Geometria Diferencial Bom

John H.
Lowenstein
Pseudochaotic Kicked Oscillators

Renormalization, Symbolic Dynamics,
and Transport
John H. Lowenstein
Pseudochaotic
Kicked Oscillators
Renormalization, Symbolic Dynamics,
and Transport
With 86 figures, 9 of them in color

Author
John H. Lowenstein
Department of Physics
New York University
2 Washington Place
New York, NY 10003, U.S.A.
Email: [email protected]
ISBN 978-7-04-032279-8
Higher Education Press, Beijing
ISBN 978-3-642-28153-2 ISBN 978-3-642-28154-9 H%RRN

Springer Heidelberg New York Dordrecht London
Library of Congress Control Number: 2012930642

c Higher Education Press, Beijing and Springer-Verlag Berlin Heidelberg 2012
This work is subject to copyright. All rights are reserved by the Publishers, 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. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered and
executed on a computer system, for exclusive use by the purchaser of the work. Duplication of this pub-
lication or parts thereof is permitted only under the provisions of the Copyright Law of the Publishers’
locations, in its current version, and permission for use must always be obtained from Springer. Permis-
sions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable
to prosecution under the respective Copyright Law.
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.
While the advice and information in this book are believed to be true and accurate at the date of publi-
cation, neither the authors nor the editors nor the publishers can accept any legal responsibility for any
errors or omissions that may be made. The publishers make no warranty, express or implied, with respect
to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

Dedicated to the memory of George Zaslavsky
Preface
The physics of a pseudochaotic kicked oscillator could hardly be simpler: a one-

dimensional harmonic oscillator, subjected to impulsive kicks in resonance with
the natural frequency, with the amplitude of the kicks a periodic, piecewise linear
function of the position. As with other nonlinear systems, the simple dynamics can
produce, over a long time, exceedingly complex behavior, typically too complex
for meaningful mathematical analysis or even reliable numerical experiments. For-
tunately, a special choice of parameters provides the key to a remarkably detailed
understanding of the long-time asymptotics. That key is renormalizability. To show
how renormalizability leads to a wealth of exact results, as well as powerful compu-
tational tools for exploring long-time asymptotics, has been my principal objective
in writing this monograph. The exposition is reasonably self-contained, with refer-
ences to the literature where supplementary details are needed. The methods used
are largely traditional ones which should be accessible to readers with a modest fa-
miliarity with the dynamics of low-dimensional Hamiltonian systems. No previous
acquaintance with pseudochaos is assumed.
This book weaves together a number of important threads drawn from a long,
productive, and continuing collaboration with Franco Vivaldi at Queen Mary, Uni-
versity of London. Our research program, initiated in 1995, has resulted in a number
of published articles on the topic of piecewise isometries and their various man-
ifestations, including Hamiltonian round-off, interval exchange transformations,
kicked-oscillator models, flights, and pseudochaos. Over the years, other researchers
have contributed significantly to the enterprise, notably Gullaume Poggiaspalla,
Konstantin Kouptsov, Sangtian Liu, and Spyros Hatjispyros.
In writing the book, I have drawn liberally from the published results, tying them
together and also adding considerable new material, and new perspectives, in order
to fashion a coherent whole. Central to the story is the unifying role of the symbolic
dynamics. A detailed description of the latter, including the previously unpublished
derivation of admissibility rules, appears in Chapter 3. In addition, the discussion
of transport in Chapter 6 is entirely new. Finally, Chapter 7 integrates the results on
Hamiltonian round-off, originally published during 1997—2000, into the full story,
viii Preface
with the inclusion of a new numerical experiment to provide fresh evidence for some
of the main conclusions.
Over the years, my collaborators and I have benefited from generous short-term
support of our research from the Engineering and Physical Sciences Research Coun-
cil (EPSRC) and The Royal Society. The continuing hospitality and support of the
Department of Physics, New York University and the School of Mathematical Sci-
ences of Queen Mary, University of London, have been crucial to both the research
program and the writing of this book.
Some of the figures in the book have been copied, with permission of the pub-
lishers, from articles in the journals Physics Reports (Elsevier), Nonlinearity (IOP),
and Communications in Nonlinear Science and Numerical Simulation (Elsevier).
Portions of the first five chapters were originally included in a series of pedagog-
ical lectures at the National University of Singapore (NUS) in August, 2006. I very
much appreciate the hospitality and financial support provided by the Institute of
Mathematical Sciences of NUS at that time.
Finally, I want to thank Valentin Afraimovich and Albert Luo for their advice
and encouragement throughout the writing of this book.
New York, August, 2011 John H. Lowenstein

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Kicked oscillators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Poincaré sections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Crystalline symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Stochastic webs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Normal and anomalous diffusive behavior . . . . . . . . . . . . . . . . . . . . . . 7
1.6 The sawtooth web map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.7 Renormalizability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.8 Long-time asymptotics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.9 Linking local and global behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.10 Organization of the book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2 Renormalizability of the Local Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.1 Heuristic approach to renormalizability . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1.1 Generalized rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1.2 Natural return map tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 Quadratic piecewise isometries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.1 Arithmetic preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.2 Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2.3 Geometric transformations on domains . . . . . . . . . . . . . . . . . . 30
2.2.4 Scaling sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2.5 Periodic orbits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2.6 Recursive tiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2.7 Computer-assisted proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3 Three quadratic models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.3.1 Model I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.2 Model II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.3.3 Model III . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
x Contents
2.4 Proof of renormalizability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

2.5 Structure of the discontinuity set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.5.1 Model I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.5.2 Model III . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.6 More general renormalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.7 The π /7 model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3 Symbolic Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.1 Symbolic representation of the residual set . . . . . . . . . . . . . . . . . . . . . 63
3.1.1 Hierarchical symbol strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.1.2 Eventually periodic codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.1.3 Simplified codes for quadratic models . . . . . . . . . . . . . . . . . . . 66
3.2 Dynamical updating of codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.3 Admissibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.3.1 Quadratic example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.3.2 Models I, II, and III . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.3.3 Cubic example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.4 Minimality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4 Dimensions and Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4.1 Hausdorff dimension and Hausdorff measure . . . . . . . . . . . . . . . . . . . 79
4.2 Construction of the measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.3 Simplification for quadratic irrational λ . . . . . . . . . . . . . . . . . . . . . . . . 81
4.4 A complicated example: Model II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.5 Discontinuity set in Model III . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.6 Multifractal residual set of the π /7 model . . . . . . . . . . . . . . . . . . . . . . 85
4.7 Asymptotic factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.8 Telescoping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.9 Unique ergodicity for each Σ (i) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.10 Multifractal spectrum of recurrence time dimensions . . . . . . . . . . . . . 94
4.10.1 Auxiliary measures and dimensions . . . . . . . . . . . . . . . . . . . . . 95
4.10.2 Simpler calculation of the recurrence time dimensions . . . . . 100
4.10.3 Recurrence time spectrum for the π /7 model . . . . . . . . . . . . . 102
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5 Global Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

5.1 Global expansivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.1.1 Lifting the return map ρ K (0) . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.1.2 Lifting the higher-level return maps . . . . . . . . . . . . . . . . . . . . . 109
5.2 Long-time asymptotics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.3 Quadratic examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.4 Cubic examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.4.1 Orbits in the (0, k, 6 ∞ ) sectors . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Contents xi
5.4.2 Numerical investigations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

5.4.3 A non-expansive sector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.4.4 Generic behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6 Transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.1 Probability calculation using recursive tiling . . . . . . . . . . . . . . . . . . . . 132
6.2 Ballistic transport in Model I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.3 Subdiffusive transport in Model II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.4 Diffusive transport in Model II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.5 Superdiffusive transport in Model III . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7 Hamiltonian Round-Off . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

7.1 Vector field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
7.2 Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.3 Localization of the vector field and periodic orbits . . . . . . . . . . . . . . . 155
7.4 Symbolic codes for walks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
7.5 Construction of the probability distribution . . . . . . . . . . . . . . . . . . . . . 159
7.6 Rotation number 1/5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
7.6.1 Recursive tiling for the local map . . . . . . . . . . . . . . . . . . . . . . . 161
7.6.2 Probability distribution P(x,t) . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.6.3 Fractal snowflakes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
7.6.4 Substitution rules for lattice walks . . . . . . . . . . . . . . . . . . . . . . 168
7.6.5 Separating out an asymptotic walk . . . . . . . . . . . . . . . . . . . . . . 170
7.6.6 Asymptotic scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
7.7 Model I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
7.8 Model II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7.9 A conjecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
Appendix A Data Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

A.1 Model I Data Tables, from Kouptsov et al. (2002). . . . . . . . . . . . . . . . 185
A.1.1 Generating domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
A.1.2 Level-0 scaling sequence domains . . . . . . . . . . . . . . . . . . . . . . 185
A.1.3 Level-0 periodic domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
A.1.4 Miscellaneous periodic domains . . . . . . . . . . . . . . . . . . . . . . . . 186
A.2 Model II Data Tables, from Kouptsov et al. (2002) . . . . . . . . . . . . . . . 186
A.2.1 Generating partition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
A.2.2 Level-0 scaling domains, sequence A . . . . . . . . . . . . . . . . . . . 187
A.2.3 Level-0 periodic domains, sequence A . . . . . . . . . . . . . . . . . . 188
A.2.4 Miscellaneous periodic domains, j > 10 . . . . . . . . . . . . . . . . . 188
A.2.5 Level-0 scaling domains, sequence B . . . . . . . . . . . . . . . . . . . . 189
A.2.6 Level-0 periodic domains, sequence B . . . . . . . . . . . . . . . . . . . 189
xii Contents
A.2.7 Incidence matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190

A.3 Model III Data Tables, from Kouptsov et al. (2002) . . . . . . . . . . . . . . 191
A.3.1 Generating domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
A.3.2 Pre scaling level L = −1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
A.3.3 Domains D j (L) for even L . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
A.3.4 Domains D j (L) for odd L . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
A.3.5 Domains Π j (L) for all L . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
A.3.6 Tiling data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
A.3.7 Section of the discontinuity set . . . . . . . . . . . . . . . . . . . . . . . . . 193
A.4 Cubic Model Data Tables, from Lowenstein et al. (2004) . . . . . . . . . 194
A.5 Inadmissibility Tables for Models II and III . . . . . . . . . . . . . . . . . . . . . 200
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
Appendix B The Codometer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Color Figure Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

Chapter 1
Introduction
In this initial chapter, we introduce a class of one-dimensional kicked oscillator

models whose crystalline symmetry and renormalizability will allow us, in the
course of this book, to explore the rich and varied relationships between dynami-
cal self-similarity at the local level and transport on the infinite phase plane.
1.1 Kicked oscillators
The dynamics of a periodically kicked oscillator is governed by a Hamiltonian of

the form (in convenient units)
1
H(x, y) = (x2 + y2 ) − F(x) ∑ δ (t − 2π nρ ), (1.1)
2 n
where the rotation number ρ is the number of instantaneous kicks per natural period.
It is assumed that the kicks are in resonance with the unperturbed oscillations, so
that ρ is a rational number, and that the derivative f (x) = F (x) is a periodic function
of the oscillator position. Hamilton’s equations of motion take the form
∂H ∂H
ẋ = = y, ẏ = − = −x + f (x) ∑ δ (t − 2πρ n). (1.2)
∂y ∂x n
Between successive kicks, the system undergoes free oscillation, depicted in the
x, y phase space as uniform clockwise motion on a circular arc of arbitrary radius and
angle 2πρ . This is followed by an instantaneous momentum shift y → y+ Δ y, where
Δ y is given by the kick function f (x). In Fig. 1.1, we illustrate such a phase-space
orbit for a 4-fold resonance and a sinusoidal kick function. The example is typical of
kicked-oscillator models, introduced in the 1980’s by Zaslavskii et al. (1986, 1991)
to model the interaction of electromagnetic waves with gyrating charged particles in
a plasma, and more abstractly, to illustrate the dynamical generation of crystalline
and quasicrystalline order in 2-dimensional phase space.
2 1 Introduction
1.2 Poincaré sections
The simplicity of the motion between kicks in Fig. 1.1(a) suggests that it might
be advantageous to adopt a stroboscopic point of view, regarding the essence of
the dynamics to be a discrete (Poincaré) map W connecting the phase-space points
(xn , yn ) at which the kicks are initiated. Explicitly, we have

x cos 2πρ sin 2πρ x
W :R →R ,
2 2
→ . (1.3)
y − sin 2πρ cos 2πρ y + f (x)
As is typical of dynamical systems with one degree of freedom and periodic forcing,
the stroboscopic phase space (Poincaré section) is partitioned into disjoint invariant
subsets. These subsets may be collections of points (periodic orbits), curves (quasi-
periodic orbits) and more complicated sets (stochastic layers) populated in part by
chaotic orbits. The discreteness of the dynamical map makes it especially easy and
efficient to visualize such sets via computer-assisted iteration.
Fig. 1.1 Quasiperiodic orbits of the sinusoidal kicked oscillator with F(x) = −a cos x, a = 0.8,
calculated over (a) 5 and (b) 50 oscillation periods. Points of the Poincaré section are shown as
large dots.
Figure 1.1(b) follows the orbit of Fig. 1.1(a) through 200 kick periods (50 os-
cillation periods). The phase-space orbit occupies an approximately annular re-
gion with numerous self-intersections. Without emphasizing (with dots) the points
of the Poincaré section, the pattern would be confusing. Further iteration would
make matters worse, since we would eventually be left with a featureless annu-
lar region overlapping those of nearby orbits. The Poincaré section, on the other
hand, reveals the topological simplicity of the orbit, which, viewed stroboscopi-
1.3 Crystalline symmetry 3
cally, fills out quasiperiodically four symmetrically placed closed curves. Choosing
other initial conditions reveals a “phase portrait” in which the curves of the exam-
ple are part of an infinite family of orbits circulating around the periodic points
(π , π ), (π , −π ), (−π , −π ), (−π , π ), as shown for example, by Fig. 1.5(b) in Sect.
1.4.
For ρ = 1/4 in general, the Poincaré map (1.3) simplifies to

x 0 1 x y + f (x)
W = = (1.4)
y −1 0 y + f (x) −x
For reasons which will become apparent below, the map W of (1.4) is often referred
to the 4-fold web map, and we will use this terminology in the remainder of the
book.
1.3 Crystalline symmetry
The 4-fold web map provides a simple and elegant theoretical laboratory for study-
ing transport in a low-dimensional Hamiltonian system. This is due in large part to
its crystalline symmetry. Specifically, suppose that the kick function f (x) has period
τ , and that , for τ > 0, f (x) is continuous from the right on [0, τ ), while, for τ < 0,
f (x) is continuous from the left on (τ , 0]. Then every point of the real plane can be
uniquely decomposed as the sum of a local vector u in the fundamental domain

[0, τ )2 τ > 0,
Ω=
(τ , 0]2 τ < 0,
and a global vector τ m, m = (m, n) ∈ Z 2 .

From (1.4), we have, for all u ∈ Ω , m ∈ Z 2 ,

0 1
W (u) + τ I · m, I= . (1.5)
−1 0
Since I4 is the identity, we get from (1.5) the discrete translation invariance of the
fourth iterate of W : for all x ∈ R 2 , m ∈ Z2 ,
W 4 (x + τ m) = W 4 (x) + τ m. (1.6)
Figure 1.3 illustrates the decomposition (1.5) for a hypothetical choice of f (x) de-
picted in Fig. 1.2.
To take full advantage of the local-global decomposition, we now introduce the
piecewise continuous local map K : Ω → Ω , defining, for all u = (u, v) ∈ Ω ,
K(u) = W (u) − τ d(u), d(u) = (τ −1 (v + f (u)), −1) ∈ Z2 ,

4 1 Introduction
Fig. 1.2 Graph of some kick function f (x) on [0, 1).
so that, for all u ∈ Ω , m ∈ Z2 ,
W (u + τ m) = K(u) + τ L̃u (m),
with the lattice isometry (rotation composed with lattice translation)
L̃u (m) = I · m + d(u).
The action of K on Ω , for the example of Fig. 1.3, is shown in Fig. 1.4. We
note that for piecewise continuous f (x), the unit square Ω is partitioned into re-
gions Ω i , i = 1, 2, . . . , ν on which the lattice translations d(u) are constant. Thus
Fig. 1.3 Action of W (x).

1.4 Stochastic webs 5
Fig. 1.4 Construction of K(u) for the map of Fig. 1.3.
L̃u depends on u only through the index i(u), u ∈ Ω i(u) , and we can express the
local-global decomposition of W (x) as
W (u + τ m) = K(u) + τ Li(u) (m), (1.7)
where
Li(u) = L̃u .
How does the crystalline symmetry of the dynamics, expressed through (1.7),
facilitate the investigation of chaotic transport? For the latter, the asymptotic long-
time behavior mimics the effects of a random walk on the plane, in which successive
steps are dictated by the results of independent coin tosses. Clearly it is the deter-
ministic chaos of the local map K which plays the role of the coin tosses, producing
a “code” sequence i 1 , i2 , . . . , which in turn determines a sequence of steps on the
infinite lattice. For asymptotically long times, it is the statistical distribution of the
code-driven lattice coordinates which may in some sense exhibit diffusive behavior.
1.4 Stochastic webs
Although the main focus of this book is on maps which possess only some of the
features of true chaos (hence the term pseudochaos), it is important, to properly
understand the motivation for this work, that we focus first on the chaotic stochastic
web map with kick amplitude
f (x) = a sin x.
While the quasiperiodic orbits shown in Fig. 1.1(b) are restricted to only four cells,
the same is not necessarily the case for chaotic orbits originating in the vicinity of
one of the points (mπ , nπ ), m + n odd. The linearized map approximating W 4 there
is a 2 × 2 matrix, with one real eigenvalue greater than 1, the other less than one, so
that these points are saddle points. By selecting initial points near one of these sad-
dle points, say (π , 0), it is easy to simulate numerically the orbits of the stochastic
layer in which it is embedded. Each saddle point acts as a kind of random gate: on
6 1 Introduction
approaching it, an orbit must “decide” whether to turn right or left. The “decision”
is deterministic, of course, but extremely sensitive to initial conditions whenever the
orbit passes very close to a saddle point. The effect is very much like a punctuated
random walk, with the orbit point moving stochastically from one saddle point’s
vicinity to another’s after circulating locally. For ensembles of extremely long or-
bits, the local sensitivity to initial conditions translates into asymptotic behavior, in a
statistical sense, resembling diffusion. For relatively small parameter a, the wander-
ing chaotic orbits are restricted to a narrow web-like region connecting the saddle
points at (mπ , nπ ), m + n odd (see Fig. 1.5).
Fig. 1.5 (a) Chaotic orbit of the a = 0.8 sinusoidal kicked-oscillator map W (104 iterations).
(b) Same orbit as (a), folded into the fundamental cell, showing structure of the stochastic web.
Also shown are selected quasiperiodic orbits surrounding the fixed points at the center and at the
corners of the cell.
Stochastic webs are present, not just for ρ = 1/4, but quite generally for sinu-
soidal kick function, rational ρ , and sufficiently small a (Zaslavskii et al., 1991).
However, the cases ρ = p/q, q = 3, 4, 6 are the only cases where the 2-dimensional
phase space has a periodic tiling by a finite number of cell shapes. This is a well-
known constraint of 2-dimensional crystallography. On the other hand, there do
exist mathematical patterns which exhibit quasi-crystalline structure (the Fourier
transform of the density of points is point like and invariant under q-fold or 2q-fold
rotations). It is an empirical fact, based on numerical simulations, that the stochas-
tic webs generated by kicked-oscillator Poincaré maps are indistinguishable from
mathematical quasicrystals (Zaslavskii et al., 1991). For the patterns of periodic
points (rather than the stochastic layers), one can in some cases prove the quasicrys-
talline symmetry rigorously (Lowenstein, 1993, 1994). Because we are specifically
interested in the interplay between local and global behavior, we will restrict our-
selves henceforth to the crystalline maps, and, in the interest of simplicity, to the
case ρ = 1/4.
1.5 Normal and anomalous diffusive behavior 7
1.5 Normal and anomalous diffusive behavior
The local and global manifestations of chaos in the ρ = 1/4 stochastic web map
have been studied numerically in great detail by Zaslavsky and Niyazov (1997), Za-
slavsky et al. (1997), and Zaslavsky and Edelman (2001, 2003). That the long-time
behavior is characterized by a diffusive power law (mean squared distance increas-
ing as Dt) was verified for a wide range of parameter values. The empirical values
of the effective diffusion constant D are shown in Fig. 1.6. The existence of sharp
peaks in the plot suggested to the authors that perhaps the function D(a) is actually
singular for certain parameter values. In order to understand the possible breakdown
of normal diffusive behavior in their data, the authors examined the Poincaré section
in the fundamental cell for parameter a = a ∗ = 6.349972 chosen to be as close as
possible to the most prominent apparent singularity. Figure 1.7(a) shows the phase
portrait of the Poincaré map folded into the the square [0, 2π ) 2. Although all but a
small fraction of the square is occupied by a chaotic “sea” of orbit points, the distri-
bution of points appears to be highly non-uniform. The orbit appears to have spent a
great deal of time in the vicinity of four small island systems, leading to a prominent
darkening of the picture in those locations. The effect comes into clearer focus by
zooming in on the darker regions to reveal several levels of island chains. This is
shown in Fig. 1.7(a—d). What distinguishes the parameter value a ∗ at which D(a)
blows up appears to be the self-similarity of the family of nested island chains.
If the diffusive behavior of the stochastic-web map breaks down at some a = a ∗ ,
what replaces it? The numerical experiments of Zaslavsky and Niyazov (1997) and
Zaslavsky et al. (1997) provide an answer: for large times, the distance of an orbit
point from the origin increases proportional to a super-diffusive power of the time:
Fig. 1.6 Diffusion constant, divided by Dql = a2 /2, as a function of parameter a for the 4-fold
sinusoidal kicked-oscillator map, from Zaslavsky and Niyazov (1997), Fig. 2.
8 1 Introduction
Fig. 1.7 Sticky chaotic orbit in the vicinity of one of the accelerator-mode islands for a = a∗ , from
Zaslavsky and Niyazov (1997), Figs. 3 and 5. In clockwise order, each frame is a zoom of the
preceding one.
|x| ∼ t μ , μ > 1/2. The kind of process which produces such behavior is evident
in Fig.1.8: globally, the chaotic orbit alternates between what appears to ordinary
random-walk wandering and long, constant-velocity flights. The latter correspond
to periods of “stickiness” in the vicinity of a self-similar island system, while the
former correspond to excursions in the chaotic sea. The intermittent combination
of the two produces an “anomalous” power law with an exponent intermediate be-
tween that of ballistic motion (μ = 1) and a normal random walk (μ = 1/2). In
the numerical investigations of Zaslavsky and Niyazov (1997) and Zaslavsky et al.
(1997), the relevant islands are accelerator modes: they are mapped from one cell to
another by the action of the 4th-iterate map, and hence correspond to motion with
constant average velocity. While trapped in the vicinity of such an island, a sticky
orbit will temporarily share that velocity. Small accelerator-mode islands embedded
in a chaotic sea are thus a good place to look for anomalous diffusive behavior. Dana
(2004) has adopted such an approach to study a family of super-diffusive orbits in a
ρ = 1/3 kicked oscillator model.
1.6 The sawtooth web map 9
35000
-35000
-35000 u 35000
Fig. 1.8 Global orbits (2 × 106 iterations) for a = 6.35. Long ballistic flights correlated to long so-
journs near self-similar island-around-island configurations, from Zaslavsky and Niyazov (1997),
Fig. 1(a).
The discovery of power-law asymptotic behavior in numerical simulations of si-

nusoidal stochastic web maps with finely tuned parameter values has opened up a
fascinating new realm to be explored both empirically and theoretically. Out of this
research have emerged some promising and novel ideas concerning dynamical trap-
ping by self-similar island structures, critical exponents for spatio-temporal scaling,
and fractional-kinetic equations for describing anomalous transport, all under the
rubric of “pseudochaos”. For a comprehensive discussion of some of the main ideas
in this rather broad field, the reader is referred to Zaslavsky (2005).
1.6 The sawtooth web map
In order to model pseudochaos with mathematical precision, one would like to con-
sider relatively simple dynamical systems which exemplify in a distilled form those
features which Zaslavsky and his collaborators found in the chaotic boundary re-
gions of finely tuned web maps, namely (1) vanishing Lyapunov exponent and sym-
bolic entropy, (2) self-similarity of the local dynamics, and (3) asymptotic diffusion
like behavior of the global orbits, with a clearly delineated link between (2) and (3).
Our choice for such a distilled model of pseudochaos is a particular class of 4-fold
web maps with piecewise linear (“sawtooth”) kick amplitudes (see Fig.1.9),
10 1 Introduction
def def
f (x) = λ (x − xτ τ ) = λ {x}τ , xτ = x/τ τ . (1.8)
Fig. 1.9 Example of a sawtooth kick function f (x).
Sawtooth kicked-oscillator models have been studied in some detail by Lowen-

stein et al. (2005), as an outgrowth of our work on piecewise isometries of convex
polygons. We found that the piecewise isometric mappings of the square could be
promoted to the local maps, on the fundamental cell, of sawtooth kicked-oscillator
maps of the plane with rotation number 1/4. Whereas Lowenstein et al. (2005) limit
their attention to sawtooth maps with quadratic irrational parameters, the present
investigation generalizes the “lifting” of piecewise isometries to include a (poten-
tially) much larger class of models satisfying the so-called recursive tiling property.
The local map, being a piecewise linear mapping of the fundamental square Ω
onto itself, is capable of generating fractal structures, thanks to the branching effect
of its discontinuities, but lacks the stretching/contraction mechanism which gener-
ates chaotic layers and exponential divergence of nearby orbits in the presence of
continuous kick functions.
1.7 Renormalizability
In rather general terms, renormalizability of a dynamical system refers to its dy-

namical self-similarity, i.e. the existence of an infinite hierarchy of scales, with
geometrical and dynamical relations assuming the same form, apart from a scale
transformation, at each level of the hierarchy (see, for example, MacKay, 1993).
In Chapter 2 of this book, we shall see how the concept of renormalizability
is realized in the context of the sawtooth web map. There, the scaling hierarchy
involves a nested sequence of convex polygonal domains, D(L) ⊂ Ω , together with
their respective first return maps ρ (L) induced by K, L = 0, 1, 2, . . .. The induced
maps have the defining property that, for all x ∈ D(L), we have
1.7 Renormalizability 11
ρ (L)(x) = K t (x),
where t, depending on x, is the smallest positive integer such that K t (x) ∈ D(L), but
K s (x) ∈
/ D(L) for all positive s < t.
The self-similarity of D(L), ρ (L) pairs, for L = 0, 1, 2, . . ., consists essentially of
the recursion relations, valid for all levels L of the hierarchy:
(i) ρ (L + 1) is induced by ρ (L) in a level-independent way.
(ii) D(L + 1) = Sω D(L), where Sω is a scale transformation by a factor ω < 1, with
respect to some limit point x ∞ ∈ Ω .
(iii) ρ (L + 1) = Sω ◦ ρ (L) ◦ Sω−1.
The existence of a sequence of D(L), ρ (L) pairs satisfying the above criteria
(what we call a scaling sequence) is not always sufficient to ensure full renormaliz-
ability. Also needed is a criterion ensuring that the scaling sequence provides com-
plete coverage of the aperiodic orbits of the model. For the sawtooth web maps, this
criterion is recursive tiling: the return orbits of ρ (L + 1) should completely tile the
region of D(L) complementary to the periodic orbits of ρ (L).
For most of the sawtooth web maps which we consider in detail in this book,
renormalization indeed takes the form of a single scaling sequence of domains
equipped with a return map hierarchy satisfying the recursive tiling property. All
of these have parameter values with particularly simple properties: they are all
in the interval −2 < λ < 2 and are irrational roots of quadratic polynomials
λ 2 + aλ + b, a, b ∈ Z. However these do not exhaust the cases of interest. In partic-
ular, we have cases where more than one scaling sequence is needed to completely
cover√the aperiodic orbits. Among the models with quadratic irrational λ , the case
λ = 3 stands out: there are two disjoint scaling sequences, with distinct limit
points and distinct temporal scaling properties, leading to fractal sets of aperiodic
points with different dimensions. Our single example of a cubic irrational λ , the
so-called π /7 model, has infinitely many disjoint scaling sequences, but even tak-
ing them together does not lead to complete coverage of the pseudochaotic web of
aperiodic orbits.
What is needed for the π /7 model, and presumably more widely among non-
quadratic models, is a renormalization scheme characterized by a catalogue of sev-
eral polygons, each of which is recursively tiled by the return orbits of domains
equivalent to the various catalogue members under translations, generalized rota-
tions and reflections, and scale transformations. The hierarchy is then described
by an infinite tree, rather than by a ladder or collection of ladders appropriate to
quadratic models. The scheme is sketched in Fig. 2.10.
Obviously not all sawtooth web maps are renormalizable, even in the more gen-
eral sense. The ones we have found are quite special, not only in the restriction to
parameter values with special arithmetic properties, but also in the choice of the
kick-function period τ and the placement of the fundamental domain Ω . In spite
of what appears to be the highly restrictive nature of this class, we will find that
there is considerable variety in the asymptotic behaviors on the phase plane which
its members produce.
12 1 Introduction
1.8 Long-time asymptotics
As we will see in Chapter 5 of this book, the hierarchical partition of phase space
provided by recursive tiling allows us not only to prove general properties of asymp-
totically long orbits, but also to compute these orbits explicitly with great efficiency
and exactitude. In particular, with the use of computer-manipulated algebraic num-
bers, the numerical errors inherent in floating point arithmetic can be completely
avoided. Moreover, with recursive tiling and hierarchical symbolic dynamics, the
time scales increase exponentially with level, making possible accelerated routes to
the asymptotic regime.
Complementing the study of the long-time behavior of aperiodic orbits is a sta-
tistical formulation, embodied in the time evolution of the moments of a probabil-
ity distribution. This is what we mean by transport. What does such a statistical
formulation look like for our renormalizable kicked oscillator models? There are
many possibilities, but, given the crystalline structure of the phase space, with τ -
periodicity both horizontally and vertically on the x, y plane, as well as the local-
global factorization of the dynamics, it is natural to adopt the following coarse-
graining: average uniformly over local motions and define a probability distribution
P((m, n),t) solely for “walks” on the integer lattice {(mτ , nτ ) : m, n ∈ Z}.
In the definition of P(r,t), the notion of random selection of initial conditions
is the usual one of a uniform distribution with respect to some invariant measures.
For the full set of orbits, Lebesgue measure (area) is the obvious choice. Since the
full area measure of Ω is occupied by periodic polygons for the renormalizable
models we are considering, it is clear that we could base our entire analysis on
periodic orbits. This gives rise to the somewhat paradoxical situation that although
the support of P(r,t), i.e. the set of lattice points r where P(r,t) > 0, grows without
bound with increasing t, the motion on any individual orbit with nonzero weight in
the averaging is periodic, hence bounded.
In some interesting cases we will be primarily interested in the statistical long-
time behavior of orbits living on specific invariant subsets of the exceptional set.
Typically the latter occupies zero area, and so, in order properly to formulate the
transport problem, it will be necessary to abandon Lebesgue measure. Hausdorff
measure restricted to the particular sets of initial conditions will turn out to be a
convenient alternative.
The picture which emerges from our analysis of long-time asymptotics is not
quite as simple as one might have anticipated. The local self-similarity does not
always lead to a global quasi-diffusivity: a number of renormalizable models have
only uniformly bounded global orbits. Furthermore, where there is global expan-
sion, it is not necessarily governed by the same power laws that one encounters
locally. The main reason is that every iteration of the global map involves a 90 ◦ ro-
tation of the plane, introducing a phase factor which must be included in calculations
of the asymptotic scaling. The criteria for self-similarity are thus distinct from those
which characterize the local scaling, and the associated power laws have different
exponents.
1.10 Organization of the book 13
1.9 Linking local and global behavior
What makes the study of the sawtooth kicked-oscillator maps so compelling is the
fact that the local renormalizability and global transport can be tied together in a
single, satisfying package. Central to this linkage is the fact that the local map K,
a piecewise linear map on the square, and the global map L, a lattice isometry,
are both governed by a single abstract map acting on symbol strings. This is not
the simple “coin-toss” coding based on the elementary partition of Ω , but rather a
hierarchical one reflecting the level structure of recursive tiling. In such a coding,
which we will discuss fully in Chapter 3, successive symbols refer to successive
levels in the renormalization hierarchy, and hence compress space and time at a
geometric rate as one proceeds along the symbol string. The dynamical map on
the symbol strings is of a lexicographic type developed by Vershik (1985), with
applications in diverse mathematical contexts. In our numerical explorations of the
long-time regime, the updating rule for symbol strings acts as an abstract machine
which we call a codometer (see Appendix B). It is reminiscent of, but more general
than, that of an ideal odometer with infinitely many registers, and will allow highly
efficient computational excursions into the long-time regime.
For the quadratic models, the story does not end here. As we shall see in Chapter
7, the symbolic dynamics of the kicked-oscillator models also underlies the lattice
dynamics of certain models of non-dissipative round-off. In spite of the common
abstract structure, the lattice transformations are vastly different from the kicked-
oscillator lattice maps, and are endowed with completely different long-time asymp-
totic behavior. By means of a substitution-based analysis of the orbits in the long-
time regime, we succeed in deriving, in one of our models, a rigorous expression
for the asymptotic power law with log-periodic modulation. By means of highly
efficient, codometer driven programs, we have verified the general form of the mod-
ulated power laws in other globally expansive models as well, including one where
the modulating factor for the second moment is log-quasiperiodic.
1.10 Organization of the book
The remainder of the book is organized as follows.

Chapter 2
We begin with a review of the concepts and formalism of scaling, renormalization,
and recursive tiling, all within the context of the local map. This includes a gener-
alization of the renormalization scheme to cover a wider class of models, including
the so-called π /7 model with its cubic irrational parameter.
Chapter 3
We establish the correspondence of hierarchical symbol sequences with points on
the discontinuity-avoiding aperiodic orbits. After introducing the dynamical map,
14 1 Introduction
in the form of an odometer-like updating scheme, the codometer, we concentrate on

the technically complicated problem of deriving intrinsic rules for the admissibility
of symbol sequences.
Chapter 4
Here we focus on the measure-theoretic and fractal (or multifractal) properties of the
residual set of the local map. The introduction of Hausdorff dimension and Haus-
dorff measure will be essential, later on, to formulating transport on interesting sets
of Lebesgue measure zero. Using a transfer matrix to link successive levels of the
renormalization hierarchy, we calculate a nontrivial spectrum of generalized dimen-
sions for the π /7 model.
Chapter 5
We now examine how the local return maps of a scaling sequence can be lifted to the
infinite kicked-oscillator phase space. In some cases the resulting global translation
vectors are globally expansive, increasing in length geometrically with increasing
level. In such cases, we prove that the discontinuity-avoiding aperiodic orbits tend
to infinity according to a power law. We study this phenomenon numerically in a
number of examples. In the particularly interesting case of the π /7 model, we see
that different initial conditions can lead to a variety of asymptotic power laws.
Chapter 6
A statistical formulation of the time evolution is presented in this chapter, based on
a coarse-grained probability distribution P((m, n),t). Once again, the renormaliz-
ability of the models and especially the recursive tiling property, serve as the orga-
nizing principles. We consider examples in which the probability is supported on
the 2-dimensional periodic domains, but also ones restricted to interesting lower-
dimensional sets. Highly efficient codometer updating allows us to investigate nu-
merically asymptotic power laws with log-periodic modulation.
Chapter 7
In this final chapter, we consider a different type of kicked oscillator dynamics,
namely that of a Hamiltonian round-off map (Lowenstein et al., 1997; Lowenstein
and Vivaldi, 1998, 2000). Once again we have a local piecewise isometry which
induces a permutation of Z 2 . Although in this class of models all orbits are periodic,
one can still describe the long-time behavior statistically in terms of periodically
and quasi-periodically modulated power laws. We study these both analytically and
with highly efficient numerical experiments.
Appendixes
For the sake of completeness, we include in the Appendix A a number of data tables
for the models which appear throughout the book. In addition, we have a short expla-
nation of our implementation of the hierarchical symbolic dynamics, the codometer,
in Appendix B.
References 15
References
Dana I. (2004) Global superdiffusion of weak chaos, Physical Review E 69, 016212.
Lowenstein J.H. (1993) Quasiperiodic structure of the stochastic web map, Physical Review E
47 (Rapid Communication), 3811–3814.
Lowenstein J.H. (1994) Quasilattice of fixed points of the fivefold stochastic web map, Physical
Review E 49, 232–246.
Lowenstein J.H. and Vivaldi F. (1998) Anomalous transport in a model of Hamiltonian roundoff,
Nonlinearity 11, 1321–1350.
Lowenstein J.H. and Vivaldi F. (2000) Embedding dynamics for round-off errors near a periodic
orbit, Chaos 10, 747–755.
Lowenstein J.H., Hatjispyros S., and Vivaldi, F. (1997) Quasi-periodicity, global stability and scal-
ing in a model of Hamiltonian round-off, Chaos 7, 49–66.
Lowenstein J.H., Poggiaspalla G., and Vivaldi F. (2005) Sticky orbits in a kicked-oscillator model,
Dynamical Systems 20, 413–451.
MacKay R.S. (1993) Renormalisation in Area-Preserving Maps, World Scientific, Singapore.
Vershik A.M. (1985) A theorem on the Markov periodical approximation in ergodic theory, Journal
of Soviet Mathematics 28, 667–674.
Zaslavskii G.M., Zakharov M. Yu., Sagdeev R.Z., Usikov D.A., and Chernikov A. A. (1986) Gen-
eration of ordered structures with a symmetry axis from a Hamiltonian dynamics, JETP Letters
44, 451–456.
Zaslavskii G.M., Sagdeev R.Z., Usikov D.A., and Chernikov A. A. (1991) Weak Chaos and
Quasiregular Patterns, Cambridge University Press, Cambridge.
Zaslavsky G.M. (2005) Hamiltonian Chaos and Fractional Dynamics, Oxford University Press,
Oxford.
Zaslavsky G.M. and Edelman M. (2001) Weak mixing and anomalous kinetics along filamented
surfaces, Chaos 11, 295–305.
Zaslavsky G.M. and Edelman M. (2003) Pseudochaos, Perspectives and Problems in Nonlinear
Science: A Celebratory Volume in Honor of Lawrence Sirovich, eds. E. Kaplan, J. Marsden,
and K. R. Sreenivasan, Springer, New York, 421–423.
Zaslavsky G.M. and Niyazov B.A. (1997) Fractional kinetics and accelerator modes, Physics Re-
ports 283, 73–93.
Zaslavsky G.M., Edelman M., and Niyazov B.A. (1997) Self-similarity, renormalization, and phase
space nonuniformity of Hamiltonian chaotic dynamics, Chaos 7, 159–181.
Chapter 2
Renormalizability of the Local Map
The local component of the sawtooth kicked-oscillator dynamics is the renormaliz-

able, piecewise linear mapping K of a square Ω onto itself,
K : Ω → Ω, K(u) = C · u mod Ω , (2.1)
where
λ 1 [0, τ )2 τ > 0,
C= , Ω=
−1 0 (τ , 0]2 τ < 0.
Here the “mod” operation is a lattice translation to enforce K(u) ∈ Ω . The current
chapter is a thorough examination of the map K for an important class of renormal-
izable models.
The piecewise affine map K on the square has an important place in the dynam-
ical systems literature of the last two decades. Some of the interesting examples
are:
• Overflow oscillations of a second-order nonlinear digital filter: Ashwin et al.
(1997), Chua and Lin (1990), Davies (1995), Wu and Chua (1993).
• Equivalence to the sawtooth analogue of the Chirikov-Taylor standard map: Ash-
win (1997), Dana (2004).
• Sticky orbits in stochastic-web maps: Zaslavsky and Edelman (2003), Lyubomu-
drov et al. (2003).
• Localization of the lattice orbits of oscillator models with Hamiltonian round-off:
Lowenstein et al. (1997), Lowenstein and Vivaldi (2000).
• Piecewise isometries of convex polygons: Adler et al. (2001), Akiyama et al.
(2008), Buzzi (2001),Goetz (1996, 1998, 2000, 2001), Goetz and Poggiaspalla
(2004), Kahng (2000, 2002), Kouptsov et al. (2002), Lowenstein (2007), Lowen-
stein et al. (2004), Poggiaspalla (2006).
The material of the current chapter is drawn in large part from Kouptsov et al.
(2002) and Lowenstein et al. (2004), as well as the electronic supplements linked
to those articles, to which the interested reader is directed for additional details. For
the sake of notational consistency with these articles, we introduce a trivial change
18 2 Renormalizability of the Local Map
of phase-space coordinates at this point, namely a canonical transformation which

essentially interchanges the roles of position and momentum, namely

x y
→ .
y −x
Adopting the transformed canonical frame through the remainder iof the book, we
redefine the web map W of (1.4) and matrix C of (2.1) to be

x y 0 1
W = , C= .
y −x + f (y) −1 λ
2.1 Heuristic approach to renormalizability
The restrictions imposed by renormalizability on the parameter λ and square do-

main Ω in (2.1) are obviously quite severe. Currently, no theorem is available to tell
us which λ are allowed, and we will not make an attempt to explore this problem.
Our aim will be a more practical one, namely to construct a class of renormalizable
local maps associated with an interesting and diverse family of kicked oscillator
models. In each case, we demand that the asymptotic long-time behavior be deter-
mined with mathematical precision. Our choice of models will thus be governed by
heuristic considerations, while our treatment of each model will proceed without
sacrifice of rigor.
2.1.1 Generalized rotations
The matrix C has eigenvalues

1
(λ ± λ 2 − 4).
2
which, for |λ | < 2, are complex conjugate points e ±iθ on the unit circle. The matrix
is thus linearly conjugate to a planar rotation by θ ,

cos θ sin θ
R(θ ) =
− sin θ cos θ
and the mapping K is linearly conjugate to a piecewise isometry on the rhombus.

If we assign to the original xy plane the metric tensor
⎛ ⎞
λ
1 −
⎜ 2⎟
Q=⎝ ⎠ (2.2)
λ
− 1
2
2.1 Heuristic approach to renormalizability 19
we can legitimately refer to C as a rotation and K itself as a piecewise isometry

(PWI for short). We shall routinely use this terminology in what follows.
At first sight we cannot rule out |λ | ≥ 2, for which the eigenvalues of C are real.
However, for reasons which will soon become clear, a rotational C turns out to be
essential to our strategy of constructing renormalizable models, and so henceforth
we shall assume |λ | < 2.
2.1.2 Natural return map tree
Renormalizability entails self-similar families of polygonal domains, but given only

the map K, how does one find those subdomains of Ω which have the desired struc-
ture? The only successful strategy we have discovered (and even then, only for care-
fully selected parameter values) is a recursive search of the natural return map tree.
Let us describe informally how this works.
The natural return map tree takes as its starting point (level 0) the partition of Ω
into a convex polygonal domains (known as atoms)
Ωi = {z ∈ Ω : ι (z) = i}, i = 0, . . . , a − 1.
We then calculate the (first-) return map ρ i on each domain Ω i . This is defined, for
almost all z ∈ Ω i , by
ρi (z) = Kt (z),
where the return time t is the smallest positive integer such that K t (z) ∈ Ωi . A con-
structive algorithm for ρ i , which emphasizes its role as a piecewise isometry on Ω i ,
is the following:
Induction algorithm. Start with a specific Ω i . Apply K and intersect the result with
each of the a atoms Ω j to obtain
Ei, j = K(Ωi ) ∩ Ω j .
/ we note that a subdomain of Ω i , namely

If Ei,i = 0,
Di = K −1 (Ei,i )
has returned to Ω i after one iteration of K, and hence qualifies as a member of the
set A (ρi ) of atoms of the (level 1) return map ρ i of Ω i . The remaining nonempty
Ei, j , with j = i, constitute a set of “in transit” domains which need to be subjected
to additional iterations of K to bring them home. In the next step, we calculate
Ei, j,k = K(Ei, j ) ∩ Ωk ,

appending the nonempty D i, j = K −1 (Ei, j,i ) to the list of atoms A (ρi ) and designat-
ing the remaining nonempty E i, j,k , j, k = i, as candidates for further mapping and
slicing.
Continuing as above, we obtain a recursive construction of a set of atoms
Di,i1 ,i2 ,...,in−1 , i1 , . . . , in−1 = i.
For each n, these are convex subdomains of Ω i which return to that polygon after a
return time of n, along the return paths
(i, i1 , . . . , in−1 ).
Specifically, the return time denotes the number of K iterations needed for first re-
turn, and the return path lists the atoms of the level-0 map visited along the way.
Note that we have no guarantee that the construction ever terminates, and so the
number of atoms of the induced map ρ i may be countably infinite.
It is clear that there are no formal differences between the original piecewise
isometry K and the return maps ρ i induced via the above construction on the atoms
of K. It is thus completely straightforward to apply the analogous construction re-
cursively to obtain a tree in which each node of level L is a dressed domain, i.e. a
pair Δ = (D, ρ ) where D is a convex polygon (perhaps including some or all of its
edges) and ρ is the (first-) return map induced by the level-(L − 1) dressed domain
containing Δ . Clearly, level 0 of the tree consists of the dressed domain (Ω , K),
while level 1 consists of {(Ω i , ρi ) : i = 0, . . . , a − 1}.
In terms of the natural return map tree, the search for renormalizability re-
duces to an attempt to find two dressed domains, Δ (0) = (D(0), ρ (0)) and Δ (1) =
(D(1), ρ (1)), the latter a direct descendant of the former, differing from it by a sim-
ilarity transformation S, i.e.
D(1) = S(D(0)), ρ (1) = S ◦ ρ (0) ◦ S −1.
Most simply, S will be a scale transformation by a factor ω < 1, but could also
involve other linear transformations which preserve the Q-metric. Having found
one such pair on the tree, we are guaranteed of an infinite sequence, namely
D(L) = SL (D(0)), ρ (L) = SL ◦ ρ (0) ◦ S−L.
In what follows we will refer to {Δ (L) : L = 0, 1, 2, . . .} as a scaling sequence and

Δ (0) as a scaling domain.
Suppose for a particular choice of (K, Ω ) we have been fortunate enough to find
a scaling sequence of dressed domains buried in the natural return map tree. Does
the model qualify as a renormalizable one? The answer will be in the affirmative
if the scaling sequence in some way accounts for all of the essential dynamics of
the model. The criterion which we adopt to make this concept precise is recursive
tiling: the return map orbits of D(0) should tile all of Ω , apart from a finite number
of periodic K-orbits, and the return map orbits of each D(L), L > 0, should tile all
of D(L − 1), again up to finitely many periodic ρ (L − 1)-orbits. We will also want
to consider renormalizable those models with a finite number of disjoint scaling
sequences which together tile the entire square, again up to finitely many periodic
orbits.
2.1.3 Examples
To illustrate the return map construction and quest for renormalizable models de-
scribed in abstract terms above, it is useful to study in detail two explicit examples.
The first of these has the deceptively simple choice λ = 1/2, with K defined on the
unit square Ω = [0, 1) 2 . The map has two atoms, Ω 0 and Ω 1 . The first 8 steps in the
construction of the return map partition of the triangular domain Ω 0 are shown in
Fig. 2.1.
Fig. 2.1 First 8 steps of the construction of the return map ρ0 on Ω0 .

After two applications of K, the image triangle overlaps the two atoms and is split
into E0,0 ⊂ Ω0 (gray) and E 0,1 ⊂ Ω1 (black). The former, mapped back with K −1 ,
gives us the first of the atoms of ρ 0 , the induced return map on Ω 0 . The latter will
require further mapping and slicing to produce the remaining members of A (ρ 0 ).
Another branching event can be seen in the eighth panel. Obviously the construction
does not halt here, and continuing the process for a large number of iterations does
not complete the job either. In fact, it is possible that it will never halt, i.e. the return
map partition may be infinite.
The return map for Ω 1 is dramatically simpler (the construction terminates after
two steps), but when we proceed to the second level of the return map tree, we find
that the construction of the return map partition of each of the atoms does not appear
to terminate after many iterations.
From the point of view of a reasonable search procedure, the model with
λ = 1/2, Ω = [0, 1)2 appears to be a most unpromising candidate. The reason is
not difficult to understand. Consider the last step of a return map construction for
a domain Ξ which actually terminates. Here we have a convex polygon D out-
side of Ξ (colored black in the figures) and a convex polygon E within Ξ (col-
ored white in the figures), such that E = K(D). A necessary condition for suc-
cessful termination is clearly that the edges of polygons D and E should have
the same relative orientations, with the orientations of E obtained from those
of D by a single rotation C. This can be checked quite easily, since every do-
main boundary, by construction, has the orientation of a line C m · φ , where m
is a nonnegative integer and φ is the line through (0, 0) and (1/2, 1). The dif-
ficulty in achieving termination can be seen in frames 2, 3, and 7 of Fig. 2.1,
where the black and white triangles have a “bow-tie” configuration, with orienta-
tions (Cm1 φ ,Cm2 φ ,Cm3 φ ) and (C m2 φ ,Cm1 φ ,Cm4 φ ), with m1 , m2 , m3 , m4 all differ-
ent. Since (m1 + 1, m2 + 1, m3 + 1) ∈ {(m2 , m1 , m4 ), (m1 , m4 , m2 ), (m4 , m2 , m1 )}, the
corresponding return orbits cannot terminate the construction.
The above discussion suggests a strategy which might enhance our chances of
finding renormalizable models by means of a systematic search of the natural return-
map tree. Suppose C q were equal to the identity for some small value of q, corre-
sponding to a rational rotation number p/q. There would then be only a small num-
ber q of possible orientations of lines, which would greatly increase the chances
for finite (perhaps even small) return map partitions, as well as for finding similar
polygons among the nodes of the tree. It is not difficult to show that the (necessarily
irrational) parameter λ = 2 cos(2π p/q) is an algebraic integer of degree n, i.e. the
solution of a polynomial equation of the form
λ n = c0 + c1 λ + · · · + cn−1 λ n−1 , (2.3)
where n, ck ∈ Z, n > 1. For small values of n, only a small number of possible

rotation numbers are allowed. Thanks to (2.3), regarded as a recursion relation for
powers of λ , the set
Q(λ ) = {a0 + a1 λ + · · · + an−1 λ n−1 : ai ∈ Q, i = 0, . . . , n − 1},

is an algebraic number field, closed with respect to all the operations of ordinary
arithmetic, including division by non-zero elements. Thus, provided that the kick-
function period τ belongs to Q(λ ), all of the coordinates of polygonal vertices aris-
ing in our return map construction will also be restricted to Q(λ ).
It is not immediately obvious that choosing λ to be a low-degree algebraic inte-
ger should help our search for dynamical self-similarity (beyond the restriction that
it places on the denominator of the rotation number). Of course, it is well known
that the lowest-degree algebraic integers, solutions of quadratic equations, enjoy
algebraic self-similarity in their continued fraction expansions. Moreover, for one-
dimensional maps analogous to piecewise isometries, namely the interval exchange
transformations, one has a powerful theorem of Boshernitzan and Carroll (1997) es-
tablishing their renormalizability for quadratic irrational parameters. Unfortunately,
no comparable theorem for two-dimensional PWI’s has been proved. However, for
two-dimensional PWI’s, the renormalizability of an important class of models with
quadratic irrational λ has been rigorously established by Kouptsov et al. (2002) us-
ing computer assisted proofs. It is here that the true advantage of the restriction to
low-degree algebraic numbers makes itself felt: it makes it possible to use com-
puter software to perform exact calculations on specific models, most of which have
exceedingly complicated multi-level return map structures, thereby verifying impor-
tant properties of each model and, by exhaustion, the entire class.
Before examining three particularly interesting models from the class of PWI’s
of the square with rational rotation numbers and quadratic irrational parameters, it
will be useful to illustrate how the systematic search for renormalizable return map
structure succeeds in a particularly simple example. The contrast with the λ = 1/2
case will be striking.
We choose a model with rotation number 1/5, namely
1 √
λ = 2 cos(2π /5) = ( 5 − 1), Ω = (τ , 0]2 , τ = −λ .
2
For simplicity, we ignore here the orbits of points on the boundaries and line of
discontinuity. The atoms may then be considered as open polygons specified by
their respective lists of vertices, in square brackets:

Ω0 = (0, 0), (0, τ ), (−τ 2 , τ ) , Ω1 = (0, 0), (−τ 2 , τ ), (τ , τ ), (τ , 0) .
The partition is shown in frame 1 of Fig. 2.2.

We now explore a pathway on the return map tree which will reveal a scaling
domain at level 2. We first construct the return map for Ω 1 . The 2-step process is
shown in Fig. 2.2. Of the two level-1 atoms, we choose to construct the return map
for the triangular one. Again, the construction requires only 2 steps to terminate,
as shown in Fig. 2.3. We are now at level 2. Next, we construct the return map of
the right-hand triangle. This time 3 steps are required, as shown in Fig. 2.4. Now
we have reached level 3. Finally, we compute, in 3 steps, the level 4 return map of
the upper triangular atom, as shown in Fig. 2.5. Comparing the level 2 and level
4 dressed domains in Figs. 2.3 and 2.5, we see that they appear to be identical,
up to a rescaling. Explicit
√ calculation confirms this and gives the scale factor as
ω = 1 + τ = 12 (3 − 5).
Fig. 2.2 (a) Construction of the return map for Ω1 . (b) Atoms of the level 1 return map.
Fig. 2.3 (a) Construction of the return map for the triangular level 1 atom. (b) Atoms of the level
2 return map (gray).
Fig. 2.4 (a) Construction of the return map for the right-hand level 2 atom. (b) Atoms of the level
3 return map (gray).
Fig. 2.5 (a) Construction of the return map for the upper level 3 atom. (b) Atoms of the level 4
return map (gray).
As a final step in checking the renormalizability of this simple model, let us

verify the recursive tiling property, both for the level 2 dressed domain Δ (0) =
(D(0), ρ (0)) and its level 4 granddaughter, Δ (1) = (D(1), ρ (1)). The return K-orbits
of the atoms of Δ (0) are seen in Fig. 2.6(a) to tile all of Ω , with the exception of a
pentagon which is mapped into itself by K. Similarly, the Δ (0) orbits of the atoms
of Δ (1) are seen in Fig. 2.6(b) to tile all of D(0), apart from two pentagons which
are mapped into themselves by ρ (0). Combining the two tilings, we display the K-
orbits of the atoms of Δ (1) in Fig. 2.6(c). Again the coverage is complete apart from
the periodic pentagons.
By going further up the scaling sequence, we reveal more and more of the self-
similar return map structure, as well as more and more of the periodic domains of
the model. To see this, we plot in Fig. 2.7 the Δ (0) orbits of the atoms of Δ (4) =
(D(4), ρ (4)), where D(4) is D(0) rescaled by a factor ω 4 , and the Δ (0) return paths
are obtained by iterating 4 times the substitutions
0 → 1, 0, 0, 1 → 1, 0, 0, 1, 1.
Fig. 2.6 (a) Tiling of Ω by atoms of Δ1 . (b) Tiling of D1 by atoms of Δ2 . (c) Tiling of Ω by atoms
of Δ2 .
Fig. 2.7 Tiling of Δ(0) by atoms of Δ(4).

We can calculate from the return paths the total return times for the Δ (4) atoms,
namely 171 and 341, approximately 4 times the corresponding return times for Δ (3),
and 16 times those of Δ (2). It appears that for n → ∞, the entire area occupied by
the periodic pentagons of various sizes is tending toward the full area of the triangle
D(0), while the complementary exceptional set, covered by the Δ (L) return orbits,
is tending toward a zero-area fractal. In fact the recursive process constructing the
exceptional set via a sequence of coverings is analogous to that of a two-dimensional
Cantor set, e.g. the well-known Sierpinski gasket, and its dimension can readily be
estimated. Since the diameters of the covering triangles are decreasing uniformly
with L as ω L , while the number NL of covering triangles can be seen to be increas-
ing roughly as 4 L , the resulting fractal dimension d, defined for such Cantor set
constructions (Falconer, 1990) by
NL ∼ 4L ∼ (ω L )−d ,
gives
log 4
d=− = 1.44 . . ..
log ω
2.2 Quadratic piecewise isometries
We now treat systematically a family of piecewise isometries K for which renormal-

ization can be carried out exactly using algebraic numbers and integer arithmetic.
The family consists of maps (2.1) with the following properties:
(i) The rotation number is a rational number p/q.
(ii) The parameter λ = 2 cos 2π p/q is a root of a polynomial x 2 + bx + c, b, c ∈ Z.
We call the conjugate root λ̃ .
(iii) The square Ω is [0, τ ) 2 (resp. (τ , 0]2 ) for positive (resp. negative) τ = λ̃ −1 .
(iv) The point (0, 0) is a fixed point.
The third and fourth defining criteria may seem, at this point, somewhat arbi-
trary. They arose from the role of this particular family in models of Hamiltonian
round-off, a subject which we shall revisit in Chap. 7. In any case, they provide a
convenient set of constraints on the otherwise unmanageably large class of piece-
wise isometries satisfying the first two criteria.
2.2.1 Arithmetic preliminaries
The parameter λ is the trace of the matrix C in (2.1), and hence
λ = ξ + ξ −1, ξ = e2π ip/q ,

2.2 Quadratic piecewise isometries 27
where, without loss of generality, we may assume that p and q are relatively prime
positive integers, i.e. their greatest common divisor is 1. The fact that λ satisfies
a quadratic polynomial equation implies that ξ satisfies a quartic equation. From
number theory (Hardy and Wright, 1979), one learns that the φ (q) roots of unity
e2π ip/q with p running over the set of positive integers with gcd(p, q) = 1 are roots
of the equation
Hq (ξ ) = 0,
where Hq is the degree-φ (q) cyclotomic polynomial
Hq (ξ ) = ∏(ξ d − 1)μ (q/d) .

d|q
Here φ (q), the Euler totient function, is the number of positive integers not exceed-
ing q which are coprime to q. Moreover, μ (n), the Moebius function, is equal to 1
if n = 1, 0 if n has a square divisor greater than 1, and (−1) k if n is not divisible by
any square greater than 1 and has k prime divisors. For q > 2, one has
Hq (x) = Gq (x + x−1 )x−φ (q)/2 .
For the specific case φ (q) = 4, one can show (Kouptsov et al., 2002) that
Hq (x) = x4 − μ (q)x3 + h(q)x2 − μ (q)x + 1,
and
Gq (x) = x2 − μ (q)x + h(q) − 2, (2.4)
where, for φ (q) ≥ 4,
⎧
⎨ μ (q)(μ (q) − 1)/2 q odd,
h(q) = μ (q)(μ (q) + 1)/2 4 |q,
⎩
−μ (q/2) 4|q.
As a consequence of G q (λ ) = Gq (λ̃ ) = 0, we have
λ̃ = μ (q) − λ .
The possible values q consistent with quadratic G q are, from the preceding, those
for which φ (q) = 4, namely 5,8,10,12. In each case there will be four p values
such that gcd(p, q) = 1. Since numerators p and q − p lead to the same value of
λ = 2 cos(2π p/q), the choices of rotation number reduce to the following eight:

p 1 2 1 3 1 3 1 5
∈ , , , , , , , .
q 5 5 8 8 10 10 12 12
As the root of a polynomial of the form x d + cd−1 xd−1 + · · · + c0 , with d and

all coefficients integers, the parameter λ is an algebraic integer. In what follows,
several sets of algebraic numbers associated with λ will play an important role,
namely the ring Z[λ ], the field Q(λ ), and the Z[λ ]-module M (τ ) defined by
Z[λ ] = {m + nλ : m, n ∈ Z},
Q(λ ) = {r + sλ : m, n ∈ Q},
M (τ ) = τ Z[λ ] = {m + nτ : m, n ∈ Z},
The last of these will be of particular importance in specifying the boundaries of

domains in return map partitions, exploiting the property of closure of the module
under multiplication by elements of Z[λ ].
We note that the principal difference between Z[λ ], a ring, and Q(λ ), a field, is
the fact that elements of the former are not required to have multiplicative inverses
in Z[λ ], and so the nonzero elements do not form a group. Those elements which
do have inverses are called units, and can be shown to be roots of polynomials √
xd + cd−1 xd−1 + ·√· · + c0 for which c0 = ±1. For example,
√ in the ring Z[ 2], the
algebraic integer
√ 2 + 1 is a unit, since its inverse, 2 − 1, is also in the ring. One
checks that 2 + 1 is a root of the polynomial x 2 − 2x − 1.
In each of the models we will study with quadratic irrational λ , all units, includ-
ing the scale factor appearing in its renormalization, will be of the form ±η n , n ∈ Z,
with a unique fundamental unit η . In contrast, the units in the cubic model which we
will introduce later in this chapter are generated by two independent fundamental
units, consistent with the more intricate renormalization which takes place in that
model.
2.2.2 Domains
For the piecewise isometry K with rational rotation number p/q, the discontinuity
set Γ consists of all forward and backward images of the intersection of the x axis
with the square Ω . Given that C q is the identity matrix, it is clear that the line
segments which comprise Γ are restricted to only q possible orientations, i.e. such
a segment would lie along a line with equation
u(m) · x = b,
where u ∈ Z[λ ]2 is one of the vectors

(i) 1
u = (C ) ·T i
, i = 0, . . . , q − 1,
0
where CT is the transpose of C. We note that in the case of even q, we have
(CT )q/2 = −1,
hence the reflection symmetry

u(i+q/2) = −u(i) , i = 0, . . . , q/2 − 1.
In our investigation of the return-map hierarchy, the primary focus will be on

piecewise isometric transformations of convex sets (of dimension zero, one or two)
bounded by discontinuity lines and boundary lines of Ω . Such a convex set will be
called a domain, and we will use the notation
D(m, s, b), m = (m1 , . . . , mn ), s = (s1 , . . . , sn ), b = (b1 , . . . , bn ), (2.5)
to completely specify, in clockwise order, each boundary line in terms of an orienta-

tion index mi ∈ {0, . . . , q− 1} (corresponding to basis vector u (mi ) ), a line coordinate
bi ∈ M (τ ), and a binary index s i = ±1 distinguishing whether b i is a lower or up-
per bound for u (mi ) · x in the domain. For positive τ , the lower bounds are always
inclusive, and the upper bounds exclusive, whereas for negative τ , the opposite is
the case:
τ > 0, si = −1 : u(mi ) · x ≥ bi , τ > 0, si = +1 : u(mi ) · x < bi ,

τ < 0, si = −1 : u(mi ) · x > bi , τ < 0, si = +1 : u(mi ) · x ≤ bi .
Figure 2.8 illustrates these conventions in one of the renormalizable quadratic mod-
els.
The rectangular coordinates of a vertex of D(m, s, b) can be calculated, using
elementary linear algebra, solving simultaneously the equations u (m j ) · x = b j and
u(mk ) · x = bk for two boundary lines intersecting at the vertex. The coordinates are
in Q(λ ) but may not be elements of M (τ ). Because of the finite number of distinct
orientations, however, they are in M (τ )/a, where a is the least common multiple
of the determinants of the 2 × 2 matrices (u (m) , u(n) ), m, n ∈ {0, 1, . . ., q − 1}.
√
2.8 Periodic domain in the model with λ = − 2, illustrating the notation D (m, s, b). Since
Fig. √
τ = 2/2 > 0 and si = 1 for all i, all of the bounds are exclusive upper bounds, and so the domain
is an open octagon.
The area of a domain D with more than two vertices is easily calculated from the
vertex coordinates. Suppose that v 0 , . . . , vn−1 are the vertices of a domain, arranged
in clockwise order. The area of the domain is then
1 n−2
A (D) = ∑ (vi+1 − v0 ) × (vi − v0 ),
2 i=1
where the cross product a × b of a pair of 2-vectors is defined as the scalar quantity
ax by − ay bx .
2.2.3 Geometric transformations on domains
All of the dynamical mappings, scale transformations, and equivalence relations of

domains relevant to our investigations belong to a group G of geometric transfor-
mations generated by a generalized rotation C (corresponding to application of the
matrix C), a generalized reflection R (provided that q is even), a scale transforma-
tion Sω (where ω is a unique element in Z[λ ]), and a translations Tδ by an arbitrary
2-vector δ in Z[λ ] 2 . The operations of these generating group elements on domains
are the following:
CD(m, s, b) = D(m , s, b), mk ≡ mk − 1 mod q

q/2 − 1 − mk, 0 ≤ mk < q/2
RD(m, s, b) = D(m̃, s, b), m̃k =
3q/2 − 1 − mk, q/2 ≤ mk < q
Sω D(m, s, b∞ + b) = D(m, s, b∞ + ω b), ω a unit in Z[λ ], ω > 0, b∞ ∈ Q(λ )2
Tδ D(m, s, b) = D(m, s, b ), δ ∈ Z[λ ]2 , bk = bk + u(k) · δ .
We note that the scale transformations act on line coordinates relative to those of the
scaling limit point. In the models under consideration, there are at most two such
points. By choosing a frame of reference with the scaling limit at the origin, we can
simplify the notation, eliminating the references to b ∞ in the scaling formulas.
The existence of inverses of the listed operations, needed to establish the group
property of G , is readily established. The fact that we only allow scaling by algebraic
units is crucial here.
From the definitions, one easily sees that the generating elements of G have the
following composition and commutation relations:
Tδ ◦ Tε = Tδ +ε , Sω ◦ Sω = Sωω , R ◦ R = 1,
Sω ◦ C = C ◦ Sω , Sω ◦ R = R ◦ S ω , R ◦ C = C−1 ◦ R,
G ◦ Td = TG(d) ◦ G, G ∈ {C, R, Sω },
where, in the last relation, G(d) refers to the pointwise action of G on the 2-vector d,
for example C(d) = C · d. Using the above formulas to reorder and combine factors,
one can write any element G ∈ G in the canonical form
G(κ , ν , d, ω ) = Sω ◦ Td ◦ Cν ◦ Rκ .
2.2.4 Scaling sequences
In Sect. 2.1 we introduced the basic ideas of return map dynamics and discussed
informally the concept of scaling sequences of dressed domains in renormalizable
models. We now provide a more precise definition and prove a theorem that will
allow us to establish their central role in the quadratic models.
Definition 2.1. Let x ∗ be a point of Ω and ω a unit in Z[λ ] with magnitude < 1.
A scaling sequence with limit point x ∗ and contraction factor ω is a sequence of
dressed domains Δ (L) = (D(L), ρ (L)), L = 0, 1, . . .} such that x ∗ ∈ D(L) and, if
{(D j (L), ρ j (L)) : j = 0, . . . , J} is the partition of Δ (L) associated with return map
ρ (L), then, for all j,
D j (L) = Sω D j (L − 1), (2.6)

ρ j (L) ◦ Sω D j (L − 1) = Sω ◦ ρ j (L − 1)D j (L − 1), (2.7)
ρ j (L)D j (L) = ρ jν −1 (L − 1) ◦ · · · ◦ ρ j0 (L − 1)D j (L), (2.8)
where ν and j0 , . . . , jν −1 are functions of j, independent of L,
ν = ν j, jt = p( j,t), t = 0, . . . , ν j − 1. (2.9)
The path function p( j,t) associated with a scaling sequence specifies the level-
independent return paths of the domains D j (L). Not surprisingly, it will be of prime
importance when we introduce a more abstract distillation of the return map dynam-
ics in the next chapter.
The spatial scaling of domains in the definition of a scaling sequence is inevitably
tied up with temporal scaling as well, as we saw in the example of Sect. 2.1. Clearly
the level-L return times T j (L) are related to those of level L − 1 by
ν j −1
T j (L) = ∑ Tjt (L − 1).
t=0
This can be reduced to a concise linear recursion relation by introducing the inci-
dence matrix Ai j , which counts the number of times the return path of D j (L) passes
through domain D i (L − 1):
ν j −1
T j (L) = ∑ Ti (L − 1)Ai j , Ai j = ∑ δi,p( j,t),
i t=0
where δi, j is the Kronecker delta. Asymptotically, the return time scales as α L , α >
0, where α is the largest eigenvalue of the transpose of the incidence matrix.
To exploit the concept of scaling sequences, we obviously need an efficient way
of identifying them in specific models. This is provided by the following lemma
(Kouptsov et al., 2002):
Lemma 2.1. Let ω a unit in Z[λ ] with magnitude < 1 and let S ω be the corre-
sponding scaling operator centered at the point x ∗ . Further, let D be a domain such
that x∗ ∈ D̄, and suppose that ρ is the first return map for D, partitioning it into
subdomains D j , j = 0, . . . , J − 1. The scaled domains D(L) = S ω L D and isometries
ρ (L) = Sω L ◦ ρ ◦ Sω −L , for L = 0, 1, . . . , form a scaling sequence with origin x ∗ and
contraction factor ω , provided that ρ (1) is a genuine return map for D(1), and the
respective partitions for levels 0 and 1 satisfy the renormalization equations (2.6–
2.8).
Proof. The proof is by induction. We assume that ρ (L) and ρ (L+ 1) are return maps
for D(L) and D(L + 1), respectively, such that the renormalization equations are
satisfied for the level-L and L + 1 partitions. We need to prove the same properties
for levels L + 1 and L + 2.
We begin with the verification of (2.8) (the other two equations follow immedi-
ately from the definitions in the statement of the lemma). Thus, using the induction
hypothesis,
ρ j (L + 2)D j (L + 2) = ρ j (L + 2) ◦ Sω D j (L + 1)
= Sω ◦ ρ j (L + 1)D j (L + 1)
= Sω ◦ ρ jν −1 (L) ◦ · · · ◦ ρ j0 (L)D j (L + 1) (2.10)
= ρ jν −1 (L + 1) ◦ · · · ◦ ρ j0 (L + 1) ◦ Sω D j (L + 1)
= ρ jν −1 (L + 1) ◦ · · · ◦ ρ j0 (L + 1)D j (L + 2).
Note that the third equality is obtained by commuting S ω through the chain of
partial return maps ρ j (L). This is not simply a matter of expressing ρ j (L) in the
form of an isometry Tδ Cν and applying the group commutation relations to obtain
Sω ◦ Td ◦ Cν = Tω d ◦ Cν Sω .
One must also check, as we have done, that the range of each mapping in the chain
is contained in the domain of the mapping to the left.
To complete the proof, we need to show that ρ (L + 2) is in fact a first return map
for D(L + 2). From (2.10) and the induction hypothesis, ρ j (L + 2) decomposes into
ν j linked orbits of the elementary map K, with a total return time of
νk −1
T j (L + 2) = ∑ Tk (L + 1).
k=0
We must show that each of the T j (L + 2) domains on the orbit, with the exception
of the initial and final one, is disjoint from D(L + 2). From the induction hypothesis
all of the relevant domains, with the exception of
Ek (L + 2) = ρ jk (L + 1) ◦ · · · ◦ ρ j0 (L + 1)D j (L + 2), 0 ≤ k < ν j − 1,
are disjoint from D(L + 1), hence from D(L + 2) ⊂ D(L + 1). The E k (L + 2) are in
D(L + 1), but
Ek (L + 2) ∩ D(L + 2) = Sω Ek (L + 1) ∩ D(L + 1) = 0.
/

2.2.5 Periodic orbits
Suppose that Π is a periodic domain contained in the lowest-level member D(0)

of a scaling sequence. This means that K τ Π = Π for some (nonzero, minimal)
return time τ . Without loss of generality, we assume Π ⊂ D j0 (0) for some j0 ∈
{0, . . . , J − 1}, so that there exists a level-0 return orbit
Π = ρ jν −1 ◦ · · · ◦ ρ j0 Π
for some finite sequence of indices j k ∈ {0, . . . , J − 1}, k = 0, . . . , ν − 1. Clearly

Π can be promoted to an infinite sequence of periodic domains Π (L) by multiple
application of the scaling operator S ω :
Π (L) = SωL Π , L = 0, 1, . . . ,
with level-L return orbit
Π (L) = ρ jν −1 (L) ◦ · · · ◦ ρ j0 (L)Π (L)
and return time

ν −1
τ (L) = ∑ Tjk (L).
k=0
Here ν and the return path j 0 , . . . , jν −1 depend on Π , but not on the level L.
We note that there is a distinction between the return time of a periodic domain
and the minimal period of its points under K. If a periodic domain Π has rotational
symmetry (i.e. C μ Π = Π for μ < q), the centroid (mean position of the vertices)
will have period τ , while the period of all other points in Π will be the least common
multiple of τ and q.
2.2.6 Recursive tiling
As pointed out earlier in this chapter, the existence of a scaling sequence is essential
to what we mean by renormalizability of a piecewise isometric system, but it may
not be the whole story. There must be some sense in which the return map dynamics
gives a complete description of the dynamics on Ω , at least on the complement of
the discontinuity set. We now turn to a precise definition of this concept, which we
call the recursive tiling property.
A piecewise isometric mapping of the square Ω is said to satisfy the recursive
tiling property if there exist finitely many scaling sequences
{(D(σ ) (L(σ ) ), ρ (σ ) (L(σ ) )) : L(σ ) = 0, 1, . . .}, σ = 0, . . . , N − 1,

(σ )
such that, for any choice of levels L (σ ) , the return orbits of the subdomains D j (L(σ ) ),
(σ )
together with those of the periodic domains Π k (M (σ ) ) of level M (σ ) < L(σ ) and
a finite number of exceptional periodic domains, completely tile Ω , apart from the
measure-zero set of domain boundaries. Since orbits on the discontinuity set can be
ignored, the task of establishing the recursive tiling property reduces to summing up
the areas of the tiles comprising the return orbits to check that the total is equal to
τ 2 , the area of Ω .
In the present chapter we will emphasize that a stronger definition of recursive
tiling is within reach, provided that one is willing to expend the (considerable) ad-
ditional effort to do a proper accounting of the return orbits on the discontinuity set.
Two quadratic models, where this has been achieved (Kouptsov et al., 2002), will
be studied in detail below. The following definitions and lemma state a concise set
of tests for the strong form of recursive tiling.
Definition 2.2. Given a line segment γ in the plane, the skew length of γ , written
L (γ ), is the Q-metric length of the segment.
Definition 2.3. The weighted perimeter function W (D) assigns to a domain D a

number in Q(λ ) as follows:
1. If D is an isolated point, W (D) = 0.
2. If D is a line segment , W (D) is its skew length, L (D).
3. If D is a nondegenerate polygon, W (D) is a weighted sum over the skew lengths
of its edges, with weighting factor 1/2 if the edge is included in D, and −1/2 if
it is excluded.
We note that with the representation of a polygonal domain as D(m, s, b), the inclu-
sion or exclusion of an edge can be read off from the sign of the relevant component
si of s.
Lemma 2.2. To prove the recursive tiling property (strong form) of a quadratic
model equipped with a scaling sequence (D(L), ρ (L)), L = 0, 1, . . . with a finite
number of associated sequences of periodic domains Π j (L), and a finite number
of non-scaling periodic domains Π j0 (Ω ) whose orbits are disjoint from D(0), it is
sufficient to verify the following sum rules:
1. Area sum, tiling of Ω .
J−1 J0 −1
∑ ν j (0)A (D j (0)) + ∑ ν̃ j0 (Ω )A (Π j0 (Ω )) = A (Ω ).
j=0 j0 =0
2. Area sum, tiling of D(0).
J−1 J˜
∑ ν j (1)A (D j (1)) + ∑ ν̃ j˜(0)A (Π j˜(0)) = A (D(0)).
j=0 j˜=0
3. Weighted perimeter sum, tiling of Ω .

J−1 J0 −1
∑ ν j (0)W (D j (0)) + ∑ ν̃ j0 (Ω )W (Π j0 (Ω )) = W (Ω ).
j=0 j0 =0
4. Weighted perimeter sum, tiling of D(0).
J−1 J˜
∑ ν j (1)W (D j (1)) + ∑ ν̃ j˜(0)W (Π j˜(0)) = W (D(0)).
j=0 j˜=0
5. Vertex coverage, tiling of Ω . The orbits of the D j (0) and Π j˜(Ω ) must cover all
vertices not lying on the excluded boundaries of Ω .
6. Vertex coverage, tiling of D(0). The orbits of the D j (1) and Π j˜(0) must cover all
vertices not lying on the excluded boundaries of D(0).
For the weak form of recursive tiling, the area sums suffice.
Proof. Thanks to the scaling of areas and lengths, it will be sufficient to prove (a)
the tiling of Ω by the return orbits of D j (0), j = 0, . . . , J and of the periodic do-
mains which do not intersect D(0), and (b) the tiling of D(0) by the return orbits
of D j (1), j = 0, . . . , J and of the periodic domains of D(0) which do not intersect
D(1). Since the union of relevant return orbits in Ω comprises a set of mutually
disjoint convex polygons, the area sum rules imply coverage up to a set restricted
to the domain boundaries, not including those which lie along the excluded bound-
aries of Ω , i.e. along x = τ or y = τ . The weighted perimeter sums, on the other
hand, guarantee full coverage of the domain boundary segments, apart from a finite
number of isolated points. Having complete coverage of such vertices completes the
proof of 100% coverage. Analogous arguments establish the complete coverage of
D(0).

For models with more than one scaling sequence (such as Model II below),
Lemma 2.2 can immediately be generalized. The modifications to the statement
and proof are obvious, and we omit the details. Another rather simple, but use-
ful, generalization (needed for Model III below) involves the technique of tele-
scoping. This refers to the situation where we have a sequence of dressed do-
mains Δ (L), L = 0, 1, 2, . . . , which possess the recursive tiling property but where
the renormalization equations relate non-consecutive levels. For example, for some
fixed k, levels L and L+ k, L = 0, 1, 2, . . . , may be so linked. In such a case, we might

have k inequivalent scaling sequences, all governed by the same scale factor ω . Pro-
vided that the full sequence satisfies the area and perimeter sum rules of Lemma 2.2,
not just for levels 0 and 1, but for levels 0, . . . , k, it is clear that the hierarchy can be
collapsed (“telescoped”) into a true scaling sequence
Δ̂ (N) = Δ (N k), N = 0, 1, 2, . . . .
This sequence will satisfy all the hypotheses of Lemma 2.2, and hence will be recur-
sively tiled with respect to the telescoped sequence of levels. Of course, the return
paths may become extraordinarily large for the telescoped hierarchy, which is why
one might prefer to work with the original one with its leap-frogging scaling rela-
tions. This will be our preference in the formulation of Model III below.
The union of all return orbits initiated in the base domain D(0) of a scaling se-
quence is an invariant set containing, in the models of interest, both periodic and
aperiodic points. The set of aperiodic points on discontinuity-avoiding orbits is ob-
viously contained in every set of the form
ν j −1

J−1
C (L) = ρ (L − 1)t D j (L),
j=0 t=0
and hence the area occupied by these points is bounded above by the area A (C (L))
for all L. But
A (C (L)) ∼ (ω 2 α )L for L → ∞,
where α is the largest eigenvalue of the incidence matrix. In all of the quadratic
models ω 2 α < 1, and so the aperiodic points, including those on the discontinuity
set, occupy a set of measure zero. Consequently, the periodic domains within a
scaling sequence have full measure. The same property extends to the full set of
periodic domains within Ω : they tile the square apart from a set of measure zero.
In what follows, we will refer to the zero-measure complement of the periodic
domains as the exceptional set, and to the subset of discontinuity-avoiding aperiodic
points as the residual set. Due to the decrease without limit of the constituent do-
mains of the covering sets C (L) as L increases, we conclude that each point in the
residual set is arbitrarily close to a point of the discontinuity set Γ , and hence the
exceptional set can be identified with the closure of Γ .
2.2.7 Computer-assisted proofs
In Kouptsov et al. (2002), nine quadratic models were shown to be renormalizable

in the sense discussed above. Eight of these comprise the family of piecewise iso-
metric models described at the beginning of Sect. 2.2, motivated by the role of these
maps in models of Hamiltinian round-off (Lowenstein et al., 1997; Lowenstein and
Vivaldi, 2000). The ninth, which differs from the rest in having the fixed point at the
center of Ω rather than at the corner (0, 0), was the subject of an important article
by Adler et al. (2001) on two-dimensional piecewise isometries.
Computer assisted proofs were given by Kouptsov et al. (2002) to establish the
existence of a complete set (in the sense of recursive tiling) of scaling sequences in
all nine cases. A summary of the most important parameters is given in Table 2.1.
Note that our parameter λ , which appears in the generalized rotation matrix C, is
the algebraic conjugate of the parameter λ of Kouptsov et al. (2002). This reflects
the shift of emphasis from round-off models to piecewise isometries and kicked
oscillator maps. The last column of the table lists the asymptotic temporal scaling
factor α , the largest eigenvalue of the transpose of the incidence matrix A.
Table 2.1 Scaling parameters for the renormalizable quadratic models of Kouptsov et al. (2002).
λ p/q τ Ω ω x∗ α
√ √
(−1 + 5)/2 1/5 (1 − 5)/2 (τ , 0]2 1+τ (0, 0) 4
√ √
(−1 − 5)/2 2/5 (1 + 5)/2 [0, τ )2 2−τ (0, 0) 4
√ √
(1 + 5)/2 1/10 (−1 − 5)/2 (τ , 0]2 2+τ (0, τ ) 4
√ √
(1 − 5)/2 3/10 (−1 + 5)/2 [0, τ )2 1−τ (τ , τ ) 4
√ √
− 2 1/8 2/2 [0, τ )2 3 − 4τ (0, τ ) 9
√ √
2 3/8 − 2/2 (τ , 0]2 3 + 4τ (τ , τ ) 9
√ √
3 1/12 − 3/3 (τ , 0]2 2 + 3τ (τ , τ ) 4
√ √
3 1/12 − 3/3 (τ , 0]2 2 + 3τ (41 + 72τ , 41 + 72τ ) 4
√ √
− 3 5/12 3/3 [0, τ )2 7 − 12τ (τ , τ ) 25
√ √
2 3/8 − 2 (τ , −τ ]2 3 + 4τ (−1 − τ , τ ) 9
Before getting into a detailed review of some of these results, it is important

to make clear the role which computer hardware and software played in obtaining
them. This is particularly important in light of our claim that the principal statements
regarding renormalization deserve to be treated as possessing a level of mathemat-
ical rigor comparable to that of other propositions proved without the assistance
of electronic devices. Computer calculations enter into our investigations at three
stages as follows
Phase 1: exploration
Each of the piecewise isometric models which we deal with is a highly complex
mathematical environment which, in the first stage, we explore using the least so-
phisticated tools: direct iteration of mappings using machine-precision floating-
point representations of the real numbers, motivated and assisted by computer
graphics. The existence (or not) of self-similarity (i.e. apparent fractal structure in
the phase portrait), as well as the likely locations of scaling fixed points, are usually,
but not always apparent at this stage.
Phase 2: derivation
This is essentially the heuristic approach described informally at the beginning of
this chapter, carried out, however, without making numerical approximations. This
entails treating τ as an abstract symbol, with polynomials of degree greater than two
reduced to the form r + sτ , r, s ∈ Q via application of the quadratic equation satisfied
by τ . In order to efficiently carry out the construction of fairly extensive branches of
the various return map trees, specialized software in the C programming language
was developed. This included a library of functions to handle the arithmetic of ar-
bitrarily large integers, as well as of algebraic numbers in Z[λ ] and Q(λ ). Other
functions were invented to handle the mapping and intersection of domains. Finally,
programs were developed to construct recursively the return map partitions of any
level L.
Phase 3: computer assisted proof
While the constructions of Phase 2 reliably produce a detailed accounting of the
scaling sequences and periodic domains of the nine models listed in Table 2.1, they
do not constitute a computer assisted proof, even if they are a meticulously executed
computational experiment. The difference lies in the relative transparency of the
arithmetic and algebraic procedures. For example, as part of a non-computational
proof, a mathematician would find acceptable carrying out “by hand” any number
of elementary operations involving the addition, subtraction, multiplication and di-
vision of rational numbers and polynomials with rational coefficients. Presumably
he or she would have no objections to having a commercially available computer
running (say) Mathematica carry out the same set of procedures as part of a proof.
Should the same level of trust extend to a custom-designed function used to calcu-
late the intersection of two domains, yielding reliable results even in cases where
the intersection is a line segment or a single point with coordinates in Q(λ )? Pre-
sumably not.
Fortunately, it is not necessary for the proof of a mathematical theorem to be
constructive. We need only be sufficiently careful in using the computer to verify
the multitude of results obtained by the less than rigorous methods of Phase 2. We
do this using a small number of very short procedures, each of which is sufficiently
simple and transparent that it could easily be performed “by hand”. The role of com-
puter is to perform thousands of these elementary operations in succession within a
fraction of a second (rather than the much more extended time period, sometimes
exceeding the age of the universe, needed for a human being to do the same). What
makes this simplicity possible is the fact that once return paths and the subdomains
of a return map partition have been specified in detail, to prove the correctness of
those return paths and partition involves only mapping and checking inclusion rela-
tions, but no intersections of domains and no branching of orbits.
For the proof of the existence and derived properties of the listed scaling se-
quences and their associated periodic domains, the following elementary procedures
(all but two of which are made explicit as Mathematica functions in an appendix of
Kouptsov et al. (2002)) are employed:
signum(a + bτ ) Returns 0 if a = b = 0, 1 if a + bτ > 0, and −1 if a + bτ < 0.

vertex(m1 , m2 , b1 , b2 ) Calculates the intersection point of the lines u (m j ) · x =
b j , j = 1, 2.
vertexconstraint(v, m, s, b) Returns −1, 1, or 0, depending on whether the
vertex v lies inside, outside, or on the boundary of the half-plane su (m) · x < sb.
vertices(D) Returns the set of vertices of the domain D, in cyclic order.
disjoint(D1 , D2 ) Returns True if domains D 1 and D2 are disjoint.
subdomain(D, D1 ) Returns True if domain D is a subdomain of domain D ,
False otherwise.
equal(D, D ) Returns True if domain D is equal to domain D , False other-
wise.
area(D) Returns the (Euclidean metric) area of domain D.
length(x, y) Returns the (Q-metric) length of the line segment connecting x
to y.
perimeter(D) Returns the (Q-metric) perimeter of the domain D.
rescale(a, D) Returns a domain equal to domain D with all bounds rescaled
by a factor a ∈ Z[λ ].
rotate(ν , D) Returns Cν D.
translate(d, D) Returns Td D.
map(ν , d, D) Returns Td ◦ Cν D.
phimap(D) Implements the map K on the domain D.
rho(L, j, D) Implements the map ρ j (L) on the domain D.
returntest1(L, j) Tests whether ρ j (L), as defined in terms of iterations of
K, is in fact the return map for D j (L). Returns True or False.
returntest2(L, j) Tests whether ρ j (L), defined in terms of iterations of
ρ (L − 1), is in fact the return map for D j (L). Returns True or False.
returntest3(L, j) Tests whether Π j (L) is a periodic domain with a specified
return path. Returns True or False.
The only one of these functions which is not straightforward to program is

signum(a + bτ ) which we must implement without resorting to a floating-point ap-
proximation of the quadratic irrational τ . The trick is to use the continued-fraction
expansion (Khinchin, 1964) of τ , whose even and odd convergent satisfy the recur-
sion relations, for k > 0,
n2k f0 n2k−1 + n2k−2 n2k+1 f1 n2k + n2k−1
= , = ,
d2k f0 d2k−1 + d2k−2 d2k+1 f1 d2k + d2k−1
with recursion coefficients and initial approximants given in Table 2.2:

Table 2.2 Recursion coefficients and initial approximants.
p/q f0 f1 n0 d0 n1 d1
1/5 1 1 −1 1 −1 2
2/5 1 1 1 1 2 1
1/8 2 2 −1 1 −2 3
3/8 2 2 1 1 2 3
1/10 1 1 −1 1 −2 1
3/10 1 1 1 1 1 2
1/12 2 2 −1 1 −1 2
5/12 2 2 1 1 1 2
The rational approximants m 2k = a + b n2k /d2k and m2k+1 = a + b n2k+1 /d2k+1

bracket a + bτ , and so to calculate signum(a + bτ ) for b = 0, it suffices to keep
iterating the recursion relation until m 2k and m2k+1 are found to have the same sign.
The latter then gives the desired sign of a + bτ .
The two functions added to the original toolbox of the “computer assistant” of
Kouptsov et al. (2002) are length and perimeter. These will be needed below in
order to establish the strong form of recursive tiling in two of the models introduced
below. In these procedures, the measurement of the Q-distance between two end-
points of a line segment is measured by first rotating the segment by application of
the matrix C until it is horizontal, then taking the difference of the x components.
This works, since C-rotation leaves the Q-distance unchanged, and on a horizontal
line Euclidean distance coincides with Q-distance.
2.3 Three quadratic models
We now examine in detail three piecewise isometric mappings of the square with
quadratic irrational parameters. The first of these, Model I, is fairly representative
of the quadratic models with rotation number p/q, q < 12. The second, Model II,
exhibits the highest level of complexity and the least transparent renormalization
structure of any of the quadratic examples studied, and thus is an excellent guard
against too facile generalizations.
The third model, Model III, originally introduced by Adler et al. (2001) and
further developed by Kouptsov et al. (2002), is unique among our examples in its
symmetrical placement of the rotation center, and provides a beautiful example of
rich dynamics on the discontinuity set. Apart from their intrinsic interest as piece-
wise isometries of the square, all three of the selected models are equally fascinating
when lifted to the infinite plane as kicked-oscillator maps. We will leave this part of
the story until later chapters.
In the following pages we present in a concise format the core data of the three
models, with accompanying figures to help the reader gain an intuitive appreciation
of the models without having to wade through the details. The full data sets needed
for computer assisted proofs of renormalizability, as well as the kicked-oscillator
2.3 Three quadratic models 41
applications of later chapters, have been placed in Appendix A. These data are in
one sense more elaborate than one might think necessary: the partition of Ω into
subdomains (the so-called generating partition) is carried out in such a way that
all two-dimensional domains are open, i.e. do not contain their any of their edges.
Similarly, the one-dimensional domains do not contain their endpoints. This greatly
facilitates the study of the discontinuity-set dynamics and is well worth the extra
bookkeeping.
2.3.1 Model I
Parameter summary
λ
√ p/q τ√ Ω ω x∞ α
(1 + 5)/2 1/10 (−1 − 5)/2 (τ , 0]2 2+τ (0, τ ) 4
Orientation co-vectors
Map K on the generating domain Ω

Level L atoms, return paths
Level L periodic domains, return paths
Incidence matrix ⎛ ⎞
1 2 2 4 4
⎜0
⎜ 0 0 0 0⎟⎟
A=⎜
⎜1 1 1 1 1⎟⎟
⎝0 1 0 1 0⎠
0 0 1 2 3
Recursive tiling of Ω by atoms of D(0)
Recursive tiling of D(0) by atoms of D(1)
2.3.2 Model II
Parameter summary
λ p/q τ Ω ω x∞ α
√ √ (τ , τ ) 4
3 1/12 − 3/3 (τ , 0]2 2 + 3τ
(41 + 72τ , 41 + 72τ ) 5

Return map partitions of D(A) (0) and D(B) (0)

Tiling of D(A) (0).1
Tiling of D(B) (0).1
1 See color figures at the end of the book.

2.3.3 Model III

Parameter summary
λ p/q τ Ω ω x∞ α
√ √
2 1/8 − 2/2 (τ , −τ ]2 3 + 4τ (−1 − τ , τ ) 9
Prescaling domain
Level L atoms
Incidence matrix
⎛ ⎞
1 13 35 39 50 49 12 23 24
⎜0 1 3 2 3 3 0 1 0 ⎟
⎜ ⎟
⎜0 0 1 0 0 0 0 0 0 ⎟
⎜0 0 1 1 1 1 0 1 0 ⎟
⎜ ⎟
A=⎜ ⎟
⎜0 0 0 0 1 0 0 0 0 ⎟
⎜0 2 3 2 3 2 0 1 0 ⎟
⎜ ⎟
⎜ 0 2 6 8 10 12 5 6 8 ⎟
⎝0 1 2 1 1 0 0 1 0 ⎠
1 1 1 1 1 1 1 1 1
Level L periodic domains
Recursive tiling
2.4 Proof of renormalizability
The proof of the renormalizability of Models I, II, and III consists of verifying, with
computer assistance, the hypotheses of Lemmas 2.1 and 2.2. The explicit specifi-
cation of the relevant domains and (alleged) return mappings of levels 0 and 1 is
contained in Appendices A.1, A.2 and A.3. For the proof of scaling, it remains only
to verify, through straightforward application of the toolbox functions, that ρ (0) and
ρ (1) are true first return maps with corresponding partitions of Ω , D(0), and D(1)
which satisfy the renormalization equations. We have done this, obtaining affirma-
tive outcomes for all relevant tests. The reader interested in the details is referred
to the appendix of Kouptsov et al. (2002), as well as to the electronic supplement
linked to that article.
For the strong recursive tiling proofs in Model I, the areas, weighted perimeters,
and single-level return times (ν j or ν̃ j˜) needed for application of Lemma 2.2 are
collected in Table 2.3. In addition, with the help of the function vertices(D), we
Table 2.3 Tiling data for Model I.

2.5 Structure of the discontinuity set 49
find that there are 38 distinct vertices of the nondegenerate polygons in the partition
of Ω which do not lie on excluded boundaries or on interior points of included
segments. The analogous vertex count for D(0) is 6. The verification of the four
sum rules and two vertex counts is then completely straightforward.
In Model III, the strong recursive tiling is proved in similar fashion, making use
of the data table A.3.6 in Appendix A. The weak recursive tiling property for Model
II is simpler, requiring only verification of the area sum rules in Lemma 2.2.
2.5 Structure of the discontinuity set
2.5.1 Model I
An important byproduct of the strong recursive tiling of Ω is a comprehensive anal-

ysis of the structure of the discontinuity set Γ . For Model I, this structure is par-
ticularly simple: Γ is the union of the orbits of a countable collection of periodic
domains, with a self-similar baseline on the boundary of Ω . As a preliminary to
establishing this, we introduce the notion of concatenation.
Definition 2.4. Let A be an oriented line segment (open, closed, or half-open), and
let Ai , i = 1, . . . , n be a collection of disjoint oriented segments and points whose
union is A. We say that A is the concatenation of A 1 , . . . , An , written
A = A 1 A2 · · · An ,
if for each i = 1, 2, . . . , n − 1, the final boundary point of A i coincides with the initial
boundary point of A i+1 . By convention, the initial and final boundaries of a point
coincide with the point itself.
We next prove the following lemma.
Lemma 2.3. Every point in the discontinuity set Γ belongs to an orbit containing a
point on the segment B = Ω 6 Ω8 = {(x, y) : x = 0, −1 < y ≤ 0}.
Proof. The definition of Γ already establishes that all of its orbits intersect the in-
cluded part of the boundary of Ω at least once. To restrict the latter to the segment
B, we easily verify:
Ω5 ⊂ K −1 (Ω6 Ω9 Ω7 ), Ω7 Ω9 ⊂ K 5 (Ω6 ).

Our main result for the structure of Γ is then

Theorem 2.1. Each point of Γ lies on a periodic orbit passing through a unique
point of B, the baseline. The latter is partitioned into the periodic points Pn =
(0, −|λ |n ), n = 1, 2, . . . and the periodic open segments I n = (Pn , Pn + 1), n = 0, 1, . . .
together with the fixed point at the origin. Each periodic point or segment lies on
the return orbit of one of the scaling or pre-scaling domains, Π j (L) or Π j (Ω ).
Specifically,
(A) P1 , P2 , P3 , I0 , I1 , I2 are, in order, Π3 (Ω ), Π0 (Ω ), Π2 (Ω ), Π5 (Ω ), K 8 Π0 (0), Π1 (Ω ).
(B) For L ≥ 2, P2L and P2L+1 lie on the orbits of Π1 (L − 2) resp. Π3 (L − 2).
(C) For L ≥ 2, I2L−1 and ILn lie on the orbits of Π0 (L − 1) resp. Π2 (L − 2).
Proof. Direct calculation establishes (A), as well as the decomposition
B = (P0 , P3 ] K 12 D3 (0) (0, 0).
From the latter, we see that B inherits the recursive tiling of D 3 (0). Specifically, for
any level L,
D3 (L) = ρ (L)2 Π0 (L + 1) Π1 (L) Π2 (L) Π3 (L) ρ (L)4 D3 (L + 1),
which follows from the data tables in Sect. A.1 of the Appendix. Statements (B) and
(C) follow easily. The geometrical relations among the various points and segments
is shown in Fig. 2.9.

Fig. 2.9 Model I scaling domains (a) whose orbits intersect the discontinuity-set baseline B,
shown in (b). The coordinates in (a) are relative to the scaling limit point at (0, τ ).
2.5.2 Model III
For Model III, the discontinuity set has a much richer structure than that of Model I.
In particular, Model I’s discontinuity set contains no aperiodic orbits, that of Model
III is host to an uncountable infinity of them (a Cantor set). The recursive inclusion
2.6 More general renormalization 51
structure of the baseline B = Ω 7 Ω8 Ω9 Ω10 is laid out in great detail

in Sect. A.3.7 of the Appendix. The data in this table were established rigorously
by Kouptsov et al. (2002) using the same type of computer assisted proof used
to prove the renormalizability of the model, making use of the same toolbox of
elementary Mathematica functions. In the table, each segment of the baseline is
listed as the image of some atom or periodic domain of the scaling sequence. From
this information, one can easily infer the length and up/down orientation of any
segment.
A notable feature of the table is the following: apart from a single image of the
prescaling domain Π 1 (−1), it is only the scaling sequence Π 3 (L) ⊂ D7 (L), L =
0, 1, 2, . . . which participates in the recursive tiling. This does not prevent full cov-
erage, up to a set of measure zero, of the baseline by periodic segments, a result
which we now verify by explicit computation. At level L ≥ 0, the baseline is tiled,
up to a set of measure zero, by the periodic domains of level less than L and sets
of νk (L) open segments congruent to D k (L), L = 1, 3, 5, 7. From the table, one can
easily read off the following recursion rules for determining ν k (L):
⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎛ ⎞
ν1 (0) 6 ν1 (L + 1) 1 0 2 1 ν1 (L)
⎜ ν3 (0) ⎟ ⎜ 9 ⎟ ⎜ ν3 (L + 1) ⎟ ⎜ 2 1 2 1 ⎟ ⎜ ν3 (L) ⎟
ν (0) = ⎜ ⎟ ⎜ ⎟
⎝ ν5 (0) ⎠ = ⎝ 9 ⎠ ,
⎜ ⎟ ⎜ ⎟⎜
⎝ ν5 (L + 1) ⎠ = ⎝ 3 1 2 0 ⎠ ⎝ ν5 (L) ⎠ .
⎟
ν7 (0) 6 ν7 (L + 1) 1 1 1 1 ν7 (L)
Since ν (0) is an eigenvector of the recursion matrix with eigenvalue 5, we immedi-

ately get
ν7 (L) = 6 · 5L, L = 0, 1, 2, . . ..
Inserting (from the data tables)the lengths
√
|Π3 (L)| = ω L+2 , |Π1 (−1)| = ω , ω = 3 − 2 2,
we can sum up to get

∞
1+ω √
ω + 6ω 2
∑ (5ω ) L
=ω
1 − 5ω
= 2,
L=0
which is equal to the length of B.

In Chapter 4, we will return to this model to study the fractal properties of the
set C obtained by removing all of the periodic domains from B. Then, in Chaps. 5
and 6, we will explore the remarkable long-time behavior of kicked oscillator orbits
which originate in C .
2.6 More general renormalization
For piecewise rational rotations of the square with quadratic irrational parameters,
renormalization built around scaling and recursive tiling provides a comprehensive
hierarchical framework for understanding the periodic and aperiodic orbits of the
model. Going beyond the quadratic models requires an expanded concept of renor-
malization. Such a framework, again based on recursive tiling, was developed by
Lowenstein et al. (2004) to handle the intricacies of the so-called π /7 model with
cubic λ = 2 cos π /7 and will be the subject of the current section. How large a class
of models can be accommodated by the formalism remains to be seen.
Before introducing the extended notion of recursive tiling, we first need to make
more precise for the idea of equivalence of domains:
Definition 2.5. Two dressed domains Δ 1 = (D1 , ρ1 ) and Δ 2 = (D2 , ρ2 ) are equiva-
lent, written Δ 1 ∼ Δ2 , if they differ by a mapping G ∈ G , where G is the group of
geometric transformations defined in Sect. 2.2.3. Specifically,
Δ1 ∼ Δ2 ⇐⇒ ∃G ∈ G , D2 = GD1 , ρ2 = G ◦ ρ1 ◦ G−1.
In our discussion of the quadratic models, the notion of equivalence entered im-
plicitly, through the definition and renormalization properties of a scaling sequence.
In all of the cases studied, the equivalence transformations were scale transforma-
tions with respect to a fixed point x ∞ , involving only the subgroup generated by
translations and scale transformations. In the more general scheme to be introduced
below, we will remove this restriction and allow our equivalence relations to involve
C-rotations and R-involution as well.
We now introduce our generalized recursion scheme. Suppose we have a cata-
logue of N inequivalent dressed domains, Δ [n] = (D[n] , ρ [n] ) n = 0, 1, . . . , N. Sup-
pose further that each catalogue member has one or more dressed sub-domains
Δ (i) = (D(i), ρ (i)), i = 0, 1, . . . , I [n] , each of which is equivalent to some member
of the catalogue. Note that the index i labels a link between parent and child dressed
domains,
Δ [n(i)] ⊃ Δ (i) ∼ Δ [h(i)] ,
with n(i) and h(i) uniquely defined by these relations. Finally, suppose that each of
the domains D[n] in the catalogue is tiled by a finite number of periodic domains to-
gether with the return orbits of the sub-domains D j (i), n(i) = n, j = 0, 1, . . . , J(h(i)).
The catalogue members and their embedded sub-domains, each equivalent to a
catalogue member, comprise the roots and first level of an infinite tree satisfying the
recursive tiling property. Clearly, each Δ (i 0 ), being equivalent to Δ [h(i0 )] , has dressed
sub-domains Δ (i 0 , i1 ), i1 running over those index values such that n(i 1 ) = h(i0 ),
and is tiled, up to periodic domains, by the return orbits of the various D j1 (i0 , i1 ).
Continuing in this fashion, one generates the entire hierarchy of level-L dressed
def
domains Δ (i) = Δ (i0 , i1 , . . . , iL ), where ik ∈ 0, 1, . . . , I with the linkage n(i k+1 ) =
h(ik ).
In Fig. 2.10, we represent schematically the tree generated by a hypothetical 3-
member catalogue, each containing 2 dressed sub-domains equivalent to members
of the catalogue. The tree is set beside the much simpler one of a quadratic model
with a single scaling sequence. Note that in the more general framework, a scaling
sequence can still make its a apearance, in the form of a path down the infinite tree
2.7 The π /7 model 53
in which all links are identical from some level onward. In this way, the quadratic
models can be incorporated into the more general scheme, with the obvious changes
of notation (e.g. D j (L) becomes D j (i0 , i0 , . . . , i0 )).
Fig. 2.10 Schematic representation of a fictitious 3-member catalogue and its recursive tree. At
left, for comparison, is represented the scaling sequence of a typical quadratic model.
2.7 The π / 7 model
We now focus on a nontrivial example of the more general recursive tiling, namely
the π /7 model with its cubic irrational parameter λ = 2 cos π /7 (local rotation num-
ber p/q = 1/14) satisfying
λ 3 − λ 2 − 2λ + 1 = 0.
The map on Ω = [0, τ ) 2 , τ = λ −1 is shown in Fig. 2.11.

As a result of a strenuous, computer-aided search of the natural return map tree,
it was discovered by Lowenstein et al. (2004) that a catalogue of four dressed
triangles, Δ [n] , n = 0, 1, 2, 3, is sufficient for a recursive tiling of Ω . The return-
map partitions of these triangles are displayed in Figs. 2.12 and 2.13. There are
15 dressed sub-triangles whose return orbits implement the full tiling of the cata-
logue: Δ (9), . . . , Δ (14) tile Δ [0] , Δ (0), . . . , Δ (2) tile Δ [1] , Δ (3), . . . , Δ (5) tile Δ [2] ,
and Δ (6), . . . , Δ (8) tile Δ [3] . Each tiling covers the parent triangle up to a finite dis-
joint union of periodic cells, all of which are equivalent to a single prototype, the
heptagon (notation of (2.5))

Π =D (8, 10, 12, 0, 2, 4, 6), (1, . . ., 1), (0, λ , −1+λ +λ 2, λ +λ 2 , −1+λ +λ 2, λ , 0) .
(2.11)
Fig. 2.11 Local map of the π /7 model.
Fig. 2.12 Catalogue members Δ [0] and Δ [1] of the π /7 model, with return map partitions.
Fig. 2.13 Catalogue members Δ [2] and Δ [3] of the π /7 model, with return map partitions.
Recursion requires that each of the dressed sub-triangles Δ (i) be equivalent to a

member of the catalogue. The correct parent-child relations, from which the infinite
return map tree can be constructed recursively, are listed, together with other basic
features of the catalogue members, in Table 2.4, and displayed graphically in Figs.
2.14, 2.15 and 2.16. Complete tables of geometrical and dynamical data needed
for the computer assisted proof of renormalizability may be found in Sect. A.4 of
the Appendix. Readers interested in the details of the proof itself can referred to
Lowenstein et al. (2004) and its electronic supplement.
Table 2.4 Catalogue of prototype dressed domains δ[n] = (D[n] , ρ [n] ), with D[n] = D(m, (1, 1, 1),
[n]
(0, 0, 1)), with an overview of the recursive tree. Here J[n] is the number of atoms D j in the level-
[n]
0 return map partition, and J˜[n] is the number of generating cells Π ˜ . The level-0 children of the
j
prototypes are listed in the fourth column, together with the equivalence relations Δ (i0 ) ∼ Δ [h(i0 )] .
Δ [n] m J [n] Subdomains J˜[n]

Δ [0] (7, 1, 9) 4 Δ (0), Δ (2), Δ (4) ∼ Δ [3] , Δ (1), Δ (3), Δ (5) ∼ Δ [1] 28
Δ [1] (7, 1, 11) 11 Δ (6) ∼ Δ [2] , Δ (7), Δ (8) ∼ Δ [1] 9
Δ [2] (7, 1, 9) 12 Δ (9) ∼ Δ [2] , Δ (10), Δ (11) ∼ Δ [1] 6
Δ [3] (7, 1, 9) 10 Δ (12) ∼ Δ [3] , Δ (13), Δ (14) ∼ Δ [1] 10
Fig. 2.14 Catalogue of the π /7 model, with dressed subdomains.
Fig. 2.15 Recursion rules for the catalogue of the π /7 model.
To begin the construction of the return map tree, we note that the first member
of the catalogue, Δ [0] , is equivalent to the subdomain Ω 2 of the square. Figure 2.17
shows the tiling of the square by the 4 return map orbits of that dressed triangle. To
proceed further down the tree, one recursively applies the 15 subdomain relations
summarized in the graph of Fig. 2.16, in which the nodes are the catalogue members
Δ [n] , n = 0, . . . , 3, and the links are the 15 inclusion relations, with arrows pointing
from parent to child. To each of these oriented links corresponds an embedding
function g(i), i = 0, 1, . . . , 14, defined by (recall h(i) = child, n(i) = parent)
Δ (i) = g(i)Δ [h(i)] ⊂ Δ [n(i)] ,

Δ (iL ) = Δ (i0 , i1 , . . . , iL ) = g(i0 , i1 , . . . , iL )Δ [h(iL )] ⊂ Δ (i0 , . . . , iL+1 ),
g(iL ) = g(i0 , . . . , iL ) = g(iL−1 )g(iL ) = g(i0 )g(i1 ) · · · g(iL ).
Fig. 2.16 Same as Fig. 2.15, in the form of a Fig. 2.17 Tiling of Ω by sub-domains of Ω2 .
graph with oriented links.
As an illustration of the recursive tiling at level L = 0, let us consider the tiling of

D[1] by the return orbits of the L = 1 subdomains D j (0), j = 0, . . . , 11, D j (1), j =
0, . . . , 10, and D j (2), j = 0, . . . , 10. Since
h(0) = 2, h(1) = h(2) = 1.
we only have to deal with the dressed triangles similar to Δ [1] and Δ [2] in this exam-
ple. In order that the reader be able to follow the return orbits by eye, we have made
a linear transformation to a coordinate system in which the metric is Euclidean, and
the similarity transformations involve true rotations and reflections. The return map
partitions and dressed subdomains for Δ [1] and Δ [2] are shown in Fig. 2.18, and the
recursive tiling of D [1] is show, color coded, in Fig. 2.19.
In the recursive tiling, the basic building blocks of the level-L partition are the
domains
DtjLL (iL ) = K t0 ρ (i0 )t1 ρ (i0 , i1 )t2 · · · ρ (iL−1 )tL D jL (iL ),
which we shall refer to as tiles. An important consequence of the recursive tiling is
that every point in the residual set Σ lies in the intersection of a nested sequence of
tiles, i.e.
{x} = DtjLL (iL ),
L
and so x is uniquely determined by the infinite sequence
σ (x) = ((i0 , j0 ,t0 ), (i1 , j1 ,t1 ), . . .). (2.12)
Note that once once j L is known, all the preceding j L are determined by the path
constraints jk = p(ik+1 , j ,t ).
k+1 k+1
The set Σ (iL ) = jL ,tL DtjLL (iL ) ∩ Σ consists of all discontinuity-avoiding orbits
which pass through D(i L ) and is clearly an invariant subset of Σ . This prop-

erty extends to the intersection Σ (i) = L Σ (iL ), for any infinite linked sequence
i0 , i1 , . . . , ∀L, h(iL ) = n(iL+1 ), i.e. to any path down the infinite tree generated by the
recursion rules of the catalogue.
The existence of a multiplicity of invariant sets implies a profound contrast be-
tween the quadratic models of Sect. 3, in which the recursion trees are simple lad-
Fig. 2.18 Return map partitions of catalogue members D[1] and D[2] into their respective atoms.
The Euclidean metric is used here to make evident the various similarity relations. Superposed are
the dressed subdomains Δ (i), i = 0, . . ., 5. (See the color figure at the end of the book.)
Fig. 2.19 Tiling of D[1] by the return orbits of D j (i), i = 0, 1, 2, j = 0, . . ., J(h(i)). The Euclidean
metric is used to make evident the similarity relations. Consistent with the previous figure, different
values i = 0, 1, 2 are coded red, green, and blue, respectively. (See the color figure at the end of the
book.)
ders, and models with more elaborate trees, such as the π /7 model. In the latter,
there are, except for level 0, where i 0 = 0, and level 1, where i 1 ∈ {9, 10, . . ., 14},
exactly 3 alternatives at each branching of the tree. Thus, the number of branches at
level L ≥ 1 is 6 × 3L−1 , which for L → ∞ tends to an uncountable infinity of alterna-
tives. To see this, we can map the sequences i 2 , i3 , . . . onto the set of infinite ternary
sequences, hence onto the continuum interval [0, 1].
The phase portrait of K on the residual set thus splits up into sectors labeled by
the code sequences i. Each sector has a dynamics which parallels, to a large extent, a
typical quadratic model with a single scaling sequence. For example, in the symbolic
representation (2.12), the indices i k are an inert background, relevant only through
the path constraint j k = p(ik+1 , jk+1 ,tk+1 ). All of the points on the forward orbit of
the point x have the same sector indices i 0 , i1 , . . .. Within a sector, the level-L return
times are determined by iteration of the incidence matrices
νk −1
A j,k (i) = ∑ δ j,p(i,k,t) , j, k = 0, . . . , J(h(i)).
t=0
A(iL ) = A(iL−1 ) · A(iL ) = A(i0 ) · A(i1 ) · · · A(iL ),

J(h(i)
Tk (i) = ∑ A j,k (i), k = 0, . . . , J(h(i)),
j=0
T (iL ) = (T0 (iL ), . . . , TJ(h(iL )) (iL )) = T (iL−1 ) · A (iL ) = T (i0 ) · A (i1 , . . . , iL ). (2.13)
The level-to-level scale factors are also determined recursively,
ω (iL ) = ω (iL−1 )ω (iL ) = ω (i0 )ω (i1 ) · · · ω (iL ).
As in the quadratic models, the scale factors are all units. As pointed out in Section
2.2.1, these are all of the form
η1m1 η2m2 ,
where η1 and η2 are so-called fundamental units. For convenience, we choose these
to be
η1 = 2+ λ − λ 2 = τ = λ −1 , η2 = −1− λ + λ 2 = (−1+ λ )−1 = 1− η1 . (2.14)
The fact that λ and τ are units is a special feature of this model, not necessarily
expected to hold for more general parameter values.
For almost all sectors i, the domains D(i L ), L = 0, 1, . . ., do not form a scaling
sequence, since the shapes and return map dynamics keep changing from one level
to the next. However, among the various sectors, there are indeed infinitely many
scaling sequences, namely those for which i is eventually periodic, i.e. of the form
. . . ia , . . . , ib , ia , . . . , ib , ia , . . . , ib , . . . = . . . (ia , . . . , ib )∞ .
References 61
In that case we have a scaling sequence whose limit point is x ∞ with code se-
quence σ (x∞ ) = i. Its temporal scale factor ω T (i) is equal to the largest eigenvalue
of (A(ia ) · · · A(ib ))T , and its geometric scale factor is ω K (i) = ω (ia ) · · · ω (ib ). By
the informal argument at the end of Sect. 2.1, the set Σ (i) is expected to be a frac-
tal of dimension − log ω T (i)/ log ωK (i). A more precise discussion will be given in
Chapter 4.
Although the countable set of eventually periodic sector codes is of negligible
size within the uncountable set of all codes, it is nevertheless of considerable im-
portance, analogous to that of the rational numbers among the real’s. Just as the
rational numbers are essential for practical numerical computation, the sectors with
eventually periodic codes are essential for the explicit construction of orbits in the
residual set.
References
Adler R., Kitchens B., and Tresser C. (2001) Dynamics of nonergodic piecewise affine maps of
the torus, Ergodic Theory and Dynamical Systems 21, 959–999.
Akiyama S., Brunotte H., Pethö A., and Steiner W. (2008) Periodicity of certain piecewise affine
planar maps, Tsukuba Journal of Mathematics 32, 1–55.
Ashwin P. (1997) Elliptic behaviour in the sawtooth standard map, Physics Letters A 232, 409–
416.
Ashwin P., Chambers W., and Petrov G. (1997) Lossless digital filter overflow oscillations: approx-
imations of invariant fractals, International Journal of Bifurcation and Chaos 7, 2603–2610.
Boshernitzan M.D. and Carroll C. R. (1997) An extension of Lagrange’s theorem, Journal
d’Analyse Mathématique 72, 21–44.
Buzzi J. (2001) Piecewise isometries have zero topological entropy, Ergodic Theory and Dynami-
cal Systems 21, 1371–1377.
Chua L.O. and Lin T. (1988) Chaos in digital filters, IEEE Transactions:Circuits and Systems
CAS-35, 648–658.
Chua L.O. and Lin T. (1990) Fractal pattern of second order non-linear digital filters: a new sym-
bolic analysis, International Journal of Circuit Theory and Applications 18, 541–550.
Dana, I. (2004) Global superdiffusion of weak chaos, Physical Review E 69, 016212.
Davies A.C. (1995) Nonlinear oscillations and chaos from digital filters overflow, Philosophical
Transactions of the Royal Society London A 353, 85–99.
Falconer K. (1990) Fractal Geometry, Wiley, Chichester.
Goetz A. (1996) Dynamics of piecewise isometries, PhD Thesis, University of Chicago.
Goetz A. (1998) Dynamics of a piecewise rotation, Continuous and Discrete Dynamical Systems
4, 593–608.
Goetz A. (2000) Dynamics of piecewise isometries, Illinois Journal of Mathematics 44, 465–478.
Goetz A. (2001) Stability of cells in non-hyperbolic piecewise affine maps and piecewise rotations,
Goetz A. and Poggiaspalla G. (2004) Rotation by π /7, Nonlinearity 17, 1787–1802.
G.H. Hardy and E.M. Wright (1979) An Introduction to the Theory of Numbers, 5th edition,
Oxford University Press, Oxford.
Kahng B. (2000) Dynamics of symplectic affine maps on tori, PhD Thesis, University of Illinois
at Urbana-Champaign.
Kahng B. (2002) Dynamics of symplectic piecewise affine elliptic rotation maps on tori, Ergodic
Theory and Dynamical Systems 22, 483–505.
Khinchin A. Ya. (1964) Continued Fractions, Dover, Mineola, New York.
Kouptsov K.L., Lowenstein J.H., and Vivaldi F. (2002) Quadratic rational rotations of the torus
and dual lattice maps, Nonlinearity 15, 1795–1842.
Lowenstein J.H. (2007) Aperiodic orbits of piecewise rational rotations of convex polygons with
recursive tiling, Dynamical Systems 22, 25–63.
Lowenstein J.H., Hatjispyros S., and Vivaldi, F. (1997) Quasi-periodicity, global stability and scal-
Lowenstein J.H., Kouptsov K. L., and Vivaldi F. (2004) Recursive tiling and geometry of piecewise
rotations by π /7 Nonlinearity 17, 1–25.
Lyubomudrov O., Edelman M., and Zaslavsky G. M. (2003) Pseudochaotic systems and their
fractional kinetics, Int. Journal of Modern Physics B 17, 4149–4167.
Poggiaspalla G. (2006) Self-similarity in piecewise isometric systems, Dynamical Systems 21,
147–189.
Zaslavsky G. M. and Edelman M. (2003) Pseudochaos, Perspectives and Problems in Nonlinear
Science: a Celebratory Volume in Honor of Lawrence Sirovich, eds. E. Kaplan, J. Marsden, and
K. R. Sreenivasan, Springer, New York, 421–423.
Wu C.W. and Chua L.O. (1993) Properties of admissible sequences in a second order digital filter
with overflow non-linearity, International Journal of Circuit Theory and Applications 21, 299–
307.
Chapter 3
Symbolic Dynamics
In the preceding chapter, we introduced a hierarchy of covering sets (recursive

tilings) which converge to the residual set of aperiodic, discontinuity-avoiding orbits
of the local map. We now want to encode the points of the residual set abstractly as
strings of symbols (codes), in such a way that application of the dynamical map is
conjugated to a relatively simple transformation of the codes, implemented by the
odometer-like machine which we will call a codometer.
One of the themes which we want to emphasize in this book is the fact that the
kicked oscillator maps and the corresponding piecewise isometries on the square
may be considered as two different representations of the same symbolic dynamics.
In some cases, we will even have a third representation, in the form of a model of
Hamiltonian round-off (see Chapter 7). For the symbolic dynamics to be a mean-
ingful underpinning of these quite different geometric mappings, it must be defined
intrinsically. That is, there must be a set of pruning rules for distinguishing the ad-
missible codes from the inadmissible ones. Much of the current chapter will be
devoted to this nontrivial task.
3.1 Symbolic representation of the residual set
There is a natural symbolic representation of points in the residual set Σ [n] , i.e. points
which lie neither in periodic cells nor on discontinuity lines of a catalogue member
Δ [n] = (D[n] , ρ [n] ). Let us introduce this coding, step by step, for a given point x.
3.1.1 Hierarchical symbol strings
Thanks to the recursive tiling of D [n] by level-0 tiles D j00 (i0 ), j0 = 0, 1, . . . , J [h(i0 )] ,
t
t0 = 0, 1, . . . , ν j (i0 ), we can identify the unique tile within which x resides, assigning
it the symbol (i0 , j0 ,t0 ). Since x is assumed not to lie on any tile boundary, we can
64 3 Symbolic Dynamics
work here, and in the rest of this chapter, entirely with open polygonal tiles, making
use only of the weak version of recursive tiling.
To pin down the location of x with higher resolution, we can go to level 1 and
locate the tile
t t t
D j01 1 (i0 i1 ) ⊂ D j00 (i0 )
which (again uniquely) contains x, adding a second symbol to the list,
x ↔ (i0 , j0 ,t0 ), (i1 , j1 ,t1 ), . . . (3.1)
indicating with dots the additional symbols to be specified as we proceed. Note that
one of the component symbols, j 0 , is redundant in the notation, due to the path
constraint built into the renormalization structure:
j0 = p(i1 , j1 ,t1 ).
Moreover, the indices i 0 and i1 are not completely independent: recall that in general
the index i stands for an inclusion of one dressed domain, equivalent to catalogue
member D[h(i)] , within another, equivalent to the catalogue member D [n(i)] , and so in
(3.1) we must have
h(i0 ) = n(i1 ).
Rather than eliminating such redundancies in the notation, we will retain them as
valuable signposts of the recursive structure.
Clearly the above method allows us to construct as small a neighborhood of the
target point x as we want, simply by finding a sufficiently small tile of the form
t ...tL t ...t t
D j0L (i0 . . . iL ) ⊂ D j0L−1L−1 (i0 · · · iL−1 ) ⊂ · · · ⊂ D j00 (i0 ).
The recursive tiling property ensures that such a tile always exists. The specification
of all such tiles for L = 0, 1, . . . , ∞ pins down the point x exactly, and so we write
the correspondence in the form
x ↔ ((i0 , j0 ,t0 ), (i1 , j1 ,t1 ), . . . , (ik , jk ,tk ), . . .), (3.2)
h(ik ) = n(ik+1 ), jk = p(ik+1 , jk+1 ,tk+1 ), k = 0, 1, 2, . . . . (3.3)

A constrained symbol sequence of the form (3.2), with (3.3), will be called a
linked symbol sequence, or, more simply, a code. Although each x ∈ Σ [n] deter-
mines a unique linked sequence, the reverse is not always the case: a linked symbol
sequence might instead determine a point on one of the boundary lines (since an
infinite sequence of nested domains can have as their limit a point on a bounding
discontinuity line). The linked sequences which correspond to points in Σ [n] will
be called admissible, while those which relate to points on boundary lines will be
called inadmissible. To meaningfully exploit the advantages of symbolic dynamics,
we will need a way to distinguish intrinsically the admissible from the inadmissible
codes. This task will occupy our efforts in Sect. 3.3.
3.1 Symbolic representation of the residual set 65
3.1.2 Eventually periodic codes
Among the admissible linked symbol sequences, an important subset is that of the
eventually periodic ones. Like the rational numbers embedded in the reals, the points
corresponding to eventually periodic codes form a countably infinite dense skeleton
within the residual set. It is these points which we normally deal with in compu-
tational experiments, primarily because they can be located precisely, as algebraic
numbers. Thus, even though Σ [n] is a set of Lebesgue measure zero so that the prob-
lem of picking out a sample point by trial and error is zero, we can nonetheless find
plenty of x ∈ Σ [n] by exploiting their eventually periodic codes.
To see how the construction works, let us suppose that we are given a period-L
linked symbol sequence,
σ = (σ1 , σ2 , . . . , σL )∞ , σk = (ik , jk ,tk ),
so that
x = lim ρ [n(i1 )]t1 ρ (i1 )t1 ρ (i1 , i2 )t2 · · · ρ (iNL−1 )tNL D0jNL (iNL ).
N→∞
where
0 def
D jkk (ik ) = D00···0
jk (i0 , . . . , ik ).
The recursive tiling allows us to write
[h(i)]
D0j (i) = g(i)D j , ρ (i) = g(i)ρ [n] g(i)−1 ,
where g(i) ∈ G , and more generally

0
D j k (i1 , . . . , ik ) = g(ik )D[h(ik )] , ρ (ik ) = g(ik )ρ [h(ik )] g(ik )−1 ,
where g(ik ) = g(i1 )g(i2 ) · · · g(ik ). Thus
x(σ ) = lim gN D0jNL (iNL ), (3.4)

N→∞
where
g = ρ [n(i1 )]t1 g(i1 )ρ t2 (i1 )g(i2 ) · · · ρ tL (iL−1 )g(iL ).
Using the group commutation relations, g ∈ G can be expanded as
g = Tδ Rα Cν Sω = Tδ g ,
where, since 0 < ω < 1, g is a strictly contracting linear transformation. Thus, from
(3.4), we have
x(σ ) = gx(σ ) = g x(σ ) + g0, (3.5)
where 0 is the origin, a fixed point of g . Since 1 − g is invertible,
x(σ ) = (1 − g)−1 g0.

This provides a straightforward algorithm for calculating x(σ ) for repeating σ .

For an eventually periodic sequence, a finite number of additional group pre-
multiplications are sufficient to yield the result.
3.1.3 Simplified codes for quadratic models
For the quadratic models we deal with in this book, the linked index sequences
i = i0 , i1 , i2 , . . . which appear in the coding are exceedingly simple. Typically, one
has
i = i0 , i1 , i1 , i1 , . . . = i0 , i∞
def
1,
where i0 corresponds to the inclusion of the base domain of a scaling sequence

in the generating dressed domain (Ω , K), and i 1 corresponds to the inclusion of one
member of a scaling sequence within another. The most complicated of the quadratic
examples introduced in Chapter 2 was Model III, where the multiindices would be
of the form
i = i0 , i1 , i2 , i3 , i2 , i3 , . . . = i0 , i1 , (i2 , i3 )∞ ,
with the respective ik corresponding to the inclusions
D(−1) ⊂ Ω , D(0) ⊂ D(−1), D(k odd) ⊂ D(k − 1 even),
D(k even) ⊂ D(k − 1 odd).

Note that if we are only interested in points of the residual set, the distinction be-
tween even and odd levels disappears, and so, with telescoping, we can use multi-
indices which conform to the simpler pattern of the other examples.
In most contexts, the multiindex i is unique and unchanging, and so it makes
sense in the quadratic models to suppress it in the symbolic representation, writing
σ (x) = ( j0 ,t0 )( j1 ,t1 )( j2 ,t2 ) · · · .
We will adopt this convention in what follows.
3.2 Dynamical updating of codes
Having introduced a representation of each point x in Σ [n] as an infinite symbol

string σ , we now want to determine how σ is transformed when x is mapped into
ρ [n] (x). We begin by focusing on the first symbol in the string, (i 0 , j0 ,t0 ), which
assigns x to the level-0 tile
D j00 (i0 ) = ρ [n]t0 D j0 (i0 ),

t def
0 ≤ t0 ≤ ν j0 (i0 ) − 1 = t0max
3.2 Dynamical updating of codes 67
Clearly, if t0 is not maximal, i.e. if t 0 < ν j0 (i0 ) − 1, then applying ρ [n] maps Dtj00 (i0 )
t +1
to D j00 (i0 ), and hence
(i0 , j0 ,t0 < t0max ) · · · → (i0 , j0 ,t0 + 1) · · · ,
where the dots correspond to an arbitrary admissible continuation to the right, the
same before and after the mapping. On the other hand, if t 0 is maximal, we have
t max
ρ [n] D j00 (i0 ) = ρ [n] ν j0 (i0 ) D j0 (i0 ) = ρ (i0 )D j0 (i0 ).
Thus, the application of ρ [n] completes the level-0 return orbit of D j0 (i0 ), bring-
ing it back into D(i 0 ). Here we are faced with the possibility of branching, since
ρ (i0 ) acts piecewise on the level-1 domains which tile D(i 0 ). The outcome is, of
course, determined by bringing into play the second symbol in the representation of
x. Specifically, if
σ = (i0 , j0 ,t0max )(i1 , j1 ,t1 ) · · · ,
where t1 is submaximal, we have
t max t1
x ∈ D j01 (i0 , i1 ),
0t1 +1
ρ [n] x ∈ ρ (i0 )D0t
j1 (i0 , i1 ) = D j1
1
.
Thus, the level-1 return orbit of the base tile D j1 (i0 , i1 ) advances by one step. The
level-0 atom within which the image tile lies is determined by the path function:
D0t1 +1 (i0 , i1 ) j1 ⊂ D p(i1 , j1 ,t1 +1) (i0 ),
and so we get the symbolic mapping
(i0 , j0 ,t0max ) (i1 , j1 ,t1 < t1max ) · · · → (i0 , p(i1 , j1 ,t1 + 1), 0) (i1 , j1 ,t1 + 1) · · · , (3.6)
where once again the dots indicate the same admissible extensions on both sides of
the arrow.
Clearly the process can be iterated ad infinitum to obtain the general rule
(i0 , j0 ,t0max ) (i1 , j1 ,t1max ) · · · (ik , jk ,tk < tkmax ) (ik+1 , jk+1 ,tk+1 < tk+1
max ) · · ·
(3.7)
→ (i0 , j0 , 0) (i1 , j1 , 0) · · · (ik , jk , 0) (ik+1 , jk+1 ,tk+1 + 1) · · ·
jr = p(ir+1 , jr+1

, 0), r = 0, . . . , k − 1, jk = p(ik+1 , jk+1 ,tk+1 + 1).
This leaves only the case where all t k are maximal, which, according to the pre-
ceding pattern, should transform as
(i0 , j0 ,t0max ) (i1 , j1 ,t1max ) · · · (ik , jk ,tkmax ) · · · → (i0 , j0 , 0) (i1 , j1 , 0) · · · (ik , jk , 0) · · ·
(3.8)
jk = p(ik+1 , jk+1

, 0), k = 0, 1, 2, . . . .
At first sight, this prescription appears to be non-unique. Fortunately, we have con-
structed all of our models in such a way that each base domain D(i L+1 ) is contained
in a single atom of D(i L ), and so jk depends only on i k+1 , not on jk+1 .
The symbolic representation of the dynamics makes very clear the invariance of
each set Σ (i0 , i1 , . . .) consisting of all those points of the residual set which share the
same linked index sequence i 0 , i1 , . . . , h(ik ) = n(ik+1 ). On each of the Σ (i), the hier-
archical coding and updating scheme just described is an example of a Vershik map
on an ordered Bratteli diagram (Bratteli, 1972; Vershik,1985; Durand et al., 1999).
This association was discussed by Poggiaspalla (2006) in the case of periodic se-
quences (i1 , i2 , . . . , iL )∞ , where the Bratteli diagram is stationary, and generalized to
the present context by Lowenstein (2007). Simple examples of such a dynamical
map are the simple odometer, where all i k and jk can be set to zero and the code re-
duces to (t1 ,t2 , . . .), and the multi-register odometer, with code ((i 0 ,t0 ), (i1 ,t1 ), . . .),
in which each level has a fixed identity i k with 0 ≤ tk < ν (ik ). Dynamical mappings
of the odometer type were used by Afraimovich et al. (2000) to model sticky orbits
in Hamiltonian systems.
The hierarchical updating scheme defined above is conveniently implemented as
a symbolic machine, the codometer. The latter is described in elementary terms in
Appendix B, with application to a familiar example, the Gregorian calendar.
3.3 Admissibility
Thus far we do not have a completely satisfactory symbolic dynamics, thanks to

the problem of admissibility, i.e. distinguishing those sequences which correspond
to points of Σ from those which correspond to points of Σ \ Σ . Geometrically, the
distinction is not difficult to specify. In each case, the code σ (x) = (σ 0 , σ1 , . . .) cor-
responds to a nested sequence of open domains D 0 ⊃ D1 ⊃ · · · , with x = limL→∞ DL .
t ,...,t
Each DL is in fact a tile D j0L L (i0 , . . . , iL ), but for the time being we dispense with
the detailed indices. For x ∈ Σ , x lies in the interior of each D L by definition, while
for y ∈ Σ \ Σ , y lies on the boundary ∂ D L of at least one DL . In fact, if y ∈ ∂ DL , y
also lies on ∂ DL+1 , and hence is a boundary point of all D K , K ≥ L.
The inadmissible sequences thus correspond to the common boundary points of
infinite nested sequences of tiles. We now wish to associate these points with infinite
sequences of tile edges, the latter being defined to include their endpoints, the tile
vertices. Suppose that x is not a vertex of any of the domains associated with the
code σ (x). Then it is clear that x lies in the intersection of a nested set of tile edges.
If, on the other hand, x is a vertex of D L , it is not obvious that x is the endpoint of a
nested set of edges E L , L ≥ L. In particular, x ∈ E L does not imply that there exists
a matching edge E L+1 of DL+1 , as one sees in Fig. 3.1. We note that this can happen
only if the vertex angle at x for D L+1 is strictly less than that for DL . But in our
3.3 Admissibility 69
models, only finitely many vertex angles are possible, and so there always exists an
Lmin and a nested set of edges E L ⊂ ∂ DL , with x ∈ EL , L ≥ Lmin .
Fig. 3.1 Matching vertices for DL and DL+1 , with (a) and without (b) matching edges.
Recursive edge-matching is thus the key to checking the tails of symbol se-
quences for admissibility. Fortunately, our description of polygonal domains in
terms of half-plane inequalities in Sect. 2.2.2 allows for a systematic enumeration of
edges which is level-independent and thus well suited to testing for matching edges.
Specifically, the a th edge of D(m , s , b ) matches the ath edge of D(m, s, b) if
ma = ma , sa = sa , ba = ba , (3.9)
or, if the parameter λ is equal to 2 cos(2π p/q) with q even,
ma = ma + q/2 mod q, s a = −sa , ba = −ba . (3.10)
None of the operations of generalized rotation, reflection, translation, or scaling

which generate the transformation group G affects the order of the indices in m, s, b,
and hence the edge enumeration is G -invariant. Since the return maps relating dif-
ferent levels are elements of G , the level-independence of the edge assignments
follows.
We now have an abstract criterion for inadmissibility, which we need to turn
into a simple rule for testing specific symbol sequences. We already have such a
rule for the linked sequences: given the kth symbol (i k , jk ,tk ), the (k + 1)st symbol
(ik+1 , jk+1 ,tk+1 ) must satisfy
n(ik+1 ) = h(ik ), p(ik+1 , jk+1 ,tk+1 ) = jk , (3.11)
but is otherwise arbitrary. This rule can be conveniently represented by a transition

graph or equivalent transition matrix of 0’s and 1’s. An inadmissible tail sequence
has in addition a transition rule expressing the edge-matching condition. For this,
we need to append to each symbol (i L , jL ,tL ) a fourth index a L labeling an edge
of the tile DtjLL (iL ). To construct all possible inadmissible tail sequences, we seek a
transition rule of the following sort: given the symbol ((i L , jL ,tL , aL ), which possi-
ble successor symbols (i L+1 , jL+1 ,tL+1 , aL+1 ) are compatible with (3.11) and cor-
respond to matching edges in the sense of (3.9) or (3.10)? Ultimately, it should be
possible to express the rule without explicit reference to the edge sequences, since
the existence of a single sequence of matching edges in the tail is enough to restrict
the limit point to a tile boundary. How to accomplish this will become clear in our
examples below.
Following the above prescription, one can generate a transition matrix T , with
entries 0 and 1, which can be used to test a code σ for admissibility one symbol
at a time, at least if the tail consists of a repeated finite symbol string, so that the
edge-matching test is guaranteed to terminate. The rows and columns of T are la-
beled by the various possible (i, j,t, a); if (i, j,t, a) can be followed by (i , j ,t , a )
in an inadmissible tail, then the row of T labeled by (i, j,t, a) has a 1 in the column
labeled by (i , j ,t , a ); if such a succession of symbols is forbidden, then the entry
is a 0. Obviously, if a row (resp. column) consists entirely of 0’s, then it and the cor-
responding column (resp. row) can be deleted, since such a symbol can be banished
by initiating the tail sequence at a sufficiently high level.
The transition matrix is easily seen to be equivalent to a graph T with oriented
links. The nodes of T are labeled by the possible (i, j,t, a), and if (i, j,t, a) can be
followed by (i , j ,t , a ) in an inadmissible tail, there is a link connecting the cor-
responding nodes. Any node without an incoming or outgoing link can be deleted,
together with its remaining links. The same reasoning allows us to further “reduce”
the graph if the latter consists of two subgraphs connected by links which all have
the same orientation, as in Fig. 3.2.
Fig. 3.2 Reduction of a graph.
In that case, the links can be deleted, splitting the graph into disconnected pieces.
Repeating the process, deleting isolated nodes along the way, one ends up with a set
of non-trivial irreducible subgraphs. For convenience in checking admissibility, one
normally converts the transition graph into a table listing the possible successors
of symbols which might appear in an inadmissible tail sequence. Let us now see
how the above scheme works in two examples. In the first, with quadratic irrational
parameter, we derive the rules graphically, while in the second, with cubic irrational
parameter, the original transition graph is huge, and it is more feasible to manipulate
the corresponding matrix, with computer assistance.
3.3.1 Quadratic example

√
Our first example is the quadratic model with parameter λ = −τ = 12 ( 5 − 1), the
inverse of the golden mean. The catalogue consists of a single triangle, with the do-
Fig. 3.3 Level-zero return map, with labeled edges. The rotation angles for domains D0 (0) and
D1 (0) are 4π /5 and −4π /5, respectively.
Fig. 3.4 Tiling of level-zero triangle by periodic pentagons and return orbits of level-one sub-
domains, with labeled edges.
main map shown in Fig. 3.3. To assist the eyes, we have used a coordinate system
in which C is a true rotation matrix. It is sufficient to consider the edge assignments
for levels 0 and 1, shown in Figs. 3.3 and 3.4, since the matching conditions are
independent of level. The reader can easily verify visually the edge-matching con-
ditions displayed in the graph of Fig. 3.5. In the latter, the node corresponding to
edge a of Dtj (0) is labeled ( jta). If this edge contains an edge a of the subdomain

Dt,t
j (1), the graph has a link directed from ( jta) to ( j t a ). For example, edge 3 of
D00 (0) contains edge 3 of D 0,2 0,1 0,1
0 (1), edge 1 of D 1 (1), and edge 2 of D 0 (1). In the
graph, this corresponds to links from (013) and (023) to (023), (111), and (012).
There is no node (003), since edge 3 of D 0,0
0 (1) is internal and cannot participate in
the true tail of a code sequence.
Fig. 3.5 Transition graph for inadmissible codes. Removal of all links crossed by the dashed lines
reduces the graph to four connected components. The surviving nodes occupy the white disks.
As explained above, we can reduce the transition graph as follows. Each of

the dashed lines splits the graph into two pieces with unidirectional links. Delet-
ing the latter leaves us with 8 irreducible subgraphs, of which 4 are isolated nodes
(013),(012),(111),and (132), which can be deleted together with their links. In the
figure the four nodes have been displayed as gray disks. The remaining irreducible
parts (with nodes represented by white disks) correspond to the following table:
( jk ,tk ) (0, 2) (0, 0), (0, 2) (1, 0), (1, 2) (1, 0) , (1, 4) (1, 4)
( jk+1 ,tk+1 ) (0, 2) (0, 2) , (1, 2) (0, 0), (1, 0) (1, 0) , (1, 4) (1, 4)
In each column, any of the symbols above can be followed by any of the symbols
below. As a simple application, let us construct the inadmissible codes which are
of the form ( j,t) ∞ . From the table, the condition ( j k ,tk ) = ( jk+1 ,tk+1 ) selects only
(0, 2)∞ , (1, 4)∞ , and (1, 0)∞ . The admissible period-1 codes are all the remaining
ones compatible with the path function p(0) = (1, 0, 0), p(1) = (1, 0, 0, 1, 1), namely
(0, 1)∞ and (1, 3)∞ .
For period n, what fraction of the codes are inadmissible? We can easily calculate
an upper bound for this ratio from the irreducible transition matrices corresponding
to the graphs of Fig.3.5, namely
⎛ ⎞
0 1 0 1
1 1 ⎜0 1 0 1⎟
(1), (1), , ⎜ ⎟,
1 1 ⎝1 0 1 0⎠
1 0 1 0
with respective eigenvalue sets {1}, {1} {2, 0} {2, 0, 0, 0}. This gives for the upper
bound on the number of inadmissible tail sequences 1 + 1 + 2 n + 2n = 2n+1 + 2.
For n = 1 this gives 6, which is twice the actual number because of the duplica-
tion of unprimed and primed edge labels. The total number of linked sequences
is given by Tr P n , with P the transition matrix corresponding to the path func-
tion. The matrix is 8-dimensional, corresponding to rows and columns labeled by
( j,t) = (0, 0), (0, 1), (0, 2), (1, 0), (1, 1), (1, 2), (1, 3), (1, 4):
⎛ ⎞
0 1 1 0 1 1 0 0
⎜0 1 1 0 1 1 0 0⎟
⎜ ⎟
⎜0 1 1 0 1 1 0 0⎟
⎜ ⎟
⎜1 0 0 1 0 0 1 1⎟
P=⎜ ⎟
⎜ 1 0 0 1 0 0 1 1 ⎟.
⎜ ⎟
⎜1 0 0 1 0 0 1 1⎟
⎜ ⎟
⎝1 0 0 1 0 0 1 1⎠
1 0 0 1 0 0 1 1
The eigenvalues of P are {4, 1, 0, 0, 0, 0, 0, 0}, and so Tr P n = 4n + 1. Thus the frac-

tion of inadmissible sequences is bounded above by (2 n+1 + 2)/(4n + 1), which for
large n tends to zero proportional to 2 −n .
3.3.2 Models I, II, and III
We have carried out a similar analysis for our three quadratic examples, Models I,
II, and III. In each case, a list of edge matches was obtained by iteratively mapping
the level-1 atoms along their level-0 return paths. In each case, the list was reduced,
with the help of graphs, to a minimal set of succession rules ( j k ,tk ) → ( jk+1 ,tk+1 )
for infinitely long tail sequences. The tables allow us to systematically test the ad-
missibility of any given eventually periodic code: if any neighboring code elements
in the tail fail to satisfy the listed succession rules, then the code is admissible and
we know that the point corresponding to the symbol sequence does not lie on any
discontinuity line.
The inadmissibility table computed for Model I is shown in Table 3.1.
Table 3.1 The inadmissibility table for Model I
( jk ,tk ) (0, 0), (0, 1) (2, 0), (2, 1) (4, 4), (4, 6) (4, 2), (4, 4)
( jk+1 ,tk+1 ) (0, 1), (2, 1) (0, 0), (2, 0) (4, 2), (4, 4) (4, 4) , (4, 6)
Note that there are two irreducible sets of code elements, separated by double
vertical lines in the table. The corresponding succession matrix has two irreducible
blocks, each of the form ⎛ ⎞
0 1 0 1
⎜0 1 0 1⎟
M=⎜ ⎝ 1 0 1 0 ⎠,
⎟
1 0 1 0
with eigenvalues (2, 0, 0, 0). This matrix counts edge matches rather than ( j,t)
matches, and since we have two distinct edge symbols (4, 4) and (4, 4) , we have
to be on the lookout for duplicate codes. Fortunately, a brief inspection of the pos-
sibilities shows that the only periodic code in this category is (4, 4) ∞ .
Thus, the number of inadmissible codes of period n is
In = 2Tr M n − 1 = 2n+1 − 1.
On the other hand, if Â is the incidence matrix (see Sect. 2.3) restricted to the open
polygonal atoms: ⎛ ⎞
1 2 4
Â = ⎝ 1 1 1 ⎠
0 1 3
with eigenvalues (4, 1, 0), the total number N n of period-n codes is given by
Nn = Tr Ân = 4n + 1.
We see that the fraction of inadmissible periodic sequences tends to zero as 2 −n as

n tends to infinity.
For n = 1, the formulas give
I1 = 3, N1 = 5,
which we can check explicitly using Table 3.1 and the Model I path function. The
list of codes ( j,t)∞ is the following, with the inadmissible one underlined:
(0, 1)∞ , (2, 0)∞ , (4, 2)∞ , (4, 4)∞ , (4, 6)∞ .
For n = 2, the trace formulas tell us
I2 = 7, N2 = 17.
Once again, use of Table 3.1 and the path function allow us to identify the admissible
and inadmissible (underlined) codes. In addition to the period-1 cases already listed,
we have
((0, 0)(2, 1))∞ , ((2, 1)(0, 0))∞ , ((0, 0)(2, 3))∞ , ((2, 3)(0, 0))∞ ,
((2, 2)(4, 0))∞ , ((4, 0)(2, 2))∞ , ((4, 4)(4, 2))∞ , ((4, 2)(4, 4))∞ ,
((4, 2)(4, 6))∞ , ((4, 6)(4, 2))∞ , ((4, 6)(4, 4))∞ , ((4, 4)(4, 6))∞ ,
confirming the results of the matrix calculations.
The other quadratic models can be handled similarly. In particular, the inadmis-
sibility tables for Models II and III, obtained using the edge-matching algorithm,
are given in Sect. A.5 of the Appendix. In each case, diagonalization of the succes-
sion matrix for inadmissible tails produces eigenvalues which are strictly less than
the largest eigenvalue of the restricted incidence matrix. The inadmissible periodic
codes thus form a fraction of all codes which tends to zero geometrically as the pe-
riod increases. It is not difficult to show that the same behavior holds if one restricts
oneself to eventually periodic codes.
3.3.3 Cubic example
Our final and most elaborate example is the 4-member catalogue of the π /7 model.
Here the number of domains is too large to calculate easily the matching edges and
to visualize accurately the structure of the transition graph. Fortunately, the transi-
tion table can still be derived with modest effort, using computer assisted calcula-
tion and manipulation of the transition matrices. The computational tools for this
are supplied in the electronic supplement to Lowenstein et al. (2004), in the form of
Mathematica data structures and functions. Our task is simplified by observing that
we may restrict ourselves to two disjoint classes of symbol-sequence tails, namely
those with i = 6∞ and those with all iL ∈ {0, 1, 2, 3, 4, 5}. The transition tables for
the two classes can be constructed and reduced separately.
Let us work out in detail the nontrivial, irreducible transition tables for the in-
admissible codes in the sector i = 6 ∞ , this time working with matrices instead of
graphs. The first step is to map the domains D j1 (i1 ), i1 = 6, around their respective
return orbits, recording the matching edges in a table. Each entry of the table corre-
[3]
sponds to a particular edge (label a 0 ) of a particular domain D j0 , and lists all edges
of the various Dtj11 (6) which are contained in it. Most pairs ( j 0 , a0 ) contain no such
edges, and we omit them from both columns of the table, leaving us with Table 3.2.
Table 3.2 The preliminary inadmissibility table for the π /7 model, sector i = 6∞
( j0 ,t0 , a0 ) ( j1 ,t1 , a1 )
(0, 1, 1) (0, 1, 1)
(0, 0, 2), (0, 1, 2) (0, 1, 2), (5, 1, 2), (5, 11, 1), (6, 1, 1), (6, 17, 1)
(5, 0, 2), (5, 1, 2), (5, 6, 2) (5,6,2),(6,6,1)
(5, 10, 1), (5, 11, 1) (5, 10, 1), (6, 16, 1)
(6,t, 1),t = 0, 1, 6, 16, 17 (0, 0, 2), (5, 0, 2), (6, 0, 1)
(6, 0, 3) (6, 0, 3)
(8, 10, 1), (8, 22, 1) (8, 10, 1), (8, 22, 1)
(8, 22, 2), (8, 34, 2) (8, 22, 2), (8, 34, 2)
The above table corresponds to an 18 × 18 transition matrix T , which we now

proceed to split into irreducible parts. In the matrix T n , nonzero entries correspond
to (n + 1)-symbol sequences compatible with the recursion rules of inadmissibility.
Specifically, if such a sequence begins with α and ends with β , then T αβ n = 0 for
some n. Thus, to reduce the graph, we can examine entries of
T = T + ε T + ε 2 T 2 + · · · = T (1 − ε T )−1 ,
with ε > 0 chosen small enough to ensure convergence. If T αβ = 0, there is no

finite inadmissibility-compatible symbol sequence connecting α to β , and we con-
clude that the corresponding graph splits into subgraphs A and B connected by links
directed from B to A, with B containing only those nodes which can initiate a path
ending at β .
Explicitly calculating T with (say) ε = .01, we note there are 2 columns with
only a single non-zero entry, and 2 pairs of columns with 2 entries. These produce 4
irreducible sub-matrices, while deleting their 6 rows and columns leaves a 12 × 12
irreducible matrix, all of whose entries are positive. Stripping off the unnecessary
edge labels, we end up with Table 3.3.
Table 3.3 The final inadmissibility table for the π /7 model, sector i = 6∞
( j0 ,t0 ) ( j1 ,t1 )
(0, 0), (0, 1) (0, 1), (5, 1), (5, 11), (6, 1), (6, 17)
(5, 0), (5, 1), (5, 6) (5, 6), (6, 6)
(5, 10), (5, 11) (5, 10), (6, 16)
(6, 0), (6, 1), (6, 6), (6, 16), (6, 17) (0, 0), (5, 0), (6, 0)
(0, 1) (0, 1)
(6, 0) (6, 0)
(8, 10), (8, 22) (8, 10), (8, 22)
(8, 22) , (8, 34) (8, 22) , (8, 34)
Here horizontal lines separate the irreducible sub-tables. From the table we can read
off the period-1 inadmissible tails, namely
(6, 0, 1)∞ , (6, 5, 6)∞ , (6, 5, 10)∞ , (6, 6, 0)∞ , (6, 8, 10)∞ , (6, 8, 22)∞ , (6, 8, 34)∞ .
From the path constraints, we are left with the following admissible period-1 tails:
(6, 5, 2)∞ , (6, 5, 4)∞ , (6, 5, 12)∞ , (6, 5, 14)∞ , (6, 8, 8)∞ ,
(6, 8, 12)∞ , (6, 8, 14)∞ , (6, 8, 30)∞ , (6, 8, 32)∞ , (6, 8, 36)∞ .
Once again, we can diagonalize the transition matrices and easily calculate an upper
bound on the fraction of inadmissible symbol sequences of period n. The ratio is
found to decrease for large n as 0.225708 n.
We can proceed in the same way for the tails with all i L ∈ {0, 1, 2, 3, 4, 5}. The
transition matrix, after pruning the rows (resp. columns) with only zeros and the cor-
responding columns (resp. rows), has dimension 673. The reduction process reveals
the simple graphical structure of Fig. 3.6.
Fig. 3.6 Transition graph for tails with iL ∈ {0, . . . , 5}. The subgraphs A (one node)and C (665
nodes) survive reduction.
where subgraph A contains a single node, B has 7, and C has 665. The single re-
peating node in A corresponds to tail (3, 2, 2) ∞ . The subgraph B and its links can be
deleted, leaving us with a final transition table with 666 = 665 + 1 rows.
A numerical calculation evaluates the largest eigenvalues of the transition and
path matrices to be 8.67058 and 93.3268, respectively, so that the fraction of period-
n inadmissible tails tends to zero for large n as 0.0929056 n.
The constraints for period-1 codes (i, j,t) ∞ , namely n(i) = h(i) and p(i, j,t) = j,
restrict us to the following possibilities as shown in Table 3.4.
Table 3.4 The period-1 codes (i, j,t)∞ of the π /7 model
i j t
1 1 1 − 3, 5 − 7, 14 − 16
1 2 8 − 13, 21 − 25, 33 − 38
1 7 0, 4, 17, 29, 41, 54, 66, 78, 91, 103, 115
2 0 0, 2 − 4, 8, 9
2 2 6, 16, 45, 69
2 3 7, 17, 26, 31, 36, 46, 55, 60, 70, 79, 84, 89, 99, 108, 113
2 4 5, 15, 25, 30, 39, 49, 54, 59, 68, 78, 83, 92, 102107, 112
2 5 1, 11, 21, 35, 45, 59, 69, 88, 98
3 1 3
3 2 1, 3
3 4 2
3 5 0
3 7 2
6 0 1
6 5 2, 4, 6, 10, 12, 14
6 6 0
6 8 8, 10, 12, 14, 22, 30, 32, 34, 36
Consulting the admissibility tables obtained in the previous paragraph, we have un-
derlined those entries (i, j,t) which were found to be inadmissible.
3.4 Minimality
For any catalogue member (D [n] , ρ [n] ), the action of the return map on each Σ [n] (i)
is not only invariant, but also minimal in the sense that, given any x, y ∈ Σ [n] (i) and
ε > 0, there exists a nonnegative integer m such that |ρ m (x) − y| < ε . This was
proved by Lowenstein (2007) making use of the property of the incidence matrices
called semi-positivity.
Definition 3.1. (Lowenstein, 2007, Definition 3) A matrix will be called semi-

positive if each of its rows has either only zero entries or only positive entries.
Theorem 3.1. (Lowenstein, 2007, Theorem 1) Let i be a linked symbol sequence

with the property that , for all positive integers K, there exists an integer P ≥ K such
that A(iK , iK+1 , . . . , iP ) is semi-positive. Let ρ = ρ [n(i)] . Given any x, y ∈ Σ [n] (i) and
ε > 0, there exists a non-negative integer m such that |ρ m (x, ) − y| < ε .
For the proof, we refer the reader to Lowenstein (2007). There it is also shown
that the multi-index incidence matrices A(i K ) of the π /7 model are all semi-positive
except for the cases i K = (0, 3, 3, . . . , 3) and i K = (3, 3, 3, . . . , 3). The only possibly
nonminimal Σ [n] are those for which i has a tail consisting of all 3’s. In fact, however,
if one examines the set of codes σ consistent with such sequences i, it turns out that
only those ending in (3, 1, 3) ∞ are standing in the way of minimality. For all others,
one can safely replace the matrix elements A(0) 0,1 and A(3)1,1 by zeros and apply
Theorem 3.1. But, according to the analysis of the previous section, the code tail
(3, 1, 3)∞ is inadmissible, i.e. it does not actually correspond to points of Σ [n] (i).
Thus all of the invariant sectors of the π /7 model are minimal.
References
Afraimovich V., Maass A., and Urias J. (2000) Symbolic dynamics for sticky sets in Hamiltonian
systems, Nonlinearity 13, 617–637.
Bratteli O. (1972) Inductive limits of finite-dimensional C* algebras, Transactions of the American
Mathematical Society 171, 195–234.
Durand F., Host B., and Skau C. (1999) Substitutional dynamical systems, Bratteli diagrams, and
dimension groups, Ergodic Theory and Dynamical Systems 19, 953–993.
Lowenstein J. H. (2007) Aperiodic orbits of piecewise rational rotations of convex polygons with
Lowenstein J. H., Kouptsov K. L., and Vivaldi F. (2004) Recursive tiling and geometry of piecewise
rotations by π /7, Nonlinearity 17, 1–25.
Poggiaspalla G. (2006) Self-similarity in piecewise isometric systems, Dynamical Systems 21,
147–189.
Vershik A. M. (1985) A theorem on the Markov periodical approximation in ergodic theory, Journal
of Soviet Mathematics 28, 667–674.
Chapter 4
Dimensions and Measures
As we have seen in Chapter 3, the pseudochaotic web (residual set) of aperiodic,

discontinuity avoiding orbits in a renormalizable piecewise isometric system, while
occupying zero area, nevertheless is far from trivial in its dynamics. In probing more
deeply, we are naturally led to ask questions of a probabilistic nature, concerning,
for example, what fraction of the web is contained in a particular subset of the
plane, or with what probability will an orbit initiated in a particular set return to
that set within a designated time? To investigate such questions requires a measure,
and so we turn our attention to the task of constructing one. Fortunately, Hausdorff
measure, widely used in the study of Cantor sets and other fractals (Falconer, 1990),
turns out to be the most natural choice.
4.1 Hausdorff dimension and Hausdorff measure
The s-dimensional Hausdorff measure of a set D in R n is defined (Falconer, 1990)

as

H s (D) = lim Hδs (D),
δ →0
Hδs (D) = inf ∑ |Ui |s : {Ui} is a δ -cover of D
i
where a δ -cover of D is a finite or countable collection of sets U i with diameter less

than δ : def
D ⊂ Ui , |Ui | = sup{|x − y| : x, y ∈ Ui } < δ .
i
For a given D, there exists a unique value of s, the Hausdorff dimension d H (D), such
that
∞ if s < dH (D),
H s (D) =
0 if s > dH (D).
If H dH (D) (D) is nonzero and finite, we define it to be the Hausdorff measure of D.
80 4 Dimensions and Measures
If Hausdorff measure is well defined (not always the case!), it is clearly invariant
with respect to isometries on R n , and inherits the convenient scaling property of
H s,
H dH (ω D) = ω dH H dH (D).
We remind the reader here that “diameter” and “isometry” are to be understood in
terms of the Q metric (2.2). Since the difference is a nonsingular linear transforma-
tion, the value of d H and the existence of Hausdorff measure are the same as for the
Euclidean metric, but obviously the measure of a particular set does depend on the
choice of metric.
4.2 Construction of the measure
We now restrict our attention to a renormalizable piecewise isometry of the square,

with a hierarchy of dressed domains Δ (i 1 , . . . , iL ) = (D(i1 , . . . , iL ), ρ (i1 , . . . , iL ))
rooted in a catalogue {Δ [n] : n = 0, . . . , N − 1}. The return map structure of the re-
cursive tiling hierarchy is characterized by an incidence matrix A j,k (i) specifying
[n(i)]
the number of times the return orbit of D k (i) visits the atom D j , where n(i) is
defined by D(i) ⊂ D [n(i)] . Scaling within the hierarchy is governed by the Q-metric
diameter ratios ω (i) = |D(i)|/|D[h(i)] |, where h(i) is the catalogue member similar
to D(i).
We would like to define a measure μ on the square which vanishes on the peri-
odic cells and boundary lines of tiles, but coincides with Hausdorff measure on the
residual set (hence on the closure of the latter, the exceptional set). We proceed to
construct such a measure as in Lowenstein et al. (2004).
We start with the set S of tiles and cells of arbitrary level, their edges and ver-
tices, and the empty set. Thanks to the recursive tiling structure, the set is a semi-
ring, since it is closed under intersection and, for all A, B ∈ S such that B ⊂ A, the
set A \ B is a finite or countable union of disjoint elements of S . We define μ by
setting μ (A) = 0 if A is a cell, edge, vertex, or empty, and assign a positive number
[n]
ψn, j (to be determined below) to each level-0 atom D j (i), extending the definition
to arbitrary tiles by assuming isometric invariance and the scaling relation
μ (D j (i1 , . . . , iL )) = ω (i1 , . . . , iL )s μ (h(D j (i1 , . . . , iL ))), (4.1)
where the “dimension” s is another parameter to be determined.

To complete the definition of μ , we apply the additivity equations expressing the
root level of the recursive tiling property,
[n]
∑ A j,k (i)μ (Dk (i)) = μ (D j ), n = 1, . . . , N, j = 0, . . . , J(n) − 1. (4.2)
i,k
n(i)=n
4.3 Simplification for quadratic irrational λ 81
Inserting the unknown ψ n, j and the scaling relation (4.1), Equation (4.2) assumes
the form of a matrix eigenvalue equation,
∑ M(n, j),(n, j ) (s)ψn , j = ψn, j , (4.3)

(n , j )
where (treating the pair (n, j) as a single matrix index),
M(n, j),(n, j ) (s) = ∑ δn,n(i) δn ,h(i) A j, j (i)ω (i)s . (4.4)

i
If the determinental condition for a nontrivial solution,
det(1 − M(s)) = 0,
has a solution s = dH , then the ψn, j can be calculated from the linear system (4.2)
up to a common normalization factor.
It is now straightforward, using the recursive tiling property to establish σ -
additivity, to prove that if all the ψ n, j are strictly positive, μ is a genuine probability
measure. The final step, to identify the restriction of μ to the residual set Σ with
Hausdorff measure (up to normalization), is more tedious, and we refer the reader
to Lowenstein et al. (2004) , Theorem 4, for the detailed proof. The main strategy is
to show that for an arbitrary tile D, there exist positive numbers κ 1 and κ2 such that
the measure H dH (D ∩ Σ ) satisfies the inequalities
κ1 μ (D) ≤ H s (D ∩ Σ ) ≤ κ2 μ (D),
and so is nonzero and finite. Since Hausdorff measure must also satisfy the additivity
condition (4.2), and the solution is unique up to normalization, we have the desired
result.
We see that the problem of constructing a suitable measure for the residual set
reduces to an eigenvalue problem for the transfer matrix M(s) with components
given by (4.4). Let us summarize the result as a theorem (after Lowenstein et al.,
2004, Theorem 4 and Corollary 5):
Theorem 4.1. Let M(s) be the transfer matrix for a catalogue with recursive tiling.
Suppose that, for some s = d H , 1 < dH ≤ 2, there exists an M(dH )-invariant vector
ψ whose components are positive. Then a finite, nonzero Hausdorff measure H Hd
can be introduced on the sigma-ring generated by the restrictions of all tiles to the
residual set. On the lowest-level restricted tiles, the Hausdorff measure coincides
with the elements of ψ , up to overall normalization. The Hausdorff dimension d H is
a solution of the transcendental equation det(1 − M(s)) = 0.
4.3 Simplification for quadratic irrational λ

For the models with quadratic irrational parameter which we have studied, the cal-
culation of Hausdorff dimension and Hausdorff measure simplifies considerably.
For most of these models, the recursive tiling structure reduces to a simple scaling
sequence, with a single scale factor ω and a single incidence matrix A. The transfer
matrix then takes the form
M(s) = ω s A,
so that the eigenvalue problem of Theorem 4.1 reduces to
A · ψ = ω −s ψ .
If diagonalization of A produces an eigenvalue a in the interval (1, 2], with corre-

sponding (non-normalized) eigenvector ψ (a), we obtain a Hausdorff dimension
log a
dH =
| log ω |
and a measure determined, up to an overall normalization factor c, by
μ (D j (0)) = cψ (a) j .
For a probability measure on the square, we have

J−1
c−1 = ∑ Tk (0)ψ (a)k ,
k=0
where Tk (0) is the level-0 return time for the kth atom of D(0).
To√illustrate the method, consider Model I from Sec. 2.3 with parameter λ =
(1 + 5)/2. Here the scale factor is ω = λ −2 , and the reduced incidence matrix,
restricted to the polygonal domains D j (0), j = 0, 2, 4,
⎛ ⎞
1 2 4
⎜ ⎟
A = ⎝1 1 1⎠
0 1 3
has eigenvalues 4, 1, 0. Only the first leads to a dimension in the correct range,
namely
log2
dH = .
log λ
The corresponding eigenvector of A is, up to normalization
ψ = (2, 1, 1).
Since the level-0 return times are (see Sect. A.1.2 of the Appendix) 6, 10, and 18,
we obtain
1 1 1
{ μ (D j (0)) : j = 0, 2, 4} = , , .
20 40 40
4.4 A complicated example: Model II 83
4.4 A complicated example: Model II
The only quadratic example which is√slightly more√

complicated than the preceding
one is Model II from Sec. 2.3, λ = 3 and τ = − 3/3, where there are two non-
overlapping pseudochaotic webs, each with its own incidence matrix. We construct
two separate measures, each coinciding with Hausdorff measure on one of the webs
and vanishing on the other. We begin with hierarchy A, with scaling limit (τ , τ ).
From (A.1), the reduced incidence matrix is
⎛ ⎞
0 0 0 0 0 4 2 5 1 0 0 0
⎜0
⎜ 0 0 0 0 24 12 25 0 0 0⎟⎟ 7
⎜ ⎟
⎜0
⎜ 0 0 0 0 5 2 4 1 0 0 0⎟ ⎟
⎜ ⎟
⎜0 1 2 1 1 1 1 1 1 1 1 1⎟
⎜ ⎟
⎜ ⎟
⎜0 0 0 1 2 1 1 1 1 1 1 0⎟
⎜ ⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 0⎟ ⎟
A=⎜ ⎟.
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 0⎟ ⎟
⎜ ⎟
⎜0 0 0 0 0 0 0 0 0 0 0 0⎟
⎜ ⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 0⎟ ⎟
⎜ ⎟
⎜0
⎜ 0 0 0 0 34 19 34 13 1 0 0 ⎟
⎟
⎜ ⎟
⎝0 0 0 0 1 1 1 1 1 1 2 1⎠
1 1 1 1 1 1 1 1 1 1 1 1
Diagonalizing this matrix, we find that the eigenvalues are 4, 1 (3 times), and 0
(8 times). The largest of these yields a Hausdorff dimension
log4 log4
dH = − = √ = 1.05 . . .
log ω log(2 + 3)
and a measure specified on the atoms by
ψ = (0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 1, 1),
with normalization factor c −1 ,
c = T7 (0) + T10 (0) + T21 (0) + T22 (0) = 161280.
Note that only the atoms D j (0), j = 7, 10, 21, 22 contribute to the measure. This can
be understood in terms of the pattern of zeros in A. By a permutation of the labels
of the atoms, the matrix can be transformed into the block form

B C
A→ ,
O D
where B is the 4 × 4 submatrix corresponding to j = 7, 10, 21, 22, and O has only
zero elements. Thus the diagonalization factorizes, with the eigenvalue 4 arising
from B alone.
For hierarchy B, with scaling limit (41 + 72τ , 41 + 72τ ), the reduced incidence
matrix, corresponding to the atoms D j , j = 0, 3, 5, 6, 8, 10, is
⎛ ⎞
0 0 0 1 0 0
⎜0
⎜ 2 2 2 0 0⎟
⎟
⎜ ⎟
⎜0
⎜ 1 1 1 1 2⎟
⎟
A=⎜ ⎟,
⎜1
⎜ 1 1 1 1 1⎟
⎟
⎜ ⎟
⎝0 1 2 3 0 0⎠
0 0 1 0 0 0
with largest eigenvalue 5 and corresponding eigenvector
ψ = (3, 20, 15, 15, 19, 3).
The Hausdorff dimension of this portion of the residual set is thus
log5 log5
dH = − = √ = 1.22 . . ..
log ω log(2 + 3)
The Hausdorff measure on the six atoms is specified by c −1 ψ , with
c = 3 T0 (0) + 20 T3(0) + 15 T5(0) + 15 T6(0) + 19 T8(0) + 3 T10(0) = 5298000.
4.5 Discontinuity set in Model III
In Chapter 2, we discussed in detail the dynamics

√ on the discontinuity set of Model
III, a piecewise isometry on (τ , −τ ] 2 , τ = − 2/2, with scale factor ω = 3 + 4τ .
The complement of the periodic orbits on the ‘base line’ B (the right-hand edge
of Ω ) was found to be a Cantor set C with apparent Hausdorff dimension d H =
log5/ log ω . Let us now verify that result using the transfer matrix method, simul-
taneously obtaining the corresponding measure on the base line.
The scaling on the discontinuity set was found to involve the 2-dimensional do-
mains D j (n), j = 1, 3, 5, 7, plus boundary points, which we can safely ignore in the
measure calculation. For both even and odd levels, the reduced incidence matrix
takes the form ⎛ ⎞
1 2 3 1
⎜0 1 1 1⎟
A=⎜ ⎝2 2 2 1⎠
⎟
1 1 0 1
4.6 Multifractal residual set of the π /7 model 85
with largest eigenvalue 5 and corresponding eigenvector
ψ = (3, 1, 3, 1).
According to Sect. A.3.7, the base line is partitioned, up to discrete points, into 3
copies of D j (−1), j = 0, 2, 6, and a single copy of D 5 (−1). Together, these are par-
titioned into a periodic segment (coinciding with D 5 (−1)) plus 6 copies of D 1 (0), 9
copies of D3 (0), 9 copies of D 5 (0), and 6 copies of D 7 (0). This gives us a normal-
ized measure with values c −1 ψ on the level-zero atoms, with normalization factor
c = 3 × 6 + 1 × 9 + 3 × 9 + 1 × 6 = 60.
4.6 Multifractal residual set of the π / 7 model
As one might suspect, it is the π /7 model, with its more elaborate renormalization
tree, which provides a more challenging application of the transfer matrix construc-
tion of Hausdorff measure on the residual set. As we have seen in Chapter 2, the
latter decomposes into an uncountable infinity of disjoint invariant sets, and the
complex scaling behavior is best described by a multifractal formalism with a con-
tinuous spectrum of dimensions rather than the one or two Hausdorff dimensions
which we encounter in the quadratic models. Nevertheless, it is instructive, to see
how our transfer matrix methods lead (Lowenstein et al., 2004) to a fairly simple
transcendental equation for the overall Hausdorff dimension.
We begin by determining the transfer matrix with the help of the tables in Sect.
A.4 of the Appendix, and Figs. 2.12, 2.13 and 2.16. We note that the catalogue
member Δ [0] appears only at the lowest level of the tiling of Ω , and so will not
contribute to the calculation of d H . Its incidence matrices A(i), i = 9, . . . , 14, will
not be needed, except at the end for imposing the normalization of the measure.
From (4.4), we have the transfer matrix, in block form,
⎛ ⎞
ω (1)s A(1) + ω (2)sA(2) ω (0)s A(0) 0
⎜ ⎟
⎜ ω (4)s A(4) + ω (5)sA(5) ω (3)s A(3) 0 ⎟
⎝ ⎠
ω (7) A(7) + ω (8) A(8)
s s 0 ω (6) A(6)
s
The zeros in the third multi-column correspond to the fact that in the renormalization
tree, there are no dressed domains similar to Δ [3] appearing as children of domains
similar to Δ [1] or Δ [2] . As a consequence, the diagonalization of M(s) factorizes. In
fact (as we shall see below), it is only the upper left, 23 × 23 submatrix which is
relevant.
The problem can be further simplified by expressing the 6 scale factors ω (i) in
terms of the fundamental units η 1 , η2 from (2.14):
ω (0) = η1 η23 , ω (1) = ω (2) = η12 η22 , ω (3) = η1−1 η23 , ω (4) = ω (5) = η22 .
The determinantal equation D(s) = det(1 − M(s)) = 0 then becomes
F(η1s , η2s ) = 0,
where F is the polynomial
F(x, y) = 1 − 6 y2 − 37 x2y2 − 51 xy3 + 9 y4 + 92 x2y4 − 29 x4 y4 + 71 xy5

−69 x3y5 − 4 y6 − 66 x2y6 + 23 x4y6 + 9 x6y6 − 20 xy7 − 48 x3 y7
−50 x5y7 + 25 x2y8 + 19 x4y8 − 9 x6y8 + 5 x3y9 − 6 x5y9 − 6 x4y10 ,
which is irreducible over the rationals. To show the existence of a root between
s = 1 and s = 2, we can evaluate D(1) and D(2) as polynomials in λ , using the
cubic equation for λ to reduce the polynomial to second degree. One finds
D(1) = 20345 + 3096 λ − 7985 λ 2 < 0, (4.5)

D(2) = −7462700 − 1803627 λ + 3299289 λ 2 > 0. (4.6)
Moreover, by bounding the derivative of D(s) in the interval (1, 2), one can prove
that the function has a simple zero s = d H there. It is then a trivial matter, using
numerical root-finding, to evaluate d H to as many digits of precision as desired.
With 20 significant digits, the result is
dH = 1.6522336518816620573... (4.7)
with the corresponding array of un-normalized Hausdorff measures of the 23 atoms

of Δ [1] and Δ [2] (displayed with 6-digit precision):
ψ = (12.2300, 6.22136, 5.56962, 3.20561, 3.20561, 2.97613, 1.30659, 1.22344,

0.428586, 0.771474, 0.771474, 8.56381, 28.5357, 29.2811, 16.4917,
20.8323, 9.14583, 5.94682, 4.40014, 2.61159, 3.27812, 2.67410, 1)
It is now a straightforward but tedious exercise to construct the normalized measure

on the atoms of all catalogue members, but we will not continue along that path.
4.7 Asymptotic factorization
In a renormalizable model with recursive tiling, the geometric scaling is exact. Each
link of the recursion tree is labeled by an index i ∈ {0, 1, . . ., M − 1}, where M is
the total number of parent-child relationships among the N members of the cata-
logue. Corresponding to each i is a unique scale factor ω (i). The diameter of any
tile Dtj0 ,...,tL (i0 , . . . , iL ), relative to that of D [h(iL )] , is given by the cumulative scale
factor along the path i 0 , . . . , iL , namely
4.7 Asymptotic factorization 87
L
ω (iL ) = ∏ ω (ik ).
k=0
The independence of t L is the result of isometric invariance, provided that we take

care to measure diameters using the Q metric.
For the calculation of measures and dimensions, the geometric scaling is, of
course, only one of the two essential ingredients, the other being temporal recursion.
Here we avoid using the word scaling, since in general there will not exist individual
factors associated with links of the tree. Rather we are dealing with J(n(i)) × J(h(i))
rectangular matrices, namely the incidence matrices A(i), which allow us to calcu-
late the level-(L + 1) return times as linear combinations of those of level L:
J(n(iL ))
T j (iL ) = ∑ Tk (iL−1 )Ak, j (iL ),
k=0
or, in matrix notation,

T(iL ) = T(iL−1 ) · A(iL ).
The cumulative incidence matrix along the path i 0 , . . . , iL is just the product, in
left-to-right order, of the matrices A(i),
A(iL ) = A(i0 )A(i1 ) · · · A(iL ). (4.8)
Summing over the row index, one obtains a list of the total ρ [n(i0 )] return times of
the tiles D j (iL ), j = 0, . . . , J(h(iL )). Summing over both indices, one gets the norm
A(iL )||, the total number of tiles in the return orbit of the base domain Δ (i L ).
||A
While the matrix character of the temporal recursion relations is a complication
in the calculation of dimensions and measures, it is not necessarily a troubling one.
This is evident in all of the quadratic models we have studied, for reasons which
are not difficult to understand. In those models, the i dependence is trivial, so that
the matrix product (4.8) become a simple matrix power A L+1 . Moreover, the in-
cidence matrix, in all of our quadratic examples, is a square matrix with a single
largest eigenvalue a > 0, and so, when applied n times on a vector ψ ψ, gives, for
asymptotically large n,
An · ψ ∼ anψ 0 ,
where ψ 0 is a vector proportional to the eigenvector associated with a. Another way
of looking at the situation is the following. Suppose we make a Jordan decomposi-
tion of A,
A = S · A0 · S−1 , (4.9)
where the invertible matrix S can be chosen so that A 0 has a triangular form with
vanishing elements below the main diagonal and the largest eigenvalue placed at
the lower right-hand corner of the matrix. For asymptotically large n, the leading
contribution to A n in Jordan form will be
⎛ ⎞
0 0 · · · 0
⎜0 0 · · · 0⎟
⎜ ⎟
⎜·
⎜ · ·⎟⎟ · S−1 = an u ⊗ v,
a S·⎜
n
⎜· · ·⎟⎟
⎝· 0 0⎠
0 0 · · · 1
where
def −1 −1
(u ⊗ v)k, j = uk v j , u = (S0,J , . . . , SJ,J ), v = (SJ,0 , . . . , SJ,J ).
Note u·v is the J, J element of the identity matrix, hence is equal to unity. Moreover,
u is an eigenvector of the leading part of A n with eigenvalue a n , since
(u ⊗ v) · u = (v · u)u = u.
It is obviously too much to expect that arbitrary products (4.8), constrained only
by the linkage condition J(h(i k )) = J(n(ik+1 )), should have the asymptotic simplic-
ity of simple powers. Nevertheless, it has been possible to extract a factorization
theorem (Lowenstein, 2007) which states that, if certain technical inequalities are
satisfied by A(i), then the product matrix A(i L ) factorizes for asymptotically large
L, with an error which decreases as δ L , with δ < 1. The theorem does not im-
ply a limiting matrix for A(i L )/||A(iL )||, but rather allows the second factor in the
u(iL ) ⊗ v(iL ) tensor product to fluctuate while the first factor approaches a limiting
value, u∞ (i). This will be enough to construct a unique invariant measure supported
on the closure of Σ (i), as well as to apply transfer matrix techniques to investigate
the multifractal dimensions and measures on the full residual set.
We begin by introducing some important definitions from Lowenstein (2007) to
be used in the factorization theorem. Since the dynamical origin of the matrices is
irrelevant here, we will use the notation B(i) instead of A(i).
Definition 4.1. Let B(i), i = 0, 1, . . . , M − 1, be a finite set of p(i) × q(i) matrices
def
with non-negative coefficients. Let B(i N ) = B(i1 ) · B(i2 ) · · · B(iN ). Define
||B(iN )|| = ∑ B(iN )kj , u(iN )k = ∑ B(iN )kj /||B(iN )||, v(iN ) j = ∑ B(iN )kj /||B(iN )||,
k, j j k
B(iN )
w(iN ) = − u(iN ) ⊗ v(iN )T ,
||B(iN )||
j
r(iN ) = max |w(iN )k |, η = max{r(i) : i = 0, . . . , M − 1}, (4.10)
j,k
Δ u(iN ) = u(iN ) − u(iN−1 ),

Δ v(iN ) = v(iN ) − v(iN ).
The factorization theorem will take as its hypothesis certain properties of the
individual matrices B(i). The most important of these is r(i), which is a measure of
4.7 Asymptotic factorization 89
how well the B(i) are already factorizable. If this quantity is sufficiently small, and
in addition the inner products v(i) · u( j) are not too small, it is likely that the family
of matrices B(i) will be well conditioned in the sense of the following definition.
Definition 4.2. A family of matrices B(i) as in Definition 4.1 will be called well
conditioned if the following relations are satisfied:
(i) For i, j = 0, 1, . . . , M − 1, such that p( j) = q(i),

def 1
ε± (i, j) = v(i) · u( j) ± (v(i) · u( j))2 − 4r( j)
2
is real, and moreover,

def
ε = max ε− (i, j) < min ε+ (i, j).
i, j i, j
(ii) The following inequalities hold:
c > 0, V > ε, δ < 1,
where
c = max{(v(i) · u( j) − ε )−1 : i, j = 0, . . . , M − 1, p( j) = q(i)},
V = min{v(i)k : i = 0, . . . , M − 1, k = 0, . . . , q(i)},
δ = max{p( j)2 r( j)(v(i) · u( j) − ε )−2 : i, j = 0, . . . , M − 1, p( j) = q(i)}.
Let us now state and prove the factorization theorem of Lowenstein (2007).
Theorem 4.2. Let {B(i) : i = 0, . . . , M − 1} be a well conditioned family of non-
negative p(i) × q(i) matrices, and let i = (i 1 , i2 , . . .) be an infinite sequence of in-
tegers ik with 0 ≤ ik < M and p(ik+1 ) = q(ik ) for all k. Then, in the notation of
Definitions 4.1 and 4.2,
||B(iN )|| = ||B(iN−1 )||||B(iN )||v(iN−1 ) · u(iN−1 ), N > 1,
and, for k = 0, . . . , p(i 1 ) − 1, j = 0, . . . , q(iN ) − 1,
r(iN ) ≤ ηδ N−1 . (4.11)

|Δ u(iN )k | ≤ cηδ N−2
, (4.12)
|Δ v(iN ) j | ≤ ε , (4.13)
Moreover,
B(iN )
= u(iN ) ⊗ v(iN ) + O(δ N ),
||B(iN )||
and there exists a vector u∞ (i) such that
|u(iN ) − u∞ (i)| = O(δ N ).

Proof. We begin by deriving some useful estimates. From the definitions, the com-
ponents of u(iN ) sum up to unity, as do those of v(i N ). Hence
∑ w(iN )kj = ∑ w(iN )kj = 0.

k j
Moreover, if a represents the arguments i 1 , . . . , iK , and b denotes i K+1 , . . . , iN , then

the following identities follow from substitution into the product formula, B(i N ) =
B(a) · B(b), and the majorization of the components of w(a) and w(b) by r(a) and
r(b):
||B(iN )|| = ||B(a, b)|| = ||B(a)||||B(b)||v(a) · u(b), (4.14)

(w(a) · u(b))k r(a)
|u(a, b)k − u(a)k | = ≤
, (4.15)
v(a) · u(b) v(a) · u(b)

(v(a) · w(b)) j r(b)
|v(a, b) − v(b) | =
j j ≤ . (4.16)
v(a) · u(b) v(a) · u(b)
|w(iN )kj | = (v(a) · u(b))−2 | ∑ w(a)kj1 v(a) j2 w(b)kj1 u(b)k2 (δkj11 δkj22 − δkj12 δkj21 )|
k1 ,k2 , j1 , j2
−2
∑ |w(a)k1 ||w(b)k1 |
j j
≤ (v(a) · u(b))
k1 , j1
≤ q(a)p(b)r(a)r(b)(v(a) · u(b)) −2 . (4.17)
Equipped with the above estimates, we now prove (4.11), (4.12), and (4.13).
The bound (4.11) clearly holds for N = 1, by definition. Moreover, if we assume
r(iN−1 ) ≤ ηδ N−2 , we can use (4.17) and the definition of δ to obtain, for any j, k,
j
|w(iN )k | ≤ r(iN−1 )δ ≤ ηδ N−1 ,
and thus,
r(iN ) ≤ ηδ N−1 .
This establishes (4.11) by induction.
To get (4.12) for arbitrary k and N > 1, we apply (4.15), (4.11) and the definition
of c:
ηδ N−2
|Δ u(iN )k | ≤ ≤ cηδ N−2 .
v(iN−1 ) · u(iN ) − ε
For the third estimate we again resort to mathematical induction. For N = 2 and
arbitrary j, we have
r(i2 ) r(i2 )
|Δ v(i1 , i2 ) j | ≤ ≤ < ε,
v(i1 ) · u(i2 ) v(i1 ) · u(i2 ) − ε
where the final inequality is a consequence of the definition of ε and its assumed
upper bound, which imply, for all j, k such that p(k) = q( j),
4.8 Telescoping 91
ε 2 − v( j) · u(k) ε + r(k) < 0.
Since ε < v( j) · u(k), we have
r(k)
< ε.
v( j) · u(k) − ε
The asymptotic factorization of B(i N )/||B(iN )|| and the convergence of u(i N ) to
u∞ (i), with O(δ N ) corrections, follow immediately from the estimates (4.11) and
(4.12).

The reader will of course be curious whether the abstract factorization theorem
has any relevance to our one nontrivial example of a recursive tiling tree, namely
that of the π /7 model. For simplicity, we restrict ourselves to the cases where i k ∈
{1, 2}, where the matrices B(i) are identified with A(i), i = 0, . . . , 5. The numerical
calculation of the quantities η , ε , c, and δ is easy, but the results are discouraging:
δ turns out to be greater than 1, and so the matrices are not well conditioned.
Fortunately, all hope is not lost. The remedy is telescoping of the recursion tree.
We break up the infinite string into K-element sub-strings, defining new matrices
B(i) as K-fold linked products of A(i) matrices. The first KN levels of the original
tree are now replaced by N levels of the telescoped tree. The reason this might
help is our conjecture, backed up by numerical experiments, that long products of
A(i) really do approach factorizability, so that taking products of K matrices should
reduce significantly the quantity r(i), and perhaps δ as well. For our π /7 example,
attempts with K ≤ 6 continue to produce δ > 1, but K = 7, with a set of 4374
incidence matrices, is a winner! In Lowenstein (2007), the relevant quantities were
calculated exactly, yielding, with 4-digit precision,
η = 7.297 × 10 −6, ε = 1.137 × 10 −4, c = 37.60, δ = 0.6183.
4.8 Telescoping
Once asymptotic factorizability has been established, the advantages of further tele-
scoping are considerable and immediate, as follows from the following theorem.
Theorem 4.3. Given well conditioned matrices B(i), i = 0, . . . , M − 1, define the
telescoped incidence matrix B L (k), k = 0, . . . , ML − 1 to be the kth member of the
set
{B(iL ) : ia = 0, . . . , M − 1, n(ia ) = h(ia−1 )},
ΔuL ,
listed in numerical order. For the new family of matrices, define u L , vL , wL , rL ,Δ
Δ vL as in Definition 4.1. Then
rL (kN ) ≤ ηL δLN−1 ,
|Δ uL (kN ) j | ≤ CηL δLN−2 ,

|Δ vL (kN ) j | ≤ εL ,
where
δL = δ L < 1, ηL = ηδ L−1 , εL = Cηδ L−1 ,
and
cη
C−1 = c−1 − (1 + ε ).
1−δ
Moreover
BL (kN )
= uL (kN ) ⊗ vL (kN ) + O(δLN ),
||BL (kN )||
|uL (kN ) − u∞ (k)| = O(δLN ),
and
min {vL (i) · uL ( j)} − εL ≥ C−1 .
n( j)=h(i)
Proof. The estimate for r(k N ) follows immediately from (4.11). For Δ u L and Δ vL ,
we use Theorem 4.2 to obtain:
cη def
|u(i(N−1)L+1 · · · iNL )k − u(i(N−1)L+1)k | ≤ cη (δ L−2 + δ L−3 + · · · + 1) ≤ =ε,
1−δ
|v(i1 · · · i(N−1)L ) j − v(i(N−1)L ) j | ≤ ε .

Thus, with the help of (4.15) and (4.16),
r(i(N−1)L )
|Δ uL (kN ) j | = |u(iNL ) j − u(i(N−1)L ) j | ≤
v(i(N−1)L ) · u(i(N−1)L+1 · · · iNL )
ηδ (N−1)L−1
≤ ≤ CηL δLN−2 ,
v(i(N−1)L ) · u(i(N−1)L+1 ) − (ε + ε + εε )
r(i(N−1)L+1 · · · iNL )
|Δ vL (kN ) j | = |v(iNL ) j − v(i(N−1)L · · · iNL ) j | ≤
v(i(N−1)L ) · u(i(N−1)L+1 · · · iNL )
ηδ L−1
≤ ≤ εL .
v(i(N−1)L ) · u(i(N−1)L+1 ) − (ε + ε + εε )

4.9 Unique ergodicity for each Σ (i)
In the models with quadratic irrational parameter, the asymptotic factorization of

the products of incidence matrices was closely tied to the existence of a unique
invariant measure supported on the closure of the residual set. Recall that the vector
u appearing in the asymptotic tensor product was proportional to an eigenvector of
the incidence matrix, which, up to normalization, provided the Hausdorff measures
of the various tiles. That a similar situation exists for each of the infinitely many
sectors Σ (i) is the content of the following theorem of Lowenstein (2007).
4.9 Unique ergodicity for each Σ (i) 93
Theorem 4.4. Consider a recursively tiled dressed domain Δ [n] . Let i be an infinite
linked symbol sequence with n(i 1 ) = n, and Σ (i) the corresponding invariant com-
ponent of the residual set. Suppose that the family of J(n(i)) × J(h(i)) incidence
matrices A(i) is well conditioned. Then ρ [n] restricted to Σ (i) is uniquely ergodic.
The measure on each tile d k (iL ) is given, up to renormalization, by the quantity
u∞ (iL+1 , . . .)k appearing in the asymptotic factorization.
Proof. We begin by defining a positive invariant function μ i on residual tiles of Σ (i)
by
u∞ (iL+1 · · · )k
μi (dktL (iL )) = μi (dk (iL )) = , (4.18)
||A(iL )||v(iL ) · u∞ (iL+1 · · · )
where the denominator has been chosen to enforce the normalization condition
∑ A(iL )kj μi (d j (iL )) = 1.

j,k
Further, on any set X which does not intersect the residual tiles, we assign μ i (X) = 0.
By Proposition 1.7 of Falconer (1990), if we can establish the additivity condition
μi (dk (iL )) = ∑ A(iL+1 )kj μi (d j (iL+1 )), (4.19)

j
we will be able to conclude that μ i can be extended to a unique measure on the Borel
sets of R2 , with support on the closure of Σ (i), and so the theorem will be proved.
Our starting point for the additivity proof is the recursive factorization of multi-
level incidence matrices:
A(iL+1 , . . . , iN )kj = ∑ A(iL+1 )kl A(iL+2 , . . . , iN )l j .

l
Summing over j yields
||A(iL+1 , . . . , iN )|| u(iL+1 , . . . , iN )k = ∑ A(iL+1 )kl ||A(iL+2 , . . . , iN )||u(iL+2 , . . . , iN )l .

l
(4.20)
We now use our asymptotic factorization result, Theorem 4.2, to extract the
asymptotic large-N behaviour,
u(iL+1 , . . . , iN )k = u∞ (iL+1 , . . .)k + O(δ N )

= ||A(iL )||v(iL ) · u∞ (iL+1 , . . .) μi (dk (iL )) + O(δ N ),
with a similar expression for u(i L+2 , . . . , iN )k . Inserting these into (4.20) and dividing
through by a common factor, we get
v(iL ) · u∞ (iL+1 , . . .)
μi (dk (iL ))
v(iL ) ·u( iL+1 , . . . , iN )
v(iL+1 ) · u∞ (iL+2 , . . .)
=
v(iL+1 ) · u(iL+2 , . . . , iN ) ∑ A(iL+1 )kl μi (dl (iL+1 )) + O(δ N ).
l
In the limit N → ∞, we get (4.19).

4.10 Multifractal spectrum of recurrence time dimensions
For the quadratic models, there is a pronounced simplicity in the asymptotic scaling,
both spatial and temporal, as one proceeds to higher levels of the renormalization.
While the characteristic lengths (e.g. diameters of tiles and cells) contract by a fac-
tor ω L , the characteristic time intervals (e.g. periods and return times) grow as a L ,
where a is the largest eigenvalue of the incidence matrix. As we saw earlier in this
chapter, this makes possible the construction of a transfer matrix T (α ) whose di-
agonalization determines the Hausdorff dimension (the eigenvalue) and Hausdorff
measure (from the eigenvector).
In the present section, we go beyond the quadratic models, making full use of the
asymptotic factorization property of incidence matrix products, including the accel-
erated convergence achieved by telescoping, to describe the asymptotic spatiotem-
poral properties of the residual set. Instead of a single dimension, characteristic of
a fractal, we find a nontrivial continuous spectrum of dimensions characteristic of a
multifractal.
Before proceeding to our multifractal analysis, we need a more precise definition
of the “characteristic time” of a set. Following Afraimovich et al. (2000), we choose
the recurrence time, defined quite generally as follows:
Definition 4.3. Let S be a set subject to iterated application of a map φ . We define
the recurrence time τ (S) as
τ (S) = min{t : t ≥ 1, φ t (S) ∩ S = 0}.

/
Within the context of recursive tiling, this notion is particularly apt, since if d j
is a residual base tile, mapped by ρ , then the recurrence time of every tile on the
return (to d(i N ) ) orbit {d t = ρ t d j , t = 0, 1, . . . , T j } is the same. To see this, we note
immediately that τ (d t ) ≤ τ (d 0 ), since there are points of d t whose forward orbits
pass through d 0 ∩ d τ (d ) and return to d t after a total of τ (d 0 ) iterations. On the other
0
hand, the fact that d t+τ (d ) ∩ d t is nonempty implies that its preimage d τ (d ) ∩ d 0 is
t t
also nonempty, and so τ (d t ) ≥ τ (d 0 ). This establishes the equality over the whole
return orbit.
We now introduce a generalized dimension which takes into account the joint
spatial and temporal scaling properties of subsets of Σ [n] . We follow the same pattern
used to introduce Hausdorff dimension, except that now we have an extra weighting
factor equal to a power of the recurrence time. For real q and α , positive δ , and
F ⊂ Σ [n] , we define
M (α , q, δ , F) = inf
Cδ (F)
∑ τ (U)−q |U|α ,
U∈Cδ (F)
4.10 Multifractal spectrum of recurrence time dimensions 95
where Cδ (F) is a finite or countable covering of F by residual tiles U of (Q-metric)

diameter |U| ≤ δ , and τ (U) is the recurrence time for U. For δ tending to zero, the
function M (α , q, δ , F) is nondecreasing. We set
M (α , q, F) = lim M (α , q, δ , F),
δ →0
where we allow the possibility M (α , q, F) = ∞.

As a function of α , M (α , q, F) satisfies:
M (α0 , q, F) = 0 =⇒ ∀α > α0 , M (α , q, F) = 0.
M (α0 , q, F) = ∞ =⇒ ∀α < α0 , M (α , q, F) = ∞.
If M (α , q, F) is not identically 0 or ∞ as a function of α , then there exists a critical

value, the recurrence time dimension α (q, F), such that

∞ α < α (q, F),
M (α , q, F) =
0 α > α (q, F).
The set function M (α , q, F) is a special case of a Carathéodory outer measure

(Pesin, 1998) and α (q, F) is the corresponding Carathéodory dimension.. This con-
cept has wide applicability in multifractal theory (Pesin, 1998), especially where it
can be shown that the dimension is independent of F, so that one can write it as
α (q). As we shall see below, this is indeed the case for the recurrence time dimen-
sion in our models with well conditioned recursive tiling.
4.10.1 Auxiliary measures and dimensions
As in our earlier discussion of Hausdorff dimension, we seek to reduce the calcula-

tion of the recurrence time dimension to a transfer matrix eigenvalue problem, since
this technique, if it is feasible, quite effectively captures the essence of the recursive
tiling structure and at the same time produces a set-independent result. Unfortu-
nately, there is no hope of an easy construction of a transfer matrix T (q, α ), thanks
to the complicated nature of the temporal recursion. There is, however, an indirect
route based on the asymptotic factorization of incidence matrix products, which was
set forth by Lowenstein (2007) and which will be reproduced in the remainder of
this chapter.
The first step is find within the recursive tiling formalism suitable proxies for
the domain diameter |D| and recurrence time τ (D) which appear in the function
M (α , q, δ , F). The relevant domains D are, of course, the residual tiles d tN (iN ).
Here we assume that we have L-level telescoped incidence matrices B(i), omitting,
for the sake of notational simplicity, explicit reference to the parameter L except
where it plays an essential role.
The choice of proxy for the diameter is fairly obvious (and was already used in
the Hausdorff case), namely the scale factor ω L (iN ), since there are positive, level-
independent constants κ − and κ+ such that
κ− ω (iN ) ≤ |d tN (iN )| ≤ κ+ ω (iN ),
and the transition from one level to the next is simple multiplicative scaling:
ω (iN+1 ) = ω (iN )ω (iN+1 ),
exactly what we need for constructing a level-by-level transfer matrix.

The natural choice of a proxy for the recurrence time is presumably
ν (iN ) = ||B(iN )|| = ∑ T j (iN ).

j
Thanks to the asymptotic large N scaling of the return times T j (iN ), it is straight-
forward to find uniform upper and lower bounds for the ratios τ (d LtN (iN ))/ν (iN ).
However, the quantity ν (i N ) does not have the kind of product expansion which
would make it ideal for inclusion in a transfer matrix. Specifically, from (4.14), we
have
N
ν (iN ) = (v(i1 ) · u(i2 ))(v(i1 i2 ) · u(i3 )) · · · (v(i1 i2 . . . iN−1 ) · u(iN )) ∏ ||B(ik )||.
k=1
This is not exactly the kind of product expansion we need, namely one involving a
string of factors of the form (v(i k ) · u(ik+1 )), but it is close, in the sense that
v(i1 i2 . . . iK ) = v(iK ) + Δv(iK ), |Δ v(iK ) j | ≤ ε ,
with ε = εL given by Theorem 4.3. This suggests a way out of our problem, namely
defining a measure and calculating the dimension using a transfer matrix constructed
as if we had the desired product structure, then taking the limit L → ∞ to get a
handle on the true outer measure and dimension. This is roughly the strategy which
we adopt below.
In order to bracket the true large-L asymptotic behavior, we introduce, for η ∈
{−, 0, +},
ν η (iN ) = (v(i1 ) · u(i2 ) − ηε )(v(i2 ) · u(i3 ) − ηε ) · · ·

N
(v(iN−1 ) · u(iN ) − ηε ) ∏ ||B(ik )||. (4.21)
k=1
Thanks to the fact that the u(i) k are positive numbers which sum to unity, we have
Δv(iK ) · u(iK+1 )| ≤ ε ,
|Δ
and so we get the bracketing inequalities

v(iK ) · u(iK+1 ) − ε ≤ v(iK ) · u(iK+1 ) ≤ v(iK ) · u(iK+1 ) + ε ,
ν + (iN ) ≤ ν (iN ) ≤ ν − (iN ), (4.22)

as well as the same estimates with ν replaced by ν 0 .
As intended, the spatial and temporal domain functions now have parallel level-
by-level factorization properties, which we can write in the form of scaling ratios,
ω (iN ) ω (iK · · · iN ) ν η (iN ) ν η (iK · · · iN )

= , = (4.23)
ω (iK ) ω (iK ) ν η (iK ) ν η (iK )
Now we are in a position to define measures μ η which will bracket the recurrence-
time outer measure and produce, via the construction and diagonalization of transfer
matrices, approximants to the recurrence-time dimension α (q). To this end, we in-
troduce the following positive invariant function on residual tiles:
μ η (α , q, L, N, d tjK (iK )) = μ η (α , q, L, N, d j (iK ))

= ∑ B(iK+1 , . . . , iN )kj ν η (iN )−q ω (iN )α , (4.24)
k,iK+1 ,...,iN
[n]
μ η (α , q, L, N, d j ) = ∑ B(iN )kj ν η (iN )−q ω (iN )α .
iN ,k
We note that N plays the role of δ in the definition of M (α , q, δ , F), with N → ∞

corresponding to δ → 0, and recall that the factor B(i N )kj counts the number of level-
N residual tiles dktN (iN ) which cover the level-K base tile d ijK (iK ). At this stage, the
telescoping depth L remains an adjustable parameter.
The extension of μ η to finite and countable unions of disjoint tiles is immediate,
employing the easily established additivity formulas
μ η (α , q, L, N, d j (iK )) = ∑ δh(iK ),n(iK+1 ) B(iK+1 )kj μ η (α , q, L, N, dk (iK+1 )),

iK+1 ,k
(4.25)
[n]
μ η (α , q, L, N, d j ) = ∑ δn,n(i) B(i)kj μ η (dk (i)).
i,k
In addition, the scaling relation
μ η (α , q, L, N, d j (iK )) μ η (α , q, L, N, d j (iK ))
= . (4.26)
ν η (iK )−q ω (iK )α ν η (iK )−q ω (iK )α
is a straightforward consequence of the factorization properties (4.23).

We now introduce the doubly indexed transfer matrix
T η (α , q, L)(i1 , j1 ),(i2 , j2 ) = δh(i1 ),n(i2 ) B(i2 ) jj21 (||B(i1 )||(v(i1 ) · u(i2 ) − ηε ))−q ω (i2 )α .
(4.27)
It is not difficult to write μ η (α , q, L, N, d j (iK )) as a matrix element of this matrix
between N-independent vectors Φ η and Ψjη , with components
Φ η (q, L)(i, j) = ν η (i)−q ,

1 K=1
Ψjη (α , q, L, iK )(i , j ) = δ j j ω (iK )
(ν η (iK )/ν η (iK ))−q K > 1.
Two theorems from Lowenstein (2007) complete the extraction of the recurrence-
time dimension. The first of these obtains the approximants α (q, L) from the
transfer-matrix eigenvalue problem:
Theorem 4.5. (Lowenstein, 2007, Theorem 9) Suppose that for all real q and α ,
η ∈ {−, 0, +}, and positive integer L, there exists a positive integer n such that all
elements of the matrix T η (α , q, L)n are strictly positive. Then, for every tile F, there
exists a real number α η (q, L) such that
⎧
⎪
⎪ ∞ α < α η (q, L),
⎨
lim μ η (α , q, L, N, F) = μ η (α , q, L, F) α = α η (q, L),
N→∞ ⎪
⎪
⎩
0 α > α η (q, L),
where μ η (α , q, L, F) has the same additivity and scaling properties as μ η (α , q, L,

N, F) and can be extended to a measure on the Borel sets of R 2 . The three dimen-
sions satisfy the inequalities
α −sign(q) (q, L) ≤ α 0 (q, L) ≤ α sign(q) (q, L).
Proof. The Perron-Frobenius theorem (see Katok and Hasselblatt, 1995) tells us that
the transfer matrix has an isolated largest eigenvalue λ η (α , q, L) associated with a
normalized eigenvector Ω η (α , q, L) with all components positive. Since every ele-
ment of the transfer matrix depends on α only through a factor ω α with 0 < ω < 1,
we infer that λ η (α , q, L) decreases monotonically from +∞ to 0 as α ranges over
(−∞, +∞), passing through the value 1 for unique α η (q, L). This gives the stated be-
havior of μ η for N → ∞, with μ η (α η (q, L), q, L, F) a finite, nonzero matrix element
of the projector along Ω η (α η (q, L), q, L). The additivity and scaling properties are
immediate consequences of (4.25) and (4.26), and the extension to a measure fol-
lows from Proposition 1.7 of Falconer (1990). The inequalities for the dimensions
are a direct consequence of (4.22) and the monotonic decrease of all transfer matrix
elements as functions of α .

The second theorem of Lowenstein (2007) establishes the connection to the

recurrence-time dimension in the limit of infinite L.
Theorem 4.6. (Lowenstein, 2007, Theorem 10) For a telescoped recursive tiling
scheme with L-fold incidence matrices satisfying the conditions of Theorems 4.2
and 4.3, and a transfer matrix satisfying the hypothesis of Theorem 4.5, the dimen-
sions α η (q, L), η ∈ {−, 0, +} have a common L → ∞ limit, which coincides with a
tile-independent recurrence-time dimension α (q).
Proof. To show that the three dimensions α η (q, L) differ by amounts which vanish
in the L → ∞ limit, we focus on the μ η measures of the lowest level tiles d [n] and
assume q > 0. The case q < 0 is strictly analogous and for q = 0, the Hausdorff case,
the three dimensions are identical for all L. We write
μ + (α , q, N, d [n] ) = ∑iN ν (iN )ν + (iN )−q ω (iN )α

− (q,L) ν + (iN ) −q −
= ∑iN ν (iN )ν − (iN )−q ω (iN )α −
ν (iN ) ω (iN )α −α (q,L) .
But, from (4.21),

−q N−1 −q
ν + (iN ) v(ia ) · u(ia+1 ) − ε
1≤
ν − (iN )
= ∏ v(ia ) · u(ia+1 ) + ε
≤ (1 + 2Cε )qN
a=1
and
ωmin
LN
≤ ω (iN ) ≤ ωmax
LN
,
where the L-independent constant C is defined in Theorem 4.3, and ω min = min{ω (i)}
and ωmax = max{ω (i)} are independent of L. Thus

L(α −α − (q,L)) N
μ + (α , q, N, d [n] ) ≤ μ − (α − (q, L), q, N, d [n] ) H(q, L) (1 + 2Cε )q ωmax ,
so that if
q log(1 + 2Cε ) def
α − α − (q, L) > −1
= ξ (q, L),
L log(ωmax )
then
μ + (α , q, L, d [n] ) = lim μ + (α , q, L, N, d [n] ) = 0.
N→∞
It follows that
0 ≤ α + (q, L) − α − (q, L) ≤ ξ (q, L),
since otherwise we could find an α between α − (q, L) + ξ (q, L) and α + (q, L)
such that μ + (α , q, L, d [n] ) = 0, in contradiction to the definition of α + (q, L). But
limL→∞ ξ (L) = 0, and hence the three dimensions have a common limit.
In the final segment of the proof, we show that the recurrence-time dimension
α (q) is sandwiched between α − (q, L) and α + (q, L) for all L, and hence coincides
with their common limit when L tends to infinity. Once again making use of the
level-independent upper and lower bounds on the ratios τ (d)/ν (d) and |d|/ω (d)
for arbitrary residual tiles d, as well as the bracketing of ν (d) between ν + (d) and
ν − (d), we know there exist positive coefficients C ± (α , q, L) such that
M (α , q, δ .F) ≥ C− (α , q, L) inf
Cδ (F)
∑ ν − (d)−q ω (d)α ,
d∈Cδ (F)
M (α , q, δ .F) ≤ C+ (α , q, L) inf
Cδ (F)
∑ ν + (d)−q ω (d)α .
d∈Cδ (F)
But the scaling relation (4.26) implies the bounds

ν − (i)−q ω (i)α
ν − (d)−q ω (d)α ≥ μ − (α , q, L, d) min ,
i, j μ − (α , q, L, d j (i))
ν − (i)−q ω (i)α
ν + (d)−q ω (d)α ≤ μ + (α , q, L, d) max ,
i, j μ − (α , q, L, d j (i))
and so there exist positive coefficients K ± (α , q, L) such that
M (α , q, δ .F) ≥ K− (α , q, L) inf
Cδ (F)
∑ μ − (α , q, L, d)
d∈Cδ (F)
M (α , q, δ .F) ≤ K+ (α , q, L) inf
Cδ (F)
∑ μ + (α , q, L, d).
d∈Cδ (F)
The inequalities imply that M (α , q, δ .F) is infinite for α < α − (q, L) and vanishes
for α > α + (q, L), and hence the critical value α (q, F) is caught between those two
F-independent values for all L. In the limit, they all coincide and are independent of
F.

4.10.2 Simpler calculation of the recurrence time dimensions
Theorems 4.5 and 4.6 suggest a straightforward algorithm for a numerical cal-
culation of the recurrence time dimension α (q): Calculate the largest eigenvalue
λ (α , q, L) of the transfer matrix T 0 (α , q) and use standard root-finding to deter-
mine the value of α where λ (α , q, L) = 1. For fixed q, repeat the experiment with
increasingly large L until the desired precision has been attained. Lowenstein (2007)
found that a simpler eigenvalue problem, using the single-index matrix
Tˆ (α , q, L)i1 ,i2 = ω (i1 )α (||B(i1 )||v(i1 ) · u(i2 ))1−q ,
produces the same dimension as the one based on T 0 (α , q, L) defined in (4.27). To

prove this, it is not necessary to use the simpler transfer matrix to construct a full-
fledged measure. In fact, that may not be possible. Rather, we need only concentrate
on a specific level-0 base tile, d [n] , defining
μ̂ (α , q, L, N, d [n] ) = ∑ ν 0 (iN )1−q ω (iN )α = ∑ Tˆ (α , q, L)N−1

i1 ,iN ν (iN )
0 1−q
ω (iN )α .
iN i1 ,iN
Now we can use the same strategy which was used in the first part of the proof of
Theorem 4.6, writing
ν 0 (iN )
μ̂ (α , q, N, d [n] ) = ∑ ν (iN )ν 0 (iN )−q ω (iN )α ω (iN )α −α (q,L) .
0 (q,L) 0
iN ν (iN )
But
N−1
ν (iN ) v(ia ) · u(ia+1 )
= ∏ v(ia) · u(ia+1 )
ν 0 (iN ) a=1
with each factor bounded above and below as

−1 v(ia ) · u(ia+1 )
(1 + ε C) ≤ ≤ (1 + ε C)
v(ia ) · u(ia+1 )
Thus
NL(α −α 0 (q,L)) μ̂ (α , q, N, d [n] ) N−1 NL(α −α (q,L))

0
(1 + ε C)−N+1 ωmin ≤ ≤ (1 + ε C) ω max .
μ 0 (α , q, N, d [n] )
Defining
μ̂ (α , q, L, d [n] ) = lim μ̂ (α , q, L, N, d [n] ),
N→∞
we have
log(1 + ε C) def
α − α 0 (q, L) > −1
= ξ2 (q, L) =⇒ μ̂ (α , q, L, d [n] ) = 0
L log(ωmax )
and
log(1 + ε C) def
α − α 0 (q, L) < − −1
= −ξ1 (q, L) =⇒ μ̂ (α , q, L, d [n] ) = ∞,
L log(ωmin )
and so the critical α value α̂ (q, L) lies in the interval
α 0 (q, L) − ξ1 (q, L) ≤ α̂ (q, L) ≤ α 0 (q, L) + ξ2 (q, L)
for all L, and hence converges to α (q) in the limit.

The reformulation of the eigenvalue problem allows us to write down immedi-
ately the recurrence-time dimension for renormalizable models with unique spatial
and temporal scale factors, ω < 1 and τ > 1. This includes not only our family of
quadratic models, but also those individual ergodic components of the cubic π /7
model which are characterized by a repeating sequence i = (i 1 , i2 , . . . , iM )∞ . In the
latter case, τ is the largest eigenvalue of the M-fold telescoped incidence matrix,
and
M
ω= ∏ ω (im ).
m=1
It is clear that for finite N, μ̂ is proportional to
∑ τ 1−q ω α ,
iN
which tends to 0 (resp.∞) for α > (q− 1) log τ / log ω (resp. α < (q− 1) log τ / log ω .
Thus the recurrence-time spectrum of dimensions is just the linear function
log τ
α (q) = (1 − q) D, D=− . (4.28)
log ω
We note that α (0) = D is just the familiar Hausdorff dimension.
4.10.3 Recurrence time spectrum for the π /7 model
The residual set of the π /7 model is characterized by infinitely many spatial and
temporal scale factors, and so we do not expect a closed-form expression for the di-
mension spectrum α (q). However, the transfer matrix method makes it straightfor-
ward to obtain quite accurate numerical values with a modest input of computational
effort. As we found in the calculation of Hausdorff dimension earlier in this chapter,
the structure of the transfer matrix is such that the eigenvalue problem factorizes
and we need only consider the two-member catalogue {Δ [1] , Δ [2] }.
Using the incidence matrices for the catalogue, namely A(i), i = 0, . . . , 5 listed
in Tables A.5 and A.6, we calculated the coefficients of Tˆ (α , q, L) for telescoping
depths L = 5, 6, 7. For each of 30 equally spaced values of q in the range 0 ≤ q ≤
1.45, we calculated (Lowenstein, 2007) the largest eigenvalue of the transfer matrix
by the method of iteration, homing in on α̂ (q, L) by means of the secant method of
root-finding. The results for L = 7 are given in Table 4.1, where we have kept only
those digits which are judged conservatively (from a comparison with the L = 6
numbers) to have converged to their infinite-L values. The range of q values was
extended to −8 ≤ q ≤ 8, using the L = 6 matrix, in order to obtain the graph of
Fig. 4.1.
Table 4.1 Selected α̂ 0 (q, L) values for L = 7.
q α̂70 (q) q α̂70 (q) q α̂70 (q)

0 1.652233652 0.5 1.047005707 1 0.4602540423
0.05 1.591027785 0.55 0.9874103750 1.05 0.4028042538
0.1 1.529960826 0.6 0.9280048043 1.1 0.3455986983
0.15 1.469037424 0.65 0.8687944874 1.15 0.2886424095
0.2 1.408262337 0.7 0.8097849509 1.2 0.2319402547
0.25 1.347640428 0.75 0.7509817413 1.25 0.1754969094
0.3 1.287176666 0.8 0.6923904094 1.3 0.1193168306
0.35 1.226876115 0.85 0.6340164934 1.35 0.06340423039
0.4 1.166743930 0.9 0.5758655008 1.4 0.007763050387
0.45 1.106785355 0.95 0.5179428886 1.45 −0.04760306470
Fig. 4.1 The α (q) curve for Example II.
It is instructive to compare the values of α̂ (q, 7), which we believe to be excel-

lent approximations to α (q), with values obtained by independent methods at three
points on the curve: q = 0, q = 1, and the point q = q 0 such that α (q0 ) = 0. From
the definition of M (α , q, d [n] ), α (0) is the Hausdorff dimension of d [n] , while α (1)
is that of the fractal set obtained by taking the N → ∞ limit of the union over all
base tiles d(iN ) with n(i1 ) = n. Earlier in this chapter, the transfer matrix method
was used to obtain α (0) as the solution of a transcendental equation of the form
α (0) α (0)
F(η1 , η2 ) = 0,
where η1 and η2 are the fundamental units of the ring Z[λ ] and F(x, y) is a polyno-
mial of degree 14. This yielded the value
α (0) = 1.6522336518816627081. . ..
The analogous calculation for the base-tile fractal gives α (1) as the solution of
α (1) α (1)
G(η1 , η2 ) = 0,
where
G(x, y) = x2 y2 + xy3 + y2 − 1.
A numerical value with 20-digit precision is easily obtained by Newton’s method:
α (1) = 0.46025404225607400229. . .
It is encouraging that the L = 7 numerical transfer matrix calculation reproduces

these dimensions with 10-digit precision.
The third point of comparison is less precise, but important for the interpretation
of the model as pseudochaotic, i. e. having a power-law increase of complexity over
time. From Table 4.1, we see that q 0 is slightly larger than 1.40. Using the same
numerical construction of α̂ (q, 7) as before, supplemented by the secant method to
home in on the zero of that function, we obtained
q0 = 1.40699563 (4.29)
with an error bounded by 10 −8 . To interpret this exponent (nonrigorously), we note

that the calculation with α set equal to zero parallels closely that of Hausdorff or
box-counting dimension, provided that we think of the inverse of the recurrence time
as the “size” of a covering set. Then t q0 is seen to be proportional, for asymptotically
large t, to the number of tiles in all the return orbits with recurrence time less than or
equal to t, a quantity which is one of various possibilities for measuring the growth
of complexity in this class of models. Another is the number of polygonal cells of
periodic orbits with return time less than t. A numerical experiment by Lowenstein
et al. (2004) produced a value for q 0 − 1 which agrees with (4.29) to 4 digits of
precision.
We note in passing that application of the above calculation of the complexity-
growth exponent q 0 to the quadratic cases, using (4.28), gives q 0 = 1. This is consis-
tent with the observation that in the quadratic models the number of periodic cells
increases, on average, proportional to the return time, as one ascends the renormal-
ization ladder of each scaling sequence.
The plot in Fig. 4.1 suggests that the function α (q) resembles a hyperbola, with
linear asymptotes for q → ±∞. This type of behavior should not come as a surprise.
After all, for q → ∞, the calculation emphasizes the residual tiles with the largest
diameters and smallest return times, which are found to scale with a temporal factor
τ+ = 4 and spatial factor ω = 2 − λ , where λ is the parameter 2 cos π /7. On the
other hand, for q → −∞, the tiles which are dominant are those with the smallest
1 √
diameters and longest return times, with scaling factors τ − = (51 + 2641) and
2
ω− = −5 + λ + λ 2 . Thus
α (q) ∼ (1 − q)β±, q → ±∞,
where
√
log 4 log(51 + 2641)
β+ = − = 0.856 . . ., β− = − = 1.304 . . ..
log(2 − λ ) 2 log(2 − λ )
It is sometimes useful to reformulate the description of a multifractal in terms of a

Legendre transform of α (q) (see Falconer, 1990, Chap. 17). One advantage of this
is that the infinitely extended asymptotes of the α (q) get mapped into the endpoints
of a finite curve, whose peak value is just the Hausdorff dimension. Specifically, we
define f (β ) parametrically by the equations
d α (q)
β (q) = − ,
dq
f (q) = α (q) + qβ (q).
References 105
and display its graph in Fig. 4.2. The interpretation of this function is roughly the
following: if one covers the residual set by tiles of diameter approximately equal
to some small δ , the number of such tiles with box-counting dimension β will be
proportional to δ − f (β ) .
Fig. 4.2 The f (β ) curve for Example II.
References
Afraimovich V., Schmeling J., Ugalde E., and Urias J. (2000) Spectra of dimensions for Poincaré
recurrences, Discrete and Continuous Dynamical Systems 6, 901–914.
Falconer K. (1990) Fractal Geometry, Wiley, Chichester.
Katok A. and Hasselblatt B. (1995) Introduction to the Modern Theory of Dynamical Systems,
Cambridge University Press, Cambridge.
Lowenstein J.H. (2007) Aperiodic orbits of piecewise rational rotations of convex polygons with
Pesin Ya. B. (1998) Dimension Theory in Dynamical Systems: Contemporary Views and Applica-
tions, University Chicago Press, Chicago.
Chapter 5
Global Dynamics
5.1 Global expansivity
In the preceding chapters, we have achieved a detailed understanding, based on the

organizing principles of renormalization and symbolic coding, of the local dynam-
ics associated with renormalizable kicked-oscillator maps. We now wish to show
how the whole recursive return map formalism can be extended (“lifted”) to the
aperiodic orbits on the infinite plane of the global map W . The local scaling, char-
acterized by the self-similar contraction of domains in the fundamental cell Ω , will
be seen to have, if certain conditions are satisfied, a global counterpart in the form
of asymptotic self-similar expansion. Just as the local contraction (within a single
ergodic component) is associated with a scale factor ω K < 1, the global expansiv-
ity will have its own scale factor ωW > 1. Given that the temporal behavior scales
with a factor ωT , it is not surprising that orbits will be found to tend to infinity for
asymptotically long times with a power-law exponent log ω W / log ωT .
5.1.1 Lifting the return map ρK (0)
We recall from Chap. 1 that the sawtooth kicked-oscillator map W can be factored
into local and global piecewise linear maps. It is convenient to break up the plane
into translates
√ of the square Ω , each such cell labeled by a Gaussian integer n +
mi, i = −1, m, n ∈ Z. Each point x of the plane is then represented by a pair [u, ζ ],
with the local point u a point of Ω and ζ ∈ Z[i] labeling the cell. The map W is then
given by
W [u, ζ ] = [K(u), iζ + Δ (u)],
the equivalent, in the new notation, of (1.7). The orbit of [u, ζ ] then consists of the
points
W n [u, ζ ] = [K n (u), in ζ + in−1 Δ (u) + in−2 Δ (K(u)) + · · · + Δ (K n−1 (u))].

108 5 Global Dynamics
Now suppose that u is a point in one of the domains (level=0 base tiles) D j (i0 )
of a recursive tiling hierarchy, i.e. performing T j (i0 ) iterations of K on u will map
the latter back into D(i 0 ) for the first time, i.e.
K Tj (i0 ) (u) = ρK (i0 )(u).
Here and in what follows we will use ρ K , in place of ρ , to denote the local return
map of Sect. 4. Then the same number of W iterations on [u, ζ ] produces
W Tj (i0 ) [u, ζ ] = [ρK (i0 )(u), iTj (i0 ) ζ + d j (i0 )] = ρW (i0 )[u, ζ ],
def
where d j (i0 ) is the lattice translation associated with the map ρW (i0 ), namely,
d j (i0 ) = iTj (i0 )−1 Δ (u) + iTj (i0 )−2 Δ (K(u)) + · · · + Δ (K Tj (i0 )−1 (u)).
One must be careful here to interpret the notation properly: ρ W (i0 ) is not in gen-
eral a true first return map, but rather a first return map modulo a lattice transla-
tion: the domain ρW (i0 )[D j (i0 ), ζ ] is identical to [ρK D j (i0 ), ζ ], translated to the cell
iTj (i0 ) ζ + d j (i0 ). This is illustrated in Fig. 5.1, which shows the respective local and
global orbits of the tile D 0 (0) of the fundamental cell in the quadratic Model I with
Fig. 5.1 Local and global orbits of the level-0 atom D0 (0). After 6 iterations of the local map K,
the image domain has returned, ‘rotated’, to D(0). After 6 iterations of W , on the other hand, the
image domain lands, with the same ‘rotation’, in D(0) + (τ , −2τ ).
5.1 Global expansivity 109
1 √
parameter (1 + 5). In the local orbit, 6 copies of the tile, rotated as well as trans-
2
lated, are located throughout Ω , sharing with the orbits of D 1 (0) and D2 (0) that
part of Ω not occupied by periodic pentagons (see Sect.2.3). The global orbit, on
the other hand, consists of translates of the 6 tiles located in a number of nearby
lattice cells.
We see here the possibility of long-distance transport on the plane under iteration
of ρW . To investigate this possibility, we must first study the lifted versions of the
higher-level return maps ρ K (i0 , i1 , . . . , iL ).
5.1.2 Lifting the higher-level return maps
To construct the higher-level maps ρ W (iL ) which elevate ρK (iL ) to the plane, we
proceed recursively. Suppose we have succeeded in defining
ρW (iL )[u, ζ ] = [ρK (iL )u, iTj (iL ) ζ + d j (iL )], u ∈ D j (iL ). (5.1)
We wish to construct, for u ∈ D j (iL+1 ),
ρW (iL+1 ) [u, ζ ] = ρW (iL )ν (iL+1 , j) [u, ζ ],
iterating ρW (iL ) along the ν (iL+1 , j)-step return path p(i L+1 , j).
For an arbitrary t-step path ( j 0 , j1 , . . . , jt−1 ), we have
ρW (iL )[u, ζ ] = [ρK (iL )u, iTj0 (iL ) ζ + d j0 (iL )],

ρW (iL )2 [u, ζ ] = [ρK (iL )2 u, iTj0 (iL )+Tj1 (iL ) ζ + iTj1 (iL ) d j0 (iL ) + d j1 (iL )],
..
.
t−1 t−1
ρW (iL )t [u, ζ ] = [ρK (iL )t u, i∑s=0 Tj0 (iL ) ζ + i∑s=1 Tj0 (iL ) d j0 (iL ) + · · · + d jt−1 (iL )].
Introducing the more compact notation

t−1
j = ( j0 , . . . , jt−1 ), Tj (iL ) = ∑ Tjs (iL ),
s=0
t−1
T(j|r) (iL ) = ∑ Tjs (iL ), 0 ≤ r ≤ t − 1, T(j|t) (iL ) = 0, (5.2)
s=r
t−1
ck,j (iL ) = ∑ δk, js iT(j|s+1) (iL ) ,
s=0
we get
ρW (iL )t [u, ζ ] = [ ρK (iL )t u, iTj (iL ) ζ + ∑ ck,j (iL ) dk (iL ) ]. (5.3)
k
For the full return path of u ∈ D j (iL+1 ), we have t = ν (iL , j) and thus, from (5.3),
Tp(i (i )
ρW (iL+1 )[u, ζ ] = [ ρK (iL+1 ) u, i L+1 , j) L ζ + d j (iL+1 ) ],
with
d j (iL+1 ) = ∑ M jk (iL+1 ) dk (iL ),
K
ν (iL+1 , j)−1
(i )
∑
Tp(i
M jk (iL+1 ) = ck,p(iL+1 , j) (iL ) = δk,p(iL+1 , j,s) i L+1 , j|s+1) L . (5.4)
s=0
Here we have introduced the partial path

def
p(i, j | s) = (p(i, j, s), . . . , p(i, j, ν (i, j) − 1)),
with the conventions
p(i, j | ν (i, j)) = 0,

/ T0/ (iL ) = 0.
Since we have previously constructed ρ W (i0 ) satisfying (5.1) with L = 0, we now

have a recursive construction of ρ W (iL ) satisfying the equation for all L.
We noted in Chap. 4 that if the sector code i is eventually periodic, i.e. has
a tail (i1 , . . . , i p )∞ , then the recursion formula (2.13) implies asymptotic temporal
scaling with a scale factor ω T > 1 equal to the largest eigenvalue of A(i p )T =
(A(i1 ) · · · A(i p ))T , provided that certain minor technical assumptions on the ma-
trices A(i) are valid. Eventual periodicity of i also leads, under appropriate circum-
stances, to asymptotic scaling of the global translation vectors. To see this, define a
vector of return times modulo 4,
τ (iL ) = (τ0 (iL ), τ1 (iL ), . . . , τJ(h(iL ))−1 (iL )), τ j (iL ) = T j (iL ) mod 4.
For sufficiently large L 0 , iL0 +np = iL0 , and the vectors τ (iL0 +np ), n = 0, 1, . . . have
the same length and are related by
τ (iL0 +(n+1)p ) = τ (iL0 +np ) · A(i p ).
But there are at most 4 J distinct τ vectors of length J, and so the sequence
{ττ(iL0 +np ), n = 0, 1 . . .} is eventually periodic with some period N, starting with
some n = n0 . From (5.4) we see that the sequence of matrices M jk (iL ) becomes
periodic, with period P = N p, starting at L 1 = L0 + n0 p,
M jk (iL+P ) = M jk (iL ), L ≥ L1 .
Introducing
M̂k = M(iL1 +k ), k = 1, 2, . . . , P,
we have the following recursion relation for the global translation vector,
5.2 Long-time asymptotics 111
d(iL ) = (d0 (iL ), d1 (iL ), . . . , dJ(h(iL ))−1 ),
d(iL1 +nP ) = (M̂P · · · M̂2 · M̂1 )n · d(iL1 ),

and so d(iL1 +nP ) will scale, for asymptotically large n as the nth power of the largest-
magnitude eigenvalue ωW of the global recursion matrix
M = M̂P · · · M̂2 · M̂1 ,
that is, if M · ξ = ωW ξ , then
d(iL1 +nP ) ∼ ωWn ξ , (5.5)
provided that the eigenvector ξ of ωW is unique and d(i L1 ) has a non-zero projection
along ξ . These last conditions are not automatically satisfied by the complex matrix
M, and so one has to be careful to verify them explicitly in specific models.
Invariant sectors of sawtooth kicked-oscillator models whose global translation
vectors satisfy (5.5), with ωW > 1 will be called globally expansive. In Sect. 5.2,
after developing some necessary machinery of symbolic dynamics, we will explore
the implications of global expansivity for the asymptotic long-time behavior of saw-
tooth kicked-oscillator maps.
5.2 Long-time asymptotics
We now return to the question of the long-time behavior of globally expansive

kicked-oscillator models. What we will be looking for is asymptotic power-law be-
havior, meaning roughly that an orbit initiated in the cell labeled by complex coor-
dinate ζ0 has points ζt which, for sufficiently large t, have |ζ t | ∼ At μ , where A and
μ are positive constants and corrections are of negligible relative size. We do not
insist that the power-law growth be uniform in time, since there are known exam-
ples where the orbit exhibits the phenomenon of recurrence, returning sporadically
to the vicinity of the origin. Thus, for t → ∞, the sequence of ratios |ζ t |/t μ have an
accumulation point, not a true limit, at A. Formally this is expressed as
|ζt |
lim sup = A. (5.6)
t→∞ tμ
To simplify our discussion power-law behavior, we will make certain basic as-
sumptions. First of all, we assume that for level L larger than or equal to some
fixed minimum L ∗ , the modulo-4 return-time vector ττ is independent of L, as are
the incidence matrix A and the global recursion matrix M. For an invariant sector
with global expansivity, these properties can be achieved by suitably telescoping the
recursive tiling structure (see Sect. 4.8). in addition, we assume that the vector of
elementary lattice displacements d(i L ) and the coefficient vectors c j of (5.2) all have
non-zero components in the direction of ξξ , the eigenvector of M with largest eigen-
value ωW > 1. Note that cj , where j is any piece of any path p(i, j), is independent
of level due to our assumption on ττ .
The principal benefit of the assumptions of the preceding paragraph is that we
have both upper and lower bounds on the exponential growth of arbitrary lattice
excursions as a function of level. Consider a piece of an orbit along a path (part
j of some p(i, j)) at level L ≥ L ∗ . If the orbit fragment starts at ζ , it will end at
iτ (j) ζ + Δ ζL (j) where
Δ ζL (j) = cj · d(iL ).
From (5.5) , the assumed non-vanishing of c j · ξ , and the fact that there are only a
finite number of possible j, we have
|Δ ζ (j)|
α− ≤ ≤ α+ , (5.7)
ωWL
where α± are positive bounds independent of L and j.

We now want to use the inequalities (5.7) to obtain an upper bound on the lattice
excursion from ζ 0 at time 0 to ζt at time t along an orbit of W . An optimal decom-
position of the orbit into a concatenation of return-orbit fragments is suggested by
the symbolic representation of Sect. 3.1. Suppose that the local components of the
initial and final points correspond, respectively, to symbolic codes
σ = ((i0 , j0 ,t0 ), (i1 , j1 ,t1 ), . . . , (iL , jL ,tL ), (iL+1 , jL+1 ,tL+1 ), . . .),
σ = ((i0 , j0 ,t0 ), (i1 , j1 ,t1 ), . . . , (iL , jL ,tL ), (iL+1 , jL+1 ,tL+1 ), . . .),
where t < t and the symbols coincide from level L + 1 onward (as they must, for
some L, if the codes correspond to points on the same orbit). To get from initial
to final code as efficiently as possible, we adopt the strategy of first applying W
ν (i0 , j0 ) − t0 times to reach (at least) level 1, then applying ρ W (i0 ) ν (i1 , j1 ) − t1 − 1
times to reach (at least) level 2, etc. until we reach level L, where application of
ρW (iL−1 ) tL − tL − 1 times brings the symbols into coincidence. Now we work our
way down the hierarchy, applying ρ W (iL−2 ) tL−1
times, ρW (iL−3 ) tL−2 times, until

finally application of W t 0 times completes the job.
Note that the levels L1 , L2 , . . . visited on the way up the hierarchy are not nec-
essarily consecutive, thanks to the possible presence of maximal symbols with
tK = ν (iK , jK ) − 1. Thus, after the first ν (i 0 , j0 ) − t0 iterations of W , if t1 is max-
imal, it will be reset to zero and t 2 incremented by one, unless it too is maximal,
etc. Thus Lk , k = 1, 2, . . . , is the kth non-zero level such that TLk is non-maximal.
Similarly, on the way down the hierarchy, the return maps are applied only at levels

Mk , k = 1, 2, . . . , where tM k
is non-zero.
The sequence of codes in the above can be followed easily by keeping track of
the changes in the t-components of the symbols σ K , k = 0, 1, . . . , L (since the i-
components do not change and the j-components are completely determined, read-
ing from right to left, by the path constraints), as shown in the following diagram.
5.2 Long-time asymptotics 113
(t0 ,t1 , . . . ,tL )

(0, . . . , 0,tL1 + 1,tL1 +1 , . . . ,tL )

(0, . . . , 0,tL2 + 1,tL2 +1 , . . . ,tL )

..
.

(0, 0, . . . , 0,tL+1 )
↓
(0, 0, . . . , 0,tL )

(0, . . . , 0,tM , . . . ,t )
n L

..
.

(0, . . . , 0,tM , . . . ,t )
1 L

(t0 ,t1 , . . . ,tL )
L(t)
To obtain an upper bound on |ζ T − ζ0 |/ωW , we apply (5.7) to each of the return-
map excursions with level ≥ L ∗ and sum up. The contributions from levels below
−L(t)
L∗ will be of order ωW . Thus

|ζt − ζ0 | m
L −L(t)
n
M −L(t) −L(t)
L(t)
≤ α+ 1 + ∑ ωWk + ∑ ωW k + O(ωW ). (5.8)
ωW k=1 k=1
Since ωW > 1 and L(t) − Lk ≥ 1, L(t) − Mk ≥ 1, we have

|ζt − ζ0 | ωW + 1
lim sup ≤ α+ .
t→∞ ωW
L(t) ωW − 1
To show that the lim sup is non-zero, let us assume the contrary. Then infinitely
L
many of the quantities |ζ k − ζ0 |/ωWk−1 , where ζk is the starting point of the level L k
piece of the orbit, are less than α − /2 − ε , where ε is an arbitrarily small positive
number. For such k values, we have
|ζk+1 − ζ0 | |Δ ζk | |ζk − ζ0 | α− α−
≥ − ≥ α− − ( − ε) = + ε.
ωWLk ωWLk L
ωWk−1 2 2
But this would imply the existence of infinitely many orbit points ζ k with |ζk −
L
ζ0 |/ωWk−1 more than α− /2, contradicting the assumption of vanishing limsup.
The temporal inequalities are easier to derive, since they involve only sums over
positive quantities. In parallel with (5.8), we have

t ωT + 1 −L(t)+1
β− ≤ lim sup L(t)−1 ≤ β+ + O(ωT ).
t→∞ ω ω T − 1
T
Introducing
log ωW
δ= ,
log ωT
we have δ L(t)−1
ωT − 1 ωW
β+−δ ≤ lim sup ≤ β−−δ . (5.9)
ωT + 1 t→∞ tδ
|ζt − ζ0 |
Combining the limsup inequalities, we see that lim sup t→∞ is finite and
tδ
non-zero. This establishes the connection between global expansivity and asymp-
totic power-law behavior that we were seeking.
5.3 Quadratic examples
In Lowenstein et al. (2005) the issues of global expansivity and asymptotic power
laws were studied thoroughly in a number of sawtooth kicked-oscillator models with
quadratic irrational parameters. The models exhibited a wide range of behaviors, in-
cluding both bounded and unbounded orbits, with the latter exhibiting logarithmic,
sub-diffusive, diffusive, and super-diffusive growth at infinity. The results are sum-
marized in Table 5.1.
Table 5.1 Summary of local and global scaling parameters for the quadratic kicked-oscillator mod-
els. The last two lines refer to the residual set and discontinuity set of Model III, respectively.
λ ω ωT ωW μ behavior
√ √
2 3−2 2 9 1 0 bounded
√ √
− 2 3−2 2 9 1 0 bounded
√ √
(−1 + 5)/2 (3 − 5)/2 4 1 0 bounded
√ √
(−1 − 5)/2 (3 − 5)/2 4 1 0 bounded
√ √
(1 + 5)/2 (3 − 5)/2 4 4 1 ballistic
√ √
(1 − 5)/2 (3 − 5)/2 4 4 1 ballistic
√ √
− 3 7−4 3 25 4 .430677 sub-diffusive
√ √
3(A) 2− 3 4 2 .5 diffusive
√ √
3(B) 2− 3 5 2 .4306770 sub-diffusive
√ √
2(A) 3−2 2 9 1 0 logarithmic
√ √
2(B) 3−2 2 9 5 .732487 super-diffusive
5.3 Quadratic examples 115
Perhaps the most √ interesting of these quadratic examples is Model II of Sect. 2.3,
with λ = −3τ = 3, for which there are two distinct scaling sequences, A and B,
each with its own scaling domain, limit point, temporal scale factor, and fractal di-
mension. In the case of hierarchy A, with limit point (τ , τ ), local geometric scale
factor ωK = 2 − λ and temporal scale factor ω T = 4, there are 23 sub-domains,
including 11 boundary segments. For simplicity, we limit our attention to those ape-
riodic orbits for which which the level-L sub-domain indices j L , L = 0, 1, 2, . . . are
restricted to the values 7, 10, 21, 22. Relabeling these 0, 1, 2, 3, respectively, we write
down (from Kouptsov et al., 2002) the return paths and return times,
p(0) = (3, 1, 0), p(1) = (3, 1, 1, 2, 0) p(2) = (3, 1, 2, 2, 0) p(3) = (3, 2, 0),
T0 (L) = T3 (L) = 26880 × 4 L + 53, T1 (L) = T2 (L) = 53760 × 4 L − 53,

so that the modulo-4 return-time array is
τ (L) = (1, 3, 3, 1),
independent of level L. The resulting global recursion matrix is

⎛ ⎞
1 i 0 1
⎜1 1−i −1 ⎟
⎜ i ⎟
M=⎜ ⎟,
⎝ 1 −i 1 + i −1 ⎠
1 0 i 1
(A)
with dominant eigenvalue ωW = 2 and corresponding eigenvector (1, 0, 0, 1). The
lowest level global displacement vector,
d(0) = (22 + 310i, 18 − 8i, 18 + 22i, −8 + 310i),
and for asymptotically large L,
(A)
d(L) = ML · d(0) ∼ 328i(ωW )L (1, 0, 0, 1).
The invariant subset of orbits is indeed globally expansive. One can, moreover, ver-
ify that the coefficient vectors c j for partial paths j have nonzero projections along
(1, 0, 0, 1), and so the orbits should display asymptotic power-law behavior, with
exponent
(A)
log ωW log2 1
μA = (A)
= = .
log ωT log4 2
The power law is thus a diffusive one.

As a simple example of an aperiodic orbit in this sector, we chose an initial point
with a period-1 admissible symbol sequence,
1 √ √
σ = (1, 1)∞ , x(σ ) = (−1422 + 821 3, 7770 − 4486 3).
39
Here we used (3.5) to calculate x(σ ) as a pair of algebraic numbers in Q(λ ). A plot
of the orbit (55344 iterations of ρ (0)) is shown in Fig. 5.2. Although it is restricted
to a zero-area pseudo-chaotic web, the orbit is clearly exploring the plane in a fairly
uniform manner, moving outward, on the average, at the rate of a random walk. The
high level of self-similarity is due to the choice of periodic symbol sequence, and
is made more evident in a log-log plot, in Fig. 5.3, of the squared distance from the
origin as a function of iteration number. √
For the scaling domain for scaling sequence B, the limit point is (41−24
√ √ 3, 41−
24 3), and although the local geometric scale factor is again 2 − 3, the temporal
(B)
scale factor is now ωT = 5. In this case, we consider all 6 polygonal sub-domains,
designated 0,3,5,6,8,10 for Model II in Sect. 2.3, relabeling them as 0,1,2,3,4,5,
respectively. Their paths and modulo-4 return times are, from Kouptsov et al. (2002),
p(0) = (3), p(1) = (3, 1, 4, 1, 2), p(2) = (3, 1, 4, 5, 4, 1, 2)

p(3) = (3, 1, 4, 0, 4, 4, 1, 2), p(4) = (3, 2), p(5) = (3, 2, 2),
τ (L) = (3, 0, 3, 3, 2, 1).
Since τ (L) is independent of L, we can calculate a level-independent global recur-

sion matrix,
⎛ ⎞
0 0 0 1 0 0
⎜ 0 −i 0⎟
⎜ 0 1 i ⎟
⎜ ⎟
⎜ 0 1−i 1 1 −1 − i i ⎟
M=⎜ ⎜ ⎟,
⎜ −i 1 − i 1 1 −1 0⎟⎟
⎜ ⎟
⎝ 0 0 1 −i 0 0⎠
0 0 1 − i −1 0 0
with dominant eigenvalue ω (B) = 2 and corresponding eigenvector (2i, −1+ i, 4i, 4i,
2 + 2i, 2). As in the A sector, the hypotheses of global expansivity and asymptotic
power-law behavior are satisfied, this time with the sub-diffusive exponent
log2 1
μB = < .
log5 2
In Fig. 5.4 we show the orbit of the point x((1, 2), (1, 1) ∞ ) (1438699 iterations
of ρ (0)). Like the A orbit, this one explores large regions of the plane, moving to
infinity at a rate slower than that of a normal random walk. The log-log distance vs.
time plot in Fig. 5.5 shows an interesting aspect of this orbit: in spite of the steady
progress toward infinity, the orbit is recurrent, i.e. it returns infinitely many times to
the vicinity of the origin.
5.3 Quadratic examples 117
Fig. 5.2 First 104 ρW (0) iterations of the aperiodic orbit starting at x(σA ), σA = (1, 1)∞ .
Fig. 5.3 Log-log plot of squared distance versus number of iterations of ρW (0).
Fig. 5.4 First 1438699 ρW (0) iterations of the aperiodic orbit starting at x(σB ), σB = (1, 2)(1, 1)∞ .
Fig. 5.5 Log-log plot of the squared distance vs. the number of iterations of ρW (0).
5.4 Cubic examples 119
5.4 Cubic examples
We now consider some more complicated, but highly instructive, examples from
the π /7 model. We recall that the set Σ of discontinuity-avoiding aperiodic orbits
decomposes into infinitely many invariant sectors, each labeled by a sequence of in-
dices i = (i0 , i1 , . . .). To explore the phenomena of global expansivity and asymptotic
power-law behavior, we restrict ourselves to eventually periodic index sequences
with period 1.
5.4.1 Orbits in the (0, k, 6∞) sectors
As our first, we consider 3 sequences of the form (0, k, 6, 6, 6, 6 . . .), k = 9, 11, 13,
where we are using the same notation as in Section 3.1.2. These sequences corre-
spond to the three branches of the recursive tiling sub-tree shown in Fig. 5.6. The
recursive localization of the base domains associated with these sequences is shown
in Fig. 5.7. The embeddings g 0 , g(6), g(9), g(11), g(13) are given, using the notation
of Sect. 3.1.2, by
g0 = T(λ −1 ,0) (−1)Sλ −1 ,
g(6) = S8+λ −3λ 2 ,
g(9) = RC 6 T(−1+λ ,1) S4−4λ +λ 2 ,
g(11) = C 6 T(0,−5+3λ ) S−9+5λ ,

g(13) = RC 11 T(1+λ −λ 2,8−λ −2λ 2) S13−4λ 2 .
Fig. 5.6 Sub-tree of the (0, k, 6∞ ) sectors.

Fig. 5.7 Embedding scheme for the (0, k, 6∞ ) sectors.
To explore the possibility of global expansivity in the sectors (k, 6, 6, . . .), we now
examine the return-time vectors for increasing level. The level-0 dressed domain is
a triangle equivalent to Δ [0] , with 4 subdomains and the return-time vector
T(0) = (2, 6, 3, 3) ≡ (2, 2, 3, 3) mod 4.
The level-1 vectors are obtained by right-multiplication by the incidence matrices

A(9), A(11), A(13):
T(0, 9) = T(0) · A(9) = (76, 226, 448, 214, 1078, 502, 952, 1654, 1978, 5056)
≡ (0, 2, 0, 2, 2, 2, 0, 2, 2, 0) mod 4,
T(0, 11) = T(0) · A(11) = (246, 810, 1764, 1074, 5250, 2466,
4158, 8034, 8634, 22062)
≡ (2, 2, 0, 2, 2, 2, 2, 2, 2, 2) mod 4,
T(0, 13) = T(0) · A(13) = (174, 576, 1260, 732, 3612, 1692,
2898, 5532, 6024, 15402)
≡ (2, 0, 0, 0, 0, 0, 2, 0, 0, 2) mod 4.
We note that already at the first level, a simplification has occurred: the entries
of the modulo-4 return-time vectors are all even. This is due to a peculiarity of
the incidence matrices A(9), A(11), A(13), namely that their third and fourth rows
are identical, combined with the fact that the third and fourth components of T(0)
also coincide. Even though these circumstances will not repeat themselves at higher
levels, the restriction to even entries cannot be undone, since multiplication of an
even integer by any integer is necessarily even. A similar “ratchet” effect will occur
if, at some level, the entries are all zeros.
For levels greater than 1, we calculate
T(0, 9, 6, 6, ..., 6) = T(0, 9) · A(6) · A(6) · · · A(6) ≡ (0, 2, 0, 2, 2, 2, 0, 2, 2, 0) mod 4,

T(0, 11, 6, 6, ..., 6) = T(0, 11) · A(6) · A(6) · · · A(6)
≡ (0, 0, 0, 0, 0, 0, 0, 0, 0, 0) mod 4,
T(0, 13, 6, 6, ..., 6) = T(0, 13) · A(6) · A(6) · · · A(6)
≡ (0, 2, 0, 2, 2, 2, 0, 2, 2, 0) mod 4,
We see that for L ≥ 2, the modulo-4 return-time vectors are independent of L, and
so we can employ (5.4) to obtain the complex global recursion matrices:
⎛ ⎞
1 0 0 0 0 0 1 0 0 0
⎜0
⎜ 0 0 0 0 0 −1 0 1 0⎟⎟
⎜ ⎟
⎜2 0 0 0 0 0 1 0 0 1⎟
⎜ ⎟
⎜0
⎜ 0 0 0 0 1 −1 0 0 0⎟⎟
⎜ ⎟
⎜0 0 0 0 0 4 −1 0 −3 0 ⎟
M(0, 9, 6, 6, ..., 6) = M(13, 6, 6, ..., 6) = ⎜ ⎟,
⎜0
⎜ 0 0 0 0 2 −1 0 −1 0 ⎟
⎟
⎜ ⎟
⎜1 0 0 0 0 0 1 0 0 0⎟
⎜ ⎟
⎜0
⎜ 0 0 0 0 6 −1 0 −5 0 ⎟
⎟
⎜0
⎝ 0 0 0 0 0 −1 0 1 0⎟⎠
1 0 0 0 0 0 1 0 0 0
M(0, 11, 6, 6, . . . , 6) = A(6) T
The matrix M(0, 9, 6, 6, ..., 6) departs slightly from the standard scenario in that
its dominant eigenvalue, ωW = 2 has a 2-dimensional eigenspace. This is not
an essential problem, however, and we find that the complex displacement vec-
tor d(0, 9, 6, 6, . . . , 6) behaves, for L → ∞, as 2 L times a unique eigenvector of
M(0, 9, 6, 6, ..., 6),
d(0, 9, 6, 6, . . . , 6) ∼ 2 L (1, −1, 2, −1, −1, −1, 1, −1, −1, 1).
Similarly,
3
d(0, 13, 6, 6, . . ., 6) ∼ 2 L × (−1, 1, −2, 5, 17, 9, −1, 25, 1, −1).
2
Since A(6)T has maximum eigenvalue

√
ωT = 8 + 2 7 = 13.2915 . . .,
the sectors (0, 9, 6, 6, . . .) and (0, 13, 6, 6, . . .) are globally expanding, and we can
expect asymptotic behavior with lattice distance increasing as t μ , with
log ωW log 2
μ= = = 0.267922 . . ..
log ωT log ωT
Since μ is smaller than the value 1/2 characterizing a random walk (normal diffu-
sion), we label the asymptotic behavior as sub-diffusive.
In the sector (0, 11, 6, 6, . . .), the displacement vectors are proportional asymptot-
ically to ωTL times an eigenvector of A(6) with all positive components. The lattice
distance increases proportional to the time, analogous to ballistic motion.
5.4.2 Numerical investigations
Now that we have a theoretical understanding of the power-law asymptotic behavior

of the aperiodic orbits in the sectors (0, k, 6, 6, . . .), it is worthwhile to see how the
concepts are realized in detail for specific initial conditions. We choose two admissi-
ble codes σA and σB in the sectors (0, 9, 6 ∞ ) and (0, 11, 6 ∞), respectively, calculating
the corresponding points in Ω using (3.5),
σA = (0, 1, 0)(9, 8, 0) (6, 8, 8) ∞, σB = (0, 1, 0)(11, 8, 0) (6, 8, 8) ∞,
uA = g0 g(9) (1 − C 10 S8+λ −3λ 2 )−1 ρ [3]8 (0)

1
= 864 + 391λ − 471λ 2, 347 − 237λ + 30λ 2 ,
71
uB = g0 g(11) (1 − C 10 S8+λ −3λ 2 )−1 ρ [3]8 (0)
1
= 279 + 624λ − 425λ 2, −2087 − 1078λ + 1241λ 2 ,
71
In our calculations, we treat λ as an abstract symbol, using the relation λ 3 =
λ 2 + 2λ − 1 to simplify polynomial expressions with exact rational coefficients. In
this way the dynamical map can be iterated arbitrarily many times without being
contaminated by round-off error. Floating point arithmetic enters only at the stage
of plotting the results.
Our first experiment was aimed at visualizing the local orbits. In Figs. 5.8 and
5.9 we see the result of iterating the local map K 10,000 times. The two local or-
bits are not qualitatively very different. Both show evidence of island-around-island
structures reminiscent of sticky orbits found in chaotic models. To see this more
clearly, we zoomed in to a small square of side 0.0013 the width of the fundamen-
tal cell, increasing the number of iterations to 10 7 to obtain high resolution. In Fig.
5.10 the island-around-island structure is quite evident. Not pictured are the islands
themselves; these are periodic heptagons of various sizes which in fact occupy 100
percent of the area, with the aperiodic orbits confined to a set of measure zero.
Now we want to see the W orbits of the same initial points on the plane. From
our analysis based on global expansivity, we expect the orbits to behave quite differ-
ently for long times: orbit A has scaling exponent 0.27, and hence should proceed
relatively slowly to infinity (much slower than a random walk, with exponent 1/2),
while orbit B, perhaps after a transient period, should move rapidly outward, with
distance from the origin increasing proportional to the time. We begin by calcu-
lating the first 10,000 points on each orbit (these are the same points whose local
coordinates we have already seen).
The resulting pictures, in Fig. 5.11, are dramatically different: the points of or-
bit A fill out a cloud of approximate radius 3 surrounding the origin, with little
discernable structure. Orbit B, on the other hand, has already become an “accel-
erator mode”, spiraling outward at an approximately constant rate in narrow strips
surrounding the x and y axes.
Fig. 5.8 Local orbit (104 iterations) of x(σA ).
Fig. 5.9 Local orbit (104 iterations) of x(σB ).

Fig. 5.10 Detail of the x(σB ) orbit, showing island-around-island structure. Actual islands are
periodic heptagons, not shown.
Fig. 5.11 Global orbits (104 iterations) of x(σA ) and x(σB ).
A clearer picture of what is going on is obtained by increasing the number of

iterations to 1 million, as shown in Fig. 5.12. Orbit A has continued to make its way
slowly outward, and some interesting geometric patterns have emerged. With a 0.27
power law, we would expect a 100-fold increase in t to produce an increase of linear
size of approximately 100 0.27 ≈ 3.5, which is consistent with the figure. Orbit B,
on the other hand, has continued its linear progress to infinity, with the maximum
distance increasing by a factor of approximately 100. Moreover, the departure from

the axes is now below the resolution of the graphics.
Finally, in Fig. 5.13 we plot the distance from the origin as a function of time, on
a log-log scale, for the two orbits. In the sub-diffusive case, we note the apparently
self-similar pattern of shifts from one level to another, including frequent returns
to the neighborhood of the origin (recurrences), with the maximum level slowly
increasing in accordance with the power law.
Fig. 5.12 Same orbits as in Fig. 5.11, with 106 iterations.
Fig. 5.13 Log-log plot of distance vs. number of iterations for the orbits of (a) x(σA ) and (b)
x(σB ).
5.4.3 A non-expansive sector
Not all invariant sectors of the π /7 kicked oscillator are globally expansive. It is
instructive to see how this can come about, even when the global recursion matrix
has an eigenvalue greater than 1 in magnitude. Our example has sector code i =
(0, 10, 1∞ ). As in the previous cases, the modulo-4 return-time arrays are built up by
repeated application of the incidence matrices:
τ (0) = (2, 2, 3, 3),

τ (0, 10) = τ (0) · A(10) mod 4 = (2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2),
τ (0, 10, 1) = τ (0, 10) · A(1) mod 4 = (0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0)
τ (0, 10, 1, 1) = τ (0, 10, 1) · A(1) mod 4 = (2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2),
τ (0, 10, 1, 1, 1) = τ (0, 10, 1, 1) · A(1) mod 4 = (0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0)
The global recursion matrix is
M = M(2) · M(1) ,
ν (1, j)−1 (l)
(l) τ p(1, j|t+1)
M j,k = ck,p(1, j) = ∑ δk,p(1, j|t) i , l = 1, 2,
t=0
where
τ (1) = (2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2), τ (2) = (0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0).
Diagonalization of M reveals eigenvalues 3, 1, −1, and 0, and so one might expect a

sub-diffusive asymptotic power law, with exponent
log3
μ= = 0.151270..., (5.10)
2 log α
√
where α = 19 + 4 22 is the maximum eigenvalue of A(1) T . On the other hand,
explicit calculation of the complex displacement vectors for increasing level gives
d(0) = (1 − i, −i, 2, 1 + i),

d(0, 10) = M(0, 10) · d(0) = (1 − 4i, 1 − 4i, 1 − 4i, . . ., 1 − 4i),
d(0, 10, 1) = M (1) · d(0, 10) = (0, 1 − 4i, 0, 0, 0, 0, 0, 0, 0, 0, 0),
d(0, 10, 1, 1) = M (2) · d(0, 10, 1) = (1 − 4i, 1 − 4i, 1 − 4i, . . ., 1 − 4i),
d(0, 10, 1, 1, 1) = M (1) · d(0, 10, 1, 1) = (0, 1 − 4i, 0, 0, 0, 0, 0, 0, 0, 0, 0)
and we see that there is no global expansivity. In fact, there is no contradiction: the
displacement vector (1 − 4i, . . . , 1 − 4i) is indeed an eigenvector of M, but corre-
sponds to the nonleading eigenvalue 1 instead of the dominant eigenvalue 3.
To see what happens to a typical orbit in this sector, we select the simplest ad-
missible code and calculate the corresponding point in Ω (this parallels exactly the
calculations of the previous examples, and so we omit the details:
σC = (0, 1, 0), (10, 7, 0), (1, 7, 0) ∞,

1
uC = (86 + 52λ − 47λ 2, −179 + 57λ + 26λ 2).
58
The first 100, 000 points of the local orbit are shown in Fig. 5.14(a), while the corre-
sponding points of the global orbit are shown in Fig 5.15(a). Further iteration does
not expand the extent of the latter at all.
Fig. 5.14 Local orbits in sectors (a) (0, 10, 1∞ ) and (b) (0, 12, 1∞ ).
Fig. 5.15 Global orbits in sectors (a) (0, 10, 1∞ ) and (b) (0, 12, 1∞ ).
In contrast to this non-expansive case is the sector (0, 12, 1 ∞ ), which has the
same modulo-4 return time arrays and global recursion matrix as (0, 10, 1 ∞ ). This
time, however, the projection of the displacement vector along the dominant eigen-
direction is non-zero, and so there is a global scaling factor ω W = 3 and a sub-
diffusive power law with exponent μ given by (5.10). The local and global orbits
are plotted in Figs. 5.14(b) and 5.15(b).
5.4.4 Generic behavior

In the preceding subsection, we have gotten a glimpse of the various possibilities
for power-law asymptotics in the countably many invariant sectors with eventually
periodic sector codes. Such behavior is not generic, of course, since the full set of
sectors is uncountable. In this final subsection, we consider the generic case, for
which the sector code is a random sequence of i L compatible with the constraint
n(iL+1 ) = h(iL ). Apart from the initial choice of i 0 (see Fig. 2.15), this is a ternary
process: given iL , the choice of i L+1 is among exactly 3 equally weighted possibili-
ties.
First of all, let us calculate the probability distribution for the various possible
modulo-4 return-time arrays τ (i L ), in the limit L → ∞. We suspect that, due to the
‘ratchet effect’ (irreversibility of the transition to (0, 0, . . . , 0)), the probabilities may
become skewed toward arrays containing only zeros. By explicit calculation, we find
that from L = 2 onward, only the following arrays are represented:
h(iL ) = 1
(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0), (0, 0, 2, 0, 0, 2, 0, 0, 0, 2, 2), (0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0),

(0, 2, 2, 0, 0, 2, 0, 0, 0, 2, 2), (2, 2, 0, 2, 2, 0, 2, 2, 2, 0, 0), (2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2),
h(i L ) = 2
(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0), (0, 0, 0, 0, 2, 0, 2, 0, 0, 0, 2, 0),

(0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0),
(0, 2, 0, 0, 2, 0, 2, 0, 0, 0, 2, 0), (2, 0, 2, 2, 0, 2, 0, 2, 0, 2, 0, 2),
(2, 0, 2, 2, 2, 2, 2, 2, 0, 2, 2, 2),
h(i L ) = 3
(0, 0, 0, 0, 0, 0, 0, 0, 0, 0), (0, 2, 0, 2, 2, 2, 0, 2, 2, 0, ),
Multiplying (from the right) each of the above by the three appropriate inci-
dence matrices (A(0), A(1), A(2) for h(i L ) = 1, A(3), A(4), A(5) for h(i L ) = 2, and
A(6), A(7), A(8) for h(IL ) = 3) we find that the above set of 14 arrays is mapped
into itself, with a transition matrix
⎛ ⎞
2 0 0 0 0 0 1 0 0 0 0 0 0 0
⎜0 1 0 1 0 0 0 1 0 0 0 0 0 0 ⎟
⎜ ⎟
⎜1 0 0 0 0 1 1 0 0 0 0 0 0 0 ⎟
⎜ ⎟
⎜0 0 0 1 1 0 0 1 0 0 0 0 0 0 ⎟
⎜ ⎟
⎜0 1 0 1 0 0 0 0 0 0 0 1 0 0 ⎟
⎜ ⎟
⎜0 0 2 0 0 0 0 0 0 0 1 0 0 0 ⎟
⎜ ⎟
⎜2 0 0 0 0 0 1 0 0 0 0 0 0 0 ⎟
R=⎜ ⎟
⎜ 0 1 0 1 0 0 0 1 0 0 0 0 0 0 ⎟.
⎜ ⎟
⎜1 1 0 0 0 0 0 0 1 0 0 0 0 0 ⎟
⎜ ⎟
⎜0 1 1 0 0 0 0 0 0 1 0 0 0 0 ⎟
⎜ ⎟
⎜0 0 1 0 1 0 0 0 0 1 0 0 0 0 ⎟
⎜ ⎟
⎜0 1 0 0 0 1 0 0 1 0 0 0 0 0 ⎟
⎜ ⎟
⎝2 0 0 0 0 0 0 0 0 0 0 0 1 0 ⎠
2 0 0 0 0 0 0 0 0 0 0 0 0 1
References 129
The level-L populations of the 14 species, n(L), is then given by
n(L) = (n1 (L), n2 (L), . . . , n14 (L)) = n(2) · RL−2,
where, by explicit calculation,
n(2) = (6, 0, 4, 0, 0, 2, 1, 0, 0, 0, 2, 0).
The L → ∞ behavior is obtained by diagonalizing the transpose of R. The largest

eigenvalue is 3, with eigenvector (2, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0) corresponding to
only arrays containing zeros. The convergence rate is determined by the second-
largest eigenvalue, 2.95263 . . .: the probability for τ (i L ) = 0 converges to zero pro-
portional to (2.95263/3) L = 0.98421 L. The constant of proportionality is found to
be 0.109446 . . ..
We conclude that for a randomly selected invariant sector, the modulo-4 return-
time array has only zeros from some level onward, with probability one. Scaling
is generic, with the power-law exponent equal to unity (‘ballistic’ case). Moreover,
the dense set of points with eventually periodic codes is drastically affected by the
same mechanism: of all the points whose codes become periodic starting at level
10 or higher, at least 90% have eventually all-zero modulo-4 return-time arrays, and
hence have orbits with μ = 1 power laws.
References
Chapter 6
Transport
In our discussions so far concerning the asymptotic long-time behavior of pseu-

dochaotic kicked oscillators, we have concentrated on individual orbits, especially
those within the web of discontinuity-avoiding aperiodic orbits. In the present chap-
ter we turn our attention to the statistical properties of various classes of orbits for
asymptotically long times. Once again, the recursive tiling property serves as a key
organizing principle.
In defining a coarse-grained probability distribution P((m, n),t), it is important
to take account of the π /2 rotation incorporated in the map W . The orbits of interest
actually spiral out to infinity rather than following a succession of small steps on
the lattice. Obviously, our probability distribution refers to lattice displacements
(m2 − m1 τ , n2 − n1τ ) : m1 , m2 , n1 , n2 ∈ Z and time differences t 2 − t1 , t1 ,t2 ∈ Z, with
the latter always assumed to be multiples of 4 (we are not interested in the position-
dependent displacements due to the π /2 rotations). Due to the invariance of the
4th-iterate map W 4 (see (1.6)), the probability distribution will then be independent
of the choice of initial lattice point, allowing us to assume, for convenience, that all
relevant orbits are initiated in the fundamental domain Ω .
With these considerations in mind, we adopt the following definition.
Definition 6.1. We define P((m, n),t) as the probability that a W orbit initiated at
time 0 at a randomly selected point in Ω will be found in the domain Ω + (mτ , nτ )
after t iterations, t an integer multiple of 4.
We note that P((m, n),t) is always positive on a finite number of lattice sites
(vanishing elsewhere), and satisfies the normalization condition, for all t,
∞
∑ P((m, n),t) = 1.
m,n=−∞
If χ (r), r = (m, n) ∈ Z2 is any real-valued lattice function, we define the expectation

value of χ (r) at time t,
χ (r) t = ∑ P(r,t)χ (r). (6.1)
r
132 6 Transport
Moments of the probability √ distribution correspond to χ ((m, n)) being a monomial

in m, n, |m|, |n|, |m| + |n|, or m2 + n2 .
In specific examples, the words “randomly selected” in the definition of P((m,
n),t) need to be interpreted in terms of some measure, which, as pointed out in
the Introduction, is not necessarily Lebesgue measure on the plane. In fact, some
of the most interesting long-time behavior is found in subsets of the exceptional
set, which for all of the models of interest is of Lebesgue measure zero. For these
subsets, Hausdorff measure, which we discussed in some detail in Chap. 4, serves
as the basis for our probability calculations.
6.1 Probability calculation using recursive tiling
The definition of the probability distribution P(r,t) given above suggests a direct
method of calculation, namely, with the help of a random number generator, to se-
lect large numbers of initial conditions in Ω and then apply the map W iteratively
on each one, recording the lattice coordinates of each orbit at a succession of time
intervals. The number of visits to each lattice site could then be accumulated and
eventually normalized to give the desired probability. Such a direct approach is un-
fortunately highly inefficient, due to the extreme degree of nonuniformity of the lo-
cal phase space. Most of the computational effort would end up iterating relatively
uninteresting low-period orbits which never get very far from the origin.
A more elegant strategy would make use of the time-t dynamical partition of
Ω . The latter consists of a set of disjoint convex polygons, with distinct t-iteration
itineraries, whose union is Ω . By construction, a polygon D in the partition can be
subjected to t iterations of W without landing on a discontinuity line, ending up in
the domain Ω + (m, n)τ . We then have an exact calculation of P((m, n),t), namely
A (D)
P((m, n),t) = ∑ A (Ω )
, (6.2)
D∈D (m,n)
t
where
Dt (m, n) = {D : W t (D) ⊂ Ω + (m, n)τ }.
While using the dynamical partition removes problems of convergence (for ev-
ery t, the calculation is exact) and eliminates overemphasis on parts of phase space
where branching has ceased, it still does not take advantage of the exponential
shrinking of the time scale which renormalizability offers. For that we turn once
again to the hierarchy of recursive tilings.
For simplicity, let us restrict ourselves to a model with quadratic irrational pa-
rameter and a single scaling sequence. Then, at level L, the square Ω is partitioned
into the orbits of the periodic domains Π k (L ), L < L, together with those of the
scaling subdomains D j (L), j = 0, 1, . . . , J. As usual, D(L) ⊂ D j0 (L − 1) for L ≥ 1
and some j0 , and D(0) ⊂ Ω k0 for some k0 . We call the partition P(L).
6.1 Probability calculation using recursive tiling 133
Let t∗ (L) be the largest number of iterations such that for all D ∈ P(L), W t (D)
does not intersect a discontinuity line for t = 1, 2, . . . ,t ∗ (L). Then we can use for-
mula (6.2) to calculate P(r,t) exactly for any t ≤ t ∗ (L). Our hope is, of course, that
for large L, the non-branching interval t ∗ (L) will scale as ωTL . This will give an ex-
ponential shrinking of the time scale for calculations of P(r,t), just as we had for
individual orbits in Chapter 5.
At first sight, the above strategy appears to be futile. Consider any one of the
level-L local return orbits K s (D j (L)), s = 0, 1, . . . , T j (L) . The first T j (L) members
of this sequence are members of P(L), but the last is not and may in fact straddle
a boundary between some D j1 (L) and D j2 (L). Since it takes only one iteration of
K to pass from K Tj (L)−1 (D j (L)) to K Tj (L) (D j (L)), it would seem that we are stuck
with t∗ (L) = 1! Fortunately, there is a loophole in this reasoning. The fact that the
adjacent domains D j1 (L) and D j2 (L) initiate return orbits with distinct level-(L − 1)
itineraries does not imply that they will separate immediately under application of
K. In fact, since both are subdomains of D(L), hence of D j0 (L − 1), they will remain
glued together for at least T j0 (L − 1) iterations. In fact, once they get mapped around
the full return orbit of D j0 (L − 1), they land in D j0 (L − 2), and so the free ride
continues for at least another T j0 (L − 2) iterations. Eventually, the forward images
of D j1 (L) and D j2 (L), still glued together, end up in Ω k0 , and there is no further
guarantee against landing on a discontinuity line of K. We conclude that
L−1
∑ T j0 (L ).
def
t∗ (L) ≥ T∗ (L) = (6.3)

L =0
The desired large-L scaling behavior, i.e.
T∗ (L) ∼ ωTL , L → ∞. (6.4)
is immediate.
Now we are in a position to write an explicit formula for P(r,t) and its moments
χ (r) t . Given t, we choose the partition P(L ∗ (t)), where L∗ (t) is given by
L∗ (t) = min{L : t ≤ T∗ (L)}. (6.5)
For any function χ on Z 2 and sequence of integer pairs
Z = (z0 , z1 , . . . , z|Z|−1 ), zi ∈ Z2 ,
we define, for nonnegative integer t ≤ T ,

|Z|−T −1
χ Z
t = ∑ χ (zi+t − zi ). (6.6)
i=0
Denote by ZD (L, j) (resp. ZΠ (L, j))) the lattice coordinates of the W orbit of the
centroid of D j (L) (resp. Π j (L)), extended to a total of T j (L) + T∗ (L) (resp. t j (L) +
T∗ (L)) iterations. If A (D) is the area of a domain D, we define
134 6 Transport
A (Π j (L )) ZΠ (L , j ) A (D j (L∗ (t))) ZD (L∗ (t), j)

χ t = ∑

∑ A (Ω )
χ t + ∑ A (Ω )
χ t .
−1<L <L∗ (t) j j
(6.7)
Here level -1 is included in the first sum to include all periodic orbits which do not
intersect the scaling domain D(0). The same formula can be used to calculate the
probability distribution P(r,t), since, by definition (6.1),
P((m, n),t) = δm δn t , m, n,t ∈ Z, t ≥ 0, (6.8)
where
1 n = n ,
δn (n ) =
def
0 n= n .
6.2 Ballistic transport in Model I
Perhaps the simplest of our examples √ with nontrivial global expansion is Model I,
with golden mean parameter λ = 12 ( 5 + 1). As a warm-up calculation to illustrate
the method, let us calculate the second moment m 2 + n2 t for t = 8. As required,
t is a multiple of 4. For this value of t, we need to use the tiling of (at least) level
L∗ (10) = 1 (see (6.5), noting that T∗ (1) = T2 (0) = 10). The tiling, shown in Fig.
6.1, partitions Ω into three K-orbits of periodic pentagons, namely Π 4 (0), Π5 (0),
Fig. 6.1 Level-1 tiling of Ω . The periodic pentagons are colored white, while the domains
D0 (1), D2 (1), and D4 (1) are colored gray, dark gray, and light gray, respectively.
6.2 Ballistic transport in Model I 135
and Π0 (−1), complemented by the three return orbits of the three scaling subdo-
mains D0 (1), D2 (1), and D4 (1). Note that we can safely omit orbits of domains of
dimension less than 2, since these do not contribute to the probability distribution.
For the pentagons, the periods are 24, 6, and 4, respectively, while for the subdo-
mains of D(1), the return times are, respectively, 16, 40, and 88. The reader can
easily locate all of these orbits on the figure.
As a prelude to the moment calculation, we calculate all six of the corresponding
global orbits on Z 2 , remembering to extend each one by T∗ (1) = 10 iterations. To
simplify the calculation, we map pointwise the centroid of each convex polygon,
since all points in the polygon have the same lattice itinerary. We get, for example,:
ZΠ (0, 5) = {(0, 0), (0, 1), (1, −1), (−1, −1), (−1, 1), (1, 1), (1, −2), (−2, 0), (0, 1),
(1, 0), (0, −1), (−1, 0), (0, 0), (0, 1), (1, −1), (−1, −1), (−1, 1)}.
For the second moment at t = 8, T∗ (1) = 10, (6.6) becomes

Tj (L)−1
ZD (L, j)
χ 8 = ∑ ((mi+8 − mi )2 + (ni+8 − ni )2 ),
i=0
Z (L, j)
with an analogous expression for χ 8 Π . Note that the number of terms in
the sum is precisely the return time (or period) of the orbit, since we get a single
contribution from each tile in the partition of Ω . The six double-brackets are easily
calculated:
ZΠ (0,4) ZΠ (0,5) ZΠ (−1,7)
χ 8 = 68, χ 8 = 8, χ 8 = 16,
ZD (1,0) ZD (1,2) ZD (1,4)
χ 8 = 36, χ 8 = 104, χ 8 = 240.
To proceed, we need the weights in the probability sum, given by the areas
A (Π4 (0)) = −(199 + 123τ )/2, A (Π 5 (0)) = −(11 + 7τ )/2,
A (Π7 (−1)) = −9τ − 29/2, A (D 0 (1)) = −13τ − 21,

A (D2 (1)) = −17τ − 55/2, A (D 4 (1)) = (89 + 55τ )/2,
as well as
A (Ω )−1 = τ −2 = 2 + τ .
Putting everything together, we get, from (6.7),
x2 + y2 8 = 54 + 32τ = 2.22291 . . .
The probability distribution at time t = 8 can be determined from (6.8). The result
is displayed in Fig. 6.2, with the shorthand notation
a = 233/2 + 72τ , b = 161 + 199τ /2, c = 89/2 + 55τ /2, d = 34 + 21τ ,

136 6 Transport
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0
0 0
0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Fig. 6.2 Probability distribution after 8 iterations of W . The symbols a–h and m–q stand for the
values listed in (6.9). The probability is zero over the entire gray region.
e = −571/2 − 353τ /2, f = −359 − 691τ /2, g = −1063/2 − 657τ /2, (6.9)
h = −2807/2 − 1735τ /2, m = −271/2 − 335τ /2, n = −453/2 − 140τ ,
p = −21/2 − 13τ /2, q = −521 − 322τ .
We now are ready for a more ambitious application of formula (6.7), namely
the calculation of the moments m 2 + n2 t and m4 t of P((m, n),t) over the entire
range of t from 0 to 218450 = T∗ (8), using the level-8 partition of Ω . The latter
is comprised of the 20 return orbits listed, together with their respective areas and
return times, in Table 6.1. For the sake of efficiency, the first step of our procedure
is to calculate and store the lattice coordinates of the various extended return orbits.
The moments are then calculated systematically using (6.7). If an exact result for a
specific time interval is desired, it is essential to employ arbitrary-precision integer
arithmetic and computer algebra. For example, as a test case, we calculated the
second moment for t = 218448, the largest multiple of 4 less than or equal to T ∗ (8),
obtaining
m2 + n2 218448 = −2277570288992466304278 − 1407615850364419027930 τ
Log-log plots for 500 t values, 932 ≤ t ≤ 218448, are displayed in Fig. 6.3 on a
logarithmic time scale. The chosen log 10 t values are evenly spaced, up to round-off.
ν
The t ν increase of (m2 + n2 ) 2 t comes as no surprise, given that all but one of the
20 lattice orbits expands outward ballistically. The exception is the tightly bound
orbit of Π5 (0), which has negligible effect on the moments.
6.2 Ballistic transport in Model I 137
Table 6.1 Areas and return times for the 20 domains participating in the calculation of moments
for t ≤ 218448 .
Domain Area Return Time

D0 (8) −14930352 − 9227465τ 218456
D2 (8) (−39088169 − 24157817τ )/2 655360
D4 (8) 31622993 + 39088169τ /2 1529168
Π4 (7) 70711162 − 87403803τ /2 436904
Π5 (7) −70711162 − 87403803τ /2 54616
Π4 (6) (−20633239 − 12752043τ )/2 109224
Π5 (6) (−1149851 − 710647τ )/2 13656
Π4 (5) −3010349/2 − 930249τ 27304
Π5 (5) −167761/2 − 51841τ 3416
Π4 (4) −219602 − 271443τ /2 6824
Π5 (4) −219602 − 271443τ /2 856
Π4 (3) (−64079 − 39603τ )/2 1704
Π5 (3) (−3571 − 2207τ )/2 216
Π4 (2) −9349/2 − 2889τ 424
Π5 (2) −521/2 − 161τ 56
Π4 (1) −682 − 843τ /2 104
Π5 (1) −38 − 47τ /2 16
Π4 (0) (−199 − 123τ )/2 24
Π5 (0) (−11 − 7τ )/2 6
Π7 (0) −29/2 − 9τ 4
Fig. 6.3 Log-log plots of m2 + n2 t and m4 t .

138 6 Transport
6.3 Subdiffusive transport in Model II
We now turn to a statistical description of the long-time behavior on the pseu-

dochaotic web (residual set) of a globally expansive model √ with quadratic irrational
parameter. Our choice is Model II, with λ = −τ −1 = 3, where we have already
found asymptotic diffusive and subdiffusive asymptotic behavior for orbits in the
parts of the web associated with the scaling sequences A (limit point (τ , τ )) and B
(limit point (41 + 72τ , 41 + 72τ )), respectively. In the present section we will fo-
cus on the latter, making use of Hausdorff measure to assign weights to the orbits
contributing to the probability distribution P(r,t).
Apart from the restricted measure, concentrated on the pseudochaotic web of
Case B, the strategy for calculating the expectation values χ t is the same as that
used in the preceding sections. In fact, it is simpler: only the second summation
(over level-L ∗ (t) tiles) is relevant, since the periodic domains do not contribute to
the measure:
Z (L (t), j)
χ t = ∑ μH (D j (L∗ (t)))χ t D ∗ , (6.10)
j
where μH is normalized Hausdorff measure (concentrated on the Case B web, with

dimension d H = − log5/ log ω ). According to Sec. 4.4, the measure assigned to the
level-zero atoms is given by
3 20
μH (D0 (0)) = μH (D10 (0)) = , μH (D3 (0)) = ,
c c
15 19
μH (D5 (0)) = μH (D6 (0)) = , μH (D8 (0)) = ,
c c
c = 5298000.
A serious practical problem which we must face in the present example is the fact
that the return times are extremely large, even for the lowest levels of the scaling se-
quence. Fortunately we can significantly improve on the mapping method of the
previous example by exploiting the fact that the lattice itineraries are entirely deter-
mined by the symbolic dynamics, i.e. Vershik updating of code sequences, with the
least significant code symbol at each step determining the current lattice displace-
ment. Only elementary operations with integers are involved (no symbolic manipu-
lation of algebraic numbers), and so the calculations can be performed rapidly using
a short program in a standard compiled computer language.
Let us carry out the above program in a numerical experiment which probes
several levels of the scaling hierarchy. We calculate χ t for
χ = rk , k = 1, 2, r = |x| + |y|,
for a sample of time intervals t k , k = 1, . . . , N, all multiples of 4, approximately

equally spaced on a logarithmic scale, with
1052 ≤ tk ≤ tmax = 72000000.

6.3 Subdiffusive transport in Model II 139
To do this, we calculate symbolically the orbits of D j (5), j = 0, 3, 5, 6, 8, 10 for

T j (5) + tmax iterations, with
{T j (5) : j = 0, 3, 5, 6, 8, 10} =
{59107675, 236507588, 295690483, 295690483, 118215350, 177323025},
using a codometer with 7 registers,

⎛⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎞
0 1 1 1 1 1 1
⎝⎝ j0 ⎠⎝ j1 ⎠⎝ j2 ⎠⎝ j3 ⎠⎝ j4 ⎠⎝ j5 ⎠⎝ j6 ⎠⎠ .
t0 t1 t2 t3 t4 t5 t6
The paths (p(1, j,t) : t = 0, . . . , ν j ) assigned to all but the zeroth register are
(6), (6, 3, 8, 3, 5), (6, 3, 8, 10, 8, 3, 5), (6, 3, 8, 0, 8, 8, 3, 5), (6, 5), (6, 5, 5).
The lengths of the paths (p(0, j,t) : t = 0, . . . , ν j ) assigned to register 0 are given by
ν0 (0) = 19459, ν 0 (3) = 76580, ν 0 (5) = 93115,
ν0 (6) = 93115, ν 0 (8) = 38918, ν 0 (10) = 58377.

For the sake of efficiency, the corresponding paths (sequences of 0’s,1’s and 2’s)
were worked out and stored in preparation for the calculation of moments.
According to the prescription (6.10), the exact moment calculation involves mea-
surements on only six orbits, namely those of D j (5), j = 0, 3, 5, 6, 8, 10. In the initial
codometer settings for these, we set j 5 = j,t5 = 0, which requires ( j k ,tk ) = (6, 0)
for k = 0, 1, 2, 3, 4. The last register is then set to a value compatible with j 5 = j, in
such a way that the full T j (5) + tmax iterations can be accommodated.
Before proceding to the calculation of moments, let us briefly examine the six
orbits on which all the measurements are to be made. Figure 6.4 provides a low-
resolution (2000 × 2000 pixels) view of the footprints of these orbits, revealing the
asymptotic large-scale self-similarity. Log-log plots of the expectation values χ t
for χ = |m| + |n| and (|m| + |n|) 2 are shown in Fig.6.5. The average slopes are con-
sistent with the power laws found in Sect. 5.3, namely
(|m| + |n|)k t ∼ t ks Gk (t)
where
log 2
s= = 0.430676558 . . .
log 5
and, as we see in Fig. 6.6, the form factors G k are apparently bounded period-1
functions of log 5 t.
Using essentially the same program for χ = δ m δn , we have also calculated the
probability distribution P((m, n),t) for a single value of t, namely 72000000. The
result is displayed with 2000× 2000 resolution in Fig. 6.7. Here we have used colors
to indicate the magnitude of the probability on a logarithmic scale.
140 6 Transport
Fig. 6.4 Footprints of the 6 orbits contributing to the calculation of moments.
Fig. 6.5 Log-log plots of the expectation values χ t for χ = |m| + |n| and (|m| + |n|)2 .
6.3 Subdiffusive transport in Model II 141
Fig. 6.6 Form factors of the expectation values χ t for χ = |m| + |n| and (|m| + |n|)2. They appear
to be tending toward log-periodicity.
Fig. 6.7 Probability distribution P((m, n),t), multiplied by the normalization factor c = 5298000,
for web B, −1000 ≤ m, n < 1000,t = 72000000. (See the color figure at the end of the book.)
142 6 Transport
6.4 Diffusive transport in Model II
Following the same pattern as in the preceding section, we now√study the statistical
long-time behavior of the other pseudochaotic √ web for √λ = 3, namely the one
whose scaling sequences tend to (τ , τ ) = (−1/ 3, −1/ 3). Once again, formula
(6.10) is applicable, this time with Hausdorff dimension d H = − log4/ log ω ≈ 1.05,
and level-0 measures
1
μH (D7 (0)) = μH (D10 (0)) = μH (D21 (0)) = μH (D22 (0)) = ,
161280
and μH (D j (0)) = 0 for j = 7, 10, 21, 22.
Our numerical calculations of the probability distribution and its moments paral-
lels that of the preceding section. Again we probe the recursive tiling hierarchy up
to level 5, but because of the slower growth relate of the scaling sequence, the re-
turn times T j (5) and critical time t∗ (5) are smaller. Our sample of 401 time intervals
extends from t = 524 through t max = 9000000. Again we generate the lattice orbits
via a 7-register codometer, with paths p(1, j) for nonzero registers given by
p(1, 7) = (22, 10, 7),

p(1, 10) = (22, 10, 10, 21, 7),
p(1, 21) = (22, 10, 21, 21, 7),
p(1, 22) = (22, 21, 7).
Fig. 6.8 Probability distribution P((m, n),t), multiplied by the normalization factor c = 161, 280,
for web A, −550 ≤ m, n < 550,t = 9000000. (See the color figure at the end of the book.)
6.5 Superdiffusive transport in Model III 143
The paths used by the zeroth register, of lengths
ν7 (0) = ν22 (0) = 26933, ν 10 (0) = ν21 (0) = 53707,
are calculated by iterating the local map and storing the resulting code sequences in
advance of the moment calculations.
Using (6.10), we have calculated the expectation values (|m| + |n|) k t , k = 1, 2,
as well as the P(r,t) for t = 9000000. This involves symbolic updating over the
return orbits of D j (5), j = 7, 10, 21, 22, with return times
T7 (5) = T22 (5) = 27525173, T10 (5) = T21 (5) = 55050187.
The results are displayed in Figs. 6.8 and 6.9.
Fig. 6.9 Form factors of the expectation values χ t for χ = |m| + |n| and (|m| + |n|)2 in web A,
plotted versus log4 t. They appear to be tending toward periodicity.
6.5 Superdiffusive transport in Model III
As we see in Table 5.1, the long-time asymptotics of Model III is globally expansive,
but with a dramatic difference between the aperiodic orbits of the residual set and
those residing on the x = 0 discontinuity line. Among the latter, one finds (Lowen-
stein et al., 2005) asymptotic superdiffusive behavior
√ for orbits which pass through
a Cantor set C of dimension − log5/ log(3 − 2 2). In the present section we study
144 6 Transport
the statistical properties of these orbits, including numerical calculations of several

moments of a suitable probability distribution.
Note that we are dealing here with a set of measure zero, not only with re-
spect to 2-dimensional Lebesgue measure on the plane, but also with respect to
one-dimensional Lebesgue measure on the discontinuity set. Not unexpectedly, we
will use Hausdorff measure on C , as constructed in Sect. 4.5, in order to formulate
an acceptable probability measure for calculating expectation values χ t .
Suppose we are interested in calculating χ t for a given t. Following Sect. 6.1,
we calculate L = L∗ (t) from (6.5) and partition C , apart from boundary points,
into subsets covered by forward images of the one-dimensional atoms D j (L), j =
1, 3, 5, 7. A recursive prescription for this covering is given in Sect. 2.5.2 and sum-
marized in Sect. A.3.7 of the Appendix. Specifically, the covering tiles are precisely
those encountered along the respective level-L return orbits of the 4 atoms. The
number N j (L) of each tile type was found in Sect. 2.5.2 to be
N1 (L) = N7 (L) = 6 · 5L, N3 (L) = N5 (L) = 9 · 5L ,
and their Hausdorff measures, from Sect. 4.5 , are

1 1
μH (D1 (L)) = μH (D5 (L)) = , μH (D3 (L)) = μH (D7 (L)) = .
20 60
Suppose we have calculated all 6 × 5 L+1 baseline segments Dα covering C at
level L = L∗ (t), choosing from each a representative point u k with algebraic number
coordinates. It is important to note that the point need not be an element of C ,
since all orbits initiated in the covering segment have the same lattice itinerary for
times no longer than t. For simplicity, we could choose u k to be the midpoint of D k .
Denoting by [u] the integer lattice coordinates of the planar point u, we have for the
χ moment for time interval t the formula
6×5L+1
χ t = ∑ μH (Dα )χ ([W t uα ]). (6.11)
α =1
Since every Dα is isometric to one of the D j (L), j = 1, 3, 5, 7, we can rewrite

(6.11) as
χ t = ∑ μH (D j (L)) ∑ χ ([W t uα ]). (6.12)
j=1,3,5,7 Dα ∼
=D j (L)
Because of lattice translation invariance of the dynamics, we can replace the

second summation in (6.12) by an expression which involves operations on a single
orbit originating in the base tile D j (L):
Tj (L)
χ t = ∑ μH (D j (L)) ∑ χ ([W k+t u j ] − [W k u j ]). (6.13)
j=1,3,5,7 k=0
W k u j ∈B
6.5 Superdiffusive transport in Model III 145
As in the example of Sect. 6.3, the exact iteration of the planar map W using
symbolic manipulation of algebraic numbers can be circumvented by making use of
the fact that the motion on the integer lattice is completely driven by the symbolic
dynamics. We have adopted this strategy in the following numerical experiment.
Our goal is to calculate χ t for
χ = rk , k = 1, 2, 4, 6, r = |x| + |y|,
for a sample of time intervals t k , k = 1, . . . , N, all multiples of 4, approximately

equally spaced on a logarithmic scale, with
1000 ≤ tk ≤ tmax = 16000000.
To do this, we calculate symbolically the orbits of D j (7), j = 1, 3, 5, 7 for T j (7) +

tmax iterations, with
{T j (7) : j = 1, 3, 5, 7} = 171874384, 315207204, 358253925, 214921105,
using a telescoped codometer with 7 registers,

⎛⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎞
i2 i3 i4 i5 i6 i7 i8
⎝⎝ j2 ⎠⎝ j3 ⎠⎝ j4 ⎠⎝ j5 ⎠⎝ j6 ⎠⎝ j7 ⎠⎝ j8 ⎠⎠ .
t2 t3 t4 t5 t6 t7 t8
Here the indices iL take on the successive values 0, 1, 2, 1, 2, 1, 2. The first register is
a telescoping of levels −1, 0, 1, and 2, with 0 ≤ t 2 < T j2 (2), in contrast to the other
registers where, as usual, 0 ≤ t L < ν jL .
The instantaneous state of the system consists of a pair (V, r) where V is a
codometer state and r = (m, n) is a point on the lattice Z 2 . Under the dynamical
map W ,
(V, r) → (V , r ),
where V is obtained from V using the rules of Vershik updating (see Sect. 3.2), and
r = (m , n ) = (n, −m) + d( p(0, j2(V ),t2 (V )) ).

Here the path function gives the current atom index (corresponding to one of the 12
generating domains Ω j ), and d( j) is the corresponding lattice displacement, namely
d(1) = d(3) = d(5) = d(6) = d(7) = d(8) = (0, 0);

d(2) = d(4) = (0, −1);
d(0) = d(10) = d(9) = d(11) = (0, 1);

For each of the intimal domains D j (7), we start with an initial state
⎛ ⎛⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎞ ⎞
0 1 2 1 2 1 2
ψ = ⎝ ⎝⎝ 8 ⎠⎝ 8 ⎠⎝ 8 ⎠⎝ 8 ⎠⎝ 8 ⎠⎝ j ⎠⎝ j8 ⎠⎠ , (0, 0)⎠
0 0 0 0 0 0 t8
146 6 Transport
with j8 and t8 chosen (non-uniquely) with p(2, j 8 ,t8 ) = j. As a prelude to the main
calculation, we prepare a list of N + 1 initial states
Ψ = {W tk ψ : k = 0, . . . , N}, t0 = 0,
with precisely the time intervals needed for our moment calculations. The idea is to
advance all of these states simultaneously, step by step,
Ψ → W s Ψ = {W tk +s ψ : k = 0, . . . , N}, s = 0, . . . , T j (7) − 1,
computing the contributions to the right-hand side of (6.13) whenever W s ψ assumes

the form (V, (m, n)) with
⎛ ⎞
0
V = ⎝ j2 ⎠ , . . . , p(0, j2 ,t2 ) ∈ {8, 10}.
t2
As discussed above, we expect there to be 468750 such events for j = 1, 7 and

703125 of them for j = 3, 5.
The results of our moment calculations for 1000 ≤ t ≤ 16000000 are displayed
in Figs. 6.10 and 6.11. The first figure shows a log-log plot of each of the moments
versus the time interval t. The average slope for the kth moment is apparently con-
sistent with power law t kν with exponent
ν = log5/ log9
Fig. 6.10 Log-log plots of the expectation values (|m| + |n|)k t , for k = 1, 2, 4, 6 and 36588 ≤ t ≤
16000000. The moments have been calculated exactly for 401 time intervals approximately equally
spaced on the logt axis. In the plot, these points have been connected by straight line segments.
6.5 Superdiffusive transport in Model III
Fig. 6.11 Form factors t −kν (|m| + |n|)k t , for (a) k = 1, (b) k = 2, (c) k = 4, (d) k = 6, over the range 36588 ≤ t ≤ 16000000.
147
148 6 Transport
found in Sect. 5.3, but there is clearly a log-periodic oscillation modulating the
power law. In Fig. 6.11, we have divided out a factor t ν to obtain a clear picture of
the form factors with evident asymptotic unit periodicity with respect to log 9 t.
A variant of the above computational experiment, which exploits the fact that
the orbits are confined to a narrow strip containing the negative n axis, allows us to
probe the evolving probability distribution in great detail. Using the same algorithm
as before, with the same set of N = 401 time intervals, we have calculated
δn t = ∑ P((m, n),t).

m
plotting the resulting function of n for each of the selected t values. In each plot,
we choose the scale of the n axis to be proportional to t ν to facilitate the interpreta-
tion of the log-periodic structures found above for the kth moments. In Fig. 6.12, we
show δn t for t = 250000. The probability distribution is characterized by a number
Fig. 6.12 Probability distribution P(n,t) in the vertical lattice coordinate n, for t = 250000. The
normalization factor is c = 60 × 55 = 187500.
of narrow peaks which change their amplitudes and positions as t increases. Each
prominent peak corresponds to significant trapping of the orbit within a small inter-
val of the axis, sometimes for thousands of iterations. This effect is made evident
in Fig. 6.13, where n has been plotted versus t for an orbit located in the region of
peak probability at time t = 250000. A piece of this orbit leading up to trapping is
shown, with successive points connected by line segments, in Fig. 6.14.
6.6 Discussion 149
Fig. 6.13 Plot of n versus t for an orbit located in the highest peak of P(n,t) in Fig. 6.12.
Fig. 6.14 Plot of the second half of the orbit of Fig. 6.13, rotated by 90◦ .
6.6 Discussion
Let us try to summarize what we have learned thus far about power-law asymptotic
behavior in our class of kicked oscillator models. First of all, we note that power
laws are associated with dynamically self-similar return map structures of the local
map, namely scaling sequences of dressed domains. This means that in a specific
model, there may multiple power laws, (at most) one associated with each inde-
pendent scaling sequence. Whether a particular scaling sequence corresponds to a
nontrivial power law is a rather delicate question, whose answer depends on the
synchronization, or lack thereof, of the global 4-fold rotation number with the local
return map periodicities. If there is such synchronization, we are able to construct
a recursion matrix whose largest eigenvalue, ω W , together with the temporal scale
factor ωT , determine the exponent of power law expansivity. The local geometric
scale factor ωK plays no role here.
In Chap. 5, we saw, both analytically and numerically, how global expansivity
leads to an asymptotic power law for individual orbits, in the form of the exis-
tence of a lim-sup for the distance divided by t μ , μ = log ωW / log ωT . In Chap.
6, these results were supplemented by high-precision numerical calculations of the
150 6 Transport
moments of a coarse-grained probability distribution. The numerical experiments

provide strong evidence for the modulated power law
rk t = t k μ (Gk (t) + O(t −ε ))

√
where r is either |m| + |n| or m2 + n2, ε > 0, and the form factor G k is an even-
tually periodic function of logt with eventual period log ω T . To prove this result
analytically remains a challenging goal for future investigations.
References
Chapter 7
Hamiltonian Round-Off
In this final chapter, we shift our attention to a variation on the theme of a resonantly
kicked one-dimensional harmonic oscillator, namely a model of an invertible, area-
preserving round-off map (Vivaldi, 1994; Lowenstein et al., 1997; Lowenstein and
Vivaldi, 1998, 2000; Lowenstein and Liu, 2003; Akiyama et al., 2008). Here the
oscillator is represented stroboscopically by application of the generalized rotation
Ψ (x) = C · x,
where
λ −1
C= ,
1 0
and λ = 2 cos(2πρ ) is a quadratic irrational and the rotation number ρ is assumed
to be rational. The corresponding round-off map is defined, for integer x, y, to be

x λ x − y
Φ = (7.1)
y x
where · is the floor function. It is an easy exercise to verify that Φ is both invertible
and the restriction to Z 2 of an area-preserving map of the plane, and so it is natural
to use the terminology Hamiltonian round-off. For a more detailed introduction to
the subject, the reader is referred to the articles cited above.
7.1 Vector field
The generalized 2π p/q rotation Ψ returns to the initial point after q iterations, while
the round-off map Φ ends up displaced by a vector
d(x) = Φ q (x) − x,
152 7 Hamiltonian Round-Off
the round-off error. A simple example with non-zero round-off error is illustrated
in Fig. 7.1. In Fig. 7.2 we show the possible displacement vectors for the round-off
models with ρ = 1/5 and ρ = 5/12, respectively.
One can interpret d(x) as a vector field on Z 2 , assigning a lattice displacement
to each point x. Starting at a given x and iteratively applying Φ q (resp. Φ −q ), one
obtains the forward (resp. backward) orbit through x. Thanks to the invertibility of
Φ , the full orbits are disjoint. By drawing the vector field displacements for all the
lattice points in a given region as line segments, one obtains a graphical representa-
tion of the orbits intersecting that region. Figs. 7.3 and 7.4 display such vector field
diagrams for rotation numbers 1/5 and 5/12, respectively.
Obviously, our round-off model is closely related to the kicked oscillators of the
preceding chapters with quadratic irrational parameters. One way of thinking about
Fig. 7.1 Five iterations of (a) Ψ and (b) Φ for p/q = 1/5.
Fig. 7.2 Possible displacement vectors for (a) p/q = 1/5, (b) p/q = 5/12.
7.1 Vector field 153
Fig. 7.3 Vector field diagram for p/q = 1/5. Only non-zero displacements have been drawn. Note
the sequences of increasingly large “snowflakes” along the positive x axis and the main diago-
nal. Note also the large amount of white space, reflecting the prevalence of period-5 orbits with
displacement vector (0, 0).
Fig. 7.4 Vector field diagram for p/q = 5/12. The period-12 orbits appear as dots which are too
scarce to have a non-zero density.
the relation is the following. The round-off map is nothing but the lattice restriction
of the map W of (1.3) with the true rotation matrix replaced by its oblique cousin
C and the kick function f (x) taken to be the fractional part of λ x. Unfortunately
there is a mismatch between the unit periodicity of Z 2 and the periodicity τ = λ −1
of f (x), and so the local-global decomposition introduced in Chap. 1 does not work.
Nevertheless, as we will see below, the essence of the dynamics is once again to
be found in a piecewise isometry of the square, with quadratic irrational parameter
conjugate to λ . This localization of the round-off map on the integer lattice will
allow us to take over without change the results of Chaps. 2 and 3 and quickly turn
our attention to the issue of the propagation of round-off error on the infinite lattice.
7.2 Localization
The localization map E (x) maps the integer lattice Z 2 into a dense subset of the
square
[0, τ )2 , τ > 0,
Ω=
(τ , 0]2 , τ < 0.
Specifically,
E ((x, y)) = (x − m0 τ , m0 − y − m1τ ), m0 = λ x,

m1 = λ (m0 − y).
(7.2)
The following theorem shows the conjugacy of the round-off map Φ on Z 2 to a
piecewise isometry Φ̃ on E (Z2 ) ⊂ Ω :
Theorem 7.1. Let λ be a quadratic irrational of the form λ = 2 cos 2πρ , ρ rational,
with λ̃ its conjugate root. Let Φ be the round-off map with parameter λ . Further,
let Φ̃ be the piecewise isometry on Ω defined by
def
Φ̃ (ζ ) = C̃ · ζ mod Ω = C̃ · ζ − (0, mτ ), (7.3)
where
0 1
C̃ =
−1 λ̃
and m is the unique integer such that the right-hand side of (7.3) lies in Ω . Then
Φ = E −1 ◦ Φ̃ ◦ E .
Proof. Thanks to the irrationality of τ , if ζ is of the form x + mτ , x, m ∈ Z, then

the choice of x and m is unique. Hence E −1 is well defined on E (Z2 ) and it will be
sufficient to verify the intertwining relation
Φ̃ ◦ E (x) = E ◦ Φ (x) (7.4)
for arbitrary x, y ∈ Z. To evaluate the left-hand side, we note that

7.3 Localization of the vector field and periodic orbits 155
λ̃ = μ (q) − λ = (h(q) − 2)τ , μ (q), h(q) ∈ Z,
so that

0 1 x − m0 τ
Φ̃ ◦ E (x) = mod Ω
−1 λ̃ m0 − y − m 1 τ

m0 − y − m 1 τ
= mod Ω
m1 − x + τ ((h(q) − 2)(m0 − y) − μ (q)m1 + m0)

m0 − y − m 1 τ
= mod Ω
m1 − x
But this is just E applied on (7.1), and so the lemma is proved.

It is easy to invert the localization map, defining E −1 on the points of Ω ∩ Z[τ ] 2 :

−1 ζ0 [ζ0 ]1
E = , (7.5)
ζ1 −[ζ1 ]1 − [ζ0 ]τ
where we define the “integer part” and “tau part” of m + nτ , m, n ∈ Z, by
[m + nτ ]1 = m, [m + nτ ]τ = n.
We note that E −1 is additive and distributive with respect to multiplication by inte-

gers:
E −1 (χ + ψ ) = E −1 (χ ) + E −1 (ψ ), E −1 (kχ ) = kE −1 (χ ), k ∈ Z.
7.3 Localization of the vector field and periodic orbits
We now want to show how the localization map E provides a remarkable organiza-
tion of the lattice orbits according to shape (a concept yet to be defined precisely).
We begin by calculating the possible vector field vectors d (i) , i = 0, 1, . . . , N for a
given model, showing that the set of lattice points assigned a specific vector field
d(i) is mapped by E into a finite number of convex polygons in Ω . The simplest
way of establishing these results is to construct the time-q dynamical partition of
Ω for the piecewise isometry Φ̃ (with parameter λ̃ ) defined in Theorem 7.1. Recall
from Sect. 6.1 that this is a partition of Ω into convex polygons D (k) , each of which
is mapped without branching for q iterations of the local map. The cumulative effect
of these mappings on D (k) is an isometry, namely
Φ̃ (ζ ) = C̃q · ζ + δ (k) = ζ + δ (k) , δ (k) ∈ Z[τ ],
where we have made use of the identity C̃q = 1. Now the vector field d assigned to
any lattice point E −1 (ζ ), ζ ∈ D(k) , is by definition given by
d(E −1 (ζ )) = E −1 (ζ + δ (k) ) − E −1 (ζ ).
Using the additivity of E −1 , we get, finally,

(k)
[δ0 ]1
d(E −1 (ζ )) = E −1 (δ (k) ) = (k) (k) .
−[δ1 ]1 − [δ0 ]τ
This establishes the following theorem:

Theorem 7.2. Suppose the convex polygon D has the property that it is mapped
without branching during q iterations of Φ̃ . Then the vector field assumes a unique
value on the set E −1 (D ∩ Z[τ ]2 ).
The localization of the vector field is illustrated in Fig. 7.5, for p/q = 1/5. Above,
one sees a 41 × 41 array of lattice points, color-coded according to the vector field
values. Below, the local images of these points in Ω are displayed, showing the
localization of the vector field (arrows) within the convex polygons of the time-q
dynamical partition.
Fig. 7.5 Localization of the vector field for p/q = 1/5. (See the color figure at the end of the
book.)
From the localization of the vector field, it is a relatively simple step to the local-
ization of walks , i.e. dynamically generated vector-field sequences
{d(k0 ) , d(k1 ) , . . . , d(kN−1 ) }.
Here we extend the notion of a walk to include all iterations of Φ , which means that,
except very close to the origin, a walk will include q disjoint branches, each of which
is a walk in the conventional sense, i.e. a sequence of small steps forming a broken
line segment on the lattice. Two walks will be considered to have the same shape
if their vector-field sequences are identical, which is easily seen to be equivalent to
congruence of their respective branches as broken line segments.
7.3 Localization of the vector field and periodic orbits 157
Consider now the E images of all the lattice points (including the initial point, but
not the final one) of an N-step walk. These clearly will inhabit convex subpolygons
of the local vector-field domains D (k) . For a three-step walk,
{d(k0 ) , d(k1 ) , d(k2 ) },
the polygons would be
P1 = Φ −2 (Φ (Φ (D(k0 ) ) ∩ D(k1 ) ) ∩ D(k2 ) )}, P2 = Φ (P1 ), P3 = Φ (P2 ).
Probably the most significant walks to be considered are the periodic orbits, since
in the models treated in this chapter the points on periodic orbits constitute the full
set of points on the lattice (for p/q = 1/5, this is strongly suggested by Fig. 7.3). The
localization of periodic orbits in the p/q = 1/5 model is nicely illustrated in Fig. 7.6.
On the left, we see all points of the period-20 orbits (5 disjoint branches, each with
Fig. 7.6 Localization of the period-20 orbits for p/q = 1/5.
Fig. 7.7 Localization of the period-9444 orbits for p/q = 5/12.

4 points arranged in a square) in a 60 × 60 portion of Z 2 centered at the point (0, 0).

Each walk is now a sequence of four different displacement vectors repeated ad
infinitum, and there are infinitely many walks with the same shape distributed fairly
uniformly over the plane. On the right is the set of E -images of the lattice points on
the left, neatly segregated into four disjoint “regular” pentagons, each one associated
with one of the vector field types. The q-fold symmetry of the polygons, visited
q times per period, is typical of this model, but is not a general rule (in another
case, a period-20 orbit might have a local orbit of 20 unsymmetrical polygons, each
visited once per period). A much more complicated orbit, in both its global and local
versions, is shown in Fig. 7.7.
7.4 Symbolic codes for walks
In our treatment of transport in kicked oscillator models, we found it useful to think

of the local piecewise isometries and related global lattice permutations as two man-
ifestations of the same symbolic dynamics. Given the existence of the localization
map E , it is rather obvious that the lattice dynamics of the round-off model with the
same local map Φ̃ must somehow be a third representation of the same symbolic
scheme. Let us now examine how this works in detail.
The simplest coding of local-map orbits is provided by the generating partition
of the square Ω into subdomains Ω i . Keeping track of the sequence {i 0 , i1 , . . .}of the
Ωik through which an orbit passes is an effective way of describing its local motion.
For the kicked-oscillator models, this sequence uniquely specified a lattice itinerary
as well. Unfortunately, the same is not true of the round-off model’s lattice motions:
for ζ 1 and ζ 2 in the same Ω i , the vector-field displacements associated with the
corresponding lattice points E −1 (ζ 1 ) and E −1 (ζ 2 ) are not necessarily the same.
A more useful coding corresponds to the finer partition provided by the q-step
dynamical partition of Ω , i.e. the partition into maximal convex polygons Ξ i which
are mapped forward q steps without intersecting a discontinuity line. By construc-
tion, the vector field is uniform on each domain of the partition, although there may
be more than one such domain with the same vector field. Now, to each sequence of
Ξi , the shape of all lattice walks whose local representatives pass through those do-
mains is uniquely determined. Here the concept of shape is in the generalized sense
introduced in Sect. 7.3.
The recursive tiling property of the local map in our renormalizable kicked-
oscillator models allowed us to introduce a highly efficient hierarchical coding and
symbolic dynamics. Much of this can be taken over to the formulation of round-
off models, with an important difference. In the models we study, all lattice points
are localized within periodic polygons, and so the intersection of E (Z 2 ) with the
residual set of the local map is empty. Nevertheless, for mappings of tiles within the
hierarchy, the symbolic dynamics remains a valuable tool.
7.5 Construction of the probability distribution 159
7.5 Construction of the probability distribution
Theorem 7.1 establishes the dynamical equivalence of round-off maps and piece-
wise isometries of the square for quadratic irrational parameters. It allows us to take
over immediately everything we have learned about renormalizability and Vershik-
type symbolic dynamics. In some ways the situation is analogous to that of the
kicked-oscillator models which have been our major concern so far. But there are
crucial differences, especially when it comes to formulating a meaningful transport
problem. In both cases, we are interested in a probability distribution P(x,t) giving
the probability that the orbit of a randomly selected initial point with lattice coor-
dinates x0 ends up displaced on the lattice by x after t iterations of the elementary
map. To remove the effects of global rotations, one adopts a stroboscopic viewpoint,
requiring that t be a multiple of 4 (resp. q) in the kicked-oscillator (resp. round-off)
case.
In the kicked-oscillator case, the local phase space is the fundamental domain of
an infinite crystalline array, and so, using translation invariance, random selection of
the initial point refers naturally to the uniform distribution of points in the square,
and convex polygons are assigned weights in the probability calculation propor-
tional to their respective areas. In the round-off case, the phase space consists solely
of the lattice points, and the local square Ω is an abstraction. Moreover, we do not
have translation invariance: the trajectories initiated at different lattice sites can have
different shapes. Nevertheless, it is possible to assign nonzero weights to lattice or-
bits according to their shapes, and moreover to identify these weights with the areas
of convex polygons in Ω . This will then allow us to construct P(x,t) along the same
lines as we used in Chapter 6 for the kicked oscillator models.
To understand how a probability measure can be assigned to a lattice, let us con-
sider the analogous one-dimensional problem (Lowenstein and Vivaldi, 1998). The
points of the non-negative integer lattice Z + = {0, 1, . . .} can be “localized” on the
unit interval by the map
η : Z → [0, 1), η (m) = {λ m} = λ m − λ m,
where λ is a positive irrational real number and {·} is the fractional part function.
For any interval I ⊂ [0, 1), the set η −1 (I) ⊂ Z+ can be assigned a density
1
D(η −1 (I)) = lim #{m ∈ η −1 (I) : m < n} ≤ 1.
n→∞ n
By Weyl’s theorem (see Hardy and Wright, 1979, p. 445), this density is equal to
the length of I,
D(η −1 (I)) = |I|.
It is a standard exercise in measure theory to promote the density function on special
subsets to a measure on the lattice Z + . The latter is additive over finite unions of
disjoint sets, but not over countable unions. For example, the set Z + has unit mea-
sure, but is the union of countably many discrete integers, each of which has zero
measure.
The reader can verify that in our round-off models, the localization map E is
essentially the 2-dimensional analogue of our 1-dimensional example. It is then
straightforward to generalize the above arguments and prove (Lowenstein and Vi-
valdi, 1998) the following lemma.
Lemma 7.1. Let D ⊂ Ω be a convex polygon. Then the density
1
D(E −1 (D)) = lim #{(l, m) ∈ E −1 (D) : −n < l, m < n} ≤ 1.
n→∞ 4n2
is equal to the area of D divided by the area of Ω .

Note that this result extends immediately to finite unions of convex polygons.
Having defined a density measure on the lattice, we are in a position immediately
to define the probability distribution for the propagation of round-off error, namely
P(x,t) = D({z ∈ Z2 : Φ t (z) − z = x}), (7.6)
with t restricted to integer multiples of q. Since the argument of D is the intersection

of a finite set of walks, namely all those which produce a net lattice displacement
of z after t iterations, the localized image of that set resides in a finite set of convex
polygons in Ω . By Lemma 7.1, the right-hand side of (7.6) is just the total area of
those polygons.
For an efficient calculation of the probability distribution and its moments, we
can turn to the procedures of Sect. 6.1 which exploit both recursive tiling and hier-
archical symbolic dynamics. There are, however, some differences worth noting. In
the kicked-oscillator case, we had two alternative ways of implementing the local-
global correspondence. In the simpler version, which we adopted, each application
of the map W corresponds both to the a unique piecewise isometry of the square and
to a lattice isometry with rotation number 1/4, whose displacement vector depends
uniquely on the domain Ω i containing the local point. Iteration of W produces a
quadruply branched generalized walk on the lattice. An alternative implementation
focuses attention on the fourth-iterate map W 4 , whose generating domain map in-
volves a finer partition of Ω . The local representation is again a piecewise isometry
of Ω , but now the corresponding lattice isometry is a translation, again depending
on the local subdomain. Now iteration of W produces four singly branched walks,
one for each value of the rotational phase.
In the round-off models, we have no local-global decomposition of the map Φ ,
only of Φ q , and so only the second scenario is available to us. Thus in our proba-
bility calculations we are obliged to use a lowest-level partition in which the vector
field vector is unique on each subdomain, and to keep track of q separate walks si-
multaneously if we want to include a complete set of initial conditions. Fortunately,
these add only minor complications to our algorithms.
7.6 Rotation number 1/5 161
7.6 Rotation number 1/5
In the remainder of this chapter√we will consider several interesting examples, start-
ing with p/q = 1/5 and λ = ( 5 − 1)/2, the one case in the literature where rigor-
ous results concerning the long-time behavior of the moments have been obtained
(Lowenstein and Vivaldi, 1998). Here we follow closely the treatment of that article.
Let us review√briefly what we have learned about this model so far. The round-off
map with λ = ( 5 − 1)/2 is defined in (7.1). Its vector field, d(x) = Φ 5 (x) − x, is
plotted, over a 200 × 200 portion of the lattice, in Fig. 7.3. We note the apparent ab-
sence of unbounded orbits (proved for the entire lattice by Lowenstein et al. (1997)
and generalized to other models by Akiyama et al. (2008), as well as the presence of
scaling sequences of periodic orbits (apparently tending toward fractal snowflakes)
arrayed along the x axis and the 45 ◦ diagonal. The localization of the vector field is
illustrated in Fig. 7.5.
7.6.1 Recursive tiling for the local map
According to Theorem 7.1, the local map is given by
Φ̃ (ζ ) = C̃ · ζ mod Ω ,
with √
0 1 −1 5+1
Ω = [0, τ ) ,
2
C̃ = , τ =λ = .
−1 −τ 2
This is the same as the local map of Lowenstein and Vivaldi (1998), apart from a
trivial interchange of the coordinates ζ 0 and ζ1 .
The generating domain and its four atoms, shown in Fig. 7.8, are
D(Ω ) = D((0, 1, 0, 1), (−1, −1, 1, 1), (0, 0, τ , τ )),

D0 (Ω ) = D((0, 1, 2), (−1, −1, −1), (0, 0, 0)),
D1 (Ω ) = D((0, 2, 1, 2), (−1, 1, −1, −1), (0, 0, 0, −τ )),
D2 (Ω ) = D((0, 1, 2, 0, 2), (−1, 1, −1, 1, 1), (0, τ , −2τ , τ , −τ )),
D3 (Ω ) = D((0, 1, 2), (1, 1, 1), (τ , τ , −2τ )).
The choice of level-0 scaling domain is not unique, but it is advantageous to

follow Lowenstein and Vivaldi (1998) (rather than Kouptsov et al., 2002) and define
D(0) = D((0, 2, 3), (1, −1, −1), (0, 0, 1 − τ )).
The level-zero subdomains D j (0) and their return times T j (0) are
Fig. 7.8 Generating partition for p/q = 1/5.
D0 (0) = D((1, 0, 2, 1, 3), (−1, 1, −1, 1, −1), (0, 0, 0, −1 + τ , 1 − τ )), T0 (0) = 4

D1 (0) = D((0, 1, 3), (1, 1, −1), (0, 0, 1 − τ )), T1 (0) = 6.
Here the coordinates are relative to the scaling limit point (τ , 1).
The level-0 tiling, including not only the return orbits of D 0 (0) and D1 (0), but
also the three period-1 pentagons
Π0 (Ω ) = D((0, 2, 4, 1, 3), (−1, −1, −1, −1, −1), (0, −τ , −τ , 0, 1)),

Π1 (Ω ) = D((1, 4, 2, 0, 3), (1, 1, 1, 1, 1), (τ , −τ , −τ , τ , 2 + τ )),
Π2 (Ω ) = D((0, 2, 4, 1, 3), (1, 1, 1, 1, 1), (τ , −2τ , −2τ , τ , 3 + τ )),
is shown in Fig. 7.9, together with the level-1 tiling of D(0). Note that the limit point
of the scaling sequence is (τ , 1), and that the level-1 scaling triangle D(1) straddles
the boundary between D 0 (0) and D1 (0). This is different from our usual convention
that D(L) lies intirely in a single D j (L − 1). We note that the level-1 tiling includes
two periodic pentagons, Π 0 (0) and Π1 (0), where, expressed in terms of coordinates
relative to the scaling limit,
Π0 (0) = D((1, 4, 2, 0, 3), (−1, −1, −1, −1, −1), (0, 2 − τ , 0, 1 − τ , 1 − τ )),
Π1 (0) = D((0, 3, 1, 4, 2), (1, 1, 1, 1, 1), (0, 3 − 2τ , 0, 2 − τ , 2 − τ )),
with minimal periods 14 and 26, respectively. The higher-level members of each
scaling sequence are obtained by rescaling of the coordinates relative to (τ , 1) by
powers of τ −2 = 2 − τ .
Once again referring to Fig. 7.9, we see that the return orbit of D 0 (1) visits D0 (0)
twice and D1 (0) once, while that of D 1 (1) visits those domains 2 and 3 times, re-
Fig. 7.9 Left: level-0 tiling of Ω . Right: level-1 tiling of D(0).
spectively. Thus the incidence matrix is

2 2
A= ,
1 3
and we get for the level-L return times,

1 2
(T0 (L), T1 (L)) = (T0 (0), T1 (0))AL = (10 × 4L + 2), (10 × 4n − 1) . (7.7)
3 3
The usefulness of recursive tiling for describing lattice walks is already evident at
level 0. Comparison of Fig. 7.9 with Fig. 7.5 shows that every domain in the level-0
tiling is assigned a unique vector-field displacement. This is a property which is not
shared by the original generating partition, but is now guaranteed for all higher-level
tilings.
7.6.2 Probability distribution P(x,t)
The probability distribution P(x,t) is normalized to unity for all t, and is initially
entirely concentrated at the origin,

1 x = (0, 0),
P(x, 0) =
0 x = (0, 0).
The relative frequencies on the lattice of the seven vector field displacements are
given by P(x, 5) (recall the constraint that t be a multiple of 5):
⎧
⎨ 14τ − 22 = 0.652 . . .
⎪ x = (0, 0),
P(x, 5) = (5 − 3τ )/2 = 0.073 . . . x = ±(1, 0), ±(0, 1),
⎪
⎩
(13 − 8τ )/2 = 0.028 . . . x = ±(1, 1).
As is evident in Fig. 7.3, nearly two thirds of the lattice points are fixed points of
Φ 5 . They form an inert background which does not participate in the time evolution
of P(x,t).
As we saw in Chap. 6, a simple way of calculating P(x,t) for small t is provided
by straightforward calculation of the dynamical partition of Ω for a given t. Each
convex polygonal domain of the partition corresponds to a distinct displacement
vector xi and a distinct area A i . One then has
P(x,t) = τ −2 ∑ Ai δ (x, xi ),
i
where δ (a, b) is equal to 1 for equal arguments and vanishes otherwise. Unfortu-
nately, this technique quickly becomes unwieldy for increasing t, and so once more
we turn to recursive tiling for relief. From (6.7) and (6.8), we have
L∗ (t)−1 1
P(x,t) = Pfix + τ −2 ∑ ∑ A (Π j (L))Q(Π j (L), x,t) (7.8)
L=0 j=0
1
+τ −2 ∑ A (D j (L∗ (t)))Q(D j (L∗ (t)), x,t)
j=0
where Q(D, x,t) counts the number of times that a forward t-step displacement
vector x is encountered along the return orbit of a tile D. Moreover, if T is the
length of the return orbit, then L ∗ (t) is chosen to be sufficiently high that each set
Φ̃ k D, k = 0, 1, . . . , T − 1 has a unique forward t-step displacement vector. Explicitly
we can choose (see (6.4) and (6.5))
L−1
L∗ (t) ≥ min{L : t ≤ T∗ (L)}, T∗ (L) = ∑ T0(L ).
L =0
Here we have been careful to choose, on the right-hand side of the last equation, the
smaller of the two return times T0 and T1 . Inserting (7.7), we have
1 L−1
3 L∑
T∗ (L) = (10 · 4L + 2).
=0
Summing the geometric series, we find

10 L 2
T∗ (L) = (4 − 1) + L, (7.9)
9 3
and so, for L > 1, we may choose
L∗ (t) = !log4 (9t/10)". (7.10)
A simplifying feature of the chosen recursive tiling is apparent in Fig. 7.9: for
t ≤ T j (L) + T∗ (L) iterations, the itineraries of D j (L) and Π j (L) are identical, and
hence
Q(D j (L∗ (t)), x,t) = Q(Π j (L∗ (t)), x,t).
This allows us to write (7.8) in a more concise form,
2L∗ (t)
P(x,t) = Pfix + ∑ bn Pn (x,t) (7.11)
n=0
where
P2L−1 (x,t) = Q(Π0 (L), x,t)/T0 (L), P2L (x,t) = Q(Π1 (L), x,t)/T1 (L),
and
b2L−1 = τ −2 A (Π0 (L))T0 (L), b2L = τ −2 A (Π1 (L))T1 (L),
b2L∗ (t)−1 = τ −2 A (D0 (L∗ (t)))T0 (L∗ (t)), b2L∗ (t) = τ −2 A (D1 (L∗ (t)))T1 (L∗ (t)).
7.6.3 Fractal snowflakes
Figure 7.3 strongly suggests that the scaling sequences of periodic orbits in the
p/q = 1/5 round-off model approach, when suitably rescaled, some sort of fractal
curve as the periods tend to infinity. What is less apparent is that this geometric
self-similarity will reflect itself in asymptotic scaling properties of the probability
distribution P(x,t) and its moments. Following Lowenstein and Vivaldi (1998), we
will now show that both geometric and dynamical self-similarity are in fact valid:
there is a precise relation between asymptotically long periodic orbits and fractal
snowflakes obtained by a Koch construction, and this geometrical property underlies
the modulated power-law behavior of moments.
We begin by showing that the periodic orbits can themselves be assembled, with-
out approximation, by a Koch-type construction. The key concept here is that of a
walk W , represented abstractly as a coded sequence of displacement vectors, i.e.
W = (dc0 , dc1 , . . . , dcN−1 )
where the vectors d i are taken from a displacement list D of M vectors, and the c i are
integers in 0, . . . , M − 1. If the set D is considered fixed, it is convenient to represent
W by the code sequence
C = (c0 , c1 , . . . , cN−1 ).
The kth vertex r k of a walk W is defined as the position (relative to the starting point)
reached after k steps:
k−1
r0 (W ) = (0, 0), rk (W ) = ∑ dc j .
j=0
We denote by W (D) the set of all walks with displacement list D. A substitu-
tion φ maps each element of {0, . . . , M − 1} into a finite sequence of elements
of {0, . . . , M − 1}, and extends naturally, by concatenation, to a map on code se-
quences.
To illustrate the usefulness of walks in the construction of fractal curves, let us
consider the well known Koch snowflake. Here the displacement list D consists of
six elementary steps,

2π k 2π k
dk = cos , sin , k = 0, . . . , 5.
6 6
We start with the three-step walk around an equilateral triangle,
W0 = (d0 , d2 , d4 ).
Applying the code substitution
φ (k) = (k, (k + 5)6 , (k + 1)6 , k), k = 0, . . . , 5,
to the sequence (0, 2, 4) produces a new sequence
φ (0, 2, 4) = (0, 5, 1, 0) (2, 1, 3, 2) (4, 3, 5, 4) = (0, 5, 1, 0, 2, 1, 3, 2, 4, 3, 5, 4),
and a new walk (around a 12-sided hexagram)
W1 = (d0 , d5 , d1 , d0 , d2 , d1 , d3 , d2 , d4 , d3 , d5 , d4 ).
Iterating the process, shown geometrically in Fig. 7.10, we obtain at the nth repeti-
Fig. 7.10 Iterative construction of the Koch snowflake.
tion an Nn = 3 · 4n -sided polygon with diameter d n = 2 · 3n−1/2 , traversed counter-

clockwise. Since
δ
Nn dn log4
= , δ= ,
N0 d0 log3
we see that in the limit n → ∞, the polygons, rescaled by 3 −n , approach a hexago-

nally symmetric fractal of dimension δ .
A pentagonal Koch-type walk K n resembling our round-off walks may be con-

structed using the displacement lists
D− = ((0, −λ 2 ), (λ 2 , 0), (λ 3 , λ 2 ), (−λ 3 , λ 3 ), (−λ 2 , −λ 3 )), (7.12)

D+ = −λ D+ = ((0, λ ), (−λ , 0), (−λ 2 , −λ ), (λ 2 , −λ 2 ), (λ , λ 2 )). (7.13)
and the substitutions φ ± defined by

φ− (k) = (k, (k + 1)5 , (k + 4)5 , k)
, k = 0, . . . , 4.
φ+ (k) = (k, (k + 4)5 , (k + 1)5 , k)
Starting with the walks on pentagons
K0± = (d0 (D± ), d1 (D± ), d2 (D± ), d3 (D± ), d4 (D± )),
the higher-level walks are generated by
K2l−1 = σ−l (K0− ), K2l = σ+l (K0+ ), l = 1, 2, . . . .
where σ± are the maps induced by the substitutions φ ± .

With suitable rescaling, the polygons traced out by the two sequences of walks
can be made to have a single fractal limit set K . Specifically, we define rescaled
position vectors kn,i , n = 0, 1, . . . , i = 0, 1, 2, 3, 4,
k0,i = λ −1 xi (K0+ ) = λ −2 xi (K0− ),

k2l−1,i = λ 2l−2 xi (K2l−1 ) + (0, 1), l = 1, 2, . . .
k2l,i = λ 2l−1
xi (K2l ), l = 1, 2, . . .
and the polygons

Kn = {kn,0 , . . . , kn,|Kn | },
where |Kn | = 5 × 4!n/2".

Note that the vertices of the sequence of K n are nested:
kn,i = kn+2m,4m i .
Moreover, as n tends to infinity, the edges of K n scale as τ −2n , while the number
of edges increases proportional to 2 2n , and so, by the usual arguments, the resulting
limit set K = limn Kn will have fractal dimension log 2/log τ .
Our aim in this section is to relate the asymptotic long-time behavior of round-off
orbits to geometrical properties of the fractal K . Specifically, for χ a continuous
function of lattice coordinate differences, we will relate moments χ t to geometri-
cal moments of K . The analogue of temporal displacement t is an “angle” α repre-
senting the fraction of a polygon’s perimeter contained between two vertices. This
is a quantity which has a chance of remaining meaningful in the limit n → ∞. For
the polygon K n , walking from vertex k n,i to vertex ki+ j corresponds to an angular
displacement
αn, j = j/|Kn |,
and we define the expectation value of our function χ for j-step displacement vec-
tors as
1 |Kn |−1
μn (αn, j ) = ∑ χ (kn,i+αn, j |Kn | − kn,i ).
|Kn | i=0
Due to the odd-even pairing of the K n and the |Kn | periodicity, the function μ n
satisfies
μ2l−1 (α2l−1, j ) = μ2l (−α2l, j ),
μn (αn, j ) = μn (αn, j+|Kn | ) = μn (αn, j + 1).
7.6.4 Substitution rules for lattice walks
The route to asymptotic self-similarity for the sequence of lattice walks generated
by periodic orbits of the p/q = 1/5 round-off map is not nearly as straightforward
as for the Koch construction. The starting point is, of course, to find a substitution
rule taking us from one level to the next. Using the vector-field displacement list
DV = ((1, 0), (1, 1), (0, 1), (−1, 0), (−1, −1), (0, −1))
turns out to be inadequate for this purpose. Instead, we use (Lowenstein and Vivaldi,
1998)
D = ((−1, 0), (0, −1), (1, 0), (0, 1), (0, 0), (−1, 1), (−1, −1), (0, −1), (1, 0), (1, 1)).
The code substitution φV from W (D) to W (DV ) is
φV (2) = φV (8) = (0), φV (9) = (1), φV (3) = (2),
φV (0) = (3), φV (6) = (4) φV (1) = φV (7) = (5),

φV (4) = 0,
/ φV (5) = (3, 2).
Here the symbol 0/ denotes the null sequence. With list D, using φ V to translate code
sequences into vector-field walks, it is possible to construct the periodic-orbit walks
of arbitrary level, starting with
W1 = (d0 , d1 , d2 , d3 , d4 ), W2 = (d5 , d9 , d8 , d7 , d6 ),
and iterating the map σ : W (D)− > W (D) induced by the code substitution φ de-
fined by

φ (k) = (k, (k + 1)5 , 5 + (k + 2)5),
k = 0, . . . , 4.
φ (k + 5) = (k + 5, 5 + (k + 4)5, (k + 3)5, (k + 4)5 , k + 5),
One obtains
W2l−1 = σ l−1 (W1 ), W2l = σ l−1 (W2 ).
From the construction, the numbers of steps of each type at level l are found by
applying the (l − 1)st power of the counting matrix
⎛ ⎞
1 1 0 0 0 0 0 1 0 0
⎜0 1 1 0 0 0 0 0
⎜ 1 0⎟⎟
⎜0 0 1 1 0 0 0 0
⎜ 0 1⎟⎟
⎜0 0 0 1 1 1 0 0
⎜ 0 0⎟⎟
⎜1 0 0 0 1 0 1 0 0 0⎟⎟,
M=⎜ (7.14)
⎜0 0 0 1 1 2 0 0
⎜ 0 1⎟⎟
⎜1 0 0 0 1 1 2 0
⎜ 0 0⎟⎟
⎜1 1 0 0 0 0 1 2
⎜ 0 0⎟⎟
⎝0 1 1 0 0 0 0 1 2 0⎠
0 0 1 1 0 0 0 0 1 2
from the right, on the initial row vectors
u1 = (1, 1, 1, 1, 1, 0, 0, 0, 0, 0), u 2 = (0, 0, 0, 0, 0, 1, 1, 1, 1, 1).
Summing over the components gives the lengths (in D steps) of the walks W n ,
Fig. 7.11 Substitution rules for construction of Wn , from Lowenstein and Vivaldi (1998), Fig. 6.
5
|Wn | = (2n − (−1)n ), n ≥ 1.
3
The corresponding lengths in D V steps are just the periods T j (L) in (7.7),written as
a single sequence, with L = n+1
2 :
1
Tn = (5 · 2n − 2(−1)n), n ≥ 1.
3
The geometric interpretion of σ in terms of vector-field sequences can be read off

from Fig. 7.11.
7.6.5 Separating out an asymptotic walk
We now want to show how iterated application of the substitution φ leads asymp-
totically to fractal snowflakes. Our strategy is to regard φ as a linear transforma-
tion in a 20-dimensional vector space of displacement-list components, with matrix
M̃ = M⊗ I2 , with M given in (7.14) and I 2 equal to the identity matrix in two dimen-
sions. Multiplication by this matrix from the left maps a column of displacements
into a column whose entries are the net displacements of the respective substituted
code strings.
Eigenvectors of M̃ correspond to displacement lists with simple scaling proper-
ties. The hexagonal Koch construction considered earlier has such a displacement
list: writing the components of the d k , k = 0, . . . , 5 as a 12-dimensional column vec-
tor, multiplication by the analogue of M̃ reproduces that vector, multiplied by 3.
The list D for the p/q = 1/5 round-off map does not scale this way, but rather is a
linear superposition of the three displacement lists, D 1 , D2 , and D3 , shown in Table
7.1. These correspond, respectively to the eigenvalues τ 2 , 1, and λ 2 of M̃. One can
verify the following decomposition of D: for k = 0, . . . , 9,
τ2 λ2
dk (D) = √ dk (D1 ) + dk (D2 ) + √ dk (D3 ).
5 5
Consider now the three sequences of walks Wn,i , n = 1, 2, . . . , i = 1, 2, 3, based
on the three displacement lists D i , but with the same hierarchy of code sequences as
Wn based on D. That Wn,1 dominates in the asymptotic regime is evident from the
following lemmas of Lowenstein and Vivaldi (1998).
Lemma 7.2. The vertices of Wn,1 form a subsequence of the vertices of a walk on
the Koch-type pentagonal snowflake K n with 5 · 4!n/2" sides.
Proof. We establish the embedding of vertices, pictured in Fig. 7.12, by showing

that there exist substitutions φ 1+ and φ1− which induce maps σ 1± from W (D) to
W (D± ) such that for l = 1, 2, . . .,
Table 7.1 The displacement lists Di , i = 1, 2, 3.
k dk (D1 ) dk (D2 ) dk (D3 )

0 (0, −1) (−1, 1) (0,1)
1 (1, 0) (−1, −1) (−1, 0)
2 (λ , 1) (0, −1) (τ , −1)
3 (−λ , λ ) (1, 0) (−τ , τ )
4 (−1, −λ ) (1, 1) (1, −τ )
5 (0, τ ) (−1, −1) (0, λ )
6 (−τ , 0) (1, −1) (−λ , 0)
7 (−1, −τ ) (1, 1) (1, −λ )
8 (1, −1) (0, 1) (−1, 1)
9 (τ , 1) (−1, 0) (λ , −1)
Fig. 7.12 Embedding of Wn,1 in Kn , from Lowenstein and Vivaldi (1998), Fig. 8.
K2l−1 = σ1− (W2l−1 ), K2l = σ1+ (W2l ) (7.15)
and, moreover, for k = 0, . . . , 9,
σ1 (dk (D)) = σ1+ (dk (D)) = σ1− (dk (D)). (7.16)
Explicitly, the substitutions φ 1± are, for k = 0, . . . , 4,
φ1− (k) = (k, (k + 1)5 , (k + 4)5 , k),

φ1− (k + 5) = ((k+2)5 , (k+3)5 , (k+1)5 , (k+2)5 , (k+3)5 , (k+4)5 , (k+2)5 , (k+3)5 ),
φ1+ (k) = ((k + 3)5 , (k + 2)5 ),
φ1+ (k + 5) = (k, (k + 4)5 , (k + 1)5 , k).
Equation (7.15) is proved by induction. It is easy to verify the equalities for l = 1,

and then proceed to higher levels using the intertwining relation
σ1± σ = σ± σ1± .
The latter is readily established directly from the explicit definitions of the respective
code substitutions.
Lemma 7.3. Wn,2 is a walk on the pentagon with vertices
(0, 0), (−1, 1), (−2, 0), (−2, −1), (−1, −1),
and so its vertices are bounded in norm by 3.
Proof. The code for Wn,2 is the same as for Wn , namely φ n (0, 1, 2, 3, 4) for odd n and
φ n (5, 9, 8, 7, 6) for even n. These code sequences are easily seen to have the simple
succession rule: for k = 0, . . . , 4, the index k is always followed by either (k + 1) 5 or
5 + (k + 1)5 , while k + 5 is always followed by either 5 + (k + 4) 5 or (k + 4)5 . The
reader will readily verify that each step on the pentagon π prescribed by the code
is followed by another step on π , either forward or backward. Thus the walk W n,2 ,
which necessarily starts with a step from (0, 0) to either (−1, 1) (n odd) or (−1, −1)
(n even), always remains on π .

√
Lemma 7.4. The vertices of Wn,3 , scaled by the factor λ 2 / 5, are bounded in norm
by 1.
Proof. The proof is analogous to that of Lemma 7.2. One shows that the vertices
of Wn,3 are a subsequence of those of a walk K̃n generated by a Koch-type con-
struction based once again on the substitutions φ ± and initial codes (0, 1, 2, 3, 4) and
(0, 4, 3, 2, 1), but with the displacement lists
D̃− = ((0, 1), (−1, 0), (τ , −1), (−τ , τ ), (1, −τ )),

D̃+ = τ D̃− = ((0, τ ), (−τ , 0), (τ 2 , −τ ), (−τ 2 , τ 2 ), (τ , −τ 2 )).
The mapping σ̃− , which takes K̃0− into K̃1 , is illustrated in Fig. 7.13. In contrast
to the case of Wn,1 , where the diameter of the polygonal path traversed by the walk
increases with n as τ n , the walks Wn,3 never leave the confines of a bounded region √
(the square [−τ , τ ]2 for n odd, and [τ 2 , τ 2 ]2 for n even). Rescaling by a factor λ 2 / 5,
we confirm the bound stated in the lemma. For further details of the proof, see
Lowenstein and Vivaldi (1998).

It is now the time to quantify the small differences among the four families of
periodic walks we have introduced, namely the vector-field walks {V n }, the scal-
ing family {Wn }, the dominant parts {Wn,1 }, and the pentagonal Koch snowflakes
{Kn }. Omitting the tedious details, we state the key lemma bounding the difference
between corresponding vertices x t (Vn ) and x j (Kn ) (Lowenstein and Vivaldi, 1998).
Fig. 7.13 Substitution rules for construction of Kn and K̃n , from Lowenstein and Vivaldi (1998),
Fig. 7.
Lemma 7.5. For all positive t, n, there exist constants ε i , i = 1, 2, 3, such that
τ2
||xt (Vn ) − √ xs(n)t (Kn )|| ≤ ε1 (2−nt)1/δ + ε2 !n/2" + ε3,
5
where s(2l − 1) = 6, s(2l) = 3.
7.6.6 Asymptotic scaling
The final steps in deriving the long-time behavior of moments involve a number of
of technical lemmas (Lowenstein and Vivaldi, 1998). We will be satisfied here with
a sketch of the main arguments and a statement of the final result.
We consider the expectation value χ t of a continuous and piecewise differ-
entiable function (random variable) χ (x) on Z 2 which is homogeneous of degree
κ > 0, i.e. for all positive ρ ,
χ (ρ x) = ρ κ χ (x).
The expectation value χ t with respect to P(x,t) can be decomposed into contribu-
tions χ n,t from walks over the respective Π n using (7.11):
2L∗ (t)
χ t = ∑ bn (t)χ n,t .
n=0
Here a walk on Π n is extended by periodicity to time t if the latter exceeds the

return time Tn . Using Lemma 7.5, the partial expectation value χ n,t can be well
approximated for large t by the corresponding rescaled expectation value on the
Koch snowflake, √
(λ n+1 5)−κ μn (αn,s(n)t ),
with error terms given by the lemma. Further approximations, with additional error
5
estimates, include: (1) replacing Tn by 2n in the definition of b n (t); (2) deleting
3
from the sum over n the terms with n < L ∗ (t); (3) relacement of μ n (αn,st ) by (see
Lowenstein and Vivaldi, 1998, Lemma 5.1, for justification)
μ ((−1)n αn,st ) = lim μn+2p ((−1)n αn,st );

p→∞
(4) making the substitution

3t 1
(−1)n αn,st = (−2)2L∗ (t)−n αt , αt = = 4{9t/10} ,
5 · 4! log4 (9t/10)" 6
where {·} is the fractional part function; (5) changing the summation variable to
m = 2L8 (t) − n; and (6) extending the summation range to infinity. The final result
of all these manipulations, accompanied by careful error accounting, is the following
theorem (Lowenstein and Vivaldi, 1998, Theorem 5.2).
Theorem 7.3. Let χ be a homogeneous random variable of degree κ > 2 − δ , where
δ = log2/ log τ . Then the following holds:
χ t = t β (G(t) + O(t −a logb t)), (7.17)
where
κ −2 β 1 1
β = 1+ , a = min , , b = 1+ ,
δ 2 δ δ
G(t) = F(αt ),
β
1 3 √ ∞
cm
F(α ) = (λ 5)2−κ ∑ β m μ ((−2)m α ),
6 5α m=0 2
√ 4
with c0 = 1, c1 = λ , and cm = 5λ for m > 1. The form factor G(t) is log-periodic,
i.e. is a function of logt with period log4.
We thus have, for the moments of the probability distribution, asymptotic power-
law t β modulated by a log-periodic form factor G(t). If we think of these formulas
as describing some sort of diffusive propagation of the round-off error, the diffusion
is decidedly anomalous, even in the special case of the second x-moment x 2 t : the
second transport coefficient, defined as
7.7 Model I 175
g2 (t) = t −1 (x2 t − x t2 ),
is just G(t), which fails to tend toward a limit for t tending to infinity, in contrast to
the convergent transport coefficient of normal diffusion.
Lowenstein and Vivaldi (1998) made a high-resolution numerical calculation of
the form factor G(t) for χ (m, n) = m 2 , with t in the range 683 ≤ t ≤ 2730. Their
efficient algorithm exploited a linear recursion relation for the 2-point correlation
function of the snowflakes Kn . The plot is shown in Fig. 7.14.
Fig. 7.14 Plot of the form factor G(t) for 683 ≤ t ≤ 2730, from Lowenstein and Vivaldi (1998),
Fig. 9.
7.7 Model I
We now explore the asymptotic long-time behavior of the probability distribution

P(x,t) and its moments using the computer assisted methods of Chapter 6. For the
sake of comparison, we will consider precisely the same local maps which we em-
phasized in the context of kicked oscillator maps, namely Models I and II, with
p/q = 3/10 and 5/12. The underlying symbolic dynamics is, of course, exactly the
same, while the lattice isometries driven by the symbolic machine are completely
different. In particular, the lattice rotation number is always 1/4 for our kicked os-
cillators, while for the corresponding round-off model it has the same denominator
as that of the local map.
We treat first Model I, using the same list of 501 time intervals and level-8 tiling
of the local square that we employed in Chap. 6. The list of participating domains
(the three scaling domains D j (8) with nonzero area, plus 17 periodic domains) are
listed with their return times and areas in Table 6.1. Choosing a representative lat-
tice point localized in each of the 20 domains, we calculated and stored the orbits
obtained by iterating the round-off map over a complete return time, augmented by
t∗ (8) = 218450. The time-t lattice displacements, for each of the 501 t values and
each orbit point were calculated and their weighted contribution to P((m, n),t) and
to the expectation values of |m|, m 2 , and m4 , accumulated. The results are displayed
in Figs. 7.15 and 7.16.
As we might have anticipated from Sect. 7.6.6, the kth moments are characterized
asymptotically by power-law growth modulated by log-periodic form factors:
k−2
|x|k t ∼ t 1+ δ Gk (t)
where δ is the fractal dimension associated with the local-map scaling (ratio of
the logarithms of the temporal and spatial scale factors), as well as the asymptotic
Fig. 7.15 Plots of the Model I form factors Gk as functions of log4 t for moments |m| t , m2 t ,
and m4 t .
7.8 Model II 177
Fig. 7.16 Evolution of the Model I probability distribution. Successive t values differ by approxi-
mately a factor of 21/4 .
scaling of the periodic orbits on the lattice, and the asymptotic logarithmic period of
Gk is one-half the logarithm of the temporal scale factor. From Fig. 7.16, we see that
the support of the probability distribution also exhibits apparent power-law growth
with log-periodic modulation.
7.8 Model II
Perhaps the most fascinating example of long-time asymptotics among the class of
renormalizable round-off models is Model II, with rotation number 5/12, thanks to
the presence of two competing pseudochaotic webs with different dimensions. From
the preceding sections, we expect the moments to behave as
1+ k−2 1+ k−2
|x|k t ∼ t δ
A GAk (t) + t δ
B GBk (t), (7.18)
where
log 4 log 5
δA = − √ = 1.05 . . ., δB = − √ = 1.22 . . .,
log(2 − 3) log(2 − 3)
This means that, for k < 2, the B term will dominate for asymptotically large t, while
the A term will be dominant for k > 2. The especially interesting case is k = 2, since
there the dependence on dimension drops out, and (7.18) reduces to

|x|k t ∼ t GAk (t) + GBk(t) .
The two form factors oscillate with incommensurate frequencies, and hence the
quasi-diffusive power law is modulated by a log-quasiperiodic factor.
The expected long-time behavior in Model II has been verified by Lowenstein
and Liu (2003) in a numerical calculation which uses essentially the method applied
to Model I in the preceding section. The authors used a list of 441 time intervals
approximately equally spaced on a logarithmic scale between 21000 and 24982500
for the two sequences of periodic orbits in the A hierarchy, with scaling limit point
(41+ 72τ , 41+ 72τ ), and 292 t values between 21000 and √ 2271000 for sequences in
the B hierarchy, with limit point (τ , τ ), where τ = −1/ 3. The log-log plots of x 2 t
and x2 t /t in Fig. 7.17 provide strong evidence for the validity of the asymptotic
logarithmic quasiperiodicity.
In the absence of a rigorous asymptotic analysis analogous to that of Sect. 7.6.6,
it is important to probe more deeply the asymptotic power laws embodied in (7.18),
making evident the remarkable competition between the two scaling hierarchies
with different fractal dimensions. The numerical experiment described below fills
in some of the missing pieces. It also provides a fitting way to wrap up the final
chapter of this book, since it exploits in a powerful and practical way its major
themes of recursive tiling and underlying symbolic dynamics.
The new investigation extends the results of Lowenstein and Liu (2003) in several
ways. First of all, the calculation treats time intervals uniformly for the two scaling
hierarchies and the nonscaling orbits as well, in the range 7200000 ≤ t ≤ 36000000,
while the previous work assumed maximum times 2271000 and 24982500 for hier-
archies A and B, respectively, and omitted the nonscaling orbits from the probability
sums. While the latter are bounded on the lattice give no significant contribution to
the asymptotics of the second moment, they do provide an important component
of the probability distribution for small excursions on the lattice. A further innova-
tion in the current experiment is the exploration of moments (|m| + |n|) k over a
sufficiently wide range of k values to test the asymptotic behavior conjectured in
(7.18), including the alternating dominance of the A and B hierarchies for k greater
than or less than 2, respectively. Finally, the computation exploits fully the sym-
bolic dynamics underlying the propagation of round-off error on the lattice: each
map iteration is executed by updating a finite codometer.
7.8 Model II 179
Fig. 7.17 Log-log plots (Lowenstein and Liu, 2003, Figs. 17, 20–22) of x2 vs. t (on the left)
and the corresponding form factors (on the right) for the scaling sequences of the 5/12 model
associated with the limit points (τ , τ ) (above) and (41 + 72τ , 41 + 72τ ) (below).
In Model II, the recursive tiling property (see Sect. 2.3) allows us to partition the
square Ω , apart from a set of measure zero, into three disjoint families of convex
polygons: Family√ A, consisting of domains associated with the scaling limit point
(τ , τ ), τ = − 3/3, Family B, consisting of domains associated with the second
scaling limit, (41 + 72τ , 41 + 72τ ), and Family C, consisting of periodic domains
not associated with either scaling limit. For arbitrary (non-negative integer) choice
of level L(A) (resp. L(B) ) in the scaling hierarchy, Family A will consist of the do-
mains D j (LA ), j = 0, . . . , J, together with all domains which are periodic under the
return maps of levels less than L (A) , with an analogous property for Family B. The
choices of L(A) and L(B) are constrained by the selection of time interval t in the con-
(A) (B)
struction of P(r,t). For times in our target range, since t ∗ (6),t∗ (5) > 36000000,
it is sufficient to choose
L(A) = 6, L(B) = 5.
The probability distribution then has the expansion
T (D)−1
P(r,t) = τ −2 ∑ ∑ A (D)δ (Φ s+t (rD ) − Φ s (rD ), r),
D s=0
where T (D) is the return time of domain D under iteration of Φ , and D runs over
the 93 domains in the following list (the chosen tiling of Ω ):
(A)
D j (6), j = 0, 2, 4, 7, 10, 19, 21, 22, 12, 14, 15, 17,
(A)
Π j (L), L = 0, . . . , 5, j = 0, . . . , 7,
(B)
D j (5), j = 0, 3, 5, 6, 8, 10,
(B)
Π j (L), L = 0, . . . , 4, j = 0, 1,
Π misc
j , j = 0, . . . , 16.
The relatively short periodic orbits, Π misc j , were handled as in Sect. 7.7. Using
a modified bisection method, a single point with algebraic-integer coordinates was
found in each of the initial domains, and the corresponding lattice points were cal-
culated using E −1 . Here we are relying on the fact that all the orbits starting at such
points have the same shape, hence lead to the same probability distribution for dis-
placements. For each initial point, two lattice orbits, displaced in time by a specific
time interval t, were generated by explicit iteration of the map Φ using exact, com-
puter assisted algebra. The area-weighted contributions to P(x,t) and (|x| + |y|) k
were recorded at each step.
For the remaining initial domains, a more efficient approach, based on symbolic
dynamics, was employed. To illustrate the method, let us consider a specific domain
D in the tiling, with return time T , and describe the determination of D’s contribution
to the second moment (|m| + |n|) 2 . For simplicity, we assume D is a periodic
(A)
domain Π j (L); the treatment of the others follows the same pattern.
The dynamical engine of our numerical calculation is a finite Vershik codometer
with registers (ia , ja ,ta ), a = 0, . . . , L + 1. Thanks to scaling, registers 1 through
5 operate identically, with the predecessor function p(i, j,t) corresponding to the
(A)
ρ (A) (L − 1) return path of the domain D j (L). Register 0 keeps track of individual
(A)
Φ iterations, with j0 labeling return orbits of the 12 domains D j0 (0), while register
6 keeps track of the domains D jL (L) visited by D in its return orbit. The initial
codometer setting is
⎛⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎛ ⎞⎞
0 1 1 1 1 1 2
⎝⎝ 22 ⎠⎝ 22 ⎠⎝ 22 ⎠⎝ 22 ⎠⎝ 22 ⎠⎝ 17 ⎠⎝ 7 ⎠⎠
0 0 0 0 0 0 0
where we have assigned indices i = 0, 1, 2 to label the three distinct register species.
After T time steps, the codometer resets automatically to its initial sequence.
At each time step, the state of register 0 is used to update an array storing the
coordinates of 12 points, (x b , xb ), b = 0, . . . , 11, one on each branch of the lattice
orbit. Since the probability and moment calculations involve only vector displace-
7.8 Model II 181
ments between orbit points on the same branch, all 12 points of the lattice array
were conveniently set to (0, 0) at the outset.
Our numerical experiments measured three sets of data:
Moments
The moments (|x| + |y|) k/4 , k = 5, . . . , 16 are area-weighted average over all do-
mains of the recursive tiling of the square. All vector dispacements were calculated
exactly and the converted to floating point approximate real numbers before sum-
mation. This avoided the need for arbitrary precision arithmetic, but introduced an
acceptable numerical error in the summation. The case k = 0 (total area) was in-
cluded as a check on accuracy and completeness. The results are displayed in Fig.
7.18 and Table 7.2. In the latter, we measured the scaling exponents associated with
times which differ by a factor 4 in case A and by a factor 5 in case B and com-
pared them with the predicted asymptotic values from (7.18). Apart from the case
k = 1, the agreement is within a few percent, improving with increasing k. It seems
reasonable to assume that increasing the range of t would improve the convergence
considerably.
Second moment for 7200000 ≤ t ≤ 36000000
The second moment was calculated for 401 points, equally spaced on a logarithmic
scale, over the range 7200000 ≤ t ≤ 36000000. Log-log plots of the second moment
and form factor, for both the A and B scaling sequences, are shown in Fig. 7.19. This
extends to a longer time interval, with greater detail, the results of Lowenstein and
Liu (2003), shown in Fig. 7.17. The results confirm the conjectured quasiperiodicity
of the form factor of the second moment.
Fig. 7.18 Log-log plot of the moments (|x| + |y|)k t , for 17 equally spaced time intervals t in the
range 7200000 ≤ t ≤ 36000000 and k = 1, 1.25, 1.5, . . ., 4.
Probability distributions
The probability distributions P ( f ) (x,t), f = A, B, C, coarse grained to display the

results at four different scales, were calculated for 17 equally spaced t values in the
range 7200000 ≤ t ≤ 36000000. The combined probability for two different scales
at t = 36000000 are shown in Figs. 7.20 and 7.21.
Table 7.2 Scaling exponents β (A) and β (B) from the data. The measured values of β (A) in the
second column are taken from the moments for t = 9000000 and t = 36000000; those of β(B) in
the fourth column use the moments for t = 7200000 and t = 36000000. The third and fifth columns
give the predicted values of the exponents, from (7.18).
k β (A) 1 + (k − 2)/δA β (B) 1 + (k − 2)/δB

1 0.161035 0.0500157 0.225791 0.181728
1.25 0.325134 0.287512 0.397758 0.386296
1.5 0.530483 0.525008 0.591461 0.590864
1.75 0.755445 0.762504 0.793556 0.795432
2 0.988014 1 0.998148 1
2.25 1.22331 1.2375 1.20318 1.20457
2.5 1.45962 1.47499 1.40808 1.40914
2.75 1.69634 1.71249 1.61278 1.6137
3 1.93326 1.94998 1.81729 1.81827
3.25 2.17031 2.18748 2.0217 2.02284
3.5 2.40743 2.42498 2.22604 2.22741
3.75 2.64462 2.66247 2.43035 2.43198
4 2.88186 2.89997 2.63464 2.63654
Fig. 7.19 Log-log plots of the second moment r2 , r = |x| + |y| and form factor r2 /t, for the A
and B scaling sequences
7.8 Model II 183
Fig. 7.20 Coarse-grained probability distribution for the round-off error for a time interval t =
36000000. There are 600 × 600 pixels, each covering 6002 lattice sites. Colors are assigned to the
indicated ranges of the total probability P per pixel. (See the color figure at the end of the book.)
Fig. 7.21 Coarse-grained probability distribution for the round-off error for a time interval t =
36000000. There are 600 × 600 pixels, each covering 62 lattice sites. Colors are assigned to the
indicated ranges of the total probability P per pixel. (See the color figure at the end of the book.)
7.9 A conjecture
Based on the analytical and numerical results of this chapter, we can feel reasonably
confident that we understand the asymptotic long-time behavior of the renormal-
izable round-off models. Only the case of rotation number 1/5 was treated with
mathematical rigor, but the picture was sufficiently clear that we are able to for-
mulate a conjecture to cover all of the relevant models. For these, every point of
the Z2 lattice is localized, via the map E , in a polygon belonging to one of a finite
number of scaling sequences, or else to one of a finite number of periodic polygons.
Each scaling sequence is associated with an infinite sequence of lattice walks which
tend in the limit to fractal snowflakes. Moreover each scaling sequence has a unique
fractal dimension δ , and contributes, for time t tending to infinity, to the expectation
value of a homogeneous random variable χ of degree κ as
κ −2
χ t = t β (Gκ (t) + O(t −a logb t)), β = 1+ ,
δ
where a and b are positive, model dependent, constants and the form factor G κ is a
periodic function of logt which can be expressed in terms of the expectation value of
χ over the corresponding fractal snowflake. Our numerical experiments in Models I
and II provide strong evidence for the validity of the conjectured modulated power
law.
References
Akiyama S., Brunotte H., Pethö A., and Steiner W. (2008) Periodicity of certain piecewise affine
planar maps, Tsukuba Journal of Mathematics 32, 1–55.
Hardy G. H. and Wright E. M. (1979) An Introduction to the Theory of Numbers, 5th edition,
Oxford University Press, Oxford.
Lowenstein J.H., Hatjispyros S., and Vivaldi F. (1997) Quasi-periodicity, global stability and scal-
Lowenstein J.H. and Liu Sangtian (2003) Propagation of round-off error in a model of quadratic
rational rotations, Communications in Nonlinear Science and Numerical Simulation 8, 215–
237.
Lowenstein J.H. and Vivaldi F. (1998) Anomalous transport in a model of Hamiltonian roundoff,
Vivaldi, F. (1994) Periodicity and transport from round-off errors, Experimental Mathematics 3,
303–315.
Appendix A
Data Tables1
A.1 Model I Data Tables, from Kouptsov et al. (2002)
A.1.1 Generating domain
Below is a table of generating domains and subdomains, and the base domain for
the scaling sequence. The coordinates for the base domain are relative to the scaling
limit point x∞ . #V is the number of vertices of a domain (see Fig. 2.8). The map K
acts on the generating subdomain Ω j as x → C··x + d j (Ω ).
Domain #V D (m, s, b) d j (Ω )
Ω 4 D ((0, 1, 0, 1), (1, 1, −1, −1), (0, 0, τ , τ )) —
Ω0 3 D ((1, 5, 7), (−1, −1, −1), (τ , 0, −τ )) (0, −τ )
Ω1 5 D ((0, 1, 2, 5, 7), (−1, −1, −1, −1, −1), (τ , τ , τ , 0, 0)) (0, 0)
Ω2 3 D ((0, 2, 6), (−1, −1, −1), (τ , 0, 0)) (0, τ )
Ω3 2 D ((0, 2, 5, 7), (−1, 1, −1, 1), (τ , 0, 0, 0)) (0, 0)
Ω4 2 D ((0, 2, 5, 7), (−1, 1, −1, 1), (−1, τ , 0, −τ )) (0, −τ )
Ω5 2 D ((0, 1, 5, 6), (−1, 1, −1, 1), (τ , 0, 0, 0)) (0, τ )
Ω6 2 D ((1, 0, 6, 5), (−1, 1, −1, 1), (−1, 0, 0, 0)) (0, 0)
Ω7 2 D ((1, 0, 6, 5), (−1, 1, −1, 1), (τ , 0, 1, 0)) (0, −τ )
Ω8 1 D ((0, 2, 5, 7), (1, 1, 1, 1), (0, 0, 0, 0)) (0, 0)
Ω9 1 D ((0, 2, 5, 7), (1, 1, 1, 1), (0, τ , 0, −τ )) (0, −τ )
D(0) 3 D ((5, 1, 7), (−1, −1, −1), (0, 0, −1)) —
A.1.2 Level-0 scaling sequence domains
j #V D (m, s, b) T j (0) d j (0)

0 4 D ((1, 3, 5, 7), (−1, −1, −1, −1), (0, −α1,1, 0, −α1,0)) 6 (0, α1,0 )
1 2 D ((0, 3, 5, 8), (−1, 1, −1, 1), (−α2,1, −α1,1 , 0, α1,1)) 19 (0, α2,1 )
2 3 D ((1, 5, 7), (−1, −1, −1), (0, 0, −α2,1)) 10 (α1,1 , 0)
3 2 D ((0, 2, 5, 7), (−1, 1, −1, 1), (−α2,1, α2,1 , 0, −α2,1)) 27 (α3,2 , −α1,1)
4 3 D ((2, 5, 8), (−1, −1, −1), (α2,1, 0, α1,1)) 18 (0, −α1,1 )
1 In these tables, we use the abbreviations αr,s = r + sτ and αr,s

L = ω L+1 (r + sτ ).
186 Appendix A Data Tables
A.1.3 Level-0 periodic domains
j #V D (m, s, b) T̃ j (0)
0 2 D ((2, 1, 7, 6), (1, −1, 1, −1), (α2,1, α5,3 , −α2,1, α3,2 )) 25
1 1 D ((2, 1, 7, 6), (1, 1, 1, 1), (α2,1, α5,3 , −α2,1 , −α5,3)) 33
2 2 D ((2, 1, 7, 6), (1, −1, 1, −1), (α2,1, −α8,5 , −α2,1, −α5,3 )) 105
3 1 D ((2, 1, 7, 6), (1, 1, 1, 1), (α2,1, −α8,5 , −α2,1 , α8,5)) 57
4 5 D ((0, 2, 4, 6, 8), (−1, −1, −1, −1, −1), (α1,1, α4,2 , α1,0, α3,2 , α2,2 )) 24
5 5 D ((1, 3, 5, 7, 9), (−1, −1, −1, −1, −1), (0, −α1,1, 0, −α1,0 , −α1,0)) 6
A.1.4 Miscellaneous periodic domains
j #V D (m, s, b) T̃ j (Ω )
0 1 D ((0, 1, 5, 6), (1, 1, 1, 1), (0, −2 − τ , 0, 2 + τ )) 13
1 2 D ((0, 1, 5, 6), (1, −1, 1, −1), (0, −2 − τ , 0 − 3 − 2τ )) 25
2 1 D ((0, 1, 5, 6), (1, 1, 1, 1), (0, 3 + 2τ , 0, −3 − 2τ )) 17
3 1 D ((0, 1, 5, 6), (1, 1, 1, 1), (0, 1 + τ , 0, −1 − τ )) 7
4 1 D ((0, 1, 5, 6), (1, 1, 1, 1), (0, 0, 0, 0)) 1
5 2 D ((0, 1, 5, 6), (1, −1, 1, −1), (0, −1, 0, −1 − τ )) 5
6 5 D ((0, 2, 4, 6, 8), (−1, −1, −1, −1, −1), (τ , 1 + τ , 1, 1, 1 + τ )) 4
A.2 Model II Data Tables, from Kouptsov et al. (2002)
A.2.1 Generating partition
Below is a table of generating domains and subdomains, and the base domains for
the scaling sequences. The coordinates for the base domain are relative to the re-
spective scaling limit points.
Ω 4 D ((0, 1, 0, 1), (1, 1, −1, −1), (0, 0, τ , τ )) —
Ω0 3 D ((0, 1, 2), (−1, 1, −1), (τ , 0, 0)) (0, τ )
Ω1 5 D ((0, 2, 0, 2, 1), (−1, 1, 1, −1, −1), (τ , 0, 0, τ , τ )) (0, 0)
Ω2 3 D ((0, 1, 2), (1, −1, 1), (0, τ , τ )) (0, −τ )
D(A) (0) 3 D ((4, 11, 7), (−1, −1, −1), (0, 0, 209 + 362τ )) —
D(B) (0) 3 D ((11, 3, 6), (−1, −1, −1), (0, 209 + 362τ , 0)) —
A.2.2 Level-0 scaling domains, sequence A
j #V D (m, s, b) d j (0) T j (0)

0 3 D ((0, 7, 4), (−1, −1, −1), (α571,989 ), (α209,362 , 0)) (−α209,362 , α362,627 ) 6773
1 2 D ((11, 0, 4, 6), (−1, 1, −1, 1), (−α780,1351, α571,989 , 0, −α571,989 )) (0, α571,989 ) 20187
2 4 D ((5, 7, 11, 2), (−1, −1, −1, −1), (α780,1351, α209,362 , 0, α571,989 )) (−α209,362 , α571,989 ) 13546
3 2 D ((6, 5, 2, 11), (−1, 1, −1, 1), (−α571,989, α780,1351 , α571,989 , −α780,1351 )) (0, α571,989 ) 20187
4 3 D ((6, 11, 2), (−1, −1, −1), (−α571,989 , −α780,1351 , α571,989 )) (0, α571,989 ) 20319
5 2 D ((6, 2, 11, 8), (−1, 1, −1, 1), (−α571,989, α571,989 , 0, −α571,989 )) (−α989,1713 , α1142,1978 ) 26647
6 2 D ((9, 1, 4, 7), (−1, 1, −1, 1), (α780,1351, −α780,1351 , 0, α780,1351 )) (−α780,1351 , α1351,2340 ) 26801
7 3 D ((4, 9, 1), (−1, −1, −1), (0, α780,1351, −α780,1351 )) (−α780,1351 , α1351,2340 ) 26933
8 2 D ((1, 3, 4, 9), (−1, 1, −1, 1), (−α780,1351, −α780,1351 , 0, α780,1351 )) (−α1560,2702 , α1351,2340 ) 53421
9 2 D ((3, 2, 11, 8), (−1, 1, −1, 1), (−α780,1351, α1351,2340 , 0, −α1351,2340 )) (−α2340,4053 , α2702,4680 ) 53575
10 3 D ((3, 8, 11), (−1, −1, −1), (−α780,1351 , −α1351,2340 , 0)) (−α2340,4053 , α2702,4680 ) 53707
11 2 D ((4, 2, 11, 8), (−1, 1, −1, 1), (α7173,12424, α3482,6031 , 0, −α3482,6031 )) (−α6031,1044 , α6964,12062 ) 608839
12 3 D ((4, 8, 11), (−1, −1, −1), (α7173,12424, −α3482,6031 , 0)) (−α6031,1044 , α6964,12062 ) 608971
13 2 D ((6, 4, 3, 10), (−1, 1, −1, 1), (α3691,6393 , α7173,12424, −α12424,21519 , −α7173,12424)) (−α12424,21519 , α7173,12424 ) 610259
14 4 D ((6, 10, 11, 3), (−1, −1, −1, −1), (α3691,6393 , −α7173,12424, 0, −α1560,2702 )) (−α5251,9095 , α1922,3329 ) 348666
15 3 D ((2, 6, 11), (−1, −1, −1), (α3482,6031 , α3691,6393 , 0)) (−α6031,10446 , α6964,12062 ) 608971
16 2 D ((11, 0, 3, 6), (−1, 1, −1, 1), (0, −α3691,6393, −α1560,2702 , α3691,6393 )) (0, −α3691,6393 ) 402015
17 4 D ((0, 3, 8, 11), (−1, −1, −1, −1), (−α3691,6393, −α1560,2702 , −α571,989 , 0)) (−α989,1713 , −α989,1713 ) 244544
18 2 D ((1, 3, 4, 9), (−1, 1, −1, 1), (−α1560,2702, −α1560,2702 , α2131,3691 , α1560,2702 )) (−α3120,5404 , α2702,4680 ) 80613
19 4 D ((4, 8, 9, 1), (−1, −1, −1, −1), (0, −α571,989, α1560,2702 , −α1560,2702 )) (−α2131,3691 , α2131,3691 ) 53630
20 2 D ((2, 1, 11, 7), (−1, 1, −1, 1), (α1351,2340 , −α1560,2702 , 0, α1560,2702 )) (−α1560,2702 , α2702,4680 ) 53993
21 3 D ((2, 7, 11), (−1, −1, −1), (α1351,2340 , α1560,2702 , 0)) (−α2340,4053 , α2702,4680 ) 53707
22 3 D ((4, 11, 7), (−1, −1, −1), (0, 0, α780,1351)) (−α780,1351 , α1351,2340 ) 26933
187
A.2.3 Level-0 periodic domains, sequence A
188
j #V D (m, s, b) T j (0)
0 4 D ((4, 11, 8, 7), (−1, −1, −1, −1), (−α10864,18817, −α6602,11435, α21157,36645 , α5251,9095 )) 2499612
1 4 D ((0, 7, 4, 3), (−1, −1, −1, −1), (α571,989, α209,362 , 0, α1351,2340 )) 81276
2 6 D ((7, 6, 3, 2, 11, 10), (−1, −1, −1, −1, −1, −1), (−α571,989, −α1351,2340 , −α4471,7744 , α571,989 , 0, −α4262,7382 )) 174436
3 4 D ((0, 1, 4, 9), (−1, −1, −1, −1), (−α780,1351, α4262,7382 , α5042,8733 , α3691,6393 )) 644484
4 3 D ((0, 4, 7), (−1, −1, −1), (−α780,1351 , α5042,8733 , −α4262,7382 )) 1092180
5 3 D ((1, 5, 10), (−1, −1, −1), (α1351,2340 , −α2131,3691 , −α7953,13775)) 1092180
6 4 D ((0, 3, 7, 8), (−1, −1, −1, −1), (−α780,1351, −α780,1351 , −α4262,7382 , −α4262,7382 )) 563340
7 4 D ((0, 3, 8, 11), (−1, −1, −1, −1), (−α3691,6393, −α1560,2702 , −α571,989 )) 351804
A.2.4 Miscellaneous periodic domains, j > 1 0
j #V D (m, s, b) T j (0)
11 4 D ((0, 1, 6, 7), (−1, −1, −1, −1), (α2,4, α2,4 , α2,3 , α2,3 )) 30
12 12 D ((0, . . ., 11), (−1)12 , (α155,269 , α1480,2564 , α192,333 , α1098,1902 , α817,1415 , α398,689 , α1405,2433 , α80,138 , α1368,2369 , α462,800 , α743,1287 , α1162,2013 )) 13001
13 12 D ((0, . . ., 11), (−1)12 , (α2,4 , −α80,138 , −α17,29 , −α44,76, −α60,104 , −α5,9 , −α84,146 , −α2,4, −α65,113 , −α38,66 , −α22,38 , −α77,133 )) 355
14 12 D ((0, . . ., 11), (−1)12 , (α2,4 , α114,198 , α9,16 , α79,137 , α63,109 , α21,36 , α110,190 , −α2,4 , α103,178 , α33,57 , α49,85 , α91,158 )) 473
15 12 D ((0, . . ., 11), (−1)12 , (−α4,6 , −α4,6, −α1,1 , −α6,10, 0, −α5,9 , −α2,4, −α2,4 , −α5,9, 0, −α6,10 , −α1,1)) 7
16 8 D ((1, 2, 4, 5, 7, 8, 10, 11), (−1)8 , (τ , −α 3,4 , −α1,1 , −α3,5 − α3,6 , −τ , −α2,4 , 0)) 9
17 6 D ((6, 7, 10, 11, 2, 3), (−1, −1, −1, −1, −1, −1), (−α41,72, −α250,434 , −α55,95 , −α113,195 , −α266,460 , α1,2 )) 15524
Appendix A Data Tables
A.2.5 Level-0 scaling domains, sequence B
j #V D (m, s, b) T j (0) d j (0)

0 3 D ((6, 10, 3), (−1, −1, −1), (0, α571,989, α209,362 )) 19459 (α209,362 , −α362,627 )
1 2 D ((6, 2, 11, 8), (−1, 1, −1, 1), (0, −α571,989, 0, α571,989 )) 57089 (0, −α571,989 )
2 2 D ((2, 1, 9, 7), (−1, 1, −1, 1), (−α1351,2340, α780,1351 , −α780,1351 , −α780,1351 )) 86787 (−α780,1351 , 0)
3 4 D ((6, 7, 11, 2), (−1, −1, −1, −1), (0, −α780,1351, 0, −α571,989)) 76580 (−α571,989 , α209,362 )
4 2 D ((6, 3, 1, 9), (−1, 1, −1, 1), (0, α780,1351, α780,1351 , −α780,1351 )) 92983 (α780,1351 , −α1351,2340 )
5 3 D ((1, 6, 9), (1, −1, −1), (α , 0, −α )) 93115 (α780,1351 , −α1351,2340 )
6 3 D ((6, 11, 3), (1, −1, −1), (0, 0, α780,1351)) 93115 (α780,1351 , −α1351,2340 )
7 2 D ((0, 4, 6, 10), (−1, 1, −1, 1), (α780,1351, −α571,989 , 0, α571,989 )) 58245 (α989,1713 , −α1142,1978 )
8 4 D ((3, 11, 8, 5), (−1, −1, −1, −1), (α209,362, 0, α571,989 , α780,1351 )) 38918 (α571,989 , −α780,1351 )
9 2 D ((0, 5, 8, 11), (−1, 1, −1, 1), (α780,1351, α780,1351 , α571,989 , −α780,1351 )) 58245 (α989,1713 , −α1142,1978 )
10 3 D ((4, 11, 8), (−1, −1, −1), (−α571,989 , −α780,1351 , α571,989 )) 58377 (α989,1713 , −α1142,1978 )
A.2.6 Level-0 periodic domains, sequence B
j #V D (m, s, b) T j (0)
0 4 D ((3, 10, 7, 6), (−1, −1, −1, −1), (α209,362, α571,989 , α1351,2340 , 0)) 233508
1 6 D ((6, 7, 10, 11, 2, 3), (−1, −1, −1, −1, −1, −1), (0, −α780,1351, −α780,1351 , 0, −α571,989 , −α571,989 )) 76580
189
A.2.7 Incidence matrices
Sequence A
⎛ ⎞
0 0 0 0 0 0 0 0 0 0 0 4 4 4 2 5 2 1 0 0 0 0 0
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 24 24 24 12 25 14 7 0 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 4 5 4 2 4 2 1 0 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 0 0 0 0 0 0⎟⎟
⎜0
⎜ 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 1 1 1 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1⎟⎟
⎜0
⎜ 0 0 0 0 1 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 1 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 1 1 1 2 1 1 1 1 1 1 1 1 1 1 1 0⎟⎟
A=⎜
⎜0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎟⎟ (A.1)
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 34 34 34 19 34 22 13 2 1 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1 0 0⎟⎟
⎝0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 2 1⎠
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
Sequence B
⎛ ⎞
0 0 0 0 0 0 1 0 0 0 0
⎜0
⎜ 0 1 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 1 0 0 0 0 0 0 0 0 0⎟⎟
⎜0
⎜ 0 2 2 2 2 2 0 0 0 0⎟⎟
⎜0
⎜ 0 0 0 0 0 0 1 0 1 0⎟⎟
A=⎜
⎜0 1 1 1 1 1 1 1 1 1 2⎟⎟ (A.2)
⎜1
⎜ 1 1 1 1 1 1 1 1 1 1⎟⎟
⎜0
⎜ 0 0 0 1 0 0 0 0 0 0⎟⎟
⎜0
⎜ 0 2 1 2 2 3 0 0 0 0⎟⎟
⎝0 0 0 0 0 0 0 0 0 0 0⎠
0 0 0 0 0 1 0 0 0 0 0
A.3 Model III Data Tables, from Kouptsov et al. (2002) 191
A.3 Model III Data Tables, from Kouptsov et al. (2002)
A.3.1 Generating domain
Below is a table of generating domains and subdomains, and the base domain for
the scaling sequence. The coordinates for the base domain are relative to the scaling
limit point x∞ .
Ω 4 D ((0, 1, 0, 1), (1, 1, −1, −1), (−τ , −τ , τ , τ )) ——
Ω0 3 D ((1, 4, 6), (−1, −1, −1), (τ , τ , −τ )) (0, −2τ )
Ω1 6 D ((0, 1, 2, 4, 5, 6), (−1, −1, −1, −1, −1, −1), (τ , τ , τ , τ , τ , τ )) (0, 0)
Ω2 3 D ((0, 5, 2), (−1, −1, −1), (τ , τ , −τ )) (0, 2τ )
Ω3 2 D ((2, 0, 6, 5), (1, −1, 1, −1), (−τ , τ , τ , τ )) (0, 0)
Ω4 2 D ((1, 0, 5, 4), (1, −1, 1, −1), (−τ , τ , τ , −1 − τ )) (0, 2τ )
Ω5 1 D ((1, 2, 5, 6), (1, 1, 1, 1), (−τ , −τ , τ , τ )) (0, 0)
Ω6 2 D ((1, 0, 5, 4), (1, −1, 1, −1), (−τ , 1 + τ , τ , τ )) (0, 0)
Ω7 1 D ((0, 1, 4, 5), (1, 1, 1, 1), (−τ , −τ , τ , τ )) (0, 0)
Ω8 2 D ((0, 1, 4, 5), (1, −1, 1, −1), (−τ , 0, τ , τ )) (0, 0)
Ω9 1 D ((0, 1, 4, 5), (1, 1, 1, 1), (−τ , 0, τ , 0)) (0, −2τ )
Ω10 2 D ((0, 1, 4, 5), (1, −1, 1, −1), (−τ , τ , τ , 0)) (0, −2τ )
Ω11 2 D ((2, 1, 6, 5), (1, −1, 1, −1), (τ , τ , −τ , 0)) (0, −2τ )
D(−1) 3 D ((1, 4, 6), (−1, −1, −1), (0, 0, −1)) ——
A.3.2 Pre scaling level L = − 1
Return map data for the pre-scaling domain D(−1). Coordinates are relative to the
scaling limit point x ∞ = (−1 − τ , τ ).
j D (m, s, b) d j (−1) T j (−1)

0 D ((3, 4, 7, 0), (1, −1, 1, −1), (−1, −1, 1, 3 + 3τ )) (−3 − 6τ , 4 + 6τ ) 13
1 D ((0, 1, 4, 5), (1, 1, 1, 1), (3 + 3τ , −4 − 6τ , −3 − 3τ , 4 + 6τ )) (7 + 10τ , −5 − 6τ ) 22
2 D ((3, 5, 7, 1), (1, −1, 1, −1), (−1, 4 + 6τ , 1, 0)) (4 + 4τ , −4 − 6τ ) 12
3 D ((0, 1, 3, 4, 6), (−1, −1, −1, −1, −1), (−1 − 2τ , 0, −1, −1, 0)) (1, 1) 2
4 D ((0, 2, 4, 6), (1, 1, 1, 1), (−1 − 2τ , −3 − 4τ , 1 + 2τ , 3 + 4τ )) (1, 0) 13
5 D ((0, 1, 4, 7), (1, −1, 1, −1), (−1 − 2τ , 0, 1 + 2τ , −1 − 2τ )) (−1 − 2τ , 1 + 2τ ) 6
6 D ((0, 2, 4, 6), (1, −1, 1, −1), (−1 − 2τ , −3 − 4τ , 1 + 2τ , 0)) (5 + 6τ , −4 − 6τ ) 13
7 D ((1, 6, 4), (−1, −1, −1), (0, 0, 1 + 2τ )) (2 + 2τ , −1 − 2τ ) 3
8 D ((4, 1, 7), (−1, −1, −1), (−1, 0, 1)) (2, 1) 3
A.3.3 Domains D j ( L ) for even L
j D (m, s, b) d j (L) T j (L)

0 D ((4, 1, 7, 6), (−1, −1, −1, −1), (−α1,−2
L , 0, α L , 0))
1,0 (α2,−2
L ,αL )
1,−2 3 × 9L
1 D ((3, 1, 7, 6), (1, −1, 1, −1), (−α1,0, α2,2
L L , α L , 0))
1,0 (α2,2
L , −α L )
2,2 36 × 9L − 4 × 5L
177 3
2 D ((1, 2, 5, 6), (1, 1, 1, 1), (α3,4
L , −α L , −α L , α L ))
2,2 3,4 2,2 (−α0,4
L ,αL )
2,2 × 9L − 11 × 5L −
2 2
3 D ((3, 1, 7, 6), (1, −1, 1, −1), (−α1,0L , 0, α L , α L ))
1,0 2,2 (−α0,4
L ,αL )
2,2 66 × 9L − 6 × 5L
4 D ((0, 1, 4, 5), (1, 1, 1, 1), (−α1,3, α2,2 , α1,3 , −α2,2))
L L L L (−α1,4
L ,αL )
1,2 96 × 9L − 10 × 5L
5 D ((0, 1, 4, 6), (1, −1, 1, −1), (−α1,2L , 0, α L , 0))
1,2 (α1,−2
L ,αL )
2,2 75 × 9L − 6 × 5L
6 D ((0, 1, 3, 6), (−1, −1, −1, −1), (−α1,2 L , 0, −α L , 0))
1,0 (−α0,2
L ,αL )
1,2 15 × 9L
7 D ((3, 5, 7, 2), (1, −1, 1, −1), (−α1,0, −α2,2, α1,0 , −α2,2 )) (−α1,4
L L L L L ,αL )
2,2 5 × 9L − 4 × 5L
8 D ((4, 1, 6), (−1, −1, −1), (α1,2 L , 0, 0)) (α2,0
L , 0) 24 × 9L
A.3.4 Domains D j ( L ) for odd L
j D (m, s, b) d j (−1) T j (L)

0 D ((4, 1, 7, 6), (−1, −1, −1, −1), (−α1,−2L , 0, α L , 0))
1,0 (α2,−2
L ,αL )
1,−2 3 × 9L
1 D ((3, 1, 7, 4), (1, −1, 1, −1), (−α1,0, 0, α1,0
L L , −α L ))
1,0 (−α0,2
L ,αL )
2,2 36 × 9L − 4 × 5L
177 3
2 D ((0, 2, 4, 6), (1, 1, 1, 1), (α1,0
L , α L , −α L , −α L ))
1,2 1,0 1,2 (−α0,4
L ,αL )
2,2 × 9L − 11 × 5L −
2 2
3 D ((3, 1, 7, 0), (1, −1, 1, −1), (−α1,0L , −α L , α L , α L )) (−α L , α L )
2,2 1,0 1,0 0,4 2,2 66 × 9L − 6 × 5L
4 D ((0, 1, 4, 5), (1, 1, 1, 1), (−α1,3, α2,2
L L , α L , −α L ))
1,3 2,2 (−α1,4
L ,αL )
1,2 96 × 9L − 10 × 5L
5 D ((3, 1, 7, 6), (1, −1, 1, −1), (−α1,0, α2,2 , α1,0, 0))
L L L (α3,0 , 0)
L 75 × 9L − 6 × 5L
6 D ((0, 1, 3, 6), (−1, −1, −1, −1), (−α1,2 L , 0, −α L , 0)) (−α L , α L )
1,0 0,2 1,2 15 × 9L
7 D ((0, 6, 4, 1), (1, −1, 1, −1), (−α1,2, 0, α1,2
L L , 0)) (−α1,4
L ,αL )
2,2 5 × 9L − 4 × 5L
8 D ((4, 1, 6), (−1, −1, −1), (α1,2 , 0, 0))
L (α2,0 , 0)
L 24 × 9L
A.3.5 Domains Π j ( L ) for all L
L j D (m, s, b) T̃ j (L)
−1 0 D ((0, . . ., 7), (−1)8 , (−1 − 2τ , 0, −2 − 2τ , −1, −1, −2 − 2τ , 0, −1 − 2τ )) 2
1 D ((0, 1, 4, 7), (1, −1, 1, −1), (−1 − 2τ , 0, 1 + 2τ , −1 − 2τ )) 8
≥0 0 D ((0, . . ., 7), (−1)8 , (α1,0
L , 0, α L , −α L , −α L , α L , 0, α L ))
0,2 1,−2 1,−2 0,2 1,0 3 · 9L
1 D ((0, . . ., 7), (−1)8 , (−α1,2
L , 0, −α L , −α L , −α L , −α L , 0, −α L ))
2,2 1,0 1,0 2,2 1,2 18 · 9L
2 D ((0, . . ., 7), (−1)8 , (α5,6
L , 0, α L , α L , −α L , α L , α L , α L ))
2,4 3,6 1,0 4,6 2,2 1,0 27 · 9L
even 3 D ((3, 0, 7, 4), (1, −1, 1, −1), (−α1,0, −α2,4 , α1,0 , −α1,0))
L L L L 84 · 9 − 4 · 5L
L
odd 3 D ((0, 1, 4, 5), (1, −1, 1, −1), (−α1,2, 0, α1,2 , −α3,4))

L L L 84 · 9L − 4 · 5L
A.3 Model III Data Tables, from Kouptsov et al. (2002) 193
A.3.6 Tiling data
A.3.7 Section of the discontinuity set

Generating domains
B = Ω78910
(−1)
Ω 7 = K 7 D4
(−1) (−1) (−1) (−1) (−1) (−1)
Ω8 = K 11 D2 K 21 D1 K 12 D6 K 8 D46 K 17 D1 K 7 D210
(−1)
Ω9 = K 11 D1
(−1) (−1) (−1) (−1)
Ω10 = K 3 D012 K 1 3D1 K 4 D645 K 11 D40
Pre-scaling domains
(−1) (0) (0)
D0 = ρ−1
4
D723 ρ−1
7
D5
(−1) (0)
D1 = ρ−1
4
D4
(−1) (0) (0) (0) (0) (0) (0) (0)
D2 = ρ−1
5
D5 ρ−1
7
D2 ρ−1
14
D5 ρ−1
13
D4 ρ−1
10
D3 ρ−1
9
D2 ρ−1
4
D1
(−1) (0)
D4 = ρ−1
8
D2
(−1) (−1)
D5 = Π1
(−1) (0)
D6 = ρ−1 D14723
Scaling sequence, L even

(L) (L+1) (L+1) (L+1)
D1 = ρL52 D5 ρL20 D2 ρL7 D54321
(L) (L+1)
D2 = ρL11 D2
(L) (L+1) (L+1) (L+1)
D3 = ρL16 D7 ρL34 D2 ρL36 D345
(L) (L+1)
D4 = ρL52 D4
(L) (L+1) (L+1) (L+1) (L+1) (L+1)
D5 = ρL13 D1 ρL63 D4 ρL27 D7 ρL45 D2 ρL47 D345
(L+1) (L+1)
ρL15 D2 ρL2 D54321
(L) (L+1) (L+1) (L+1) (L+1)
D7 = Π3 (L) ρL13 D2 ρL18 D1 ρL68 D4 ρL32 D7
(L+1) (L+1)
ρL50 D2 ρL52 D3
Scaling sequence, L odd

(L) (L+1) (L+1) (L+1) (L+1) (L+1)
(L+1) (L+1) (L+1) (L+1)
ρL45 D2 ρL13 D14723 ρL29 D4 ρL27 D5
(L) (L+1)
D2 = ρL41 D2
(L) (L+1) (L+1) (L+1)
D3 = ρL32 D5 ρL34 D4 ρL18 D327
(L) (L+1)
D4 = ρL18 D4
(L) (L+1) (L+1) (L+1) (L+1) (L+1)
(L+1) (L+1) (L+1) (L+1)
ρL37 D2 ρL21 D5 ρL23 D4 ρL7 D32741
(L) (L+1) (L+1)
D7 = Π3 (L) ρL39 D2 ρL D14723
A.4 Cubic Model Data Tables, from Lowenstein et al. (2004)
Table A.1 Data for dressed subdomains. Each Δ (i) is generated from its equivalent catalogue
member Δ [h(i)] by application of the transformation Rκ ◦ C−ν ◦ T−d ◦ Sω −1 . Algebraic numbers
a + bλ + cλ 2 are represented here as z[a, b, c].
Δ (i) κ ν d ω
Δ (0) 1 8 (z[1, −1, 0], z[−1, 0, 0]) z[−3, 5, 6]
Δ (1) 0 13 (z[2, 4, −3], z[2, −3, 1]) z[−5, 6, 8]
Δ (2) 0 8 (0, z[5, −3, 0]) z[−14, 20, 25]
Δ (3) 1 13 (z[−9, −2, 4], z[−4, −1, 2]) z[−17, 25, 31]
Δ (4) 1 3 (z[−1, −1, 1], z[−8, 1, 2]) z[−11, 16, 20]
Δ (5) 0 4 (z[16, 4, −7], z[8, 3, −4]) z[−14, 20, 25]
Δ (6) 0 3 (z[2, 1, −1], z[2, 1, −1]) z[−2, 3, 4]
Δ (7) 0 9 (z[−11, −3, 5], z[−5, −1, 2]) z[−2, 3, 4]
Δ (8) 1 4 (z[8, 3, −4], z[−3, 4, 5]) z[−2, 3, 4]
Δ (9) 0 0 (0, 0) z[0, 1, 1]
Δ (10) 1 7 (z[1, 1, −1], z[1, −1, 0]) z[−2, 1, 2]
Δ (11) 0 0 (z[−2, −1, 1], 0) z[0, 1, 1]
Δ (12) 0 0 (0, 0) z[−2, 3, 4]
Δ (13) 1 7 (z[6, 2, −3], z[3, 0, −1]) z[−3, 4, 5]
Δ (14) 0 0 (z[−5, −1, 2], 0) z[0, 1, 1]
A.4 Cubic Model Data Tables, from Lowenstein et al. (2004) 195
[n]
Table A.2 Data for partition of prototype domain D[n] . The return map for the atom D j is
parametrized by ν and d.
Atom m b ν d
[0]
D0 (9, 11, 1, 3, 7) (1, λ , 0, −1, 0) 2 (1, 1)
[0]
D1 (1, 11, 7) (0, 1, 0) 0 (1, 0)
[0]
D2 (4, 1, 10) (0, 1, 0) 3 (1, 1)
[0]
D3 (9, 4, 1) (1, −λ , 0) 3 (2, λ )
[1]
D0 (3, 1, 11, 7) (−2 − λ + λ 2 , 0, 1, 0) 10 (−2 − λ + λ 2 , −1)
[1]
D1 (11, 1, 3, 7) (2 + λ − λ , 0, −5 − λ + 2λ )
2 2
10 (−5 − λ + 2λ 2 , −2 − λ + λ 2 )
[1]
D2 (10, 13, 1, 5, 7) (5 + λ − 2λ , −3 + λ , 0, −2 + λ , 0)
2 2
12 (−2 + λ , 3 − λ 2 )
[1]
D3 (4, 1, 12, 7) (−3 − 2λ + 2λ , 0, 2 − λ , 0)
2
0 (0, −2 + λ )
[1]
D4 (10, 1, 4, 7) (2 + λ − λ 2 , 0, −2 − λ + λ 2 , −2 + λ ) 0 (−2 + λ , 0)
[1]
D5 (10, 6, 4, 0) (2 + λ − λ 2 , −2λ + λ 2 , −2 − λ + λ 2 , 2 − λ ) 2 (−1 − λ + λ 2 , 2λ − λ 2 )
[1]
D6 (4, 12, 10) (−2 − λ + λ , 7 − 2λ , 2 + λ − λ )
2 2 2
0 (3 + 2λ − 2λ 2 , 3 + 2λ − 2λ 2 )
[1]
D7 (10, 6, 1) (5 + λ − 2λ , 3 − λ )
2 2
8 (0, −5 − λ + 2λ 2 )
[1]
D8 (3, 1, 11, 7) (−8 − λ + 3λ , 0, 3 + 2λ − 2λ )
2 2
10 (−8 − λ + 3λ 2 , −3 − 2λ + 2λ 2 )
[1]
D9 (13, 10, 5) (2λ − λ 2 , 2 + λ − λ 2 , −7 + 2λ 2 ) 12 (−7 + 2λ 2 , −2λ + λ 2 )
[1]
D10 (1, 10, 7) (0, 8 + λ − 3λ 2 , 0) 12 (3 + 2λ − 2λ 2 , 6 + 2λ − 3λ 2 )
[2]
D0 (9, 11, 13, 1, 5, 7) (1, −λ + λ , 1, 0, 1 − λ , 0)
2
4 (−λ + λ 2 , −1 + λ )
[2]
D1 (4, 9, 0) (λ − λ , 1, 1 − 2λ + λ )
2 2
2 (1, −1 + 3λ − λ 2 )
[2]
D2 (1, 9, 7) (0, 1, −1 + 2λ − λ ) 2
0 (−1 + 2λ − λ 2 , 0)
[2]
D3 (9, 12, 1, 4, 7) (1 − 2λ + λ , −1 + λ , 0, −1 + 2λ − λ , 0)
2 2
0 (2λ − λ 2 , 0)
[2]
D4 (3, 1, 11, 7) (2λ + λ 2 , 0, 1 − 2λ + λ 2 , 0) 2 (1, 2λ − λ 2 )
[2]
D5 (1, 9, 7) (0, 2 − λ , 0) 0 (−1 + λ , 0)
[2]
D6 (2, 12, 7) (−1 + 2λ − λ , −1 + λ , 0)
2
12 (1 + λ − λ 2 , 1 − λ )
[2]
D7 (10, 2, 5, 7) (2λ − λ , −2 + λ , 3 − λ , 0)
2 2
10 (−1 − λ + λ 2 , 1 + λ − λ 2 )
[2]
D8 (1, 11, 7) (0, −λ + λ , −2 − λ + λ )
2 2
6 (−λ + λ 2 , 2 + λ − λ 2 )
[2]
D9 (11, 6, 1) (−λ + λ 2 , −1, 1 + 3λ − 2λ 2 ) 10 (−2 − 3λ + 2λ 2 , 1 + 4λ − 3λ 2 )
[2]
D10 (8, 6, 0) (−1 − 3λ + 2λ 2 , −1, 2 + λ − λ 2 ) 12 (3 − 4λ + λ 2 , 3λ − 2λ 2 )
[2]
D11 (2, 12, 7) (−2 + λ , −3 + λ , 0) 2
12 (2 + λ − λ 2 , 3 − λ 2 )
[3]
D0 (9, 1, 3, 7) (1, 0, 1 − λ , 0) 8 (0, −1)
[3]
D1 (10, 13, 1, 4, 7) (−1 + λ , 2 + λ − λ 2 , 0, 1 + λ − λ 2 , 0) 12 (0, −2 − λ + λ 2 )
[3]
D2 (10, 6, 1) (−1 + λ , −2 − λ + λ 2 , 0) 0 (−3 + λ 2 , 0)
[3]
D3 (11, 1, 5, 7) (−1 − λ + λ 2 , 0, 3 − λ 2 , 0) 10 (−1 − λ + λ 2 , 1 + λ − λ 2 )
[3]
D4 (9, 12, 1, 4, 7) (2 − λ , −3 + λ , 0, −2 + λ , 0)
2
0 (−1 + λ , 0)
[3]
D5 (3, 1, 11, 7) (−3 − 2λ + 2λ 2 , 0, 2 − λ , 0) 2 (1, 3 + 2λ − 2λ 2 )
[3]
D6 (1, 9, 7) (0, 8 + λ − 3λ 2 , 0) 0 (−7 − λ + 3λ 2 , 0)
[3]
D7 (2, 12, 7) (−2 + λ , −3 + λ , 0) 2
12 (2 + λ − λ 2 , 3 − λ 2 )
[3]
D8 (10, 2, 5, 7) (3 + 2λ − 2λ 2 , −8 − λ + 3λ 2 , 14 + 3λ − 6λ 2 , 0) 10 (−10 − 3λ + 5λ 2 , 6 + 2λ − 3λ 2 )
[3]
D9 (2, 12, 7) (−8 − λ + 3λ 2 , −14 − 3λ + 6λ 2 , 0) 12 (7 + 2λ − 3λ 2 , 14 + 3λ − 6λ 2 )
[n]
Table A.3 Generating cells Π j constructed from the prototype Π of (2.11) by a transformation
[n]
C −ν ◦ T−d ◦ Sω −1 . The cell is mapped into itself after T̃ j iterations of ρ [n] . Algebraic numbers
a + bλ + cλ 2 are represented here as z[a, b, c].
[n] [n]
Πj ν d ω T̃ j
[0]
Π0 1 (z[5, 1, −2], z[1, 0, 0]) z[−2, 3, 4] 3
[0]
Π1 0 (z[13, 5, −7], z[8, 1, −3]) z[−20, 29, 36] 18
[0]
Π2 0 (z[−74, −17, 32], z[8, 1, −3]) z[−409, 591, 737] 2160
[0]
Π3 0 (z[−9, 1, 2], z[−1, −3, 2]) z[−45, 65, 81] 51
[0]
Π4 1 (z[−75, −25, 37], z[−28, −2, 10]) z[−409, 591, 737] 264
[0]
Π5 1 (z[−380, −99, 172], z[−28, −2, 10]) z[−1328, 1919, 2393] 861
[0]
Π6 1 (z[−14, −3, 6], z[−6, 2, 1]) z[−45, 65, 81] 117
[0]
Π7 1 (z[60, 10, −24], z[33, 11, −16]) z[−182, 263, 328] 111
[0]
Π8 1 (z[−5, −8, 6], z[−6, 2, 1]) z[−25, 36, 45] 33
[0]
Π9 1 (z[−49, −16, 24], z[−15, −2, 6]) z[−81, 117, 146] 63
[0]
Π10 1 (z[1, 5, −3], z[3, −3, 1]) z[−6, 9, 11] 2
[0]
Π11 1 (z[−11, −19, 14], z[4, −9, 4]) z[−126, 182, 227] 396
[0]
Π12 0 (z[9, −7, 1], z[2, 8, −5]) z[−56, 81, 101] 40
[0]
Π13 0 (z[−52, −20, 27], z[2, 8, −5]) z[−409, 591, 737] 272
[0]
Π14 0 (z[−4, −16, 10], z[−7, 13, −5]) z[−126, 182, 227] 76
[0]
Π15 1 (z[−7, 4, 0], z[0, −5, 3]) z[−14, 20, 25] 12
[0]
Π16 0 (z[−9, 12, −4], z[2, −10, 5]) z[−14, 20, 25] 24
[0]
Π17 0 (z[10, 14, −11], z[8, −8, 2]) z[−101, 146, 182] 84
[0]
Π18 0 (z[14, 28, −20], z[8, −8, 2]) z[−510, 737, 919] 368
[0]
Π19 0 (z[184, 67, −94], z[8, −8, 2]) z[−1656, 2393, 2984] 1183
[0]
Π20 0 (z[−25, 10, 2], z[8, −8, 2]) z[−227, 328, 409] 149
[0]
Π21 1 (z[13, −9, 1], z[6, 6, −5]) z[−56, 81, 101] 73
[0]
Π22 1 (z[−5, −8, 6], z[−3, 2, 0]) z[−25, 36, 45] 27
[0]
Π23 1 (z[43, 14, −21], z[−3, 11, −5]) z[−510, 737, 919] 3072
[0]
Π24 0 (z[6, 9, −7], z[4, −4, 1]) z[−56, 81, 101] 165
[0]
Π25 0 (z[−3, 14, −7], z[4, −13, 6]) z[−31, 45, 56] 46
[0]
Π26 0 (z[−8, −3, 4], z[−2, 3, −1]) z[−25, 36, 45] 20
[0]
Π27 0 (z[−1, −5, 3], z[−4, 6, −2]) z[−8, 11, 14] 6
Table A.4 Continuation of Table A.3.
[n] [n]
Πj ν d ω T̃ j
[1]
Π0 1 (z[−9, −2, 4], z[−4, −1, 2]) z[−2, 3, 3] 1
[1]
Π1 0 (z[−30, −14, 17], z[−25, −4, 10]) z[−81, 117, 146] 77
[1]
Π2 1 (z[−12, −4, 6], z[−6, 0, 2]) z[−9, 13, 16] 5
[1]
Π3 1 (z[−20, −5, 9], z[−9, −2, 4]) z[−3, 4, 6] 1
[1]
Π4 1 (z[42, 11, −19], z[−3, 0, 1]) z[−65, 94, 117] 37
[1]
Π5 0 (z[22, 4, −9], z[3, 2, −2]) z[−7, 10, 13] 1
[1]
Π6 0 (z[28, 6, −12], z[3, 2, −2]) z[−36, 52, 65] 3
[1]
Π7 1 (z[−31, −8, 14], z[−14, −3, 6]) z[−16, 23, 29] 1
[1]
Π8 1 (z[−32, −11, 16], z[0, 2, −1]) z[−81, 117, 146] 2
[2]
Π0 1 (z[2, 1, −1], z[1, 0, 0]) z[−1, 2, 2] 1
[2]
Π1 0 (z[16, 9, −10], z[7, −2, −1]) z[−20, 29, 36] 6
[2]
Π2 1 (z[−4, −3, 3], z[−3, 4, −1]) z[−25, 36, 45] 32
[2]
Π3 0 (z[−4, 2, 0], z[1, −4, 2]) z[−3, 4, 5] 2
[2]
Π4 0 (z[−8, −5, 5], z[−5, 1, 1]) z[−25, 36, 45] 2
[2]
Π5 1 (z[7, −2, −1], z[2, 3, −2]) z[−11, 16, 20] 4
[3]
Π0 1 (z[2, 1, −1], z[1, 0, 0]) z[−1, 2, 2] 1
[3]
Π1 0 (z[15, 4, −7], z[8, 1, −3]) z[−20, 29, 36] 34
[3]
Π2 0 (z[46, 10, −20], z[17, 5, −8]) z[−36, 52, 65] 2
[3]
Π3 0 (z[−198, −51, 89], z[17, 5, −8]) z[−263, 380, 474] 22
[3]
Π4 0 (z[−50, −16, 24], z[−5, 1, 1]) z[−81, 117, 146] 8
[3]
Π5 0 (z[−12, −1, 4], z[−1, −3, 2]) z[−9, 13, 16] 2
[3]
Π6 1 (z[−44, −17, 23], z[−20, −1, 7] z[−81, 117, 146] 64
[3]
Π7 1 (z[11, 3, −5], z[1, 0, 0]) z[−4, 6, 7] 2
[3]
Π8 1 (z[5, 1, −2], z[1, 0, 0]) z[−2, 3, 4] 2
[3]
Π9 1 (z[−14, −3, 6], z[1, 0, 0]) z[−9, 13, 16] 2
Table A.5 Incidence matrices of the π /7 model.
⎛0 3 0 0 0 0 0 0 0 0 0 0⎞
⎛0 0 0 0 0 0 0 0 0 0 0⎞
⎜0 0
⎜ 0 0 0 0 0 0 0 0 0 0⎟ ⎟
⎜3
⎜ 9 21 27 27 33 27 63 15 21 21 ⎟
⎟
⎜ ⎟ ⎜ ⎟
⎜0 0
⎜ 0 0 0 0 0 0 0 0 0 0⎟ ⎟ ⎜0
⎜ 6 17 22 22 27 22 54 12 17 17 ⎟
⎟
⎜ ⎟ ⎜ ⎟
⎜0 2
⎜ 1 0 0 0 0 0 0 0 0 0⎟ ⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0⎟ ⎟
⎜ ⎟ ⎜ ⎟
⎜0 2 1 0 0 0 0 0 0 0 0 0⎟ ⎜0 0 0 0 0 0 0 0 0 0 0⎟
⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
A(0) = ⎜ 2 0 0 2 1 0 3 0 4 10 9 0 ⎟ A(1) = ⎜ 0 0 0 0 0 0 0 0 0 0 0⎟
⎜ ⎟ ⎜ ⎟
⎜1 1
⎜ 1 1 1 1 1 1 1 1 1 1⎟ ⎟ ⎜0
⎜ 0 0 0 0 0 0 0 0 0 0⎟ ⎟
⎜ ⎟ ⎜ ⎟
⎜0 0
⎜ 0 0 0 0 0 0 0 0 0 0⎟ ⎟
⎜1
⎜ 2 4 5 5 6 5 11 3 4 4 ⎟ ⎟
⎜ ⎟ ⎜ ⎟
⎜0 0
⎜ 0 0 0 0 0 0 3 6 6 0⎟ ⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0⎟ ⎟
⎜ ⎟ ⎜ ⎟
⎝0 0 0 1 0 0 2 1 1 7 6 4⎠ ⎝0 0 0 0 0 0 0 0 0 0 0⎠
0 0 0 1 0 0 2 1 1 7 6 4 00 0 0 0 0 0 0 0 0 0
⎛ ⎞
⎛ 6 24 0 0 0 0 0 0 0 0 0 0 0
57 72 72 87 72 168 42 57 57 ⎞ ⎜2
⎜0 0 ⎜ 8 19 24 24 29 24 56 14 19 19 ⎟
⎟
⎜ 0 0 0 0 0 0 0 0 0⎟ ⎟ ⎜
⎜0
⎟
⎜ ⎟ ⎜ 0 0 0 0 0 0 0 0 0 0⎟ ⎟
⎜1 2 4 5 5 6 5 11 3 4 4 ⎟ ⎜ ⎟
⎜ ⎟ ⎜0
⎜ ⎟ ⎜ 3 8 10 10 12 10 24 6 8 8 ⎟ ⎟
⎜1 5 12 15 15 18 15 35 9 12 12 ⎟ ⎜ ⎟
⎜ ⎟ ⎜1 3 7 9 9 11 9 21 5 7 7 ⎟
⎜ ⎟ ⎜ ⎟
⎜1 5 12 15 15 18 15 35 9 12 12 ⎟ ⎜ ⎟
⎜ ⎟ ⎜0 0 0 0 0 0 0 0 0 0 0⎟
⎜ ⎟
A(2) = ⎜ 1 3 7 9 9 11 9 21 5 7 7 ⎟ A(3) = ⎜
⎜1
⎟
⎜
⎜0 0
⎟ ⎜ 2 4 5 5 6 5 11 3 4 4 ⎟ ⎟
⎜ 0 0 0 0 0 0 0 0 0⎟ ⎟ ⎜
⎜0
⎟
⎜ ⎟ ⎜ 0 0 0 0 0 0 0 0 0 0⎟ ⎟
⎜0 0 0 0 0 0 0 0 0 0 0⎟ ⎜ ⎟
⎜ ⎟ ⎜0
⎜ ⎟ ⎜ 0 0 0 0 0 0 0 0 0 0⎟ ⎟
⎜0 0 0 0 0 0 0 0 0 0 0⎟ ⎟ ⎜ ⎟
⎜ ⎜0 0 0 0 0 0 0 0 0 0 0⎟
⎜ ⎟ ⎜ ⎟
⎝0 0 0 0 0 0 0 0 0 0 0⎠ ⎜ ⎟
⎝0 0 0 0 0 0 0 0 0 0 0⎠
0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0
⎛ ⎞
1 2 4 5 5 6 5 11 3 4 4
⎛ ⎞ ⎜
4 16 38 48 48 58 48 112 28 38 38 ⎜0 0 0 0 0 0 0 0 0 0 0⎟
⎟
⎜ ⎟ ⎜ ⎟
⎜0 0⎟ ⎜0 0 0 0 0 0 0 0 0 0 0⎟
⎜
0 0 0 0 0 0 0 0 0
⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎜0 0 0 0 0 0 0 0 0 0 0⎟ ⎜0 0 0 0 0 0 0 0 0 0 0⎟
⎜ ⎟ ⎜ ⎟
⎜3 ⎟ ⎜ ⎟
⎜ 9 21 27 27 33 27 63 15 21 21 ⎟ ⎜0 0 0 0 0 0 0 0 0 0 0⎟
⎜ ⎟ ⎜ ⎟
⎜ ⎟
⎜0
⎜ 5 13 16 16 19 16 38 10 13 13 ⎟
⎟ ⎜0 0 0 0 0 0 0 0 0 0 0⎟
A(4) = ⎜ ⎟ A(5) = ⎜
⎜
⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0⎟ ⎟ ⎜0 0 0 0 0 0 0 0 0 0 0⎟
⎟
⎜ ⎟ ⎜ ⎟
⎜0 0⎟ ⎜0 0 0 0 0 0 0 0 0 0 0⎟
⎜
0 0 0 0 0 0 0 0 0
⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎜1 2 4 5 5 6 5 11 3 4 4⎟ ⎜1 1 1 1 1 1 1 1 1 1 1⎟
⎜ ⎟ ⎜ ⎟
⎜ ⎟
⎜0
⎝ 0 0 0 0 0 0 0 0 0 0⎟ ⎠ ⎜0
⎜ 1 2 2 2 2 2 4 2 2 2⎟
⎟
⎜ ⎟
0 0 0 0 0 0 0 0 0 0 0 ⎝0 0 1 2 2 3 2 6 0 1 1⎠
0 0 0 0 0 0 0 0 0 0 0
Table A.6 Continuation of Table A.5.
⎛ ⎞ ⎛ ⎞
1 2 2 4 16 8 11 24 22 55 4 16 38 48 48 58 48 112 28 38 38
⎜0 0 0 0 0 0 0 0 0 0 ⎟ ⎜0 0 0 0 0 0 0 0 0 0 0 ⎟
⎜ ⎟ ⎜ ⎟
⎜0 0 0 0 0 0 0 0 0 0 ⎟ ⎜0 0 0 0 0 0 0 0 0 0 0 ⎟
⎜ ⎟ ⎜ ⎟
⎜0 0 0 0 0 0 0 0 0 0 ⎟ ⎜3 9 21 27 27 33 27 63 15 21 21 ⎟
⎜ ⎟ ⎜ ⎟
⎜0 0 0 0 0 0 0 0 0 0 ⎟ ⎜0 5 13 16 16 19 16 38 10 13 13 ⎟
A(6) = ⎜ ⎟ A(7) = ⎜ ⎟
⎜0 0 0 3 12 6 6 18 12 30 ⎟ ⎜0 0 0 0 0 0 0 0 0 0 0 ⎟
⎜1
⎜ 1 1 1 1 1 1 1 1 1 ⎟
⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0 ⎟⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 ⎟
⎟
⎜1
⎜ 2 4 5 5 6 5 11 3 4 4 ⎟⎟
⎝0 1 0 0 3 1 4 5 9 24 ⎠ ⎝0 0 0 0 0 0 0 0 0 0 0 ⎠
0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
⎛ ⎞
3 11 25 31 31 37 31 71 19 25 25
⎜0 4 10 12 12 14 12 28 8 10 10 ⎟
⎜ ⎟
⎜1 3 7 9 9 11 9 21 5 7 7⎟
⎜ ⎟
⎜0 0 0 0 0 0 0 0 0 0 0⎟
⎜ ⎟
⎜0 0 0 0 0 0 0 0 0 0 0⎟
A(8) = ⎜ ⎟
⎜0 0 0 0 0 0 0 0 0 0 0⎟
⎜ ⎟
⎜0 0 0 0 0 0 0 0 0 0 0⎟
⎜0
⎜ 0 0 0 0 0 0 0 0 0 0⎟⎟
⎝0 0 0 0 0 0 0 0 0 0 0⎠
0 0 0 0 0 0 0 0 0 0 0
⎛ ⎞
8 23 44 23 113 53 98 173 203 518
⎜8 23 44 23 113 53 98 173 203 518 ⎟
⎜
A(9) = ⎝ ⎟
2 7 16 5 29 13 28 45 59 152 ⎠
2 7 16 5 29 13 28 45 59 152
⎛ ⎞
15 77 183 227 227 271 227 527 139 183 183
⎜ 15 77 183 227 227 271 227 527 139 183 183 ⎟
A(10) = ⎜
⎝ 5
⎟
25 59 73 73 87 73 169 45 59 59 ⎠
5 25 59 73 73 87 73 169 45 59 59
⎛ ⎞
21 69 150 93 453 213 357 693 741 1893
⎜ 24 79 172 105 513 241 406 785 843 2154 ⎟
⎜
A(11) = ⎝ ⎟
10 33 72 43 211 99 168 323 349 892 ⎠
10 33 72 43 211 99 168 323 349 892
⎛ ⎞
111 519 1227 1527 1527 1827 1527 3543 927 1227 1227
⎜ 125 585 1383 1721 1721 2059 1721 3993 1045 1383 1383 ⎟
⎜
A(12) = ⎝ ⎟
51 239 565 703 703 841 703 1631 427 565 565 ⎠
51 239 565 703 703 841 703 1631 427 565 565
⎛ ⎞
18 60 132 72 360 168 294 552 612 1566
⎜ 16 53 116 67 331 155 266 507 553 1414 ⎟
A(13) = ⎜
⎝ 7
⎟
23 50 31 151 71 119 231 247 631 ⎠
7 23 50 31 151 71 119 231 247 631
⎛ ⎞
84 396 936 1164 1164 1392 1164 2700 708 936 936
⎜ 79 371 877 1091 1091 1305 1091 2531 663 877 877 ⎟
⎜
A(14) = ⎝ ⎟
37 173 409 509 509 609 509 1181 309 409 409 ⎠
37 173 409 509 509 609 509 1181 309 409 409
A.5 Inadmissibility Tables for Models II and III
Model II, Sector A

( jk ,tk ) (7, 1) (7, 2) (10, 4) (7, 0) (22, 0), (22, 2) (10, 2) (19, 2) (21, 2)
( jk+1 ,tk+1 ) (7, 2) (7, 2), (10, 4) (7, 1) (22, 2) (7, 0), (22, 0) (10, 2) (19, 2) (21, 2)
Model II, Sector B

( jk ,tk ) ( jk+1 ,tk+1 )
(0, 0) (6, 3)
(3, 0), (3, 1) (3, 3)
(3, 3), (3, 4) (3, 1)
(3, 0) , (3, 2), (3, 3) (3, 1), (5, 1), (6, 1)
(3, 1) , (3, 2) , (3, 4) (3, 3), (5, 5), (6, 6)
(5, 3), (5, 4), (5, 6) (3, 4) , (5, 6), (8, 1), (10, 1)
(5, 0), (5, 1), (5, 5), (5, 6) (3, 4), (5, 6) , (6, 7), (10, 2)
(6, 0), (6, 1), (6, 6), (6, 7) (0, 0), (3, 0), (5, 0), (6, 0)
(6, 0) , (6, 2), (6, 3), (6, 4) (3, 0) , (6, 0) , (8, 0)
(8, 0) (3, 2) , (5, 4), (6, 4)
(8, 1) (3, 2), (6, 2)
(10, 1), (10, 2) (5, 3)
Model III
( jk ,tk ) ( jk+1 ,tk+1 )
(0, 1), (1, 7) , (1, 10), (1, 15), (1, 17),
(0, 0), (0, 1)
(2, 7), (2, 15), (2, 17), (2, 25), (2, 32)
(0, 1) , (1, 1), (1, 3), (1, 11), (1, 8),
(0, 0) , (0, 1)
(2, 18), (2, 26), (2, 1), (2, 8), (2, 16)
(0, 0) (1, 12), (2, 27), (2, 17)
(1, 16), (1, 17) (1, 2)
(1, 1), (1, 2) (1, 16)
(1,t), t = 2 , 3, 4, 9 − 12 (1, 16), (1, 9), (1, 14), (2, 24), (2, 31), (2, 14)
(1,t), t = 0, 7 , 8, 9 , 14, 15, 16 (1, 2), (1, 4), (1, 9) , (2, 19), (2, 2), (2, 9)
(2,t), t = 18, 19, 24 − 27 (0, 0)
(2,t), t = 0, 1, 2, 31, 32 (0, 0) , (1, 0), (2, 0)
(2,t), t = 0 , 7 − 9, 14 − 17 (0, 0) , (1, 0), (2, 0)
References
Appendix B
The Codometer
In this appendix we describe in concrete terms the software device, the codometer,
which we use in this book to implement the symbolic dynamics appropriate to recur-
sive tiling. Our intent here is to give the reader an intuitive feeling for the relatively
simple and concrete machinery which we have found so useful in our computer
“experiments”.
The general setup is shown in the figure. Our codometer consists of a linear array
of registers, labeled k = 0, 1, 2, . . . K, each assigned a pair of integers: a cycle j k ∈ Jk
and a phase tk ∈ Tk ( jk ) = {0, 1, . . ., ν ( jk )}. We represent the state of the codometer,
i.e. the set of current register values, by the expression

j0 j1 j2 j
··· K .
t0 t1 t2 tK
The set Jk of possible cycles is assumed to be finite, but otherwise arbitrary. For
each cycle jk , the lower and upper limits for the corresponding phase t k are stored
in memory. Also stored in memory are the values, for k > 0, of the predecessor
functions pk ( jk ,tk ), which are essential both to set and to update the state of the
codometer.
The codometer’s inner workings only permit states for which the cycle and phase
values neighboring registers satisfy the constraint
jk−1 = pk ( jk ,tk ).
202 Appendix B The Codometer
Setting the state is initiated by entering j K and tK in the highest register. The
codometer then supplies j K−1 using the predecessor constraint, while the user
chooses tK−1 from the set TK−1 ( jK−1 ). This process continues until all the regis-
ters are set.
We now describe the updating process. If t 0 is less than its maximum value,
we simply increment it to t 0 + 1. If, on the other hand, the first r phases are all
maximal, but the r + 1st is sub-maximal, then we increment t r and reset the registers
r − 1, r − 2, . . . , 0, in that order, to obtain

j0 j1 j jr jr+1 j
· · · r−1 ··· K .
0 0 0 tr+1 tr+1 tK
where

jr−1 = pr ( jr ,tr + 1), jk−1 = pk ( jk , 0), k = r − 1, r − 2, . . ., 1.
This updating rule uniquely specifies the stepwise incrementation of states, with the
exception of the possibility that all t k are maximal. The device may include such
a rule in its program, or else restrict use to applications where the maximal state
never arises (easily arranged in practice by choosing the number of registers, or the
maximum value of the last register, large enough).
As a simple application, let us program a 4-register codometer as a “perpet-
ual” Gregorian calendar, whose registers k = 0, 1, 2, 3 are designed to keep track
of days, months, years, and centuries, respectively. Following the usual convention,
we will reset the registers for days and months to 1, and the registers for years and
centuries to 0. This is a trivial modification easily incorporated into the codome-
ter’s program. For register 0, which keeps track of the days of the month, we have
j0 = 28, 29, 30, 31, and t 0 runs from 1 to j 0 , respectively. For register 1 (months
of the year), j 1 = 0 (common year), 1 (leap year), with t 1 running from 1 to 12 in
both cases. For register 2 (years of the century), we have j 2 = 0, 1 corresponding
to whether the century phase t 3 is divisible by 4 or not, with t 2 running from 0 to
99. Finally, register 3 simply counts the centuries: j 3 = 0,t3 ≥ 0. The upper limit on
t3 could be whatever limit fits our concept of a “perpetual” calendar. To complete
the programming of the codometer, we need to specify the predecessor functions
p1 , p2 , p3 consistent with the dictates of Pope Gregory. First, we have
{p1 ( j1 ,t1 ) : t1 = 1, . . . , 12} = {31, 28 + j 1, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31}.
Then, p2 ( j2 ,t2 ) = 0 if j2 = 0 and t2 = 0, or t2 is not divisible by 4; otherwise,

p2 ( j2 ,t2 ) = 1. Finally, p3 ( j3 ,t3 ) = 0 if t3 is divisible by 4; otherwise, p 3 ( j3 ,t3 ) = 1.
As an example, consider the date December 31, 1999. To enter the register data,
we first set j3 = 0 (no choice) and t 3 to the number of the century, 19. Since 19
is not divisible by 4, this forces j 2 = p3 (0, 19) = 0. To specify the year, we set
t2 = 99. Since the latter is not divisible by 4, the predecessor constraint forces j 1 =
p2 (0, 99) = 0, corresponding to a common year. Next we insert the month, t 1 = 12,
which forces j0 = p1 (0, 12) = 31. Finally, we choose the day, t 0 = 31, and we are
Appendix B The Codometer 203
done. The resulting state is

31 0 1 0
.
31 12 99 19
Updating by one day, to January 1, 2000, we follow the prescription described above
(details left as an exercise for the reader) to obtain

31 1 0 0
.
1 1 0 20
The codometers used in this book to study the aperiodic orbits of piecewise
isometries, kicked oscillators, and nondissipative round-off maps have infinitely
many registers, at least in theory (in computer simulations, of course, we deal only
with a finite resolution, hence a finite number of levels of a scaling hierarchy, mod-
eled by a codometer with finitely many registers). What makes this situation man-
ageable is the fact that there are only a small finite set I of possible register types.
In a codometer, the kth register is assigned a fixed index i k ∈ I , and the predecessor
constraint does not depend on k, only on i k . For the quadratic models discussed in
this book, the number of distinct register types is typically 2 or 3: the zero register
is distinguished from the rest, which, for a single scaling sequence, can be taken
to be of the same type. For the cubic π /7 model, there are 15 distinct types, and
each codometer has infinitely many registers and is assigned a fixed index string
i = i0 i1 i2 . . .. In the book, we represent its state as
⎛⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎞
i0 i1 i2 i3
⎝⎝ j0 ⎠ ⎝ j1 ⎠ ⎝ j2 ⎠ ⎝ j3 ⎠ · · · ⎠ .
t0 t1 t2 t3
Index
A long-time behavior, 119, 122, 125

admissibility, 64, 68 minimality, 78
π /7 model, 75 multifractal spectrum, 94
edge matching, 69 nonexpansive sector, 125
quadratic example, 70 recurrence time spectrum, 102
tables, 200 recursive tiling, 56
transition graph, 70 residual set, 57
transition matrix, 70 return map tree, 56
algebraic integer, 27, 28 scaling sequences, 60
algebraic number field, 23 symbolic dynamics, 75
asymptotic factorization, 86 transfer matrix, 85
asymptotic long-time behavior, 12, 111, 114, unique ergodicity, 92
150 cyclotomic polynomial, 27
B D
ballistic transport, 129, 134 diffusion constant, 7
baseline, 49, 51 singular, 7
diffusive behavior, 7
C anomalous, 7
Carathéodory dimension, 95 normal, 7
catalogue, 11, 52 discontinuity set, 28, 34, 36, 40, 49
chaos, 5 Model I, 49
chaotic transport, 5 Model III, 50, 193
coarse graining, 12, 131 domain, 28
code, 64 dressed domain, 20, 31
codometer, 13, 63, 68, 201 dynamical self-similarity, 10
complexity, 104
computer assistant, toolkit, 38 E
computer assisted proof, 36 equivalence of domains, 52
concatenation, 49 eventually periodic codes, 65
crystalline symmetry, 3 exceptional set, 26, 36, 80
cubic π /7 model, 53
admissibility, 75 F
generic long-time behavior, 127 factorization theorem, 88
global expansivity, 119, 122 field Q(λ ), 22, 28
Hausdorff dimension, 85 first return map, 10
invariant sectors, 60 form factor, 139, 148, 150, 174–176, 178, 179
205
206 Index
log-quasiperiodic, 178 level 0, 107

fundamental domain, 3 linked index sequence, 66, 68
fundamental domain Ω , 4 linked symbol sequence, 64
fundamental units, 28, 60, 85 local map K, 3, 17
localization, 154
G long-time asymptotics, 111
generalized rotation, 18 π /7 model, 119
generating partition, 41 cubic examples
geometric scaling, 86 (0, k, 6∞ ) sectors, 119
global dynamics, 107 generic, 127
global expansivity, 12, 107, 111 nonexpansive sector, 125
global recursion matrix, 110, 111 quadratic examples, 114
global scaling, 110 Model II, web A, 115
globally bounded orbits, 12 Model II, web B, 116
group G of geometric transformations, 30, 52 long-time behavior, 131
H M
Hamiltonian round-off, 151 metric tensor Q, 18
Hausdorff dimension, 79 minimality, 78
π /7 base tile, 103 Model I
π /7 model, 85
ballistic transport, 134
Model I, 82
data summary, 41
Model II, 83
data tables, 185
Model III Cantor set, 84
discontinuity set, 49
quadratic λ , 81
Hausdorff dimension,measure, 82
Hausdorff measure, 79, 144
moment calculation, 135, 136
construction, 80
probability calculation, 136
Model I, 82
reduced incidence matrix, 82
Model II, 83
Model II
Model III Cantor set, 84
codometer, 139
quadratic λ , 81
data summary, 43
scaling, 80
data tables, 186
I diffusive transport, 142
Hausdorff dimension,measure, 83
incidence matrix, 31
invariant measure, 12 moment calculation, 139, 142
island-around-island self-similarity, 7 probability calculation, 139, 142
reduced incidence matrix, A, 83
K reduced incidence matrix, B, 84
subdiffusive transport, 138
kick function
piecewise linear, 9 Model III, 143
sawtooth, 9 superdiffusive transport, 143
sinusoidal, 5 Cantor set, 84, 143
kicked oscillator, 1 data summary, 46
equations of motion, 1 data tables, 191
Hamiltonian, 1 moment calculation, 145
rotation number, 1 probability calculation, 148
modulated power law, 139, 150, 174, 176, 178
L module M (τ ), 28
lattice map multifractal spectrum, 94
complex, 107
lifting the return map N
higher level, 109 natural return-map tree, 19, 53
examples, 21
Index 207
O S
odometer, 68 sawtooth web map, 9
scaling domain, 20
P scaling sequence, 11, 20, 31
path constraint, 64, 69 sector code
path function, 31 eventually periodic, 110
periodic domain, 33 skew length, 34
piecewise isometry, 18 sticky orbit, 7
Poincaré map, 3 stochastic web, 6
Poincaré section, 2 crystalline, 6
probability distribution, 131, 132 quasicrystalline, 6
direct calculation, 132 stochastic web map, 5
moments, 133 power law, 8
using dynamical partition, 132 subdiffusive transport, 138
using recursive tiling, 132 superdiffusive transport, 8, 143
probabiliuty distribution P((m, n),t), 12 Model III, 143
pseudochaos, 5, 9 symbolic dynamics, 63
pseudochaotic web, see residual set π /7 model, 75
quadratic example, 70
Q symbolic representation, 63
quadratic models, 26, 40 synchronization, 149
long-time behavior, 114
R T
recurrence time, 94 telescoping, 35, 91
recurrence time dimensions, 94, 98 temporal recursion, 87
recurrence-time spectrum, 104 transfer matrix, 81, 97
π /7 model, 102 π /7 model, 85, 97
recursive tiling, 10–12, 20, 25, 33, 34 for recurrence time dimensions, 98
π /7 model, 56 for single scaling sequence, 82
strong, 34, 48, 49 Model III Cantor set, 84
weak, 35, 49 transport, 12, 131
reduced incidence matrix ballistic, 134
Model I, 82 subdiffusive, 138
Model II, A, 83 superdiffusive, 143
Model II, B, 84
Model III Cantor set, 84 U
renormalizability, 10, 11, 13, 17–20, 22, 23, unique ergodicity
25, 26, 36 π /7 model, 92
computer assisted proof, 36 unit, 28, 60
cubic irrational parameter, 11 fundamental, 28
more general, 51
proof, quadratic models, 48
quadratic irrational parameter, 10
V
recursive tiling lemma, 34 vector field, 151, 152
scaling lemma, 32 Vershik map, 68
residual set, 36, 57, 63, 79
return map, 19 W
return path, 20 walk, 156
return time, 20 same shape, 156
ring Z[λ ], 28 web map, 3, 5
round-off error, 152 global, 3
round-off map, 151 local, 3
as kicked oscillator map, 152 weighted perimeter function, 34
Color Figure Index
Tilingof ofAl(O)
0.0008
j color
0
2 •
I;]
4 0
7 0
0.0006
10 B
12 0
••
14 0
15
••
17
19
•
21
22
0.0004
0.0002
Color figure of page 4S

210 Color Figure Index
j col or
-0.0002
0
3 •
0
5 0
DI
•
6
8
10
0
-iH)004
-iJ.OOO6
-iJ.OOO4 -0. 0003 -0.0001 o

Color figureof page 45
Color Figure Index 2lJ
r,,---o:::"liOlll"'"\ .1(0) _ .1 (2)
.1( 2) _ .1 [11
.1(3) - .1 [2]
Fig, 2.18 Return map partitions of catalogue members D [l ] and D [2] into their respective atoms.
The Euclidean metric is used here to make evident the various similarity relations. Superposed are
the dressed subdomains J. (i). i = 0, .. . , 5.
212 Color Figure Index
Fig. 2.19 Tiling of D!11by the return orbits of Dj (i), i = 0, 1,2, j = 0, ... ,J (h(i»). The Euclidean
metric is used to make evident the similarity relations. Consistent with the previous figure , different
values i = 0, 1, 2 are coded red, green, and blue, respectively.
Color Figure Index 213
Color key:
o 500 1000 1500 2000 2500 3000 3500 4000+

p«m,n),I) x C
Fig. 6.7 Probability distribution P « m, n ),I ), multiplied by the normalization factor c = 5298000,
for web S , - 1000 :5 m ,n < 1000,1 = 7200000o.
Color key
o 100000 200000+
p« m,n),I) x C
Fig. 6.8 Probability distribution P «m , n ),t ), multiplied by the normalization factor c = 16 1280,
for web A, - 550:5 m ,n < 550,1 = 900000O.
214 Co lor Figu re Index
Fig. 7.5 Localization of the vector field for plq = 1/ 5.
- 1 8‫סס‬OO
Color key :
-" - lJ- 11- 11- 11l-9 - 8 - 7 - 6- ,

Log\O<:PI3)
Fig. 7.20 Coa rse-graine d probabili ty distribution for the round-off error for a time interval I =
36000ooo. There are 600 x 600 pixels, each covering 6Wlattice sites. Colo rs are assigned to the
indicated ranges of the total probability P per pixel.
Color Figure Index 215
Color key:
I I ..... I_ I_
_ 15 14 LJ 12_11 _1(1 _9
Log 1O<Pf3)
Fig. 7.21 Coarse-grained probab ility distribution for the round-off error for a time interval t =
]600000O. There are 600 x 600 pixels, each covering 62 lattice sites. Colors are assigned to the
indicated ranges of the total probability P per pixel.

Geometria Diferencial Bom

Uploaded by

Copyright:

Available Formats

Geometria Diferencial Bom

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Geometria Diferencial Bom

Uploaded by

Copyright:

Available Formats

John H.

Pseudochaotic Kicked Oscillators

With 86 figures, 9 of them in color

ISBN 978-3-642-28153-2 ISBN 978-3-642-28154-9 H%RRN

Library of Congress Control Number: 2012930642

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

The physics of a pseudochaotic kicked oscillator could hardly be simpler: a one-

New York, August, 2011 John H. Lowenstein

2 Renormalizability of the Local Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.4 Proof of renormalizability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

4 Dimensions and Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

5 Global Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

5.4.2 Numerical investigations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

7 Hamiltonian Round-Off . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

Appendix A Data Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

A.2.7 Incidence matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190

Appendix B The Codometer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201

Color Figure Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

In this initial chapter, we introduce a class of one-dimensional kicked oscillator

1.1 Kicked oscillators

The dynamics of a periodically kicked oscillator is governed by a Hamiltonian of

1.2 Poincaré sections

1.3 Crystalline symmetry

and a global vector τ m, m = (m, n) ∈ Z 2 .

K(u) = W (u) − τ d(u), d(u) = (τ −1 (v + f (u)), −1) ∈ Z2 ,

Fig. 1.2 Graph of some kick function f (x) on [0, 1).

so that, for all u ∈ Ω , m ∈ Z2 ,

W (u + τ m) = K(u) + τ L̃u (m),

with the lattice isometry (rotation composed with lattice translation)

L̃u (m) = I · m + d(u).

Fig. 1.3 Action of W (x).

Fig. 1.4 Construction of K(u) for the map of Fig. 1.3.

W (u + τ m) = K(u) + τ Li(u) (m), (1.7)

1.4 Stochastic webs

1.5 Normal and anomalous diffusive behavior

The discovery of power-law asymptotic behavior in numerical simulations of si-

1.6 The sawtooth web map

Fig. 1.9 Example of a sawtooth kick function f (x).

Sawtooth kicked-oscillator models have been studied in some detail by Lowen-

In rather general terms, renormalizability of a dynamical system refers to its dy-

1.8 Long-time asymptotics

1.9 Linking local and global behavior

1.10 Organization of the book

The remainder of the book is organized as follows.

in the form of an odometer-like updating scheme, the codometer, we concentrate on

The local component of the sawtooth kicked-oscillator dynamics is the renormaliz-

K : Ω → Ω, K(u) = C · u mod Ω , (2.1)

of phase-space coordinates at this point, namely a canonical transformation which

2.1 Heuristic approach to renormalizability

The restrictions imposed by renormalizability on the parameter λ and square do-

2.1.1 Generalized rotations

The matrix C has eigenvalues

and the mapping K is linearly conjugate to a piecewise isometry on the rhombus.

we can legitimately refer to C as a rotation and K itself as a piecewise isometry

2.1.2 Natural return map tree

Renormalizability entails self-similar families of polygonal domains, but given only

/ we note that a subdomain of Ω i , namely

Ei, j,k = K(Ei, j ) ∩ Ωk ,

Di,i1 ,i2 ,...,in−1 , i1 , . . . , in−1 = i.

K(u) = W (u) − τ d(u), d(u) = (τ −1 (v + f (u)), −1) ∈ Z2 ,

Di,i1 ,i2 ,...,in−1 , i1 , . . . , in−1 = i.