Both pairs
have one chaotic and one regular orbit. One of the pairs of orbits is located in the HH
model and the other in the NFW model. The computing time of the LIs for the HH
model (i.e. 4 LIs) is 0m15.204s., and for the NFW model (i.e. 6 LIs) is 3m27.703s.
The energy conservation is E  1013 .
The results are shown in Table 6.3. The value N is the number of d.o.f. of the
system. For instance, in the HH model (second column in Table 6.3), we compute 4
LIs and 3 GALIs: the GALI2 , the GALI3 and the GALI4 .
The RLI is not the least consuming indicator so far, according to the information
given in Table 6.3. The FLI/OFLI, the MEGNO, the SALI and the SD might be
more desirable options (in that order). Nevertheless, the easy implementation of the

Table 6.3 Ratios of the Indicator(s) Ratios (HH model) Ratios (NFW model)
computing times for several
indicators, including the RLI, (2N-1) GALIs 2.026 1.024
for the HH and the NFW (2N) LIs 1 1
models RLI 0.892 0.443
SD 0.613 0.375
SALI 0.582 0.372
MEGNO 0.53 0.174
FLI/OFLI 0.432 0.151

5
Here, the LIs are the numerical approximations of the spectra of Lyapunov Characteristic
Exponents.
6
Further information on the LP-VIcode can be found at the following url: http://www.fcaglp.
unlp.edu.ar/LP-VIcode/.
206 Z. Sndor and N. Maffione

RLI helps to reduce significantly the computing time of the LIs, and its ratio is not
much larger than those of the SD or the SALI.

6.3 Application of the RLI to Planetary Systems

The ongoing discovery of exoplanetary systems is certainly the most rapidly grow-
ing field of astronomy. Up to now the number of known planetary systems is almost
200. On the other hand, the masses and orbital elements of the planets detected can
be determined only with uncertainties. Therefore it is of high importance to provide
stability estimates of these systems. Before the launch of the space missions devoted
to detect terrestrial planets (CoRoT, Kepler), only the massive giant planets were
discovered mainly by the radial velocity method. One of the major applications
of the RLI was to study the stability of the still hypothetical terrestrial planets
in planetary systems containing at least one giant planet [38]. Another possible
application area of the RLI is to study the stability of different orbital solutions
of resonant exoplanetary systems provided by radial velocity observations. In this
review we shortly describe the stability studies done for the resonant planetary
system HD 73526 [39]. Finally, we present the applicability of the RLI to map
the high order resonances in the restricted three-body problem, which might have
relevance when studying the behavior of Kuiper belt objects [15]. We note that close
relatives to our investigations are the works of [50] computing the stability maps of
the system 55 Cancri by using various indicators (LI, SALI and FLI), and of [7]
studying the dynamical stability of the Kuiper-Belt using the LI indicator.
In all of the above problems the mean motion resonances (MMR) play an
essential role, therefore we shortly summarize their properties. A MMR occurs
between two bodies orbiting a more massive body if their orbital periods can
approximately be expressed as a ratio of two positive integer numbers, T1 =T2 D
. p C q/=p, where T1 and T2 denotes the orbital periods of the two bodies,
respectively. A MMR can be characterized by studying the behaviour of the resonant
angle, which in the model of the restricted three-body problem for an inner MMR is

 D . p C q/0  p  q$; (6.5)

while for an outer MMR can be written as

 D . p C q/  p0  q$; (6.6)

where  and $ are the mean orbital longitude, and longitude of the pericenter of
the massless body, while 0 is the mean orbital longitude of the massive body. If 
librates with a certain amplitude, the two bodies are engulfed in the . p C q/ W p
MMR. In this way almost the same orbital configuration of the bodies involved in
the given MMR is repeated. Depending on the relative positions of the bodies, this
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 207

configuration can be protective, or can result in unstable orbits, see for more details
in [32].

6.3.1 Stability Catalogue of the Habitable Zones

of Exoplanetary Systems

The main idea behind the stability catalogue was to map the regions of a planetary
system that can host dynamically stable terrestrial planets [38]. The dynamical
stability of a terrestrial planet is one of the strongest requirements for a habitable
planetary climate. The most important requirement for the habitability of a planet is
to contain water in liquid phase on its surface. A region around a star, in which an
Earth-mass planet could be habitable in the above sense is called as the habitable
zone (HZ), see [23] and [25] for more details.

6.3.1.1 Used Models and Initial Conditions

The stability of terrestrial planets can be studied by using different approaches:

(i) by detecting the stable and unstable regions of the parameter space of each
exoplanetary system separately or (ii) by using stability maps computed in advance
for a large set of orbital parameters. In the stability catalogue we presented such
stability maps also showing how to apply them to the exoplanetary systems under
study. This second approach has the advantage that the stability properties of a
terrestrial planet can be easily reconsidered when the orbital parameters of the
giant planet of an exoplanetary system are modified. This is very often the case,
since the orbital parameters of the giant planets are quite uncertain, and due to the
accumulation and improvement of the observational data, they are subject to change
quite frequently. Instead of the re-exploration of the phase space of each individual
exoplanetary system after possible modification of the orbital parameters of the
giant planet, the stability properties of the investigated planetary system can be
easily re-established from the already existing stability maps. These stability maps,
which form a stability catalogue, can also be used to study the stability properties
of the habitable zones of known exoplanetary systems.
The majority of planetary systems which are detected so far consists of a star and
a giant planet revolving in an eccentric orbit. Therefore, we used a simple dynamical
model, the elliptic restricted three-body problem in which there are two massive
bodies (the primaries) moving in elliptic orbits about their common center of mass,
and a third body of negligible mass moving under their gravitational influence (for
details see [45]). In our particular case the primaries are the star and the giant planet,
and the third body is a small Earth-like planet, being regarded as massless. We
note that among the extrasolar planetary systems there is a high rate of multiple
planet systems. Thus, a more convenient model for the stability maps would be
208 Z. Sndor and N. Maffione

the restricted N-body problem (with N-2 giant planets, N  4). The presence
of additional giant planets certainly enhance the instabilities induced by just one
massive planet, turning the HZ of the system more unstable. The main source of
instabilities are the mean motion resonances (MMR) between the massive giant
planet and the Earth-like planet. Thus, by mapping these resonances, we can find
the possible regions of the instabilities in the HZs. On the other hand, the dynamical
model with one giant planet also offers the most convenient way to display the most
important MMRs as a function of the mass ratio of the star and the giant planet, and
of the eccentricity of the giant planet.
In the catalogue of dynamical stability one important quantity is the mass
parameter of the problem  D m1 =.m0 C m1 /, where m0 is the stellar, and m1 is
the planetary mass. The mass parameter has been changed between broad limits
(104  102 ) with various steps of , in total the different stability maps have
been calculated for 23 values of . The giant planet was placed around the star in
an elliptic orbit, with semi-major axis a1 , eccentricity e1 , argument of periastron !1
and mean anomaly M1 . The semimajor axis a1 was taken as unit distance a1 D 1
during all simulations. The eccentricity e1 was changed between 0.0 and 0.5 with a
stepsize of 5  103 . The argument of periastron was fixed at !1 D 0 , while the
mean anomaly M1 was changed between 0 and 360 with M1 D 45 . The test
planet was started in the orbital plane of the giant planet with an initial eccentricity
e D 0, argument of periastron ! D 0 , and mean anomaly M D 0 . The semi-major
axis a of the test planet was changed in two different intervals: (i) for orbits of the
test planet inside the orbit of the giant planet between 0.1 and 0.9 with a stepsize
of a D 103 and (ii) for outside orbits between 1.1 and 4.0 with a stepsize of
a D 3:625  103 . Further details and the complete catalogue can be found at
http://astro.elte.hu/exocatalogue/index.html.

6.3.1.2 Stability Maps

Due to the very good visibility of the outer MMRs, we first display the case when
the semi-major axis of the test planet is larger than the giant planets semi-major
axis (a > 1). Additionally, in the stability map shown in Fig. 6.14 the values
 D 0:001 and M D 0 were kept fixed. For each value of the RLI a gray shade
has been assigned. White regions correspond to small RLI values, thus they are
very stable. The MMRs appear either as light strips in the dark, strongly chaotic
regions or as the well-known V-shaped structures representing the separatrices
between resonant and non-resonant motion. The inner regions of the resonances
may be lighter than the lines of the bounding separatrices indicating regular motion
in a protective resonance (e.g. the 2:5 MMR). Near the separatrices the motion is
always chaotic, moreover at some MMRs even the inner part of the resonance is
chaotic as indicated in the case of the 1:3 MMR, for instance. By increasing the giant
planets eccentricity e1 many resonances overlap giving rise to strongly chaotic and
thus very unstable behaviour. The reason of this phenomenon is that by increasing
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 209

2:3 1:2 2:5 1:3 2:7 1:4 2:9 1:5 1:6 1:7

Fig. 6.14 Stability map of the outer MMRs for the Earth-like planets in the elliptic restricted
three-body problem for  D 0:001 and M D 0 . White colour denotes ordered motion, light grey
strips and V shapes the different resonances, while black the strongly chaotic regions

e1 the apocenter distance of the giant planet also increases, therefore the giant planet
perturbs more strongly the outer test planet.
In stability maps displayed in Fig. 6.15 the semi-major axis of the test planet is
smaller than that of the giant planet (a < 1). The two panels for the mass parameter
 D 0:001 show the stability maps for the two different starting positions of the
giant planet, M1 D 0 (upper panel) and M1 D 180 (lower panel), respectively.
Between the test planet and the giant planet, a large number of inner MMRs can
be found, which dominate the stability maps. Inside the resonances the stable or
chaotic behaviour of the test planet depends on the initial angular positions of the
two planets. This is clearly visible by comparing the two panels. On the other hand
the location of the MMRs is not altered, since this depends on the ratio of the
semi-major axes of the two planets. In the lower panel of Fig. 6.15 several MMRs
(5:2, 5:3, 3:2) are stable, which is not the case in the upper panel of Fig. 6.15.
This is due to the fact that the relative initial positions determine the places of
conjunctions of the two planets. If they meet regularly near the pericenter of the
giant planet, the motion of the test planet becomes chaotic, while it can remain
regular if the conjunctions take place near the apocenter of the giant planet. The
effect of the initial phase difference between the planets is important, therefore a
bunch of stability maps have been prepared for more initial values of the mean
anomaly M1 of the giant planet. These stability maps can be found in the online
exocatalogue (http://astro.elte.hu/exocatalogue/index.html). By increasing the value
of the mass parameter  it becomes clearly visible that the larger mass of the giant
planet results in stronger perturbations, and therefore more enhanced chaotic region.
210 Z. Sndor and N. Maffione

Fig. 6.15 Stability map of the inner MMRs for the Earth-like planets in the elliptic restricted three-
body problem for  D 0:001, M D 0 (upper panel), and M D 180 (lower panel), respectively.
We note the different character of the 5:2, 7:3, and 3:2 MMRs depending on the orbital positions
between the test particle, and the perturbing body. White colour denotes ordered motion, light
grey strips and V shapes the different resonances, while black the strongly chaotic regions. We
displayed the properly scaled HZs of three exoplanetary systems in the stability maps. Studying
the figures one can conclude that the HZ of the Solar system is stable, the HZs of Eridani and
HD 114729 are marginally stable meaning that they are filled with several MMRs

6.3.1.3 Stability of Terrestrial Planets in the Habitable Zones

In this section, we show how to use the catalogue to determine the stability of
hypothetical Earth-like planets in exoplanetary systems. As an example, we consider
the case of HD10697, where a1 D 2:13 AU, e1 D 0:11 and  D 0:0055. Figure 6.16
shows a stability map, calculated for  D 0:005 for inner orbits of the test planet.
This corresponds to the minimum mass of the giant planet (minimum masses are
used throughout). The stability of a small planet (starting with e D 0) in the
system HD10697 can be studied along the line e1 D 0:11. One can see that for
small semimajor axes, a < 0:33a1 D 0:729 AU the parameter space is very stable.
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 211

9:1 7:1 7:3

8:1 6:1 5:1 4:1 3:1 5:2 2:1

HD 10697

Fig. 6.16 Stability map for inner orbits of an Earth-like planet, when  D 0:005 and M1 D 0 .
The line e1 D 0:11, corresponds to the system HD 10697. The scaled HZ of this system is between
0:39 < a < 0:77. Its inner part is stable (containing only a few weakly chaotic MMRs), while the
outer part is in the strongly chaotic regions

When a > 0:33a1 several resonances appear, among which the most important
are the 5:1, 4:1, 3:1 and 2:1 MMRs. For a > 0:73a1 D 1:55 AU, a strongly
chaotic region appears. The classical HZ of this system is between 0.85 and 1.65 AU
therefore in Fig. 6.16 the scaled classical HZ is located between 0:85=a1 D 0:39
and 1:65=a1 D 0:77 (shown as a rectangle, elongated in horizontal direction). One
can see that the inner part of the classical HZ contains ordered regions, but stripes
of certain resonances are also present. The outer part of the classical HZ is in the
strongly chaotic region.

6.3.2 Stability of Resonant Exoplanetary Systems

A significant amount of multiple extrasolar planetary systems contain pairs of giant

planets which are orbiting in MMRs. The in situ formation of resonant planetary
systems is very unlikely, since each resonance requires certain ratio of the semi-
major axes. The more favorable scenario is the type II migration of giant planets
being embedded in the still gas rich protoplanetary disc. Type II migration appears
when a massive giant planet carves a gap in the gaseous protoplanetary disc
practically inhibiting the gas flow through the gap. In that case the planets semi-
major axis is changing according to the viscous evolution of the protoplanetary disc,
see more about the topic in [3].
If the migration of two giant planets is convergent (e.g. the difference between
their semi-major axes is decreasing) the phenomenon of the resonant capture will
occur between them, and the two planets can migrate very close to their host
star. The efficiency of migration is excellently demonstrated by hydrodynamic
212 Z. Sndor and N. Maffione

simulations modeling the formation of the resonant system around the star GJ 876
[24]. There are other planetary systems in which the giant planets reached the 2:1
MMR through type II migration such as HD 128311 [35], and HD 73526 [39].
The detection of giant planets is based on the radial velocity method. To calculate
the orbital elements of the planets of a multiple system is not an easy task. Although
there are well-known and widely used algorithms to provide reliable orbital fits,
the orbital elements obtained not always result in a stable configuration for the
planetary system. Regarding the resonant system of giant planets around HD 73526,
[47] published orbital elements and planetary/stellar masses which resulted in stable
orbits of the giant planets over 1 million years. On the other hand, the orbits
are chaotic, as was clearly visible from numerical integrations of the three-body
problem using the given elements as initial conditions. Since chaotic behavior
may be uncommon among the resonant extrasolar planetary systems, and may not
guarantee the stability of the giant planets for the whole lifetime of the system (being
certainly longer than 1 million years), we searched for regular orbital solutions for
the giant planets as well [39]. As a first attempt to study the degree of the chaoticity,
we mapped the parameter space around the solution of [47]. We have calculated the
stability properties of the a1  a2 , e1  e2 , M1  M2 , and $1  $2 parameter planes,
where a is the semi-major axis, e is the eccentricity, M is the mean anomaly and $
is the longitude of periastron of one of the giant planets.
In Fig. 6.17 the stability structures of the parameter planes for the semi-major
axis and the eccentricities are displayed. During the calculation of a particular
parameter plane the other orbital data have been kept fixed to their original values.
On each parameter plane the stable regions are displayed by white, the weakly
chaotic regions by grey, and the strongly chaotic regions by black colors. The values
of the corresponding orbital data are marked on each parameter plane. By studying
the stability maps, one can see that the orbital elements given by Tinney et al. [47]

Fig. 6.17 Stability maps calculated around the orbital elements of [47]. The structure of the
stability maps indicates that the orbital elements (marked by +) are in a weakly chaotic region.
Here white colour refers to ordered, lighter grey shades to weakly chaotic, and darker shades to
unstable regions
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 213

Fig. 6.18 Stability maps calculated around one set of orbital elements given in [39] It can be seen
from the stability maps that the orbital elements (marked by +) are embedded well in the stable
region marked by white colour

are located in a weakly chaotic region, which explains the irregular behavior of the
planetary eccentricities. We again stress that this does not automatically imply the
instability of their fit, however by using these orbital data the system yields chaotic
behavior and can be destabilized in longer timescales. Studying the stability maps it
can also be concluded that the fit cannot be easily improved by the simple change of
one of the orbital elements. The parameter plane is almost entirely weakly chaotic,
there exists only a narrow strip of ordered behavior. After obtaining completely new
sets of orbital elements using the Systemic Console [30] the stability maps around
these fits were recalculated. In Fig. 6.18 parameter planes for the semi-major axis
and the eccentricities are displayed, now around one of the stable orbital solutions.
It can be clearly seen that the new orbital solution is well inside the region for
ordered motion, which provides stability for the whole lifetime of the system. Based
on the above example, it can be concluded that the RLI performs well in stability
investigations of extrasolar planetary systems.

6.3.3 Application of the RLI to Study Libration Inside High

Order MMRs

The most recent application of the RLI in dynamical astronomy has been presented
by rdi et al. [15] investigating the dynamical structure of high order resonances in
the elliptic restricted three-body problem. This study has relevance in dynamics of
Kuiper-belt objects. Moreover, it proved through numerical integration of a large set
of orbits that the RLI excellently indicates the libration of the resonant angle inside
a MMR. In what follows, first we present the results of rdi et al. [15] obtained for
214 Z. Sndor and N. Maffione

Fig. 6.19 The 8:5 inner MMR. Left panel: the two regions of libration as displayed by the libration
amplitude of the resonant variable. The exact location of the MMR is at ar D 0:7310 in normalized
units. Right panel: the RLI map around the resonance, in which the two regions of libration can
be clearly seen. The continuous curve denotes the unit apocentre distance, while the dotted and
dashed curves the unit apocentre distance decreased by 1, and 1.5 Hills radius, respectively

the 3rd order 8 W 5 inner resonance, in which case according to Eq. (6.5) the resonant
variable is

 D 80  5  3$; (6.7)

where  and $ are the mean orbital longitude, and longitude of the pericenter of
the inner body, while 0 is the mean orbital longitude of the outer body.
A simple, but computationally demanding way to map the neighbourhood a
MMR on the a  e plane is to calculate the libration amplitude of  for orbits whose
initial semi-major axis and eccentricity values are taken from a grid, and the test
particle has been started from pericentre. Such calculation can be seen on the left
panel of Fig. 6.19. Using the same grid resolution and initial conditions the RLI
values were also calculated, see the right panel of Fig. 6.19. The exact location of
the resonance (when  D 0 ) is marked with the vertical dashed line in the left panel
of Fig. 6.19. The colour code corresponds to the value of the librations amplitude:
the darker the colour, the smaller the amplitude.
The 8 W 5 MMR has an interesting structure, there is a libration for low values of
the eccentricities e < 0:2, and also for very high values 0:75 < e < 0:85. Although
the picture obtained by the RLI is more detailed, also showing some other neighbour
MMRs, gives back the region of libration very well.
The RLI has also been applied to study the high order outer MMRs, in which
case the massless bodys orbit is outside the massive planets orbit. In the study of
rdi et al. [15] the outer resonances of the Sun-Neptune system have been studied in
two different models. These are the restricted three-body problem, in which only the
gravitational effects of the Sun and Neptune were included, and also in a model in
which the four giant planets were also taken into account. We will present here the
results obtained for two different MMRs, namely the 8:5 and 7:3 outer resonances
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 215

Fig. 6.20 The 8:5 outer MMR. Left panel: the two regions of libration displayed by the libration
amplitude of the resonant variable. The exact location of the MMR is at ar D 1:3680 in normalized
units. Right panel: the RLI map around the resonance, in which the regions of libration are clearly
visible. The continuous curve indicates the unit pericentre distance, while the dotted curve the
unit pericentre distance increased by one Hills radius. The dash-dotted curve shows the place of
Uranus-crossing. The filled circles mark circulating TNOs

in the model of the restricted three-body problem. We note that the resonant variable
of an outer . p C q/ W p MMR is given by Eq. (6.6).
Similarly to the 8 W 5 inner MMR, the neighbourhood of the exact 8 W 5 outer
MMR (ar D 1:3680 in normalized units) has been mapped on a dense grid of the
a  e plane (the test particle has been started from apocentre) and the libration
regions are marked with different shades corresponding to the amplitude of . There
are two regions of libration in this resonance located at relatively high eccentricities
(0:35 < e < 0:7, 0:75 < e), see the left panel of Fig. 6.20. Using the same a  e
grid the RLI values have also been calculated, see the right panel of Fig. 6.20. At
the location of the exact resonance there is a nearly vertical dark strip indicating
weak chaotic behaviour, and also the fact that in this configuration the 8 W 5 MMR
is not protective for low values of the test particles eccentricity. The origin of the
strong chaotic behaviour is due to the crossing of the orbits of the test particle and
the perturbing body (Neptune in this case). There are different lines plotted to the
RLI map. The continuous curve is the pericentre distance, the dotted curve is the
pericentre distance increased by the Hills radius of Neptune, and finally, the dash-
dotted curve is the place of Uranus crossing. This latter curve may indicate that
the high eccentricity libration regions of Fig. 6.20 being present in the model of the
restricted three-body problem, might vanish in more advanced models including the
planet Uranus, for instance. Finally, we roughly compare the 8 W 5 outer resonance
with the behaviour of some of the existing TNOs having inclination i < 25 , see the
black filled circles for the corresponding a and e values of these objects. All of the
TNOs displayed are circulating, and really they occupy the low eccentricity regions
of the a  e plane. We note, however, that the a and e values of the TNOs have
different epochs, so their positions does not reflect the actual state of the system. On
216 Z. Sndor and N. Maffione

Fig. 6.21 The 7:4 MMR. Left panel: the two regions of libration displayed by the libration of the
resonant variable. The exact location of the resonance is at ar D 1:4522. Right panel: the RLI map
around the resonance. The triangles mark librating, while the filled circles circulating TNOs. For
the description of the curves see the previous figures

the other hand, it can be clearly seen in Fig. 6.20 that the TNOs at the 8 W 5 MMR
have circulating resonance variable.
In order to have a more complete picture of the high order outer resonances,
we also summarize the case when a MMR has a protective character, and the
resonant variable of bodies lying in its vicinity can both librate and circulate. A
good candidate for this purpose is the 7 W 4 3rd order outer MMR. The exact
resonance is at ar D 1:4522 (normalized units). Similarly to the 8 W 5 outer MMR,
this resonance also has two regions of libration, but in this case the lower region
of libration allows libration of test particles having low eccentricities, see Fig. 6.21.
The right panel shows the dynamical structure of the resonance, and also TNOs
found at this resonance. Most of them have circulating , but there are a few of
them in the librating region, too. Studying the right panel of Fig. 6.21, one can see
that the librating TNOs are clearly below the Neptune-crossing line, while the region
of libration extends above it.
As an overall conclusion we can state that the RLI is a very reliable tool in
detecting the positions of the lower and higher order MMRs being present in various
planetary systems.

6.4 Discussion and Summary

In this chapter we summarize the basic properties of the recently introduced

chaos detection method, the RLI. The RLI is based on the time evolution of the
infinitesimally small tangent vector to the orbit, which is provided by solving
numerically the variational equations. Hence, the RLI belongs to the family of
the socalled variational indicators. Although the definition of the RLI is based on
that of the LI, in this review we give evidence that the distinction between regular
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 217

and chaotic motion is much clearer with the RLI, which makes it a more reliable
alternative than the LI.
According to the comparative study with some widespread variational indica-
tors, the RLI shows convincing performances in the experiments and considerably
improves the performances of the classical LI. In generality, indicators like the
FLI/OFLI or the MEGNO (actually there is a strong relationship between both
indicators, see [31] for further details) are usually believed to be better options for a
general analysis of the structure of the phase space. Therefore, our study reinforces
the fact that the RLI can also be used as an alternative technique, which operates
with reasonable computing times to make conclusive pictures of the dynamics
despite the complexity of the problem.
Based on the comparative work presented in Sect. 6.2, we can summarize both
the advantages and disadvantages of the RLI. In what follows we list its favorable
properties/advantages, and also add that in which dynamical systems have been done
the corresponding simulations.
Having determined a reliable threshold in the HnonHeiles model, the RLI (and
the FLI/OFLI) shows the best approximation rates in the ordered/chaotic regions.
We note that the methods SALI and GALI also estimate the true percentage of
the ordered/chaotic orbits but with a slight slower way.
Comparing to the other chaos indicators also in the HnonHeiles model, the RLI
detects much faster the orbits from the large chaotic sea (e.g. the c-cs orbit),
than the other indicators.
Studying a 4D symplectic mapping the RLI have been compared to the indicator
MEGNO(2,0). In this case both the RLI and the MEGNO(2,0) reveal the fine
structure of the phase space very accurately (much better than the LI, for
instance). As a result of this experiment we also conclude that the RLI is a very
effective tool in the characterization of a large array of initial conditions.
In a complex 3D potential resembling a triaxial galactic halo (the so called NFW
model), the RLI (together with the MEGNO, the OFLI, and the GALI5 ) identifies
the chaotic orbits within a Hubble time (13 Gyrs). These indicators show that
chaotic orbits can be identified within a physically meaningful time (i.e. the age
of the Universe), which is important when studying the dynamics of a galaxy.
On the other hand, when applying the RLI one should be aware that:
Among the presented CIs, in the HnonHeiles model the RLI identifies the so
called c-sl orbit in the slowest way.
In the study performed in the 4D symplectic mapping, the RLI cannot really
distinguish between chaotic and sticky orbits. This is a disadvantage if one is
interested in detecting the sticky orbits. On the other hand, if we are interested
in detecting all chaotic orbits (including sticky orbits, as well), the application of
the RLI might be useful.
Regarding the time needed to calculate the RLI, we can conclude that it is not
the least time consuming indicator. The FLI/OFLI, the MEGNO, the SALI and
the SD might be more desirable options if the computation time really matters.
218 Z. Sndor and N. Maffione

We note, however, that with the current generation of fast computers this option
became less important.
In the last section of the current work we summarize the application of the RLI
to planetary systems, which is its major application area. These studies include
the development of a stability catalogue of hypothetical terrestrial exoplanets in
extrasolar planetary systems, stability studies of resonant planetary systems and the
investigation of high order mean motion resonances having relevance in studying
the dynamics of the Kuiper-belt objects. We find that the RLI is an efficient and
reliable numerical tool to map and characterize the dynamical structure of various
mean motion resonances.
We note that the RLI (together with the SALI) has also been applied to map the
stability regions of the Caledonian symmetric four-body problem, [46]. Since the
preceding studies are mainly related to detecting the chaotic behaviour restricted
four-body problem does not really belong to this line, thus to keep the length of the
present study tractable, we omitted its presentation here.
We would also like to remark that the very simple computation of the RLI from
the widespread well-known LI and its better performances reported in very different
scenarios, make the RLI a serious candidate to replace the LI in a variety of fields,
and not only in dynamical astronomy. For instance, in a paper published in a journal
of Chemical Physics, the RLI is used as the default chaos indicator in the Lyapunov
weighted path ensemble method. One of the capabilities of the method is to identify
pathways connecting stable states which are relevant in the context of activated
chemical reactions (see [18]).
Finally, the reliability of the RLI as a chaos indicator has been strongly
demonstrated throughout this study, and as a result, the choice of the RLI to analyze
a general dynamical system is well-founded.

Acknowledgements The authors thank the invitation and support of the scientific coordinators
of the international workshop on Methods of Chaos Detection and Predictability: Theory and
Applications: Georg Gottwald, Jacques Laskar and Haris Skokos and the hospitality of the Max
Planck Institute for the Physics of Complex Systems where the meeting took place. ZsS is
supported by the Jnos Bolyai Research Scholarship of the Hungarian Academy of Sciences. NM
is supported with grants from the Consejo Nacional de Investigaciones Cientficas y Tcnicas de la
Repblica Argentina (CCT - La Plata) and the Universidad Nacional de La Plata.

References

1. Barrio, R.: Sensitivity tools vs. Poincar sections. Chaos Solitons Fractals 25, 711726 (2005)
2. Barrio, R., Borczyk, W., Breiter, S.: Spurious structures in chaos indicators maps. Chaos
Solitons Fractals 40, 16971714 (2009)
3. Baruteau, C., Masset, F.: Recent development in Planet Migration Theory. In: Souchay, J. et al.
(eds.) Tides in Astronomy and Astrophysics. Lecture Notes in Physics, vol. 861, pp. 201253.
Springer, Berlin, Heidelberg (2013)
4. Belokurov, V. et al.: An orphan in the Field of Streams. Astrophys. J. 658, 337344 (2007)
6 The Relative Lyapunov Indicators: Theory and Application to Dynamical. . . 219

5. Benettin, G., Galgani L., Giorgilli, A., Strelcyn, J.: Lyapunov characteristic exponents for
smooth dynamical systems; a method for computing all of them. Meccanica 15, Part I: theory,
920; Part II: Numerical Applications, 2130 (1980)
6. Carpintero, D., Maffione, N., Darriba, L.: LPVIcode: a program to compute a suite of
variational chaos indicators. Astron. Comput. 5, 1927 (2014)
7. Celletti, A., Kotoulas, T., Voyatzis, G., Hadjidemetriou, J.: The dynamical stability of a Kuiper
Belt-like region. Mon. Not. R. Astron. Soc. 378, 11531164 (2007)
8. Christodoulidi, H., Bountis, T.: Low-dimensional quasiperiodic motion in Hamiltonian sys-
tems. ROMAI J. 2, 3744 (2006)
9. Cincotta, P., Sim, C.: Simple tools to study global dynamics in non-axisymmetric galactic
potentials I. Astron. Astrophys. 147, 205228 (2000)
10. Cincotta, P., Giordano, C., Sim, C.: Phase space structure of multidimensional systems by
means of the mean exponential growth factor of nearby orbits. Physica D 182, 151178 (2003)
11. Contopoulos, G., Voglis, N.: A fast method for distinguishing between ordered and chaotic
orbits. Astron. Astrophys. 317, 7381 (1997)
12. Cooper, A.P., et al.: Galactic stellar haloes in the CDM model. Mon. Not. R. Astron. Soc. 406,
744766 (2010)
13. Darriba, L.A., Maffione, N.P., Cincotta, P.M., Giordano, C.M.: Comparative study of varia-
tional chaos indicators and ODEs numerical integrators. Int. J. Bifurcat. Chaos 22, 1230033
(2012)
14. Darriba, L., Maffione, N., Cincotta, P., Giordano, C.: Chaos detection tools: the LPVicode
and its applications. In: Cincotta, P., Giordano, C., Efthymiopoulos, C. (eds.) Chaos, Diffusion
and Non-integrability in Hamiltonian Systems-Application to Astronomy, pp. 345366. Uni-
versidad Nacional de La Plata and Asociacin Argentina de Astronoma Publishers, La Plata
(2012)
15. rdi, B., Rajnai, R., Sndor, Z., Forgcs-Dajka, E.: Stability of higher order resonances in the
restricted three-body problem. Celest. Mech. Dyn. Astron. 113, 95112 (2012)
16. Fouchard, M., Lega, E., Froeschl, Ch., Froeschl, Cl.: On the relationship between fast
lyapunov indicator and periodic orbits for continuous flows. Celest. Mech. Dyn. Astron. 83,
205222 (2002)
17. Froeschl, Cl., Gonczi, R., Lega, E.: The fast lyapunov indicator: a simple tool to detect weak
chaos. Application to the structure of the main asteroidal belt. Planet. Space Sci. 45, 881886
(1997)
18. Geiger, P., Dellago, C.: Identifying rare chaotic and regular trajectories in dynamical systems
with Lyapunov weighted path sampling. Chem. Phys. 375, 309315 (2010)
19. Gmez, F., Helmi, A., Cooper A., Frenk, C., Navarro, J., White, S.: Streams in the Aquarius
stellar haloes. Mon. Not. R. Astron. Soc. 436, 36023613 (2013)
20. Guzzo, M., Lega, E., Froeschle, C.: On the numerical detection of the effective stability of
chaotic motions in quasi-integrable systems. Physica D 163, 125 (2002)
21. Hnon, M., Heiles, C.: The applicability of the third integral of motion: some numerical
experiments. Astrophys. J. 69, 7379 (1964)
22. Ibata, R.A., Irwin, M.J., Lewis, G.F., Stolte, A.: Galactic halo substructure in the sloan digital
sky survey: the ancient tidal stream from the Sagittarius Dwarf Galaxy. Astrophys. J. 547,
133136 (2001)
23. Kasting, J.F., Whitmire, D.P., Reynolds, R.T.: Habitable zones around main sequence stars.
Icarus 101, 108128 (1993)
24. Kley, W., Lee, M.H., Murray, N., Peale, S.J.: Modeling the resonant planetary system GJ 876.
Astron. Astrophys. 437, 727742 (2005)
25. Kopparapu, R.K., Ramirez, R., Kasting, J.F., et al.: Habitable zones around main-sequence
stars: New estimates. Astrophys. J. 765, (2013), article id. 131, 16 pp.
26. Laskar, J.: Frequency analysis of a dynamical system. Celest. Mech. Dyn. Astron. 56, 191196
(1993)
27. Maffione, N., Darriba, L., Cincotta, P., Giordano, C.: A comparison of different indicators of
chaos based on the deviation vectors: application to symplectic mappings. Celest. Mech. Dyn.
Astron. 111, 285307 (2011)
220 Z. Sndor and N. Maffione

28. Majewski, S.R., Skrutskie, M.F., Weinberg, M.D., Ostheimer, J.C.: A two micron all sky survey
view of the Sagittarius dwarf galaxy. I. Morphology of the Sagittarius core and tidal arms.
Astrophys. J. 599, 10821115 (2003)
29. Manos, T., Skokos, Ch., Antonopoulos, Ch.: Probing the local dynamics of periodic orbits by
the generalized alignment index (GALI) method. Int. J. Bifurcat. Chaos 22, 1250218 (2012)
30. Meschiari, S., Wolf, A.S., Rivera, E., Laughlin, G., Vogt, S., Butler, P.: Systemic: a testbed
for characterizing the detection of extrasolar planets. I. The systemic console package. Publ.
Astron. Soc. Pac. 121, 10161027 (2009)
31. Mestre, M., Cincotta, P., Giordano, C.: Analytical relation between two chaos indicators: FLI
and MEGNO. Mon. Not. R. Astron. Soc. Lett. 414, 100103 (2011)
32. Murray, C.D., Dermott, S.F.: Solar System Dynamics. Cambridge University Press, Cambridge
(1999)
33. Navarro, J., Frenk, C., White, S.: The structure of cold dark matter halos. Astrophys. J. 462,
563575 (1996)
34. Navarro, J., Frenk, C., White, S.: A universal density profile from hierarchical clustering.
Astrophys. J. 490, 493508 (1997)
35. Sndor, Z., Kley, W.: On the evolution of the resonant planetary system HD 128311. Astron.
Astrophys. 451, L31L34 (2006)
36. Sndor, Z., rdi, B., Efthymiopoulos, C.: The phase space structure around L4 in the restricted
three-body problem. Celest. Mech. Dyn. Astron. 78, 113123 (2000)
37. Sndor, Z., rdi, B., Szll, A., Funk, B.: The relative lyapunov indicator: an efficient method
of chaos detection. Celest. Mech. Dyn. Astron. 90, 127138 (2004)
38. Sndor, Z., Sli, ., rdi, B., Pilat-Lohinger, E., Dvorak, R.: A stability catalogue of the
habitable zones in extrasolar planetary systems. Mon. Not. R. Astron. Soc. 375, 14951502
(2007)
39. Sndor, Z., Kley, W., Klagyivik, P.: Stability and formation of the resonant system HD 73526.
Astron. Astrophys. 472, 981992 (2007)
40. Skokos, Ch.: Alignment indices: a new, simple method to for determining the ordered or
chaotic nature of orbits. J. Phys. A Math. Gen. 34, 1002910043 (2001)
41. Skokos, Ch.: The Lyapunov Characteristic Exponents and Their Computation. Lecture Notes
in Physics, vol. 790, pp. 63135. Springer, Berlin, Heidelberg (2010)
42. Skokos, Ch., Antonopoulos, Ch., Bountis, T., Vrahatis, M.: Detecting order and chaos in
Hamiltonian systems by the SALI method. J. Phys. A Math. Gen. 37, 62696284 (2004)
43. Skokos, Ch., Bountis, T., Antonopoulos, Ch.: Geometrical properties of local dynamics in
Hamiltonian systems: the Generalized Alignment Index (GALI) method. Physica D 231, 3054
(2007)
44. Skokos, Ch., Bountis, T., Antonopoulos, Ch.: Detecting chaos, determining the dimensions of
tori and predicting slow diffusion in Fermi-Pasta-Ulam lattices by the Generalized Alignment
Index method. Eur. Phys. J. Spec. Top. 165, 514 (2008)
45. Szebehely, V.: Theory of Orbits. The Restricted Problem of Three Bodies. Academic,
New York (1967)
46. Szll, A., rdi, B., Sndor, Z., Steves, B.: Chaotic and stable behaviour in the Caledonian
symmetric four-body problem. Mon. Not. R. Astron. Soc. 347, 380388 (2004)
47. Tinney, C.G., Butler, R.P., Marcy, G.W., et al.: The 2:1 resonant exoplanetary system orbiting
HD 73526. Astrophys. J. 647, 594599 (2006)
48. Vogelsberger, M., White, S., Helmi, A., Springel, V.: The fine-grained phase-space structure of
cold dark matter haloes. Mon. Not. R. Astron. Soc. 385, 236254 (2008)
49. Voglis, N., Contopoulos, G., Efthymiopoulos, C.: Detection of ordered and chaotic motion
using the dynamical spectra. Celest. Mech. Dyn. Astron. 73, 211220 (1999)
50. Voyatzis, G.: Chaos, order, and periodic orbits in 3:1 resonant planetary dynamics. Astrophys.
J. 675, 802816 (2008)
Chapter 7
The 0-1 Test for Chaos: A Review

Georg A. Gottwald and Ian Melbourne

Abstract We review here theoretical as well as practical aspects of the 0-1 test
for chaos for deterministic dynamical systems. The test is designed to distinguish
between regular, i.e. periodic or quasi-periodic, dynamics and chaotic dynamics.
It works directly with the time series and does not require any phase space
reconstruction. This makes the test suitable for the analysis of discrete maps,
ordinary differential equations, delay differential equations, partial differential
equations and real world time series. To illustrate the range of applicability
we apply the test to examples of discrete dynamics such as the logistic map,
PomeauManneville intermittency maps with both summable and nonsummable
autocorrelation functions, and the Hamiltonian standard map exhibiting weak chaos.
We also consider examples of continuous time dynamics such as the Lorenz-96
system and a driven and damped nonlinear Schrdinger equation. Finally, we show
the applicability of the 0-1 test for time series contaminated with noise as found in
real world applications.

7.1 Introduction

The 0-1 test for chaos was developed in a series of papers [19, 20, 22] to distinguish
between regular and chaotic dynamics in deterministic dynamical systems. Rather
than requiring phase space reconstruction which is necessary to apply standard
Lyapunov exponent methods to the analysis of discretely sampled data, the test
works directly with the time series and does not involve any preprocessing of the
data. The test requires only a minimal computational effort independent of the
dimension of the underlying dynamical system under investigation.

G.A. Gottwald ()

School of Mathematics & Statistics, The University of Sydney, Sydney, 2006 NSW, Australia
e-mail: [email protected]
I. Melbourne
Mathematics Institute, University of Warwick, Coventry CV4 7AL, UK
e-mail: [email protected]

Springer-Verlag Berlin Heidelberg 2016 221

Ch. Skokos et al. (eds.), Chaos Detection and Predictability, Lecture Notes
in Physics 915, DOI 10.1007/978-3-662-48410-4_7
222 G.A. Gottwald and I. Melbourne

The test has found applications in a wide range of fields. Besides general
studies of dissipative [12, 35, 67] and Hamiltonian [72] dynamical systems and
multi-agent systems [39], the test has found its way into as disparate areas as
engineering [42, 43, 55], electronics [65], finance and economics [28, 3638, 68
70], geophysical applications [7, 47, 48, 60], hydrology [32, 40], epidemology
[8, 9, 50] and traffic dynamics [34]. In particular its application to non-smooth
processes [2, 42, 43], to systems with fractional derivatives and delays [3, 5, 71],
and to nonchaotic strange attractors [18] are notable as those are not amenable to
standard methods employing Lyapunov exponents. The test has also been used to
analyse systems with non-local operators in integro-differential equations [62] and
integro-partial differential equations [10]. Moreover, it has been used to analyse
experimental data and observations [13, 33, 34, 37, 38].
The remainder is organised as follows. In Sect. 7.2 we briefly describe the test.
The algorithm is then presented in Sect. 7.3 where we discuss several implementa-
tions of the test. The theoretical underpinning of our test is explained in Sect. 7.4.
This is followed by numerical results in Sect. 7.5 illustrating the efficiency of our
test to deal with intermittent maps, chaos in thin separatrix layers in Hamiltonian
systems, partial differential equations and data contaminated by observational noise.
We conclude with a summary in Sect. 7.6.

7.2 Description of the Test

The input of the test is a one-dimensional time series .n/ for n D 1; 2; : : : We use
the data .n/ to drive the 2-dimensional system

p.n C 1/ D p.n/ C .n/ cos cn;

q.n C 1/ D q.n/ C .n/ sin cn ; (7.1)

where c 2 .0; 2/ is fixed. Define the (time-averaged) mean square displacement

1 X 
N
M.n/ D lim p. j C n/  p. j/2 C q. j C n/  q. j/2 ; n D 1; 2; 3; : : :
N!1 N
jD1

and its growth rate

log M.n/
K D lim :
n!1 log n

Under general conditions, the limits M.n/ and K can be shown to exist, and K
takes either the value K D 0 signifying regular dynamics or the value K D 1
signifying chaotic dynamics.
7 The 0-1 Test for Chaos: A Review 223

A brief explanation of the rationale behind the test is as follows. (The math-
ematics is described more carefully in Sect. 7.4.) In the regular case (periodic or
quasiperiodic dynamics) the trajectories for the system (7.1) are typically bounded,
whereas in the chaotic case the trajectories for (7.1) typically behave approximately
like a two-dimensional Brownian
p motion with zero drift and hence evolve diffu-
sively (i.e. with growth rate n). A convenient method for distinguishing these
growth rates, bounded or diffusive, is via the mean square displacement M.n/ which
accordingly is either bounded or grows linearly. The diagnostic K 2 f0; 1g captures
this growth rate.
To summarise, we have the following two scenarios:

Underlying dynamics Dynamics of p.n/ and q.n/ M.n/ K

Regular Bounded Bounded 0
Chaotic Diffusive Linear 1

In the following section we describe the test in more detail focusing on the
practical issues in the implementation of the 0-1 test.

7.3 Description of the Algorithm

The test can be readily implemented in a few lines of code. We briefly describe its
implementation and refer the reader to [22] for more details. Given a time series
. j/ for j D 1; : : : ; N we perform the following sequence of steps:
1. For c 2 .0; /, we solve the system (7.1) to obtain

X
n X
n
pc .n/ D . j/ cos jc; qc .n/ D . j/ sin jc (7.2)
jD1 jD1

for n D 1; 2; : : : ; N. Typical plots of p and q for regular and chaotic dynamics

are given in Fig. 7.1 which clearly illustrates the bounded motion of p and q for
underlying regular dynamics and asymptotic Brownian motion for underlying
chaotic dynamics.
2. To analyse the diffusive (or non-diffusive) behaviour of pc and qc we compute

1 X
N
Mc .n/ D . pc . j C n/  pc . j/2 C qc . j C n/  qc . j/2 / : (7.3)
N jD1

To assure the limit N ! 1 we require n  N. Hence we calculate Mc .n/ only

for n  N0 where N0  N. In practice we find that N0 should not be chosen
much larger than N=10.
224 G.A. Gottwald and I. Melbourne

Fig. 7.1 Plot of p versus q for the logistic map xnC1 D xn .1  xn /. Left: Regular dynamics at
 D 3:55; Right: Chaotic dynamics at  D 3:97. We used N D 5000 data points

In [23] a modified mean square displacement

Dc .n/ D Mc .n/  Vosc .c; n/ (7.4)

was derived which exhibits the same asymptotic growth rate as Mc .n/ but with
better convergence properties. The correction term

1  cos nc
Vosc .c; n/ D .E/2
1  cos c
is readily estimated from the time average of the observable

1 X
N
E D lim . j/ :
N!1 N
jD1

Note that the asymptotic growth rates of Mc .n/ and Dc .n/ are the same.
In Fig. 7.2 we show the two mean square displacements Mc .n/ and Dc .n/ for
the logistic map xnC1 D xn .1xn / with  D 3:97 (which corresponds to chaotic
dynamics) and an arbitrary value of c D 0:9. The subtraction of the oscillatory
term Vosc .c; n/ clearly regularizes the linear behaviour of Mc .n/. This allows for
a much better determination of the asymptotic growth rate Kc of the mean square
displacement which is described in the next step.
3. The contrasting behaviour of the translation variables pc and qc as seen in Fig. 7.1
can be distinguished by the asymptotic growth rate Kc of the mean square
displacement (or of the modified mean square displacement Dc .n/). We present
here two different methods to compute Kc , namely the regression method and the
correlation method.
7 The 0-1 Test for Chaos: A Review 225

Fig. 7.2 Plot of the mean square displacement versus n for the logistic map with  D 3:97
corresponding to chaotic dynamics. We used N D 2000 data points and computed Mc .n/ and
Dc .n/ for n D 1; : : : ; 200 and an arbitrary value of c D 0:9

Regression method The regression method consists of linear regression for the
log-log plot of the mean square displacement (cf. Fig. 7.2). For the original
mean square displacement Mc .n/, the asymptotic growth rate Kc is given by the
definition

log Mc .n/
Kc D lim :
n!1 log n

Numerically, Kc is determined by fitting a straight line to the graph of log Mc .n/

versus log n through minimizing the absolute deviation.1 It is recommended to
minimize the absolute deviation rather than employing the usual least square
method as the latter assigns a higher weight to outliers. Outliers are typical for
small values of n since the linear behaviour of the mean square displacement is
only given asymptotically.
As seen in Fig. 7.2, Dc .n/ exhibits far less variance than Mc .n/ so it is natural
to apply the regression method to Dc .n/. However, since Dc .n/ may be negative
due to the subtraction of the oscillatory term Vosc .c; n/, we need to set

Q c .n/ D Dc .n/ C a min jDc .n/j ;

D
1nN0

1
One may either use off the shelf routines provided for example in Numerical Recipes [61] or
build-in routines in MATLAB [49].
226 G.A. Gottwald and I. Melbourne

where a > 1 (in the simulations presented here we chose a D 1:1) to obtain the
asymptotic growth rate

Q c .n/
log D
Kc D lim :
n!1 log n

Again, Kc can be determined numerically by regression (minimizing the absolute

Q c .n/ versus log n.
deviation) for the graph of log D
Correlation method In the correlation method, we form vectors
 D .1; 2; : : : ; N0 / and  D .Dc .1/; Dc .2/; : : : ; Dc .N0 // (alternatively, Mc .n/
could be used instead of Dc .n/). Recalling the definition of covariance and
variance of given vectors x, y of length q
1 Pq 1 Pq
cov.x; y/ D q jD1 .x. j/  xN /.y. j/  yN /; where xN D q jD1 x. j/ ;
var.x/ D cov.x; x/ ;

we define the correlation coefficient

cov.; /
Kc D corr.; / D p 2 1; 1 :
var./var./

This quantity measures the strength of the correlation of Dc .n/ with linear
growth. The correlation method greatly outperforms the regression method (see
Figs. 7.3 and 7.4 below), but assumes that the dynamics is such that with
probability 1 we have Kc D 0 or Kc D 1. This is justified for large classes of
dynamical systems [23].
4. Steps 13 need to be executed for various values of c. In practice, 100 choices of
c is sufficient. We then compute the median of these values of Kc to compute
the final result K D median.Kc /. The values of c are chosen randomly in
the interval c 2 .=5; 4=5/ to avoid resonances. Resonances occur when the
dynamics involves a periodic component with frequency ! implying a term in
the Fourier decomposition of the observable  proportional to exp.i!k/. In
this case there is a resonance at c D ! leading to pc .n/  n and qc .n/  n
and hence Mc .n/  n2 (and Dc .n/  n2 ) implying Kc D 2 for the regression
method and Kc D 1 for the correlation method. Note that for c D 0 the test
would yield a resonance irrespective of the underlying dynamics (which is why
this value should be excluded). See [19, 22] for more details on resonances.
In Fig. 7.3 we show Kc versus c for the logistic map for regular and chaotic
dynamics. Resonances are clearly visible for the periodic case, with Kc D 2
for the regression method and Kc  1 for the correlation method.
Our test states that a value of K  0 indicates regular dynamics, and K  1
indicates chaotic dynamics. This is exemplified in Fig. 7.4 where K is shown as
a function of the parameter  of the logistic map.
7 The 0-1 Test for Chaos: A Review 227

Fig. 7.3 Plot of Kc versus c for the logistic map calculated using the regression method (top)
and correlation method (bottom). We used here N D 5000 data points, and 1000 equally spaced
values for c. Left:  D 3:55 corresponding to regular dynamics; Right:  D 3:97 corresponding
to chaotic dynamics

Remark on Finite Size Problems There are finite size issues that are inherent to all
methods for chaos detection, namely that the length of the time series is sufficiently
long to capture the dynamics across the whole of the attractor. Specifically, for the
0-1 test the determination of the mean square displacement requires n  N0  N,
and the test relies on asymptotic behaviour of the (non)-diffusive behaviour of p and
q which for too small time series data length may not yet be dominant. Concerning
the latter point it is pertinent to mention that even in cases of time series which
are too short to allow for convergence of K to either 0 or 1, strong indications for
the presence or absence of chaos can be found by looking at the behaviour of K
with the length of the time series used to determine K. Figure 7.5 shows typical
decreasing/increasing behaviour of K near parameter values of the logistic map at
the so called edge of chaos, indicating regular or chaotic dynamics, respectively.
This was discussed at length in [21, 22].
228 G.A. Gottwald and I. Melbourne

Fig. 7.4 Plot of K versus  for the logistic map with 3:5    4 increased in increments of
0:001. We used N D 2000 data points. Shown are results when K is calculated via the regression
method (dashed line, blue) and when K is calculated via the correlation method (continuous line,
red). The horizontal lines indicate the limiting cases K D 0 and K D 1. We used 100 randomly
distributed values of c, and the mean square displacement was determined using Dc .n/

0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
1000 2000 3000 4000 0.5 1 1.5 2
x 104

Fig. 7.5 Plot of K versus the length of the time series N for the logistic map near the edge of
chaos. We used 100 randomly distributed values of c, and the mean square displacement was
determined using Dc .n/ with N0 D N=10, and the correlation method was used to determine Kc .
Left:  D 3:569 corresponding to regular dynamics; Right:  D 3:571 corresponding to chaotic
dynamics

7.4 Theoretical Framework for the 0-1 Test

Systems of the type (7.1) were studied extensively in [1, 14, 52, 53, 56]. The
motivation there was to understand growth rates of trajectories in systems with
Euclidean symmetry. A large class of discrete time systems with planar Euclidean
7 The 0-1 Test for Chaos: A Review 229

symmetry are given by skew product equations of the form

x.n C 1/ D f .x.n//;
#.n C 1/ D #.n/ C h.x.n//; (7.5)
p.n C 1/ D p.n/ C .x.n// cos.#.n//   .x.n// sin.#.n//;
q.n C 1/ D q.n/ C .x.n// sin.#.n// C  .x.n// cos.#.n//:

Here f W X ! X defines the base dynamics (perpendicular to the symmetry vari-

ables) while #.n/ represents two-dimensional rotations and .p.n/; q.n// represent
planar translations. It is assumed that the functions h; ;  W X ! R are smooth.
In [56], it was shown that if the dynamics on X is periodic or quasiperiodic, then
typically the translation variables p.n/; q.n/ remain bounded. However, sufficiently
chaotic dynamics on X leads to diffusive behaviour in the translation variables.
(See [14, 53] for the case of uniformly hyperbolic dynamics, and more recently [25]
for nonuniformly hyperbolic dynamics.) Using these results, and computing the
growth rate K of the mean square displacement as described in Sect. 7.2, we obtain
with probability one the growth rates K D 0 and K D 1 respectively in these two
situations. It is pertinent to stress that the test does not rely in anyway on a possible
exponential decay of the auto-correlation function; the test holds for polynomial
decay as well and even does not require summability of the auto-correlation function
[24, 25] (see also Sect. 7.5.1.1). Also it is irrelevant whether the dynamics is mixing.
For example, in the case of the logistic map mentioned in Sect. 7.5.1, the attractor
is almost always a periodic sink or a strongly chaotic attractor consisting of a finite
union of intervals permuted cyclically by the dynamics. In both cases, the attractor
is mixing up to a finite cycle but is generally nonmixing. More importantly, in the
case of continuous time dynamics, it is rarely the case that mixing can be established
but the 0-1 test is still valid.
It should be noted that a similar dichotomy holds in the absence of the rotation
variables, except that the bounded/diffusive behaviour is superimposed on a linear
drift. The rotation symmetry kills off the linear drift [15, 56], rendering the
bounded/diffusive dichotomy more readily detectable.
The idea behind the 0-1 test is to adjoin rotation and translation variables #, p, q
to a given (but unknown) dynamical system f W X ! X generating data .n/, thus
producing a system with Euclidean symmetry to which the above theoretical results
apply. Note that choosing h  c,   0 and making the identification .x.n// D
.n/ reduces the skew product system (7.5) to the 2-dimensional system (7.1) used
in the 0-1 test.
The original version of the test [19] used generic choices of h so that certain
theoretical results of Field et al. [15] could be applied in justifying the test.
The current version is much more effective for noisy data [20] but the original
theoretical justification for the test no longer applies. Nevertheless, it transpires
that the simplified nature of the equations in (7.1) enables certain improvements
to the theoretical underpinnings for the test, as described in [23]. The structure of
230 G.A. Gottwald and I. Melbourne

the simplified equations means that they are amenable to techniques from Fourier
analysis. In particular, there are connections with power spectra as described in the
next subsection. In the aforementioned cases of periodic/quasiperiodic dynamics
and uniformly/nonuniformly hyperbolic dynamics, we typically obtain K D 0
and K D 1, respectively. Here typically is in the sense of probability one: for
almost every choice of c. As mentioned previously, we take the median value of K,
computed with 100 randomly chosen choices of c, to circumvent the issue regarding
bad choices of c. Moreover, the considerations in [23] lead directly to the modified
mean square displacement Dc .n/ which we have seen leads to improved results
(there is no analogue of this modification for the original test).

7.4.1 Connection with the Power Spectrum

Consider a discrete dynamical system f W X ! X with ergodic invariant measure .

Given a square-integrable observable v W X ! R, the power spectrum S is defined
to be the square of the Fourier amplitudes of v f j per unit time2
Z Xn1 2
1
S.!/ D lim eij! v f j d; ! 2 0; 2: (7.6)
n!1 n X jD0

It was proven in [51] that the power spectrum has a broadband nature and is nowhere
zero for a large class of dynamical systems, including slowly mixing systems such
as PomeauManneville maps provided the auto-correlation function is summable.
(For a discussion of recent results in the case of nonsummable autocorrelations, see
Sect. 7.5.1.1.)
A simple short calculation shows that
Z X
n1
1 1
S.c/ D lim j eijc v f j j2 d D lim Mc .n/ ; (7.7)
n!1 n X n!1 n
jD0

implying that

Mc .n/ D S.c/n C o.n/ : (7.8)

This may give the wrong impression that the 0-1 test for chaos is simply evaluating
the power spectrum. From (7.8) one can only conclude that if the power spectrum is
nowhere nonzero (S.c/ 0 for all c), then the asymptotic growth rate of the mean
square displacement becomes Kc D 1 for all c. On the other hand, if S.c/ D 0 for all
c, it does not automatically follow that Kc D 0 (for example, the o.n/ term could be

2
One may use e2ij!=n rather than eij! for a rescaled domain.
7 The 0-1 Test for Chaos: A Review 231

of the form n= log.n/ implying Kc D 1). However in [23] it was rigorously proven
that for a large class of dynamical systems, the o.n/ terms are such that for chaotic
dynamics one obtains Kc D 1 and for regular dynamics Kc D 0.
It is pertinent to mention the computational advantage of the 0-1 test which
extracts in a single number K the property of the power spectrum which is relevant
for underlying chaotic or regular dynamics, i.e. whether it is everywhere or nowhere
nonzero. The test completely bypasses the explicit computation of the power
spectrum which would require considerably more data. Moreover, K can be plotted
against a parameter of the system as in Fig. 7.4 and the convergence of K can be
seen against the number of iterates N as in Fig. 7.5. There do not exist analogous
plots for the power spectrum.

7.5 Numerical Examples for the 0-1 Test

We now illustrate the applicability of our test to be able to distinguish regular

dynamics from chaotic dynamics in discrete and continuous time systems, dissi-
pative and Hamiltonian systems, noise free and noise contaminated data.

7.5.1 Discrete Time Systems

One of the simplest families of dynamical systems that exhibits regular and chaotic
dynamics is the logistic map f W 0; 1 ! 0; 1 given by f .x/ D x.1  x/. Here,
 2 0; 4 is a parameter. It is well-known that there is a unique attractor for each
value of  and that the basin of attraction is of full measure in 0; 1. For almost
every value of , the attractor is either a periodic orbit or a strongly chaotic attractor.
Throughout the earlier sections, this family of maps was used as an illustrative
example for various features of the 0-1 test, see Figs. 7.1, 7.2, 7.3, 7.4 and 7.5.
We now proceed to explore two further families of discrete time systems.

7.5.1.1 PomeauManneville Map

A prototypical family of maps exhibiting intermittency and weakly chaotic dynam-

ical systems with sticky equilibria are PomeauManneville intermittency maps
xnC1 D f .xn / with f W 0; 1 ! 0; 1 given by
(
1
x .1 C 2 x / 0x 2
f .x/ D 1
(7.9)
2x  1 2
x1
232 G.A. Gottwald and I. Melbourne

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 200 400 600 800 0 200 400 600 800

Fig. 7.6 Time series of the PomeauManneville map (7.9). Left: Strongly chaotic case with  D
0:2. Right: Weakly chaotic case with  D 0:7

where  is a parameter [44, 59]. For  2 0; 1/ there exists a unique absolutely

continuous invariant probability measure (SRB measure). When  D 0 the map
reduces to the doubling map with exponential decay of correlations. For  2 .0; 1/
the decay of correlations is polynomial with rate 1=n.1= /1 which is summable
for  < 12 and nonsummable for  2 12 ; 1/ [30]. For  > 0 the fixed point at
0 is indifferent (f 0 .0/ D 1) and plays the role of the sticky regular dynamics
leading to laminar behaviour interspersed with intermittent chaotic bursts. This is
illustrated in Fig. 7.6, where we show a trajectory of the PomeauManneville map
in the strongly chaotic case with  D 0:2, where the correlations are summable,
and in the intermittent weakly chaotic case with  D 0:7, where the correlations are
nonsummable.
It is well-known [17] that for such intermittent systems the usual central limit
theorem breaks down for  2 . 12 ; 1/ leading to fluctuations of Lvy type rather
than of Gaussian type. For mathematically rigorous results on this, see [27, 54, 74].
Despite this, our test for chaos is still able to detect chaos in the weakly chaotic
case with nonsummable correlations. A proof of this statement is currently work in
progress, but the underlying reason, as discussed in related results in [24, 25], is that
the anomalous diffusion is suppressed due to the rotation symmetry induced by the
presence of c in (7.2). As a result of this, when the dynamics is trapped near the
indifferent fixed point, the dynamics appears regular and therefore leads to bounded
dynamics of the translation variables p and q as seen in Fig. 7.7 for n D 1 C xn . In
Fig. 7.8 we show the asymptotic growth rate as a function of  .

7.5.1.2 Standard Map

We now consider the area-preserving Standard map [6, 41]

nC1 D n C sin.n / (7.10)

7 The 0-1 Test for Chaos: A Review 233

Fig. 7.7 Typical plots of the translation variables p and q driven by an observable n D 1 C xn
of the PomeauManneville map (7.9) for c D 2:1375. Left: Strongly chaotic case with  D 0:2.
Middle: Weakly chaotic case with  D 0:7. Right: Zoom for weakly chaotic case with  D 0:7
showing the trace of the laminar phases of the PomeauManneville dynamics in form of bounded
circles

0.8
correlation method
regression method
0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1

Fig. 7.8 Plot of K as a function of  for an observable n D 1 C xn of the PomeauManneville

map (7.9), using the regression method (crosses) and the correlation method (circles). We used
N D 10;000 data points and 100 randomly distributed values of c

nC1 D n C nC1 : (7.11)

Figure 7.9 shows the trajectories of 100 randomly chosen initial conditions after a
transient of 10;000 iterates for D 0:9. The phase space consists of regular islands
embedded in chaotic layers. In contrast, at D 0:3 there is a hyperbolic fixed
point at the origin and the asymptotic dynamics occurs in a thin separatrix layer
as seen in Fig. 7.10 for 100 randomly chosen initial conditions 0 2 0; 0:03 and
0 2 0:03; 0:03. This thin separatrix layer contains complex structures with many
tiny islands embedded within a chaotic sea [29]. This case exhibits weak chaos in the
sense of [73] with small Lyapunov exponents which may be difficult to distinguish
from those corresponding to regular orbits.
234 G.A. Gottwald and I. Melbourne

Fig. 7.9 Standard map (7.11) exhibiting chaos with D 0:9

0.03

0.02

0.01

0
0.02 0 0.02

Fig. 7.10 Standard map (7.11) exhibiting weak chaos in a small separatrix layer with D 0:3.
The right figure is a zoom near the hyperbolic point showing the enlarged stochastic layer

In Figs. 7.11 and 7.12 we show how the 0-1 test is able to detect regular and
chaotic orbits, even in the weakly chaotic case. We have chosen 10002 initial
conditions and run them for 10;000 steps. We used the correlation method and
applied it to the modified mean square displacement Dc .n/.

7.5.2 Continuous Time Systems

We have so far formulated the 0-1 test for discrete time systems. For continuous
time series .t/, we obtain a discrete time series .t1 /, .t2 /, .t3 /; : : : for given
discrete times 0 < t1 < t2 < t3 <    to which the test for chaos may be
7 The 0-1 Test for Chaos: A Review 235

Fig. 7.11 Contour plot of K for the standard map (7.11) exhibiting chaos with D 0:9. We used
10002 equally spaced initial conditions and calculated K via the correlation method from Dc

Fig. 7.12 Contour plot of K for the standard map (7.11) exhibiting chaos with D 0:3. We
used 10002 equally spaced initial conditions and calculated K via the correlation method from Dc

applied as in previous sections. The sequence tj , j  1, has to be chosen in a

deterministic manner to assure that the time series .tj / is deterministic. One may
choose the tj as the intersection times with a cross-section. In this case the time
series .tj / corresponds to observing a Poincar map. A second, perhaps more
usual, approach is to take tj D js where s > 0 is the sampling time. The time
series .tj / D . js / corresponds to observing the time-s map associated with
the underlying continuous time system. Contrary to the case of observing a Poincar
map, in the latter approach one is faced with a well-known oversampling issue: If s
is too small, then the system is oversampled and this often leads to incorrect results
236 G.A. Gottwald and I. Melbourne

[22]. Although oversampling is a practical problem for data series of finite size, it
should be emphasized that theoretically the test works for all sampling times s in
the limit N ! 1. We now present numerical results for an ordinary differential
equation and a partial differential equation where care has to be taken to overcome
the issue of oversampling for the realistic case of finite data series.

7.5.2.1 Rssler Equations

To illustrate how the issue of oversampling manifests itself in the 0-1 test for chaos
and how to overcome it, as proposed in [22], we consider here the 3-dimensional
Rssler system [64]

xP D y  z
yP D x C ay
zP D b C z.x  d/ : (7.12)

For the values a D 0:432, b D 2 and d D 4, the system exhibits chaos with a
maximal Lyapunov exponent of about max  0:1 (we use the natural logarithm).
We have integrated this system with a fourth-order RungeKutta scheme with
variable step-size and recorded 100;000 data points each  t D 0:01 (i.e. 1000
time units) after disregarding a transient behaviour of 50 time units to allow for the
dynamics to settle on the attractor. A plot of the dynamic in the x-y-plane is provided
in Fig. 7.13.

6
4 2 0 2 4 6

Fig. 7.13 Phase portrait for the Rssler system (7.12). The short trajectory segment (blue) was
sampled at s D 0:5
7 The 0-1 Test for Chaos: A Review 237

4
0 20 40 60

Fig. 7.14 Plot of the observable .t/ D x.t/ for the Rssler system (7.12). The finely sampled
data (crosses) are sampled at s D 0:05 time units. The coarsely sampled data (filled circles) are
sampled at s D 0:35 time units

Figure 7.14 shows an oversampled and a sufficiently coarsely sampled observ-

able for the Rssler system (7.12). The finely sampled time series (s D 0:05) yields
K  0 whereas the coarsely sampled data (s D 0:35) yields K  1 already despite
using only 1=7th of the data.
A good choice of the sampling time s can often be obtained by visual inspection
as in Fig. 7.14. A more quantitative method is to use the e-folding time of the auto-
correlation function or to use the first minimum of the mutual information [16, 31].
For the data depicted in Fig. 7.14 these method yield s D 1:15 and s D 1:50,
respectively. However, we observed here that the smaller sampling time s D 0:35
already yields K  1 with the advantage of being a longer data set. In general,
the optimal sampling time will depend on the dynamical system and the time series
under consideration. We refer the reader to [31] for a discussion on optimal time
delays in the context of phase space reconstruction.
In the following we show how the issue of oversampling arises in the 0-1 test. For
continuous time systems, the (time-averaged) mean square displacement is defined
as
Z
1 T
Mc .t/ D lim .p.t C /  p.//2 C .q.t C /  q.//2 d ;
T!1 T 0

which, for a time series sampled with sample time s , is approximated by

1 X 
N
Mc .n/ D lim ps . j C n/  ps . j/2 C qs . j C n/  qs . j/2 s2 :
N!1 N
jD1
238 G.A. Gottwald and I. Melbourne

Similarly the power spectrum for the time-continuous case discretizes to

1 X 2i  j 2
n1

S./ D lim E e s . j/ s2 ; (7.13)
n!1 n
jD0

where s D 1=s is the sample frequency. For chaotic systems the power spectrum
decays for large frequencies , and so for frequencies larger than some max the
power spectrum is zero for all practical purposes.
Comparing (7.13) with the power spectrum (7.6) for discrete-time data, we
identify

c D 2 ;  2 0; max  : (7.14)
s

Sampling at the Nyquist rate with s? D 2max corresponds to c 2 .0; / as before.
However, oversampling at a higher frequency s > s? , restricts the effective choices
?
of c to c 2 .0; c? / where c? D ss  < . In this case, the test for chaos may
incorrectly classify the dynamics of a chaotic system as regular, since it is possible
that more than half of the randomly chosen values of c 2 .0; / will lie in .c? ; /
yielding a median K D 0. Note that once a sampling time s 0 is fixed, the
problem of oversampling cannot be alleviated by increasing the length N of the
time series.
We illustrate this using the Rssler system (7.12) sampled with s ranging from
s D 0:05 up to s D 1. In Fig. 7.15 the median of the asymptotic growth rate

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1

Fig. 7.15 Plot of K as a function of the sample time s for the Rssler system (7.12). At the
finest sampling rate s D 0:05 we recorded N D 100;000 data points. Results are shown for the
correlation method (circles) and the regression method (crosses)
7 The 0-1 Test for Chaos: A Review 239

0.8

0.6

0.4

0.2

0
0

Fig. 7.16 Plot of Kc as a function of the frequency c for the Rssler system (7.12). From left to
right we used s D 0:15, s D 0:25 and s D 0:35. The corresponding values of the growth rate
(calculated using the correlation method) are K D 0:005, K D 0:5 and K D 0:97, respectively. At
the sampling rate s D 0:05 we recorded N D 100;000 data points

K is shown as a function of the sample time. For data that is too finely sampled,
we obtain K D 0 although the dynamics is actually chaotic. Figure 7.16 illustrates
how the range of effective values of c depends on the sampling time s . The linear
scaling of the range of c for which Kc  1 as suggested by (7.14) is clearly seen.
The pronounced dips of Kc for certain values of c are caused by near resonances
which occur in the chaotic Rssler system for our parameter values caused by
regularly appearing revolutions of the dynamics as illustrated in Fig. 7.13. Note that
the presence of resonances does not affect the value of the median K as seen in
Fig. 7.15.

7.5.2.2 Partial Differential Equations

We apply now our test to the driven and damped nonlinear Schrdinger equation

{qt C qxx C 2jqj2 q D { q C "e{.!tC / ;

which describes a plasma resonantly driven by a capacitor with frequency ! and

damped via collisions [4, 11, 57]. For q D Q exp.{.!t C // we solve

{Qt C Qxx C 2jQj2 Q D !Q  { Q C " : (7.15)

It is well known that for given system length L the system (7.15) undergoes
a period doubling bifurcation route into chaos [57] for increasing values of the
driving amplitude ". We present here results for a system with length L D 80,
240 G.A. Gottwald and I. Melbourne

Fig. 7.17 Hovmller diagram of jQ.x; t/j for regular dynamics with " D 0:095 (left) and chaotic
dynamics with " D 0:2 (right) for the driven and damped Schrdinger equation (7.15)

 D 0:11, ! D 1 for " D 0:095 and " D 0:2 for regular and chaotic dynamics,
respectively. The system (7.15) is integrated with a second-order in space and time
finite difference CrankNicolson solver where the nonlinear term is treated with an
AdamsBashforth scheme. We use nx D 256 grid points and anpintegration time
step of dt D 0:0001 and evolve from an initial condition q D { 2 C 0:1 cos.kx/
with k D 15 with reflective (von Neumann) boundary conditions. In Fig. 7.17
we show Hovmller diagrams of jQ.x; t/j for regular and chaotic dynamics. We
construct observables by evaluating the field Q.x; t/ at spatial locations xj D jdx with
dx D L=nx and j D 1;    ; nx . In particular, we consider the following observables

X
nx
1 .t/ D jQ.xj ; t/j;
jD1

2 .t/ D jQ.L=2; t/j;

5
X
3 .t/ D jQ.xj? ; t/j:
jD1

For the last observable 3 .t/ we randomly choose five locations xj? from the nx D
256 spatial gridpoints at time t D 0.
The observables 1;2;3 are sampled time every 0:3 time units with a total of
10;000 snapshots taken. This sampling time is sufficiently large to avoid the
oversampling effects for continuous time systems which would lead to K  0
irrespective of the underlying dynamics, as discussed in Sect. 7.5.2.1 (see also [22]).
Using the correlation method on Dc .n/, for each of the three observables we obtain
values of K smaller than 0:0017 in the regular case with " D 0:095, and values of K
within 0:003 from K D 1 for the chaotic case.
We note that although the nonlinear Schrdinger equation (7.15) is formally
infinite-dimensional, its dynamics evolves on a finite dimensional attractor [11].
7 The 0-1 Test for Chaos: A Review 241

7.5.3 Data Contaminated by Noise

For a test for chaos to be able to analyse real world data one needs to show its
capability to be able to process observations contaminated by noise. In the following
we revisit the example of measurement noise in an 8-dimensional Lorenz-96 model
studied in our previous work [20]. There it was shown that our 0-1 test for chaos is
far superior to traditional methods using phase space reconstruction and Lyapunov
exponents [63, 66], without preprocessing the data with standard noise reduction
methods [31].

7.5.3.1 Lorenz-96 System

We revisit the tough test case of analysing quasi-periodic dynamics with measure-
ment noise [20]. In particular we study the Lorenz-96 system [45]

zPi D zi1 .ziC1  zi2 /  zi C F i D 1;    ; D (7.16)

with periodic ziCD D zi . This system is a toy-model for midlatitude atmospheric

dynamics, incorporating linear damping, forcing and nonlinear transport. In the
atmospheric context one usually uses D D 40. This particular value for D is
chosen such that the spacing between adjacent grid points zi roughly equals the
Rossby radius of deformation at midlatitudes where the circumference of the earth
is roughly 30;000 km. The dynamical properties of the Lorenz-96 system have been
investigated, for example, in [46, 58]. We use D D 8 modes where quasi-periodic
windows were found to alternate with chaotic dynamics [58].
In the previous examples, the test was able to distinguish sharply between regular
and chaotic dynamics in noise free data. However such good performance of the
test for noise free data is detrimental for noise contaminated dataeven small
amounts of noise would be detected and noisy regular dynamics would be falsely
classified as chaotic. The sensitivity of our test was enhanced by the subtraction of
the oscillatory term (cf. Eq. (7.4) and Fig. 7.2), and the application of the correlation
method. In [20] our test employed directly the mean square displacement Mc .n/
including the oscillatory term rather than the modified version Dc .n/. Furthermore,
it used the regression method rather than the correlation method to calculate Kc
from the mean square displacement Mc .n/. We showed that in this case our test
greatly outperforms methods involving phase space reconstruction and Lyapunov
exponents in distinguishing quasi-periodic dynamics from chaotic dynamics when
10 % measurement noise was added to the observations.
Here we will revisit our test including a method to deal with measurement noise
proposed in [22]: If we assume that the noise is independent of the dynamics, i.e.
pure measurement noise, and also independent of the forcing F, we can decompose
the linear part of the modified mean square displacement as

Dc .n/ D .Sdyn .c/ C Snoise .c//n C o.n/ : (7.17)

242 G.A. Gottwald and I. Melbourne

Here Sdyn .c/ is the variance associated with the underlying deterministic dynamics
to be analysed and Snoise .c/ is the variance associated with the measurement noise.
We can estimate the variance associated with the measurement noise Snoise .c/ by
estimating Dc .n/ from the noisy observations at a parameter F where the dynamics
is known to be regular with Sdyn .c/ D 0. We then estimate Snoise .c/ to compensate
for the linear growth of Dc .n/ due to the noise. We remark that this is not always
possible and requires (at least) that gauge experiments can be performed. For
example, this method cannot help with studying the regularity of planetary motion
for noisy observations. If gauge experiments are not possible, our test can still
be used to analyse noise-contaminated experimental data using the formulation
proposed in reference [20] as done, for example, in [13, 33, 34, 37, 38]. Once
Snoise .c/ is estimated the test can proceed with

O c .n/ D Dc .n/  Snoise .c/ n

D (7.18)

as described in the previous sections. In particular we can employ the more

sensitive correlation method. Note that the linear growth term of the mean-square
displacement will not be entirely eliminated (unless by chance Snoise .c/ is correctly
estimated from the data) but the proposed scheme controls its magnitude. This
allows the test to analyse observational data of finite length; for unlimited noise-
contaminated data, one would, of course, obtain K D 1, irrespective of the
underlying deterministic dynamics.
In Fig. 7.18 we show results of our test using first Mc .n/ and the regression
method as in [20] and second using D O c .n/ and the correlation method. We use
a fourth-order RungeKutta scheme with a time step of dt D 0:05 to generate
observations

.t/ D .1 C / .z2 C z3 C z4 / : (7.19)

The measurement noise  D u is drawn from a uniform distribution u  U 1; 1.

In Fig. 7.18 we show results for noise-free observations with  D 0, and for
measurement noise levels of 10 % and 20 % with  D 0:1 and  D 0:2, respectively.
Observations are taken every 2:5 time units and a total of N D 100;000 observations
are taken.
We gauge the variance due to the noise at F D 5:25. We obtain Snoise D 0:3
and Snoise D 0:62 for a noise level  of 10 % and 20 % observational noise,
respectively. Both methods detect the quasi-periodic windows well for 10 % mea-
surement noise (we remark that methods relying on phase space reconstruction and
maximal Lyapunov exponents were not able to accurately distinguish quasi-periodic
dynamics from chaotic dynamics [20]). The distinction between quasi-periodic
dynamics and chaotic dynamics with a noise level of 20 % is less clear in the test
employing Mc .n/ but still remarkably good when using D O c .n/. We propose this
example which involves noise contaminated quasi-periodic dynamics as a challenge
for other tests.
7 The 0-1 Test for Chaos: A Review 243

0.8

0.6

0.4

0.2

0.2
5.25 5.3 5.35 5.4 5.45 5.5

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0

0.2 0.2
5.25 5.3 5.35 5.4 5.45 5.5 5.25 5.3 5.35 5.4 5.45 5.5

Fig. 7.18 Plot of K versus F for the Lorenz-96 system (7.16) for 5:25  F  5:5 increased in
increments of 0:005. We used N D 100;000 data points sampled at 2:5 time units. Top: Noise free
data; Bottom left: 10 % measurement noise; Bottom right: 20 % measurement noise. K is calculated
via the regression method for Mc .n/ (crosses, blue) and for the correlation method for Dc .n/ with
subtracted noise variance (circles, magenta). The horizontal lines indicate the limiting cases K D 0
and K D 1. We used 100 randomly distributed values of c

7.6 Summary

We have described the 0-1 test for chaos, focusing on its implementation and several
practical issues as well as on the theoretical justifications outlining the realm of
validity of the test. We have illustrated the versatility and efficiency of our method
by treating the notoriously difficult case of weakly chaotic separatrix layers in the
standard map as well as analysing measurement noise contaminated data.
The advantage of our method lies in (a) its computational low cost and ease
of implementation, (b) its generality of applicability independent of the nature of
the dynamical system and its dimension, (c) its working directly with the time
series without the need for phase space reconstruction, (d) its ability to detect
weak chaos and (e) its ability to detect regular behaviour within noisy data. In
particular, we mention the 8-dimensional Lorenz-96 model contaminated by noise
(see Sect. 7.5.3). We are not aware of any other method that comes close to matching
the effectiveness of our test for this example.
244 G.A. Gottwald and I. Melbourne

The theoretical justification of the 0-1 test depends on the nature of attractors for
general smooth (or piecewise smooth) dynamical systems. In [23] we challenged
the skeptical reader to construct a robust smooth example where the test fails. So
far, no such example has come to light. This was explored further in [26] where
we formulated a conjecture which, roughly speaking, states that for typical smooth
dynamics, either Kc D 1 for almost every c or Kc D 0 for almost every c.

Acknowledgements GAG acknowledges support from the Australian Research Council. The
research of IM was supported in part by the European Advanced Grant StochExtHomog (ERC
AdG 320977).

References

1. Ashwin, P., Melbourne, I., Nicol, M.: Hypermeander of spirals; local bifurcations and statistical
properties. Physica D 156, 364382 (2001)
2. Bernardini, D., Rega, G., Litak, G., Syta, A.: Identification of regular and chaotic isothermal
trajectories of a shape memory oscillator using the 01 test. Proc. Inst. Mech. Eng. K: J. Multi-
body Dyn. 227(1), 1722 (2013)
3. Cafagna, D., Grassi, G.: An effective method for detecting chaos in fractional-order systems.
Int. J. Bifurcat. Chaos 20(3), 669678 (2010)
4. Cai, D., McLaughlin, D.W.: Chaotic and turbulent behaviour of unstable one-dimensional
nonlinear dispersive waves. J. Math. Phys. 41(6), 41254153 (2000)
5. Cao, J., Syta, A., Litak, G., Zhou, S., Inman, D., Chen, Y.: Regular and chaotic vibration in a
piezoelectric energy harvester with fractional damping. Eur. Phys. J. Plus 130(6) (2015)
6. Chirikov, B.V.: A universal instability of many-dimensional oscillator systems. Phys. Rep. 52,
263 (1979)
7. Chowdhury, D.R., Iyengar, A.N.S., Lahiri, S.: Gottwald Melborune (01) test for chaos in a
plasma. Nonlinear Process. Geophys. 19(1), 5356 (2012)
8. Dafilis, M., Frascoli, F., McVernon, J., Heffernan, J.M., McCaw, J.M.: The dynamical
consequences of seasonal forcing, immune boosting and demographic change in a model of
disease transmission. J. Theor. Biol. 361, 124132 (2014)
9. Dafilis, M., Frascoli, F., McVernon, J., Heffernan, J., McCaw, J.: Dynamical crises, multista-
bility and the influence of the duration of immunity in a seasonally-forced model of disease
transmission. Theor. Biol. Med. Model. 11(1), 43 (2014)
10. Diddens, C., Linz, S.J.: Continuum modeling of particle redeposition during ion-beam erosion.
Eur. Phys. J. B 86(9), 113 (2013)
11. Eickermann, T., Grauer, R., Spatschek, K.H.: Identification of mass capturing structures in a
perturbed nonlinear Schrdinger equation. Phys. Lett. A 198, 383388 (1995)
12. Erzgrber, H., Wieczorek, S., Krauskopf, B.: Dynamics of two semiconductor lasers coupled
by a passive resonator. Phys. Rev. E 81, 056201 (2010)
13. Falconer, I., Gottwald, G.A., Melbourne, I., Wormnes, K.: Application of the 01 Test for chaos
to experimental data. SIAM J. Appl. Dyn. 6, 395402 (2007)
14. Field, M., Melbourne, I., Trk, A.: Decay of correlations, central limit theorems and
approximation by Brownian motion for compact Lie group extensions. Ergodic Theor. Dyn.
Syst. 23, 87110 (2003)
15. Field, M., Melbourne, I., Trk, A.: Stable ergodicity for smooth compact Lie group extensions
of hyperbolic basic sets. Ergodic Theor. Dyn. Syst. 25, 517551 (2005)
16. Fraser, A.M., Swinney, H.L.: Independent coordinates for strange attractors from mutual
information. Phys. Rev. A 33, 11341140 (1986)
7 The 0-1 Test for Chaos: A Review 245

17. Gaspard, P., Wang, X.-J.: Sporadicity: between periodic and chaotic dynamical behaviours.
Proc. Natl. Acad. Sci. USA 85, 45914595 (1988)
18. Gopal, R., Venkatesan, A., Lakshmanan, M.: Applicability of 01 test for strange nonchaotic
attractors. Interdiscip. J. Nonlinear Sci. 23(2), 023123 (2013)
19. Gottwald, G.A., Melbourne, I.: A new test for chaos in deterministic systems. Proc. R. Soc. A
460, 603611 (2004)
20. Gottwald, G.A., Melbourne, I.: Testing for chaos in deterministic systems with noise. Physica
D 212(12), 100110 (2005)
21. Gottwald, G.A., Melbourne, I.: Comment on Reliability of the 01 test for chaos. Phys. Rev.
E 77, 028201 (2008)
22. Gottwald, G.A., Melbourne, I.: On the implementation of the 01 test for chaos. SIAM J. Appl.
Dyn. 8, 129145 (2009)
23. Gottwald, G.A., Melbourne, I.: On the validity of the 01 test for chaos. Nonlinearity 22,
13671382 (2009)
24. Gottwald, G.A., Melbourne, I.: A Huygens principle for diffusion and anomalous diffusion in
spatially extended systems. Proc. Natl. Acad. Sci. USA 110, 84118416 (2013)
25. Gottwald, G.A., Melbourne, I.: Central limit theorems and suppression of anomalous diffusion
for systems with symmetry (2013, submitted)
26. Gottwald, G.A., Melbourne, I.: A test for a conjecture on the nature of attractors for smooth
dynamical systems. Chaos 24, 024403 (2014)
27. Gouzel, S.: Central limit theorem and stable laws for intermittent maps. Probab. Theor. Relat.
Fields 128, 82122 (2004)
28. He, K., Xu, Y., Zou, Y., Tang, L.: Electricity price forecasts using a Curvelet denoising based
approach. Phys. A Stat. Mech. Appl. 425, 19 (2015)
29. Howard, J.: Discrete virial theorem. Celest. Mech. Dyn. Astron. 92(13), 219241 (2005)
30. Hu, H.: Decay of correlations for piecewise smooth maps with indifferent fixed points. Ergodic
Theor. Dyn. Syst. 24, 495524 (2004)
31. Kantz, H., Schreiber, T.: Nonlinear Time Series Analysis. Cambridge University Press,
Cambridge (1997)
32. Kedra, M.: Deterministic chaotic dynamics of Raba River flow (Polish Carpathian Mountains).
J. Hydrol. 509, 474503 (2014)
33. Krese, B., Govekar, E.: Nonlinear analysis of laser droplet generation by means of 01 test for
chaos. Nonlinear Dyn. 67, 21012109 (2012)
34. Krese, B., Govekar, E.: Analysis of traffic dynamics on a ring road-based transportation
network by means of 01 test for chaos and Lyapunov spectrum. Transp. Res. Part C Emerg.
Technol. 36, 2734 (2013)
35. Kulp, C.W., Smith, S.: Characterization of noisy symbolic time series. Phys. Rev. E 83, 026201
(2011)
36. Kr, R.: Chaotic analysis of the GDP time series. In: Zelinka, I., Chen, G., Rssler, O.E.,
Snasel, V., Abraham, A. (eds.) Nostradamus 2013: Prediction, Modeling and Analysis of
Complex Systems. Advances in Intelligent Systems and Computing, vol. 210, pp. 353362.
Springer International Publishing, Berlin (2013)
37. Kr, R.: Finding chaos in finnish gdp. Int. J. Autom. Comput. 11(3), 231240 (2014)
38. Kr, R., Kratochvil, .: Analyses of the chaotic behavior of the electricity price series. In:
Sanayei, A., Zelinka, I., Rssler, O.E. (eds.) ISCS 2013: Interdisciplinary Symposium on
Complex Systems. Emergence, Complexity and Computation, vol. 8, pp. 215226. Springer,
Berlin, Heidelberg (2014)
39. Leon, F.: Design and evaluation of a multiagent interaction protocol generating behaviours with
different levels of complexity. Neurocomputing 146, 173186 (2014)
40. Li, X., Gao, G., Hu, T., Ma, H., Li, T.: Multiple time scales analysis of runoff series based on
the Chaos theory. Desalin. Water Treat. 52(1315), 27412749 (2015)
41. Lichtenberg, A., Lieberman, M.: Regular and Chaotic Dynamics. Applied Mathematical
Sciences. Springer, New York (1992)
246 G.A. Gottwald and I. Melbourne

42. Litak, G., Syta, A., Wiercigroch, M.: Identification of chaos in a cutting process by the 0-1 test.
Chaos Solitons Fractals 40, 20952101 (2009)
43. Litak, G., Radons, G., Schubert, S.: Identification of chaos in a regenerative cutting process by
the 0-1 test. Proc. Appl. Math. Mech. 9(1), 299300 (2009)
44. Liverani, C., Saussol, B., Vaienti, S.: A probabilistic approach to intermittency. Ergodic Theor.
Dyn. Syst. 19, 671685 (1999)
45. Lorenz, E.N.: Predictability - a problem partly solved. In: Palmer, T. (ed.) Predictability.
European Centre for Medium-Range Weather Forecast, Shinfield Park, Reading (1996)
46. Lorenz, E.N., Emanuel, K.A.: Optimal sites for supplementary weather observations: simula-
tion with a small model. J. Atmos. Sci. 55(3), 399414 (1998)
47. Lugo-Fernndez, A.: Is the loop current a chaotic oscillator? J. Phys. Oceanogr. 37(6), 1455
1469 (2007)
48. Martinsen-Burrell, N., Julien, K., Petersen, M.R., Weiss, J.B.: Merger and alignment in a
reduced model for three-dimensional quasigeostrophic ellipsoidal vortices. Phys. Fluids 18,
057101 (2006)
49. MATLAB: version 7.10.0 (R2010a). The MathWorks Inc., Natick, MA (2010)
50. McLennan-Smith, T.A., Mercer, G.N.: Complex behaviour in a dengue model with a seasonally
varying vector population. Math. Biosci. 248, 2230 (2014)
51. Melbourne, I., Gottwald, G.A.: Power spectra for deterministic chaotic dynamical systems.
Nonlinearity 21, 179189 (2008)
52. Melbourne, I., Nicol, M.: Statistical properties of endomorphisms and compact group exten-
sions. J. Lond. Math. Soc. 70, 427446 (2004)
53. Melbourne, I., Trk, A.: Statistical limit theorems for suspension flows. Israel J. Math. 144,
191209 (2004)
54. Melbourne, I., Zweimller, R.: Weak convergence to stable Lvy processes for nonuniformly
hyperbolic dynamical systems. Ann. Inst. H. Poincar (B) Probab. Stat. 51, 545556 (2015)
55. Nair, V., Sujith, R.: A reduced-order model for the onset of combustion instability: physical
mechanisms for intermittency and precursors. Proc. Combust. Inst. 35(3), 31933200 (2015)
56. Nicol, M., Melbourne, I., Ashwin, P.: Euclidean extensions for dynamical systems. Nonlinear-
ity 14, 275300 (2001)
57. Nozaki, K., Bekki, N.: Low-dimensional chaos in a driven damped nonlinear Schrdinger
equation. Physica D 21, 381393 (1986)
58. Orrell, D., Smith, L.: Visualising bifurcations in high dimensional systems: the spectral
bifurcation diagram. Int. J. Bifurcat. Chaos 13(10), 30153028 (2003)
59. Pomeau, Y., Manneville, P.: Intermittent transition to turbulence in dissipative dynamical
systems. Commun. Math. Phys. 74, 189197 (1980)
60. Prabin Devi, S., Singh, S.B., Surjalal Sharma, A.: Deterministic dynamics of the magneto-
sphere: results of the 0-1 test. Nonlinear Process. Geophys. 20(1), 1118 (2013)
61. Press, W.H., Teukolsky, S.A., Vetterling, W.T., Flannery, B.P.: Numerical Recipes 3rd Edition:
The Art of Scientific Computing, 3rd edn. Cambridge University Press, New York (2007)
62. Radons, G., Zienert, A.: Nonlinear dynamics of complex hysteretic systems: oscillator in a
magnetic field. Eur. Phys. J. Spec. Top. 222(7), 16751684 (2013)
63. Rosenstein, M.T., Collins, J.J., De Luca, C.J.: A practical method for calculating largest
Lyapunov exponents from small data sets. Physica D 65, 117134 (1993)
64. Rssler, O.: An equation for continuous chaos. Phys. Lett. A 57(5), 397398 (1976)
65. Swathy, P.S., Thamilmaran, K.: Dynamics of SC-CNN based variant of MLC circuit: an
experimental study. Int. J. Bifurcat. Chaos 24(02), 1430008 (2014)
66. Takens, F.: Detecting strange attractors in turbulence. In: Dynamical Systems and Turbulence,
Warwick 1980 (Coventry 1979/1980). Lecture Notes in Mathematics, vol. 898, pp. 366381.
Springer, Berlin (1981)
67. Tsai, T.-L., Dawes, J.H.: Dynamics near a periodically-perturbed robust heteroclinic cycle.
Physica D 262, 1434 (2013)
68. Webel, K.: Chaos in German stock returns - new evidence from the 01 test. Econ. Lett. 115(3),
487489 (2012)
7 The 0-1 Test for Chaos: A Review 247

69. Xin, B., Li, Y.: 0-1 test for chaos in a fractional order financial system with investment
incentive. Abstr. Appl. Anal. 2013, 876298 (2013)
70. Xin, B., Zhang, J.: Finite-time stabilizing a fractional-order chaotic financial system with
market confidence. Nonlinear Dyn. 79(2), 13991409 (2014)
71. Yuan, L., Yang, Q., Zeng, C.: Chaos detection and parameter identification in fractional-order
chaotic systems with delay. Nonlinear Dyn. 73(12), 439448 (2013)
72. Zachilas, L., Psarianos, I.N.: Examining the chaotic behavior in dynamical systems by means
of the 01 test. J. Appl. Math. 2012, 681296 (2012)
73. Zaslavskii, G.M., Sagdeev, R.Z., Usikov, D.A., Chernikov, A.A., Sagdeev, A.R.: Chaos and
Quasi-Regular Patterns. Cambridge University Press, Cambridge (1992)
74. Zweimller, R.: Stable limits for probability preserving maps with indifferent fixed points.
Stoch. Dyn. 3, 8399 (2003)
Chapter 8
Prediction of Complex Dynamics: Who Cares
About Chaos?

Stefan Siegert and Holger Kantz

Abstract We compile knowledge on limitations to prediction of the time evolution

of complex systems. Although such systems are typically highly chaotic, the inverse
of the maximal Lyapunov exponent, the Lyapunov time, is not the time scale beyond
which predictions fail. Instead, as the example of weather forecasting will show,
predictions can be successful on lead times which are several orders of magnitude
longer. We analyze the reasons which prevent errors from growing exponentially
fast with a rate related to the maximal Lyapunov exponent. Moreover, we advocate
that standard practices from weather forecasting should be transferred to other fields
of complex systems predictions, which includes a statement about the uncertainty
related to the actual prediction and a performance measure on past predictions so
that a decision maker can assess the potential quality of a forecasting scheme.

8.1 What Do We Want to Predict?

Throughout human history we find reports about attempts to predict the future. The
targets of prediction were typically quite complex, such as future personal fate, the
outcome of a war, or about the richness of harvest. Most ancient prediction methods
(e.g., oracles of Greek and Roman antiquity) are nowadays not any more considered
as scientifically sound. The skeptic, however, might think similarly about, e.g.,
the statements of economic research institutes which regularly publish predictions,
precise up to tenths of percentages, of economic growth or tax volume change
without providing any estimates of prediction uncertainty and without explicitly
mentioning the scenarios for which such predictions were made [1].

S. Siegert ()
College of Engineering, Mathematics and Physical Sciences, University of Exeter, Laver
Building, North Park Road, Exeter EX4 4QE, UK
e-mail: [email protected]
H. Kantz
Max Planck Institute for the Physics of Complex Systems, Nthnitzer Str. 44, 01187 Dresden,
Germany
e-mail: [email protected]

Springer-Verlag Berlin Heidelberg 2016 249

Ch. Skokos et al. (eds.), Chaos Detection and Predictability, Lecture Notes
in Physics 915, DOI 10.1007/978-3-662-48410-4_8
250 S. Siegert and H. Kantz

It is, however, fact that a large part of our daily decisions are based on or at least
influenced by predictions. It is another fact that most of these predictions concern
very complex systems where our understanding is limited. In this article, we want to
discuss the issue of predictions and predictability of complex systems, analyzing the
sources of prediction errors, suggesting optimal prediction strategies, and arguing
for a fair presentation of not only the prediction result, but also of its uncertainty.
We will use weather forecasts as our reference, which allows us to illustrate most
of our statements. Weather forecasts are the most sophisticated and most developed
forecasts known to us, based on profound physical understanding of the system and
its dynamics, on an excellent network of observations of the current state, and on a
huge number of people and computers generating the daily forecasts and further
improving the forecast quality. Also, weather forecasters have done pioneering
work in assessing the uncertainty of their forecasts (e.g., by ensemble forecasts
[2]), in developing scoring schemes (e.g., [3, 4]), and improving predictions by
postprocessing model output and by blending new measurements and past forecasts
into new initial conditions (data assimilation).
Weather forecasts are omnipresent. Actually, weather forecasting has become a
considerable business. Many sectors of economy, not only agriculture and traffic, but
also renewable energy production and energy consumption are strongly dependent
on weather and therefore benefit from accurate predictions [5]. Even if for most of
us it would not be a real problem to be mislead by an inaccurate forecast, we are
aware of, and often like to complain about lack of forecast accuracy. Considering
huge effort spent for weather prediction, in terms of persons, measurement stations,
computer power, we conclude that forecasting the weather is very difficult. The most
evident reason for this seems to be the underlying chaoticity or the turbulent nature
of air flow.
Weather hence is a typical complex phenomenon, and its complexity has several
aspects. First of all, as said, the global hydrodynamic transport equations for air, the
Navier Stokes equations, yield turbulent solutions with positive Lyapunov exponents
and hence with sensitive dependence on initial conditions. Secondly, turbulence is a
multi-scale problem, related to the fact that the atmosphere is a continuum system
in 3-d space, with a very small cut-off scale where molecular dissipation sets in.
Thirdly, beyond that, weather is a consequence of the interaction of many different
physical (and chemical and biological) processes which take place on very different
time and length scales. On the other hand, weather is a natural system of which we
believe that in principle we can set up model equations for all these sub-processes.
Economy and finance, to mention another field where predictions are highly
desirable and part of the daily business, suffers from the lack of first-principles
models and from the lack of observation data to assess the current state of the
system: many relevant economic data are acquired with some time lag (such as
unemployment rates, tax revenue, economic growth), and many relevant data are
not publicly available. But in addition, these systems are controlled by humans.
This has two striking consequences: Firstly, humans can introduce innovations and
hence are able to change the rule of the game by their knowledge and activities. This
means that model equations might become invalid in the course of time, and even
8 Prediction of Complex Dynamics: Who Cares About Chaos? 251

new degrees of freedom might be introduced. Secondly, humans are responding

to forecasts. Whereas the weather does its thing regardless of what we predict
about it, economy and finance will react to forecasts. This makes it hard to verify
forecasts, and even harder to control the system. Nonetheless, it is plausible that
also economy shares relevant aspects of complexity with weather: The sensitive
dependence on tiny details in initial conditions, the existence of micro- and macro-
scales, the interaction of subsystems.
In the first part of this article, where we will be concerned with model based
forecast, we clearly have to require the existence of reasonable mathematical
models of the phenomenon to forecast. In the second part, where we discuss data
based forecasts, more general situations of complex systems are addressed. For the
atmosphere, chaos is considered to be one of the main limitation of predictability
[6]: Chaos is defined by the mixing property and hence is apparent through temporal
irregularity of the solutions. The lack of predictability is evident from the linear
instability with respect to tiny errors in initial conditions: as it is well known,
such errors can grow exponentially fast in time, if there is at least one positive
Lyapunov exponent. This exponential error growth is so fast that it always will win
over an improved precision by which the initial condition is determined. For long
and inspired by Lorenz pioneering work [6], this was considered to be the main
limitation for accurate predictions in complex systems.
When we speak of complex systems in this article, we mean really complex
systems such as the atmosphere, large economic or social systems, or technical
systems such as the power grid or telecommunication networks. But also biological
systems such as ecosystem dynamics and the irregular fluctuations of physiological
parameters such as blood pressure or heart rate have been discussed in the context
of chaos and have been subject to prediction attempts (e.g., [7]). Such systems are
not only chaotic, but in addition they have a huge dimensional phase space and
are coupled to some environment, which often cannot be explicitly modelled or
observed. They are composed of different sub-systems, which introduces a strong
inhomogeneity among the different degrees of freedom, and they typically also act
on a wide range of different time and length scales. All this together introduces
types of instability and limitations to successful predictions which go far beyond
the consequences of chaos.
In this article, we will pick up arguments showing that the issue of whether or not
such a complex system is chaotic, and in particular the value of its largest Lyapunov
exponent, is not really relevant for the accuracy of many types of predictions: Both
the difficulties in forecasting and the methods to perform good forecasts do not
really care about linear instability. We will present both, obstacles against good
forecasts, and the current answer to this, where we will distinguish between data
based forecasts and model based forecasts. Most of our considerations are not at all
newan accessible and more comprehensive discussion of some of them can be
found in the book by L. Smith [8].
252 S. Siegert and H. Kantz

8.2 Sources of Unpredictability

8.2.1 Chaos

As is well known, chaotic dynamics suffers from sensitive dependence on initial

conditions. This creates an evident source of unpredictability due to the exponential
divergence of nearby trajectories: Given a chaotic dynamical system f and two initial
conditions x0 and y0 with a very small Euclidean distance, jjx0  y0 jj D  1, we
consider the two trajectories emerging from x0 and y0 as x.t/ and y.t/. If a system has
at least one positive Lyapunov exponent max > 0, then jjx.t/y.t/jj ' exp.max t/
for large t and for Lebesgue-almost all x0 and y0 with jjx0  y0 jj D  1. (Remark:
In order to be chaotic, a system with continuous time must possess an at least 3-
dimensional phase space, i.e., x and y are vectors.) There is, however, a transient
time during which the distance between x.t/ and y.t/ might grow faster or slower
than the asymptotic exponential growth, depending on the direction of the initial
perturbation and on the structure of the local Jacobians. The exponential growth is
a limiting behavior in the double and non-exchangeable limit jj jj ! 0 and t ! 1.
It is evident that even with perfect model equations the prediction of every
natural system will suffer from exponential error growth: We can never assume
to know the true state of the observed system with arbitrary precision, so that
our initial condition inserted into the model equations will always deviate from
the true state. Whatever tiny this errors is, after some time the exponential factor
exp.max t/ will be large enough to make it visible and to eventually make our
forecast trajectory diverge fully from the true one. Actually, the time 1=max is often
called the Lyapunov time and determines the order of magnitude of the time when
this error growth is dominant. If we assume an initial error of size , then it will be of
size unity after  ln times the Lyapunov time. This illustrates that when increasing
the precision of the knowledge of the initial condition by orders of magnitudes, the
time till the error reaches order one increases only linearly [9].
In weather forecasting, the error in initial conditions is not a simple consequence
of measurement inaccuracies. It is much more determined by the lack of suitable
observations: Whereas a weather model, let it be global or regional, due to limita-
tions of computer power currently represents the continuum of the atmosphere by
up to 1010 degrees of freedom1 located on a 3-dimensional grid in space. The World
Meteorological Organization lists only of the order of 105 measurement stations
which provide direct observations, in addition there are satellite data which are often
projections along the vertical axis. From these observed data, a model state has to
be constructed which has a much higher dimension. The logic behind this is, see
below, the belief that the true attractor dimension, and hence the effective number
of degrees of freedom, is much lower than 1010 , so that much less observations

1
For example, the ECMWF model in the current 40th cycle is a T1279 model with 91 vertical
layers.
8 Prediction of Complex Dynamics: Who Cares About Chaos? 253

might suffice in order to determine where on the attractor the new initial condition
should be located. How does one constrain the initial condition to the attractor?
This is implicitly done by techniques summarized as data assimilation [10]: Data
assimilation tries to merge the information obtained from the new, most recent set
of measurement data with the model trajectory of the past few days in order to
create a model vector which is both close to the model attractor and is close to the
observed data. Despite great successes, this inherent lack of observation data (and
its very inhomogeneous distribution across the Earth) causes considerable errors of
initial conditions, which, since they are constructed using the model dynamics, also
are affected by model errors.
An additional problem lies exactly in these model equations: We can not assume
that they are perfect. In the best case, there is simply a small error in parameter
values contained in these equations. Model errors would first (say, after one time
step) introduce a tiny error in our forecast even if the initial condition were perfectly
known. Then, this deviation between forecast trajectory and true trajectory will grow
again exponentially due to chaos irrespectively of the model error. Hence, the effect
of small model errors is similar to the effect of errors in the initial conditions, if the
motion is chaotic.
There is an additional and potentially very severe problem with model errors:
Lack of structural stability. In physicists words, the system dynamics is structurally
stable, if a small change of model parameters can be compensated by a coordinate
transform which should be close to identity. That is, the properties of the attractor
such as invariant measure, Lyapunov exponent, KS-entropy do not change. Almost
all studied low-dimensional models are non-hyperbolic, which means that they are
not structurally stable. Hence, a tiny perturbation of some model parameter might
result in a very different dynamical behavior, e.g., the wrong model might possess
an attracting periodic orbit, whereas the true system is chaotic. A prominent
example for such behavior is the logistic equation and the Hnon map, which have
tiny periodic windows all-over their parameter space. There is common believe (and
this is formalized in the chaotic hypothesis [11]), that high dimensional systems,
even if not structurally stable in the strict mathematical sense, behave practically like
hyperbolic systems. And indeed, free running GCMs seem not to lock into periodic
behavior when parameters are slightly detuned.
Arent the last few paragraphs a demonstration of how relevant chaos is in
limiting predictions of complex systems? Yes, they are, and the reasoning above
has been stressed many times during the past 50 years.
However, when we aim at predictions of really complex systems, i.e., systems
much more complex than a simple chaotic few-degrees of freedom system, then
several additional aspects arise which might be more relevant to forecasts than the
systems chaoticity, and these are essentially all consequences of the phase space
being very high dimensional.
254 S. Siegert and H. Kantz

8.3 Beyond Chaos

In this section we will discuss why very often chaos, although present in the system
for which one intends to make forecasts, is not the limiting aspect. Most of these
arguments are not new, but we are not aware that they have been put together before
as we do it here.

8.3.1 Outside the Linear Regime

The exponential divergence and thereby exponential error growth is only present
when errors are tiny, in the so called linear regime. We argue that in many real
world prediction tasks, one works outside this regime, and we refer here again to
our favorite example, weather forecasts.
In Fig. 8.1 we show root mean squared (rms) prediction errors of temperature
forecasts for Hanover, Germany, for four different forecast schemes, as a function
of lead time. The benchmark of forecasting is the so called climatology c.d/, d D
1; : : : ; 365. Climatology is the average value at a given day d in the year, averaging
over many such days d from past years. As a function of the date d, climatology
c.d/ shows the seasonal cycle. It is the systematic part of temperatures, whereas
the deviation of the actual temperature on some day, T.d/, from c.d/ is due to the
specific weather at the particular time. Actually, T.d/  c.d/ is called temperature

Fig. 8.1 Forecast error of temperature as a function of lead time: Shown are the root mean squared
prediction errors of the three forecast schemes persistence (future temperature is identical to
todays temperature), climatology (future temperature is the multi-annual mean temperature for
that day of the year), and a model forecast obtained from the NCEP/NCAR re-forecast project,
for temperature predictions in Hanover (Germany) averaged over the years 19812010 with and
without re-calibration of the model output
8 Prediction of Complex Dynamics: Who Cares About Chaos? 255

anomaly and its prediction is the nontrivial part of a temperature forecast. Therefore,
the rms error of climatology represents the standard deviation of the distribution of
true temperature values around the long year average, averaged over every day in
the year. Roughly, this standard deviation is constant over the year for the Hanover
station, so that the rms error of climatology represents the uncertainty of what we
need to forecast: Knowing the date, one knows that the true temperature will be
distributed around c.d/ with a distribution whose standard deviation is, as we see
from Fig. 8.1, about 4 K. Actually, this distribution is in rather good approximation
symmetric and Gaussian (not shown here, see [12]). In other words, if we subtract
the seasonal cycle, the task of weather forecasting is to predict which value from
this approximately Gaussian distribution will be realized. Its standard deviation
therefore sets the scale for the dynamical range of the quantity to be predicted. This
range is much smaller than the range of the temperatures proper, since the latter can
be seen as the superposition of a specific weather related value over the seasonal
cycle.
The rms error of the calibrated weather model, i.e., a weather model output with
adjustments to compensate for systematic errors observed in the past, is about 2 K,
i.e., one half of the rms error of climatology. In the context of time series analysis
[13], one normalizes the rms prediction error of a forecast scheme by the standard
deviation of the quantity to be predicted, and in this sense the average error of the re-
calibrated weather model output is about 50 %. Is that good or bad? That is not our
concern here, although we should say that there are better models: We use data from
a global weather model from a reanalysis/reforecast project [14], which provides us
with the excellent statistics of more than 40 years of forecast/analysis pairs with
the very same model. The local prediction suffers from the fact that this model
is not perfectly up to date, and that it is a global model. State-of-the-art regional
weather models predict temperature anomalies with about 25 % error. Our concern
is that if at the shortest lead time of 24 h the rms error is already 50 % (or 25 %) of
the standard deviation of the signal, then we are definitely out of the linear regime
of the underlying dynamics. Trajectories are such far away from the assumed true
trajectory that we are already closer to the saturation regime than we are to the
exponential divergence regime. Hence, chaos is not the main cause of further error
growth when we increase lead time up to 10 days.
We can illustrate the fact that typical distances between trajectories are so large
that exponential divergence is not any more relevant in two more ways. Both are
related to the fact that for state-of-the-art weather predictions, ensembles forecasts
are performed [2]. This means that a deterministic model of the atmosphere is run
with several slightly different initial conditions. Why this is done will be discussed
more thoroughly in Sect. 8.4. Here we want to stress that although the number of
ensemble members is limited to 1050 by the demand of saving computation time
to do real time predictions, such ensembles often include for every perturbed initial
condition x C u the perturbation into the opposite direction, x  u (here, x is
the high-dimensional state vector, u is a unit vector in state space, and is the
amplitude of the perturbation). If errors stayed in the linear regime, then one of
the two trajectories emerging from these two initial conditions would be redundant,
256 S. Siegert and H. Kantz

and in particular their average would exactly yield the unperturbed solution. That
they are both included in the ensemble shows that generally they do not produce
redundant information, simply because they soon leave the linear regime around the
unperturbed solution.
Error growth as a function of error magnitude has been characterized by the
finite size Lyapunov exponent of Boffetta et al. [15]. There it was shown for much
less complex systems than weather that scale dependent measures of instability can
be employed to characterize the growth of finite errors and that indeed such error
growth might depend strongly on the magnitude of the error, as opposed to the
mathematical limit of infinitesimal errors.
That ensemble members do not diverge from each other exponentially fast can
also be visualized directly. In Fig. 8.2 we show 13 traces of temperature over a lead
time of 10 days, together with the observed value. These traces are the temperature
values of 13 different initial conditions of a weather model, and on visual inspection
there is no evidence that they diverge exponentially from each other.
One could argue that local stretching rates or finite time Lyapunov exponents
(FTLE) [16] do fluctuate along a trajectory, as a consequence of the fact that
instability varies over the phase space. Indeed, this is a relevant aspect also for
forecasting: There are situations, where a forecast can be more precise, and others
where it is less precise, as a function of the current state of the system. However, if
one averages FTLEs over many points in phase space, their average is usually close
to the maximum Lyapunov exponent (they are identical to the maximum Lyapunov
exponent only if the most expanding direction is not a function of the phase space
point), whereas the average over the spread of many ensembles does not yield an

Fig. 8.2 Sketch of an ensemble forecast: Shown are temperature forecasts for the city of Hanover,
initialized at December 31, 2007, up to 10 days into the future. The bold black line represents
the actual measurements, whereas the 13 thin lines are model forecasts with 13 slightly different
initial conditions, as they are produced by the NCEP/NCAR ensemble forecast within the
reanalysis/reforecast project [14]. Notice that the divergence of nearby trajectories is in no stage
exponential in time but instead very irregular
8 Prediction of Complex Dynamics: Who Cares About Chaos? 257

estimate of the maximal exponent, simply because they are not reflecting the growth
of infinitesimal errors.
In summary, the uncertainty of the initial condition is so big that two initial
conditions which are, within these uncertainties, both candidates for the true
atmospheric state, depart from each other not exponentially fast but with some
very complicated behavior which is slower than exponential. Without any doubt,
this complicated behavior has its origin in complicated, nonlinear dynamics, or
even in the aperiodic nature of chaotic flows. Nonetheless, exponential growth of
infinitesimal perturbations due to chaos is irrelevant, and the error growth in a
complex nonlinear stochastic process would look very much alike.

8.3.2 Dilute Phase Space

For the simulation of a PDE system on a computer, one needs some coarse graining,
i.e., one projects the equations of motion from an infinite dimensional phase space
to a finite dimensional one. It has been shown in many studies [17, 18] that the
attractor dimension of such approximations as, e.g., represented by the Kaplan
Yorke dimension, approach a constant when spatial resolution of the discrete system
is improved beyond a certain level. In other words, the positive Lyapunov exponents
remain essentially unchanged and only negative ones are added, if one increases the
phase space dimension by better spatial resolution beyond a physically motivated
minimum. A typical complex system such as the dynamics of the atmosphere
has an attractor dimension which is indeed large, but finite, as to be compared
to the in principle infinite phase space dimension (see, e.g., [19]). The study
in [20] might serve as an example of a quantitative result: a quasi-geostrophic
model with 1449 degrees of freedom is studied, its KaplanYorke dimension is
about 300. It is expected that this discrepancy between phase space dimension and
attractor dimension is much larger for more realistic models. As mentioned before,
operational weather models today have of the order of 1010 degrees of freedom.
Nobody has been able to compute the KaplanYorke dimension of such a model,
but it seems to be consensus that it is by several orders of magnitude smaller than
1010 .
In order to illustrate the effect of such sparseness of the unstable phase space
directions among all possible directions, we introduce here a very simple model
in discrete space and time (coupled map lattice) with the following properties: Its
phase space dimension can be adjusted as the parameter N of coupled maps, it is
chaotic, its maximal Lyapunov exponent is independent of N, and so is its very
small attractor dimension. The model consists of one chaotic map (Ulam map) at
lattice position k D 1, and of N  1 linear maps (most of them being stable, some
are unstable) at the other lattice points. We introduce a kind of convective transport
by a skew coupling. The dynamics hence reads as follows:

x1;nC1 D 1  2x1;n
xk;nC1 D ak xk1;n for k D 2; : : : ; N; (8.1)
258 S. Siegert and H. Kantz

where the parameters ak are uniformly distributed random numbers from the interval
0:87; 1:07. The maximal Lyapunov exponent is 1 D ln 2 from the chaotic driving
at lattice site k D 1. Along with the trajectory in phase space, we iterate tangent
space vectors by the linearized dynamics. We start a perturbation at n D 1,
i.e., far in the past, which, in the long time limit, provides a numerical estimate of
the maximal Lyapunov exponent, but in addition yields the instantaneous stretching
rate from one time step to the next. We additionally start ensembles of random
perturbation vectors at different points along the reference trajectory. Initially,
these random perturbations grow much more slowly than with the instantaneous
stretching rate. Notice that this is in contrast to observations of super-exponential
growth of such transients in low dimensional systems [20]. Only after a transient
are these random perturbations dominated by the most unstable direction in tangent
space. Their growth is then governed by the same instantaneous stretching rate as
given by the perturbation started far in the past. In Fig. 8.3 we illustrate this: We
show the average of the logarithm of the local stretching rates as a function of the
number of time steps after initialization of an ensemble of perturbations, where the
average is taken over many ensemble members and over very many points along the
chaotic trajectory. As expected, after several time steps, it converges to the maximal
Lyapunov exponent. However, and this is the main result of this exercise, during the
first few iterations, the growth is much slower and can even be negative (shrinking
of the perturbation vectors). Moreover, the transient time needed till the growth rate
approaches its asymptotic value (for this system: ln 2) increases with system size N.
Hence this is an illustration of what is commonly known from ensemble weather

0.8
average local stretching exponent

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0 maximal Lyapunov exponent
-0.1
0 2 4 6 8 10 12 14 16
time steps after initialization

Fig. 8.3 Averaged over ensemble members and many starting points for perturbations, the curves
represent the instantaneous stretching rate of infinitesimal randomly initialized perturbations as a
function of iterations past initialization for the system (8.1). Asymptotically, the growth rate coin-
cides with the maximum Lyapunov exponent max D ln 2, but during the very first iterations there
might even be shrinkage. The curves represent system sizes N D 10; 100; 1000; 10;000; 100;000
from left to right. Hence, the effect is the more pronounced, the higher the dimension of phase
space
8 Prediction of Complex Dynamics: Who Cares About Chaos? 259

predictions: If one wants to have perturbations growing from the very first moment,
one has to chose the initial perturbation vectors in a very special way, namely they
should be aligned with the most unstable directions in phase space.
This system, Eq. (8.1) is, admittedly, very specific, even though we argue above
that it shares essential features with PDE systems. However, it has been observed
in many studies on specific model PDEs or coupled ODEs that Lyapunov vectors
localize in real space [21, 22], i.e., that at a given time, only a few components
of a Lyapunov vector are considerably different from zero. Hence, the projection
of a vector with random components onto such a localized vector is small, more
precisely, is the smaller the larger the phase space dimension. A simple estimate
which considers the normalization of the vector in N-dimensional space shows p
that on average, the projection onto a localized vector has a magnitude of 1= N,
which means that it takes a time t D ln N=2max till this perturbation reaches unity.
This ln N behavior can also be observed in Fig. 8.3: Every order of magnitude by
which we increase N leads to a constant shift of the curves to the right. This is
plausible since in our system Eq. (8.1) the Lyapunov vector of the maximal exponent
is trivially localized, namely it is the vector .1; 0; : : : ; 0/.
Back to the issue of chaos: In a complex system with very many degrees of
freedom, a random error on the initial condition will not grow exponentially fast
during the first few time steps. Only asymptotically, infinitesimal errors will be
governed by the maximal Lyapunov exponent. Therefore, if we assume that a typical
initial condition used for forecasting is a random perturbation of the unknown true
state of the system, the prediction error should grow less fast than expected if the
maximum Lyapunov exponent is known.

8.3.3 Multi-Scale Dynamics

The other reason why chaos is not the true limitation of weather predictions lies
in the multi-scale aspect of weather. The most unstable structures of atmospheric
dynamics are eddies of boundary layer turbulence, i.e., the gusty motion of air in
interaction with the rough surface of the Earth, driven by shear and by the heating
from the ground. Their instability determines the value of the largest Lyapunov
exponent of this dynamical system atmosphere, and there are estimates that it is in
the range of 1/s to 100/s [23]. On the other hand, our weather instead is determined
by large scale atmospheric conditions such as low and high pressure systems, whose
lifetime is of the order of several days and which typically move not faster than about
a few hundred kilometers per day. Hence, these structures are much more stable,
such that weather forecasts are meaningful on lead times which are by orders of
magnitude larger than the Lyapunov time of atmospheric dynamics. Indeed, these
features of multi-scale dynamics in weather were first discussed in a quantitative
way by Lorenz [24].
It is worthwhile to reformulate this issue: The maximal Lyapunov exponent is a
well defined mathematical concept, involving two limits. The initial perturbation
260 S. Siegert and H. Kantz

should be infinitesimal, and the growth of errors is studied in the infinite time
limit. By this procedure, the linear subspace with maximal linear instability is
identified and characterized. In a chaotic system with a few degrees of freedom,
this maximally unstable subspace will after short time dominate the growth of
perturbations. In a very high dimensional but homogeneous system such as the
coupled map lattice Eq.(8.1), the transient time till a tiny random perturbation will
grow with the maximal Lyapunov exponent might be quite long as shown before.
The dynamics of the atmosphere is, in addition to living in an infinite dimensional
phase space, spatially inhomogeneous, in particular in vertical direction. Hence,
strongly chaotic motion of many small scale degrees of freedom seems to have only
a weak effect on the large scale dynamics. It has not been investigated so far but it is
plausible that Lyapunov vectors corresponding to the largest exponents are localized
close to the ground. Hence, the multi-scale nature of atmospheric processes is
another reason why the Lyapunov time is not a limitation to the prediction of those
quantities in which we are interested in weather forecasts.

8.3.4 Model Errors and Noise

The more complex a natural system is, the less probable is it that the model
equations used for its mathematical description are perfect. For models of the
atmosphere it is very well known that they are imperfect in many respects. As
in every model, relevant parameters in the model equations are known only with
finite accuracy, and some of them might be even poorly known. In addition, many
parameters of the model might vary slowly in time. In weather models, this is,
e.g., change of albedo (the reflectivity of the Earths surface) and of the hydrology
(water storage and evaporation of the soil) due to change of land use. Hence, model
parameters need regular updates, which are not always available.
A severe shortcoming of weather models is the fact that due to lack of
computational resources one is forced to represent the 3-dimensional continuum
of the atmosphere by a set of grid-points in space. None of the physical processes
living on scales which are smaller than the grid spacing can be resolved in such a
model. Grid spacings of operational models vary between about 1 km to hundreds
of kilometers, depending on whether the model represents a small region or the
whole globe. That is, even models with the currently achievable best resolution
are hardly able to represent individual clouds, whereas the global models will even
skip large orographic structures such as narrow mountain ranges and islands. All
physical processes on the smaller scales have to be parameterized. This means
that they are not represented by their own dynamical variables, but that instead
their action on the resolved variables is approximated as a function of the latter
themselves (also known as closure ). Hence, the model variable at a grid-point lacks
the interaction with potentially fast fluctuating quantities; only their average effect
can be captured by deterministic parametrization. The missing fluctuations can, in
the simplest approximation, be understood as some kind of noise which acts on
8 Prediction of Complex Dynamics: Who Cares About Chaos? 261

the true, natural variable at a grid point and which is missing in the model. It is,
however, evident that introducing a noise term in the model equations, as, e.g., by
stochastic parametrization, would not solve the prediction problem: In order to make
accurate predictions, the model noise should have exactly the same realization as the
natural fluctuations have. An illustration of this statement is the prediction of a small
particle in water: We know that this undergoes Brownian motion. But every single
simulated Brownian path would be a bad prediction for the observed Brownian path,
since in the prediction we are unable to guess the correct realization of the noise
which would be needed to generate the observed path. The best prediction with least
rms error would be to predict the particle being at rest. Hence, stochastic terms in
atmospheric models are useful for long term modelling (one might get much better
statistics of all sorts of events), but every single stochastically driven trajectory has
no evident predictive power.
However, even if the individual prediction cannot become better by noise, one
can use the concept of stochastic driving to explore the uncertainty of a prediction
induced by unresolved degrees of freedom. The benefit from stochastic terms
modelling the action of unresolved variables is in ensembles of forecasts [2, 25]:
If one has a good estimate of the magnitude and the temporal correlations of these
noise terms, one can use an ensemble of stochastically driven trajectories where
each ensemble member has a different noise realization, and thereby construct a
forecast distribution. This distribution then reflects the uncertainty of the forecasts
due to unresolved degrees of freedom.
It is inevitable that models resolve only a part of all relevant degrees of freedom,
so that the unresolved degrees act as some kind of noise on the resolved ones. This
is true for many fields of science, such as weather, economy, traffic, ecology, human
health, or biology. In addition, in many cases there are influences which cannot
be modelled deterministically. In economics as well as in traffic, one might model
the unpredictability of human behavior stochastically. Also the impact of natural
disasters on both could be considered as a stochastic processes. These noises as
well as those from unresolved degrees of freedom further limit predictability.
Actually, it is the natural tendency of researchers to make models as realistic
as possible, resolving as many degrees of freedom as can be handled (from the
numerical point of view). However, simply making one part of a model more
realistic does not always improve predictions, as new dynamical instabilities might
be introduced. Although we currently do not have a good concept for this problem,
with our experience in dynamics we can state that every refinement of an existing
model needs to be checked for its improved performance, one must not take that
improvement for granted.
The conclusion from these considerations is that ignored degrees of freedom and
noises in the real system cause additional prediction errors. These will grow slower
than exponentially fast in time; for a pure random walk approximated by a fixed
point they would grow like square root of time.
262 S. Siegert and H. Kantz

8.4 Best Practice Predictions

Given these many uncertainties, one should follow some best practice in predictions.
As we complained earlier in this text, it is bad practice that when forecast results
are communicated to the public, including political decision makers, usually only
numbers are given as if they were unquestionable and precise. Decision makers
have to use these numbers without any information about their reliability. This is
intolerable and also unnecessary, as the example of weather forecasting shows.
So firstly, every individual prediction should not only report the best guess of
the value of an entity, but also the best guess for the uncertainty of this value.
In an ideal case, one could issue a probability distribution of which value is how
probable, the forecast probability. This is the probability for the future to attain a
certain value, given our rough knowledge about the system and its current state. So
even in perfectly deterministic settings, such a forecast probability is different from
a -distribution. For the purpose of giving simplified messages to the end user, an
error bar representing the standard deviation of this probability might be sufficient.
In model based forecasts, the forecast probability can be constructed from
ensemble forecasts, as it was outlined in Sect. 8.3. The ensemble of different forecast
values then yields, after applying a technique called kernel dressing, a forecast
distribution: A probability distribution of the prediction target, conditioned on our
lack of knowledge about the system. Kernel dressing means that every forecast
value of an individual ensemble member is replaced (or dressed) by a probability
distribution which has its maximum at the predicted value. The shape and the width
of this kernel has to be chosen appropriately [26]. The forecast distribution is the
normalized sum of all of these probability distributions.
Model errors are also partially represented by this method, but can also partly
be removed by post-processing: If predictions are made with sufficiently high
frequency, one can study the prediction-observation pairs of past predictions in a
statistical way and correct, by re-calibration of the model output, for systematic
deviations: One desirable property of ensemble forecast is unbiasedness, i.e.
having a mean-zero prediction error. Another desirable property is a flat Talagrand
histogram [27]: If one considers the rank of the observation among an ensemble
of forecasts, this rank should be uniformly distributed, leading to a flat Talagrand
histogram. Flatness of the histogram is a necessary condition for the forecast to be
reliable [28]. A systematic bias of the ensemble which is reflected by an imbalance
of outliers to the larger and outliers to the smaller values (tilted diagram), can be
removed directly by a seasonal shift of the model forecasts. A systematic violation
of reliability, as indicated by a non-flat, [-shaped or \-shaped Talagrand histogram,
can partly be corrected by adjusting the kernel width in ensemble dressing, thereby
adjusting the ensemble spread. Clearly, from the physical point of view it would be
nicer to improve the model itself in order to avoid systematic forecast errors.
In situations where debatable values for certain model parameters have strong
impact on the prediction result, also these must be mentioned together with the
prediction. In climate projections, this is done by simulations for different scenarios
8 Prediction of Complex Dynamics: Who Cares About Chaos? 263

of greenhouse gas emissions. But also in every forecast of economic growth some
assumptions about, e.g., the oil price have to be made, and these assumptions must
be communicated to forecast users.
Finally, a decision maker who has competing forecasts at disposal wants to know
the performance of each forecaster in the past. Hence, generally accepted skill scores
for predictions such as the rms prediction error for deterministic predictions, or the
Brier score [3] for probabilistic predictions, should be computed and published by
the forecaster.
In particular in weather forecasting, where at least one prediction per day is
issued, every weather service could report, e.g., a score for its predictions as a
function of the past years. We are only aware of the ECMWF providing information
on performances [29], whereas many public weather services and commercial
weather platforms in the internet do not. But also repeated economic forecasts such
as of the expected growth of GNP or of tax revenue could be scored and would, after
a decade or more, present a nice picture of how well forecasts have been in the past.

8.5 Data Driven Forecasts

In this section, we want to tackle the issue who cares about chaos from the point
of view of data based forecasts. It is a common expectation that a mathematical
model of a complex system will usually outperform any purely data based prediction
scheme. The reason why this is usually indeed the case lies in the additional
information contained in the model equations, if they are constructed on the basis
of first principles, independent of a training data set. Nonetheless, a model based
prediction of a really complex system requires a huge effort. In weather prediction,
this is a whole industry developing the models, running a supercomputer, setting
up and operating an observation network. Instead, data based prediction requires
much less effort and relies essentially on a few input data sets in order to produce
a forecast, and often can be optimized by a single person on a small computer.
Moreover, there are many forecast issues where detailed mathematical models are
lacking, so that the only chance to predict the future is based on recordings of past
data.
In Fig. 8.1 we presented the performance of two data driven forecasts, which
were both not particularly successful: Climatology was called a prediction which
learns the seasonal cycle from many past observations and hence makes a prediction
which depends on the calendar day for which the prediction is made, but not on the
actual weather situation. It serves as a benchmark in Fig. 8.1. We also show the
performance of persistence. Persistence means that the system has the tendency to
remain in the same state as it is. Hence, a predictor exploiting persistence simply
predicts for the future the value of the last available observation. As we see, this is
not too bad 1 day ahead, but after 3 days this prediction is worse than climatology.
Although being a function of the current state (represented by the last observation),
264 S. Siegert and H. Kantz

the persistence predictor ignores dynamics. In the following we present a predictor

exclusively derived from data which depends on the state and includes dynamics.
Much more promising than persistence and climatology are data driven forecasts
which rely on a principle which in atmospheric sciences is called the Lorenz
method of analogues. It was introduced by Edward Lorenz [30] and suggests that
if we find an atmospheric state in the past which is similar to the present state, then
the future evolution of the atmosphere will follow approximately the evolution of
this similar past state. Hence, one can read off from the past how the future will
be. We should remark that this is actually how experience works, and that this was
the way of weather forecasting in times before numerical weather models. But here
we (as most surely also E. Lorenz) mean a much more detailed analogue than one
on synoptic scales, which is found by a computer comparing the actual situation
to a very large data base of past situations, and hence can also cope with sensitive
dependence on initial conditions.
This method has been specified in many different ways, and it is worth to
mention Farmer and Sidorowich [31] who formalized this for chaotic deterministic
dynamics. However, using the simplest version, called zeroth order by Casdagli
[32], is at the same time also the optimal predictor for stochastic processes, provided
one is able to reconstruct their Markov property. Therefore, the issue of a chaotic or
a stochastic origin of observed irregularities in data is not an issue. In the following
two subsections we will discuss this in more detail.

8.5.1 Data Driven Forecasts for Deterministic Dynamics

Determinism means that the initial condition uniquely determines the trajectory
which evolves from it. As a consequence of this, deterministic trajectories of the
same dynamics cannot intersect each other: If they intersect, then their further
evolution must be identical (because the intersection point must have a unique
future), and if the dynamics is time invertible, then also the past of the intersection
point must be unique, hence the two trajectories are identical for all times.
Data driven prediction for deterministic systems then consists of two parts: First,
prepare the data in such a way that they represent uniquely the state of the dynamical
system in the above sense. Second, find an algorithm which identifies the unique
future of a new initial condition on the basis of already observed pairs of initial
condition and future.
In practice, both parts can only be achieved approximately. For the identification
of the state vector, Takens time delay embedding [33] is a mathematically rigorous
concept for low dimensional systems. For high dimensional systems, it still serves as
a useful guideline, despite practical limitations. Given a time series of observations
equidistant in time, x1 ; : : : ; xN , and choosing a dimension m of the phase space to be
reconstructed, one defines delay vectors sn WD .xn ; xnd ; xn2d ; : : : ; xn.m1/d / with
the time lag d in an overlapping way. Hence, the observed time series x1 ; : : : ; xN is
represented by a sequence of N  md vectors sn . The embedding dimension m and
8 Prediction of Complex Dynamics: Who Cares About Chaos? 265

the time lag d are parameters by whose variation the prediction performance can be
optimized, for guidelines for their proper choice see, e.g., [13]. An independent test
to assure the suitability of the embedding parameters m and d consists of computing
the fraction of false nearest neighbors [34]: If two points in phase space are close
to each other (i.e., they are neighbors), then, due to continuity of the dynamics,
their near future under the dynamical evolution should be neighbors as well. If the
embedding dimension is too small, points which are quite distant from each other
in the true phase space might appear to be close due to projection. However, it is
improbable that their future values are also close under the very same projection.
Hence, the method of false nearest neighbors counts how many pairs of nearby
points sn , sn0 have future points snC1 , sn0 C1 which are not close, in order to detect
the effects of insufficient embedding dimension. If additional input variables exist,
a mixed multivariate/time delay vector might serve best. This has to be optimized
through minimization of the prediction error.
After one has decided on the form of the state vector, the algorithm for actual
prediction following the concept of analogues can be approached: Given a state
vector as input to the prediction scheme, one searches in a training set of historic
recordings for similar state vectors. Similar here means that one defines a measure
of distance in embedding space such as the Euclidean norm of the difference vector
and collects all states from the data base where this distance towards the new input
is smaller than a threshold . The smaller the more similar are these states, but
the lesser are found in the finite data base. Hence, is a parameter which balances
statistical robustness (many neighbors) with accuracy (good analogues) and has to
be adjusted a posteriori.
Since the state vectors are constructed from a time series, one can read, from this
time series, a future value of this observable at arbitrary lead time. Given the lead
time, we call the corresponding time series value the future of this state vector.
The prediction is then a suitable function of the futures of all the similar states.
In the simplest case, this function is simply the mean of these futures, which was
called 0th-order model in [32]. Instead, these futures can be seen as a sample
of a forecast distribution, similar to the ensemble from Sect. 8.4. Their empirical
standard deviation is a measure for the prediction error to be expected and quantifies
the uncertainty of this forecast, and by kernel dressing the set of futures can be
converted into a forecast distribution, so that this data driven forecast complies with
our requirements for state of the art forecasts in Sect. 8.4.
Such a local prediction is very flexible. The collection of all input/forecast
pairs can model an arbitrary nonlinear relationship between input and output. The
parameters which are embedding dimension m, time lag d, neighborhood diameter
, can be intuitively understood and therefore they can be adjusted using intuition.
The local approach has the drawback that it fails when data are sparse, because
it cannot extrapolate. The consequence of sparse data is that either the number of
neighbors is so small that the forecast suffers from statistical fluctuations, or that
one has to enlarge the size of the neighborhood to unsuitably large values. This
leads to using analogues which are not really analogues, so that this might replace
statistical errors by systematic errors.
266 S. Siegert and H. Kantz

8.5.2 Data Driven Forecasts for Stochastic Data

If we assume that the observed data stem from a stochastic process, a precise
prediction is impossible. Nonetheless, there exists an optimal predictor, and even
the prediction of the uncertainty of the forecast is possible. Let us assume that
the stochastic process is Markovian. This means that knowing the current state of
the process, xt , the conditional probability p.x0 ; t C tjxt ; t/ to find any other state
x0 at t C t is well defined through the transition probability from xt to x0 over
a small time step t. For a Markov process, the set of all transition probabilities
defines completely its dynamics (e.g., correlation functions) and its asymptotics
(e.g., stationary distribution) [35].
Hence, knowing the current state, p.x0 ; t C tjxt ; t/ gives us the full information
about what happens at time t C . In a stochastic setting, no further information
exists. If we intend to issue a sharp forecast value and we measure the forecast
quality by the rms prediction error, then the best forecast is the mean value of
p.x0 ; t C tjxt ; t/, which can be easily verified by minimization of the rms error
on a large sample of forecast trials. The uncertainty of this forecast can be readily
quantified by the standard deviation of this distribution. But as well, one can use
p.x0 ; t C tjxt ; t/ for a probabilistic forecast. Hence, knowing p.x0 ; t C tjxt ; t/, one
can make a forecast which complies with Sect. 8.4.
How can we determine p.x0 ; t C tjxt ; t/ from time series data? One possible
approach is again by the principle of analogues. First we have to find a good
approximation to the state vector xt . A good approach in practice is to use a
time delay embedding as in the deterministic case, i.e., in the case of a scalar
time series we use st D .xt ; xtt :xt2t ; : : :/, where the optimal value of the
embedding dimension is found by minimizing the prediction error with respect
to this parameter. Then, the principle of analogues assumes that all state vectors
which are similar to st will be subject to a very similar transition probability. I.e.,
p.x0 ; t C tjst ; t/  p.x0 ; t0 C tjst0 ; t0 /, if jst  st0 j  1. This includes the assumption
of stationarity, since st and st0 are not just different state vectors, but also observed
at different times. The analogue-concept together with the stationarity assumption
then yields the following result: The futures of all sufficiently close neighbors of st
form a random sample according to p.x0 ; t C tjst ; t/. The empirical mean of this
sample is an estimator of the mean of p.x0 ; t C tjst ; t/, and the sample variance
is an estimator of its variance. And as in the deterministic model based forecast,
where an ensemble was interpreted as a random sample of a forecast distribution,
we can exploit the sample of future values here in a probabilistic way. A variant of
this concept was first used in [36] for modelling, and the predictor was introduced
in [37].
Now back to the title of this article: When we perform data based forecasts,
the algorithm to perform the forecast is independent of whether we believe that
the underlying dynamics is deterministic or stochastic, since the forecast schemes
of Sects. 8.5.1 and 8.5.2 are identical, even if their motivation is quite different.
8 Prediction of Complex Dynamics: Who Cares About Chaos? 267

The formal reason behind this lies in the fact that a deterministic system is a
limiting case of a Markovian stochastic process: the transition probability is a Dirac-
distribution.

8.6 Conclusions

There are many sources for prediction errors, if we try to predict the future of
complex systems. Very different from low-dimensional systems, prediction errors
in real prediction tasks, i.e., starting from estimated model states, are usually not
dominated by the exponential growth of errors on the initial condition, even if the
dynamics of the system is chaotic. We discussed the sources for prediction errors
and also showed several reasons why the exponential error growth with a rate given
by the maximal Lyapunov exponent is usually not relevant. In weather forecasting,
current models possess predictive skill beyond the annual cycle up to about 1014
days into the future, which is by many orders of magnitude larger than the Lyapunov
time, which is determined by the instability of air flow on very small spatial scales.
Nonetheless, the nonlinearities of a complex system are essential for predictive skill:
among others, they are responsible for the fact that prediction uncertainties are state
dependent, i.e., that there are situations where predictions are more precise and other
situations where they are less precise, in the very same system.
An essential message of this contribution is also that we advocate a best practice
for predictions which should be routinely applied also outside the realm of weather
predictions. It is highly desirable that together with every prediction, a quantitative
assessment of its expected precision should be issued, alongside with results of
scoring schemes applied to past predictions. Only then a decision maker can assess
how reliable an individual forecast is and compare the average past performance of
different forecasters.

References

1. See, e.g., the web-site http://www.pwc.co.uk/economic-services/global-economy-watch/gew-

projections.jhtml of PricewaterhouseCoopers as an example of economic forecasts two years
into the future.
2. Toth, Z., Kalnay, E.: Ensemble forcasting at NMC: the generation of perturbations. Bull. Am.
Meteorol. Soc. 74, 2317 (1993)
3. Brier, G.: Verification of forecasts expressed in terms of probability. Mon. Weather Rev. 78, 1
(1950)
4. Christensen, H.M., Moroz, I.M., Palmer, T.N.: Evaluation of ensemble forecast uncertainty
using a new proper score: application to medium-range and seasonal forecasts, Q. J. R.
Meteorol. Soc. (2014). doi:10.1002/qj.2375
5. Mailier, P.J., Jolliffe, I.T., Stephenson, D.B.: Quality of weather forecasts. Online publication of
the Royal Meteorological Society (2006). Available through the publication lists of its authors,
e.g. http://empslocal.ex.ac.uk/people/staff/dbs202/publications/booksreports/mailier.pdf
268 S. Siegert and H. Kantz

6. Lorenz, E.N.: Deterministic nonperiodic flow. J. Atmos. Sci. 20, 130 (1963)
7. Glass, L., Belair, J., Milton, J.J., an der Heiden, U. (eds.): Dynamical Disease: Mathematical
Analysis of Human Illness. AIP, New York (1997)
8. Smith, L.A.: CHAOS A Very Short Introduction. Oxford University Press, Oxford (2007)
9. Shaw, R.: Strange Attractors, Chaotic Behavior, and Information Flow. Z. Naturforsch. 36a,
80112 (1981)
10. Kalnay, E.: Atmospheric Modeling, Data Assimilation, and Predictability. Cambridge Univer-
sity Press, Cambridge (2003)
11. Gallavotti, G., Cohen, E.G.D.: Dynamical ensembles in stationary states, J. Stat. Phys. 80,
931970 (1995)
12. Siegert, S., Brcker, J., Kantz, H.: Skill of data based predictions versus dynamical models -
case study on extreme temperature anomalies (2013). http://arxiv.org/abs/1312.4323
13. Kantz, H., Schreiber, T.: Nonlinear Time Series Analysis. Cambridge University Press,
Cambridge (2004)
14. Hamil, T.M., Whitaker, J.S., Mullen, S.L.: Reforecasts, an important data set for improving
weather predictions. Bull. Am. Meteorol. Soc. 87(1), 3346 (2005)
15. Boffetta, G., Giuliani, P., Paladin, G., Vulpiani, A.: An extension of the Lyapunov analysis for
the predictability problem. J. Atmos. Sci. 55, 34093416 (1998)
16. Paladin, G.: Intermittency and equilibrium measures in dynamical systems. In: Jullien, R.,
Peliti, L., Rammal, R., Boccara, N. (eds.) Universalities in Condensed Matter. Springer
Proceedings in Physics, vol. 32. Springer, Berlin, Heidelberg (1988)
17. Doering, C.R.: Exact Lyapunov dimension of the universal attractor for the complex Ginzburg-
Landau equation. Phys. Rev. Lett. 59, 2911 (1097)
18. Yang, H.-L., Radons, G.: Geometry of inertial manifolds probed via a Lyapunov projection
method. Phys. Rev. Lett. 108, 154101 (2012)
19. Lions, J.L., Manley, O.P., Teman, R., Wang, S.: Physical interpretation of the attractor
dimension for the primitive equations of atmospheric circulation. J. Atmos. Sci. 54, 1137
(1997)
20. Vannitsem, S., Nicolis, C.: Lyapunov vectors and error growth patterns in a T21L3 quasi-
geostrophic model. J. Atmos. Sci. 54, 347 (1997)
21. Pikovsky, A., Politi, A.: Dynamic localization of Lyapunov vectors in spacetime chaos.
Nonlinearity 11, 1049 (1998)
22. Yang, H.-L., Radons, G.: Universal features of hydrodynamic Lyapunov modes in extended
systems with continuous symmetries. Phys. Rev. Lett. 96, 074101 (2006)
23. Basu, S., Foufoula-Georgiou, E., Port-Angel, F.: Predictability of atmospheric boundary-layer
flows as a function of scale. Geophys. Res. Lett. 29, 2038 (2002)
24. Lorenz, E.N.: Predictability of a flow which possesses many scales of motion. Tellus 21, 289
(1969)
25. Palmer, T.N.: A nonlinear dynamical perspective on model error: a proposal for non-local
stochastic dynamic parametrization in weather and climate prediction models. Q. J. R.
Meteorol. Soc. 127, 279 (2001)
26. Silverman, B.W.: Density Estimation for Statistics and Data Analysis, 1st edn. Chapman and
Hall, London (1986)
27. Hamill, T.M.: Interpretation of rank histograms for verifying ensemble forecasts. Mon. Weather
Rev. 129, 550560 (2001)
28. Gneiting, T., Raftery, A.E.: Strictly proper scoring rules, prediction, and estimation. J. Am.
Stat. Assoc. 102, 359378 (2007); Brcker, J.: Reliability, sufficiency, and the decomposition
of proper scores. Q. J. R. Meteorol. Soc. 135, 15121519 (2009)
29. Buizza, R., Houtenkamer, P.L., Toth, Z., Pellerin, G., Wei, M., Zhu, Y.: A comparison of the
ECMWF, MSC, and NCEP global ensemble prediction systems. Mon. Weather Rev. 133, 1076
(2004)
30. Lorenz, E.N.: Atmospheric predictability as revealed by naturally occurring analogues. J.
Atmos. Sci. 26, 636 (1969)
8 Prediction of Complex Dynamics: Who Cares About Chaos? 269

31. Farmer, J.D., Sidorowich, J.J.: Predicting chaotic time series. Phys. Rev. Lett. 59, 845848
(1987)
32. Casdagli, M.: Nonlinear prediction of chaotic time series. Physica D 35, 335 (1989)
33. Takens, F.: Detecting Strange Attractors in Turbulence. Lecture Notes in Mathematics, vol.
898. Springer, Berlin (1981); Sauer, T., Yorke, J.A., Casdagli, M.: Embedology. J. Stat. Phys.
65, 579 (1991)
34. Kennel, M.B., Brown, R., Abarbanel, H.D.I.: Determining embedding dimension for phase-
space reconstruction using a geometrical construction. Phys. Rev. A 45, 3403 (1992)
35. Gardiner, C.W.: Handbook of Stochastic Methods. Springer, Berlin (2009)
36. Paparella, F., Provenzale, A., Smith, L.A., Taricco, C., Vio, R.: Local random analogue
prediction of nonlinear processes. Phys. Lett. A 235, 233 (1997)
37. Kantz, H., Holstein, D., Ragwitz, M., Vitanov, N.K.: Markov chain model for turbulent wind
speed data. Physica A 342, 315 (2004)
Erratum to: Chaos Detection and Predictability

Charalampos (Haris) Skokos, Georg A. Gottwald, Jacques Laskar

Erratum to:
Ch. Skokos et al. (eds.), Chaos Detection and Predictability,
Lecture Notes in Physics 915, DOI 10.1007/978-3-662-48410-4
Springer-Verlag Berlin Heidelberg 2016

The Editor name Charalampos Haris Skokos was not correct. It is now corrected
to Charalampos (Haris) Skokos throughout the book.
The publisher apologizes for having published an early draft edition of the preface.
This has been updated to the final version.

The updated original online version for this book can be found at
DOI 10.1007/978-3-662-48410-4

Ch. (Haris) Skokos

Department of Mathematics and Applied Mathematics, University of Cape Town, Rondebosch,
South Africa
e-mail: [email protected]
G.A. Gottwald
School of Mathematics and Statistics, University of Sydney, Sydney, Australia
J. Laskar
Observatoire de Paris, IMCCE, Paris, France

Springer-Verlag Berlin Heidelberg 2016 E1

Ch. Skokos et al. (eds.), Chaos Detection and Predictability, Lecture Notes
in Physics 915, DOI 10.1007/978-3-662-48410-4_9