BEYOND

QUASICRYSTALS

Les Houches, March 7-18, 1994

Editors
Fran~oiseAxel, Universire Paris VII-Denis Diderot
Denis Gratias, CECN-CNRS, Vitry

Springer-Verlag Berlin Heidelberg GmbH

ISBN 978-3-540-59251-8 ISBN 978-3-662-03130-8 (eBook)
DOI 10.1007/978-3-662-03130-8

This work is subject to copyright. AlI rights are reserved, whether the whole or part of the
material is concemed, specificalIy the rights of translation, reprinting, re-use of illustrations,
recitation, broad-casting, reproduction on microfilms or in other ways, and storage in data banks.
Duplication of this publication or parts thereof is only permitted under the provisions of the
French and German Copyright laws of March lI, 1957 and September 9, 1965, respectively.
Violations falI under the prosecution act of the French and German Copyright Laws.

© Springer-Verlag Berlin Heidelberg 1995

Originally published by Springer-Verlag Berlin Heidelberg New York in 1995
 The school was supported by:

Universite Joseph Fourier de Grenoble

Centre National de la Recherche Scientifique (CNRS-fonnation pennanente)

Direction des Recherches et Etudes Techniques (DGA-DRET)

Commission of the European Communities - DG xn

Minisrere des Affaires Etrangeres

 Obituary

Professor Gerhard Fritsch, who lectured in our School "Beyond

Quasicrystals" on transport properties in quasiperiodic media, died
suddenly and unexpectedly on April 14, 1994.

Gerhard Fritsch was born on July 3, 1940 in Marktredwitz in Nothern

Bavaria where he also spent his school time. Afterwards he studied
physics at the Technical University (TU) of Munich and got there his
PhD in 1969 with a work on "Measurements on the thermic and electric
conductivity of sodium directly at the melting point".

From 1969 to 1975 he worked as Scientific Research Assistant at the Physics

Department of the TU Munich. From 1975 to 1977 he was Research
Associate at the University of lllinois (USA). In 1977 Gerhard Fritsch
was appointed as a Senior Research Scientist at the Physics Department
of the TU Munich and became in 1979 full Professor for Experimental
Physics at the University of the Armed Forces Munich. In 1984 he spent
a year at the University of California at Los Angeles (USA).

Besides the stimulating work in his Institute Professor Fritsch was

committed to the development of the knowledge in modem physics for
teachers. He served as co-editor of the German journal "Physik in
unserer Zeit" for more than two decades.

In research, he combined interests for fundamental topics as the transport

properties of amorphous and quasi crystalline alloys or the determination
of the diffuse X-ray scattering of sodium with interests for more applied
studies as measurements of the properties of materials by means of micro-
beam X-rays.

To his wife, who was with him in Les Houches and to their young daughter,
we present the expression of our heartfelt sympathy.
 Foreword

This book is the collection of most of the written versions of the

Courses given at the Winter School "Beyond Quasicrystals" in Les Houches
(March 7-18, 1994).
The School gathered lecturers and participants from all over the
world and was prepared in the spirit of a general effort to promote
theoretical and experimental interdisciplinary communication between
mathematicians, theoretical and experimental physicists on the topic of the
nature of geometric order in solids beyond standard periodicity and
quasi periodicity.
The overall structure of the book reflects the wish of the editors to
pose this fundamental question of geometric order in solids from both the
experimental and theoretical point of view.
The first part is devoted more specifically to quasicrystals. These
materials were the common starting point of most of the audience and
present a first concrete example of a non-trivial geometric order. We chose
to focus on a few fundamental aspects of quasicrystals related to hidden
symmetries in solids which are not easily found in standard textbooks on the
topic, not to reach an exhaustive survey which is already available
elsewhere.
Y. Meyer redevelops here a few of the ideas presented in his 1972
book "Number theory and harmonic analysis". It is a unique occasion for
most physicists of the quasicrystal community to discover Y. Meyer's very
early synthetic definition of a "quasicrystal". Basics of geometry, structural
and dynamical properties of quasicrystals are given by T. Janssen, whereas
the more experimental crystallographic aspects are developed by W. Steurer
and P. Mikulic. "Special" symmetries in quasicrystals form the basis of the
lectures by J. Patera, P. Kramer and A. Janner, respectively on
"Pentacrystals", on non commutative geometry, the other on multimetrical
analysis. All three extend the field of crystallography by introducing more
general concepts than simple isometries.
The fundamental question of understanding the propagation of non-
trivial order by local interaction "matching rules" in quasi crystals is
discussed by A. Katz in connection with the problem of atomic diffusion by
P. Kalugin. Defects in aperiodic materials from quasicrystals to amorphous
are presented and compared by N. Rivier.
The second part deals with deterministic aperiodic order "beyond"
quasi periodicity. Deterministic sequences are a major example of quasicrystal
generalization. Different properties are presented here from basics in the
VI

Courses by M. Mendes France and J.-P. Allouche, to automatic walks by M.

Dekking with possible relationships to diffusion. Their Fourier transforms
are discussed in great details by M. Queffelec based on the general notions
of measures and substitution dynamical systems. V. Berthe presents and
discusses their entropies, Z. Y. Wen the subwords appearing in the
Fibonacci substitution and its generalizations. Algebraic properties appear
in the Course on trace maps by J. Peyriere.
Realizations in physics of such deterministic aperiodic sequences are
at present of two kinds
- SchrOdinger type equations having such a deterministic sequence for a
potential. This very important topic is presented in the Courses by A. Siito
and H. Kunz.
- Multilayers systems which are now being studied mostly by diffractionnists
Finally, an introductory Course with basics for multifractal analysis
is given by J. Peyriere.
Almost all the authors have made at our request a gigantic effort,
for which we are happy to heartily thank them, to render the contents of
their Course accessible to non specialists: in the present instance, this
School was directed to experimental physicists with the hope to develop the
basis for joint investigations on long-range order in solids. We also are
particularly grateful to the following colleagues who generously provided
help in shaping this book:
E. Cockayne, R. Collela, M. Dekking, F. Delyon, F. Denoyer, F. Gahler,
A. Janner, T. Janssen, P. Kramer, M. Mendes France, M. Queffelec, S. Van
Smaalen.
This very intense moment of scientific communication between
mathematicians, physicists and material scientists, and this book would
never have existed without the constant encouragement of Michele Leduc
(Directeur du Centre de Physique des Houches), her very competent and
helping staff in Les Houches Mesdames G. Chioso et B. Rousset, and
Madame Grosseaux's efficiency and patience (Les Editions de
Physique/Springer-Verlag); we would like them to accept the expression of
our sincere gratitude.
But particular acknowledgements, particular thanks, are due to
Madame Fran~oise Kakou. She took care of the entire administrative
organization of the School, of the preparation of this book and its annexes
with her usual perfect skill, competence, efficiency and kindness, without
which we would not have succeeded.
Readers, be merciful: if you find errors of any kind, please write to
us in view of future printings ...

Fran~oise Axel Denis Gratias

 CONTENTS

Quasicrystals

COURSE 1

Quasicrystals, diophantine approximation and algebraic numbers

by Yves Meyer
1. Introduction.................................................................................... 3
2. Almost-periodic functions. Poisson summation formula and algebraic numbers.. .... 5
3. Model sets and quasicrystals. ................................................................ 9
4. Quasicrystals and diophantine approximation. .. .. . . . . . .. .. . . . . .. . .. . .. . . . . .. . .. . . . .. . .. .. . 10
5. Poisson summation formula and quasicrystals ............. ..................... .......... 14
6. Conclusion ...... ...... ......... ...... .......................................................... 15

COURSE 2

The pentacrystals
by J. Patera
1. Introduction... . . . .. . .. . .. . . .. . . .. . .. . . .. .. . .. . .. .. . . .. .. . . .. .. . . . . . .. . .. . .. . . .. . .. .. . . .. . .. .. .. 17
2. Preliminaries.......... ......................................................................... 18
3. The pentacrystal map. ......................................................................... 20
4. DefInition of quasicrystals. . . . .. . .. . . .. .. . .. .. .. . . . .. . .. . .. .. .. .. . .. . . .. . .. . . .. . .. . . . .. .. . .. . . 22
5. Phasons......................................................................................... 23
6. Quasiaddition .................................................................................. 24
7. Examples....................................................................................... 25

COURSE 3

Elements of a multimetrical crystallography

by A. Janner
Abstract. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . ... . . . . .. . . . . . ... . . . . . .. . . . .. . . . . . . . . . . . . .. . . . . . .. .. . . . . . 33
1. Introduction.................................................................................... 33
2. Close-packed structures.. . . . . . .. . . . .. ... . .. . . .. .. .. .. . .. .. . . . . . .. .. . .. . . .. . .. .. .. . .. . .. . . . . .. 35
2.1 The 2-dimensional case.................................................................. 35
2.2 The 3-dimensional case.................................................................. 35
3. Multimetrical symmetry of the 20 hexagonal lattice...... .. ....... ........ ..... ...... ..... 37
4. Binary integral quadratic forms and quadratic fIelds. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5. IndefInite ternary integral quadratic forms.................................................. 41
6. Quadratic forms oflattices of3D close-packed structures.. .. .. .. .. ... .. .. .. . .. .. . . .. .. .. 44
6.1 Reduced metric tensors.................................................................. 44
6.2 IndefInite binary quadratic forms....................................................... 45
6.3 IndefInite ternary quadratic forms ...................................................... 46
7. Multimetrical point group of the hexagonal close-packed lattice. . . ... . . . . .. . . . . . . . . . . . . . 47
8. Multimetrical space groups of crystal structures...... ......... .... ............... ..... .... 49
VIII

8.1 Hexagonal close-packed structures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

8.2 The Wurtzite structure. .................................................................. .... 51
9. Concluding remarks........................................................................... 52

COURSE 4
Non-commutative models for quasicrystals
by P. Kramer and J. Garcia-Escudero

1. Why non-commutative models for quasicrystals ? .................................... .... 55

2. Free groups and their automorphisms ................. . . . . . . . . . . . . . . . . . . .. . . .. .. . .. . . . . . . . . 56
3. Non-commutative crystallography..................... ... .................................. 57
4. Structure and geometry of the group Aut(F2) .. • • .. • • • • • .. • • .. • .. • .. • .. • • .. • .. • .. .. • .. • • • • • 58
5. Free groups and automorphisms for n > 2 ................................................ 66
6. Non-commutative models and symmetries for 2D quasiperiodic patterns.............. 66
7. Automata for the triangle and Penrose patterns. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
8. Survey of other results.. .. . .. . . . . .. . .. . .. . .. . .. . . .. . . . . .. . . .. . . .. . .. . .. . .. . .. . . .. .. . .. . . . . . . . . 72

COURSE 5

From quasiperiodic to more complex systems

by T. Janssen

1. Structures........ .................................................................. ... .......... 75

1.1 Introduction ............................................................................... 75
1.2 Classes of quasiperiodic structures. .. . .. . . . . . . .. . . .. . .. . .. . .. . . .. . .. . . . . . .. . . . . . . .. .. . . 78
1.3 Embedding of quasiperiodic systems.................................................. 81
1.4 Superspace groups.................................................... ................... 84
1.5 Action of symmetry groups in 3-dimensional space.............. . . . . . . . . . . . . . . . .. . . 86
1.6 Scale symmetries......................................................................... 89
1.7 Hierarchy of structures .................................................................. 91
1.8 Physical origin of quasiperiodicity . . .. . . .. . . .. . .. . .. . . .. . . . . .. . . . . .. . . .. . . . . . . . . . . . . .. . 93
2. Diffraction............. . . .. . . . .. . . . . . .. . .. . . . . .. . . .. . . .. . .. . .. . . . . . . . . . . . . .. . .. . . . . .. . .. . . . . . .. . 94
2.1 Structure factor. . .. . . . . .. .. . . . . . .. . .. . .. . .. . .. . . .. . .. . . . . . . . . . . . .. . . .. . . . . .. . .. . .. . . . . . . . . 94
2.2 Structure factor of quasiperiodic structures .............................. . . . . . . . . . .. . . 97
2.3 Influence of symmetry ................................................... ................ 99
2.4 Thermal vibrations ....................................................................... 100
2.5 Disorder................................................................................... 101
3. Phonons........................................................................................ 104
3.1 Phonons in IC phases.................................................................... 104
3.2 Spectra..................................................................................... 110
3.3 Phonons in quasicrystals ................................................................ 111
3.4 Neutron scattering from quasiperiodic structures..................................... 116
4. Substitutional chains.......................................................................... 122
4.1 Introduction .................................................................. ............. 122
4.2 Atomic surfaces........................................................................... 124
4.3 Fractal atomic surfaces................................................................... 127
5. Electrons........................................................................................ 132
5.1 Models..................................................................................... 132
5.2 Spectra..................................................................................... 135
5.3 Wave functions........................................................................... 136
 IX

COURSE 6
Matching rules and quasiperiodicity: the octagonal tHings
by A. Katz
1. Introduction........ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
2. Quasiperiodic tHings..................................... ..................................... 142
2.1 Quasiperiodicity ......................... ................................................. 142
2.2 The atonric surfaces ............... ......... ... ......... ... ... ... ... ...... ......... ... .... 143
2.3 The cut algorithm... ... ...... ......... ... ................ ............ ... ........ ... ....... 144
2.4 Canonical or "Penrose like" tilings. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
2.4.1 Definition............................................................................ 144
2.4.2 The oblique tiling................................................................... 145
2.4.3 Octagonal tHings. ........................... ... ............ ... ..... .......... ....... 146
3. The composition-decomposition method................................................... 148
3.1 Self-sinrilarity............................................................................. 148
3.2 Inflation and quasiperiodicity............................ ............................... 150
4. The method of forbidden planes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4.1 Position of the problem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4.2 Non-transversality conditions. ...... ... ................................................. 153
4.3 The forbidden planes..... ...................................................... ... ....... 154
5. Decoration of the tiles....... ........ .................. ............ ...... ... ................... 156
5.1 A simple case. .......... ........ ..................... ......... ... ... ... ............... .... 156
5.2 The Ammann decoration of vertices......... ........................... ... ... ....... ... 159
6. The main theorem............................ ....................................... ... ....... 162
6.1 Position and intersections of the forbidden planes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
6.2 Systems of data.................................................. ... ........ .............. 163
6.3 Propagation of order. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
6.4 Proof of the theorem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
6.4.1 The pushing procedure............... .................................... .......... 166
6.4.2 The cone of planes................................................................. 168
6.5 Quasiperiodic tilings and "special tilings" .................... ........................ . 170
7. Generalised Ammann tilings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
7.1 Definitions................................................................................. 174
7.2 Symmetry considerations.................. .......................................... .... 174
7.3 Setting the method........................................................................ 175
7.3.1 Systems of data in £1........................................................ ....... 177
7.4 Reduction to "bad prisms" .......................................................... .... 177
7.5 Proof of the theorem.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.6 Order in generalised Ammann tilings of the [ust kind ............ . . . . . . . . . . . . . . . . .. . 181
7.7 Generalised Ammann tilings of the second kind: an example of weak: rules. . . . . . 183
8. Conclusion..................................................................................... 188

COURSE 7
A mechanism for diffusion in quasicrystals
by P.A. Kalugin............................................................................... 191
COURSE 8
Experimental aspects of the structure analysis of aperiodic
materials
by W. Steurer
1. Introduction................................................................................... . 203
2. What are aperiodic materials ? ............................................................... 204
3. Experimental probes for distinguishing between crystals and aperiodic structures.... 206
x
3.1 Diffraction methods ...................................................................... 207
3.2 Imaging techniques....................................................................... 211
3.3 Spectroscopical methods ........................... ..................................... 212
4. Structure detennination methods ............................................................ 212
4.1 The maximum-entropy method (MEM)................................................ 213
4.2 How many reflections have to be measured 1......................................... 214
4.3 Aperiodic sequences and their Fourier transforms......................... .......... 216
4.4 Symmetry-minimum function and Patterson Deconvolution........................ 218
4.5 Quasicrystals versus twinned approximants ........ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
5. Quantitative aperiodic-crystal structure analysis......... ........................... ....... 226

COURSE 9
Scattering on aperiodic superlattices
by P. Mikulfk

1. Introduction.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
2. Theories of X-ray diffraction ................................................................ 230
2.1 Kinematical theory of X-ray diffraction ............................................... 230
2.1.1 The Fraunhofer approximation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
2.1.2 Stationary phase method........................................................... 232
2.2 Dynamical theory................................. ... ..................................... 232
2.2.1 Semi-kinematical approximation.............................................. .... 233
3. Diffraction on multilayers .... ................................................................ 234
3.1 Periodic lattice............................................................................ 236
3.2 The Fibonacci lattice..................................................................... 237
3.2.1 Finite length of the Fibonacci lattice............ .................................. 239
3.2.2 Calculations in the semi-kinematical approximation............... ............. 240
3.2.3 Maxima of the diffraction curve of Fibonacci SL............................... 241
3.3 Other aperiodic superlattices ............................................................ 242
4. X-ray reflectivity ........ .................................... .................................. 243
5. Gratings ........................................................................................ 244
6. Conclusion.................. ................................................................... 246

COURSE 10
Defects in quasicrystals, in systems with deterministic disorder and
in amorphous materials
by N. Rivier

1. Introduction.................................................................................... 249
2. Defects in condensed matter.................................................................. 252
2.1 Homotopy description of defects. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
2.2 A simple method for calculating homotopy groups.................................. 260
2.3 Defects in crystals: dislocations........... ... ... ... ... ........................ ... ....... 261
3. Defects in quasicrystals....................................................................... 263
3.1 The oblique tiling......................................................................... 265
3.1.1 Tilings in 2d and their topological defects....................................... 265
3.1.2 The oblique tiling............................................. ... ... ................ 267
3.1.3 Defects in quasicrystals; summary ......................... ...... ............ .... 269
3.2 Observation of defects................................................ ................... 270
3.2.1 Phase contrast and diffraction contrast microscopy............................ 271
4. Defects in amorphous materials. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
4.1 Symmetry of disorder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
4.2 Glass as a fibre bundle. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
 XI

4.3 Curved space, packing atoms and frustration.... ............ ... ... ...... ... ..... ..... 276
4.4 Iterative decurving.. .. ....... ............... ...... ... ......... ... ... ... ... ... ... ..... ..... 277
4.4.1 Three-dimensional materials. .... .... ..... ... ..... . ... ... ... ... ... ... ... ..... ..... 279
4.4.1.1 The laves phase. .... .... .... .... ..... ... .... .... .. . ... ... ..... . .... ... ... . 280
4.4.1.2 Hierarchic glass. ..... .... ......... ... ... .... . ... ... ... ... ... ... .. .. . .. ... . 281
4.4.1.3 Spatial disorder... . ... ... ... . . .... ... .... . ... . .. . ... ... .. . .. . .. ... .. . .. .. . 282
4.4.2 The role of entropy.... ... .... . .... ... ... . . .... ... ... . ... ... . .. . .. ... ..... ... ... ... . 285
5. Conclusions.................................................................................... 286

... and Beyond

COURSE 11
Automata and automatic sequences
by J.-P. Allouche and M. Mendes France

Part I. Introduction. Substitutions andfinite automata. .... ... ..... ...... .......... ... ... .... 293
1. The Fibonacci sequence ...................................................................... 295
2. The Prouhet-Thue-Morse sequence......................................................... 296
3. The paper folding sequence.. ..... .... ...... ... ... .... .... .... .... ........... ... ... .......... 297
4. Automatic sequences: definition; a zoo of examples. A waming . .. ... ... .. ... ... .. ... .. 298
5. Where fmite automata enter the picture.. .... .. . . ... ... .... . .... ... ... ... .. . ..... .. ... .. ... . 303
6. How random can an automatic sequence be ?.. ... .... ..... ... ... ... ... ... ... ............. 304
7. Miscellanea................................. ............... ..................................... 305
8. Appendix: automata generating the five examples of section 4 ... . . . . . . . . . . . . . . . . . . . . . . . 308
Part II. Further properties ofpaperfolding .................................................... 312
1. Direct reading and reverse reading.......................................................... 312
2. Words and diagrams...................... ................ .................................... 313
3. The folding operators................ ..... .......................................... .......... 314
4. The dimension of a curve. . ... ... ... . ..... ... . . .. . . ... ... ..... . ... ... ... ... ... ... .. . .. .. . .. .. . 318
5. Paperfolding and continued fractions....................................................... 320
Part III. Complements........................................................................... 324
1. Repetitions in infinite sequences............................................................. 324
1.1 The beginning of the story. ... ... . . .... .... .... . ... ... . . . ... ... ... ... ... ... ... .. ... .... . 324
1.2 Why study repetitions ?............ .......................... ...... ........ .............. 325
1.3 More examples... .. . ... ... . .. . . . ... .... . . ... ... . ... . .... ...... ... ... ... .. . . . .. . .. .. . .. ... . 325
2. Multidimensional morphisms and (finite) automata....................................... 326
2.1 An example ..................................... '" ... . ..... ... . ... . . . . . . .. ... ... ... .. .. . . . 326
2.2 Properties. . .. .. . .. . .. . . .. . .. . .... ... . .. ... .... . ... . ... . .... . ... ... . .. . .. ... ... ... .. . . . .. .. . 327
3. Links with cellular automata................................................................. 327
Part N. Fourier Analysis. . ... ... ... ... . . . ... ... ... ... . .. . . .... .. ... . .. . ... ... .. .. . .. . .. .. . .. .. . . 331
1. Fourier-Bohr coefficients..................................................................... 332
2. Bessel's inequality and Parseval's equality.............. ............ ...... ........ ........ 333
3. The Wiener spectrum and the spectral measure.... .............................. .......... 334
4. Analysing the spectral measure ............................. " ........ " . . . . . . .. ... .. . . . .. .. . 337
5. Appendix I: the spectral measure of the Thue-Morse sequence................ .......... 337
6. Appendix 11: the spectral measure of the paperfolding sequence. .. . .. . . . .. . .. ... .. . . . .. 343
XII

Part V. Complexity of infinite sequences.... ............. ... ... ... ... ... ... ... ... ... .. ....... . 346
1. Stumllan sequences and generalizations.......... ..... ... ... ... ... ... ...... ... ...... ....... 347
2. Complexity of automatic sequences. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
2.1 An upper bound.... .... .. ... ................ ........ ... ... ... ... ... ......... ... .. .. .. ... . 348
2.2 Complexity function of some automatic sequences ... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
2.3 The case of non-constant length morphisms .. .. ... .. .... ... ... .... ... .. ..... .. ... .. .. 349

Part VI. Opacity of an automaton........ ... ... ... ..... . .. .. ... ... .... ..... . ...... .... . .. .. .. .. . 352
1. Automata revisited. ..... .. .. ... .. ..... ........ ... ... ... ... .. . .. . ... ... ... ... ... ... .. . ... ... .... 352
2. Computing the opacity.. ... . ........... ... ... ... ... ..... ..... ......... ..... .. ... .. .. ... .. .. .. . 355

Part VII. The Ising automaton .. ..... ... .......... ... ... ... ...... ... ... ... ...... ... ... ... ... .... 358
1. The inhomogeneous Ising chain...... ... .............. ..... .......... .... ......... ... ... .... 358
2. The induced field.. ..... .... ......... ...... ... ... ... ... .. . .. .. ........ ...... ... ... .. .... ... .. .. 359
3: The Ising automaton.......................................................................... 362
4. An ergodic property..... ... .... .. ... ... ....................... ......... .... ... ... .. .. .. ... .. . . 364
5. Opacity of the Ising automaton.............................................................. 366

COURSE 12
Spectral study of automatic and substitutive sequences
by Martine Queffelec
Introduction.. ... . .. .... .. ... ...... ... ....... ...... ...... ... .. ... ... ........ .. . ..... ... .. ... .. .. .. . 369
1. Measures on T .. ............ ... ... ........ ... ...... .... .. ..... ... .... ... .... ... .. . .. .. .. ... .. .. 370
1.1 Basic definitions and notations..... ...... ... ...... ... ... .... .. . ... . .. . .. . . ... ... .... .. . 370
1.2 Discrete and continuous measures. .. .......... ..... ... . .. . ... ... .... ... .. . . ... .. ... .. .. 371
1.3 Singularity and absolute continuity... ... ......... .... ........ .... ............... ....... 372
1.4 Constructive examples... ... ..... ... ....... ... ... ... ...... ... ... ... ...... ... ... .......... 373
2. Tools from ergodic theory.... ... ... .......... ... ..... ... ... ... ... ... .... ...... ..... .......... 376
2.1 For mutual singularity of measures............ ... ... ... ... ... ... . ........ . .. ... .. .. .. . 376
2.2 For dichotomy properties............... ... ... ... ............ ... ... ... ... ... ... .. .... ... . 379
3. Correlation measures............... .......... ... ... ... ............ ...... ... ...... ... ...... .... 380
3.1 Correlation of sequences..................... ... ............ ... ... ... ...... ... ...... ... . 380
3.2 Classification............. .......... ... ............... .................. ... ... ...... ... .... 381
4. Substitutive sequences - automatic sequences........... ............... ... ... .. .... ... .... 382
4.1 Basic notations...... ... ......... ... ... ....................... .......... ... ... .. .. ..... ... . 382
4.2 Basic definitions...................... ... .................... .......... ... ... ......... ... . 383
4.3 Examples... .................. ... ..................... ... ... ... ... ......... .. .. ..... ... .... 383
5. Substitution dynamical systems............................................................. 385
5.1 Defmitions and basic results .................... ... ...... ... .................. ... ....... 385
5.2 Examples .................................................................................. 386
5.3 Spectrum of ~......................... ................ .................... ... ...... ... .... 388
5.4 Discrete substitutions.................................................................... 390
6. Substitutions of constant length............................................................. 391
6.1 Examples .............................. ... ...... ... ............ ... ......... ... ......... .... 391
6.2 Discrete substitutions of constant length. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
6.3 Spectrum of a substitution oflength q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
6.4 Dynamical system as extension of the odometer....... ...... ... ............ .......... 396
7. Substitutions of non-constant length. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
7.1 Description of eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
7.2 Various situations for A = {a,b} ....................................................... 399
7.3 Miscellaneous. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
7.4 Conclusion. ...... ............ ... ... .............. ..... ....................... ... ...... .... 410
 XIII

COURSE 13
Random and automatic walks
by F.M. Dekking

1. Introduction.................... ......... ... ......... ...... ..................................... 415

2. Automatic walks............................................. .................................. 416
3. Scaling structure and self-similarity..... ......... ........................................... 420
4. Quasi lattice walks ................................................. '" ...... ................... 421
5. Dynamical systems, random and automatic..................... .......................... . 424
6. Generalized random walks................................................................... 425
7. Mean square displacement. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428

COURSE 14
Singular words, invertible substitutions and local isomorphisms
by Wen Zhi-Ying

1. Introduction .... , .............................. ......... ...... .................................. 433

2. Singular words ................................................................... . . . . . . . . . . . . . 434
3. Invertible substitutions.................................................. ... ... ... ............ . 436
4. Fibonacci-chain as a periodic chain with discommensurations . .................. ... .... 438

COURSE 15
Entropy in deterministic and random systems
by V. Berthe

1. Introduction....................... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
2. Thermodynamical entropy.................................................................... 442
3. Information theory ............................................................................ 442
3.1 Entropy of a single event. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
3.2 Entropy of an experiment.... ........... ......... ......... ............................... 443
3.3 Concavity of the function L....... ........... .................. ......................... 444
3.4 Marginal and conditional entropy. ..... .............................. ... ................ 445
3.5 Entropy of a finite curve. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
3.6 The sequence of block entropies....... .................. ........................ ....... 447
4. Topological and measure-theoretic entropies............................... ................ 449
4.1 Topological entropy of a sequence..................................................... 449
4.2 Measure-theoretic entropy of a sequence. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
4.3 Measure-theoretic entropy of a partition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
4.4 Topological entropy of an open cover................................................. 452
4.5 Variational principle. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
5. Entropy and spectral properties............................... ............................... 454
6. Some examples of computation of block entropies........................................ 455
6.1 Ultimately periodic and "random sequences" . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
6.2 Sturmian sequences............... ....................................................... 457
6.3 Block frequencies for some automatic sequences .................................... 459
6.4 Conclusion........................ . ............................................. .......... 460
 XIV

COURSE 16
Trace maps
by J. Peyriere

1.Introduction.. . . .. .. . . . . . . . . . . . . . . . . . . . .. . .. . .. . .. . . . .. . .. . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
2. Some identities for 2 x 2 - matrices ......................................................... 466
3. Reviews and notations for free groups ................ ........................ ............. 470
3.1 Free semi-group generated by ;t.............................................. ...... .... 470
3.2 Free group generated by ;t.............................................................. 470
3.3 Representations in SL (2, «::)...................................... ...... ................ 471
3.4 Endomorpbisms .......................................................................... 471
4. Trace maps (two letter alphabet).............................. ................ ..... ...... .... 472
4.1 Definition of trace maps .............................................. ... ...... .......... 472
4.2 First properties of trace maps .... ...... .............................. ... .... ....... ..... 474
4.3 Further properties of trace maps........................................................ 475
5. Trace maps (n letter alphabet).............................. ... ..................... .......... 476
5.1 Three letter alphabet...... ...... .............................. ... ... ... ... ................ 476
5.2 n letter alphabet....... .................. .................................... ............. 477
6. Comments...................................................................................... 477

COURSE 17
Schrodinger difference equation with deterministic ergodic
potentials
by A. Stito

1. Introduction.............. . . . .. . . . . . . . . . . . . . . . .. .. .. . .. . .. . .. . . . . . . . .. . .. . .. . . .. . .. .. . . . . . . . .. . . 483
2. Main examples................................ .......................................... ....... 484
3. General results on the Schrodinger difference equation............ ......... ............ . 488
3.1 Basic observations. ...................................................................... 488
3.2 Transfer matrices......................................................................... 489
3.3 Lyapunov exponent...................................................................... 492
3.4 Scattering problem: Landauer resistance.......... ................................. .... 493
4. Schrodinger equation with periodic potentials. . .. . .. . .. . . . .. .. . . . .. . .. . . . . .. . . . . . . . .. . . . . . 495
5. Spectral theory........................................................... ...... ................ 498
5.1 SchrOdinger operator, [2(Z)-space and spectrum..................... ......... ...... . 498
5.2 Point spectrum............................................................................ 499
5.3 Cantor sets. ......... ........................ ......... ............... ...... ......... ....... 502
5.4 Continuous spectrum................................................................... . 502
5.5 Spectral projections...................................................... ... ............. 503
5.6 Measures .................................................................................. 506
5.7 Cantor function.. ................................................... ...... ......... ....... 509
5.8 Spectral measures and spectral types................................ ................... 510
5.9 A spectral measure for Ho ....................... ........................... ............. 514
5.10 [2(Z) versus [2(N)...................................................................... 515
5.11 Asymptotic behaviour of generalized eigenfunctions. Subordinacy .... . . . . . . . . . . 515
6. Schrodinger equation with strictly ergodic potentials. . . . . . . .. .. . .. . .. . .. . . . . . . . . . . . . . . . . . 517
6.1 Strict ergodicity. . . .. .. . . . . . . . . .. . .. . .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 517
6.2 The spectrum of H (00) =H 0 + V (00) ................................................. 518
6.3 Integrated density of states...................... ........................ ... .... .. ...... . 520
6.4 IDS and Lyapunov exponent............................................... ...... ....... 521
 xv

6.5 Results on the set IE : "«E) =O} ..... ......... ...... .... ..... ... ......... ...... ....... 522
6.6 The role of periodic approximants................. ..................................... 524
6.7 Gordon-type theorems... .... ..... ... ... ......... ...... .... ........ ........ .......... .... 524
6.8 Kotani theorem for potentials of finite range ................................ '" ..... .. 525
6.9 Gap labelling. . . ..... ... ... . . . ... . . . ... .. . ... .. . .. .... .. . .. . ... .... . . . .. ... ... .... . . .. .. . . . 526
7. SchrOdinger equation with Sturmian and substitutional potentials...................... 527
7.1 Fibonacci potential . . . . .. . . . . .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . .. . .. . . . . . . . . . . 527
7.2 General Sturmian potentials.............. ..... .......................................... 528
7.3 Period doubling potential... ...... ...... ...... ...... .... ..... ...... ....... .......... ..... 530
7.4 Thue-Morse potential. ... ......... ....... ............ ......... .... ....... ........ .... .... 530
7.5 Systematic study of substitutional potentials ........................... " .... . . .. ... . . 531
8. Solutions of the problems........... ... .. .... . ... ... . .... . . .... . ... . ... ..... .. . ... . .. . ... ... . 532

COURSE 18

Schrodinger equation in a hierarchical potential

byH. Kunz

1. Introduction............. . . .... .. .. . ..... ....... ..... .. ..... .. ...... . .. . . . . . .... ... ... . . ... .. . .. 551
2. Hierarchical potentials .............................................................. . . .. . .. . . . 554
3. Spectral properties of the Schrodinger equation. . . . .. . . . . .. .. . . .. .. . .. . .. .. .. .. . .. . .. . . .. . 557
4. Liapunov exponents.. ...... ...... ...... ..... ..... ......... ..... .......... ........... .......... 558
5. Spectral measures........................................ ..................... ................ 558

COURSE 19
Introduction to multifractal analysis
by J. Peyriere
1. Introduction.. .. . .. . . .. . .. . . . .. .. . .. .. .. .. .. . .. .. .. . . .. .. .. .. .. .. . .. .. .. .. .. . .. . . .. .. . . .. .. . . .. . 563
2. Short recalls on measures.... . . . ..... . .. .... .. .... . . . .... . . ... .. . ... . . . . . . ... ...... ... .. ... . . 563
2.1 Borel sets and measures... . ...... ...... ...... ...... . ........ .... . . . . .... ........ ..... . . . 563
2.2 An example: the trinomial measures.... .. ...... ......... ... .... .. ...... ...... .. . ...... . 564
3. Several notions of dimensions......... .................. ............... ...................... 565
3.1 Definitions........ ............ ............ ... ..... .... ......... ...... ...... ...... .......... 565
3.1.1 Hausdorff measures and dimension.. . . . . .. . . . . . .. . . . . . . . . . .. . .. . . .. . . . . . . . . . . . . . 565
3.1.2 The Bouligand-Minkowski dimension (or "box dimension")................. 566
3.1.3 Packing dimension................................................................. 566
3.2 Estimating the Hausdorff dimension.. ...... ......... ...... ....... ...... .... ... ........ 567
3.2.1 Comparison of these dimensions................................................. 567
3.2.2 Lower bound for dim.............................................................. 567
3.2.3 Example................ ............. ..... ...... ............... ............ .......... 568
4. The multifractal formalism... . . . .. . .. . . .. .. . .. . . .. .. .. .. . . . . .. . . . .. . . . . . .. .. . . .. . .. . . . .. . . . . . 569
4.1 Definitions...... ..................... ........... ....... ............... ...... ....... ........ 569
4.1.1 The pointwise HOlder exponent ............................... '" ......... ... ... . 569
4.1.2 The t function. ... .... .. .... .. . .. .. . . . . ......... . . .... .. . ... . . . . ..... ...... ... ... ... . 569
4.1.3 The multifractal spectrum...... .. ............. . .. ... . .. ... .... ..... . . ..... . . ... ... . 570
4.1.4 The multifractal formula. .... .. ... . .. ... . . . ... . . . ... . .. . .. ... . .... . .. . ... . .. . . . . . . . 570
4.2 A counterexample of the multifractal formalism. .. . ...... . ... . . . .. .. . .. .... . . . ... . .. . 570
4.3 The trinomial measures..... ........................... ........................... ....... 571
 XVI

4.4 A few results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573

4.5 Negative dimensions..... ...... .......................................................... 574
4.6 Concluding remarks......... ............................................................. 574
Appendix 1. Legendre transforms.............................................................. 575
Appendix 2. Large deviations...... .............................. ............................... 575

Annexes
Current deterministic sequences......................................................... 585

Glossary......................................................................................... 595

Extra references... ..................................................................... ....... 613

Quasicrystals...
 COURSE 1

Quasicrystals,
Diophantine approximation
and algebraic numbers
Yves Meyer

CEREMADE
Universite Paris-Dauphine
Place de Lattre de Tassigny
75775 Paris Cedex 16 - FRANCE

Abstract. Quasicrystals can be characterized by a remarkable Diophantine

approximation property. This permits to define the dual quasicrystal A· as the
collection of y in IRn such that leiYoZ -11 ~ 1 for each x in the given quasicrystal
A. In many cases one obtains A·· = A and this duality is nicely related to the
spectral properties of quasicrystals.

1. Introduction.

In the late sixties, I was studying some problems in harmonic analysis in

which Pisot numbers were playing a key role. A crucial issue was to relate the
arithmetical structure of a given set A of real numbers to the properties of the
vector space consisting of all trigonometric sums f(x) = L.\EA c.\ei.\z whose
frequencies A belong to A. The model case is the trivial observation that f(x)
is 21T-periodic when A = Z.

This was the problem. Then I discovered some remarkable patterns A for
4 Y. Meyer

which all the above mentioned trigonometric sums satisfy

(1) sup If(x-r)- f(x) I ~ € sup If(x) I

z z

for each r belonging to a Delaunay set A;. Here € is an arbitrary positive

number and the Delaunay set A; does not depend on f. (The definition of a
Delaunay set is given at the next page and (1) is detailed in section 4). Using
a clever lemma by N. Varopoulos, I proved the equivalence between (1) and a
Diophantine approximation property satisfied by A. I could characterize these
sets A by a deceivingly simple additive property (definition 2 below). I wrote
two books on the subject and gave an invited lecture at the ICM70 (Nice) on
these patterns A which I named "model sets" .

Fifteen years later on, two physicists working at Ecole Polytechnique, Michel
Duneau and Andre Katz told me that my "model sets" anticipated Penrose
tiling and quasicrystais.
Fran~ise Axel convinced me to lecture on my book during this conference
on quasicrystais. Indeed most of the content of my book was still unknown to
the participants.
Let me use this opportunity to offer the reader a leasury tour and let him
have distinct viewpoints on quasicrystals. If needed, the proofs can be found
in [11 or [21. Th~ notes will be divided into four parts. In the first one, some
basic facts on almost periodic functions, Poisson summation formula and Pisot
or Salem numbers will be stated, for the reader's convenience.
In the second section, the following definition will be used.

Definition 1. A subset A of RR is a De1aunay set if there exist two radii

R2 > RI > 0 such that each ball with radius RI. whatever be its location, shall
contain at most one point in A while each ball with radius R 2 , whatever be its
location, shall contain at least one point in A.

The. first requirement can be given the following equivalent formulation :

there exists a positive r such that ~ E A, ~' E A and ~ =F~' imply I~'-~I ~ r.
A collection of points fulfilling the second condition is relatively dense (this
terminology was introduced by Besicovitch).

Returning to quasicrystais, they will be defined by several equivalent prop-

erties. The first definition stresses the connection between a quasicrystal and
a lattice. A lattice A contained in m,R is an additive subgroup. This can be
written A-A c A where A-A is the collection of all ~-~', ~ E A, ~' E A and
C denotes inclusion. But a lattice A is also a Delaunay set. We are ready for
the first definition.
 QUASICRYSTALS, DIOPHANTINE APPROXIMATION... 5

Definition 2. A quasicrystal A is a Delaunay set in m.n such that A-A c A+F

where F is a finite set.

This approach is great since it is elegant and concise. But for everyday life
one needs a more handable definition. This second definition is the celebrated
"cut and projection" definition which was proved in [2] to be equivalent to
definition 2, modulo finite sets.

In the third section, the same quasicrystals will be characterized by the same
Delaunay property and by a remarkable Diophantine approximation property.
This approach leads to the notion of the dual quasicrystal. The dual lattice
A* is the collection of all y E m.n such that exp (iy . A) = 1 for every A in the
given lattice A. In the case of arbitrary quasicrystal this definition will force y
to be O. Indeed, the dual quasicrystal r* is defined as
A* = {YEm.nj lexp(iY'A)-II~I, AEA}.
In many cases, we will obtain (A*)* = A as expected.

In the last section, a Poisson summation formula will be given in the qua-
sicrystal setting. Roughly speaking the Fourier transform of a sum of Dirac
masses D>., A E A, is a sum of Dirac masses Dy where y runs over the dual
quasicrystal A*. This statement becomes a true fact if the Dirac masses D>.,
,X E A, are suitably weighted. We then consider the measure IJ. = E>'EA W('x)D>.
which is almost-periodic in a sense explained in the following section. The
Fourier transform fJ, of IJ. (in the distributional sense) is a sum E~ 11m where
11m = EYEA* W~(Y)Dy and A~ = A* + ... + A* (m times). But the dual
weights w~ have a rapid decay as m tends to infinity and only finitely many
layers are observed in the diffraction pattern of a quasicrystal.
All these results can be found in my two books [1] and [2) which were
published in 1970 and 1972.

2. Almost-periodic functions, Poisson summation formula and

algebraic numbers.

Once for all, the Fourier transform of a function / E Ll(JRn ) will be defined
by

(2.1) j(~.) = f e- iz'{ /(x) dx

and the Fourier inversion formula reads

(2.2)
6 Y. Meyer

m.
A continuous function I : n - <V is almost-periodic (in the sense given by
Bohr) if for each positive e there exists a relatively dense set Ae of e-almost
periods T for f. These two concepts (relatively dense and e-almost period) are
now defined.
m.
A subset A c n is relatively dense if there exists a positive R such that
each ball with radius R (whatever be its center) contains at least a point A in
A.
An e-almost period T for I is defined by

sup I/(x+T) - l(x)1 :5 e.

zER"

A more constructive approach to almost periodic functions is the following.

First let us observe that any finite trigonometric sum P(x) = E~EF c(A)ei~oz
is almost-periodic (F is a finite subset of m.n). Next if h, j = 0,1,2, ...
is a sequence of almost-periodic functions and if III - hlloo - (j - +00) °
then I is also an almost-periodic function. Here and in what follows, 11/1100 =
SUPzER"I/(x)l·
Finally a function I : m.n - <V is almost-periodic if and only if, for each
e> 0, there exists a finite trigonometric sum Pe(x) = E~EF(e) C(A, e)ei~oz such
that II/-Pelloo :5 e.
This approach will become a more constructive one if one knows how to
construct F(e) and Pe(x) when I is a given almost-periodic function. For
this purpose, we need to define the Fourier coefficients of an almost-periodic
function.
Let B(x, R) be the ball centered at x, with radius R and let IB(x, R)I = enRn
be its volume. Then the limit

(2.3) M(f) = lim IB(x,R)I- 1 [ I(y)dy

R-++oo JB(z,R)
is attained uniformly in x for an almost periodic function I.
Moreover, for each w E m.n, e- iwoz I(x) is also an almost-periodic function
and the Fourier coefficients 'Y(w) of I are defined as

(2.4) M(e- iwoz I(x» = 'Y(w) .

If I is almost-periodic, so is 1/12 and one has

(2.5) MCl/12) = E h(wW .

wES
 QUASICRYSTALS, DIOPHANTINE APPROXIMATION... 7

This implies that the set S of those w for which 'Y( w) =F 0 is at most a numerable
set. This set S is named the spectrum of I and one would expect that, in some
sense,

(2.6) I(x) = L 'Y(w) e iw · z .

wES

Summation procedures generalizing Cesaro summation have been fixed to give

a meaning to (2.6). There exists a family (3s(e,w) of weight factors such that
o ~ (3s(e,w) ~ 1, lime!O (3s(e,w) = 1 for each wE S, (3s(e,w) = 0 if w does not
belong to the finite set F(e) and finally II/-Pelloo --t 0 when e tends to 0 and
when

(2.7) Pe(X) = L{3s(e,w)-y(w)eiw .Z •

wES

As indicated by the subscript S, (3s(e,w) does not depend on I but only on

the additive properties of S.

Let us return to the connection between quasicrystals and almost-periodic

functions which has been alluded to in the introduction. If A is a quasicrystal,
then for any positive e there exists a second quasicrystal A; (named the e-dual
of A) such that

(2.8) for A E A and YEA;.

This implies

(2.9) II/(x-Y)-/(x)lIoo :5 Cell III 00 , yEA;

where I is any almost-periodic function whose spectrum lies in A and C is an

absolute constant (1/100 = SUPzERn I/(x)I). The key point is the fact that the
e-almost periods yEA; do not depend on I.

A second relation between quasicrystals and almost-periodic structures is

provided by the notion of an almost periodic Radon measure.

Definition 3. A (unbounded) Radon measure p. on JRR is almost periodic if

the convolution product p. * 9 = I is an almost periodic function whenever 9 is
a compactly supported continuous function.

If p. is an almost-periodic measure, the closed graph theorem implies that

SUPzERn fB(z,R) Idp.1 = C(R) < 00 for any R > O. Therefore p. * 9 is an almost-
periodic function for any 9 in the Schwartz class S(JRR ).
8 Y. Meyer

In a sense to be made precise, the theory of quasicrystals amounts to studying

those almost-periodic measures f./, which have the property that their (distribu-
tional) Fourier transform p, is also an almost-periodic Radon measure.

Roughly speaking f./, will be a (weighted) sum of Dirac masses 0)" where A
runs over a quasicrystal A and p, will be the corresponding (weighted) sum of
Dirac masses Oy where y runs over the dual quasicrystal A* .

A trivial example is given by the ordinary Poisson formula. In that case, A

is a lattice. A lattice r c R n is a discrete subgroup with compact quotient.
In other words r = A(zn) where A is an n x n invertible matrix. We define
vol (r) as the volume of any fundamental domain of r. Then vol (r) = Idet A I.
The dual lattice r* c R n is defined by eill ·:Z: = 1 for every x E r and every
y E r*.
Then the Poisson summation formula reads

(2.10) vol (r) L fh) = L ih)

,.Er ,.Er o

whenever f belongs to the Schwartz class S(JRn ). In other words, the Fourier
transform of f./, = L,.Er 0,. is P, = vol (r*) L,.ErO 0,. which provides a triv-
ial example of an almost-periodic measure f./, with an almost-periodic Fourier
transform.

This section will be ended with the definition of some remarkable algebraic
integers, the so-called Pisot and Salem numbers.

Definition 4. A Pisot number is a real number 0 > 1 with the following two
properties

(2.11) o is an algebraic integer of degree n ~ 1

(2.12) the conjugates O 2 , ••• , On of 0 satisfy
102 1< 1, ... ,IOnl < 1.

For example, 2, 3, ... are Pisot numbers and the second condition is vacuous
in that case. When the degree nofO exceeds 1, wehaveOn+alOn - l +. +a n = 0
where al E Z, ... , an E Z. Then O 2 , I •• , On are the other solutions to this
equation and can be either real or complex numbers.
Salem numbers are defined in the same way. One keeps condition (2.11) but
replaces (2.12) by 102 1:$ 1, ... , 10ni :$ 1 with, at least, equality somewhere.
Then n is even, O2 = i and 103 1= ... = 10ni = 1 ([4]).
 QUASICRYSTALS, DIOPHANTINE APPROXIMATION... 9

3. Model sets and quasicrystals.

A model set A c m." is defined the following way. One considers a lattice
D c m." x m.m where m is an integer and m = 0 is not excluded. If (x, y) =
X E m." x m.m, we write x = PI (X) and y = P2(X}.
Let us assume that PI: D -+ PI(D} c m." is a 1-1 mapping and that P2(D}
is a dense subgroup of m.m. The latter condition is always satisfied when m = O.

Definition 5. Keeping these notations, let B be any bounded set in m.m.

Then the model set A defined by D and B is the collection of all A = PI (d)
such that d E D and P2(d} E B.

The reader is referred'to definition 4, page 48 of [2].

When m = 0, then D = A is a lattice in m.". If m ~ 1 and if 1'2 : D -+ P2(D}
is not a 1-1 mapping, one easily shows that in a suitable coordinate system
A = Al X A2 where Al C m."l is a lattice and A2 C m."2 is a model set. We
have n = ni + ~, n2 < n and the mapping 1'2 associated to A2 is 1-1. It
means that the general definition of model sets can be reduced to the special
case when 1'2 : D -+ P2(D} is a 1-1 mapping. In that special case, any subset
A' of a model set A is still a model set.

Let us remind the reader of the definition of a Delaunay set A. Each ball
of radius Rl whatever be its location should contain at most one point in A
and each ball of radius R2 whatever be its location should contain at least one
point in A (R2 > RI > O).

Definition 6. A quasicrystal is a subset A of JR." fulfilling the following two

conditions

(3.1) A is a Delaunay set

(3.2) there exists a finite set F contained in m."
such that A-A C A+F.

Before going further let us make a few remarks on that definition. From
(3.1) and (3.2) one immediately deduces that A+F is still a Delaunay set. So is
A-A and moving one step further one obtains that A-A-A is also a Delaunay
set.
If A is a Delaunay set and if A - A - A is also a Delaunay set, then one has
A-A C A+F. The proof is obvious. Let A E A, A' E A; then A-A' = A"-tr where
A" E A and Irl ~ R2 (since A is a Delaunay set). Therefore r = A-A'-A" E
A-A-A and Irl ~ R 2 • Finally rEF which is a finite set since A-A-A is a
Delaunay set.
10 Y. Meyer

We can conclude

Lemma 1. A quasicrystal is a Delaunay set A such that A - A - A is also a

Delaunay set.

Property (3.1) in the definition of a quasicrystal implies that for each R> 0
and each Ao E A, the set F(Ao, R) = {A-Ao j A E A and IA-Aol ~ R} is finite.
But property (3.2) tells us that there are only finitely many such finite sets
F(Ao, R) when Ao runs over A, R being kept fixed. This is exactly a Penrose
tiling property.
The following theorem relates model sets and quasicrystals.

Theorem 1. Let A eRn be a model set such that the corresponding bounded
set B has a non empty interior. Then A is a quasicrystal.
Conversely if A is a quasicrystal, there exists a finite set F and a model
set Ao such that A c Ao + F. Moreover this model set Ao corresponds to a
bounded set Bo with a non empty interior.

This theorem is better understood if the following definition is used.

Definition 7. Two subsets A and A' ofRn are equivalent modulo finite sets
if there exist two finite sets F and F' such that A C A' + F' and A' C A + F.
One then writes A '" A'.

If one keeps Pl, P2 and D fixed and lets B vary, all the model sets AB
associated to bounded sets B with non empty interior will be equivalent. It
is also true that any set which is equivalent to a quasi crystal will still be a
quasicrystal. This remark is no longer true for model sets. A counterexample
is given by Z + {O, J2} which is not a model set while Z is a model set. That
is the reason why theorem 1 does not say that any quasicrystal is a model set.
If two equivalent model sets are identified, one is reduced to the case where
B is the unit ball in definition 5.

4. Quasicrystals and Diophantine approximation.

In this section A will denote a model set for which the corresponding bounded
set B is Riemann-integrable. It means that the Lebesgue measure of B (the
closure of B) equals the Lebesgue measure of int (B) (the interior of B).
Let D* c R n x R m be the dual lattice of D. It means that D* is the
collection of all (e,1]) E R n x R m such that x . e+ Y . 1] E 211"Z for each
 QUASICRYSTALS, DIOPHANTINE APPROXIMATION... 11

(x, y) ED. Then PI : D* -+ PI(D*) is still a 1-1 mapping while P2(D*) is still
dense in IRm.

A first observation is an asymptotic form of the Delaunay property. If A is a

model set for which B is Riemann-integrable, then each ball B(x, R) centered
at x with radius R contains exactly cRn + o(Rn) points in A where c = :'rD
and where the little 0 is uniform in x.
If c>., ,\ E A, is any bounded sequence, its mean value over A (if it exists)
will be defined as the uniform limit as R tends to infinity of

E>.EB(:r:,R) c>.
(4.1)
U {B(x,R)nA} .

With these definitions in mind, one can state the following theorem ([2], page
159).

Theorem 2. Let A be a model set for which B is Riemann-integrable with a

non empty interior. Let w E m.n and let us assume that w ¢ PI (D*). Then the
mean value over A of eW .>' is O.
If w = PI (d*) and d* E D*, then the mean value of eiw .>' over A is given
by IBI- 1 i B (P2(d*)) where IB is the indicator function of B, iB denotes the
Fourier transform of this indicator iwlCtion and IBI is the Lebesgue measure
ofB.

Theorem 2 paves the road to the definition of the dual lattice. Let us assume
that leW.>' -11 ::; 1 uniformly over A. Then Re( eW.>') ~ 1 and by convexity
the same inequality will be valid for any average of eiw .>., ,\ E A. Therefore
wE PI(D*) (otherwise the mean value of eW.>' would be 0).
If w = PI(d*) while'\ = Pled), then

(4.2) w· ,\ + P2(d) . P2(d*) E 21l"Z

and

(4.3) exp (iw .,\) = exp (-iP2(d) . P2(d*)) .

Since P2(D) is dense in m.m, so is BnP2(D) in B and Iexp (iP2 (d)·P2(d*)}-1 I ::; 1
for P2(d) E B implies lexp{iy ·P2{d*)-ll::; 1 for any y E B.
Since B has a non empty interior, that forces P2(d*) to belong to a bounded
set B*.
Therefore w = PI(d*) belongs to a model set A*(B*). Let us observe that
le iw ·>'-11 ::; p < J2 uniformly over A would also lead to wE A*(B*) but this
argument breaks down if p > J2.
12 Y. Meyer

We do not know the optimal value of p ensuring that leiw'~-11 ~ p uniformly

over a given model set A implies w E A* where A-is also a model set. Let us
assume that B is a closed convex set, that B is Riemann integrable and 0 E B.
Then leiw'~-11 ~ 1 uniformly over A is equivalent to w = PI (d-), d* E D*
and Iy' P2(d*) I ~ 'Tr/3 for each y E B.
Let B* be the dual convex set defined by Z E B* iff sUPIIEB Iz . yl ~ 'Tr /3.
Then P2(d*) E B-.
A full symmetry between a model set A and its dual model set A* is then
given by the following theorem.

Theorem 3. H B is Riemann integrable, closed, convex and symmetric around

0, if A is the model set defined by {oX = Pled) j dE D and P2(d) E B}, then
the dual A* = {y E m.n j lei~'11-11 ~ 1 , A E A} is also a model set given by

and (A *)* = A.

Let us now compute the E-dual set A; defined by leill'~-11 ~ E, A E A, when

E E (0,1). We immediately obtain

(4.4) y E' A; <=> y = PI (d*) , d* E D* and P2(d*) E'1B*

where 0 < '1 < 1 and sin ('1l-) = ~.

Therefore A; is also a model set.
This observation paves the way to the following definition.

Definition 8. A subset A ofm.n is harmonious if for each E > 0, the E-dual

set A; is relatively dense in m.n •

Any finite set F is harmonious. IT A is harmonious and if F is finite, then

A+F is still harmonious. IT A is harmonious, so is A-A. Finally any harmonious
set is uniformly discrete: there exists a positive 6 such that A E A, A' E A and
A' :F Aimply lA' -AI ~ 6.
Any model set is harmonious and theorem 1 implies that any quasicrystal is
harmonious.
Let us prove the following observation.

Theorem 4. Let A be a Delaunay set in m.n • Then A is a quasicrystal if and

only if A is harmonious.

IT A is harmonious and is a Delaunay set, then ~ = A - A - A is also

harmonious. Therefore ~ is uniformly discrete and is a Delaunay set. Lemma
1 shows that A is a quasicrystal.
 QUASICRYSTALS, DIOPHANTINE APPROXIMATION... 13

Let us return to the relationship between the arithmetical properties of a set

A and the uniform almost periodicity of the linear space C A of all trigonometric
sums I(x) = E~eA c~ ei~':r: whose frequencies belong to A.
I wanted to know whether for each positive e there exists a Delaunay set TE
such that

(4.5) rETE =? II/(x-r)- I(x) II 00 :5 ell III 00 , IE CA .

If one picks I(x) = ei~.:r:, then one obtains lei~'T' -11 :5 e for each rETE'
Therefore A is harmonious. Conversely if A is harmonious, a clever lemma due
to N. Varopoulos implies that II/(x -r) - I(x) 1100 :5 Cell/lioo where C is an
absolute constant and I E CA.

The relationship between Pisot numbers, Salem numbers and harmonious

sets is given by the two following theorems.

Theorem 5. LetO > 1 bearealnumber. Then thesetS = {1,O,02, ... ,(Ji, ... }
is harmonious if and only if (J is a Pisot or is a Salem number.
The proof is given in [2], p. 37.
A simple corollary is the following observation.

Theorem 6. If A is a quasicrystal, if (J > 1 is a real number and if (J A c A,

then (J is either a Pisot number or is a Salem number.
Conversely for each dimension n and each Pisot or Salem number (J, there
exists a quasicrystal A in m.n such that OA c A.

Let us prove theorem 6.

Let ~ E A, ,xo =I- O. Then OA c A implies that ,xoS is contained in A when
S is defined as in theorem 5. But a subset of a harmonious set is obviously a
harmonious set. Therefore ,xoS is a harmonious set in m.n and also in the line
,xom.. Finally S is harmonious and theorem 5 can be used .
. Conversely let 0 be either a Pisot or a Salem number and let K = Q«(J) be
the algebraic number field generated by 0. Let O2 , ••• , On be the conjugates
of 0 and 0'2 : K -+ <t, ... , Un : K -+ <V be the Q-isomorphisms defined by
0'2«(J) = O2, ... , O'n«(J) = On).
The ring n of all algebraic integers contained in K admits a direct sum
decomposition given by

(4.6) n = Zwl + ... + ZWn •

Let A be the collection of all algebraic integers ,x E n such that I0'2 (,x) I :5
1, ... , 100n(,x) I :5 1. Then A is a model set for which m = n - 1, D = n,
1'2 = (0'2, ... , Un) and PI is the identity mapping.
14 Y. Meyer

Moreover A is closed under multiplication : if A E A and A' E A, then

AN EA. Therefore 8A c A.
We have obtained a one dimensional model set A such that 8A c A. If one
needs an n-dimensional model set, it suffices to take A x ... x A (n times).

5. Poisson summation formula and quasicrystals.

Let A be a model set defined as above by D c m,n X m,m and B c m,m.

Let us assume this bounded set B Riemann integrable and let cp denote any
c~(m,m) function vanishing outside B.

Using this function cp, weights W(A) are defined on the model set A by
w(Pl(d)) = cp~(d)), d E D. If cp were the indicator function of B (which
is not smooth), we would have W(A) = 1 on A and w(Pl(d)) = 0 if PI (d) ¢ A.
With these notations, one obtains

Theorem 7. Let JL be the sum E~EA w(A)6~ of Dirac masses over A where
the weighting factors are defined as above. Then the Fourier transform in the
distributional sense {.t of JL is given by

(5.1) (.t = Lw~(d*»6plW)

D'

where the dual weights w~(d*)) are defined by

(Y) = (21r)n A( )
(5.2) W vol D cp -Y , Y =P2(d'") , d'" E D*.

If one could find a testing function cp E c~(m,n) with a compactly supported

Fourier transform, {.t would then be supported by the dual quasicrystal A". This
is not the case. But let us consider the case where B is a convex set, symmetric
around 0 and Riemann-integrable. Then the dual quasicrystal A'" is defined by
the dual convex set B'". For each integer m ~ 1, let us define A:'n with mB'"
instead. of B'", PI, P2 and D'" being kept fixed.
Then theorem 7 yields {.t = Em>l Vm where Vm is supported by A:'n and
where this series has a rapid decay at infinity. More precisely

(5.3)

for every ball B and each integer N.

 QUASICRYSTALS, DIOPHANTINE APPROXIMATION... 15

This rapid decay explains why only finitely many layers A:n are observed in
the diffraction pattern of a quasicrystal.

6. Conclusion.

To my opinion, the most sensible definition of a quasicrystal is given in terms

of Diophantine approximation : a quasicrystal A is a Delaunay set such that
the e-dual sets A~ are also Delaunay sets whenever 0 < e ::; 1. The e-dual set
A~ is defined by leiY'~-II::; 10, A E A, y E A~.

This approach is appealing since it gives an interesting heuristic to the

diffraction pattern. Indeed if the weights W(A), A E A, are nonnegative, the
corresponding sum E~EA w(A)ei~'1I attains its largest value on A* and is ev-
erywhere larger than a half of this largest value on the dual quasicrystal A* .
This explains why the Fourier transform of the measure JL in theorem 6 peaks
on the dual quasicrystal A* .
The Diophantine approximation definition of quasicrystals also explains the
role played by Pisot and Salem numbers as inflation factors of quasicrystals.
This property is used in a systematic way in J. Patera's work where the Pisot
number T = ¥ plays a key role.
Finally this Diophantine approximation characterization is exactly what is
needed for proving that almost-periodic functions with frequencies in a given
quasicrystal A behave like periodic functions with frequencies in a given lattice.

The second definition of a quasicrystal is the celebrated "cut and projection"

algorithm. I was led to this definition when I tried to generalize examples of
the type Am = {Ek>Oek(}k ; ek E Z , lekl ::; m} where m ~ 1 is a given
integer, (} is a Pisot number and ek = 0 when k is large enough. In such an
example each A E A is an algebraic integer and its conjugates A2,"" An will
satisfy

The third definition of a quasicrystal is given by A-A c A+F (F finite) and

A is a Delaunay set. This approach to quasicrystals is striking by its elegant
simplicity. But this approach is deceiving since it does not directly yield the
Diophantine approximation characterization of quasicrystals. One has to go
through the "cut and projection" method (theorem 1).
One is tempted to relate this additive characterization of quasicrystals to
the definition of pentacrystals which was introduced by J. Patera. Patera
defines a pentacrystal by the following two conditions : A is a Delaunay set
 16 Y.Meyer

and r2 A-T A = A where T = l t2V5. Pentacrystals are examples of quasicrystals

and T is an example of a Pisot number.
Admiring the spectacular examples of quasicrystals which were designed by
J. Patera, I better understand why my early work was never given the credit
it deserved. A systematic search for all quasicrystals was needed for solving
the problem I was attacking. The specific quasicrystals designed by J. Patera
would have been useless in my own work. They are based on a specific Pisot
number while I needed any Pisot number.

References.

[1] Y. Meyer, Nombres de Pisot, nombres de Salem et analyse harmonique.

Lecture Notes in Math., 117 (1970) Springer-Verlag.
[2] Y. Meyer, Albegraic numbers and harmonic analysis. (1972), North-Holland.
[3] J. Patera, The pentacrystals. These proceedings.
[4] R. Salem, Algebraic numbers and FouriJr analysis. Boston, Heath, (1963).
[5] R. V. Moody, Meyer sets and the finite generation of quasicrystals. (Preprint).
[6] J.P. Gazeau, Quasicrystals and their symmetries. (Lectures given at the
lIIrd International School on Symmetry and Structural Properties of Con-
densed Matter, POZNAN 1994, preprint).
[7] J.P. Gazeau, Lectures on Quasicrystals and their symmetries. Charles Uni-
versity, Prague, June 1994.
[8] A. Katz, A short introduction to quasicrystallography. From Number theory
to physics, (M. Waldschmidt & al. ed.) Springer-Verlag, Berlin (1992).
 COURSE 2

The pentacrystals
J. Patera

Departement de matMmatiques et de statistique

and Centre de recherches matMmatiques,
Universite de Montreal, C.P. 6128,
succ. Centre-Ville, Montreal, Quebec, H3C 3J7
Canada

1. INTRODUCTION

The purpose of this talk is to bring to your attention recent new general ap-
proaches to the structure of what we call pentacrystals [1,2,3,4]. A pen-
tacrystal is any quasicrystal whose points can be written, relative to some
basis {Ul, ... , un} of a real n-dimensional Euclidean space lRn, with coefficients
in Q[J5]. the quadratic extension of the rational number field Q. In these lec-
ture notes all quasicrystals are pentacrystals even if they do .not display local
5-fold symmetry.
In [1] (see also [5]) by starting from the root lattice Q of Es, the largest
of the exceptional simple Lie algebras, and treating it as the generic "mother
of all pentacrystals" - one gains an appealing unifying view of the diversity
of special cases found in the literature. Indeed, the number of special cases
and the avenues through which quasicrystals in dimensions 2, 3, and 4 arise
from Es is rather large and· has not yet been investigated systematically. The
principal tool which allows one to achieve this is a pair of (Q[v'5]- )conjugate
maps 11"11 and 11"1., the conjugation being the involution of the field Q[ v'5] which,
in particular, transforms the constants T = !(1
+ J5) and u = !(1- v'5) into
each other. The maps are defined for our purposes in Figure 1.
In [2] a new algebraic binary operation, denoted by I- and called quasi ad-
dition, is defined (Section 6 below). It operates on L F , the subset of points in
lRn with coordinates from IF = Q[v'5]. Although the true significance of this
18 J. Patera

operation most likely exceeds the goals and aims of the study of quasicrystal, it
leads here to a natural definition of a quasilattice as any set of points which has
the Delaunay property (cf. Section 2) and which is closed under quasiaddition.
It turns out that virtually all common quasi crystals are closed under I- and
that it is not necessary for a quasicrystal to display a 5- or 100fold symmetry
to be closed under the quasiaddition.
In addition let me also point out related work in [3,4,6]. More precisely in
[3] the wealth of the structure of quasicrystals is made evident by (i) introducing
the notion of (finite) colorings of quasicrystals, (ii) showing that there is an
unlimited number of different colorings possible for each quasicrystal, and (iii)
actually classifying the colorings with the property that each subset of points
of a given color is closed under quasiaddition. Hence any such one-color subset
is a sub-quasicrystal.
Reference [4] is a Macintosh software called simpLie in which the methods
and operations of [1,2,3] are being implemented. All the examples here and
in [1 - -5] were calculated and printed from simpLie.
In [6] the shelling problem is solved for the Es-quasicrystals. Supposing that
the origin is a quasicrystal point, one considers concentric shells containing the
quasi crystal points. The problem consists in counting the number of points in
each shell.

2. PRELIMINARIES

Let us recall the following facts while introducing some notation.

1. One has the real numbers JR, the rational numbers Q and integers Z as
well as the extension IF = Q[ y5j of Q by v'5. Thus a generic element of IF
is of the form a + by5, where a, b E Q. The constants T = ~(1 + V5) and
a = ~(1 - V5) and the identities

T + a = 1, T2 = 1 + T, a2 = 1 + a, aT = -1, (1)

are frequently used. Subsequently a crucial role is played by the conjugation

in IF denoted by prime:

X' = (a + bv'5)' = a - bv'5, a,bEQ. (2)

Note that T' = a.

2. We will use the quaternions lHIF over IF. Any x E lHIr can be written as

The use of quaternions is not indispensable. Superficially they serve here only
as a model of a 4-dimensional Euclidean space (in the same way that the com-
plex plane is commonly used as a model of a Euclidean plane). Indeed, one
 THE PENTACRYSTALS 19

can consider (3) as a vector in R4 given relative to an orthonormal basis. How-

ever, there are marvelous sets of special quaternions with properties begging to
be used in the context of a study of quasicrystals; beautiful consequences and
deeper insight into the problem might otherwise be obscured.
The quaternions come equipped with their standard operations:

In-conjugation of x : x= (Xl, -X2, -xa, -X4) (4)

multiplication in lBl : xy E lBl, x,y E lBl (5)
1
scalar product in lBl, : X· Y = "2(xfi + yx) E F (6)
Inp-norm: N(x) = X· x= Ixl 2 = xx E F. (7)

3. Important but little known is the rational form (x,y).,. relative to r on

In" defined as follows. If x . y = a + rb, where a, b E Q, then

(x,y).,. = a. (8)

Analogously one defines a rational form relative to 0'. Obviously both forms
take only rational values on In,.
4. Next is a special set of 120 quaternions of norm 1, called icosians [6].
The set I of icosians consists of the following elements of In,:

1
"2(±1, ±1, ±1, ±1) (9)
(±1, 0, 0, 0) & all permutations (10)
1
'2(0, ±1, ±O', ±r) & all even permutations (11)

The icosians form the group Ha, IHal = 120, under the quaternionic multipli-
cation. It is isomorphic to the binary icosahedral group.
5. Finally, we use the isometric isomorphism 11"11 between the space R S
spanned by the roots of Es and In, [7]. See Figure 1 for its explicit definition.
The mapping 11"11 establishes the correspondence between the root system !:::J. of
Es and the icosians and their r-multiples

11"11 :!:::J. - - IUrI

(12a)
(al.8) = 2(11"11 (a), 11"11 (.8).,., <a,{3 E!:::J..

Similarly one defines

11".1 :!:::J. - - IuO'I
(12b)
(al{3) = 2(11".1 (a), 11".1 ({3»cr.
Note that one has (11"11 (a»' = 1I".1(a).
20 J. Patera

al a2 a3 a4 al a2 a3 ra4

~ - 7r1l

~
a7 a6 as as ral ra2 ra3 a4

a1= ~(-a,-7,0,-1), 7a1 = ~(1,-7 2,0, -7),

a2= ~(O, -a, -7, 1), 7a2= ~(O, 1,-72 ,7),
a3= ~(O, 1, -a, -7), 7a3= ~(O, 7,1, _7 2 ),
a4= ~(O, -1, -a,7), 7a4= ~(O, -7, 1,72 ).

Fig. 1. - Mapping 11"11 of the simple roots of Es into specific icosians (quaternions).
The conjugate map 11".1. is obtained from 11"11 by replacing u by 7 and 7 by u. The
Coxeter (or Dynkin) diagram of type Es serves here as a short way to describe a
non-orthogonal basis {a1,"" as} in an Euclidean space: Each node stands for a
normalized basis vector, two nodes which are directly linked indicate the angle 1200
between the corresponding basis vectors, two nodes which are not directly connected
stand for a pair of mutually orthogonal basis vectors.

3. THE PENTACRYSTAL MAP

The pentacrystal map 11"11 is the 'projection' part of the 'cut & projection'
method for construction of quasicrystals. It is useful to introduce it indepen-
dently of the narrow purpose of quasi crystal construction. There is a natural
generalization to dimension 4 of the 2D and 3D cases, familiar in physics and
first pointed out by Elser and Sloane [13]. Once such a generalization is set
up explicitly, it is extremely useful because the well known lower dimensional
cases become special cases of the 4-dimensional 'mother of all pentacrystals'.
The map realizes the reduction of the orbits of the point group of the Es-
lattice (the Weyl group W(Es) of order 696729600 generated by eight 'elemen-
tary' reflections in the planes orthogonal to the basis vectors a1, ... ,as defined
by the diagram on Figure 1) to the orbits of the largest non-crystallographic
irreducible finite group H4 (see for example [1,9]) acting in lHIJF which is of
order 14400.
An explicit realization of the map 1r1l is given in Fig. 1. There one finds
the simple roots of Es, basis of the space VQ of all linear combinations of sim-
ple roots with rational coefficients, mapped into specific icosians at. a2, a3, a4
and their 7-multiplies, basis of the target (quasicrystal) space I. The rational
form (-, ')7" of (8) assures that the relative lengths and angles of Es roots are
 THE PENTACRYSTALS 21

preserved under 11"11. Over the real number field lR the quasicrystal space R
is 4-dimensional, ai and rai being linearly dependent, while over the rational
numbers Q the space is 8-dimensional because then ~ and r~ are linearly
independent.
In addition to the 1I"1I-map, we have as well its F-conjugate, the 1I".L -map
acting from VQ to R. It is obtained by replacing r by u everywhere in Fig. l.
Any point X of VQ ,

X =:= C1a1 + ... + csas, C1, ... , Cs E Q, (13)

is mapped by 11"11 and by 1I".L as follows

11"11 (X) = (C1 + rC7 )a1 + (C2 + rcs)a2 + (C3 + rcs)a3 + (rc4 + cs)a4'
(14)
1I".L (X) = (C1 + UC7 )a~ + (C2 + ucs)a~ + (C3 + ucs)a~ + (UC4 + cs)a~,

where at, ... ,a4 are the F -conjugates of the icosians al. ... , a4.
Subsequently we will need to find the length of vectors X E VQ and 1I"1I(X)
and 1I".L (X). It is therefore useful to have the matrices of scalar products ofthe
corresponding basis vectors:
2 -1 0 0 0 0 0 0
-1 2 -1 0 0 0 0 0
0 -1 2 -1 0 0 0 0
0 o -1 2 -1 0 0 0
((aj I ak)) = 0 0 0 -1 2 -1 o -1 , (15)
o 0 0 0 -1 2 -1 0
o 0 0 0 0 -1 2 0
o 0 0 0-1 0 0 2

«aj I ak» = ~ (-~2-1_~ -~0_~0) , ( 2-1 0 0)

«aj I al.J) = ~ -~ _~ -~ _~ . (16)
o 0 -7" 2 o 0 -iT 2

The matrix (15) coincides with the well known Cartan matrix [10] of the Lie
algebra Es. The target space of the maps 11"11 and by 1I".L is in JH[F. Hence any
quasicrystals constructed with the maps would be 4-dimensional. It can also
be shown [1] that the underlying symmetry group is H4. Although it is not
common to speak of Cartan matrices of non-crystallographic Coxeter groups,
it is justifiable [9] to call either of the matrices (16) the Cartan matrix of H 4 •
Let us now single out the two important subcases of 11"11 and 1I".L maps lead-
ing to well known icoshedral quasi crystals in 3 dimensions and to decagonal
quasicrystals in 2 dimensions:
Putting C1 = C7 = 0 in (14) is equivalent to striking out the nodes of Fig. 1
labeled by al, a7, and al. ra1 thus reducing the dimension of the space VQ
from 8 to 6 and the Es-diagram to that of Ds (cf. Example 1 and Eq. (29)
below). Note also that the quatemionic target space becomes the space of pure
quatemions (x = -x). At the same time the underlying symmetry group H4
is reduced to its subgroup H 3 , the binary icosahedral group [1].
22 J. Patera

Setting Cl = C2 = C6 = C7 = 0 in (14) effectively cuts from Fig. 1 the two

vertical pairs of nodes on the left of the diagram. The remaining diagram is of
type A4 (cf. (28)). Resulting quasi crystals are 2-dimensional with symmetry
group H 2 , the dihedral group of order 10.
Among the other possible restrictions on the pentacrystal maps let us also
mention the following three:
If Cl = C3 = Cs = C7 = 0 in (14), we have in Fig. 1 a disconnected diagram
of two vertically aligned pairs opf nodes (the diagram of type 4A 1 ) (cf. (28)).
Resulting quasicrystals have rectangular local symmetries.
For C3 = C4 = Cs = Cs = 0 in (14), we have in Fig. 1 a disconnected pair of
A2 diagrams (cf. (28)), and we get quasicrystals with hexagonal symmetries.
The maps 11'11 and by 11'1. for I-dimensional quasicrystals arise by retaining
only the subdiagram of (any)one pair of vertically aligned disconnected nodes
in (14).

4. DEFINITION OF QUASICRYSTALS

For any fixed convex polytope P E lllI we define a penta crystal as the set EP of
points x,

determined by the condition:

EP = {x I x' E Pl· (17a)

According to (14), x and x' can be interpreted as a result of the projections

11'1I(X) and 11'1. (X) of X E]R4 respectively. The polytope P is called the accep-
tance domain of EP.
Among the properties of P with important consequences for our construc-
tion are the convexity (the closure of pentacrystals under the quasiaddition of
Section 7 would fail without it) and simple connectedness (to guarantee the in-
flation invariance (21) below). However to make things simpler, we now replace
P by a ball of radius r and denote the pentacrystal by Er,

(17b)

where N(X') is the norm of x' calculated either from its quaternionic definition
(7) or using the scalar product matrix in (16).
Directly from the definition follow the properties:

ErE r c E rs ; (18)
Er + E r c Er + s ; (19)
E r cE S
if r < s; (20)
if x E Er then TX E Er/T c Er. (21)
 THE PENTACRYSTALS 23

This property is called the inflation symmetry of Er with respect to the origin.
Note that for it to be present in the general case, the acceptance domain P
must be simply connected.
The quaternions can be percieved as points of a four dimensional real space.
Consequently the quasicrystals of (17) are four dimensional. Our definition
makes no reference to the 8-dimensional space VQ. However, in view of the
maps ?Tn and ?T.l of (14) and Fig. 1, the points x and x' of (17) can always be
taken to be the result of the mappings

?Tn (X) = x, ?T.l(X) = x', X e VQ • (22)

In the 'cut and projection method', the cut is the inequality on Ix/l and the
projection is the ?Tn-map. Moreover, rather than choosing X from the 8-
dimensional space VQ , one may consider it to belong to one of the subspaces
described at the end of Section 3. In this way one gets a variety of lower di-
mensional quasicrystals with or without the 5-fold symmetry. Some of these
options remain largely unexplored. Equally unexplored are the possibilities of-
fered by the fact that X need not belong to a root lattice of Es. It suffices that
XeVQ .
Finally let us underline that for a realistic definition of the quasicrystal
an additional requirement is indispensable. The number of points which the
definition (17) would have us to consider is infinite. It includes all the points of
the quasi crystal and all the points falling outside of the acceptance domain. In
real life one can see, and indeed there could exist, only finite quasicrystals, each
occupying a volume R. Consequently the acceptance domain condition must
hold only for points x in R. It is remarkable that our definition of quasicrystals
is easily amendable to this requirement.
In practice one always works with the definition

EP(R) = {x I x e R and x' e P}, (23)

where P and R are given convex simply connected domains.

The symmetry between P and R in (23) is evident: there are two quasicrys-
tals simultaneously defined: points x in R, (quasicrystal EP(R)) and points x'
in P (quasicrystal ER(P)). It is often convenient to choose P and R of the
same shape if not of the same size. Both quasicrystals coincide when P = R.
It is conceivable that the duality between EP(R) and ER(p), and through
it the finite size of real life quasiperiodic materials, is of fundamental nature.

5. PHASONS

The origin, x = 0, is a point of the quasicrystals (17) and (23). As a con-

sequence [11 one gets the global H 4 -symmetry of the quasicrystals. That
symmetry disappears when the origin is shifted off quasicrystal points.
24 J. Patera

Let () E V be a fixed vector which we call phason, and let us generalize the
quasicrystal definition (17a) using the projections 1I"1I«(») and 1I".L«(») as follows.

(24)

Clearly Eo of (24) is E r of (17b). The acceptance domain of E~ for () =I- 0 is

not centered at the origin anymore. The fixed point of the H 4-symmetries is
now 11"11 «(») and, in general, it does not belong to E~.

6. QUASIADDITION

The quasiaddition is a recently defined binary operation [2] in the quasicrystal

space whose importance to the theory of quasicrystals cannot be exaggerated.
In particular all our pentacrystals tum out to be closed under it.
Suppose Lr is a linear space over F of dimension n < 00. The quasi addition
in L, is defined by
(25)
The non-commutativity and non-associativity of quasi addition can be veri-
fied using the definition in a straightforward way for all x, y, u E Lr as well as
the following properties:

Tx(x) := x I- x = x; (200)
x I- (x I- y) = x I- y; (26b)
TxT; = TvT;; (26c)
(x + u) I- (y + u) = (x I- y) + u. (26d)

The closure of E P under I- is equally easy to check. The acceptance domain

condition,
P 3 (x I- y)' = r,2 x' - r'y' = (1 + r')x' - r'y', (27)
turns into the convexity condition for all x, y E EP. Indeed, the coefficients of
x' and y' are both positive and their sum is 1. Consequently for (x I- y)' to be
in P, the convexity of P is sufficient.
The emergence of this operation immediately suggests many interesting ques-
tions. In order to finish let me list some-of them:
• What is the structure of a set of seed points from which repeated application
of quasiaddition produces a 'reasonable looking' quasicrystal in a space of
given dimension n ~ I?
• What can one say about the Fourier image of the orbits of quasi addition
from the previous point?
• Description of equivalent sets of seed points. What are the minimal sets of
seed points for each n?
• Study of quasicrystals in terms of the properties of the seed points.
 THE PENTACRYSTALS 25

• How does one decompose a given quasicrystal into orbits of quasiaddition

(corresponding to some minimal sets of seed points)?
• Can one invent local growth mechanisms using quasi addition?
• Using a union of several quasiaddition orbits corresponding to different
'atoms', what new structures can arise?
• What identities other than (26c) are there for the operators Tx?

7. EXAMPLES

In order to be able to draw examples in a plane, we are limited to quasi crystals

in 2 dimensions.
EXAMPLE 1. We thus consider the subdiagrams of the Es-diagram of Fig. 1
consisting of two pairs of vertically aligned nodes. There are three different
diagrams which one obtains. Using the standard nomenclature in Lie theory,
these are the diagrams of types
o 0
o 0 (28)
Al + Al + Al + A l .
For quasicrystals in 3D, one would have the following subdiagrams to consider:
0-0 0
0-0 0
D s,
EXAMPLE 2. To each of these diagrams corresponds a root system of the Lie
algebra of that type. Its root system is a finite subset of points in the root
lattice Q of E s , and its root lattice is a (root) sublattice of Q. Using the
sublattices of (28) or (29) in our definition of quasicrystal, we get quasi crystals
of dimension 2 or 3 of the type indicated by the corresponding diagram.
In this example let us see the effect of the mapping 7!'1I on the three root
systems of (28). The results in Fig. 2 indicate the type of local symmetry
displayed by the quasicrystals of the same type. The symmetry in the case of
A 4 , 2A 2 , and 4Al is respectively 10-, 6-, and 4-fold.
EXAMPLE 3. Examples of 2D quasicrystals (17) of the three types (28) are
found in Fig. 3. Note the global rotational symmetry around the origin indi-
cated by the cross.

EXAMPLE 4. The A 4-quasicrystal of the previous example is shown in Fig. 4.

The origin is indicated by a cross. This time there is a non-zero shift <I> of the
origin. In one case the origin is shifted to a point of the weight lattice of A4
which does not belong its root lattice. The basis of the weight lattice are the
26 J. Patera

0 1+02 0~03
O2
• • • • 01+0~03

04· • 0)+04

0, 03
• • + • •
• •
• • • •
• •

fJ2. • 0
• fJ,+fJ2
O2 • • 0,+02 • -y

• • + .
0, fJ,
• • • +
0
•
fJ

• •
•
• •
•

Fig. 2. - The root systems of A4. 2A2. and 4Al after the projection 11"11. The Es sim-
ple roots of Fig. 1 are renamed according to the conventions usual for each subalgebra.
Only positive roots are labeled.
 THE PENTACRYSTALS 27

... :O::: : :::O:n

. :::.::::.: ::::::.::: :.::: .
••••••••••••• 0 ........

~ ::!:!!: !:: ~:! :: ! :: !:! ::!:!!: ! :: !.

.: :::::::0:: : : 0:::::::0
................... .
::::0::::: :::
.......... .

:: :-: :::.::.::: :.::.: .::.::::.: :.::: !-:::

........ 0 ....

:": :_::": "::" :::: 0: :0: 0::0::: ;.: :0: 0::-::":

.::::: ::::::::::: : .::: ::::.:::::::::::: :::::
••••• _ •••••••• • • 0 ......... . . _ ................. .

f~; ~ ;~ ~;~ ~; ~ ;~~; ~ ;~ ~~~ ~; ~ ;~ ~;~ ~; ~;~ ~; ~ ;~ l~~ ~; ~ ;~

•••••• 0 _ . . . . . . . . . ...... . .............. 0 •••• • • __ ••• • •
I

;)~ 1~ 1:1 ~ 1~ll~ 1~l l~ll~ 1~ll~1 l~ 1~ll~ 1~ll:ll~ 1~

~ :~ :~:~:::::::.::
~~ ~ :~ ~: ~ ~~::.:::::::
~~~ ~: ~ :~ ~:~.::::
~~ ~ .:::::::
:~ ~: ~ ~~ ~:~. ~/
-i :: ! :~ i: ! :: !:! ;: ! :; !:!:: ! :! i: ! :: r
.:::.::::. :::::::.:: :: .::: :.,.. ....
. .: . .:-:: ::: :: ::: :~i :: :.. /'

Fig. 3. - Circula r window view of quasicr ystals E lo of types

A4. 2A 2. and Ai re-
spectively.
 loS
3'§
t.~
3 ~
+ ...
•·.···::~-
.......; •::v~
..:..: .:...::··;;::· ·'::-i~ .....::/~:;:~:~:~~~
~, 3§
-.._............
::.;: I.- .•. -
:. I···:'·~·~·
*•• e._ .,
·':.. it:::···::-r::'·~:~·'::-i,.::...
I:~I·~ ..
••- .... . ... ~:,:.::4!:
'I:.·:·.:,·:·. ·:·.i,:~!.::.:. ::.!~:,:!=::.:~'~
.. :'.::·::.':·.j.·:· ••:.,,; .. :.~_
II ~
0&
:: r::·-'::·-;
III
.,:~;.::,::::~~:~~::::.::,;~:~:::::::::~;,::.::::~=:.
I ~ :':~.~:'i::i.· ;.: .;.~:.: .::;. ~:,;:: i:::': ~.(::. :.: i.·;·:·:
.. •
.. ..
.,.:~i.::· -::.i.::;.o:::~:::.i.:::'.=::ci.:::.;'::·.
-.: :::,:.:::.:,::: ::.:.:: .:::~. ::.:.:: :::,:.::~:,
··::,!'::'·
·::.i ••
:::--.
.d'iil
... =
.~ 0
.
:•••••:::i::·'::'!
~
:'·::i.:··:••:. i:: .':i:: ••:'.~ i ~··:··:o:··:: ~: ..:.. :. i~ .": •• ::i:1.
•• 1: • • • , • • • • • • • • • • • • • • • • , ••• , • •
•••• 'e::-:". • :-'••• : C=:···::'·· :-·~::··-::
·el·:'- : ..... : .,:.'•••'!'::'·::i.
: •••'.:.. ··'::'!,:.
: ••••• : • .: S ... ~
~ g 0
:-= .... ••• :
••:•• ::.":••:..:•.:.,
••:...:: ••:....... :•••••:,•.:.,.=-: ..... I·~!~::::::~:::~j,i::::::·i~:::~i~:::~!·:::::::·!~:::~::::::~!::
-••' . :.. : •••••
:••:••:'.::. ~ :.~:_':*1.- c:. :•. :... :
•••:••
•••·_O-' :.;f•••••• _*tI•••••:~ci.·.·. .1"'
. . . ·····0····
.. :·'..I,:···.i.··'
... .·::··~C:.'··.i. ··· .···.i.··".•....•
::'·:: '..••..•....•"'•....;-
••i.·:".::· J!l~~
i ... i .. ·'.:'.':.'......... ···,;.···.0
e:):· ::::': ~:~:'::::.: ._, :::.....:::,:, ••~:::::::
• I . . ..11 ••• ~•• , ••:::I ~:,:

a •••••
::~: ~ :,::.::: :,:~: , ...
:~.
··· • :····:···.i: :.,·-=:' i.···:···
.. . ::···::,·' .., •.• :'·::i.·
.."fIIII ..• .• -. i::·'::'!': ft... ••.. :···.
~'O=
....... 0
E-.h
'._ •• 1 ._ • .)... , •• I ~••:.::::, ':~i ::.•.::........
r::··::::··'
:r••••:,=-;-::'
•••••••"··::'
••••:""::··:=-i ::•••••••••••
:: r::'·c··:::·••
~
~..
~ -=:c-•• •••::••• :·• ::-.-:: .-:::-. ::.,.:: .-=:,••:::••,::. ::.,.:: .-:::-. ::.,~
:: .-·:'··::::·
••,.~.-:.~........... ••••••••••• '~
, •• •-:~,.
.•••• :'--:\ .•/:l:·••.·:'.-:· •..• :••.,.~.-:.,.•• :•..• ·:l:·••.·:'.-:~,~ ,:::,::·:·::.i~:':::::::':::i.::·:==:,::::~:,:~:·::.i~:'::::~:':~i.::·:·::.: .~ .8 .8
~ :••••.: :••..'.:-.
!.. :,.!=.~\: •• !.. :...:.. !.. :.. !t.:••,: ••!.•:,. .:.<!-.: ..: •• :•••: ••!.. :,~ I..~'" ·tIII· .....•• :'I•• , ••' -.0 ••••• ~ •• , ......... ' •• , ........... ......
·: .::.~:r: •• ::.::.,: •• : .:,.:: ••: •• :.-:•• ::.,:.-:•• :,-:: ••: •• :•• :~::.:::r:
l(l ......
=
....; : ••••• : .-:.,••• ! ••.,.:. : •.••. : .-:.:-. : ••••• : .-:.' ••• ! •••'.:. :
~~:::.i.::~. ·::o·:::.~::· .-::.i.:::.:::::.:::.i.::· .·::,;.:::.o:::.:::-i.:::.
9' .!0 ....
::I
:i'::···:!~··:··;,!:··'::-r::'·::!::·,::-{::···::'i-::··~!::·,:=-i-::'·:::·
~:::~::::,:~;h::::~:::~;':::::::';~:~:::::::~:~;'::·::::~~:~~::::' .: .-:::-. ::•••:: •::,••::••.,::-. ::.,.:: .-::c-.:::••,:: •::.,.:: .-:::-. : ~Z Q)
.8
.., ... :,:.::_.:.:
:.:..•. ::::,:::.:: .. ~.:::-:.::·
.... ... ,.:::I:::i:::':::·:::·:·::,.:.:.: ••::••
::•....:::..r::'·~:::·
.. .. .-, .." .. .. . i'l::···::'r::'-':::·I::i'::-··:~'··:"'::···::r::'·:::~·':':-i~::···::"
~:::;:::::~i'::·:::;~~:~:~::::':~;·::·::::~~:~~;:::·:;,;~:~:::::: . "0 3' !
:,.::, ..:-.:
···':=-i ::.,:- :..:,.::.:: ..:..: .:..::. :-.:..:..~.,:. : .:,.
~.. ~.. ~ 'o:t: .0
':~r::'··:~~r·:·-i,::···:· i-::'·:::~-':: ·:i-::'·::·
••••••••••'::·-:":"·: + '.5...
···:~I·~:'·.,::·::·':':- .II!
.... '....... ' _.. i.···
....-.:;'..;.----~.....-·.i.··'....C:.'••-i.··· .-
...... ~ i ·:···:·-:··:··~·::-'::
~ -i
0
~ ••••• :r........... ... " •••••••••"": ••••• •••• •••
i::.'::'-'::'.::i ...•:.. ~~:•.....
>
...:.... •••••:.::••••• :I·::i::·':
'.:'.':.'.!-··~ci:'·::i
': ••'.:.. .:··:··:.i:
~'
:·':c:··:'·::i::··:··~ci::'
o ~ i
I ~:'.::i::.'::'.~:
:•••••: .':.' •• :"=.'.:'.":.'•••: •. ~~-
.":~:.::'::::.:~:~:'::::.::I:~:~:::::::~:~:'::.::::':~. ..·~;';::::::';~:::~;~:~:~;~:::::;~:::~i':::::::'::: . -g.
-:·.~::•••::,!~:,.:::~.,::,:~::.-·::,!.::·.t::·.,
-:.• .::c..::: •.':: •::.,.:: ..:::.. ::.,.:: • .~.:::::::~:~;.::.::::~~:~~::::.::.;~:~::::.:.. .~ ~ ~
. •••.:•..~"
• I·.
: .::•.:.0··
•••
:·.:,·::.::·l •
-t ••••
• 'I••_: ..
'I•• I ••' ...
...! •
....:,.::...:..•••:..:.:..i··:'·::·::
-::'·::~C:··:··i :
~
::.,:.... ·'~ '"
i~]
",0-

N (.)
0+.8
I ...... =
3 ....
=
to-
..,po
.!!P II .t;:'6'0
00 r.t..o&o
N
 THE PENTACRYSTALS 29

/ .:.: :.:.{::.. :~::-:':.: .

... ............
..
.... . ......
.....
.. ...
................ ..
.........-;--.--
......... ... : .. .
:.....................
:~

.. .... ...... .... .. .. .

...... .... ......
....... .. .....
.... ......
\\
.........
.. .... .... .. .... .. .. .
.... ......
. .... .... .. .... .. .. .
...................
.. ......
.../. . ...... ... ............. ..... ....
.. ......
"I .. .... ...... .... .. .... ...... .... ..
t .. .... .... ...... .... .. .. ..
...... .. ..
.... ............... ....
... .. ... ... .. ... ...... ... .. ... .. .... . .
~
:.. •.........: :.-••:: ...-.: .. ..
r
.. ...... .... ...... .. .... .. .... ...... .... ...... ....
f ........-........: .................... . ............:.... ....... .... .. . ..
~

\ ... .. .. .. ...: ....: ........-: ........ . ......: ...: ......:: ..........: .. .

.... .... ...... ...... .... .. ..
.. ....:....
...........:: ................
_. •••• •• •••• •••• •• •••• • "I
~ • • • • • • • • • • • • • • • • • • • • •* ./ . '•
.... .. .... .. .... .
... .. ............
.... .... .... . .
.. .. ...... .. .: ....... .........
r ~::~~:::~:::~~::~'
* •• * • • *. * • • • •
.. .... ..
...........
...............
....... .......
••••

. ...... ..
\
.. ....
· ..... ......
....
... ....... . ..... .
. : .....: ....
•••
-........ /
•••••• y
-.!~~. ~!.--

.. .... .... ..
/ ..
"" ...... .
. .. .... .. . .. .
..............
.. ....
....... .... .. ..... .. .... ......
. . .. ....
.......
. .... ...... .. ... · ... ... ... ....
.. .. ... ..
.... ... ... ..
.. •• + ••
. .. .... .... .
.... .. * • *.
• •..
... ... •......
· ... *.
• • * ••••••• ••
• . ••
.. ... ... .
• •• • •

.... .... ..... ..

.....
... ..... ... . ..
... ..
\
..
.... ......
\
\\ .... .... ..
. .. . .. .. . .. .. .. ..
. .........
. ... ... ...

• + •

Fig. 5. - Three examples of generation of a quasicrystal by quasiaddition (25) are

shown. The seed points (black squares) are vertices of regular polygons: decagon,
hexagon and square. Two stages of the generation are shown for each of the examples.
30 J. Patera

fundamental weights Wi, .•. , W4' Note the reduction of the global rotational
symmetry around the origin to a 5- fold one. In the second case the origin is
in a generic point which not in the weight lattice. There is no global rotation
symmetry left.
EXAMPLE 5. In Fig. 5 we show two stages of multiple quasiaddition for three
sets of seed points: vertices of a regular decagon, hexagon, rectangle. The seed
points are indicated by black squares. The generation proceeds by quasiaddi-
tion of vectors connecting the origin and the seed points to the known points
of the quasi crystal according to (25).
The examples of this notes were calculated and printed using version 2 of
the software [4].
As a final remark let me point out a new mechanism for generation of qua-
sicrystals [13] in essentially any dimension, which is unlike any previously
proposed method. It is based on the results of Y. Meyer [14,15] and the
amenability property of certain polytopes [13,16]. The process begins with
the initial seed cluster of sites from which growth proceeds radially outwards.
In its simplest formulation, which can already produce an infinite variety of
quasi crystals, the only global information is supplied by the choice of one or
several constants determining a continuous background field for the growth
process, which then proceeds locally by attempting to mimic this field. In
scope the process can generate most of the possibilities offered by the 'cut and
project' method, but has the advantage of avoiding its awkward non-physical
aspects. More precisely we have a notion of an ideal local configuration Cile)
of points, for example vertices of a icosadodecahedron, or the decagon in 2D.
Starting from a finite seed (for example, a single point), each existing point
of the growing quasicrystal views itself as a center of a potential ilc. Exactly
which points of the ilc can actually be accepted as new quasicrystal points
depends on how well some local phase conditions match some globally imposed
continuous background field.

References

[I] Moody R. V., Patera J. J. Phys. A: Math. Gen. 26 (1993) 2829.

[2] Berman S., Moody R. V. J. Phys. A: Math. Gen. 27 (1994) 115.
[3] Moody R. V., Patera J. Can. J. Phys. 72 (1994) 442.
[4] Moody R. V., Patera J., Rand D. simp Lie, Macintosh software for rep-
resentations of simple Lie algebras v. 2 {Publications CRM, Montreal,
1994}.
[5] Moody R. V., Patera J. "The Es family of quasicrystals" in Noncompact
Lie Groups and Some of Their Applications, E. A. Tanner and R. Wilson
Eds. (Kluwer Academic Publ., Boston, 1994) p. 341.
[6] Moody R. V., Weiss A. J. of Number Theory 47 (1994) 405.
[7] Witt E. Abh. Math. Sem. Univ. Hamburg 14 (1941) 289.
 THE PENTACRYSTALS 31

[8] Tits J. J. Alg. 63 (1980) 56.

[9] Champagne B., Kjiri M., Patera J., Sharp R. T. Description of reflection
generated polytopes using decorated Coxeter diagrams, CRM-preprint
1877 (1994).
[101 Moody R. V., Patera J. J. Phys. A: Math Gen. 25 (1992) 5089.
[111 Moody R. V., Patera J. Can. J. Math. (to appear).
[12] Moody R. V., Patera J., "A Coxeter diagram description of faces of
Voronoi cells and their duals" in Proc. of the Colloq. on Group Theor.
Meth. in Phys., Salamanca, Spain, Anales de Ffsica, Monografias vol 1
(1992) p. 438.
[13] Moody R. V., Patera J. Local Dynamical Generation of Quasicrystals,
preprint (1994).
[14] Meyer Y. Algebraic numbers and harmonic analysis (North Holland, New
York, 1972).
[15] Meyer Y. "Quasicrystals, Diophantine approximations and algebraic num-
bers" in these Proceedings.
[16] Chen L., Moody R. V., Patera J. Non-crystallographic root systems and
quasi crystals (in preparation).
 COURSE 3

Elements of a multimetrical crystallography

A. Janner

Institute for Theoretical Physics,

Toernooiveld, NL-6525 ED Nijmegen,
The Netherlands

ABSTRACT

Different metrics can be assigned to a given lattice of symmetry translations

in such a way that the lattice is' left invariant by circular and by hyperbolic
rotations, respectively. Therefore, a crystallographic point group can be de-
fined having as generators elements of orthogonal groups of same dimension
and different signature. This explains the term multimetrical crystallography.
Combining lattice translations with those point groups one obtains multimet-
rical space groups, which can be interpreted as symmetry groups of point-like
crystal structures. The hexagonal close-packed structure and the Wurtzite
crystal structure are discussed from this point of view and their multimetrical
symmetry derived on the basis of the theory of integral binary and ternary
quadratic forms. The connection with tbe group of unit.s of quadratic fields is
briefly explained.

1. INTRODUCTION

The motivation for extending crystallography to multimetrical symmetry finds

its roots in the belief, that the basic laws of crystallography are dimension
independent. Accordingly, a quasicrystal described as a lattice-periodic struc-
ture in a higher-dimensional space (the superspace)[l] shares the properties of
a normal crystal in space and conversely. In other words, the idea is that one
34 A. Janner

can consider for a crystal in space the same type of symmetry transformations
as one has in the superspace description of a quasi crystal [2].
Considering first the structural relations bet.ween a quasi crystal in space and
in superspace one recognizes that lattice translational symmetry in superspace
leads by projection to a Z-module of symmetry translations in space [3-6].
In real quasi crystals that symmetry is broken at large distances (because the
crystal is finite) and at short distances (because the at.oms are of finite size
and form a discrete structure). The first limitation is expressible in terms of a
crystal growth form (in space), the second one in terms of an atomic surface
(in the internal space). Both together define a superspace region separating
the infinite set of space-group equivalent positions into a finite set of real ones,
describing the location of existing atoms, and the remaining virtual ones. In
a normal crystal such a distinction exists in space already and it is considered
a surface effect. In a quasicrystal two are the surface effects: one at large
distances, implying finiteness, and one at short distances, implying discreteness.
In the same way as the crystal growth form reflects at large distances ro-
tational point group symmetries, the boundaries of the atomic surface reflect
at short distances in space scaling symmetries (if present) modulo the set of
Z-module translations [7-9]. In the superspace all these are crystallographic
point group transformations represented on a lattice basis by integral invertible
matrices. By the rules of the superspace approach [10-14], rotations in space
are represented by circular rotations in the superspace, whereas pure scalings
appear as hyperbolic rot.ations[3], the combination of the two giving rise to
crystallographic multimetrical transformations[15]' which in space appear as
homotheties [16].
The term multimetrical arises from the property of circular and hyperbolic
rotations to leave invariant a positive definite and an indefinite quadratic form,
respectively, both defined on the same (affine) lat.tice of symmetry translations.
This ensures the compatibility of the two (or more) quadratic forms and implies,
in particular, that these real quadratic forms are equivalent to integral ones.
Therefore, a multimetrical cryst.allographic group, to be well defined, requires
the choice of a set of quadratic forms assigned to a given lattice of symmetry
translations. Many choices are in principle possihle and one has to learn how
to select the relevant ones, from t.he mathematical and from the physical point
of view.
The statement that a quasi crystal in superspace is like a crystal in space im-
plies the existence (and the relevance) of cryst.allographic multimetrical sym-
metries for normal crystals as well [2,6]. \Vhat this means is presented here on
the basis of ideal close-packed crystal structures. A first attempt to identify
possible physical consequences due to the multimetrical symmetry of a crystal
has been published elsewhere [17]. Here the considerations are restricted to the
geometrical level only. Even so, the whole still has a pronounced exploratory
character.
 MULTIMETRICAL CRYSTALLOGRAPHY 35

2. CLOSE-PACKED STRUCTURES

2.1. The 2-dimensional case

The centers of close-packed circles of a given diameter a in a plane form an

hexagonal lattice Ahex with basis vectors ai, az given with respect to an or-
thonormal basis el, e2 by:

al = a[l,O], (1)

)
and metric tensor
go(a) = a 2 (
_~1 -~1 . (2)

The interstitial positions with 3-fold site symmetry, at (~+nl )al +(~+n2)a2 ""
(~, ~) and (~+nl )al +(~+n2 )az "" (~, ~), are denoted by Band C, respectively,
whereas the lattice points nlal + nZa2 = (nl,n2) "" (0,0) are at the positions
A.
It is sometimes convenient to consider the hexagonal lattice as a centered
orthorhombic one spanned by the orthogonal basis vect.ors bl , b2 where

bl = al = a[l, 0], (3)

The inverse basis transformat.ion is t.hen

(4)

2.2. The 3-dimensional case

Starting from a close-packed planar layer of spheres wit.h given diameter a
at positions A at height 0, two following close-packed st.acking of layers are
possible: one at B and one at C, both with centers at a height of Aa. A
close-stacked structure of equal spheres is t.hen given by a sequence of layers of
type A, Band C at neighboring dist.ance Via,
the only condit.ion being t.hat
adjacent layers never are of the same type.
Two are the most important cases: the hexagonal close-packing defined by
the period-two sequence .. . ABAB ... (or ... BCBC ... , or .. . ACAC ... ) and
the cubic close-packing defined by t.he period-three sequence ... ABC ABC . ...
In the first case, the centers of t.he spheres define an hexagonal lattice of
translational symmetry A hcp wit.h basis

al = a[l, 0, 0], (5)

where the components are expressed with respect to an orthonormal basis

el, e2, e3·
36 A. Janner

In the second case, the translational symmetry is given by a face-centered

cubic lattice, which is also expressible as a centered hexagonallatt.ice with al,
a2 as above and
a3 = a[O, 0, v'6], (6)

representing the body diagonal of the cube. The centering posit.ions are at
(~, ~, ~) and (~, ~, ~), with components now expressed with respect to the
hexagonal lattice basis (a) = al, a2, a3·
Between two adjacent layers there are three types of interstitial positions,
two with tetrahedral and one with octahedral nearest neighboring coordination,
respectively, symbolized correspondingly by 'V, 6 and D.
Many inorganic crystals can be described as a slightly deformed close-packed
structure with (partial or full) occupation of tetrahedral and/or octahedral
interstitial sites [18]. In Table I a number of close-packed non-metallic com-
pounds is given with octahedral and/or tetrahedral coordination of met.al ions.
Given is the Strukturbericht type (SB), the chemical formula of a representa-
tive compound, the close-packing formula (nb.' no, n'V, n. )cp, with nb., no, n'V
and n. the occupation ratio of the tetrahedral 6, octahedral 0, t.et.rahedral \l
interstitial sit.es and n. of the close-packed posit.ions, respectively. The sub-
script C P characterizes the stacking order of the close-packing: hexagonal (h),
face-cent.ered cubic (e) and ot.her (he). In general, the real structures deviate
somewhat from the ideal close-packed case considered here.
The translational symmetry of an ideal close-packed structure can be given
in terms of a centered hexagonal lattice with basis vectors al, az as in Equation
5 and a3 = a[O, 0, N ~], or as a centering of an orthorhombic lattice with basis
vectors

bl = a[I,O,O], b2 = a[O,.J3,O], b3 fi
= a[0.0,NV3], (7)

with N the integer of the stacking periodicity. Labeling by n the n-th stacked
layer of the close-packing starting with a layer (A) = (AD), we denote by
An, Bn and C n the corresponding representatives of t.he close-packing, and by
6A n , 6Bn , 6Cn , \l An, \l B Il , 'VCn and DAn, 0 Ell, DC" the corresponding
interstitial positions. So for example in the (a)-basis one can have

(8)

and so on. Of course, in a given structure, only part of these possible positions
is occupied by atoms.
 MULTIMETRlCAL CRYSTALLOGRAPHY 37

Table I

SB Formula Name Close-packing Formula

B1 NaCI Rock-Salt (0101 )c

B3 ZnS Zincblende (1001)c
B4 ZnS Wurtzite (1001)h
B8 NiAs Niccolite (0101)h
C1 CaF2 Antifluorite (IOll)c
C6 CdI2 (O!OI)h
D51 a- A12Oa Corundum (O~OI)h
D57 'Y-AI2Oa 'Y-Alumina (L2~I)c
Ell CuFeS2 Chalcopyrite (1001)c
E21 CaTiOa Perovskite (OiOI)c
E22 FeTiOa Ilmenite (01 01h
E5 1 (Fn, Mn)Nb 20 6 Niobit.e (01 OI )h
H11 MgAI 2 0 4 Spinel (~'i~l)c
S1 2 Mg 2 Si04 Olivine (S!sl)h

3. MULTIMETRICAL SYMMETRY OF THE 2D HEXAGONAL

LATTICE

One can forget what has been said in the introduction about the scaling sym-
metry of quasicrystals and simply considers [2] an hexagonal lattice Ahe:z:. One
defines on it, in addition to t.he positive definite metric tensor go given in Equa-
tion 2, an indefinite metric tensor g2 for the same basis (given in Equation I):

(9)

by choosing for the "pseudo-time" axis the vector eo. with negative squared
length -I, instead of the vector e2 of the orthonormal basis considered above.
(One could also consider gl = -g'). choosing the "pseudo-time" axis along e1).
In this way Ahe:z: becomes a multimetricallattice invariant not only with respect
to a 6-fold circular rotation R by an angle ¢ = arccos ~, but also with respect
to an hyperbolic rotation L by an angle X with cosh X = 2. With respect to
the corresponding metric one has

RUoR = go, (10)

38 A. Janner

where tilde denotes transposition. Indeed, on the orthonormal basis one has

4 -~ )
R(e) = ( 1
'2
'
L(e) = ( J32 J323) (11)

implying

R(e) ( ~ ) = ( 1)= a1 + a2, R(e) ( 4)= ( -~ ) = -a1 (12)

L(e) ( ~ ) =( J )= 3a 1 + 2a2, L(e) ( 4) (J) = a1 + a2,

(13)
so that on a lattice basis both are integral and with det.erminant 1:

R(a) =( ~ -1 )
o ' L(a) = ( ; ~). (14)

Both rotations, Rand L, are generated by reflections, Euclidean and hyperbolic

reflections, respectively, according to:

(15)

with mo(a) = (-~ ~) a reflection leaving both 90 and 92 invariant. There

is no non-singular met.ric tensor left. invariant. by the product transformation
RL, which is a parabolic one:

R(a)L(a) = (~ ~). (16)

As one can verify in the ·present case, a parabolic transformation has the same
invariants as t.he ident.ity (trace two and determinant one). RL is, nevert.heless,
a meaningful symmetry transformation. In the quasicryst.al case a homothety
(combination of a rot.ation with a scaling) is described in t.he superspace by a
product of a circular and an hyperbolic rotation. The absence of an invariant
metric for RL, whereas each factor has its own natural metrical invariance, has
been the main reason for introducing the concept of a multimetrical t.ransfor-
mation. A precise definition has been given in a previous publication [15] (still
at a working level only). Here it is sufficient to recognize the concept at an
intuitive level allowing to verify t.hat, having defined for the hexagonal lattice
Ahex the metric tensors 90 and 9'], this multimetricallattice is left invariant by
the point group J{ generat.ed by R, Land mo :

J{ =< R, L, Ino > . (17)

Before going over to the much richer multimetrical symmetry of the 3-dimensional
hexagonal lattices, it is convenient to consider the problem in the frame of the
binary quadrat.ic forms with integral coefficients.
 MULTIMETRICAL CRYSTALLOGRAPHY 39

4. BINARY INTEGRAL QUADRATIC FORMS AND

QUADRATIC FIELDS

Consider the real metric tensor 9 = (a, b, c) left invariant by a crystallographic

point group transformation A and a lattice A spanned by the basis vectors
aI, a2 having 9 as metric tensor: gij = ajaj, so that

AgA = g, with A E Gl(2, Z) and 9 = (b/~ bl ;). (18)

In the Euclidean case (g positive definite) A is of finite order and in that case
there are always non-trivial invariant. solutions g. That is not. necessarily so in
the the multimetrical case, as we have seen for ,4 = RL. If there are non-trivial
solutions 9 = (a, b, c), then also rational ones and thus also for a, b, c rational
integers and relatively prime. We can, therefore, restrict the considerations to
that case.
The squared lengths of the latt.ice vectors 1110.1 +1120.2 = (n1' 11.2) with integers
11.1,11.2 define a binary quadratic form f =
(a, b, c) (see, for example, the book
by Buell [19] and references therein):

(19)

with 0.= =
911, b = 2912, C 922. In our case the form is integral (a, b, c E Z) and
primitive ( gcd(a, b, c) = 1). Such forms have been ext.ensively investigated by
Gauss [20] and the geomet.rical point of view has been developed by Minkowski
[21]. (See the book by Hancock [22] for further developments). The discrimi-
nant of f is ~ = b2 - 4ac. If ~ < O. the form is positive definite and if ~ > 0,
then f is indefinite. An element of Gl(2, Z) with determinant 1 leaving f (and
thus g) invariant is called an automorph of the form (a., b, c), or also orthogonal
with respect to g. All such automorphs can be written as:
l(n-bu)
A =( 2
au ~(n
-cu )
+ b1l) , (20)

with 11., U integral solutions of the Pell equation:

n2 - ~u2 = 4. (21)
Disregarding the trivial cases A =
±l. if the discriminant ~ is negative, 9 then
defining the Euclidean metric case, this equation is only solvable for ~ = -4
(the square lattice case) and for ~ = -3 (t.he hexagonallatt.ice case). We are
here interested in the indefinite forms (.6. > 0), the automorphs A are then
hyperbolic rotations, and in that case the Pell equation always has integral
solutions (n, u). Among these there is a fundamental solution (no, uo) with
no, Uo > 0 and (no + uo..Ji5.)/2 minimal. All other solutions are of the form

11 + uv'X = (nO + UOv'X)k (22)

2 ± 2
40 A. Janner

and correspond to the k-powers

(23)

of the automorph A obtained from the fundamental solution of the Pell equa-
tion. This brings us to the theory of real quadratic fields Q( VD) with elements

(24)

where D is the square-free part of a non-square integer. (See for example

[23-25]). The conjugated element of r is

The norm N orm( 7') and t.he t.race Trace( r) are defined by

Norm(7') = 1'1", T1'ace(1') = l' + 1". (26)

The quadrat.ic integers v and the quadratic units c of Q( VD) depend on the
discriminant of Q( VD) which is d = 4D for D == 2,3 mod 4 and d = D
for D == 1 mod 4. One has Z( VD) if D == 2,3 mod 4 and Z( ~ + 'If) if
D == 1 mod 4. In all the cases, however, the quadratic integers can be written

a+bVd
v= wit.h a == db mod 2, a,b E Z. (27)
2
The units c are the integers with norm Norm(c) = ±l. Accordingly, units are
in 1-to-1 correspondence wit.h the integral solut.ions (n, u.) of the equation:

(28)

The posit.ive units (N orm(c) = 1) are solution of t.he Pell equation, called now
Pell plus equation, and the negative units (No1'7n(c) = -1) of the so-called
Pell minus equation. As a square factor in ~ can be absorbed in the integer u,
one sees that the automorphs A considered above are in 1-t.0-1 correspondence
wit.h positive unit.s of the given quadratic field. Negative unit.s (if admitted by
the quadratic field) lead to negautomorphs N, which are elements of GI(2, Z)
with det N = -1 transforming 9 into -g (and f into - f)

NgN = -g, N E GI(2, Z) (29)

and that explains the name. The negautomorphs have the same matrix form
as the automorphs in Equation 20 when expressed in terms of solutions of the
Pellminus equation. The fundamental solution (no, ti~) of the Pell (plus/minus)
equation then leads to the fundamental unit co, all other units c being express-
ible in powers of co:
c = ±co,
k
k E Z. (30)
 MULTIMETRICAL CRYSTALLOGRAPHY 41

The existence of negative units depends on the quadratic field and so also that
of negautomorphs. The classical example of a negautomorph is the Fibonacci
transformation which corresponds to the fundamental unit of the quadratic
field Q( -/5):

1 + -/5 r.;
€o = 2 E Q(v5), f = (1, -1, -1), (31)

The hexagonal lattice case considered above with indefinit.e 92 = (2, -2, -1),
or g1 = (-2,2,1), has a quadratic form with discriminant Ll d = 12 and the =
corresponding quadratic field Q( va)
only has positive units. The fundamental
unit is €o = 2 + -/3, to which corresponds the hyperbolic rotation A = L
considered in the previous section :

€o -_ 4 +2v'i2 E Q( v.),
r.i3) f = (2, -2, -1), A= (~ ~). (32)

In the general case of a multimetrical crystallography, one has to take into ac-
count the peculiar properties of negautomorphs, possible combined. in various
ways, with automorphs.

5. INDEFINITE TERNARY INTEGRAL QUADRATIC FORMS

The 3-dimensional Euclidean crystallography is based on the automorphs A of

positive definite ternary quadratic forms
3
f(ni. n2, n3) = L gikni'TIk, 7li. 'TIk E 71 (33)
i,k=l

with gik = g/d elements of the positive definite metric tensor 9 of the ternary
quadratic form left invariant by the automorph A

.4.gA = g. A E Gl(3, 71). (34)

All such automorphs form the point. group J{ symmetry of a lattice A with
metric t.ensor g. As well known, J( is a finite subgroup of Gl(3, 71). This point
group induces an equivalence relation among metric tensors sharing the same
J(, so that one can always restrict the considerations to integral quadratic forms
for which the gil· are rational integers. (Due to the symmetric character of the
gik. for i f k the rational integer condition is only imposed to 29ik. but. that is
a pure conventional matter).
A multimetrical crystallographic extension requires indefinite quadratic forms
as well and in that case the point group J{ (generated by the corresponding au-
tomorphs) is of infinite order. When J( is reducible (i.e it leaves a 2-dimensional
subspace invariant) such an extension can be obtained from indefinite binary
42 A. Janner

quadratic forms in the way discussed in the previous section: then automorphs
are units of the corresponding quadratic field(s) and one has to consider ne-
gautomorphs as well. In the irreducible case, however, only automorphs are
possible because det( -g) = - det g, so that the elements of f{ correspond to
positive units of the real ternary quadratic form f and [( is the group of units
of f.
In this section the considerations are restricted to indefinite ternary integral
quadratic forms. The group of units of these forms are known as discussed
by J. Mennicke [26], who was mainly concerned with the abstract structure of
these groups for non-zero ternary forms. (A quadratic form f is a nOll-zero
form if the equati9n f = 0 has trivial solutions, only).
Denoting by F the subgroup of index two of the orientation-preserving trans-
formations of f{ / (1, -1), Mennicke derived the following theorem on these
groups of units.
Theorem 1 (Mellllicke) With the exception of a finite number of groups, all
orientation-preserving groups of units of non-zero ternary forms are subgroups
of a particular one, which belongs to the form f =
[1, -3, -3] defined by
f(nl,n2,n3) = ni -
3n~ - 3n§ With nl,n2,113 E lL a1ld which can be presented
as an abstract group by F =
{A, B, 51A 6 B4 = = =
52 AB5 I}, =
where F is a Fuchsian group.

In the present context, one deals with more general subgroups of GI(3, lL)
generated by several groups of units (positive definite and indefinite ones) and,
at present, we are unable to discuss the abstract structure of the multimetrical
point groups involved. Moreover, the arithmetic aspects of these groups are
our main interest. In that respect, the pioneering work of Fricke and Klein [27,
28] is relevant for approaching the problem. The more recent. present.ation by
Magnus [29] helps to bridge the hundred years gap.
Considered are ternary integral quadratic forms

f = [-p, q, r],
for positive, square-free and pair-wise coprime rational integers p, q, r E lL

(p, q) = (p, I') = (q, I') = 1, p,q,I' > O. (36)

The procedure for getting the automorphs A of f leaving invariant the diagonal
metric tensor g with gl1 = -P, g22 = q, !J33 = r involves the following steps (see
Fricke-Klein [27] pp. 501-634, and in particular p. 537) :
1st step: Decompose p, q, r into all possible positive factors:

(37)

2nd step: For each decomposition find the integral solutions a, b, c, d of the
First quaternary Pell equatw1I (s = 1):

a 2 plqlrl - b2 p2ql1'2 + C'2pIQ21'2 - d2p2Q21'l =4 (38)

 MULTIMETRICAL CRYSTALLOGRAPHY 43

satisfying the congruences

aplqlrl == bp2Qlr2 == cpIQ2r2 == dp2Q2rl (mod 2). (39)

Each such integral solution yields an automorph Al of f = [-p, Q, r]

according to
Al=

Hadrl + ber2)
1'-(a 1plrl - b2p2r2) - 1
%Caepi + bdp2)

(40)
In this case
(s = 1). (41)
3rd step: In addition to t.he solutions obtained in t.he 2nd st.ep find for each
decomposition of the first step t.he integral solut.ions a, b, c, d of the Seco1ld
quaternary Pell equatio1l (s = 2):

a2plqlrl - b2p2q17'2 + c2plq2r2 - d2p2Q2rl = 2. (42)

Note that in this case the congruences indicated in Equat.ion 39 are auto-
matically sat.isfied. Each integral solution yields an aut.omorph A.2 which
is given by:

PI (a 2 ql"1 + e2q2r2) - 1 q(adl'l + bel'2)

( p(ad"l - bera) ql(a 2 pI"1 - b2p2"2) - 1
p(abql + edq2) q(aepl + bdp2)

(43 )
and in this case

(s = 2). (44)

The first. and the second quat.ernary Pell equations ensures t.hat. the matrices
AI, A2 are element.s of GI(3, 7l). The trace of these aut.omorphs determines
the character of the proper transformations as an elliptic, or a parabolic, or an
hyperbolzc aut.omorph, as indicat.ed in Table II.
44 A. Janner

Table II

Automorph Elliptic Parabolic Hyperbolic

Order 2 3 4 6 00 00
Trace -1 0 1 2 3 n>3

6. QUADRATIC FORMS OF LATTICES OF 3D CLOSE-PACKED

STRUCTURES

6.1. Reduced metric tensors

Considered are the lattices of close-packed structUl'es of identical atoms repre-

sented by spheres with a diameter a. and stacking periodicity N (where N ~ 2).
It is convenient to express the lattice of symmetry translations as centering of
a primitive lattice, common to all periodic close-packed structUl'es after nor-
malizing 0. to 1. To disregard centering means to consider as equivalent the
met.ric tensors which are related by a rational hasis transformation S :

9 '" g' ---- g' = 5gS with S E Gl(3, Q). (45)

In that way, however, only part of the symmetry of t.he centered lattice is found.
It is thus a first approach where one disregards t.he stacking order which can
give rise to additional cent.ering(s). Starting from an hexagonal lattice with
basis 0.1,0.2,0.3 as in subsection 2.2, that basis is rationally equivalent to the
orthorhomhic one b1 , b2 • b3 wit.h diagonalmet.ric tensor

(46)

which is again rationally equivalent to a reduced metric tensor [/0 (P, 'I, I']
where p, 'I, r are pair-wise coprime positive integers none of which is divisible
by the square of a prime numher. Fricke and Klein showed that if p, 'I, l' do
not satisfy these conditions, there is a rationally equivalent set Po, '10, 1'0 which
does [27]. Indeed

1. A common factor 11 can be eliminated by normalization:

p = npo, 'I = nqo, I' = 111'0, ( 47)

as (p, q.l'] and [Po, qo."o] share the same point group symmetry.
 MULTIMETRlCAL CRYSTALLOGRAPHY 45

2. A square factor n 2 ( say in p) can also be eliminated by putting

p= n 2 po (48)
the matrix S of the corresponding basis transformation involving only n
(or lIn) and 1.
= = =
3. If (p, q) n whereas (p, r) (q, ,,) 1, one goes over t.o Po, qo, "0 :
(49)
by combining Equat.ions 47 and 48.
The reduced met.ric tensor of the primitive close-packed lattice PCP is:
go = [3, 1,2] (50)

because [1,2,3N2] g [3,9,2N2] g [3,1,2]. The mult.imetrical set. of metric

tensors one can associate, in a straightforward way, to the primitive close-
packed lattice Pcp are:
go = [3,1,2)' gl = [-3,1,2)' g2 = [3, -1,2)' g3 = [3,1, -2), (51)
so that Ppc becomes a mult.imet.ricallattice. For the indefinite met.ric tensors
we are exactly in the ternary cases considered by Fricke-Klein and discussed
in the previous section. (Note that. t.he case [-p, q, r] appears in t.he work of
Fricke-Klein in the notation [p, q, r]. Here we had t.o change that convention
for obvious reasons).
Before dealing with the ternary forms, it is convenient t.o consider the re-
ducible cases, which are expressible in terms of binary quadratic forms. For
these reduced metric tensors we adopt the orthorhombic p-basis, which in the
Euclidean case is defined by:
(52)

6.2. Indefinite binary quadratic forms

The reduced binary quadratic forms we have to consider are [-3, 1] = (-3,0,1)
in the (e1,e2)-plane, [-3,2] = (-3,0,2) in the (e1,e3)-plane and [-1,2] =
(-1,0,2) in the (e2,e3)-plane,
In the first case, with f =
(a,b,c) =
(-3,0,1), the discriminant is ~ 12 =
and the quadratic field is Q(V3). The fundamental unit co = 2+V3 is positive,
with n = 4 and u = 1. The corresponding solution to the Pell plus equation
(given here for u = -1 for ensuring compatibility wit.h t.he conventions of the
previous subsection) leads t.o an aut.omorph Az which is an hyperbolic rotation
around f3 given with respect to the p-basis (where now e1 -+ eo with f5 ~ -1)
by the matrix:

A:(p) = ( 023 012 1°0) . (53)

46 A. Janner

In the second case f = (-3,0,2), the discriminant is Ll = 24 and the quadratic

field is Q( V6). The fundamental unit co = 5 + 2V6 is again positive and
n = 10, u = 2 yields a corresponding solution of the Pell plus equation and an
automorph Ay (again given here for negative u = -2), which is an hyperbolic
rotation around f2 expressed with respect to the same I)-basis as above by the
matrix:

(54)

In the third case f = (-1,0,2), one has Ll = 8, the quadratic field is Q( J2)
which has a negative fundamental unit co =
1 + J2 obtained from the solution
n = 2, U = 1 of the Pell minus equation leading to a negautomorph N r , whose
square is an hyperbolic rotation Ax = N; around fl. With respect to the
])-basis (where now f2 - eo is the "pseudo-time" axis and for u = -1) one

i
finds:
o
NAp) = ( 1 (55)
1

Note, however, that automorphs of the reduced form (corresponding to a lattice

called primitive) are not necessarily automorphs (i.e. integral matrices) for the
non-reduced form (i.e. a centered lattice) and conversely. This is apparent if
one considers Euclidean point-group transformations. The four-fold rotational
symmetry of the fcc-lattice of the cubic close-packed struct.ure is not a symme-
try of the hc])-Iattice of the hexagonal close-packed st.ructure, and the six-fold
symmetry rotation of the latter is not a point symmetry of the former: both
do not occur as Euclidean symmetry of the reduced Ppc lattice. So, as al-
ready pointed out, what is presented here is a first approach only, rich enough,
however, for revealing hidden relations between Euclidean and non-Euclidean
crystallographic symmetries.

6.3. Indefinite ternary quadratic forms

Despite the fact that the groups of units of t.ernary indefinite integral quadratic
forms are in principle known, they have not yet been worked out explicitly for
the three reduced metric cases considered in the previous subsection.
For getting an example of a non-reducible 3-dimensional point symmetry
of infinite order, the attention is focussed to parabolic autol1lorphs. We have
seen t.hat. paraholic elements also occur in 2-dimensions as a multimetrical
symmetry, but not as metrical ones (i.e. as automorphs of a binary quadratic
form). Therefore, parabolic automorphs of ternary forms are necessarily non-
reducible.
Consider the form f = [-3,1,2] corresponding to the reduced metric tensor
gl. Within the principal type characterized by the decomposition P1 = q1 =
 MULTIMETRICAL CRYSTALLOGRAPHY 47

r1 = 1 in the s = 1 case, implying thus

P2 = 11 = 3, q2 = q = 1, 1'2 = r = 2, (56)

according to Equation 41 and Table 2 t.he parabolic transformation only occurs

for a = 2, The quaternary Pell equat.ion for this case (Equation 38 ) reduces
to the diophantine equation

(57)

which has b =
1, c =
3, d 2 as an int.egral solution. The corresponding
automorph is then given by:

Ao(p) = (~~ -~ -4 )
o . (58)
12 6 -5

This concludes the present exploration, which was int.ended t.o include in the
determinat.ion of muIt.imet.rical symmet.ries of exist.ing crystal struct.ures 3-
dimensional non-reducible non-Euclidean point group t.ransformations as well.
The derivation given is suited for any ideal close-packed stru('ture with a finite
stacking periodicity. In the present paper, the general scheme will be illust.rat.ed
by looking at the hexagonal close-packed case (N =
2). Results obt.ained for
N = 3 (the cubic close-packed ('ase) and N = 6 have been presented at. t.he Les
Rouches meeting and will be worked out elsewhere.

7. MULTIMETRICAL POINT GROUP OF THE HEXAGONAL

CLOSE-PACKED LATTICE

In the procedure adopted here, we consider the 3-dimensionallattice Ahcp of an

ideal hexagonal close-packed structure and add to the elements of the Euclidean
point group /(0 = 6/111111111 these words in the fundamental generat.ors of the
indefinite binary and ternary quadratic forms derived above which become
elements of G/(3, Z) when expressed in the hexagonal lattice basis a1, a2, a3
given in Equation 5. The relation between the hexagonal a-basis and the
reduced p-basis (given in Equation 52 ) is

so that disregarding the common factor 7:i t.he basis transformation is given

U
by

A(a) = S-l A(p)S, wit.h (60)

s=
48 A. Janner

and we get for the matrices of the fundamental units of the various indefinite
metric tensors derived above:

1
Nx(a) = ( 0
0
1
.1)
j .
5
Ay(a) = ( ~ (61)
o ~
<4
1

Az(a) = ( 230 011 10)

0 , A,(a) = ( -: -~ ]) (62)

Inspection leads to the following generat.ors of the mult,imet.rical point group

of the ideal close-packed mult.imet.ricallatt.ice Ahc/,:

D, C D'
1. Euclideal1 :

(l
-1 1
Rz(a) = 0 mAa) = ~l 1 (63)
0 0

m,(a) = U D' U j)-1

-1
0
m,(a) =
0
1
0
(64)

autommph, of g,(a) = i0) . .

1 -21
a' ( -'2
0
1
1
0
Q [3,1,2].
.

2. Iudefiuite:

Lz(a) = Az(a) = ( 3 1 0)
2 1 0 ,
80o ) ,
001 49
(65)
29 4 -32)
pea) = A6(a) = ( -16 -3 16 , (66)
21 3 -23

69 -28 -112)
El(a) = AoA;l(a) = ( -10 3 16 , (67)
45 -18 -73

E,(a) = .4;'.4,(a) = ( -~ -! _~) (68)

6
with hyperbolic (L., Ly), p(ar~~Oli~ (P a)l1d ellipt.ic of order t.wo (E 1 , E 2 )

automorphsofg 1 (a)=a 2 t! ~ 9[-3,1,2].

 MULTIMETRICAL CRYSTALLOGRAPHY 49

3. Indefinite:

U
8
16 )
L",(a) = N:(a) = 17 32 (69)
9 17

hype<boJic autommph of g,ca) =

1
a' ( -2 1
1
-~
-2
0)
~
Q
'" [3,-1,2].
0 0
Putting all these results together, noticing the relation E2 = my m z L z , one
obtains a multimetrical point group [{ of the hexagonal close-packed lattice
with associated metric tensors go(a),gl(a),g2(a) :

(70)

where ]{ is expressed in terms of generators. Only the relevant metric tensors

involved in the generators have been indicated.

8. MULTIMETRICAL SPACE GROUPS OF CRYSTAL

STRUCTURES

The multimetrical symmetry of a crystal structure is a multimetrical space

group G having as space group symmetry the Euclidean subgroup Go of G. It
is not easy to derive the multimetrical space groups for a given point group
K, because the latter is of infinite order. It is fairly easy, however, knowing
the generators of J( and the Euclidean symmetry space group Go to get a
multimetrical space group of the strueture, in particular when the atoms are
at parameter-free positions. That multimetrieal spaee group is, however, not
necessarily the full symmetry group for the chosen set of metric tensors.
We will now analyze the multimetrieal symmetry of the hexagonal close-
packed structure of identical atoms, neglecting possible (small) distortions. As
already mentioned in the introduetion, a multimetrieal symmetry is always
broken in real structures, because only compatible with point-like atoms. As
an example of a structure with atoms at Wyckoff positions of the spaee group
involving a continuous parameter 11. we will consider the Wurtzite structure
(ZnS) adopting not only the ideal close-packed structure for the S-anions,
requiring % = Ii, but also an ide~l tetraiJedral position of t.he Zn-cations,
requiring 1.1 = ~. Both condit.ions are very accurately verified in Z71S, but both
do not follow from the space group P6 3 1HC of the Wurtzite. Indeed, the ob-
served cj a rat.io [30] is 1.6363, whereas t.hat of the closed-packed approximation
assumed here is 1.6330. Furthermore, the Zn atom is at the center of a regular
tetrahedron and that implies the u-value given above.
These examples should suffice to demonstrate the great compatibility of pos-
itive definite and indefinite crystallography for the appropriate lattice of sym-
metry translations. More than that, the additional non-Euclidean symmetries
50 A. Janner

characterize in fact Euclidean structural regularities non-expressible in terms

of Euclidean symmetries and which appear as accidental for the space group
of the crystal.

8.1. Hexagonal close-packed structures

In many metals with hexagonal close-packed structure the ratio %is very closed
to the ideal value of Ii adopted here. The space group is P6 3 /mmc with
atoms at the Wyckoff position 2c and fractional coordinates Xl = ~, ~,~ and
X 2 -- 3'
2 1 3
3' 4·
It is remarkable, that also wit.hin the l11ultimetrical extension presented here,
the point group of this structure is the whole point group symmetry of the
lat.tice, the largest possible point group symmetry thus, and that the only non-
primitive translations occurring are equivalent either to (0,0,0) or to (0, 0, ~).
The first case occurs, for example, with t.he parabolic element P, t.he second
one with the hyperbolic element. L;c. Indeed:

PX 1
29
= ( -16 -3
4 -32) (
16 ~ 1.) ( -44 ++ 1.)
~ (1.)
~ = Xl, (71)
21 3 -23 4 3+ 4 4

and correspondingly PX 2 = -\2 (mod Ah~p), whereas

Lx X 2 ""
( =-:4~
3
)

(72)
so that

All together, the generat.ors of the mult.imet.rical space group G of an hexagonal

close-packed structure (with metric tensors go, gl, g2 defined on Ahep as above)
are:

1 1 1 1
G = < Ahep, {R z 10,0'"2}' {m;cIO, 0, O}, {myIO, 0'"2}' {111 z 10, 0'"2}' {L;cIO, 0'"2}'
1 ~
{LYIO,O,O},{LzIO,O,~},{PIO,O,O},{EdO,O,O} >. (/4)

Note, however, that because the full multimetrical symmetry point group of
Ahep, even for the chosen set of met.ric tensors, has not yet been derived, the
multimetrical symmetry of the close-packing can be larger (it is very likely so).
The symmetry group given here is already larger than the one presented 111
previous publications [17].
 MULTIMETRICAL CRYSTALLOGRAPHY 51

8.2. Tile Wurtzite structure

The Euclidean space group of the Wurtzite (hexagonal ZnS) is Go = P63mc =
C:". Two are the occupied Wyckoff positions:
Atom X at 26 w.ith coord.inates ~", v and i, i,
~ + v, where v:: ~
Atom Y at 26 with coordmates 3' 3' u and 3' 3'"2 + u, where u - s·
In the ideal approximation adopted, the atoms X form an hexagonal close-
packing with ratio ~ = Vi and we already know the mult.imetrical symmetry
of this partial structure. The atoms Yare at the cent.er of a regular tetrahedron,
formed by the surrounding X atoms, leading to a parameter value u = ~ and
the mirror m z is no more a member of the point. group.
In the multimetrical extension of P63mc (with metric tensors as above) the
addit.ion of the point. group generat.ors Lx, L y , P and EI requirf's a discret.iza-
tion of the values for the parameter tt in order to ensure that t.he number of
equivalent positions at the Wyckoff letter 6 remains t.wo, despite t.he infinit.e
order of these generators. No condit.ion is imposed by Lz which leaves t.he

°°
hexagonal axis point.-wise invariant.. As already derived in previous a publica-
tion [17], the elements Lx and Ly lead t.o the congruence 16" == (mod 1),
whereas the new symmetry element.s P and EI require 8tt == (mod 1). In-
deed, considering P as an example, one has:
' (1
{PI 0,0,0 }}'1= 2 1: 1 2 .
12+ a -32tt,-8+ a +1611,9+ a -23111-(a'a,tt)=}'1 (75)

As (32,16,24) = 8(4,2,3), the condition for tt becomes Sit == °

(mod 1).
Instead of t.he Euclidean Wyckoff posit.ion 26 for P6 3 111c (given is the mult.i-
plicity, t.he Wyckoff letter, the sit.e symmetry and t.he fract.ional coordinates):
121
2 b 3m
a' a' 2 + II,
(76)

we now have for t.he multimetrical space group "P63111CLr,:tLyLz,:tPEI" the

corresponding posit.ions - -
12n 121 n
2 bn
a'a'8 a'a'2+8
(77)
for any integer 11. (mod 8). For t.he close-packed ions X =
S one has n = 0,
whereas the tetrahedrally coordinated Y = Zn are at b3 wit.h 11. =
3. Accord-
ingly, the mult,imet.rical space group G of t.he idl'al Wurt.zit.e st.ructure (for t.he
chosen set. of metric tensors) is genf'rat.ed by:

G =

This is, again, not necessarily the full multi metrical space group of the struc-
ture, which is expected to he larger. It. is remarkable t.hat t.he ot.her tet.rahedml
52 A. Janner

positions (at 2b wit.h u = ~) and the oct.ahedral ones at 2a wit.h 0,0,:; and
0,0,4 + z for z = ~, as discussed in subsection 2.2, are also left. invariant by
the same group.

9. CONCLUDING REMARKS

Multimetrical crystallography has still a pronounced exploratory character. It

is, however, already so rich in structural relations that it appeared better to
limit here the considerations t.o the direct space and t.o simple examples of
hexagonal close-packed st.ructUl'es. Possible consequences of t.hese symmetries
for cryst.al diffraction and analogous derivation of mult.imet.rical symmetries
in the cubic close-packed case, even if shortly present.ed in Lp~ Houches, have
been, therefore, omit.ted.
A multimetrical space group contains the usual space group of t.he crystal as
subgroup, therefore, the consequences of mult.imet.rical symmetries can never
be less than t.hese following from t.he Euclidean symmetry only. One expects,
of course more, and not only at t.he pUl'e geomet.ricallevel present.ed here. The
matter is int.rinsically not. easy and delicat.e in t.he sensp that. one has t.o chose
an appropriate set. of metric tensors and, furthermore, one st.ill does not know
how important are t.he symmet.ry-breaking effect.s. Preliminary invest.igat.ions
seems t.o indicat.e that. mult.imet.rical symmetry can possible explain accident.al
degeneracies observed in band st.ructure calculat.ions for t.he elt'd,rons and in
phonon spect.ra [Ii]. It, is cert.ainly too early for claiming a relevance of mul-
timet.rical cryst.allography in crystal st.ruct.ure det.erminat.ion and in solid state
physics, but. it.s int.erest. is evident..

Acknowlcdgmcnts

Thanks are expressed t.o G.J. Heckman for valuahle comnwnts and crit.ical
remarks and t.o Fran<;oise Axel and Denis Gratias for t\wir kind invit,at,ion t.o
present. t.his t.opic at. t.he school.
 MULTIMETRICAL CRYSTALLOGRAPHY 53

References

[1] Janssen, T., Acta Cryst. A 42 (1986) 261-27l.

[2] Janner, A., Phys. Rev. Letters 67 (1991) 2159-2162.
[3] Janner, A., "Metrical aspects of quasicrystal embedding in superspace" , Ge-
ometry and Thermodynamics. Common Problems of Quasi-Crystals, Liq-
uid Crystals, and Incommensurate Systems. Proceedings of a NATO Ad-
vanced Research Workshop, Preveza September 4-8, 1989, J .-C. Toledano,
Ed. (Plenum Press, New York and London, 1990) pp. 49-65.
[4] Janner, A., Acta Cryst. A 47 (1991) 577-590.
[5] Janner, A., Acta Cryst. A 48 (1992) 884-90l.
[6] Janner, A., Acta Cryst. A (1995), To appear.
[7] Godreche, C., Luck, J.M., Janner, A. and Janssen, T., J. Phys. I France 3
(1993)1921-1939.
[8] Luck, J.M., Godreche, C., Janner, A. and Janssen, T., J. Phys. A: Math.
Gen. 26 (1993) 1951-1999.
[9] Zobetz, E., Acta Cryst. A 49 (1993) 667-676.
[10] Janner, A. and Janssen, T., Phys. Rev. B 15 (1977) 643-658.
[11] Janssen, T., Janner, A., Looijenga-Vos, A. and Wolff, P.M. de, "Incom-
mensurate and commensurate modulated structures", International Tables
for Crystallography. Vol. C, Mathematical, Physical and Chemical Tables,
A.J.C. Wilson, Ed. (Kluwer Acad. Publ., Dordrecht, 1992) pp.797-835.
[12] Wolff, P.M. de, Acta Cryst. A 30 (1974) 777-785.
[13] Wolff, P.M. de, Acta Cryst. A 33 (1977) 493-497.
[14] Janssen, T. and Janner, A., Adv. Phys. 36 (1987) 519-624.
[15] Janner, A., Phys. Rev. B 43(1991) 13206-13214.
[16] Janner, A., Phase Trans. 43 (1993) 35-47.
[17] J anner, A. and Nusimovici, M.A., "Can multimetrical symmetry help to ex-
plain accidental degeneracy in crystals ?" , Theories of Matter: A Festschrift
for Professor Joseph L. Birman, A. Solomon, M. Balkanski and H-R. Trebin,
eds. (World Scientific, Singapore, 1994) pp. 98-118.
[18] Wells, A.F., Structural Inorganic Chemistry (At the Clarendon Press, Ox-
ford, 1950).
[19] Buell, D.A., Binary Quadratic Forms (Springer, Berlin, 1989).
[20] Gauss, C.F., Disquisitiones Arithmeticae (Springer, New York, 1985).
[21] Minkowski, H., Geometrie der Zahlen (Teubner, Leipzig, 1910).
[22] Hancock, H., Development of the Minkowski Geometry of Numbers (Vol. 1
and 2) (Dover, New York, 1939), Reprinted edition 1964.
[23] Dickson, L. E., Introduction to the theory of numbers (Dover Publ., New
York, 1957).
[24] Hasse, H., Vorlesungen liber Zahlentheorie (Springer, Berlin, 1964).
 54 A. Janner

[25] Cohn, H., A second course in number theory (J. Wiley, New York, 1962).
[26] Mennicke, J., Proc. Roy. Soc. Edinburgh, Sect. A 41 (1968) 309-352.
[27] Fricke, R. and Klein, F., Vorlesungen tiber die Theorie der automorphen
Funktionen (Band 1 u. 2) (Teubner, Stuttgart, 1897), Reprinted edi·tion 1965.
[28] Klein, F. and Fricke, R., Vorlesungen iiber die Theorie der elliptischen Mod-
ulfunktionen (Band 1 u. 2) (Teubner, Stuttgart., 1890), Reprinted edition 1966.
[29] Magnus, W., Noneuclidean Tesselation and Their Groups (Academic Press,
New York, 1974).
[30] Ulrich, F. and Zachariasen, W., Z. Krist. 62(1925)260-273.
 COURSE 4

Non-commutative Models for Quasicrystals.

P. Kramer(i), J. Garcia-Escuderoe)

(i) Institut fiir Theoretische Physik der Universitiit Tii bingen ,

Auf der Morgenstelle 14, D 72076 Tiibingen, Germany
e) Departamento de Fisica, Universidad de Oviedo,
A vda. Calvo Sotelo, 33007 Oviedo, Spain

1. WHY NON-COMMUTATIVE MODELS FOR QUASICRYS-

TALS?

A powerful approach to quasicrystals arises from the use of n-dimensional

crystallography: Given a non-crystallographic point group H in R 3 , look for
the lattice r of minimal dimension n which has H as (part of) its pOint group.
The n-dimensional representation of H must be reducible and contain a 3D
representation. Split R n into this representation space R3 and its orthogonal
complement, construct discrete quasiperiodic patterns by projection from the
lattice to R3 and use the orthogonal complement for windows and for coding.
Periodicity and tranlational symmetry in R3 are absent by construction. This
scheme allows to treat diffraction properties by lifting the Fourier coefficients
to the reciprocal lattice rR.
Another aspect of n-dimensional crystallography is scaling. Scaling sym-
metry arises from in general Non-Euclidean point symmetries of r which are
compatible, in the simplest case commute, with the representation of H. This
property is necessary, but not sufficient to establish an inflation symmetry of
the discrete quasiperiodic patterns associated with the projection from a lattice.
Within the frame of a lattice embedding for a quasiperiodic structure, what
is required is a certain selection of objects which are to be projected into the
physical space and at the same time preserve certain point symmetries. In the
lattice embedding, this selection is described by appropriate windows in the
56 P. Kramer and J. Garcia-Escudero

orthogonal complement. A very different way of selecting objects is offered by

the combinatorics on words. It uses the algebra of the free monoid generated
by a finite alphabet. Substitutional sequences are generated by typical substi-
tution laws. Many fascinating aspects and mathematical properties of these
sequences are treated in other articles of this volume.
It is here that non-commutative crystallography appears as a possible frame
for quasicrystals. In non-commutative crystallography we drop the commuta-
tive aspect of the lattice. This is done by making a distinction between two
different paths connecting a fixed pair of an initial and a final lattice point.
An algebraic description for the path can be given in terms of the free group,
where a path is composed by concatenation from elementary directed paths
< Xl, ... Xn >, called the generators of the free group, and where inversion
of the direction determines the inverse element. The selection of a path can
now be achieved by constructing the orbit under an automorphism of the free
group.
The first question is if there is a natural correspondence of this non-commu-
tative scheme to the crystallographic scheme. This question will be addressed
by the use of two fundamental homomorphisms: the first one relates the free
group by abelianization to the commutative group (lattice translations), the
second one relates the group of automorphisms of the free group to the one
of the commutative group. Both these homomorphisms are compatible in a
natural fashion. There arises the second question if the path picture can be
endowed with point symmetry as was the case for the lattice. This question will
be analyzed by use of automorphisms of the free group. It will be shown that
point symmetry groups can operate on the paths without any reference to a
lattice. This will be demonstrated for dimension 2. Through the example of the
Fibonacci sequence, it will be seen that inflation symmetry in its simplest form
appears in the free group. We wish to emphasize that, by these homomorphism,
we obtain guidelines and control of non-commutative crystallography through
the well-known concepts and results of standard commutative crystallography.
What is required for this approach is a careful study of the group of auto-
morphisms of the free group, first analyzed by Nielsen in 1924. It turns out
that the case n = 2 shows a great simplicity, whereas a much more complex
structure appears already for n = 3.
There are many relations to the combinatorics on words mentioned above
and its monoid structure. The substitution matrix used in relation with free
monoids is again related to abelianization. We shall mention in some places
the relations between these fields.

2. FREE GROUPS AND THEIR AUTOMORPHISMS.

We start with an algebraic description of free groups. The symbol Fn stands

for the free group with n generators < Xl, X2, ••• , Xn >. The unit element of
Fn is the empty word, there is an inverse Xi ~ (Xi)-l for each generator. So
 NON-COMMUTATIVE MODELS 57

Fn has more algebraic structure than the monoid underlying the combinatorics
on words. The elements of Fn are words in the generators and in their inverses,
obtained by (associative) concatenation of the generators. From the inverse of
all generators one finds the inverse of any word, for example (xIx;lx3)-1 =
-1 -2
X3 X2Xl .
An automorphism is an invertible map Fn --+ Fn which respects the group
multiplication and the unit element. It is easy to see that an automorphism is
completely characterized by giving the images of the generators < Xl, ... ,Xn >.
For the multiplication of automorphisms we follow Nielsen and write for two
automorphisms <Pl, <P2
e: Xl X2 X3 Xn
<Pl : 01 (Xl .•. Xn) 02(Xl .. • X n ) 03(Xl ... Xn) On(Xl .. . X n )
<P2 : f3l (Xl ... Xn) f32(Xl ... Xn) f3a(Xl .. . Xn ) f3n(Xl ... Xn)
(<PltP2) : 01 (f3l ... f3n) 02(f3l ... f3n) 03(f3l ... f3n) On(f3l ... f3n)
(1)
In this and similar expressions we always write in two lines the preimages and
the images of the generators under an automorphism. The letter e denotes
the unit automorphism. Clearly all automorphisms of Fn form a group of
automorphisms, which Nielsen 1924 denoted as: (In := Aut(Fn ).
The group Aut(Fn) itself is an infinite group but is finitely generated. As an
illustration we give Nielsens set of generators < P, Q, (1, U > which act as
e: Xl X2 X3 Xn-l Xn
P: X2 Xl X3 Xn-l Xn
Q: X2 X3 X4 Xn Xl (2)
(1: (xd- l X2 X3 Xn-l Xn
U: XIX2 X2 X3 Xn-l Xn
Nielsen gave a large set of defining relations between these four generators.
New generators and relations can be given which allow for an algebraic and
geometric interpretation of the structure and subgroups of Aut(Fn).

3. NON-COMMUTATIVE CRYSTALLOGRAPHY.

Classical crystallography in n dimensions starts out from a translation group,

which geometrically forms a lattice, and algebraically is the commutative group
zn whose elements are vectors with n integer components, with addition as
group composition. Since we want to remove the commutative property, it is
better to replace addition by multiplication. If < bl , ... ,bn > denotes a formal
lattice basis, we describe a translation (ml, ... ,mn ) by the symbol

(3)

All factors in this product commute. The automorphisms of zn are given by

n x n matrices with integer entries and determinant ±1, they form the group
58 P. Kramer and J. Garcia-Escudero

GI(n, Z) = Aut(zn). The group GI(n, Z) itself is an infinite group. Point

and scaling symmetry in a lattice arise from specific subgroups of GI(n, Z).
Usually they are associated with a specific Euclidean or Non-Euclidean metric.
For example it is well-known that finite point groups conserve a Euclidean
metric, and that the passage from their form in GI(n, Z) to their form in the
orthogonal group O(n, R) determines a lattice basis.
A homomorphism between two groups is a map compatible with the group
multiplication: images of products are the products of the images. We shall
now describe the relation of the non-commutative scheme to the commutative
one of crystallography by homomorphisms between the groups. The relation of
the free group Fn to the commutative group zn arises through a first homomor-
phisms ham1 called abelianization: Replace Xi by bi , disregard the order of the
bi, and determine from any word wE Fn an element of zn by power counting.
For example we have ham1 : xix2xl1x;-2 -t b1b;-1. This abelianization on the
level of Fn produces a second, compatible homomorphism ham2 between the
groups of automorphisms: Power counting for the n images of the generators
Xi under a fixed automorphism ¢ E Aut(Fn) yields the n x n entries of a matrix
from GI(n, Z). It follows from the invertibility of the automorphism that the
matrix must belong to this group. By compatibility we mean the following:
Consider a word w E Fn and a fixed automorphism ¢ E Aut(Fn). Then for
any pair (w, ¢) we have

(4)

In the monoid structure used for the analysis of sequences, the matrix resulting
from abelianization is called the substitution matrix. This matrix will in general
not belong to GI(n, Z), it may not be invertible.
The two homomorphisms allow one to speak of non-commutative crystal-
lography: For a given lattice vector, its preimages in Fn under ham1 are all
connected paths between the initial and final lattice point. For a given element
of GI(n, Z) which describes a point or scaling symmetry, we can now examine
its preimages in Aut(Fn) under ham2. For a finite crystallographic point group
the question arises if there are preimages in Aut(Fn) which form a finite, per-
haps isomorphic group. To examine these questions we start from the lowest
dimension 2.

4. STRUCTURE AND GEOMETRY OF THE GROUP Aut(F2 ).

The group F2 with two generators < Xl, X2 > will be used to illustrate the non-
commutative scheme and the relations to 2D commutative crystallography. For
Aut(F2 ) we introduce three new generators < C2, C3, 0'1 > which, together with
their images in GI(2, Z) under ham2 given in the right-hand column, are defined
 NON-COMMUTATIVE MODELS 59

by
Aut(F2) Ol(2, Z)

[ ~]
1
e: Xl X2
0

C2 : X1 X2 x 2-1 [ 1
0 -~ ] (5)

C3 : X2-1 Xl
-1
[ 0
-1 -~ ]
[ ~]
-1
a1 : Xl-1 X2
0
We define the following relations for these generators:

Q1, Q2, Q3 (C2) 2 = (C3) 2 = (ad 2 = e, (6)

Q4 (C2C3)3 = e,
Q5 (c3ad2 = e,
Q6 a1c2(a1c3)2c2al = c2(al c3)2 c2 .
and claim that Aut(F2) can be presented by generators and relations as

(7)

From relations Ql, Q2, Q3 it follows that the new generators of F2 are all invo-
lutions, i.e. their squares are the unit automorphism. With these new genera-
tors and relations it is possible to detect finite subgroups of the infinite group
Aut(F2). Two finite crystallographic Coxeter groups appear:

< C2, C3; (C2)2 = (C3)2 = (C2C3)3 = e >, (8)

< C3,al; (C3)2 = (a1)2 = (c3ad4 = e >
Their relations are part of the 6 relations given for Aut(F2 ). So it can be
seen that finite point symmetry can exist beyond commutative crystallography.
Next we consider the typical Fibonacci scaling automorphism which is of the
form
(9)

The Fibonacci automorphism is the product </> = C3al C2. By iteration it gener-
ates the Fibonacci words in F2 . The Fibonacci words form an orbit under the
powers of this special automorphism.
All automorphisms admit a graphical and geometric representation in terms
of affine reflections in the plane R2. This affine approach uses only the dual
space and hence is metric-free. Let r be a fixed reflection vector and CPr a linear
60 P. Kramer and J. Garcia-Escudero

Fig. 1. - The path representation of the generators < Xl, X2 > of F2 as vectors, the
triangle T, and its images under the affine reflections which represent the generators
< C2, C3, 0"1 > of Aut(F2}. Broken lines mark reflection lines.

form on R2 restricted by q)r(r) = -2. Then an affine reflection is defined by

the linear involutive map

(r, q)r) : x -+ x' = x + q)r(x)r. (1O)

All points on the line q)r{Y) = 0 are fixed. In a basis formed from r and from
one vector along this fixed line, x and x' have opposite signs of their component
along r. Represent the elements Xl, X2 as two vectors in the plane, with the
initial point of Xl at the end point of X2' Represent the inverse by an inversion
of the direction. Close the triangle T formed by the initial and end points.
We shall give now a linear representation of Aut(F2 ) in terms of affine re-
flections. This linear representation is fixed once we have given the images of
the generators. Consider first three affine reflections whose fixed line passes
through the midpoint of an edge and the opposite vertex of the triangle, and
whose reflection vector coincides with the edge.
These affine reflections interchange pairs of vertices, and map the triangle
into itself. Two of them represent the generators C2, C3. For the generator
0"1, the fixed line contains the edge formed by the vector X2, the reflection
vector is Xl, and the affine reflection produces a new triangle with one new
vertex. It is clear that the generator 0"1, as well as its conjugates under the
other generators, allows one to propagate the triangle over the plane by affine
reflections. Consider now the possible new vertex positions of all images of
the triangle obtained by this propagation. For these we claim: The possible
 NON-COMMUTATIVE MODELS 61

Fig. 2. - Images of the triangle T under some automorphism of Aut(F2). Two typical
fixpoints of order q = 3,4 are marked. The point marked 6 has order 6 in Gl(2, Z)
but infinite order in Aut(H).

vertices for all these propagated triangles are on the lattice spanned by the
vectors Xl, X2. In Fig.1 we represent the generators of Aut( F2). The fixed
reflection lines are indicated by broken lines.
The linear representation gives a graphical illustration of the algebraic form
for the automorphisms. Its linear form just described on a lattice actually
implies already a commutative scheme, and so all conclusions drawn from the
graphs must be controlled by algebraic multiplications of the automorphisms.
In deriving the graphs from the algebraic expressions, we allow for shifts of each
vector along its direction and use these shifts to find the correct concatenation
for the images of Xl, X2.
We return to the subgroups of finite order and to their graphical represen-
tation. Clearly the first three affine reflections described above generate the
Coxeter group A2 , with a fixpoint on the intersection of the .three invariant
lines. The element of order four belongs to the finite Coxeter group B 2 , it
results by propagating the triangle around a single fixed vertex. So one sees
that, in a scheme using affine reflections, different point symmetries like 3-fold
and 4-fold rotations can coexist.
In Fig.2 we indicate some fixpoints of order 3 and 4 of this type. These
different point symmetries are strictly affine and metric-free. If we wish to
give them orthogonal form corresponding to Euclidean metric, we must make
62 P. Kramer and J. Garcia-Escudero

Fig. 3. - The six class representatives for elements of Aut(H) of finite order q =
2,3,4 are shown in two rows by giving the orbits of the triangle T under powers of
each element.

a choice of the original triangle: a regular triangle for the group A 2 , a triangle
with two orthogonal edges of equal length for B 2 • It is not possible to introduce
a scalar product <, > such that all linear forms cf>r for the generators of Aut(F2 )
can be rewritten as Weyl reflections of the form

(11)

Now we determine the conjugation classes and representatives for all ele-
ments of finite order in Aut(F2 ). The possible orders are q = 2,3,4. Their
 NON-COMMUTATIVE MODELS 63

I
I
I
I

Fig. 4. - The first two images <jJT, <jJ2T under the Fibonacci automorphism show that
the triangle T is stretched and propagated in the direction of an eigenvector.

representatives are given algebraically as

Aut(F2 ) GI(2, Z)

[ ~]
1
e: Xl X2
0

q = 2: C3 : X 2-1 -1
Xl [ 0
-1 -~ ]
[ ~]
-1 -1
0"1 : Xl X2
0

V2 : Xl
-1 -1
X2
[ -1
0 -~ ]
(12)

[ ~]
-1 -1
W2: Xl XlX2 X l
2

q = 3: C2 C3 :
-1 -1
X 2 Xl Xl
[ -1
1 -~ ]
[ ~]
0
q = 4: 0"1 C3 : X2
-1
Xl
-1

In Fig.3 we give these representatives generated by affine reflections. There is

64 P. Kramer and J. Garcia-Escudero

no element of order 6 in Aut(F2 }, the preimage of the standard 6-fold rotation

element in Gl(2, Z} is of infinite order. The graph for the obvious candidate
for order 6 is shown in Fig.2. To compute it algebraically we introduce a
short-hand notation for conjugation in a group by

(13)

If one introduces a simple automorphism p E Aut(F2 } which under hom2 pro-

duces an element of order 6, one obtains

Aut(F2 } Gl(2, Z}

[ ~]
1
e: Xl X2
0

[ ~]
-1 1 (14)
p: X2 X 1 Xl
-1

[ ~]
p6 :
1
(xd K (X2}K
0

where K = (x2xd-1 (X1X2) is a commutator of group elements. The image of p

in Gl(2, Z) is of order 6. In contrast, the power 6 of this automorphism differs
from the identity, and in fact the automorphism is of infinite order.
The graphical representation of the Fibonacci automorphism yields a propa-
gation of the triangle whose first steps are shown in Fig.4. If we would put the
triangle into the square lattice, which is often used with the Fibonacci word,
we would see that the triangle is stretched and asymptotically propagated into
the direction of one of the eigenvectors of the corresponding matrix in Gl (2, Z).
We turn to some infinite subgroups of Aut(F2 }. Consider the subgroup X3 <
Aut(F2 } generated by

Each of these group elements is an involution. We claim that this subgroup is

normal in Aut(F2 }. For the proof we compute the conjugates of these group
elements with the generators of Aut(F2 } and obtain from the relations Q1 ... Q6

8i : 81 82 83
(8i}C 2 : 81 83 82
(16)
(8i}C3 : 83 82 81
(8i}O"' : 82 8 1 8 2 82 83

The relation Q6 is crucial for these results. In Aut(F2 ) it is possible to con-

struct the homomorphism ham2 : Aut(F2 } --+ Gl(2, Z} in full detail. For this
construction we require its kernel ker(ham2}, that is, the invariant subgroup
of Aut(F2} mapped into the identity of Gl(2, Z} under ham2. In the case
 NON-COMMUTATIVE MODELS 65

Fig. 5. - The generators < it, t2 > of the normal subgroup Inn(F2) <l Aut(F2)
translate the triangle T along the directions of the vectors Xl, X2 by twice the length
of these vectors.

n = 2, this invariant subgroup coincides with the group of inner automor-

phisms Inn(F2)' Since F2 is free, its inner automorphisms are generated by
conjugations with the generators < Xl, X2 > of F2. It is possible to construct
explicitly in Aut(F2) two elements

tl = (C3C2CTl C3CTd 2, (17)

t2 (CTI C3CTl C3C2)2 ,

which from the maps

e: Xl X2 (18)
-1
tl : Xl Xl X2 Xl

t2 : X2-1 XlX2 X2
are seen to generate the inner automorphisms. It follows that these two auto-
morphism generate a normal subgroup of Aut(F2) which is isomorphic to F2.
The generators tl, t2 can be constructed from the involutions 81,82,83 as

(19)

and so we have a normal subgroup chain Aut(F2) I> X3 I> F2 •

In terms of the graphs for the triangle, the three involutions 81,82,83 reflect
the triangle in its three vertices, and the automorphisms tl, t2 simply translate
the triangle along the direction of Xl, X2 respectively by twice the length of
these vectors, as is shown in Fig.5. These two translations commute only in
the graphical representation which corresponds to abelianization, they do not
commute algebraically.
66 P. Kramer and J. Garcia-Escudero

5. FREE GROUPS AND AUTOMORPHISMS FOR n > 2.

The analysis of Fn , Aut(Fn} for n > 2 shows both similarities and a richer
structure. The finite Coxeter group Bn E Aut(Fn} is implicit in the Nielsen
analysis. The finite Coxeter group An E Aut(Fn } is given in [4]. So there
are finite groups in Aut(Fn } which provide schemes for non-commutative point
symmetry. It is found that the affine reflections have a natural generalization
to dimension n. An affine reflection is determined by a vector r and by a linear
form CPr which corresponds to a hyperplane not containing r. The triangle
is replaced by a simplex with n + 1 vertices, which are permuted under the
group An '" Sn+1. These subgroups of Aut(Fn } appear within an infinite
subgroup which maintains many features found in Aut(F2 }. This subgroup
includes all permutations of the vertices of the simplex. In addition it contains
affine generalizations of 0'1 whose reflection hyperplane contains a face of the
simplex, and whose reflection vector corresponds to an edge of the simplex not
on this face. It is clear then that 0'1 and its conjugates propagate the simplex,
with all vertices on a lattice. The two finite Coxeter groups An, Bn belong to
the infinite subgroup. They allow to implement finite point group symmetry
in the non-commutative scheme. The infinite subgroup has a normal subgroup
generated by the reflections of the simplex in its n+ 1 vertices. Products of pairs
of these reflections translate the simplex along its edges by twice the length of
the edges, where the n basic translations generate a subgroup isomorphic to
F 3 . This infinite subgroup generalizes many properties of Aut(F2 }.
Beginning with n = 3, there appear new elements in Aut(Fn) outside this
infinite subgroup. There role is under study first of all for Aut(F3 }. There
are 12 new relations due to Nielsen which must be rewritten and interpreted
in the new generators. The kernel of hom2 now is bigger than Inn(F3}' New
subgroups appear as for example the braid group B 3 .
In the non-commutative analysis applied to scaling and inflation with n > 2,
there are up to now only tentative results. The non-periodic triangle and
Penrose patterns can be generated in a modified form by automorphisms of
free groups, see [5, 6]. The modification implies first of all that elements of the
free group be interpreted as triangles rather than as paths. Secondly it appears
that rules and brackets of a formal grammar are needed in order to describe
the production of the 2D quasiperiodic patterns. This topic is treated in the
following two sections.

6. NON-COMMUTATIVE MODELS AND SYMMETRIES FOR

2D QUASIPERIODIC PATTERNS.

In [5, 6] we have extended to 2D the very well known interpretation of the

Fibonacci chain as an orbit of an automorphism of a free group. We now
describe the main features of this approach.
 NON-COMMUTATIVE MODELS 67

Fig. 6. - The tiles aI, acute triangle, and b1 , obtuse triangle.

In order to do this we implement an automorphism with a parenthesis struc-

ture which can be introduced by means of a formal grammar. A formal grammar
is a quadruple
G =< ~T,~N,S,P > (20)
where ~T is the alphabet of terminal symbols and ~N the alphabet of nonter-
minal symbols, S E ~N is the start symbol or sentence symbol and P is a finite
set of production or substitution rules. The specific grammars we will use are
context free grammars [8].
The free group we use is F40 =< a" b" a~, b~ > were i E ZIO. The correspond-
ing automorphism group is denoted by <1>40. The elements ai, a~ on one hand,
and bi , b~ on the other hand represent tiles with the same shape but they are
distinguished by a colour: white for ai, bi , black for a~ and b~. The tile at is an
acute isosceles triangle with (length of side)/(length of base)=T= (1 + J5) /2,
and bi represents an obtuse triangle with (length of side)/(length of base)=l/T.
The sides of bi have the same length as the basis of ai.
In Fig.6 we have drawn the tiles al and b1 . The tile at(bi ) is obtained from
al (bd by a rotation of angle 21T(i - 1)/10 through the vertex indicated.
We now describe the generation of the Triangle and Penrose patterns, de-
noted for short by (TP), (PP), compare [3]. Let wTP,w PP E <1>40 be defined
in the following fashion:
wTP(ai) = a~_3bi+5ai' wTP(bi ) = a~+3bi+l
wTP(a~) = at_3b~+3a~+4 ' wTP(bi) = ai-3b~+3
wPP(ai) wPP(bi ) = (21)
= ai+2a~+l bi - 1 , a~+l bi - 1
wPP(ai) = a~_2ai-l b~+5' wPP(bi) = ai-5b~+l'

The first application of these automorphisms on a given generator is converted

into a gluing of triangles by the following prescription: if two letters follow
one another, the corresponding oriented triangles are glued face to face. This
prescription is unique. In the second application of the automorphism we use
this prescription for the new triangles, which appear now scaled by a factor
T, and glue them face to face, disregarding their internal composition into the
initial tiles. In the words of F 40 produced in this second step, not all consecutive
letters correspond to glued tiles. To characterize the geometric hierarchy of this
gluing process in the word structure, we use a formal grammar. We take the
alphabets ~T =< ai, bi , a~, b~,), (> and ~N =< Ai, B i , A~, B~ > with i E ZIO.
68 P. Kramer and J. Garcia-Escudero

Fig. 7. - Production of the Triangle pattern by an automorphism.

Fig. 8. - Production of the Penrose pattern by an automorphism.

Every element belonging to L,N can be used as start symbol S. We take S = AI '
The set of production rules for the Triangle pattern are

r
I--t (A~_3Bi+5Ai)
pTP _ B: I--t B
(Ai+3 i+l)
II b,
aj }
(22)
- A~ I--t (Ai-3B~+3A~+4) I a~
B'
• I--t (Ai-3B~+3) I b~
and for the Penrose pattern

r
I--t (Ai+2A:+l Bi-d
ppp _ Bi I--t (A~+l Bi-d
II bi
aj }
(23)
- A~j
I--t (A~_2Ai-IB~+5) I a;
B'
• I--t (Ai-5B~+l) I b:
The vertical bars separate two alternative possible images under the map given
by the arrows. Consider the following derivation for a part of the Triangle
pattern:

Al I--t (A~B6AI) I---t (24)

((A5B~ A~)(A~B7 )(A~B6AI)) I---t ((a5b~ a~)(a~b7 )(a~b6ad)

The string ((a5b~ a~)(a~b7 )(a~b6ad) represents a part of the Triangle pattern:
the acute triangle a5b~ a~, the obtuse triangle a~b7 and the acute triangle a~b6al
can be glued face to face in a unique way.
In Figs. 7 and 8 we show the patterns obtained in the first three generations
for the Triangle and Penrose patterns respectively. Observe that the pairing of
 NON-COMMUTATIVE MODELS 69

two obtuse triangles already appears in the second generation of the Penrose
pattern while it is forbidden in the Triangle pattern.
It is possible to combine the generating automorphisms with the generators

m
of a group of reflections containing fivefold symmetry. Consider the Coxeter
group H =< R 1, R 2\(R 1R 2)5 = R~ = = e >. We define the following action
of H on F40 :

R1(ai) a~_i; R1(bi) = b;_i; Rl(a~) = a9-i;Rt{b~) = b3 -, (25)

R2(ai) = a;_i; R 2(bi) = b;_i; R2(aD = a7-i; R2(b~) = b1- i

Then if we use the multiplication law given in [17], it can be proved by induction
on n that for k=1,2:

(R 1R2)2n(wTP )nRk = Rk(WTP)n (26)

Rk(WPP)n = (wPP)n Rk

for the Triangle and Penrose pattern respectively.

7. AUTOMATA FOR THE TRIANGLE AND PENROSE PAT-

TERNS.

There exists a class of automata which are specific for the context free lan-
guages: the push-down automata [7]. The generating automaton contains an
input tape, a memory tape and a computing unit with two operating heads:
one to write on the input tape and the other to read, erase and write on the
memory tape. The symbols written on the input tape belong to the vocabulary
~T, those written on the memory tape belong to ET U EN. In addition there
is one special symbol a written on the memory tape. The computing unit can
be in one of the following internal states:

(27)

A physical situation of the automaton is a triple{x, [y], z} where x E ET, [y] is

an internal state and z is a or a string of symbols belonging to ~T U ~N' A
situation of the automaton is either the triple of a physical situation {x, [y] ,z}
or a triple obtained by replacing the x or z (or both) by the neutral word e. If
the automaton is in the physical situation {x, [y] , z} then it is at the same time
in the four situations:{x, [y] ,z}, {e, [y] ,z},{x, [y] ,e} and {e, [y] ,e}.
An instruction is a quintuple of the form:{[ul,v} ~ {a, [bl,c}, where
{a,[bl,c} is a situation ,[u] an internal state and v E ~T U ~N U {a}. If
the automaton is in the internal state [u] and at the same time it is scanning
the symbol v written on the memory tape then it can choose the instruction
{[ul, v} ~ {a, [bl, c}.In that case
a) Internal state: the computing unit switches to the state [b]
70 P. Kramer and J. Garcia-Escudero

b )Input tape: According as a E ET or a == e then the corresponding head

writes the symbol on the left of the last written symbol on the input tape or
else doesn't move the tape and writes nothing.
c) Memory tape: the head scanning the memory tape acts upon it in the
following way:
• if c = a then it moves the tape so that the head is scanning the symbol
to the left of v and it erases v ;
• if c = e then it dosn' t move the tape and writes nothing;
• if c is a string of symbols then it writes the string to the right of the last
symbol and moves the tape so that the head is scanning the right most
symbol written.
At the end the automaton will be in a state that we label So with the symbol
a written on the extreme left of the memory tape.
The automaton generating the Triangle pattern has the following set of in-
structions:
1 : {[O], a} f-----t ie, [1], Ad;
2i: {[I] ,AI} f-----t ie, [Ai], a}
3. : {[I], Bd f-----t ie, [Bi], a};
4,: {[I] ,AD f-----t {e,[A~],a}
5i : {[I] , B:} f-----t {e, [B:], a} ;
6i : {[AI], e} f-----t {e, [1] , (A~_3B'+5Ai)}
7i : {[AI], e} f-----t {e,[l],ad;
8j:{[Bi],e} f-----t {e, [1], (Ai+3BHd}
9. : {[Bi], e} f-----t ie, [1], bi};
lOi : {[A~], e} f-----t {e, [1] , (Ai- 3B:+ 3Ai+4)} (28)
11i : {[Ai], e} f-----t {e,[I],aD;
12i : {[Bn ,e} f-----t {e, [1] , (Ai-3Bi+3)}
13i : {[Bi], e} f-----t ie, [I],bD;
14i : {[I] , ad f-----t {ai, [1], a}
15i : {[I] , bi } f-----t {bi,[I],a};
16i : {[I] , aD f-----t {ai, [1], a}
I7i : {[I] ,bD f-----t {bi,[l],a};
18: ([I], (} f-----t {(,[l],a}
19: ([I],)} f-----t {), [1], a} ;
20: {[I] ,a} f-----t {[So] ,a}
Now we want to show how the automaton starts generating the strings. For
instance we suppose the string to be generated is ((a5b~a~)(a~lJ.r)(a~b6ad). If
the first internal states of the automaton are
[0], [1], [Ad, [1], [1], [AI], [1], [1], [Ad, [1], [1] (29)
and the rules that the automaton chooses are
(30)
 NON-COMMUTATIVE MODELS 71

then the strings written on the memory tape are

a ,aA1, a, a(A~B6Ad,a(A~B6Al' a(A~B6' a(A~B6(A~B6Al)' (31)
a (A~B6 (A~B6Al' a(A~B6 (A~B6, a(A~B6(A~B6al' a(A~B6(A~B6
and the string written on the input tape is al)).
In a similar fashion we can get the rules for the automaton generating the
Penrose pattern:
1 : {[O], a} t-----+ {e,[I],Ad;
2, : {[I] ,At} t-----+ {e,[AI],a}
3i: {[I], B t } t-----+ {e, [Bi],a};
4.: {[I] ,AD t-----+ {e,[A~],a}
5, : {[I], B:} t-----+ {e, [B:],a};
6i : {[At] ,e} t-----+ {e, [1] , (Ai+2A~+1 B t- 1)}
7i : {[Ai], e} t-----+ {e, [1], ad;
8t : {[B t ], e} t-----+ {e, [1], (A~+l Bi-d}
9i : {[Bi], e} t-----+ {e, [1], bd;
10 1 : {[A~], e} t-----+ {e, [1], (A~_2Ai-lB:+5)}
(32)
l1t : {[A~], e} t-----+ {e,[I],a~};
12i : {[B~], e} t-----+ {e, [1] , (Ai-5B~+l)}
13i : {[B~], e} t-----+ {e, [1], bD;
14i : {[I], ad t-----+ {adI],a}
15 t : {[I],b i } t-----+ {bdI],a} ;
16i : {[I], aD t-----+ {a~,[I],a}
17i : {[I],b~} t-----+ {b~, [1], a};
18: ([I], (} t-----+ {(,[I],a}
19: ([I],)} t-----+ {), [1], a} ;
20: {[I] ,a} t-----+ {[So] ,a}
Not all the words belonging to the language generated by the grammars defined
in the previous section represent pieces of the patterns, only those appear with
the same derivation length for the elements of EnET. It is possible to introduce
a formal grammar in order to avoid this problem [6].
A OL-grammar [8] is a triple G =< E, P, Q > where E is a finite alphabet,Q
is the start string whose elements belong to E and P is a set of productions
or substitution rules of the form a t-----+ /3 where a E E and /3 is a string with
elements in E. We take the alphabet E =< ai, bi , a~, b~, ), (> with i E ZIO.
Every element belonging to E can be used as the start symbol Q. The set of
production rules for the Triangle pattern are
ai t-----+ (a~_3bi+5ai)
bi t-----+ (a~+3 bi+l )
a~ t-----+ (ai-3b~+3a~+4)
p TP = b~ t-----+ (ai-3b~+3)
(33)
(t-----+ (
) t-----+)
72 P. Kramer and J. Garcia-Escudero

and for the Penrose pattern

aj f----t (aj+2a~+1 bj-d

bj f----t (a~+ 1 b, - d
ppp = a~ f----t (a~_2ai-l b:+5)
(34)
b: f----t (ai-5b:+l)
(f----t (
) f----t)

All the words belonging to the language generated by these grammars represent
parts of the patterns.

8. SURVEY OF OTHER RESULTS.

We introduced and illustrated the concepts and basic results of non-commu-

tative crystallography mainly for the dimension n = 2. For the 2D patterns
we found it necessary to introduce the new terms of a parenthesis structure
and of formal grammars. In this section we describe background and refer-
ence material and indicate various lines of application and generalizations. For
the general concepts of n-dimensional crystallography we refer to [1, 18]. For
Coxeter groups see [9]. Crystallographic symmetries and scaling for quasicrys-
tals are treated in [10]. An example of the lattice embedding for quasiperiodic
patterns with lO-fold symmetry is given in [3].
For definitions, results and many examples in the combinatorics on words,
based on the free monoid structure, we refer to [16). The free group and its
group of automorphism, the Nielsen generators and relations, and the relation
to Gl(n,Z) are described in [17]. The new generators for Aut(F2 ) and the
affine reflections are introduced in [14]. Note that some changes were intro-
duced in the present version. New aspects of the automorphisms arise if one
turns from the geometry to the representation of observables. For ID general-
ized Fibonacci sequences one can represent the transfer matrices for an electron,
propagating through intervals on the line which represent the sequence, by two
elements from SU(I, 1), compare [2,12)' and analyze states, backscattering and
transmission problems. Given a homomorphism from F2 to SU(I, 1), automor-
phisms of F2 lead to mappings between pairs of group elements in SU(I,I)
which are called induced automorphisms. This scheme admits a geometric and
graphical representation of Aut(F2 ) in terms of propagating triangles, which
has been called Fricke-Klein geometry. It places the triangles on hyperboloids
but locally preserves the representation by reflections developed in the previous
sections, and also preserves the non-commutative structure, see [13, 14]. If the
energy of the electron is varied continuously, on can reach points where the
transfer matrices commute. These points are significant for the transmission
properties, as is shown in [2]. A crucial role in this analysis is played by the
commutator of the two basic transfer matrices, which becomes the unit element
in the commutative case. The traces and trace maps, treated in other parts
 NON-COMMUTATIVE MODELS 73

of this volume, have a simple interpretation in terms of this geometric scheme:

they are trigonometric functions of the arc length in hyperbolic triangles. A
similar analysis for spin systems involves the group SU(2) and leads to spher-
ical triangles [12). In this case the periodic orbits on all finite subgroups of
SU(2) under the induced Fibonacci automorphism are given in [15).

References

[1) Brown H., BUlow R., Neubiiser J., Wondratschek H., Zassenhaus H.,
Crystallographic Groups of Four-Dimensional Space,
(Wiley, New York, 1978)
[2) Baake M., Joseph D. and Kramer P., Phys. Lett. A168 (1992) 199-208
[3) Baake M., Kramer P., Schlott mann M. and Zeidler D.,
Int. J. Mod. Phys. B4 (1990) 2217-68
[4) Garcia-Escudero J. and Kramer P., Anales de Fisica, Monografias 1, voU
(1993) 339-42, Madrid
[5) Garcia-Escudero J. and Kramer P., J. Phys. A26 (1993) L1029-35
[6) Garcia-Escudero J. and Kramer P., Proc. Int. Wigner Symposium (1993),
Oxford
[7) Gross M. and Lentin A., Introduction to Formal Grammars, (Springer,
Berlin, 1970)
[8) Hopcroft J. E. and Ullmann J. D., Einfiihrung in die Automatentheorie,
formale Sprachen llnd Komplexitatstheorie,
(Addison-Wesley, Bonn, 1988)
[9) Humphreys J. E. , Reflection Groups and Coxeter Groups,
(Cambridge University Press, Cambridge, 1990)
[10) Janner A., Acta Cryst. A47 (1991) 577-590
[11) Kramer P., Anales de Fisica, Monografias 1, vo1.2 (1993) 370-3, Madrid
[12] Kramer P., J. Phys. A26 (1993) 213-228
[13) Kramer P., J. Phys. Lett. A26 (1993) L245-L250
[14) Kramer P., J. Phys. A27 (1994) 2011-22
[15) Kramer P. and Wagner H. (1994), preprint Tiibingen
[16] Lothaire M., Combinatorics on Words,
(Addison-Wesley, Reading, 1983)
[17] Magnus W., Karras A. and Solitar D., Combinatorial Group Theory,
(Dover, New York, 1976)
[18] Schwarzenberger R. L. E., N-Dimensional Crystallography,
(Pitman, San Francisco, 1980)
 COURSE 5

From Quasiperiodic to More Complex Systems

T. Janssen

Institute for Theoretical Physics,

University of Nijmegen and University of Utrecht
6525 ED Nijmegen, The Netherlands.

1. STRUCTURES

1.1. Introduction

For a long time solid state systems were considered as either crystalline, i.e.
lattice periodic and with space group symmetry, or amorphous. In the former
case the diffraction peaks are infinitely sharp for a perfect infinite crystal, in
the latter there are no sharp peaks. The presence of some disorder does not
eliminate sharp Bragg peaks as long as long-range order is preserved. Moreover,
the sharp Bragg peaks lie on a lattice, the reciprocal lattice, such that each peak
can be labelled with three integer coefficients. It was also known that sometimes
additional peaks appear, satellites. For example, a crystalline material with
76 T. Janssen

magnetic moments may show an ordering of the moments that can be described
as a helix. If the period of this helix is not a period of the underlying lattice, the
total system consisting of crystal and magnetic moments does not have lattice
periodicity. This was not a problem as long as one did not want to describe
the symmetries of crystal and magnetic system simultaneously. Moreover, for
a long period it was not clear whether there was a common (super)lattice for
crystal and magnetic moments, or not.

At the end of the sixties satellites were observed also in other systems,
such as thiourea[27] and NaN0 2 [22]. These satellites pointed to an additional
modulation of the structure with an own period. It was de Wolff who pointed
out in 1972 that such a modulated crystal with no strict lattice periodicity can
be described in usual crystallographic terms if tine uses not a three-dimensional,
but a four-dimensional space group. He showed this on the example of the
modulated -y-phase of dehydrated Na 2 C03 [14, 1, 79].

Higher-dimensional crystallographic groups have been studied since the

end of the 19th century. Bieberbach[ll] studied the structure of space groups
in arbitrary dimension. Ascher and Janner[2, 3] studied the algebraic struc-
ture of such groups. Four-dimensional crystallographic transformations were
presented by Goursat[30], Heesch[35] and Hermann[38], general algorithms to
derive space groups in arbitrary dimension if the point groups are known were
given by Zassenhaus[86], by Brown[15] and by Fast and Janssen[24]. A list
of all four-dimensional space groups that are simultaneously subgroups of the
Poincare group (the inhomogeneous group of relativistic dynamics) was derived
by Fast and Janssen. Finally a full list of all four-dimensional space groups
was published by Brown, Biilow, Neubiiser, Wondratschek and Zassenhaus[16].
Crystallographic point groups in more than four dimensions were derived by
Plesken[64]. It turned out that the 4-dimensional groups needed by de Wolff
for the description of incommensurate soda were exactly the groups studied by
Janner, Janssen and Fast. The lists could then be applied to incomensurate
structures.

The idea of using higher-dimensional space groups for the description of

systems without lattice periodicity is simple. For an ordinary crystal all diffrac-
tion peaks can be labelled with three indices. This means that the position of
the diffraction peaks can be written as
k = hii* + kb* + lC", (1.1)
for integers h, k, l. In the -y-phase of N a2 C0 3 there are satellites and a general
diffraction vector is
k = hii* + kb* + lC" + m(o:ii* +-yC"). (1.2)
The peaks at positions with m = 0 are called main reflections and the oth-
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 77

ers with m i- 0 are called satellites. The coefficients 0: and "I are irrational
numbers. Therefore, there is not a three-dimensional basis for which ii*, b*, i!'
and if = o:ii* + "Ii!' have simultaneously integer coordinates. This implies that
the system does not have lattice periodicity, because there is no 3-dimensional
basis such that each basis vector has a scalar product equal to zero modulo 1
with the 4 basis vectors in reciprocal space. However, one may consider the
Bragg peaks as the projection on three-dimensional space of Bragg peaks be-
longing to a four-dimensional reciprocal lattice. Thus k is the projection of
a four-dimensional reciprocal lattice vector ks which has integer coordinates
e,
h, k, m with respect to a basis in 4 dimensions. To these four-dimensional
Bragg peaks corresponds a four-dimensional lattice periodic structure. It will
be shown later on that the real physical structure is just the intersection of this
four-dimensional periodic structure with a three-dimensional hyperplane that
can be identified with the physical space.

The case of "I - N a2C03 is an example of a structure with sharp Bragg

peaks that can be labelled with a finite number of integer coefficients, i.e.
n
k = L h.ii;, (h. E Z). (1.3)
.=1

For "I - N a2C03 n = 4. If n = 3 and the 3 basis vectors iii are not in one
plane the structure is lattice periodic. A basis for its lattice is given by ii. such
that
(1.4)
In general, a structure with Bragg peaks satisfying (1.3) is called quasiperiodic.
This term stems from the theory of dynamical systems. A system is quasiperi-
odic in time if there are two or more mutually incommensurate frequencies. In
modulated crystals there are no incommensurate frequencies but incommensu-
rate basis vectors in reciprocal space. The number of rationally independent
basis vectors is called the rank of the system. The set of all vectors (1.3), that
mathematically has the structure of a Z-module, is called the Fourier module.
If one considers a reciprocal lattice in n dimensions with basis vectors that
project on a three-dimensional subspace, the physical space, as the basis vec-
tors of the Fourier module, there is a one to one correspondence between the
vectors of the Fourier module and those of the n-dimensional reciprocal lattice.

The idea that a quasiperiodic system of rank four corresponds to a lattice

periodic structure in 4 dimensions, and, more generally, that a quasiperiodic
system of rank n can be embedded as n-dimensional structure with space group
symmetry has been used extensively in structure determination of quasiperiodic
systems. In the early seventies these were mainly incommensurate modulated
strucures. These can equally well be described by means of irreducible rep-
resentations of space groups, and this was done usually by physicists. Later
78 T. Janssen

-
1)001)
•
• t .~
-- • •
.
t
•
•·
--

t I'
•
•·
•• .1
•~ • ~ (!)~
•, :::=>

•
•• • •
•
.•
4,
.'
•
•
•
••
•
.1

Fig. 1. - Diffraction pattern for a quasiperiodic structure. The peaks are arranged
in lines. .

followed intercalates and misfit structures, and finally in 1984 quasicrystals.

Was incommensurability in the early days considered as something rather ab-
normal, one now knows hundreds of different quasiperiodic systems. Also in
absolute quantity quasi periodicity is not rare. Many minerals in the earth
crust, such as plagioclase and feldspar, occur as incommensurate phases. In
the following we shall discuss the various ways of describing these structures.

1.2. Classes of Quasiperiodic Structures

The first materials to be treated as quasiperiodic were incommensurate modu-

lated crystal phases (IC phases)[80J . These systems show a basic structure that
has three-dimensional space group symmetry (and thus lattice periodicity) and
a periodic modulation such that the periodicity of the modulation is incom-
mensurate with the lattice of the basic structure. Consider a lattice periodic
structure with atoms of species 1,2, ... , s on positions

~j = ii + fj, j = 1,2, ... ,8. (1.5)

 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 79

Here the vectors n form a lattice. The vector 'G indicates the position of the
j-th atom in the unit cell. Now suppose there is a displacive wave described
by the vector functions l(x) with l(x) = l(x + 1). This means that the
positions of the modulated structure are

(1.6)

for some wave vector if. If the components of the wave vector with respect to the
reciprocal lattice are rational, there is still lattice periodicity, but the system
is quasiperiodic if one of these components is irrational. A simple example is
the sinusoidal modulation

Such a modulation is called displacive. In general, the displacement field may

have more than one modulation wave vector:

(1.8)

with vector functions l that are periodic in each of the d arguments. Exam-
ples of such displacive IC phases are 'Y - Na2C03[14, 1, 79], K2Se04[39, 5], a
whole series of A 2 BX4 compounds [with A = Rb, K, ammonium, tetramethy-
lammonium, etcetera, B = Se, Zn, Co, Mn, Fe, etcetera, X = Br, CI, ...][63,
41], biphenyl[17]' ThBr4 [10]. An example of rank higher than 4 is TTF-
TCNQ[20].

Another type of modulation is an occupation or density modulation. Here

the positions in the basic structure are taken with a certain probability by one
out of several atomic species (Cu or Au in CuAuI I [61, 83]) or by molecules in
one out of several orientations (N02 molecules pointing to the right or to the
left in NaN0 2), and this probability is periodic with a periodicity that does
not belong to the lattice of the basic structure. For example, the orthorhombic
compound NaN0 2 has a probability of finding the N0 2 molecule in one of
two orientations given by the function p(if.n) with if = 0:0,*. The function
p satisfies p(x) = p(x + 1), is 1/2 above a certain temperature TI, 1 below
another temperature T2 <Tl and 1/2+€+6cos(21rx) for T2 <T<Tl. In the
intermediate phase the structure is an occupation IC phase.

A second bigg class of quasiperiodic systems is formed by the composite

structures. Here one can distinguish two or more subsystems that are, in a
certain approximation, lattice periodic. If, however, the lattice parameters of
the subsystems are mutually incommensurate the full system does not have
lattice periodicity: a translation that leaves one subsystem invariant does not
leave necessarily another subsystem invariant. Each subsystem diffracts with
diffraction vectors belonging to its reciprocal lattice. Therefore, the diffrac-
tion vectors of the full system belong to linear combinations of the reciprocal
80 T. JanSSeIl

·········f· ··+···I ···t ···~···· ~···· . ~ . ~ .. :.... ®... ..

.•.
0 0 0 0 0 0 0 0

_.. '! '"'t,·· .. ... -' .... .... .. . .~ : ..• . ®.. e

.. .
. ~., ~ - ~- .~ ~ ~ ~ ~ .
~ ~

. ..
.. ~ . ~ .; ...$ . ~ ...
. . . . .
... '.'. . 't,
·· . ~.... ~... , ~.- .. ~ "' ~"" 0 0 0 0 0 0 0 0 0 0

.. " 6 " '! ,.. .~ .... ~... .~ ... .~ .... ~- _. . ~ .. ~ ~ ... Q Q Q ~ ~ ~

.. . ~ .. :.... . ®.. 0 0 0 0

··t··· .... ... .. ...

. ..
.. . . ·8 .. · e e e
. ~
.. ... .. -+- . ~ ~, ~.- ~ "~'''. ~ ~ ~

... .•... .; -- .,... ~ ' ... ~.... ~ . , .•.... . ~

"~" ' .. ~ .. 0 0 0 0 0 0 0

Fig. 2. - 3 types of quasiperiodic crystals: displacively modulated, occupation mod-

ulated and incommensurate composite. For the first the wave vector is horizontal, the
displacement vertical. In the second the radius of the circles is a measure for the prob-
ability of finding atom species A. In the composite structure 2 two-dimensionallatice
periodic structures show intergrowth. They have incommensurate lattice parameters.

lattices of the subsystems. Because these do not have a common basis, as a con-
sequence of the incommensurability, one needs more than three indices to label
the diffraction spots. As an example consider an orthorhombic system with
lattice parameters a, band c. Suppose that in the channels in the x-direction
there are chains with another type of atoms with equilibrium distance d. So
the atoms of system 1 are at positions (n}a, n2b, n3c) and those of subsystem
2 at (n 4 d, ~ + n5b, ~ + n6c) . The diffraction vectors then are at positions

k = (hla + mid, klb, flc) (1.9)

and these form a Fourier module of rank 4. This situation occurs, for example,
in TTF 7 I 5 - x [82J and Hg 3 - x AsF6 [33,36, 65J where I, resp. Hg, atoms are situ-
ated in channels in a host lattice formed by a periodic TTF, resp. AsF6, array.
Other such composite structures are intercalates (where a material intercalates
between the layers of a layered structure), misfit structures, and monolayers
on a substrate, for example Ar on graphite. In the case of an adsorbed mono-
layer the distances between the adsorbed atoms depend on pressure, and are,
generally, incommensurate with the periodicities of the substrate.

Finally, in 1984 an intermetallic alloy, AIMn, was discovered with a

diffraction pattern consisting of fairly sharp peaks and having the symmetry of
an icosahedron[71J. This point group symmetry is incompatible with a lattice
periodic arrangement in 3 dimensions. Later other materials like AIMnPd and
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 81

AICuFe[74, 75] were found with really sharp peaks. In Cartesian coordinates
the diffraction vectors were of the form
k = (hi + h4T, h2 + h5T, h3 + h6T)
with T = (J5 + 1)/2 the golden rule. Models for quasicrystals are quasiperi-
odic coverings of space with copies of a finite number of building blocks. For
example, a three-dimensional version of the Penrose tiling with two types of
building blocks (a thick and a thin rhombohedron with a volume ratio of T),
having icosahedral symmetry, has diffraction vectors as in (1.2). The building
block,> can be filled with atoms in a certain arrangement, and form the gener-
alizations of the unit cell in a lattice periodic structure. Also other symmetry
groups have been found, with decagonal, octagonal or dodecagonal symmetry.
These quasicrystals have diffraction patterns of rank 5 or 6. Notice that qua-
sicrystals are not the only quasiperiodic structures. As we have seen above
there are many more.

These classes of what can be called quasiperiodic structures are not mu-
tually exclusive. Most occupation modulated structures have also a displacive
modulation, the subsystems in a composite structure are generally not really
lattice periodic, but displacive Ie phases, and one has at least found one exam-
ple of a quasicrystal with a modulation. Also other systems, like long-period
structures, systems with anti-phase boundaries, and helical magnetic structures
are often quasiperiodic.

1.3. Embedding of Quasiperiodic Systems

There is a simple way to construct from a quasiperiodic three-dimensional

structure of rank n a lattice periodic structure in n dimensions. Suppose the
Fourier module of the quasiperiodic structure has elements
n
k = 2>.a;.
,=1
(1.10)

The n basis vectors in physical space VE are projections of n linearly indepen-

dent vectors (it;, b:), where the vectors b: are (n-3 )-dimensional vectors in an
additional space VI. The n vectors span the reciprocal lattice E*, and to each
vector (1.10) in the Fourier module corresponds a vector
n n
ks = Lh.(a;, b;) = Lhia;•. (1.11)
i=l .=1
The lattice E* is the reciprocal lattice for the direct lattice E with basis vectors
asi. Because the rank of the Fourier module is n, the projection of E* on
physical space is one-to-one.
82 T. Janssen

A point in physical spaee corresIxmds uniquely to a point in the unit ('ell

of E. We write i E VI';' as (f,0). Then
n
(i,O) = L~,as, mod E, (1.12)
,=1
where the lattice eoordinates 0 ~ ~, < 1 follow from

~. = Frac[i.ii;]. (1.13)

If the quasiperiodic crystal is described by its (electron) density function p(T)

the points in the unit cell corresponding to an i E VE form a dense set, and
in this way a function in the unit cell is defined. Due to the periodicity in
n-dimensional space, it gives a periodic function in n-dimensional space that
corresponds to the three-dimensional function p.

An equivalent embedding is the following. The density function can be

expanded in a Fourier series, for which the wave vectors belong to the Fourier
module M* of the quasiperiodic crystal.

p(T) = L p(k) exp(ik.T}. (1.14)

kEM*

Then a function in n-dimensional space is defined by

ps(i,rl) = L p(k)exp(ik.r+k/.i/), (1.15)

kEM*

where (k, k/) E E* is the unique vector that is projected on k. By construction

this function Ps has lattice periodicity E.

If the structure consists of point atoms, the density function is a set

of delta peaks. The resulting embedding then consists of a periodic array
of (n - 3)-dimensional hyperplanes. For example consider the 3-dimensional
displacively modulated structure with atoms at

ii + !cos(27rif.ii). (1.16)

Then the embedded structure is a periodic array of lines, parametrized with a

parameter t and given as

iE = ii + !cos(27r(if.ii+t)), (1.17)

Clearly this pattern of lines is invariant under the 4 translations

(a, -if.a), (b, -if.b), (2, -if.C), (0,1).

 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 83

,, ,,
,, ,, ,
,, ,,
,, ,,
,, ,, ,, ,
,, ,, ,,
, ,, ,,
,, , ,,
,, ,, ,
,, ,, ,,
,, ,
,
,, ,,
,, ,, ,, ,
,, ,, ,,
, ,, ,
,, , ,,
,, ,, ,,
,, ,,
,, ,,
, ,, ,,
,, , ,,

I
I I I

I J I
Fig. 3. - Embedding of a number of quasiperiodic crystals in higher-dimensional
space: displacively modulated, occupation modulated, incommensurate composite
and quasi crystalline. For the occupation modulated structure the thickness of the
line is a measure of the probability. The subsystems for the composite structure are
displacively modulated. The atomic surfaces of the quasicrystal are bounded.

These 4 vectors span the lattice E.

Notice that the density function determines completely the embedded

function in n dimensions, and vice versa. The function peT) is simply the
restriction of Ps to physical space. Therefore, the information content of the
n-dimensional description is exactly the same as that of the 3-dimensional
description. It is just the way of presentation.

For point atoms, the (n-3)-dimensional hyperplanes (Le. the lines when
n=4) are called the atomic surfaces. For Ie phases and composite structures
these are without boundaries. For quasicrystals there is an infinite set of dis-
joint atomic surfaces. Examples are given in Fig.3. Structure determination for
quasiperiodic crystals comes down to determining the positions and the forms
of the atomic surfaces in the unit cell. This is especially a convenient way
84 T. Janssen

of presenting the problem if the form can be described by a small number of

parameters. For example, this is the case for a displacive IC phase if the mod-
ulation is sinusoidal. Then only amplitude and phase of the modulation, and
direction of the displacement enter as parameters. Sometimes the number of
parameters may become large, if the form of the atomic surface is complicated.
This is not necessarily the case for those models where one has found fractal
atomic surfaces. Although these have a rich structure, it does not mean that
that cannot be expressed by a small number of parameters.

1.4. Superspace Groups

A lattice periodic structure in n dimensions has necessarily an n-dimensional

space group as symmetry group. If {Rslas } is a space group element, for which
the action on a point T" is given by

a density function Ps is invariant if

If Ps is the embedding of the density function of a quasiperiodic function and is

invariant under the space group element one says that {Rslas } is a symmetry
transformation for p( f') in 3 dimensions. The orthogonal transformations Rs
form the point group. Generally there are limitations for this point group. The
reason is that the structure in n dimensions is constructed from the projection
of Bragg peaks on physical space. This means that the length scale in VI is
not fixed. Therefore, symmetry operations can only be relevant if they do
not mix VE and VI. For this reason an n-dimensional space group is called
a superspace group if its point group leaves a three-dimensional hyperplane
invariant that has an incommensurate orientation, i.e. for which the projection
of the reciprocal lattice on this plane is one-to-one. Consequently, point group
elements occurring in a superspace group can be written as pairs (R, RI) of
a three-dimensional orthogonal transformation R and an (n - 3}-dimensional
orthogonal transformation RI. Then an arbitrary element of a superspace
group can be written as

(1.18)

Invariance of a quasiperiodic crystal under a superspace group occurs if for the

embedded density function one has

(1.19)
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 85

Such a symmetry relation can also be written for the Fourier transform.
The Fourier components of Ps are the Fourier components of the 3-dimensional
quasiperiodic structure. Moreover, if Rs belongs to the point group it leaves
the reciprocal lattice invariant. Because of the one-to-one correspondence this
means that the Fourier module is invariant under R. The symmetry relation
(1.19) can then be written as

(1.20)

This means that for oi-thogonal transformations R in the point group such that
Rk = k the Fourier component p(k) vanishes if

(1.21)
This leads to systematic extinctions just as in ordinary crystallography. As an
example consider a superspace group element with orthogonal transformation
consisting of a mirror in VE combined with the identity in VI, and a translation
(iis3 + iis4 )/2.

Any vector ks = h2a:2 + h3a:3 + h4ii: 4 is left invariant by the orthogonal trans-
formation. Therefore p(k) = 0 if ks.iis =f. 0 mod 7l., i.e. if (h3 + h4)/2 =f. 0 mod
1, or in other words p(k) = 0 for all k =[Oh2h3h4] with h3 + h4 odd. The pres-
ence of such systematic extinctions leads on the other hand to the conclusion
that such a superspace group symmetry is present.

For structure determination it is, therefore, of great importance to have

all the possible superspace groups. The practical problem here is the rapid
increase of the number of groups if the number of dimensions increases. In
1,2,3 and 4 dimensions, the number of space groups is, respectively 2, 17,
230 and 4895[16]. Although superspace groups form a subset, their number is
already 791 in dimension 4[81]. (In fact this number is larger than that of the
corresponding 4-dimensional space groups because one has to use a different
equivalence relation). It does not make sense to derive all possible superspace
groups up to a certain dimension, especially because the number of groups used
till now for the characterization of investigated materials is much smaller. A
better solution is to use computer programs to derive those groups that are
relevant for the material one wants to study. Only special cases, like the 4-
dimensional superspace groups for Ie modulated phases and the superspace
groups for quasicrystals up to dimension 6 have been tabulated[42].

The description in superspace and the characterization of the symmetry

by superspa<:e groups has been incorporated in a number of computer programs
86 T. Janssen

for structure determination of incommensurate and other quasiperiodic mate-

rials. The first one was written by Yamamoto[84]. Other programs are those of
Petricek and Paciorek. A great number of structure determinations have been
performed in this context.

1.5. Action of Symmetry Groups in 3-dimensional Space

Although the description of quasiperiodic systems in a higher-dimensional space

is convenient for crystallography, a description of the symmetry that remains
fully in physical space has also advantages. Here we shall comment on the
relations between the two approaches. Because the information contained in
the unit cell in n dimensions is exactly the same as that in the infinite three-
dimensional structure the two approaches are basically the same.

Incommensurate modulated structures have been treated often by means

of irreducible representations of three-dimensional space groups. This was the
language used by solid state physicists dealing with lattice dynamics. An irre-
ducible representation of a space group is characterized by the star of a wave
vector (the image of a vector in the Brillouin zone under the point group, mod-
ulo the reciprocal lattice) and a label for an irreducible, in general projective,
representation of the point group of the wave vector (i.e. the subgroup of the
point group leaving the vector invariant, modulo reciprocal lattice vectors).
This is not the place to discuss in detail irreducible representations of space
groups. We briefly consider the procedure. Suppose the elements of the space
group are {Ria}, where R is an element of the point group, and ii a translation
(not necessarily a lattice translation). Consider a vector k in the Brillouin zone
(the unit cell of the reciprocal lattice). The star of k is the set of different
vectors (modulo the reciprocal lattice A") Rk with R in the point group. The
little group of k is the subgroup of the space group consisting of all elements
{Ria} for which Rk = k mod A". The little representation then is given by

(1.22)

Here r II is an irreducible repres{'ntation of the point group of k. In general, it

is not an ordinary representation, but on{' that satisfi{'s

w(R. S) = ('xp{i(Rk - k).as). (1.23)

with factor system w. Only if w is identically unity does one have an ordinary
represmtatioll. This happens in the CIl..'I(' of symmorphic space groups, or wh{'n
the wav", v('('tor k lies in th{' int{'rior of th{' Brillouin zon{'. The corr{'sponding
irreducible repn's{'ntat.ion of the full space group th{'n is obtain{'d by induc-
tion. This implil's that til<' diIlll'nsion of the irreducible represC'ntation is thl'
 FROM QUASIPERlODIC TO MORE COMPLEX SYSTEMS 87

dimension of r v(K) and the number of points of the star of k. Consider the
situation that the irreducible representation r v is one-dimensional, i.e. consist-
ing of phase factors exp( i4>( R)). An arbitrary function, e.g. a density function,
can be decomposed in components belonging to one irreducible representation.
In the case of a one-dimensional r v these components can in turn be written
as sum of terms belonging to one of the branches of the star of k. In particular,
this is the case for a modulation of a basic structure with space group symme-
try G. Then for a component jJ belonging to an irreducible representation of
G one has
jJ(k) -+ exp(i4>(R, k))exp(ik.a)jJ(k); Rk = k modA*, (1.24)
and
jJ(k') -+ exp[i,p(R, k')]jJ(Rk'); Rk' =F k' mod A*. (1.25)

The action of {Ria} E G on the component jJ(k) is just a phase shift. This is
exactly as the action of a superspace group element on jJ(k).

For a density function of a quasiperiodic structure the Fourier compo-

nents jJ(k) are just the Fourier components of the corresponding periodic n-
dimensional structure. As said before these satisfy the following relation, if
one considers an element {(R,RI)I(a,aI)} of the superspace group:
(1.26)
For an Ie modulated phase the elements {Ria} belong to the space group of
the basic structure. That implies that the action of such an element, which
satisfies Rk = k on jJ( k), is just a phase shift exp(ikI .aI)' This shows the
equivalence of the description by means of a superspace group and that using
irreducible representations. However, generally, there is no basic structure and
irreducible representations cannot be used in an <!.I?vious way.

The relation (1.26) between Fourier components of a quasiperiodic system

with superspace group symmetry G can be written in an alternative way.
jJ(k) = exp[iRk.ii)] exp(i4>(R, k))jJ(Rk). (1.27)
For 4>(R, k)=O this reduces to the usual condition for having a symmetry trans-
formation {Ria}. The relation (1.27) is a generalization ofthat condition. The
usual condition is only satisfied modulo a phase factor, which corresponds to
a shift of the modulation wave with respect to the lattice. Such a generalized
symmetry condition has been used in several contexts. For systems with elec-
tromagnetic potentials it is known as a compensating gauge trans/ormation[40].
From the properties of the translation parts of space group elements it follows
that 4>( R, k) is linear in k and satisfies

4>(R, k) + 4>(8, Rk) = 4>(R8, k) mod 211". (1.28)

88 T. Janssen

Conversely these properties follow from quite general considerations. Suppose

the density component p(k) corresponds to the order parameter of a phase
transition to an IC phase. It then belongs to an irreducible representation
characterized by (kv). Purely from the symmetry point of view the component
p(2k) may belong to (2k, 11), where 11 and v have no relation. However, in Lan-
dau theory of phase transitions the relevant component p(2k) is the square of
the order parameter and is coupled to p(k). This fixes the irredudble represen-
tation 11 at 2k. If under a symmetry operation from the space group jJ(k) gets
a minus sign, the component p(2k) gets a + sign. If one uses the superspace
group this is automatically taken into account, because the factor exp[iRlkI .aI]
for p(k) is replaced by exp[2iRlkI .aI] for p(2k). Along these lines one can prove
the linearity of cI>(R, k) from Landau theory. In the superspace formulation it
is just a consequence of the n-dimensional symmetry transformation[43].

One may go one step further and consider the condition

p(k) = exp[iw(R, k)]p(Rk) (1.29)

as a generalized symmetry condition for 'pOint group transformations. _For a

lattice periodic structure the phase W(R, k) is just the scalar product Rk.a for
some vector a. If such a vector can be found such that the relation holds for
every component p(k), one has found a symmetry operation, which, of course,
corresponds to a space group element {Ria}. Bienenstock and Ewald[12] have
presented this alternative view on symmetry of crystals, but they were well
aware of the fact that both approaches are equivalent. The same procedure
has been used by Mermin et a1.[67, 68, 60] to rederive higher-dimensional space
groups, but in the form of eq. (1.29).

In conclusion, one may state that the superspace symmetry of the em-
bedded structure corresponding to a quasiperiodic system can be interpreted
as (non-Euclidean) transformations in physical space. They are non-Euclidean
because a shift of the modulation wave with respect to the lattice does not leave
the interatomic distances the same. The supers pace transformations consist of
a Euclidean transformation combined with a shift in the phase of the Fourier
components. For a modulated structure this is just the translation of the mod-
ulation function with respect to the basic structure. In that case there is a
direct relationship with irreducible representations of three-dimensional space
groups.

As ~n example consider an IC phase with basic structure with space

group Pcmn. Consider the case where there is a displacive modulation in the b-
direction with wave vector 'Y~. The point group ofthis wave vector is the group
mm2, generated by a mirror perpendicular to the a-axis and one perpendicular
to the b-axis. Because the displacement is in the b-direction one has r(mx)=l,
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 89

r(my)= -1. (Because k is in the interior of the Brillouin zone the factor system
w is identically unity). The modulation therefore can be characterized by an
irreducible representation of Pcmn with labels k = "(i!' and the label v for the
irreducible representation of mm2 with character x(l,m x ,m y ,2 z ) = (1,1,-1,-1).

As superspace group the point group is generated by mx with RI=1

(because mxk = k), by my with RI=1 (because myk = k), and by m z with
RI= -1 (because mzk = -k). The modulation function gets a minus sign under
my, which can be compensated by a shift of 1/2 in the phase of the modulation.
Therefore, the 3 generators of the 4-dimensional superspace group (next to the
translations) are

{(m x ,+I)

{(my, +1)

{(mz, -1)
This group has symbol Pcmn(1sI). The s in the symbol means a shift of 1/2
in the 4-th component, which is associated with my.

The superspace group symmetry implies the following relations for the
Fourier components jJ(k) with k = hklm:
jJ([hklmJ) = jJ([hklmJ)(-I)hH
= jJ([hklmJ)(-I)k
jJ(hklm]) ( _1)h+k+l
which give the systematic extinctions
jJ([Oktm]) = 0 for lodd; jJ([hkOO]) = 0 for h + k odd.

1.6. Scale Symmetries

Points obtained from the origin by a lattice translation group have trivial scal-
ing properties: if one multiplies all position vectors by a common integer factor,
the transformed points belong to the lattice as well. For more complicated pe-
riodic structures this may be the case as well. For example, the N aCt structure
with N a atoms on positions (nI, n2, n3) and Ct atoms at (nl + ~, n2 + ~, n3 + ~)
is invariant under scaling by an odd integer factor. Quasiperiodic tHings may
also have a scaling property, but in general, the scaling factor is not an integer.
For example, the Fibonacci chain with atoms at positions
Xn = 1- T + n(3T - 4) + (T - 2)Frac(nT), (T = (v5 + 1)/2), (1.30)
90 T. Janssen

is invariant under multiplication by r: rXn = X n ,. This transformation

maps the endpoints of the Fibonacci tiling on other end points. One can write
the positions also as m + nr (n, m E Z), where m is a function of n. Then
multiplication yields

m + nr ---+ n + (m+n)r

i.e. m' = nand n' = m + n. This can be interpreted as a basis transformation

for the embedding, which has basis vectors (1,1 - r) and (r - 1, 1). The basis
transformation is represented by the matrix

(1.31)

It has eigenvalues rand 1 - r, the corresponding eigenspaces are VE and VI,

respectively. This basis transformation of the two-dimensional lattice generates
an infinite group. On the line, however, the multiplication by r only yields a
semi-group. Janner has pointed out that a characteristic difference between
the infinite point group generated by S and a finite point group of a super-
space group is that the former does not leave a positive definite metric tensor
invariant. Consider the transformation Ef2. It leaves invariant the metric tensor
g:
-2a a)
g= ( a 2a ' (1.32)

which is indefinite. Notice that S does not leave a metric tensor invariant.
The orbits of the transformation Ef2 fallon hyperbolas. In physical space (here
one-dimensional) the scale transformations form a semi-group, because there is
not always an inverse. This semi-group corresponds to a group of basis trans-
formations in superspace (here two-dimensional). Such scaling transformations
occur also for tilings in 2 and 3 dimensions. They are associated with infla-
tion rules, and correspond in the embedding with basis transformations of the
n-dimensional lattice. For example, for the standard octagonal tiling, the set
of vertices is transformed onto itself by a multiplication by a factor of 1 + ../2.
This is one of the eigenvalues of the basis transformation

S = ( i
-1
1 0
1
1
o
1
1
1
-~1
1
)
'
(1.33)

which has eigenvalues 1 ±../2, The semi-group of scale transformations in the

plane stems from a group of basis transformations of the 4-dimensionallattice.

Just as space groups are subgroups of the affine group, one can con-
sider such subgroups which have as point group one of these infinite groups
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 91

Fig. 4. - Scale invariance of the standard octagonal tiling. When the pattern of the
tiling given by thin lines is multiplied by a factor 1 + ,f2 each vertex goes to another
vertex.

of basis transformations. Also in this case one can have symmorphic and non-
symmorphic groups. The latter have elements {Sla} with a point group element
S of infinite order and a translation part a that cannot be transformed to a
lattice transformation by a shift of origin. Such scale-space groups can be non-
trivial extensions of a lattice translation group and a point group of infinite
order.

1. 7. Hierarchy of Structures

For a long time lattice periodicity was considered to be characteristic for crys-
tals. Only in the last two decades one has investigated more general crystals,
the quasiperiodic structures. Many examples have already been found in na-
ture. One can ask whether it is possible to have still more general types. If
one introduces a measure for the complexity one can construct a hierarchy of
structures. The order can be seen in the diffraction pattern, and one possible
definition of crystal is "a system with delta peaks in the diffraction pattern".
However, it is possible to construct deterministic models that do not have this
property.

An ordinary crystal is lattice periodic. It has then a unit cell, that charac-
terizes the structure completely. Moreover, there exists also a reciprocal lattice,
92 T. Janssen

with a unit cell that is called the Brillouin zone. The unit cell needs only to
exist in an average sense. The degree of order in the unit cell can vary a lot.

Quasiperiodic structures are most easily characterized by their Fourier

spectrum. A structure which shows delta peaks in the Fourier transform at the
positions of a Fourier module of finite rank is quasiperiodic. A lattice periodic
structure then is quasiperiodic of rank three. Equivalently one may say that a
d-dimensional structure is quasiperiodic of rank n if it is the restriction of an
n-dimensional lattice periodic structure to a d-dimensional hyperplane.

Almost periodic structures are characterized by a density function p(T)

such that for every e > 0 there is a relatively dense set of translations li such
that Ip(f) - p(f + li)1 < e. Relatively dense here means that there is a radius
~ such that in each sphere of radius R£ there is a translation vector li. The
theory of almost periodic functions was developed by H. Bohr[13]. An almost
periodic function has a Fourier transform consisting of delta peaks. Quasiperi-
odic systems are special cases of almost periodic systems. Any quasiperiodic
function is almost periodic.

If the conditions for almost periodicity are only satisfied locally, i.e. when
there is a radius A such that for each e > 0 there is a relatively dense set of
translations li such that Ip(f) - p(r + li)1 < e for each f inside a sphere with
radius A, the system is called weakly periodic. In general, the Fourier transform
is not a delta peak, but a continuous function.

Still more general, are systems with local correlation. The latter exist
always. Completely random systems are unphysical. There is a hierarchy in the
structures in the sense that lattice periodicity implies quasiperiodicity, which
in turn implies almost periodicity and weak periodicity.

Up to almost periodic systems there are Bragg peaks (delta peaks) in

the Fourier transform. In real systems there is in addition always an abso-
lute continuous part, because of size and temperature effects. Between almost
and weakly periodic systems are situated systems with a singular continuous
Fourier transform (see the section on diffraction). One-dimensional examples
of all these structur~ can be constructed from substitutional chains (see sec-
tion 4). Such chains are sequences of letters from an alphabet with p letters. If
one replaces the vertices of an equidistant array by atoms corresponding to the
letters in the sequence, or the letters of the sequence by intervals of different
lengths, one obtains a model system that belongs to one of the classes men-
tioned above. Physical realizations of such models have been made in the form
of multilayer systems[6, 73]. Whether more general systems than quasiperiodic
systems occur in nature is not clear at the moment.
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 93

A brief summary is given in the following table.

Type Characteristic Examples

Lattice 3-dimensional lattice Ordinary crystals
Periodic Unit cell Plastic phases
Brillouin zone
3
k = I:i-l hiCii
~

Quasiperiodic No lattice IC phases

No Brillouin zone Incommensurate composites
k = I:~-l hia: Quasicrystals
Almost No finite rank module Regular fractals
Periodic Bragg peaks
Critical No Bragg peaks Thue-Morse chain
Singular continuous spectrum
Weakly Absolute continuous spectrum
Periodic Locally "almost periodic"
Amorphous Only finite range ordering
Random No correlation (non-physical)

1.8. Physical Origin of Quasiperiodicity

In the last two decades one has discovered a great number of quasiperiodic
structures. The reason for having a thermodynamically stable state that is
quasiperiodic is as difficult to give as one for periodic systems. There is no a
priori reason why the ground state would be one of the two. However, one may
say that, generally, the reason for quasiperiodicity in crystals is competition
between forces or incompatibility of length scales.

The origin of an incommensurate magnetic structure can be seen from

a model proposed by Elliott[23]. If one considers a linear chain of Ising spins
with first and second neighbour interaction, i.e. with energy

L[J1SnSn- 1 + J2 SnSn- 2 ] (1.34)

n

the system shows in molecular field approximation a phase transition towards

an ordered phase with sinusoidally varying magnetic moments with wave vector

(1.35)

or an (anti}ferromagnetic ordering for (lJd4J2 i > 1. Later two- and three

dimensional models with similar competitive interactions turned out to have
incommensurate phases if there was a certain balance between the coupling
94 T. Janssen

parameters. A well known example of such a model is the ANNNI (axial

next-nearest neighbour Ising) model. Its phase diagram shows many incom-
mensurately varying magnetic phases[7, 70, 85].

For dielectrics a simple model that gives, again in molecular field approx-
imation, structural phase transitions and a rich phase diagram with IC phases
is the DIFFOUR model (discrete frustrated ¢4 model)[46, 47, 48]. It consists
in its simplest form of a linear chain of particles with one degree of freedom,
e.g. a rotation or a displacement X n , and with potential energy

(1.36)

Competition between the terms with (3 and 6 can give rise to IC phases.

A model for a monolayer adsorbed on a substrate is the Frenkel-Konto-

rova model[25]. The substrate is modelled by a sinusoidal background, the
monolayer is a linear chain with harmonic interactions.

v = L)~(Xn - Xn-l - b)2 + Acos (27rx n la + ¢)]. (1.37)

n

This model has been investigated in detail by many authors. If alb is irrational,
the ground state is incommensurately modulated. The form of the modulation
function depends on the parameter A. For A large enough the modulation
function is not a smooth function[4].

The origin for the existence of quasicrystals is less clear. A possible

explanation is a kind of Hume-Rothery mechanism[26]. However, the question
of relative stability of quasicrystalline and crystalline phases remains open.

2. DIFFRACTION

2.1. Structure factor

Scattering experiments from crystalline materials are determined by the struc-

ture factor. If the positions of the atoms are Tn and the atom at position Tn
has an atomic scattering factor In the structure factor is defined as

(2.38)
n
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 95

AID
6
5
4
3
2
1

0 1 1 !
6 4 3
-1

-5 -4 3 2 1 0 1 2 3 4 5 BID

Fig. 5. - Phase diagram for the DIFFOUR model. High values of AID correspond
to high temperatures. P: paraphase, which is unmodulated. The fractions give the
modulation wave vector of the modulated phase with respect to the original chain. In
the hatched regions IC phases can be found. Along the instability line of the P-phase
all values of the wave vector occur, from 0 to 1/2.

This is the Fourier transform of a density function consisting of D-functions on

the position Tn. The values of Ii for which the structure factor is nonzero is
called the Fourier spectrum. In a scattering experiment from a powder the
measured quantity is the structure factor integrated over a sphere of radius
H. The radial structure factor can, generally, be written as the sum of three
components.
S(H) = S(H)point + S(H)ac + S(H)sc, (2.39)

called respectively the point, absolute continuous and singular continuous part.
The integrated structure factor is the function

F(H) = foB S(H)dH. (2.40)

96 T. Janssen

For a lattice periodic crystal the Fourier spectrum belongs to the re-
ciprocal lattice, the structure factor consists of isolated Bragg peaks and the
function F(H) is piecewise constant with isolated jumps. One has

S(H) = -
47r
1/ '"' -
dO ~ b(H - 27rK)-
KEA-
-1L: 13 exp[iH.~l.·
S
s

j=l
- (2.41 )

A quasiperiodic structure has a Fourier spectrum that belongs to the Fourier

module. The function S(H) now consists of a dense set of Bragg peaks. The
same is true for an almost periodic function. On the other hand, an amor-
phous structure has only an absolute continuous component, and F(H) is a
continuous, monotonically nondecreasing function. A real lattice periodic or
quasiperiodic structure does not have a pure point spectrum. Because of disor-
der and thermal vibrations in real crystals, there is also an absolute continuous
component, which is the diffuse scattering background. This diffuse scattering
contains also structural information.

As discussed in the preceding section, there is a class of structures without

Bragg peaks, but having a singular continuous component possibly in combi-
nation with an absolute continuous part. The singular continuous component
is characterized by the fact that F(H)sc is a Cantor function, or as physicists
call it sometimes a devil's staircase. The diffraction pattern of such a structure
will show peaks with a certain width, but these are due to the finite resolution.
The measured intensity is the average over a neighbourhood. For increasing
resolution more and more peaks will be seen with less and less intensity.

Experimentally it is not always easy to classify a structure along these

lines. Even a lattice periodic system has a finite coherence length, which means
that delta peaks (Bragg peaks) are an idealization. Other complicating factors
are the finite size of a real crystal, defects and impurities. If one observes
satellites this is generally only possible because satellites go down in intensity
for increasing order. That means that in practice one observes only a discrete
set of peaks. Whether their positions are commensurate or incommensurate is
often difficult to say unambiguously, because of the uncertainty in their posi-
tions. Even if the satellites move with respect to main reflections as function
of temperature this is not a watertight proof. There are examples, where the
length of the modulation wave vector as function of temperature shows many
plateaus, and resembles a Cantor function. In principle, the components of the
modulation wave vector could be always rational. The same applies to qua-
sicrystals where the difference between a really quasiperiodic strcuture and a
lattice periodic approximant becomes negligible if the unit cell of the approxi-
mant becomes very large. Actually the problem is not very much different from
the situation in an ordinary lattice periodic system, where it is also difficult
to distinguish an orthorhombic structure from a monoclinic one with an angle
that is within the accuracy equal to ninety degrees, at least if the filling of the
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 97

0.2 0.4 0.6 0.6

Q-VECTOR

Fig. 6. - Integrated diffraction intensity for a critical structure with singular con-
tinuous spectrum. It is a Cantor function. At the top two diffraction patterns are
given with different resolution. They are the diffraction intensity averaged over a
neighbourhood.

monoclinic unit cell has also pseudo-orthorhombic symmetry. In all these cases
one works with an idealized picture.

Therefore, one will never be able to prOve experimentally the existence

of an almost periodic, but not quasiperiodic structure. Nevertheless it may
be convenient in certain cases to describe a structure as almost periodic, if
there is for example a regularity in the intensities of peaks at k = C" /2 n if n
becomes large. It is certainly interesting to keep mOre general structures than
quasiperiodic systems in mind.

2.2. Structure factor of quasiperiodic structures

Because the Fourier components of a quasiperiodic structure are equal to those

of the embedded periodic structure in n dimensions, one can calculate them
from the contents of the unit cell. In n-dimensional space the normalized
98 T. Janssen

structure factor is

(2.42)

where the sum runs over all s particles in the unit cell. For a quasiperiodic
structure one has s atomic surfaces OJ in the unit cell. The atomic surfaces
have a local scattering function Ii (1"1) and are described by positions J (f'I), v
vectors parallel to physical space. For ideal quasicrystals il:J does not depend
on 1"1 because the atomic surfaces are flat and parallel to VI. Then for fl in the
Fourier module, i.e. a projection of a reciprocal lattice vector in n dimensions,

(2.43)

Here 0 is the sum of the volumina IOjl and (fl,flI) is the reciprocal lattice
vector corresponding to fl.

For a displacively modulated structure all atomic surfaces have infinite

extent and the expression becomes

~ = sO
S(H) 1~ -
L...Jliexp(iH.fj)
J=1
1 -
Q
-
exp[iH,'uj(f'I) +iHI.1"I]d1"I. (2.44)

Here iij is the modulation function. In the case of a rank four Ie phase this
becomes
8 1
S(fl) = ~ Lliexp(ifl.fj) [ exp(ifl.Uj(t) + 2rrimt)dt, (2.45)
s j=1 io
for an m-th order satellite.

For quasicrystals with flat atomic surfaces there is no displacement field

and the expression for the structure factor becomes

(2.46)

The simplest example is the Fibonacci chain with flat atomic surfaces consisting
of a line of length 7 in the vertices of a lattice spanned by (1, 1-7) and (7 -1, 1).
The Fourier module is spanned by (2+7,1-27)/5 and (27 -1, 2+7)/5. Then

IS(hI,h 2)1 = [ -I
7
j+r
-r/2
/2 exp(2rri(hl(I-27)+h2(2+7))t/5)dt [ = I -sin y I
Y
(2.47)
for y = rr[ hI (1 - 27) + h2 (2 + 7) l/5. This is the well known expression for the
Fourier transform of the Fibonacci chain.
 FROM QUASIPERlODIC TO MORE COMPLEX SYSTEMS 99

Ii II
I I I

II'
!i I
-ci

. i' I

c
ci~~~~~~~~~~~~~
-10.0 -S.O 0.0 S.O 10.0
a

Fig. 7. - Diffraction pattern for a Fibonacci chain. All peaks are at positions m+nr.
Because the intensity drops with increasing perpendicular component, the set of peaks
that can be measured above a certain threshold is discrete.

2.3. Influence of symmetry

The intensity of a Bragg peak in the diffraction pattern is given by I(k) =

IS(k) 12. The structure factor S( k) is the Fourier transform of a density function.
If the latter is invariant under a space group element (in arbitrary dimension,
i.e. 3 for lattice periodic and n for quasiperiodic systems) {Ria}, it satisfies
the relation
S(k) = exp(iRk.a) S(Rk). (2.48)

This implies that

1) the diffraction pattern is invariant under the point group K: I(k) = I(Rk)
for each R E K;
2) if Rk = k and k.a i- 0 mod 211' the intensity I(k) vanishes.
The latter is called a systematic extinction. For lattice periodic crystals these
correspond to the existence of screw axes or glide planes. In quasiperiodic
crystals the intensity vanishes if Rk = k for the 3-dimensional vector k and k.a+
kl.iiJ i- 0 mod 211', where (k, k/) is the n-dimensional reciprocal lattice vector
whose projection is k and (a, a/) is the translation part of the n-dimensional
superspace group element.

As an example consider the modulated IC ~ZnCl4' one member of the

large family of A 2 BX4 structures with an IC phase. The superspace group
is four-dimensional and has symbol Pcmn:lsl. Four basis vectors for the 4-
100 T. Janssen

dimensional lattice are

(ii,O), (b,O), (2, --y), (0,1),
where ii, band c form an orthogonal basis in 3 dimensions. Three other gener-
ators of the superspace group are

Here the translation parts are given with respect to the 4-dimensional basis.
To the three generators corrspond the following extinction rules for
k= [hkem] = hii* + kb* + lC' + mil

Rk = k ~ [Oklm] : e f:. even

~ [hOem]: m f:.. even
~ [hkOO]: h + k f:. even.
From the occurrence of such extinction rules one can conclude that the structure
has the considered symmetry.

Also the existence of scale symmetry can help in giving information about
the diffraction pattern. Consider as an example the Sierpinski gasket. This is
not a quasiperiodic system. It can be constructed as follows. Start with an
equilateral triangle of edge a. Then connect the middles of the edges and
multiply the figure with a factor of two, from one of the corners. The result
consists of 4 equilateral triangles, three top up and one top down. Connect the
three midpoints of each top up triangle and multiply again by a factor of two.
In this way the minimal distance between neighbours remains a. The number
of points in the n-th step is N n with Nn+l = N n + 3n . The structure factor in
the n-th step can be found recursively. If el and e2 are two vectors from the left
corner of the original, pointing along the edges and of length a the recurrence
relation is
(2.49)
where aj = ii.e'j (j = 1,2). Sn+l consists of three copies of Sn with ap-
propriate phase factors. The term with a minus-sign compensates for double
counting. Because the structure is limit periodic and therefore almost periodic
the diffraction pattern has Bragg peaks.

2.4. Thermal vibrations

Although perfect almost periodic functions (and a fortiori lattice periodic and
quasiperiodic ones) have a Fourier transform consisting of Bragg peaks, real
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 101

•
• • •
• • o:e:' ; ;.~".
>-0 e.. .: . • ...:~
I...;
· .. .. i:..
-::. .~.' •.•..
.. :!= •.~:
0'

o
... •
.•.... .
••••••
•
.
• o.

..=t .:. .=t:.

•• •

Le.:.-·
~...... . ...:..•.. .-.
"11111"
50 100 150 10
X Q-X

Fig. 8. - Sierpinski gasket and its diffraction pattern

crystals show another behaviour. The finite size of samples will lead to a
broadening of the peaks. Moreover, thermal vibrations will lead to a diffuse
scattering background. If one considers the oscillations around the equilibrium
positions as a Gaussian smearing out of the positions the density function for
a modulated IC phase can be written as

p(T) = 2;/[/,6(1"- ii - T; - 11, (q.ii) - V)exp(-iff Bj(ii)v)]dv. (2.50)

R,
The tensor B, is the thermal displacement tensor. The Fourier transform for
an IC phase of rank 4 becomes

F(k) = th
,=1
exp(ik.1"j) ! exp[ik.l1,(t) + ik1.t -;;r Bj(t)kf4]dt. (2.51)

For vectors k in the Fourier module this gives the intensity of the ideal static
structure diminished by a factor determined by the tensor B,. This suppression
factor is the Debye-Waller-factor.

2.5. Disorder

In the literature one can find fierce discusions about the possibility that the
structures found in nature, especially the quasicrystals, are not quasiperiodic
102 T. Janssen

at all, but either random systems, like random tilings, or periodic systems with
possibly large unit cells. In order to be able to discuss the differences we have
to consider the influence of disorder on diffraction.

First look at a finite system with N lattice sites, and suppose that n of
these sites are unoccupied, in a random fashion. The intensity of the scattering
then is given by
(2.52)
nm
where In is the scattering power of the atom at site nand rnm the connecting
vector between the sites of particles nand m. In the case of n vacancies, and
under the assumption that all atoms are equal the scattering can be approxi-
mated by

J(H) (N - n? 12 + (N;; n) n#m

L exp(iH.f'nm)
2 12 (2.53)

N - n)2 2,", . ~ ~ n(N - n) 2

(~ I L....-exp(zH.rnm ) + N I,
nm
where the first term in the last expression gives Bragg peaks at the reciprocal
lattice sites in the limit N ~ 00, and the second term is responsible for a
diffuse background, an absolute continuous spectrum. In this case, where there
is no correlation between the vacancies, the diffuse scattering does not have
any structure. This is different, when there is correlation. For example, when
the vacancies occur mainly in pairs along the z-axis the diffuse term becomes

n(N -n)
N 12 (cos(Hz) + 1),
which leads to a clustering of the diffuse scattering around the Bragg peaks.
The latter seem to have broad shoulders. Therefore, disorder in an otherwise
perfect crystal may blur the diffraction image.

If one goes to completely random systems, like gases, the scattering is

given by

J(H) L fJ J (exp(iH.r,j») (2.54)

J
tJ

L0]
I.JJ ~
7rr,]
dB sin B27rr;J exp( iH r '] cos B) (2.55)

'"' f-/sinHr,] (2.56)

L....- '] Hr,] .
']
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 103

1.2
"fort.15" - -
1

0.8

0.6

0.4

0.2

o
o 5 10 15 20 25 30 35 40

Fig. 9. - Diffraction from a chain with correlated vacancies

In the continuum limit this becomes

where the reduced radial distribution G(r) is given by

G(r) = 411"r[p(r) - Po]. (2.58)

The function gives the probability of finding a particle at a distance r from an-
other one. This fully disordered system has a scattering function that oscillates
with broad peaks.

However, a random system may lead also to rather sharp peaks. Consider
for example a series of length a and b. Hendricks and Teller[37] have calculated
the diffraction from a random stacking of layers of thickness a or b. They
show that there are both broad and narrow peaks. Rather sharp (not delta)
peaks are found for wave vectors that are simultaneously multiples of 211"1 a
and 211" lb. The intensity is then relatively high. Peaks with a larger width
104 T. Janssen

2.5
"fort.11" -
2

1.5

0.5

o
o 1 2 3 4 5 6 7 8
Fig. 10. - Fourier transform of random sequence of intervals with lengths 1 and T - 1
with frequencies T - 1 and 2 - T, respectively. Near the points where m ~ nT the
peaks are sharp and intense (after Hendricks and Teller).

are usually less intense. This makes why it is difficult to distinguish between
quasiperiodic quasicrystals, where always a certain degree of disorder will be
present, and random tiling systems. The situation with sharp strong peaks with
icosahedral symmetry has been found in a random packing system of vertex
sharing icosahedra by Stephens and Goldman[72].

3. PHONONS

3.1. Phonons in Ie phases

In lattice periodic systems the problem of lattice vibrations is enormously sim-

plified by the use of translation symmetry. The harmonic oscillations around
equilibria of an ordinary crystal are characterized by the irreducible represen-
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 105

tations of the space group, and these are labelled by (stars of) wave vectors
in the Brillouin zone. AB a consequence lattice vibrations labelled by different
wave vectors are decoupled and the, in principle infinite, problem is reduced to
the diagonalization of a finite-dimensional dynamical matrix. The dimension
is three times the number of particles in the unit cell.

For quasiperiodic crystals of rank higher than three there is no Brillouin

zone in three dimensions. One can introduce a Brillouin zone in n-dimensional
space but then the number of particles in the unit cell is the number of points
on the atomic surfaces, i.e. infinite. So, generally, the problem remains infinite-
dimensional. The existence of superspace group symmetry may give selection
rules, but it does not give the same simplification as in three dimensions.

Quasiperiodic structures consisting of point atoms can be embedded in

n-dimensional space as lattice periodic structures. The positions of the qua-
sicrystalline structure are the intersections of the atomic surfaces with physical
space. Oscillations around these equilibrium points can be described by a dis-
placement vector field on the atomic surfaces, that is time dependent. For the
simple case of a modulated Ie phase of rank four, the embedded structure is

(it + r, + h(q.it + r), r). (3.59)

The displacement of the points of the atomic surface for the atom at it +
fj in the basic structure and with internal coordinate r then are given by
UnjT exp(iwt)+c.c. In the harmonic approximation one gets for the equations
of motion
mJ W2UnJT = L4'(nj,n'j'i r )Un'J'T. (3.60)
n'j'

Here 4' is the dynamical matrix, the second derivative of the potential energy
with respect to the displacements UnjT. Lattice periodicity in 4 dimensions
allows to use Bloch's theorem. Then
-
UnjT -- mj
-! exp(-zk.n-zk/r)Uj(r
.- _. - --
+ q.n). (3.61)

According to Bloch's theorem the function Uj is periodic. The equations of

motion for the vector function tJj then become

w2tJJ(r) = L(mjmJ,)-!V(it,jj'ir)tJj,(r +q'.it). (3.62)

nJ'
where
V(it-it',jj',r) = 4'(nj,n'j'ir).
The vector function tJJ (r) corresponds to the eigenvector of an oscillation mode
in a normal crystal. There the eigenvector €(klli j) gives the displacement of
106 T. Janssen

particle j in the unit cell, when the crystal is in the vibration mode character-
ized by k, 1/. Here OJ (T) is the displacement of particle j in the unit cell for
phase T in a vibration mode kl/. A more complete labelling would require a
symbol OJv (T).

The equations of motion (3.62) can be studied by approximations. One

is a perturbative approach, in which the amplitude Ih I of the modulation plays
the role of the perturbation parameter. A second possibility is to study a series
of commensurate approximants to the incommensurate modulation vector if.

A simple one-dimensional model is the following. Consider a chain of

particles with harmonic nearest neighbour interactions.

(3.63)
n

where an is a quasiperiodic sequence of real numbers. One possibility, the

modulated spring model[51J, is

an = a[l + 8cos(qn + 4»]. (3.64)

Another model, that is closer to a quasicrystal, considers a chain with 2 kinds

of springs.
if nT < T mod 1
(3.65)
if nT ~ T mod 1
If one approximates q by L/ Nor T by L/N = Fn+d Fn (the Fibonacci numbers)
the equations of motion

(3.66)

form a finite set if one uses

Un+N = Un exp(ikN) , (-7f/N < k:5 7f/N). (3.67)

For each approximant one can calculate numerically the spectrum. If there is
a limit for L/N --+ q it is considered to be the spectrum of the quasiperiodic
system. In fig.ll the spectra are shown for the model (3.64), for all approx-
imants L/N with N :5 50[51]. From the figure the self-similar properties are
evident, although a rigorous proof of this property is not known.

For the model with springs (3.65) the situation is essentially the same.
In fig.12 the integrated density of states is given for two approximants to T.
The integrated density of states I(w) is the number of modes with frequency
smaller than w. There is a limiting behaviour which is a Cantor function. This
implies that the spectrum is singular continuous.
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 107

, /
\
\ /
\ /

"I
\ /
/i l
/1'
,~rt1'·
\.
,
\ ~:' /
I

Fig. 11. - Spectra modulated spring model for approximants q = LIN with 0:<:;
N :<:;50 (6 = 1). For each of these values the spectrum is plotted vertically.

The modulated spring model does not show some modes that are typi-
cal for quasiperiodic structures. The reason is that there is no mechanism for
producing an Ie phase in the model. For that one can consider the DIFFOUR
model (1.36) . In the high temperature phase (Le. for high values of the param-
eter a) the equidistant array is stable. The dispersion curves for the oscillations
around these equilibrium points follow immediately:

w~ = a + 2{3 cos(k) + 26 cos(2k). (3.68)

If the temperature (or a) decreases there is an instability for k = cos- 1 ({3/ 46)
(or for k equal to 0 or 11"). After the phase transition one has, generally, an
incommensurate phase. The equations of motion

(3.69)
108 T. Janssen

~3
"
3Y89

~
4 ---w 2

Fig. 12. - Integrated density of states modulated spring model. Compared are two
approximants to the golden mean. In the quasiperiodic limit this integrated density
is a Cantor function.

can be studied again by means of a series of approximants. This means one

considers values of the parameters for which Xn = Xn+N is a solution, in which
case one may use Un+N = Un exp( ikN). The dispersion curves of one such an
approximant are given in fig.13 [471

The number of curves is equal to the number of particles in the unit

cell, i.e. N. There are two branches for which the frequency tends to zero if
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 109

3.5 . . - - - - - - - - - - - - - - ,

3.0 t---========:::f

2.5t::===~

1.5

1.0

0.5

0.0 -+-::::;':"--r~--''---'''---r---r----r"---.---1
0.0 0.2 0.4 0.6 O.B 1.0
-q

Fig. 13. - Dispersion curves for an N = 11 approximant, which is a solution of the

DIFFOUR model. The 2 branches with zero frequency are the phason and acoustic
branch.

k goes to zero. One is the usual acoustic branch. Its slope is the velocity of
sound. The eigenvector of the corresponding modes have the property that for
neighbouring sites the values of f(kv;j) are almost the same. The other branch
has a very different character. The eigenvector at w = 0 corresponds to a
shift of the modulation function Xn with respect to the lattice. It is a so-called
pltason mode. Strictly speaking, if one considers an approximant, the frequency
w is not exactly equal to zero for k = O. However, for N not too small the
"phason gap" here is very small. In the limit N -+ 00 the phason gap is zero.
This corresponds to the fact that for an incommensurate configuration {Xn}
the energy does not change if one displaces the origin of the modulation. One of
the other branches depends strongly on temperature. At the phase transition
it has also w = 0, but this frequency rises sharply with increasing value of
Ti-T. Its eigenvector corresponds mainly to an oscillation in the amplitude of
the modulation (i.e. f(n) is proportional to xn). This branch is the amplitudon
branch.
110 T. Janssen

3.2. Spectra

From Fig.11 it is clear that spectra have self-similarity properties, in the q, w-

plane. Also for fixed q this is the case. To find a characterisation of the spectra
one plots the Lebesgue measure of the spectra (Le. the sum of the widths
of all bands) against the order of the approximant. This comes down to a
characterisation of the limit spectrum for the quasiperiodic structure, if one
maps the logarithm of the total measure as function of the logarithm of the
resolution. In Fig. the logarithm of the spectral measure is plotted against
the order of the approximation, i.e. the logarithm of the size of the unit cell
(Np ~ ~p if ~ is the scaling factor of the quasiperiodic structure). It shows
that the measure of the quasiperiodic spectrum is zero, and this limit is reached
exponentially with an exponent f: if ~i (i = 1, ... , N p ) are the band widths in
the p-th approximation
Np
log(2: ~i) = constant - flog N p • (3.70)
i=l

f is the scaling factor for the spectral measure. For the separate bands there
are also scaling factors. For example, for the top band of the modulated spring
model. One finds a similar behaviour for the other bands. A plot of the
logarithm of the band width against the logarithm of the length of the unit cell
shows that the (Np-th) band behaves as ~Np rv exp( -fNp log N p ). In general,
one often finds such a scaling behaviour:
(3.71)
Such scaling band structures can be analyzed by means of a multifractal anal-
ysis [49, 32]. Suppose the fraction of bands with scaling exponent between f
and df is n(f). Suppose this distribution function scales as well:
(3.72)
Narrow bands can be distinguished from broad bands in the partition function

(3.73)
Large (3's emphasize broad bands, large negative values of (3 rather small bands.
The summation has to be restricted to nonzero bands. This is indicated by I
in the summation. The partition function can be approximated as follows.

(3.74)
For p tending to infinity the maximal contribution comes from the value € of f
for which
(3.75)
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 111

100~--------------------------~

10- 1

10-2

10-3
2 4 6 8 10 12
step

Fig_ 14_ - Logarithm of the spectral measure against the order of the approximant,
for the modulated spring model with 6=1-

By a Legendre transformation one can transform

(3.76)

into
F((3) = S(€) - (3€. (3.77)
This means that the logarithm (Sp(€)) of the distribution function np(€) can
be found from

(3.78)

If the limit of Sp exists for p tending to infinity, it gives the distribution ~f

scaling exponents. In Fig.15 this has been done for the modulated spring
model, for 3 values of p[44J. Clearly the limit of S(€) exists, which implies
self-similarity and a singular continuous function. In section 5 we will come
back to the scaling exponents.

3.3. Phonons in quasicrystals

The problems for determining the eigenvibrations for a quasicrystal are in prin-
ciple the same as those for an IC phase. To understand the vibrations of a real
quasicrystal in detail one will have to consider models with interactions that
are realistic for these intermetallic alloys. We shall start, however, with simple
112 T. Janssen

0.6
-e-
0.5

0.4

S 0.3

0.2

0.1

0.0
0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8
E
Fig. 15. - Entropy function 8(10) for the modulated spring model. (6 = 1)

models, with harmonic interactions. We will argue that the essential physics of
lattice dynamics in quasicrystals is in the special structure, in other words the
topology, rather than in details of the interaction. We consider a quasiperiodic
array of atoms at positions rn
with finite range harmonic interaction. Then
the potential energy is given by

v = ~ L Q mn [If'm + um(t) - ~- un(t)1 - If'm - ~112, (3.79)

<m,n>

where the sum runs over 1st, 2nd, ... , n-th neighbour pairs. The coupling
constant Q mn depends on the separation of the neighbours.

To attack the problem of vibrations in systems without lattice periodicity

one can use several strategies. One is the cluster method. In this method one
considers a finite piece of the infinite quasicrystal model. This piece can be
cut out from the quasicrystal (free boundary conditions) or left inside it, but
with all atoms outside the chosen piece frozen (fixed boundary conditions). For
increasing sizes of the cluster one calculates the eigenmodes.

A different approach is that via approximants. The lattice of the em-

bedded quasiperiodic structure has mutually incommensurate components in
additional space VI. These irrational numbers may in a systematic way be
approximated by rational numbers. For such values the intersection of the
n-dimensional lattice with physical space becomes a three-dimensional lattice,
and the approximant structure has lattice periodicity. Consider, for example,
as model for an icosahedral phase a three-dimensional Penrose tiling. It has
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 113

rank 6, and the embedding has a 6-dimensionallattice with basis vectors

(1,7,0,1,7,0) (-1,7,0,0, -7,0), (0,1,7, -7,0, -1)
(7,0,1,0, -1,7), (7,0, -1,0, -1,7), (0,1, -7, -7,0,1).
If one replaces the irrational value 7 = (J5 + 1)/2 by approximants 2/1, 3/2,
5/3, 8/5, 13/8, ... the corresponding lattices have a nontrivial intersection with
physical space. The resulting structure has therefore lattice periodicity, with
a unit cell that goes to infinity with increasing order of the approximant. The
same lattice periodicity is obtained if one replaces only the irrational numbers
7 in the internal coordinates (the coordinates in additional space). Because
the components in physical space remain the same the resulting tiling has the
same tiles as the 3D Penrose tiling. If one would have replaced 7 everywhere
by Fn+l/ Fn the tiles of the approximants would have been deformed. If 7
is approximated by L/N the 3-dimensional structure is periodic with a cubic
unit cell. For each of the approximants one can reduce the infinite-dimensional
vibration problem to a finite one, albeit that the size of the dynamical ma-
trix increases strongly. The number of atoms in the unit cell rises with 7 3n .
However, because there is only interaction between low-order neighbours, most
of the matrix elements (recall that the ma, nf3 matrix element is the second
derivative of the potential energy with respect to the a-component of position n
and the f3-component of position m) are zero. There are special techniques that
use the sparseness of these matrices. Other related methods are the moments
methods[8, 9, 34, 76]. We shall not go into details concerning the techniques
used to diagonalize the dynamical matrices but concentrate on the results of
the calculations for two models, the tw(}-dimensional standard octagonal tiling
and the three-dimensional Penrose tiling. These have been studied by Los,
Giihler and myself.

The first model to be considered is the standard octagonal tiling model.

It consists of a quasiperiodic tiling of the plane by squares and rhombs with
an angle of 45°, with point group symmetry D8 of the diffraction patterns, and
with atoms on the vertices. Unit cells of the approximants are squares, and the
number of atoms per unit cell in the p-th approximant grows with (1+v2)2p.
The simplest approximant has 7 atoms in the unit cell. The 4-th approximant
has already 1393 atoms per unit cell, which requires diagonalization of a 2786-
dimensional matrix.

The results of a comparison of the various approximants are the following [58].

1. The density of states becomes quickly very spiky, with many (pseud(}-)-
gaps, where the density of states (nearly) vanishes.
2. The pseudogaps have a scaling property: if there is one at frequency w,
there is also one, albeit considerably less deep, at w / (1 + y'2). This would
imply that there is structure in the DOS at arbitrarily low frequency.
114 T. Janssen

o
N

o
~+-------------------------------.
-2.0 2.0
x

Fig. 16. - Approximant with N =7 of the standard octagonal tiling.

3. The pseudo-gaps disappear as soon as there is some disorder. This can

be studied by allowing a number of so called phason jumps, local vertical
displacements of the atomic surfaces. This leads eventually to a random
tiling system.
4. There is an average slope of the DOS curve at low frequencies. The
average DOS is linear at low frequencies, just as one would find for a
lattice periodic structure.
5. The dispersion curves have also a scaling behaviour. The Brillouin zone
of an approximant is a square. The dispersion can be visualized by plot-
ting the frequencies wv(k) for k along a path in the Brillouin zone. In
the figure this path is from r = [00] via X = [10] and M = [11] back
to r. If one looks at the lowest 20 branches of the dispersion curves for
different approximants, the curves are very similar up to a scale factor in
the frequency. This implies that the horizontal, almost dispersionless, op-
tical branches occur at arbitrarily low frequency in the final quasiperiodic
structure. The consequence of this fact for inelastic neutron scattering
will be discussed in the next subsection.

A second model studied, also in collaboration with Los and Gahler, is the
3-dimensional Penrose tiling obtained from a 6-dimensional periodic structure
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 115

'"
~~~~~~~~~~~
-17.0 17.0
x

Fig. 17. - Approximant 1393 for the octagonal tiling

with 3-dimensional triacontahedra in the lattice points. In the vertices of the 3-

dimensional tiling one puts atoms which are connected with harmonic springs,
up to third neighbours[54, 55, 56, 57J. The approximants here have as symmetry
group a tetrahedral space group. The number of atoms grows with powers of 7 3 .
Considered were approximants where 7 was approximated by 1/1, 2/1, 3/2, 5/3
and 8/5. The results confirm the findings for the octagonal tiling. The effects
of quasi crystallinity are, however, less pronounced in three dimensions.

1. The DOS has a lot of structure, but much less so than in the two-
dimensional case. There are no clear pseudo-gaps, except at low frequen-
cies. The DOS differs from the second approximant on strongly from that
for the first approximant, the 1/1 structure, which is an f.c.c. structure.
2. At low frequencies there is a scaling structure in the DOS, but deviations
from the classical w2 behaviour in 3 dimensions become rapidly smaller
when approaching zero frequency. Therefore, although there is some ef-
fect, it is small.
3. Disorder smears out the structure in the DOS. This is in agreement with
experiment where for poor quality quasicrystals one does not observe any
fine structure.
4. The harmonic force interactions have a stretching and a bending compo-
nent. When changing the ratio of these stregths the peaks in the DOS
shift.
116 T. Janssen

1.0 1.0
FREQUENCY
Fig. 18. - Density of states of the octagonal model.

5. The dispersion curves show again scaling behaviour. The Brillouin zone
is cubic for the approximants. If one plots the lowest branches along a
path from r = [0(0) via X = [~OO) and a corner M back to r, or via R,
M = and X, one finds almost the same curves for different approximants,
apart from a scale factor in the frequency. Also here this implies that
in the quasiperiodic limit there are optical branches at arbitrarily low
frequencies.
6. The lattice dynamics of icosahedral quasicrystals has been studied also
by Benoit et al.(66) and Hafner et a1.[31, 78). The latter use more real-
istic potentials, which describe interactions in intermetallic alloys better.
The results are, however, strikingly similar, which shows that the lattice
dynamics is for a large part determined by the topology of the structure.

3.4. Neutron scattering from quasiperiodic structures

The existence of propagating waves and of optical branches at low frequency

does not directly imply that they can be distinguished easily in experiment. To
see the effect in inelastic neutron scattering one has to calculate the dynamic
structure factor. For an ordinary, i.e. lattice periodic, crystal the differential
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 117

~~----------~----------~~--------------~
r x M r

Fig. 19. - Dispersion curves octagonal tiling for approximant 3/2.

cross section has the form

d2q = kf (271")3 "'"' "'"' 8(ii + k _ 271"K) IH- (ii)1 2 Sv(k,w), (3.80)
dndE ki v L..J ~ kv
KEA' kv

where ii = kf - ki is the scattering vector, v the volume of the unit cell, A*

I I
the reciprocal lattice of the crystal, H kv (ii) a matrix element, given below,
Sv(k,w) the thermal Bose factor, and where the sum runs over all vectors kin
the Brillouin zone and all branches v. The matrix element has in this case the
form
Hkv(ii) = t
;=1 m;
b:/2e-W3(il)eiil·r,ii.€(kV;j). (3.81)

Here the sum runs over all particles in the unit cell, labelled by j and at
positions fj, W; is the Debye-Waller factor, and €(kv;j) the eigenvector of the
eigenmode labelled by k and v.

These formulas can be generalized to quasiperiodic crystals. There is

118 T. Janssen

-,
:0-0

o
:-~
o
o
I

0.0 1.0 1.0 1.0

FREQUENCY

Fig. 20. - Density of states of the 1/1 (f.c.c.) and the 5/3 approximants to the
icosahedral model (dashed and solid lines, respectively).

now no longer a reciprocal lattice, nor a Brillouin zone. There are, however,
branches emanating from strong Bragg peaks at positions of the Fourier module
vectors, which are the projection of n-dimensional reciprocal lattice vectors.
For an modulated IC phase of rank 4 one has

UnjT rv .n j kv (q.n+T
exp (t'k--)U --) = exp ('k--)~Akv
t .n ~ mJe tmT . (3.82)
m

The static modulation is given by the modulation function f~(T) which is pe-
riodic in T. Consequently one can write

exp[iH.J;(T)] = LB lj exp(ilT). (3.83)

f.

Analogously to the lattice periodic case one gets now[21]

d2 a kf (27f)3 ~ ~ _ _ _ 1 _1 2 -
df!dE = ki -v- _~ L:-- 8(H + k - 27fG) Hkv(H) Sv(k,w). (3.84)
GEM' kv

In the case of IC phases the matrix element H kv is

(3.85)
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 119

ft
/./;;~<i'.J

f/
;"'.f""r'"'
t :l
/\ . /\.i-v
".
0.0 1.0 I.. 1.0
FREQUENCY

Fig. 21. - Density of states icosahedral model, for two choices of force constants.
The model is a 3/2 approximant. Dotted: only stretching forces, solid: stretching
force ~ and bending force i.

This implies that the inelastic neutron scattering has components coming from
the real dynamics via the factors Alj and from the static deformation (via
Btj). This will lead to more pronounced scattering along lines in reciprocal
space emanating from Bragg spots with a high intensity. The vectors of the
Fourier module are here
G = ho'* + kb* + if! + rn,q.

Also for quasicrystals is it possible to derive such an expression for the

differential cross section. The general expression is again eq. (3.84). For the
case of one atomic surface in the n-dimensional unit cell the expression for the
matrix element is
B -
H~kv = ' " ~e-W'ein.r, ' " Bj
L...J tm;; L...J Hr-Qr
H.Akv.e-
Q3
iHrT
'
(3.86)
j=1 V"·3 QEE·

where bj is the atomic scattering factor of atom species j, mj its mass, Wj

the Debye-Waller factor, fj the position of the j-th atomic surface, BK the
Fourier transform of the characteristic function of the atomic surface and AQj
the Fourier coefficients of the displacement Uj ( T) on the atomic surface.
120 T. Janssen

o
o
•

''""
"'.
'"

Fig. 22. - Dynamic structure factor for the Fibonacci chain.

By means of these expressions one can calculate the differential cross

section for coherent inelastic neutron scattering. In Fig.22 this is presented
for the one-dimensional case of a Fibonacci chain. Actually what is shown is
the dynamic structure factor without the prefactors, for temperature zero and
without Debye-Waller factor. One notices the following points.

1. For low frequency there are clear dispersion curves with small width. In
an ideal lattice periodic system the width would be zero.
2. In principle, from each point of the static Fourier module start dispersion
curves. Although the Fourier module is dense this does not mean that
the dispersion curves cann ot be distinguished, because the intensities of
branches starting from different Bragg peaks differ strongly in intensity.
It can be shown that the intensities of curves starting from different Bragg
peaks increase with increasing strength of the Bragg peak. In the static
diffraction pattern one only sees a discrete set of peaks above a certain
threshold. In the dynamic structure factor one sees the curves emanating
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 121

from this discrete set. In the figure there are strong spots near 3.0 and
5.0.

3. The spectrum has an infinite number of gaps. These can be seen in the
dynamic structure factor as well.

4. For increasing frequency the intensities of the curves decrease and their
widths increase. There are two reasons for that behaviour. As seen
above there are bands of optical branches at arbitrarily low frequency.
Their intensities go down, however, for decreasing frequency. For higher
frequencies one passes through regions of such optical bands. A second
reason is the form of the eigenvectors. One knows that the states of
the Fibonacci chain are critical. The eigenvectors of the approximants
approach these critical states. For each step in the approximation, the
structure is periodic, and therefore the states extended: the absolute
value of the displacement field is the same in each unit cell. For low
frequencies the relative number of atoms participating in the oscillation is
large. For increasing frequency the eigenvectors are much more localized
in the unit cell. Its Fourier transform has therefore contributions at small
(reciprocal lattice) vectors. This is reflected in an increasing width.

5. The width occurs in the wave vector. A width in energy would be the
consequence of finite lifetimes. This happens if anharmonic terms are
taken into account.

A similar behaviour can be observed for the dynamic structure factor of

the 3D Penrose model. For low frequencies dispersion curves are well defined.
For increasing frequency there is loss of intensity and increase of width. High
intensity along curves is found for the curves emanating from strong peaks.
Therefore, if in the experiment one goes from one strong peak to another the
intensity is high along a curve that starts at one peak at zero frequency, in-
creases till the midpoint and decreases to frequency zero at the other peak.
The form of this curve is a parabola, like in a periodic crystal. There the max-
imum is reached between two Bragg peaks at the Brillouin zone center. Here,
there is no zone center. The maximum in frequency is now reached at a point
that is at the boundary of a so-called pseudo-Brillouin zone. It is a part of
reciprocal space bounded by planes halfway between two strong Bragg peaks,
and perpendicular to the connecting line.
122 T. Janssen

4. SUBSTITUTIONAL CHAINS

4.1. Introduction

Although physical quasiperiodic and more general structures are three-

dimensional, some of the characteristic properties can be studied on simple
one-dimensional models. The advantages of such an approach are that calcula-
tions and proofs are, generally, simpler, and that there is an infinite family of
such models, as we shall see. The disadvantages are that, as already mentioned
in the section about phonons, many effects that are found in one-dimensional
models are less pronounced or even absent in higher dimensions, and that some
of the special features are due to the connectedness, which is trivial in one di-
menSlOn.
The family of one-dimensional models referred to in the preceding para-
graph is that of substitutional chains. This subject has been treated in detail in
other contributions in this volume. Here we shall only mention some relevant
facts to fix the notation.

We consider an alphabet A consisting of digits 0, 1, ... , r - 1 and the free

monoid A * based on A. Furthermore, there is a substitution u : A -. A * , which
induces a mapping from A * onto itself by
U(WIW2) = u(wt}u(w2),
where the latter means concatenation. The n-th iterate of u is un. We suppose
that the limit

exists. It is a fixed point: u(w) = w. We denote the number of a's in z E A*

by Izla. Then for each word z there is a length vector
L(z) = (lzlo, ... , Izlr_lI)T,
where T means transposition. The substitution matrix Mu is the matrix with
matrix elements
Mij = lu(j - l)li-l, i,j = 1, ... ,r. (4.87)
For two digits one has the 2 x 2 matrix

( IU(O)lo IU(l)lo)
lu(O)11 lu(l)11 .
We denote the highest eigenvalue of Mu by A. If Sn (w) is the word consisting
of the first n digits of w we define the frequencies di by
di = lim ISnli. (4.88)
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 123

Then the sum of the frequencies is unity. In general, the substitution ma-
trix does not fix the substitution, only its relative frequencies. The vector
(do, ... , dr_1)T is a right eigenvector of the substitution matrix M.

To a word w we associate a chain in the following way. We consider

intervals of length li (i = 0, 1, ... , r - 1) and positions

r-l

Xn L ISnlili
i=O
(4.89)

on the line. If
A = lim Xn (4.90)
n ..... oo n
exists it is the inverse density, and the chain can be considered as a modulated
crystal with modulation
Un = Xn -nA. (4.91)

One of the first questions one can ask is about the behaviour of this modulation.
The asymptotic behaviour has been discussed by Dumont and Thomas.

For r = 2 we introduce a normalized modulation by

(4.92)

For each finite piece of the chain one has a maximum and a minimum of the
modulation.

var(n) = M(n) - m(n). (4.93)

If A is the second eigenvalue one has the following properties for this variation
function.

1. if IAI < 1 (Pisot property) the variation var(n) is bounded;

2. if IAI > 1 the variation is unbounded;
3. if IAI = 1 the variation is bounded if and only if Vn is periodic;
4. if IAI < 1 and IDet M".I=1 the fixed point w is quasiperiodic.

Zh.Y.Wen and Janssen[77] have studied the way the variation var(n) goes to
its limit (in the case of bounded variation) or to infinity. A number of examples
is given in Figs.
124 T. Janssen

2.0 40 6.0 60 10.0 2.0 4.0 8.0 8.0 100

n n

Fig. 23. - Variation of some substitutional chains. Left: solid line 0 -+ 01, 1 -+ 0;
dashed 0 -+ 001, 1 -+ 10. Both are quasiperiodic chains. Right: solid 0 -+ 001,
1 -+ 110 (non-Pisot, loXl = 1), dashed 0 -+ 0100, 1 -+ 1011 (non-Pisot).

4.2. Atomic surfaces

A linear chain corresponding to a fixed point of a substitution can be embedded

in a higher-dimensional space. If the chain is quasiperiodic this follows from
the general theory, but also in the case of a chain that is not quasiperiodic one
can give an embedding. Consider an r-dimensionallattice spanned by the basis
vectors ei (i = 1,2, ... r). One starts in the origin and proceeds in the n-th
step according to the n-th letter in w. If it is j the n-th step is along ej. The
constructed path fluctuates around the eigenspace of A. If the substitution has
the Pisot property the fluctuation is bounded. If one considers the projection of
the path on the span VI of all eigenspaces with eigenvalue different from A this
set is restricted to a bounded region in (r -1 }-dimensional space. The closure of
this set is called the atomic surface. In the case of a quasiperiodic chain one can
attach a copy of the atomic surface to each vertex of the r-dimensional lattice
and get the quasiperiodic structure from the intersection of the periodic array
of atomic surfaces with physical space, which is the eigenspace VE belonging
to A. In ([59]) these have been analyzed by Luck et al .. A brief summary is
given here. We notice that points X on the path on the r-dimensional lattice
~ correspond one-to-one to vertices of the chain. If we denote the set of all
points of the path that correspond to left points of an interval fi by Pi, the
path is the union of all these sets: .
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 125

Each subset Pi of the path P is projected on VI. One has for the atomic surface

L. = Projection Pi on VI. (4.94)

To find the structure of the atomic surface we consider the substitution

matrix, which is an integer r-dimensional matrix, as a basis transformation of
the r-dimensional lattice. Because w is a fixed point of Ma every point X E P
goes to a point of P:
X E P ~ MaX E P.
On VI the matrix Ma acts as a contracting linear transformation. This implies
that the atomic surface is mapped into itself by Ma.

Consider as an example the Fibonacci chain. It is produced from the

substitution
0-(0) = 01, 0-(1) = 0
with substitution matrix
Ma = (~ ~),
which has A = T and>' = 1 - T. If X E Po one has either X E MaPo or
X E Ma P1 . On the other hand, if X E PI then it is not included in MaP, but
X E MaP + eo. The reason is that 0 appears at the left side in both 0-(0) and
0-(1), but 1 occurs only in 0-(0), but is then separated from the left side by a
"0". Therefore, one has

Lo >.Lo U >'L1
L1 >.Lo + elO,
where eIO is the component of eo in VI. In other words the vector (Lo, L 1) is a
fixed point for the contracting nonlinear mapping

LOI >.Lo U >'L 1

L1' >.Lo + elO,
If one starts with two arbitary intervals on the line VI repeated application of
this mapping yields the subsets L i , and therefore also the atomic surface L.
In the case of the Fibonacci substitution one obtains in the limit two intervals
with a ratio of T, and with total length the projection of the unit cell of ~ on VI.
It should be noted that the decomposition of the atomic surface corresponds
to a labelling system for the points. Each point has a nonnegative number
of pre-images. This number can be used as a label. In general, one needs an
additional label if more than one point does not have a pre-image, but for the
present Fibonacci chain there is no need for that. It is connected with the fact
that every positive number can be written in a unique way as the sum of a
number of decreasing Fibonacci numbers (28 = 21+5+2, etcetera).
126 T. Janssen

---------',-
____ •. ___ ._____ I
--------r--------------
,
________
,
------- - - - - - ________ L ___________ _

, - - - -"T- - - - ----

-- -- -
,, _ _ _ .L. _ _ _ _

, - '"""T"'---

- --'-- ,,
-.,-

-1.0 -0.5 0.0 0.5 1.0

X

Fig. 24. - Orbit of intervals under the nonlinear map for the Fibonacci substitution.
Starting from 2 arbitrary intervals LA and LB the orbit has in the n-th step Fn
intervals. Fn-l of them form LA and Fn- 2 form LB. To avoid overlap the B-atomic
surface is shifted slightly upward. X gives the position in internal space.

The same atomic surface is found, of course, for the square 0'2 of the
substition. The fixed point for 0'2(0) = 010, 0'2(1) = 01 is the same as that
for o'. However, a substitution 0 -+ 001 and 1 -+ 10 which has the same
substitution matrix gives a completely different atomic surface. This can be
seen from the variation var(n). For the squared Fibonacci substitution this
variation tends to the limit 1 (in appropriate units), but the other chain gives
in the limit a value of T. Analysis shows that the former atomic surface is
a connected interval of unit length, whereas the second one is a fractal of
Hausdorff dimension 0.9157 ... and also of measure one. The last point follows
from the fact that the average densities of the two chains are the same.
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 127

4.3. Fractal atomic surfaces

The foregoing shows that not only the substitution matrix Mu is important.
This matrix determines the bounded or unbounded character of the variation,
and it fixes the Fourier module. The structure of the atomic surface, and
therefore the Fourier transform, depend on the specific substitution. The square
of the Fibonacci substitution has matrix

(4.95)

With such a substitution matrix correspond 6 substitutions:

0"(0) 010 010 001 100 001 100

0"(1) 01 10 01 10 10 01

Substitutions 1 and 2, 3 and 4, and 5 and 6 are mirror images of each

other. The resulting chains can therefore be identified. Moreover, the substitu-
tions 1 and 4 are also simply related. Therefore, the first 4 substitutions behave
in the same way (they have a connected atomic surface oflength 1), and the last
two behave in the same way (with a fractal atomic surface of extension r). In
general, one can partition binary (r = 2) substitution chains into classes each
with a characteristic atomic surface in the one-dimensional VI. Generically
these surfaces are fractals. As seen in the section on diffraction, the structure
factor is closely related to the Fourier transform of the characteristic func-
tion on the atomic surface(s}. For fractals this Fourier transform decays much
slower. As a result the structure factor of a chain with fractal atomic surface
has many more peaks above the threshold than that for a smooth surface.

A similar situation occurs for r > 2. As an example we consider two

ternary substitutions[59]. The matrix

is the substitution matrix for 24 substitutions, the permutations of

0"(0) = 01, 0"(1) = 02, 0"(2) = 012.

It has 3 real eigenvalues, A = Al = 2.247 .. , A2 = -0.802 .. and A3 = 0.555 ...

Because all except A are smaller than unity, the substitutions have the Pisot
property and the atomic surfaces are bounded. Because Det(M}=-l the corre-
sponding chains are quasiperiodic. A basis for the 3-dimensional space is given
128 T. Janssen

by the 3 eigenvectors ~ of M. The reciprocal basis consists of the vectors

e:.Then the projection of a point r of the path in 3 dimensions on the space
spanned by e2 and e3 is
(f'.~)e2 + (f'.e;)e3'
A number of examples of the atomic surfaces is given in Fig.25. There are 11
classes of substitutions. The action of M on the atomic surface is given by the
eigenvalues. It is an anisotropic contraction.

A second example is the matrix

oo 1)
0 ,
1 1

which has two equivalent substitutions. The substitutions have the Pisot prop-
erty, because the eigenvalues are

A = Al = 1.466 ... , A2 A*3 -0.233 ... ± 0.793 ... i.

The action of M on the atomic surface L is a contraction with factor

0.826 and a rotation over 0.5901T. Just as for the Fibonacci chain one can
construct the corresponding nonlinear contracting map. Because a 0 appears
as first letter in 0'(2), 1 as first letter in 0'(0) and 2 as first letter in 0'(1) or
after a 0 in 0'(2) one has

Lo = T~, Ll = TLo, L2 = TLI U (T~ + elO), (4.96)

where T is a the scale-rotation in VI. If one starts with the projection of the
unit cell as l, partioned in 3 Li's for which we take the projection of three
faces of the unit cell, a number of steps in the iterative procedure are shown in
Fig.27. One sees that the interior of the atomic surface is filled, but that the
border looks like a fractal. That it is really a fractal is seen from the following
reasoning. The nonlinear map on the Li's induces the following map of the
borders. If the edges corresponding to the three basis vectors ei are denoted
by a,{3,'Y one has
a -+ 'Y - {3, {3 -+ a, 'Y -+ {3. (4.97)
This substitution has matrix

N (: H)
with eigenvalues 1.32.. and -0.66 .. ±0.56 ... i. This matrix is obtained from
M-l by replacing all matrix elements by their absolute values. Because the
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 129

0120120 01 02201
2 2
1.5 1.5
1 1
0.5 0.5
o o
-0.5 -0.5
-1 -1
-1.5 -1.5
-1.5-1-0.500.5 1 1.5 2 -1.5-1-0.500.511.52

01 20021
2
1.5
1
0.5
o
-0.5
-1
-1.5
-1 .5-1 -0.5 0 0.5 1 1.5 2

Fig. 25. - Atomic surfaces for ternary substitutional chains with a substitution ma-
trix with real eigenvalues. Above the figure are given the substitutions 0"(0), 0"(1),
0"(2). '

number of edges increases with J.L"i and the total length with P'21 n the border
has a Hausdorff dimension

dim 1.47 ...

130 T. Janssen

0. 8r---~--~---.---r---'----r---~--~---r--~
" fo rt.19 " -
311132
0.6

0. '

0.2

o ..... .

- 0.2

- 0.4

-0.6

- 0 . 8 '-----'--__...L..__ _~_ _- - ' -_ _____'_ _ _ _' _ __ __'___ _...L..__ _.....&...._ ____l

-\ - 0 .8 - 0.6 - 0. ' -0.2 0.2 0 .4 0.6 0.8

Fig. 26. - Atomic surface for ternary chain with complex eigenvalues. Also the pro-
jection of the unit cell of the lattice is shown.

As seen in the previous case, in general the atomic surface itself may also have
a fractal structure.

Not only in chains can one find fractal atomic surfaces. The problem of
finding self-similar tilings is much more difficult in two and three dimensions. If
one wants to construct a tiling by means of inflation rules, one needs the set of
tiles and a recipe to decompose each tile in copies of smaller versions of the tiles
(smaller by a factor A). For a one-dimensional tiling this suffices, but in higher
dimensions one has to check that the procedure of decomposition followed by
a dilatation by A can be iterated infinitely many times. This topological con-
straint restricts the possibilities strongly. Therefore, there are infinitely many
one-dimensional tiles, but only a small number of known higher-dimensional
ones. If one has found such a quasiperiodic tiling, one can lift the vertices of
the tiling to points in n-dimensional space. The scaling transformation cor-
responds to a basis transformation of the n-dimensional lattice. This basis
transformation has eigenvalue A and a two- or three-dimensional correspond-
ing eigenspace, to be identified with physical spaces. All other eigenvalues are
smaller than unity in absolute value, and the span of all the corresponding
eigenspaces is the internal space. The points of the n-dimensional lattice that
project on vertices of the tiling are projected on the internal space. The closure
of this set is the atomic surface. Some of the well studied tilings have simple
atomic surfaces, a regular octagon for the standard octagonal tiling, a rhombic
triacontahedron for the 3D Penrose tiling. The 2D Penrose tiling needs 4 atomi~
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 131

Fig. 27. - Steps of the nonlinear mapping for the substitution 0, 1, 2 -+ 1,2,02

surfaces, in the form of regular pentagons. One has found also atomic surfaces
that are fractal with octagonal[69, 29, 19], pentagonal[87] and dodecagonal[28]
symmetry. Although these systems are quasiperiodic, they have a richer, or
more complicated, structure than the tilings one usually studies. Occurrence
of real crystals with fractal atomic surfaces is possible, but has not (yet) been
found.
132 T. Janssen

5. ELECTRONS

5.1. Models

To show that quasiperiodic systems and their generalizations can have physical
properties that differ from those of lattice periodic systems, something that we
have already discussed in the section on phonons, we shall look at the problem of
electrons in such systems. To keep the argument as simple as possible, we shall
consider only one-dimensional chains. That means we consider the problem of
Schrodinger operators for substitutional chains. This subject is treated in more
generality by Siito in this volume. Here we are interested in spectra and wave
functions especially when comparing different types of chains.

There exist several simple one-dimensional models for electrons. The

model that has received the most attention in theoretical studies is the tight-
binding model. This is a linear chain of atoms with site potentials with one
energy level, and with overlap between the wave functions which allows hopping
between neighbouring sites. The Schrodinger equation can be written as
(5.98)

where tn is the hopping frequency between sites nand n + 1, En is the energy

level at site n, and {en} is the wave function, defined on the sites n. E is the
eigenvalue of the Hamiltonian
(5.99)
The hopping terms and energy levels form a sequence of numbers which cor-
responds to the fixed point of a substitution. The spectrum of H in that case
has been discussed for several types of chain in the chapter by Siito.

Another model is the modulated Kronig-Penney model. It has the Hamil-

ton operator
(5.100)
n
Here p is the momentum operator, {xn} the positions in a chain, and an the
strength of the delta potential at site X n . As an example, electrons in a mod-
ulated Ie phase have been studied on the modulated KP model[52, 53] with
an = 1 and
Xn = n+6cos(nq+¢). (5.101)

There are several methods of studying the spectra and wave functions of
the Schrodinger operators given above. One is similar to that used for phonons
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 133

211

11

T
o 211
q

Fig. 28. - Spectra for q = LIN with 0::; N::; 50 in the Kronig-Penney model

in quasiperiodic systems. If a substitutional chain is quasiperiodic, its basis

vectors of the higher-dimensional lattice can be approximated by nearby vec-
tors such that the deformed lattice has an intersection with physical space.
For a substitutional chain the physical space is the eigenspace of the highest
eigenvalue. Suppose the components of the eigenvector are ai, then the num-
bers ai/a) are irrational. If one approximates these by rational numbers the
corresponding line goes through a lattice point and the structure is periodic.
134 T. Janssen

For example, for the Fibonacci chain the eigenvector is (1, l/T), which can be
approximated by (Fn' F n - 1). For the periodic chain one may diagonalize the

[~;':e~(ik) 1
finite matrix
...
... re~(-ik)
tl 0
tt E2 t2
H = " t2 E3 ... (5.102)
.. . . ..
0 0 ... EN
if the chain has periodicity N (tn+N = tn, En+N = En). This matrix can be
diagonalized using its sparseness.

A second method is the one using a trace map [50]. One writes the
Schrodinger equation in the form

( c",c",+1) =
(~-
1 tn 0
tn-l
tn
) (c",)
c",-l = Tn
(c",)
c",-l . (5.103)

The spectrum consists of those values of E for which the trace of TIn Tn tends
to a value between -2 and +2. The traces can more rapidly be calculated if
the chain has self-similarity. Here one has to distinguish between the diagonal
model (t n = 1) and the off-diagonal model (En constant). For the diagonal
model the procedure is exemplified on the Fibonacci chain. Because of the
substitution character one may write

T(n) = IT Tm
Fn
= T(n-2)T(n-l). (5.104)
m=l

Because the determinant of the matrix Tm is +1, one may use the relation
A2 B = Tr(A)AB - B (5.105)
to derive the following relation for the traces xn=Tr(rn):
Xn+l = XnXn-l - Xn-2· (5.106)
This so-called trace map is a nonlinear mapping in 3 variables. It gives in
variables x, y, z the mapping
x, y, z -+ y, z, yz - x. (5.107)

The chain starts with abaab . .., which implies that

T(l)
= ( ~-El o
-1 ) = A,

T(2) AB with B = ( ~- E2
-1
0 ),
T(3) ABA.
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 135

One can take ro = B. Hence

Xo =E - €2, Xl =E - €l, X2 = (E - (1)(E - (2) - 2,
from which follows the orbit. An analysis of the orbits under this mapping
yields in principle the spectrum. Especially fixed points of the non-linear trace
map are of importance. For example, the Fibonacci trace map has a 6-cycle
starting from (a, 0, 0), i.e. a fixed point for its 6-th iterate. Analysis of the
linearized map around this fixed point gives information about the scaling
properties of the bands. One can use the trace map also to scan the spectrum
rapidly.

5.2. Spectra

The spectra of the Schrodinger operators for substitutional chains are calcu-
lated as limits of the spectra for its approximants. Suppose we consider a
systematic series of approximants, labelled by n. One first information is the
way the spectral measure behaves when n tends to infinity. Define
Nn
Bn = L~~n), (5.108)
i=l

as the total spectral measure, i.e. the sum of all widths of the bands (their
number is N n ) in the n-th approximant. In Fig. 29 is plotted the logarithm of
Bn against n for two cases. One is the Fibonacci diagonal tight-binding model,
the second one is the analogous curve for the modified square of the Fibonacci
chain (0"(0) = 001,0"(1) = 10). In both cases the spectral measure tends to
zero. However, whereas both go exponentially to zero(Bn = exp( -na)) , the
spectrum for the second case does this much more rapidly.

A similar calculation can be performed for a specific band. If one does

this for the two substitutions and the top band, one sees that for the Fibonacci
chain the width of this band scales

(5.109)

whereas the corresponding band for the other case goes faster to zero than
exponentially. This can be taken as a sign for an eigenvalue, i.e. that there
is a point spectrum. One may analyze this by looking at the behaviour of the
function
€~/kn _~IOg~~n). (5.110)

For the Fibonacci chain this converges to a function €~oo) on the interval [0,1].
For the second case there is only convergence in certain points. It should
136 T. Janssen

,,
,,
,
~ ,,
,,
,
A,
,,
,,
,
3.0 6.0
N

Fig. 29. - Total spectral measure for electrons in two different substitutional chains.
Solid lines: substitution a -+ 00, b -+ a, dashed lines: a -+ aab, b -+ 00; full circles:
site energies I:i = ±1, open circles: site energies ±O.5.

be stressed that these are numerical results. However, it indicates that the
behaviour and the scaling properties of electronic spectra for chains with the
same substitution matrix may be very different[62].

A behaviour that differs from that of the quasiperiodic Fibonacci chain

as well has been found in other substitutional chains.

5.3. Wave functions

A simple characterisation of a wave function is its localization, or inverse par-

ticipation ratio. For a wave function {en} this is defined as

(5.111)
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 137

1.9
1.8
1.7
1.6
1.5
1.4
1.3
1.2
1.1
1
0.9
o 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 30. - The function €., on [0,1] for the substitutional chain a - ab,b - a in a
mnnber of approximants.

For an approximant with N sites in the unit cell this becomes liN for the case
that en is constant, and I for the case that the wave function is only nonzero
at one site. Here also there is a .big difference between the two cases with
different atomic surfaces. This can also be seen if one plots the values ICiI or
the logarithm of this quantity against ilNn . For a Fibonacci chain the wave
functions are critical, which implies self-similarity. For the modified squared
Fibonacci chain, with a fractal atomic surface, there are such critical states also,
for bands for which a scaling exponent exists. However, when one considers
wave functions for bands going faster to zero than exponentially, the character
is different. If one compares the wave functions for successive approximants,
one sees that the wave function is localized near a fixed point on the scale i INn
on the interval [0,1). Its width decreases on this interval with liNn. This means
that the peak keeps in real space its width, but that its position tends to infinity
with respect to the beginning of the chain. In the plot of the logarithm log ICiI
against ilNn one sees an exponential falloff, and a self-similar structure. The
wave function can be called pseudo-localized. This means that on the energy
scale pseudo-localized states are mixed with a dense set of critical states. A
more detailed analysis has been given by Paasschens [62).
138 T. Janssen

2.5

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1
Fig. 31. - The function c:'" for approximants to the chain a --+ aab, b --+ 00.

In higher-dimensional spaces the behaviour of electrons is not yet known.

It will be interesting to see whether critical and pseudo-localized states exist,
and if so, what the character is of the mobility edge, the boundaries on the
energy scale between localized and non-localized states. In lattice periodic
crystals with disorder there are mobility edges near the shoulders of the bands
in the DOS. Anyway, in one dimension, there is already a marked difference
between quasiperiodic crystals with a smooth and with a fractal atomic surface.
The same seems to hold for other substitutional chains, where the difference
cannot be characterized by an atomic surface.

References

[IJ Aalst W. van, Hollander J. den, Peterse W.J.A.M., Wolff P.M. de, Acta
Cryst. B32, 47 (1976).
[2] Ascher E., Janner A. , Helv. Phys. Acta 38, 551 (1965).
[3] Ascher E., Janner A. Commun. Math. Phys. 11, 138 (1968).
[4] Aubry S., Le Daeron P.Y. Physica D 8, 381 (1983)
[5] Axe J.D., lizumi M., Shirane G. Phys. Rev. B22, 3408 (1980).
 FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 139

[6] Axel F., Terauchi H., Phys. Rev. Lett. 66, 208 (1991).
[7] Bak P., Boehm J. von, Phys. Rev. B 21, 5297 (1980).
[8] Benoit C., J.Phys. C 20, 765 (1987).
[9] Benoit C., Poussigue G., Azougarh A. J.Phys. C.M. 2, 2519 (1990).
[10] Bernard L., Currat R., Delamoye P., Zeyen C.M.E., Hubert S.,
Kouchkovsky R. de, J.Phys. C 16, 433 (1983).
[11] Bieberbach L., Math. Ann. 70, 297 and 401 (1911).
[12] Bienenstock A., Ewald P.P., Phys. Rev. 6, 720 (1961).
[13] Bohr H., Acta Math. 45, 29 (1924), 46, 101 (1925), 52, 127 (1928).
[14] Brouns E., Visser J.W., Wolff P.M.de Acta Cryst. B 25, 2665 (1969).
[15] Brown H., Math. of Compo 23, 499 (1969).
[16] Brown H., Biilow H., Neubuser J., Wondratschek H., Zassenhaus H.,
"Crystallographic Groups of Four-Dimensional Space", J. Wiley & Sons
(New York, 1978).
[17] Cailleau H., Moussa F., Mons J., Sol. St. Comm., 31, 521 (1979).
[18] Cailleau H., Moussa F., Zeyen C.M.E., Bouillot J., Sol. Stat. Comm. 33,
407 (1980).
[19] Cockayne E., Phys. Rev. B49, 5896(1994).
[20] Comes R., Shirane G., Shapiro S.M., Garito A.F., Heeger A.J., Phys.
Rev. B14, 2376 (1976).
[21] Currat R., Janssen T. in " Solid State Physics 41", (Ehrenreich H., Turn-
bull D. eds.), Academic Press (New York, 1988), 20I.
[22] Durand D., Denoyer F., Currat R., Lambert M., "Incommensurate
Phases in Dielectrics 2" (eds. R. Blinc, A.P. Levanyuk), Elsevier (Ams-
terdam, 1986), lOI.
[23] Elliott R.J., Phys. Rev. 124, 346 (1961).
[24] Fast G., Janssen T., J. Compo Phys. 7, 1 (1971).
[25] Frenkel Y.!., Kontorova T., Zh. Eksp. Teor. Fiz. 8, 1340 (1938).
[26] Friedel J., Denoyer F., C.R. Acad. Sci. Paris Ser. II 305, 171 (1987).
[27] Futama H., Shiozaki Y., Chiba A., Tanaka E., Mitsui T., Furuichi J.,
Phys. Lett. 25A, 8 (1967).
[28] Ga.hler F., Dissertation ETH Zurich (1988).
[29] Godreche C., Luck J.M., Janner A., Janssen T., J.de Physique 3, 1921
(1993).
[30] Goursat E., Ann. Sci. Ecole Norm. Super. Paris 6, 9 (1889).
[31] Hafner J., Krajci M., Europhys. Lett. 21, 31 (1993).
[32] Halsey T.C., Jensen M.H., Kadanoff L.P., Procaccia I., Shraiman B.L,
Phys. Rev. A33, 1141 (1986).
[33] Hastings J.M., Pouget J.P., Shirane G., Heeger A.J., Miro N.D., McDi-
armid A.G., Phys. Rev. Lett. 39, 1484 (1977).
[34] Haydock R., Heine V., Bullet D., Kelly M.J., in " Solid State Physics 35",
Academic Press (New York 1980).
[35] Heesch H., Z. Kristallogr. 73, 325 (1930).
[36] Heilmann LU. Axe J.D., Hastings J.M., Shirane G., Heeger A..J., Mac-
Diarmid A.G., Phys. Rev. B 20, 751 (1979).
140 T. Janssen

[37) Hendricks S., Teller E., J. Chern. Phys. 10, 147 (1942).
[38) Hermann C., Acta Cryst. 2, 139 (1949).
[39) lizumi M., Axe J.D., Shirane G., Phys. Rev. B 15, 4392 (1977).
[40) Janner A., Janssen T. Physica 53, 1 (1971).
[41) Janssen T., Ferroelectrics 66, 203 (1986).
[42) Janssen T., Physics Reports 168, 57 (1988).
[43) Janssen T., Janner A., Physica A 126, 163 (1984).
[44) Janssen T., Kohmoto M., Phys.Rev. B 38, 5811 (1988).
[45) Janssen T., Los J., Phase Trans. 33 , 29 (1991).
[46) Janssen T., Tjon J.A., Phys. Rev. B 24, 2245 (1981).
[47) Janssen T., Tjon J.A., Phys. Rev. B 25, 3767 (1982).
[48) Janssen T., Tjon J.A., J. Phys. C 16,4789 (1983).
[49) Kohmoto M., Phys. Rev. A37, 1345 (1988).
[50) Kohmoto M., Kadanoff L.P., Tang Chao, Phys. Rev. Lett. 50, 1870
(1983).
[51) Lange C. de, Janssen T., J. Phys. C 14, 5269 (1981).
[52) Lange C. de, Janssen T., Phys. Rev. B 28, 195 (1983).
[53) Lange C. de, Janssen T., Physica A 127, 125 (1984).
[54) Los J., Janssen T., J.Phys. Cond.Matter 2, 9553 (1990).
[55) Los J., Janssen T., Gahler F., J. Non-Cryst. SoL, 153/4, 581 (1993).
[56) Los J., Janssen T., Gabler F., J.Physique France 3, 107 (1993).
[57) Los J., Janssen T., Gahler F., J. de Physique 3, 1431 (1993).
[58) Los J., Janssen T., Gahler F., Int.J.Mod.Phys. B 7, 1505 (1993).
[59] Luck J.M., Godreche C., Janner A., Janssen T., J.Phys.A 26, 1951
(1993).
[60] Mermin N.D., Rev. Mod. Phys. 64, 3 (1992).
[61] Okamura K., Iwasaki H., Ogawa S., J. Phys. Soc. Japan 24, 569 (1968).
[62] Paasschens J., Janssen T., preprint (1994).
[63] Pater C.J. de, Axe J.D., Currat R., Phys. Rev. B 19, 4684 (1979).
[64] Plesken W., Pohst M., Math.Comp. 31, 536 and 552 {1977}.
[65] Pouget J.P., Shirane G., Hastings J.M., Heeger A.J., Miro N.D., McDi-
armid A.G., Phys. Rev. B 18, 3645 (1978).
[66] Poussigue G., Benoit C., de Boissieu M., R. Currat, J.Phys. CM 6, 659
(1994).
[67] Rabson D.A., Ho Tin-Lun, N.D.Mermin, Acta Cryst. A 45, 538 (1989).
[68] Rabson D.A., Mermin N.D., Rokhsar D.S., Wright D.C., Rev. Mod. Phys.
63, 699 (1991).
[69] Sasisekharan V., Baranidharan S., Balagurusamy V.S.K., Srinivasan A.,
and Gopal E.S.R., Pramana 33, 405 (1989).
[70] Selke W., in "Phase Transitions and Critical Phenomena Vo1.15",
C.Domb and J.L. Lebowitz, eds., Academic Press, (New York, 1992),
1.
[71] Shechtman D., Blech I., Gratias D., Cahn J.W., Phys. Rev. Letters 53,
1951 {1984}.
[72] Stephens P.W., Goldman A.I., Phys. Rev. Lett. 56, 1168 (1986).
 COURSE 6

Matching Rules and Quasiperiodicity:

the Octagonal Tilings
A. Katz

Centre de Physique Theorique, Ecole Polytechnique

91128 Palaiseau Cedex
France

1. INTRODUCTION

This lecture discusses one of the most important question raised by the discov-
ery of quasicrystals: the onset of quasiperiodic order. In fact, one of the main
problems about quasicrystals is to understand the simple possibility of a non
periodic long range order, since no two atoms have exactly the same environ-
ment up to infinity. One possible solution to this problem is to consider that
the order stems from privileged local configurations and is able to propagate
throughout the structure. This point of view deals with the existence of local
constraints which would enforce the quasiperiodic order: these are the so-called
"local rules" , or "matching rules" in tiling language.
Although the atomic structure of quasicrystals is not linked with quasiperi-
odic tilings in the same strong way that crystals are linked with periodic tilings,
it makes sense to approach the problem of ordering of quasicrystals through the
simpler and schematic theory of matching rules for tilings. The reason is that
experiments suggest that real quasicrystals share a main geometrical property
with the kind of "canonical tilings" which will be studied here: within experi-
mental resolution. it seems that there is only a finite number of different atomic
environments (up to a given distance) for atoms in the quasicrystal. This "rigid-
ity" feature, which is described by the "flatness" of the atomic surfaces to be
defined below, obviously occurs also for tilings, so that the notion of matching
142 A. Katz

rules for tilings, which prescribes how the tiles should be joined together, is
not so far from the notion of local rules (L. Levitov [1]) which prescribes for a
quasicrystal what are the allowed local neighbourhoods of atoms.
In other words, we shall consider here quasiperiodic tilings as abstract pat-
terns carrying the long range order typical of the atomic structures of qua-
sicrystals.
Thus, from now on, we shall deal only with tilings. For the sake of simplicity,
we shall develop our arguments for two-dimensional tilings, and even more
specifically in the case of the octagonal and related tilings. But, as will be
made clear, the approaches explained in this lecture are by no means limited
to the two-dimensional case.
Let us now define more precisely this notion of matching rules. For tilings,
they consist in decorations (typically, arrowing of the edges of the tiles for
a two-dimensional tiling) together with a "recipe" which prescribes how the
decorations of adjacent tiles should fit with each other [1][2][3][4]. We say that
a tiling admits matching rules when such a set of local constraints enforces the
quasi periodicity of the tiling (a global property) as soon as they are satisfied
everywhere in the tiling of the whole space.
The very existence of such rules is by no means obvious. Matching rules
for the original Penrose tilings were derived in several forms by Roger Penrose
using the self-similarity (inflation and deflation) of his tilings. For the sake
of completeness and reference purpose, we shall give a short account of this
approach, referring for instance to the paper [3] by F. Gahler for more de-
tails about what he called the composition-decomposition method (the author
thanks F. Gahler for his help in the redaction of this section). The remainder
of the paper is devoted to a quite different approach initiated by the author in
[2] and developed subsequently in collaboration with L. Levitov.

2. QUASIPERIODIC TILINGS

We will focus our attention on the best known class of quasiperiodic systems,
which are the canonical or "Penrose-like" tilings. They are named after Roger
Penrose, who discovered a strikingly simple non-periodic tiling of the plane
with five-fold symmetry [5] later studied in detail by N. G. de Bruijn [6]. Al-
though the original construction relied mainly on self-similarity properties (see
section 3.1) and the quasiperiodicity of the tilings was recognised only later by
crystallographers and solid state physicists, these tilings are in fact the simplest
non-trivial quasiperiodic sets of points that one can imagine.

2.1. Quasiperiodicity

Let us first recall briefly the definition of quasi periodicity, as developed by H.

Bohr [7][8][9] and A. S. Besicovic [10]: a function (of d real variables) on an
affine (d-dimensional) space Ec is said to be quasiperiodic if it is the restriction
 MATCHING RULES 143

Fig. 1. - Atomic surfaces in the 2 dimensional case; the cut space Ec is the horizontal
line; the atomic surfaces 0" are line segments transversal to Ec and attached at the
lattice nodes of A.

to Ec (embedded as an affine subspace) of a periodic function of n real variables

defined in a higher-dimensional space IRn. Of course, if the embedded space
Ec (which will be referred to as the "cut") is rationally oriented with respect to
the lattice of periods of the periodic function (i. e., if Ec is parallel to a lattice
subspace) then the restriction of this function to Ec is also periodic. But if the
direction of the cut is irrational, that is, if the vector subspace parallel to the
cut contains no point of the lattice besides the origin, then the restriction is
not a periodic function, but a strictly quasiperiodic one.

2.2. The Atomic Surfaces

To describe a quasiperiodic set of points rather than a quasiperiodic function,

the natural generalisation is to attach to each of these points a Dirac delta and
to consider them as the d-dimensional restriction p, of a n-dimensional periodic
measure, say if, instead of the restriction of a periodic (smooth) function. Let
A be the lattice of periods of this periodic measure. The carrier of if defines
a A-periodic geometric locus (j which decomposes, in the elementary domains,
into identical pieces 0". The manifold 0" is called the "atomic surface" and (j is
the periodic set of atomic surfaces as shown on Fig. l.
The intersections between the cut and the atomic surfaces, and therefore
the resulting structure, are well defined each time the cut is transversal to
the atomic surfaces. The transversality condition means that the cut does not
intersect any atomic surface on its boundary or on a point where the tangent
space to the atomic surface is parallel to the cut. This is not a strong restriction;
it simply means, in a first approach, that the atomic surface boundaries are not
a too complicated set, such as for example, a fractal set. Observe that since
the cut Ec is of dimension d and the boundaries of the atomic surfaces are
(piece-wise smooth) sub manifolds of dimension (n - d - 1) or less, their non-
144 A. Katz

intersection is a generic property in mn , which means that almost all choices

of Ec will work.

2.3. The cut algorithm

To generate the quasiperiodic set of points, we simply collect the intersection

points ofu with Ec. We designate by Ell the vector subspace ofmn defining the
direction of the cut Ec and by E..L the perpendicular subspace. This algorithm
is nothing but a slight generalisation (first advocated by P. Bak [U]) of the
former "hyperspace" formalism developed by Janner and Janssen [12] from the
pioneer work of de Wolff [13] for modulated and incommensurate structures.
For a given generic direction of the cut, the technique requires an additional
set of parameters which specifies the relative location of the cut space Ec with
respect to A. Indeed, one generates infinitely many "different" quasiperiodic
set of points by shifting the cut space: observe that for a generic atomic surface,
two parallel cuts yield isometric structures if and only if they are mapped on
each- other by a translation belonging to the lattice, up to a translati.on parallel
to the cut. Such a translation is a vector in mn which projects on the subspace
orthogonal to the direction of the cut, on a vector belonging to the projection
of the lattice. Thus we see that the different structures are classified by the
quotient of this subspace by the projection of the lattice. Since this last set is
countable, we see that by shifting the cut we generate an uncountable infinity
of different (non-isometric) structures.

2.4. Canonical or "Penrose like" tilings

2.4.1. Definition
Let us start with the general definition of what we call tHings of the Penrose
type [14] or canonical tHings: they are obtained by the cut method with one
atomic surface u per unit cell defined by the projection on E..L along Ell of
the unit cell "tn of A. To construct a d-dimensional canonical tiling, consider
in mn the simple cubic lattice Zn generated by the canonical orthonormal
basis of mn , which spans the unit cube "tn. Then choose any d-dimensional
subspace Ell in mn , and denote by E..L the orthogonal subspace. We define the
atomic surface u by projecting orthogonally "tn on E..L. This yields a (n - d)-
u
dimensional polyhedron, and the corresponding lattice of atomic surfaces in
mn is obtained by copying this polyhedron at each vertex of 7l. n.
e
u
The vertices of our tHings are the intersections of with any d-dimensional
plane cut Ec parallel to Ell, and which is everywhere transversal to U, i.e.,
which does not intersect any of the boundaries 8u~ for E Zn.e
The simplest example of this construction is obtained with n = 2 and d = 1,
and is depicted on Fig. 1. For an irrationally oriented Ell, we get a quasiperiodic
tiling of the cut Ec by means of two segments, which are the projections of the
two edges of the unit square "t2. Although this construction may look rather
trivial, it deserves attention because the most important features of this class of
 MATCHING RULES 145

tilings already appear in this simple case and may be discussed in a dimension-
independent way.

2.4.2. The oblique tiling

To prove that we actually get a tiling by means of the projections of the edges
of the square, the best way is to construct the so-called oblique tiling [15] or
Kljjtze decomposition [16].
The idea is the following: consider any tiling offfin and any plane cut through
this tiling. Each time the cut is generic, that is, intersects transversally the
boundaries of the tiles, the traces of the tiles on the cut make up a covering
of the cut, without overlappings or holes. But this covering is not a tiling in
general, since there is no reason for the traces of the tiles to belong to a finite
set of shapes. For instance, consider a cut Ec with an irrational slope through
the standard square tiling of the plane: since there is no minimal distance
between the vertices of the tiling and the cut, there is no minimal length for
the segments of the induced covering of the cut, and this entails that there are
infinitely many different lengths in this covering, which therefore is not a tiling.
However, it is possible to adapt the shape of the tiles of a periodic tiling of
ffin to the direction of the cut, in order to obtain only a finite number of shapes
in the generic cuts: the trick is to make the boundaries of the tiles parallel to
either the direction of the cut Ell or to the orthogonal subspace EJ.. For our
low dimensional case, the construction of this oblique tiling is the following:

Fig. 2. - Construction of the oblique tiling for the linear quasiperiodic chain. The
"oblique tiles" are the sum of a (horizontal) tile of the quasiperiodic chain and of its
so-called "existence domain" (here a vertical segment)
146 A. Katz

let us start with the unit square, spanned by the canonical basis (c1' c2) of rn?
Defining 7I"1I(resp. 71"1.) as the orthogonal projection on EII(resp. E1.), we set
ei = 7I"1I(ci) and e~ = 7I"1.(ci) (i = 1,2), in such a way that Ci = ei + e~. Let
us distort the edge C1 along the broken line defined by the segment e~ follo~d
by the segment e1. We see that the union {e~, c2} U (e~ + {e1' c2}) of the two
parallelograms spanned by {el, c2} and {e~, c2} is still a unit cell for the lattice
71. 2. Now, let us proceed to the same decomposition for the vector C2 and each
parallelogram: we get a new unit cell of 71. 2 made of four subcells spanned by
{ei, e2}, {ei, ~}, {e1' e2} and {eI, e~}. But the two subcells spanned by {ei,
~} and {e1' e2} are flat and we can omit them, so that we obtain finally only
two subcells whose union is a fundamental domain of 71. 2. The corresponding
tiling of the plane (gray area on Fig. 2) is the oblique tiling.
Observe that whatever the order of the decomposition, the resulting tiling is
the same. Since each tile is the sum of the projection of a basis vector on Ell
and of the projection of the other on E1., it is clear that any cut Ec parallel
to Ell which does not intersect the lattice inherits a tiling by means of the
two projections e1 and e2, which is our quasiperiodic tiling: in fact, the pieces
of the boundaries of the tiles of the oblique tiling which are parallel to E1.
(which are called the existence domain of the corresponding tile) have by their
very construction an union identical to the lattice (j of atomic surfaces (which
appear as the existence domains of vertices).
Due to its recursive character, the same argument works in any dimension n.
Since we double the number of subcells each time we operate the decomposition
Ci = Ci + c~, we end with 2 n subcells. But only those which are spanned by d
projections ei of basis vectors on Ell and (n - d) projections e~ on E1. have a
non-zero volume and their number is (~). As in the low dimensions case, one
easily verifies that the traces of these subcells on E1., which are parallelohedra
spanned by (n-d) projections e~, exactly cover the atomic surface 0"0 = 7I"1.(rn)
attached to the origin, so that our construction yields in the general case a tiling
of the d-dimensional cut Ec by means of the projections of the (~) d-dimensional
facets of the hypercube 'Yn.

2.4.3. Octagonal tilings

The canonical octagonal (or Ammann) tiling, shown in Fig. 3, was first in-
troduced by R. Ammann [17] and F. P. M. Beenker [18], and is obtained in a
straightforward way as Penrose-like tilings with n = 4, d = 2, and the direction
of the pair (Ell, E1.) prescribed by the following symmetry considerations:
Consider a regular octagon in the Euclidean plane and choose four of its
vertices, no two of them being opposite as shown on Fig. 4. Consider the
four vectors joining the center of the octagon to these vertices. There exists an
unique embedding of the plane in JR 4 such that the canonical orthonormal basis
of JR4 projects orthogonally on our four vectors. Now consider the symmetry
group 8mm' of the octagon. Since it permutes the vertices of the octagon,
we can define a 4-dimensional action of this group by the condition that it
 MATCHING RULES 147

Fig. 3. - A sample of the Ammann octagonal tiling.

permutes the basis vectors in the same way as the vertices of the octagon, and
since this action involves only signed permutations, the lattice Z4 spanned by
the basis is preserved by this action. Then we decompose this representation of
8mm' into irreducible representations and find two of them, one carried by our
embedded plane, which is identified with Ell, and the other by the orthogonal
plane, identified with E.l.
It is easy to see that the prototypic atomic surface 7r.l (/4) is a regular oc-
tagon, and that the six 2-dimensional facets of 1'4 fall under 7r1l on two orbits
of tile~: two squares (with orientations differing by a 7r/4 rotation), and four
rhombi with an acute angle of 7r / 4, again mapped on each other by rotations
which are multiples of 7r / 4. Observe how the basis vectors {e}, e2, e3, e4} project
onto the parallel and perpendicular spaces. On the parallel space, they are de-

E.l'

Fig. 4. - The orthogonal projection on Ell and on E.l of the standard orthonormal
basis Ci, i = 1, ... ,4 of JR4.
148 A. Katz

duced from each other by a rotation of 11"/4. In the perpendicular space, the
projections of the same vectors are mapped on each other by a rotation of 511"/4
(or -311"/4) as shown on Fig. 4.

3. THE COMPOSITION-DECOMPOSITION METHOD

3.1. Self-similarity

In some cases, there exist lattice-preserving linear transformations which com-

mute with the two projections (11"11, 1I"..l)' Such transformations are elements M
of Gl(n, Z) which preserve the subspaces Ell and E..l.
Given a tiling associated to the canonical atomic surface 1I"..l (rn), one can
construct a new atomic surface as 1I"..lM{rn), by taking the projection of the
image through M of the unit hypercube. Since M preserves Ell and E..l, it
transforms any cut Ec parallel to Ell into a parallel cut M(Ec) which carries
the image under the restriction of M to Ell of the tiling carried by E c .
Of special interest is the case where M operates on Ell and E..l by homoth-
eties, since in this case the image of the tiling under M is a tiling of the same
type, but at a different scale:

IAILI = 1

This happens in particular when there is an invariance point group of the high-
dimensional lattice, such that Ell and E..L are the only two invariant subspaces,
carrying irreducible non-equivalent representations of the invariance group.
Let us give an explicit example with the octagonal tiling. A simple exam-
ination of the projection of the canonical basis onto Ell and E..l shows that
el + e2 + e3 = (v'2 + 1) e2 in Ell while e~ + e~ + e3 = - (v'2 - 1) e~ in E..l. This
suggests to construct the matrix:

M = (J2 + 1)11"11 - (J2 - 1)1I"..l = ( ~ : [ - ~1 )

-1 0 1

M is easily seen to have all the required properties: it belongs to Gl(4, Z),
its determinant is 1 and it commutes, by construction, with the action of the
octagonal group, so that it reduces on Ell to a dilatation of ratio (v'2 + 1) and
on E..l to a contraction of ratio (1 - v'2). Notice that such a matrix M exists
also in the icosahedral case. The corresponding "inflation" and "deflation"
ratios are 2 ± J5.
Now, our general argument shows that if we replace our original atomic
surface by an octagon (v'2 - 1) times smaller, then we will find in any cut an
octagonal tiling scaled by a factor (J2 + 1). In particular, if we consider both
atomic surfaces: our original one containing the smaller one, we see that we
 MATCHING RULES 149

Fig. 5. - A sample of the octagonal tiling, showing some deflated tiles.

can "extract" from any tiling the vertices of a larger one, by discarding all the
vertices which correspond to intersections of the cut with the large octagon,
falling outside the small one, as shown on Fig. 5. In the cut, one can describe
this operation as the regrouping of clusters of tiles to form larger tiles, and
this is called a deflation. Since the matrix M is invertible, this process may be
done in the reverse way: it is possible to "dissect" the tiles of a given tiling, in
order to obtain a tiling of the same type, but with an edge length shortened by
a factor (v'2 - 1). This is called an inflation, because it enlarges the number
of the tiles.

Observe that for these considerations we are not interested in comparing the
tiling carried by Ec and M(Ec), because the "absolute" position of the cut is in
general difficult to assess (due to the so-called local isomorphism property, see
for instance [19]) unless the tiling has special (global) symmetry properties. On
the contrary, we are interested in comparing two tilings carried by the same cut.
This entails that the position of the small atomic surface inside the large one is
irrelevant: whatever this position, the discarding process explained above will
lead to a "deflated" tiling. We conclude that generally the inflation/deflation
constructions are not uniquely defined, although they usually become unique
if some further natural requirements are imposed.
150 A. Katz

....
•
•..
,4 !"

'

Fig. 6. - Composition and decomposition for the Ammann decoration of the octag-
onal tiling

3.2. Inflation and quasiperiodicity

In order to describe these additional requirements, let us first consider the infla-
tion process. A natural requirement is that the (globally defined) inflation can
be described by a "local" dissection of tiles into smaller ones, which means that
the dissection follows rules (the decomposition rules) which are purely local in
the sense that the dissection of each tile depends only on a finite neighbour-
hood of the tile. If we require furthermore that the decomposition rules are
compatible with symmetry, by treating symmetry-related tiles in a symmetry-
related way, inflation usually becomes unique. In the case of the octagonal
tiling, this latter requirement implies that the octagonal atomic surface for the
initial tiling and its inflation are concentric.
For what follows, it is important that the global and the local definitions of
inflation define the very same inflation process. The local definition of inflation
can therefore be used as a tool to generate the tiling: one simply starts with a
finite seed, for instance reduced to one tile, and repeatedly applies the inflation
or decomposition procedure, followed each time by a rescaling so as to maintain
the same size for the tiles. In this way, a tiling of any desired size may be
constructed. This method has been used in particular when no global methods
were available (R. Penrose [5J, L. Danzer [20]) or when the atomic surfaces are
fractal (P. Stampfli [21J, E. Zobetz [22]).
So far, we have considered a local definition only for inflation. In many cases,
however, deflation also allows for a local definition, through local rules following
which one can build "supertiles" from clusters of original tiles (composition
rules). As for the decomposition rules, locality means here that the composition
rules should depend only on a finite neighbourhood of the cluster. However,
locality of deflation is a much less trivial property than locality of inflation.
Now, if both inflation and deflation allow for a local definition, they can
be used to prove that a properly chosen set of matching rules enforces the
quasi periodicity of any tiling in which these rules are satisfied everywhere.
 MATCHING RULES 151

Let us illustrate this on the case of the octagonal tiling carrying the Ammann
decoration depicted on Fig. 12 in section 5 where the origin of this decoration is
explained. Given the Ammann decoration, the composition and decomposition
rules can be described in a very local way, as shown on Fig. 6. In fact, one can
show (see [27]) that any tiling admitting the Ammann decoration (i. e., satisfy-
ing the matching rules) has a unique composition and a unique decomposition
(whether it is quasiperiodic or not). This is the key property for proving that
the matching rules enforce the quasi periodicity of the tiling.
Let T be any tiling carrying the Ammann decoration, and let us consider a
finite, but otherwise arbitrary patch P from T. We can apply repeatedly the
(unique) composition process to P, until it is so small that it is obviously a piece
of a quasiperiodic tiling (notice that all the vertex neighbourhoods allowed by
the matching rules do occur in the quasiperiodic tHings, as can be checked by
inspection). But since composition and decomposition are unique and inverse
of each other, we can conclude that P itself must be a piece of a quasiperiodic
tiling. Since this is true for any finite P, we have shown that that T must be
quasiperiodic.
The procedure described above is actually a method (see [3]) to prove that
a certain set of matching rules enforces quasiperiodicity. It has been explicitly
or implicitly used by many authors, including N. G. de Bruijn [6], J. E. S.
Socolar [23], F. Gahler [3], R. Klitzing, M. Schlott mann and M. Baake [24] [25]
and probably others. The key requirement for this method to work is that both
inflation and deflation are unique and local, and that they are defined for all
tilings satisfying the matching rules.

4. THE METHOD OF FORBIDDEN PLANES

We switch now to a quite different approach to the theory of matching rules.

4.1. Position of the problem

Let us first observe that, given a canonical tiling and its prototiles as suitable
projections of facets of the hypercube, it is quite possible to describe any tiling
made of the same prototiles with a cut through the same lattice of atomic
surfaces used for obtaining the quasiperiodic tiling: in fact, we can distinguish
between the subspace En on which is built the tiling and the cut Ec used to
select the vertices of the high-dimensional lattice which are projected on the
vertices of the tiles. As soon as the direction of El., considered as the carrier of
the atomic surfaces and the direction of the projection on En, is irrational with
respect to the lattice, it is clearly possible to lift in an unique way each vertex
of an arbitrary tiling on a vertex of the lattice. Then it is possible to dissect
the tiles in simplices (triangles in two dimensions, tetrahedra in three and so
one) and to lift these simplices to linear affine simplices of lRn in order to get
a "faceted" cut which selects the relevant vertices (if it happens that this cut
152 A. Katz

intersects additional atomic surfaces, it is quite clear that it is always possible

to distort locally the cut in order to remove the extra intersections). Of course,
the cut is not uniquely defined: any cut which goes through the same set of
atomic surfaces selects the same set of vertices and defines the same tiling, as
exemplified on Fig. 7.
Since our atomic surfaces are topologically balls, it is easy to show that two
cuts define the same tiling if it is possible to distort continuously one cut until
it becomes equal to the other without crossing the boundary of any atomic
surface: in mathematical terms, we say that a tiling is defined by a homotopy
class of cuts in the complementary of the boundaries of the atomic surfaces.
The special property of our quasiperiodic tilings is that this homotopy class
contains a plane (up to now, we have considered mainly these planes), so that
the problem of proving that a given matching rule enforces the quasiperiodicity
is to show that every homotopy class of cuts compatible with this rule contains
a plane. As we shall see, the direction of this plane cannot be arbitrary.
Our strategy to deal with this problem is best explained by referring to the
low dimensional model of Fig. 1. We accept now as a cut any line projecting
one-to-one on Ell along El. and we may distort it as we want as long as the cut
does not cross any endpoint of an atomic surface. How could we impose through
local constraints that such a homotopy class of cuts contains a straight line? A
first step in this direction is provided by the oblique tiling (see Fig. 2). As is
easily seen, the requirement of getting whole tiles upon projection (i. e., to get
only the two projections of the edges of the canonical square in the projected
structure on Ell) is achieved if we impose that the cut does not intersect the
components of the boundaries of the oblique tiles which are parallel to Ell. Of
course, this constraint does not forbid the cut from wandering very far, but
nevertheless results in a kind of local channelling of the cut, of which we are
looking for a global counterpart.

Fig. 7. - The curved cut drawn here can be continuously distorted to a straight line
while generating the same tiling: all curves generating the same tiling belong to the
same homotopy class (see text).
 MATCHING RULES 153

The most natural idea is then to enlarge the line segments that the cut is
forbidden to cross and to see what happens: in fact, there are only two possible
situations. The first one corresponds to rational values for the slope of Ell' In
such a case, for a certain finite enlarging, the endpoints of our line segments
will connect and the re..<;ulting lines divide the plane into parallel stripes. A
homotopy class of cuts is made of all possible cuts confined in one of these
stripes and contains evidently straight lines. The corresponding tilings are
obviously periodic.
The other case corresponds to irrational values for the slope of Ell' Then,
whatever large is the finite enlargement we make on the segments, they will
never connect and nothing happens: we do not obtain a global channelling in
this way.
These considerations may look somehow disappointing, since all we get is that
we may define matching rules in this way only for periodic tHings. However,
they are the real key to the theory of matching rules in higher dimension, where
the "special" directions which allow this machinery to results in matching rules
are not only the rational ones. We shall now discuss that point.

4.2. Non-transversality conditions

In higher dimension too, we can make such a rough classification: if Ell is a

d-dimensional subspace in IRn, we classify the directions of Ell by the rank of
the sublattice of 7l. n E IRn which falls on Ell' This rank is d for a "completely
rational" subspace associated with periodic tilings, and more generally it is the
number of dimensions along which the tiling is periodic. In the "completely
irrational" case in which we are interested, cuts parallel to Ell go through at
most one vertex of 7l. n . Observe that this is equivalent to the fact that the
projection of 7l. n on E.l along Ejl is uniformly dense (and notice that this is
independent on the direction of E.l, which is involved here only as the quotient
IRnjEII)'
However, this rough classification does not describe sufficiently the situation
and we have to introduce a finer one. The idea is to consider the projection
on E.l of the lattice subspaces of dimension t greater than or equal to (n - d)
(the dimension of E.l). For an arbitrary direction of Ell, the projection covers
E.l. But for special directions of Ell, there may exists one or several lattice
subspaces whose projections "loose dimensions" and does not cover E.l. Of
course, this is the case for the completely rational case, but the important
point is that such situations may also occur in the completely irrational case.
In fact, the corresponding property of Ell is that it is not transversal to a lattice
plane and this does not require Ell to be itself, partially or completely, rational.
This fact is indeed widely used in quasicrystallography of icosahedral phases:
all 2-dimensional rational planes spanned by any two integer vectors and e
M(e) of 7l. 6 , where M is the inflation matrix (section 3.1), are not transversal
to Ell: their projection on Ell are equal to their traces on Ell' For example,
5-, 2- or 3-fold symmetry 2-dimensional planes project on Ell along lines and
154 A. Katz

not on planes as would be expected in a generic case. This exceptional non-

transversality is the basis of the definition of "rows" and "reticular planes" in
the quasicrystallography of the icosahedral phases (see for instance [26]).
In such a non-transversal situation, the projection of 7l.. n on El. acquires
some extra structure: not only this projection is dense in El., but its trace ()n
the projection of the non transversal lattice subspace is also somewhere dense
since this projection, which contains a module of rank t, has a dimension lower
than t.
This is the situation in which non trivial matching rules may exists. Rather
than developing the general theory in an abstract way, we shall explain with
sufficient details the case of the octagonal tiling, which is the most simple case
but contains all the main ideas, and moreover is very close to the icosahedral
case of physical interest.

4.3. The forbidden planes

Like in the low dimensional model, let us consider the oblique tiling associated
with the octagonal tiling and explore it by shifting the cut. When the cut is
in a generic position, then it intersects the boundaries of the oblique tiles only
on parts parallel to El. and we see in the cut a regular tiling. But when we
let the cut hit the boundary of one atomic surface, then it intersects around
that point parts of the boundaries of the oblique tiles which are parallel to Ell:
namely we see in such a special cut "a flipping hexagon, at the instant where it
flips". As already observed, to forbid a cut to cross such parts of the boundary
of the oblique tiles forces it to select whole tiles.
Now, the remarkable point is that in such a special cut, we find not only one
flipping hexagon, but a whole infinite family aligned along a so-called "worm"
(Fig. 8). This comes from the fact that there are translations in Z4 which
project on El. along the boundary of the atomic surface, and these projections
fill densely the carrier of this boundary.
To be more specific, let us consider the segment of the boundary of the
octagon parallel to e~. Its carrier is the projection of a plane P parallel to the
lattice plane spanned by (cI, C2 - c4). But this carrier is also the intersection
of P with El., so that we find in P a rectangular lattice of segments parallel
to e~.
On the other hand, our spec~al cut intersects P along a line parallel to el.
Thus we get exactly the situation of our low dimensional toy-model: a lattice
(here rectangular) of atomic surfaces (the segments e~ bounding the octagon)
irrationally oriented with respect to the lattice, and a cut (here a line parallel
to el) also irrationally oriented: we conclude that we have a quasiperiodic
distribution of intersections between the cut and the segments, which, viewed
in the four dimensional space, is a set of intersections of the special cut with
boundaries of octagons. They are aligned along a straight line and separated
by finite distances. Each of them belongs to pieces of boundaries of oblique
tiles, whose traces in the cut build the "flipping hexagons". These hexagons
 MATCHING RULES 155

Fig. 8. - Infinite family of flipping hexagons defining a "worm" of the first kind in
the octagonal tiling

are separated by One or two squares.

Observe that this highly non generic situation is the consequence of two
distinct features of the construction of the octagonal tiling.
On one hand, the two planes El. and Ell' whose direction is forced by the
octagonal symmetry, are non transversal to a set of lattice planes, specially
those of the kind (ell e2 - e4) , but they are infinitely many of them (and in
fact we shall need another family to make the matching rules) . These planes
project On lines On El. and the two-dimensional lattice which they contain
densely fills the projection. This non-transversality property accounts for the
fact that in the plane P, the cut yields a one-dimensional trace and that the
given intersection between the cut and the boundary of an atomic surface yields
a whole lattice of such intersections.
On the other hand, it is because the boundary of the atomic surface is made
of segments contained in planes like P that we find in P a lattice of segments.
Would this segment be replaced by a curve in El., the plane P would still touch
a lattice of atomic surfaces as soon as it touches one, but it would touch them
on points instead of whole segments, so that the trace of the cut in P would
not intersect more than one of them.
Now we shall apply the naive idea stated in the introduction to this section:
we shall "fill the holes" between the flipping hexagon by means of decorations
which result in the dissection of the tiles of the oblique tiling and thus in an
enlargement of the number of their boundaries. Deferring the explicit construc-
tion to the next section, let us first examine what will be the topological result
of this construction.
In fact, the geometry of the boundaries "parallel to Ell" of the oblique tiling
is not very simple. They are three dimensional facets extending for two dimen-
156 A. Katz

sions in Ell and for the other one in EJ... To see how they glue together, let us
move our special cut: for a given position, we see in the cut a continuous worm
made of flipping hexagons and of squares coming from the decoration. When
we move our cut, we recover a regular cut unless we move along ei, in which
case we keep the worm, whose precise structure changes with the shift of the
cut along ei, according to the change of the structure defined by the restricted
cut construction in the plane P.
For each position of the cut, the worm covers completely its symmetry axis,
which is a line parallel to el so that when we shift the cut along ei, we see
that the sets of facets covers completely a plane spanned by ei and el, which
is parallel to our plane P. Thus we see that this set of facets is a complicated
"thickening" of a very simple object: a two-dimensional lattice plane, suitably
shifted. Finally, observe that it is equivalent to ask for the non-intersection
of the cut with the set of facets, or with the plane on which this set can be
"retracted" (Le., flattened).
Thus the final form of our constraint will be the non intersection of the
cut with a family of lattice planes, which we call forbidden planes. Before
explaining the consequences of such a constraint, let us describe how a suitable
decoration of the tiles together with a matching rule may be equivalent to the
constraint for the cut not to cross the forbidden planes.

5. DECORATION OF THE TILES

In this section, will shall describe the decoration scheme first devised by R.
Ammann for his octagonal tilings. This is not the only possible decoration
(decorations are never unique) and perhaps not the simplest. However, this
choice presents the advantage of involving many different topics of this kind of
construction.

5.1. A simple case

Although there is no mathematical difficulty in the constructions that we are

making, their description may be obscured by the fact that they rely entirely
on the geometry of the oblique tiling, which is intrinsically four dimensional
and not so easy to visualise.
Our first step is to explain why there are holes in the worm which appears in
the special cuts going through the boundary of one octagon (and thus of a whole
infinite quasiperiodic set of them). Recall that the boundaries of the oblique
tiling we are interested in are made of the sum of a tile and of a segment of the
boundary of its existence domain. Consider for instance the square belonging
to the flipping hexagon spanned by (e2, e3, e4), which is the "vertical" hexagon.
This square is (e2, e4) so that its existence domain is the square (ei, e~). If the
cut touches an atomic surface along ei, say on the point x, the relevant piece
of the boundary of the oblique tiling is the cube (e2, e4, eD.
 MATCHING RULES 157

Now, consider one ofthe squares (el, e3) touching this flipping hexagon along
the worm (Fig. 9). It has a vertex v shifted from the flipping point x bye2,
so that the cut pierces the atomic surface of v in the existence domain (~, e~)
of our square on a point on a diagonal of this existence domain. Since this
diagonal does not belong to any boundary of the oblique tiling, our square is a
regular one and does not belong to the forbidden set: in that sense, it is a hole
in the worm.
Thus we see that in order to fill this kind of holes and make the whole worm
a forbidden set, we have to dissect the oblique tiles corresponding to squares
into several sub-tiles in such a way as to introduce as a new type of boundary
the sum of a square of the tiling with the diagonals of its existence domain.
This is readily achieved through the simplest decoration of the edges of the
tilings, which consists in attaching an orientation to them.
Consider the existence domain of the origin of an edge, say e3, which is a
"horizontal" hexagon inscribed in the bottom of the octagonal atomic surface.
Let us cut this hexagon into two parts along its long (horizontal) diagonal
(Fig. 10) and let us decide to associate an arrow pointing upward to each edge
whose origin falls in the bottom half of the hexagon, and pointing downward
for those whose origin falls in the other half. Now observe that this decoration
is compatible with the symmetries of this edge: namely, the vertical symmetry
axis preserves each decorated subdomain, while the associated operation in En
exchanges the two sides of the edge (so that the arrow have to be symmetric
with respect to the direction of the edge). Concerning the horizontal symmetry
axis, the situation is slightly more subtle. This axis maps the existence domain
of the origin of the edge to the existence domain of its extremity, while the
associated operation in En exchanges the two end-points of the edge. To get a
symmetry compatible decoration, we thus have to put a downward arrow in the
upper half of the upper hexagon and an upward one in the other half. Then we
verify that we have actually defined an orientation for the edge, since a given
edge is oriented in the same way, viewed from its origin or viewed from its
extremity. Then we can set the decoration of the other edges, using associated
symmetry operations in El.. and in En.

Fig. 9. - The vertices of the squares close to a flipping hexagon fall on the diagonal
of their existence domains.
158 A. Katz

E.L

i-
t
Fig. 10. - The orientation of the edges corresponds to the dissection of their corre-
sponding existence domains.

o
E,L ~I

o
o
o
Fig. 11. - The decoration of the square induced by the decoration of the edges (left);
the unique decoration of the rhombus (right).

Now we come to the corresponding decoration of the tiles themselves, which

amounts to superimpose the atomic surfaces corresponding to the vertices of
the tiles, shifting them as required, and to study how the existence domains of
tiles are partitioned in existence domains of decorated tiles (Fig. 11). What we
find is that the existence domain of the rhombi are not partitioned, so that we
get immediately only one kind of arrowed rhombi, with the arrows pointing to
the obtuse angles. On the contrary, the domain of the squares are partitioned
in four decorated subdomains by their two diagonals, and that is what we were
looking for. Observe however that, although there are four subdomains, there
is only one decorated square, characterised by the fact that parallel edges are
given the same orientation. But this unique square is attached in different
orientations to each of the subdomains.
It remains to see what is the matching rule and how it works. The matching
rule itself is trivial in this context: it requires to fit tiles together with the
same orientation for the shared edges (what is trivial is that within the cut
 MATCHING RULES 159

construction, each edge is in fact selected only once, together with its arrow
if it is arrowed). But we have to show why it is forbidden to cross the new
"internal" boundaries of the decorated oblique tiling (recall that it is forbidden
to cross the original boundaries in order to get only whole tiles in the cut).
Now, the point is simply that if we allow a cut to enter into an oblique tile
through an edge of a given orientation, and then to cross the internal boundary
corresponding to this orientation, then it will leave the oblique tile through an
edge having the opposite orientation, and we would get a wrongly decorated
tile. If we insist to make a tiling with only our "well decorated" tiles, then we
forbid the cut to cross these new boundaries. Notice in particular that in this
context, a "fault" is not the mismatch of decorations, which means nothing as
long as we use continuous cuts. It is the occurrence of an improperly decorated
tile.
To sum up, we have set simple decorations of the edges, obtaining one dec-
orated rhombus and one decorated square, such that making a tiling with
these two tiles alone and matching decorations corresponds to use a cut which
does not intersect the family of forbidden planes made of four lattices of 2-
dimensional lattice planes spanned respectively by (el, e2 - e4), (e2, el + e3),
(e3, e2 + e4) and (e4, e3 - ed·
Alas, this is not sufficient to get the octagonal quasiperiodic tilings (although
the edge decoration yields matching rules! see section 7). The reason for that
is quite clear. Recall that the directions of Ell for which we may have matching
rules are required to be non transversal to a set of lattice planes. But there
are infinitely many directions of planes which are non transversal to the four
above-mentioned directions. More precisely, there is a one parameter family
of them, which all possess the symmetry of the square, and only two of them
have in addition the full octagonal symmetry. Thus, in order to get matching
rules for the octagonal tiling, we need to add more forbidden planes in order
to get a set such that our E.L and Ell will be the only planes non transversal
to this set.

5.2. The Ammann decoration of vertices

The strategy is clear: pick suitable other lattice planes, and devise decorations
such that using only these new decorated tiles corresponds to not crossing these
planes, i. e., defines them as forbidden planes.
In order to illustrate various aspects of this theory, and to honour the clever-
ness ofR. Ammann (see [17), we shall sketch the description of his decorations
(and finally show that they enforce the quasiperiodicity, a result which may also
be obtained from a self-similarity approach, as we have already seen).
First, the choice of the new planes. Besides the planes which project on
En and E.L on the same directions as the basis vectors, it is natural to try
those which fall on the bisectors of them, because these bisectors are simple
directions (diagonals of the rhombi) and correspond also to the diagonal of the
octagon. Thus we choose the planes spanned respectively by (el + e2, e3 - e4),
160 A. Katz

(C2 + C3,C1 + c4), (c3 + C4,C2 - c1), (c4 - C1,C3 - C2). It is a matter of simple
algebra to verify that the only planes simultaneously non transversal to the
whole set of eight directions of planes are our El. and Ell'
Second, the decoration. It is a vertex decoration, defined through a partition
of the atomic surface (existence domain of a vertex) much like the previous edge
decoration was defined through a partition of the existence domains of edges.
Thus, we divide the octagon in eight sectors by the diagonals running from
a vertex to the opposite one, and we attach to each sector a mark, arbitrary
excepted for the fact that it should have the same symmetry as the sector (a
symmetry axis parallel to a basis vector). Let us choose for the moment the
generic shape depicted on the upper left of Fig. 12. We complete the decoration
of the other sectors using symmetry.
Third, we examine how many different decorated tiles we get. As usual (see
[19]), we superimpose suitably shifted copies of the octagon and we study the
decomposition of the existence domain of each tile in subdomains of decorated
tiles. What we find is that there are many of them, so that we get a lot of
decorated tiles and this does not look very pretty. However, this complication
is inessential and we can simplify the decoration using a "reduction trick" that
we shall now explain.
Let us proceed by steps. Initially, we have a generic mark attached to each
vertex and the matching rule consists in putting tiles around each vertex in
such a way that the marks coming from the different tiles coincide. Observe
that since the mark is attached to the vertex, it overhangs the edges of the
tiles. We can give an alternative form to the matching rule by attaching to
each vertex of tile only the trace of the mark on this tile, and ask to recompose
the shape of the mark around each vertex. Clearly, these two forms of the
matching rule are equivalent as long as the trace of the mark on the tile gives
sufficient information to reestablish the whole mark, and that is the case if the
shape of the mark is generic enough.
Next, we can modify and specialise the shape of the mark in such a way
that some traces of our mark become identical. To see the effect of this proce-
dure, consider two adjacent subdomains of existence in the existence domain
of a tile. The internal boundary between the subdomains is there because at
some vertex the tile bears the mark in two different orientations for the two
subdomains. If we modify the mark up to identify its trace on the tile for these
two different directions, then we can no longer make any distinction between
the two subdomains and this amounts to rub out the corresponding internal
boundary.
This is a way to simplify the decoration and to decrease the number of
decorated tiles. Of course, one has to be careful not to go too far: to put
a disk on each vertex is obviously equivalent to no decoration at all .... A
remarkable property of the octagonal tiling is that there is a class of shapes for
the mark, among which the Ammann's "large arrow" , which yield an effective
decoration with only one decorated square and two decorated rhombi (which
are mirror images of each other, see Fig. 12). We get only one decorated
 MATCHING RULES 161

D
Fig. 12. - A generic vertex decoration for the octagonal tiling (upper left). The
Ammann arrow is a special distortion of this generic mark (upper right) which leads
to three kinds of decorated tiles. Complete Ammann and edge decorations of the
octagonal tiling (bottom).

square, because we have rubbed out all the internal boundaries. Those which
survive in the existence domain of the rhombus are easily seen to be the two
diagonals: two opposite triangles correspond to the same decorated tile, and
the two diagonals are symmetry axes associated in Ell to two symmetry axes.
Each of them exchanges the two decorated rhombi.
Now, we have to see that these internal boundaries of the existence domains
of the rhombi are sufficient to provide us with the required supplementary set of
forbidden planes. To fix the notations, let us consider a special cut piercing an
ea
atomic surface somewhere on the diagonal spanned by e~ -e~ (or +e~). Then
this cut will pierce also a whole quasiperiodic set of atomic surfaces along the
same diagonal. Since this cut hits the internal boundary (the long diagonal) of
the rhombus (e~, e~), which is the existence domain of the rhombus (e3, e4), we
see in the cut these rhombi as traces of the boundaries of the decorated oblique
tiling, and similarly for the rhombi (el' e2) through the short diagonal of its
existence domain (ea,e~). To see that this set ofrhombi covers completely its
axis (a line parallel to el-e2), let us build the restriction of our four dimensional
cut construction to the lattice plane (c3 +c4, C2 -cd which intersects our special
cut along a line. We just have to observe that the trace of the atomic surface in
this plane, which is its diagonal, is exactly the projection on the trace of El. of
the unit rectangle of the two dimensional lattice to conclude that we are again
exactly in the situation of the low dimensional model and that this restricted
cut construction yields a tiling of the trace of the special cut by means of the
two projections of the basis vectors, which are nothing but the diagonals of our
rhombi.
162 A. Katz

Fig. 13. - A worm of the second kind in the octagonal tiling.

Thus we get a "worm of the second kind" made uniquely of rhombi, which
does not have holes and covers completely its axis (Fig. 13). Moving the special
cut along the direction ei - e~ of the diagonal shows as previously that the
whole plane (c3 + C4, C2 - cd is contained as a retract in the boundaries of
the decorated oblique tiling. On the other hand, the fact that this plane is
forbidden results of the same considerations already made: would the cut cross
the internal boundary, this would result in an improperly decorated rhombus.
Finally, the whole set of decorations for the octagonal tiling is the superim-
position of the orientation of the edges (Ammann used half-disc rather than
arrows) and of the vertex decoration. The matching rules require that two ad-
jacent tiles give the same orientation to their common edge and that the mark
on the vertices fit together to form the large arrow. We know that this set of
constraint is equivalent to the non-intersection of the cut and the forbidden
planes. We shall now prove that this entails the quasiperiodicity of our tilings.

6. THE MAIN THEOREM

For this, we first need more information about the position of the forbidden
planes with respect to the lattice, and about their intersections.

6.1. Position and intersections of the forbidden planes

Setting once for all the positions of the atomic surfaces such that their centers
are on the vertices of the lattice 7l. 4 , let us begin by the second set of planes
corresponding to the vertex decoration: since they intersect octagons along a
diagonal, they go through the vertices of 7l. 4 • It is a matter of simple algebra to
 MATCHING RULES 163

study their intersections. One finds that they intersect all four on the vertices
of the lattice and on the body centers, and that in addition each pair of planes
which are orthogonal in both projections in E.l and Ell intersects on a face
center.
Concerning the first family, consider a "horizontal" worm parallel to el. The
corresponding forbidden plane is parallel to (cl, C2 -c4) but does not go through
the lattice from which it is shifted twice: first, this plane passes on the boundary
of the atomic surface and thus is shifted from its center by (-e~ + e~ - e~)/2.
Second, it is shifted to the axis of the worm. To com pute this shift, observe that
after the first shift the plane goes through the flipping points of the hexagons
of the worm, so that we have to shift it to the middle between the flipping
point and its image under the flip, which are symmetrical with respect to the
axis of the worm. Since the jump vector is e3 - (e2 + e4), the last shift is
(-e2 + e3 - e4)/2 so that the total shift from the vertex is (-c2 + C3 - c4)/2.
Now the plane itself contains cI/2 and we conclude that these planes go through
the body centers of the lattice. Then one computes easily that they intersect
all four on the body centers and that like in the previous case the "orthogonal"
pairs intersect also on face centers. Finally, any two planes coming from the
two families intersect only on a body center.
The conclusion is thus that we have three kinds of intersections: pairs of
"orthogonal" planes coming from both families on the face centers, quadruplets
of the second family on the vertices, and octuplets of all planes on the body
centers.
Our problem is now mathematically clean: we have to show that any cut
(everywhere transversal to E.l) which does not intersect any forbidden plane
is homotopic to a plane parallel to Ell. To begin with, we give the following
description of the homotopy classes in question.

6.2. Systems of data

Let us describe the cut by its intersection with each of the fibers of the pro-
jection ?Til on Ell (the fiber above the point x E E" is the plane parallel to
E.l which intersects E" on x). It is very important to observe that, due to
the non-transversality of the forbidden planes with respect to E.l, the trace
of the forbidden planes on each fiber ?T,,-l(x) is not a point for any x as is
the generic case for two (two-dimensional) planes in ffi4. On the contrary, we
have for each forbidden plane the following situation: the plane projects on E"
along a line, and if x does not belong to this line, then there is no intersection
between the plane and the fiber. When x belongs to this line, then the forbid-
den plane intersects the fiber along a line. Finally, it is easy to observe that
when x is the intersection of the projections of several forbidden planes, then
the corresponding lines in the fiber also intersect on a common point.
This situation allows a very simple description of the homotopy class of the
cut, which is transversal to E.l and thus intersects each fiber on exactly one
point: we have only to specify, for every x E Ell, through which connected
164 A. Katz

component of the complementary of the trace of the forbidden planes goes the
cut. This is what we call a system of data. The data are of three possible types:
first, for almost all x, the fiber does not intersect any forbidden plane and there
is only one connected component to choose (the whole fiber). Second, for a x
which lies on the projection of a single forbidden plane, the fiber is divided
in two half-planes by its intersection with the forbidden plane and we have
to choose one of them. Finally, if several forbidden planes are involved, they
define in the fiber a set of sectors around their common intersection and we
have to choose one of them.
It is now almost obvious that systems of data defined by continuous cuts
are in one-to-one correspondence with the homotopy classes of cuts. In fact,
consider two cuts defining the same system of data. This means that in every
fiber their traces belong to the same connected component of the complemen-
tary of the traces of the forbidden planes. Since these components are convex,
we can draw in each fiber the line segment between the traces of the two cuts,
bel'lg sure that this segment does not intersect the forbidden planes. Then, we
can interpolate between the two cuts along these segments in order to define a
homotopy from one cut to the other. Reciprocally, if for at least one fiber the
two cuts fall in different connected components, then it is quite clear that it is
not possible to distort continuously one to the other without crossing one or
several forbidden planes.
Thus we have shown that the homotopy classes of cuts are classified by the
associated system of data. Our task will be now to characterise those systems
of data which stem from continuous cuts.

6.3. Propagation of order

The number of possible systems of data seems enormous, since it corresponds

to making a choice for each x E Ell' However, we have already seen that for
almost all x there is in fact no choice since the corresponding fiber does not
intersect any forbidden plane. A further step in -that direction is to observe
that for a continuous cut, we make in fact only one choice for each forbidden
plane. To see this, consider the three-dimensional space defined as the union of
the fibers 7r1l-1(x) for x belonging to the projection of a forbidden plane. This
three-dimensional space contains this forbidden plane which disconnects it so
that the trace of the cut in this space (a curve) must stay in one half-space
in order not to cross the forbidden plane. Thus the choices in each fiber are
correlated and this is best seen in projection onto E1.: 7r1. maps the forbidden
plane onto a straight line and the image of the previous curve must stay on one
side of this line. It is now obvious that any system of data compatible with a
continuous cut (we shall say just compatible) must correspond to the "same"
half plane all along the projection on Ell of any forbidden plane, the correlation
between half-planes being established through their projection on E1..
The next step is to investigate how the choices are or are not correlated
for intersecting forbidden planes. As observed above, there are three cases to
 MATCHING RULES 165

==P ----~~----

Fig. 14. - Basic mechanism of the propagation of order: the given choices corre-
sponding to the forbidden planes (1) and (2) induce the choice of the sector on the
central picture for which the choices of the forbidden planes of type (3) is Wldeter-
mined but imposes those of type (4), (5) and (6) planes.

consider. In any case, the choice to be made for constructing a system of data
is the choice of a angular sector in the fiber, among 4 possibilities in the first
case, 8 in the second and 16 in the third.
Here again, there is an obvious constraint: the chosen sector must be con-
tained in all the half-planes selected with respect to the lines intersecting on
the considered point. This requirement has no special consequences in the case
of pairwise intersection (the sector is then a quadrant which is exactly the in-
tersection of the two half-planes associated to the two lines) but the situation
is quite different for the other cases. To understand that point, refer to Fig. 14
and let us suppose that we know what is the half-plane selected along the two
lines (1) and (2), as indicated. Then we see that we cannot say anything about
the line (3), for which both choices are compatible, but that we can assert what
is the half-plane selected "above" the lines (4,5,6), for which only one choice is
compatible with the data already known .
This elementary property is in fact the very source of the propagation of the
local order which eventually results in the global quasiperiodicity of the struc-
ture. The global consequence of this propagation is contained in the following
theorem which expresses what are the compatible systems of data, in the sense
that they are induced by a continuous cut.

Theorem 1 The compatible system of data are those for which the closures of
the projections of all selected half-planes on E.l have a non-empty intersection.

As we shall see, this theorem entails that any tiling fulfilling the matching
rules is either quasiperiodic or is the limit of a sequence of quasiperiodic tilings
(and such a limit is to a large extent undistinguishable of a strictly quasiperiodic
tiling). But let us begin by the proof of the theorem.
166 A. Katz

Fig. 15. - To each bad triangle we can attach a bad strip.

6.4. Proof of the theorem

6.4.1. The pushing procedure

Since we want to prove that an infinite family of half-planes have a non empty
intersection, we need some tool for reducing the discussion to finite families.
Such a tool is conveniently provided by the classical Helly's theorem:

Theorem 2 (Helly) Consider in IRn a family:F of convex sets (which must

be finite for arbitrary convex sets but may be arbitrary for compact convex sets)
such that every finite subfamily of p ~ n + 1 sets has a non-empty intersection.
Then the whole family :F has a non-empty intersection.

Since our half planes are not bounded and their family is infinite, we shall
need some care when using this theorem. Nevertheless, it allows us to restrict
the discussion to pairs and triplets of half planes (in our case n = 2). The
only case where two half planes may not intersect is when their boundaries are
parallel. Let us call bad strip such a pair of half planes. Similarly, let us call
bad triangle a triplet of half planes which do not intersect on the triangle that
they define.
Our task is to show that for a compatible system of data, there is no bad
strip and no bad triangle. But it is immediate that we can associate a bad strip
to any bad triangle: recall that the forbidden planes intersect either by pairs,
by quadruplets or by octuplets, and in the first case, the two are orthogonal
in both projections on En and E.L. Thus a bad triangle involves necessarily at
least two vertices where 4 or 8 planes intersect, for which we can attach a bad
strip (in two or three ways) to the bad triangle, as depicted on Fig. 15.
Finally, it remains to show that there are no bad stripes. Let us suppose that
such a bad strip exists. We shall show a contradiction with the continuity of
the cut through the construction of a family of forbidden planes about which
we know what are the selected half planes. Observe that there are two kinds of
bad stripes: those which are parallel to a plane of the type £1, £2 - £4 and those
which are parallel to a plane of the type £1 + £2, £3 - £4. Since the argument is
almost the same in both cases, we shall deal in detail only with the first one.
 MATCHING RULES 167

u- -

Fig. 16. - The pushing procedure (see text).

For this, consider the Fig. 16, which is drawn in E.L and where the bad strip
is horizontal. Let us denote by the same letter U the "upper" plane of the bad
strip (parallel to (cI, C2 - c4)) and its projection on E.L, and similarly by D· the
other plane of the strip. We can find a vertical forbidden plane, say V (parallel
to (c3, C2 +c4)), which intersects the upper line U on a "eightfold" intersection
(one half or one quarter of the planes have this property). Let the selected half
plane associated to this vertical forbidden plane be the one on the right (which
means that the would-be cut would cross each fiber intersecting this vertical
plane on the right side of the intersection line).
We can get other vertical forbidden planes "selecting their right side" through
the following pushing procedure described on Fig. 16. Consider the plane la-
belled PI going through the intersection x and recall that x is the projection
on E.L of a body center of Z4 : x = 1I'.L(e). In the fiber going through e,we
find the traces of 8 planes, among which those of U, V and Pl. As already
mentioned, the half plane selected by Pl must intersect the intersection of the
half planes selected by V and U. We see that there is no choice and that the
upper-right half plane is selected.
Following PI, we arrive to its intersection with D, "above" the point y of E.L.
We are either on a fourfold or on a eightfold intersection, and we can consider
the plane P2. The same argument shows that P2 selects its lower-right half
plane. Along P2 we find its intersection with U above z E E.L, and through
this intersection we can consider the vertical plane V' parallel to V. Then we
immediately see that the half plane selected by V' is on the right. Observe
at this point that we need to start with a bad strip of non-zero width for this
argument to be non trivial, and this requires to consider in E.L the closure
of the projections of selected half planes, although in each relevant fiber we
consider the open selected half-plane, simply because our hypothesis is that
the cut never intersects any forbidden plane.
168 A. Katz

This is the pushing procedure. We can iterate it to exhibit a family of for-

bidden planes parallel to V, of which we know that they select the half plane
on the right in each fiber that they meet. Moreover, we have three possible
choices for the direction PI of the "pusher", namely those which projects on
El. on e~ (this is the direction of plane spanned by (e4. el - e3)), on (e~ - e~)
«e4 - e3, el + e2)) and on e~ + e~ «el + e4, e2 + e3)).
Let us discuss the distribution of this family of planes.

6.4.2. The cone of planes

In order to compute the distances over which we push the vertical plane, let
us assume that the distance between U and D (that is, the thickness of the
bad strip) is m~ + n( ~ + e~). Then simple geometry shows that the distance
between two consecutive planes of the family is:

• d = 2m~ + 2n(e~ - e~) when we push with (e4, el - e3)

• d' = (4n - 2m)ei + (2m - 2n)(~ - e~) for (e4 - e3, el + e2)
• d" = (2m + 4n)ei + (2m + 2n)(~ - e~) for (el + e4, e2 + e3)
Now, starting with a given plane V, we can reach any plane sitting at a
distance from V which is a linear combination with integer positive coefficients
of these three distances. Since d" = 2d+d', we can omit the third one. Observe
that the positiveness of the coefficients is very important and corresponds to
pushing the planes "on the side they select" in order to flush the would-be cut
to infinity.
Thus we find a cone of planes. To describe this cone precisely, let us consider
its trace on the plane spanned by (ell e2 - e4) orthogonal to V (Fig. 17).
The whole lattice of forbidden planes parallel to V intersects this plane on a
rectangular lattice C of ratio ../2. Choosing our V as the origin, consider the
two vectors u = 2mel + 2n(e2 - e4) and v = (4n - 2m)el + (2m - 2n)(e2 - e4)
corresponding respectively to d and d'. Our family C of planes is the trace on
the cone defined by the two positive half lines generated by these vectors, of
the sublattice of C which they span.
Observe now that the traces on this plane of El. and Ell are respectively the
first and the second diagonals of the plane. In fact, we must have on El. that
the projection of e2 - e4 is -../2 times that of ell and on Ell the ratio of the
two projections is +../2, whence the two diagonals.
The crucial remark is that whatever the thickness of the bad strip is, i. e.,
whatever the values of m and n are, the edges of this cone fall on the two
opposite sides of El.' To see this, observe that we have measured the thickness
upward, so that m - n../2 is positive. Then the projection of u and v on El.
are respectively 2(m - n../2)ei and 2«-2n - m) + (m + n)../2)ei = 2(../2-
1)(m - n../2)ei, and point both to the interior of the cone. On the contrary,
the projections of u. and von Ell are 2(m + n../2)el and 2«-2n - m) - (m +
n)../2)el = 2(../2 + 1)( -m - n../2)el, which point on opposite sides.
 MATCHING RULES 169

/
'v':~:::
// .•...
.........
.•...•.. .
.........
.•...•.. .
... ..•...
.•...•.. .
.•.. .•...
.•.. .•...
.........
... ...
.•.. .•... .. ..•.. .•.. ..•...•.. .•..
································E
••• , •••••••••••••
......
"
, ••••••••••••••••• •••••• .1.••

Fig. 17. - The set of planes about which the pushing procedure allows to assert which
half plane they select, is the trace of a lattice inside a cone. These planes project
discretely on E.J.. and densely on En.

For the other kind of bad strip, say, parallel to el + e2, e3 - e4, the drawing
of the cone is made in this plane where the relevant forbidden planes form a
square lattice. The traces of El. and Ell are here tilted by 31l"/ 4 with respect to
this lattice, but besides this details the main conclusion is the same: the two
half lines which bound the cone project one upon the other on El. while their
projections on Ell cover Ell'
In other words, the qualitative situation is always that of Fig. 17. We see
that the projection of C on El. is discrete, although the projections become
closer and closer when one goes farther inside the cone. But the projection of
the same family C on Ell is everywhere dense, and this allows us to conclude
in the following way:
Consider an generic point x in Ell, and the fiber above it, with the trace of
the cut. Since the projection of C is is everywhere dense in Ell, we can find a
sequence of lines in this projection converging to a limit going through x. This
sequence is the projection of a sequence in C which escape to infinity in the
cone, since C is a part of a lattice and does not have any accumulation point.
This means that we can find as close to x as we want, a fiber in which the trace
of the cut is pushed on the right as far as we want. This is a contradiction with
the continuity of the cut and we have shown that a compatible system of data
does not admit bad stripes (nor bad triangles as a consequence).
170 A. Katz

Fig. 18. - Once four forbidden planes have been found which form a parallelogram,
the fact that there are no bad stripes nor bad triangles forces the choice for every
forbidden plane whose projection on E.l does not cross the parallelogram.

Now we can finish the proof of our theorem. What remains to do is to apply
ReIly's theorem. For that, consider any forbidden plane and its associated
selected half plane in EJ.. There exists a parallel plane whose selected half
plane is on the opposite side: if that was not the case and siQce there is no
bad stripes, the cut would be entirely at infinity. Thus we have found a strip
of finite width such that we know for any forbidden plane parallel to this pair
and falling outside the strip, that the associated selected half plane is the one
which contains the strip.
Selecting another direction of forbidden planes, we find in the same way an-
other strip which intersects the previous one on a parallelogram. Now we know
for any forbidden plane whose projection on EJ. does not cross this parallel-
ogram, that the selected half plane contains the parallelogram: otherwise we
get a bad triangle, as depicted on Fig. 18.
Finally, we have to discuss the {infinite!} subfamily of forbidden planes whose
projection on EJ. intersects the parallelogram. But since they project at finite
distance of each other, it is no longer necessary to associate to each of them
a whole half plane and we can work with compact convex sets. Consider in
EJ. a disc D large enough to contain the parallelogram, and lift it along 7r J.
on each fiber. Then consider the intersections of our half planes with the discs
instead of the half planes themselves, for all fibers belonging to the subfamily
F' under consideration. Observe that the projection on EJ. of these bounded
convex sets intersects by pairs and by triplets, otherwise we would get a bad
strip or a bad triangle. Then applying Helly's theorem completes the proof of
theorem 1.

6.5. Quasiperiodic tHings and "special tHings"

Let us now turn to the consequences of theorem 1 on the quasiperiodicity of

well-decorated tilings {Fig. 19}. We have just shown that any continuous cut
 MATCHING RULES 171

defines a system of data such that the intersection of the corresponding (closed)
half planes is non-empty. Observe first that these intersections are always
reduced to a single point. In fact, two distinct points in E.L are always separated
by the projection of some forbidden plane, since each lattice of forbidden planes
projects on a dense set of lines. Thus two distinct points correspond to two
different systems of data. Then the conclusion depends on whether this point
x does or does not belong to the projection of some forbidden plane.
The simple case is when the point x falls in the complementary of the projec-
tion of the forbidden planes. Then we can consider the plane cut parallel to Ell
and going through x: it is a regular cut which intersect the atomic surfaces on
their interiors and defines a quasiperiodic tiling. This is the generic case.
Observe that in this case, the point x is as well the intersection of all the projec-
tions of the open selected half-planes, in such a way that we have a one-to-one
correspondence between quasiperiodic tHings and connected components of the
complementary of the projection on E.L of the forbidden planes.
But this family of quasiperiodic tilings does not exhaust the set of all well-
decorated tilings, and we have to consider the cases where the point x falls
on the projection of one or several forbidden planes. We shall call special the
tHings of this kind. For the sake of simplicity, let us first suppose that x belongs
to the projection D of one forbidden plane P only. Such a point x corresponds
to two distinct systems of data, differing by the choice relative to the forbidden
plane P. It is clearly impossible to choose a "perfectly" plane cut to define
the corresponding tHings: the only possible plane cut goes through P and we
have to distort it (as slightly as we want) on one side or on the other, above
the projection of P on Ell'
These well-decorated special tHings, which are not as "strongly" quasiperi-
odic as the previous ones, may be conveniently described in our case as limits
of quasiperiodic tHings in the following way:
Consider in E.L a smooth arc parameterised by the interval [0, 1], transversal
to the projections of all forbidden planes and such that the image of 0 belongs
to the projection D of a forbidden plane. Consider on this arc a sequence
of points Xn converging to the image x of 0 and such that no point in this
sequence falls on the projection of a forbidden plane. Then each of these points
Xn corresponds to a well-behaved quasiperiodic tiling Tn. We define the limit
tiling T corresponding to x as a simple limit: a vertex belongs to T if and only
if it belongs to Tn for infinitely many n (observe that we have constrained the
sequence Xn to belong to a smooth arc in order to avoid pathologies such that
sequences oscillating or spiralling towards their limit).
Because of the local constancy of local patterns with respect to shifts of
the cut along E.L, it is quite obvious that the limit obtained in this way is
a well-decorated tiling (any bounded region in T occurs in infinitely many
Tn). It follows in particular that these special tHings have no particular local
property. But they do have a special global property, namely they contain
a finite number of infinite worms. For instance, if the limit x belongs to the
projection D of only one forbidden plane P, then we find in the limit tiling only
172 A. Katz

Fig. 19. - A sample of octagonal tiling, decorated with Ammann rules.

one infinite worm (whose axis is the projection of P on Ell) and if we consider
now another sequence x~ converging to x from the other side of D, we find
in the corresponding limit tiling the "flipped" version of the worm. Of course,
the superimposition of these two limits is just the trace of the oblique tiling in
the singular cut defined by x. The axis of the worm is almost a symmetry axis
of the tiling, which is broken only along the worm: if P belongs to the second
kind of forbidden planes, the two limit tilings share the same sets of vertices
and only the vertex decoration along the worm is "flipped" between the limits.
If P belongs to the first kind of forbidden planes, then the limit tiling is defined
by a plane cut which go through the boundary of a family of atomic surfaces,
and the taking of a limit is a consistent way to select half of the concerned
atomic surfaces.
If now the intersection x of the closure of the projection of all the selected
half-planes belongs to several images of forbidden planes, we must be careful
not to "pinch" the cut between forbidden planes upon flipping worms. Recall
that, although we have taken the closure of the projection of the selected half-
planes in order to develop our pushing argument, the cut is not allowed to
intersect the forbidden planes. Observe first that we have to study only a finite
number of different tilings. In fact, such an x is either the intersection of two,
 MATCHING RULES 173

four or eight projections of forbidden planes, and there is only a finite number
of orbits for such a point under the projection of the 4-dimensionallattice. But
recall that two tilings whose associated plane cuts are mapped on each other by
a translation belonging to the high-dimensional lattice are just the same tiling,
up to a global translation which is the projection on Ell of the high-dimensional
translation.

Consider a special tiling for which the point x falls on the intersection of only
e
two projections of forbidden planes. Let be the intersection of these planes
e
(we know that is a face center), and consider the fiber parallel to E.L going
through e. We see that we can flip freely each of the worms associated to the
two forbidden planes. In fact, each choice of passing the cut on one side or the
other of each forbidden plane correspond to a quadrant in this fiber, and the
cut is never pinched. But the situation is quite different when x belongs to the
intersection of four or eight forbidden planes (then eis respectively a vertex or
a body center of the 4-dimensionallattice). Defining one of the corresponding
special tHings as a limit, the traces of the associated sequence of cuts in the
fiber going through eis a sequence of points converging to the intersection of
the traces of the forbidden planes. This sequence of points is contained in one
of the eight or sixteen sectors defined in this fiber and the tiling of the worms
in the limit tiling is prescribed by this sector. We get complete worms only for
the forbidden planes bounding this sector, and we can flip one of them. This
operation just yields the limit tiling obtained in a sector adjacent to the initial
one. Finally, we see that we access only eight different special tilings associated
to fourfold intersectioIl$ and sixteen in the case of eightfold intersections.

Let us now explain in which sense these special tilings are "less" quasiperi-
odic than regular ones. Recall that in the standard computation of the Fourier
transform, we take the multiplicative product of the Lebesgue measure carried
by the plane cut and of the measure carried by the atomic surfaces. But such
a multiplicative product of generalised functions exists only under a transver-
sality condition which is precisely violated when the cut intersects the atomic
surface on its boundary. Thus the computation breaks down and this notion
of quasiperiodicity does not apply to the special tilings involving coming on
the projection of forbidden planes of the first kind. Observe on the other hand
that nothing dramatic happens to the special tilings: since the ambiguous ver-
tices (those on the worm) constitute a vanishingly small fraction of the set of
vertices of the tiling, the statistical notions such as autocorrelation functions,
which define the physically relevant quantities, are exactly the same for special
tilings and for regular tHings.
174 A. Katz

7. GENERALISED AMMANN TILINGS

7.1. Definitions

As already explained, we need two sets of forbidden planes in order to imple-

ment the Ammann decorations of the octagonal tHings: the first set corresponds
to the edge part of the decoration, and the second set to the vertex part. We
shall now study what happens when one keeps only one of these two sets, and
the associated decoration. As we shall see, the corresponding decorated tHings
are still highly ordered: they are either quasiperiodic or periodic, up to some
discrepancies that we shall explain.
We shall call generalised Ammann tHings of the first kind those for which
we keep only the edge decoration to define the matching rule, and accordingly
generalised tHings of the second kind those with only the vertex decoration.
Let us first discuss which symmetry properties may remain for the generalised
tHings, provided that they are ordered enough to keep some symmetry.

7.2. Symmetry considerations

Recall that the octagonal group acting in the plane is a Coxeter group generated
by two reflections, and we may take for instance as generating mirrors of this
action in Ell the two lines spanned respectively by el and el + ea. This group
contains two subgroups isomorphic to the symmetry group of the square, which
is still a Coxeter group. We may take as mirrors generating the first subgroup
the two lines spanned by e1 and e2, and as mirrors generating the second one
those spanned by el + e2 and e2 + e3. We have a similar construction in El..
These two subgroups are linked with the two kinds of forbidden planes in
the following way: recall that we have lifted in JR4 the action of the octagonal
group, so that we have in particular an action of the two subgroups, and we
know how to transform any 2-dimensional plane embedded in JR4 under these
actions: in other words, we have an action of the octagonal group and of its
subgroups on the Grassmannian manifold g~ of the 2-dimensional planes of
JR4.
Then it is easy to see that the planes simultaneously non-transversal to the
four directions of the first set of forbidden planes are invariant under the action
of the first subgroup, and their set is a one-dimensional smooth submanifold
SI of g~. We have a similar result for the second set of forbidden planes and
our second subgroup, which yields another smooth submanifold Sa of g~. The
two curves S1 and Sa intersect on two points in g~. The corresponding planes,
which are invariant through both subgroups, are invariant through the whole
octagonal group and are obviously our Ell and El..
Finally, the canonical cut construction of the octagonal tHings generalises
as follows: setting once for all El. as the plane carrying the atomic surface,
we consider a plane cut non intersecting the first set of forbidden planes. The
direction of such a plane is surely non-transversal to the four directions of for-
 MATCHING RULES 175

bidden planes, so that the direction of the cut belongs to 51. Setting as atomic
surface the projection of the unit hypercube on E.L along the cut, we take the
intersections of the cut with the lattice of these atomic surfaces and project
these points on Ell along E.L to get the vertices of a generalised Ammann tilings
of the first kind. It is made of the same rhombi and squares as the octago-
nal tiling and possesses the symmetry of the square to the same extent that
the Ammann tHings are octagonal. Also, it admits the edge decoration by its
very construction. We set an obviously similar construction for the generalised
tilings of the second kind.
A remarkable property of both 51 and ~ is that these two curves con-
tains infinitely many points corresponding to planes with a completely rational
direction. For such a direction, the cut construction yields periodic tHings.
Otherwise, we get quasiperiodic tilings.
However, we are concerned by the converse problem: we would like to show
that any tiling admitting the edge decoration alone (or the vertex decoration
alone) is (up to some discrepancies for special tilings) a periodic or quasiperiodic
tiling. For the sake of clarity, we shall work out the construction for the case of
the generalised Ammann tHings of the first kind (edge decoration). Also, from
now on, we shall consider as forbidden planes the first family only.

7.3. Setting the method

This problem is clearly more difficult than for the octagonal case. In that
latter case, we had for symmetry r{'..asons only one candidate for the direction
of the plane cut we were looking for inside the homotopy class of cuts defining
a given tiling: namely the direction of Ell. This is no longer the case, since any
direction belonging to the curve 51 of the Grassmannian manifold is a possible
candidate, and this forces us to refine our method.
We shall still describe a genera!" cut by its intersections with fibers parallel
to E.L. Since the generalised tilings are still built by projecting the selected
points on Ell, we may as previously limit our considerations to cuts everywhere
transversal to E.L, (in order to get a one-to-one projection of the cut on Ell
along E.L) so that the cut yields exactly one intersection in each fiber. Then,
since E.L is non-transversal to the directions of the forbidden planes, we define
the same notion of system of data as in the octagonal case: nothing is changed
from that point of view.
Our problem is again to show that a compatible system of data is associated
to a homotopy class of cuts which contains a plane. In the octagonal case, we
mapped the forbidden planes on E.L along Ell, which was the only direction
to be considered, and developed the globalisation argument in E.L. This is the
point we have to generalise in order to take into account that we have now a
whole one-parameter family 51 of possible plane directions for the cut.
Let us consider the 3-dimensional space £1 = E.L X [,1, where [,1 = m is a
local chart of 51, which we define as follows: take anyone of the four directions
of forbidden planes and consider in this plane the traces of the planes belonging
176 A. Katz

to Sl. On account of the non-transversality, these traces are straight lines and
we take as coordinate along Sl the tangent of the angle <p between the trace
of Ell and that of the given plane; because of the symmetry, this definition
does not depend on the chosen direction of forbidden planes (up to a sign).
We restrict the domain S'1 of the chart to angles <p E [-7r/4, +7r/4] because
for larger angles we get overlapping tiles. It is easy to see from elementary
geometry that when we vary <p, the projections on E.L of the basis vectors
keep the same relative angles, the length of the projections of e1 (resp. e2)
remains equal to that of e3 (resp. e4), and the ratio of these lengths varies like
tan( <p + 7r / 4). For <p = ±7r / 4, one or the other pair yields a null projection,
so that the atomic surface become a square and the corresponding tilings are
simple square tHings built with one or the other of our square tiles.
The 3-dimensional space £1 is to be considered as a stack of copies ~I" of
E.L, each attached to the direction in Sl along which we project JR4 on E.L:
the fiber ~o above <p = 0 is the one for which JR4 is projected along Ell, and is
just E.L as we have used it when dealing with the octagonal case.
Now, we define the map
F1 : JR4 x S~ 1-+ £1
which send a pair (~, <p) E JR4 x S'1 on the projection of ~ on E.L along the
direction designated by <p, so that the image is a point in the fiber ~I'" Then,
due to the non-transversality of the forbidden planes to any plane in Sl1 the
image by F1 of any forbidden plane P is a straight line in each fiber ~I'" and
due to the choice of the parameterisation tan( <p) along .e1, these lines align
along a plane in £1 when <p describes Sf. Thus, the image P = FI(P, S1) is a
two dimensional plane in £1.
Let us describe briefly the geometry of this collection of planes. Consider first
the images of two parallel forbidden planes: their traces in each fiber ~I" are
parallel lines, but due to the irrational direction of E.L, the distance between
these lines changes with the projecting direction <p, so that there are no parallel
planes in the collection. Second, observe that two planes P and pI intersect
along a line contained in a fiber ~I" if and only if, first, they come from parallel
forbidden planes, and second, the direction <p corresponds to a rational plane.
Then we find a whole infinite family of planes intersecting on the same line,
coming from an affine sublattice of parallel forbidden planes. Moreover, in such
a fiber ~I" where <p is the coordinate of a rational plane, we find a whole lattice
of parallel lines on which intersect the images of sublattices of forbidden planes:
in these fibers, the traces of our planes build a periodic decomposition of the
fiber. On the contrary, the fibers corresponding to irrational directions in Sl
are densely filled by the traces of the planes, in the way we are used to for the
projection on E.L along Ell, which is now found as ~o·
On the other hand, consider the intersection D of two planes P and pI
coming from non-parallel forbidden planes P and P'. Its trace on each fiber
~I" is the intersection of the corresponding projections of P and pIon E.L.
But since non-parallel forbiq,deh planes are transversal, they actually intersect
 MATCHING RULES 177

in JR4 and this trace is the projection of the intersection ~ = p n pl. Thus the
whole line D is the image .rl(~' S1). In particular, the discussion of section 6.1
applies and we see that such a line D is either the intersection of two planes
p and pI (if ~ is a face center) or of four such planes if ~ is a body center.
Finally, observe that, once again because the direction of E.l. is irrational, for
no pair of lines (D, D') are parallel.

7.3.1. Systems oE data in £1

Our next task is to translate in the space £1 the notion of system of data, since
the globalisation procedure using Helly's theorem will take place in £1. Let us
return in JR4 considered as usual as the trivial fiber bundle E.l. x Ell' in which
we describe a cut by considering for each x E Ell its trace on the corresponding
fiber. Recall that for non-transversality reasons the forbidden planes disconnect
the fibers that they intersect, so that we characterise the homotopy class of a
cut by specifying the relevant connected component in each fiber. Recall also
that the choice made at one point x E Ell relatively to one forbidden plane
is forced (by the non-intersection condition) all along the projection of this
forbidden plane on Ell, so that the choice is in fact attached to each forbidden
plane rather than to points of its projection on Ell. Now we can translate such
a choice in £1 simply by observing that it is quite independent of the direction
cp of the projection: we copy this choice in each fiber 4.>", of £1 and this amounts
to select one of the two half-spaces defined in £1 by the image of the forbidden
plane (observe that the domain S1' of the chart £1 on SI is small enough to
allow this procedure).
The strategy to prove that the edge decoration is sufficient to enforce periodic
or quasiperiodic order (up to something) will as in the octagonal case consist in
showing that a system of data is compatible with its definition by a continuous
cut if and only if the associated half-spaces selected in £1 have a non-empty
intersection.
We shall prove the following:

Theorem 3 The compatible systems of data are those for which the closures
of all selected half-spaces in £1 have a non-empty intersection.

7.4. Reduction to "bad prisms"

In view of the application of Helly's theorem in £}, we have to show that, for a
system of data defined by a continuous cut, any pair, triplet and quadruplet (£1
is 3-dimensional!) of selected half-spaces in £1 has a non-empty intersection.
Observe first that since there is no pair of parallel planes, the intersection of
two selected half spaces is never empty. Next, it is immediate that in order for
a triplet of half-spaces to have an empty intersection, the pairwise intersections
of the corresponding planes must be three parallel lines, in which case each
of them is contained in a fiber 4.>", and the three planes come from parallel
forbidden planes. We shall refer to this configuration as a "bad prism" .
178 A. Katz

Now we want to show that to any "bad quadruplet" of half-spaces, we can

attach a bad prism (just as we have attached a bad strip to any bad triangle
in section 6.4.1). There are several configurations to study and we shall just
sketch the arguments.
Let us classify the possible configurations according to the number of parallel
lines among the intersections of our four planes. If we have four parallel lines,
then the configuration is a prism with a quadrangular basis and any triplet
of planes among the four defines a bad prism. If we have three parallel lines,
then the configuration is a bad prism crossed by the forth plane and we simply
retain the bad prism. The situation with only two parallel lines just does not
exist, so that it remains to examine tetrahedral configurations.
To deal with them, we use two facts: first, considering any edge of the
tetrahedron which is not contained in a fiber ~'P' we know that it comes from
an intersection of forbidden planes. Depending on the intersection and of the
half-planes selected in the corresponding fiber of lR4 , we may be able to find
one or several other forbidden planes going through the same intersection about
which we know which are the selected half-planes (see Fig. 14). These planes
are mapped in el onto planes containing the given edge. Second, considering
a "bad tetrahedron" and any other plane crossing it, we can build another
bad tetrahedron bounded by this last plane (whatever the selected side is).
This allows to get rid of some special configurations, and the inspection of all
possible cases shows that it is always possible to associate a bad prism to a bad
tetrahedron.

7.5. Proof of the theorem

The reduction to bad prisms allows us to work out the proof in a planar section
of el . Suppose for instance that the direction of forbidden planes involved in
the bad prism is (el' e2 - e4)) and consider the plane in el spanned by £1
and the direction e~ (or ~ + e~) in EJ... In this plane, the trace (and the
projection) of the bad prism is the bad triangle depicted on Fig. 20. The
vertices of the triangle project on £1 on 'PI. 'P2, 'P3 such that 'PI < 'P2 < 'P3
(inequalities are strict because there is no "vertical" plane P in e1 : such a
plane would be the image of a forbidden plane located at infinity in lR4 ). Let
us choose a cp Ej'P}' 'P2[ (or, equivalently 'P Ej'P2, 'P3[) corresponding to an
irrational plane (in order to avoid superimposition of projections of forbidden
planes) and consider the situation in the fiber ~'P (Fig. 21): whatever the choice
of cp is, the forbidden planes whose images in e1 build the bad prism project
on two bad stripes (PI, P2 ) and (P3 , P2 ) to which we can apply our pushing
procedure.
This is the key of the proof: in the octagonal case, we made use of only one
bad strip, but we had two rationally independent distances to push a forbidden
plane along it, using as the "pusher" two forbidden planes of the two different
kinds. Now, we have only the four forbidden planes of the first kind, so that we
have only one choice for the "pusher" and the whole procedure generates only a
 MATCHING RULES 179

Fig. 20. - The traces of the planes Pi, P2 and P3 forming a bad prism define a bad
triangle in the relevant plane of &1.

v V' V"

Fig. 21. - The trace of the bad prism in any fiber <1>'1' with CPl < cP < CP3 contains
two bad stripes, so that the pushing procedure allow to push the plane V on two
distances d and d'.

half one-parameter sublattice of forbidden planes "about which we know which

half-plane is selected" . Recall that this does not result in any contradiction,
for which we need a suitably oriented cone of planes. This difficulty is solved in
the present framework by the occurrence of two bad stripes in the construction,
which together allow to recover a whole cone of planes.
In fact, there is nothing more to say about the pushing procedure. The only
point to be studied is the relative position of the cone of planes with respect
180 A . Katz

__ ~~ __ ~ ________________ ~3

........ :, ..... .... ................ .

.. . ..... . \.,,: . .... .................. .
~

t?i~,l~ ~ ::~ ~ ~ ~1>; :~;.~ ~ ~ ~ ~ ~ ~ j ~ j ~ ~ ~ ~ ~ ~

~.~3: >'<, ~; ::::~ ~ ??~<~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~

•! ~ti,[li-";;;".,!">.,'"
::::
, ,
: :: ::::: >;;
: : : : : : ::'~,~
j,"i
: .:: : ...... :
.................... ~;,' ..... , . . . 1.'
",

Fig. 22. - The three forbidden planes Pl, P2 and P3 are located in ]R4 in positions
corresponding to the slope of their images P l , P 2 and P 3 in £1, in such a way that
in projection on Ell' P2 falls between Pl and P3 •

to the pair (E,L, Ell)' For this, let us first consider simultaneously the two
figures Fig. 20 and Fig. 22, which represents through its trace in the plane
(e3, e2 + e4) C R4 the lattice offorbidden planes (eb e2 - e4) among which are
the three planes PI. P2 , P3 forming our bad prism.
In order to correlate the lines in Fig. 20 and the corresponding points in
Fig. 22, observe that the slope of the lines in the first figure corresponds to the
distance to E,L of the corresponding vertices in the second one: the vertices
whose projection on E,L moves quickly with the direction cp of the projection
(and which corresponds to large slopes on Fig. 20) are those which are far from
E,L on Fig. 22. Thus we see that the projections on Ell along E,L of the vertices
PI. P2 , P3 are such that the projection of P2 lies between those of PI and P3 •
This piece of information will be sufficient for us. Let us return to Fig. 16 to
observe that, since our "pusher" makes an angle of 7r / 4 with both the pushed
plane V and the direction of the bad strip, the components of the displacement
in the coordinates (el' e2 - e4) are just twice the components of the width of
the bad strip in the coordinates (e3, e2 + e4), so that the basis of our cone of
planes is obtained by copying in the plane (el' e2 - e4) of Fig. 17 the two vectors
(PI - P2 ) and (P3 - P2 ) read on Fig. 22, and doubling them. Then the remark
above entails that this cone contains E,L, and thus that the vertices it contains
project densely on Ell, which was the point to be proved.
 MATCHING RULES 181

The end of the proof runs exactly like in the octagonal case and does not need
details: the pushing procedure exhibits a contradiction with the continuity of
the cut, so that a compatible system of data does not have any bad prism, and
applying Helly's theorem then concludes the proof of theorem 3.

7.6. Order in generalised Ammann tHings of the first kind

Let us now turn to the consequences of this theorem for the ordering of the
well-decorated tilings. As for the octagonal case, we shall distinguish between
the generic situation and the special tilings.
The generic case is such that the intersection of the (closed) half-spaces de-
fined by the system of data does not intersect any image P of a forbidden
plane, so that it is also the intersection of the corresponding family of open
half-spaces. Such an intersection is a connected component of the complemen-
e
tary in 1 of the images P = F1 (P, S1) of all the forbidden planes P. Let
us describe these connected components. The main point is that all of them
are entirely contained in a single fiber cP"". In fact, the collection of planes P
contains planes as close to the fibers as we want (coming from forbidden planes
as far as we want in IR4) so that any two points in different fibers can be sep-
arated by such a plane and do not belong to the same connected component.
Now there are two cases: in fibers corresponding to irrational directions, the
traces of the planes P are dense and the connected components are reduced
to single points. In fibers corresponding to rational directions, the planes P
fall on four discrete grids which generate a periodic decomposition of the fiber
(which we refrain from calling a tiling for the sake of clarity). In both cases, the
system of data may be defined by a plane cut parallel to the plane designated
by the coordinate rp of the fiber cPcp which contains the intersection, and going
through the corresponding point of E-.l (in the irrational case) or anywhere in
the interior of the corresponding cell (in the rational case). The correspond-
ing generalised Ammann tHings are respectively quasiperiodic and
periodic, as illustrated on Fig. 23.
Concerning the special tilings, it is still true that they correspond to inter-
sections of closed half-spaces entirely contained in a single fiber cP"", but now
the situation is rather different when rp corresponds to a rational direction and
when it corresponds to an irrational one.
If rp corresponds to an irrational direction, the discussion of section 6.5 ap-
plies completely: we find a finite number of infinite worms, which we can flip
according to a non-pinching restriction. Averaged physical quantities do not
distinguish between quasiperiodic and special tilings corresponding to the same
rp.
The family of special tHings is paradoxically much more larger in the periodic
case, where rp correspond to a rational direction of the cut. As mentioned above
the forbidden planes project on such a fiber cP"" on the boundaries of a periodic
cellular decomposition of the plane, in which each cell corresponds to a periodic
tiling. Of course, two cells mapped on each other by the translation group of
182 A. Katz

Fig. 23. - Two examples of generalised Ammann tilings of the first kind: (top) pe-
riodic tiling corresponding to cp = arctan(3 - 2\1'2); (bottom) quasiperiodic tiling
corresponding to cp = 15°.

the periodic decomposition of ~rp define the same tiling, so that we find in
fact finitely many different periodic tilings associated to the angle cp. Let us
consider the special tilings associated to an edge of a cell: this line is the image
of a whole sublattice of parallel forbidden planes. They are (with one exception
to be dealt with below) associated with a whole lattice of worms which we can
flip independently. If we start from a regular tiling associated with the interior
of a cell, and we flip all the worms associated with a given edge of this cell, we
end of course with the regular tiling associated with the adjacent cell. But we
can flip any arbitrary subset of worms, still keeping a well-decorated tiling.
If now we come on a vertex of a cell, there are again two cases: if this point
is the image of a face center (in which case this vertex is the intersection of
 MATCHING RULES 183

two orthogonal lines and belongs to four cells), then we find two perpendicular
lattices of worms which do not "interact", in the sense that here again we
can flip any arbitrary subset of both family of worms, still keeping a correct
decoration. When the vertex is the projection of a body center, the situation
is much more complicated: let wand w' be the two families of relevant worms,
which form an angle of 7r / 4. It is obviously possible to flip an arbitrary subset
of either w or w', but let us try to combine flips from the two families: let
us flip first a worm from w. This operation breaks all the worms of w' on
their intersections with the flipped one and produces pieces of worms of a third
direction w", but these are broken on their intersections with all the other
worms of w. Thus it seems that when one flips one worm, only the worms
parallel to this one remain complete. It is unclear in this case whether it is
possible or not to find well decorated special tilings besides the above-mentioned
ones.
Finally, let us say a word about the strangest tiling of all, which is obtained
for tp = ±7r/ 4 and has no worms at all: it is the regular square tiling! Without
insisting, let us mention that it is possible to produce a well decorated special
tiling associated to it by taking a limit of tiling along a sequence with tpn -- 7r / 4,
depicted on Fig. 24.
As a conclusion on this section, let us remark that the notion of matching
rules (even strong like these) seems not to fit so easily with the notion of
(quasi)periodicity. Everything goes smoothly as long as regular tHings are
concerned. But we cannot escape the special tilings (recall that they are locally
the same as regular ones, so that no local matching rule can reject them), and
in the periodic case, special tilings may differ from a periodic one for a finite
fraction of their vertices, so that physical averaged quantities may reveal that
they are less ordered than periodic ones.

7.7. Generalised Ammann tilings of the second kind: an example

of weak rules

One can develop the same theory with the vertex decoration and the second
family of forbidden planes. We shall not repeat the arguments, leaving them
as an exercise to the reader. Let us only set up the framework.
We need now an auxiliary space £2 = E.L X £2, where £2 is a local chart on
~ defined in a quite similar way as £1: considering the traces of the planes
belonging to S2 on the direction of any forbidden plane of the second kind, we
observe that these traces are lines so that we define the coordinate 1/1 of a plane
in S2 as the angle between Ell and the given plane of ~, and we set on £2
the parameter tan(1/1). Notice that now the domain Sf1 of the chart is limited
to 1/1 E [-7r/8, +7r/81, in order to avoid overlapping tiles. Here again, simple
geometry shows that when we vary 1/1, the lengths of the projections on E.L of
the basis vectors remain the same, but the pair (e2' e4) rotates relatively to the
pair (el, e3) of twice the variation of 1/1. For the extremal values 1/1 = ±7r/8, the
two pairs become collinear and we get as atomic surface a square whose edges
184 A. Katz

~1t t');
~)/, ;'\~
t;(y. ~~
t;(y. ~~
t;(jI ~),
~)I, ~
j" (t
~~ ~)/,
~~ t!~

Fig. 24. - Special tiling associated to the regular square tiling. It is obtained by
taking a limit for cp -+ 11'/4, such that the limit cut goes through a vertex of the
periodic decomposition of the fiber <1>,../4. If the limit cut goes only through an edge,
one of the worm is sent to infinity, and of course if the limit cut goes through the
interior of a cell, we get the regular square tiling.

are the sum of the projections of two basis vectors. The existence domains of
two rhombi over the four vanish, so that we get tilings made of the two squares
an two of the rhombi. These tilings are periodic.
We show on Fig. 25 some examples of generalised Ammann tilings of the
second kind.
The important point is that the matching rules thus obtained are only weak
rules, in the sense of Levitov ([I]).
To elucidate this point, let us return to the very beginning of our theory
(section 4.3). By its very definition, the forbidden set touches the atomic sur-
faces on their boundaries. But the forbidden planes of the first kind on which
we have retracted this forbidden set no longer intersect the atomic surfaces.
However, we have completely forgotten about atomic surfaces during the de-
velopment of the theory, dealing only with the forbidden planes. In fact, our
rules for the octagonal tilings and for the generalised Ammann tilings of the
first kind are strong rules only if the following property holds: it is not possible
to distort a cut so as to switch an intersection point from a given atomic surface
to a neighbouring one, without intersecting a forbidden plane.
To verify that this the case for the first family of forbidden planes and not
for the second, the simplest way is to refer to the "closeness property" of the
atomic surfaces. This property plays an important role in the geometry of
quasicrystals ([28],[19]), but we do not need to explain it in the large and we
shall only make use of the following simple argument:
Consider the lattice of octagonal atomic surfaces, appended by there centers
 MATCHING RULES 185

Fig. 25. - Two examples of generalised Ammann tilings of the second kind: top: the
periodic tiling obtained for 1/J = 1r /8; bottom: a quasiperiodic tiling corresponding to
1/J = 10°.

to the vertices of the four dimensional simple lattice 7/.. 4 . It is possible to

"complete" these atomic octagons in order to get a manifold without boundary.
This is done in two steps. The first one consists in gluing along the edges of the
octagons four sets of rectangles of the type {ei, e2 + e3 + e4}. The two edges
parallel to e~ are glued on two atomic surfaces, and we are left with the four sets
of edges of the type e2 + e3 + e4. But it is easy to see that these segments form
the edges of a lattice of small octagons contained in planes parallel to Ell. We
just add these octagons as our second step and we are done: the "completed"
atomic surfaces have no longer any boundary.
In fact, this procedure is interesting because it changes nothing to the con-
struction of the tiling: the new pieces are non transversal to Ell, so that a
186 A . Katz

Fig . 26. - A (slightly metaphoric) description of the closure of the atomic surfaces
of the octagonal tiling: the first step consists in gluing rectangles to the edges of the
main octagon in El.; the second step closes the surface by adding a small octagon
contained in Ell' The main octagons in El. are disconnected from each others by the
forbidden planes whose traces are drawn on the figure.

generic plane cut never intersect them. The geometry of the complete atomic
surface is illustrated on Fig. 26.
Now, we are lucky, because the forbidden planes also intersect the completed
atomic surfaces in a non-transversal way, namely along line segments. On
Fig. 26, one can see that, first, the four forbidden planes of second kind in-
tersect the original atomic surface (the large octagon) along the diagonals and
then cross the small octagon also along the diagonals while, second, the four
forbidden planes of the first kind divide in two the rectangles before crossing
the small octagon. All eight planes intersect on the center of the small octagon
which is located on the body center of the lattice (the center of large octagons
being on the vertices of this lattice).
Thank to this highly non generic configuration, it is easy to discuss our
question. Consider a cut intersecting the atomic surface of Fig. 26 inside the
large octagon, and let us try to move this point: its trajectory defines a line
on the atomic surface and it is quite obvious that it is impossible to make such
a line between two large octagons without intersecting one of the forbidden
 MATCHING RULES 187

Fig. 27. - The vertex rules are only weak: it is possible to Hip a hexagon while
keeping a well-decorated tiling; however, such a Hip does not open new possibilities
besides the reverse ftip .

planes of the first kind, whose traces surround completely the large octagon.
This confirms that the rules involving the forbidden planes of the first kind
are strong ones (for regular tilings). But now consider the second family of
forbidden planes: we see that their traces on the completed atomic surfaces
build closed polygons made of four pairs of segments joining the center of a
large octagon to one of its vertices, then to the center of the small octagon
sharing this vertex. We see that a cut constrained not to exit of this polygon
may wander from one large octagon to a neighbouring one, but cannot go
further. Thus the matching rules for the generalised Ammann tilings of the
second kind are only weak: given such a regular tiling, it is possible to flip
some hexagons without violating the rule. Such a flip creates a new hexagon,
but you cannot flip it without violating the rule: a cut compatible with the
forbidden plane may only oscillate between two and only two atomic surfaces.
This is illustrated on Fig. 27, where one can see that the characteristic feature
of the "flippable" hexagons is that their two tips (right angles) bear the same
trace of the large Ammann arrow (a square), which is not always the case for
tips occupied by a pair of rhombi.
188 A. Katz

8. CONCLUSION

Let us first stress that nothing in the methods we have explained is limited
to the two-dimensional case. For instance, the same ideas work for devising
matching rules for the 3-dimensional (icosahedral) Penrose tilings (for which
the method of forbidden planes was initially developed).
Now, one can draw several conclusions from this theory. The main one is
that (at least in the author's mind ... ) the existence of matching rules is no
longer mysterious. But this approach solves also a long-standing conjecture
in this field, which was that matching rules are intrinsically linked with self-
similarity. The case of generalised Ammann tHings shows that this is not the
case: for almost all cp (along St) and 1/1 (along ~), the quasiperiodic tilings do
not have any self-similarity property. On the contrary, the real geometric key
to matching rules seems to reside in non-transversality properties.
In the present framework, non-transversality is involved in three different
ways. First, the direction of the cut must be non-transversal to a sufficient
number of lattice subspaces. This corresponds to the existence of worms and
is absolutely inherent to this approach. Second, we have described the cut
through its intersections with fibers (here parallel to E.d which intersect non-
transversally the forbidden planes, so that they are disconnected by these in-
tersections. This allows a straightforward description of the homotopy classes
of cuts. This context could perhaps be slightly relaxed: in situations where
there does not exist a common fiber simultaneously non-transversal to all the
forbidden planes, one could perhaps work with several fibers, each of them
being non-transversal to a sufficient subset of forbidden planes.
The third way non-transversality is involved concerns the shape of the atomic
surfaces: as we have seen, they must be bounded along the intersections of their
carrier (here E.d with the forbidden planes. This is the second condition to
get worms, and is also inherent to this approach. In view of the structure
determination of quasicrystals (which amounts essentially to construct a set of
atomic surfaces) this constraint is of primary importance. In fact, if you trust
that matching rules have something to do with real quasicrystals, you can
restrict your search for atomic surfaces to a limited class of polyhedra, which is
of course an enormous restriction. Thus, besides its own geometrical interest,
we may hope that this theory will help solving the structure of quasicrystals.
 MATCHING RULES 189

References

[1) L. S. Levitov, Commun. Math. Phys. 119 (1988 ) 627.

[2) A. Katz, Commun. Math. Phys. 118 (1988) 263.
[3) F. Gahler, Journal of Non-Crystalline Solids 153 & 154 (1993) 160.
[4) J. E. S. Socolar, Commun. Math. Phys. 129 (1990) 599.
[5) R. Penrose, Mathematical Intelligencer 2 (1979) 32.
[6) N. G. de Bruijn, Nederl. Akad. Wetensch. Proc. Ser. A 43 (1981) 39.
[7) H. Bohr, Acta Math. 45 (1924) 29.
[8) H. Bohr, Acta Math. 46 (1925) 101.
[9) H. Bohr, Acta Math. 47 (1926) 237.
(10) A. S. Besicovitch, Almost periodic functions, Cambridge University Press,
Cambridge, (1932).
[11) P. Bak, Scripta Met. 20 (1986) 1199.
[12) A. Janner and T. Janssen, Phys. Rev. B 15 (1977) 643.
[13) P. M. de Wolff, Acta Cryst. A30 (1974) 777.
[14) A. Katz and M. Duneau, Journal de Physique 47 (1986) 181.
[15) C. Oguey, M. Duneau and A. Katz, Commun. Math. Phys. 118 (1988) 99.
[16) P. Kramer, J. Math. Phys. 29 (1988) 516.
[17) B. Griinbaum and G. C. Shephard, THings and Patterns, W. H.Freeman,
San Francisco, (1987)
[18) Beenker F. P. M., Algebraic theory of non-periodic tHings by two simple
building blocks: a square and a rhombus (Eindhoven, TH-Report 82- WSK-
04,1982).
[19) A. Katz and D. Gratias, in Lectures on Quasicrystals (Aussois 1994),
edited by F. Hippert and D. Gratias, Les Editions de Physique, Paris
(1994).
[20] L. Danzer, Discrete Math. 76 (1989) 1.
(21) P. Stampfli, Helv. Phys. Acta 59 (1986) 1260.
(22) E. Zobetz, Acta Cryst. A48 (1992) 328.
[23] J. E. S. Socolar, Phys. Rev. B 39 (1989) 10519.
(24) R. Klitzing, M. Schlott mann and M. Baake, Int. J. Mod. Phys. B 7 (1993)
1455.
(25) R. Klitzing and M. Baake, Journal de Physique I 4 (1994) 893.
(26) T. Kupke and H. R. Trebin, Journal de Physique I 3 (1993) 564.
(27) R. Ammann, B. Griinbaum and G. C. Sherphard, Discrete Comput. Geom.
8 (1992) 1.
[28) P. A. Kalugin, Europhys. Lett. 9 (1989) 545.
 COURSE 7

A mechanism for diffusion in quasi crystals

P. A. Kalugin

Centre de Physique Theorique, Ecole Poly technique

91128 Palaiseau Cedex
France.

Within ten years after their discovery quasicrystals (QC) evolved from poor
quality micron-sized grains to thermodynamically stable materials with struc-
ture coherence length as large as that in good crystals. The X-ray diffraction
experiments give us today the intensities of hundreds of independent reflec-
tions. This would largely suffice to determine the structure of any crystalline
metallic alloy, but our knowledge of the structure of QC is still far from being
complete. Whatever the reasons for this situation may be, it might be useful
to have an alternative source of structural information. The atomic diffusion,
being strongly structure dependent, is of great interest from this point of view
[1].
According to present views, there are several mechanisms for bulk self-
diffusion in solids. All of them contribute to the diffusion, but as a rule one
mechanism dominates over the others, depending on the structure of the mate-
rial. For instance, in metals diffusion is mostly due to the presence of vacancies
in the lattice, while in silver halogenides the interstitial diffusion dominates over
all others. No matter what the mechanism of diffusion in crystals is, it always
requires the presence of defects. The energy cost of these defects suppresses,
by virtue of Arrhenius law, the diffusion rate at low temperatures:

D = Doexp( -EfT) (1)

The activation energy E in Arrhenius law includes two terms: one is the energy
of creation of a defect, and the other is the activation energy for its displace-
ment. For pure metals with simple lattices both terms are of order (1-1.5) eV

(*) Permanent address: Landau Institute for Theoretical Physics, Academy of

Sciences, 117334 Moscow, eEl
192 P.A. Kalugin

and the total activation energy is (2-3) eV [2]. Thus, the diffusion in these
materials is hardly observable at room temperature.
In the case of quasi crystals the mechanism of diffusion can be based on a
completely different principle than the diffusion of vacancies. The point is that
the presence of vacancies is not necessary for atomic jumps in QC, where the
jumps are possible even without defects [3, 4, 5]. Moreover, the presence of a
finite fraction of atoms which are able to jump to a neighbouring equilibrium
position is of crucial importance for the very existence of QC. These jumps
are related to the so-called phason degree of freedom, which is intrinsic for
any incommensurate structures, to which belong quasicrystals. This degree of
freedom should not be frozen at least at high temperatures, close to the melt-
ing point. This requirement is obvious for the random model of quasi crystals
[7, 8]' for which QC are thought to be stabilized entropically. In fact, it is
reasonable to speak about the entropy related to the fluctuations of the pha-
son degree of freedom only if it takes part in real dynamics. The situation is
more complicated for the deterministic models of QC. Following these models,
the quasiperiodic structure is stabilized by atomic interactions, which impose
some sort of matching rules for each atomic species. It is known, however,
that matching rules are by no means growth rules i. e., they cannot enforce the
defect-free growth of QC. In particular, there should be finite phason gradients
frozen during the solidification because of the anisotropy of the growth. Nev-
ertheless, no phason stress is observed in as-cast specimens. This should mean,
that this stress is annealed during the the solidification i. e., at temperatures
close to the melting point [6].
All the aforesaid allows us to conclude that the activation energy for the
atomic jumps which are related to the phason degree of freedom in quasicrys-
tals should not be high. In what follows, the scenario for diffusion based on the
dynamics of these low-energy defects is constructed. For the sake of simplicity
we shall restrict our consideratioFls to the tutorial structure model of QC -
the octagonal tiling. This tiling covers the plane by squares and 45°-rhombi
with no gap or overlap. There exist two versions of the octagonal tiling: the
deterministic one, or Ammann-Beenker tiling [12], and the random octagonal
tiling [11]; they describe, respectively, the structures of deterministic and ran-
dom quasicrystals. In atomistic models, the vertices of the octagonal tiling are
usually treated as atomic positions. Consider now the hexagonal pattern which
occurs frequently in the octagonal tiling (Fig. 1). The internal vertex in the
hexagonal cage can occupy anyone of two allowed positions, which are sepa-
rated by the distance (V2 - 1) times the minimal inter-atomic spacing. The
jump of the central atom from one equilibrium point to the other constitutes
the elementary step in the phason motion. The fraction of vertices of this kind
can be easily calculated in the case of Ammann-Beenker tiling, and it appears
that (V2 - 1), or 41% of atoms are in such "ambiguous" positions!
Although jumping back and forth, the central atom on Fig. 1 remains en-
closed into the hexagonal cage. The situation may be, however, more compli-
cated, as shown on Fig. 2. The sequence of eight elementary steps shown here
 DIFFUSION IN QUASICRYSTALS 193

Fig. 1. - The elementary step in the dynamics of the octagonal tiling.

returns the tiling to its initial state, but three atoms inside the octagonal cage
permute their positions. The possibility for atoms to be permuted is necessary,
but not yet sufficient for real transport of matter; in order to describe diffu-
sion we need a sequence of steps, allowing us to push certain atoms on long
distances. It appears, however, that there exist quasiperiodic structures, for
which the dynamics of Fig. 2 does not give rise to permutations of atoms, and,
hence, to any diffusion. For instance, consider the same standard octagonal
tiling, but now place the atoms on the centres of the rhombi and squares in-
stead of on the vertices. These points correspond to the vertices of the dual
grid [10], and cannot be permuted in any way. There displacement over a long
distance requires a phason shift of the same order of magnitude as the dis-
placement. Clearly, this means that the dynamics of Fig. 2 does not give rise
to any diffusion in this case. Thus the question arises: what is the property of
the quasiperiodic structure responsible for diffusion? For the models of random
tiling the situation is much complicated, and we try to answer this question for
the deterministic models of QC.
In the deterministic models the structure of quasicrystal is described by a
set of atomic surfaces periodically arranged in the "superspace" of high di-
mension. The positions of atoms are given by the points of intersections of
this set with the physical subspace Ell. For the octagonal tiling the physical
space is two-dimensional, while the atomic surfaces form a periodic lattice in
the four-dimensional space Ell E9 E.l, where E.l is called "perpendicular" or
"phason" space. As we have seen, the main difference between the model with
central positions of atoms and the model with atoms placed at the vertices is
that in the first model no permutation of atoms is possible. This property is
closely related to the topology of the atomic surfaces for these two models. In
order to clarify this point, it is convenient to "complete" the atomic surfaces
according to [3, 4, 5], by parts non-transversal to the physical space, in a way
which respects the jumps of atoms. This transforms the atomic surfaces into
194 P.A. Kalugin

a set of manifolds without boundary. Within this framework, it is clear that

two points of the quasiperiodic structure can be permuted if and only if they
belong to the same connected component of the above-mentioned manifold.
For the model of centers the atomic surfaces, after being completed, are still
disconnected and different centres of tiles correspond to different connected
components. On the other hand, the atomic surfaces for vertices (after being
completed) are all glued in only one connected component, such that any two
points on them can be connected. As a consequence of this topology, some ver-
tices undergo a permutation when the phason coordinate of the whole structure
makes a loop (see Fig. 3). As an example one can consider the triple of internal
vertices on Fig. 2. Thus, we have formulated the geometric criterion for the
possibility for the quasi crystalline mechanism of diffusion to exist. Fortunately,
all existing realistic models for the structure of deterministic QC have atomic
surfaces which allow permutations of atoms.
Let us now describe in more detail the model to be used in the study of
diffusion in QC. Consider some sort of matching rules for Ammann-Beenker
tiling, but now place the system in a thermostat at temperature T [9J. What
this means, is that the matching rules can be violated, but any defect in them
costs some energy Ed, and the weight of the corresponding configuration in the
Gibbs ensemble is lowered by the factor exp( -EdiT). The advantage of the
proposed model is that it combines random and deterministic quasicrystals.
Indeed, at zero temperature no violations in matching rules is allowed and only
perfect Ammann-Beenker tilings are present in the Gibbs ensemble. On the
other hand, in the limit of infinite temperature the weights of all tHings become
equal, which corresponds to the random octagonal tiling.

@@
@ \(3)
(2)

\ 3

~ @
~~
Fig. 2. - The sequence of local rearrangements of the octagonal tiling, giving rise to
the permutation of vertices 1, 2 and 3.
 DIFFUSION IN QUASICRYSTALS 195

Fig. 3. - A closed loop of the phason coordinate in EJ. may give rise to the displace-
ments of atoms. Three cases are illustrated: a) atomic surfaces which contain the
loop completely correspond to atoms which do not jump at all; b) atomic surfaces
for which the loop crosses only a single edge correspond to isolated jumps, the atom
going back and forth when the loop crosses the edge; c) the interesting phenomenon
occurs in the third case, when the loop encloses a vertex of (an octuplet of) atomic
surfaces: after one pass, a triplet of atoms is permuted, as the result of the sequence
of rearrangements shown on the Fig. 2. (In the real four-dimensional situation eight
atomic surfaces are connected around each vertex, and the direction of the cut, shown
here as a dashed line, is two-dimensional; this is only a metaphoric illustration.)

As already noted, the dynamics of Fig. 2 does not correspond to transport

of atoms on long distances. Thus, we have to construct a more complicated
sequence of steps, allowing unbounded displacements of atoms. Because any
isolated atomic jump creates defects in the matching rules, these displacements
require overcoming the corresponding energy barrier. Thus, the "wandering"
atom must be surrounded by a finite number of violations of matching rules.
Obviously, the lowest possible energy of this cloud of defects which is still
sufficient to mobilise the atom, plays the role of an excitation energy for this
diffusion process. In the case of non-interacting defects in matching rules,
this energy is determined by its number. This number strongly depends on
the definition of what the defects in matching rules are, so we only roughly
estimate it from above, and thus demonstrate that a finite number of defects
is sufficient.
As may be seen from Fig. 3, the proposed diffusion mechanism involves pha-
son fluctuations. Although these fluctuations must be local, we first solve the
following auxiliary problem just in order to estimate their amplitude: what
is the amplitude of fluctuations of the global phason coordinate, which is suf-
ficient for displacements of atoms on long distances? In other words, let us
196 P.A. Kalugin

study the rearrangements of atoms under a global uniform phason shift. For
the Ammann-Beenker octagonal tiling, the atomic surface (or "acceptance do-
main") is a perfect octagon with an edge length equal to 1 [6]. First, suppose
that the amplitude of the phason shift (i. e., the size of the loop) is very small.
In this case the only atoms which undergo permutations are those lying close
to the vertices of the acceptance domain (see Fig. 3). The corresponding per-
mutations are shown on Fig. 2. When the amplitude is increased, the density
of the permuted triples also increases. After passing a first threshold (which is
easily seen to allow loops large enough to encircle two neighbouring vertices of
the octagon) the permutations become more complex because certain vertices
are shared by two or more triples. In particular, the vertices on the boundary
of the octagon on Fig. 3 are now brought in motion, so that they exchange
their positions with the inner vertices. The cluster of permuted atoms become
larger and larger, and at some critical value of the phason magnitude, an in-
finite cluster appears: this is quite analogous to the crossing of a percolation
threshold, but instead of random media this percolation occurs in a regular,
though quasiperiodic environment. The percolation threshold depends on the
shape of the region in E.L where the phason shifts are allowed; keeping in mind
the qualitative character of the consideration, we calculate its value for a re-
gion having the shape of perfect octagon. In this particular case the infinite
percolation cluster is a self-similar fractal set, which simplifies the calculation.
It appears that the percolation occurs if the region where the phason shifts are
allowed is J2 smaller than the acceptance domain. The corresponding infinite
percolation cluster is shown on Fig. 4.
Global uniform phason shifts, being unphysical, cannot be considered as a

Fig. 4. - A part of the critical infinite cluster of positions available for one atom
when the whole structure undergoes phason shifts of critical amplitude
 DIFFUSION IN QUASICRYSTALS 197

realistic mechanism for diffusion. One can note, however, that the motion of
any given atom depends only on the rearrangements in its nearest vicinity. In
principle, this fact allows us to construct a non-uniform phason shift, which
has a proper magnitude near a given atom and decreases to zero at some finite
distance. When the distorted region is taken too large or too small, the energy
of distortion increases, because of the increasing number of defects in one case
and because of stronger distortion of the structure in the other. Thus, there
exists an optimal size for the distorted region, which depends on the details
of the structure. In the case of the octagonal tiling under consideration, this
size is determined by the requirement of "non-destruction" of tiles. Indeed, as
only tHings with no gap or overlap are included into the Gibbs ensemble, it
is implicitly assumed that other tilings have infinite energy. Thus, we have to
create a local phason distortion having a phason gradient as large as possible
although still compatible with the "non-destruction" condition.
The definition of what the phason gradient is when only a finite region is
distorted is ambiguous, because of the discrete character or the distortion. It
is possible, however, to define the phason coordinate of the octagonal tiling
as a continuous variable, at least formally. In order to do this we use the
dual grid description of the octagonal tiling. More specifically, instead of the
tiling itself one can consider its dual graph [10). It is known that this graph
can be represented by four beams of equally spaced parallel lines. The angles
between different beams are all multiples of 45 0 • The spacing between the lines
in a beam is determined in such a way that the density of the vertices in the
dual graph is equal to the density of tiles in the original tiling; this spacing
is equal to 2. These lines can also be considered as the lines of intersection
of the physical space Ell and four grids of parallel 3D subspaces in E.l EEl Ell'
Now, we can continuously distort the physical space Ell, as shown on Fig. 5,
in order to have a given value of the phason coordinate at the centre of the
distorted region. The phason gradient, as discussed above, cannot be large,
because otherwise the resulting grids do not correspond to any tiling. The
most restrictive condition for the phason gradient comes from the fact that
the signs of intersection indices for grids should not change. This condition
is fulfilled when the absolute value of the phason gradient is smaller then the
tangent of the smallest angle between Ell and any of the 2D spaces appearing
in intersections of 3D grids. For the octagonal tiling this tangent is equal to
Y2 - 1. On the other hand, the critical amplitude of phason fluctuations, is as
mentioned above Y2 smaller than the "radius" of the acceptance domain i. e.,
is equal to (.../2 + .;'2)/2. This means that if the radius of the distorted region
on Fig. 5 is equal to (1 + 1/ .;'2)3/2, the phason coordinate at the centre can
reach its critical value, without creating too strong a phason gradient.
Now, we can explicitly construct a sequence of local rearrangements, giving
rise to the displacements of certain atoms on long distances. As was already
shown, global phason shifts with amplitude greater than or equal to the crit-
ical one can give rise to such a displacement. If we consider now only one
"travelling" atom, we can restrict the phason shift to a neighbourhood of it, as
198 P.A. Kalugin

shown on Fig. 5. This means that the phason coordinate h of the vertex of the
cone on Fig. 5 is equal to the current value of the global phason shift. On the
other hand, the physical coordinate of the centre of deformation must follow
the moving atom. In the dual-grid interpretation this means that the vertex
of the cone on the Fig. 5 must belong to the cell corresponding to the atom
under consideration. As the distorted region is finite, it contains a finite num-
ber of violations of matching rules. The exact value of this number depends on
the kind of matching rules considered. One can, however, estimate the total
number of atoms brought in motion in this process. In order to do this, let us
associate each vertex of the octagonal tiling with a tile in the following way.
Select a direction in the plane of tiling, which is non-transversal to the edges
of tiles. Then associate a given vertex with the tile lying in that direction
and containing this vertex. As can be easily seen, this establishes a one-to-one
correspondence between the vertices of the octagonal tiling and those of the
dual grid. Note now, that no more than three lines out of any beam of straight
lines constituting the dual grid can intersect the distorted region. Thus, for
any given pair of beams, there are no more than 9 intersection points inside the
distorted region . As there are 6 such pairs, the total number of points brought
in motion is bounded from above by 6·9 = 54. It should be remembered,
however, that this is only a rough estimate. The important point is that there
exists a critical number nc of violations of matching rules, sufficient to make
atoms mobile.
Consider now the temperature dependence of the diffusion coefficient under
the assumption that the diffusion is mostly due to the proposed mechanism.
At low temperature, the diffusion involves only the processes with the lowest
excitation energy. In the model under consideration this is the energy of the
cloud of nc defects in matching rules, to which one should add the activation
energy for each atomic jump f.j. From the previous considerations, it is obvious

Fig. 5. - The distortion of the physical space localized in a finite region. The ampli-
tude of distortion reaches the value h at the centre of the region.
 DIFFUSION IN QUASICRYSTALS 199

log D

------- 1/ T

Fig. 6. - The expected shape of Arrhenius plot: the activation energy E (1) varies
from fcj + ncfcd in region 1 to fcj in region 2.

that the number nc and the activation energy take the same values throughout
the sample. This means that the diffusion process is single-activated, and,
being local, must obey the Arrhenius law (1). The total activation energy E in
(1), under the assumption of non-interacting defects in matching rules, is given
by

where Ed is the energy of creation of one defect. At higher temperatures, the

behaviour of the diffusion coefficient deviates from the Arrhenius law. There are
two reasons for this deviation. First, when the temperature increases, processes
involving a number of violations of matching rules larger than nc contribute to
the diffusion. As discussed below, they can have larger pre-exponential factor
Do, and thus become leading in some temperature interval. This gives rise to
effective increase in the value of E (see Fig. 6). On the other hand, when the
temperature becomes high enough so that the quasicrystal becomes closer to
a random tiling than to a regular structure, the second term tends to vanish,
because the density of defects in matching rules is close to saturation and
they do not limit the diffusion anymore. The details of the crossover between
"deterministic" and "random" regimes (dotted curve on the Fig. 6) might reveal
evidence of the hypothetic phase transition between deterministic and random
quasi crystals [14].
The pre-exponential factor Do in the Arrhenius law (1) is, roughly speaking,
given by the product of the square of the jump length a and the frequency of at-
tempts to overcome the energy barrier v, which is close to the Debye frequency.
There are, however, two important corrections to this term ([2]). First, as not
only the jumping atom, but the whole system undergoes thermal fluctuations,
the energy E in the Arrhenius law might depend on the temperature as well.
This dependence is described by the correction - T b.S, where b.S is the entropy
200 P.A. Kalugin

gain when the jumping atom is at the top of the barrier. Being independent of
the temperature after division by T, this term is usually grouped with the pre-
exponential factor. As the entropy gain is usually positive and equals several
k, this correction may increase diffusion by a factor about 10 2 . Although the
same effect should exist in quasi crystals, the distortion of the structure at the
moment of passing through the energy barrier is much smaller and the entropy
gain, too. The second correction is due to the so-called correlation effects.
These effects describe correlations between the successive jumps of a diffusing
atom, usually enhancing the probability of the reverse jump. This correction,
being rather moderate in the case of the vacancy diffusion (a factor 0.8 for FCC
lattice, for example) might become important for the present mechanism of dif-
fusion. The point is that the unbounded motion of one atom is accompanied by
complex correlated rearrangements of the atoms in its surrounding. The com-
puter simulation in [13] shows that the correlation effect lowers the diffusion
coefficient by the factor 0.025 in the case of the octagonal tiling. One should
remark, however, that the complexity of the diffusion process may be lower, if
we accept more violations in the matching rules than minimally needed.
In conclusion, one should remark that in quasi crystals may exist other dif-
fusion mechanisms, for instance the usual vacancy diffusion. This means that
the possibility of observing the proposed unusual diffusion process depends on
whether it can dominate over all others. If we consider the exponential factor
in the Arrhenius law, it is likely to be smaller than that for the vacancy diffu-
sion. The point is that the activation energy f.j cannot be higher than that for
jumps to vacant positions in a crystal, because in QC the distances of jumps
are thought to be shorter. The energy of the defects in matching rules cannot
be high neither, because this energy is due to misarrangements of atoms in the
second coordination shell. On the other hand, the pre-exponential factor, for
the reasons discussed above, may to be 2 or 3 orders of magnitude smaller than
that in crystals. Thus, the non-usual diffusion is the most likely to dominate
at low temperatures or, as may be seen from Fig. 6, at the temperature just
above that of the transition between regular and random quasi crystals.

References

[1] Kalugin P. A., Katz A., Europhys. Lett. 21 (1993) 92l.

[2] Flynn C. P., Point Defects and Diffusion (Clarendon Press, Oxford, 1972).
[3] Frenkel D. M., Henley C. L., Siggia E. D., Phys. Rev. B 34 (1986) 3649.
[4] Kalugin P. A., Europhys. Lett. 9 (1989) 545.
[5J Katz A., in Proceedings of the Adriatico Anniversary Research Confer-
ence on Quasicrystals (World Scientific Publishing Co., Singapore, 1990).
[6] Katz A., in From Number Theory to Physics (M. Luck, P. Moussa, M.
Waldschmidt and C. Itzykson, eds.), (Springer-Verlag, 1992).
[7] Elser V., Phys. Rev. Lett. 54 (1985) 1730.
[8] Shaw L. J. and Henley C. L., J. Phys. A 24 (1991) 4129.
 DIFFUSION IN QUASICRYSTALS 201

[9] Kalugin P. A., JETP Lett. 49 (1989) 406.

[10] De Bruijn N. G., Netherl. Acad. Wetensch. Proc. Ser. A 43 (1981) 39.
[11] Li W., Park H., Widom M., J. Stat. Phys. 66 (1992) 1.
[12] Beenker F. P. M., Algebraic theory of non-priodic tilings by two simple
building blocks: a square and a rhombus (Eindhoven, TH-Report 82-
WSK-04, 1982).
[13] Joseph D., Baake M., Kramer P., Trebin H.-R., Europhys. Lett. 27 (1994)
451.
[14] Dotera T. and Steinhardt P. J., Phys. Rev. Lett. 72 (1994) 1670.
 COURSE 8

Experimental aspects of the structure analysis of

aperiodic materials
Walter Steurer

Laboratory of Crystallography
ETH-Zentrum, CH-8092 Zurich
Switzerland

1. INTRODUCTION

The main difference between the analysis of periodic and aperiodic structures is that
in the latter case not only the short-range order, in particular the atomic structure of
one unit cell, but also the type of long-range order has to be determined, while the
lattice periodicity even of bad quality regular crystal structures is taken for granted.
This means that, for instance, the analysis of a typical intermetallic structure with
"" 10 16 atoms in the sample is reduced to the problem of finding the positions of the
"" 10 2 atoms located in one unit cell. Even in the case of extremly large virus
structures the number of atoms per unit cell does not exceed ""106 •
Without the knowledge of the global ordering principle, however, structure
analysis can become extremly difficult. Then the main problem is to find the rules
determining the long-rang order if there is any. The simple crystallographic rule of
thumb that crystal structures are only well ordered if their diffraction patterns do not
show any diffuse scattering can not be applied any more. Infinitely many structures
are imaginable which are perfectly deterministically ordered and have discrete and
continuous, or only continuous, components in their Fourier spectra [1].
Real crystals can never be ideally ordered because they are finite, their atoms are
vibrating and statistically distributed point defects exist in thermodynamic
equilibrium. Crystals with only equilibrium defects are called perfect crystals,
silicon and germanium are examples of the most perfect real crystals. One should
keep in mind, however, that even a single crystal of extremely pure 99.9999%
silicon, with dimensions suited for an X-ray structure analysis, contains ",,10 12
foreign atoms.
The possibilities and limits of today's state-of-the-art structure analysis to
identify and determine new types of aperiodic structures, and some experimental
aspects which are important especially for the long-range order analysis of aperiodic
structures as well as problems arising from pseudosymmetry are discussed.
204 W. Steurer

2. WHAT ARE APERIODIC MATERIALS?

According to the traditional crystallographic definition, an ideal crystal corresponds

to an infinite 3-dimensional translationally periodic array of identical structure
motifs. Its symmetry can fully be described by one of the 230 3-dimensional space
groups [2]. Mathematically, a periodic structure can be generated by the convolution
of a function describing a structure motif with a lattice function. The structure motif
corresponds, for instance, to the electron density function p(r) of one unit cell. The
lattice function g(r) corresponds to a set of &functions at the nodes r =i~ kja, of a
lattice A with basis a" i =1..3, and k j ell (1L is the set of positive and negative
integers including zero). In reciprocal spac~ this convolution corresponds to a
multiplication of the Fourier transform of the lattice function, i.e. the reciprocal
lattice A * with o-functions at the reciprocal-lattice nodes H =,~ hJa~ , withhj ell ,
and reciprocal basis vectors a~, i =1..3 , with a,.a; = Oij' and the Fourier transform
F(H) = Iv p(r)e ,uH.rdr of the structure motif. This product is called the weighted
2

reciprocal lattice, the weights are the structure factors F(H). For a comprehensive
description and discussion of the properties of the reciprocal space, see [2]. Thus,
the characteristic of an ideal crystal in direct and reciprocal space is the existence of
a lattice. In direct space, this lattice is decorated with identical structure motifs
preserving translational and point symmetry, in reciprocal space it is weighted with
structure factors preserving the point symmetry. This corresponds to a Fourier
spectrum containing a countable infinite number of discrete o-peaks.
Preserving the essential part of the above mentioned ideal-crystal
characterization, namely the lattice symmetry, a d-dimensional (d~n) (aperiodic)
crystal section, shortly (aperiodic) crystal, can be defined as ad-dimensional
(ir)rational section of a n-dimensional generalized crystal with n-dimensional
translational symmetry. Cutting the hypercrystal with the d-dimensional direct
physical space is equivalent to a projection of the weighted n-dimensional reciprocal
lattice upon the d-dimensional reciprocal physical space. The resulting dense set of
reciprocal lattice points defines a Z-module of rank n. The concept of describing an
aperiodic structure as section of a n-dimensional crystal makes only sense, however,
if it simplifies the representation of its structural order. Thus, with special
restrictions considering the density distribution of the atomic surfaces in the n-
dimensional unit cell, (in)commensurately modulated, composite, quasiperiodic or
fractal structures, for instance, can be obtained from irrational cuts [4].
In order to to make a clear dividing line between the aperiodic materials which
are object of this paper and the ones which are not, a rough classification of all
possibilities to arrange an infinite number of structure motifs is given:

Deterministic structures exhibit ideal long range order. There exists a unique
construction process, fixing type and position of each structure motif with a finite
number of parameters. According to their Fourier spectra, two different classes can
be distinguished.
 EXPERIMENTAL ASPECTS 205

(a) F(H);I!: 0 only for H =i~ hja: and h j eZ

discrete or countable dense Fourier spectra.
Examples: aperiodic (i.e. incommensurately modulated, quasiperiodic,
composite structures, .. ) and periodic crystals
(b) F(H);I!: 0 for all H =i~ hJa~ and h j elR
continuous, continous with discrete components, or singular Fourier spectra
Examples: fractal structures, structures generated by automata etc. which do
not belong to type (a)

Non-deterministic structures exhibit no long-range order. The Fourier spectra are

continuous.

(a) Statistical distribution with short-range order

Examples: glasses, molecular liquids, gases
(b) Statistical distribution without short-range order
Examples: monoatomic liquids, gases
(c) Arbitrary distribution
Examples: solids far from the equilibrium state

All real structures correspond to intermediate states between these limiting types due
to their finiteness, thermal vibrations of atoms, and equilibrium and non-equilibrium
defects. The finiteness of the structures leads in reciprocal space to a finite full width
at half-maximum (FWHM) of Bragg peaks due to the convolution of the B-peaks
with the Fourier transform of the crystal size-function. The random deviations from
the ideal atomic positions give rise to diffuse scattering phenomena like thermal
diffuse scattering (TDS), for instance. Thus, real (aperiodic) crystals can be
characterized as partly-detenninistic structures:

Partly-deterministic structures exhibit long-range order for time- and space-

averaged structure motifs. The averaging allows a unique assignment of
averaged structure motifs to average-structure positions. This corresponds to
Fourier spectra as defined for deterministic structures convoluted with crystal-
size functions, and which show additional continuous components.

Thus, only partly- or non-deterministic structures are experimentally accessible, and

in the following only partly-deterministic structures are dealt with.
In the case of the determination of regular, periodic structures the translational
symmetry is used as hard constraint. This means, that only the structure of one unit
cell is determined and the lattice symmetry is taken for granted. In reciprocal space
this assumption corresponds to a condensation of Bragg reflections with finite
FWHM to B-peaks located accurately on the reciprocal lattice nodes. This
extrapolation to the existence of an ideal crystal is generally out of the question even
when very bad quality samples (high mosaicity, micro-domain structure, defects, .. )
are investigated.
The situation is the same in the case of real aperiodic structures once the type of
idealized aperiodic ordering is "known". Then again this global ordering principle is
taken as hard constraint. The question, for instance, whether a structure is
206 W. Steurer

commensurately or incommensurately modulated can only be answered within a

given experimental resolution. Experimentally, the ratio of the modulation wave
length to the period of the underlying lattice can always be expressed as a rational
number. Speaking of incommensurately modulated structures means that the
experimental findings can be better understood and interpreted assuming an
incommensurate modulation. An incomensurate charge density wave can be moved
through an ideal crystal without changing the energy of the crystal, for instance.
This is not the case for a commensurate one. In some cases, the modulation period
changes with temperature in discrete steps ("devil's staircase") generating a series of
commensurate superstructures ("lock-in structures"), in other cases, one finds a
continuous variation within experimental resolution [5,6]. Then, one can call it an
incommensurately modulated structure.
If one cuts an ideal crystal of silicon, for instance, into two parts, both pieces will
have the same crystal structure. If one does the same for a glassy material or a
quasi crystal this will not be the case, the structures in the two pieces are not
superposable anymore. This may be an academic problem for macroscopic samples
but not for nano-crystals. Considering a nano-crystallite containing only a few unit
cells, nobody would doubt having a small part of a crystal. In the case of a nano-
quasicrystallite, however, it would be difficult to identify its structure as part of a
quasi crystal. Anyway, in both cases it would be better to speak of a large atomic
cluster because periodicity or quasiperiodicity are properties of the global ordering
of matter. This problem is actually present in the common nano-domain structures of
decagonal phases [7,8] which are often incorrectly called "twinned approximant
structures". They should better be interpreted in terms of five-fold orientational
ordering of very small domains (nanodomains) or clusters. It does not make much
sense to characterize nanodomains as "crystalline" or "quasicrystalline", it is much
more important what type of long-range order they show.
To find out whether a "quasicrystal" is perfectly quasiperiodic, in average
quasiperiodic or something else, requires the information from experiments being
sensitive to the global structure, i.e. diffraction experiments considering the accurate
distribution of the diffracted intensities. The limiting factors are the maximum
spatial and intensity resolution of the diffraction and detection equipment, and the
size and quality of the sample. The resolution which is available today on standard
synchrotron-beam-line equipment is sufficient to test at least whether the ordering of
atoms in a sample reaches the same degree of perfection as for silicon, for instance.
Of course, the higher the quality of the samples the more it is necessary to account at
least for some dynamical diffraction effects as reflection broadening and
displacement.

3. EXPERIMENTAL PROBES FOR DISTINGUISHING BETWEEN

CRYSTALS AND APERIODIC STRUCTURES

The number of all imaginable aperiodic structure types (quasicrystals, fractals,

automata generated structures, ... ) which can occur in real samples is limited by the
ratio between the size of structural building units and the size of the domains with
perfect structure. A rational approximant, for instance, is per definitionem a
translation ally periodic crystal. In the case of, for example, the (987,61 D)-rational
 EXPERIMENTAL ASPECTS 207

approximant (Le. 1: = 1 +j' is approximated by the ratio ~~b )of icosahedr~l Al-
Cu-Li, however, one single unit cell is larger than the average =IJlm domain size of
a pretty good crystal. There is no translational periodicity present anymore, and it
does not make sense to differentiate between crystals and quasi crystals in such a
case.
It is also easy to calculate that for an average size of the structure building
elements in quasicrystals of 20 - 30 A, and a domain size in a quite perfect sample of
=IJlm = 10000 A, only a few hundred clusters line up per direction. This is enough
to obtain quite sharp Bragg reflections in the case of periodic or quasiperiodic
ordering but not enough to realize more complex structures as, for instance,
structures with complicated fractal atomic surfaces in the n-dimensional description
[1,9] .

3.1 DitTraction Methods

Diffraction techniques (Tab. I) are today the only methods allowing a quantitative
determination of the full structure (atomic coordinates, probability density function
of the atoms from the Fourier transform of the temperature factor, occupancy factor
of atomic positions) of any material. As long as the kinematical diffraction theory,
perhaps with some corrections for dynamical effects (extinction), holds for the
interpretation of the experimentally accessible intensity data, methods for ab initio
structure analysis are available (statistical direct methods, Patterson syntheses, etc.).
If it fails then the trial-and-error refinement of structure models is the only way. The
more perfect the crystals become the less dynamical diffraction effects can be
neglected. This may be a serious problem for synchrotron radiation experiments on
perfect samples. A comparison of intensity data collected from a single crystal of
decagonal AI-Co-Ni on a synchrotron beam-line and on a X-ray tube, respectively,
proved that the strong reflections of the synchrotron data set were diminished to
=10% of the values of the X-ray tube data set due to anisotropic extinction. The
existence of dynamical effects (anomalous transmission) was also demonstrated on a
rather perfect icosahedral AI-Mn-Pd sample with coherence length of nearly IOJlm
[10].

Table 1.- Diffraction techniques with optimum sample size and diffraction theory needed for
evaluation indicated (SAED: selected area electron diffraction, CBED: convergent beam
electron diffraction).

method sample diffraction theory ab initio structure

diameter solution works
neutron diffraction =l-lOmm kinematical yes
X-ray diffraction =1 - 100 /l1Il kinematical (dynamical) yes
SAED =1 urn dynamical not yet
CBED >10 A dynamical not yet

The kinematical [3,11,12] theory bases on the assumption that in a diffraction

experiment the wave fields of the primary beam and the diffracted beams do not
208 w. Steurer

interact with each other in the material. It can be shown that under these
circumstances the diffracted intensity is proportional to the square of the modulus
of the structure factor. The Fourier transform of the reduced intensities corresponds
to the weighted vector diagram of the structure (Patterson function). It is possible to
calculate normalized structure factors for Bragg reflections and to employ statistical
direct methods, or to derive the structure factor moduli of diffuse intensities and to
use the maximum-entropy method or other techniques for ab-initio structure
analysis. The kinematical theory explains the fundamental properties of a diffraction
experiment, the geometry and discretness of the diffracted beams as well as their
intensities, but only approximately and within limits depending on the type of
radiation used.
The dynamical theory [3,11,12,13] takes the interactions with energy exchange
between all wave fields in a sample into account. It bases in the case of X-ray
diffraction on Maxwell's equation, and in the case of electron and neutron
diffraction on Schroedinger's equation for a crystalline medium. The geometrical
diffraction conditions (Laue equations) differ only slightly from those calculated by
the kinematical theory. Due to the refraction of the waves at the crystal-vacuum
interface, the moduli of the wave vectors of the incident wave and the diffracted
wave do not coincide anymore. Thus, Ewald's construction can be used but the
diffraction condition can be fulfilled even if a reciprocal lattice point does not
exactly intersect the Ewald-sphere. The main difference to the kinematical theory
lies in the calculation of the intensities. According to the dynamical theory, the
intensities are proportional to first power of the structure factor moduli. All real
crystals, however, lie somewhere in between the ideal values calculated from the
kinematical and from the dynamical theory. For the strongest reflections, a sample
may behave like a perfect crystal for others like an imperfect one. Crystals with a
mosaic block distribution allowing the application of the kinematical theory are
called ideally imperfect crystals.
The size Ak (extinction length) of a crystal (mosaic block, domain with perfect
structure), for which the kinematical approximation holds, can be calculated from

Ak- 0.7cose
- (e 2 I mc 2 )1 F(H) I'AiD.

with scattering angle e, charge e and mass m of the electron, light velocity c, wave
length A., and volume of one unit cell D. [12]. Thus, for large structure factors (strong
reflections), small unit cells, and long wave lengths, Ak becomes small. For MoKa-
radiation and the very strong 111 reflection of silicon, we get Ak =7 flm, for
instance. The extinction lengths for electron diffraction are much smaller and adopt
values of =1 nm for 100 kV electrons.
In any real diffraction experiment coherent and incoherent superposition,
respectively, of beams diffracted from different regions of a sample occur. Within
perfectly ordered domains (mosaic blocks) well defined phase relationships exist
between the wave fields. Thus, each domain with K structure elements with partial
structure factors Fk(H), gives qualitatively the same Fourier spectrum
 EXPERIMENTAL ASPECTS 209

Between the domains, however, there can be random phase relationships due to
randomly varying domain wall dimensions or small angle domain boundaries. These
average out the interference interaction between the wave fields emanating from the
different domains. Consequently, the wave fields stemming from different domains
can be summed up without considering any phase relationship. They have to be
incoherently superposed to each other, and instead of the complex structure factors
pomain(H) , the intensities Idomain(H) =pomain(H)F* domain(H) have to be summed up
giving

for a perfect crystal. The difference between the integral diffracted intensity of a
perfect (single domain) and an imperfect (multidomain) crystal equals the integral
intensity of the diffuse scattering. With increasing ratio of the scattering
contributions from domain boundaries the amount of diffuse scattering increases.
In real materials, there may exist twin- or antiphase-domain boundaries which do
not disturb the underlying lattice periodicity. This is, for instance, the case in the
domain structures resulting from martensitic (diffusionless) transformations where
the strain energy can be decreased by subdividing the crystal into twin domains to
compensate for the individual strain fields. Thus, there are in some cases well
defined phase jumps between the domains and the superposition of diffracted beams
from the domains occurs coherently leading to superstructures or incommensurately
modulated structures(e.g., Au-Cu anti-phase domain structures).
The nanodomains resulting from quasicrystal-to-crystal transformations appear
also to have, at least partly, non-statistical phase relationships to each other. They
have, therefore, to be dealt with coherent superposition as long as the domains are
small enough. Several HRTEM and high-resolution X-ray diffraction studies [7,14]
show that all intermediate states between coherently and incoherently diffracting
domain structures may occur depending on the chemical composition and thermal
history of the samples.
X-ray and neutron single-crystal diffraction are the methods of choice to
determine unknown (aperiodic) crystal structures quantitatively, provided that one
has crystals of sufficient size (Fig. 1). In standard experiments, the intensities
(proportional to the squares of the structure amplitudes) can be measured with
typical estimated standard deviations of ",,1 %, lattice parameters to ",,0.01 % and the
reciprocal space resolution is of the order ",,103 A·I. In particular cases, the X-ray
Pendellosung-method allows an accuracy of ",,0.05 % for structure amplitudes [15],
the X-ray Bond-method ""0.0001 % for peak positions and lattice parameters
[16,17], and on X-ray three-crystal diffractometers a resolution of ",,5xlO·s )..-1 can be
obtained [14].
Powder diffraction experiments on polycrystalline samples have the
disadvantage that the 3-dimensional diffraction patterns of all the individual single
grains in a material are angularly averaged and the reciprocal space information is
reduced to one dimension. In particular, the very large number of weak reflection
intensities present in quasicrystals or even more in fractally ordered or disordered
phases can not be resolved any more even in high-resolution experiments (FWHM
",,0.02°, CuKex). With known structures, powder diffraction may be used as "finger-
210 W. Steurer

print method" to identify and characterize samples during phase transformations, for
instance. For the identification of new types of ordering, however, it is less suited
due to also the lack of symmetry information. There is one example known that by
using single-crystal diffraction instead of powder techniques, the quasicrystals
would have been discovered much earlier. In the fifties, namely, equilibrium phase
was found in the system AI-Cu-Li, with "weak, fairy simple, but not cubic powder
pattern" [19] which was thirty years later identified by single crystal diffraction as
icosahedral AI6CuLi3 [20], the first stable quasicrystal.

Fig. l(a). - HRTEM image of decagonal AI70.5MnI6.5PdI3 with 2 A point-to-point

resolution and (b) projected electron density map as determined from X-ray structure analysis
on the same scale. The edge length of the outlined pentagon amounts to approximately 20 A
[18].

In the case of selected area electron diffraction (SAED) the intensities of weak
reflections and of diffuse scattering may strongly be enhanced due to dynamical
(multiple scattering) effects compared to the values expected from kinematical
theory. Since the mathematical relationship between crystal structure and diffracted
intensities can no more be simply expressed in terms of the Fourier transform
conventional structure determination techniques can not be applied. A quantitative
evaluation of the diffracted intensities is extremely time consuming and
complicated, and beside trial-and-error no further structure determination methods
are established yet. Generally, structure amplitudes can be measured by SAED at
least as accurate as by X-ray techniques with the exception of the X-ray
Pendellosung-method. Phases may be measured in favorable cases by means of the
three-beam technique with an unrivalled accuracy of far better than I % [21). To get
very quickly qualitative information about metrics and symmetry of the reciprocal
 EXPERIMENTAL ASPECTS 211

lattice from small areas of a new compound, SAED, carried out in the electron
microscope, is the method of choice. The spatial resolution, however, is low,
comparable to that of a conventional precession photograph taken on a conventional
X-ray tube.
Convergent beam electron diffraction (CBED) is often used to derive the point
symmetry (not only the Laue class) from extremely small areas (>10 A) of a
sample. It can also be used to identify the symmetry relationships between twin
domains, for instance, or to check whether the microscopic symmetry agrees with
the macroscopic one.' A CBED study of a decagonal phase, for instance, was
published recently [22]. The lattice parameters can be determined by CBED with an
accuracy of ::::0.1 %. Quantitative structure refinements are also possible under
certain conditions [21].

Fig. 2. - HRTEM image of a domain structure with overall pseudo-tenfold diffraction

symmetry of monoclinic (m) and orthorhombic (0) Al-Mn-Pd approximant phases [24].

3.2 Imaging Techniques

High-resolution transmission electron microscopy (HRTEM) can give very

impressive pictures related to the local projected structures of samples of arbitrary
order (Fig. I). The big advantage of this technique is that non-periodic features like
212 w. Steurer

aperiodic structures, crystal defects or complicated domain structures can also be

imaged (Fig. 2). With a lateral resolution up to =1-2 A, the type of ordering can be
studied very well, supposing it is not biased by the projection over the =100-500 A
thick sample. Reliable structure models can only be obtained by simulation
calculations from HRTEM images [23]. HRTEM is the method of choice and is
absolutely essential to detect new types of aperiodic structures in metallic and
inorganic materials and to get an impression of the kind of ordering.
Scanning tunneling microscopy (STM) or atomic force microscopy (AFM) can
give information on the local surface structure of a sample. One must consider,
however, that surface structure and bulk structure are not necessarily identical.
These techniques are good probes for obtaining qualitative information on complex
ordering phenomena of any kind. An example of a STM study of a decagonal phase
is given in the paper [25].

3.3 Spectroscopical Methods

EXAFS (extended X-ray absorption fine structure), ME (Mossbauer effect)

spectroscopy, and NMR (nuclear magnetic resonance), to mention only the most-
often used spectroscopical techniques in quasicrystal structural studies, can give
very accurate information on the local configuration of structural units, averaged
over the whole sample, however. They are not suited for identifying new unusual
long-range order but they may be very powerful in checking structure models
differing in local atomic surroundings. Since quasicrystals consist of the same
structure motifs as their crystalline approximants, the differences between both
structure types can hardly be detected with spectroscopical methods. More recent
work using one of these techniques in quasicrystal structure analysis were
undertaken, for instance, to study the local order in icosahedral AICu(Ru,Fe) by
EXAFS [26], in icosahedral and crystalline AI-Cu-Fe-V by ME spectroscopy [27],
and in ferromagnetic icosahedral phases by NMR [28]. An extensive review on the
application of ME spectroscopy to the study of quasicrystalline materials was
published recently [29].

4. STRUCTURE DETERMINATION METHODS

In nearly all structure analyses of incommensurately modulated phases (IMP),

composite crystals (CC), and quasicrystals (QC), of the last twenty years the higher-
dimensional embedding approach [4,30] has been used. The starting points of the n-
dimensional structure refinements are the estimated positions of the hyperatoms (Le.
the convolution of an atom in physical space with the (n-3)-dimensional atomic
surface in complementary space) in the n-dimensional unit cell. The positions can be
determined from the average structure in the case of incommensurately modulated
phases or composite crystals, or via the interpretation of the n-dimensional Patterson
function according to some model. There are attempts to find starting models by the
derivation of structure factor phases via direct methods (approximant embedding,
statistical techniques, maximum-entropy method) or experimental techniques
(neutron contrast variation by isotopic or isomorphous substitution, X-ray A.-
method, multi-beam techniques) but in most cases to date the trial-and-error method
is used [31].
 EXPERIMENTAL ASPECTS 213

Since by phasing of structure factors employing one of the above mentioned

methods, correct phases can be obtained only for a subset of all intensity data, a
successive structure refinement is indispensable. The main problem in n-
dimensional analysis is the parametrization of the hyperatoms: one needs n-tuples of
parameters for their coordinates, parameters for the description of their shapes and
the assignment of different types of atoms to regions corresponding to different
types of vertices in the physical space structure. One further needs up to six
parameters to account for anisotropic thermal vibration of the atoms in physical
space and also parameters for the description of random fluctuations in
perpendicular space.
The problem is that, with the choice of a particular parametrization of a n-
dimensional structure model for least-squares refinements, the possible solutions are
restrained to slight modifications of this starting model. This will become a big
problem, in particular, for the description of fractal or other non-dense atomic
surfaces. The procedure would become more flexible, however, if changes in the
parametrization of hyperatoms were also allowed during the refinements. The most
elegant way to do something in this direction is offered by the maximum-entropy
method (MEM) which will be discussed in the following section.

4.1 The Maximum-Entropy Method (MEM)

MEM represents a reconstruction technique giving the least-biased deduction

compatible to a given information. Applying this method to the problem of Fourier
inversion, the best obtainable electron density map can be derived from a noisy
and/or incomplete set of diffraction data [32,33]. In the usual method of Fourier
transforming the structure factors ("Fourier synthesis"), all non-observed structure
factors are set to zero and series truncation effects (riples, dummy maxima and
minima) appear in the electron density maps. Thus, both observed and unobserved
structure factors influence the results. MEM calculations, however, may rather be
compared to constrained many-parameter least-squares refinements. A quasi-
continuous electron density is modelled to get the best fit of calculated structure
factors to observed ones. Another application of MEM is the ab-initio phase
determination of structure factors [34]. A first example of the application of MEM in
the course of a n-dimensional structure analysis of a real quasiperiodic phase was
published recently [35].
For the calculations the Shannon/Jaynes configurational entropy

s =- ~ Pn(r)ln [Pn(r) / Pn_l(r)]

with probability Pn(r) =p(r) / r. p(r) , prior probability Pn _l(r) ='t(r) / r. 't(r) ,
electron density p(r) and prior electron density 't(r) at a particular pixel with
position vector r, has to be maximized under the constraint that the sum over the
squared differences between observed and calculated structure factors becomes a
minimum. The unconstrained maximization of S would result in an uniform electron
density distribution p(r). Thus, using some approximations [32] one gets for the
electron density
214 W. Steurer

with Lagrange multiplier A., and weights of the structure factors inversely
proportional to their experimental variances 0 2(0). The summation runs over all H,
the electron density p(r) is given on a regular grid defined in n-dimensional or
physical space. The structure factors are calculated by Fourier transforming the
electron density

Fcal.(O) =V~ p(r) exp (27tiH.r)

summing up over the asymmetric unit of grid points in the crystal volume V. Of
course, symmetry in direct and/or reciprocal space, respectively, has to be taken into
account. The iterations can be started with uniform electron density on all grid
points. In the case, where one wants to calculate accurate electron density maps
from known structure factors, the above given formula is sufficient (Fig. 3). For ab
initio structure analysis, the situation becomes much more difficult and one has to
include additional constraints to the calculations [36].

4.2 How many reflections have to be measured?

Theoretically, aperiodic crystals have an infinite number of reflections within a

given diffraction angle contrary to periodic crystals. The number of reflections to be
included in a structure analysis of a periodic crystal may be very high, one million
for a virus crystal, for instance, but there is no ambuiguity in the selection of
reflections: all reflections are used which are enclosed by a limiting sphere in
• -I
reciprocal space, given by 0::;; sine I A.::;; 0.7 A • for instance. All reflections, the
observed and the non-observed ones are needed to fit a reliable structure model [37].
How shall one proceed in the case of an aperiodic crystal? It would not be
possible to collect all X-ray intensity data within 0::;; sine I A.::;; 0.7 A-I, for a
quasicrystal, for instance. The way out is to define a limiting hypersphere in n-
dimensional reciprocal space and, again, to include observed and unobserved data in
the structure refinements. One should keep in mind. however, that the indexing of
reflections is the first model introduced in a structure analysis. If there does not exist
a unique way of indexing, an alternative can be to perform the data collection pixel
by pixel, with pixel size determined by the resolution of the diffraction equipment.
The assignment of indices and a reciprocal lattice, if any, to the intensity data could
be performed in the next step by fitting a model reciprocal lattice to the observed
reflection positions.
What happens if not all reflections are included in a structure analysis? How
important is the contribution of high order reflections which are weak but densely
distributed? Some answers to these questions are given on the example of the
Fibonacci sequence. An infinite model structure consisting of Al atoms with thermal
parameter B=l A2, and distances S=2.5 A and L='tS, was used for the calculation
(Tab. II).
 EXPERIMENTAL ASPECTS 215

2.5,--------------,. 2.5-r-- - - - -- - - - . . , .
H HJ.
2.0

..: .. ", .
:.
...
, ','
1.5
', ,' : , ...... .

1.0 . . : .:.. ' ' ..... ,:'

. .
"

', _ -,
. -,. - ":
" , ,'

o.
0.0 0.5 1.0 1.5 2.0 H " 2.5 0.0 0.5 1.0 1.5 2.0 H "2.5

a d

b e
'V

~ o

., ~
(9)

• • 0
~ ~ ~

C f
Fig. 3. - Comparison of electron density maps calculated from a full data set (a) by Fourier
transform (b) and MEM (c), from a data set truncated in typical way (d) by Fourier transform
(e) and MEM (f). Artifacts in the Fourier maps due to series termination are marked by
arrows.
 216 w. Steurer

Table 11.- Intensity statistics of the Fibonacci chain for total 161822 reflections with -
lOOO$;hr:;lOOO, and O$;sinSlN:;2 k 1 • In the upper lines, the number of reflections in the
respective intervals is given, in the lower lines the partial sums L I(H) of the intensities I(H)
are given as percentage of the total diffracted intensity. The F(OO) reflection is not included in
'the sums.

F(H)IF(H)max ~ 0.1 ~0.01 ~ 0.001 < 0.001

< 0.1 <0.01
0~sine/A<0.2 17 148 1505 14511
1: I(H) 52.53 % 2.56% 0.27% 0.03 %

O.~sinelA<O.4 11 107 1066 14998

1:1(H) 27.03 % 2.03 % 0.19% 0.02%

0.4~sinelA<0.6 9 64 654 15456

1: I(H) 9.84% 0.96% 0.12% O.oI %

0.6~sinelA<0.8 6 27 326 15823

1: I(H) 2.94% 0.34% 0.07% O.oI %

0.8~sine~2 1 35 338 96720

1:1(H) 0.23 % 0.79% 0.06% 0.01 %

It turns out that 92.57 % of the total diffracted intensity of 161822 reflections is
included in the 44 strongest reflections, and 99.25 % in the strongest 425 reflections.
It is remarkable, however, that in all the experimental data collections of icosahedral
and decagonal quasicrystals collected so far, rarley more than ten reflections along
reciprocal lattice lines corresponding to netplanes with Fibonacci sequence-like
distances could be observed. The consequences for structure determinations with
such truncated data sets are illustrated in Fig. 4 where the respective Fourier
syntheses are shown.

4.3 Aperiodic Sequences and Their Fourier Transforms

To get an impression to what extent experimental diffraction patterns of aperiodic

structures and approximants differ from each other, X-ray diffraction patterns
(according to the kimatical theory) of several I-dimensional structures were
calculated simulating the best resolution today obtainable at synchrotron storage
rings (AI atoms with thermal parameter B=1 A2, and distances S=2.5 A and L='tS).
The influence of the degree of perfection on the diffraction patterns was also studied
by simulating three different crystal qualities by introducing random domain walls
at different distances:

(a) Perfect crystal: coherence length 17711 atoms. This corresponds to a

domain size of 61190 A (=6Ilm).
 EXPERIMENTAL ASPECTS 217

LS L
l
.1, J 1 J ,I.
(a)
J ,I,
r
r
~
r
-
-
l $L -
-

J
•
(b )
; r

l'S l

'- L-..I '-- .-J ~

- ' '--' '--

(c) r

(d)

Fig. 4.- Fourier transforms of different sets of structure factors calculated for an ideal
Fibonacci-sequence: (a) all structure amplitudes with IF(H)I~O. IIFmax(H)I, (b) with
IF(H) I~ 0.01 IFmax(H)) , and (c) with IF(H) I ~ O.OOl\Fmax(H)I . An example with missing
large structure factors 0.11 Fmax (H) I ~ F(H) I ~ 0.011 Fmax(H) \ ' as this may be the case in real
diffraction experiments due to strong, non-correctable extinction effects, is shown in (d).
218 W. Steurer

(b) Good crystal: coherence length 987 atoms, corresponding to a domain

size of 3410 A(""0.3 /lm).

(c) Bad crystal: coherence length 200 atoms, corresponding to a domain

size of 690 A (""0.07 /lm).

A number of F-type icosahedral phases (AI-Mn-Pd, AI-Mg-Re, AI-eu-Fe, ... ) can be

grown with very high quality, thus classifying them for category (a). Decagonal
phases and most icosahedral phases are of quality type (b), and metastable
quasicrystals of type (c). High quality means large coherence length, and does not
say anything about the type of ordering (strictly quasiperiodic, random-tiling-like, .. ).
How would X-ray diffraction patterns of samples having these three different
domain sizes look with an experimental resolution obtainable from high-resolution
experiments (beam divergence 0.003°). In Fig. 5 crystals of the types (a) and (c)
of the Fibonacci sequence and a (3,2)-approximant are compared with each other.
One sees, that even in the worst case significant differences continue to exist.
In Figure 6 the simulated diffraction patterns of five different good quality
crystals (type (b)) are compared. In all structures there exist only two next-
neighbour distances S and and L='tS, their relative frequencies are always I: 't. The
sequence with fractal atomic surface in the higher-dimensional description [9] was
generated by the substitution rule S~SL and L~LLS. The composite crystal
structure corresponds to a I-dimensional projection of a 3-dimensional LLL... host
structure with channels including SSS ... chains.
Going from the Fibonacci-sequence to the fractal sequence one finds a
characteristic increase of the intensities in-between Bragg reflections with large
intensities. In a low resolution X-ray photograph this would be interpreted as diffuse
scattering. Thus, it is essential to perform high resolution measurements to find the
actual differences. In all cases, also in the random sequence, the strongest reflection
intensities nearly coincide, indicating that the local order is very similar in all
sequences. Performing a low resolution experiment without distinguishing
Fibonacci-sequence, approximant, fractal or random tiling, one would find the
fundamental atomic distances S, L and the characteristical short-range order.
Well, one is able to measure the diffraction spectra with high resolution in a
quasicontinuous scan along the reciprocal lattice line, and to calculate its Fourier
transform. The resulting Patterson function of the quasiperiodic, fractal and the
random structure are quite similar. Is it possible to reconstruct a structure model
from it unambigously? Beside MEM in connection with special constraints, a second
powerful method can be applied to these aperiodic structures.

4.4 Symmetry-Minimum Function and Patterson Deconvolution

This part will focus on a real space approach. Structure solution by unravelling the
autocorrelation function of the crystal structure in the unit cell, the Patterson
function [38]

p(u) = f
v
p(r)p(u+r) dr
 EXPERIMENTAL ASPECTS 219

(a)
1.0*105
1.89
I
1 I
1.9

5.0*10'

(b)
1.0*105
1.89
I
1 I
1.9

5.0*10'

0.0*1 00 -+-.L1...LLL.J....,I-L.l.....l_1-..l...,.L-J-L.l.....l-'--...,-L-~--l"""T'"-'--~---l

~C)
1.0*105
1.89 1.9

5.0*10'

0.0*1 0°-+..w...I....L....J.....l..+....L....'-'-~-4............L.........l.....J......,...L....--~-.I......;i"""''----4
(d)
1.0*105 •
I """'~I roC'

1.89 1.9

5.0*10'

0.0*1 O°-l-.;-l....L.J......L...J.....,I-.J..-'-'--_L...,L-...L.JL...........I...-,,--,-'--_~-'-...,....-'-_ _---',

o 2 3 4 H 5
Fig. 5. - The calculated diffraction pattern of 17711 atoms arranged with distances according
(a) to a perfect Fibonacci-sequence and (b) to a perfect (3,2)-approximant. Decreasing the
coherence length to 200 atoms the diffraction patterns (c) for an imperfect Fibonacci-
sequence and (d) for an imperfect (3,2)-approximant are obtained. The insets show enlarged
parts of the diffractograms.
 220 W. Steurer

6.0*10 3

4.0*103

2.0*103

o 2 3 4 H 5

(b)

lM!lJ\Al
6.0*103' -

4.0*103'-

r
2.0*10 3'-

0.0*10·
I I I I II
I I
o 2 3 4 H 5

(c) ,

6.0*103

4.0*103

2.0*103
 EXPERIMENTAL ASPECTS 221

(d)

6.0*1

4.0*103

2.0*103

0.0*100~-•

(e)
~
.

UI1J]
6.0*10 3

4.0*103

r
2.0*10'

0.0*10°..,---
o 2 3 4 H 5

Fig. 6. - Calculated diffraction patterns, patterson functions (upper inset) and electron density
functions (lower insert) of (a) a (3,2)-approximant, (b) a composite structure, (c) a Fibonacci-
sequence, (d) a fractal sequence, and (e) a random sequence.

where p{r) is the scattering density in the real crystallographic cell. The Patterson
function can be calculated from the experimentally derived diffraction intensities
alone (convolution theorem) by
222 w. Steurer

which is a major advantage for quasicrystal work. A systematic way to analyse the
Patterson map is based on the interatomic vectors between symmetrically equivalent
atomic positions, the Harker vectors [39]

where Ri is the rotational part and ti the translation part of the ith symmetry operator
C. of the space group [1]. Once these Harker vectors are identified in the Patterson
map, they can be directly used to determine the original atomic position r. In
practice, the identification of the Harker vectors depends largely upon the quality of
the Patterson map and the complexity of the structure involved. For X-ray data in
particular, the limits of attainable resolution (minimum d-spacing) will affect both
the resolution of atomic peaks in the unit cell as well as peaks in the Patterson cell.
The width of a Patterson peak will also increase with increasing thermal motion.
Hence, it is preferable to utilize the information in the entire Patterson map rather
than just the identifiable peaks. Instead of deriving atomic positions from a selection
of Harker vectors, all positions r in the crystal cell are tested as trial atomic
positions. The symmetry minimum function (SMF) [40-42]

SMF(r) = u!'in n\P(r - C. r)

1-1 I

inspects the unique Harker vectors (i=l, ... ,p) in the Patterson map (the origin vector
is not included); the quantity mi takes the multiplicity of the Harker vectors into
account. The minimum function ensures that for an acceptable atomic position r all
the Harker vectors are above background level. The peaks in the SMF are the
possible atomic positions. Obviously, the chances of obtaining a large value in the
SMF at an incorrect atomic position decrease rapidly with an increasing number of
unique Harker vectors.
A crystal structure, its enantiomorph structure and the structure shifted by a
permissible origin shift have exactly the same vector set (homometric structures),
and hence the same Patterson map. These ambiguities are also present in the SMF
map. Mathematically, these ambiguities are described by the Euclidean normalisers
of the space groups and were first derived by Hirshfeld who called it Cheshire
groups [43]. From a set of symmetrically equivalent peak positions in the SMF,
only one position will be selected as a representative. However, the representatives
are not necessarily related to each other by the same origin, and hence cannot be
used directly for a structure refinement. We have to select a single image from all
the origin-shifted images of the crystal structure present in the SMF.
In principle, a single image of the crystal structure can be retrieved from the
Patterson function by selecting a pivot peak from the SMF with position r' (i.e. a
possible atomic position) and comparing it with other possible atomic positions r by
searching the Patterson function for interatomic vectors r-r'. Such vectors between
atoms which are not symmetrically equivalent are called cross vectors and provide
information additional to that of the Harker vectors. Again, each position r in the
crystal cell is tested with respect to the fixed position r' and its space group
equivalent copies C i r+ by analysing all the cross vectors r - C. r+ (i=l, ... ,p) with
the minimum function
 EXPERIMENTAL ASPECTS 223

(a) (b) o 0
. o 0
.
0 0 0 0
0 0 0 000 000 0
0
000 0 000 0
0 @ @ 0000
000 0
0000
000 0
0 @ @ og g 0 og g 0

••
0 o 0 0 0 o 0
0

0 o 0 o 0
0 0 0 0
000 0 000 0
@ @
000 0 000 0
@ @ 0000 0000
000 0 000 0
o g g 0 0 000
0 0 0 0 0
o 0 o 0
0

(c) (d)

0
"0
® I o 0

•
0 ®
!!J 0
~ ~ •
•® 0o 0
0 0 @

•® [J
11 0 e ~ •
~ e
"I:l 0
0 0 i
•i
® 0

0 • ~ 0

• 0
0 @ 0
@ 0 0

Fig. 7. - Decagonal Al-Fe-Pd: characteristical (1011O)-section through the a) 5-dimensional

Patterson map, b) symmetry minimum function (SMF), c) image seeking minimum function
(IMF) with pivot element (0.4,0.0,0.25,0.4,0.0), and d) image seeking minimum function with
pivot element (0.2,0.0,0.25,0.2,0.0).
224 w. Steurer

The function IMF is called the image seeking minimum function because it locates
the image of all symmetrically equivalent positions Cj r* (i=l, ... ,p) relative to a
trial position r. By selecting and fixing r*, all but one of the ambiguous origin
shifted images are eliminated. If the structure is not centrosymmetrical, one of the
two enantiomorphic solutions can be eliminated by suitable selection of a second
position r;

How this procedure works is demonstrated on the example of decagonal AI-Fe-Pd.

For decagonal quasicrystals there exist characteristic 2-dimensional cuts through the
5-dimensional hypercrystal exhibiting all centres of the hyperatoms. Based on the
Patterson function calculated on such a (10110) cut, all the above functions can be
reduced from the 5-dimensional space group P105/mmc to the 2-dimensional plane
group p2mg (Fig. 7). The SMF shows every possible atomic site in this structure.
One easily can recognize the superposition due to the origin shifts. The two structure
solutions of the IMF correspond to the choice of the strongest SMF maxima as pivot
elements.

4.5 Quasicrystals Versus Twinned Approximants

It has been shown by many HRTEM and high-resolution X-ray experiments that it
may be rather difficult to prepare good quality single crystals of quasiperiodic
materials. Decagonal phases, probably stable at high temperatures only, are
especially difficult to quench. Samples studied at room temperature, therefore,
correspond often to not very well defined intermediate states of the transformation
process between decagonal and crystalline phase. Naturally, the question arises as
to which results can be expected if one takes such an intermediate state phase for a
quasicrystal and performs a n-dimensional structure analysis on it.
This situation has been simulated on the example of decagonal AI-Co-Ni. It has
been demonstrated by HRTEM [7], how the high- and the low-temperature phases
are structurally related to each other. Both structures are built from the same
columnar clusters, which are quasiperiodic ally ordered at high temperature, and in
domains randomly or periodically ordered at low temperature. The domains,
however, show globally fivefold orientational order, and the domain boundaries
appear to be at least partly in non-random phase relationships.
The unit cell dimensions of the low-temperature approximant have been derived
from several experiments to a=61.024(5) A, b=83.99(1) A and c=4.0799(5) A in the
orthorhombic setting [14]. We derived a structure model for this approximant from
the structure of the decagonal phase [35]. It was asssumed that the structure building
elements (columnar clusters) have the same dimensions in the quasicrystal and the
approximant. In Figure 8 the simulated diffraction patterns of a single crystalline
approximant and a fivefold twinned phase are shown in comparison with a simulated
diffraction pattern of decagonal AI-Co-Ni.
 EXPERIMENTAL ASPECTS 225

. . . .. .. . .
. . . .. . . . . . . .. .. . .. .. . . .
.. ... .. .
. ..
.
.. . . .
.. : .: ::~.. ;.: ... . .
. . ' '
.. . ,', ..
, :.· /·,·: :·:;\·~~w~/~. . : , >:· .
.. . . . .. . . . . .
. . . ..
. ,' .'
.. . . . . .
., ', '

. .. . .
"

. . . .

a b c

Fig. 8 X-ray precession photographs calculated with laboratory resolution from the single
crystalline approximant with a=61.024(5) A, b=83.99(l) A and c=4.0799(5) A(a), the fivefold
incoherently twinned approximant (b), and for decagonal AI-Co-Ni (c).

The tenfold pseudosymmetry of the approximant is obvious, even the systematic

extinctions of the odd layer (not shown here) are fulfilled within the experimental
resolution. Thus, the differences between approximant and quasicrystal structure
appear to be small. What happens if one collects the intensity data on a four-circle
diffractometer in the laboratory around the reciprocal lattice points given by a Z-
module? In Fig. 9 ahistogram is given to demonstrate how many reflections from
100.------------------------------------------------

90

80

70

-c
Q)
~ 50
60

Q)
a.. 40 a Q,002A' '

30

20

10

0
12 S":5 I , . 11011U13 1 "11S1'17 1 ".2021ZZ23%4252121282t3031n~"':)S3eS131"40"'4l4.S44U

b Clustering of twin reflections in groups of N (N=1 ... 45)

Fig. 9. - (a) Quintuplet of twin reflections (marked 1, ..,5) calculated for synchrotron
resolution superposed with the one calculated for laboratory resolution. X indicates the
position of the quasicrystal reflection. (b) Histogram indicating the number of reflections
(shaded bars) and the total intensity of reflections (black bars) collected under a single peak
applying laboratory resolution.
226 w. Steurer

the superposed five reciprocal lattices of the approximants are subsumed in each
case under one "quasicrystal" reflection under laboratory conditions (IOn-scans,
MoKa). It is obvious that nearly 95% of all reflections are measured in quintuplets.
Thus, if the approximant is pseudo-decagonal then a rotation of its reciprocal lattice
by multiples of 36° leaves the reciprocal lattice nearly invariant. The superposition
of the five reciprocal lattices always collects five different reciprocal lattice points in
a small area. To check the situation quantitatively, a data collection has been
simulated and a n-dimensional structure refinement was performed mistaking the
twinned approximant as a quasicrystal. The refinements converged to a final
weighted R-factor of less than 1% giving nearly the ideal model-quasicrystal
parameters. Thus, from a pragmatic point of view, it is encouraging that a low
resolution diffraction experiment on a twinned approximant can give nearly the
same structural results as if a quasicrystal were used.

5. QUANTITATIVE APERIODIC· CRYSTAL STRUCTURE ANALYSIS

There exist strict recommendations for conventional structure analyses what kind of
quantitative information is to be explicitely given in publications (Notes for Authors
[44 D. Aperiodic crystal structure analyses have to build up on eighty years of
experience in the art of structure solution and presentation, and follow these guide-
lines by analogy as far as possible. A few of the most important points are listed
below, which should be observed performing n-dimensional structure
determinations and publishing the results:

(I) Symmetry and metrics: n-dimensional space group symmetry, way of indexing,
systematic absences, and n-dimensional unit cell parameters are to be given.

(2) Full data set: The data set should include all observed and unobserved (!)
reflections with n-dimensional diffraction vectors

giving an isotropic resolution function in the n-dimensional space comparable to that

of conventional structure determinations. The range of indices hi, i=I, .. ,n, and the
number of observed and unobserved reflections is to be listed.

(3) Reliability factors: The internal R-factor should be given if symmetrically

equivalent intensities are averaged. The quality of the n-dimensional least-squares
I I
structure refinements should be illustrated by a IFob.(H) I/ Fca1c(H) plot and/or a R-
factor statistics with R-factors calculated as a function of IHoll and IHnl, for
instance.

(4) Refined model parameters: Site symmetries, n-dimensional coordinates, shape,

dimensions, chemical compositions of the hyperatoms, temperature factors, and all
other refined parameters should be given with their estimated standard deviations.
 EXPERIMENTAL ASPECTS 227

(5) Residual densities: The residual densities calculated from the n-dimensional
difference Fourier synthesis after the final refinements are to be given.

The aim is that the information needed to evaluate or repeat a structure analysis is
given in the publication (or in additional material deposited) as it is demanded for
standard structure determinations. In this way also aperiodic-crystal structure
analysis will become reproducible and the results simpler to compare.

Acknowledgements

The author wishes to thank Dr.T.Haibach and Dr.M.Estermann for their excellent
contributions to section 4.

References

[1] Luck,J.M., Godreche, Janner & Janssen,T., 1.Phys.A: Math.Gen. 26 (1993)

1951.
[2] International Tablesjor Crystallography, Vol.A: ed. by T.Hahn, Kluwer
Academic Publishers: DordrechtIBostonILondon ( 1992)
[3] International Tablesjor Crystallography, Vol.B: ed. by U.Shmueli, Kluwer
Academic Publishers: DordrechtIBostonILondon ( 1993)
[4] Janssen, T., Physics Reports 168 (1988) 55.
[5] Cummins, H.Z., Phys.Reports 185 (1990) 211.
[6] van Smaalen, S., Crystallogr.Rev. 4 (1995) 79.
[7] Hiraga, K., Lincoln, F.J. & Sun, W., Mater. Trans.JIM 32 (1991) 308.
[8] Hiraga, K., Sun, W., Lincoln, F.J., Kaneko, M. & Matsuo, Y.,
lpn.J.Appl.Phys.30 (1991) 2028.
[9] Zobetz, E., Acta Crystallogr. A49 (1993) 667.
[10] Kycia, S.W., Goldman, A.I., Lograsso, T.A., Delaney, D.W., Black, D.,
Sutton, M., Dufresne, E., Bruning, R. & Rodricks, B., Phys.Rev. B48 (1993)
3544.
[11] Cowley, J.M., Diffraction Physics (North-Holland, Amsterdam, 1975).
[12] Vainshtein, B.K., Fundamentals of Crystals (Springer, Berlin Heidelberg,
1994).
[13] Pinsker, Z.G., Dynamical Scattering of X-rays in Crystals (Springer, Berlin
Heidelberg, 1978)
[14] Kalning, M., Kek, S., Burandt, B., Press, W. & Steurer, W., 1.Phys.: Condens.
Matter 6 (1994) 6177.
[15] Cummings, S. & Hart, M., Aust.l.Phys. 41 (1988) 423.
[16] Bond, W.L., Acta.Crystallogr. 13 (1960) 814.
[17] Galdecka, E., Acta Crystallogr. A49 (1993) 106.
[18] Beeli, C. & Nissen, H.-U., 1.Non-Cryst.Solids 153 & 154 (1993) 463.
[19] Hardy, H.K & Silcock, J.M., 1.Inst.Metal. 84 (1956) 423.
[20] Saintfort, P. & Dubost, B., 1.Phys.France 47 (1986) C3-321.
[21] Spence, J.C.H., Acta Crystallogr. A49 (1993) 231.
[22] Tsuda, K., Saito, M., Terauchi, M., Tanaka, M., Osai, A.P., Inoue, A. &
Masumoto, T., lpn.J.Appl.Phys. 32 (1993) 129.
228 w. Steurer

[23] Beeli, C. & Horiuchi, S., Phil. Mag. B70 (1994) 215.
[24] Daulton, T.L. & Kelton, K.F., Phil.Mag. B68 (1993) 697.
[25] Kortan, AR, Becker, RS., Thiel, F.A & Chen, H.S., Phys.Rev. B64 (1990)
200.
[26] Sadoc, A, Berger, C. & Calvayrac, Y., Phil.Mag. B68 (1993) 475.
[27] Stadnik, Z.M., Mii11er, F., Stroink, G. & Rosenberg, M.,Hypeifine Interactions
94 (1994) 2255.
[28] Shinohara, T., Sci.Rep.RITU A39 (1994) 165.
[29] Dunlap, RA & Lawther, Mater.Sci.Eng. 10 (1993) 141.
[30] deWolff, P.M., Acta Crystallogr. A30 (1974) 777.
[31] Steurer, W., Z.Kristallogr. 190 (1990) 179.
[32] Sakata, M. & Sato, M., Acta Crystallogr. A46 (1990) 263.
[33] Livesey, AK. & Skilling, J., Acta Crystallogr. A41 (1985) 113.
[34] Bricogne, G. & Gilmore, C.J., Acta Crystallogr. A46 (1990) 284.
[35] Steurer, W., Haibach, T., Zhang, B., Kek, S. & Luck, R, Acta Crystallogr.
B49 (1993) 661.
[36] Bricogne, G., Acta Crystallogr. 049 (1993) 37.
[37] Watkin, D., Acta Crystallogr. A50 (1994) 411.
[38] Patterson, A. L., Phys. Rev. 46 (1934) 372.
[39] Harker, D., J. Chern. Phys. 4 (1936) 381.
[40] Simpson, P.G., Dobrott, RD. & Lipscomb, W., Acta Crystallogr. 18 (1965)
169.
[41] Kraut, J., Acta Crystallogr. 14 (1961) 1146.
[42] Estermann, M.A, Nucl.Instr.Meth.Phys.Res. A354 (1995) 126.
[43] Hirshfeld, F.L., Acta Crystallogr. A24 (1968) 30l.
[44] Notes for Authors: Acta Crystallogr. C51 (1995) 159.
 COURSE 9

Scattering on aperiodic superlattices

P. Mikulik

Department of Solid State Physics

Masaryk University
611 37 Erno
Czech Republic

1. Introduction

X-ray elastic scattering techniques used for the nondestructive characterization

of aperiodic superlattices are studied in this paper. A superlattice (SL) is a
layered system formed by a multilayer (ML) deposited on a substrate. Accord-
ing to the given rule for the layer sequence the multilayer can be either periodic
or aperiodic. Due to the lateral translational invariance, the ML can be viewed
as an artificial 1D superstructure in the growth direction.
The aperiodic SLs grown so far were based on semiconductor materials, usu-
ally GaAs/ AlAs. The first report about a periodic semiconductor SL was pub-
lished by Esaki, Chang and Tsu in 1971 [1]. In 1985, Merlin and co-workers
[2] reported X-ray diffraction measurements on an aperiodic Fibonacci super-
lattice for the first time. They have described the diffraction peak positions in
the reciprocal space by two integers, as opposed to one integer in the case of
a periodic SL. X-ray structural analysis of Fibonacci SLs grown by molecular
beam epitaxy (MBE) were carried out in Ref. [3]. Todd et al. [4] simulated the
influence of the fluctuation of the layer thickness in the Fibonacci SL. They
have found that the profile of the most intense peaks is not disrupted, whereas
the intensity of the low intensity peaks is either enhanced or decreased. In
1988 the so-called configuration and environmental Fibonacci SL was studied
by Terauchi et al. [5]. In 1990, diffraction on third-order Fibonacci SL was
reported [6]. Today, different generalizations of the Fibonacci SL are being
230 P. MikuHk

studied [7, 8]. Among the non-quasiperiodic SLs, the Thue-Morse superlattice
has been deeply investigated. The first reported high resolution X-ray diffrac-
tion measurement was performed by Axel and Terauchi [9]. Raman scattering
on Fibonacci, Thue-Morse and random superlattices was compared in Ref. [10].
The paper is organized as follows. In the second part the X-ray diffraction
theories for studying SLs will be presented. In the third part both the kine-
matical theory and the semi-kinematical approximation of dynamical theory
will be used to calculate both the positions and the intensities of the diffrac-
tion peaks of the Fibonacci SL by the Fourier transform treatment and by an
exact treatment. The diffraction curve of Fibonacci SL will be compared with
that of a periodic SL. Diffraction patterns of other types of aperiodic superlat-
tices will be reviewed briefly. The fourth part concerning X-ray reflectivity will
demonstrate the equivalence between the reflectivity and diffraction. In the
last section the idea of a Fibonacci grating will be proposed and the reciprocal
space intensity distribution will be kinematically calculated.

2. Theories of X-ray diffraction

Let us first give a short introduction to the diffraction theories which will be
later applied to the detailed study of the diffraction pattern of the quasiperi-
odic Fibonacci superlattice. Both the kinematical and dynamical theory will
be treated. Their regions of validity and their connections with the Fourier
transform will be discussed.

2.1. Kinematical theory of X-ray diffraction

In the kinematical theory the response of a crystal to an incoming wave

Dinc(r, k o ) is given by a formula equivalent to the Huygens principle used
in optics [11]

(1)

The function of response T( r) = -relC p( r) is proportional to the electron

density of the crystal p( r). The detector position is denoted by r D, rei is
the classical electron radius and C is the polarization factor. The approach of
the kinematical theory is similar to the Born approximation used in quantum
mechanics as both theories neglect the multiple interaction between scatterers.

2.1.1. The Fraunhofer approximation

The Fraunhofer approximation is usually used to simplify the calculation of the
diffraction integral (1). It is valid if the dimension of the sample is much less
than the distance from the sample to the source and to the detector. Performing
 SCATTERING ON APERIODIC SUPERLATTICES 231

(a) (b)
=
y

Fig.!. - Direct (a) and reciprocal (b) space of a crystalline layer with thickness L.
The reciprocal space consists of a periodic distribution of the truncation rods, which
are infinite for an infinitely thin layer and finite for a layer with finite layer thickness.

the so-called plane wave approximation the diffracted amplitude is proportional

to the Fourier transform (FT) of the electron de~sity

Ddif (rD, k, k o) = _Dinc(O, k o) relCJdr p(r)e-·(Ic-Ico)r ~ FT{p}(k-ko) (2)

rD

and it depends on the difference ofthe exit and incident wave vectors q = k-k o.
The diffracted intensity is proportional to the square of the modulus of the
Fourier transform.
Consequently, it is useful to interpret the diffraction pattern of the superlat-
tices by means of the FT of its density. The Fourier transform of any function
r
f( 1') will be further defined FT {f}( q) == (q) == f dr f( 1') e- iqr .
Let us ignore the restriction to small crystals and look at what this approach
gives. The density distribution of a superlattice is composed of the substrate
density psub and the multilayer density PML' Since the FT of a sum of functions
is a sum of the FTs, the contributions of the substrate and the ML to the
total diffracted amplitude can be treated separately. If we treat the substrate
as a perfect infinite crystal, its density is given by a periodic repetition of
the density of one unit cell. Its FT is nonzero only in discrete points of the
crystal reciprocal lattice. The diffracted intensity is non-zero for discrete sets
of wave vectors fulfilling the well-known Bragg law 2 dhkl sin ()B = A, and it
is proportional to the square of modulus of the structure factor of the unit
cell. The density of an mjinztely thzn layer, which represents a two-dimensional
crystal, is given by a periodic repetition of the density of the 2D unit cell. Its
Fourier transform in 3D reciprocal space is formed by the set of infinite sticks
located at points of the 2D reciprocal lattice.
The finite size of a real crystal is taken into account by multiplying the
density of the infinite crystal by the shape function n( 1') (equals 1 inside and
o outside the crystal), p(r) = n(r).pOO(r). It turns the Fourier transform into
the convolution p*(q) = n*(q) (9 poo*(q).
In the particular case of a 2D layer with finite thickness (Fig. 1) the finite
size effect cuts the infinitely long sticks into jinzte truncatzon rods or, likewise,
232 P. MikuHk

finite truncation rods are generated by the one-dimensional expansion of the

reciprocal lattice points of the infinite 3D crystal.
The shape function treatment is valid only if the FT of the infinitely large
crystal has an atomic spectral measure (it is given by delta functions), as in
periodic and quasiperiodi.c crystals. Since the Fourier transform of the shape
function n· is proportional to the crystal volume and poo • is volume inde-
pendent. the diffracted intensity shows the volume square scaling. For certain
aperiodic crystals, such as a 1D Thue-Morse crystal, the FT of the infinitely
long lattice does not exist and the intensity of its diffraction curve scales dif-
ferently for different wave vectors.

2.1.2. Stationary phase method

It was stated in the previous section that the validity of the Fraunhofer ap-
proximation is restricted to the diffraction on small crystals. The superlattices
are laterally large and can be supposed to be laterally infinite. To solve the
kinematical diffraction integral for laterally infinite multilayers the stationary
phase method can be used [12].
The stationary phase method is a mathematical method to solve certain
types of integrals [11]. In our case it reduces the diffraction integral (1) over
the whole volume of the crystal into the integration over the classical path of
the diffracted beam. The contributions of the other beam paths are cancelled.
The amplitude of the wave diffracted with a scattering vector in the vicinity of
the reciprocal lattice vector h is then given by the expression

Dh = Do - i211T
k e/C
hz
1 ML
'('- (L +L)
dz Ph(Z) e' .. ".-"0 R •.) z , (3)

which has the form of the Fourier transform of Ph(Z) (the Fourier coefficient
of the electron density Ph along the growth direction :;). The wave vector of
the diffracted wave kh fulfills the condition of elastic scattering Ikhl = Ikol
and the in-plane Bragg law'khll = kOIl + hll' The solution of the diffraction
integral within the stationary phase method is similar to that given within the
semi-kinematical approximation of the dynamical theory. However, using the
kinematical theory, the effects of refraction and absorption are not involved
here.
The preceding analysis allows us to describe qualitatively the expected diff-
raction pattern of SLs. It consists of the Bragg spots of the substrate and
of the truncation rods of the ML with an intensity modulation related to the
structure of the multilayer. For example, the distribution of the intensity along
the truncation rod for a quasiperiodic ML consists of a dense set of maxima.

2.2. Dynamical theory

The kinematical theory is valid as long as the diffracted wave is so weak that the
attenuation (extinction) of the transmitted and diffracted beam is negligible.
 SCATTERING ON APERIODIC SUPERLATTICES 233

If the diffraction is strong, the dynamical theory should be employed. For

example the .diffraction curve of the semi-infinite substrate of the Fibonacci SL
should be calculated dynamically. The dynamical theory of X-ray diffraction
calculates the diffracted wave field by solving the Maxwell equations in the
crystal [11, 13]. Here only the main results of this theory will be summarized.
In the so-called two beam case the incoming wave nearly fulfills the Bragg
condition for only one set of net planes. Then two interacting wave fields
propagate in the crystal: the transmitted and the diffracted one.
The dependence of the diffracted intensity outside the semi-infinite crystal
on the angular departure from the Bragg angle is described by the Darwin-
Prin's rocking curve [11, 13]. Its main features are its finite width in contrast
to a delta peak expected by the kinematical theory and a shift of its maximum
from the exact Bragg position due to refraction.
The lattice mismatch between different layers creates a strain field in the
SLs. In epitaxial semiconductor SLs the lateral lattice parameters are adapted
to the substrate lattice. This induces a distortion of the elementary unit cells
of the layers. The deformation field f(::) in 'a ML with the total thickness
L is f( =) = f{ ~c( z) d=, where the local deformation in the lattice (lattice
parameter c(=)) with respect to the substrate (lattice parameter c) is ~c(z) =
(c(=) - c)/c.
The Taupin equation [14] can be used to find a recurrent formula for the
diffraction amplitude of a strained multilayer system [15]. Unfortunately, the
dynamical formulae do not relate easily to the intensity distribution in the
reciprocal space. However, it can be shown that if the total thickness of the
NIL is less than the so-called extinction length, the diffracted amplitude is small
and dynamical effects are small. The extinction length for usually employed
semiconductor materials and non-grazing diffraction geometries is of the order
of micrometers. It therefore follows that for the calculation of the diffraction
curve of a thick superlattice the dynamical formulae must be used, whereas for
a thin one the semi-kinematical approximation is valid [16].

2.2.1. Semi-kinematical approximation

The semi-kinematical approximation is valid for thin and weakly diffracting
multilayers, since the dynamical effects can be neglected there. The Taupin
equation can be linearized. Its solution for the relative diffracted amplitude
X = D di / / Dine of a superlattice with thickness L

X = Xsubei(uL+"Yhh./(Lll/'Yh + i2J(C 1L d= Xh(Z) eih:/(z) . ei(u/,'Yh) Z (4)

'Yh 0

is the sum of the substrate and multilayer diffracted amplitudes. The latter is
proportional to the one-dimensional Fourier transform in the growth direction
of the Fourier coefficient of the electron susceptibility 'Xh, modulated by a
phase shift hzf( z) caused by the local change in the reciprocal lattice due to
the local lattice deformation f( z). The Fourier coefficients of susceptibility
234 P. Mikuh'k

Xh, which are used in the dynamical theory, are proportional to the Fourier
coefficients of the electron density Ph used in the kinematical theory. The
wave vector length in vacuum is K = 271"/ A. The wave vector transfer in the
= =
growth direction in the Bragg position is hz (kB - kOB)z -2K cosw sin BB.
The angle between the diffracting net planes and the crystal surface is w, the
direction cosines of transmitted and diffracted beams are 10 = sin(w + BB) and
Ih = sin(w - BB), respectively. The wave vector departure on the substrate
u = u( 11) = J{ 11 sin 2BB + H<Xo(l-,h/'YO) depends on the angular departure of
the incoming wave from the Bragg angle 11 = B - BB. The intensity diffracted
by the crystal is then R(I1) = I!h/'YoIIX(I1)j2.
The linearized Taupin equation for superlattices with a multilayer having N
layers can be expressed in a compact way by a discrete summation

(5)

where we have denoted

- Fk = Xh,k dk sinc (¢k/2) eit/>k/2 the structure factor of the k-th layer (thick-
ness dk), where sinc(z) == sin(z)/z,
- ¢k = Ukdk/'Yh the phase shift on the k-th layer,
- uk(l1) = U(I1) + Ihhz6.Ck the wave vector departure on the k-th layer,
- Xh,k the Fourier coefficients of susceptibility in the k-th layer.

In Eq. (5) it is useful to introduce the quantity

N
:F = L Fk e- icp • , (6)
k=l

which will be referred to as the structure-geometric factor.

In summary, the basis of the semi-kinematical approximation lies in the
matching of the dynamically calculated diffraction amplitude of the substrate
Xsub with the "kinematically" (i.e. having a form of a Fourier transform) cal-
culated diffraction amplitude :F of the multilayer, and including all the effects
of refraction, absorption and lattice deformation.

3. Diffraction on multilayers

It was shown above that in the kinematical theory (Eq. (3)) or in the semi-
kinematical approximation (Eq. (4)) the diffracted amplitude of a multilayer is
given by the Fourier transform in the growth direction of a variable P propor-
tional to the electron density

p*(q) = J dz p(q, z) e-iq,z . (7)

 SCATTERING ON APERIODIC SUPERLATTICES 235

For convenience, we denote k to be the z component of q, k == qz, in the

whole of this chapter.
Let us suppose that the multilayer is formed by two building layers A and
B which together form the lattice Q. This can be represented graphically as
follows:

Lattice Q: A B A A B A B A

Lattice QA: • • • • •
Lattice QB: • • •

The A layers form the sublattice QA with lattice points at ZAn, the B layers
form the sublattice QB with lattice points at ZBn. Their layer thicknesses are
dA., dB, the densities are PA, PB, and the structure factors of the layers are
F..t == P'A, FB == PH' respectively. The density of such a ID crystal is the sum
of the densities of both sublattices

(8)
n n

The Fourier transform (7) then becomes

p.(q)
n n

(9)

Both the A and B layers contribute to the FT by the product of the structure
factor F (it has to be calculated numerically) and the phase sum S (it can be
calculated analytically for certain layer sequences). Therefore the main task is
to determine the sum of the phase contributions over both types of layers

(10)
n

where =n is generated by any rule for both lattices QA, QB.

The construction rule for any lattice Q can be given by various methods. An
infinite lattice Q is conveniently defined by its generations Qn and the rules for
each generation Qn must be defined. The corresponding lattice point positions
=n must be found afterwards.
This Fourier transform treatment will be compared in the following section
with the exact analytical formula for periodic and quasiperiodic Fibonacci MLs.
236 P. MikuHk

3.1. Periodic lattice

In the language of the physics of quasicrystals, some equivalent definitions

of the infinite periodic lattice P=ABABABABABABAB ... by means of its
generations Pn can be found:
1. recurrent: P 1 =AB, P n =Pn-1P n - 1,
. 2-- 1
2. the first analyttcal: P 1 =AB, P n =(Pt) ,
3. deflation (substitutional): P1(A,B)=AB, Pn(A,B)=Pn_1(AB,AB),
4. the second analytical: P n = ~1~2'" ~2-, where the objects ~j are
defined by

l(j + 1)/2J - li/2J = {~ ~~~ ~~ ~ ~ :

The function l x J denotes the integer part of x.

The layer positions are easily found ZAn = den - 1), ZBn = d A + den - 1).
The superlattice period is d == d A + dB. The phase sums SA, SB of the infinite
sublattices P A and PB are given by the geometrical progression

L [e-ikdr- 1= ~ L b(k - krr:) , (11)

n m

SB(k) (12)

The Fourier spectrum consists of a countable set of delta peaks distributed

equidistantly in the reciprocal space. Their positions k m = 27Tm/d are de-
scribed by one integer m.
The diffracted amplitude (9) of the infinite lattice P is then

pOO*(q)
m
FA(k m ) + FB(k m ) e-ikmdA
(14)
d
Let us now assume a periodic multilayer with finite thickness L . The shape
function treatment can be applied now. It gives

p*(q) (15)
m

(16)

The sum in Eq. (15) can be represented graphically (Fig. 2). At any recipro-
cal space point k all the multilayer Bragg peaks k m contribute to the diffracted
amplitude p* (k) with a weight given by its structure factor pr;;* multiplied by
the Fourier coefficient of the shape function Q*.
 SCATTERING ON APERIODIC SUPERLATTICES 237
1

0.1 1\ 1\ ~

0.01 fI 1\ fI {I,
/1\
fI fI 1\ 1\
11 1\
1\ {I,
vV' II t\
0.001
/ " (I, 1\ "

V '\ '~
",
\
/

1// II
0.0001 "-
"-~
wave vector transfer k [arb. units]

Fig. 2. - Graphical representation of the sum in Eq. (15) for the diffracted amplitude
of a periodic multilayer with finite thickness. The contributing maxima are distributed
equidistantly in the ID reciprocal spa.ce.

The previous shape function Fourier transform treatment can be compared

with the explicit calculation. The structure factor of the finite periodic lattice
P consisting of N = L / d layers is
N-l
P:xpl [FA + FBe- ikdA ] . L e- ikdn
n=a
[F F -ikdA] sin(kd(N - 1)/2) -ilcdN/2
A + Be . sin(kd/2) e . (17)

It can be easily shown that near the Bragg peaks k m only one term in the sum
in (15) dominates so that formula (15) is equivalent to the formula (17).

3.2. The Fibonacci lattice

The Fibonacci lattice F=ABAABABAABAAB ... is a well-known quasiperi-

odic lattice. Many equivalent construction rules have already been found [17]
1. recurrent: Fa =B, Fl =A, Fn =Fn- 1F n - 2 ,
2. deflatzon (substztuttonal): Fa =B, Fn(A,B)=Fn_1(AB,A),
3. analytical: Fn = ~1~2 ... ~Jk+~ with the following definition of ~j

LU + l)rJ - lJrJ = {~ i~~ ~~ ~ ~ :

The golden mean is r = xq-l. The formula for the Fibonacci lattice point
positions is also well known [i7] Zn = dB[n + (d A - dB)/dB . LrnJ].
238 P. MikuHk

The calculations of the phase sum (10) of the Fibonacci lattice have already
been worked out using different methods (the cut and projection method has
been established by Zia and Dallas [18] and Elser [19], the modulated phase
method by Levine and Steinhardt [20]). The resulting formula is

S(k, dA, dB) = f 2:

B pq
sine q,;q e- iiflpq / 2 6(k - kpq ) , (18)

where the maxima positions are K:pq = 27r(p/r + q)/d, and the phase q,pq =
27rq - kpq(d A - dB). Here p, q are integers and the superlattice period d '=
dA + rdB·
For the multilayer consisting of two different building layers the particular
phase sums SA, SB have to be calculated. The decomposition of the Fibonacci
lattice F into sublattices FA, F B is
sublattice FA: F = rlA'--B""I'--A"T'"IA-B--r-IA-B-rl-A"T'"IA-B--r-IA-rl-A-B"T'"IA--'I ...
sublattice FB: F = ~I BAA I BA I BAA I BAA I BA I··.
The sub lattice FA is the Fibonacci lattice of groups of layers 1 AB 1 and ~,
and the sublattice F B is the Fibonacci lattice of groups of layers 1 BAA 1 and
1BA I· This represents the self-similarity of the Fibonacci lattice in the direct
space. Rescaling the thicknesses allows one to express the sublattice phase
sums SA, SB by means of the sum (18)

SA(k, d A , dB, k;q(dA, dB), q,;q(dA, dB)) = S(k, dA + dB, dA) (19)
e- ikdA SB(k, dA, dB, k:q(dA, dB)' q,:q(d A , dB))
e- ikdA S(k, 2d A + dB, dA + dB) . (20)

The reciprocal space maxima of the phase sums for both sublattices

k;q kpq(d A + dB, dA ) = 27r(p + qr)/d , (21)

k:q = kpq (2d A + dB, d.4. + dB) = rk;q = k;'p_q (22)

are rescaled by the golden mean r. This represents the self-similarity of the
diffraction pattern in the reciprocal space. The particular phases are

(23)

The final formula for the diffracted amplitude (9) of the infinite Fibonacci
lattice is therefore

(24)
pq
 SCATTERING ON APERIODIC SUPERLATTICES 239
1 r-------------------------------------~

0.1

0.01

0.001

1e--l

le-5

wave vector transfer k [arb. units]

Fig. 3. - Graphical representation of the sum in Eq. (26) for the diffracted amplitude
of a quasiperiodic multilayer. The contributing maxima fill the ID reciprocal space
densely.

Compared to the structure factor of a periodic lattice (14) that of a Fibonacci

5L (25) exhibits a modulation by an additional sine function. The cut and
projection method shows that this is caused ·by the projection of a periodic
lattice in the higher-dimensional space into a quasiperiodic lattice in a lower-
dimensional space.

3.2.1. Finite length of the Fibonacci lattice

A finite thickness L of the Fibonacci lattice could be considered by employing
the shape function treatment. Calculation of the convolution gives

p*(q) = n* 0 poo* = L p~* . Lsinc (k - ~pq)L e- i (k-k,q)L/2 . (26)

pq

This formula can be represented graphically (Fig. 3) as in the case of the

periodic lattice. In contrast to the periodic lattice, the diffraction maxima of
the Fibonacci lattice are distributed densely.
The previous shape function Fourier transform treatment involved the ap-
proximation of cutting the finite part of the Fibonacci lattice from the infinite
lattice. Furthermore, there is a sum in (26) over an infinite number of pairs of
indices which cannot be calculated exactly, but only over a small number of the
pairs. For this reason, it would be useful to compare this approximate formula
with the exact one. The latter requires the calculation of the diffracted am-
plitude of the finite Fibonacci lattice layer by layer according to the recurrent
definition of the Fibonacci lattice Fk =Fk-1Fk-2. Consequently, the formula
240 P. MikuHk

for the diffracted amplitude of Fibonacci lattice is also recurrent

(27)

The first two diffracted amplitudes equal the structure factors of the building
layers :Fa = FB , :F1= FA and the phase ~n-2 = (d A fn-1 + dB fn-2) k. The
=
Fibonacci numbers fk are II 0, h = 1, fk =
fk-1 + fk-2·
The comparison of both formulae in the semi-kinematical approximation
was reported in Ref. [21] and a good agreement of both treatments has been
found. The transformation from the kinematical theory to the semi-kinematical
approximation will be treated in the following section.

3.2.2. Calculations in the semi-kinematical approximation

The kinematical theory for the calculation of the diffracted amplitude of the
Fibonacci lattice has been used until now. To calculate more precisely a Fi-
bonacci superlattice angular intensity profile measured by the diffractometer
along the truncation rod, the semi-kinematical approximation should be used.
Instead of the Fourier transform (10) the structure-geometric factor (6) has to
be calculated. Both formulae are equivalent if we replace the products d.4 k, dBk
of the kinematical theory by layer phase shifts <PA, <PB of the semi-kinematical
approximation (notation according to 2.2.1). The phase shift on each layer is
not separated into the product of a unique angularly dependent wave vector
and a variable depending only on the structural parameters of the layer. How-
ever, for infinite Fibonacci SL a possible approximation exists allowing one to
make a suitable decomposition enabling the use of the former results.
Let N; (N!!) denote the number of layers A (B) between the substrate and
the n-th layer. The phase shift <Pn of the n-th layer in Eq. (6) is expressed
by the layer phase shifts <PA,<PB to be <Pn = N;;+1<PA + N!!+l<PB. A similar
equation holds for the positions of layers Zn = N;;d A + N!!d B . The ratio of
occurrences of the layers B with respect to those of layers A in the infinite
Fibonacci lattice is the golden mean T = '4,-1. Under the approximation
that the same condition holds for each layer - n (N!! / N;; = T) the phase is
re-expressed

(28)

which is in the form consistent with the kinematical theory. The formula for
the wave vector

(29)

consists of two terms. The first one is the angularly dependent wave vector
departure on the substrate and the second is the shift from the exact Bragg
position due to the average multilayer deformation.
 SCATTERING ON APERIODIC SUPERLATTICES 241

....,>.
';a 0.1
..,v 0,1 0,-1
-1,0
.5 1.0
0.01 1.1 1.-1 -1,1
"..,
<l)

u
<tI 2,-2 1,-2 -2,2
-1

~ 0.001 -1,2
:..a 0,-2
0,2

le-4

le-5

le-6 ......._ -
-5000 -4000 -3000 -2000 -1000 0 1000 2000 3000 4000 5000
"., [arc sec]
Fig. 4. - Diffraction curve (diffracted intensity vs. angular departure 11 from the
Bragg angle) on the 13-th generation of the Fibonacci superlattice (substrate GaAs,
layer A is 42 A GaAs, layer B is 25 A AlAs, diffraction (111)). The diffraction peaks
according to the formula (30) are labelled by two integers p, q.

The exact diffracted intensity pattern in the semi-kinematical approximation

is given by the formula (5), where the structure-geometric factor (6) is calcu-
lated by (27). The diffraction (111) on the 13th generation of Fibonacci SL is
presented in Fig. 4.

3.2.3. Maxima of the diffraction curve of Fibonacci SL

According to the kinematical theory the positions of the maxima of the diffrac-
tion curve of Fibonacci SL are kpq =
27r(p + qr)/d. Using the relation (29) the
angular positions of maxima can be found [21)

-xo(l- ~)
T)pq = 2 sin 2(JB

In Fig. -! the peaks are labelled by p, q. The angular position of the main
= =
(zeroth order, p q 0) satellite is then

hZAfh dA~C.4. + dBr~cB

(31)
K sin 2(JB d

and depends on the thicknesses d A , dB and on the compositions ~CA, ~CB'

Furthermore, the angular departure of the pq-th satellite from the main one

~TJpq = T)oo - T)pq =

A' I'hl
-d ~2(J (p + qr) (32)
sm B
242 P. MikuHk

depends only on the averaged layer thickness d [22].

The formula (30) is well suited for fitting the measured diffraction curves
in order to estimate the real superlattice structure parameters. By fitting
the relative peak distances the averaged SL period d can be estimated. By
fitting the shift of the zeroth order satellite with respect to the substrate peak
position the ratio dB/d A , keeping d fixed, is expressed as the function of the
lattice parameters. To estimate them, it is necessary to fit the whole diffraction
curve.

3.3. Other aperiodic superlattices

The Fibonacci superlattice, which is the most common example of a quasiperi-

odic lattice, has been mainly considered so far. As already mentioned in the
introduction, other types of aperiodic lattices are also studied. Therefore a
short look at their diffraction properties will now b~ presented.
The straightforward generalizations of a Fibonacci lattice are the so-called
families An, Bn. The layer sequence in the family An is defined by the sub-
stitutional rule A -- AB n , B -- A, for family Bn holds A -- AnB, B -- A.
Their diffraction properties were studied by Severin [7]. He has found that
the diffraction properties of these two families are very different. By a matrix
multiplication scheme he has shown that the Bn family should be called an
aperiodic crystal, whereas the An family behaves like a quasicrystal.
For a general family defined by the substitution rule A __ Am Bn B --
1 1,

Am Bn the limiting numbers ml ... n2, for which the family is quasiperiodic,
2 2

can be found by means of the properties of the Pisot numbers [23]. Another
type of a generalization of Fibonacci lattice was reported in Ref. [8]. They d~alt
with the I-component Fibonacci structures obtained by a certain substitution.
The structure factor of ID two-letter substitutional systems were studied by
Kolar [24]. His approach involved different scattering factors and lengths of
the constituent tiles and he found that up to three indices should be used to
describe the peak positions. A three indices notation for peak positions has
been found necessary also for generalized Cantor superlattices [25].
The Thue-Morse sequence is automatic and non-quasiperiodic. There are
2n layers in the n-th generation stacked according to the substitution rule
A - AB, B - BA, starting from the sequence AB. Its Fourier transform has a
singular continuous measure. The analytical formula for the diffracted intensity
of point scatterers under a certain arrangement of this sequence has already
been reported [26], as well as the formula for the peak positions. The most
interesting feature of this lattice is a lack of intensity scaling proportional to
the second power of the total thickness, as in the case of quasiperiodic lattices.
The scaling exponent is different for different wave vectors. References [27-29]
have been devoted to a detailed analysis of the diffraction spectra of Thue-
Morse superlattices. The different physical properties (scattering powers and
thicknesses) of the constituent layers have been taken into account. The peak
positions in an infinite SL are labelled by three indices [24, 27]. Taking into
 SCATTERING ON APERIODIC SUPERLATTICES 243

account the finite sample thickness and the usual resolution of the measuring
equipment, the fine structure is smeared out and a description of peak positions
by two indices is satisfactory [28, 29].
Diffraction properties of abstract generalized Rudin-Shapiro sequencies have
been studied in Ref. [30]. The Fourier transform of this sequence has an abso-
lutely continuous spectral measure resulting in bounded intensity and no peak
stability under size changes. This resembles a random disordered system much
more than a crystalline system.
The diffraction spectra mentioned above were treated in the framework of
the kinematical theory. The dynamical theory was employed to analyze the
infinite Thue-Morse lattice by La Rocca et al, [31].

4. X-ray reflectivity

In contrast to X-ray diffraction, X-ray reflectivity is sensitive only to the mean

electron density distribution, hence it is used for the structural characterization
of both amorphous and crystalline SLs. The vector of the momentum transfer
between the incoming and outgoing wave lies near the origin of the reciprocal
space. In the case of specular X-ray reflection the momentum transfer is parallel
to the qz direction (which is parallel to the growth direction z), in non-specular
X-ray reflection it also has a non-zero in-plane component.
The calculation of the reflectivity curves is well-known from optics [32]. Par-
ratt [33] has proposed a similar recursion algorithm to calculate the X-ray
specular reflectivity. The recurrent formula is similar to the dynamical solu-
tion of diffraction on superlattices. If the reflected intensity is small (the angle
of incidence is much greater than the critical angle of total external reflection),
the semi-kinematical approximation can be introduced. The resulting explicit
formula for the reflected amplitude is
N 1-1
}( = L ne-''I'' , 'PI = - L 2k zj d j (33)
1=1 J=l

Here N is the number of layers, rl is the Fresnel coefficient of the I-th interface,
kZJ is the normal component of the incident wave vector and dj is the thickness
of the j-th layer.
The form of Eq. (33) is equivalent to that in the semi-kinematical approxi-
mation of X-ray diffraction (Eq. (5)). Consequently, the main features of the
reflection curve of an aperiodic superlattice are similar to those of the diffrac-
tion curve. For example, the maxima of the reflectivity curve of the Fibonacci
superlattice are given by the resonance condition

2nk z ,pq 21r(p + qr)/d , (34)

where the :: component of the vacuum wave vector depends on the angle of
2;
incidence () as kz = y'1- (cos ()/n)2, where n is the mean refraction index.
244 P. MikuHk
1

>.
.;s~

....S:::
<I) 0.1

...
'"0
<I)

u
<I)
<:I: 0.01 0.2
...
<I) 2.-1
3.-2

0.001
1,2
0.4
0.0001

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

angle of incidence [deg]

Fig. 5. - X-ray reflectivity curve of the 13-th generation of the Fibonacci superlattice
(substrate GaAs, layer A is 42 A GaAs, layer B is 25 A AlAs). The reflectivity peaks
according to the formula (34) are labelled by two integers p, q.

The reflectivity curve of the 13-th generation of the Fibonacci superlattice is

shown in Fig. 5.
Near the critical angle the intensity is nearly unity and the semi-kinematical
approximation is no longer valid. The reflectivity curve must be calculated
c

dynamically and the thickness dependence on the reflected intensity of the

Fibonacci superlattice can be studied. As the multilayer thickness increases,
the width of all peaks decreases. The most intense peaks near the critical angle
become saturated or grow very slowly, whereas the low intensity kinematical
peaks grow according to thickness squared.

5. Gratings

Up to now attention has focused on scattering by multilayers which represent

the artificial ID crystal in the growth direction. Other types of artificial ID
lattices exist such as lateral grating structures. The gratings are usually grown
and etched on the surface of a superlattice (Fig. 6). A grating consists of wires
which are infinite in one direction parallel with the surface and limited in the
other two directions.
According to the study of the kinematical diffraction on SLs by means of the
FT of the crystal density, a similar approach to treat the diffraction spectra
of gratings can be applied [34]. The density of the grating can be split into
density of the multilayer PML, shape function of one wire n and the distribution
 SCATTERING ON APERIODIC SUPERLATTICES 245

Fig. 6. - Drawing of a Fibonacci grating.

function of the wires Pd(X) =Ln b(x - xn) . The density of the grating is then

(35)

Diffracted amplitude is proportional to the Fourier transform of the density

(36)

According to the rule for the wire distribution, the intensity in the qx direction
in the reciprocal space is modified. The distribution of the intensity in the
reciprocal space in the case of the periodic grating consists of the periodic
distribution of the truncation rods in the qx direction. The peak positions are
qh ,m ,p= k - ko =h + qz,m + qx ,P' where h is the wave vector transfer at the
exact Bragg position, qz ,m determines the m-th maximum on the truncation
rod and qx,p determines the position of the p-th truncation rod. Similarly, the
peak positions of the Fibonacci grating (Fig. 7) are expected to be qh ,m,pq =
h+qz ,m +qx ,pq, where the distribution of the truncation rods in the qx direction
is labelled by two integers p , q causing a dense distribution of the truncation

---------j
' I
1
1
1

/
, '\
h/ /
/ /
/ /'
Fig. 7. - Outline of the diffraction pattern of a quasiperiodic grating. The method
of labelling the peak maxima is shown .
246 P. MikuHk

rods in the qx direction. However, the Fibonacci gratings have not yet been
grown, so the experimental test of this prediction has not yet been carried out.

6. Conclusion

X-ray elastic scattering techniques for structural characterization of aperiodic

superlattices have been presented. Concerning X-ray diffraction, both kinemat-
ical and dynamical theories have been treated. Intensities and peak positions of
the diffraction pattern of the Fibonacci superlattice, using the kinematical the-
ory and the semi-kinematical approximation ofthe dynamical theory, have been
calculated. Two methods for calculating the diffracted amplitude within the
kinematical theory and the semi-kinematical approximation have been shown.
In the first method. the Fourier transform of the infinitely thick multilayer has
been restricted by a shape function of a multilayer with a finite thickness. In
the second method the diffracted amplitude has been calculated layer by layer.
Numerical calculations have shown good agreement between the two methods.
The diffraction properties of other types of aperiodic superlattices have been
briefly reviewed.
In the study of the X-ray reflectivity the similarities with the diffraction
theory have been noticed, which allows one to use similar formulae for the
calculation of the amplitude of the reflected wave. Furthermore, grating struc-
tures, which are another representation of artificial one-dimensional lattices,
have been briefly presented and the quasiperiodic grating has been proposed.
The kinematical calculation of the reciprocal space diffracted intensity for the
Fibonacci grating has been treated and a dense set of truncation rods has been
predicted.

Acknowledgments

The author would like to thank Dr. T. Baumbach (Institut Laue-Langevin,

Grenoble, France) and Dr. V. Holy (Masaryk University, Brno, Czech Repub-
lic) for helpful discussions.

References

[1} Esaki L., Chang L.L., Tsu R., Proc. 12th Int. Conf. Low Temp. Phys.
(Tokyo, Japan, 1971) p. 55l.
[2] Merlin R., Bajema K., Clarke R., .luang F.Y., Bhattacharya P.K., Phys.
Rev. Lett. 55 (1985) 1768.
[3] Terauchi H., Sekimoto S., Kamigaki K., Sakashita H., Sano N., Kato H.,
Nakayama M., J. Phys. Soc. Jap. 54 (1985) 4576.
 SCATTERING ON APERIODIC SUPERLATTICES 247

(4) Todd J., Merlin R., Clarke R., Mohanty K.M., Axe J.D., Phys. Rev. Lett.
57 (1986) 1157.
(5) Terauchi H., Noda Y., Kamigaki K., Matsunaka S., Nakayama M.,
Kato H., Sano N., Yamada Y., J. Phys. Soc. Jap. 57 (1988) 2416.
(6) Terauchi H., Kamigaki K., Okutani T., Nishihata Y., Kasatani H.,
Kasano H., Sakaue K., Kato H., Sano N., J. Phys. Soc. Jap. 59 (1990)
405.
(7) Severin M., J. Phys.: Condo Matter 1 (1989) 67i1.
(8) Hu A., Wen Z.X., .Jiang S.S., Tong W.T., Peng R.W., Feng D., Phys.
Rev. B48 (1993) 829.
(9) Axel F., Terauchi H., Phys. Rev. Lett. 66 (1991) 2223.
[10) Merlin R., Bajema K., Nagle J., Ploog K., J. Phys. (Paris) 48 (1987)
C5-503.
(11) Azaroff L.V., Kaplow R., Kato N., Weiss R.J., Wilson A.J.C.,
Young R.A., X-ray diffraction (McGraw-Hill, New York, 1974) p. 217,
p.222.
(12) Holy V., Kubena J., AbramofE., Lischka K., Pesek A., Koppensteiner E.,
J. Appl. Phys. 74 (1993) 1736.
(13) Pinsker Z.G., Rentgenovskaja kristallooptika (Nauka, Moscow, 1982)
p. 167.
(14) Taupin D., Bull. Soc. Fr. Mineral. Crystallogr. 87 (1964) 469.
(15) Bartels W.J., Homstra J., Lobeek D.J.W., Acta Cryst. A42 (1986) 539.
(16) Speriosu V.S., J. Appl. Phys. 52 (1981) 469.
(17) Steinhardt P.J., Levine D., The Physics of Quasicrystals .(World Scien-
tific, Singapore, 1987).
(18) Zia R.K., Dallas W.J., J. Phys. A. (UK) 18 (1985) L341.
(19) Elser V., Acta Cryst. A42 (1986) 36.
[20) Levine D., Steinhardt P.J., Phys. Rev. B34 (1986) 596.
[21] yIikulik P., Holy V., Kubena J., Ploog K., Acta Cryst. (1995), in press.
[22) Tapfer L., Horikoshi Y., 14th Symp. GaAs and reI. compo (Heraklion,
Greece, 1987).
[23) Bombieri E., Taylor .1 .E., Can temp. Math. 64 (1987) 241.
[24) Kolar M., Phys. Rev. B47 (1993) 5489.
[25) Jarrendahl K., Birch .1., Dulea M., Arwin M., Sundgren J.-E., Thin Sol.
Films 264 (994) 120.
[26) Cheng Z., Savit R., ~Ierlin R., Phys. Rev. B37 (1988) 4375.
[27) Kolar M., Phys. Rev. Lett. 73 (1994) 1307.
[28) Axel F., Terauchi H., Phys. Rev. Lett. 73 (1994) 1308.
[29] Peyriere .J.. Cockayne E., Axel F., J. Phys. (Paris) 5 (1995) 127.
[:30) Axel F., Allouche .1 .-P., Wen Z.-Y., J. Phys. 4 (1992) 8713.
[:31) La Rocca G.C., Tapfer L., Cingolani R., Ploog K., Phys. Rev. B45 (1992)
12198.
[32) Born ~I., Wolf E., Principles of Optics (Pergamon Press, 1993) p. 51.
[:3:3] Parratt L.G., Phys. Rev. B95 (1954) 359.
[34J Gailhanou M., Baumbach T., Marti U .. Silva P.C .. Reinhart F.K ..
Ilegems M., Appl. Phys. Lett. 62 (1993) 1623.
 COURSE 10

Defects in quasicrystaJs, in systems with

deterministic disorder and in amorphous materials

Nicolas Rivier

Laboratoire de Physique Theorique,

Universite Louis Pasteur,
67084 Strasbourg, France

1. INTRODUCTION

What are (topological l ) defects in matter ? Mathematically, they can be

defined as singularities of a function 4>(x) associating a quantity, number,
vector, etc., 4>E M to any point x inside the material. 4>(x) represents the
configuration of matter in the material. It is better to speak of topological
entanglement instead of singularity, since we will encounter defects which
are not singular. The aim of a general theory of defects is to associate
topological entanglements to the set M of microscopic states of matter, and
to classify them.

Physically, defects are visible objects which can move in response to

outside sollicitations, and determine strength, toughness or resistance of
the material. The physical properties of a perfect, defect-free material are
the result of a collective organisation. Thus, a perfect crystal would be very

1 Typical examples are dislocations in crystals, or disclinations in liquid crystals. Point

defects can also be regarded as topological defects. They are classified by the second
homotopy group 1t2 [1], or, in the framework of gauge theories, by the "extra-matter" (Wey!)
tensor. See, for example [37], pp.64-66, or [68], §12-13.
250 N. Rivier

stiff under shear, and fail dramatically when the shear stress exceeds the
threshold needed to make a whole plane of atoms slide over the next. By
contrast, the dislocations (line defects) in an imperfect crystal can easily
glide as a response to shear. The response of the material to shear' is
determined by motion of dislocations, and by impurity atoms (point defects)
which can prevent this motion by pinning down the dislocations. The
physical properties of a material are modified quantitatively and controlled
by its defects.

The aim of these lectures is to uncover which types of defect occur in a

given type of condensed matter, crystalline, quasicrystalline or amorphous
(disordered). A technique based on homotopy theory [1-7] (§2)1 is general
enough to classify defects, at least in principle. In practice, however, this
must be qualified:

1) Homotopy theory relies on a continuous space {x}, in which matter (the

field 4>(x» lies condensed. This is a questionable assumption for tilings, and
at best misleading in covalent glasses (averaging at an early stage may
eliminate the defect itselO.

2) The second qualification is that homotopy theory classifies (line) defects

by the ru:4fi that they upset2 . What if matter is amorphous, if its global
symmetry H is trivial, as in glasses? There are line defects in glasses, in
spite of their complete lack of order. In quasicrystals, the essential symmetry
is inflation, but it is hidden in the irrational slope of the oblique tiling

1 The original papen are [1-3]. Reviews by Mennin [4] and by Michel [5]. There is an elementary
introduction by Toulouse [6].
Classification of ordered phmJ. by orbits of groups has been done first by Golo and Monastynky [7].
For a simple introduction to homotopy, see Stewart's book [27].
2 For example, a dislocation is a translation defect in crystals. It is dermed by a Burgers
vector, which is an elementary translation of crystal which it dis-locates. This statement
is expressed mathematically by the expression "1 (OIH)=H", which we will encounter in
§2.2. Here OIH is the set of possible local values, orientations, etc. of the parameter
describing the state of the material at every point (the 'order' parameter). The material,
when perfectly ordered, has a symmetry or order characterized by the group H of
transformations leaving it unchanged. Line defects are classified by the first homotopy
group or fundamental group of OIH, which is, generally, the group of symmetry H itself
(possibly extended (lifted».
For a visual introduction to dislocations and topological defects in crystals, see the
article of Bragg and Nye [64]. For disclinations in natural structure, see that of Harris
[65].
 DEFECTS IN QUASICRYSTALS ... 251

separating high-dimensional superspace (with translational symmetry) into

parallel (physical space) and perpendicular (atomic surface) components.
Defects (dislocations) and symmetry (translation) are rather dull in
superspace. It is only in real space that they appear in their full
'questionnable shape' (phason flip, full or partial dislocation), with
remarkable metallurgical properties (hardness, followed by catastrophic
brittle-ductile transition at high temperature [8]). We will see that defects
should be studied in the (oblique) tiling, where they are elementary and
topological. Metric symmetry H is too fine to classify all defects, topological
groups suffice. This will enable us to talk about defects in topological,
disordered and discrete tilings or foams, and to deal with discrete space. The
symmetry then is local instead of global.

3) Homotopy theory proceeds against the empirical and historical trend. It

postulates a field cI»(x) first, i.e. begins by stating the microscopic elements
of the condensed matter under scrutiny. It then proceeds to classify the
defects which a dense assembly of such microscopic objects supports, in an
elegant and direct fashion. However, historically, macroscopic properties,
dominated by defects, led to the microscopic constituents, not the other way
round. Recall liquid crystals, where the observation of the defects led
Georges Friedel [43] to their microscopic structure (nematic and smectic) at
the turn of the century, or antiferromagnetism, for which microscopic
evidence (by neutron scattering) only appeared in 1949, nearly twenty years
after Neel's thesis [63]1.

These notes are organised accordingly: The first part sketches the classic
homotopy theory of defects in ordered condensed matter in a continuum. We
will proceed to quasi crystals, and study defects in (oblique) tilings; then in
any tHings (the defects are the same, local and independent of order).
Finally, we will see how disorder and its local invariance enables us to deal
with the discrete scaffolding connecting the microscopic object (atoms,
molecules, spins) constituting matter and with its defects.

Excellent reviews exist for each section. Homotopy theory of defects is

covered at any level by ref. [1-7]. Defects in quasicrystals have been studied
in detail, whether as phason flips (basculons) [9,10], or as dislocations [11-
16, 19]. The oblique tiling is now well understood, at least by its main
authors [17,18]. Topological defects in glasses, amorphous materials [20] and
foams [21,20] have also been reviewed. Motivated by the absent or
complicated global symmetry and by the topological stability and possible

1 A single picture to launch two Nobel prizes (N6el's and Shull's), quoted by a third [63].
252 N. Rivier

disorder in our materials, we gather here these various techniques, albeit at

a much lower level.

Recurrent in the description of any topological defect is the notion of a

closed contour surrounding it and causing a finite change of the local state
of the material For example, the local velocity in a fluid rotates by 27t
along any contour surrounding a vortex. This is expressed mathematically
by the notion of fibre bundle [66,37] : A closed contour (in base space)
around a defect results in a change of some local parameter (on the fibre),
which characterises the defect. For example, a staircase is a defect of a
building, which is a fibre bundle with the floor plan as base space and the
elevation as fibre. Base space can be continuous or discrete (as in glasses),
concrete (ordinary euclidean space occupied by matter) or abstract (it may be
the atomic surface in quasicrystals or even the shape of an amoeba [67]). We
will proceed from physical examples. The common underlying concept will
hopefully become apparent, and the material will always described first and
foremost by its visible defects.

2. DEFECTS IN CONDENSED MATTER

Condensed matter presents visible and permanent (in solids) or slow-moving

(in fluids) configurations of atoms, spins, etc .. It has a permanent or long-
lasting geometry, which is visible and accessible experimentally.

A closed contour in R3 (whether a loop (one-dimensional) or an envelope

(two dimensional» encloses a configuration which mayor may not contain a
defect. It is possible to define invariants which are independent of the
actual contour and measure only the defect which it encloses. For example,
a charge or a current line correspond to configurations of the
electromagnetic field (divergence or rotational, respectively) which are
singular. The integral of the electric field over any envelope, of the
magnetic field along any closed contour, respectively, is a measure of the
defect, charge or current respectively, which they enclose. And this measure
is independent of the actual contour, as long as it encloses the defect.

Ordered condensed matter is described and classified by crystallography,

essentially the compatibility of rotation and translation symmetries, or, in
quasicrystals, or rotation and inflation symmetries [22,23]. It is therefore
based on global and constructive symmetries, on the action of a group of
transformations which leaves the crystal strictly invariant, i.e exactly
superposable on the original. By contrast, disordered condensed matter has
 DEFECTS IN QUASICRYSTALS ... 253

a different type of translation and rotation symmetry, which is a local,

gauge invariance. The translated glass is not exactly superposable on the
original, but the difference is not physically relevant or observable. In
ordered condensed matter, the symmetry is metric, although only its
topological features are relevant1 In disordered matter, it is of course only
topological.
Defects upset this symmetry. Let us see a few examples.

- A 2d magnet, when ordered, has all its spins aligned. Small fluctuations, in
the absence of topological defects, are described by a phase angle 9(s:) in
each point, with the magnetisation M(s:)=grad 9(s:). There are two types of
defects, divergence (monopole) and vortex, whereby CD = curl M O. Then, '*
grad 9(s:) is not defined at s:=O, the core of the defect .
- Dislocation in a crystal. The contour consisting of an equal number of up
and down, left and right steps (anholonomic coordinate dX<l) would close in
the absence of defect. The torsion, or lack of closure is called the Burgers
vector. Note that it is a symmetry element of the perfect crystal.

In 3d, the dislocation, enclosed by a contour, is a line defect. A

monopole, enclosed or bounded by a surface, is a point defect. A domain
wall, bounded by two points, a zero-dimensional contour, is a wall defect.

In general, the dimensionality of space d, the dimensionality of the

defect dd and that (dr=n) of the contour r enclosing the defect (the
codimension of the defect) are related:

n=d-dd-l (1)

2.1 Homotopy description of defects [1-6]

If matter occupies continuous space. Rd, it can be described by an "order"

parameter w(s:), called Ginzburg-Landau order parameter or Higgs field. cl»(s:)
is a map from Rd into M. M is the manifold of all 'orientations' or non-
vanishing values of the order parameter. At different points, the order
parameter may take different values, and this inhomogeneity in condensed

1 Classification of the space groups in crystals relies only on the existence of a minimal
distance between atoms. One can go quickly a long way by finding all discrete subgroups
of SO(3) [24]. The same holds for quasicrystals (combination of rotations and inflation
symmetries) [22,23].
254 N. Rivier

matter costs some energy (proportional to la4>(x)12). 4>(:1:) describes an actual

configuration of matter, including its defects. A defect is a singularity of
4>(:1:).

Often M can be specified further as the set GIH of minina of a Landau

free energy v(T<Tc,4». G is the full symmetry group of the disordered phase
at high temperature T, and H is the broken symmetry group, a subgroup of
G that leaves invariant an uniform configuration 4>. H is the space group of
crystallography, that subgroup of the high-temperature symmetry group G
which survives symmetry breaking below Tc.1

- For example, in a 2d magnet2 , 4>(:1:) is the planar magnetisation M(:I:), and

M = 8 1 (a circle). The configuration of the magnetization around a (point)
defect is either polar (a divergence as in Fig.I) or rotational. A closed
contour on the plane corresponds (through the map 4>(:1:» to a closed
contour in M. Contours in M which can be continuously deformed into each
other describe the same defect, labelled by 7tI (M), the fundamental group or
first homotopy group of M. For example, a contour not surrounding any
defect on the plane corresponds to a loop on the circle M which can be
shrunk to zero. By contrast, any contour around a defect like that of Fig.I

Fig.I - A defect in a planar ferromagnet (left). Right: The image by 4>(:1:), on

the circle M, of two closed contours in R2. The contour not surrounding the
defect can be shrunk continuously. That surrounding the defect is
unshrinkable. It winds once around the circle M and this rotation
characterizes the defect.

1 The original papers [1-7] assumed that matter had to be ordered. It is in fact not
necessary; homotopy theory has been applied successfully to spin glasses [25] and
glasses [26].
2 G=SO(2), H=I, M=SO(2)=SI
 DEFECTS IN QUASICRYSTALS ... 255

corresponds to a closed loop winding Z times around the circle M. It is clear

that the fundamental group, and defects in a planar magnet, are given by
the positive and negative) integers Z, and that the divergence or rotational
of the magnetization is quantized (because the structure of condensed
matter is extended and must fit without mismatch) and additive. Thus, for a
2d ferromagnet, 1tl(M=SI) = Z. The defect is a point (a divergence of the
magnetisation field) in d=2, n=1.
cII(x) also maps a closed n-contour in Rd, r n",sn into a closed n-contour r n'
in M. Now defects of codimension n (enclosed by an n-dimensional contour)
are classified by the group 1tn(M)=[Sn,M]. The elements of this group are
the classes of closed n-dimensional paths in M, in which paths are identified
which can be smoothly deformed into one another (homotopic paths). These
classes form a group. Combination of closed paths is clearly a closed path as
long as an arbitrary, common base point is chosen. The identity contains the
trivial, infinitesimally small loop which takes no time to traverse, as well as
all loops which can be deformed into it. The inverse of a loop 1-1 is that loop
traversed backward (or any other loop homotopic to it), because the
combination 101- 1 is smoothly deformable to the trivial loop. The group
operations are illustrated in Figs.2 and 3. In particular, line defects in R3
(d=3, dd=l, thus n=l from eq.(1» are classified by 1tl(M) = [Sl,M], the
fundamental group of the manifold M (oflocal matter configurations).l

- Examples: 3d nematic liquid crystal. The order parameter is the director2, a

e o
vector without arrow, or a spin with its two directions identified. The manifold M is

,
Fig.2 - Cartoon showing that 1t1 (RP2) = Z2. Antipodal points, identified, are
used to turn the path around. (After [27]).

1 Warning: In this chapter, 1t1 denotes a homotopy group (for which it is the standard
notation), rather than a projection, as elsewhere in this volume.
2 G=SO(3), H=Doo, M=SO(3)/Doo=RP2
256 N. Rivier

a sphere (specifying the possible orientations of a spin) with antipodal points

identified, i.e. the real projective plane RP2. Now 7t1(RP2)=Z2. (Fig.2). The two defects
can annihilate (to see it, the third dimension must be used)
- 3d magnet, 7t1 (M=s2)=I, there are no line defects in 3D (Heisenberg) magnets. There
can be monopoles as we shall see.

7ti(M) is always abelian for htl, but 7t1(M) can be non-abelian. This leads
to topological obstruction between two line defects moving against each
other. This obstruction is physical. It occurs in biaxial nematics [28,29],
where 7t1(M)=Q, the group of quaternions. M=SO(3)1D2 ",SU(2)1D2''', where
H=D2 is the point group or group of covering transformations of the order
parameter, i.e. rotations by 7t about each axis, and D2"={rotations by n7t
about each axis} is D2 lifted so that rotations by 27t equal (-I) and rotations
by 7t about the three axes are the three Pauli matrices. D2" has 8 elements,
and is isomorphic to the quaternions, D2""'Q =7t1 (M).I Then, two defects
labelled by p,q E Q cannot cross each other if qpif'pq. Here, qp and pq
represent defect p in front and on the back (before and after the
hypothetical crossing. Fig.3d). Actually, the two defect can cross each other,
but at the cost of an additional defect linking the original defects, labelled
by K(p)p-I=qpq-Ip-I (Fig.3), which is the commutator of the two defects p and
q [30,31,32]. rlA<rIBtl=qpq-1 is the action of q on p when the origin of
contour p is passed around q (Fig.3b). For the quaternion group, the only
non-trivial commutator qpq-Ip-I is -1, and the additional defect is a 27t-
di sclinati on. If 7t1(M) is abelian, the additional defect qpq-Ip-I=I is trivial.
Non-commutativity will also be the local signature of disorder in glasses
(§ 4.3, Fig.14). If it is reassuring that the the algebra of crossing and of the
action of the additional defect is a commutator of closed paths qP=1C(p)p-l.pq,
it is not something that one would write down by inspection.

The classes of closed paths form a group. Each class characterises a defect
uniquely. Take two contours rIA and rIB about the same line defect. The
composition rlA<rlBtl",1 is clearly shrinkable (FigA), so that rlA""rIB.

Similarly, an envelope S2 surrounds a point, so that 7t2(M)=[S2, M]

classifies point defects.
- Example : 3d ferromagnet: 7t2(M=S2) = z.

1 This results from the exact sequence (2) and from the fact that G"=SU(2) has been lifted
to be simply connected. See § 2.2.
 O)f
DEFECTS IN QUASICRYSTALS ... 257

CjsP --'>'
d~(p)
~llr 911r
b)
~-' = d:~(,)P'
91 \p
;:

9
IK~ p 4

tJII
,)

~')
11
(jP ~

"
II

~}M9
~
I :l
") ><\1 r><1
9P

-----7'

t
Fig.3 - For two line defects p and q in 3d to go across each other without
hindrance (d) , their labels (elements of 7tl(M» must commute.
a) When the origin of the contour encircling defect p is moved around defect
q, the labelling of the defect changes from p to lC(p). This move corresponds
to the observation of defect p before and after it crosses q.
b) The additional defect (thick line) linking p and q after their crossover is
lC(p)p-l (operating from right to left).
c) Shows that qP=lC(p)q, or lC(p)=qpq-l. The additional defect is the
commutator qpq-lp-l.
258 N. Rivier

FigA - A unique contour

characterises a line defect :
-1
r 1A ",r 1B since r 1A (rlB) ",1 is
shrinkable.

Fig.5 - Grain boundaries (arrays of dislocations) in a daisy. Inset: A

dislocation is a dipole pentagon-heptagon. The pentagon/heptagon is a
positive/negative disclination.
 DEFECTS IN QUASICRYSTALS ... 259

Two points, SO surround a wall, and 1tO(M) = [SO ,M) characterises wall
defects. 1t0 is usually only a set, not a group. Walls may have their own
internal structure, for example, grain boundaries betwen crystals are
quasicrystalline [33). Quasicrystalline grain boundaries can also be seen in
composite plants (a crystallography of identical florets, with visible spirals,
the parastichies, taking the part of reticular planes in crystals) [34). (Fig.5).

- Example: Ising magnet, cl>(:a:)=(±l}, M=Z2· 1t0(Z2)=Z2.

1t3(M) classifies textures, nonsingular, yet topologically entangled

configurations.
- Example : Nematic liquid crystals, M=RP2=S2/Z2 (identification of
antipodal points; The order parameter is a thin stick with an orientation
but no direction). 1t3(RP2) = [S3,S2/Z2) = Z . The celebrated map from S3 into
S2 is Hopfs, who described the textures [S3 ,S2) and predicted that non-
trivial textures could be seen by the linkage of the locus of points with
director in one given orientation, inverse image cl>-l of one point of RP2,
which form two loops each corresponding to opposite directions. These
linked loops were observed by Bouligand [35) (Fig.6)l, and interpreted later
by homotopy theory [36).

Fig.6 - The linked 'olympic rings',

evidence of a Hopf texture, observed in a
liquid crystal under polarizing microscope
by Bouligand [35). One black line is the
locus of "up" directors, the other, the
locus of "down" directors. Their linkage
is the index of the texture. The liquid
crystal (MBBA) is a long pitch
cholesteric rather than a nematic (see
footnote this page). (Courtesy :
Y. Bouligand, with permission from
Journal de Physique [35)).

1 The liquid crystal used by Bouligand was a cholesteric, rather than a simple nematic
(with a long helix as order parameter, rather than a straight segment). Cholesterics have
the same M=Q as biaxial nematics. However, they do not like to have singularities on
the direction of the order parameter, so that, at low energy, one observes non-singular
configurations of directors, i.e. Hopf textures.
260 N. llivier

Examples (all 3d) M 1tO 1t3

Ising magnet Z2 Z2
planar (xy) magnet 81 Z
Heisenberg magnet 82 Z Z
(Hopf)
nematic RP2 Z2 Z Z
dislocated crystal RlZ+R/Z+R/Z Z+Z+Z
(transl.
group)
glass1 80(3) Z2 Z

Trivial homotopy groups are denoted by -. In the case of dislocations, note

that line defects are characterised by the symmetry which they upset. Any
element of RlZ+RIZ+RIZ represents an arbitrary displacement, reduced to
the crystal unit cell (fundamental domain). This is in fact a fairly general
observation, justified in § 2.2.

2.2 A simple method for calcu1ating homotopy groups

It is based on the exact sequence of homomorphisms (Le. maps) between the

homotopy groups of the fibre bundle H -+ G -+ GIH (fibre -+ total space -+
base space), namely

... -+ 1tn+ 1(O/H) -+ 1tn(H) -+ 1tn(O) -+ 1tn(O/H) -+ 1tn-l (H) -+ ... (2)

A sequence A (f)-+ B (g)-+ C of group homomorphisms is exact if Im(f)=Ker(g).

The trick is to select a fibre bundle such that as many as possible of the 1t'S
in the exact sequence are trivial:
- (i) G is usually a connected (1tO(G)=I) Lie group (with 1t2(G)=1 (Cartan»,
which can be lifted to its universal covering group GA, such that 1tl(GA)=1.
A circle can be lifted on a spiral, Le. on R.
- (ii) H is often a discrete group (the point or space group in
crystallography), thus 1tiiFO(H)=1, 1tO(H)=H.

Then,
if 1 (f)-+ B (g)-+ 1, B=1
if 1 (f)-+ B (g)-+ C (h)-+ 1, B=C and map g is 1:1.

1 Here, glass has been modelled by an elastic continuum with trivial space group. See
§4.2 for a more realistic description (as a random covalent network). Defects are
unchanged.
 DEFECTS IN QUASICRYSTALS ... 261

Thus, if M=GIH=GA/HA, 7tl(M)=7tO(H)=HA, line defects are labelled by the

symmetry which they upset, and 7t2(M)=1, there are no topologically stable
point defects. This enables us to complete the table above. (Recall that
80(3)=8U(2)/Z2, 8 2=80(3)/80(2), RP2=82/Z2, 8 1=RlZ).

2.3 Defects in crystals: Dislocations

Line defects are labelled by the symmetry H which they upset. Namely, 7tl is
the space group H (strictly, the space group HA lifted (extended) so that the
lifted group GA of all possible excitations (phonons, phasons, etc.) is
connected and simply connected. (§ 2.2). Let us see how this works out in
crystals.

Line defects in 3D crystals are dislocations and disclinations (rotation

dislocations), corresponding to the translation and rotation symmetries of
the material. Also, as two opposite rotations constitute a translation, a
dislocation can be decomposed into a dipole of two, opposite, topological
disclinations (Fig.5).

cI>(y) is the map from the actual, stressed and deformed state {y} of the
crystal into its natural state where all elastic distortions have been relaxed
(while topological defects, called plastic strains in elasticity, remain). cI>(y)
(defined by dXa.=cI>a. i dyi) contains all the local information. The metric of
the actual state gij=cI>a.icl>a.j =Oij-2eij, is related to its strain ,tensor eij- The
map is invertible everywhere, except at the core of the defect. The
coordinates are {yi} in the actual state, and the steps {dXa.} in the natural
state. {dXa. }are anholomic, non-integrable, and the lack of closure or
mismatch of the image r' of a closed contour r is the Burgers vector ba.
labelling the dislocation. It is, as stated, an element of the translation
group, an integral number of steps {dXa.} in the natural state.

(3)

2T<Xij is, physically, the density of dislocations. It measures nonintegrability

or torsion. A similar contour made around a disclination yields the density
of disclinations, the curvature Ra.~ij. R and T obey Bianchi and torsion
identities, stating, respectively, that disclinations close as loops, and that
disclination lines may end on dislocations (or, alternatively, that a
dislocation line is a double line of two disclinations of opposite sign). Note
262 N. Rivier

that disclinations cost a prohibitive energy in hard, three-dimensional

crystals. Not so in two dimensions, or in soft crystals, or in glass where their
stress is screened by dislocations. (See [37] for a review).

In this case, the base space of the fibre bundle is the actual state of the
crystal. The total space is the natural state, and the fibre is the set of all
possible Burgers vectors, the translation group of the crystal. We see that if
a closed contour r in base space (in the actual crystal) encloses a
dislocation, it fails to close in total space by a discrete amount, the Burgers
vector, which is a translation on the fibre. We see also that a path in the
actual crystal can be naturally lifted in total space by eliminating elastic
strain, through the connection, the steps (dXa ).

The two descriptions of defects, Burgers contour and homotopy theory,

are equivalent, even though for the former the image contour f' does not
close, whereas it does for the latter. The equivalence is given by the exact
sequence (2) of homotopy groups of lifted groups Gh and Rh. Defects are not
simply a singularity in some pattern, but an effect of a physically broken
symmetry.

Figure 7, Pieranski's gravity rainbow, illustrates the actual state of a

soft crystal deformed by gravity, but without defects (or nearly so). This fact
makes the map cI>(y) invertible and integrable everywhere, and the natural
coordinates (Xa) holonomic. In the absence of defects, the actual state of
the two-dimensional crystal is a conformal transformation of the perfect
triangular lattice [38].

··.........
....
....... ...............
.'... ......................
. ........... . .. .....
............
............. '.
'
·. ......... ..............
~
~ ' ,
'.'
' ' '.
:.::............... I, :::.:: I' .....I, II':!':'::'::': ......... II II.
'.' •••••••• ••••• • •••••••• ••••• . ' • • • • • • • • '. "l
~:.:. ~ ~.:
::.::.:::::......:.:. ~ ~~::! :.:::::::..... ',_:.:!:: ::::::::::::.......I.: ::::
:::: :'.:.:.:.:.:.:.:.:.~'.~~:::!:::::::~.:.:.:.:.:.:.:.'.\~~::;:::: i:'::::.:.:.:.:.:.:....~\~:::
t:.:::::::::::::............... 1
_' .::... ·i···~-:···~~~·::,~i:
.::::::::::::...... .. t
~ ~: t ::::::::::.": ...............:. ~ ~! !~:::
11 ·············:·.:·····ll··llffl;i~:~~\'0.\~illi!li#.4'ffb~»::~~~\R\
··················· "
...·::.·.·..... ..
...............................................
~ ! •• :':'.: •• :···········\~i·:

tI 'I.~~IWI '. .I!);mIImi~ailllMMlrtl!l{wiiiiiIDf!'l.ll'J;ti1

Fig.7 - Piotr Pieranski's gravity rainbow: A soft, (almost) defect-free crystal

under gravity. It is a conformal map w=(-i/a)ln(az) of a perfect triangular
lattice (z), the natural state of the crystal.
 DEFECTS IN QUASICRYSTALS ... 263

The reader will have recognized in eq.(3) the analogue of Ampere's law in
electromagnetism, with b, 2T and cI» playing the parts of magnetic flux,
magnetic field and electric current, respectively.

3. DEFECTS IN QUASICRYSTALS

We begin with a justification for working with a one-dimensional

quasicrystal, represented by a two-dimensional oblique tiling. This can be
omitted by the converted or at first reading.

The d- (= 6-, 5-, 4- or 2- dimensional) superspace is a lattice Ad. It has

translation sym-metry and its defects are dislocations. Disclinations are, as
we have argued, energetically too costly to occur in three dimensions and
above. We might think that dislocations, of codimension 1 (enclosable by a
Burgers contour), are d-2-dimensional objects (Eq.1). This is not so, because
superspace decomposes into parallel (real) and perpendicular (atomic surface)
subspaces, intersecting at points (the atoms themselves). Thus, a one-
dimensional path (for example, a segment in superspace with some
component along real space) is a path in real, parallel space Ell, which is
lifted in superspace to become part of a Burgers contour in Ad. DiBloca.tionB
in Ad must therefore be surrounded by a contour in real space, and a.re
therefore objectB of dimenBion dl,2, namely line defects for icosahedral
quasicrystals (as observed [15,16,39]), points for octagonal, decagonal and
dodecagonal quasi crystals, and textures for chains. 1

A consequence of the zero-dimensional intersection between parallel and

perpendicular spaces is that we have a fibre bundle 2 , of total space Ad, with
Ell as base space and the atomic surface E.l as fibre. Figure 8 is a schematic
illustration for d=4 (a caricature of an octagonal quasicrystal). Of course, the
lift from a path in E II to A d is rather tricky: it contains the full tiling
geometry and matching rules. Also, the lifted path is extremely jagged in
the fibre (see the simplest A 2 example below); it has been described

1 It follows that the (d-2)-dimensionaI dislocation core in superspace intersects Ell

perpendicularly. KIernan [14] argues that it must be so because physical strain in
quasicrystals is isotropic. It is in fact another immediate consequence of atomic surfaces
intersecting Ell at points.
2 The fibre bundle is dermed (in an example) in §4.2. For an introduction (without any
examples from condensed matter). see Bernstein and Phillips [66]. See [37] for
applications of fibre bundle and gauge theory to (disordered) condensed matter. An
excellent illustration of fibre bundle is in swimming [67].
264 N. Rivier

masterfully by Katz [10, §IV, Fig.21,40], but the principles are all there and
easily represented in 2d.

~I
Fig.8 - Cartoon of parallel and perpendicular spaces and their intersection,
showing that they constitute a fibre bundle.

Alternatively, one could regard the infinite periodic minimal surface in

E.L formed by the atomic surface [40], as the base space of the bundle and
real space En as its fibre. This provides a direct interpretation of the phason
flip as a translation on the fibre upon closed contour in the atomic surface,
but I find it counterintuitive to have to jump about in real space when
following a continuous path in base space. Why should one care whether an
extremely jiggly path closes or not? Yet, when it does not, we have a
phason flip, somewhere. This fibre bundle representation is implicit in
Oguey [40], and explicit in Kleman [14] and in Katz and Gratias [10,
unpublished part of the lecture].1

The final, overriding reason for describing defects in 2D, with one-
dimensional real space (quasiperiodic sequence) and one-dimensional atomic
surfaces, is that Pleasants [22,23] (see also [33]) has shown that
quasicrystals are direct products in real space of inflatable, substitution
sequences, on directions separated by an angle 27r1s, with s=5, 8, 10 and 12

1 The atomic surface is an edge-on edge tiling by octagons. The octagon is the
fundamental regions or unit cell of the tiling. This tiling {8,3} (notation defmed in note
2, following page) is hyperbolic, and covers an infmite, periodic minimal surface
(IPMS) without overlap [48,40], so that the same point on E.L may correspond to
different points in Ell, depending on the path taken to reach it from the original
octagonal fundamental region. This is a phason flip [9]. Fundamental translations on
IPMS do not all commute [48].
 DEFECTS IN QUASICRYSTALS ... 265

only. For each sequence, an oblique tiling can therefore be constructed with
the true atomic surface along E.1 intersecting the real line at points.

3.1 '!be oblique tiling

The actual quasicrystal lives on a line En of irrational slope. Coordinates in

superspace are (XII, X.1) forming naturaUy the oblique tiling of Kramer and
Katz [17,18]1. In 2d, it is the tiling (4.8.8).2 It has vertex coordination z=3,
and is therefore topologicaUy3 and dynamically stable (which is not the case
of the Four-Corner boundary between Utah, Colorado, New Mexico and
Arizona). The oblique tiling has its edges on Ell and E.1, most of which are
atomic surfaces and phason flips (basculons). (Fig. H).

Because superspace has translation symmetry, defects are dislocations (in

sUlerspace) of Burgers vector b, an integer combination of unit vectors of
A . Because real space has an oblique slope in Ad, b has always both
"phonon" bn and "phason" b.l components. These have been seen by
diffraction contrast analysis [15,16,39]. (§ 3.2).

We have argued that dislocations are the only topological defects present
(at low energy) in the oblique tiling representing superspace. However, in
contrast with crystals, dislocations find it hard to move in quasicrystals,
because their mixed character (b=bll+b.1) forces them to climb as wen as
glide. We will show that two dislocations combine into the traditional
"phasons" in physical space. In 2d (one dimensional quasicrystal), the
phason flip is a pair of topological dislocations. (Fig.H). The phason flip is
the jump of one atom. It is also an elementary, local transformation in the
oblique tiling.

3.1.1 Tilings in 2d and their topological defects

THings consist of vertices, edges and tiles covering R2. Topological stability
imposes vertex coordination z=3. Thus, the tiling is naturally edge-on-edge,

1 The difficulty is that it involves more than three dimensions. We will see that parallel
Ell and perpendicular E.1subspaces intersect at points only, which greatly simplifies
matters.
2 There are two notations for tilings. Here (4.8.8) indicates that one quadrilateral tile and
two octagons meet at any vertex. The other (Schlaefli's) notation (n,z) indicates that all
tiles are n-gons, with z meeting at any vertex. An hexagonal tiling of the euclidean plane
is denoted by (6.6.6) or (6,3).
3 Topologically unchanged by small deformations.
266 N. Rivier

without adjustment. A random tiling is called a froth. A consequence of z=3

is that the average number of edges per cell is <n>=6.1

There are only two, local, elementary topological transformation in 2D:

neighbour exchange (T1) (Fig.9), and disappearance of a triangular cell (T2),
or its inverse, cell division or mitosis. Elementary transformations shuffle
the local variable n, and are source of defects and of entropy. If the tiling
(6.6.6) = {6.3}2 is defect-free, a dipole 517 is a dislocation (Fig.5). It is a l2ml
topological defect, which is defined even in a random froth where there is
no regular structure to dis-locate. It can be formally introduced in any tiling
by successive cellular divisions (Fig.5). A pair of dislocations can be created
physically by a local shear (5\717\5) or by a single mitosis (7\515\7) (Fig.10).3
Note that they have inverse polarity [41]. Dislocations glide in response
to further shear, and climb away from each other through further divisions,
leaving behind an additional row of cells, a painless mode of growth for
biological tissues (none of the cutting and reglueing of the Volterra
process). In random froth, shear is usually applied in a given direction,

,, ,
X,
,, ,

Fig.9 - The elementary topological transformation T1 (neighbour exchange)

is a pair of dislocations. (A dislocation in a froth is a pair
pentagon/heptagon. See Fig.5).

1 This follows immediately from Euler's relation between numbers of tiles, vertices and
edges, IT I-I E I+ I V 1=0(1) (2 for a sphere, 1 for a plane (a sphere from which the tile at
infinity has been removed), and valences relations 21 E I =Z I V I =<n> I T I, in the
"thermodynamic" limit I E 1=00.
2 See note 2, preceding page.
3 s\7/N can also be produced formally by two semi-infmite arrays of mitoses, each
constituting one dislocation. The directions of the dividing ·cells and the small jog
between twe two arrays indicates clearly that the Burgers vector is not exactly
perpendicular to the dipole 5{1.
 DEFECTS IN QUASICRYSTALS ... 267

Fig.10 - Pair of dislocations produced by local shear (T1, left) or by a cell

division (right). Note the opposite polarity in the two cases. (. = pentagon,
x = heptagon).

whereas the easy glide line fluctuates. This is why random froths do not
deform plastically, except in localised regions (shear bands at large stresses)
[42]; two dislocations diffuse away from each other as a response to applied
shear.

Thus, topological defects in tilings and froths are created and moved by
elementary, local topological transformations [21,41] There are only two
elementary topological transformations in 2D, and only one of them, the
neighbour switch Tl, occurs in systems with conserved number of tiles.

3.1.2 The oblique tiling (4.8.8)

Oblique tilings have "parallelipipedic" tiles. Edges are parallel to either E.l
(constituting the atomic surfaces) or to Ell, where some of them are the
actual edges of the quasiperiodic tiling (in d>2). Because atomic surfaces
correspond to real atoms, all perpendicular edges exist physically: They
constitute (in 2D) or bound the atomic surfaces. Some parallel edges are
phason flips, as we shall see. (Fig.ll). They link two atomic surfaces,
ensuring that matter does not disappear. The tiling has translation
symmetry in superspace. This fact determines the nature of topological
defects, which can then be decomposed along the parallel and perpendicular
directions of the tiling. The inflation symmetry of the quasicrystal is due to
268 N. Rivier

the irrational orientation of the tiles with respect to the superlattice Ad. It
is not directly relevant to defects classification.

Because En and E.1 intersect on points only (they are indeed the atoms
in quasicrystals), the edges in the tiling belong either to parallel or to
perpendicular space, not to both. This is obvious in 2D, but somewhat
surprising in higher dimensions.

Consider now the the 2D oblique tiling (4.8.8) specifically. It is a froth,

not adjusted. (The only adjusted, z=4 tilings have parallel space horizontal,
vertical or diagonal with respect to A2. Any oblique tiling (even rational
approximants) will have z=3). The Burgers vector of any dislocation in A2
has components along both parallel and perpendicular directions. Motion of
a single dislocation is very difficult: Successive tile division is not physical;
parallel division is not observable, perpendicular division introduces new
atoms. Under applied shear, the dislocation will have to climb along E.1 in
order to move in real space En (a physical consequence of the oblique slope
of the tiling). This is why quasi crystals are very hard below the brittle-
ductile transition (T<Tbd, with Tbd high compared to the crystalline alloy of
similar composition) [8].

However, a pair of dislocations 4/8\8/4-+7\515\7 can be created locally in

response to a shear 011.1 of the atomic surfaces along parallel space (or of the
strip of the cut-and-projection method). This neighbour exchange in the
froth is a phason flip (Fig.l1). Phason flip is therefore the only local defect
possible at low energy in quasicrystals. Note the orientation of the
dislocations constituting the pair, which is caused by shear (Tl) rather than
mitosis (T2-1). Note also that the phason flip is a motion in both parallel and
perpendicular space. Only those points at the edge of the atomic surface in
E.1 can jump.

A stronger applied shear forces the two dislocations of the phason flip
apart in En. However, they cannot glide away, since a climb along E.1 is
required for any move. The flip (not the atoms) diffuses away, like a blob of
ink in water. Several diffusing phasons will ultimately percolate, a
mechanism suggested by Kalugin and Katz [9] to explain the catastrophic
strain softening of quasicrystal above Tbd. Indeed, melting occurs almost at
once above Tbd: Compare TbdlTmelting in quasicrystals (Tbdtrmelting=0.9),
and in crystals (O.l<Tbdtrmelting< 0.6) [8].

Thus, the oblique tiling has phason flips as inevitable defects, thereby
explaining the hardness of quasicrystals at low temperatures, their sharp
 DEFECTS IN QUASICRYSTALS ... 269

brittle-ductile transition and their very low melting transition for such hard
materials. Indeed, phason flips are the agents of melting in quasi-crystals,
instead of oscillating atoms (in the old Lindemann model) or gliding
dislocations (in more modern theories) in crystals. In quasicrystals, the
agents are unambiguous and their effect is catastrophic.

Fig.ll - The phason flip is a T1 transformation in the oblique tiling, i.e. a

pair of topological dislocations. Left: The oblique tiling [17,18]. Thick lines
represent atomic surfaces, dashed lines, potential phason flips.

3.1.3.Defects in quasicrystals; summary

At the end of the day, the oblique tiling exhibits, almost at a glance, all the
main features of defects (dislocations and phason jumps) in quasicrystals.

1) En and E.1 intersect at points, the atoms, regardless of the dimension.

2) Atomic surfaces are connected by phason jumps and constitute an infinite
periodic hyperbolic surface (lPHS). The connection is made by linking
opposite end points (1d.1), edges (2d.1) or facets (3d.1) of the atomic surface
which serves as fundamental region of the IPHS.
3) Dislocations in Ad have mixed (II and .1) character. To move they would
have to climb in E.1 while gliding in En. They are defects of dimension dll-2.
4) A phason jump is a dipole of two neighbouring dislocations. Their mixed
character stabilizes the phason jump. Dislocations cannot simply glide away
from each other, as they do in foams. Physically, this is responsible for the
high ratio.Tbdtrmelting in quasi crystals and for their catastrophic softening
270 N. Rivier

at high stresses.
5) Because in IPHS two translations across the fundamental domain do not
commute, a closed loop in Ell may not close in E.l.. It has been pointed out
by lO~man [14] that this lack of closure is independent of the particular
geometry of the atomic surface (the fundamental region of the IPHS). It is
therefore an universal characteristics of (the irrational slope) of
quasicrystals and of their approximants.

3.2 ObeerYation of defects

Historically, defects have been identified well before having the microscope
power necessary to see them. Moreover, their observation has led to the
correct determination of the microscopic molecules constituting the matter
before actually seeing them. Examples are Newton rings, Franklin's oil film,
liquid crystals (Georges Friedel and F. Grandjean 1 ), Faraday rotation for
magnets, and, more recently, Kolosov's [44] bent, conformal lattices (which
have yet to be observed at an atomic scale). Quasicrysta1s are exceptions.
Their defects, like their physical properties (the celebrated quasi-frying pan
[45]) have been discovered much later than their diffraction pattern, high
resolution microscopy images, and even the position of all the atoms [46].

Polarized light is the classical tool to observe defects in liquid crystals.

The director d = 4»(x) rotates the direction of polarization of light (E). Thus,
polarized light will not traverse two parallel polarizers with liquid crystal
inbetween, except through those points x where d(x) is parallel to E. By
continuity, the loci of these points are lines closing as loops. In a Hopf
texture 7t3(Rp2)=Z=(lkJ, the two distinct closed loops corresponding to d
parallel and antiparallel to E are linked, and the linking number is precisely
the index (7t3(Rp2)=Z) of the texture. These "olympic rings" (Fig.6) have
been seen first by Bouligand [35,36]. Note that the points at infinity where
the director has the same orientation have been identified, so that liquid
crystals illustrate Hopfs map S3-+S2. The two distinct loops correspond to
opposite directions of the director~ The dimension of a texture defect is -1
(dr=3=d). It is an entangled, non-singular, yet topologically stable

1 "n est remarquable que, avant toute ~tude structurale aux rayons X. ce soit l'~tude des
d6fauts de structure qui ait amen~ Georges Friedel, avec Grandjean, A proposer la
structure moMculaire de ces stases. Le nom m@me de n~matique vient des fils mobiles
visibles dans les liquides; ce sont des d6fauts lin~aires ...qu'on appellera plus tard des
dislocations de rotation". [43, p.11l]. These defects adom the T-shirt of the programme on
Topological Defects at the Isaac Newton Institute, Cambridge 1994.
 DEFECTS IN QUASICRYSTALS ... 271

configuration. The linking number of Fig.6 measures the entanglement and

immediately demonstrates its topological stability.

3.2.1.Phase contrast and diffraction contrast microscopy

The latter technique has been used to identify dislocations in quasi crystals
[15,39]. We recall that a perfect crystalline lattice {rj} diffracts X rays into
sharp Bragg peaks, which constitute its Fourier image. The Fourier
transform of a lattice (a pattern both discrete and periodic) is a lattice, the
reciprocal lattice {G}, since the Fourier transform of a periodic pattern is
discrete, and Fourier transform is an involution.

Concentrate on one Bragg reflection G.ri=27tn, and on the transmitted

beam. Tilt the crystal very slightly. The reflection by the perfect crystal is
lost, and all the intensity is in the transmitted beam.

Now, suppose that there is a dislocation. It strains (slightly) the crystal,

with a distortion proportional to the Burgers vector b. The position of an
atom is then ui=ri+ab. We look at the same Bragg reflection G. Ui= 2xn,
now again visible in the slightly deformed im.d slightly tilted crystal. Hence,
G. b=O [47]. The strain due to the dislocation is seen as a dark contrast in
the transmitted beam.

Consider now the case of the quasicrystal. We have

G. b = 0 = Gn. bn + G.l. b.l (4)

which may imply that,

("strong" extinction),

thus G.l. b.l = 0, and the Bragg reflection shows full-fledged dislocations in
both parallel and perpendicular spaces. Alternatively,

("weak" extinction).

To distinguish between the two alternatives, look at the next Bragg

reflection in the family (by dark field imaging: Fourier transform again one
single Bragg reflection in reciprocal space. This illuminates the region in
the real material which made up the Bragg peak). It is at 'tGIi (inflation
symmetry. The corresponding wave vector in perpendicular space is (lI't)G.l).
272 N. llivier

If the extinction is strong, the dark contrast in real space remains.

If the extinction is weak, the contrast disappears, since now 'tGU. bU +

(lh)G1.. b1. ~ O. A good illustration of weak and strong extinction conditions
can be seen in [16]. Experimentally, the Burgers vector of the dislocation is
larger in perpendicular than in parallel space.

4. DEFECTS IN AMORPHOUS MATERIALS

In §2, we dealt with crystals, with a well-defined, global symmetry
(translation and rotation). The only problem was to find the relevant
manifold M. Defects are given by the homotopy groups Kn(M), under the
assumption that matter lies in a space (Rd) which is continuous.

In §3, Quasicrystals still had a global symmetry (inflation and rotation).

But now real space E II is only an irrational part of superspace Ad.
Identification of defects can be done in the oblique tiling: Defects are local,
geometrical defects of the tiling. Yet, it is more physical to work on EU only
as base space, and stick atomic surfaces E 1. as fibres. A defect is then
labelled by a closed contour r in EU, lifted (r') in superspace where it does
not close, exactly like ordinary dislocations (eq.3). The lift is effected by a
connection (gauge field). In quasi crystals, this connection is by no means
obvious, since a smooth contour in En jumps around in E.L.

In this section, we consider dB.u and amorphous materials, which have

no global, generative symmetry, only local gauge invariance, expressing the
fact that real changes from one point in space to the next are physically
irrelevant and average out over a short range. The best example is a
(random) forest in which one is lost; it has objective physical translation
invariance (one is lost) in spite of the fact that every tree and its immediate
environment is different from another. One is forced to work on a fibre
bundle.

There are two approaches :

1) (§4.2) The local, gauge symmetry, enables us to construct a fibre
bundle, with a natural connection (the covalent bond). The defects are
tunneling modes, elementary excitations of glass at low temperature.
Tunneling takes place between two, equivalent, classical configurations,
localised around a geometrical loop, the odd line, which is the core of the
defect.
 DEFECTS IN QUASICRYSTALS ... 273

2) (§4.3) One constructs the glass by sequence(s) of decurving operations.

Decurving operations provide the connection. These sequences are
algebraically similar to substitutional word sequences (discussed elsewhere
in this volume), and, as for word sequences in quasicrystals, they have a
geometric representation, but in 3D, and random, at two levels: The
sequence itself is random, and different random sequences may be used
randomly in different regions of space. The boundaries between regions are
easily connected by a few Tl transformations. (See Fig.14 below). Defects
are localised pelotes of negative and positive disclinations created by these
Tl.

4.1 Symmetry of disorder

Structurally, a glass (for example, the plain window glass Si02) is a
continuous random network, a (random) regular graph, whose vertices are
the tetravalent Si atoms (regularity z=4 is imposed by chemistry) and whose
edges are covalent electronic bonds, decorated by one oxygen which plays no
topological role. A vertex is therefore a four-legged reference frame
(Vierbein), with random orientation towards its neighbours. The symmetry
is a physical or objective invariance (homogeneity) despite different local,
subjective environments. This is gauge invariance, for which fibre bundle is
the natural geometric framework [37].

The same holds for the 2d random tilings and froths seen in §2 (with z=3
now imposed by randomness). A random structure remains macroscopically
and physically invariant in spite of, and because of defects (local, elementary
topological transformation like the neighbour exchange Tl discussed above),
which are the 'collisions' enabling the topological 'gas' to explore its phase
space (here, the cell shape n) and maximise its entropy [20]. Thus, the (local,
gauge) symmetry of random structures is preserved by defects. By contrast,
the global symmetry of crystals is upset by defects. Defects in amorphous
materials contribute to their symmetry and preserve it.

4.2 Glass as a fibre bundle

Gauge invariance is the symmetry of disorder. Glass is a disordered solid,

and can be represented as a fibre bundle, because a gauge field,
geometrically, is a connection in a fiber bundle. A fibre bundle consists of a
base space (the regular graph of Si atoms and bonds), a total space (position
and relative orientation of the Vierbein), and a map that projects every
point of the total space onto a point in the base space. The set of all points
274 N. llivier

in the total space that are mapped onto the same point in the base is called
the ~ (set of permutations of the legs of a Vierbein - see below). There is
one, unique representative fibre for the whole bundle.

The connection. or P,U" field describes how the orientation of the fiber
changes as one goes along a path in base space, or how the path in base
space is li1lwl into a path of the fibre bundle. In glass, this connection is
natural and physical one: It is carried by the bond separating two Vierbein.
This bond can be bent, but not stretched at low energy.

A closed contour r in base space (a ring of bonds) is lifted into a contour

r' in total space. If r' is closed, no defect is enclosed, and the bundle is
trivial within r. Ifr' does not close, the missing element is a displacement
on the fibre. We return to the same point in base space, in a solid with a
given, frozen configuration, but its labels have been entangled (like our
tendons when we rotate our hand by 21t). r encloses (in base space) a defect
labelled by the displacement on the fibre, also called large gauge
transformation.

We now show that even rings in the continuous random network are
trivial, while odd rings surround defects, which are loops (dimension 1). The
existence of these lines, and the dynamics of the defects (tunneling modes)
can be and has been seen experimentally [49,50].

First, odd rings are threaded through by closed loops (odd lines) which do
not thread through even rings [51]. Lines traversing odd rings are
continuous. Odd lines play therefore, in random structures where there is
no meter to measure the Burgers vector, the structural part of dislocations.
One proves this statemment like any conservation law: any line coming into
an arbitrary region must get out.

[Consider an arbitrary, closed surface S inside the material, intersecting

the network at its vertices (this can always be arranged by small
deformation of S). S is tiled by some vertices, edges and rings of the
network. On any S, there is an even number of odd rings, providing an exit
for any odd line entering S: Ascribe a weight Je=-l to every edge e on S,
and an index If = nee f J e = ±l (even/odd ring) to every ring f on S. Then
nfe s Gf = nee s J e 2 = 1 (Je enters twice in the product because every edge
separates two faces on S).].

Maximum entropy [50] indicates that the odd lines are semi-dilute in the
glass: Odd loops are small but intermixed, taking maximum advantage of
 DEFECTS IN QUASICRYSTALS ... 275

mixing and configuration entropies. There is only one characteristic length

scale, the average size of a loop equals the average distance between
different loops.

We want now to lift this index in total space r':

- An edge e, linking vertex i to j, is lifted as a reflection-rotation of the
Vierbein. The reflection is through a mirror perpendicular to the linking
bond. Rotations are due to bending the bond and twisting around it. (If this
twisting is by any· multiple of 27t13, it is an even permutation of the 3
unshared legs. If, however, one deals with a z=3 glass (amorphous As),
twisting by 27t12 is an odd permutation of the 2 unshared legs; all the
arguments below do not apply, and there are no tunneling modes in a-As
[49,50]).

- A ring of n edges is lifted as a covering transformation of the Vierbein - a

permutation of its 4 legs. This permutation is also the product of n
successive reflection-rotation. If n is even, the permutation is even, and we
can label edges consistently (choose an algorithm such that the permutation
is the identity). This is the lowest energy state of an even ring. If n is odd,
the permutation is odd and we cannot label edges consistently. The edges
must be permuted (a large gauge transformation, or displacement on the
fibre) after circumnavigation about an odd ring. The odd ring surrounds the
core of a line defect, the odd line. (It is clear that the same permutation
characterises all odd rings belonging to the same odd line, since they are
connected by even rings which are labelled consistently).

It is only the ~ of permutation which characterises the phuical

configuration of the defect: If we measure two rings Q and R, one after the
other, the physical configuration is independent of the order, thus
P=QR=RQ=P, thus P=RPR-l, VR.

There are ~ classes of odd permutations of 4 legs I(a~» and l(a~'Y~»,

with 6 permutations each. Permutations, like labelling of the legs of the
Vierbein, are not relevant physically, and there is a relabelling G which
transform one class into the other, GI(a~»= l(a~'Ya». G is a large gauge
transformation. The physical configuration of the glass, and its elementary
excitations must be gauge-invariant: They are the tunneling modes

I±> = (1{1/2) [1I(a~»±I(a~'Ya»] (5)

observed experimentally [52,53].

276 N. Rivier

Note that if homotopy and fibre bundle representations of topological

defects are fairly similar (they both rely on a mapping, resp. lifting, of a
closed contour r in R3, resp. base space), they have specific differences. The
base space can be discrete, whereas homotopy theory requires a continuum.
The lifted contour does not close, and it is this lack of closure on the fibre
which identifies the defect. (In homotopy theory, r' always closes in M, and
its homotopy class identifies the defect). Defect types are obtained from non-
commutativity of physical connections on r. (By contrast, in homotopy
theory, the 7ti(M) come from an exact sequence of various maps between
different homotopy groups (eq.2) which is much more general but less
intuitive physically or even geometrically).

4.3 Curved space, packing atoms and frustration.

We turn now to a different geometric representation of amorphous

materials. Instead of networks of covalent bonds, consider random packings
(of atoms or grains).l Their structure will be givan by random sequences of
decurving operations (eq. 10).
Packing (spherical) atoms in 3D is the realm of metals, intermetallic
compounds, icosahedral and dodecahedral quasicrystals and metallic glasses.
It is done best (but not perfectly, as we shall see) in the Frank-Kasper, or
tetrahedrally-dose-packed phases (t.c.p.) [54].

The obvious packing unit is a regular tetrahedron, which has a dihedral

angle of 27t/5.1 = cos- 1(1/3). This means that 5.1 regular tetrahedra fit about
one edge (like the segments of an orange). This is doubly frustrating: The .1
wastes space between atoms, and 27t over the nearest 5 is not a
crystallographic rotation. 2 Kleman and Sadoc [55] pointed out that
frustration is relieved simply by embedding the structure in positively
curved space. Positive curvature takes care of the .1 (the lips of an orange
with 5 regular tetrahedral segments, ajar in Euclidean space, are just closed
by curvature), and the five-fold symmetry is crystallographic in the finite
curved space. The structure is a scaffolding for the hypersphere S3, polytope
{3,3,5},3 consisting of 120 spheres (vertices), 600 regular tetrahedra, 5 of

1 The two representations are dual. Precisely, a froth is the topological dual of a packing
[20].
2 In 2D, there is no problem (euclidean tiling {3,6} by equilateral triangles).
3 Schlaelli's notation {a,b,c} denotes a polytope made of c polyhedra {a,b} sharing each
edge. {a,b} is a polyhedron made of a-gons, b sharing a vertex. Thus {3,3}=tetrahedron,
{3,5}=icosahedron, {5,3}=dodecahedron, {3,4}=octahedron and {4,3}=cube.
 DEFECTS IN QUASICRYSTALS ... 277

which are incident on each of the 720 edges, and 12 on each vertex, making
up an icosahedral coordination shell. Polytope {3.3,5} is therefore the 'ideal'
close-packed structure.

But it only exists in curved space, and materializes to us mortals (to

pursue this Platonic discourse) as various structures (some ordered, some
random) in Euclidean space. To obtain these structures, one must either
project (as in the .original myth of the Cave) [56], or, better, decurve
iteratively, without [57] or with disorder [58].

Note that in all the t.c.p., quasicrystals and metallic glasses, the local
environment is still very much icosahedral. Iterative decurving will preserve
this local environment as much as possible. It is done in sequences which
can be either random but spatially uniform, or also random in space [58]. A
product of de curving operations is a decurving operation.

We will see that the whole structure is described by a sequence of

decurving operations, or a tree of sequences if the structure is spatially
random. Two sequences at different points differ by the order of two
de curving operations. They bifurcate at some level in the tree. A decurving
operation is similar to the substitutions studied elsewhere in this book. In
glasses, however, the sequence(s) is (are) random. To offset this apparent
complication, we have a visible record of the whole sequence of decurving
operations.

Most of the physical properties (energy, density) of the materials, like

the Perron-Frobenius eigenvalue of the full decurving matrix, are
independent of the order of decurving operations of the sequence. They are
homogeneous in the glass. Defects, inhomogeneities are commutators
between decurving operations which are more localized if they occur later in
the sequence. Conversely, an early commutator will give rise to an extended
defect. These defects are observed in glass as tunneling modes [52,53].

4.4 Iterative decurving

Let us illustrate this sequence of operations by a two-dimensional example.

Consider a dome for an athletic stadium, supported by a network of steel
rods. The steel rods make up a polyhedral scaffolding, naturally part of a
regular polyhedron. Mechanical rigidity imposes triangles as polygonal
units. The largest regular polyhedron made of triangles is the icosahedron
{3,5}. This may be large enough for a tent, not for a football stadium; the
bars would be much too long to remain rigid. So decorate each triangle as in
278 N. Rivier

Fig.12, keeping the rod length constant, and rescale the structure. The
radius of curvature of the dome is increased, the curvature of the structure
decreased. Because triangles are subdivided into triangles (keeping the
structurally rigid triangular motive at every stage of the decoration), the
procedure can be repeated iteratively, and the curvature eventually
decreased to zero.

This example (Fig.12) is Buckminster Fuller's geodesic dome, a scaffolding

of arbitrary size and curvature. Only the original vertices are 5-valent. All
additional vertices, at all stages of iteration, are 6-valent, and are therefore
cores of neiatiye disclinations, decuryini the original icosahedron {3,5}. One
could imagine a different decurving algorithm, for example dividing the
edges into 3 parts instead of 2, and transforming a triangle into 3 triangles
and one hexagon (a kagome network). In 2d, the hexagon is then centered,
divided into 6 triangles and the procedure can be repeated. The final
structure is still a geodesic dome, but with a different amplification or
decurving ratio at each stage (a=3 instead of 2). This is a particular feature
of two-dimensional space. In three dimensions, the two decurving modes do
not produce the same structure, and this gives us a means of producing a
(hierarchically) disordered structure (see eq. (9) below).

The vector n(i) = (II of 5-valent vertices, II of 6- valent vertices) summa-

rizes the state of the structure at stage (i) of decurving. The initial
structure vector is n(O) = (12,0). Successive stages are related by the
transfer matrix a,with Ou =1,012 =0, 021 =512,022 =4, as

Fig.12. Geodesic dome mode of decurving (a=2) in 2d. Circled vertices are
negative disclinations. Right : a=3 and 5.
 DEFECTS IN QUASICRYSTALS ... 279

j, k=(5,6). (6)

The Perron-Frobenius (largest) eigenvalue of n is 9=4, and the total number

of vertices N(i) = n(i)5+n(i)6 increases asymptotically as 12x4i. This is exactly
as in the substitutions discussed elsewhere in this volume. The vector n(i)
has for componants I w II, the numbers of occurrences of the letter I (5- or 6-
valent vertex) in the word w (the triangular scaffolding representing the
structure at decurving stage (i». The transfer matrix n is the matrix of
substitution for the particular mode of decurving. The new feature is that
the mode of decurving, n(i)(x), can change at every stage (i), or for different
triangles (tetrahedra in 3d) (x) in space.

Because decurving deals with triangles, tetrahedra or simplices at every

stage, it can be iterated, and preserves closest possible local packing.
Accordingly, the resulting structures are called t.c.p. (tetrahedrally-close-
packed) or Frank-Kasper phases [54]. The only effect of frustration is carried
by disclinations (points (vertices) in 2d, closed lines in 3d), which are defects
with a tension, trying to minimise their (low) energy. The network (in 3d) of
disclinations is a super-scaffolding containing the structural information. It
only connects vertices which are not 12-valent (bold vertices of Fig.13).
Vertices of coordination 16, 15 and 14 are nodes of 4, 3 and 2 disclinations,
respectively. These are the only possible, topologically stable vertices in the
absence of disorder. 1

The construction is analogous to Regge's method in general relativity

[59]: d-dimensional space-time is decomposed geometrically into flat simplices.
Curvature (caused by gravitational mass) is carried by (d-2)-simplices (points
in 2d, lines in 3d, triangles in 4d), where d is the dimension of space-time. It
is a defect of codimension 1 (eq.(I». (Regge's empty space is flat, covered by
simplices which are perforce not equilateral, except in 2d. By contrast, the
ideal space of a packing is curved, and covered by regular tetrahedra).

4.4.1.Three-dimensional materials
The material is represented by polytope {3,3,5} decurved by decoration, or,
more simply, by its disclination network or by its structure vector n. Even
though the starting polytope is always the same, the decoration algorithm is
not unique. It can be varied from stage to stage of the decurving process
(giving rise to a hierarchically scrambled, but spatially ordered structure) or
from region to region in space (giving rise to a spatially disordered
structure) [58].

1 Disorder brings in positive disclinations and exotic (non t.c.p.) coordinations [58].
280 N. Rivier

Clearly, if one uses different modes of decurving for different triangles,

orland at various stages, the matrix n(i+l) = n(x), and n(i)(x) depend on the
spatial coordinate x of a triangle (or simplex) at the ith generation, as well
as on (i). The structure is described locally by the decurving action on the
(i)-triangle at x, n(i+l)(x) and by its coordination n(i)(x), more uniformly at
lower (i) and with finer details obtained at higher iteration levels.

In 3d, one starts with polytope (3,3,5), and decurve systematically by

decorating the tetrahedra. Decurving now generates disclination lines (the
common edges of 6 tetrahedra instead of 5). The tension of the disclination
lines carries most of the decurving energy. These lines close as loops or meet
at vertices, so that there are, besides the icosahedral z=12 vertex of the
curved space, only 3 possible vertices, z=14, 15 and 16, meeting points of 2, 3
or 4 disclination segments, respectively. Others are neither topologically,
nor mechanically stable: They would split into canonical vertices by an
infinitesimal deformation. Thus the structure at decurving stage (i) is
summarized by a 4 component vector n(i) = (#12,#14,#15,#16), with n(O) =
(120,0,0,0) as initial condition. The transfer matrix depends on the mode(s)
of decurving. For example:

4.4.1.1. The Laves phase : We start with polytope (3,3,5), and decurve
systematically by dividing every edge of the tetrahedra into a=3 equal parts
in a kagome pattern as in Fig.13. Each tetrahedron is therefore divided into
4 tetrahedra and one larger polyhedron called Friauf-Laves (FL) polyhedron.
Add a new vertex at the centre of the FL polyhedron. The centres of two
adjacent FL's, connected by an hexagonal face, have 6 tetrahedra incident
on the edge joining the two centres (Fig.13). This edge is therefore a
negative disclination (with 6 instead of 5 incident tetrahedra), and the
center has coordination 16 (it is a node of 4 disclinations). Each old
tetrahedron generates 4+4x6/2=16 tetrahedra at the next stage [it retains
its 4 vertices with the same coordination, and adds 6x2 new vertices on its
edges, of coordination 12 if the old edge is not a disclination, 14 if it is, and
one vertex of coordination 16 at the center (of the old tetrahedron and of
the new Friauf-Laves polyhedron)). It is easy to write the transfer matrix n,
relating 12, 14 and 16-coordinated vertices at consecutive stages of
decurving :

13 12 12
Q= (0 3 4 ) (7)
5 6 8
 DEFECTS IN QUASICRYSTALS ... 281

Fig.13 - Decoration of a Friauf-Laves polyhedron (truncated tetrahedron)

occurring in the a=3 de curving mode (left). Negative disclinations join bold
vertices of coordination 16.

with Perron-Frobenius eigenvalue 6=20:;a 3 . N(i) increases very rapidly

(N)10 7 for i=5); the structure is effectively flat after a few iterations. Note
that there are only 12, 14 and 16-coordinated vertices at every stage (15
would also be allowed in t.c.p.), so that every structure at each stage of the
decurving is a t.c.p. phase [54].

The final, flat structure is called Laves phase [58,54]. It could also have
been obtained by one single decurving operation with a=oo (odd).

4.4.1.2 Hierarchic glass: The same type of decurving can be done for any odd
a. a=1 labels the identity decurving, a=oo yields the structure of the cubic
Laves phase. One can choose a sequence of odd a randomly, either
uniformly, or randomly for different tetrahedra in space, to obtain random
structures.

The sequence {a (i)} =a 1.a2 .... } of decurving operations is chosen at

random. The decurving matrix nin(i) is a product of 3x3 matrices (7), since
only 12-, 14- and 16-coordinated vertices are produced at each stage. Each
decurving matrix depends on a, but is algebraically always the same (it is
the sum of 4 basic matrices, with a-dependent coefficients, not writen down
explicitly here [58]). Thus, nin(i) has also the same algebraic form

(8)

where 6i=(17/24)ai(ai2-1)+aiis the Perron-Frobenius (largest) eigenvalue of

n, proportional to the volume amplification ratio a 3 , the ratio between
282 N. Rivier

numbers of atoms in the structure after and before the de curving step.
M (ai}) (given in [58]) is the only coefficient depending on the order of the
decurving sequence {ai}. ~=O for a single operation. The basic matrices
(together with the null matrix 0) form a closed algebra under multiplication,
whose only non-commutative component is proportional to
C=[X,Z]=[C,Z]=[X,C]. I is the identity matrix, and X2=X, Z2=Z,
C 2 =o=ZX=ZC=cx.

The structure is given by the Perron eigenvector ~, IIiW(i) ~=9 ~,

corresponding to the Perron-Frobenius eigenvalue 9=(IIi9(i), which is
independent of the order of the a's. Accordingly, the distribution of 9 is
lognormal [58]. Both strain energy (given by the length of the disclination
network) and the total number of atoms depend on 9 only, and are
independent of the order of the de curving operations. After 5 decurving
operations, the structure has over 107 atoms and is practically flat.

Any sequence of odd a's corresponds to a sequence of decurving

operations, and yields a (homogeneous but usually not invariant by
translation) t.c.p structure, characterized by the Perron eigenvector ~ of
nin(i) (which, unlike its eigenvalue, depends on ~, i.e. weakly on the order
of decurving).

All these hierarchic structures ~ are homogeneous but not

translationally invariant. The different structures are distinguished chiefly
by the last decurving operation. Differences are local, they quickly average
out, as in all random structures.

4.4.1.3. Spatial disorder: In 3D, different decurving operations n do not

commute in general and tetrahedra decurved by n(a)n(~) are structurally
different from those decurved by n(~)n(a), even though they have the same
density (same number of vertices), matching edges, and nearly matching
faces. (Fig.14).

Suppose that a tetrahedron at x is decurved according to a=3 first and

a=5 next, and the surrounding tetrahedra according to a=5 first and a=3
next. The space curvature, density of atoms, energy density and Perron-
Frobenius eigenvalue are the same everywhere (they depend on the product
of the a's only) and the glass retains its over-all uniformity. But, locally, the
structure is not uniform, because the decurving operations do not commute;

(9)
 DEFECTS IN QUASICRYSTALS ... 283

Fig.14 - Interface between two tetrahedra, one decurved by 0(3) 0(5), the
other by 0(5) 0(3). Only 12 vertices (bold) would have occupied slightly
different positions in the two modes of decurving. They have been moved to
intermediate positions, but connected according to the respective mode of
decurving. The two configurations differ by a few flipped edges (T1). Inside
the tetrahedra, the structures (j.I.) are different, but they have the same
number of atoms, energy and the same Perron-Frobenius eigenvalue 9 of the
decurving matrix.

The triangular interfaces between tetrahedra can be patched up with small

displacements of their atoms (the bold vertices in Fig.14), which are linked
according to decurving (35) or (53). The two possible configurations are
related by flipping edges, i.e. T1 transformation which we meet again as
essential defect. Vertices on the edges of the tetrahedra are independents
of the order of decurving operations. The patch-ups are local defects
("pelotes" (wool-balls) of positive and negative disclinations)l, not necessary
for over-all decurving, but constituting the local geometrical signature of
disorder [58,60].

1 There are a few flipped edges (Tl) on the interfaces between regions decurved in
different order (Fig.14). The atoms involved in the Tl have exotic coordinations, and are
on a pelote of disclinations surrounding the region decurved in commuted order. This
pelote is a small ball of positive and negative disclination loops which is mechanically
and topologically stable. Indeed, it has elements in tension (positive disclinations) and
in compression (negative disclinations, locally overdoing the decurving), like the
spokes and the rim of a bicycle wheel, and is ideal to absorb external stresses, as random
materials do very efficiently. These pelotes, like all defects, cost some energy
proportional to the total length of their disclinations loops.
284 N. Rivier

One obtains a spatially disordered structure, as tetrahedrally close-

packed as possible (by construction, through systematic decurving), given by
an infinite sequence of local decurving operations,

(10)

The structure is given by the Perron eigenvector ~(lI:). Again, energy

density, density of atoms and Perron-Frobenius eigenvalue 9 are
independent of the order of decurving operations, and therefore
homogeneous throughout the material. This enables us to define a tree of
decurving sequences all with the same 9. Two structures at the same level
(same stage of decurving) are related by commutation of two (or more)
decurving operations earlier in the sequence. The tree is ultrametric (the
distance between two structures at the same decurving stage is given by the
number of decurving operations to their last common ancestor). The glass
represented by the ultrametric tree has uniform density and energy
density. 1

There are two different atomic rearrangements at the interface,

degenerate in energy, separated by an energy barrier, and related through
flipping (T1) a few bonds (Fig.14). These constitute the tunneling modes [58]
discussed earlier (§4.2). If the commutation of operations occurs at an earlier
stage the tunneling modes are more extended, involve more atoms with a
higher barrier and a smaller tunneling rate than if it occurs at a later stage
of decurving.

Sequences differing through early decurving operations show long-

ranged structural inhomogeneities, whereas those differing through later
stage operations correspond to short-ranged inhomogeneities. Early defects
are amplified in size through later decurving operations, but their elastic

1 One can speculate on the first branchings in the tree. We would need as many as
possible of the 600 initial tetrahedra to differ by one transposition from the decurving
sequences of its neighbours. Thus. the dual polytope (5.3.3) must have every edge
(separating two neighbouring tetrahedra in (3.3.5)) labelled by a transposition (a.b.... )
or by the identity permutation 1. Every pentagonal circuit returns to the same decurving
sequence. so that the product of its edges is the identity. Since a transposition is an odd
permutation. and a2 =1. [a.b]~. faces have 4 (in the order aabb or aaaa, with 1 anywhere
inbetween). 2 or zero transpositions. The 30 edges of each of the 120 dodecahedra of
(5.3.3) carry at most 6 different transpositions. and at least 6 are the identity. We expect
at most 6.120/3=240 branches in the tree initially.
 DEFECTS IN QUASICRYSTALS ... 285

energy is better screened. The interest of the de curving approach is that, to

a simple and natural geometrical picture (Figs.12-14), it associates an even
simpler local algebraic structure, which has gauge invariance and
corresponding topological defects [58].

4.4.2 The role of entropy

The decurving method gives an algorithm (one infinite, ultrametric tree) for
the structure of glass. If the algorithm was finite, the structure would be
deterministic, not random. The canopy of the tree contains all the
tetrahedra in the final, decurved structure. It occupies three-dimensional
space. Branching is binary: the two branches differ by one transposition in
the sequence of decurving parameters (a). The entropy could, in principle,
be calculated, since the structure is flat after 5 iterations, so that further
decurving operations serve only to randomise better the structure.

There is a large number of decurving sequences, all with approximatively

the same energy. Physically, polymorphism is found in atomic packings (t.c.p.
quasicrystals and polymorphs) and covalent materials ("chair" and "boat"
rings). This suggests that entropy is the main, and probably the sole agent
responsible for disorder. One would need to adjust carefully (use the same
decurving mode everywhere at all stages) in order to build a crystalline
structure. Any energy difference is small, local, and associated with one
particular disordering operation.

The vertex coordination vector J.L(x) is a local structural parameter (one

cannot call it an "order" parameter). It is the eigenvector of decurving
operators obeying a non-commutative, albeit very simple algebra. Almost all
structural information is in the disclinations, which act on a curved, finite
reference state (S3, or its scaffolding (3,3,5)), and not on a Euclidean lattice.
They appear in two distinct guises: As a network of negative disclinations,
whose only role is decurving, and in local pelotes of positive and negative
disclinations, which are signatures of disorder. In t.c.p. phases, the network
of disclinations is regular, albeit complicated. In glasses, the network is
random, and it may be difficult (given only the final structure) to
disentangle it from the pelotes which disorder it.

The systematic decurving procedure restricts disordering to a few possible

local operations at every stage. Thus, the specific geometry of the pelote
and the exotic coordination polyhedra are determined from the ordered
structure. In this respect, the pelote is a genuine structural defect. Its size
286 N. Rivier

depends on the stage at which the disordering commutation occurred.

Iterative decurving preserves closest packing, locally, at all stages; it

has thus lowest possible energy. When done at random, it has also high
entropy. There are several possible modes of decurving, all costing
approximatively the same energy (proportional to the length of the
disclination network), which can be selected locally. Regions with different
modes of decurving can be patched up. The patch-up involves local
elementary topological transformations (T1), which again appear as the
agents of disorder, increasing the entropy at no or little energy cost.

5. CONCLUSIONS

We began these lectures with a critical assessment of the general and

elegant classification of defects by homotopy theory. Quasicrystals and
glasses appear, by their discrete or disordered nature, to escape such
classification, even though they have observable defects (phasons and
tunneling modes). So, if quasicrystals are already beyond homotopy
classification, which defects lie beyond quasicrystals ...?
This led us into a detailed analysis of the defects in quasicrystals and in
glasses. They have several features in common with homotopy theory :

- Elementary defects are identified as topological entanglements over

contours in the real material.
- They can be represented as a non-trivial fibre bundle.

The active element is the compound defect: The T1, local, elementary
topological transformation creates and moves a pair of dislocations (a
dislocation loop in 3d). It is recurrent in these lectures, as T1 in froths, as
phason jump in quasicrystals (a T1 in superspace), or as pelote of positive
and negative dislocations in glasses. The compound defect is localised in a
small region in space. By its topological and mechanical stability, it can
screen external stresses. Thus, its own (small) energy remains also
concentrated in a small region in space.

Last, but not least, it is extremely satisfactory that not only

quasicrystals, but glasses are material examples of the mathematical ideas
(sequences of substitutions) presented elsewhere in this volume. Besides
paper-folding, of course, 'the only known material which freezes at ninety
degrees exactly' (M. Mendes-France, Les Houches 1989 [61]).
 DEFECTS IN QUASICRYSTALS ... 287

Acknowledgements
The early redaction of these notes was made at the Isaac Newton Institute
in Cambridge, during the programme on Topological Defects. Hospitality of
the Institute and discussions, particularly with M. Kleman, H.-R. Trebin
and M. Monastyrsky, are gratefully acknowledged. I am also grateful for
critical comments and ideas from F. Axel, T. Aste, B. Dubertret, from
participants to the workshop in Les Houches and to a lecture course in
Strasbourg.

References

[1] Toulouse G. and Kleman M., J.Physique Lett. 37 (1976) 149.

[2] Rogula D., "Large Deformation of Crystals, Homotopy and Defects",
Trends in Application of Pure Mathematics to Mechanics, G. Fichera,
ed., Pitman, NY 1976.
[3] Volovik G.E. and Mineev V.P., Zh.E.T.F.Pis'ma 23 (1976) 647.
[4] Mermin D., Rev.Mod.Phys. 51 (1979) 591,
[5] Michel L., Rev.Mod.Phys. 52 (1980) 617.
[6] Toulouse G., "A lecture on the topological theory of defects in ordered
media: How the old theory was leading to paradoxes, and how their
resolution comes within the larger framework of homotopy theory", in
Modern Trends in the Theory of Condensed Matter, A. Pekalski and J.
Przystawa, eds., Springer 1980, p.189.
[7] Golo V.L and Monastyrsky M.I., Phys. Letters A 66 (1978) 373.
Monastyrsky M.I., Topology of Gauge Fields and Condensed Matter,
Plenum, New York, 1993.
[8] Bresson L., "Mechanical properties of quasicrystals", in Lectures on
Quasicrystals, F. Hippert and D. Gratias, eds., Editions de Physique,
Les mis 1994, ch.13, p.549.
[9] Kalugin P.A. and Katz A., Europhys. Lett. 21 (1993) 921.
[10] Katz A. and Gratias D., "Tilings and quasicrystals", in Lectures on
Quasicrystals, F. Hippert and D. Gratias, eds., Editions de Physique,
Les mis 1994, ch.3, p.187.
[11] Kleman M. and Pavlovitch A., J. de Physique (Coll) 47 (1986) C3-229.
[12] De P. and Pelcovits R.A., Phys. Rev. B 35 (1987) 8609.
[13] Kleman M., "Imperfections in quasicrystals: Dislocations", in
Quasicrystalline Materials, C. Janot and J.M. Dubois, eds., World
Scientific 1988, p.318
[14] KIernan M., "Dislocations and disvections in aperiodic crystals", J.
Physique I France 2 (1992) 69.
288 N. Rivier

[15] Wollgarten M., Gratias D., Zhang Z. and Urban K., Phil. Mag. A64
(1991) 819.
[16] Wollgarten M. and Urban K., "Dislocation and plasticity of
quasicrystals", in Lectures on Quasicrystals, F. Hippert and D. Gratias,
eds., Editions de Physique, Les Ulis 1994, ch.12, p.535.
[17] Oguey C., Duneau M. and Katz A., Comm. Math. Phys. 118 (1988) 99.
[18] Kramer P., J. Math. Phys. 29 (1988) 516.
[19] Frenkel D.M., Henley C.L. and Siggia E.D., Phys. Rev. B 34 (1986)
3689.
[20] Rivier N., "Order and disorder in hard sphere packings" in Disorder
and Granular Media, D. Bideau and A Hansen, eds., Elsevier 1993,
ch.3, p.55.
[21] Weaire D. and Rivier N., "Soap, cells and statistics - Random patterns
in two dimensions". Contemp. Phys. 25 (1984) 59.
[22] Pleasants P.A.B., "Quasicrystallography: some interesting new
patterns". Banach Center Publ., 17 (1985) 493
[23] Lunnon W.F. and Pleasants P.AB., J. Math. pures and appl. 66 (1987)
217.
[24] Doubrovnine B.A., Novikov S.P. and Fomenko A.T., Geometrie
Contemporaine, Mir, Moscow 1982, §20.
[25] Toulouse G., Physics Reports 49 (1979) 267.
[26] Rivier N., "Glasses: a suitable case for homotopy?". Phil.Mag.A 45
(1982) 1081.
[27] Stewart I.N., Concepts in Modern Mathematics, Penguin 1975.
[28] Yu L.I. and Saupe A, J. Amer. Chern. Soc. 102 (1980) 4879.
[29] De'Neve T., Kleman M. and Navard P., J. Physique II 2 (1991) 187.
[30] Poenaru V. and Toulouse G., J. Physique 38 (1977) 889.
[31] Jiinich K and Trebin H.R., in Physics of Defects, R. Balian, M. Kleman
and J.P. Poirier, eds. (North Holland 1981), p,421.
[32] Kutka R., Trebin H.R. and Kienes M., J. Physique 50 (1989) 861.
[33] Rivier N., "The topological structure of grain boundaries", in Number
Theory and Physics, J.M. Luck, P. Moussa and M. Waldschmidt, eds.,
Springer Proc. in Phys. 47 (1990) 118.
[34] Rivier N., "A botanical quasicrystal". J. de Physique (ColI.) 47 (1986)
C3-299
[35] Bouligand Y., J. Microsc. Spectrosc. Electron. 3 (1978) 373.
[36] Bouligand Y., Derrida B., Poenaru V., Pomeau Y. and Toulouse G., J.
Physique 39 (1978) 863.
[37] Rivier N., "Gauge theory and geometry of condensed matter", in:
Geometry in Condensed Matter Physics, ed. J.F. Sadoc (World
Scientific, Singapore, 1990) p.1.
 DEFECTS IN QUASICRYSTALS ... 289

[38] Rothen F., Pieranski Piotr, Rivier N. and Joyet A., "Cristaux
conformes". European J. Phys. 14 (1993) 227.
[39] Wang Z.G., Xai M.X. and Wang R., "Observation and simulation of a
small dislocation loop in Al-Pd-Mn face-centered icosahedral
quasicrystal". Phil. Mag. Letters 69 (1994) 291.
[40] Oguey C., "Periodic surfaces in the description of quasicrystals". J. de
Physique (ColI) 51 (1990) C7-273.
[41] Rivier N., "Dislocations in two-dimensional systems, topology and
thermodynamics". in Dislocation 93, J. Rabier, A. George, Y. Brechet
and L. Kubin, eds., Solid State Phenomena 35-36 (1994) 107.
[42] Georges J.M., Meille G., Loubet J.L. and Tolen A.M., Nature 320
(1986) 342.
[43] Friedel J., Graine de Mandarin, Odile Jacob 1994.
[44] Kolosov V.Yu., "Nontranslational atom ordering in crystals growing in
amorphous films". In Aperiodic '94, G. Chapuis, ed., 1995.
[45] Dubois J.M., Kang S.S. and von Stebut J., J. Mater. Sci. Lett. 10
(1991) 537
[46] De Boissieu M., Guyot P. and Audier M., "Quasicrystals,:
quasicrystalline order, atomic structure and phase transitions", in
Lectures on Quasicrystals, F. Hippert and D. Gratias, eds., Editions de
Physique, Les Ulis 1994, ch.l, p.l.
[47] Hirsch P.B., Howie A. and Wheelan M.J., Phil. Trans. Roy. Soc. A252
(1960) 499.
[48] Sadoc J.F. and Charvolin J., "The crystallography of the hyperbolic
plane and infinite periodic minimal surfaces". J. de Physique (ColI) 51
(1990) C7-319.
[49] Rivier N., "Glass: elementary excitations and geometry; from bonds to
tunnelling modes". Phil. Mag. B69 (1994) 911.
[50] Rivier N., "Odd rings and tunneling modes in glasses". J. Non-cryst.
Solids 182 (1995) 162.
[51] Rivier N., "Disclination lines in glasses". Phil.Mag.A 40 (1979) 859.
[52] Hunklinger S. and Raychaudhari A.K., Progr.Low T.Phys. IX, ed.D.F.
Brewer (North Holland, Amsterdam 1986) p.265.
[53] Phillips W.A., ed., Amorphous Solids. Low-Temperature Properties
(Springer 1981).
[54] Shoemaker D.P. and C.B., Acta Cryst. B 42 (1986) 3.
[55] KIernan M. and Sadoc J.F., J. Physique Lett. 40 (1979) 569.
[56] Rivier N. and Lawrence A.J.A., "Disorder as projection from ideal,
curved space". J. Physique (ColI.) 46 (1985) C8-409.
[57] Sadoc J.F. and Mosseri R., J. Physique 46 (1985) 1809.
[58] Sadoc J.F. and Rivier N., "Hierarchy and disorder in non-crystalline
structures". Phil. Mag. B55 (1987) 537.
290 N. llivier

[59] Misner C.W., Thorne K.S. and Wheeler J.A., 1973, Gravitation,
Freeman, San Francisco, ch. 42.
[60] Rivier N. and Sadoc J.F., "The local geometry of disorder". J. Non-cryst
Solids 108 (1988) 282.
[61] Mendes-France M., in Number Theory and Physics, J.M. Luck
P. Moussa and M. Waldschmidt, eds., Springer Proc. in Phys. 47 (1990)
195.
[62] Shull C.G. , Strauser W.A. and Wollan E.O., Phys. Rev. 83 (1951) 333.
[63] Van Vleck J.H., Physics Bulletin 19 (1968) 167.
[64] Sir Lawrence Bragg and Nye J.F. , "A dynamical model of a crystal
structure". Proc. Roy. Soc. 190 (1947) 474. Reprinted in The Feynman
Lectures in Physics, 11-30-9, Addison-Wesley 1964.
[65] Harris W.F., Scient. Amer. 237 (1977) 130.
[66] Bernstein H.J. and Phillips A.V., "Fiber bundles and quantum
theory". Scient. Amer. 246 (1981) 94.
[67] Shapere A. and Wilczek F., "Gauge theory of swimming",
Amer. J. Phys. 67 (1989) 514; Phys. World 2 (1989) 36.
[68] Tonnelat M.-A., Les Theories Unitaires de I'Electromagnetisme et de la
Gravitation, Gauthier-Villars, Paris 1965.
... and Beyond
 COURSE 11

Automata and automatic sequences

J .-P. Allouche e) and M. Mendes France (2)
e) CNRS, LMD, Luminy Case 930,
F-13288 Marseille Cedex 9, France
(2) Universite Bordeaux I, Matbematiques,
351 cours de la Liberation,
F-33405 Talence Cedex, France

In the following pages we discuss infinite sequences defined on a finite alpha-

bet, and more specially those which are generated by finite automata. We have
divided our paper into seven parts which are more or less self-contained. Need-
less to say, we feel that the order we propose is the most natural one. References
appear at the end of each one of the parts which implies some redundancy.

PART I. Introduction. Substitutions and finite automata

1. The Fibonacci sequence
2. The Prouhet-Thue-Morse sequence
3. The paperfolding sequence
4. Automatic sequences: definition; a zoo of examples. A warning
5. Where finite automata enter the picture
6. How random can an automatic sequence be?
7. Miscellanea
8. Appendix: automata generating the five examples of Paragraph 4
References

Part II. Further properties of paperfolding

1. Direct reading and reverse reading
2. Words and diagrams
3. The folding operators
4. The dimension of a curve
5. Paperfolding and continued fractions
References
294 J.-P. Allouche and M. Mendes France

Part III. Complements

1. Repetitions in infinite sequences
2. Multidimensional morphisms and (finite) automata
3. Links with cellular automata
References

Part IV. Fourier analysis

1. Fourier-Bohr coefficients
2. Bessel's inequality and Parseval's equality
3. The Wiener spectrum and the spectral measure
4. Analyzing the spectral measure
5. Appendix I: the spectral measure of the Thue-Morse sequence
6. Appendix II: the spectral measure of the paperfolding sequence
References

Part V. Complexity of infinite sequences

1. Sturmian sequences and generalizations
2. Complexity of infinite sequences
References

Part VI. Opacity of an automaton

1. Automata revisited
2. Computing the opacity
References

Part VII. The Ising automaton

1. The inhomogeneous Ising chain
2. The induced field
3. The Ising automaton
4. An ergodic property
5. Opacity of the Ising automaton
References
 AUTOMATIC SEQUENCES 295

PART I
Introduction. Substitutions and finite automata
The discovery of quasicrystals in 1984 by Shechtman et al [34] led to renewed
studies of the Penrose tiling with its fivefold symmetry. In a regular Penrose
tiling one can observe infinite arrays or worms of short and long ties distributed
according to the Fibonacci sequence. Similar sequences, which are generated
by substitutions (or substitution rules), became very popular among physicists.
Instead of substitutions computer scientists rather speak of morphisms whereas
physicists also use the term inflation rules.

1. THE FIBONACCI SEQUENCE

Let us first recall what the Fibonacci morphism is and how it generates the
Fibonacci sequence. One starts with the alphabet (finite set) {O, I}. One then
considers the set {O, I} * of all finite words on this alphabet, a finite word being a
finite string (for example 0101110) which may be empty. The concatenation of
words, consists in assembling words together one after the other, thus creating
a longer word. For instance 011.1010 = 0111010. The set {O, 1}* endowed with
the operation of concatenation is called the free monoid generated by {O, I}.
We now define the map u on {O, I} by:

u(O) = 01, u(l) = O.

This map can be extended "morphically" to a map from {O, 1}* to itself, for
example:

u(0101110) = u(O)u(l)u(O)u(l)u(l)u(l)u(O) = 0100100001.

The image of an infinite sequence is defined the same way.
Let us compute now the iterates of the map u applied to the one-letter word
0:
u(O)= 01,
= u(Ol) = 010,
u 2 (0)
u (0) = u(010) = 01001,
3

The reader will have noticed that each new word begins with the previous one,
which implies that the sequence of words uR(O) converges to an infinite word

w=0100101001001 .. ·

called the Fibonacci sequence. This infinite word is by construction a fixed

point of u: u(w) = w.
296 J.-P. Allouche and M. Mendes France

Note that a rule as the one above is a growth rule in the terminology of D.
Gratias, and not a matching rule. A matching rule is local, but both types of
rules are drawing rules.
That such a rule gives - or may give - a quasicrystalline structure will be
discussed all along this volume. A quasicrystalline structure is defined either
via the notion of quasilattices, see the contribution of Y. Meyer [26] in this
volume, or via Fourier transforms. It is somewhere between periodicity and
chaos (or randomness).
Note also that these morphisms are particular cases of formal grammars
which were invented as toy models for the study of languages. Needless to say,
computer scientists were among the first to develop the theory.

2. THE PROUHET-THUE-MORSE SEQUENCE

Let us make a (seemingly small) change in the definition of the Fibonacci

morphism, and let us consider the morphism defined by:

o---+ 01, 1 ---+ 10.

Iterating this morphism starting from 0 yields:

o
01
0110
01101001
0110100110010110

hence, as previously noticed, we obtain an infinite sequence which is a fixed

point of the morphism. This sequence is the celebrated Prouhet-Thue-Morse
sequence, (see [29], [36J, [37J, [27]). The sequence is also known as the Thue-
Morse sequence or the Morse sequence. We shall freely use any of these de-
nominations.
What makes this sequence drastically different from the Fibonacci sequence
is that the images of 0 and 1 have the same length, namely 2. The importance
of this observation will be emphasized in two ways:
- denoting by (Un)nEN the Prouhet: Thue-Morse sequence, it is not difficult
to prove that Un = 1 if and only if the number of 1's in the binary expansion of
n is odd. Such a simple property is due to the underlying binary expansion in
the morphisms of length 2. If one wants a similar arithmetic definition for the
Fibonacci sequence, one should introduce the Fibonacci expansion of integers
which is not as straightforward;
 AUTOMATIC SEQUENCES 297

- write down the Fibonacci sequence in a way which emphasizes the definition
as a fixed point of the morphism u:
01 0 01 01 0 01 0 01
01001010
where each block 01 or 0 is written above the letter it originates from. This can
be seen as a renormalization process. The constant length Prouhet-Thue-Morse
morphism can be seen as a constant renormalization process:
01 10 10 01 10 01 01 10
o 1 1 0 100 1
3. THE PAPERFOLDING SEQUENCE

In this section we will discuss a slightly more general way of generating

sequences: the morphisms will be followed by a literal map.
Let us examplify this with the paperfolding sequence. This sequence can be
generated by regularly folding a piece of paper onto itself and by considering,
after unfolding, the sequence of folds ("mountains" and "valleys") created on
its edge (see for instance [18] and [21], see also Part II of this paper). This
sequence can also be defined as follows.
Consider the alphabet A = {a, b, c, d} and the morphism on A defined by
a --+ ab
b --+ cb
c --+ ad
d --+ cd.
This morphism admits a fixed point, obtained as usual by computing the iter-
ates of the morphism applied to the letter a:
abcbadcbabcd ...
One then applies the literal map (which can be also seen as a morphism of
length 1 onto a different alphabet) defined by:
a --+ 1
b --+ 1
c--+O
d --+ 0
thus obtaining the sequence
110 1 1 0 0 1 1 100
which is the paperfolding sequence on the alphabet {O, I}. The sequence is
obtained as the projection of a fixed point of a morphism of constant length.
298 J.-P. Allouche and M. Mendes France

4. AUTOMATIC SEQUENCES: DEFINITION; A ZOO OF EXAM-

PLES. A WARNING

DEFINITION. Let k be an integer greater than or equal to 2. A sequence

U = (Un)nEN with values in the alphabet A is said k-automatic if there exists:
• an alphabet B,
• a uniform morphism of B"', of length k (i.e. such that the image of each
letter in B is a word of length k in B"'),
• an infinite sequence with values in B, say v = (Vn)nEN, which is a fixed
point of this morphism,
• a map tp from B to A such that "In EN, tp(vn ) = Un.

Examples
- 1) The Prouhet-Thue-Morse sequence defined above is a 2-automatic se-
quence.
- 2) The paperfolding sequence defined above is a 2-automatic sequence.
- 3) The Rudin-Shapiro sequence (see [33], [31]) can be defined as follows.
Let A =
{a, b, c, d}. Consider the morphism on A:
a --+ ab
b --+ ac
c --+ db
d --+ de

Then define tp: A -+ {-I, +1} by:

a --+ +1
b --+ +1
c --+-1
d --+-1
The image of the fixed point of the above morphism by tp is the Rudin-Shapiro
sequence
+1 + 1 + 1 - 1 + 1 + 1 - 1 + 1 + 1 + 1 + 1 - 1 ...

Prove that the Rudin-Shapiro sequence cannot be obtained as the image of

the fixed point of a morphism (of length 2) on an alphabet with less than 4
elements.
This sequence is 2-automatic. It has been considered both by Shapiro [33]
and Rudin [31] in the hope of mimicking randomness in the following sense.
Consider an arbitrary sequence of +l's and -l's, say a = (an)n, and put
 AUTOMATIC SEQUENCES 299

Then one has:

where the quadratic norm is defined by:

1 ) 1/2
11/112 = ( 11/(OWdO

It is not hard to show that for a periodic (or even almost-periodic) sequence a
the order of magnitude of MN(a) is actually N (check for instance this claim
for the constant sequence an = 1 for all n). On the other hand it can be shown
that for Lebesgue-almost all sequences a one has, when N goes to infinity:

which can be reinterpreted by saying that, for a "random" sequence, the order
of magnitude of MN(a) is between,fN and v'NlogN.
The Rudin-Shapiro sequence is an example of a clearly deterministic se-
quence for which one can prove the existence of a positive constant C ::; 2 +.j2
such that:

This property of the Rudin-Shapiro sequence to simulate randomness will

also be seen later on (see the contribution of M. Queffelec [30] in this volume).
Let us see how the proof of ( *) works. We first show that the Rudin-Shapiro
sequence (an)n satisfies:

If (un)n is the fixed point of the above morphism on the four-letter alphabet
{a, b, c, d} the very definition of being a fixed point implies that:

where the maps a and (3 are defined by:

a(a) = a, a(b) = a, a(c) = d, a(d) = d,

(3(a) = b, (3(b) = c, (3(c) = b, (3(d) = c.
One then notices that:

<p 0 a = <p, <p 0 {32 = -<p 0 {3, <p 0 {3 0 a = <po

300 J.-P. Allouche and M. Mendes France

Hence, \In ~ 0:

a2n = <p(U2n) = <po:(un ) = <p(un ) = an,

a4n+1 = <p(u4n+d = <pj3o:(un ) = <p(un ) = an,

a4n+3 = <p(U4n+3) = <pj32(un ) = -<pj3(un ) = -<p(u2n+d = -a2n+l·

For all n ~ 0 define the two-dimensional vector Vn = ( an ).

a2n+l
Then:

where Mo and Ml are the 2 x 2-matrices:

Now let
2K_l
S(K,O):= L Vne2i1fn9.
n=O
One has:
2K +1 _l
S(K + 1,0) = L Vne2ilrn9
n=O
2K_l 2K_l
= E V2ne2ilr(2n)9 + E V2n+le2ilr(2n+l)9
n=O n=O
2K_l
= M(O) L V n e 2ilr (2n)9
n=O
= M(O)S(K, 20),

where
e 2i7r9 )
M(O) = ( 11 _e2i7r9 .

Therefore:
S(K, 0) = M(0)M(20) ... M(2 K- 1 0)S(0, 2K 0)

= M(0)M(20) ... M(2 K- 1 0) (~) .

Hence, taking the L 2-norm:
K-l
IIS(K,0)112 S; J2 II IIM(2i O)1I2.
i=O
 AUTOMATIC SEQUENCES 301

But the L 2 -norm of a matrix M = (~ ~z), where z is a complex number

of modulus 1 is easy to compute: this is the square root of the spectral radius
of the matrix t M M = (~ ~), hence IIMII2 = v'2. Thus for all ():

2K_l

L ane2ilm6 ~ IIS(K, (})1I2 ~ v'22K/2.

n=O

In other words the inequality M N ((}) ~ CVN holds if N is a power of 2 with

C = v'2. It is then possible to extend the inequality to arbitrary N provided
the value C = v'2 is replaced by C = 2 + J2.

- 4) The period-doubling sequence occurs when iterating a map of the unit

interval. Consider a one-parameter family of unimodal maps showing a Feigen-
baum cascade. Choose the value of the parameter where the "chaos" begins.
Then the kneading sequence of the point where the function has its maximum
is universal (even in cases where the Feigenbaum constant is not obtained) and
equal to the period-doubling sequence. For a general reference on this subject
see [15] and the included bibliography. This sequence can be defined as the
fixed point of the morphism
0--+01
1 --+ 00;

it is therefore a 2-automatic sequence. Note that this sequence is actually

connected with the Thue-Morse sequence (see [5], this seems also to be implicit
from reading between the lines in [24]).

- 5) The Hanoi sequence underlies the game known as the Towers of Hanoi.
In this puzzle there are three pegs and N disks of increasing diameters. At
the beginning all the disks are stacked on the first peg in increasing order, the
topmost being the one with the smallest diameter. Then, at each step, one may
take a disk on the top of a peg and put it on another peg provided that it is
never stacked on a smaller disk. The game ends when all the disks are on a new
peg (necessarily in increasing order). Denote by a, b, c the moves consisting in
taking the topmost disk from peg I, (resp. peg II, III), and putting it on peg
II, (resp. III, I), and by a, b, c the inverse moves. It can be shown that there
exists an infinite sequence on the six-letter alphabet A = {a, b, c, a, b, c} such
that its prefixes of length 2N - 1 give for each N the minimal strategy to move
N disks from peg I to peg II if N is odd and from peg I to peg III if N is even.
Furthermore this sequence can be obtained (see [4]) as the fixed point of the
302 J.-P. Allouche and M. Mendes France

morphism defined on A by:

a--ac
b--cb
c--ba
a--ac
b--cb
c--ba.
It is hence also a 2-automatic sequence. For a survey and a "direct" proof of
automaticity, see [6]. Another general survey on the towers of Hanoi is [23].
Warning
The definition ofthe Fibonacci sequence (0 - - 01, 1 - - 0) uses a morphism
of non-constant length. Hence the question arises whether this sequence can
be generated in another way (constant length morphism followed by a literal
map) i.e. whether it is k-automatic for some k ::::: 2. The answer is NO: it can
be shown that, for all k ::::: 2, this sequence is not k-automatic.
In the same vein consider the "cyclic" towers of Hanoi, where the towers are
disposed on a circle and where only clockwise moves are permitted (using the
aboves notations this means that one may use only the moves a, b and c). Then
there exists, as previously, an infinite sequence which is the limit of the finite
sequences of moves for N disks. This sequence is given (see [6]) by the following
morphism (of non-constant length) on the alphabet A = {f,g, h, u, v, w}:
I--Ivl
g--gwg
h--huh
u--Ig
v--gh
w--hl,
followed by the projection:

u--c

w--a.
But it can be shown (see [1]) that, for all k ::::: 2, this sequence is not k-
automatic.
Many other sequences discussed by physicists are not k-automatic: for in-
stance the so-called generalized Fibonacci sequences, generated by 0 __ 01 a+1,
 AUTOMATIC SEQUENCES 303

1 --+ 01 a, are k-automatic for no value of k ~ 2. Another example of a se-

quence which is k-automatic for no value of k ~ 2 is the "circle sequence": this
sequence is associated to the substitution a --+ cae, b --+ aeeae, e --+ abeae;
(actually the substitution has no fixed point, but its square has two fixed
points).

5. WHERE FINITE AUTOMATA ENTER THE PICTURE

In this section we give the reason why the terminology "automatic" is used.
A finite automaton (with output function) - also called "uniform tag system" by
Cobham in [14] - consists of states, transition functions and an output function.
It reads words written with an input alphabet and computes an output for any
given word. More precisely, if k ~ 2 is an integer, a k-automaton consists of:
• a finite set S of states; one of the states is distinguished and is called the
initial state. S can be thought of as the set of vertices of a graph;
• k (transition-) maps from S to S labelled 0, 1, "', k - 1, represented as
oriented edges (arrows) of the above graph;
• an output function <p, i.e. a map from the set of states S to some set U.
Such a machine works as follows: the input alphabet is {O, 1,···k - 1}.
One chooses to read words on this alphabet either from left to right (direct
automaton) or from right to left (reverse automaton). Given an input word
one reads it and feeds the automaton with its letters interpreted as transition
maps, starting from the initial state and following the arrows in order. Having
finished to read the given word, one finds oneself on a state. Take the image of
this state by the output function. Hence to each word on {O, 1,···k - 1} one
associates an element of U.
The machine generates an infinite sequence ('U,,),,>o as follows: to each input
n EN, or rather to its k-ary expansion, the automaton associates an element
'Un E U.

Example
Take k = 2, hence the input alphabet is {O, 1}. Let S = {A, B, C} and let A
be the initial state. The arrows of .the graph below define the transition maps.
Put <p(A) = <p(B) = 0, <p(C) = 1.

o 1
304 J.-P. Allouche and M. Mendes France

Choose to read words from right to left. The input word 001110 yields:
001110.A = 0011l.A =001l.B =OOl.A =
OO.B =
O.C =
Aj and finally
<p(A) = O.
To the sequence of binary integers corresponds the sequence of outputs:
0--0
1--0
10--0
11--0
100--0
101--1
110--0
111--0
1000 --0
1001-- 0
1010 - - I
1011--0

We shall give in the last paragraph of this chapter more examples of au-
tomata which actually generate the five automatic sequences discussed in Para-
graph 4.
Finally, and to conclude the section, we mention a theorem of Cobham's
[14] which asserts that a sequence is generated by a k-automaton as above if
and only if it is the literal image of a fixed point of a morphism of length k.
Furthermore automatic sequences can be generated either by direct or reverse
reading automata.

6. HOW RANDOM CAN AN AUTOMATIC SEQUENCE BE?

The automatic sequences occur in many mathematical fields (number theory,

harmonic analysis, fractals ... ) but also in theoretical computer science, or even
in music (see [7] for example). The "philosophical" reason, if we may say, is
that they lie between order (periodicity) and chaos (randomness), but that
they are however easy to construct and ... theorems can actually be proved.
Now, what happen~ if one perturbs an automatic sequence?
Firstly one can change the name of the letters without changing the struc-
ture of the sequence: for instance the Rudin-Shapiro sequence above has been
defined on the alphabet {-1, +1}. As far as automaticity is concerned it is
perfectly irrelevant to take {-I, +1}, {O, I} or say {a, b}.
Secondly the family of k-automatic sequences is closed under simple oper-
ations such as shifting, changing finitely many terms, adding two sequences,
etc.
 AUTOMATIC SEQUENCES 305

How "random" can an automatic sequence be? Of course we should ex-

clude the periodic or ultimately periodic sequences. The remaining automatic
sequences:
- have zero entropy (whatever the definition of entropy is, see the contribu-
tion ofV. Berthe [9] in this volume), hence are deterministic;
- have a low factor complexity (see Part III of this paper), more precisely
the number of blocks of length n in an automatic sequence is O(n), whereas
for a "random" binary sequence one has 2n blocks;
- form a countable set, (hence there are very few of them, contrarily to what
would be expected from "random" sequences);
- form a set of measure zero, (of course almost all the sequences on a finite
alphabet are expected to be "random");
- BUT they may have an absolutely continuous spectrum, such as the Rudin-
Shapiro sequence for which the spectral measure is actually equal to the Le-
besgue measure, just as would be expected for a random sequence (see the
contribution of M. Queffelec [30] in this volume).

7. MISCELLANEA

• Define a sequence of words on the alphabet {O, 1} by Wo = 0, Wn+1 = Wn Wn ,

"In;::: 0, where W is obtained from W by replacing the O's by l's and vice versa.
Hence WI = 01, W2 = 0110, Wg = 01101001, ...
Prove that the sequence of words (wn)n converges to the Prouhet-Thue-
Morse sequence, (all these words are prefixes of our Prouhet-Thue-Morse se-
quence). For other such examples see [4], and for general results see [32].

• Define a sequence of words on the alphabet {O, 1} by Wo = 1, wn+1

w n lf(w n ), "In ;::: 0, where f(w) is obtained from W by reading W backwards
then replacing the O's by 1's and vice versa. Prove that the sequence of words
one obtains converges to the paperfolding sequence. This construction is known
as "perturbed symmetry", (see [10]).

• A deep theorem of Cobham (see [14], see also [11]) asserts that the only se-
quences which are both k- and i-automatic for two multiplicatively independent
integers k and i (i.e. such that the ratio ~~~~ is irrational) are the ultimately
periodic sequences. There is yet no general result for the non-constant length
substitutions. Try to guess such a result (there is a conjecture by Hansel [22],
and a partial solution has been given quite recently by F. Durand). Before
guessing, consider the sequence u = (un)n which is the fixed point of the mor-
phism
0---+12
1 ---+ 102
2 ---+ O.
306 J.-P. Allouche and M. Mendes France

Now prove that this sequence can be obtained by taking the fixed point begin-
ning by 1 of the following morphism (this result is due to Berstel [8]):

0---+12
1 ---+ 13
2 ---+ 20
3 ---+ 21

and projecting it modulo 3

o ---+ 0, 1 ---+ 1, 2 ---+ 2, 3 ---+ O.

The same sequence

1021201020···

has a remarkable property which is worth mentioning: it has no squares in it

(i.e. no two consecutive identical blocks), see Part III ofthis paper. For surveys
on similar topics see also [8] and [3].
Here is yet another property of the sequence which links it to the Prouhet-
Thue-Morse sequence. Add 1 modulo 3 to each term, obtaining:

2102012101···.

This is the sequence of lengths of the strings of 1's between two consecutive O's
in the Morse sequence

011010011001011010010

• Let A be defined as the smallest (for the lexicographical order) subset of N

which contains 1 and such that n E A =::} 2n rt. A. Hence the elements of A
in increasing order are:

1 3 4 5 7 9 11 12 13 15 16 17 19

Now take the first differences to get:

211222112112···.

The reader might now consider again his favourite Prouhet-Thue-Morse se-
quence:
0110100110010110100···
and write down the lengths of the blocks composed only with O's or only with
1'so He will obtain:

121 1 222 1 121 1 2

 AUTOMATIC SEQUENCES 307

which is - up to the first term - the sequence of first differences above. To finish
with these two sequences of l's and 2's, the patient reader can prove that the
first one is the fixed point of:
2 --+ 211
1--+2
whereas the second one is the fixed point of:

1 --+ 121
2 --+ 12221.

To read more on these sequences see [2] and [35]. Note that it has been
proved independently by Mkaouar and Shallit that both sequences above are
not 2-automatic .
• Another intriguing sequence is the self-running Kolakoski sequence. Consider
any infinite sequence of l's and 2's, say for example:

u : 2 111 2 1 22 1 ....

Let K(u) be the sequence formed by the lengths of the strings of l's and the
strings of 2's. In our example:

K(u) : 1 3 11 2

The Kolakoski sequence is the only sequence beginning by 2 which satisfies the
remarkable property K(u) = u.

u: 22 --.,...
--.,... 11 --.,...
2 --.,...
1 --.,...
22 1
K(u): 2 2 1 1 2

It can be shown that this sequence can be obtained by iterating the following
map:
11 --+ 21
12 --+ 211
22 --+ 2211
21 --+ 221.
Note that this map is not a morphism (there is some kind of "context-depen-
dence") and that it is not known whether this sequence can be obtained as the
projection of the fixed point of a morphism (even with non-constant length).
For more properties of this sequence see [25], [19], [20], [38], [16], [17], [12] and
[28].
This last example may open the way to new constructions of sequences which
are still algorithmically (easy) computable but not "too simple" , hence again
between order and chaos. An interesting paper of Wen and Wen [39] discusses
308 J.-P. Allouche and M. Mendes France

the composite iteration of two different morphisms ruled by a third morphism,

(see also [16] and [17]).

8. APPENDIX: AUTOMATA GENERATING THE FIVE EXAM-

PLES OF SECTION 4

Ocw. 1l~o
1) A reverse automaton generating the Prouhet-Thue-Morse sequence:

Output function: rp(A) = 0, rp(B) = 1.

Can you find what the direct automaton is?

2) A reverse automaton generating the paperfolding sequence:

1 o

Output function: rp(A) = 1, <p(B) = 1, rp(C) = 1, rp(D) = O.

The direct automaton is given in Part II.

3) A reverse automaton generating the Rudin-Shapiro sequence:

o()fo 1 0 ~ 1 1 ~ 0 1 ~o
Output function: rp(A) = +1, rp(B) = +1, rp(C) = -1.
The value rp(D) is irrelevant. What is the direct automaton?

4) A reverse automaton generating the period-doubling sequence:

 AUTOMATIC SEQUENCES 309

Output function: <p(A) = 0, <p(B) = 1, <p(C) = 0, <p(D) = 1.

5) A reverse automaton generating the Hanoi sequence:

Output function: <p(A) = any value, <p(B) = any value,

<p(C) = C, <p(D) = c, <p(E) = "ii,
<p(F) = "ii, <p(G) = b, <p(H) = b,
<p(J) = a, <p(K) = b, <p(L) = b,
<p(M) = c, <p(N) = c, <p(P) = a.
310 J.-P. Allouche and M. Mendes France

References

[1] Allouche J.-P., Note on the cyclic towers of Hanoi, Theoret. Comput.
Sci. 123 (1994) 3-7.
[2] Allouche J.-P., Arnold A., Berstel J., BrIek S., Jockusch W., Plouffe S.,
Sagan B. E., A relative of the Thue-Morse sequence, Discrete Math.
(1994) to appear.
[3] Allouche J .-P., Astoorian D., Randall J., Shallit J., Morphisms, square-
free strings and the tower of Hanoi puzzle, Amer. Math. Monthly 101
(1994) 651-658.
[4] Allouche J .-P., Betrema J., Shallit J., Sur des points fixes de morphismes
d'un monoide libre, RAIRO, Informatique Theorique et Applications 23
(1989) 235-249.
[5] Allouche J.-P., Cosnard M., Iteration de fonctions unimodales et suites
engendrees par automates, C. R. Acad. Sci., Paris, Serie A 296 (1983)
159-162.
[6] Allouche J.-P., Dress F., Tours de HanoI et automates, RAIRO, Infor-
matique Theorique et Applications 24 (1990) 1-15.
[7] Allouche J.-P., Johnson T., Finite automata and morphisms in assisted
musical composition, J. New Music Research 24 (1995) to appear.
[8] Berstel J., Sur la construction de mots sans carre, Seminaire de Theorie
des Nombres de Bordeaux, Expose 18 (1978-1979) 18-01-18-15.
[9] Berthe v., this volume.
[10] Blanchard A., Mendes France M., Symetrie et transcendance, Bull. Sci.
Math. 106 (1982) 325-335.
[11] Bruyere V., Hansel G., Michaux C., Villemaire R., Logic and p-recogni-
zable sets of integers, Bull. Belg. Math. Soc. 1 (1994) 191-238.
[12] Carpi A., Repetitions in the Kolakoski sequence, Bull. EATCS 50 (1993)
194-196.
[13] Cobham A., On the base-dependence of sets of numbers recognizable by
finite automata, Math. Systems Theory 3 (1969) 186-192.
[14] Cobham A., Uniform tag sequences, Math. Systems Theory 6 (1972)
164-192.
[15] Collet P., Eckmann J.-P., Iterated maps on the interval as dynamical
systems (Progress in Physics, Birkhauser, 1980).
[16] Culik II K., Karhumaki J., Lepiso A., Alternating iteration of morphisms
and the Kolakoski sequence, in: Lindenmayer systems: impacts in theo-
retical computer science, computer graphics and developmental biology,
G. Rozenberg and A. Salomaa Eds (Springer Verlag, 1992) p. 93-103.
[17] Culik II K., Karhumaki J., Iterative devices generating infinite words, in:
STACS 1992: 9th Annual Symposium on Theoretical Aspects of Com-
puter Science, A. Finkel and M. Jantzen Eds (Lecture Notes in Computer
Science, 577, 1992) p. 531-543.
[18] Davis C., Knuth D. E., Number representations and dragon curves, J.
Recr. Math. 3 (1970) 161-181; 133-149.
 AUTOMATIC SEQUENCES 311

[19] Dekking F. M., Regularity and irregularity of sequences generated by

automata, Seminaire de Theorie des Nombres de Bordeaux, Expose 9
(1979-1980) 9-01-9-10.
[20] Dekking F. M., On the structure of self-generating sequences, Seminaire
de TMorie des Nombres de Bordeaux, Expose 31 (1980-1981) 31-01-31-
06.
[21] Dekking F. M., Mendes France M., van der Poorten A., FOLDS!, Math.
Intell. 4 (1982) 130-138, 173-181 and 190-195.
[22] Hansel G., private communication.
[23] Hinz A. M., The tower of Hanoi, Enseign. Math. 35 (1989) 289-321.
[24] Jonker L., Periodic orbits and kneading invariants, Proc. London Math.
Soc. 39 (1979) 428-450.
[25] Kolakoski W., Self-generating runs, Prob. 5304, sol. in Amer. Math.
Monthly 73 (1966) 681-682.
[26] Meyer Y., this volume.
[27] Morse M., Recurrent geodesics on a surface of negative curvature, Thans.
Amer. Math. Soc. 2 (1921) 84-100.
[28] Paun G., How much Thue is Kolakoski?, Bull. EATCS 49 (1993) 183-
185.
[29] Prouhet E., C. R. Acad. Sci., Paris 33 (1851) 225.
[30] Queffelec M., this volume.
[31] Rudin W., Some theorems on Fourier coefficients, Proc. Amer. Math.
Soc. 10 (1959) 855-859.
[32] Shallit J., A generalization of automatic sequences, Theoret. Comput.
Sci. 61 (1988) 1-16.
[33] Shapiro H. S., Extremal problems for polynomial and power series (The-
sis, M.I.T., 1951).
[34] Shechtman D., Blech I. A., Gratias D., Cabn J. W., Metallic phase with
long-range orientational order and no translational symmetry, Phys. Rev.
Letter 53 (1984) 1951-1953.
[35] Tamura J., Some problems having their origin in the power series
E:=l z[Qn j , in: Reports of the meeting on analytic theory of numbers
and related topics, (Gakushuin University, 1992), p. 190-212.
[36] Thue A., Uber unendliche Zeichenreihen, Norske vid. Selsk. Skr. I. Mat.
Kl. Christiana 7 (1906) 1-22.
[37] Thue A., Lage gleicher Teile gewisse Zeichenreihen, Norske vid. Selsk.
Skr. I. Mat. Kl. Christiana 1 (1912) 1-67.
[38] Weakley W. D., On the number of Coo-words of each length, J. Combin.
Theory Ser. A 51 (1989) 55-62.
[39] Wen Z.-X., Wen Z.-Y., Remarques sur la suite engendree par des substi-
tutions composees, Ann. Fac. Sci. Toulouse Math. 9 (1988) 55-63.
312 J.-P. Allouche and M. Mendes France

PART II
Further properties of paperfolding
In this part we shall discuss some unrelated properties of the paperfolding
sequence and its generalizations. We shall be concerned among other things
with dimensions and continued fractions.

1. DIRECT READING AND REVERSE READING

Consider the two automata A' and A" where A' (resp. A") is the initial
state.

Output function: tp'(A') = tp'(B') = a, tp'(C') = tp'(D') = b.

1 o

Output function: tp"(A") = tp"(B") = tp"(C") = a, tp"(D") = b.

The input instructions are to be read from left to right for the automaton
A' (direct reading) whereas for the automaton A" we are to read from right
to left (reverse reading). For example "nineteen"=10011 puts A' onto state B'
 AUTOMATIC SEQUENCES 313

=
and A" onto state Gil. In both cases the output is a <p' (B') = <p" (Gil). We
leave it as an exercise to verify that for all n = 0, 1,2· ..

where A~ (resp. A~) is the state of the automaton A' (resp. A") at step n.
This illustrates a general result that we had already mentioned in Part I:
automatic sequences can be generated either by direct or reverse reading au-
tomata.

2. WORDS AND DIAGRAMS

Let {L, R}* be the set of words on two symbols L (left) and R (right).
To each word w E {L, R}* corresponds a diagram on the latti~e Z2 in the
following way. Starting from the origin follow the horizontal unit interval. At
the endpoint turn left or right according to the instruction given by the first
letter of the word w. Having then followed the vertical segment (upward or
downward) read out the second letter of w which instructs you to turn left or
right on a horizontal segment, etc... For example the diagram corresponding
tow=LRLLis
_ _ _...,L

R.--_ _--IL

The length of the word is Iwl = 4 and the length of the diagram is Iwl + 1 = 5.
Similarly L4 = LLLL represents

and LRL3 RL represents

314 J.-P. Allouche and M. Mendes France

Lr--_ _..,L

R
LI-~~R~-....JL

The first and the third examples show self-avoiding curves: edges are de-
scribed only once. The associated words are said to be self-avoiding. The third
example illustrates a closed curve or cyclic curve: the two ends coincide. We
then speak of cyclic words.
Later we shall encounter infinite length diagrams which correspond to infinite
sequences of R's and L's.

3. THE FOLDING OPERATORS [10]

Folding a word or its associated diagram is done in three steps.

Step 1. Fold the polygon w onto the unit segment by shrinking all angles
to O. Hence L3 becomes

Step 2. Multiply the length of the edges by 2 and fold the collapsed diagram
onto itself either in the positive direction (operator F+) or in the negative
direction (F_ )

• •

Step 3. Unfold to 90 0 • We thus obtain a new diagram Wi twice as long as

the original one which we denote F+{w) or F_{w). Clearly 1F±{w)1 = 21wl-1.
As an example F+{L 3 ) is the diagram
 AUTOMATIC SEQUENCES 315

L r--_ _---,L

L-.._ _ =R L
L R

so that

The following result can easily be established.

THEOREM 1. Let w = AoAl '" A n - 1 E {L,R}*. Then

where
L if n == 0 (mod 2),
x= { R if n == 1 (mod 2).

Similarly

where
R jf n == 0 (mod 2),
y= {
L if n == 1 (mod 2).

The operation which is involved in this theorem is the so-called shuffle

product III (Russian letter "Sha"): if w = Ao Al ... A n - 1 An and w' =
Bo Bl ... Bn-l (it may happen that An is the empty letter), then

The theorem asserts

F+(w) = (L R)X III w
F_(w) = (RL)X III w

where it is to be understood that (L RY (resp. (R L)X) is the alternating word

LRLR ... (resp. RLRL) of length Iwl + 1.
Let S be the set of self-avoiding words and let C be the set of cyclic words.
316 J.-P. Allouche and M. Mendes France

THEOREM 2. (Davis and Knuth).

We refer to [3] for the proof.

Suppose we start out with the empty word 0. Then, in ac.cordance with
Theorem 1

As 0 is obviously self-avoiding (it is the unit interval!) so is F± (L or R) and

by induction, given any sequence Cl C2" 'Cn of (±) signs

is self-avoiding.
To an infinite sequence C = (cl C2 ..• ) of signs correspond an infinite diagram
called a dragon curve F£(0). The above discussion shows that dragon curves
are self-avoiding.
The figures below represent respectively Fe(0) where C = (+_)00 and C =
(+)00. (There are two dragon curves corresponding to C = (+_)00; we only
show one of them.)
AUTOMATIC SEQUENCES 317
318 J.-P. Allouche and M. Mendes France

The reader will no doubt have noticed that the infinite sequence F( +)"" (0) is
the paperfolding sequence discussed earlier in Paragraph 1 and in Part 1. If c: is
an arbitrary (±) sequence, we shall say that Fe (0) is a generalized paperfolding
sequence. It is easily seen that if Ao Al A2 . .. is the original paperfolding
sequence (the regular paperfolding) then

A2n is the alternating sequence (L R)oo ,

{
A 2n +1 = An.

More generally it Bo BI B2 ... is any paperfolding sequence then

B2n is an alternating sequence,

{
B 2n +1 is a paperfolding sequence.

This property characterizes the family of paperfolding sequences.

Exercise. The generalized paperfolding sequence Fe(0) is automatic if and
only if c: is ultimately periodic.
By simple inspection one feels that dragon curves are somehow space-filling
=
(space Z2), or that at least they fill portions ofZ2. This is indeed the case and
we shall show in the next paragraph that all dragon curves are 2-dimensional.
First of all we must define the concept of dimension.

4. THE DIMENSION OF A CURVE [4], [11].

Let r be a locally rectifiable unbounded curve in the plane. All arcs of r

have finite length unlike fractal Weierstrass curves, Peano curves or Brownian
curves. (See Mandelbrot [8] for the properties of the last mentioned curves.)
Consider a portion rL of r of length L. Let c: > 0 and let r(L,c:) be the
area of the c:-sausage of r L.

2c:

Denote by D( L) the diameter of r L:

D(L) = max{dist(M,N) 1M E rL, NErd.

By definition the dimension of r is

· (r) - I' . f log r(L, c:)

d1m
L_oo 1ogD ().
- lmln L
 AUTOMATIC SEQUENCES 319

It can be shown that dim (r) is independent of e > 0 and that

1 ~ dim (r) ~ 2.

A straight line, an infinite branch of an algebraic curve have dimension 1.

=
The spiral p exp 0 is also one-dimensional. Let a > O. The spiral p =OOt
has dimension 1 + l/a if a ~ 1 and dimension 2 if a ~ 1.
THEOREM 3. (Mendes France and Tenenbaum).
All dragon curves are two-dimensional.
Let us sketch the proof which seems quite relevant here since it involves
sequences related to the Rudin-Shapiro sequence. (For details see [12].)
The Nth vertex of a given dragon curve r has coordinates

YN= E tl:,
I:<N/2

where SI: and tl: are equal to ±1; :EN is the algebraic sum of the right and
left moves whereas YN is the sum of ups and downs. Both sequences (sn) and
(t n ) are closely linked to the Rudin-Shapiro sequence. In particular for the
dragon curve F(+_)",,(0) = LRRLRR··· the sequence (sn) is exactly the
Rudin-Shapiro sequence. The sequence (t n ) is closely related.
It can be shown that for all dragon curves the Rudin-Shapiro property holds,
namely

L SI: ~ (2 + .../2)JN/2,
I:<N/2

E tl: ~(2+.../2)JN/2.
I:<N/2

Therefore the diameter D( N) satisfies

for some constant a.

On the other hand, as the dragon curve is self-avoiding,

r(N,e) '"""' 2eN

so that
dim(r) ~ lim log(2eN) - 2
N -00 log a.,fN - .
320 J.-P. Allouche and M. Mendes France

But the dimension is at most 2 therefore

dim(r) = 2, QED.

The entropy of a curve (see V. Berthe [1]) is linked to the dimension. The
complete relationship is discussed in Mendes France [11]. The higher the en-
tropy, the higher the dimension, so in some sense the dimension measures the
complexity of a curve (this is one of Mandelbrot's principles). In that respect
dragon curves are among the most complex self-avoiding curves on &:;2.

5. PAPERFOLDING AND CONTINUED FRACTIONS

The continued fraction expansion of a real number :r: reads

1
:r: = ao + ----::1.--- = [ao, at, a2,···]
al+---

where the so-called partial quotients an ~ 1 are integers; ao may be negative.

If :r: is irrational the expansion is infinite whereas if :r: is rational the expansion
terminates. Due to the trivial equality
1
a = a-I + l' a ~ 2,

rational numbers have two expansions which only differ by the last two partial
quotients. The rank n of the last partial quotient an is called the depth of :r:.
The above remark allows us to assume with no loss of generality that the depth
is always even.
In this paragraph we shall discuss the continued fraction expansion of the
number

where g ~ 2 is a given integer (the case g = 2 is slightly more complicated so

we shall suppose g ~ 3). The following analysis is essentially due to Kmosek [6]
and Shallit [13]. See also the papers of Mendes France [9], Blanchard, Mendes
France [2] or Dekking, Mendes France, van der Poorten [5].
Let p/q be a rational number in the interval (0,1). We assume p and q
coprime. It can be shown that if
p
- = [0, aI, a2, ... ,an], (n even),
q

then
 AUTOMATIC SEQUENCES 321

If w = al a2 ... an is the word associated to p / q then the word corresponding

to p/q + l/q2 is
( --+) =wpw
S -pw --+ --+ +-

where W is the word W reversed and where p denotes the "operator" which
perturbs the last letter of wand the first letter of w.
P is the inverse operator:
it adds -1 (resp. +1) to the last letter of the first word (resp. first letter of
the last word).

Consider

Z2 = -g1 + 2"
1
g
= [0, g - 1, g + 1],

Z3 =1
-g + 1 1
2" + 2'2 = S-p(g -1,g + 1)
g g
= [0,g-l,g+2,g,g-I].

Then
Z4 = S-p [g - 1, g + 2, g, g - 1]
= [O,g-l,g+2,g,g,g-2,g,g+2,g-l]

and

Z = Zoo = Sr:;(g -1,g + 1)

= [O,g -1,g + 2,g,g,g - 2,g,g + 2,g,g - 2,g + 2,g,·· .].

The partial quotients of Z are completely determined by the algorithm. Their

values are 9 - 2, 9 -1, g, 9 + 2. Mahler [7] proved that z is transcendental and
this is quite an achievement since many of the easier proofs in transcendental
number theory are based on the speed with which the partial quotients tend
to infinity (think of Liouville numbers).

Let us analyse the structure of the sequence of the partial quotients of z.

Put w = 9 - 1, 9 + 1. Then z = Sr:; (w)

--+--+~-+-++-~-+--+---++-+----+
z=wpwpwpwpwpwpw···

The sequence of w's and p's is obviously periodic but more surprisingly (?) the
sequence of arrows is the regular paper-folding sequence. As it is automatic one
can correctly guess that the sequence of partial quotients is indeed automatic.
We show below the direct reading automaton which generates the sequence of
partial quotients.
322 J.-P. Allouche and M. Mendes France

o o

Output function: <p(A) = 9 + 2, <p(E) = g,

<p(B) =g, <p(F) = 9 + 2,
<p(C) = g, <p(G) = 9 - 2,
<p(D) = 9 - 2, <p(H) = g.

References

[1] Berthe V., this volume.

[2] Blanchard A., Mendes France M., Symetrie et transcendance, Bull. Sci.
Math. 106 (1982) 325-335.
[3] Davis C., Knuth D. E., Number representations and dragon curves, J.
Recr. Math. 3 (1970) 161-181; 133-149.
[4] Dekking F. M., Mendes France M., Uniform distribution modulo one: a
geometrical viewpoint, J. reine angew. Math. 129 (1981) 143-153.
[5] Dekking F. M., Mendes France M., van der Poorten A., FOLDS!, Math.
Intell. 4 (1982) 130-138, 173-181 and 190-195.
[6] Kmosek M., Report by A. Schinzel, Oberwolfach 28 may - 2 june 1979.
[7] Mahler K., Ein Analog zu einem Schneiderschen Satz, Proc. Akad.
Wetensch. Amsterdam 39 (1936) 633-640 and 729-737.
[8] Mandelbrot B., Les objets fractals (Nouvelle Bibliotheque Scient. Flam-
marion, Paris, 1975).
[9] Mendes France M., Sur les fractions continues limitees, Acta Arith. 2
(1973) 207-215.
 AUTOMATIC SEQUENCES 323

[10] Mendes France M., Paper folding, space-filling curves and Rudin-Shapiro
sequences, Contempory Mathematics 9 (1982) 85-95.
[11] Mendes France M., The Planck constant of a curve, in: Fractal Geometry
and Analysis J. Belair and S. Dubuc Eds (Kluwer Acad. Publ., 1991) p.
325-366
[12] Mendes France M., Tenenbaum G., Dimension des courbes planes, pa-
piers plies et suites de Rudin-Shapiro, Bull. Soc. math. France 109 (1981)
207-215.
[13] Shallit J., Simple continued fractions for some irrational numbers J.
Number Theory 11 (1979) 209-217.
324 J.-P. Allouche and M. Mendes France

PART III
Complements
In this chapter we will give some complements on finite automata: a quick
survey of the notion of "repetitions" in a finite or infinite word, the multidimen-
sional generalization of morphisms of constant length and of finite automata
and finally a short account on differences and similarities between finite au-
tomata and cellular automata.

1. REPETITIONS IN INFINITE SEQUENCES

1.1. The beginning of the story

The pioneering work of Thue (see [18] and [19]) stems from the following
easy observation:
OBSERVATION. Every infinite word on a two-letter alphabet contains at least
a (non-empty) square, (i.e. a block ww where w is a non empty word).
Proof
As the proof is very easy we will give it: let {a, b} be the two-letter alphabet.
Let us try to construct a "long" word with no square in it. With no loss of
generality we can suppose that the infinite word begins with an a. The following
letter cannot be an a (since aa is a square), hence our word begins with abo The
next letter cannot be a b (since the word would contain the block bb), hence
the word begins with aba. Now if we add a letter to aba the new word would
necessarily contain a square, as it would be either abaa or abab.
Note that the above reasoning shows that every word of length ~ 4 on a
two-letter alphabet contains a square. .
A natural question hence arises: is it possible to find an infinite sequence on
a three-letter alphabet without squares in it? One can also ask whether there
exists an infinite sequence on a two-letter alphabet without cubes, (i.e. without
blocks www where w is a non-empty word). The answer to both questions is
YES, see [18], [19] and the survey [8] with its bibliography:
THEOREM.
- The fixed point of the morphism
0-+12
1-+ 102
2-+0
has no square. This sequence is 2-automatic as mentioned in Part 1.
- The Thue-Morse sequence has no cube. More precisely this sequence has
no overlap, i.e. no word wwz, where z is the first letter of the word W. This
sequence is linked to the preceding one as we have seen in Part 1.
 AUTOMATIC SEQUENCES 325

Remark
Note that the Thue-Morse sequence contains arbitrarily long squares: indeed
it is a fixed point of the morphism (1' defined on {O, I} by:

(1'(0) = 01, (1'(1) = 10.

This sequence begins by 0110··· hence by (1'1:(0)(1'1:(1)(1'1:(1)(1'1:(0) for every k ~

2. Hence it contains the square (1'1:(1)(1'1:(1).

1.2. Why study repetitions?

When he asked the above questions concerning repetitions in sequences,
Thue added that he had no particular applications in mind but that the problem
seemed interesting and not trivial; there would certainly be applications in the
future.
And indeed this question has been studied very precisely by computer sci-
entists, see for example [11], see also [8]. The notion of morphism was again an
essential tool: our leitmotiv is that sequences generated by morphism are both
easy to construct and not necessarily trivial.
The existence of repetitions is somehow linked to randomness: a random
sequence is expected to be normal, (i.e. all possible blocks of given length
occur with the same frequency), hence a sequence without squares (or cubes
or ... ) cannot be random.
We finally mention a paper of Bovier and Ghez [9] on the discrete Schrodinger
equation with automatic potential where, curiously enough, one needs the au-
tomatic sequence of potentials to begin with a square. For more information
on "automatic Schrodinger equations" one can read the contribution of Siit&
[17] in this volume, one can also read [6], [7].

1.3. More examples

We quote here· two results which are interesting in that they lead to other
developments that we will not describe here.
- The paperfolding sequence has no square ww with Iwl ~ 5. This result
is announced in [1] and more details can be found in [2]. This implies that
this sequence does not contain any 12th power, i.e. that no block has the
form W 12 , but a more precise result is proved in [2]: the paperfolding sequence
has no fourth power. The result holds for any paperfolding sequence. For the
Rudin-Shapiro sequence(s) some results of the same kind are also given in [2].
- The Fibonacci sequence has no fourth power. It does however contain cubes
since it begins by:

010010100(10010)(10010)(10010)···

Considering this sequence one might also define fractional powers: for example
10010010 is a power 2+ ~ = ~, as it can be written (100)(100)(10). Considering
326 J.-P. Allouche and M. Mendes France

the supremum of the fractional powers which occur in an infinite word permits
us to define a real maximal (positive) power occurring in a sequence. For
the Fibonacci sequence one can prove that this maximal power is equal to
2 + T = 3.618· .. , see [13].

2. MULTIDIMENSIONAL MORPHISMS AND (FINITE) AUTO-

MATA

Going back to the Penrose tiling, remember that the underlying morphism
is the Fibonacci morphism:

0-01, 1-0.

This morphism is a one-dimensional morphism and it might seem odd to de-

scribe a two-dimensional pattern by means of a one-dimensional tool. Among
the possible generalizations of morphisms we will describe the two-dimensional
substitutions and the associated two-dimensional finite automata. Although
these morphisms are not suitable to describe the Penrose tiling, they are a
natural generalization of the uniform morphisms and have many interesting
properties.

2.1. An eXaIllple

Consider the alphabet {O, I} and the following map:

0
- 0 0 0
0 0 0
0 0 0
1
- 1 0 1
0 1 O.
1 0 1

Starting from the letter 1 and iterating this map by replacing each letter by a
3 x 3 square gives:

1 0 1 0 0 0 1 0 1
0 1 0 0 0 0 0 1 0
1 0 1 0 0 0 1 0 1

1
-
1 0 1
0 1 0
1 0 1 - 0 0 0
0 0 0
0 0 0
1 0 1
0 1 0
1 0 1
0 0 0
0 0 0
0 0 0 -
1 0 1 0 0 0 1 0 1
0 1 0 0 0 0 0 1 0
1 0 1 0 0 0 1 0 1
 AUTOMATIC SEQUENCES 327

An infinite number of iterations leads to the double infinite sequence:

1 0 1 0 0 1
0 0 1
0 1 0 0 0 0
0 1 0
1 0 1 0 0 1
0 0 1
0 0 0 1 0 1 0 0 0
0 0 0 0 1 0 0 0 0
0 0 0 1 0 1 0 0 0
1 0 1 0 0 0 1 0 1
0 1 0 0 0 0 0 1 0
1 0 1 0 0 0 1 0 1

which is by construction a fixed point of this morphism.

2.2. Properties
This construction, which can be extended to any dimension, gives rise to
multi-indexed sequences with many arithmetic properties (see [14] and [15]).
Note in particular that all lines, columns and diagonals of the above sequence
are 3-automatic. Quite obviously the same construction can be done with any
integer d ~ 2 instead of 3.
These sequences can be used in Physics: an application to the Robinson tiling
is given in [5]. One could also imagine a two-dimensional discrete Schrodinger
equation with potentials generated this way, thus permitting to study the eigen-
frequencies of, say, a two-dimensional quasi crystalline mattress of springs or of
a quasicrystalline trempolino. To our knowledge this has not yet been studied.
Another application to Physics is the construction of classical fractal figures:
first note that the constant renormalization we alluded to in Part I, paragraph
2, can be interpreted as selJ-similarityfor one-dimensional automatic sequences;
in the same way double automatic sequences are self-similar. Take the above
example, replace the 1's by black squares and the O's by white squares; after
each iteration renormalize the pattern of black and white squares. This gives
the well-known Sierpinski triangle. Compare with the classical construction
which can be seen as "dual". Many such fractals can be found in the Appendix
by Shallit to the paper [14], see also [16].
A last example is given by ... the handkerchief-folding, [14].

3. LINKS WITH CELLULAR AUTOMATA

Cellular automata are often used in Physics. They have been introduced by
von Neumann and they are discrete tools to mimick phenomena which can be
continuous. From the point of view of their computing power they are equiva-
lent to Turing machines, hence they are strictly more powerful in general than
finite automata. Many books on cellular automata have been published. See
328 J.-P. Allouche and M. Mendes France

for example [10]. For an example of a two-dimensional cellular automaton as a

model of a chemical reaction one can read [4], where the Greenberg and Hast-
ings model of the Belouzov-Zhabotinski-Zaikin oscillating reaction is discussed
in detail.
Let us give a quick definition of cellular automata. A cellular automaton
consists of:
- a lattice, Zd. Each vertex is called a site or a cell;
- a (usually finite) neighbourhood of sites Vi for each site i E Zd, such that
Vi = Vj + i - j, (translation invariance). The most frequently used neighbour-
hoods in two dimensions are the von Neumann and the Moore neighbourhood,
(see below);
- a set of states Q, (usually finite);
- a map I, called the transition function, from QV to Q, where v is the
cardinality of each Vi .

......- - i - -... ......--i--~

von Neumann's Moore's

neighbourhood of neighbourhood of
i in dimension 2 i in dimension 2
The evolution of the cellular automaton is the following: one assigns to each
site i an element xr in Q. The set

is called the initial configuration. For each i one then defines xt as the image
by I of (XJ)jEV•. Hence the set

C1 = {x;, i E Zd}

is constructed by parallel updating. One writes C1 = I(Co). Similarly C2 =

I(Cd, ... , Cn +! = I(Cn ). The index of iteration n is sometimes called the
(discrete) time-scale.
For many results on the dynamical behaviour of cellular automata we refer
to [12] and its bibliography.
In the case where the lattice is one-dimensional, one usually draws a two-
dimensional pattern each line of which is a configuration Ci. A simple class in
one dimension is the class of linear cellular automata: the values of the sites
 AUTOMATIC SEQUENCES 329

are taken in 7l../ d7l.. the ring of integers modulo d and the transition rules are
linear.
For example the cellular automaton given by the Pascal's triangle rule mod-
ulo d is linear. The two-dimensional pattern it generates is exactly the double
sequence U = =
«um,n)m,n) where um,n (r;:) mod d, hence the Sierpinski tri-
angle modulo d (up to rotation). It has been proved in [3] that the sequence U
is a double automatic sequence if and only if d is equal to a prime power pic.
The sequence U is then p-automatic.

References

[1] Allouche J.-P., Suites infinies a. repetitions bornees, Seminaire de Theorie

des Nombres de Bordeaux, Expose 20 (1983-1984) 20-01-20-11.
[2] Allouche J .-P., Bousquet-Melou M., Facteurs des suites de Rudin-Shapiro
generalisees, Bull. Belg. Math. Soc. 1 (1994) 145-164.
[3] Allouche J.-P., von Haeseler F., Peitgen H.-O., Skordev G., Linear cel-
lular automata, finite automata and Pascal's triangle, Discrete Appl.
Math. (1995) to appear.
[4] Allouche J.-P., Reder C., Oscillations spatio-temporelles engendrees par
un automate cellulaire, Discrete Appl. Math. 8 (1984) 215-254.
[5] Allouche J.-P., Salon 0., Robinson tilings and 2-dimensional automata,
in: Quasicrystals, Networks and Molecules of Fivefold Symmetry, I. Har-
gittai Ed (VCH Publishers Inc., 1990) p. 97-105.
[6] Axel F., Allouche J.-P., KIernan M., Mendes France M., Peyriere J.,
Vibrational modes in a one dimensional "quasi-alloy", the Morse case,
J. Phys. Colloques, 1986, Workshop on aperiodic crystals, Les Houches
(1986) 11-20.
[7] Axel F., Peyriere J., Spectrum and extended states in a harmonic chain
with controlled disorder: effects of the Thue-Morse symmetry, J. Stat.
Phys. 57 (1989) 1013-1047.
[8] Berstel J., Sur la construction de mots sans carre, Seminaire de Theorie
des Nombres de Bordeaux, Expose 18 (1978-1979) 18-01-18-15.
[9] Bovier A., Ghez J .-M., Spectral properties of one-dimensional Schro..
dinger operators with potentials generated by substitutions, Comm.
Math. Phys. 158 (1993) 45-66.
[10] Burks A. W., Essays on cellular automata (University of Illinois Press,
1970).
[11] Cassaigne J., Motifs evitables et regularites dans les mots (These, Uni-
versite Paris 6/7, 1994).
[12] Goles E., Martinez S., Neural and automata networks (Mathematics and
its applications, 58, 1991).
(13] Mignosi F., Pirillo G., Repetitions in the Fibonacci infinite word,
RAIRO, Informatique Theorique et Applications 26 (1992) 199-204.
330 J.-P. Allouche and M. Mendes France

[14] Salon 0., Suites automatiques a multi-indices, Seminaire de Thiorie des

Nombres de Bordeaux, Expose 4 (1986-1987) 4-01-4-27; followed by an
appendix by Shallit J., 4-29A-4-36A.
[15] Salon 0., Suites automatiques a multi-indices et algebricite, C. R. Acad.
Sci. Paris, Serie I 305 (1987) 501-504.
[16] Shallit J., Stolfi J., Two methods for generating fractals, Camp. and
Graphics 13 (1989) 185-19l.
[17] Siit8 A., this volume.
[18] Thue A., Uber unendliche Zeichenreihen, Norske vid. Selsk. Skr. I. Mat.
Kl. Christiana 7 (1906) 1-22.
[19] Thue A., Lage gleicher Teile gewisse Zeichenreihen, Norske vid. Selsk.
Skr. I. Mat. Kl. Christiana 1 (1912) 1-67.
 AUTOMATIC SEQUENCES 331

PART IV
Fourier analysis
We present a rather naive point of view on the Fourier analysis of complex
valued sequences. For a deeper understanding the reader is referred to Martine
Queff6lec's contribution to this volume [8].

1. FOURIER-BOHR COEFFICIENTS

Let I = (f( n)) be an infinite sequence of complex numbers such that for all
A E 1R the limit
N-l
1(>.) = N-+oo
lim N1 L l(n)exp(-2i7rAn)
n=O

exists. This hypothesis is rather strong and eliminates such sequences as

I(n) = exp(2i7rlogn), (n ~ 1),

for which the limit does not exist for A = O.

Observe that whenever the hypothesis is fulfilled I(A + 1) = 1(>.) so that it
is enough to restrict A to the interval [0,1[.
From now on we shall always assume the existence of 1(A) for all A E [0,1[.
Under this assumption it can be shown that the spectrum of I (the Fourier-Bohr
spectrum)
S(f) = {A I i(>.) =P O}
is at most countable. We use the notation

I(n) '" L 1(A)exp2i7rAn.

AES(J)

Warning: The series may well diverge. We do not specify in which order the
terms are to be written down. Even if the the series converges (whatever we
mean by convergence!) its sum may be different from f.
We now present some examples which can be considered as exercises.
Example 1. I(n) = exp(2i7rlogn).
As we already mentioned, 1(0) does not exist.
Example 2. Let a E 1R and define I( n) = (-1 )[an) where [... ] denotes the
integer part. One can show that the spectrum is

S(I) = {~+ k a IkE 1Z} (mod 1).

332 J.-P. Allouche and M. Mendes .France

Observe that when a = p/q, (p, q) = 1, is rational, the spectrum boils down to
the finite set
{ ~ + ; I k = 0, 1, ... ,q - 1} (mod 1).

When a is irrational

(_1)[an] '" :f- L

t'll" kEZ
_1_ exp 2i'll" (k
2k + 1
+ !)2 an.
Exercise: what is the corresponding formula when a is rational?
Example 3. Let f(n) = (_1)6(n) be the ± Thue-Morse sequence (s(n) is the
sum of the binary digits of n). Then 1(>.) = 0 for all A hence

f(n) '" o.

This result may be seen as a consequence of our Appendix 1. However a direct

computation is possible.
Example 4. The Rudin-Shapiro (±) sequence has empty spectrum:

f(n) '" O.

EX8U1ple 5. Let a be irrational. Consider f(n) = (_1)[an21. Then as in the

preceding cases, the spectrum is empty and therefore

f(n) '" O.

This example is worked out in the paper of J. Bass [1].

Example 6. Let f(n) be the regular paperfolding sequence with values ±1.
Then

2a+ 1
if \ - 2v +2 '
,,- ">
v_ 0, 0 ~ a < 2 +
v 1

if not

and

L
00

f(n - 1) '"
v=o

Examples 3 and 6 will be worked out explicity in the appendices (Paragraphs

5 and 6).
 AUTOMATIC SEQUENCES 333

2. BESSEL'S INEQUALITY AND PARSEVAL'S EQUALITY

Define the norm of a sequence 1= (f(n» by

( N-1 )1/2
11111 = l~..:~p ~ ~ I/(nW

This is really a semi-norm since it may well happen that

I =1= (0,0,0, ... ).
11/11 = °
and yet

Here are two such examples:

I -- (1 '2'3'4'5'
! ! ! ! ... 'n'
! ... )
and
~
1= (1, 0, 1, 0,
~
°, .--"'--.
1, 0,0,0, 0, 1, 0,··· , °,
.--"'--. .---.
1, 0,··· )
In this last example, the density of l's vanishes.
From now on we identify two sequences I and g as soon as III - gil = 0, so
that the two above examples are considered as null sequences.
Bessel's inequality asserts that in all cases

L 1j(,\W ~ 11/112.
>.es(J)

A series of sequences Uk(.) defined on N is said to converge to a sequence

s( .) with respect to the norm II .. ·11 if

In other words

We then write
L uk(n).
00

s(n) ~
k=O

This must not be confused with a pointwise convergence.

334 J.-P. Allouche and M. Mendes France

THEOREM 1. (Parse val [9])

~ I<A)exp2i1rAn ~ I(n)
)..ES(J)

if and only if
~ 11(AW = 11/112.
)..ES(J)

In this case we say that 1 is an almost-periodic sequence in the sense of Besi-

covitch; 1 is represented by a sum of exponentials.
We leave it to the reader to verify that (-1 )[an] and that the paperfolding
sequence are Besicovitch almost-periodic. Obviously the Thue-Morse sequence,
the Rudin-Shapiro sequence and (_1)[an2] (a irrational) are not since Parseval's
equality does not hold.

3. THE WIENER SPECTRUM AND THE SPECTRAL MEASURE

by
The correlation function ,f which we assume to exist for all h E Z is defined
,f
N-l
1
(h) = lim N
N-+oo
~ l(n)/(n + h).
n=O
A deep result of Bochner's and Herglotz's asserts the existence of a bounded
increasing function u f such that

du f is known as the spectral measure of I.

Let E be a measurable subset of the unit interval. Its u f measure is denoted

In particular, let A EjO, 1[. The measure of the set {A} is

11
Similarly
u f {I} = lim
o'\.o 1-0

and
 AUTOMATIC SEQUENCES 335

One should not confuse the measure u f {A} of the set {A} with the value
u f (A) of the function u f at A. There is of course a trivial relationship between
both

u f {.A} is the length of the discontinuity at A. In particular if u f is continuous

at A then u f {A} = 0 and conversely.
The set of A's such that u f {A} ::j:. 0 is called the Wiener spectrum and we
denote it W(f). It is obviously at most countable, (indeed the number of steps
is at most countable).
Lebesgue's decomposition theorem asserts that u f is the sum of three in-
creasing functions

or
du f = du1t + dU!c + dU!g,
where u1t is a step function, u!c is absolutely continuous (it is the primitive
of its derivative) and U;g is continuous singular (its derivative vanishes almost
everywhere with respect to the Lebesgue measure).
THEOREM 2. (Bertrandias [2])

This important result shows that the Fourier-Bohr spectrum of f is a subset

of the Wiener spectrum: S(f) ~ W(f). In particular if u f is continuous, then
i(A) = 0 for all A, but the converse is not true.
Here are some examples:

f(n) = exp(2i1l"cw2), a irrational du f (z) = dz, the Lebesgue measure,

(see [1]);
f(n) is the Thue-Morse sequence du f is purely singular,
(see Appendix 1)j
f( n) is the Rudin-Shapiro sequence : du f (z) = dz, the Lebesgue measurej
f(n) = exp 2i1l"an dul(z) = Ca , the Dirac mass at aj
f( n) = exp 2i1l"vIn du l (z) = Co.

If f is a Besicovitch almost-periodic function

f(n)~ L j(A)exp2i7fAn,
>.es(f)
336 J.-P. Allouche and M. Mendes France

then
d(1'l(x) = L I!(AWC,\.
'\ES(f)

By definition a mean-periodic sequence f is a sequence whose. spectral mea-

sure is a sum of Dirac masses. A Besicovitch almost-periodic function is there-
fore mean-periodic. The converse is false as the example f(n) = exp 27rifo
shows.
In 1958 J. Bass [1] defined the family of pseudo-random functions and se-
quences. His definition, modified by J.-P. Bertrandias [2] is that f(n) is pseudo-
random if (1'1 is continuous. This is known to be equivalent to the condition

l.e. rl (h) tends in average to 0 as h increases to infinity.

We conclude this paragraph by a diagram which illustrates the relationship
between the different classes we discussed.

almost all (±)-sequences

Besicovitch almost-periodic: (-1 )[anJ

pseudo-random: e 2i ,..an'

mean almost-periodic: e2i ,....;n

 AUTOMATIC SEQUENCES 337

4. ANALYZING THE SPECTRAL MEASURE

We are given the sequence 1= (f(n)) and we assume that the Fourier-Bohr
coefficients and the correlation exist.
Case 1. Parseval's equality holds. Then the spectral measure du l is discrete
(i.e. a sum of Dirac masses or Bragg peaks, u l is a step function):

du l = L 11(.~)128>..
>.es(J)

I is Besicovitch almost-periodic.
Case 2. Suppose S(f) = 0. A theorem of Wiener's asserts that

where u l {A} is the mass at A. Hence we have a first important result which
we already mentioned in Paragraph 3:
II
lim
N_oo
~ ~
N L.J
Irl (nW = 0,
n<N

then du l is continuous, and conversely.

The question arises to decide whether du l is absolutely continuous or purely
singular (or sum of both). A well known result of Riemann and Lebesgue shows
that if du l is absolutely continuous then

tends to zero as n increases to infinity. Therefore if

limsuplrl(n)1 > 0,
n-oo

there is a nonzero singular component to du l .

We shall illustrate these ideas in the following appendices.

5. APPENDIX I: THE SPECTRAL MEASURE OF THE THUE-

MORSE SEQUENCE

Let f = (f(n)) be the ± Thue-Morse sequence. We show that its spectral

measure is singular continuous.
338 J.-P. Allouche and M. Mendes France

We first notice that f(2n) =

f(n) and f(2n + 1) = - f(n), (1(0) = 1). We
shall compute the correlation function
1",,-
,(h) = lim -
N-oo N
~ f(n) f(n + h).
n<N

One of the difficulties is to prove that ,(h) actually exists. Obviously ,(0) = l.
Define
8(N, h) = L J(n) f(n + h)

!mi!
n<N

L J(2m) f(2m + h)

= + f(2m+ l)f(2m + 1+ h) + 0(1).

!
m<~

L f(2m)f(2m + 2k + 1)
m<!:l.
8(N,2k+1)= ~
+ m~~ J(2m + 1)f(2m + 2k + 2) + 0(1)

=- E J(m)f(m + k) - E J(m)f(m + k + 1) + 0(1).

m<~ m<~

Hence

(1) 8(N,2k+1)=-8(~,k)-8(~,k+1)+0(1).
Similarly

8(N,2k) = 8 (~, k) + 8 (~, k) + 0(1)

(2)
= 28 ( ~ ,k) + 0(1).
In our first recurrence relation choose k = 0, divide by N and let N go to
infinity. Define

,*(h) = lim sup ~ E f(n)f(n + h),

N_oo n<N
{
,*(h) = l~~~f ~ L f(n)f(n + h).
n<N
 AUTOMATIC SEQUENCES 339

Then

-y*(1) = -~ - ~-Y*(1),
{
-y*(1) = -~ - ~-Y*(1).

These two equations with two unknowns show that

-y*(1) = -y*(1) = -~.

Therefore -y(1) exists and its value is -1/3.

We now assume -y(k) exists for all k < 2K. Then equation (2) implies that
-y(2K) exists and
-y(2K) = -y(K).

Now equation (1) shows that -y(2K + 1) exists and

1
-y(2K + 1) = -2(-y(K) + -y(K + 1)).

We know there exists a measure du such that

Let us show that du is a continuous measure. It suffices to show that

1
lim N
N ..... oo
I: 1-y(kW = O.
k<N

Put

r(h) = Nlim
..... oo
N1 I: r(n)-y(n + h)
n<N
340 J.-P. Allouche and M. Mendes France

(it can easily be established that the limit exists for all h E 2). Define

T(N, h) = L '7(n)-y(n + h)
n<N
= L '7(2n)-y(2n + h) + L '7(2n + 1)-y(2n + 1 + h) + 0(1),
n<!f n<.!f
T(N,2k) = L '7(2n)-y(2n + 2k) + L '7(2n + 1)-y(2n + 1 + 2k) + 0(1)
n<!f n<!f

Therefore
1 1 1 1
r(2k) = 2 r(k) + 4 r (k) + gr(k + 1) + gr(k - 1)
311
= 4r(k) + gr(k + 1) + gr(k - 1).

For k = 0,
311
reO) = 4r(0) + gr(l) + gr( -1).
But i' is real valued, so that r is real valued hence r(-I) = r(1). Therefore

(3) reO) = [(1).

On the other hand

_!T (N2' k) _!T (N2 ' k+ 1)

T(N,2k+l)= {
2 2

- ~T (~, k) - ~T (~, k - 1) + 0(1),

hence
111
r(2k + 1) = --r(k)
2
- -r(k + 1) - -r(k -
4 4
1).
For k = 0,
reO) = -3r(1).
 AUTOMATIC SEQUENCES 341

This, compared with equality (3), implies that f(O) = 0 and this in turn implies
that d(1 is a continuous measure.
Remember that !i(.~)! :::; J(1{A} = 0 so that the Fourier-Bohr spectrum is
empty.
So far we have thus established that the spectral measure d(1 is continuous
and that the correlation, satisfies the equations

,(0) = 1
(4) { ,(2h) = ,(h)

,(2h + 1) = -~b(h) + ,(h + 1)].

limsup!,(n)! ~ ~.
n--+oo 3

This shows that d(1 has a nonzero singular component d(1sg

We shall now prove that d(1ac = O. To simplify notations put

and
'l(n) = 10 1 e 2imx d(11(X),
The following equalities hold

,(2n) = 11 e2i7r2nx d(1( x)

= 12 e2i7rntd(1 (~)

= 11 e 2imt d(1 (~) + 12 e2i7rntd(1 (~)

In the last integral put t = u + 1

,(2n) = 11 e2itrnud(1 (~) + 11 (u; 1)

e2imtd(1

= 11 e2itrnt d [(1 (~) + (1 C~ 1) ].

342 J.-P. Allouche and M. Mendes France

But ,(2n) = ,en), hence

duet) = du (~) + du (t ~ 1) ,

The left hand side represents an absolutely continuous measure and the right
hand side a purely continuous singular measure. Therefore both vanish. In
particular

Reading the above calculation backwards, we see that

and in the same fashion we can prove that

(5)

Remember that dU1 is absolutely continuous hence

n --l- 00,

so that in particular

as n --l- 00.

But ,l(2 n ) = ,1(1) hence '1(1) = O. Equation (5) then implies with n = 0
that
,1(0) = 0 = 11 dU1(X).

Therefore dU1 = O. We conclude that du is indeed a purely continuous singular

measure.
These results were known to Kakutani [4] and Mahler [5). See also [3].
 AUTOMATIC SEQUENCES 343

6. APPENDIX 2: THE SPECTRAL MEASURE OF THE PAPER-

FOLDING SEQUENCE

Let f = (f(n)) be the (±) paperfolding sequence:

f(2n) = (-It
{
f(2n + 1) = f(n).

We shall prove that the sequence is Besicovitch almost-periodic.

To do so we compute the Fourier-Bohr coefficients

ip..) = lim
N-oo
~
N
L: f(n)exp(-2i7rn~).
n<N

Define
S(N,~) = L: f(n) exp( -2i7rn~)
n<N

L: f(2m) exp( -2i7r2m~)

{ m<N/2

+ L: f(2m + 1)exp(-2i7r(2m + 1)~)

m<N/2

L: (_l)m exp( -4i7rm~)

=
1 :>. L: f(m)exp(-2i7r2m~).
m<!:i.

+e- 2i
m</f

The first sum is easily seen to be equal to

L: eXp2i7rm(~-2~)=~ X(~)+O(l),
m<N/2

where
X(~) = {I if 2~ E ~ + IZ,
o if not.
To simplify notations we put e(x) = exp2i7rx. The second sum is

e( -~)S ( ~, 2~) ,
344 J.-P. Allouche and M. Mendes France

so that
S(N, A) = e( -A)S ( ~, 2A) + ~ X(A) + 0(1).
By induction

S(N,A)=e(-A-2A)S ( 2N2 ,2 2 A) N X(A) + 2N (2A)e(-A) +0(1),

+2 2X

=N f;k-1 2i~1 X(2i A)e«(1- 2i)A) + e«l - 2k)A)S ( ~, 2k A)

+ k.O(l).
Choose k = 1 + [log N / log 2]. Then the sum S( N2- k , 2k A) is empty.

k-1 1
S(N, A) =N L 2i+1 X(2i A)e«l - 2i)A) + O(log N).
j=O

Divide by N and let N go to infinity:

j(A) = f X2~2~;)
i=O
e«l - 2i)A).

If A is irrational or rational with denominator other than a power of 2, j( A) = O.

Suppose A = 2~t1 where a and f are positive integers. Then

X(2i2a+l) =X(2a+.l)
2l 2l -)
= {I if j=f-2,
0 if not.

The Fourier-Bohr coefficient boils down to

f~(2a+l)
2l
__1_ (1_2l - 2)2a+l) f
- 21-1 e 21 ' ~ 2

=Z
.( _1)a-1 ( . 2a +
21 - 1 exp 2z1r~
1) .

Then
 AUTOMATIC SEQUENCES 345

It is now easy to see that Parseval's equality holds. Indeed

2l - 1 _1

L>. 11(,xW =L
l~2
L 22(~-1)
a=O
2l - 1 1
= L 22(l-1) =L 2l - 1 = 1 = 11/112.
l~2 l~2

This proves that the regular paperfolding sequence is Besicovitch almost-perio-

dic. Its spectral measure is

This result was first published in [6] where it is proved that all paperfolding
sequences have the same spectral measure (but the Fourier-Bohr coefficients
differ).

References

[1] Bass J., Suites uniformement denses, moyennes trigonometriques, fonc-

tions pseudo-aleatoires, Bull. Soc. math. France 87 (1959) 1-69.
[2] Bertrandias J .-P., Espaces de fonctions bornees et continues en moyenne
asymptotique d'ordre p, (These), Bull. Soc. math. France, Memoire 5
(1966) 3-106.
[3] Coquet J., Kamae T., Mendes France M., Sur la mesure spectrale de
certaines suites arithmetiques, Bull. Soc. math. France 105 (1977) 369-
384.
[4] Kakutani S., Strictly ergodic symbolic dynamical systems, in: Proceed-
ings of the 6th Berkeley symposium on mathematical statistics and prob-
ability [1970 Berkeley] vol. 2 (Berkeley, University of California Press,
1972) p. 319-326.
[5] Mahler K., On the translation properties of a simple class of arithmetical
functions, Jour. Math. and Phys. 6 (1927) 158-163.
[6] Mendes France M., van der Poorten A. J., Arithmetic and analytic prop-
erties of paperfolding sequences (dedicated to Kurt Mahler), Bull. Austr.
Math. Soc. 24 (1981) 123-13l.
[7] Meyer Y., Le spectre de Wiener, Studia Mathematica 27 (1966) 189-201.
[8] Queffelec M., this volume.
[9] Wiener N., Generalized harmonic analysis, Acta Math. 55 (1930) 117-
258.
346 J.-P. Allouche and M. Mendes France

PART V
Complexity of infinite sequences
Another approach to the order versus chaos paradigm consists of the study
of the factor complexity of sequences. A factor of a (finite or infinite) sequence
is a finite block of consecutive letters.
DEFINITION. The complexity function (or for short the complexity) of an in-
finite sequence u with values in a finite alphabet A is the function n --+ pu(n),
where Pu (n) denotes the number of factors of length n in the sequence u.
Remarks
- Let u be any infinite sequence on the (finite) alphabet A. Then obviously
'tin ~ 1, 1 ~ pu(n) ~ (#A)n. The function p~ is clearly increasing.
- If u is a "random" sequence, one expects it to be normal, i.e. every possible
block of length n occurs with frequency 1j(#A)n. In particular the complexity
of such a sequence is maximal and is equal to (#A)n. On the other end of the
spectrum it is not difficult to check that an ultimately periodic sequence has an
ultimately constant complexity, (the converse holds with a weaker hypothesis,
see below). Complexity distinguishes between normality and periodicity, and
much more.
Note however that there exist algorithmically constructed sequences which
are normal: the Champernowne sequence defined, say in base 2, by writing
down consecutively the binary expansions of the integers,

o 110 11100 101110 1111000 1001 1010

is known to be normal.
- Using the above definition of complexity one defines the topological entropy
of a sequence u as
h( u ) -- l'1m logpu(n) .
n ..... oo nlog#A

The limit always exists. Clearly 0 ~ h(u) ~ 1. If the structure of the sequence
is simple, one expects h(u) = O. On the contrary a complex structured sequence
will have maximal entropy h(u) = 1. Needless to say for an automatic sequence
h( u) =
0 and for a normal sequence h( u) = 1. For any a E [0, 1] there exist
sequences u for which h( u) = a. The notion of entropy will be discussed in
more details by V. Berthe in this volume [12].
- Many other notions of complexity exist for a sequence, a very interesting
one being the Kolmogorov-Chaitin-Solomonoff complexity which compares a
given finite or infinite sequence to the length of the smallest possible "program"
which generates it, see [21].
 AUTOMATIC SEQUENCES 347

1. STURMIAN SEQUENCES AND GENERALIZATIONS

We begin this paragraph with a theorem which shows that, if the complexity
of a sequence is "not too large" , then the sequence is periodic from some point
on, and this implies in turn that the complexity is ultimately constant.
THEOREM. [24], [25] and [15]
Let U be a sequence such that 3n ~ 1, pu(n) ~ n. Then the sequence u is
ultimately periodic, (hence Pu is ultimately constant).
In view of this theorem the minimal possible complexity among the sequences
=
which are not ultimately periodic is given by p( n) n + 1, 'rIn ~ 1. With n 1 =
one sees that, if there exist such sequences, the alphabet must have cardinality
2. Such sequences actually exist, as shown by the following result. They are
called Sturmian sequences.
THEOREM. To each Sturmian sequence (Un)n on the alphabet {O, I} corre-
sponds a unique couple (a, f3) E [0,1[2, a irrational such that

-either 'rIn, Un = L(n+l)O'+f3J - LnO'+f3J,

-or 'rIn, Un = r(n+l)O'+f31- rnO'+f31,

where LxJ is the usual integer part and rx1= - L-xJ.

A nice survey on Sturmian sequences is [11].

Remarks
- The above theorem shows that Sturmian sequences can also be constructed
by intersecting a two-dimensional lattice by a straight line with irrational slope.
This implies that they can also be generated as billiard trajectories. In par-
ticular our friend the Fibonacci sequence is a Sturmian sequence. One should
remark that very few Sturmian sequences can be generated by morphisms: the
set of morphic sequences is countable, whereas the set of Sturmian sequences
is not, since each of them is determined by a couple (a, f3). To know precisely
which Sturmian sequences are fixed point of morphisms, one can read [16].
- One can study billiard trajectories in more than two dimensions. Some
results for the complexity are known: for instance the three-dimensional irra-
tional trajectories of billiard balls have complexity n 2 + n + 1, see [8] and [9],
where interesting conjectures are given.
- One can play billiard in a triangle or in a convex polygon instead of playing
in a square. The following has been proved in [19]: consider a polygonal billard
with angles 711", ~11", ... , 9f1l", where gcd(al, a2,··· , ad, r) = 1, and start with
any angle a (with a countable number of exceptions), then the complexity is:
= =
p(n) nr+(d-l)r, for n ~ no(O'). In particular p(n) 4n+8 for a right-angle
triangle with two equal sides and p( n) = 3n + 6 for an equilateral triangle.
348 J.-P. Allouche and M. Mendes France

- Other sequences with geometric properties have been studied: sequences

with complexity 2n + 1, see [29], [10], [18] and [31], sequences with complexity
2n, see [30] and [1]. Another class is defined as follows. Let ~ be the first
difference operator: for a sequence u, the sequence ~u is defined by (~u)n =
Un+! - Un. Now let a be an irrational number. Instead of considering the
Sturmian sequence ([(n + l)a] - [na]) = (~u)n' where Un = [na], consider
~2v, where Vn = [n 2a]. The complexity of this new sequence is given by [7]:

p(n) = (n + 1)(n + 2)(n + 3).

6

2. COMPLEXITY OF AUTOMATIC SEQUENCES

2.1. An upper bound

The first result was given by Cobham [14]. It states that the complexity of
an automatic sequence is not too large.
THEOREM. [14}
If U is an automatic sequence, there exists a constant depending on U such
that:
'rIn ~ 1, pu(n) ~ en.

We list below some special results.

2.2. Complexity function of some automatic sequences

- If u is the Prouhet-Thue-Morse sequence, its complexity Pu is given by a
... rather complicated expression, see [22], [23] and [13], from which we shall
only retain that the sequence n -> Pu (n + 1) - Pu (n) is a 2-automatic sequence.
This result will be revisited below.
- If u is the (regular) paperfolding sequence, then its complexity function
satisfies 'rIn ~ 7, pu(n) = 4n. Actually this result holds for any paperfolding
sequence [2], and one can also prove that there is a kind of synchronization
between the factors of different paperfolding sequences [5]. Note that the pa-
perfolding sequences can be obtained as "Toeplitz sequences", see [3] for a
survey on Toeplitz sequences, and that the complexity of a class of Toeplitz
sequences has been studied in [20].
- For the Rudin-Shapiro sequence one has p( n) = 8n - 8, 'rIn ~ 8. This result
can be extended to two generalizations of the Rudin-Shapiro sequence, see [2]
and [6], see also [4]: let us just say here that one obtains again ultimately affine
complexities.
- We quote finally in this paragraph the paper [26], where it is shown that,
if u is an automatic sequence which satisfies some mild conditions, then its
complexity function Pu has the property that (Pu (n + 1) - Pu (n))n is also an
 AUTOMATIC SEQUENCES 349

automatic sequence. The simple fact that this last sequence takes only finitely
many values is by no means trivial.

2.3. The case of non-constant length morphisms

After the Fibonacci sequence which is Sturmian as we saw above, one can
ask for the complexity function of the fixed points of non-constant length mor-
phisms. The problem is more difficult, reflecting the fact that these sequences
are more "complicated" than the automatic sequences.
It has been shown that the complexity of a sequence u fixed point of a
non-constant length morphism satisfies p(n) = 0(n 2 ). In the case where the
morphism is "primitive" (i.e. if there exists a power of the morphism such
that the image of every single letter contains all the letters), this bound can
be lowered to p(n) = O(n). The most precise result can be found in [27] and
[28] where it is proven that the complexity of any fixed point of a morphism
has one of the following orders of magnitude:

3A,B > 0,3/ E {1,n,nloglogn,nlogn,n 2 }, '<In ~ 1, A/(n) ~p(n) ~ B/(n).

Kolakoski's sequence, which has already been discussed in Part I, is probably

not generated by a non-constant length morphism. Its complexity is not known.
It has been conjectured to be of the order of n q , where q = l~:~' see [17].

References

[1] Alessandri P., Codage des rotations, Memoire de D. E. A. (Universite

Claude Bernard, Lyon I, 1993).
[2] Allouche J .-P., The number of factors in a paperfolding sequence, Bull.
Austr. Math. Soc. 46 (1992) 23-32.
[3] Allouche J .-P., Bacher R., Toeplitz sequences, paperfolding, Hanoi towers
and progression-free sequences of integers, Enseign. Math. 38 (1992)
315-327.
[4] Allouche J .-P., Bousquet-Melou M., Facteurs des suites de Rudin-Shapiro
generalisees, Bull. Belg. Math. Soc. 1 (1994) 145-164.
[5] Allouche J .-P., Bousquet-Melou M., Canonical positions for the factors in
the paperfolding sequences, Theoret. Comput. Sci. 129 (1994) 263-278.
[6] Allouche J .-P., Shallit J., Complexite des suites de Rudin-Shapiro genera-
lisees, Journal de Theorie des Nombres de Bordeaux 5 (1993) 283-302.
[7] Arnoux P., Mauduit C., Meigniez G., Sur la complexite de certaines
suites arithmetiques, in preparation.
[8] Arnoux P., Mauduit C., Shiokawa I., Tamura J. I., Complexity of se-
quences defined by billiards in the cube, Bull. Soc. math. France 122
(1994) 1-12.
350 J.-P. Allouche and M. Mendes France

[9] Arnoux P., Mauduit C., Shiokawa I., Tamura J. I., Rauzy's conjecture
on the cubic billiards, Tokyo J. Math. 17 (1994) 211-218.
[10] Arnoux P., Rauzy G., Representation geometrique de suites de complex-
ite 2n + 1, Bull. Soc. math. France 119 (1991) 199-215.
[11] Berstel J., Seebold P., Mots de Sturm - un survol, Preprmt (1994).
[12] Berthe V., this volume.
[13] BrIek S., Enumeration of factors in the Thue-Morse word, Discrete Appl.
Math. 24 (1989) 83-96.
[14] Cobham A., Uniform tag sequences, Math. Systems Theory 6 (1972)
164-192.
[15] Coven E. M., Hedlund G. A., Sequences with minimal block growth,
Math. Systems Theory 7 (1973) 138-153.
[16] Crisp D., Moran W., Pollington A., Shiue P., Substitution invariant cut-
ting sequences, Journal de Theorie des Nombres de Bordeaux 5 (1993)
123-137.
[17] Dekking F. M., On the structure of self-generating sequences, Seminaire
de Theorie des Nombres de Bordeaux, Expose 31 (1980-1981) 31-01-31-
06.
[18] Ferenczi S., Les transformations de Chacon: combinatoire, structure geo-
metrique, lien avec les systemes de complexite 2n + 1, Bull. Soc. math.
France (1995) to appear.
[19] Hubert P., Complexite des suites definies par des billards rationnels,
Bull. Soc. math. France (1995) to appear.
[20] Koskas M., Complexity functions for some Toeplitz sequences, Theoret.
Comput. Sci. (1995) to appear.
[21] Li M., Vitanyi P., Kolmogorov complexity and its applications, in: Hand-
book of Theoretical Computer Science, Volume A: Algorithms and Com-
plexity, J. van Leeuwen Ed (MIT Press, 1990) p. 187-254.
[22] de Luca A., Varricchio S., On the factors of the Thue-Morse word on
three symbols, Information Processing Letters 27 (1988) 281-285.
[23] de Luca A., Varricchio S., Some combinatorial properties of the Thue-
Morse sequence, Theoret. Comput. Sci. 63 (1989) 333-348.
[24] Morse M., Hedlund G. A., Symbolic Dynamics, Amer. J. Math. 60 (1938)
815-866.
[25] Morse M., Hedlund G. A., Symbolic Dynamics II. Sturmian trajectories,
Amer. J. Math. 62 (1940) 1-42.
[26] Mosse M., Notions de reconnaissabilite pour les substitutions et com-
plexite des suites automatiques, Prepublication du LMD 93-21 (1993).
[27] Pansiot J .-J., Bornes inferieures sur la complexite des facteurs des mots
infinis engendres par morphismes iteres, Lecture Notes in Computer Sci-
ence 166 (1984) 230-240.
[28] Pansiot J .-J., Complexite des facteurs des mots infinis engendres par
morphismes iteres, Lecture Notes in Computer Science 172 (1984) 380-
389.
 AUTOMATIC SEQUENCES 351

[29] Rauzy G., Suites it termes dans un alphabet fini, Seminaire de TMorie
des Nombres de Bordeaux, Expose 25 (1982-1983) 25-01-25-16.
[30] Rote G., Sequences with subword complexity 2n, J. Number Theory 46
(1994) 196-213.
[31] Santini M.-L., Echanges de trois intervalles (These, Universite Aix-Mar-
seille II, 1994).
352 J.-P. Allouche and M. Mendes France

PART VI
Opacity of an automaton
1. AUTOMATA REVISITED

In this part all the automata we consider are 2-automata. The arrows are
named (+) and (-).
An automaton A is determined by a finite set S (set of states), an initial
state A E S and a map r : {-, +} x S -> S represented by (±) arrows which
join two states:
A = (S, A, r).
=
Given an infinite sequence c (cn), Cn =
± 1, it acts on A in the following way.
co puts the automaton in state coCA). Then C1 sends coCA) onto c1(co(A)),
etc ... The sequence of states

is an infinite sequence on the alphabet S denoted Ac.

Vtle now define the output function

which maps Ac on cp(Ac) coordinate-wise.

We propose to measure the discrepancy between the input sequence c and
the output sequence cp(Ac).
Given two real bounded sequences 8' = (8~) and 8" = (8::) we define their
distance by
1/2

118' - 8"11 = lim sup

N--+oo
(
~
N
L
n<N
18~ - 8~12 )

This norm was already discussed in Part IV.

IIcp(Ac)-cll is to be thought as the distorsion of the sequence c when mapped
through (A, cp). We minimize the distorsion by choosing cp as "best as we can":

The inf is to be taken over all maps cp: S ->~. The opacity of A is by
definition

(1) w(A) = eEUP

sup inf IIcp(Ac) -
cP
.::11
 AUTOMATIC SEQUENCES 353

where the sup is over the set of all (±) sequences € which are periodic from
some point on (U P stands for Ultimately Periodic).
The UP restriction is a technical requirement which may perhaps not be
necessary. At the time of writing, Nathalie Loraud [1] is studying the pertinence
of the restriction. She proposes to modify the definition of the opacity:

1/2

w(A) = sup lim sup

£ N ..... oo
(
N1 inf L
'P n<N
1<p(Ae) - el 2 )

(notice the permutation of the two operations lim sup and inf). With this
N 'P
definition she can show that sup and sup coincide. She can also prove that
£EUP £
w(A) =w(A).
The question remains: is it true that

w(A) = sup inf 1I<p(Ae) - ell?

£ 'P

We shall not discuss the matter and we maintain our original definition (1).
Let us describe some simple properties of the opacity. Trivially w(A) ~ o.
We also remark
infll<p(Ae) - ell ~ 11- ell = 1
'P

since one can choose <p to be the constant 0 function. Therefore

o ::; w(A) ::; 1.

An automaton A for which w(A) = 0 is said to be transparent. If w(A) =1
then A is completely opaque.

Example 1. The identity automaton

+~ +~-
Consider for example the input sequence

e: +--++
The automaton sends it on

Ae: ABBAAB ...

354 J.-P. Allouche and M. Mendes France

Apart from the symbols, both sequences are identical. Define

<peA) = +1, <pCB) = -1.

Then for all e, <p(Ae) = e hence w(A) = O.
Example 2. The constant automaton

Put <peA) = a. Then

W(A)2 = sup inflimsup ~

eEUP a N
L la - enl
n<N
2.

In the lim sup choose en = (_l)n:

lim sup ~ L la - (_1)nI2 = lim sup ~ L (a 2 + 1- 2a(-lt)

N n<N N n<N

= a 2 + 1.
Trivially
inf(a 2 + 1) = 1,
a

hence
w(A) ~ 1.
But w(A) ~ 1 therefore w(A) = 1.
The output sequence gives no information on the input sequence, and it is
then not surprising to discover that this automaton is completely opaque.

Example 3. The Thue-Morse automaton

+()D. ~+
 AUTOMATIC SEQUENCES 355

= =
Put SO(A) a, SO(B) b. If the input sequence is C = (cn) then it can be easily
seen that the output sequence is
1 1
tin = 2(a + b) + 2(a - b)7rn,
where
= COC1·· ·Cn·
7rn

Let us choose as above Cn = (_I)n. Then

liti - cl1 2 = lim sup - 1L (1 (a + b) + 1-(a - b)7rn - )2
-2 Cn
N N n< N 2
Observing that

lim ~ Lcn = lim ~ L 7r = lim ~ LC n n 7r n = 0,

we easily see that

Then

hence
w(A) = l.
This automaton is also completely opaque though it is possible to reconstruct
C from A g •

We shall see below how to compute w(A) in many cases; we leave it to the
reader to show that w(A) can take all values of the form vp where p is any
rational number in the interval (0,1).
2. COMPUTING THE OPACITY

An automaton is said to be strongly connected if from any two states, there

exists at least one path of arrows joining the first state to the second.
356 J.-P. Allouche and M. Mendes France

The above automaton is not strongly connected since no arrows go to A.

An automaton is homogeneous if each state has exactly two incident arrows.
This is not the case in the preceding automaton. Automata in examples 1, 2,
3 are both strongly connected and homogeneous.
We now define strong states on a path. Let P be a closed path on an
automaton. Its length l(P) is the number of states it contains, counted with
multiplicity.
A state B on the path P is a strong state if it is attained twice at least, once
through a (+) arrow and once through a (-) arrows on P. v( P) counts the
number of strong states on P.
THEOREM. The opacity of a strongly connected homogeneous automaton A is
given by the formula

w(A) = sup
p

Given any rational number p in the interval (0,1) there exist a strongly con-
nected homogeneous automaton A such that w(A) = .,[p.
We shall not give the proof of this result (see [2]). It can be shown that the
result extends to a larger class of automata, namely those for which the states
are of three kinds:
a) exactly two incident arrows, one (+) and one (-),
b) any number of (+) arrows and no (-) arrows,
c) any number of ( -) arrows and no (+) arrows.
Let us illustrate this theorem by computing the opacity of the following
automaton

The shortest closed path P which contains both the + and - arrows incident
to Cis

for which l(P) = 4 and v(P) = 1 thus

f2;;1
w(A) = VT = ..j2'
 AUTOMATIC SEQUENCES 357

We shall apply these results to the Ising chain in Part VII.

References

[1] Loraud N., Contribution it l'etude de l'opacite d'un automate, Work in

progress.
[2] Mendes France M., Opacity of an automaton. Application to the inho-
mogeneous Ising chain, Comm. Math. Phys. 139 (1991) 341-352.
358 J.-P. Allouclle and M. Mendes France

PART VII
The Ising autOlnaton
A large part of this chapter (Paragraphs 1 and 2) is due to B. Derrida who
introduced us to the idea of induced field and who showed us how to compute
it. See [6] and the forthcoming paper [8].

1. THE INHOMOGENEOUS ISING CHAIN

An Ising chain consists of N + 1 sites in a row occupied by N + 1 particles

with spin ±l. At site q E {O, 1 ... , N} the spin is denoted

rrq E {-I, +1}.

A configuration is a finite sequence of spins

Let
e = (.~o,el'··· ,eN-I) E {-l,+l}N.
be a given sequence which could be thought of as the distribution of two sub-
stances in an alloy. The Hamiltonian or energy of a configuration rr is defined
by
N-l N
1f t (rr) = -J L eqG"qG"q+1 - H L G"q,
q=O q=O

where J > 0 is a given parameter known as the coupling constant or binding

energy and where H is a real parameter called the external field.
One of the main problems is to find the equilibrium state (or fundamental
state) of the chain i.e. to discover the configuration G" which minimizes 1f«G").
Let T > 0 be the temperature; put (3 = l/T. The fundamental axioms of
Statistical Mechanics show that the probability of configuration G" at tempera-
ture T is

where Z({3, N), the partition function, is defined by

Z({3, N) =

Suppose there exist L fundamental states which minimize 1f< (G"). Then it is
easy to see that as T vanishes,

1/ L if G" is fundamental,
JiZ~PT(rr) ={ 0 if not.
 AUTOMATIC SEQUENCES 359

If L = 1, then, at vanishing temperature, the chain is at equilibrium with

probability 1. Order prevails.
At high temperature, (T / +00, (J '\. 0),

lim PT(a)
T/+oo
=2 :+1'
All configurations have equal probability: the system IS III a state of total
disorder.

2. THE INDUCED FIELD

We shall now study the sign of the spin aN of the particle at site N (the
very last one) at temperature T = O. Our strategy is based on the simple fact

lim [Z«(J, N)]I/(3

(3---++=
= exp( -'Jig (0:)),
where 0: is one of the equilibrium configurations.
The discussion which we reproduce here in this section is well known to
physicists. See for example [6].
Define the two conditional partition functions

Z+«(J, N) = L exp( -(J'Jig(a))

O'E{±}N
O'N=+1

Z-«(J,N) = L exp(-(J1ig(a)).
O'E{±}N
O'N=-l

The probability that aN = +1 at temperature Tis

Z+«(J, N)/Z«(J, N).

Similarly
Z-«(J, N)/Z«(J, N)
is the probability that aN = -1 at T.
If the ratio

is larger that 1 then the event aN = +1 occurs with probability larger that 1/2;
if the ratio is less than 1 then aN = -1 prevails. We are thus led to compute
the ratio Z+ / Z- .
360 J.-P. Allouche and M. Mendes France

Define the vector

Z+({3, N))
V(N) = ( Z-({3,N) .

Then
V(N) = M(€N_dV(N - 1),
where the "transfer matrix" is defined by

To simplify notations we have put

Notice that T'\. 0 ¢=> z / +00.
0' = 2: and z = e(jJ.
By induction

V(N) = M(€N-t}M(€N-2) ... M(€o)V(O),

where
eXP(-H))
V(O) = ( exp(+H) .

The product of the transfer matrices is a 2 x 2 matrix the entries of which are
"polynomials" in z. As T decreases to 0, z increases to infinity. Only highest
degrees are relevant:

The symbol", signifies that the first (resp. second) coordinate has order za q
(resp. zb q ).
Similarly

therefore

(1) { .,+> : ~; + max{a, + e,. b. - e,}

bq+1 - -2"0' + max{a q - €q, bq + €q}.

On the other hand

 AUTOMATIC SEQUENCES 361

tends to +00 or 0 as z tends to +00 according to the sign of

Hence the sign of ON (notice the subscript!) imposes the value of the spin UN
at equilibrium:

If ON = 0 we agree to say that the spin UN is not determined; it can be chosen

at will +1 or -1. (This is not entirely correct. If aN = bN one should then
look at the coefficients of zaN in both the components of V(N). The larger
one should then impose the value of UN. This case is so singular that we shall
adopt the above "convention".)
The quantity ON is called the induced field at site N.
The two equalities (1) imply

Changing the sign of a = 2H / J only changes the sign of the sequence (Oq)
so without loss of generality we shall assume from now on that a ~ 0, (H ~ 0).
The above induction formula has a special feature namely that if a > 0, and
if q > 4/a, then for almost all c, Oq is independent of 00. If on the other hand
a = 0 we loose no information on sgn Oq by fixing 00 arbitrarily.

It is convenient to choose 00 = a + 2. We are now left to study the sequence

(on) defined by

(2)

As an easy exercise, the reader may verify that for a = 0, (H = 0), these
equations reduce to

and for a ~ 4, (H ~ 2J),

On+1 = a + 2cn.
How are we to int.erpret the sequence 00, 01, 02, ... , On, ... ? The sign of
On is the value of the spin Un at equilibrium under the influence of the sites
n-1, n-2, ... ignoringsitesn+1, n+2, ... InparticularuN=sgn(oN).
But it is not true that UN-1 = sgn(oN_1) since the site N - 1 is influenced by
the sites N - 2, N - 3, ... and also by the site N. T. Kamae pointed out that
finding the values of Un for all n ~ N is now actually possible. The details will
appear in a forthcoming paper [8].
362 J.-P. Allouche and M. Mendes France

3. THE ISING AUTOMATON

The rest of this chapter is devoted to the study of the sequence 8 = (8 n )

which we denote 8cx =
(8::) or 8CX(e) =
(8::(e)) if we wish to emphasize the
dependence of 8 on a or e = (en).

We first observe that for all n

More precisely, if a = 0 then

8n E {-2, +2}

and if a> 0

8n E {( i + 1) a - 2 I i = 0, 1, 2, .. . , [4/ a]} U{2 - (i - 1) a I i = OJ 1, 2, .. . , [4/ a ]} .

More is true: a being fixed, the map

is described by an automaton Acx and an output function 'Pcx which we discuss

now. We agree to call the couple (A,» 'Pcx) the Ising automaton, so we have a
family of Ising automata depending on the external field a = 2H / J.

Case 1. a = O. The Ising automaton is the Thue-Morse automaton with

output function 'Po(A) = 2, 'Po(B) = -2.

Case 2. a> O. Put m = [4/aJ. The Ising automaton has 2(m + 1) states.
 AUTOMATIC SEQUENCES 363

= = =
)Oa(A) 0'+2, )Oa(B i ) iO' - 2, i 1,2"" ,m+ 1,
)Oa(Cj) = 2 - (j -1)0', j = 1,2,··· ,m.

Remark
As a tends to 0 the number of states of Aa goes to infinity. On the other
hand, Ao has only two states so there seems to be some discontinuity as a van-
ishes. The "paradox" can be solved in the following way. Let n be somewhere
between 0 and [4/0'], 0 ~ n ~ [4/0']. In the vicinity of "n" the automaton Aa
looks like
364 J.-P. Allouche and M. Mendes France

and
CPa(Bn) = -2 + nO'
{
CPa(Cn) = 2 - (n - 1)0'.
When a vanishes, CPa tends to
cpo(Bn) = -2 for all n
{
cpo(Cn ) = 2.
This automaton which contains infinitely many states collapses and becomes
indiscernable from Ao!

4. AN ERGODIC PROPERTY
THEOREM 1. For all a and for almost all sequences c
1 N-l
lim N "(Aac)n = o'.
N-oo L...J
n=O

The proof we propose here is shorter than in [9]. Let T be the shift operator
on {-I, +1}N: T( cn) = (cn+ d. Let A be the initial state of the automaton
Aa. Define the operator T on the space {-I, + I}N X S:
T«c), B) = (T(c), coB).
Here S is the set of states of Aa. Then
r(c, A) = (Tnc, Cn-1Cn-2" ·coA).
The operator T is ergodic with respect to the canonical measure on the space
{-I,+I}N x S. Define
F: {-I,+I}N x S -d~
(6, B) I--> CPa(B).
The ergodic theorem asserts that, for almost all c E {-I, +1}N,

J~oo ~ L F(r(c, A)) = JCPa(B) dB = #~ Bes

n<N
L CPa(B).
Therefore
lim N1
N_oo
L CPa(cn-l cn-2 ... coA) = a, QED.
n<N

Remark
At the end of the proof we observed that, for almost all c,
. 1 N-l 1
J~ N L (AaC:)n = #S Bes
n=O
L CPa(B).
The equality shows that the dynamical average coincides with the spatial av-
erage.
 AUTOMATIC SEQUENCES 365

THEOREM 2. For all 0: > 0 and all sequences €

1 N-1 1
liminf N
N->oo
2::)A.€)n ~ -20:·
n=O

Proof. Recall

and

Hence

QED.

The following theorem strengthens Theorem 2.

THEOREM 3. Let 0: > 0 be given and let {3, {3' be two real numbers such that

max {~, 0: - 2 } ~ {3 ~ {3' ~ 0: + 2.

Then there exists a sequence € E {-1, + 1}N such that

1 N-1
liminfN "(Aa€)n = {3,
N->oo L..J
n=O
N-1

lim sup ~ L(Aa€)n =(3'.

N->oo n=O

Proof. If 0: ~ 4 then
6n+1 = 0: + 2€n
Then the result is easily established.
Let us suppose 0 < 0: < 4. Consider a closed path P(p, q) on the automaton
Aa which loops p times around A, then joins B 1, loops q times around B1 and
B2 and finally goes back to A. The average of 6n on the closed path is

p(o: + 2) + qo: + 0(1)

p+ 2q + 0(1)
366 J.-P. Allouche and M. Mendes France

Let PIl and qn be two infinite sequences of integers and define

1 = limsupPn/qn.
n-+oo

Clearly 0 ~ ~ ~ 1 ~ +00.
Consider an infinite path P which consists in the union of P(P1,q1), P(P2,q2),
... , P(Pn, qn), .. '. P determines a sequences € such that

N-1 -
.
I1m 1 '"
sup N ~
i:
un E
()
= (0 + 2)A + 0
N-+oo n=O A+ 2
N-l ()
. f -.!..
· III
I1m N
'"
~ un
i: (
€
) _
-
0 +\ 2 ~2+ 0
N-+oo n=O ~ +
Choose ~ and 1 such that

(0+2)1+0 =/3'
A+2
(0+2)~+0 =/3 QED.
~+2 '

Remark
By choosing ~ = 0 one obtains a second proof of Theorem 2.

5. OPACITY OF THE ISING AUTOMATON [9]

The Ising automaton is not homogeneous but the theorem of Part VI can
be applied, (see comment following the theorem): it is a simple matter to
maximize "j2v(P)/f.(P).
THEOREM 4. The opacity of AD is 1. If 0> 0, the opacity of Aa is

~'-1
--,
m'
where m' = max{l, [4/o]}.

Many of the results in this part have already been published in different
journals by several authors. We should point out in particular the papers of
Allouche, Mendes France [1], Derrida [6], Mendes France [9]. In [1] we show
that if € is an automatic sequence then the induced field is also automatic.
This result has been generalized by Dekking [5] to sequences which are literal
images of fixed points of any morphism, (not necessarily of constant length).
Theorems 2 and 3 have not appeared in print.
 AUTOMATIC SEQUENCES 367

The general theory of the Ising model originates with Ising [7]. The literature
concerning the Ising model is so vaste that it would be too long to list all the
books and articles. We single out the work of Baxter [2], Biggs [3], Cipra [4],
Thompson [10].

References

[1] Allouche J.-P., Mendes France M., Quasicrystal Ising chain and au-
tomata theory, J. Stat. Phys. 42 (1986) 809-82l.
[2] Baxter R. J., Exactly solved models in statistical mechanics (Academic
Press, 1982).
[3] Biggs N. L., Interaction models, London Mathematical Society (Lecture
Note Series, 30, Cambridge University Press, 1977).
[4] Cipra B. A., An introduction to the Ising model, Amer. Math. Monthly
94 (1987) 937-959.
[5] Dekking F. M., Iteration of a map by an automaton, Discrete Math. 126
(1994) 81-86.
[6] Derrida B., Products of random matrices and one dimensional disor-
dered systems in: Nonlinear Equations in Field Theory (Lectures Notes
in Physics, 226, Springer-Verlag, 1985).
[7] Ising E., Beitrag zur Theorie der Ferromagnetismus, Z. Physik 31 (1925)
253-258.
[8] Kamae T., Mendes France M., A continuous family of automata: the
Ising automata, Submitted (1994).
[9] Mendes France M., Opacity of an automaton. Application to the inho-
mogeneous Ising chain, Comm. Math. Phys. 139 (1991) 341-352.
[10] Thompson C. J., Mathematical statistical mechanics (Princeton Univer-
sity Press, 1972).
 COURSE 12

Spectral study of automatic

and substitutive sequences
Martine Queffelec

Department of Mathematics
Universite des Sciences et Technologies de Lille
F-59655 Villeneuve d'Asq Cedex, France

A natural question about a sequence taking its values in a finite alphabet,

is the following: how far from a random sequence is it and which appropriate
quantitative analysis can we imagine?
-A combinatorial analysis consists in computing the complexity function [4],
or the block frequency [6).
-Another point of view, studied first by physicists, lies in the description of
the spectrum of the Schrodinger operator, whose potential is the given sequence
[62).
-A third direction is a statistical one, that we develop below: the frequency
of occurrences of two given letters a and b whose distance is fixed can be
computed and used to measure the disorder of the sequence; we deduce a
spectral classification of sequences, in particular of automatic and substitutive
sequences.

INTRODUCTION

Denote by U = {un} a bounded sequence of complex numbers.

Physicists consider the Fourier transform of u, (indeed of the finite sequence
{Un} with n < N) which is the sum

SN(t) =L une int

n<N
and the normalized energy
370 M. Quetfelec

For mathematicians, the N first terms of a sequence are not significant with
respect to spectral properties, and they rather study, provided it exists, the
limit

where the notation weak - lim is used for the limit of the Fourier coefficients
of the trigonometric polynomial.
For "good" sequences, the limit exists and defines a well-known mathemati-
cal object: the Fourier transform of a positive measure on l' (see the definition
below). In this case, we say that the sequence {-hISN(tW} converges weakly
to a positive measure on T. The spectral properties of the sequence are thus
the spectral properties of this measure. When the sequence is the fixed point of
some primitive substitution, it is convenient to consider its closed shift-invariant
orbit, which is a dynamical system, and thus make use of ergodic techniques.
We begin by recalling elementary facts on Fourier transforms of positive
measures on 1', and then, we restrict our attention to automatic and
substitutive sequences.

1. MEASURES ON T
1.1. Basic definitions and notations.

1.1.1. T denotes as usual, the set of complex numbers of modulus 1, identified

with IR/21l'.lZ or [0, 21l').
We restrict our attention to positive, bounded measures a on T i.e. measures
satisfying
0::; a(B) ::; a(T) < +00
for every Borel set Bin T (see [62,681 for definition). M(T) is the collection of
such measures. The reference for this paragraph is [321.
1.1.2. If a E M(1'), the Fourier transform of a is the bounded sequence of
complex numbers {o-( n )}nEZI', where

o-(n) = i etntda(t) n E .lZ

We see easily that

100(n)1 ::; 110'11 = 0'(1')
0-( -n) = o-(n)
A very useful characterization of M(T) is given by Bochner's theorem:
1.1.3. Result: a E M(T) if and only if {o-(n)}nEZI' is a positive definite
sequence i.e.
L
t)tko-(j - k) ~ 0
l::;),k::;m
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 371

for any complex numbers tl ... t m .

Observe that the necessary condition is evident, since the sum

L t]tk8-(j - k)
l~],k~m

is nothing but

1.2. Discrete and continuous measures

1.2.1. If Xo E 1', the Dirac mass on Xo is denoted by bxQ •

We shall say that a is a discrete measure on T and write a E Md(1') if a is
a positive combination of Dirac masses:

an 2: 0, I:a n < +00, Xn E l' .

Question : How to identify a discrete measure?

The answer is given by the following.

1.2.2. Result: a E Md(1') if and only if {8-(n)}nELZ' is an almost-periodic

sequence, i.e., for all c > 0, there exists an infinite set Eo of integers
kl < k2 < ... with bounded gaps such that

sup 18-(n + k) - 8-(n)1 ~ c

n

for k in Eo.

1.2.3. On the other hand, we shall say that a is a continuous measure on l'
and write a E Mc(1') if a{x} = 0 for all x E 'f.
Question : How to identify a continuous measure ?
The answer is given by Wiener's criterion.[32]

1.2.4. Result: a E Mc(1') if and only if

N
lim
N--+oo 2 + 1 L 18-(n)12 = 0
N1
-N

Actually, more generally for a E M(1')

1 N
lim 2N ~18-(nW = ~(a{x})2
N--+oo +1 ~ ~
-N xET
372 M. Queffelec

Indeed, (T{x} is 0 for all but a countable number of values of x.

1.2.5. Lebesgue's decomposition theorem

This theorem, and the following 1.3.7, are very important in measure theory,
and valid in more general situations. They can be found in [58].
Every (T E M(T) can be uniquely decomposed as a sum

(T = (Te + (Tpp

where (Te E Me(T) and (Tpp E Md(T).

For a positive bounded measure /1 on JR, this result can be expressed in
terms of the positive non decreasing function F, defined by

F(x) = /1(] - 00, x])

and called the distribution function of /1, which characterizes /1. It asserts that
F can be uniquely decomposed as a sum Fe + Fpp where Fe is continuous and
Fpp a step function [57].

1.3. Singularity and absolute continuity

1.3.1. From Wiener's result, (T E Me(T) if

lim &(n) = o.
Inl ...... oo

But the converse is false.

We denote by Mo(T) those measures (T such that &(n) tends to 0 when Inl
goes to infinity (Rajchman's measures).
We shall see later that Mo(T) is a proper subset of McCr).

1.3.2. Absolutely continuous measures

(T is said to be absolutely continuous (with respect to the Lebesgue measure
m on T) and we write (T < < m, if (T(B) = 0 for every Borel set B of 0 Lebesgue
measure.
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 373

1.3.3. By the Radon-Nikodym property [58)

0" < < m if and only if 0" = 1 . m where 1 is an integrable non-negative
function (f E L1 (1'),2: 0).

1.3.4. If 0" « m, then 0" E Mo(l') : this is the Riemann-Lebesgue lemma

[32)
But again, as we shall see, L1(T) (or the set of absolutely continuous
measures) is a proper subset of Mo(l').

1.3.5. Analogously, we- define for 0", /-L E M(I')

O"«/-L ifandonlyif O"=I'/-L where IEL 1 (/-L) and 12:0

and we say "0" and /-L are equivalent"

when both 0" < < /-L and /-L < < 0" are satisfied.

1.3.6. Singularity
0" and /-L E M(I') are mutually singular and we write 0" ..1 /-L if and only if 0"
and /-L have disjoint Borel supports.
0" is singular if, by definition, 0" ..1 m.
Singularity of measures cannot be read so easily on the Fourier transform
as properties such as continuity of measures. We shall use for this purpose
techniques from ergodic theory.

1.3.7. More about decomposition

Every 0" E M(l') can be uniquely decomposed as a sum

where O"pp E Md('I'), O"ac « m, O"cs continuous singular.

If /-L is a bounded positive measure on 1R, this result can be read on the
distribution function F defined in 1.2.5, in the following way: a positive
continuous nondecreasing function F can be uniquely decomposed to a sum
Fac + Fcs, where Fac is absolutely continuous, and Fcs continuous, differentiable
m-almost everywhere with F;s = 0 m-almost everywhere (Fcs singular).
Note that the existence of a non-constant continuous singular function (for
example the Cantor-Lebesgue singular function, called the "devil's staircase")
is somewhat surprising, whence the word! [57,58)

1.4. Constructive examples

1.4.1. Classical Riesz product p

Let {aJ}J?:o a sequence of real numbers, laJI ~ 1. P N is the non-negative
valued, trigonometric polynomial
N-1
PN(t) = II (1 + aJ cos3Jt)
J=O
374 M. Queffelec

whose expansion is
PN(t) = ~ PN(n)e tnt
Inl<3 N
Every n admits a unique decomposition in base 3.

n = ~ cJ J3 with cJ E {O, ±1}

J<N

if Inl < 3N so that, PN(n) can be computed

PN(n) = II (~ )1<,1 if n =L cJ 3J
J<N J<N

For each n, limN ..... oo PN(n) exists.

This means that the sequence of measures, absolutely continuous measures
with respect to the Lebesgue measure, (PN . m), converges weakly to a
probability measure p E M(T) with Fourier coefficients

p(O)=1
p(n) = II(~ )1<,1 if n =L cJ 3J

P is called "Riesz product" [66].

Result:
1. p is continuous
2. p E Mo(T) ¢:} lim J -+ oo aJ = 0
3. P J.. m ¢:} 2: laJ I2 = +00
The classical Riesz product provides thus an example of a measure in Mc(T),
which is not in Mo(1') (by taking some constant sequence aJ = r, 0 < r ~ 1),
or an example of singular measure in Mo(T) (by choosing aJ = k).
This establishes the assertion that L 1 (m), Mo(T) and Mc(T) are all different.
For constant sequence {aJ },the result above can be deduced from ergodic
theory.
1.4.2. Special Riesz products.
In the spectral study of substitutions will appear some generalized Riesz
products of the following form

p = N-+oo
lim PN · m

with
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 375

The problem is to decide the nature of such a measure. For example,

1
II 2j1+e,zntjZ =
00 00

II(1+cos2nt)
n=O n=O

is nothing but Do while

1
II 2j1 - II (1- cos2nt)
00 00

e,Z"tjZ =
n=O n=O

is a continuous and singular measure p whose Fourier coefficients satisfy the

relations
(i) p(O) = 1
(ii) p(n) = p( -n)
(iii) p(2n) = p(n)
(iv) p(2n + 1) = -~(p(n) + p(n + 1))
Note that, from (iii), p cannot be in Mo('f), a fortiori p is not absolutely
continuous [46,52]. This measure will appear later as the correlation measure
of the Thue-Morse sequence.

1.4.3. Bernoulli measures on JR and Pisot numbers.

A measure is said to be pure when it is (purely) discrete, or (purely)
continuous, more precisely, if its Lebesgue's decomposition reduces to a single
part.
We quote the Bernoulli measures as interesting examples for which discussion
about decomposition is not far from being complete [45].
Consider a real number () > 1, JLIJ = ~(Oo +OIJ-n) where, as we recall, Ox is the
Dirac mass concentrated at x, and, using the star notation for the convolution,

* * JLz * ... * JLn * ...

00
JLIJ = JLn = JLl
n=l

JLIJ has a finite support contained in [0, IJ~l]

-For () = 2, JLIJ is the Lebesgue measure on [0,1].
-For () > 2, JLIJ is continuous and singular.
-There exist algebraic () E]I, 2[ such that JLIJ is absolutely continuous [23].
-JLIJ E Mo(JR) if and only if () is not a Pisot number.
A Pisot number () is a real algebraic integer () > 1 (= root of an equation
()k+ ak_l()k-l + ... + ao = 0, a J E LZ) whose conjugates (= the other roots of
the equation) belong to the unit disc {jzj < I}.
Pisot numbers enjoy a striking property in Diophantine approximation
theory. Roughly speaking, if () is a Pisot number, ()k is close to an integer,
for large k. More precisely, jj()kjj = O(rk) for a r in ]0, 1[ where jjxjj is the
distance from the real x to the nearest integer. Now, let () be a real number
greater than 1, and A > O. If the series
376 M. Queffelec

then () is a Pisot number. This is used to get the last assertion above [45].
The classical triadic Cantor measure can be viewed as a special Bernoulli
measure

It is a singular measure, supported by the triadic Cantor set ( Details can be

found in [62] ).
For complements on this study, see [10,23,27,35,50].

2. TOOLS FROM ERGODIC THEORY

Ergodic theory is the study of transformations from the point of view of

dynamical properties connected with asymptotic behavior.
We shall be concerned with ergodic theory in two situations:
-depth study of invariant measures in this paragraph (definition below)
-dynamical studies of sequences in section 5.
Note that ergodic theory will be helpful to solve measure theoretical
problems, when we rather could expect the reciprocal!

2.1. For mutual singularity of measures

2.1.1. We begin with classical definitions in ergodic theory. The references for
these preliminaries is [5,49].
Let X be a set and A a a-algebra on X (that is a collection of subsets of X
containing X, closed for countable union and by taking the complement).
2.1.1.1. A transformation T on X is a measurable mapping from X to X, i.e.
if A is any element of A, T- 1 A also belongs to A, where

T-1A = {x E X,T(x) E A}
the pre-image of the set A.
Let IL be a probability measure on (X, A). The sets of A are measurable.
The following property is very important and arises naturally:
2.1.1.2. The measure IL on (X, A) is said to be T-invariant (or T preserves the
measure IL), if
for AEA
For example the Lebesgue measure m on T is Ra- invariant where Ra is the
rotation by O! and O! any element of T. (See example below).
Often one is given X and T, but not necessarily the T-invariant probability.
Among all the possible T-invariant probabilities, the most interesting are the
T-ergodic ones.
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 377

2.1.1.3. First we recall Birkhoff's ergodic theorem

This theorem answers a physical question: X may be the collection of all the
possible states of some physical system, and T describes its temporal evolution
from second to second; the equilibrium measure f-l of the system is T-invariant.
Given x E X and A a subset of X, how often does the orbit of x,

visit the set A ?

Result: Let T be a measure-preserving transformation on the probability
space (X, A, f-l) and let A be any element of A. Set

Then, for f-l- almost every x in X,

exists.
(If A is a finite set, #A is the number of elements of A).
Birkhoff's theorem: Given (X, A, f-l, T) as above, set

SN(X) = L f(Tnx)
n<N

where f is f-l-integrable on X. Then

lim N1 SN(X)
N-+oo
= f*(x)
exists f-l-almost everywhere, where f* is f-l-integrable on X, invariant by T, and

2.1.1.4. Now f-l is said to be T-ergodic if the T-invariant sets of A are trivial
for f-l , that is

T-1A =A imply f-l(A) = 0 or

where AC is the complement of the set A in X. For example the Lebesgue

measure m on T is Ra-ergodic if and only if a is irrational.
If f-l is T-ergodic, the limit in the ergodic theorem is f-l-almost everywhere
constant, and equal to f-l(A) in the first result, and to Ixf df-l in the second
one.
378 M. Queffelec

Spatial mean and temporal mean are then identical p,-almost everywhere.
2.1.1.5. The following property of a transformation T is quite natural: suppose
X to be some volume of water and oil, of p,-measure 1, the oil occupying the
part A of X before shaking or mixing (operation T). After n iterations of T,
the oil occupies the part Tn A of X. B being any part of X, the proportion of
volume of oil in B, p,{B)-lp,{Tn An B), is expected to be almost the same as
in X, that is p,{A). This is the mixing-property:
p, is said to be T-mixing (or T is p,-mixing) if
lim p,{T-nA n B) = p,{A)p,{B)
n-+oo

for, A,B E A.
Note that, by choosing for A any T-invariant set of A and B = AC, we get
p,{A)p,{AC) = 0, so that "T-mixing" implies "T-ergodicity". This property is
much stronger than ergodicity since, for example, the Lebesgue measure m on
T is never Ra-mixing.
2.1.2. Result: Let T be a transformation on X, p, and v, two probability
measures on X.
If p, and v are T -invariant and T -ergodic, then

or p,l-v

2.1.3. Example: The following transformation on T will playa very

important role for substitutions of constant length:
Let q E lN, q ~ 2.

Tq is the transformation x -+ qx (mod 211")

The properties such that "Tq-invariance" and "Tq-mixing" of measures on T
can be read in a nice way on the Fourier transform of measures.
2.1.3.1. U E M{T) is said to be Tq-invariant or q-invariant if

u{qn) = u{n) n E 7L

2.1.3.2. u E M(T) is said to be Tq-mixing or q-mixing (and thus q-ergodic) if

lim u(aqn + b) = u(a)u(b) a, bE 7L
n-+oo

2.1.4. Application
2.1.4.1. For Irl ~ 1 consider Pr the Riesz product

Pr = II (1 + r cos 3n t)
n~O
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 379

These probability measures are 3-invariant and 3-mixing as easily seen on

the Fourier transform of Pr and the characterizations 2.1.3.1., 2.1.3.2.
So is the Lebesgue measure m on 'f.
Whence,
Pr 1. Ps r ::f=. s
Pr 1. m
because they are different.
2.1.4.2. In the same way (but rather less obvious),

P= II (1- cos2nt) is singular

n2:0

2.2. For dichotomy properties

We are looking for ergodic properties which induce purity laws (or dichotomy)
for those measures f.L enjoying them: for example

f.L is (purely) discrete

or
f.L is (purely) continuous

Such a property is the 1O-ergodicity

2.2.1. Let 10 a countable subgroup of 1'.

cr E M(T) is 1O-ergodic if A E 8(1'), A +d = A, dE 10, implies cr(A) = 0 or
cr(AC) = o.
2.2.2. Result: Let cr be a 1O-ergodic probability measure: then cr is discrete
or cr is continuous.
A more precise result has been established by Brown and Moran in [10] (see
also [29]).

2.2.3. Examples: We shall deal with special arithmetic groups in connection

with the transformations T q •
10q denotes the group of q-adic rational numbers in l' or, explicitly

10 q = {27r. ~ , k :::; 2n, n = 0, 1, ... }.

2.2.3.1. Fact: The generalized Riesz products

380 M. Queffelec

are IDq-ergodic probability measures and thus discrete or continuous.

As particular cases,

II ~11 + je, 2nt l2 = II (1 + cos(2nt + 2371"))

n;:::O n;:::O

is a continuous probability measure, while

II ~11 + e, 2nt l2 = II (1 + cos 2 t) = 8 n 0

n;:::O n;:::O

2.2.3.2. Fact: The classieal Riesz products

II (1 + rcos3nt) r =I- 0 Irl ~ 1

n;:::O

are continuous probability measures.

ID-ergodicity may be used for mutual singularity.

2.2.4. Result: Let (J' and /L be two ID-ergodic and ID-quasi invariant probability
measures.
Then we have

or
(J'.l./L
(Note that (J' is ID-quasi-invariant if (J'(A) = 0 implies (J'(A + d) = 0, dE ID ).
This result will be used later to determine the mutually singular components
of the spectrum of substitutions.
For complements see [27,47].

3. CORRELATION MEASURES

3.1. Correlation of sequences

3.1.1. Definition: Let U = {Un} be a bounded complex sequence.

We suppose that the correlation function of U i.e. the function r of the
integer k defined by

exists for all k E IN . Then u admits a unique correlation measure (J'u E M(T)
whose Fourier coefficients are
o-u(k) = rk k;:=: 0
o-u( -k) = '?k
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 381

Observe that
l/a,,11 = 11,,(0) = N--oo
lim N1 'L..-J
" lu n l2 .
n<N
The following result is nothing else than a re-formulation of the definition:

3.1.2. Result: au is a weak-limit of positive polynomials; more precisely

au =weak- lim
N->oo
~I
N ~
'" u n etnt l 2 .m
n<N

where m is the Lebesgue measure on T.

3.1.3. A difficult problem is to obtain estimates of the exponential sums

The behavior of such sums is well-known for the classical Rudin-Shapiro

sequence. Allouche and Mendes France [2] and Lesigne [36], give such
estimates for specific sequences: Paper folding and Morse sequences, using
their arithmetical definition (see also [21,37,42,43]).

3.2. Classification

Observe that the constant sequence u with Un :::::: 1 admits as correlation

measure au = Do. Now, if u is a periodic sequence, au a finite discrete measure.
So the question arises: which u have a discrete correlation measure? or a
continuous correlation measure?

3.2.1. Result: Suppose that the bounded sequence u admits a unique

correlation measure au. Then,
au is discrete if and only if 11 is mean almost-periodic that is
-1",
lim N ~ IU n
N->oo
- 11n+k1 :::; E
n<N

for a relatively dense set E, of integers k.

3.2.2. Sequences with continuous correlation measures, are far from almost-
periodic sequences and they are called pseudo-random sequences, since a
random sequence should have a continuous (even Lebesgue) correlation. There
exists no simple characterization for pseudo-random sequences, but a necessary
condition on 11 is the following.

3.2.3. Result [54]: If au is continuous, tllen

382 M. Queffelec

3.2.4. Examples [4]

3.2.4.1. The sequence

is mean almost-periodic and admits discrete correlation measure.

3.2.4.2. Such are, too, paper folding sequences [44].
3.2.4.3. On the opposite side Un = e2 ,,,.,,,n 2 , with a ¢ Q is pseudo-random
because au is easily computed to be the Lebesgue measure.

3.2.5. For general sequences (with unique correlation) we can estimate the
discrete part of au with the following, whose 3.2.3. is a corollary.
Result: with same notations

J~CXl ~ I L une-n"' l ~ (CTu{a} )1/2

n<N

and CTu possesses a discrete mass at any a E T for which the first mean does
not vanish.
3.2.6. Observation: For specific sequences, like substitutive sequences, it
can be proved (once more with ergodic theory) [52]

J~CXl ~I L une-nal = (CTu{a})1/2

n<N

3.2.7. Now, if CTu «: m, u is "in some sense" random. We should say u is

quasi-random.
For complements see [1,14,33,51,52,53,54].

4. SUBSTITUTIVE SEQUENCES - AUTOMATIC SEQUENCES

4.1. Basic notations

• A is a finite alphabet; sometimes A = {O, 1, ... ,s - I}.
• ( is a substitution on A, that is a mapping A -+ A* = {the words on A},
extended to A * by concatenation.
• M = M(() is the matrix of the substitution, more briefly the (-matrix i.e.
the s x s matrix whose coefficient m'J = 1((j)I" the number of i in the word
((j).
• L(((j)) is the composition column vector of the word ((j) so that L,(((j)) =
1((j)l, is exactly m'J.
• If w E A *, is a word with letters in A,
L(w) is the composition vector of w
Iwl is the length of w
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 383

and we have
L(((w))) = M. L(w)

-----
(1 ... I)M = (1((0)1,1((1)1, ... , ((8 - 1)1)
• (n denotes the n-iterate of (. Then

(see the introduction in [4]).

4.2. Basic definitions

4.2.1. ( is primitive if, for some n, Mn > 0 (all entries are strictly positive) or
equivalently, for some n, each word (n(j) contains any i E A.

4.2.2. By the Perron - Frobenius' theorem, M, in this case, admits a dominant

positive eigenvalue 8 (= PF eigenvalue) which is an algebraic number ( a proof
of this theorem can be found in [52]). "Dominant" means IAI < 8 for any other
eigenvalue A of M.

4.2.3. ( is said to be "of constant length" (= q) if

1((j)1 = q for every j E A

and then 8 = q. Otherwise ( is said to be "of non-constant length".

4.2.4. Suppose that 0 is the first letter of the word ((0), and lim n-+ oo I(n(j)1 =
+00 for every j E A. By iterating ( on 0, we get a sequence u E AIN satisfying
((u) = u, for short u = (00(0).
A substitutive sequence u E A IN , is the fixed point of some substitution on A.

4.2.5. An automatic sequence v is the image of some substitutive sequence u

by projection of letters:

v = h(u) where h: A -+ CD

Question: What about the correlation measure of such sequences (provided

it exists) ?

4.3. Examples

4.3.1. Let A = {O, 1,2, 3} and the Rudin-Shapiro substitution (

((0) = 01
((1) = 31
((2) = 02
((3) = 32
384 M. Queffelec

We put u = (00(0).
If we consider the projection h defined by
h(O) = h(3) = a
h(l) = h(2) = b
the sequence v = h( u) is the periodic sequence a b a b a b ... with correlation
1
(Tv = 2(00 + 0,..)
Choose now h
h(O) = h(l) = 1
h(2) = h(3) = -1
We then get for h(u), the Rudin-Shapiro sequence r with correlation measure-
as we shall see later -
(Tr =m
m being as usual the Lebesgue measure on T.
Consequence: The spectrum of the substitution ( - in some sense to be
given later - contains
a discrete measure
and
the Lebesgue measure

4.3.2. Let A be {O, 1, 2, 3} and the substitution

«0) = 01
«1) = 3
«2) = 0
«3) = 32
As usual u is the sequence (00(0).
The projection h defined by
h(O) = h(3) = a
h(l) = h(2) = b
leads to the Fibonacci sequence on {a,b}, whose definition is
«a) = ab
«b) = a
Consequence: The dynamical system associated to ( defined in the next
paragraph admits as a factor ([49]), the Fibonacci dynamical system. It can be
deduced that the Fibonacci correlation is a component of the spectrum of (.
For complements see [12,13,52,55,65].
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 385

5. SUBSTITUTION DYNAMICAL SYSTEMS

5.1. Definitions and basic results

( is a primitive substitution on the finite alphabet A.

u = {un}n>O is a fixed point of (.
T is the uy;jlateral shift on A IN defined by

X is the closed orbit of u in A IN , that is

x = closure(Tnu, n ~ 0)

Roughly speaking, X is the collection of sequences, each of them being close

to some shifted Tnu.
We give now two precise characterizations of these sequences in X.

5.1.1. Symbolic definition of X.

We denote by Lu the set of all the words occurring in u. Then

x E X if and only if, for every i, j ~ 0,

X,X,+l ... x] E Lu

or more briefly, Lx C Lu.

Any word occurring in the sequence x can be found somewhere in the given
u. We shall write X[,,]] for the word X'X'+l ... x ] '
5.1.2. Topological definition of X
AIN may be equipped with the metric

d(x,y) =sup{2- n , n satisfying Xn =f. Yn} and

d(x,y) =0 if x=y

Obviously, given c: > 0, d(x,y) < c: if and only if X[O,n] = Y[O,n] for some
n=n(c:). Now

x E X if and only ifx = lim Tn,u in AIN

]~OO

for some sequence (n]) going to 00 with j, i.e. lim] ~oo d( x, Tn, u) = o.
X is a compact subset of AIN, and T continuous on X.
Both definitions 5.1.1. and 5.1.2. are of course equivalent.
(X, T) is the dynamical system associated to u ([52]).
5.1.3. Minimality - unique ergodicity

5.1.3.1. (X, T) is minimal if, by definition, X is the closed orbit of any sequence
x E X; or, in a more symbolic way, Lx = Lu for any x E X.
386 M. Queffelec

5.1.3.2. There always exists a Borel probability on the compact X, preserved

by T.
(X, T) is uniquely ergodic if, by definition, there exists a unique T-invariant
probability measure on X, which turns out to be ergodic.
5.1.3.3. Result: Under primitivity, the dynamical system (X, T) is minimal
and uniquely ergodic (= strictly ergodic) and depends only on ( (no more on
the fixed point u).
In this particular case, Birkhoff's theorem leads to a more precise result.
5.1.3.4. limN fr 2::n<N f(Tn+k u)= Ix fdlL uniformly in k where
f is continuous on X
IL is the unique invariant probability on X

Now consider w E .cu. The cylinder [w] is the set of all x E X beginning
with w, that is
X[O,lwl[ =w
[w] is an open compact set in X and letting in 5.1.3.4. f = l[wJ, the indicator
function of the set [w] , we obtain with Perron Frobenius' result:
1
N#{n < N,Un+kUn+k+l" ,un+Hlwl-l = w}
tends (uniformly in k) to a positive limit IL([W]), where # stays for the number
of elements.
This proves:
5.1.3.5. Result: Under primitivity, every word w occurring in U (or in any
x EX) occurs with a positive and uniform frequency IL([W]).

5.2. Examples

5.2.1. Frequency of letters

A={O,1, ... ,s-1}

the vector IL([O], 1L[1], ... , IL[S -1]) of frequencies of letters in U is the eigenvector
of M(() associated to the Perron-Frobenius eigenvalue e, normalized by

IL[O] + 1L[1] + ... + IL[S - 1] = 1

• Fibonacci on {a, b}
((a) = ab
((b) = a

M(()=(~ ~) and e = T = 1 + J5
2
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 387

We obtain lL[a] = ~, lL[b] = fr

• Thue-Morse on {a, b}
((a) = ab
((b) = ba

M(() = (~ D, () = 2 and lL[a] = lL[b] = 1/2

• Period-doubling on {a, b}

((a) = ab
((b) = aa

M(() = (~ ~) and lL[a] = ~, lL[b] =~

• Rudin-Shapiro on {O, 1, 2, 3}

IL[O] = 1L[1] = 1L[2] = 1L[3] = 1/4

5.2.2. Frequency of 2-words

If ( is a primitive substitution on A, we can define a new primitive
substitution (2 on the alphabet fh of 2-words in u by splitting the word ((w),
w E [22, in 2-words, in this way: if

((w) = ((wowt) = ((wo)((wt)

= aOal·· .alwol-lbobl' .. blwtl- l

(2(W) is by definition the following word, of length Iwol, whose letters belong
to [22:

The so-obtained substitution (2 has the same Perron-Frobenius eigenvalue

as ( and the eigenvector of M((2) associated to B is the vector of frequencies
of the 2-words in u ([52]).
• Fibonacci on {O, I}
[22 = {OO, 01, 10} = {a, b, e} and ((00) = 01.01 so that (2(a) = (01)(10) = be;
(2(b) = be too and ((10) = 0.01 so that (2(e) = a.

and IL[OO] = 1~2T' 1L[01] = 1L[1O] = H~2T'

388 M. Queffelec

• Thue-Morse on {O, I}

0 0 1 0)
1 1 0 1
fh = {OO, 01,10, ll} and M(2) =( 1 0 1 1
o 1 0 0

whose normalized eigenvector associated to the Perron-Frobenius eigenvalue,

2, is (*,~,~, ~) whence the frequencies of the 2-words in the Thue-Morse
sequence. This study can be achieved in a comprehensive way (see [6,19])

5.3. Spectrum of (

5.3.1. For spectral study, it is better to deal with two - sided sequences; u can
be defined to the left in such a way that (X, T) remains minimal and uniquely
ergodic.
Let ab a 2-word in u such that

«(a) begins with a

«(b) (or (k(b) for a suitable k) ends with b

Then lim (n(b). (n(a)is a candidate, (or lim (kn(b).(kn(a» is a candidate.

n--+<x> +-- ----+ n-+ 00 +--- --+
This is the bilateral sequence whose term of index 0 is the first term of (00 (a).
5.3.2. Operator U
The operator U defined on L2(X, It) by

Uf(x) = f(Tx)
is an unitary operator.
We want to describe the operator U, up to a unitary operator.
5.3.3. Spectral measures
5.3.3.1. 1£ f E L2(X, It), the sequence ({Uk f, f) hEYZ is positive-definite, and
thus, the Fourier transform of some measure OJ E M(T) (Bochner'S result)

5.3.3.2. If f is continuous on X, by the ergodic theorem

 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 389

5.3.3.3. More generally, if f,g E L2(X,p,), the spectral measure aj,g is a

complex, bounded measure defined by

so that aj,J = aj.

5.3.3.4. Special spectral measures
• Take f(x) = Xo or f = 71"0 the projection on the o-indexed component of
x.
Then aj = au the correlation measure of u.
• Choose now f = I[a] , g = I[b] , a, b E A. Then we get the correlation
measure of the occurrences of the letters a and b in u, aa,b, whose Fourier
coefficients are

O-a,b(k) = N--+oo
lim N1 #{n < N, unH = a, Un = b}

• Consider v an automatic sequence, v = h( u) where h: A -+ <C. It is readily

proved that v, in turn, admits a unique correlation measure

av = L h(a)h(b) aa,b
a,bEA

5.3.4. The spectral study of ( is the spectral study of the unitary operator
U (or spectral study of the transformation T), that is the description - when
possible - of the spectral family {aj,g}.
By density argument, the spectral measures (<Tl[wJ' W E .c u ) contain all the
information of the spectrum of (. The problem will be to reduce this family of
generators.

5.3.5. General facts

5.3.5.1.There exists FE L2(X,p,) (F can even be chosen bounded) such that

aF dominates the whole family aj,g , that is

laj,gl «aj «aF denoted by am,

which is called the maximal spectral type.

5.3.5.2. Let f be an eigen function for U i.e.

Uf = f T = e't f
0

Then o-j(k) = e'kt = 8t(k)

and e't is an eigenvalue of U if and only if am {t} > O.
390 M. Queffelec

5.3.5.3. The components of the spectrum of ( are the mutually singular

measures in the decomposition of am.
Note that 80 is always a component of(, (or 1 is always an eigenvalue ofU).
5.3.5.4. If (Y, S, v) is a factor of the system (X, T, JL), with maximal spectral
type a', then a' « am.
So, for exhibiting components of am, it could be ingenious to exhibit factors
of (X, T, JL).

5.4. Discrete substitutions

5.4.1. We begin with equivalent definitions

Definition: "( is discrete" means (X, T) is a discrete dynamical system
and the following statements are equivalent:
a) Every sequence x E X is mean-almost periodic.
b) All the correlation measures a x, x EX, are discrete measures.
c) am E Md(T).
Indeed, for a finite-valued sequence, there is no need to look at the whole
family {ax}:

5.4.2. Fact: (is discrete if and only if au E Md(T).

5.4.3. Now we are able to give some techniques to prove or disprove the property
that ( be discrete:
5.4.3.1. If au is known to be pure, it is sufficient to exhibit a positive mass at
some point x.
5.4.3.2. Otherwise, in view to establish the almost-periodic character of the
Fourier transform au, we enumerate the coincidences (or the non-coincidences)
between u and Tku (k E E relatively dense to be given).
5.4.3.3. We may compare the dynamical system (X, T) with a well - known
discrete system (with discrete maximal spectral type) for example,
• (T, Ra), a ¢ 211'Q where Ra is the rotation on T (see II)

x --+ x +a (mod 211')

with eigenvalue group 7Za and maximal spectral type am ,..., EnE2Z 2-lnI8na'
• The 2-odometer: (7Z 2, X)
7Z2 is the group of 2-adic numbers identified to {O, l}IN, Z E 7Z2 can be
written Z = E~o zk2k and X is the mapping Z --+ Z + 1 (with carryover).
The eigenvalue group of X is ID2 (dyadic rational numbers) and the maximal
spectral type am ,..., EdEID2 2- d 8d •
• The q-odometer: (7Z q , X), for q ~ 2, is defined analogously, with q in place
of 2.
Actually a general result asserts:
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 391

Every ergodic discrete transformation is metrically isomorphic to some

ergodic rotation on a compact abelian monothetic group, with same eigenvalue
group whence the following method:
- we first determine the eigenvalue group r of (. r is countable and G = r
is a compact group,
- we construct a mapping ¢: X -4 G, one-to-one JL-almost everywhere,
exchanging JL and me, T and R g : Z -+ Z + 9 on G, that is

¢(JL)= me
¢(Tx) = ¢(x) + 9

This method, which seems rather "sophisticated", is the best (or the only one?)
when CT u is not computable (very often).
For complements see [15,16,18,30,31,33,39,40,41,49,52,55].

6. SUBSTITUTIONS OF CONSTANT LENGTH

In this part, we focus on substitutions of constant length for which spectral

study is rather easy ([18,39,42]).
We suppose ( primitive on A and one-to-one on the letters (Le. (a) and
(b) are different words for different letters a and b). For convenience, (0)
begins with 0 and u = (00(0).
( is assumed to have length q which means

1(a)1 = q (a E A)

6.1. Examples

• Thue-Morse substitution on {a,b}:

(a) = ab
(b) = ba

• Period-doubling substitution on {a, b}:

(a) = ab
(b) = aa

• Rudin-Shapiro substitution on {a, b, e, d}:

(a) = ae
(b) = de
(e) = ab
(d) = db
392 M. Queffelec

• Regular paper-folding substitution on {a, b, c, d}:

(a) = ab
(b) = cd
(c) = ad
((d) = cd
• Abelian generalized Morse substitutions such that:
On {a,b}:
(a) = ab ... b
(b) = ba ... a
of length q.
On {O, 1, 2, 3},
(0) = 013
(1) = 102
(2) = 231
(3) = 320
of length 3.
• Non-abelian generalized Morse substitutions such that
On {a,b,c},
(a) = aab
(b) = bcc
(c) = cba
• Generalized Rudin-Shapiro substitutions such that

A = {0,1,2, ... ,15}

= {(a, b), 0::; a, b < 4} identifying 4a + b with (a, b)

(a, b) = (0, b)(l, a + b)(2, 2a + b)(3, 3a + b)

In all these examples the spectrum can be described, but with different
techniques, that we explain now. We begin with the discrete case which can
be dealt with readily.

6.2. Discrete substitutions of constant length

6.2.1. Consider the period-doubling substitution

a -+ ab, b -+ aa

Since u = (CO(a), u = (n(u) for every nand u is the concatenation of words

 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 393

Now, if Uk E (n(a), Uk+J2n E (n(a), a = a or b because ( has length 2. We are

thus led to enumerate the coincidences between (n(a) and (n(b).
Iterating ( on (:), we see already that this number of coincidences is
exactly 2n +1 - 1 after n iterations:

b
- - ab

aa
abaa

abab
- abaaabab

abaaabaa
- ...

so that lim f;; Ek<N IUk - Uk+J2" I = O.

This proves that U is mean almost-periodic with Ee = {j2 n , n = n( g)} and
(Tu is discrete (consequence ofresult IV.2.1. and characterization VI.4.2.).

6.2.2. Definition: (admits a coincidence if (k, for s,ome k, has a column of

identical values.

6.2.3. M. Dekking proved the following general result [18]:

( is discrete if and only if ( admits a coincidence.
When the height is reduced to 1 (see [18] for this combinatorial definition),
the dynamical system of the discrete substitution ( of length q is isomorphic
to the q-odometer (system defined in 5.4).

6.2.4. Examples
6.2.4.1. The period-doubling substitution satisfies this condition, the coinci-
dence being evident on the definition; so it is discrete and as a consequence,
the correlation measure of the fixed point sequence is discrete too.
6.2.4.2. Consider now the 3-letters alphabet A = {a, b, c} and the substitution
( defined by :
(a) = ab
(b) = cb
(c) = ba

so that
abc b
= b a c b
c b a b

and the associated dynamical system (X, T) is isomorphic to the 2-odometer.

6.2.5. Question: How to compute (Tu for a discrete substitution?

Answer: by using 2.2.6.
394 M. Queffelec

6.3. Spectrum of a substitution of length q

6.3.1. Results
6.3.1.1. The spectral family {0'J,g} enjoys a self-similarity property whose
description follows.
Let (1 E M(T); cut T = [0,211') into q intervals h = [2;k,
211"(~+1) [, contract
(1 on 10 and reproduce it on the others identically. These operations are

(1 -(1q on 10 where (1q(1) = J

IJ(qx)d(1(x)
q-1
and (1q - 1I'q((1) = L (1q * (ht"
k=O

where * stands as usual for the convolution product. The result means: if
= (11, 1I'q((1) = (1g where g is computable from f.
(1

As a consequence:
6.3.1.2. Since 60 is a spectral measure and

q-1
1I'q(60 ) =L 62;k
k=O

there always exists a discrete part in the spectrum of (, which is L: 2- d 6d , d E

IDq group of q-adic rationals.
In other terms, the q-odometer is a factor of the dynamical system (X, T)
(the q-odometer has been defined in VI.4.3.3.).
6.3.1.3. The continuous part, when existing, is most often singular.
But, in some cases, the spectrum may possess a Lebesgue component.
6.3.1.4. The spectral family {(1I,g} can be deduced from the measures
This matrix of measures is a matricial Riesz product and can
{(1a,b} a,bEA'

be computed in the given examples. Then, as explained in VI.3.3.4., we are

able to compute any correlation measure (11) where v = h(u), and h a mapping
from the alphabet A into ce. Nevertheless the spectrum is sometimes obtained
more easily by isomorphism of dynamical systems.
6.3.2. Examples
6.3.2.1. Thue-Morse
The spectral components of (are: all the Dirac masses at dyadic points
00

{6d}dE ID 2 for the discrete part, and p = I1(1 -

cos 2 n t) which is continuous
o
and singular. The measure p is the correlation measure of the ±1-valued Thue-
Morse sequence and this has been first proved by Kakutani [31].
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 395

6.3.2.2. Rudin-Shapiro
The spectral components of ( are {Od}dE ID 2 and the Lebesgue measure m.
The correlation measure of the ±1-valued Rudin-Shapiro sequence has been
proved to be Lebesgue measure m, by arithmetical considerations (Kamae-
unpublished). See also 4.2. later.
6.3.2.3. Regular paper-folding substitution
The spectrum is the same as above and, by identifying the letters a and b to
0, e and d to 1, the so-obtained regular paper-folding sequence can be shown
to be almost-periodic. This is proved more directly in [44].
6.3.2.4. Abelian generalized Morse substitutions
For the substitution ( on the 3-letters alphabet
((0) = 013
((1) = 102
((2) = 231
((3) = 320
the spectral components are {Od} dEIDa and

II ~ 11 + e t3nt _ et3ntl2
n

II ~ 11 - e t3nt + e t3nt 12
n

II ~Il- e t3nt _ et3ntl2

n

which are continuous, singular and mutually singular.

For
((a)=ab ... b
((b) = ba ... a
of length q, the spectral components are {Od}dEID q and the continuous singular

II ~Il- e ttqn _ e2ttqn - ... _ et(q-l)tqn 12

n q

More generally, in the abelian case (when the column of ( are permutations
of A which commute) the continuous part is singular.
6.3.2.5. Non-abelian generalized Morse substitutions
The permutations commute no more. There is no comprehensive result in
this case, but, a Lebesgue component may be obtained. This is the case with
((a) = aab
((b) = bee
((e) = eba
396 M. Queffelec

6.3.2.6. The generalized Rudin-Shapiro substitution on A = {O, 1, ... ,15}

admits as spectral components

continuous and singular.

Observe that all the measures appearing in these examples, are q-mixing and
IDq-ergodic.
Open question: Which substitutions have a Lebesgue component?

6.4. Dynamical system as extension of the odometer

As remarked in 6.3.1.2., the dynamical system (X, T) associated to the

substitution ( of length q may be viewed as an extension over the q-odometer
by a cocycle (see [49]). We just construct the cocycle in two examples.
This involves the arithmetical property of sequences arising from substitution
of constant length.

6.4.1. Thue-Morse sequence on {±1}

6.4.1.1. The Thue-Morse sequence on ±1 is recursively defined by

Or mn = (_1)S2(n) where S2(n) is the sum of the 2-digits of n.

If we represent mn by the pair (n,c:), n E lN, c: = ±1,
m n +1 = (n + 1, <pen). c:) where <p is by definition the Morse-cocycle; actually
<p allows one to get mn+1 from m n .
Starting from mo = (0,1) we obtain successively

(0,1) -+ (1, <p(0). 1) -+ (2, <p(1 )<p(0). 1) -+ ... -+ (n, <pin) (0). 1) -+ ...

where
<p(n)(z) = <p(z)<p(z + 1) ... <p(z + (n - 1))
(cocycle equation)
By identification, we find easily.
<pen) = (_l)k+l if k is the index of the first 0 in the 2-adic decomposition
of n.
This process extends to the Morse dynamical system by extending <p to LZ 2
6.4.1.2. Result:
The Morse system (X, T) is isomorphic to the skew-product (LZ2 x
LZ/2LZ,T",) with transformation T", defined by

T",(z,c:) = (z + 1,<p(z).c:)
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 397

[10] [lItO]
r-, n
I I II
I I II
I I II
I I II
I I II
I I II
'--_....1 UI
[0] [IIO]

and co cycle <.p: LZ2 ---- LZ /2LZ defined by

<.p(Z) = (_1)~)Z+l)'- I>·

The graph of <.p, on LZ2 identified with {O, l}IN, has the following form:

6.4.2. Substitutions of length 2

For such a substitution, we extend ( to a continuous mapping on X and we
have the relation

It follows that the system (X ((), T) is isomorphic to an extension of the 2-

odometer by the co cycle <.p i.e. a map from LZ2 to the group {±1}, satisfying

The co cycle is thus well known as soon as <.p is defined on the cylinder [0], the
dyadic integers with first component 0 (see VL1.3.4.)
For example, the Rudin-Shapiro co cycle is such that

<.p(Z) =1 if Z E [00]
-1 if Z E [Ol]

Now, <.p is defined everywhere by

k ~ 1
k=O

k being the index of the first 0 in the sequence (Zi). We recover the arithmetical
definition of the classical Rudin-Shapiro sequence {Tn}

where f(n) = #{ll in the 2 - adic expansion of n}.

398 M. Queffelec

This isomorphism can be used with success to provide the spectrum of the
Rudin-Shapiro dynamical system, as observed by Host (private communica-
tion).
The following spectral result ([27]) makes clear the purity property of the
correlation of {m n } or {Tn}.
Result: Suppose (Y, S, U) is a discrete system, with D as countable
eigenvalues group.
ip is a co cycle Y --; G and (X, T, /l) the extension by ip of (Y, S, v)
1- the measure IJ with o-(n) = Jy ip(n)dv (n E 22) is D-ergodic.
2- the maximal spectral type IJm of (X, T, /l) is'" 2:dED 2- d IJ * bd.
We can deduce from this that the correlations of {m n } and {Tn} are D-
ergodic and thus pure.
For complements see [4,9,12,24,31,41,42,49,67].

7. SUBSTITUTIONS OF NON-CONSTANT LENGTH

Let ( be a primitive and non-periodic substitution on A, one-to-one on the

letters, and with non-constant length vector

(1((0)1,1((1)1,"',1((8 - 1)1)

It follows that the Perron-Frobenius eigenvalue () is not necessarily an integer

but more generally, an algebraic integer (root of a polynomial with integral
coefficients, the coefficient of the term of highest degree being equal to 1).
The spectral study in this case is far from being complete and there no more
exists evident similarity for the spectral family {IJ j,g}. An important result
is the description of the eigenvalues of ( and thus, of the discrete part of the
spectrum [26,40].

7.1. Description of eigenvalues

This discussion partially appears in [7], without rigorous proofs (see also [25]).

7.1.1. Suppose first that the alphabet A = {a, b}. The result is the following:
oX = e't is an eigenvalue of ( if and only if

n--+oo n--+oo

Observe that the discrete part of the spectrum depends only on the matrix M
since, as we mentioned,
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 399

7.1.2. Examples.
7.1.2.1. Fibonacci: ((a) = ab, ((b) = a and () = T = 1±2V5
>. is an eigenvalue of ( if and only if

where Fn is the nth Fibonacci number and the eigenvalue group G is exactly

7.1.2.2. () unitary (quadratic) Pisot that is ()()' = ±1 where ()' is the conjugate
of (). G is the group generated by e2rrtp [aJ or e21rtp [bJ.
7.1.2.3. () (quadratic) Pisot.
G contains the group generated by e2rrtp [aJ but may be strictly larger.

7.1.3. Result for a general alphabet.

7.1.3.1.>' = e tt is an eigenvalue of ( if and only if, for a E A,

lim >.I(n(all
n~oo
= cp(a)
where cp is a coboundary on A, whose definition follows:
7.1.3.2. cp is a coboundary on A if there exists a complex-valued mapping f on
A, If I = 1 such that
cp(a) = f(b)f(a)
for every 2-word ab E Lu (see 5.1.1.)
Note that, for a 2-letters alphabet, the only coboundaries are trivial ones
and equal to 1, whence the formulation 7.1.1.

7.2. Various situations for A = {a, b}

We shall encounter, already with a 2-letters alphabet, the following situations
for the spectrum:
- Purely discrete spectrum.
- Mixed spectrum (that is a non-trivial discrete part together with a
continuous part) and then, an absolutely continuous component may be found.
- Purely continuous spectrum (apart from 80 ) and in this case, there is no
component in Mo('f) [20]. We recall that 80 always appears in the spectrum
as explained in 6.3.5.4.

7.2.1. Continuous spectrum

We denote by M = M (() the matrix of ( ; d, the determinant of M ; T, the
trace of M ; ()' the conjugate of ().
400 M. Queffelec

7.2.1.1. Consider now the following somewhat generic example

a ---- ababa
b ---- abb

M = (~ ~ ), T = 5, d = 4, (j = 4, (j' =1.

The integers Pn = l(n(a)l, qn = l(n(b)1 satisfy

PI = 5, ql = 3
2qn - Pn = 1 for every n 2: 1

So that limn e,tpn = limn e,tqn = 1 implies e't = 1, and the discrete part of the
spectrum reduces to 80 .
7.2.1.2. Result :
Suppose ( defined on {a, b} satisfies
a) T and d are mutually prime.
b) PI and ql are mutually prime too.
c) WI 2: 1
Then the spectrum of ( is continuous (apart from 80 ),
Proof: From a) and b) we deduce, as in the above example, that Pn and
qn are mutually prime for every n : the relations

give

(1) {
Pn+l : PIPn + mlO(qn - Pn)
qn+l - qlqn + mOI(Pn - qn)

where M = ( moo mOl) and PI = mOO + mlO, ql = mOl + mll'

mlO mll
Moreover, by the Cayley Hamilton equation, one has

Pn+l : TPn - dpn-l

(2) {
qn+l - Tqn - dqn-l

If some prime number P =1= 1 divides both Pn and qn, for some n, then P divides
Pn+l and qn+l by (1), thus dpn-l and dqn-l by (2).
If P does not divide d, P divides Pn-l and qn-l ; by continuing, P should
divide Po and qo which is impossible.
Now, P divides d and cannot divide T from assumption a). Using (2) with
n - 1, P must divide Pn-l and qn-l, and descending till P2 + d = TPI,
q2 + d = Tql, we get that P divides PI and ql mutually prime by b); a
contradiction .
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 401

We now deduce that G, the eigenvalue group, reduces to 1. If e 2Jr ,t E G,

limn e2Jr ,tPn = 1 and

with Tn E 7Z, and En = dist(tPn, 7Z) tending to O. We derive from (2)

Or Tn+l + dTn-l - TTn = TEn - dEn-l - En+l.

But, for n :::: no, ITEn - dEn-l - En+11 < ~ and

Tn+l = TTn - dTn-l

En+l = TEn - dEn-l

By solving this recurrence, we have for n :::: no

Thus a j3 = 0 since limn En = 0 and () > WI

= :::: 1. This proves that
tpn = Tn (n:::: nO) and in a similar way tqm = Tm (m:::: mo). Combined
with the first part, this gives the result.
7.2.1.3. Examples and comments
This result, once more, depends only on M provided that ( is non-periodic.
So, non-periodic substitutions with following matrices, have continuous
spectra

But the order of letters is relevant for the description of the continuous part:

o ----> 01010 0 ----> 00011

leads to ( : and TJ :
1 ----> 011 1 ----> 110
Both substitutions have continuous spectrum but are quite different: they
have different complexity function (already on 3-words).
If O2 = {OO, 01,1O,11}, (2 and TJ2 the extended substitutions on O2 ,

M((2) __ (0221 ~; 0~1 0~1) admits 4,1,0,0 as eigenvalues

402 M. Queffelec

~ 0~ ~ D
while

M(v,) admit, 4,1,1,-1

Open question : Which exact role play these additional eigenvalues?

7.2.2. Mixed spectrum

7.2.2.1. We explore first the example borrowed from [47]. Consider

with integral eigenvalues 3,0 and non-periodic substitutions with M as (-matrix

0-+01 0-+01 0-+01

1 -+ 1001 1 -+ 1010 1 -+ 1100

Result: The first two are discrete substitutions; the third one has mixed
spectrum.
Proof: Coding the first substitution by

a = 01 b = 10,
we get a factor of the initial dynamical system, generated by the substitution
a -+ aba
, which is a discrete one. More precisely the factor is isomorphic to
b -+ baa
the 3-odometer. By decoding, (X, T) is itself isomorphic to

(7Z/27Z x 7Z 3 ,r)

• For the second substitution, (X, T) '" (7Z /27Z x 7Z 3 , r) the 3-odometer
a -+ abb
coming from the substitution b -+ bba
• For the third substitution (with different complexity) we exhibit a
continuous factor, in coding by a = 01, b = 11, c = 00. This yields a new
substitution on {a, b, c}
a -+ abc
b -+ bcbc
c -+ aa
whose associated dynamical system is isomorphic to the following one, because
b is always followed by c :
a -+ abc
b -+ bcb
c -+ caa
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 403

This last substitution of length 3 admits a continuous-even Lebesgue- compo-

nent (see the example of non-abelian Morse substitution).
7.2.2.2. Dekking proved a more general result on A = {O, 1,'",8 - I} [18].
Result: If (1((0)1, 1((1)1"",1((8 - 1)1) is a left O-eigenvector of M, the
system is isomorphic to the system defined by some substitution of constant
length.
Example: The substitution ((0) = 12, ((1) = 102, ((2) = 0 fulfills the
hypothesis of the above result.
7.2.2.3. Remark: Returning to the case of a 2-letters alphabet, this occurs
only if 0' = O. Actually, with usual notations

(p, q)M = (1, I)M2 = (1, I)(T M - dl) by the Cayley-Hamilton equation
= T(p, q) - d(l, 1)

while (O',O')M = 0'(1, I)M = O'(p,q) so that

(p - O',q - O')M = (T - O')(p,q) - 00'(1, 1)
= O(p - 0', q - 0')

Suppose now that (p, q) is a left O-eigenvector of M. Necessarily 0' = O.

Open question: what about the reciprocal ?

7.2.3. Discrete spectrum

We focus, in this section, on the quadratic unitary Pisot substitutions:
A = {a,b}, 0 > 1 admits as conjugate 0' with IB'I < 1 and 00' = ±1.
Pisot numbers have been already discussed in 1.4.3. Because of their
specific property in Diophantine approximation theory, they playa great role in
harmonic analysis ( thin sets, Bernoulli measures ... )[45], and their connection
with substitutions has been already observed in [7,38,56]. We do not know if any
Pisot substitution is discrete, even in the quadratic case, but the result is true
for a large class of them, which contains generalized Fibonacci substitutions.
7.2.3.1. Fibonacci substitution revisited
((a) = ab, ((b) = a and u = (OO(a). Remember that

o = T = 1 + y'5, 0' = _.!. = 1 - y'5

2 T 2

are the Fibonacci numbers.

How to prove that ( is a discrete substitution?
404 M. Queffelec

We list below some possibilities:

(i) Directly, by estimating the number of non-coincidences between u and
Tku, k E Eo, where

Eo = {IC(w)l, wE .cu } for fixed n = nee) .

Then prove that u is mean almost-periodic.

(ii) This can be done in an algorithmic way using an observation of Host
[52].
Fact: Denne

( is discrete iflimk--+oo Dk = 0, the convergence being geometric.

This means, we need look only at Tku, k E {1(n(a)I}.
Now, TI(k(a)lu = (k(Tu), and we enumerate the non-coincidences between
u and Tu to get the non-coincidences between (k(U) = u and (k(Tu).

u ab a ab
a ab a ab ab
Tu : ba a ~~ a ~ a
~ ~
It can be proved easily that the non-coincidences come exactly from the non-
coincidences between (k(ab) and (k(ba).
Let Nk be this number of non-coincidences between (k(ab) and (k(ba).
Expanding
(k+l(ab) = (k(aba) = (k(a) (k(ba)
(k+1(ba) = (k(aab) = (k(a) (k(ab)
we see that Nk+1 = Nk = ... = 2.
It follows that
. 1
Dk = hm N#EN,k
N--+oo

= lim N1 . NkL(k(ab)(U[O,N[)
N--+oo

recalling the notation Lw (U[O,N[) for the number of occurrences of w in U[O,N[.

so that Dk tends to 0 in a geometric way and ( is discrete.

(iii) Computation of the correlation measure of u = (00(0).
The discrete masses of (Fu are taken from (t5 21r k-r' k E ZZ). Then,applying
observation (2.6) for correlation measures,

(Fu{2?rkT} = lim (_1_ 1 '""' u n e- 2nn k-r1)2 = O!k

N--+oo FN L...J
n<FN
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 405

because the limit exists

and, by proving 2: O:k = ~, au has to be discrete, of the form

(iv) Geometric representation.

We follow the scheme given in (5.4.3.3.). The group G of eigenvalues of ( is
generated by

So, we aim to construct an isomorphism between the associated dynamical

system
(X,T,JL) and ([O,l),R",m)
where R" is the rotation x --+ x + 0: on [0,1).
Consider the sequence (tn) obtained by projection of the Fibonacci tiling:

tn = o:#{k < n,Uk = a} + (0: -l)#{k < n,uk = b}

1 1
= T2La(U[O,n[) - :;:Lb(U[o,n[)
so that (tn) is bounded.
Define now f: X --+ IR by

f(u) = 0, f(Tnu) = tn
f(Tx) =f(x) + 0: if Xo = 0
f(x) + 0: - 1 if Xo = 1

so that 9 = e 211rJ satisfies g(Tx) = e 211r"g(x).

It can be proved that 9 is continuous and one-to-one (JL-almost everywhere)
providing the expected isomorphism.
Note that
g(x) EXo if Xo = 0
Xl if Xo =1
where
7.2.3.2. Result for quadratic unitary Pisot substitutions
A = {O, I} and G, the eigenvalue group, is generated by e2 11"t/L[O] or e211"t/L[I].
By the geometric method above, the following partial result can be proved
[28,56).
406 M. Queffelec

Result: If (0) and (1) begin (or end) with identica11etters, tIle system
is discrete.
Example 1:
(0) = Onl
(1) =0
admits a discrete spectrum for n ~ 1.
Observe that M = (~ ~) and (J, (J', the roots of P(X) = X 2- nX -1 =0
satisfy
() > 1 > 18'1
Example 2:
(0) = On+1-Pl OP
(1) = IOn
admits a discrete spectrum for any n 1 and 1 ::; P ::; n.
. (n+l n)
~

In thIS case, M = 1 1 and (J , ()', the roots of

P(X) = X2 - (n + 2)X + 1 = 0
satisfy
(J> 1 > (J' > 0

7.2.3.3. Algorithm for the other cases

We consider first the example of the square-Fibonacci modified substitution

(0) = 001
(1) = 10
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 407

M admits two fixed points u = (<Xl(0) and v = (<Xl(l), that we write:

u 001 0 01 10 0 0 1 001 10 100 01

v 100 0 10 01 0 0 1 100 01 001 10

Definition: wand w' being two words on A, (~,) is called a balanced

pair if L(w) = L(w') and the pair is minimal with this property.
In the decomposition of ( ~) in balanced pairs, (~) and (~) appear.
Result:
1) If () is a quadratic unitary Pisot number, then the set of balanced pairs
in (~), denoted by A, is finite.

2) ( is discrete if (and only if) (~) or (~) belongs to A.

Example:
((0) = On+l1 OP
((1)=10n
admits a discrete spectrum for n ~ 1.

7.3. Miscellaneous

We just give examples illustrating different cases.

7.3.1. Not unitary quadratic Pisot number

Consider

((0) = 001 with matrix ( 21 02) so that

((1) = 00

() = 1 + V3, ()' = 1 - V3 are the roots of X 2 - 2X - 2.

We prove that ( is discrete, by applying the method (ii) used for the
Fibonacci substitution.
We enumerate the non-coincidences between u and TP'u where Pk = l(k(O)1

u : (k(O) (k(O) (k(l) (k(O) (k(O) (k(l)

TP'u : (k(O) (k(l) (k(O) (k(O) (k(l) (k(O) ...
'--v---" '--v---"

It suffices to calculate Nk, the number of non-coincidences between (k(Ol) and

(k(lO); in order to establish a relation between Nk+l and N k , we consider

(k+l(Ol) = (k(OO) (k(100)

(k+l (10) = (k (00) (k (001)
408 M. Queffelec

and Nk+l = Mk the number of non-coincidences between (k(001) and (k(100).

Now
(k+l(001) = (k(OO) (k(100) (k(100)
(k+l(100) = (k(OO) (k(OOI) (k(001)

so that Mk+l = 2Mk

and Nk+l = Mk = 2k+l
We conclude by computing

Dk = lim N1 #{n
N~oo
< N,u n =j:. u n +Pk }

= NkJL[(k(OI)]

= (~)kJL[OI]
It follows that limk Dk = 0 geometrically and ( is discrete.

7.3.2. Remark
When IAI > 2, and () is a unitary Pisot number, (may happen to be discrete
only if the characteristic polynomial of M is equal to the minimal polynomial
of (). As counter-example, take A = {O, 1,2, 3} and
(:0 -+ 01
1 -+ 2
2 -+ 23
3-+0
The characteristic polynomial of M, P(X) = (X2 - X + 1)(X2 - X-I) has
additional roots c, c' of modulus 1, together with T and T', while the minimal
polynomial is (X2 - X-I).
The dynamical system (X((), T) is not discrete but a continuous extension
by a co cycle of the discrete one (X(1]), T) where 1] is the Fibonacci substitution.

7.3.3. Chacon substitution

Chacon in [11] constructed a sequence of 0 - 1 by juxtaposition of blocks
B n , with the following rule:
Bo = 0
Bn+l = Bn B n 1Bn
The sequence w is the limit of
o
0010
0010001010010
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 409

We define, as usual, X = X(w) = {x = (Xn)nEZZ/X[s,tj E .c w }.

The dynamical system is topologically isomorphic to a substitution dynam-
ical system (X((), T) where
A = {a,b,c}, ( defined on A by
((a) = ab
((b) = cab
((c) = ccab
Indeed, ifu = (OO(c), w = h(u) where h:A -+ {0,1} commutes with the shifts,
and
h(b) = h(c) = 0
h(a) = 1
To the opposite, because a is always followed by b, u is obtained from w
by changing 1 in a, 0 in b if 0 follows 1, in c otherwise. This furnishes the
isomorphism [22].
Result: The spectrum of ( is continuous.
This is an immediate consequence of the result 1.3.
Let cp be a coboundary of (: because cc is a 2-word of u,
cp(c) = f(c)f(c) = 1
Also ab and ba are in .cu and
cp(a) = f(b)f(a)
cp(b) = f(a)f(b) with If I = 1
so that cp(a)cp(b) = 1.
Observe now that Pn = l(n(a)l,qn = l(n(b)I,Tn = l(n(c)1
satisfy Pn + qn - Tn = 1.
If e't is an eigenvalue of the system,

tends to cp(a)cp(b)cp(c) = 1
and e d = 1.

7.3.4. Higher dimensions

We just quote recent results of Solomyak [60] concerning Pisot substitutions
on arbitrary alphabets.
Suppose A = {1,2, ... ,m} and define
(1: 1 --+ 12 ... m
2--+1
410 M. Queffelec

1 1 0 0
1 0 1
o
1 0 1
1 0 o

and Perron-Frobenius eigenvalues (}1

2---+1. .. 13
'-v-'
a2

m-l---+1. .. 1m
'-v-'
am-l

m ---+ 1. .. 1
'-v-'

with a1 2: a2 2: ... 2: am 2: 1 and Perron - Frobenius eigenvalue (}2.

Result: (}1 and (}2 are unitary Pisot numbers, and (1 and (2 have discrete
spectrum.
These results have been obtained by using the adic version of a substitution
dynamical system [60], and the property of (}1 and (}2 to be unitary Pisot
numbers, is proved in [8].
Observe that (2(a) C [1] and ({"-1(a) C [1] so that a conjecture could be
the generalization of 2.3.2. for larger alphabets.
For complements see [7,38,43,59,63].

7.4. Conclusion

A general result about Pisot substitutions must be accessible, especially in the

quadratic case. Now, for any substitution, the description of the continuous
part of the spectrum would be very interesting but cancellations may happen
to be inextricable. Other classes of deterministic sequences (often with overlaps
with the class of automatic sequences) have been studied with few partial
spectral results: Keane's generalized Morse sequences [33,34], cutting sequences
[17,61]' q-multiplicative sequences [14,51], .... or the curious sequences obtained
by composing two different substitutions [65].
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 411

8. References

[1] ALLOUCHE J.P., LIARDET P. : Generalized Rudin-Shapiro sequences,

Acta Arith. 60 {1991} 1-27.
[2] ALLOUCHE J.P., MENDES FRANCE M. : Suite de Rudin-Shapiro et
modele d'Ising, Bull. S.M.F. 113 {1985} 273-283.
[3] ALLOUCHE J.P., MENDES FRANCE M. : Quasi-crystal Ising chain
and automata theory, J. Stat. Phys. 42 {1986}.
[4] ALLOUCHE J.P., MENDES FRANCE M. : this volume.
[5] ARNOLD V.I., AVEZ.A. : Problemes ergodiques de la mecanique
classique, Gauthier- Villars Paris 1967.
[6] BERTHE V. : this volume.
[7] BOMBIERI E., TAYLOR J.E. : Quasicrystals, tilings and algebraic
number theory: some preliminary connections, Contemporary Math. 64 {1987}
241-264.
[8] BRAUER A. : On algebraic equations with all but one root in the interior
of the unit circle, Math. Nachr. 4 {1951} 250-257.
[9] BRILLHART J., CARLITZ L. : Note on the Rudin-Shapiro polynomials,
Proc. Amer. Math. Soc. 25 {1970} 114-118.
[10] BROWN G., MORAN W. : A dichotomy for infinite convolutions of
discrete measures, Proc. Camb. Phil. Soc. 73 {1973} 307-316.
[11] CHACON R.V.:Weakly mixing transformations which are not strongly
mixing, Proc. Amer.Math.Soc. 22 {1969} 559-562.
[12] CHRISTOL G., KAMAE T., MENDES FRANCE M., RAUZY G. :
Suites algebriques, automates et substitutions, Bull. S.M.F. 108 {1980} 401-
419.
[13] COBHAM A. : Uniform tag sequences, Math. systems theory, 6 {1972}
164-192.
[14] COQUET J., KAMAE T., MENDES FRANCE M. : Sur la mesure
spectrale de certaines suites arithmetiques, Bull. S.M.F. 105 {1977} 369-384.
[15] CORNFELD J.P., FOMIN S.V., SINAI Y.G. : Ergodic theory. Springer
Verlag, 1982.
[16] COVEN E.M., HEDLUND G.A. : Sequences with minimal block growth,
Math. Systems Th. 7 {1973} 138-153
[17] CRISP D., MORAN W., POLLINGTON A., SHIUE P. : Substitution
invariant cutting sequences, J.theorie des nombres de Bordeaux 5 {1993} 123-
143.
[18] DEKKING F .M. : The spectrum of dynamical systems arising from
substitutions of constant length, Zeit. Wahr., 41 {1978} 221-239.
[19] DEKKING F.M. : On the Prouhet-Thue-Morse measure, Acta Univ.
Carolinae, mathematica et physic a, 33 {1992} 35-40.
[20] DEKKING F.M., KEANE M. : Mixing properties of substitutions, Zeit.
Wahr., 42 {1978} 23-33.
412 M. Queffelec

[21] DEKKING F.M., MENDES FRANCE M. : Uniform distribution

modulo 1; a geometrical viewpoint, J.Reine Angew. Math. 329 (1981) 143-
153.
[22] FERENCZI S. : Les transformations de Chacon: combinatoire, structure
geometrique, lien avec les systemes de complexite 2n + 1, to appear in
Bull.S.M.F. 123 (1995).
[23] GARSIA A. : Arithmetic properties of Bernoulli convolutions, Trans.
Amer. Math. Soc. 102 (1962) 409-432.
[24] HELSON H., PARRY W. : Cocycles and spectra, Arkiv for Mat. 16
(1978) 195-206.
[25] HOF A. : Quasicrystals, Aperiodicity and Lattice Systems, Thesis
Rijksuniversiteit Groningen (1992)
[26] HOST B. : Valeurs propres des systemes dynamiques definis par des
substitutions de longueur variable, Erg. tho et dyn. Sys. 6 (1986) 529-540.
[27] HOST B., MELA J.F., PARREAU F. : Analyse harmonique des
mesures, Asterisque (1986) 135-136.
[28] HOST B., PARREAU F. : Private communication.
[29] HOST B., PARREAU F. : The generalized purity law for ergodic
maesures: a simple proof, Colloq. Math. 60/61 (1990) 205-212.
[30] JACOBS K., KEANE M. : 0-1 sequences of Toeplitz type, Z. Wahrsch.
Verw. Geb. 13 (1969) 123-131.
[31] KAKUTANI S. : Strictly ergodic symbolic dynamical systems, Proceed-
ings of the sixth Berkeley Symposium on Mathematical statistics and Proba-
bility (1967) 319-326.
[32] KATZNELSON Y. : An introduction to Harmonic Analysis, Wiley
(1968).
[33] KEANE M. : Generalized Morse sequences, Zeit. Wahr. 10 (1968)
335-353.
[34] KWIATKOVSKI J. : Isomorphism of regular Morse dynamical systems
induced by arbitrary blocks, Stud. Math. 84 (1986) 219-246.
[35] LEDRAPPIER F. : Des produits de Riesz comme mesures spectrales,
Ann. LH.P. section B, 4 (1970) 335-344.
[36] LESIGNE E. : Le thooreme ergodique ponctuel Ie long de la suite de
Morse, Preprint.
[37] LESIGNE E., PETERSEN K. : Bounded exponential sums, Osaka J.
Math. 27 (1990) 277-280.
[38] LUCK J.M., GODRECHE C., JANNER A., JANSSEN T. : The nature
of the atomic surfaces of quasi-periodic self-similar structures, J. Phys. A:
Math. Gen. 26 (1993) 1951-1999.
[39] MARTIN J.C. : Substitution minimal Hows, Amer. J. Math. 93 (1971)
503-526.
[40] MARTIN J. : Minimal Hows arising from substitutions of non-constant
length, Math. Syst. Th. 7 (1973) 73-82.
[41] MARTIN J.C. : Generalized Morse sequences on n-symbols. Proc.
A.M.S. 54 (1976) 379-383.
 AUTOMATIC AND SUBSTITUTIVE SEQUENCES 413

[42] MAUDUIT C. : Automates finis et ensembles normaux. Ann. Ins.

Fourier 36 (1986) 1-25.
[43] MAUDUIT C. : Sur l'ensemble normal des substitutions de longueur
quelconque, J. Number Theory, 29 (1988) 235-250.
[44] MENDES FRANCE M., VAN DER POORTEN A.J. : Arithmetic and
analytic properties of paper folding sequences, Bull. Austral. Math. Soc. 24
(1981) 123-131.
[45] MEYER Y. : Algebraic numbers and Harmonic analysis, North Holland
(1972).
[46] MEYER Y., WEISS B. : Les produits de Riesz sont des Bernoulli shifts,
Seminaire de probabilites de Rennes, (1974).
[47] MICHEL P. : Coincidence values and spectra of substitutions. Zeit.
Wahr. 42 (1978) 205-227.
[48] MOSSE B. : Notions de reconnaissabilite pour les substitutions et
complexite des suites automatiques, (submitted to Theor.Comput.Sci.)
[49] PARRY W. : Topics in ergodic theory. Cambridge University Press,
(1981)
[50] PEYRIERE J. : Etude de quelques proprietes des produits de Riesz,
Ann. lnst. Fourier 25 (1975) 127-169.
[51] QUEFFELEC M. : Mesures spectrales associees a certaines suites
arithmetiques, Bull. S.M.F. 107 (1979) 385-421.
[52] QUEFFELEC M. : Substitution dynamical systems - Spectral analysis.
Lecture Note in Math. n01294, Springer-Verlag (1987).
[53] QUEFFELEC M. : Une nouvelle propriete des suites de Rudin-Shapiro,
Ann. Inst. Fourier 37 (1987) 115-138.
[54] RAUZY G. : Proprietes statistiques des suites arithmetiques, P.U.F.
ColI. Sup (1976).
[55] RAUZY G. : Suites a termes dans un alphabet fini, Sem. theorie des
nombres, Bordeaux, expo 25 (1982-1983).
[56] RAUZY G. : Structure des systemes dynamiques definis par substitu-
tions (unpublished).
[57J RIESZ M., NAGY B.SZ. : Lecons d'analyse fonctionnelle, Gauthier-
Villars (1972).
[58] RUDIN W. : Real and complex analysis, Mac Graw Hill (1968).
[59] SENECHAL M., TAYLOR J.E. : Quasicrystals : the view from Les
Houches, Math. Intelligencer, 12 (1990) 54-64.
[60] SOLOMYAK B. : Substitutions, adic transformations and beta-
expansions, Contemp. Math. 135 (1992) 361-372.
[61] SERIES C. : The geometry of Markoff numbers, Math. Intelligencer 7
(1985) 20-29.
[62] SUTO A. : this volume
[63J WEN Zhi-Ying: Substitutions, fractals et dimensions. Habilitation
Univ. Paris-Sud (1990).
[64] WEN Zhi-Ying : this volume.
414 M. Queffelec

[65] WEN Zhi-Ying, WEN Zhi-Xiong: Des remarques sur les suites
engendn§es par des substitutions composees, Ann. Fac. Sci. Toulouse, 9 (1988)
56-65.
[66] ZYGMUND A. : Trigonometric series. Cambridge University Press,
London, (1959).
[67] GODRECHE C., LUCK J.M. : Multifractal Analysis in reciprocal
space and the nature of the Fourier Transform of self-similar structure,
J.Phys.A.Math.and Gen. 23 (1990) 37-69.
[68] PEYRIERE J. : this volume.
 COURSE 13

Random and autonlatic walks

F.M. Dekking

TU Delft, FacilIty of Mathematics and Informatics,

Dcpartmcnt, of St.at.istics, Proballility and Operations Research
Mekclweg 4,
2628 CD Delft,
The Net,her/ands

1. INTRODUCTION

The symmetric simple random walk in Zd describes the movement of a particle

on the d-dimensional lattice which, indepcndent.ly of its past, moves at time
instants 1,2, ... with equal probability to one of its 2d neighbours. Here, the
particle is at the origin at time O. The fundamental question about the be-
haviour of the particle is whether its walk is recurrent: i.e., it returns to the
origin infinitely often (with probability one), or transient, i.e., eventually the
walk will leave any ball around the origin (with probability one). The answer
has already been given by Polya in 1921: if d ::; 2 then the walk is recurrent, if
d ~ 3 then it is transient [25].
Now consider the paper/olding walk introduced in the lecture of Mendes
France [21] &'3 the paperfolding curve. Here a particle moves in Z2, taking
left and right turns according t.o whether t.here is a 0 or a 1 in the paperfolding
sequence.
Can we interpret it as a random walk? That is, can we find a notion of which
on the one hand the simple symmet.ric random walk and the other hand the
416 F.M. Dekking

Fig. 1. - Symmel.ric simple random walk in IR2

Fig. 2. - Paperfolding walk

paperfolding walk arc special cases? Can we SHY that the paperfolding walk is
transient? In t.hat case it would be a counterexample to (a generalization of)
Polya's theorem (cf. [4]). The basic quest.ion adressed in these lectures will
be how to compare the structures ill Figure 1 and Figure 2. This has specific
relevance to the subject of quasi cryst.als, as we shall show that the Penrose
pattern can be obtained as points visit.ed by a walk in a general framework
which encompasses both the (simple) random walk and the paperfolding walk.

2. AUTOMATIC WALKS

To motivate the forthcoming definit.ions, we consider again the paperfolding

walk. In Mendes France's lectures t.his was described by the paperfolding se-
quence - a sequence on t.he t.wo symhols 0 and 1. However, in our context it
is somewhat more natural t.o descl'ibe it with four symbols, where the symbols
code for the four possible steps to neighbouring points in 1l2. Let A = {a, b, c, d}
be a 4 letter alphabet. We identify a with the vector (~), b with (~), c with
«/) and d with C\). Then t.he (part of the) paperfolding walk in Figure 2
can be coded with the word

abcbcdcbcdadcdcb. (1)
 RANDOM AND AUTOMATIC WALKS 417

There is a way 1.0 obtain this word, and words describing any finite length part
of the paperfolding walk, by a subst.it,II(,ion u on t.he set of words A*. This
substitution is given hy

u(a) = ab, u(b) = cb, u(c) = cd, u(d) = ad. (2)

The word in (1) is equal to u3 (a) as one can quickly check. That indeed the
iterates of u will describe any finite lengt.h part of the paperfolding walk (of
length 2") can easily be deduced from the geomet.ric recursion given in Mendes
France's lecture [21]. What we have done here is to associate an endomorphism
u of A* to a geometrica.l object in ~d (t.his is very reminiscent of the program
of Kramer's lectures [19]- see also [15]). To ma.ke t.his more formal, we can say
that we have considered a homomorphism

where homorphism means that for all words v, wE A*

J(t/tII) = J(v) + J(tII).

The typical example is given by (,he alphabet A = {a, b, c, d}, the space ~2,
and

If u is the infinite sequence which is the fixed point of (J" starting with a, where
(J" is the substitution in (2), then

is the nth point visited by the paperfolding walk.

A second example is given by the alphabet, A2 = {a, b}, the space JR, and the
Prouhet-Thue-Morse substitution U2 defined by

If we now define 12 : A; -+ IR. by

heal = 1, h(b) = -1

then Sn( tP») = h( ui

t/~2») gives us a sequence of points in Z which might
2 ) •••

be called the Prouhet-Thue-Morse walk (here again t,(2) = abbabaab· .. is the

fixed point of U2 starting wit.h a). Rather amazingly this walk only visits 3
418 F.M. Dekking

Fig. 3. - The 2D-Fibonacci walk (first 25 steps).

different points: S,,(u) E {-I, 0, I} for all 71 2': O. Even a superficial knowledge
of the Prouhet-Time-Morse sequence enables one to see why this must be true.
A third example is given by the same alphabet {a, b}, t.he space ]R2 and the
Fibonacci substitution U3 defined by

u3(a) = ab, u3(b) =a.

Now we consider h : {a, b} * -+ JR2 defined by

h(a) = G), h(b) = C)·

The walk S" (tP» = h( 1I~3) ... u~,3» generated by the fixed point u(3) of U3 is
very transient (see Figure 3), and well known in quasi crystal theory.
Generally, if we are given an alphabet A, a substitution u on A*, and a ho-
momorphism f : A* -+ JRd, then we call the sequence (S,,(u»,,~o an automatic
walk in JRd, where u E AN is any fixed point of u. Especially one dimensional
walks (d = 1) have been ext.ensively studied in the literature. We shall describe
the main results of [13] and [:31].
Let Al = l'vlo be the mat.rix of the subst.itution u, i.e.,

111 = (m.d.,tEA with lII.t = lu(t)I.·

The entries of Al count the number of occurrences of the symbol s in the word
 RANDOM AND AUTOMATIC WALKS 419

O'(t). For example, the paperfolding subst.itution 0' (sec (2» has matrix

( 11 01 00 01)
/11 0 = 0 1 1 0 .
o 0 1 1

An often exploit.ed fact is that if M is primitive (i.e., a power of M is strictly

positive) then M has a real eigenvalue 0 (t.he Perron-Frobcnius eigenvalue)
such t.hat IAI < 0 for all ot.her eigellvalucs A of AI, and there is a. unique st.rictly
positive eigenvect.or 'l' with AI'l' = O'l' and 2:,'EA 'l'(s) = 1.
Theorem ([13]) If (5,,(11») is an aut.omat.ic walk in lit, t.hen as n -> 00

Here ('l!, f) is the scalar product 2:'EA 1l1(.s)f(s). Furthermore, O2 is second

largest eigenvalue of M (supposed to be uniqlle and larger than 1), 0' + 1 is the
multiplicit.y of O2 in t.he minimal polynomial of M 0 and F : [1,00) -+ IR is a
bounded cont.inuolls funct.ion.

This result is somewhat less complct.e t.han it, might seem to be, as it. docs not
tell whell the function F is identically equal 1.0 O. Especially in the interesting
ca.'le ('l!, J)=0, where the growth of S .. is no longer linear, a supplementary
analysis is required, This call be found ill [31] for two Ictt,er alphabets. We
confine ourselves to giving an example inspired by their result.s. Let A = {a, b},
and let 0' be defined on A* by

O'(a) = (/.aba, O'(b) = babb.

Then AI = (~;), so 0 = 4, O2 = 2. Let. t.he homomorphism f be defined by

f(a) = +1,I(b) = -1. Then the walk (S,.(u(l) is transient, where u(l) =
aabaaabababb ... is one fixed point. of 0', bul t.he walk (5,,(u(2)) is recurrent,
where ,P) = babbaabababb . .. is the other fixed point of 0'.
It appears t.hat. implicit.ly [31] uses a kind of self-silllilarit.y of the walks,
which is forced in the case of a t.wo lett.er alphabet and (1l1, I) = O. This is
an indication why results are less complete in the general case (cf. e.g. [4],
(11], [8]). This self-similarity will he explained in t.he next, section, where we
shall also apply it to give a proof of the t.ransience of the tt(1Lwalk and the
recurrence of the 1I.(2L walk above.
420 F.M. Dekking

Fig. 4. - A non self-similar aut.omatic walk. (The subst.itut.ion u is given by U(CI) =

Clda, u(b) = bbe, u(e) = cell, u(d) = dba, t.he fixed point is u = bbeb ... )

3. SCALING STRUCTURE AND SELF-SIMILARITY

We call an automat.ic walk (.)n (11» general.ed by a fixed point 11 of a substitution

u : A* -> A* and a homomorphism f : A* -> rn:d sclf-similar if t.here exists a
linear map L = £" such t.hat for all sEA

f(u(05)) = £(1(05». (3)

Remark on the nomenclat.ure: some aut.hors might. call I.he walk self-affine,
since £ need not be a similarit.y tra.nsformation.
We have already seen examples of such walks: the paperfolding walk is self-
similar with £ given by the matrix n-D, and the Prouhet-Thue-Morse walk
is self-similar in a very trivial way, t.he map £ being Illultiplicat.ion by O. The
2D-Fibonacci walk generated by t.he Fibonacci sequence u(3) = abaababa ...
and t.he map f«(I) = (1,0),/(1» = (0,1), is also self-similar with £ given by the
matrix n ~).
At this point we would like to Illention that automatic walks have an obvious
extension to walks generated by free group substitutions. Any endomorphism
u of the free group Fr on 7' generat.ors wit.h any homomorphism f : Fr ->
rn:d determines an aut.omatic walk (Sn(lI» as before, if 11 is a fixed point of
the endomorphism. For an interesting example take F2 =< a, b >, f(a) =
(~),/(b) = (~) and the endomorphism u: F2 -+ F2 given by u(a) = ab,u(b) =
a-lb. These free group endomorph isms are useful to generate self-similar tilings
(cr. [2], [3], [16]) .
=
We now return to the example u(a) aaba, u(b) babb, f(a) = 1, feb) =
-1 =
at the end of the previous section. It is easily checked that the associated
automatic walks are self-similar, with L being multiplication by 2. Note that
 RANDOM AND AUTOMATIC WALKS 421

(3) implies

for all n ~ 1. So in this example

This can be used to prove the recurrence of (Sn (U(2»)):

I(qn(ba» = I(q" (b)qn(a» =

I(qna) + I(q"b) = 2n - 2n = O.

Hence (Sn(U(2»)) returns to 0 infinit.ely often. The proof of the transience of

(Sn(u(1») is a little bit more involved. Here one shows with induction that if
=
v(n) = qn(a) and w(n) q7l(b) then

1«
_ vI n) ... v (n)
k
< 2" ,
_ _2n ~ w~n) ... w~n) ~ 0 for all 1 ~ k ~ 4n ,

and then that

S,,(tP») ~ 1 for all 11 ~ 1.

Actually it is conceptually easier to follow this proof using the self- similarity
of the graph «n,
Sn))I!~O of the automat.ic walk (with scaling map given by the

matrix (~ ~)).

4. QUASI LATTICE WALKS

Let u E AN and 1 : A* --+ IRd a homomorphism. We call the walk (S,,(u»n~O

quasi-lattice if the set of points

A = {SI! ( u) : n ~ O}

has half of the Delaunay property, i.e., if

37' >0 such that Card(Br(x) n A) ~ 1 't/x F IRd

where Br(x) denotes the hall of radius 7' cent.ered at x. This means that any
two different points of the walk are at least distance 7' apart.
It is tempting (cf. lectures of Meyer [20], Patera [24]) to define a quasi crystal
as a set of points obtained from a self-similar automatic walk with the complete
Delaunay property (i.e., not only (*) holds, but also there exists r' such that
422 F.M. Dekking

Card(Brl(x)nA) ~ 1 for all x E plus some irrationality condition, as e.g.

]Rd) -
requiring that the Perron-Frobenius eigenvalue of Mo is irrational.
In fact one can obtain the Penrose pattern exact.ly in this way. The walk
associated to the following substitution ([5]) will generate the Penrose triangle
pattern (as first constructed by Robinson ([30])). Let

A {aO,O,k, aO,I,k, al,O,k, al,I,1;, aO,O,l:>

• •
aO,I,k> •
al,O,A:> •
al,I,A:> b0,0,1;,
bO,I,1;, bl,o,k, bl,I,A:, Vk : k :::: 0,1, ... , 9}.

Define! : A· ->]R2 (identified with C) by

,d ::::
• llr'tlr
!(a.,.,k) :::: f(a .. k) :::: f(b., f(vA:) :::: e"TO k:::: 0, ... ,9.

Consider further the following symmetries

1I"1(ao, .. ) 1I"1(al .... ) :::: ao ....

1I"2(a,0,) a',I, 1I"2(a.,I,) :::: a ,0 ..
1I"3(a ... ,k) a .. ,k+1 k :::: 0, ... ,9 (addition mod 10)
J-l3( a ... ,k) :::: a· .. ,9-k

where 1Tt, 1T2, 1T3 are extended in the obvious way to the symbols a·,.,k and b.,.,k
and /13 to the Vk. The subst,it,ut.ion (J' will then be defined by

(J'( ao,o,o) a~, 1 ,8 a I ,I ,2 bo,0,0

IT( a~,o,o) V8a~,0,3a I ,0,9 bo,l, I

(J'( bo,o,o)
(J'(vo) V9 V I,

complet.ed by the symmetry requirements

where r: A· -> A· is defined by r(wI ... w"'):::: w m . "WI.

See Figure 5. for the corresponding walk.
We return to the quasi lattice walks. There is a simple sufficient condition,
duc 1.0 F. Parrcau (priva(.c cOll1l11unicat.ion), to obtain quasi-lattice walks on
Z(r), where r :::: ~(1 + .;5). Let A :::: {a, 1, 2,3,4}. We call a substitution
(J' : A* -+ A* symmetric if (J' commut.es with the Nielsen automorphism of A*,

i.e. (J'Q:::: Q(J', where Q(j) :::: j + 1 (mod 5).

 RANDOM AND AUTOMATIC WALKS 423

Fig. 5. - Penrose triangle pattern as automatic walk, obtained from IT (8 th iteration)

defined in Section 4. (algorithm of [5])
a.<;

Proposition Let u be a symmctric subst.itution on A* = {O, 1,2, 3,4}*, and

e
let = e27r ;/5. Let

be the generating function of the first. column of Mo. If

IP(OI > 1 alld W(e)1 < I,

then u generates a quasi lattice walk with f : A* -+ <C defincd by

I(j) = {i,j = 0, ... ,1.
Example Let u be given by

u(O) = 01, er(l) = 12, er(2) = 23, er(3) = 31, u(4) = 40.
424 F.M. Dekking

Fig. 6. - The quasi taU,ice walk associat,ed to the example at the end of Section 4
(wHh a detail).

Then P(z) = 1 + z. So

IP(~)I = r> 1, IP(e)1 = -r' < 1.

lIenee (S,,(u» is CJuasi latt.ice, where 11 = 01121223·,·.
5. DYNAMICAL SYSTEMS, RANDOM AND AUTOMATIC

We return to t.he problem described in the' int.roduction: how is it possible to

see both random walks and aut.omatic walks as instances of the same structure?
We shall see here that a solut.ion is given by dynamical systems. Recall from
the lecture by Queffelec ([27]) that t.he symbolical dynamical system is the set
AN of all infinit.e sequences of symbols from the alphabet A, with the map T the
shifl defined by (Txh = Xk+1 for all k ~ 0, and a (Borel) probability measure
11 .
The dynamical system associated to random walk is the one where Jl IS
product measure, i.e. if

is a cylinder of lengt,h k, I,hen

where (PI,P2," .Pr) is a given probabilit.y dist.ribution on A =

{I, 2, ... , "}.
= =
E.g. for symmet,ric simple random walk on Z, " 2 and (I J I, P2) (~, ~).
 RANDOM AND AUTOMATIC WALKS 425

The dynamical system associated to an aut.omat.ic walk is t.he one where the
measure equals Jio, t.he unique shirt. invariallt. lI1easure on t.he closed orbit

where u is a fixed point of u. (To have this uniqueness we assume that the
matrix Mo of u is primitive). Now we have

(4)

for all cylinders tV, i.e. the probabilit.y of a cylinder is just the frequency with
which t.he corresponding word til occurs in t.he sequence 11. (sec e.g. [26])
For an example, consider \.he Prouhe\.-Thuc-l....lorse sequence u = 011010011001 .
and the word tV = O. Since tt is a conca\.ena.tion of 01 and 10, we have

Ittl tt2 ... tl.2m 10 = 111., Ittl U2 •.. tt2m+llo =m or m +1

l.e., the frequency of 0 in ttlU2.·.ttm differs less than 2111\1 from ~ for all
m = 1,2, .... Therefore /1'0([0]) = ~.

6. GENERALIZED RANDOM WALKS

We still have to see how a dynamical sys\.em gives rise \.0 a generalized random
walk (G.R.W.). This is simple: any (measurable) funct.ion f : A -+ rn.: d gives
rise to a generalized random walk (5,,(x)),,~o, J: E AN by

5,,(J:) = f(.1:I) + f(X2) + ... + f(x,,).

Note t.hat, we obtain a realizatiol/ (S,,(x)) of t.he G.R.W. (5.. ) for each x E AN.
The realizations are obt.ained according t.o the measure It.
A G.R.W. (5,,) is called recurrcllt if

II {x E AN: :1 71 5" (x) = Q} = 1,

ot.herwise (5,,) is called [ral/siell/.
For ergodic G .R. W. 's t.he same dichot.omy holds as for ordinary random
walks:
Thcorcm ([29], [I)) If/I is crgod ic t.hen
(5,,) recurrent <=> JI.{X EX: 5,,(;:) Q oo-ofl.en} = =1,
426 F.M. Dekking

(Sn) transient. ¢:> It{X E X :11 SIlex) 11-> oo} = 1.

For G.R.W.'s on the illteger lattice ~d we define t.he range Rn by

U .. (x) = Card{SI(x) ..... S,,(x)} n = 1,2, ...

This is the total number of different point.s visited by the walk in the first n
steps.
Theorcm Let II be ergodic, and let. W = {x EX: S,,{x) ;f Q. 'rIn ~ 1}.
Then as n -> 00

1 n)
-(R
71
= -1 71
J U,,(x)dIJ(X) -+ Il(W), a.nd

1
-Un -+ II(W) I" - almost surely.
71

Since W is a subset of the complemellt of the set {x EX: S,,(x) = Q. oo-often}

a G.R.W. (S,,) is recurrent if and only if JI(W) = 0, so this theorem gives us a
crit.erion for recurrcnce.
With aid of this crit.erion one can prove the following rat.her classical result.
Theorcm An ergodic G.Il.W. (S,,) with (lSI!) < 00 in IR is recurrent ¢:>

(St) = f f(Xt)d/l(X) = O.
In case (S.. ) is associat.ed to a primit.ive substit.ut.ion (T we have

(St) =L f(s)Jl([s]) =L f(s)lI1(s) = (I,II1),

where 111 is t.he normalized eigenvector for t.he Perron-Frobenius eigenvalue 0 of

Ma ·
This theorem tells us that t.he G.R.W. generated by the substit.ution u given
by
u(a) = aaba, u(b) = babb
considered at the end of Section 2, with f(a) = 1,J(b) = -1 is recurrent. This
means that the behaviour of the fixcd point tI(2) = babbaaba ... is typical, and
that of u(1) = aabaaaba ... is atypical.
Aft.cr t.his ollc-dimensional cxample wc consider now the paperfolding walk.
Recall that A = {a,b,c,d},u(a) = ab,u(b) = cb,u(c) = cd.u(d) = da and
f(a) = (~) = - fCc), feb) = m= - fed). A simple frequency computation
(recall (8» shows that

I' ([a]) = Jl([bJ) = p([c]) = Il([d]) = 4'1

 RANDOM AND AUTOMATIC WALKS 427

Fig. 7. - A realization of t.he paperfolding G.R.W., i.c. a walk (Sn(X», where x E

Orb(u), u the 4-lcHer papcrfolding sequcnce.

hence (SI) = O. But the paperfolding generalized random walk is transient as

we shall show. This means that the theorem above is the best possible, and
that Polya's result is different. for G.Jl.W.'s: here the crit.ical dimension is 1.
Paradoxically, the introduct.ion of correlations makes that a drunk man can be
lost easier.
We shall now prove that the paperfolding G.Jl.W. is transient. Let x E
SuPP(tt,,) (the smallest closed set with /t,,-meaSlire 1). Then

/1,,([£1 ... xt) > 0 for all k -> 1,

since otherwise the complement of the cylinder [XI ... X.l:J would be a closed set
of /-,,,-measure 1. But 1.!Jen it follows f!'Om (4) that XI .. . x.!: occurs in tll" .tt",
for In large enough. This implies t.hat the walk f(£I), f(Xl)+ f(X2), . .. , f(xt}+
... + f(Xk) is a (shined) "subwalk" oft.he walk f(uJ),f(lId+ f(U2), ... ,f(ud+
'" + f(lI",). But 1/115 walk obviously does not. visit, a point, more than twice,
(Maybe this is not so obvious, see [lOJ for a proof.) lIence the same is true for
f(Xl),' .. , f(xJ) + ... + f(xd· Since k was arbit.rary, t.his shows that (Sn(x»
is t.ransient.
Consider for any primitive substit.ut,ion 0' the "mother walk" (Sn(tl»n~o of a
fixed point tt of 0'. We sa.y (Sn (tI» duplicates if3 ZI, Z2 E Zd and 3 N, AI with N >
AI such that
428 F.M. Dekking

Using the Range-Theorem one can show the following (see [4] for a special
case).
Theorem A self-similar G.R.W. generated by a primitive suhstitution (J' is
transient if and only if (Sn(u)) does not duplicate.

There is no algorithm known to decide whether (S,,(u)) duplicates or not.

In many examples however this is very easy, e.g. in t.he 2D-Fibonacci walk
it is obvious that. every point is visited only once, hence (Sn(ll» does not
duplicate, and this G.R.W. is t..ransient.. SOllie special ca.'5es (generalizations of
the paperfoldiug walk) are covered by [11].

7. MEAN SQUARE DISPLACEMENT

One of t.he most int.eresting quant.it.ies of a (generalized) random walk is its

mean square displacemellt defined by

MSD J
= (II 5" If) = II SIl(X) If dJl.(x).
(We assume throughout t.his section t.hat. this quant.it.y is finite). For exam-
ple consider simple (or more complicated) random walk in ~d. Write /(s) =
(h(x), ... ,fd(S». Let us suppose (51) = O. Then

So, by independence,

lIence
as n -> 00, (5)

and the critical exp(Jllell tn, defined by

o
(II 5 n 112)! /
as Il -+ 00 (6)
P
n '" 00 /3<n
 RANDOM AND AUTOMATIC WALKS 429

Fig. 8. - Rudin-Shapiro walk (first. 15 sl.eps).

equals 0:' = ~.
IIencc all mcan zcro random walks havc thc samc critical cxponent. and the
diffusion consta1lt (1f + ... + (1~ given by (5) cxist.s. What. about the aut.omatic
walks?
Lct us considcr thc examplc of thc Prouhct.-Thuc-f\lorsc walk on ;:z (thc
substitut.ion is givcu by (1«(1) = (Ib, (1(b) =
ba and t.he cmbedding map by
/(a) = +1,/(b) = -1). Now all realizat.ions (S' .. (x»,x E Orb(u) have the
property 1S',,(x)1 :::; 2 for all n ~ O. lIencc t.hc critical cxponent 0'=O. But.

(5~n+l) =1 for all 71

(I.his is easy t.o prove), a.nd

2
(52") =:34 for all 11 (7)

(this is harder to prove).

So here the diffusion constant does 1Iot exist. The computation in (7) is
based on [7] where a rat.her precise descript.ion of t.he mcasurc 1107 is given
«(1 t.he Prouhet.-Thue-Morsc substit.ut.ion): all cylinders of a fixed lengt.h with
positive II,,-measurc have eithcr measure i2-n or measure ~2-n for some n.

Large critical exponents may also occur. One can show t.hat t.he 2D-Fibonacci
walk has crit.ical cxponcnt 1. (Since all steps arc in positive direct.ions only, this
walk is lIlaximally diffusive). We furt.hermore conjccture t.hat. e.g. t.\1f' G.R.W.
in )R2 (f as before) generat.ed by the substit.ut.ion IT(a) =
aba, (1(b) =
bcb, (1(c) =
cdc, (1(d) = dad, has crit.ical exponent. log 5/ log 9.
This raises the question of which automatic walks bchave likc mC'an zero
random walks in t.hc scnsc t.hat. t.heir crit.ical exponent is equal to ~. Thcre
is a rclation with the spectral mcasure of the dynamical system associated to
the substitut.ion. Recall from the Icctures by M. Qllcffelec (see also [26]) that.
430 F.M. Dekking

the spect.ral measure of the Rudin- Shapiro sequence is Lebesgue measure, as

it is for t.he Bernoulli dynamical system which is associatcd to random walk.
We conjecturc t.hat indeed the Rudin-Shapiro G.R.W. has critical exponent
4. Here we mean by t.he Rudin-Shnpiro G.R.W. the G.R.W. generated by the
substitution u given by

u(a) = ab, u(b) = ad, u(c) = cd, u(d) = cb

where f : {a,b,c,d} -+ ~2 is as before. The automatic walk (Sn(tt» with

tt the fixed point It = abad ... hns a very regular structure (see Figure 7 for

the pat.h t.raced hy (S,,(tI)~~o). In fact the Rudin-Shapiro automatic walk has
been studied before, e.g. in [22] nnd [17]. In particular, if t.he last formula in
[17] would have been correct., we would have a simple proof of 0' ~ for the 4
Rudin-Shapiro G.R.W. Here follows a proof of this upperbollnd that involves
some more computat.ions. First. not.e that for all I.: ~ 1

II f(uks) 11= (J2l s = a,b,c,d (8)

Since any word tv of lengl.h 2k occuring in tI cont.ains al. lenst one complete
u k - 1 (s), one complete u k - 2 (s), etc., (8) implies that for aliI.: ~ 1

lIence, if for an arbitrary 11, we writ.e its binary expansion 11 = L~=o ck2k, then
(using the Cauchy-Schwartz inequalit.y)

N
IIf(wt}+ ... +f(w,,)1I < L cd 1 + v2)(v2)k
k=O

< (1+ ,(2)",N + l~ t. <,2'

(1 + v2)VN+lVii.
It follows that. for any x E X t7 one has (uniform in J:)

II Sn{r.) 112= O(n log n) as 11 --+ 00.

Bul. then (II S" 112) = O(n log n), and 0' ~ ~. Lower bounds on 0 can not
be obt.ained by uniform f'st.illlat.iolls as nbove - it. is not. difficult. to see that It
 RANDOM AND AUTOMATIC WALKS 431

=
contains arbitrary long words tv with few) Q. Still, the computat.ions will be
rat.her combinatorial hecause by unique' ergodicit.y of It", one has

JII S~(x) II d,t(X) = N-oo

N
lim N1 "II
L..,
"=0
SIl(r" u) 112

This expression also makes it clear why there is a connection with spectral
measure.
All this is closely relat.ed to t.he calculat.ions performed by Kamae and Keane
([18]), who show that. a diffusion limit of a one-dimensional Rudin- Shapiro
walk has covariance function equal t.o that of Brownian mot.ion. However, their
computat.ions do not. seem to yield t.he lower bound on a.

Refercnccs
[1] Berbee H.C.P., Recurrence and transience for random walks with sta-
tionary increments, Z. Wahrsch. tlerw. Geb. 56,531-536(1981).
[2] Dekking F.M. Replicat.ing superfigures and endormorphisms of free
groups. J. Combin. Theory Ser. A. 32 (1982), 315-320.
[3] Dekking F.M. Recurrent set.s, Adv. in Malh. 44 (1982).78-104.
[4] Dekking F.M. On transience and recurrence of generalized random
walks. Z. Wahrsch. verw. Geb. 61 (1982),459-465.
[5] Dekking F.M. Recurrent set.s: a fractal formalism. Report 82-32. Tech-
nische IIogeschool Delft, Delft., 1982.
[6] Dekking F.l'vl. On the dist.ribution of digit.s in arit.hmet.ic sequences.
Sem. Theoric Nombres Bordeaux An. 82-83 (1983). 3201-3211.
[7] Dekking F.M. On the TIme-Morse measure. Acta Unitl. Carolin.-Math.
Phys. 33 (1992). 35-40.
[8] Dekking F.M. Marches automatiques. J. Tldorie Nombres Bordeaux 5
(1993). 93-100.
[9] Dekking F.M. It.Nat.ion of maps hy an all\.oma\.on. Diurele Math. 126
(1994),81-86.
[10] Davis C. and Knuth D., Number represent.ations and dragon curves I,
II, J. Recreatlollal Malh. 3 (1970). 61-81,133-119,
[11] Dekking F.M., Mendes France, M., v.d. Poorten. A. Folds! Math.
Intelligencer 4 (1982), 130-138, 173-181, 190-195.
[12] Dekking F.M. et Mendes France M., Uniform distribut.ion modulo one:
a geometrical viewpoint., J. Reille Angew.Math. 329 (1981), 143-153.
432 F.M. Dekking

[13] Dumont. J .-IVI. et Thomas A., Systcmes de numerat.ion et fond-ions

fractales rclatif.c; aux substil,utions, Th. COII/p. Science 65 (1989),
153-169.
(14] Dumont .J.-M., Summation formulae for substitutions on a finite al-
phabet. Number Theory and Physics (Eds: J.M. Luck, P. Moussa, M.
Waldschmidt). Springer Lect. Not.es Physics 47 (1990). 185-194.
[15] Garcia Escudero .J. and Kl'amer P., An int.crprct.ation of 2D-quasi pe-
riodic pat.t.el'lls in t.erms of aut.omorphisms of frec groups, J. Phys. A.
Malh. Gcn. 26 (1993) L1D29-1035.
(16] Ito S., Weyl aut.omorphisms, subst,it.utions and fractals (1989), Preprint.
[17] Kahane J.P., IIelices et. quasi-helices, Math.Anal. and Appl.E, Adv. in
Mat.h. Suppl. St.udies 7E (1981),487-433.
[18] Kamae T. and Keane M. A class of determinist.ic self- affine processes,
Japan J. Appl. Malh. 7 (1990), 183-195.
[19] Kramer P., t.his volume.
[20] Meyer Y., this volume.
[21] Mendes France 1'1'1., t.his volumc.
[22] Mendes-France 1\1. and Tenenbaulll G. Dimensions des courbes planes,
papiers plies et suit.es de Rudin-Shapiro, Bull.Soc.Math.France 109
(1981 ).
[23] Mendes France M. et Shalit.t. .J.O., Wirebending and cont.inucd frac-
l.ions. J. Combinal07'ial Theory Ser A 50 (1989), 1-23.
[24] Patera J ., this vol ume.
[25] Polya G., Uber eine Aufgabe der Wahrscheinlichkeitsrechnung betref-
fend die Irrfahrt, illl St.rassennet.z, Malh.Allal. 84 (1921), 149-160.
[26] Queffelec M., Su bst.it.ut.ion dynamical systems-spect.ral analysis. Leet.
Notes Mat.h. 1294 (1987), Springer Verlag, Berlin.
[27] Queffelec M., t.his volume.
[28] Rauzy G. Nombl'es algebriques et. substitutions. Bull. Soc. Math.
France 110 (1982), 147-178.
[29] Schmidt. K., Lectures on cocycles of ergodic transformat,ion groups.
Lect. in Mat.h 1., Mac Millan. 1977, Delhi-Eombay-Calcutt.a-Madras.
[30] Robillson R.M., ComllH:'nt.s 011 t.he Penrose t.iles, Univcrsit.y of Califor-
nia, Berkeley, 1975.
[31] Wen Z.-X ct Wen Z.-Y, M1lrches sur les arbres homogenes suivant une
suit.e substitut.ive. J. Theone Nombrcs Bordcaux 4 (1992), 155-186.
 COURSE 14

Singular words, invertible substitutions and

local isomorphisms
Wen Zhi-Ying

Dcpar/.JI]cnt of Mat/lCma/.ics, Wll/Jan Vnivcrsi/.y, lVu/Jan 430072, P.R. China

1. INTRODUCTION

One-dimensional (ID) quasiperiodic Schrodinger equations have been studied

by many authors in recent years (sec for example [21]). In particular, much
attention has been focussed on quasiperiodic potentials that. are derived from
the Fibonacci sequence, providing a kind of prot.otype models for studying
quasiperiodic systems. Starting from the Fibonacci sequence, some authors
have proposed diverse generalisat.ions [22]. On t.he other hand, the combina-
t.orial properties of the Fibonacci infinite word are of great interest in mathe-
matics and physics, such as fractal geometry, quasicrystals, number theory, for-
mal language, computational complexity et.c. See [1,3,4,5,7,11,12,13,14,17,22].
Moreover, t.he propert.ies of t.he suhwortls of t.he Fibonacci infinit.e word have
been studied extensively by many aut.hors [3,7,13]. In this lecture, we shall
present some new properties of the subwords of the Fibonacci word. This lec-
t.ure consists of three part.s. In t.he first part., after recalling some preliminary
remarks on t.he Fibonacci word, we int.rouuce the singular words and discuss
their elementary properties. Then we est.ablish two decompositions of the Fi-
bonacci word in singular words and t.heir consequences. By using t.hese results,
we discuss the local isomorphislll of t.he Fibonacci word and the overlap prop-
erties of the factors, and we give a necessary and suflicient condition for local
434 Wen Zhi- Ying

isomorphism. In t.he sccond part., WI' introduce the class of invert.ible substitu-
tions which is a natura.l g('Jlcralisa.tion of t.he Fibonacci substitution, and which
is also an important. class of ID quasipcriodic chains in physics [21]. We study
first the group structure and local isomorphism of the invcrtible substitutions,
then we study further the propert.ies of the singular words of t.he invertible
substitutions. In the third part wc return again t.o the Fibonacci sequence.

2. SINGULAR WORDS

Let A = {a, b} be a two letter alphabet, and let 0' be t.he Fibonacci substitution
dell ned by
O'(a) = ob, O'(b) = a.

We write Fn = O'n(o) ror the nth itcrat.ion of 0', e.g.,

Fo = a, FI = ab, F2 = aba, F3 = abaab, F4 = abaababa.

As is well known, the length IFni of Fn equals In, the nt/I Fibonacci number.
We also denote Foo for the Fibonacci sequence, thc fixed point of 0'. (which
start.s wit.h Fn for all n).
We shall stutly the words of lengt.h 11. One way t.o obtain new words from a
wortl t' = 111 tl2 •.• tin is by conjugat.ion. The kIll coujugale of tI is defined by

for k = 0, 1, ... , n - 1.
The set {Ck (tI) : 0 ~ k < n} of all conjugates of a word v is denoted as C( v).
Some words in t.he Fibonacci sequence have many conjugates:
Property 1. All conjugates of Fn are different, i.e. Card C(FII ) = In.
Let nf" be the set of all words of length In occurring in the Fibonacci sequence.
By Property 1 all conjugates of F" are in nf ". How many other words are there
in nf .. ? Surprisingly t.here is only one word tvn missing.
Define W-I = =
a, tlIo b, tvl =
aa and for 11 ~ 1

(1)
=
For example, W2 bab, tv3 = aabaa, tv4 babaabaabab. =
Obviously tv" rt. C( F,,), since it contains one more or one less letter a than the
word F".
Property 2. nJ" = C(F,,) U {w Il } .
 SINGULAR WORDS 435

This property makes tII n st.and out as a very special word, it will be called the
singular' word of lengt.h n.

We mention a direct recursive way to obtain w". Since q2(a) = q(a)q(b) =

q(a)a, we have F"+l F"Fn - 1= =
Fn-lFn-2Fn-l, and therefore

(2)
This immediately shows that Wn is a palindrome for all n (a word v = VI ••• Vk

is called a palindrome if v =v= VkVk-l" .vd. It also follows directly by

induction from (2) that

It follows by (1) that F 2n + 1 = a-1W2n+lb = W-lWOtlllW2 .• ,W2n-lb and we

obtain that the Fibonacci sequence is a concatenation of its singular words:

Theorem 1. F00 = n;: -1 Wj.

With some more work (see [20]) one can deduce from (2) an even more striking
decomposition into singular words. For each n ~ 1 we consider the alphabet
Ln consist.ing of t.he two words W'l-l and tLln+l. We define a subst.itution <Pn
from {a,b}* to E:, by /f'n(o) tlI,,+l, =
<Pn(b) = tlIn-l. We call t.he sequence
z = <p,,(Foo ) the Fibonacci word over·L".

TheorClll 2. Foo = n;'':~l tlIjlLln,lZlWn,2Z2·· .Wn,kZkWn,k+l ... where each

Wn,k = tvn and ZI, Z2 ..• is the Fibonacci word over Ln'
For example if n = 2 we obtain t.he following decomposition:

abaa bab aabaa bab aa bab aabaa bab aabaa ...

lU2,2

One can show moreover that t.he words w" only appear as a tv",k and hence
singular words of the same order are separat.ed posit.ively. This property plays
an important role in the st.udy of radors the Fibonacci sequence. Here we mean
by a factor a word occurring in another (finite or infinit.e) word.
Example 1. Power of the factors.
We can state rather precisely how often patterns in Foo can repeat:
Theorem 3.
1. tv~ does not occur in F oo ,
2. for 0 ~ k ~ fn - 1, (CA·(Fn»2 does occur in F oo ,
3. if u does occur in Foo with f .. -l < lui < In t.hen u 2 does not occur in F oo ,
436 Wen Zhi- Ying

4. If 0::; k ::; 1,,-1 - 2, then (Ck(F.. ))3 does occur ill Foo, but for In-1 - 2 <
k < In (Ck(Fn))3 does not occur in Foo
5. no u 4 can occur in Foo.
The proof follows easily from Propert.y 2 and Theorem 2.
Example 2. Local Isomorphism.
= =
Let u U1 U2 ... tl" ... and v V1 V2 ... Vn ... be two infinite words over A. We
say that u and v are locally isomorphic if any factor (or its mirror image) of u
is also factor of v and vice versa.
Theorem 4. 1) If we change a finiLe number of letters of F 00, then the obtained
infinite word l":x, is not locally isomorphic Lo Foo.
2) Let tt E A*, t.hen Foo is locally isomorphic to uFoo if and only if 3 m > -1,
such that W2n+1a or tlJ 2n b st.arts wiLh tt.
3) For any k ~ -1, define Tk(Foo) = Xk+! ... Xk+n"" then Tk(foo) is locally
isomorphic to Foo (this has nothing to do with Theorem 2, but follows directly
from the recurrence of Foo).
Example 3. Overlap of the subwords of Foo.
Let x = X1X2'" x" ... be an infinite word and let W = X",Xk+! ... Xk+p be a
factor of x, where k, p ~ 1. If there is an integer e, 1 ::; e ::; p, stich that

then we say that w has overlap with p - e as length of overlap.

Theorem 5.
1) Wn has no overlap,
2) if In-1 - 2 ::; 1: ::; In, then Ck(F.. ) has no overlap,
3) if 1.. -1-2 ~ k> 0, then CA:(Fn) has overlap oflengLh 1.. -2, and Ck(Fn ) =
uvu, lui = In-2, Ivl = 1.. -3.

3. INVERTIDLE SUDSTITUTIONS

With the alphabet A = {a, b} we also consider the free group F2 =< a, b >
generated by the leLters a and b. A substitution 7r : A* --+ A* can be naturally
extended to F2 by defining 7r(s-l) = (7r(s»-l. If 7r is a substitution, and
7r is also an clement of AliI. (F2)' i.e., t.here exists a p : F2 --+ F2 such that

p7r = trp =
Id on F2 , then 7r is called an invc7'liblc substitution (here Id is
the substitution given by Id(a) =
a, Id(b) =
b). The Fibonacci substitution 0'
given by
(T(a) = ab, (T(b) = a
 SINGULAR WORDS 437

is an example of an invert.ible substitution: choose 11" defined by

Then

1I"a(0.) 1I"(ob) = 1I"(a)1I"(b) = bb-Ia = a

11" a ( b) 11"( (l) = b

a1l"( a) d(b) = a
a1l"(b) a(b-Ia) = a(b- a(0) = (a(b))-lab = a-lab = b.
1

Since a is an invertible substit.ution, we are led to consider a natural generaliza-

tion of the infinite Fibonacci sequellce: the sequences generated by invertible
subst.itut.ions.
There is an old result. charact.erising all aut01110rphisms of F z . Let cp and t/J be
defined hy

Theorcm (Nielsell)([D]) Aut.(F2 ) =< a, cp, 1/' >.

The following result characterizes all invert.ible substitutions.

Theorem G. ([H))) The set. ofinvcrt.iblc substitut.ions is equal to < a,a',cp >.
Here a' is the substit.ut.ion given by a'(o) = ba, a'(b) = a.
Rcmark From the proposil.ion and the t.heorem above, we see that. to study
the properties of factors of an infinite word generated by an invertible subst.i-
tution, it suffices to consider t.he element.s of < a, cp >, that is, the invertible
substitutions of the following form,

where nZ, ... nk_1 ~ 1,nl,nk ~ O.

As in the case of Fibonacci, we can introduce also the singular words for an
invertible substitution, and obt.ain some similar pl'Operties in a similar way.
Theorem 7. Let. TI, T2, be invert.ible subst.it.utions the following assert.ions are
equivalent.:
(1) M r , = M:~ for SOIllC k ~ 1;
(2) <l>r, = <I>~~), where <l>r, and <l>rl are the t.mce maps associated to Tl and TZ
(see the lecture of Peyricrc [10D, <1>(") is the 71th iteration of <1>;
(3) 3w E A", such tha.t. TI(a) = w-IT~«(J)W and TI(b) = w-IT;'(b)w,
438 Wen Zhi- Ying

(4) FT. and FTl are locally isomorphic.

Corollary Let TI, T2 be invertible substit.ut.ions, then
FT. and FTo are locally isomorphic implies that FT. and FTl have the same
energy spectrum (see lecture of Siit.o [15]).

Theorem 8. If T is an invert.ible subst.itut.ion then FT is St.urmian sequence,

i.e., let
nil (FT) = {w E F T, I wi = 11}, then nn (FT) = n + 1.
Comparison of properties of chains that are generated by various types of sub-
st.itution rules T.

Fibonacci T invertible T not invertible

Pisot property yes yes yes no
det Mr -I ± 1 ±I f; ±l f; ±l
average lattice and
bounded modulation yes yes yes yes no
cut and project.
(extension of at.omic surface =1) yes yes# no# no# no
trace map preservcs volulllc yes yes no no no
St.urmian sequence yes yes no no no

The properties labeled with # are conjectures.

4. FIBONACCI-CHAIN AS A PERIODIC CHAIN WITH DIS-

COMMENSURATIONS

Let F"" be t.he Fibonacci sequence. Now we split 1"00 int.o words of length 1m
from C(P",) and t.he lct.ter a (if 111 odd) or b (for III even) according t.o the
following rule, which we call t.he periodic approximation algorit.hm (of order
m). Suppose t.he first. J..: lct.t.ers have already been partit.ioned into words. If
XA·+IXH2 .. ';/;H1", belongs t.o C(F",), we add t.his word t.o the partit.ioning,
otherwise we add simply XA·+l.
For example for 111 2 = (1"2 = aba) t.he periodic approximation algorithm
yields t.he splitting

Foo = (aba)(aba)(baa)(boa)(b)(aba)(aba)(boa) ...

Now suppose we code the words frolll C( F",) in t.his split.t.ing by let.t.ers XI, X 2, ...

and b by t.he let.t.er Y.

E.g. if m. = 2, abo, ~ Xl, baa -+ X 2 , Ii ~ Y t.hen the coded sequence is
 SINGULAR WORDS 439

By using the singular words, we can prove a general theorem, which shows how
the coded sequence call be generated by a sllbstit,ution (1",. For the case m 2,=
we have

(12(X) = X4y
(12(Y) =X
x --+ aba, Y -t b.

This new chain can be seen as a modulation of the original quasiperiodic struc-
ture. Such ordered defects in the periodic chain which give the chain the same
incommensurahilit.y as the Fibonacci chain can be compared with discommen-
suration in incollllllensurable crystal phases (sec [16]).

Acknowledgement. This work was support.ed part.ially by NSFC of China for

the author and by the Dutch Stichting Fundamenteel Onderzoek der Materie
(FOM) for ",nY. The author t.hanks to Prof. .J. Peyriere and Prof. F. Axel
for their kind helps.

References

[1] J.P. Ailouche, Arithmetique et aut.omates finis, Asterisque, 147-148(1987)

13-26.
[2] J.P. Allouche and J. Peyriere, Sur une formule de recurrence sur les traces
de produit.s de matrices associees a. certaines substitutions, C.R. Acad.Sc.Paris
302 (II) (1986) 1135-1136.
[3] J. Uerstel, Mot. de Fibonacci, SClIlinaire d'informatiqlle theorique, L.I.T.P.,
Paris, Annee (1980/1981) 57-78.
[4] E. Dombieri alld J .E. Taylor, Which distribut.ion of mat.t.er diffract? An
init.ial investigation J. Physique, 47 (1987) 19-28.
[5] J. Karhumaki, On the cube-free w-words generated by binary morphism,
Discr. Appl. Math. 5 (1983) 279-297.
[6] M. K6sa, in 'Problems and Solut.ions', BATCS, Bull. 32 (1987) 331-333.
[7] F. Mignosi and G. Piriilo, Repet.it.ions in the Fibonacci infinite word, 11IJor.
Theo1·. ci A JIJI I. vol. 26, no. 3 (1992) 199-204.
[8] F. Migllosi et P. Seebold, Morphism('s st.urmienes et regie de Rallzy, J. de
ThCo1'ie des NOlllbres de Bordeaux 5 (1993) 221-233.
[9] J. Nielsen, Die Isomorphismcll del' freien gruppell, Math. Ann. 91 (1924)
385-397.
[10] J. Peyriere. This volume.
[11] J. Peyriere, Z.X. Wen et Z.Y. Wen, Polynomes associes aux endomor-
phismes des groupes libres, L 'Ellseigllemeni Mathematiquc, 39 (1993) 153-
177.
440 Wen Zhi-Ying

[12] M. QlIcffclec, SlIust.itu(.ion dynamical systems- spectral analysis, Lecture

Notes in Math. ]294 (1987), Springer-Verlag, Berlin.
[13] P. Seebold, Proprietes combinatoires des mots infinis ellgendre par certains
morphismes, These de doctoral" Rapp. Tec. L.I.T.P., 14 (1985).
[14] P. Seebold, Fibonacci morphisms and sturmian words, Theoret. Comput.
Sci, 88, (1991) 365-384.
[15] A. Siit.o, Schrodinger opera\.ors wit.h detcrminist.ic disorder. This volume.
[16] Z.-Y. Wcn, T. Janssen and F.IVI. Dekking, Fibonacci chain as a periodic
chain" with discommensllrations, J. Phys. A: Math.Gell. 27 (1994) 1-12.
[17] Z.-Y. Wen and Z.-X. Wen, The sequence of substitution and related topics,
Adtl. Malh. (China), 18 (1989) 123-145.
[18] Z.-Y. Well and Z.-X. Wen, SOllie st.udies of fact.ors of infinite words gener-
ated by invertible substit.utions, 5th Conference Formal Power Series and
Algebraic Combinat.orics, Florence, (1993) 455- 466.
[I9] Z.-X. Wen and Z.-Y. Wen, Local isomorphisms of invertible substitions,
C.R. Acad.Sci. Paris, 318, Scrie I, (1994) 299-304.
[20] Z.-Y. Wen and Z.-X. Wen, Some properties of the singular words of the
Fibonacci word, Eump. J. Combillatorics, 15 (1994).
[21] Z.-Y. Wen, F. Wijnands and J.S.W. Lamb, A natural class of generalized
Fibonacci chains, t.o appear in J. Phys. A: Math.Gen. (1994).
[22] F. Wijnantls, I~ncrgy spcctra aJl(I local isomorphism for one-dimensional
quasiperiodic potentials, J. Phys. A: Math. Gen. 22 (1989) 3267-3282.
 COURSE 15

Entropy In deterministic and random systems

V. Berthe

Laboratoire de Mathematiques discretes,

CNRS-UPR 9016,
Case 930, 163 avenue de Luminy,
F-13288 Marseille Cedex 9,
France

1. Introduction

The purpose of this course is to introduce some notions of entropy: entropy in

information theory, entropy of a curve and topological and measure-theoretic
entropies. For these last two notions, we will consider in particular dynamical
symbolical systems in order to present some measures of disorder for sequences.
We will allude then to the problem of classification of sequences with respect to
their spectral properties thanks to the entropy. For this purpose, we will inuo-
duce the sequence of block entropies for sequences taking their values in a finite
alphabet: we will then compute explicit ely the block frequencies (or in other
words, the measure of the associated dynamical system) for some examples of
automatic sequences (Prouhet-Thue-Morse, paper folding and Rudin-Shapiro
sequences) and for Sturmian sequences (these are the sequences with minimal
complexity among all non-ultimately periodic sequences; in particular, we will
consider some generalized Fibonacci sequences). But, in order to understand
the intuitive meaning of the notions of topological and measure-theoretic en-
tropies, we will begin by defining the Shannon entropy of an experiment.
442 V. Berthe

2. Thermodynamical entropy

The concept of entropy was fist introduced in thermodynamiC's by Clausius in

1854. In statistical mechanics, the entropy of a system is equal to the logarithm
of the number n of accessiblp microstates corresponding to a macroscopical
state of this system:
S = klnn,
where k is Boltzmann's constant. It is assumed here that all the microstates of
an isolated system are equiprobable. Hence the probability of the macroscopical
state increases with the number of corresponding microstates. Entropy is thus
a measure of the probability for a macroscopical state to be realized. The
Second Law of Thermodynamics, i.e. the increase of the entropy of an isolated
system, means that the evolution of the system is towards the state of maximal
probability, which is also the state corresponding to the maximal number of
microstates, that is to say the state of maximal disorder.
Let us suppose now that the isolated system is made of N indentical molecules.
Let us decribe then this system by counting the number N, of molecules in each
of k states in the phase space. There are n = Nl!N~: .. Nk! such ways to realize
this distribution. It can easily be shown, by using Stirling's formula, that the
entropy is thus equal, for large N, to:

S = -kN '"'
L.....,;P' logp" where p, = N,
N·

This is the kind of formula we will meet in the definitions of entropy below.

3. Information theory

Entropy is known in physics as a measure of randomness or disorder. It can also

be considered as a measure of information. Namely, consider an experiment,
say, the roll of a die. Randomness and uncertainty have the same measure.
We thus call entropy a measure of the uncertainty about the outcome of this
experiment. But the amount of uncertainty before the roll corresponds to the
amount of information one receives from the outcome of this roll. Therefore,
entropy is defined in information theory as a measure of the uncertainty about
the outcome of an experiment or equivently as a measure of the information
yielded by the happening of this experiment.
A first measure of entropy was introduced by Hartley in 1928 ([18]): he
considered distributions of equiprobable events. Wiener defined then, in 1948,
the entropy of a single event ([42]). Finally, Shannon introduced, also in 1948,
a measure of information for finite probability distributions. This is the notion
we will deal with mostly, but we will begin by the Wiener notion of entropy
which is the most intuitive way of measuring information. For introductions to
the notion of entropy in information theory, see for instance [1], [32], [37].
 ENTROPY 443

3.1. Entropy of a single event

Following Wiener ([42]), consider now the entropy Hw of a single event A with
probability p(A) =J. O. For instance, the event A can be "having an ace" , when
you roll a die. The entropy is here a measure of the information we receive
when the event A occurs (and also a measure of how unexpected the event A
is). We suppose that the entropy H depends only on the probability p(A) of
the event. The function H should hence have the following properties: it must
be positive (knowing something about an event is information) and additive
(the information yielded by the occurrence of two independent events is the
sum of the informations obtained from each experiment). We thus have:

1. Hw(p) 2': 0,
2. Hw(pq) = Hw(p) + Hw(q).
It is easily shown that the only functions which satisfy Properties 1 and 2 are
the functions defined by:

f(x) = -Alog(x), for all x E ]0,1]' with A 2': O.

These functions are defined up to a positive multiplicative constant. We will

hence normalize the entropy by assigning the unit value to the equiprobable
case, hence: Hw(I/2) = 1.
The quantity Hw(p) = -log2(p) is hence the only function which satisfies
all the required conditions and will be called the Wiener entropy of a single
event.
Let us note that the logarithm appears here again, in quite a natural way.

3.2. Entropy of an experiment

Consider now the entropy of an experiment E with outcomes AI, A 2 , .. , An of

probabilities PI,P2," ',Pn, with :EP. = 1. For instance, E is a roll of a die
with n faces and PI,P2,'" ,Pn are the probabilities of the different faces. We
will suppose here again that entropy only depends upon the probabilities.
Shannon defines in [38] the entropy H s of the experiment E as the average
value of the entropies of the single events AI, A 2, .. , An, weighted according to
their probabilities. Note that Shannon has obtained this definition axiomati-
cally by deriving it from "intuitive" properties that a measure of information
should have, like for the case of a single event.
Let
L(x) = { ~xlog2(x) ifxE]O,I],
for x = o.
We then have:
n n
Hs(E) = HS(PI,P2," ',Pn) = - LPk log2(Pk) = LL(Pk).
k=l k=l
444 v. BertM

The function H s is called the Shannon entropy of a finite probability distribu-

tion.
One can ask the reasons for the choice of base 2 for the logarithm. Such a
normalization corresponds to the choice of the unit of information: the binary
entropies will be measured in bits as the natural entropies measure information
in natural units or nats.

Remark The first measure of entropy, introduced by Hartley in 1928 ([18])

was the following: the Hartley measure h of the entropy of an experiment with
n outcomes is defined as HH(E) = logn. This entropy depends only on the
number of events and not upon their probabilities. For instance, Boltzmann's
formula corresponds to this conception. These two notions coincide if all the
states are equiprobable.

Consider now the case n = 2 (for instance, flipping a coin with a given coin).
We have
Hs(E) = H(p) = -plog2P - (1 - p) log2(1- p).
It is easily seen that the entropy of an experiment of outcomes of probabilities
o and 1 is equal to O. This seems rather natural because there is no uncertainty
about the outcome in this situation. The entropy is maximum in the equiprob-
ability case (p = 1/2), which corresponds to the case of maximal uncertainty.
We will see in the next section that this result holds generally for all finite
probability distributions.

3.3. Concavity of the function L

The function L is concave: for all Xk in [0,1] and all (AI,···, An) such that
Ai 2: 0 and LA. = 1, Jensen's inequality
n n
L(L: AkXk) 2: L AkL(Xk)
k=l k=l
holds. To check this, note that the function L has a negative second derivative.
In particular, we have by putting Xk = Pk and Ak = lin, for all k:

n
Note that L Pk = 1. We thus have the following property:
k=l
Hs (PI, ... , Pn) :::; log2 n.
This inequality means that the entropy is always smaller than the entropy of
the equiprobability case. This agrees with the fact that the uncertainty about
the outcome of an experiment is maximal when all the outcomes are equally
probable.
 ENTROPY 445

3.4. Marginal and conditional entropy

We have measured here the information obtained with a unique experiment.

Suppose now that we have two experiments E and F which are not independent:
let A 1 , A 2 , •• , An and B 1 , B 2 , .• , Bm be the possible outcomes ofthe experiments
E and F respectively, with probabilities P1, P2, ... ,Pn and q1, q2, ... , qm·
Suppose that we know the result of the experiment E. One can then ask
what the information yielded by F would be. It would be equal to
Hs(E, F) - Hs(E),
i.e. the information obtained by the event of both experiments (Hs(E, F))
minus the information yielded by the first (Hs(E)). Now, let qJk be the con-
ditional probability of Bk under the condition A J • It is easy to verify that:
n
Hs(E, F) - Hs(E) =L PJ Hm (qJ!, ... ,qJm),
J=1
m
with Hm(qJ1' %2,···, qJm) =LL(qJk). This quantity is called the conditional
k=1
entropy of the experiment F with respect to the experiment E. We denote it
by He{F/E). We thus have:
n
He{F/E) = LPJHm(%l,··· ,qJm),
J=1
and we obtain the following chain-rule:
Hs(E,F) = Hs(E) + He{F/E).
The entropy of the experiment E is called marginal entropy in contrast with
the conditional entropy. The chain-rule expresses that the joint entropy of two
experiments equals the marginal entropy of the first one plus the conditional
entropy of the second with respect to the first.
From the concavity of the function L, we deduce the following inequality,
with equality if and only if the experiments are independent:
He{F/E) :5 Hs(F),
or in other words that conditionality reduces entropy. This seems rather natural
because knowledge concerning the outcome of an experiment cannot increase
the uncertainty in the outcome of another experiment.
We deduce from this that the entropy of a joint event is smaller than or equal
to the sum of the individual marginal entropies; this last inequality is called
independence bound on entropy:
Hs(E, F) :5 Hs(E) + Hs(F).
Note that we have equality here if and only if He{F/ E) = Hs(F), i.e. if E and
F are independent.
446 V. BertM

3.5. Entropy of a finite curve

Let us see now how to apply these notions to curves and, in the next section,
to sequences.
Mendes France defines in [4] the dimension of a curve. But he also associates
a notion of entropy with the curves. He defines in [22], [23], [24] or [26], for
instance, the entropy of a finite curve r £ as:

+00
H(r£) = - LPnlogPn,
n=l

where Pn is the probability that an infinite straight line cuts r £ in exactly n

points. Thus, H (r £) is the amount of information yielded by the experiment
"an infinite straight line is drawn on the plane" and the outcomes are the num-
bers of intersection points. This entropy is hence a measure of the complexity
of plane curves.
A natural question to ask is whether this measure takes finite values, i.e.
whether there is an upper bound for the entropy. Thus, Mendes France has
shown the following result:

Theorem 1 We have:
2L (3
H(rd::; log(-C ) + - ( 3 - '
£ e-1

where (3 = log 2£~£CL' L is the length of r£ and C£ is the "perimeter" of the

curve, or in other words, the length of the convex hull of r £.

The function H(r£) = log(~:) + b is called the maximal entropy of the

curve and corresponds in a way to a topological entropy, a notion that we will
discuss later on.
Note that Mendes France gives the following thermodynamical interpretation
of the coefficient {3: he defines the temperature of a curve as T = ~. Thermo-
dynamical quantities can be defined too, for example pressure or volume of a
curve, which satisfy an equation of state and even an Heisenberg uncertainty
principle (see for instance, [24] or [26]).
Define now, still following Mendes France, the entropy of an unbounded curve
r:
. . H(r£)
H(r) = h m m
l fL'
-
£ ..... +00 og
where r £ is a finite portion of length L of the curve with the same origin.
Notice the normalization obtained by dividing by log L. We obtain then the
information per unit of length. Similarly, one defines the maximal entropy as:

A H(r£)
H(r) = liminf - I L·
£ ..... +00 og
 ENTROPY 447

vVe have:
0:::; H(r) :::; if(r) :::; l.
An infinite straight line, an exponential spiral of equation p = eO are of zero
entropy, i.e. deterministic. The spiral p = 00., where 0: > 0, has entropy 1;0.'
The dragon curve (see [4]) has entropy equal to if = ~, which is the highest
value for a self-avoiding curve. The spiral p = 10g(O + 1) is of entropy 1, i.e.
chaotic.
This notion of entropy is to be connected with the dimension of a curve
([4]); namely, the dimension d of a curve satisfies for a large class of them:
d::::: l!H' where H denotes the entropy. The meaning of this inequality is that
the dimension increases with the entropy, or in other words, that the entropy
and the dimension increase with the disorder of curves.

3.6. The sequence of block entropies

The purpose of this section is to introduce the block entropies for sequences
with values in a finite alphabet.
Let us recall that the frequency P(B) of a block B in a given sequence is
defined as follows: it is the limit, if it exists, of the number of occurrences of
this block in the first N letters of the sequence divided by N.
Let u be a sequence with elements with values in the alphabet A = {I, ... , d}.
We suppose that all the block frequencies exist for u. Let
P(Xl" ,xnx)
P(XIXI ... xn) = P( ) ,
Xl" 'Xn
where Xl ... Xn is a block of non-zero probability and X a letter. The intuitive
meaning of P(XIXI ... Xn) is that it is the conditional probability that the letter
X follows the block Xl ... Xn in the sequence u. We are going to associate with
the sequence u two sequences of block entropies (Hn)nEIN and (Vn)nEIN.
Let En be the experiment "choosing a factor of length n of the sequence" .
The outcomes are the factors oflength n with probabilities P(XI ... xn). Denote
the entropy of En by Vn . We have, for all n::::: 1:

Vn = LL(P(Xl" ·xn)),
where the sum is over all the factors of length n and with L(x) = -x logd(x),
for all X i- 0 and L(O) = O. Note that we change here the normalization by
taking the base d logarithm.
Let F be the experiment "choosing a letter of the alphabet A" and Hn be
the conditional entropy of F with respect to En. We have:

(1)
where the sum is over all the blocks of length n of non-zero probability and

H(Xl" ·xn) = L L(P(X/Xl" ,xn)).

xEA
448 V. BertM

Thus. Hn is the entropy of the next symbol when the preceding (n - 1) letters
are known, i.e. Hn measures the uncertainty about what the next symbol will
be, if we are told the preceding letters.
Now, let us apply to these two sequences the theorems we have seen in
information theory in order to deduce some basic properties for them. Let Ho
be the marginal entropy of F. We have:

Ho = L L(P(x)).
xEA

Obviously, Ho < 1. Thus, we obtain: 0 ::; Hn ::; Ho < 1. Furthermore we

clearly have
Hn = Vn+ 1 - Vn,
for all n, by putting Vo O. This property corresponds to the chain-rule.
Thus, (Hn)nEIN is the discrete derivative of (Vn)nEIN. Note that (Vn)nEIN is an
increasing sequence, since Hn 2: 0, for all n.
It can be shown that (Hn)nEIN is a monotonic decreasing sequence of n (see,
for instance [11]). The intuitive meaning of this is that the uncertainty about
the choice of the next symbol decreases when the number of known preced-
ing symbols increases; in other words, conditional entropy decreases when the
conditioning increases.
From the decreasing behaviour of the positive sequence (Hn)nEIN, we deduce
the existence of the limit lim Hn. We have Hn = Vn+l - Vn . By taking
n-++oc
n-l
Cesaro means, i.e. by considering (E Hk)/n, we obtain:
k=O

·
I1m H n= I·1m -Vn
n--++oo n--++oo n

We will see below, that this limit corresponds to the measure-theoretic entropy
of the dynamical system associated to the sequence u.
Finally, let us note that the sequence (Hn)nEIN of conditional block entropies,
measures in some way the properties of predictibiIity of the initial sequence u.
Other measures of predictibility are to be found in [12], [13], for predictions
with automata, in [35] for the notion of "noise" and also in a quite similar way,
in [25] and [4], for the opacity of an automaton.

These sequences have been first introduced by Shannon in [38], who wanted
to measure the entropy of the English language. Namely, consider a source
emitting a sequence of letters like, for example, a telegraph. If the letters are
independent and with the same probability, the entropy will be:

H = log2 26.

But if the source emits an English text, the letters do not come with the same
frequency (the letter E occurs more frequently than Q) and the probability
 ENTROPY 449

that a U. for instance, immediately follows a Q. is larger than the probahility

that an X comes after a Q. Therefore, Shannon has defined the entropy of nth
order of the English prose by putting:

Vn = - L P(C) log2(P(C)) = L L(P(C)),

where enis the family of all strings of n letters and P( C) denotes the probabil-
ity of the block C. Such a quantity corresponds to the entropy of an artificial
language approximating the natural language in the following way: the only
constraint that rules this language is that the probabilities (included the con-
ditional ones) are the same up to order n as those of the natural language.
Shannon gives examples of such approximations in [38].
Shannon estimates the entropy of the English language to be about 1. The
difference with the case of independent letters with same probabilities (H =
log226) comes from the redundancy of English, which is according to Shannon
roughly about 50%: half what we say or write is determined by the structure
of the language (see [38] for more details).

Burrows and Sulston have also studied these conditional block entropies se-
quences in 1991. One motivation could be to find a measure of disorder for
sequences which would be convenient to distinguish between sequences accord-
ing to their spectral properties (see [14]).
We will see in the next section some examples of computations of these
conditional block entropies for some automatic sequences and for the Sturmian
sequences.

4. Topological and measure-theoretic entropies

Let us introduce now two measures of disorder for sequences with values in a
finite alphabet: the topological entropy is defined from the notion of complex-
ity and the definition of measure-theoretic entropy uses the block frequencies.
We will then consider the general case of a dynamical system and give the
corresponding definitions.

4.1. Topological entropy of a sequence

The complexity of a sequence u is defined as the function p( n) which counts

the number of distinct blocks of length n which occur in this sequence (see [4]
or [3]). It is a combinatorial notion. The topological entropy ([2]) is then the
exponential growth rate of the complexity as the length increases:

H top () - l' logd(p(n))

U - 1m .
n-++oo n
450 V. BertM

It is easy to check that this limit exists because of the subadditivity of the
function log(p( n)) :

log(p(n + m)) :::; log(p(n)) + log(p(m)).

We obviously have: 0 :::; Htop(u) :::; ~ = 1.

Consider now the entropy of a substitutive sequence (see [4]). The complexity
satisfies: p( n) :::; Cn 2 , for all n, C being some constant. Thus the entropy of a
of substitutive sequence is zero. As a particular case, the entropy of a periodic
sequence is also zero.
Consider now a Bernoulli sequence, i.e. a sequence corresponding to the
d
Bernoulli scheme (PI, P2, .. " Pd), with p, =f- 0 and LP' = 1: the frequencies of
i=I
the letters are independent and given by the probabilites (PI, P2, .. " Pd), as for
instance, a sequence obtained by tossing a coin iteratively with a given coin.
Thus, the number of factors of length n is equal to d n and the entropy equals
l.
These examples show that the topological entropy cannot distinguish be-
tween periodic and substitutive sequences nor between Bernoulli sequences.

4.2. Measure-theoretic entropy of a sequence

We will hence put a metrical structure in order to make this measure of disorder
more precise. We will therefore consider block frequencies for sequences.
Let U be a sequence with values in A = {l"", d} and whose frequencies
exist for all blocks. We have seen that the sequences (Hn)nEI'< and (~)nEI'<
have the same limit H(u). This limit is called the measure-theoretic entropy of
the sequence u:
H(u) = lim Vn = lim Hn.
n--->+oo n n--->+oo

It is the limit of the entropy per symbol of the choice of a block of length n,
when n tends to infinity.
Some examples of computation of the measure-theoretic entropy are to be
find in the following.

4.3. Measure-theoretic entropy of a partition

Let us consider in this section the general case of a dynamical system and
give the corresponding definitions of metrical and topological entropies. Let
(X, T, J-l) be a dynamical system where X is a metrical compact set, J-l a proba-
bility measure and T a continuous invertible measure-preserving tranformation.
For more details, the reader is referred, for instance, to [16], [34], or [41].
Let P = {PI, ... , Pr } be a finite partition of X. For x EX, let k, be the
unique integer such that T' (x) belongs to Pk " Consider now the first n points
 ENTROPY 451

of the orbit of x under T:

x, Tx,"') Tn-I(x).

They belong successively to

It is convenient to define the name of x by: n( x) = (ko, ... , k n - I , .. l Thus, it

is a way of coding the trajectory of x under T.
Consider now the new partition

being the first terms of any name of x E X}.

Let us define the join of two partitions A = {AI, ... , Ar} and B = {BI' ... , Bs}
by:

Let us note that A V B is also a partition. We have obviously

pn =p V T- I P V··· V T-(n-I) P.

How to define the entropy of such a partition? We recall that p, is a measure

of probability. A partition can be seen as an experiment. We can therefore
consider the sets p. of the partition P as the outcomes of probability p,(P.) of
the experiment "determining to which set of the partition a point x belongs" .
The atoms in P and T- 1 P have the same measures. Hence the partition
T- 1 P corresponds to the replication of the experiment associated to P and the
partition pn corresponds to n replications. Naturally, the experiments P and
T- 1 P are not necessarily independent.
Hence, let us define the entropy of the partition P by:

E L(p,(P.)) .
r

H(P) =
• =1

Thus, the quotient H(:n) is the amount of information per unit of time about
the name of a point x of X, or in other words, the information per replication.
It can be shown easily that H(pn) is subadditive, so lim H(pn) exists.
n->+oo n
This limit, i.e.

H(P V T- 1 P V ... V T-(n-1) P)

H(P,T) = n->+oo
lim
n

is called the entropy of the transformation T with respect to the partition P.

452 V. BertM

But the choice of P can considerably reduce the uncertainty, for instance. if
P = {X}. Therefore, the measure-theoretic entropy of T is defined to be the
maximal uncertainty over all finite partitions P, Le.:
H(T) = supH(P, T). (2)
p

Thus, the entropy H(T) measures the uncertainty about the way T moves the
points of X.
Now, the question is how to compute the entropy of this system. Namely, the
expression (2) is not easy to use. But, a classical result shows that the supre-
mum in (2) is obtained for the partitions, which are generators with respect to
T: this is the Kolmogorov-Sinai Theorem (see [20] and [39]); a generator is a
partition such that if x =I- y, then n(x) =I- n(y), or in other words, such that the
codings under this partition, of the orbits of two distinct points are different.
Furthermore, the Krieger Generator Theorem [21] asserts the existence of a
finite generator if the transformation T is ergodic and if the entropy is finite.

We consider now the particular example of a symbolical dynamical system.

For more details about what follows, the reader is referred to [33] and [34]. Let
u be a sequence with values in A = {I,.··, d} and whose frequencies exist for all
blocks. Let T be the one-sided shift, i.e. (Tx)n = Xn+l. Let us suppose that u
is recurrent, Le. every factor of u appears infinitely often; thus T is onto on the
orbit closure Orb(u) of u under T in AN, endowed with the product of discrete
topologies. Furthermore, we suppose that T is injective (hence invertible) on a
set of full measure: this is the case, for instance, if the complexity of u is sub-
affine or ifp.. (n+l)-p.. (n) is bounded. Let J.L be the unique T-invariant measure
obtained by assigning to the cylinders their frequencies: J.L([w]) = pew). It is
easily seen that the partition P = {[I]'···, [d]} is a generator. Furthermore,
we have pn = {[B]; B block of length n} and we check that H(pn) = Vn .

4.4. Topological entropy of an open cover

Let us now consider the topological notion corresponding to a partition, namely

open covers of the compact space X, in order to define the topological entropy
of a dynamical system (see [2]). Let a be any open cover of X. Let us recall
that an open cover of the space X is a collection of open sets (O')'EI such that
Xc UiEIOi.
Let N(a) denote the number of sets in a finite subcover of a with smallest
cardinality. We denote the join of two covers as :
a v,B = {A n B; A E a, BE ,B}.
The topological entropy of T with respect to a is then defined as:
Ho:(P, T) = lim 10g(N(a V T-1a V ... V T-(n-l)a)) ,
n-+oo n
 ENTROPY 453

and the topological entropy of T is given by

Htop(T) = supH(o:, T), (3)
l>

where the supremum is taken over all open covers.

4.5. Variational principle

Let us come back to the sequences of conditional block entropies (Hn)nEN and
(Vn)nEN. What relation exists between these two sequences?
We have, for all k 2: 0: Hn = Vn +1 - Vn and Vo = O. We deduce from this
n-l
that L Hk = Vn . The sequence (Hn)nEN is decreasing. Thus we obtain that:
k=O
n-l
nHn::; LHk = Vn = LL(P(Xl'''Xn )),
k=O
By concavity of the function L, we have, for all n 2: 1: Vn ::; logd p( n). Thus,
we deduce the following proposition:
Proposition 1 We have Hn ::; logd~(n)), for all n 2: 1.
We hence have:
.
hm Hn =
.
hm
Vn ( ) ( )
- = H u ::; Htop u =
.
hm
logd(p(n)) .
n--++oo n--++oo n n--++oo n
This inequality is a particular case of a basic relationship between topological
entropy and measure-theoretic entropy called the variational principle (see, for
a proof, [27]): if T is a continuous map of a compact metric space then
Htop(T) = sup{H(T); for J-L being any measure invariant under T}.

The two limits lim Hn and lim logd(p(n)) are distinct, in general. But,
n-+oo n--++oo n
for instance, if the system (Orb(u), T) is uniquely ergodic (see [34]), we have
naturally equality between these two limits.
Here is a case where these two limits are distinct. Consider a sequence
corresponding to the Bernoulli scheme (p, 1 - p), with p =J 0 and p =J !.
We recall that the topological entropy equals 1. We have, for every word B:
P(B) = p/B/a (1 - p)/B/b, where IBlx denotes the number of occurrences of the
letter x in the word B. We thus have, for all n 2: 1:

Vn = L(P) + L(1 - p) = -plog2P - (1 - p) log2(1 - p).

n
Thus, we obtain, as expected:

lim Hn = L(p) + £(1- p) < 1 = lim log2(p(n)).

n--++oo n--++= n
454 v. BertM

The notion of measure-theoretic entropy for a sequence seems consequently to

be more precise. But we will see in the next section that in the cases we deal
with mostly, we consider deterministic sequences, i.e. sequences with zero
entropy. Determinism means that there is no uncertainty in the choice of the
next letter, in other words, past determines future. Therefore neither metrical
nor topological entropy can distinguish between these sequences. That is why
we will consider in the sequel the rate of convergence of the sequence H n towards
its limit (the measure-theoretic entropy) and not only this limit.

5. Entropy and spectral properties

Let us now consider the question of classifying dynamicals systems up to iso-

morphism. We will not give detailed definitions here but they can be found for
instance in [34]. Let us recall that roughly, a dynamical system is said to be
a factor of another dynamical system if the first one can be "constructed" in
the second one, i.e. if there exists a map from the first one into the second one
preserving measure and transformation.
How does entropy react with factorisation? Intuitively, one can see that
there is a loss of information. In fact, it can be shown easily that entropy
decreases with factorisation (see for instance [11]). If two dynamical systems are
metrically isomorphic, then each of them is a factor of the other. Hence entropy
is an isomorphism invariant. Two isomorphic dynamical systems are spectrally
alike, i.e. they are spectrally isomorphic for the structure of Hilbert space,
which can be associated to a dynamical system, as explained in [34]. Hence,
a spectral invariant (like ergodicity or mixing properties) is an isomorphism
invariant. A natural question then arises: can we find non-spectral invariants?
The measure-theoretic entropy introduced by Kolmogorov to this effect in 1958
(see [20]) allows us to answer this question: we have seen that the entropy is
an isomorphism invariant; so, two Bernoulli schemes with different entropies
are not isomorphic but are always spectrally isomorphic. Hence a necessary
condition for two Bernoulli schemes to be isomorphic is that they have the
same entropy. In fact, the converse is true. This question remained open a long
time but Ornstein solved it in 1970, by giving in [30] a complete classification
of the Bernoulli shifts up to isomorphism:

Ornstein's theorem Two Bernoulli schemes are isomorphic if and only if

they have the same entropy.

In the case of discrete spectrum, the classification is relatively easier: we have

equivalence between metrical isomorphism and spectral isomorphism. Namely,
Halmos and Von Neuman showed in 1942 that the eigenvalues allow us to say
whether two ergodic transformations with discrete spectrum are isomorphic or
not (see [17]).
 ENTROPY 455

Discrete spectrum classification Theorem Two ergodic measure-preserving

transformations with discrete spectrum are spectrally isomorphic if and only if
they have the same eigenvalues. Furthermore, if they are spectrally isomorphic
then they are isomorphic.

In the case of discrete spectrum, the entropy is equal to O. Thus the entropy
is not convenient to distinguish between systems with low disorder.
Note that the entropy is in general equal to 0 in the examples we deal with
mostly. We have namely the following property(see for instance, [31] or [32]):

Zero entropy If the entropy of an invertible measure-preserving transfor-

mation T is strictly positive, then T has countable Lebesgue spectrum.

Thus, we have the following corollary:

Corollary 1 If T is of finite multiplicity or of continuous singular spectrum
or of discrete spectrum, then the entropy of T is equal to O.
Therefore, either low complexity sequences (like substitutive sequences, for
instance) or the examples of transformations given by M. Queffelec in [34], as
irrational rotations of the circle, q-odometers, or the Chacon transformation,
or also, Besicovitch almost-periodic and mean almost-periodic sequences ([4])
have zero entropy.

6. Some examples of computation of block entropies

We will see in this section that one can compute explicitely the block frequencies
and consequently the block entropies (Hn)nEIN, defined in the section 2.6, for
some examples of automatic sequences (Prouhet-Thue-Morse, paperfolding and
Rudin-Shapiro sequences) and for Sturmian sequences: these are the sequences
with minimal complexity among all non-ultimately periodic sequences (see for
instance, [4] or [41]); in particular we will consider some generalized Fibonacci
sequences. We will finally address the following question: given a sequence u,
can we deduce from the rate of convergence of the sequence of conditional block
entropies whether an atomic structure associated to u is "quasicrystalline" or
not? Or, in other words, can we deduce from the block entropies, spectral
properties of the initial sequence?
This question was put forward by Burrows and Sulston (1991) who have in-
troduced this measure of disorder in the study of quasiperiodic structures. By
computing the first and second order entropies HI and H2 for the Prouhet-
Thue-Morse sequence and for some generalizations of the Fibonacci sequence.
they have obtained the following comparison of disorder: among the sequences
they have studied, the sequences which are quasiperiodic (or of discrete spec-
trum) have entropy of first and second order lower than those which are not
purely discrete. But these entropies HI and H2 are not sufficient. for instanc£',
to distinguish between the Rudin-Shapiro sequence and a normal s£'quence. i.e.
456 v. BertM

a sequence such that all blocks of same length have the same frequency. Thus it
is interesting to obtain entropies of all orders and to compare them. Therefore
we need to compute the block frequencies.
M. Queffelec has shown how to obtain the block frequencies of a substitutive
minimal sequence by using the matrix of the associate primitive substitution
and the Perron-Frobenius Theorem (see [34] and for more details, [33]). We
will deduce here the block frequencies of all orders from a finite number of small
length block frequencies, by using recurrence formulas between the frequency
of a block and the frequencies of its pre-images by the substitution. But this
method does not work for Sturmian sequences, because they are generally not
substitutive. The idea here, due to Dekking [15], will be to use the Rauzy
graph of words [36], which we define in the following. Dekking obtains in [15],
a precise description of the frequencies of the words occurring in the Fibonacci
sequence.

6.1. Ultimately periodic and "random" sequences

Consider first the following two extreme cases: the case of minimal disorder,
i.e. the case of ultimately periodic sequences, or in other words, of sequences
which are periodic from some index on, and the case of maximal disorder, i.e.
the case of "random" sequences. Let us note that it is the same thing, in terms
of frequencies, to consider ultimately periodic sequences and purely periodic
sequences.
The following result can easily be shown (see [8]):

Proposition 2 Let u be a ultimately periodic sequence of period n. We have:

Hk = 0, for all k ~ n.
Namely, there is no uncertainty at all in the choice of the next letter. The
converse is not true. Suppose, for instance, that the frequencies of the letters
are equal to 0 or 1. We then have Ho = o. The sequence (Hn)nEIN being a
decreasing sequence, we obtain H n = 0, for all n.
But if the sequence is minimal, we obtain the following property:

Proposition 3 Let u be a minimal sequence such that Hko = 0 for some inte-
ger k o. Then u is a periodic sequence of period p(ko ), where p(ko) denotes the
complexity of order k o·

The proof of this statement comes from the fact that the frequencies are strictly
positive in a minimal sequence.

Consider now a "random" sequence or in other words, a normal sequence:

all the blocks of given length have the same frequency. Hence the conditional
probabilities P(x/ B) are equal and Hn = 1, for all n ~ o. It can easily be
shown, by using (1) that the converse is true. Thus, we have the following
proposition:
 ENTROPY 457

Proposition 4 We have Hn = 1 for all n. if and only if the sequence u is a

normal sequence.

Therefore, in these two extreme cases, the sequence (Hn)nEIN gives a charac-
terization of the ultimately periodic and the "random" sequences.

6.2. Sturmian sequences

The Sturmian sequences are the sequences with minimal complexity among
all non-ultimately periodic sequences (see [4]). Thus, it is rather interesting
to measure the disorder of such sequences. The Fibonacci sequence and the
generalized Fibonacci sequences which are defined as the fixed points of the
substitutions: a(a) = anb et a(b) = a, with n ~ 1 are some examples of
Sturmian sequences. Let us recall that a Sturmian sequence is the itinerary of
the orbit of a point p of the unit circle under a rotation of irrational angle a,
with respect to complementary intervals of length a and 1 - a of the unit circle
(see [28] and [29]). We have the following result:

Theorem 2 Let u be a Sturmian sequence with angle a. Let m be greater than

1. Let El.
'II and £1
q2
be two successive m- Farey points such that: El.
ql
< a < £1.
q2
The frequencies of the blocks of length m are the following:

More precisely, there are

• m - q2 + 1 blocks of length P2 - aq2,

• m - ql + 1 blocks of length aql - PI
• and (ql + q2) - m - 1 blocks of length a(ql - q2) + P2 - Pl.

Therefore the values of the conditional block entropies satisfy:

H m - l = L(P2 - aq2) + L(aql - pd - L(a(ql - q2) + P2 - pd,

with L(x) = -x log2(x).

Let us recall the definition of an m-Farey point: it is an element of the

interval [0, 1] of the form ~, where p ~ 0, 1 ~ q ~ m and gcd(p, q) = 1. Two m-
Farey points El.qlI et £1
q2
are successive if and only if: P2ql - Plq2 = 1. The points
El. and £1
q 2 depend on m and can be given explicitly, by using the continued
2
91
traction extension of the angle a.
This theorem can be proved either by using the combinatorial definition of
the Sturmian sequences (P( n) = n + 1) or the dynamical characterization (Stur-
mian sequences are irrational rotations). It is easily seen that block frequencies
correspond to intervals of the unit circle. This theorem is thus another formu-
lation of the 3-distance Theorem (see, for instance [40]):
458 V. BertM

3-distance Theorem Let a be any irrational number. Let us place the

points {a}, {2a}, ... , {na}, where {a} denotes the integral part of a, on the
line segment [0,1]. The (n + 1) segments found have at most three different
lengths, one being the sum of the two others.

The tool of the combinatorial proof here is the Rauzy graph of words. The
Rauzy graph of words of length n associated to a sequence is the orientated
graph whose nodes are the factors of length n of the sequence and whose arrows
are defined as follows: there is an arrow from U to V, if there exists a word W
such that

U = xW and V = Wy, with x,y E {a,b} and xWy is a factor of the sequence.

Suppose now that we have a Sturmian sequence. From the complexity

(p(n) = n + 1), we deduce that there is only one factor of length n with
two right extensions. Let us call this factor Rn. We define, similarly, Ln as the
factor of length n with has two left extensions. Thus, the graph of words of
given length has the two following forms, according as Ln = Rn or Ln =1= Rn:

J
1 1

( 2 C ·
J
Gn _ _ _ _~_----- Dn Gn =Dn

l __:f----) ( _____
3 __ _

Let U be a node of the graph. Let U+ denote the number of arrows which
come to U and U- the number of arrows which go out U. We have the following
lemma:
Lemma 1 If u+ = 1 and V- = 1, then the words U and V have the same
frequency.
Namely, the factor U has only one right extension, that we denote x, and
similarly, the factor V has only one left extension y. Therefore, we have the
following equalities between the frequencies:

P(U) = P(Uy) = P(xWy) = P(xV) = P(V).

We deduce in particular from this lemma that all the words of the branch
(1), except Ln and Rn, have the same frequency, that similarly, all the words
of the branch (3), except Ln and Rn, have the same frequency and finally that
all the words of the branch (2), Ln and Rn included, have the same frequency.
 ENTROPY 459

Thus we obtain that the frequencies of the factors of given length of a Sturmian
sequence take at most three values.
More precisely, Theorem (2) can be shown by studying the lengths of the
branches of the graph, whose evolution is precisely described in [7].

6.3. Block frequencies for some automatic sequences

Let us consider now some automatic sequences which are of distinct spectral
types (see, for instance, [4] or [34]):
• the Prouhet-Thue-Morse sequence, which has singular continuous spec-
trum,
• the Rudin-Shapiro sequence, which has Lebesgue spectrum,
• the paperfolding sequence, which has discrete spectrum.
Let HT, HR and H P be respectively the sequences of conditional block en-
tropies for the Prouhet-Thue-Morse, the Rudin-Shapiro and the paper folding
sequences. We expect the following inequality between HT, HR and H P :

H t: ::; H;: ::; H:;, for every n,

or in other words, we expect, for instance, from the paperfolding sequence to

show more order than the Prouhet-Thue-Morse sequence with respect to this
particular measure of disorder.
Let us recall that these three sequences are deterministic, that is to say
of zero entropy, so the sequence (Hn) converges towards 0 for each of theses
sequences.
We have the following result (see [9] and also [15], for the Prouhet-Thue-
Morse sequence):

Proposition 5 The frequencies of blocks of length m, with 2k + 1 ::; m ::; 2k+1,

take the following two values:

• 3.1k, ~, for the Prouhet- T'hue-Morse sequence, if m 2: 2,

• ~, 16~2k' for the Rudin-Shapiro sequence, ifm 2: 8,
• ~, S.1k' for the paperfolding sequence, if m 2: 7.
We can deduce from this, the expression of the conditional block entropies:
Proposition 6 Let H'!:, H:; and Ht: be respectively the conditional block en-
tropies for the Prouhet- Thue-Morse, the Rudin-Shapiro and the paperfolding
sequences. We have:

n =
• HT 3.2k, J'lor
_4_ 2k +1 <
- m <
- 3.2k-l ,
HT
n
= _2_ lor
3.2k' J'
3.2 k - 1 +1<
-
m<
-
2k+1 and k >
-
1.
• H:; = f,., for 2k + 1 ::; m ::; 2k+1 and m 2: 8,
460 V. BertM

• H nP = -.L
2"' J'lor 2k -< m <
-
2k + 1 - 1 and m -> 7.

We can notice the following relationship between H;; and H:::

Proposition 7 We have H:: = H;;+v for all n.

We will not give a detailed proof but the methods used here in the compu-
tation of the frequencies are similar to those used in the computation of the
complexity (see [4]): we use for the Prouhet-Thue-Morse sequence the fact that
each word of length greater than 4 comes from a unique word by the substitu-
tion (which is called the pre-image). For instance, the word aba comes from the
words ab and ba but baba comes only from the word bb. The substitution is here
of length 2. Hence, the number of occurrences of a word (of length greater than
4) in the first 2n letters of the sequence is equal to the number of occurrences of
its pre-image in the first n letters. We deduce from this that the frequency of a
block is equal to half the frequency of its unique pre-image by the substitution.
It is easy then to compute the frequencies by using this induction formula. The
same situation occurs for the fixed point of the substitution which generates
after a letter to letter projection, the Rudin-Shapiro sequence. Furthermore we
have a bijection between the factors of length greater than 8 of the fixed point,
and the factors of same length of the projection: namely, these two sequences
have the same complexity function for n greater than 8 (see [4]). Hence, we
can deduce using this bijection, the frequencies in the Rudin-Shapiro sequence
from the frequencies in the fixed point. For the paper folding sequence the same
situation holds: we have a bijection between the factors of length greater than
7 of the fixed point and the factors of same length of the projection (see [4]).
But there is a slight difference concerning the properties of "recognizability" of
the fixed point: some words can arise from two different words by the substi-
tution. In fact, these words which have more than one pre-image are exactly
the special (or expansive) ones, that is to say the words which have at least two
right extensions in the sequence. Despite this difficulty, it is possible to find
here again recurrence formulas.

Remark This method works also for the generalized Rudin-Shapiro se-
quences which count the number of occurences of the pattern 1 * ... * 1 in
the binary expansion of every integer (see [5] and [6]). We obtain, if d is the
length of the pattern * ... *, that the conditional block entropies are ultimately
equal to 2d times the corresponding entropies of the classical Rudin-Shapiro
sequence.

6.4. Conclusion
Let us come back to the initial question of the comparison of block entropies for
these sequences. It can be seen, by computing the first values of the conditional
block entropies, that we have inequalities between the values of conditional
 ENTROPY 461

block entropies of the same order for the first values. We have namely: HP ~
HT ~ HR, for n ~ 8. But, for n ~ 9, this ordering does not hold. In particular,
we have:
Hf = H: = 1/8 < H'f = 1/6.
From Proposition 7, we deduce that H! ~ H;; and that for almost n, this
inequality becomes an equality. More precisely,

P = !!.!
• for n = 2k , we have H n ·2 '
• but for n =I- 2k, we have H! = H;;.
Furthermore, we see that there is a kind of shuffle between the values of HR
(and consequently of HP) and the values of HT: .
• for 2k + 1 ~ n ~ 3.2 k - 1, we have: H;: = ~H;; = ~H::,
• and for 3.2k-l + 1 ~ n < 2k +1, we have: H;: = ~H: = ~H::.
In particular, these three sequences of conditional block entropies converge with
the same rate towards O.
We conclude from this that this measure of disorder cannot allow us to
distinguish between deterministic sequences even if they have different spectral
properties.

Acknowledgments

We would like to thank J.-P. Allouche for his many useful comments and Michel
Mendes France who read carefully a previous version of this paper.

References

[1] Aczel J. and Dar6czy Z., On measures of information and their charac-
terizations (Mathematics in science and engineering, Academic Press 115,
1975).
[2] Adler R. L., Konheim A. G. and McAndrew M. H., Topological entropy,
Trans. A mer. Math. Soc. 114 (1965) 309-319.
[3] Allouche J.-P., Sur la complexite des suites infinies, Bull. Belg. Math. Soc.
1 (1994) 133-143.
[4] Allouche J.-P. and Mendes France M., this volume.
[5] Allouche J.-P. and Liardet P., Generalized Rudin-Shapiro sequences, Acta
Arith. 60 (1991) 1-27.
[6] Allouche J.-P. and Shallit J., Complexite des suites de Rudin-Shapiro
generalisees, Journal de Theorie des Nombres de Bordeaux 5 (1993) 283-
302.
[7] Arnoux P. and Rauzy G., Representation geometrique de suites de com-
plexite 2n + 1, Bull. Soc. math. France 119 (1991) 199-215.
462 V. Berthe

[8J Berthe V., Fonctions de Carlitz et automates. Entropies conditionnelles

(These, Univ. Bordeaux I, 1994).
[9J Berthe V., Conditional entropy of some automatic sequences, J. Phys. A:
Math. Gen. 27 (1994) 7993-8006.
[lOJ Berthe V., Frequences des facteurs des suites sturmiennes, Preprint 94.37
Laboratoire de Mathematiques discretes.
[l1J Billingsley P., Ergodic theory and information (John Wiley and Sons, New
York, 1965).
[12J Broglio A., Prediction par automates (These, Aix-Marseille I, 1991).
[13J Broglio A. and Liardet P., Predictions with automata, Contemporary
Mathematics, Amer. Math. Soc. 135 (1992).
[14J Burrows B. L. and Sulston K. W., Measures of disorder in non-periodic
sequences, J. Phys. A: Math. Gen. 24 (1991) 3979-3987.
[15J Dekking F. M., On the Prouhet-Thue-Morse Measure, Acta Universitatis
Carolinae, Mathematica et Physica 33 (1992) 35-40.
[16J Dekking F. M., this volume.
[17J Halmos P. R. and Von Neumann J., Operator methods in classical me-
chanics II, Ann. Math. 43 (1942) 332-350.
[18J Hartley R. V., Transmission ofInformation, Bell System Tech. J. 7 (1928)
535-563.
[19J Keane M., "Ergodic theory and subshifts of flnte type", Ergodic Theory,
Symbolic Dynamics and Hyperbolic Spaces, T. Bedford, M. Keane, C.
Series (Eds) (Oxford University Press, 1991) pp. 35-70.
[20J Kolmogorov A.N., New metric invariants of transitive dynamical systems
and automorphisms of Lebesgue spaces, Dokl. Akad. Nauk SSSR 119
(1958) 861-864.
[21] Krieger W., On entropy and generators of measure-preserving transforma-
tions, Trans. Amer. Math. Soc. 149 (1970) 453-464, Erratum: 168 (1972)
519.
[22] Mendes France M., "Chaotic curves", Rythms in Biology and Other Fields
of Applications, Proc. Journ. Soc. Math. France, Luminy 1981 (Lecture
Notes in Biomathematics 49, Springer Verlag, 1983) pp.352-367.
[23] Mendes France M., "Dimension and entropy of regular curves", Fractals,
non-integral dimensions and applications, G. Cherbit Ed. (John Wiley and
Sons, 1990) pp. 222-230.
[24] Mendes France M., Folding paper and thermodynamics, Phys. Rep. 103
(1984) 161-172.
[25] Mendes France M., Opacity of an automaton. Application to the inhomo-
geneous Ising chain, Comm. Math. Phys. 139 (1991) 341-352.
[26] Mendes France M., ''The Planck constant of a curve", Fractal geometry
and Analysis, J. Belair and S. Dubucs Eds. (1991) pp. 325-366.
[27] Misiurewicz M., " A short proof of the variational principle for a 7l~ action
on a compact space", Int. Conf. Dyn. Systems in Math. Physics, Societe
Mathematique de Prance, Asterisque 40 (1976) 147-158.
[28] Morse M. and Hedlund G. A., Symbolic dynamics, Amer. J. Math. 60
(1938) 815-866.
 ENTROPY 463

[29] Morse M. and Hedlund G. A., Symbolic dynamics II: Sturmian trajecto-
ries, Amer. J. Math. 62 (1940) 1-42.
[30] Ornstein D. S., Bernoulli shifts with the same entropy are isomorphic,
Adv. in Math. 4 (1970) 337-352.
[31] Parry W., Topics in ergodic theory (Cambridge University Press, 1981).
[32] Petersen K., Ergodic theory (Cambridge studies in advanced mathematics;
2, 1991).
[33] Queffelec M., Substitution dynamical systems. Spectral analysis (Lecture
Notes in Mathematics 1294, Springer-Verlag, 1987).
[34] Queffelec M., this volume. .
[35] Rauzy G., Nombres normaux et processus deterministes, Acta Arithmetica
14 (1976) 211-225.
[36] Rauzy G., Suites a termes dans un alphabet fini, sem. de Theorie des
Nombres de Bordeaux (1983) 25-01-25-16.
[37] Roman S., Coding and information theory (Springer-Verlag, 1992).
[38] Shannon C. E., A mathematical theory of communication, Bell System
Tech. J. 27 (1948) 379-423, 623-656.
[39] Sinai Ya. S., Flows with finite entropy, Dokl. Akad. Nauk SSSR 125 (1959)
1200-1202.
[40] Slater N. B., Gaps and steps for the sequence nO mod 1, Proc. Cambridge
Phil. Soc. 63 (1967) 1115-1123.
[41] A. A. Siito, this volume.
[42] Wiener N. Cybernetics, or Control and Communication in the Animal
and the Machine (Hermann, Paris; Technology Press MIT, Cambridge,
Massachussets; Wiley, New York, 1948).
 COURSE 16

Trace maps
.J. Peyricre

Universite Paris-Sud
MatMmatiques, bat. 425
Unite associee au CNRS 757
91405 ORSAY CEDEX

1. INTRODUCTION

In several problems arising in the quasicrystal theory one is faced with the
problem of computing as simply as possible the traces of matrices defined by
recursion using a substitution scheme. To be more specific, let us take an
example.

Given two 2x2-matrices A and B, define A J and BJ by recursion in the

following way
Ao =A,
Bo =B,
A J+1 = AJAJBJ ,
B J+ 1 = BJAJ ,

how to compute tr A J , the trace of A J ?

Of course, one could think of operating by brute force, by just expressing

the eight entries of A J + 1 and B J + 1 in term of those of A J and BJ , and then
getting, by elimination, a recursion relation linking nine successive values of
tr A J . As a matter of fact, on the one hand, this method is not so bad: had
we considered a recursion involving n matrices, we should have obtained a
recursion relation, the length of which grows linearly in n, for the traces. On
the other hand, eliminating variables could be an untractable operation, even
466 J. Peyriere

when using a. ('omputt'J' 'llgebra software. Besides, this Illethod gives no ideas
about the algebraic properties of these recursions.
Another way of operating, which is developed here, is to exploit polynomial
identities in rings of matrices. This will provide an effective algorithm for con-
structing such recursion relations for traces, the so called trace maps. Besides,
these trace maps exhibit very interesting algebraic properties.
These trace maps have been widely used and studied from the point of view
of iteration. But applications as well as the dynamical properties of trace maps
are not within the scope of this course.

2. SOME IDENTITIES FOR 2x2 - MATRICES

In this section upper case letters will stand for 2x2 - matrices the entries of
which are complex numbers. The basic identity is given by the Cayley-Hamilton
theorem:
A2 - (tr A)A + I detA = 0 (1)
where I is the identity matrix of order 2.
As a consequence, one has
An = Pn(tr A, detA)A + qn(tr A, det A)I,
where Pn and qn are polynomials, independent of A, with integer coefficients. If
A is invertible, such a formula is also valid for negative n. The polynomials Pn
and qn are closely related to the Chebyschev polynomials of the second kind.
One has det A = >.p, = [(>. + p,)2 - (>. - p,)2] /2, if>. and p, are the eigenval-
ues of A. Therefore the Cayley-Hamilton relation can be rewritten as
1
A2 - A tr A + "2 [(tr A)2 - tr A2] I = o.
This form allows bilinearization: writing this formula for A, B, and A + B,
one gets
AB + BA = tr AB - (tr A)(tr B) + A tr B + Btr A (2)
(we dropped the identity matrix I as, from now on, we identify scalars and
scalar matrices).
We now deal with complex matrices having determinant 1 (i.e. elements of
8L(2, <C» for the sake of simplicity.
Proposition 1. Ifml,nlom2,n2," ',mk,nk is a sequence of integers, there
exist four polynomials p, q, r, and s in three variables with integer coefficients
such that, for any pair of matrices A and B in 8L(2, <C), one has
AmI Bnl A7n2 B n2 ... Amk Bnk = p(tr A, tr B, tr AB)
+ q(tr A, tr B, tr AB) A
+ r(tr A, tr B, tr AB) B
+ 8(tr A, tr B, tr AB) AB
 TRACE MAPS 467

Proof Using rcpeatedly the Cayley-Hamilton theorem for A and B, one is left
with a linear combination with polynomial coefficients of f, A, B, AB, BA,
and of products of the form ABA· .. or BAB . ". Then by using the Cayley-
Hamilton theorem for AB and BA, one is left with a linear combination of f,
A, B, AB, BA, ABA, and BAB. We conclude by using formula (2) and again
the Cayley-Hamilton theorem for A and B.

Corollary. Given Tn'S Rnn n's as in proposition 1, there exists R unique poly-
nomial P in three variahles with integer coefficients such that, [or any pair
(A, B), one has

Proof The existence results from the above proposition. The uniqueness fol-
lows from the fact that (tr A, tr B, tr AB) can assume any value (x, y, z). To
see this, just take

A = (~1 ~ ) and B = (z ~ A~)

with A(Z + A) + 1 = O.
Remark. Had not we restricted the determinants of A and B to be 1, the trace
of the product above would have been expressed as a polynomial in the five
variables tr A, tr B, tr AB, detA, and detB.

Examples. A and B stand for elements of SL(2, <C), x = tr A, y = tr B, and

z = tr AB.

1) Using formula (2), A(AB + BA) = (z - xy)A + yA2 + xAB. Then

ABA = (z - xy)A + y(xA -1) +xAB - (xA -l)B. Finally, we get the formula

ABA = zA+B-y (3)

which will be useful later.

2)
(AB - BAf = (AB)2 + (BA)2 - AB2 A - BA2 B
= z(AB + BA) - 2 - yABA - xBAB + A2 + B2
= z(z - xy + yA + xB) - 2 - y(zA + B - y)
- x(zB + A - x) + xA + yB - 2
= x 2 + y2 + Z2 - xyz - 4

This result is not surprising because tr(AB - BA) = O. The polynomial

A(X, y, z) = x 2 + y2 + Z2 - xyz - 4 (4)

468 J. Peyriere

will play an important role in the sequel. The above formula says that the
determinant of AB - BA is -~(x, y, z). This is an easy exercise in linear
algebra to show that det(AB - BA) = 0 if and only if the matrices A and B
have a common eigendirection.

Therefore A and B have a common eigendirection if and only if ~(x. y, z) = o.

3)

ABA- l B- 1 = AB(x - A)(y - B)

= (AB)2 + xyAB - yABA - xAB2
= zAB - 1 + xyAB - y(zA +B - y) - xA(yB -1)
= zAB + (x - yz)A - yB + y2 - 1

Therefore tr ABA-l B-1 = ~(x, y. z) + 2. This is the Fricke formula.

We now turn our attention to formulae involving more than two elements of
5L(2, <c). As a consequence of formula (2), we have the following proposition.
Proposition 2. If {AJ}l<_J_ <n are elements of 5L(2, <c), then
1° any product constructed from these matrices or their inverses, can
be written as a linear combination of the 2n matrices Ail At2 ... Atk (0 ~ k ~ n,
il < i2 < ... < ik) e) the coefficients of which are polynomials in the 2 n - 1
variables tr A l A 2 •• ·Aik (1 ~ k::; n, i l < i2 < ... < ik),
'Z' the trace of such a product can be expressed as a polynomial with
integer coefficients in the 2n - 1 traces defined above.
We now turn to formulae which involve three matrices AI, A 2, and A3 in
5L(2, <c). Let Xl, X2, X3, Yl, Y2, and Y3 denote the traces of AI, A 2 , A 3, A 2A 3,
A3Al' and A l A 2.
Define the following polynomials:

(5)

q(X, Y) = x~+x~+x5+y~+y~+y~-XlX2Y3-X2X3Yl-X3XlY2+YlY2Y3-4 (6)

where X stands for the collection of x's and similarly for Y.
Proposition 3. One has
1° tr(AlA2A3) + tr(AlA3A2) = p(X, Y)
2° tr(AlA2A3) tr(AlA3A2) = q(X, Y)
Proof. To prove assertion 1°, write AlA2A3 + AlA3A2 = Al (A2A3 + A 3 A 2)
and use formula (2).

e) Of course, for k = 0, this product should be interpreted as the identity matrix.

 TRACE MAPS 469

To prove assertion 2°, write [AI (A2A3) All A3A2 = AIA2 (A3AIA3) A2 and
use formula (3) twice. We obtain

By reducing further, we get

AIA3A2 tr(AIA2A3) = -2 + x5 + Y~ - YIX2X3 + (Xl - Y2X3)AI + X2 A2

-(X3 - YIX2)A3 + (YIY2 + Y3 - XIX2)A IA2 + Y2 A IA3 - yl A2A3 ,

from which assertion 2° follows by taking the trace.

Corollary. tr(AIA2A3) and tr(AIA3A2) are the roots of the equation

Z2 - p(X, Y)z + q(X, Y) = o.

This leads to define a polynomial in seven variables

A(X, Y, z) = z2 - p(X, Y)z + q(X, Y) (7)

The preceding corollary means that variables x's, y's, and z = tr(AIA2A3)
are not independent. Indeed the set of polynomials P in seven variables with
integer coefficients such that, for any triple {A J h<_J_<3 of elements of SL(2, <C),
P(trAI,trA2,trA3,trA2A3,trA3AI,trAIA2,trAIA2A3) = 0 is an ideal (2)
containing A. It can be shown that this ideal is generated by A. Therefore, the
polynomial the existence of which is asserted in proposition 2-20 is not unique
when n > 2. In the case n = 3 it is defined up to a multiple of A.

Proposition 4. Let z stand for tr A I A 2A 3. One has

2AIA2A3 = Z - XIYI - X3Y3 + XIX2 X3 + (YI - X3X3)AI - Y2A2

+(Y3 - XIX2)A3 + X3 A I A2 + x2AIA3 + x 1 A2A3 .

Proof In Al (A 2A 3), commute Al and A2A3 by using formula (2). We get,

among other terms, -A2A3AI. By using formula (2) twice, on can make Al to
jump over A3 and A 2. So AIA2A3 can be written as -AIA2A3 plus a linear
combination of I, AI, A 2, A 3, and products of two such matrices.

This last proposition gives a significant improvement of proposition 2 when

n is larger than 3.

e) An ideal is a set 'I of polynomials such that, if P and Q are in 'I, so is P + Q,

and, if P E 'I, the product of P and of any polynomial is in 'I.
470 J. Peyriere

_,_
Corollary. Ifn is larger tlJan 3 and {Aih< <n are elements of 8£(2, C), then
1° any product constructed from these matrices or their inverses, can
be written as a linear combination of the matrices I, Ai (1 ~ i ~ n), and A'lAi2
(1 ~ i 1 < i2 ~ n), the coefficients of which are polynomials in the variables
trA, (1~i~n),trA'lA'2 (1~il<i2~n),andtrA'lA'2A'3 (1~i-1<
i2 < i3 ~ n).
2" the trace of such a product can be expressed as a polynomial with
rational coefficients in the n(n2 + 5)/6 traces defined above.

As a matter of fact, one can go further reducing the number of traces needed.
It results from [11] that the trace of a product of the kind considered above can
be expressed as a rational fraction in the variables tr A. (1 ~ i ~ n), tr A.A,
(i < j, 1 ~ i ~ 3, 1 < j ~ n), and tr A 1 A 2A 3 .

3. REVIEWS AND NOTATIONS FOR FREE GROUPS

Let A = {a}, a2, .. " an} be a finite set called alphabet.

3.1. Free sellli-group generated by A

Let A* be the set of words over the alphabet A. The product WlW2 of two of
its elements is just the word obtained by putting the word W2 after the word
WI' This operation is called concatenation. It is associative and has a unit
element, the empty word, denoted by 10.

3.2. Free group generated by A

We perform the same construction as above with the alphabet {a}, a2, .. " an,
all, a2"l , ... , a;;: I}, but we introduce the following simplification rules:

We obtain a group e) which we will denote by r A.

(3) Strictly speaking, elements of this group are not words, but equivalence classes
of words.
 TRACE MAPS 471

3.3. Representations in 8L(2, <C)

A representation of A* or rAin 8L(2, <C) is a mapping cp from A* or r A into

8L(2, <C) such that

for any WI and W2 in A* or r A.

Such a representation is determined by the values AJ of cp( aJ ) for j =
1,2,···, n. Computing cp(w) simply consists in replacing each letter in W by
the corresponding matrix.

3.4. Endomorphisms

A map a from r A to r A is an endomorphism of r A if, for any WI and W2 in

r A, one has

In the case where none of the words a( a J ) contains negative powers, a is an

endomorphism of A* and is called a substitution on the alphabet A.
Obviously, an endomorphism a is determined by a(a J ) (for j = 1,2, ... ,n).
Hereafter we shall identify an endomorphism a and the collection oJ words
(a(ad, a(a2),"" a(a n )).
For instance (4), a = (ab, a) means that a is the endomorphism, so called
the Fibonacci substitution, such that a(a) = ab and a(b) = a. In this case, as
an example, let us compute a(aba- I ):

a(aba- I ) = a(a)a(b)a(a)-I
= aba(ab)-l = abab-Ia- I

The composition of endomorphisms is simply the composition of maps. This

is illustrated by the following examples.

- (ab, ba) 0 (ab, a) = (abba, ab)

- (ab, a) 0 (b, b-Ia) = (b, b-Ia) 0 (ab, a) = (a, b) (this means
that, as an endomorphism, the Fibonacci substitution is invertible).

3.4.1. Matrix of an endomorphism

If W is an element of r A and a E A, Wa stands for the sums of exponents
of a in W (one can easily get convinced that Wa only depends on wand not
of the particular word used to represent it). Moreover, Wa is independent of

(4) In the case of a two letter alphabet, we prefer to denote a and b the generators
instead of al and a2.
472 J. Peyri(~re

the order of the factors in w. The mapping W t----+ W = (w a1 ,···, waJ (where
aI, a2, ... , an are the elements of A) is a group homomorphism of r A onto 7Z n
known as the homomorphism of abelianization.
For instance,
(aba-I)a = 0 (aba-I)b =1
(ab- I a- 2)a = -1 (ab- I a- 2)b = -1

The n x n matrix (j whose entry of indices (i,j) is (1(a J )a. is, by definition,
the matrix of the endomorphism (1.
For instance, the Fibonacci substitution matrix is (~ ~), and the one of

the Prouhet-Thue-Morse (PTM) substitution, (ab, ba), is (~ ~).

One has, for any (1 and w,
(1(W) = (jw

and, for any pair of endomorphisms,

4. TRACE MAPS (TWO LETTER ALPHABET)

In this section, the free group on two generators a and b will be denoted simply
by r.

4.1. Definition of trace maps

The following proposition is mainly a reformulation of the corollary to propo-

sition 1.

Proposition 5. For any W E r, there exists a unique polynomial Pw with

integer coefficients such that, for any representation tp de r dans SL(2, <C), one
has
trtp(w) = Pw(trtp(a), trtp(b), trtp(ab)).
Moreover, if WI = W2, the polynomial PW1 - PW2 is divisible by>..
Proof. We only have to prove the second assertion. Consider a representation tp
such that the matrices A = tp(a) and B = tp(b) share an eigenvector. As these
matrices are simultaneously trigonalizable, the trace of a product of A's and B's
does not depend on the order of factors. This means that tr tp( WI) = tr tp( W2).
 TRACE MAPS 473

In other terms, we have PWI (x, y, z) = PW2 (x, y, z) as soon as .x(x, y, z) = O.

The conclusion then follows from the irreducibility of .x.

The following notation will prove to be convenient: if cp is a representation

of r in 5£(2, ([;), set

Tcp = (trcp(a), trcp(b), trcp(ab)) (8)

With this notation, the equation of definition of Pw is trcp(w) = Pw(Tcp).

Let a be an endomorphism of r. We define the trace map associated with

a to be
(9)

It can be considered as well as a map from ([;3 to ([;3.

Let us compute the trace map for the PTM substitution a = (ab, ba). One
has to consider a representation cpo Set A = cp(a), B = cp(b), x = tr A, y =
tr B, and z = tr AB. Then we have cp(a(a)) = AB, cp(a(b)) = BA, and
cp(a(ab)) = AB2A. Therefore Pa(a) = Pa(b) = Z. Then AB2A = yABA-A2 =
y(zA + B - y) - xA + 1 from which we get Pa(ab) = xyz - x 2 - y2 + 2. At la8t,
cI>a = (z, z, xyz - x 2 - y2 + 2).

Remark. The reader could ask whether is was necessary to replace lower case
letters by upper case ones in the previous calculation. Indeed, this is not
compulsory. If we analyze what we have done, we have just considered a and
b to be generators of an algebra on the ring LZ[x, y, z], subject to the following
relations:

a2 - xa + 1 = 0, b2 - yb + 1 = 0, and ab + ba = z - xy + ya + xb .

We call this algebra, the Procesi-Razmyslov algebra on a two letter alphabet.

Then proposition 1 can be interpreted as giving a homomorphism of the group
algebra(5) of r to the Procesi-Razmyslov algebra.

Here are a few examples of trace maps.

(5) This algebra is the set of finite formal linear combinations of elements of r
endowed with the bilinear multiplication which extends the product in r.
474 J. Peyriere

(a, b) (x,y,z)
(a-I,b- I ) (x,y,z)

(b, a) (y,x,z)
(ab,b- I ) (z,y,x)
(b, a-I) (y,x,xy - z)

(ab, a) (z,x,xz-y)
(b,b-Ia) (y,xy - z,x)

(ab,ba) (z, z, xyz - x 2 - y2 + 2)

(a 2 b,ba) (xz - y, z,x 2yz - x 3 - xy2 - yz + 3x)

4.2. First properties of trace maps

Proposition 6. For any endomorphisms a and T of r, we have <Paor = <Pro<P a .

Proof We have the following characterization of <P".: for any representation 'P,
T('P 0 a) = <P".(T'P).
The proposition then results from

Corollary. For any endomorphism a of r, and for any wE r, one has

Proof Let T be the endomorphism (w, b). Then P".(w) is the first component
of <P".or, i.e. the first component of <Pr composed with <P"..

As a consequence, if a has a fixed point w, the corresponding trace map <Pa

leaves the surfaces of Pw(x, y, z) = constant globally invariant.
If a is invertible, then <P". 0 <P ".-1 = id. Taking the jacobian, we get

As these determinants are polynomials with integer coefficients, we must have

det <P~ == 1 or det <P~ == -1.
 TRACE MAPS 475

As an example, consider the PTM substitution for which we have

00
<I>' = ( 0 0
yz - 2x xz - 2y

The corresponding determinant is 0, so the PTM substitution is not invertible.

Proposition 7. For any endomorphism a ofr, there exist a polynomial Qa

with integer coefficients such that

Proof Let x, y, and z be such that >'(x, y, z) = 0 and z =I- O. Choose unimodular
matrices A and B such that tr A = x, tr B = y, and tr AB = z, and consider
the representation <p defined by <p(a) = A and <p(b) = B. As A and B share an
eigenvector, so do <p(a(a)) and <p(a(b)). So we have >'(T(<p)) = O. Therefore
>'(x, y, z) = 0 implies>. (<I>a(x, y, z)) = O. Since>. is irreducible, it divides >.o<I>a.

As a consequence, any <I>a leaves globally invariant the surface n the equation
of which is >'(x, y, z) = O. Moreover, the restriction of <I>a to n only depends
on a.

Lemma. For any a, we have Qa(O, 0, 0) = 0 or l.

Proof This is checked by testing on matrices (~1 ~) and (~ ~) .

Proposition 8. If a and Tare endomorphisms, we have

Proof We have

Corollary. If a is invertible, then Qa == 1 and <I>a leaves globally invariant

each surface >'(x, y, z) = constant.

4.3. Further properties of traces maps

- <I>a = id if and only if a is an inner automorphism of r or an inner

automorphism composed with (a-I, b- 1 ).

- Qa == 1 if and only if a is invertible.

476 J. Peyriere

- Qa == 0 if and only if a is not one-t(}-one.

- Qa(O, 0, 0) equals 0 if det a is even, 1 otherwise (this is a precise

form of the lemma preceding proposition 7).

- <1> a (0, 0, 0) = (0,0,0) if and only if det a is even.

- grad Qa(O, 0, 0) = O.
- If det a is odd, the hessian matrix of Qa at (0,0,0) is diagonal with
non-positive entries.

5. Trace maps (n letter alphabet)

5.1. Three letter alphabet

If rp is a representation of rAin SL(2, <C), we define T rp to be the following

collection of traces:

and recall the definitions of several polynomials

A(X, Y, z) = Z2 - p(X, Y)z + q(X, Y)

where

and

(as previously, X stands for the collection of x's, and similarly for Y).

Let V be the hyper-surface in <c 7 the equation of which is A(X, Y, z) = O. It

can be seen that any point of V is of the form T 'P.

Proposition 2 (or the corollary to proposition 4), shows that, for any w ErA,
there exits a polynomial Pw such that tr rp(w) = Pw(T rp) for any representation
rp. As we have already observed, this polynomial is no longer unique. It is
indeed defined up to the addition of a multiple of polynomial A (or, more
pedantly, modulo the ideal I generated by A).

Now, if we have an endomorphism a of r A, we choose a collection of poly-

nomials
 TRACE MAPS 477

This ~q defines a map from (:7 to V, the restriction of which to V does not
depend on the different choices. Indeed, this is this map from V to V which
is the trace map and which we call ~q. As previously, T(<p 0 0") = ~q(T<p) for
any <p, and ~qOT = ~T 0 ~q.

In order to show that, as previously, there exists an algebraic sub-manifold

n de V which is globally invariant under any ~q, we need the following lemma
of which we omit the proof.
Lemma. Three matrices Al. A 2, and A3 in 8L(2, (:) have a common eigendi-
rection if and only if

A(X,Y,z) =0,
>'(Xl. X2, Y3) = >'(X2' X3, yI) = >'(X3, Xl, Y2) = 0,
and
p(X, y)2 - 4q(X, Y) = 0,
where
z= tr AIA2A3 ,
X = (Xl, X2, X3) = (tr AI, tr A 2, tr A3),
and

.:r
Let n be the manifold associated with the ideal generated by the polyno-
mials A, >'(Xl. X2, Y3), >'(X2, X3, YI), >'(X3, Xl. Y2), and p2 - 4q.
Then an argument similar to the one used in the proof of proposition 7 shows
that n is invariant under any ~q.

5.2. n letter alphabet

When n is larger than 3, some complications occur and the situation is less
easy to describe. In view of the corollary to proposition 4, we need n(n2 +
5)/6 variables. The ideal I of relations between these variables is no longer
principal. The trace maps take the variety V of I to itself. They still leave
globally invariant a sub-variety n of V defined by an ideal .:r
the definition
of which comes from expressing that n elements of 8L(2, (:) have a common
eigendirection.

6. COMMENTS

Fricke's formula and the corollary to proposition 3 (Fricke's lemma) appear in

[4], but were also stated by Vogt in 1889.
478 J. Peyrii~re

Proposition 2_10 has been stated by Fricke [4] and proved by Horowitz [5].
Since then, it has been rediscovered several times: Allouche and Peyriere [1]
for n = 2, for general n by Kolar and Nori [9] (although they gave a formula
involving a number of traces much larger than 2n -1), and Peyriere et al. [17].

Traina [20, 21] gave an efficient algorithm for computing Pw in the case of a
two letter alphabet; also in this case Wen Z.-X. and Wen Z.-Y. [25] determined
the leading term of Pw ' Procesi [18] and Razmyslov [19], instead of consid-
ering relations between traces only, used more general polynomial identities.
This gives simple algorithms for computing polynomials Pw with an arbitrary
alphabet. This is this method which is exposed here.

Proposition 4 and its corollary appear in Avishai, Berend and Glaubman [2].

The trace map appears in Horowitz [6]. It has also been rediscovered number
of times: by Kohmoto et al. [7] in the case of Fibonacci, by Allouche and
Peyriere [1] and Peyriere [16] for n = 2, by Peyriere et al. [17] for n > 2.

Proposition 7 essentially appears in Horowitz [6]. It has also be rediscovered.

Kolar and Ali [8] conjectured it after having used a computer algebra software.
The proof given here appears in Peyriere [16].

Results in the section entitled "Further properties of trace maps" can be

found in Horowitz [6] and Peyriere et al. [17].

The structure of the ideal I, for a four letter alphabet, is studied by Whitte-
more [22] and completely elucidated by Magnus [11] for an arbitrary alphabet.
It also results from Magnus [11] that, for a n-Ietter alphabet, one can use 4n-5
variables only in trace maps with the counterpart that CPu is a rational map
instead of being a polynomial one.

For a study of the quotient modulo I (the ring of Fricke characters) see
Magnus [11].

Polynomial identities for p x p matrices are studied by Procesi [18], Razmys-

lov [19], and Leron [10]. Wen Z.-X. [23, 24] gives some algorithms for getting
such identities. He also constructs a trace map for 3 x 3 matrices and a two
letter alphabet.

For basic references on free groups, see [12-15].

 TRACE MAPS 479

7. BIBLIOGRAPHY

[1] J.-P. ALLOUCHE & J. PEYRIERE Sur une formule de recurrence sur les
traces de produits de matrices associes a certaines substitutions. C. R. Acad.
Sc. Paris 302 (II) (1986), 1135-1136.

[2] Y. AVISHAI, D. BEREND, and D. GLAUBMAN Minimum-Dimension

Trace Maps for Substitution Sequences. Phys. Rev. Lett. 72 (1994), 1842-
1845.

[3] Marc CULLER & Peter B. SHALEN Varieties of group representations and
splittings of 3-manifolds. Ann. of Math. 117 (1983), 109-146.

[4] R. FRICKE and F. KLEIN Vorlesungen iiber die Theorie der automorphen
Fun ction en. B. G. Teubner, Leipzig 1897. Reprint: Johnson Reprint Corpora-
tion (Academic Press), New York 1965.

[5] R. D. HOROWITZ Characters of Free Groups Represented in the Two-

Dimensional Special Linear Group. Comm. Pure and Applied Math. XXV
(1972), 635-649.

[6] R. D. HOROWITZ Induced automorphisms on Fricke Characters of Free

Groups. Trans. Amer. Math. Soc. 208 (1975), 41-50.

[7] M. KOHMOTO, L.P. KADANOFF, and C. TANG Localization Problem in

One Dimension: Mapping, and Escape. Phys. Rev. Lett. 50 (1983), 187(}-1872.

[8] M. KOLAR & M. K. ALI Trace maps associated with general two-letter
substitution rules. Phys. Rev. A 42 (1990), 7112-7124.

[9] M. KOLAR & F. NORI Trace maps of general substitutional sequences.

Phys. Rev. B 42 (1990), 1062-1065.

[10] U. LERON Trace Identities and Polynomial Identities of n x n Matrices.

J. of Algebra 42 (1976), 369-377.

[11] W. MAGNUS Rings of Fricke Characters and Automorphism Groups of

Free Groups. Math. Z. 170 (1980), 91-103.

[12] W. MAGNUS, A. KARRASS, and D. SOLITAR Combinatorial Group

Theory. Interscience Publishers, New York 1966, and Dover, New York 1976.

[13] B. H. NEUMANN Die Automorphismengruppe der freien gruppen. Math.

Ann. 107 (1933), 367-386.
480 J. Peyriere

[14] J. NIELSEN Die IsomorplJislllen der allgemeinen unendlichen Gruppe mit

zwei Erzengenden. Math. Ann. 78 (1918), 385-397.

[15] J. NIELSEN Die Isomorphismen der Ereien Gruppen. Math. Ann. 91

(1924), 169-209.

[16] J. PEYRIERE On the trace map Eor products oE matrices associated witb
substitutive sequences. J. Stat. Phys. 62 (1991), 411-414.

[17] J. PEYRIERE, Z. X. WEN. and Z. Y. WEN Polynomes Associes aux

Endomorpbismes de Groupes Libres. L'Ens. Math. 39 (1993), 153-175.

[18] C. PROCESI Tbe Invariant Tlwory oE n x n Matrices. Advances in Math.

19 (1976), 306-381.

[19] JU. P. RAZMYSLOV Trace Identities oE Full Matrix Algebras over a Field
oE Cbaracteristic Zero. Izv. Akad. Nauk SSSR ser. Mat. 38 (1974) No.4 (in
Russian); English translation: Math. USSR Izvestija 8 (1974), 727-760.

[20] C. TRAINA Representation oEtbe Trace Polynomial oECydically Reduced

Words in a Free Group on Two Generators. Ph.D. Thesis, Polytechnic Institute
of New York.

[21] C. TRAINA Trace polynomial Eor two generator subgroups of 8L(2. <c).
Proc. Amer. Math. Soc. 79 (1980), 369-372.

[22] A. WHITTEMORE On Special Linear Cbaracters oE Free Groups oE Rank

n ~ 4. Proc. Amer. Math. Soc. 40 (1973), 383-388.

[23] Z.-X. WEN Diverses etudes sur l'automaticite. These, Universite Paris-Sud
1991.

[24] Z.-X. WEN Relations polynomiales entre les traces de produits de matrices.
Comptes Rendus Acad. Sci. Paris, 318 5erie I (1994), 99-104.

[25] Z.-X. WEN and Z.- Y. WEN On the leading term and tbe degree oE tbe
polynomial trace mapping a.ssociated with a substitution. J. Stat. Phys. 75
(1994), 627-{)41.
 COURSE 17

Schrodinger difference equation with

deterministic ergodic potentials
Andras Siito

Research Institute for Solid State Physics, POB 49, H-1525 Budapest 114,
Hungary. Email: suto@power.szfld.kfld.hu
Institut de Physique Thoorique
Ecole Poly technique FederaJe de Lausanne
CH-1015 Lausanne, Switzerland

We review the recent developments in the theory of the one-dimensional

tight-binding Schrodinger equation for a class of deterministic ergodic poten-
tials. In the typical examples the potentials are generated by substitutional
sequences, like the Fibonacci or the Thue-Morse sequence. We concentrate on
rigorous results which will be explained rather than proved. The necessary
mathematical background is provided in the text.

Contents

1. Introduction
2. Main examples
3. General results on the Schrodinger difference equation
3.1 Basic observations
3.2 Transfer matrices
3.3 Lyapunov exponent
3.4 Scattering problem: Landauer resistance
4. Schrodinger equation with periodic potentials
482 A. Siito

5. Spectral theory
5.1 Schrodinger operator, i2(Z)-space and spectrum
5.2 Point spectrum
5.3 Cantor sets
5.4 Continuous spectrum
5.5 Spectral projection
5.6 Measures
5.7 Cantor function
5.8 Spectral measures and spectral types
5.9 A spectral measure for Ho
5.10 i2(Z) versus i 2(N)
5.11 Asymptotic behaviour of generalized eigenfunctions. Subordinacy
6. Schrodinger equation with strictly ergodic potentials
6.1 Strict ergodicity
6.2 The spectrum of H(w) = Ho + V(w)
6.3 Integrated density of states
6.4 IDS and Lyapunov exponent
6.5 Results on the set {E : -y(E) = O}
6.6 The role of periodic approximations
6.7 Gordon-type theorems
6.8 Kotani theorem for potentials of finite range
6.9 Gap labelling
7. Schrodinger equation with Sturmian and substitutional potentials
7.1 Fibonacci potential
7.2 General Sturmian potentials
7.3 Period doubling potential
7.4 Thue-Morse potential
7.5 Systematic study of substitutional potentials
8. Solutions of the problems

References
 SCHRODINGER DIFFERENCE EQUATION ... 483

1. Introduction

The one-dimensional SchrOdinger equation

(-d 2 /dx 2 + V(x))1/J(x) = E1/J(x)
has long been serving as a laboratory for studying a large variety of problems
in Quantum Mechanics. Most often it appears as the eigenvalue equation of
a one-electron energy operator, and then V(x) means the potential energy of
the electron. In Atomic or Molecular Physics V (x) is typically fast-decaying or
vanishes outside a bounded set, while in applications to Solid State Physics it
is an extended function. Our concern will be the Schrodinger equation of Solid
State Physics in the tight-binding approximation: the real axis is replaced by
Z, the lattice of integers, and the differential operator by the second order dif-
ference operator. The tight-binding Schrodinger equation describes an electron
in a: lattice in which the attraction of the ion cores slows down the interatomic
motion of the electron: the kinetic energy has an a priori upper bound which
does not exisL in the continuous case.
The potentials we are going to consider are on the midway between peri-
odic and random; in physics literature they are sometimes called 'aperiodic' or
'disordered deterministic'. They all fall into the family of the so-called strictly
ergodic sequences. A typical subfamily which appeared regularly in the lec-
tures of this Winter School are the substitutional sequences as, for example,
the Fibonacci sequence. As we will see later, the periodic and the uniformly
almost periodic sequences are also strictly ergodic, but a random potential is
usually 'only' ergodic.
In this field physicists and mathematicians have been working parallelly dur-
ing about the last fifteen years. The ambition of this course is to resume and
explain the contribution of the latter for an audiance unfamiliar with the under-
lying mathematics. Most of the time, mathematicians consider the Schrodinger
equation directly on the infinite chain Z or on the semi-infinite chain N. The
equation is then the 'eigenvalue equation' of an infinite matrix which is named
after Jacobi. If the diagonal (the potential) is an ergodic sequence, the spec-
trum happens not to be a discrete point spectrum. It is rather often - at least
within the family of potentials we are going to consider - singularly continuous
and is, as an ensemble of points, a (generalized) Cantor set. These notions
and many others, necessary for the understanding of the specific results will be
introduced in the forthcoming sections.
We start, in Section 2, by the presentation of the main examples and the
definition of strict ergodicity. The subject is developped in Sections 3, 4, 6 and
7. In Section 3 we give the basic ingredients of the theory of discrete Schrodin-
ger equation in one dimension and introduce the notions of the transfer matrix
and Lyapunov exponent. A subsection is devoted to the scattering problem and
Landauer resistance. Section 4 contains some of the fundamental results for
periodic potentials. Section 6 summarizes the general results obtained for one-
dimensional Schrodinger operators with strictly ergodic potentials. Section 7 is
484 A. Siito

a review of our actual knowledge about models with specific potentials, given
by Sturmian and substitutional sequences.
Section 5 is an introduction into spectral theory. It is far more the longest
part of the course, with the definitions of spectrum, essential spectrum, eigen-
value, point and continuous spectrum, spectral projections, measures and spec-
tral measures, Cantor sets etc., illustrated with examples. Subordinacy and the
classification of the spectrum according to the asymptotic behaviour of the so-
lutions of the Schrodinger equation is also discussed here.
The widely studied Almost-Mathieu or Harper equation is not presented in
full detail but appears in examples throughout the whole text, and papers on
it are included among the references.
Some more or less simple facts are announced in the form of problems. The
reader may only retain the statements or, if he wishes, check his understanding
by proving them. The solutions can be found in Section 8.
These notes are not written in a purely mathematical style. Proofs, unless
they are really elementary, are replaced by explanations. This suffices for the
physicist who whishes to understand the results but has no ambition to pro-
vide mathematical proofs of his own. The interested reader can find excellent
monographs which present the subject with full mathematical rigor. Also, one
should remember that nothing can replace the lecture of the original research
papers. A rather detailed, although certainly incomplete, list of references is
provided at the end of the notes. I thank Fran~ois Monti for helping me to
construct this list.
It is a pleasure to thank Fran~ois Delyon, Fran~ois Monti and Charles Pfis-
ter, who accepted to read the manuscript; their remarks brought considerable
improvement to the final text.

2. Main examples

The tight-binding Schrodinger equation in one dimension is written as

(2.1)

Here n runs over the integers, V = {Vn } is a real sequence (the potential),
E is a real number (the energy), although sometimes it is useful to take it
complex. The problem is to find the 'physically allowed' values of E, that is,
those admitting at least one 'not too fast increasing' solution, and to describe
these solutions. The following potentials will be considered:
(I) Vn = V{W)n = Acos 211"{na+w) , A =F 0. The corresponding discrete Schro..
dinger equation is called the Almost-Mathieu equation (because the continuous
version is the Mathieu equation). The case A = 2 is the Harper equation.
(2) Vn = V{W)n = AXA{na + w), where A =F 0, A is the union of semi-open
intervals of [0,1) and XA{x) = 1 if the fractional part of x falls into A and is
zero otherwise. These are the so-called circle-potentials. An important special
 SCHRODINGER DIFFERENCE EQUATION ... 485

case is when A is a semi-open interval of length a: The {O, 1}-valued sequences

X[O,a)(na + w} = Lna + wJ - L(n - l}a + wJ

X(O,aj(na + w} = rna + wl - r(n -1}a + wl (2.2)

for irrational a are called Sturmian sequences. The notations L·J and r·l mean
rounding downwards and upwards, respectively. If a = (.j5 - 1}/2, we obtain
the Fibonacci sequence. Generalized Fibonacci sequences (obtained through
the substitution ~(a) = anb and ~(b) = a with n ~ I) are also Sturmian.
According to an equivalent definition, Sturmian sequences are the sequences
with minimal complexity among the non-ultimately-periodic sequences (see
e.g. [1], [23]).
(3) V is a substitutional sequence:
Let ~ be a primitive substitution ([114]) on a finite alphabet A = {al, ... , aT}' f
a non-constant real function on A. Suppose that ~(a) = ... a and ~(b) = b... for
some a, b E A and let w = uv where u and v are the left- and right-sided fixed
points of~: u = C"'(a} = ... a and v = ~OO(b} = b.... Then V(w}n = f(w n }.
A comment on the concatenation uv will follow in Section 6. Generalized
Fibonacci sequences have a natural definition on Z via the formula for Sturmian
sequences.
The potentials defined in (1 }-(3) are strictly ergodic, that is, minimal and
uniquely ergodic:
(i) V is minimal.
Let T be the left shift, and denote V(Tw} the shifted potential: V(Tw}n =
V(w}n+1. The notation (T applied to w) may seem curious, however, in all
cases Tw can be given a well-defined meaning. For potentials (I) and (2)

Tw =w + a (mod I)

In case (3), Tw is the left-shifted sequence,

The orbit of V(w}, Le., the sequence of translates, V(Tkw} for k = ... , -1,0,1, ... ,
has pointwise convergent subsequences: W = {Wn } is called a pointwise limit
of translates of V(w} if for a suitably chosen sequence {k i }, V(Tk'w}n tends to
Wn for every n. Clearly, W = V(W'}, where w' is the limit of Tk·w in cases (I)
and (2) and is a pointwise limit of translates of w in case (3).
The orbit of V together with all its pointwise limits is called the hull of V and
is denoted by O(V}. This is a kind of closure of the set of translates. The hull
can be constructed for any infinite sequence. If V is periodic with period length
L, the hull is a set of L elements. The potentials of (I) and (2) are periodic
if a is rational. If a is irrational, the hull of V = V(w = O} is {V(W)}wE[O,1)!
This is the typical situation: the hull is much larger (uncountable) than the set
of translates (countable). If S is an arbitrary sequence and S' is a pointwise
limit of translates of S, the hull of S' is usually different from the hull of S.
486 A. Siit6

(Example: Sn = 1 - lin, n = 1,2, .... The translates have a single pointwise

limit, S~ == 1. The only element of O(S') is S' itself.)
The potentials defined in (1)-(3) have the property that for any W E O(V),
O(W) = O(V). This is what the minimality of V means. The V-independent
hull is denoted by o. Clearly, every element of 0 is minimal and therefore the
'flow' (0, T) itself can be called minimal: every trajectory is dense in 0 (its
closure is 0) or equivalently, the only closed and shift-invariant subsets of 0
are the empty set and o.
Problem 1. (Cf. Queffelec [113]) Let 8 = {8n}~00 be a bounded complex
sequence, 0(8) its hull,
0(8) = {Tk8h=~00
where the bar means closure with respect to pointwise convergence. The fol-
lowing are equivalent.
1. 8 is minimal, Le., for every t E 0(8), O(t) = 0(8).
2. For every t E 0(8), 8 E O(t).
3. 8 is almost-periodic in the metric
00

n:-oo

Le., for each t: > 0 there exists an te < 00 and a sequence {ndk=~oo ofintegers
with gaps bounded by t e , 0 < nk+1 - nk < t e , such that for all k
00
d(T k8,8) =
n L Tlnl18 n+nk - 8n l ~ t:.
n=-oo

A sequence is called recurrent if it is the pointwise limit of its own translates.

A minimal sequence is recurrent but a recurrent sequence may not be minimal.
A counterexample is provided by the sequence of digits 1234567891011121314 ... ,
obtained by concatenating the positive integers, written, e.g., in base 10. It is
the pointwise limit of its own left shifts, but is not minimal: words are repeated
with increasing gaps. The hull of this sequence is the set of all sequences
composed of the digits 0,1, ... ,9.
The different meanings, in different cases, of the label w should not disturb
the reader. In (1) and (2), w is a number of the interval [0,1); in (3) it is an
element of the hull of uv, and uv is also a minimal sequence.
(ii) V is uniquely ergodic.
Recall that the dynamical system (0, T, p), where p is a T-invariant probability
measure, is ergodic if for any p-integrable function f
N

J~oo N- 1 ~ f(Tnwo) = In f(w)dp(w)

for p-almost every Wo (Le., apart from a set of wo's of zero p-measure). Choosing
f to be the characteristic function of a set A, it is seen that ergodicity implies
that (in fact, is equivalent to) p(A) = 0 or 1 for every T-invariant set A.
 SCHRODINGER DIFFERENCE EQUATION ... 487

The unique ergodicity of V means that there exists a unique shift-invariant

probability measure p on n(V) (p(T- 1 A) = p(A) for any measurable A en).
Problem 2. If s is a uniquely ergodic sequence with shift-invariant probability
measure p then (n(s),T,p) is ergodic.
In case (1) (Almost-Mathieu), the most natural way to obtain this measure
is to define, for 0 ~ a < b < 1,

p(Aab) =b- a , where Aab = {V(W)}a<w<b (2.3)

and to extend p to countable unions and intersections of the sets Aab by using
the properties of measures.
For substitutional potentials the usual construction is to assign probabilities
to the so-called cylinder sets and to extend p to countable unions and inter-
sections of cylinders. If v = (VI, V2, ... , Vk) is a finite collection of real numbers
and n is an integer, the cylinder [v, n] is a family of potentials taken from the
hull,
[v,n] = {V En: Vn+i = V.,i = 1, ... ,k} (2.4)

Noticing that the relative frequency of the 'word' v is the same in all V E n,
the probability p([v, n]) can be defined as this relative frequency.
For circle potentials a shift-invariant p can be derived naturally either through
Eq. (2.3) or through word frequencies. For the Almost-Mathieu potential cylin-
ders can also be defined (by replacing, in Eq. (2.4), Vn+. = v. by Vn+i Eli,
where Ii is an interval), and one can obtain a T-invariant probability measure
through them. Uniqueness implies that the different constructions yield the
same measurable sets and measures.
Physicists never worry (do not even have to know) about the hull; math-
ematicians all the time struggle with it. The reason is that in laboratory or
computer experiments one deals with finite samples, and these may come from
any element of the hull: therefore it is sufficient to keep just one. (However, the
physicist's sample averaging is nothing else than averaging over the hull with
the measure p.) The mathematician is naturally led to the notions of hull and
strict ergodicity by working on the infinite system. He (she) has to consider
Vasa p-distributed random variable on n. The results she (he) obtains are
valid either for a given V E n or for every V E n or, most often, for p-a.e.
(almost every) V E n, i.e., p( {V E n: the result is not valid}) = O. Notice that
the distinction between 'every' and 'almost every' is by no means harmless.
Indeed, because p is a continuous probability measure (that is, p( {V}) = 0
for everyone-point subset of n), any countably infinite subset of the hull has
zero probability. A surprising new result about an even larger (uncountable)
set of vanishing p-measure, for which a claim (localization) is not valid, will be
quoted in Section 5.2.
Problem 3. If p is a shift-invariant probability measure on n and the trajec-
tory of every point is infinite then p is continuous.
488 A. Siit{)

3. General results on the Schrodinger difference equation

3.1. Basic observations

The first two terms of Eq.(2.1) correspond to the action of the kinetic energy
operator on the wave function, and any physicist would prefer to replace them
by
-'ljJn-1 - 'ljJn+1 + 2'IjJn
This is, however, a difference of no importance: Omitting 2'IjJn means shifting
the zero of the energy by 2, and changing the sign of the first two entries in
Eq.(2.1) amounts to a unitary transformation, thus leaving the spectrum of
the energy operator unchanged. Therefore a global change of sign of V is also
of no importance: we get the mirror image of the spectrum. This simple fact
can be expressed in a more shocking form: If V gives rise to a localized state
at energy E, - V will give rise to a localized state at energy -E. Or still in
another form: There can exist extended states at energies everywhere below
the potential, E < Vn for all n.
Without saying otherwise, we always consider the Schrodinger equation on
the infinite lattice Z. Equation (2.1) is a homogeneous linear second order
difference equation. As a consequence, for any (complex) E the solutions form
a two-dimensional linear vector space:
• If 'ljJ1 and 'ljJ2 solve (2.1), then 'IjJ = C1'IjJ1 + C2'IjJ2 solves it for any complex
C1 and C2.

• There exist two linearly independent solutions 'ljJ1 and 'ljJ2.

Two nontrivial (=F 0) solutions are linearly dependent if and only if they are
constant multiples of each other. When saying that Eq.(2.1) has a unique
solution of some property, we will understand 'unique linearly independent'.
Writing (2.1) for 'ljJ1 and 'ljJ2 (not necessarily linearly independent), multiply-
ing the first equation by 'IjJ~ and the second one by 'IjJ; and subtracting we find
that
.1,1 .1,2 .1,1.1,2 _ .1,1.1,2 .1,1 .1,2 _ _ ./,1.1,2 .1,1.1,2
'Pn+1 'Pn - 'Pn 'Pn+1 - 'Pn 'Pn-1 - 'Pn-1 'Pn - ... - 'P1 'PO - 'PO 'P1

This constant is the Wronskian and will be denoted by W['ljJ1, 'ljJ2]. As a conse-
quence,
• 'ljJ1 and 'ljJ2 are linearly independent solutions if and only if W['ljJ1 ,'ljJ2] =F O.
• There can be at most one solution going to zero as n -+ 00. If 'ljJ1 is such
a solution, then for any other (linearly independent) solution 'IjJ

l'ljJnl 2 + l'ljJnHl 2 -+ 00 as n -+ 00

The same conclusion holds for the limit n -+ -00. As a consequence,

there can be at most one solution decaying on both sides, and if such a
solution exists, all the others blow up on both sides.
 SCHRODINGER DIFFERENCE EQUATION ... 489

3.2. Transfer matrices

We may rewrite the Schrodinger equation in the redundant vectorial form

where
(3.1)

By iterating,

and similarly,

The matrices T are called transfer matrices. Their determinant is 1: this is

clear for Tn and its inverse and holds for the others because the determinant
of a product of matrices is the product of the determinants.
The determinant of Tl--+n can be seen as the Wronskian of the two 'standard'
solutions 'IIi and 'ljJ2 with initial conditions

'ljJt = 1 0 = 'ljJf
'ljJJ = 0 1 = 'ljJ5

In terms of these solutions the transfer matrix can be written as

'ljJl
Tl--+n = Tl--+n ( 'IjJ~ (3.2)

showing that indeed,

det Tl--+n = W['ljJl ,'ljJ2] = 1
The general form of the transfer matrix from site m to site n is

T m--+n = A = (~ ~) , ad - be = 1 (3.3)

Here a, b, c, d are polynomials of E with real coefficients and of degree :::; In -

ml + 1. In particular,
tr A = a + d = Eln- m l+1 + ...
For real energies the transfer matrices are elements of the group SL(2,JR).
The characteristic equation of a matrix A in this group reads
det(). - A) = ).2 - ). tr A +1= 0

Therefore the two roots satisfy the relations ).1).2 = 1 and Al + >'2 = tr A. We
can distinguish four cases:
490 A. Siito

1. The two roots are real, lAd = IA;-11 > 1, the corresponding eigenvectors
are also real and nonorthogonal, in general. Itr AI > 2, the matrix is
hyperbolic.
For example, the 'hyperbolic rotations'

( coshx sinhx)
sinhx coshx

are special (symmetric) hyperbolic matrices. Al,2 = exp(±x), the corre-

sponding eigenvectors are the transposed of (1 ± 1).
2. Al = A2 = ei<p, tp =J mr. There are two complex eigenvectors. tr A =
2 cos tp, the matrix is elliptic. As an example, the ordinary rotation matrix

( c~sx -sinx)
smx cos x

is elliptic. Al,2 = exp(±ix), the eigenvectors are the transposed of (1 =fi).

3. Al = A2 = ±1. There is a unique, real, eigenvector (the geometric multi-
plicity is 1). Itr AI = 2, the matrix is parabolic. An example is

with b =J O. The unique eigenvector is the transposed of (1 0).

4. A = ±I where I is the unit matrix.
Some assertions will concern the norm of transfer matrices. The natural
matrix norm
IIAII = sup IIAull/liuli
associated with the vector norm Ilull = (IU112 + IU212)! satisfies

IIAII2 = max. eigenvalue of A * A = 211AII~ 1 + (1411AII~ - 1)! = IIA -111 2 ~ 1

(3.4)

Here IIAIi2 is the trace norm,

IIAII~ = tr A* A = a 2 + b2 + c2 + d2 = IIA-lll~ ~ 2 (3.5)

(cf. Eq.3.3), which proves also the last inequality in (3.4).
Let us return to the trace of the transfer matrix as a function of E.
trTl--+n(E) = En + ... = Pn(E)
is an nth degree polynomial with n distinct real roots and the property that
 SCHRODINGER DIFFERENCE EQUATION ... 491

In Figure 1a we represented the generic situation for n = 5: There are n disjoint

intervals of the energy within which the transfer matrix is elliptic, separated
by intervals of hyperbolicity. Figure 1b shows an atypical case for n = 4: Some
intervals of hyperbolicity disappear because the derivative of the trace vanishes
at the same E where the trace takes on the value 2 or -2.
Problem 1. For A ESL{2,1R), show that Itr Akl ~ 2 if and'only if Itr AI ~ 2.
Plot trTLn{E) on the top of Figure 1.
t r( Tl-+ 5(E»)

hyperbolic

----I----\-----I-+---\------+----r--- parabolic
elliptic

--I'-------\-----+--+-----+----f---T--- parabolic
hyperbolic

tr(T (E»)
1-+4

hyperbolic
parabolic

..
elliptic

parabolic
-2
hyperbolic
G) T1 -+ 4 (E)= - 1

Fig. 1. - Trace of the transfer matrix as a function of E. a) Generic situation for

n = 5; b) Atypical case for n = 4.
492 A. Siito

3.3. Lyapunov exponent

Consider
1
'Yn = ~ In IIT1-+nll
This number is nonnegative (because the norm is 2:: I) and depends on E and
the parameters of the potential. From 'Yn we obtain, as n -+ ±oo,

'Y± = lim sup 'Yn , 'Y ±

n~±~ -
= lim inf 'Yn
n~±~

We find the same numbers by using the trace- or any other norm. In fortunate
cases the limits exist (i.e., lim sup and liminf coincide) and give 'Y± = 'Y±(E, V}.
These we call the Lyapunov exponents. If 'Y+ = 'Y-, the common value is
denoted by 'Y(E, V}.
To elucidate the meaning of the Lyapunov exponent, recall that

d. equations (3.1), (3.2) and (3.5). Let, e.g., n -+ +00 and suppose that 'Y+
exists. If 'Y+ = 0, every solution is subexponential. Suppose that 'Y+ > O. Then
asymptotically

for i = 1 or 2 or both and, since any solution is a linear combination of the two
standard ones, the same asymptotics is valid for all solutions, except perhaps
a unique one (which is asymptotically smaller). The Multiplicative Ergodic
Theorem due to Osceledec [104] and Ruelle [123J asserts that there is indeed a
unique solution 1/J with an exponential decay:

If both of 'Y± are positive, it may happen that the unique solution decaying
exponentially to the left coincides with the unique solution decaying exponen-
tially to the right: we have an exponentially localized state. If this is not the
case, for the given energy all solutions increase exponentially at least in one di-
rection. Such an energy is considered as 'physically forbidden' and the solution
physically irrelevant, because there can be no electronic state inside the solid,
corresponding to it.
Let us make this statement somewhat more explicit. We may ask: Provided
that the electron is in the interval [-L, L], what is the probability of finding it
at site n? The answer is l1/JnI2/L~=_L l1/JkI 2. This number is of the order of
unity close to L or -L and exponentially decaying towards the inside. So in
any finite sample, at the given energy, the electron cannot penetrate into the
bulk, it is exponentially localized at the boundaries. Also, one can learn from
generalized eigenfunction expansion that solutions increasing faster than any
power of Inl will not appear in the expansion of wave packets or computation
of matrix elements and are therefore of no relevance to Physics.
 SCHRODINGER DIFFERENCE EQUATION ... 493

3.4. Scattering problem: Landauer resistance

Suppose we have a finite sample: a potential given at n = 1, ... , L. Our aim

is to discuss scattering of an electron of energy E on the sample. We devise
our scattering experiment so as to have access to all energies. To this end, the
potential is extended to Z with constant values,

if n:::; 0
if n > L

chosen so that IWl -w21 < 4. Then the scattering problem is defined for energies
satisfying the inequalities

(Notice that E < Wi is allowed, see the remark in the first paragraph of Section
3.1.) The general solution has the form

(3.6)

(the plane waves extend to n = 1 and n = L) where the real wave numbers k i
are related to E and w, via the equations

2 cos k t =E - Wt , i = 1,2
There is no loss of generality, if we suppose that 0 < ki < 1r. With this
convention, the scattering problem amounts to set A = 1, B = r, C = t and
D = 0 and to solve for t and r. The Landauer resistance of the sample of
length L is defined as RL = Irl2 Ijtl 2 .
'ljJ with the scattering boundary condition is a complex solution of the Schr6-
dinger equation. Therefore, its complex conjugate is a linearly independent
solution to the same (real) energy. The Wronskian between 'ljJ and 'ljJ* is a
constant times the current density: its constancy physically means current
conservation. Compute the Wronskian at n = 0 and at n = L. One finds

so that
Itl2sink2 = (1-lrI2)sinkl
This relation is well-known in the special case of kl = k 2 • Because the sines
are positive, Irl cannot exceed 1.
We want to express RL in terms of the elements a, b, c, d of the transfer matrix
Tl--+L' (Here we use the notations of Eq.(3.3)). This can be done directly by
solving the linear problem

)=(~ ~)(
494 A. SiitO

but there is a more elegant way to do it. We can express the Landauer resistance
also in terms of another type of transfer matrix, usually applied for the Schro..
dinger equation in the continuum. This transfer matrix,

connects the coefficients A and B to C and D, see Eq.{3.6}:

In contrast with Tl-+L' Y depends on kl and k 2 • Let now 'Ij; denote the solution
with A = 1 and B = 0; it follows that C = a and D =,. 'Ij;*, the complex
conjugate of 'Ij;, is also a solution with A = 0 and B = 1 and, thus, with C = (3
and D = 8. We find therefore

, = (3* , 8 = a*

Furthermore, the Wronskian of'lj; and 'Ij;* computed at n = 0 and at n =L

yields

and thus
det Y = lal 2 - 1(31 2 = sin kl j sin k2
The scattering problem

trivially gives
1 sin kl
r = -(3*ja* , t =- --
a* sin k2
{3.7}

from which
RL = 1(31 2 s~n2 k2
sm 2 kl
Another expression of RL in terms of the trace norm of Y can also be obtained.

With this we find

RL = ~ilYlI~s~n2 k2 _ ~ s~nk2
4 sm 2 kl 2 sm kl
These formulas are frequently used with kl = k2, see e.g. [46], [85], [82]. It
remains to connect the two different kinds of transfer matrices. This can be
done by simple inspection and results
e-,k2(L+l)
e-,k 2 L
 SCHRODINGER DIFFERENCE EQUATION ... 495

Solving this equation for a and f3 and sustituting into Eq.(3.7) we obtain
t = -2i sin kle-ik2L(ae-ikl + b - eik2(ce-tkl + d))-l (3.8)
ae ik1 +b- eik2(ceikl + d)
r = - ae- tk1 + b - e tk2 (ce- tk1 + d)
(3.9)

and
(3.10)
Recall that RL depends on E via a, b, c, d which are real polynomials of E. The
ordinary choice for the potential outside the sample is Wl = W2 = o. With this
we get kl = k2 = k, 2 cos k = E and can 'test the sample' at energies between
-2 and 2.
A nice and simple form is obtained for kl = k2 = 7r /2 which corresponds to
Wl = W2 = E:

A truly intrinsic resistance is the minimum of RL over all kl' k2 which, in

general, cannot be given in a closed form. The minimum over k2' when kl is
fixed at 7r /2, is still easy to find:

a-ib
min RL(7r/2, k2 ) = RL(7r/2, arg --.d) V 2 + b2 - V
= (1/4)(a c2 + d 2 )2
~ c-z
If the Lyapunov exponent is positive, the asymptotic behaviour of the resistance
is the same for all kl' k2: RL grows with L like exp{2'Y+L}.

4. Schrodinger equation with periodic potentials

Periodic differential equations have a fully developed beautiful theory (Floquet

theory, see e.g. [45]) which can be adapted without any difficulty to the case
of difference equations. Moreover, strictly ergodic sequences are well approx-
imated by periodic ones, and most of the results discussed in Sections 6 and
7 deeply rely upon this fact. In this section we therefore survey the periodic
case.
Let the potential be periodic with period L ~ 1 finite:
Vn+L = Vn for any integer n

Consequently,
Tm+L-+n+L = Tm-+n , all m, n
We exploit this by writing n = kL+m where k = k(n) and 0:$ m = m(n) < L
are uniquely determined and, e.g. for n > 0,

IIt n = ( 1/J~:l )= Tl-+n lIto = Tl-+m(Tl-+dklIto

496 A. Siito

According to the value of E we can distinguish several cases (see also Figure
I):
(I) The transfer matrix is elliptic or hyperbolic,

where A and A-1 are the eigenvalues. In this case there exists a matrix S
(generally nonunitary) such that

S-1T1-+LS = (~ A ~1 )

Therefore we can write

The initial condition

yields the solution

while with

we get
.T.o2
'i.'n = /\\-k(n)T1..... m(n) S ( 01 )

Because m(n} is an L-periodic function, A±k in 1}1,2 multiplies an L-periodic

function of n. There are two cases:
(1.1) The transfer matrix is elliptic,

A=et'T/ , trT1-+L(E} = 2 cos 1] , TJ=Pinteger X7r

The two solutions ¢1,2 are Bloch waves and E is in one of the (at most L)
'stability intervals'. The Lyapunov exponent exists, 1'+ = 1'- = 1'(E} = o. The
Landauer resistance Rn is a bounded quasiperiodic function of n (periodic, if
TJ is a rational multiple of 7r).
(1.2) The transfer matrix is hyperbolic, say, IAI > 1. The two solutions can be
written as
~1 = ALn/LJ p ~2 = A-Ln/LJ q
o/n n , o/n n

with L-periodic sequences Pn, qn. The energy is in an open - maybe semi-infinite
- interval. The finite intervals are called 'forbidden gaps', they separate the
stability intervals. The Lyapunov exponent 1'(E} = L- 1 1n IAI. The Landauer
 SCHRODINGER DIFFERENCE EQUATION ... 497

resistance Rn grows exponentially (with exponent 2')').

(2) The transfer matrix is parabolic, e.g. with Al = A2 =1

The solution can therefore be written as

The initial condition

gives rise to the periodic solution

and any other solution oscillates and grows linearly with a slope b/ L. The
energy is a boundary point of a stability interval. The Lyapunov exponent is
zero, the Landauer resistance has a quadratic increase.
(3) Accidentally it may happen that TI-+L(E) = ±I. Then the general solution

\lin = (±1)k(n)T1-+ m (n) \11 0

is periodic for all initial conditions. This situation is quoted as the 'closure of
a forbidden gap' (d. Figure 1b). In most concrete cases, it is difficult to say
whether it occurs or not. If all the gaps are open, there are L - 1 of them,
separating L disjoint stability intervals. Some examples :
(i) For the Almost-Mathieu equation with 0: = p/q rational and qW/7r non-
integer all the (q - 1) gaps are open. If W = 0, p is odd and q = 4m, then one
gap is closed (Bellissard, Simon [21], Choi, Elliot, Yui [31]).
(ii) For circle potentials (2.2) with rational 0: all the gaps are open.
(iii) We can obtain periodic potentials by iterating a substitution on a letter a
finite number of times and repeating the result periodically. Doing this with the
period-doubling substitution (O'(a) = ab, O'(b) = aa), for the resulting equation
all the gaps are open (Bellissard, Bovier, Ghez [15], [27]). For the Thue-Morse
substitution (O'(a) = ab, O'(b) = ba) the trace map ([9])

X n+2 - 2 = (xn+1 - 2)x; (n ~ 1)

with Xn = tr TI-+2ft reveals that Xn is a 2 n degree polynomial of E, and Xn+2 - 2

has double zeros at the 2n simple zeros of X n . This shows (see Fig. 1) that at
least one fourth of the gaps is closed! In fact, 1/4 is the precise proportion of
the closed gaps ([12]; see also [10] for the corresponding phonon problem).
(iv) For the best periodic approximations of a hierarchical potential all the
gaps are open (Kunz, Livi, Siito [90]).
498 A. SiltO

We close this section with the conclusion that the physically relevant solu-
tions belong to energies which fall into the stability intervals:

These solutions are the well-known Bloch waves. The stability intervals are
nothing else than the bands in the spectrum of the periodic Schrodinger oper-
ator.

5. Spectral theory

Let us come back to the question we asked already in the previous sections.
The Schrodinger equation on IRd or 'I} can be solved for any E. Which are
the physically relevant solutions? What can be said about the corresponding
energies? Spectral theory provides the following answer:
(i) To compute matrix elements of functions of the energy operator, it suffices
to consider only the polynomially bounded solutions of the SchrOdinger equa-
tion.
(ii) Every subexponential solution (growing slower than any exponential) be-
longs to an energy which is in the spectrum of the energy operator.

5.1. Schrodinger operator, [2-space and spectrum

Define an operator H on the sequences 1/J = {1/Jn} by

Then Eq.(2.1) is equivalent to H1/J = E1/J. We call H the Schrodinger operator;

it can be represented with the infinite tridiagonal (Jacobi) matrix (denoted by
the same letter)

1 Vn -
1
1

) (5.1)

Consider

This is a separable Hilbert space (it contains a countable dense set) with the
inner product (cp, 1/J) = E:'=-oo cp~ 1/Jn, and H is a selfadjoint operator on it. In
what follows, we use the notation 1i for a separable Hilbert space. We call the
elements of 1i vectors. The norm of a vector 1/J is defined by 111/J1I2 = (1/J,1/J). It
should not be confounded with the norm of the ((:2 vectors which we introduced
in Eq.(3.1) and denote by capital Greek letters.
 SCHRODINGER DIFFERENCE EQUATION ... 499

Suppose, for the sake of simplicity, that V is a bounded sequence. The

spectrum (j(H) of H is the set of E values for which E - H has no bounded
inverse: One can find some vector "ljJ such that the equation

(E - H)'P ="ljJ (5.2)

has no vector solution for 'P (but may have a solution outside 1i, see Problem
5.2.2) or it has more than one vector solutions. In the second case (E - H)'P = 0
has a nontrivial vector solution, that we call an eigenvector of H, belonging to
the eigenvalue E. (Recall that the spectrum of a finite matrix A is the set of
its eigenvalues, and

det(A - A) = 0 if and only if (A - A)-l does not exist.)

The spectrum is real, nonempty and closed, i.e., contains all of its limit
points.
Example. Let us decompose H into off-diagonal and diagonal parts, H =
Ho + V. Then (j(Ho) = [-2,2]. If V is L-periodic,

dH) = {E: ItrTl-+dE)1 $ 2}

In either cases, the Schrodinger equation has no vector solution, thus there is
no eigenvalue in the spectrum.
If E is an isolated point in (j(H), then E is an eigenvalue.
The essential spectrum, (jess(H), of H is the set of all non-isolated points
of (j(H), together with the isolated eigenvalues of infinite multiplicity (do not
occur in one dimension). Clearly, this is a closed set. An important character-
ization of the essential spectrum is due to Weyl:
E is in (jess(H) if and only if it is an approximate eigenvalue of H in the follow-
ing sense. One can find a sequence 'P n of vectors such that II'Pnll = 1, 'P n goes
to zero weakly (in £2 (Z) this means that 'Pk' --+ 0 for each k fixed: either 'P n
spreads over larger and larger domains or it remains well localized but 'escapes'
to infinity) and
(5.3)
where En are vectors, liEn II --+ o.
If"ljJ is not an eigenvector but it is a subexponential solution of H"ljJ = E"ljJ,
one can construct such a 'Weyl sequence', so E E (jess(H). If"ljJ is polynomially
bounded, we call it a generalized eigenvector (eigenfunction) of H, and E a
generalized eigenvalue.
Problem 1. Construct a Weyl sequence for a subexponential solution "ljJ which
is not an eigenvector. Hint: Replace"ljJk by 0 for k < 0 and k > n and normalize.

5.2. Point spectrum

The point spectrum of H is the set of all eigenvalues of H. It is denoted by

(jpp(H). The point spectrum may be empty, finite or count ably infinite. The
500 A. Siito

pure point subspace of 1i is the subspace spanned by all the eigenvectors of H.

It is denoted by 1ipp .
We say that H has a pure point spectrum, if its eigenvectors form a basis in
the Hilbert space, that is, 1i = 1ipp . In this case the spectrum is the closure
of the set of eigenvalues:

rr(H) = rrpp(H) = rrpp(H) U {limit points of rrpp(H)}

Example. A nontrivial counterexample known for every physicist is the hy-

drogen atom. Here the Hilbert space is L2(1R3). The energy operator of the
hydrogen atom has infinitely many bound states (eigenvectors) which, how-
ever, do not span 1i. So 1ipp is an actual part of 'N. rrpp is a discrete set of
negative numbers with a unique accumulation point, o. Any E > 0 belongs
to the spectrum but is not an eigenvalue, only a generalized eigenvalue. The
corresponding generalized eigenvectors are bounded: they are products of a
plane wave with a confluent hypergeometric function. The generalized eigen-
vectors ('scattering waves') can be used to build up wave packets, i.e., square
integrable functions (vectors of 1i) which are orthogonal to every eigenvector.
Problem 1. Find a possible Weyl sequence to show explicitly that 0 is in the
essential spectrum of the energy operator of the hydrogen atom.
rrpp is called a dense point spectrum if it is nonempty and has no isolated
point. If, for a bounded recurrent potential, there is a point spectrum, it is a
dense point spectrum (see Problem 6.2.2).
Examples of dense pure point spectra
1. Let H = V, Le., a diagonal matrix with Vn = cos 21l"na, where a is irrational.
V has a pure point spectrum, dense in [-1,1]:

The eigenvector belonging to the eigenvalue Vn is on, the unit vector concen-
trated on the site n. The set of eigenvectors is the canonical basis in £2(Z).
Problem 2. Choose E in [-1,1] but E i= Vn , any n. Use the definition of the
spectrum to show that E E rr(V).
2. Anderson localization. Let Vn be identically distributed independent random
variables. Let V(w) = {V(W)n} denote a realization, H(w) = Ho + V(w).
(i) Pastur [106]: There exists a set ~ clR such that

• rr(H(w)) = ~ with probability 1.

• ~ contains no isolated points.
• Any E E ~ is not an eigenvalue with probability 1.
(ii) Kunz, Souillard [91](simplified): Let the probability distribution r(x) of
the Vn's be continuous, and r(x) i= 0 if and only if a < x < b. Then
• the spectrum of H(w) is pure point with probability 1.
• ~ = [a - 2, b + 2]
 SCHRODINGER DIFFERENCE EQUATION ... 501

• All the eigenvectors are exponentially localized.

(iii) Carmona, Klein, Martinelli [28]: Let Vn be Bernoulli-distributed,
V, _ {O with probability p
n - b with probability 1 - P
Then the spectrum is pure point with probability 1,
I: = [-2,2] U [b - 2, b + 2]
and all the eigenvectors are exponentially localized.
3. Almost Mathieu equation (oX > 0).
Choose w at random in the interval [0, 1) (according to the uniform distribu-
tion). There is an appealing nonrigorous argument (Aubry-Andre duality, [4]),
according to which the spectrum should be pure point for oX > 2. This is not
quite true, but seems to hold for good Diophantine a and almost every w:
(i) If a is irrational, the spectrum is independent of w, O"(H(w)) = I:(a, oX).

°
(ii) Sinai [131]' Frohlich, Spencer and Wittwer [50]: If a is a good Diophan-
tine number, namely, there exists some constant c > such that
min{na - lnaJ, rnal - na} ~ c/n 2 for all n f:. ° (5.4)
and oX » 1, the spectrum of H(w) is pure point for almost every w and the
eigenvectors decay exponentially. Sinai's proof yields also that I:(a, oX) is a
Cantor set (see below).
Aubry, Andre [4], Thouless [138], [139]: The Lebesgue measure m of the spec-
trum is positive,

m(I:(a oX)) > 210X _ 21 m(I:(a, oX)) > oX - 2

,- ' E u - El - oX + 2
(Eu and El are the upper and lower boundary of the spectrum, respectively.)
Recently it was shown by Last ([93], [94]) that for good Diophantine a, as for
example in (5.4),
m(I:(a, oX)) = 210X - 21
(iii) Jitomirskaya [69], [70] brought sensible improvements to the above re-
sults. She simplified the proof, weakened the Diophantine condition on a, put
an explicit bound on oX (oX > 15), and showed that for all oX > 2 and for almost
every w the closure of the point spectrum has the same Lebesgue measure as
the spectrum itself.
(iv) One may think that there is only some technical difficulty to extend the
above results on localization from a.e w to every w. This is not true: Recently,
Jitomirskaya and Simon [71] proved that for any oX > 2 and irrational a there is
an uncountable set of w's which is dense in the interval [0,1] (i.e., any point in
[0,1] is a limit point of this set) and for which the spectrum of H(w) is purely
singular continuous (see later). Clearly, this set is of zero Lebesgue measure,
if for the given oX and a localization occurs for a.e. w. This result provides
an example of a singular continuous measure the support of which is a thick
(positive-measured) Cantor set.
502 A. Siito

5.3. Cantor sets

Take a closed interval Co of the real line. Cut off a finite number of open
intervals which have no common boundary points with each other and with
Co. What remains is a closed set C 1 without isolated points: a union of a finite
number of closed intervals. Repeat the same procedure with C1 to obtain C2 ,
and so on: continue it indefinitely, by following the rule that no interval is left
untouched. Let C denote the resulting set. It has the following properties:

1. C is closed. Indeed, each Cn is closed, Co ::> C 1 ::> ... , therefore C =

n~=oCn' and any intersection of closed sets is closed.

2. C is nonempty. Clearly, the boundary points Cni of Cn are not removed

during the construction, therefore C contains them for all n.
3. There can be no interval in C, according to the construction rule.
4. There is no isolated point in C. In fact, it is clear from the construction,
that the set of points Cni is dense in C.

A set with the above four properties is called a Cantor set. The boundary
points of CC, the complement of the Cantor set C, form a countable dense set
in C, but the Cantor set itself is uncountable. CC is an infinite union of open
intervals, the closure of which is the whole real line. What is the Lebesgue
measure of C? Suppose that m(Cn+1) = xnm(Cn ), then

II Xn
00

m(C) = m(Co)
n=O

This number is greater than zero if and only if

00

Llnxn >-00
n=O

which is true if and only if

00

For Cantor's 'middle thirds' set Xn = 2/3 and thus the Lebesgue measure is
zero.

5.4. Continuous spectrum

Let 1icont = 1i*p, the subspace orthogonal to 1ipp . By definition, H has

no eigenvector in 1icont . If the only element of 1icont is the null vector, the
spectrum is pure point. If 1icont contains a nonzero vector, H has (also) a
 SCHRODINGERDIFFERENCE-EQUATION ... 503

continuous spectrum O'cont(H). Indeed, the restriction of H to 1-lcont, denoted

by HI1-l cont , is a selfadjoint operator whose spectrum is nonempty, and

As a matter of fact, the dimension of 1-lcont is either 0 or infinite (1-l cont is an

H-invariant subspace (H1-l cont C 1-lcont ), and H has eigenvectors in any finite
dimensional invariant subspace) and in the second case O'cont(H) is a nonempty
closed set without isolated points. Thus we have the decomposition

O'(H) = O'pp(H) U O'cont(H)

but notice that the closure of O'pp (even O'pp itself) may overlap with O'cont.
Just as to the points of O'pp(H) there correspond the H-eigensubspaces of
1-lpp , to certain uncountable subsets of O'cont(H) there correspond infinite-
dimensional H-invariant subspaces of 1-lcont . The connection is made by spec-
tral projections.

5.5. Spectral projections

If H has a pure point spectrum, it can be decomposed as

H = L EP({E})
EEO'pp(H}

where P( {E}) is the orthogonal projection to the eigensubspace belonging to

the eigenvalue E (for the one dimensional Schrodinger operator all these sub-
spaces are one dimensional). The orthogonal projections are selfadjoint and
idempotent,
P({E})t = P({E}) = p({E})2
The orthogonality of the eigensubspaces belonging to different eigenvalues can
be expressed as
P({E})P({E'}) = 8E,E,P({E})
and, if the spebtrum is pure point,

L P({E}) = I = identity
EEO'pp(H}

Physicists' notation:
If E is a nondegenerate eigenvalue, 1/J the normalized eigenvector, then
P( {E}) = 11/J)(1/J1
If E is degenerate,
504 A. Siito

where one has to sum over an orthonormal basis in the eigensubspace.

In general, there is a fundamental relation between certain subsets of the
spectrum and orthogonal projections to H-invariant subspaces. Let ~ = (a, b]
be a half-open interval of the real line, Xa its characteristic function, i.e.,
Xa (E) = 1 if E E ~ and 0 otherwise. This function satisfies

(5.5)

One can define an operator Xa(H) (formally obtained by substituting E with

H) as a strong limit (see below) of polynomials of H, which inherits the prop-
erties (5.5) of the real-valued function Xa and is, therefore, an orthogonal pro-
jection; we denote it by P(~). Similarly, orthogonal projections can be defined
for more complicated sets, called Borel sets, as, for example, countable unions
and intersections of half-open intervals (closed and open sets are Borel sets),
and the family of the corresponding projections have nice algebraic properties:

P(0) = 0 , P(lR) = I

P(~l n ~2) = P(~dP(~2)

and for ~1 and ~2 disjoint sets

The last equality holds also in a stronger form: If ~i' i = 1,2, ... is an infinite
sequence of pairvise disjoint Borel sets then
N
P(U~l ~t) = s-lim L P(~i)
t=l
The sum converges in the strong sense, that is, for any £(J E '}{

N
L P(~i)£{J - P(U~l ~t)£{J as N - 00
t=l
The above properties of P are characteristic to probability measures; the only
difference is that P(~) are operators, so P is a projection-valued probability
measure. This measure is called the (spectral) resolution of the identity, its
values on Borel sets are the spectral projections.
The spectral projections live on the spectrum of H:

P(~) = P(~ n a(H))

If P(~) "# 0, the linear subspace P(~),}{ is nontrivial and invariant under H
(because H commutes with P(~)).
Consider now the converse relation: Let A be an H-invariant (closed) sub-
space and denote [A] the orthogonal projection onto A. [A] may not be a
 SCHRODINGER DIFFERENCE EQUATION ... 505

spectral projection. For example, if 'ljJ is an eigenvector belonging to a de-

generate eigenvalue E and A is the one-dimensional subspace spanned by 'ljJ,
[A] = 1'ljJ) ('ljJ1 is not a spectral projection. The smallest spectral projection larger
than [A] is P({E}) (for projections Q1, Q2, Q1 < Q2 means Q1Ji C Q 2Ji).
Similar situation may occur in the continuous spectrum. The operator H is
called multiplicity free if the orthogonal projection onto any H-invariant sub-
space is a spectral projection. Clearly, [Ji pp ] and [Ji cont ] are always spectral
projections:

Further important examples will be given later.

For any vector 'ljJ one can define a real positive measure J.l1/J by setting

for Borel sets ~. This is called the spectral measure associated with 'ljJ. Spectral
measures play an important role in the computation of averages and transition
amplitudes: The matrix elements of functions of the energy operator can be
obtained as integrals with respect to these measures. The spectral projection

PE = P( ( -00, E])
is a monotonically increasing function of E in the sense that for E1 < E2
PE2 - PEl = P((E1,E2]);::: 0
(P ;::: 0 means ('ljJ,P'ljJ) ;::: 0 for all 'ljJ E Ji, which holds because ('ljJ,P'ljJ)
(P'ljJ, P'ljJ).) Therefore, with

dPE = PE+ dE - PE = P«E, E + dE])

one can write down the spectral decomposition of H in the general case:

H= J EdPE

This equation has the following meaning. For any 'ljJ E Ji and any function f,
continuous on a(H),

('ljJ, f(H)'ljJ) = Jf(E)dJ.l", (E)

where
J.l",(E) = J.l",(-oo,E]) = IIPE'ljJ1I 2
This function increases monotonically with E and is upper semicontinuous (in
jumps takes on the higher value). It fully determines the spectral measure J.l",.
In the above equation the integration is done with this measure.
Here we open a parenthesis on measures. The main line of the discussion
continues in Section 5.8.
506 A. Siito

5.6. Measures

Let f-t(x) be a real monotonically increasing upper semicontinuous function on

JR., finite at any finite x. With this function is associated a measure on the
Borel sets ~ of JR.:

f-t(~):= L df-t(x) = J Xil(x)df-t(x)

which is nothing else than the total variation (increase) of f-t on ~, written as
the Stieltjes-Lebesgue integral of Xil(X) with respect to f-t(x). (The notation
should not confound the reader: f-t with a real number argument means the
function, with a set argument the measure. In this way, f-t(x) = f-t(( -00, xl)
and f-t({x}) is the measure of the one-point set {x}.) The set function thus
obtained satisfies, indeed, the properties of (positive) measures, namely, it is
nonnegative, vanishing on the empty set and countably additive:
00

f-t(U~l~') = Lf-t(~.)
.=1
if the ~. are pairwise disjoint sets.
Conversely, a positive measure f-t which is finite on the semi-infinite intervals
(-00, x] can be used to define a monotonically increasing upper semicontinuous
function, denoted also by f-t, by setting

f-t(x):= f-t((-oo,x])

This is called the distribution function of the measure f-t.

One can define the derivative of the measure f-t with respect to the Lebesgue
measure at the point x as

ddf-t
m
(x) = limf-t(J)fm(J)
J!x
~ 00
provided the limit exists. The limit is taken on open intervals containing x
and shrinking to x. If the distribution function is continuous and differentiable
at x (perhaps with infinite derivative) then its derivative f-t'(x) = (df-tfdm)(x).
Oppositely, if (df-tfdm)(x) finitely exists then f-t(x) is continuous and differen-
tiable at x and f-t'(x) = (df-tfdm)(x). In the points of discontinuity of f-t(x),
f-t'(x) does not exist, but (df-tfdm)(x) exists and is infinite. The set

S = {x : (df-tfdm)(x) does not exist finitely or infinitely}

is a part of the set where f-t'(x) does not exist. Therefore we can apply a
theorem ([124], Ch. IV, Theorem 9.1), valid for functions which are locally of
bounded variation (like f-t(x)), to conclude that both the f-t and the Lebesgue
(m-) measure of S is vanishing, so (df-tfdm)(x) is well-defined apart from a set
of zero f-t- and Lebesgue measure.
 SCHRODINGER DIFFERENCE EQUATION ... 507

Let A be a Borel set. We say that a measure t-t is concentrated on A if

t-t(AC) = 0 (AC = R \ A, the complement of A). The smallest closed set on
which t-t is concentrated is the support of t-t, supp t-t. So if t-t is concentrated on
A then supp t-t C iI, the closure of A.
An x E R is a point of increase of the distribution function t-t if for every
c > 0, IL(X + c) > t-t(x - c).
Problem 1. supp t-t is the set of points of increase of t-t( x).
A Borel set A is called an essential support of t-t if t-t( A C) = 0 and for any Ao C
A such that t-t(Ao) = 0, m(Ao) = 0 as well. So the support is unique and closed,
the essential support is unique only apart from sets of zero t-t- and Lebesgue
measure and is normally not closed. Countable unions and intersections of
essential supports are essential supports. An essential support may not be a
subset of the support. An example is Au B where A is an essential support
and B a nonempty Borel set with m(B) = 0 and B n supp t-t = 0. However,
if A is an essential support then An supp t-t is again an essential support, the
closure of which is the support.
It follows from the definitions, that any measure t-t has an essential support
on which (dt-t/dm)(x) exists (maybe infinite).
Any measure t-t can be decomposed into pure point and continuous part,
t-t = t-tpp + t-tcont· These are respectively characterized by the equations

that is, t-tpp is concentrated on a countable set of points, and

t-tcont({X}) = lim
dO
t-t([x - c, x + c]) = 0,
the t-tcont-measure of every x E R is zero. If the function t-t is continuous, the
corresponding measure is purely continuous; if it makes jumps of heights at in
the points X t ,
dt-tpp(x) = L
at8(x - xt)dx (5.6)

For t-tpp there exists a smallest essential support: the set of all points of dis-
continuity of the function t-t.
If t-t and v are two measures, we say that t-t is absolutely continuous with
respect to v, and write J.l «: v, if t-t(fl) = 0 whenever v(fl) = O. In words, if
v is concentrated on a set, then t-t is concentrated on the same set. If t-t and
v vanish on the same Borel sets, i.e., they are mutually absolutely continuous,
we call them equivalent, and write t-t "-' v. Equivalent measures have the same
support and essential supports. Most often v = m, the Lebesgue measure. If
we say only that t-t is absolutely continuous, we mean it with respect to the
Lebesgue measure. Clearly, if t-t is absolutely continuous, it is continuous.
Problem 2. If two measures are absolutely continuous and have a common
essential support then they are equivalent.
508 A. Siito

Problem 3. Find an absolutely continuous and a pure point measure whose

supports are not essential supports. Find a measure, whose support is the
smallest essential support.
A function f defined on IR is said to be absolutely continuous on a finite or
infinite interval J if to every c > 0 there exists a 8 > 0 such that for any finite
collection of disjoint open subintervals (al' bl ), (a2' b2), ... , (an, bn ) of J

L If(b.) - f(a.)1 < c whenever L(b. - a.) < 8 (5.7)

Since 8 does not depend on n, the property holds also if the number of intervals
is infinite.
Problem 4. A measure is absolutely continuous if and only if its distribution
function is absolutely continuous on every finite interval.
We say that two measures J.t and v are mutually singular, and write J.t .1 v,
if there are disjoint Borel sets A and B such that J.t is concentrated on A and v
is concentrated on B. Most often v = m and we simply say that J.t is singular.
Any pure point measure is obviously singular.
Problem 5. Any essential support of a singular measure is of zero Lebesgue
measure.
A measure is said to be singular continuous (with respect to the Lebesgue
measure) if it is singular and continuous. An example is given in the next
subsection. Thus, J.t has the following decompositions into pairwise mutually
singular measures:

J.t = J.tpp + J.teont = J.tae + J.tsing = J.tae + J.tse + J.tpp

Problem 6. Show that any measure uniquely determines its pp, ac and sc
parts.
According to the value of (dJ.t/dm)(x), one can find essential supports for
the different components as follows (taken over from Gilbert and Pearson [52]).
Let
B = {x E IR: (dJ.t/dm)(x) exists, 0 ~ (dJ.t/dm)(x) ~ oo}
The sets M, Mac, Msing, M se , Mpp defined below are essential supports respec-
tively for J.t, J.tae, J.tsing, J.tse, J.tpp:

M = {x E B: 0 < (dJ.t/dm)(x) ~ oo}

Mac = {x E B: 0 < (dJ.t/dm)(x) = J.t'(x) < oo}
Msing = {x E B: (dJ.t/dm)(x) = oo}
Mse = {x E B: (dJ.t/dm)(x) = 00, J.t({x}) = O}
Mpp = {x E B: (dJ.t/dm)(x) = 00, J.t( {x}) > O}

Obviously, Mac, Mse and Mpp are pairwise disjoint sets. They are subsets of
supp J.t and Mpp C supp J.tpp (Mpp is the smallest essential support of J.tpp),
but Mac, Msing and Mse may not be parts of the supports of the respective
 SCHRODINGER DIFFERENCE EQUATION ... 509

measures. As mentioned earlier, by intersecting them with the corresponding

supports, we still get essential supports. It may happen, however, that Mac
is nonempty but J.Lac = 0: surely this is the case if m(Mac) = 0; or Msc is
nonempty, even uncountable, but J.Lsc = O. For more details see [52).

5.7. Cantor function

The Cantor function is a continuous monotonically increasing function which
grows from 0 to 1 exclusively in the points of the middle-thirds Cantor set

{L 3- x
00

C= n n : Xn = 0 or 2}
n=l

In the interval [0,1) it is given by the formula

= L 2- = L 3-
00 00

a(x) n- 1 x n if x n xn , Xn E {O, 1, 2}
n=l n=l

with the remark that triadic rationals are taken with their infinite representa-
tion. More interesting, it is entirely determined by the conditions

1. 2a(x) = a(3x) , 0 ~ x ~ 1/3

2. a(x) +a(l- x) =1
3. a( x) is monotonically increasing
4. a(x) = 0 if x < 0 and a(x) = 1 if x> 1
from which the moments or the Fourier transform of a can be computed. For
example, the latter is

J e'txda(x) = e,t/2 IT
n=l
cos t/3 n

The function is represented by a 'devil's staircase'. The measure a is singular

continuous, because it is continuous and its support, C, is of zero Lebesgue
measure. Observe that the Fourier transform does not tend to zero when It I
goes to infinity (check it with t = 27r3 k ) as it would do if a were absolutely
continuous.
The support of a singular continuous measure may have positive Lebesgue
measure. Define, for example, an odd function (J.L( -x) = -J.L(x)) on IR by
setting
= LTq L 2- Pa(x -
00 00

J.L(x) p/q) ifx~O

q=l p=o
This is a continuous function increasing in every point of lR: J.L( x + e) - J.L( x) > 0
for every real x and positive e. J.L(x) is therefore the distribution function of
510 A. Siito

a continuous measure f.t whose support is R Furthermore, U rEQ (C + r) is a

Borel set (a countable union of closed sets) of zero Lebesgue measure and is an
essential support for f.t; thus, f.t is singular continuous. This situation may occur
for Schrodinger operators with unbounded potentials. If Vn = tan no: and 0: is a
Liouville number (an irrational number which is extremely well approximated
by rationals), all the spectral measures are purely singular continuous, and
there are spectral measures the support of which is IR ([130]).
Problem 1. Give an example of a pure point and a singular continuous
measure having a common essential support.

5.8. Spectral measures and spectral types

In the comparison of two measures, f.t. and v, yielding f.t «: v, f.t '" v or f.t .1
v, either or both can be projection-valued, and a projection-valued measure
has support, essential support and uniquely determined pp, ac and sc parts,
like real measures. In particular, for any vector 1/J, f.t", «: P, where P is the
spectral resolution of the identity. Similar relations for the pp, ac, sc parts will
be written down below. Let us start with the decomposition of the spectral
resolution of the identity:

P = Ppp + Pcont = Pac + Psing = Ppp + Pac + Psc

Is it true that the different terms are spectral projection-valued measures? The
answer is yes, but for the moment it is not even clear that Pac(~) and Psc(~)
project onto H-invariant subspaces.
By definition,

P( {E}) = P(~ n lTpp(H»

which is a spectral projection. The smallest essential support of Ppp is lTpp(H).

Notice that
Ppp(~) = P(lTpp(H»P(~) = ["lipplP(~)

where the two projections commute. Let 1/J E "lipp . Then

f.t",(IR) = 111/J112 = IIPpp(IR)1/J1I2 = L f.t",({E})

EEO'pp{H)

i.e., f.t", is a pure point measure and f.t", «: P pp .

Also, by definition, for any Borel set ~,

P( {E}) = P(~ \ lTpp(H»

is a spectral projection. The support of Pcont is lTcont(H). Notice that

 SCHRODINGER DIFFERENCE EQUATION ... 511

where the two projections commute. Let 'Ij; E H cont . Then for any E E lR

i.e., J.L1jJ is purely continuous and J.L1jJ « Pcont ·

A purely continuous spectral measure can still be decomposed into abso-
lutely and singular continuous parts. We would like to perform the analogous
decomposition on Hcont, (Jcont(H) and Pcont .

Problems.

1. J.L1jJ is concentrated on a Borel set A if and only if P(A)'Ij; = 'Ij;. As a

consequence, if'lj; E P(A)H, supp J.L1jJ cA.
2. J.LH1jJ « J.L1jJ. In particular, if J.L1jJ is absolutely or singular continuous then
J.LH1jJ is absolutely or singular continuous, respectively. (Remark. We
know already that if J.L1jJ is pure point, J.LH1jJ is also pure point because
'Ij; E Hpp and HHpp C H pp .)
3. If J.L1jJ = (J.L1jJ )pp + (J.L1jJ )ac + (J.L1jJ ).c then there exist orthogonal vectors 'lj;PP,
'lj;ac and 'lj;sc such that 'Ij; = 'lj;PP + 'lj;ac + 'lj;sc and (J.L1jJ )pp = J.L1jJPP, (J.L1jJ )ac =
J.L1jJac and (J.L1jJ )sc = J.L1jJsc. Hint. Choose disjoint essential supports and
apply 1.
4. Let 'lj;i be a sequence of vectors converging to a vector 'Ij; (Le., 11'Ij;' - 'lj;11 ->
0). Suppose that J.L1jJ' are singular measures. Then J.L1jJ is singular. Hint.
(i) J.L1jJ' -> J.L1jJ on Borel sets. (ii) Let Ai be an essential support to J.L1jJ"
then UA, is an essential support to J.L1jJ.

It follows from Problem 2 that the vectors generating a purely ae spectral

measure form an H-invariant subspace Hac; similarly, the vectors giving rise
to a purely se spectral measure form an H-invariant subspace Hsc. According
to Problem 3, Hac and H.c are orthogonal and span H cont . Thus Hac and H.c
are necessarily closed and

H cont = Hac EB Hsc

(Problem 4 shows also explicitly that H. c is closed.) Now we can consider

the restrictions of H to Hac and to H.c: these are selfadjoint operators, their
spectra are (Jac(H) and (J.c(H), respectively. Absolutely continuous and singu-
lar continuous spectral measures are concentrated respectively on (Jac(H) and
(J.c(H).
In summary, we obtained the following decompositions:

Correspondingly, for any 'Ij; E H,

512 A. Siito

and
J.L1/> = J.L1/>PP + J.L1/>ac + J.L1/>.c
What about the decomposition of Pcont ? Because

where the two projections are orthogonal, we get

with commuting operators in the products.

Problem 5. Show that the projection-valued measures [?-lac]P and [?-lsc]P are
absolutely and singular continuous, respectively.
Since the decomposition of Pcont into ae and se parts is unique, we obtained
that
(5.8)
So Pac(~) and Psc(~) project onto H-invariant subspaces within ?-lac and ?-lsc,
respectively. It remains to show that they are spectral projections.
Let A and B be disjoint essential supports for Pac and Psc, respectively,
which are disjoint also from eTpp(H). Then

= Pac (A) = peA) and [?-lsc] = Psc(B) = PCB)

[?-lac]

Indeed, for any '¢ E ?-l, [?-lac]'¢ = '¢ac while

P(A)'¢ = p(A)'¢ac + P(A)('¢SC + '¢PP) = Pac(A)'¢ + (Psc(A) + Ppp(A)),¢

= Pac(A)'¢ = Pac(JR)'¢ = '¢ac
The proof is similar for [?-lsd. The result shows that Pac(~) and Psc(~) are
the spectral projections peA n~) and PCB n ~), respectively. Using Eq.(5.8)
and the definition of J.L1/>, it is immediately seen that J.L1/J «: Pac if'¢ E ?-lac and
J.L1/> «: Psc if '¢ E ?-lsc.
At last,

where
eT(H) = suppP
eTpp(H) = ess.supp Ppp
(5.9)
eTac(H) = suppPac
eTsc(H) = sUPPPsc
It is also important to know, how to reconstruct the spectrum from real
measures. Clearly, if J.L is any real measure equivalent to P then in Eq.(5.9) P
can be replaced by J.L.
Problem 6. Let J.L be a measure and J.L1/> «: J.L for all '¢ E ?-l. Then P «: J.L.
Problem 7. J.LOt1/>l +131/>2 «: J.L1/>1 + J.L1/>2.
 SCHRODINGER DIFFERENCE EQUATION ... 513

Problem 8. Let '¢1, '¢2, ... be a normalized basis in 1£ and c. > 0 for i = 1,2, ...
such that E:1 c. < 00. Then

In the particular case of 1£ = (2(Z) and H the Schrodinger operator, we have

a much simpler result. Let 80 and 81 be the unit vectors concentrated on 0 and
1, respectively. Then
P '" J.l60 + J.l61 (5.1O)
Call J.l the measure on the right side. J.l« P is obvious, one has to show
P « J.l. Let ~ be such that J.l(~} = O. Then P(~}8° = P(~}81 = 0 and, as a
consequence,
= p(~}Hn81 = 0 for n = 1,2, ...
p(~}Hn8°

l,From this we can conclude that P(~} = 0 because of the following.

Problem 9. The set of vectors {Hn8°,Hn81}~=0 is a basis in (2(Z).
Clearly, Eq.(5.1O} holds if 0 and 1 are replaced by any two successive integers.
The sum of spectral measures is, in general, not a spectral measure.
Problem 10. J.l.pl+.p2 = J.l.pl + J.l.p2 if and only if the two vectors are in
orthogonal H-invariant subspaces.
Therefore, for the Schrodinger operator the right member of the relation
(5.10) is not a spectral measure, in general (but it is a spectral measure for the
free Laplacian, see Section 5.9).
If H has a pure point spectrum, and for the basis in Problem 8 we choose
an orthonormal set of eigenvectors, E c.J.l.p' is equal to the spectral measure
belonging to E o:.'¢', where 10:.1 2 = c•. It is interesting to remark, that there
always exist spectral measures equivalent to P, even if the spectrum is not pure
point. The construction is suggested by Problem 10. Let A be an H-invariant
subspace. A vector '¢ E A is a cyclic vector in A if any tp E A can be obtained
as
N
tp = N-+oo
lim "" O:Nn(Hn'¢}
~
n=1
In this case, J.l'P « J.l.p for all tp E A, as one can see from Problems 2, 4 and 7.
Not all H-invariant subspaces contain cyclic vectors (a counterexample is a
subspace belonging to a degenerate eigenvalue), but 1£ can always be written
as
1£ = EBA. , (5.11)
a finite or infinite direct sum of orthogonal H-invariant subspaces with cyclic
vectors. Let '¢. be a normalized cyclic vector in the ith subspace and 0:. =F 0,
E 10:.1 2 < 00. Any tp E 1£ can be written as tp = E 13.tp· where tp' EA•.
According to Problem 10,

J.l'P = J.lEI3.'P· = L 113.1 2 J.l'P' « L IO:il 2 J.l.p. = J.lEa •.p·

514 A. SiltS

and, according to Problem 6, P '" J.LEa,,,,"

The decomposition (5.11) is nonunique. We can start with any 'ljJI, choose
Al as the smallest subspace containing Hn'IjJi for every nonnegative integer
n, choose any 'ljJ2 orthogonal to AI, and so on... . The smallest number of
subspaces we need for the decomposition may be called the multiplicity of the
spectrum of H. Problem 9 shows that the multiplicity of the spectrum of a
Schr6dinger operator on £2(1.) is 1 or 2. If the multiplicity is 1, that is, there
exists a cyclic vector (in H), H is multiplicity free also in the sense defined in
5.5.
Multiplicity may not be uniform on the spectrum, therefore it is appropriate
to define it also locally: Let ~ be a Borel set such that P(~) # 0, and Ht:.. =
P(~)H. The decomposition (5.11) can be performed on Ht:.., and the smallest
number of subspaces we need to it is called the multiplicity of the spectrum on
~. If ~ is an eigenvalue, we obtain the usual notion of multiplicity.

5.9. A spectral measure for H o

Let H = £2(1.), H = H o = T+T- I (recall: T is the left shift). We are going to

compute explicitly J.Loo. The way to proceed is to find a measure J.L such that

(6°, H n6°) = JEndJ.L(E) for all n

Then we can identify J.L with J.Loo.

For any unitary operator U

Choose for U the Fourier transformation which maps £2(1.) onto L~er.[O, 1]: For
any 'IjJ E £2(1.)
L
00

(U'IjJ)(x) = 'ljJne21rmx , X E [0,1]

n=-oo

In particular, (U60)(x) == 1 and UT±IU-I correspond respectively to the mul-

tiplication by exp(=r=211"ix); therefore UHU- I corresponds to the multiplication
by 2 cos 27rx. This gives

(6 0,Hn60 ) = r (2 cos 211"x)ndx = 2 ior! (2 cos 27rx)ndx = 12 En 11" ~

io
l

4 - E2
-2

where we applied the substitution E = 2 cos 211"x. From here we recognize

dJ.Loo(E) = { ~E/1I"v'4 - E2 ifEE(-2,2)

otherwise
By integration,
11.
J.Loo(E) = - + - arcsm(E/2) if -2 S; E S; 2
2 11"
 SCHRODINGER DIFFERENCE EQUATION ... 515

and 0 below -2 and 1 above 2. This function is manifestly absolutely con-

tinuous, thus Ho has an absolutely continuous spectrum in [-2,2). Because
of the shift-invariance of H o, J..L51 = J..L50 and thus P '" J..L50' This shows that
u(Ho) = [-2,2) and the spectrum is purely absolutely continuous.
Notice that 80 is not a cyclic vector: H n 8°, n = 1,2, ... generate only the
'even' subspace,
£ = {1/J : 1/J-k = 1/Jk for all k}
This subspace is Ho-invariant and the orthogonal projection onto it is not a
spectral projection (cannot be written as a function of Ho only). Therefore Ho
is not multiplicity free. In fact, the spectrum is uniformly of multiplicity 2.

5.10. f2(Z) versus f2(N)

Consider the matrix (5.1) of the SchrOdinger operator on f2(Z) and replace the
matrix elements H_1,0 = HO,-l = 1 by zeros. What kind of effect this tiny
modification can have on the spectrum of the modified operator H'? According
to the'classical' Weyl theorem ([115), Vol. IV),

that is, apart from isolated eigenvalues, the two spectra coincide. The matrix
of H' is of block-diagonal form, so H' can be written as H' = HI EBHr , where HI
and Hr are selfadjoint operators acting on f2( -N\ {O}) and f2(N), respectively.
Thus,
U(H') = u(HI) U u(Hr)
For strictly ergodic potentials uess(Hd = uess(Hr), and, hence, = uess(H).
The fate of spectral types under such a 'finite-rank' perturbation is partly
uncertain. Point and singular continuous spectra may change into each other
(Gordon [55), del Rio, Makarov, Simon [122)). However, the absolutely contin-
uous spectrum is robust,

(5.12)

and this holds, whatever be the (selfadjoint) boundary condition at 0, defin-

ing HI and Hr. The reason is that the singular spectrum is sensitive to the
boundary condition at 0, while the absolutely continuous spectrum can be
fully characterized by the large-time behaviour of propagating wave packets
and, hence, is insensitive to local perturbations. In fact, it is stable under more
general (trace class) perturbations (see e.g. [77), Chapter X, Theorem 4.4).

5.11. Asymptotic behaviour of generalized eigenfunctions. Subor-

dinacy

We started Section 5 with a remark about the sufficiency, for Physics, of polyno-
mially bounded solutions. This remark is based on the following deep theorem:
516 A. Siito

The spectral resolution of the identity has an essential support £ such that for
every E E £ there exists a solution 1/J of the SchrOdinger equation satisfying
the following condition: For any 6 > d/4 (d is the space dimension), with a
suitably chosen positive constant c = c( 6, E),

(5.13)

The most general form of this theorem (going beyond Schrodinger operators)
can be found in Berezanskii's book [22]. The Schrodinger case in the continuum,
with precise conditions, assertions and proof is described in Simon's survey
paper [128] (also in [29]). A simple proof for the (2(Z) case, due to Lacroix, is
presented in the book of Bougerol and Lacroix [25].
Clearly, the same result is valid separately for Ppp , Pac and Psc and is non-
trivial for Pac and Psc. It, however, does not permit to distinguish between
these two cases.
It is expected that knowing the whole family of solutions for a given energy
E, one can decide whether E is in the spectrum and which spectral type it
belongs to. The right notion to deal with this question is subordinacy. It was
introduced by Gilbert and Pearson [52] for the Schrodinger differential equation
on the half-line, extended by Gilbert [51] to the problem on JR, and by Khan
and Pearson [78] to the discrete equation on N. A discussion of it can be found
in Pearson's book [108]. Results for the equation on Z can be safely deduced
from these works.
For a two-sided complex-valued sequence 1/J and an integer N let
N
lI'l/JilN = (L* l1/JnI 2)1/2
n=O

The star means that for N < 0 the summation goes over n = 0, -1, ... , N.
Fix E. A solution 1/J of the Schrodinger equation with energy E is said to be
subordinate at +00 (-00) if for every linearly independent solution tp,

as N -+ +00 (-00). Clearly, there can be at most one subordinate solution

at either side, and it suffices to check the condition on a single tp linearly
independent of 1/J. Subordinate solutions for real energies are real.
Below, by 'solution' we mean a solution of the Schrodinger equation for given
E. Consider the following sets.
• M' = JR \ {E E JR : there exist two solutions, one which is subordinate at
+00 but not at -00, the other which is subordinate at -00 but not
at +oo}.
• M~c = {E E JR : every solution is nonsubordinate at +oo}
U{E E JR: every solution is nonsubordinate at -oo}
• M: ing = {E E JR : there exists a solution which is subordinate at ±oo}
 SCHRODINGER DIFFERENCE EQUATION ... 517

• M~c = {E E IR: there exists a solution which is subordinate at ±oo but

is not in l2(Z)}
• M~p = {E E IR : there exists a solution which is subordinate at ±oo and
is in l2(Z)}
Then M' is an essential support for P and M~p = upp(H). (M' is the disjoint
union of M~c' M~c and M~p and is a subset of u(H).) Furthermore, Pac ::j:. 0 if
and only if m(M~c) > 0, in which case M~c is an essential support for Pac (and
so is M~c n uac(H)). If Psc ::j:. 0, M~c is an essential support for Psc (and so is
M~c n usc(H)). If J is a real interval, Pac(J) = Ppp(J) = 0 and M~c n J is an
uncountable set, then Psc(J) > O.
The above characterization shows that physicists are not quite right when
saying that in the singular continuous spectrum 'all the solutions are critical'.
In fact, there is one solution (the subordinate one) which is less (or more)
'critical' than the others and shows a clear analogy with an eigenvector. Notice,
however, that a subordinate solution may not decay to zero at ±oo.
The union giving M~c corresponds to the union (5.12). On the intersection
of the two parts of M~ the multiplicity of the spectrum is 2, otherwise it is 1
(Kac, [75], [76]).
Recently, AI-Naggar and Pearson [3] developed further the characterization
of the absolutely continuous spectrum in the case of the differential equation
on the half-line. Due to Eq.(5.12), their result applies to the l2(Z) case in the
form presented below.
For fixed real E a complex solution 1/J of the Schrodinger equation with energy
E is called rotating at +00 (-00) if
N N
L* 1/J~/ L* l1/Jnl 2 -+ 0
n=O n=O

as N -+ +00 (-00). Notice that a complex solution for real E is the linear
combination of two real solutions with complex coefficients. Let

M::C = {E E IR : there is a solution which is rotating at least at one of ±oo}.

Then, if m(M:c) > 0, Pac ::j:. 0 and M:c is inside an essential support of Pac.
Unfortunately, it is not known so far whether M::C itself is an essential support
for Pac.

6. Schrodinger equation with strictly ergodic potentials

6.1. Strict ergodicity

The first step is to check minimality and unique ergodicity of the sequence
defining the potential. We discuss this question only briefly; the reader may
consult some other lectures of this School ([1], [40], [114]; see also [95]) and
518 A. Siito

references therein. The minimality of a sequence (and then that of the hull)
is equivalent to the infinite repetition, with arbitrary precision and bounded
gaps, of every finite segment. In the case of sequences taking values from a
finite set, once minimality is known, unique ergodicity means that the word
frequencies exist (the defining limits converge).
The cosine potential of the Almost-Mathieu equation is a uniformly almost
periodic sequence: For any e there exists a sequence of integers, {n.}, with
bounded gaps such that for all i and all n

(see e.g. [24]). This implies also strict ergodicity. A weaker form of almost
periodicity (in the mean- or Besicovitch sense, see [24], which is equivalent
to the existence of an atomic Fourier transform of the sequence, with square
summable coefficients) also implies strict ergodicity. Sturmian sequences, or
more generally, sequences generated by the circle map with A being the union
of a finite number of intervals of the type [a, b) (cf. point (2) in Sec.2) belong
to this class. The Fourier coefficients of the I-periodic function XA(x) are easy
to compute. It is worth noticing that in this case the Fourier series converges
everywhere but may not represent the sequence in a finite number of points. If
this occurs, it is the function XA, and not the series, which defines a strictly
ergodic sequence.
Some substitutional sequences, which are not Sturmian, also admit an atomic
Fourier transform. Examples are the regular paper-folding and the period-
doubling sequences. As one-sided sequences, these are strictly ergodic.
In general, for every primitive substitution ~ one can build up two-sided
minimal sequences, which are also uniquely (and, hence, strictly) ergodic ([56],
[102]; see also [39]). The construction goes by concatenating a left- and a
right-sided fixed point of some power of~. More precisely, one can find two
letters a and b and an n ~ 1 such that ~n(a) = ... a, C(b) = b... and, with
'" = ~n, both u = ",OO(a) = ... a and v = ",OO(b) = b... contain the word ab
and the two-sided sequence uv is minimal and uniquely ergodic. Among oth-
ers, this holds for the Thue-Morse and the Rudin-Shapiro sequences, although
their Fourier transform is not atomic. Different constructions of a two-sided
sequence (for example, via symmetric extension) may violate minimality (the
starting sequence may not be recurrent). If both sides are related to the same
substitution, the essential spectrum will not suffer, but the singular spectral
measures can be seriously perturbed. (Compare with Section 5.10.)

6.2. The spectrum of H(w) = Ho + V(w)

Let us start with three remarks.
1. The shift T is defined on £2(Z) by (T'IjJ)n = 'ljJn+1' This is a unitary
operator, therefore Tn HT-n is unitarily equivalent to and, thus, has the same
spectrum as H for any finite n. This obviously holds for any potential. Notice
 SCHRODINGER DIFFERENCE EQUATION ... 519

that THoT- 1 = Ho and TV(w)T-l = V(Tw), thus

TnH(w)T- n = H(Tnw)

2. If Vk is a sequence of potentials converging pointwise to a bounded po-

tential V, then Hk = Ho + V k tends strongly to H = Ho + V, i.e., for any fixed
'¢ E £2(/£), Hk'¢ --+ H'¢. Indeed,

II(H k - H),¢112 = L: IV: - Vn12 1'¢n1 2 --+ 0 as k --+ 00

n

3. If the bounded selfadjoint operators Hk strongly converge to the bounded

selfadjoint Hand L). is an open interval not intersecting O'(Hk) for any k, the
spectral projections X6.(Hk) also strongly converge to X6.(H) ([77), Theorem
VIII.1.15); on the other hand, they all vanish and, hence, X6.(H) = O. This,
however, means that L). does not intersect O'(H). In short,

(6.1)

Problem 1. If V(w) is minimal, O'(H{w)) is independent of w. Hint. Use

the above remarks to prove that for any wand w', O'{H{w)) ::) O'{H{w')) and
O'{H{w')) ::) O'{H(w)).
The w-independent spectrum is denoted by O'{H).
Minimality also implies that there is no isolated point in the spectrum. Even
less is sufficient:
Problem 2. If the potential is bounded and recurrent then O'{H) = O'es.{H).
Hint. If E is an eigenvalue and '¢ the corresponding eigenvector, cpn = T-k n ,¢
with suitably chosen almost-periods k n is a Weyl sequence.
We emphasize that the proofs of the above problems do not use ergodicity,
only minimality or recurrence. This is the more interesting, because the same
statements were verified for p-a.e. w by using ergodicity without minimality,
see Section 5.2 for the random case.
For one-dimensional ergodic potentials Pastur proved that any E is p-almost
surely not an eigenvalue (Sec. 5.2). In fact, for this holding true, we need
less than ergodicity: If 0 is a shift-invariant set of potentials and p is a shift-
invariant probability measure on 0, which is defined on cylinder sets, any fixed
E is not in O'pp(H(w)) with p-probability 1. Indeed,

TX{E}(H(w))T- 1 = X{E}(H(Tw))
and

j trX{E}(H(w))dp(w) = jL:(on,X{E}{H(w))On)dp(w)
= L: j(OO,X{E}(H{T-nW))OO)dP{w) = f
n=-oo
j(OO,X{E}(H(W))hO)dP(W)
520 A. SiitO

All the integrals exist, the first integral takes value in [0,1] (X{E}{H{w))=0
or it projects to a one-dimensional subspace of &(Z)) and the last sum gives 0
or 00. The value of the first integral is therefore 0, implying X{E}{H{w» = 0
for p-a.e. w.
The content of the above statement is that O'pp{H{w» changes when w
changes. On the other hand, the closure of O'pp{H{w» may not change:
For ergodic potentials there are closed sets Epp , Eac and Esc such that

O'pp{H{w)) = Epp
O'ac{H{w» = Eac
O'sc{H{w)) = Esc

for p-a.e. w (Kunz, Souillard [91]). This theorem is the analogue of the as-
sertion of Problem 1, but is a great deal more subtle than the constancy of
the spectrum. Moreover, the result surely does not hold for all w, even if the
potential is strictly ergodic: a counterexample of [71] was mentioned in Section
5.2.

6.3. Integrated density of states

For finite systems the integrated density of states (IDS) as a function of E

counts the number of eigenvalues per unit volume below E. Let H~~2 denote the
restriction of H{w) to the interval [-L, L] with Dirichlet boundary condition.
The IDS for H~~2 is

Nw,L(E) = 2L ~ 1 x ( number of eigenvalues :$ E of H~~b (6.2)

For uniquely ergodic potentials Nw,L has an w-independent limit, N{E), when
L goes to infinity. As a limit of monotonically increasing functions, N{E) is
monotonically increasing and can be shown to be continuous. It is, therefore,
the distribution function of a continuous measure on JR, denoted also by N. Fur-
thermore, it can be shown that suppN = O'(H). The use of Dirichlet boundary
condition in the construction is not exclusive: any other boundary condition
yielding a Hermitian restriction Qf H{w) leads to the same IDS. (This, however,
may not be true in higher dimensional spaces.) Obviously, the eigenvalues have
to be counted with multiplicity.
The IDS for uniquely ergodic potentials can also be obtained by dealing
directly with the infinite system. Fix any V (w) and let J.L6n (E), nEZ, be the
distribution functions of the spectral measures for H(w), associated with the
canonical basis in [2(Z). Their average

1 L
NdE) = 2L + 1 L
n=-L
J.L6 n (E) ,
 SCHRODINGER DIFFERENCE EQUATION ... 521

which is the distribution function of a measure (NL), converges to N(E) as L

goes to infinity ([8]).
A part of the above results is easy to understand. If E is outside the spec-
trum, fo(w) = /-t6o(E) is a continuous function on the hull and, because of
unique ergodicity, its average along trajectories exists and is independent of
w: this is N(E). If the spectrum is a Cantor set, the spectral gaps are dense
everywhere (see Section 5.3) and N(E) has a unique extension from the gaps
to IR into an upper semicontinuous function. This, of course, does not explain,
why is the spectrum a Cantor set and why is N(E) continuous. The IDS is
constant in the gaps with values taken from a well-defined set (see Section 6.9).
Recall that the (pointwise) convergence of the distribution functions does
not imply the convergence of the measures on every Borel set. For example, if
H(w) has a pure point spectrum, for each E E upp(H(w)),

Nr({E}) ~ 1/(2L + 1) -+ 0 = N({E})

but NL(upp(H(w))) = 1 for each L while N(upp(H(w))) = O. In general, the

infinite summation, which occurs because of the u-additivity of N L , does not
commute with the limit L -+ 00.
The above remark makes less surprising the observation, that the type of the
measure N may have nothing to do with the type of the spectral measures. For
example, in the case of random potentials, for p-a.e. w the spectrum is pure
point while H(w) generates the same, continuous, IDS. In particular, N is
absolutely continuous for smooth p. For Bernoulli distribution and sufficiently
large potential strength N contains a singular continuous component with a
support of positive Lebesgue measure ([28]). If the spectrum is a Cantor set
of zero Lebesgue measure, a continuous N is necessarily singular continuous:
examples are presented in Section 7. If the IDS is singular continuous, the
(differential) density of states can be given no meaningful definition.

6.4. IDS and Lyapunov exponent

In Section 3.2 we introduced the Lyapunov exponent, more precisely, 'h(E, V)

and, ± (E, V) for an arbitrary potential V. According to a theorem by Fiirsten-
berg-and Kesten [47], in the case of ergodic potentials there exists a function
,(E) such that for every fixed E, for p-a.e. w the four numbers coincide to give

. 1
,(E,w)
.
= Inl-+oo
hm - IlIn IIT1-+n(E,w)1I
n

and ,(E,w) = ,(E). A detailed study of many related questions, as, for
example, the uniformicity in E of the convergence can be found in Goldsheid's
work [53].
Let

S {(E,w): ,(E,w) does not exist or =f. ,(E)}

522 A. SiitO

SE = {W: (E,w) E S}
SW = {E: (E,w) E S} (6.3)

According to [47], for every E, p(SE) = o. By the Fubini theorem, (mxp)(S) =

o and, for p-a.e. w, m(SW) = O.
In the theorem of Furstenberg and Kesten the restriction to p-a.e. w is
essential, even if the potential is strictly ergodic. A counterexample is provided
by the Almost-Mathieu equation for oX > 2 and a a Liouville number, where
for any E E (T(H) the set SE, defined above, is nonempty ([8]). However,
for potentials generated by primitive substitutions, Hof [63] proved that S is
empty.
There is a remarkable formula connecting the Lyapunov exponent to the IDS,
found by Herbert and Jones [58] and Thouless [137]:

"'I(E) = J In IE - E'ldN(E') (6.4)

The main observation leading to this formula is the following. For fixed w
let H[~~11 denote the restiction of H(w) to the interval [1, L] with Dirichlet
boundary condition. It has an L x L matrix with V(wh, ... , V(W)L in the
diagonal, 1 everywhere above and below the diagonal and 0 elsewhere. Let 1/J
be the solution of H(w)1/J = E1/J with initial condition 1/Jo = 0, 1/J1 = 1. Then

det(E - H&~11) = 1/JL+l

Indeed, both members are polynomials of E of degree L, they have the same
roots and the same principal coefficient. Let E 1 ,L, ... , EL,L be the roots, then
L
L- 1 1n 11/JL+l(E)1 = L- 1 2)n IE - Ei,LI
<=1

When L ---+ 00, the two sides go to the respective sides of Eq.(6.4), apart from a
set of (E, w) of zero m x p measure. Equation (6.4) holds, in fact, for all E E C.
The proof is based on the subharmonicity of "'I(E) and is due to Herman [59]
and Craig and Simon [35].

6.5. Results on the set {E : "'I(E) = O}

Let us recall that for periodic potentials

u(H) = uac(H) = Go where Go = {E: "'I(E) = O}

For H o, the pure kinetic energy operator, we still have u(Ho) = [-2,2].
For almost periodic potentials Deift and Simon [38] found

m(Go ) ~ 4
 SCHRODINGER DIFFERENCE EQUATION ... 523

with equality if and only if the potential is constant. In fact, this result extends
to any ergodic potential (Last [92]).
Ishii [68] and Pastur [106] obtained the following result:
Let V(w) be a bounded ergodic potential and pw denote the spectral resolution
of the identity for H(w). Suppose that m(Go) > o. Then for p-a.e. w, P.:'c is
concentrated on Go.
The proof is easy: Choose a typical w, such that m(SW) = 0 (cf. Eq.(6.3)).
The complement of Go can be written as Go = Au B, where

A = {E: ,),{E,w) = ')'{E) > O} , B = Go n SW

Now P::'c(B) = 0 because P::'c is absolutely continuous and B is of zero Lebesgue
measure. On the other hand, for every E E A, either there is no polynomially
bounded solution or there is an exponentially localized solution. Therefore
P::'c{A) = 0 and, hence, P::'c{G o) = o.
There is a different way to formulate this result. We can drop out of Go any
set of zero Lebesgue measure and close the rest: P:;;' is still concentrated on
this set which is now closed and, therefore, contains O"ac(H(w)). A particular
closed set to which this remark applies is the essential closure of Go.
The essential closure of a set A is
A ess = {E : m(A n (E - c, E + c)) > 0 for all c > O}

A short inspection may convince the reader that :::t ess is indeed a closed set, it
is inside the closure of A, and what it does not contain from A is a set of zero
Lebesgue measure:
m(A \ :::tess) = 0
Thus, the Ishii-Pastur theorem can be brought into the form
0"ac(H(w)) C G oSs

holding for p-a.e. w.

What do we drop from Go when we take its essential closure? Imagine that,
apart from an absolutely continuous spectrum, H has also a singular continuous
spectrum on a disjoint set D of zero Lebesgue measure. If ')'(E) = 0 on D, D
belongs to Go but not to its essential closure. Similar is valid for a point
spectrum on which the eigenvectors are not exponentially localized.
Most remarkably, the converse of the Ishii-Pastur theorem is also true and
we have the following.
For p-a.e. w
(6.5)
Moreover, Go is an essential support for P.:'c.
This theorem was proven by Kotani for the differential equation ([86], [87])
and adapted by Simon to the difference equation ([129]).
Remarks.
(i) O"ac(H(w)) is the support of P:;;', see Eq. (5.9), but it may not be an essential
524 A. BiitO

support. In fact, what we add to Go when taking its essential closure, may be
a set of positive Lebesgue measure, see the example of Problem 5.6.3.
(ii) Since uac(H(w)) n Go is also an essential support for P:C, we have, in
particular ,
m(uac(H(w)) n Go} = m(Go}
for ~a.e. w.
(iii) The main content of the above theorems is that the absolutely continuous
spectrum of H(w} is ~almost surely nonempty if and only if m(Go} > o. The
'if' is clear from the preceeding remark; the 'only if' holds because m(Go} = 0
implies that the essential closure of Go is empty.

6.6. The role of periodic approximants

Given a bounded aperiodic potential V, one can define a sequence of periodic

approximants, V k , in such a way that Vk converges to V pointwise. According
to the second remark made in Section 6.2, Hk = Ho + Vk then tends strongly
to H = Ho + V and Eq.(6.1} holds for the spectra. Our aim is to minimize
the covering set on the left side of this equation, by choosing the best periodic
approximants.
The best periodic approximants of potentials of the type V (w}n = g( no: +w),
where g is a period-l function and 0: is irrational, are obtained by replacing
0: with its best rational approximants, O:k. For example, if 0: = (v'5 - 1}/2,
O:k = Fk-tlFk' where Fk is the kth Fibonacci number.
If V is a substitutional potential, V k can be chosen to be the periodic repe-
tition of the sequence !(wn ), evaluated on the word 1Jk(a)1Jk(b} (cf. Sections 2
and 6.1).

6.7. Gordon-type theorems

Minimality sometimes implies that the potential repeats itself (exactly or with
very good precision) on three neighboring intervals, one of which starting with
1, and this holds for an increasing sequence of interval lengths. In such cases,
it can be shown that the SchrOdinger equation has no solution decaying at the
infinity, and thus the spectrum is purely continuous. The first proof of this
kind was given by Gordon [54]. Suppose that

Then

and, applying the two sides of the Caley-Hamilton equation

(6.7)
 SCHRODINGER DIFFERENCE EQUATION ... 525

to the vectors IIt-L and lito (cf. Eq.(3.1», one finds that for any solution 'IjJ of
the Schrodinger equation

Clearly, if (6.6) holds for an increasing sequence L n , no solution can decay on

both sides, so there can be no eigenvector to H. Let us see two ex.amples.
(i) Avron,Simon [8]: For the Almost-Mathieu equation, for A > 2 and a Li-
ouville number (for every positive integer k there are integers Pk and qk such
that la - Pk/qkl ~ k- qk ), for a.e. w the spectrum is purely continuous. In
this case, Eq.(6.6) does not hold exactly but with an extremely large precision.
The spectrum is, in fact, purely singular continuous: this comes from the Ishii-
Pastur-Kotani theorem, because 'Y(E) > 0 for all E ([4]).
(ii) Delyon, Petritis [43]: For circle potentials (see Section 2) with A being an
interval, for every A, a.e. wand a.e. a the spectrum is purely continuous. In
order that Eq.(6.6) hold true, some weak condition on the continued fraction
expansion of a has to be imposed. This limits the result to almost every a.
A variant of the Gordon theorem uses only two intervals. Suppose that in
Eq.(6.6) the second equality is verified for an increasing sequence Ln. This sit-
uation arises, for example, with substitutional potentials, if the substitutional
sequence starts with a square. Applying the two sides of Eq.(6.7) to lito, one
obtains
1
max{ltr AL IIIIItLil , 1I1It 2L 11} ~ "2ll l1t oll
This yields the absence of decaying solutions at +00 for energies such that

tr Tl-+L" (E)
is a bounded sequence. In some cases one can show that this set of energies is
just the spectrum (see [135] and Section 7).

6.S. Kotani theorem for potentials of finite range

Any periodic potential on £2(Z) is of finite range, i.e., takes on a finite number
of different values. Any Schrodinger operator on £2(Z) with a periodic potential
has a purely absolutely continuous spectrum. Therefore, the following theorem
by Kotani [88] may be surprising. Let V(w) be an ergodic nonperiodic finite-
range potential, p the corresponding measure on the hull of V. Then, for p-a.e.
w, H(w) has a purely singular spectrum.
This theorem is at the origin of many results on Schrodinger operators with
circle- and substitutional potentials; we discuss them in the next section. It has
long been an open question, whether the restriction to p-a.e. w can be dropped,
and whether the spectrum is purely singular continuous for every w, if the
potential is strictly ergodic. A recent result by Hof, Knill and Simon [64] goes
in this direction: The authors show that for strictly ergodic potentials, either
H(w) has pure point spectrum for all w or there is an uncountable dense set,
526 A. SiitO

although of zero p-measure, in the hull for which the spectrum is purely singular
continuous. This latter case is shown to be realized for circle potentials if a
is irrational and A is a half-open interval. This includes Sturmian potentials.
Furthermore, for potentials generated by primitive substitutions, the Lyapunov
exponent -y( E, w) is independent of w (Hof, [63]); Kotani's theorem then implies
that the spectrum is purely singular for all w.

6.9. Gap labelling

Gap labelling is a book-keeping for spectral gaps. The IDS naturally assigns a
number to each gap: the (constant) value N(E) for E in the gap. The question
is then to characterize this set of numbers. The first example of gap labelling
for an almost periodic Schrodinger equation was given by Johnson and Moser
[73] in the continuum case. The method used in the discrete case is quite
different.
Problem 1. Let V be an L-periodic potential and suppose that H = Ho + V
has no missing gap. Show that N(E) = kj L, k = 1,2, ... , L - 1 in the gaps.
For strictly ergodic potentials, the values of the IDS in the gaps are taken
from (but do not necessarily exhaust) a set S, determined by an algebraic
theory, the K-theory. The application of K-theory to gap labelling was devel-
oped by Bellissard, Lima, Testard, Bovier and Ghez ([19], [16]). A detailed
discussion can be found in [13], [14].
Sometimes it is possible to describe S quite explicitly. For example, according
to Bellissard, Bovier and Ghez ([14], [13], [16], [27]), for circle and substitu-
tional potentials S is the module (smallest additive group) containing the word
frequencies of V(w), restricted to the interval [0,1). This module contains the
integers (the sum of the frequencies of cylinders with common base (n and k in
Eq.(2.4» is 1), therefore S is a group with respect to addition modulo 1, like
for periodic potentials.
For Sturmian potentials this gives

S = {ka + m(l- a) : k,m E Z} n [0,1) = {ka + m: k,m E Z} n [0,1)

Notice that for a = KIL, where K and L are relatively primes, this gives

I 2 L-l} ,
S= { L'L' ... '-Y

in accordance with the assertion of the Problem above and the example (ii)
in Section 4. If a is irrational, S is countably infinite and dense in [0,1]. If
there exist, indeed, gaps corresponding to values in a subset S' of S which is
still dense in [0,1], the spectrum is necessarily a Cantor set and the IDS is
continuous.
In our discussion of the periodic Schrodinger equation (Section 4) we in-
troduced the notion of a missing gap (or closure of a gap). When there are
 SCHRODINGER DIFFERENCE EQUATION ... 527

no intervals in the spectrum, as is often the case with strictly ergodic poten-
tials, what one can unambiguously assert is the absence of missing gaps or the
completeness of gap labelling. Surely, there is no missing gap if

RangN == {N(E): E is in a spectral gap} =S , (6.8)

the set of all the admissible values. (Notice that each admissible value is taken
on in at most a single gap.) As in the periodic case, there is no general method
to check whether or not all the gaps open. This is known to hold true in a few
cases, like the period doubling potential or the Fibonacci potential with A > 4.
Example. For the classical Cantor set, in the kth gap on the nth level (n =
0,1, ... j k = 1, ... , 2n) the Cantorfunction a(x) = (2k-l)/2 n+1. These numbers
form an additive group modulo 1. The same values are taken by the IDS of the
hierarchical Hamiltonian ([90), see also [89]) in the domain of the parameters
where the potential is limit periodic. In this case the gap labelling is complete.

7. Schrodinger equation with Sturmian and substitutional poten-

tials

7.1. Fibonacci potential

Fibonacci substitution was the first to be used to define a two-valued potential
(see Section 2) and to study the spectral problem of the corresponding SchrO-
dinger operator (Kohmoto, Kadanoff and Tang, Ostlund et al. [79], [80], [81],
Casdagli [30), Siit8 [135], [136]). This is also the first example where the trace
map ([2], [84], [111]) was fully exploited.
The Fibonacci sequence, as any minimal sequence, has almost-periods: these
are the Fibonacci numbers Fn (Fo = FI = 1, Fn+! = Fn + F n - I ). Choose, for
example,
Vn = AWn + l)aJ - LnaJ) (7.1)
then the almost-periodicity is expressed by the equations

Vi+Fn = Vi if n 2 3 and 1 ~ 1 ~ Fn
Vi-F2n = Vi if n 2 1 and 1 ~ 1 ~ F2n+!
The transfer matrices over the almost-periods replace the single transfer matrix
TI--+Lin the L-periodic case. For an L-periodic potential

(7.2)
Let

For the Fibonacci potential Eq.(7.2) is replaced by

528 A. Siitc5

This, together with det Mn = 1, implies for Tn = tr Mn the recurrence relation

The initial conditions (with the choice (7.1))

T-1 =2 , TO =E , T1 = E - >.
contain only 2 parameters, and this indicates that the recurrence must have a
nontrivial invariant. Indeed,

From the trace map and the invariant, the following results can be deduced
([135), [136]).

1. a(H) = {E : {Tn(E)}~=1 is bounded}

(Remember that for an L-periodic potential a(H) = {E : ItrT1 ..... L(E)1 S
2}.)
2. The spectrum of H(w = 0) is purely continuous (the proof is of Gordon-
type with two intervals).
3. "'t(E, w = 0) = 0 for all E E a(H).
4. The Lebesgue measure m(a(H)) = 0 (found by confronting the Ishii-
Pastur-Kotani theorem with Kotani's theorem for potentials of finite
range). As a consequence, a(H) is a Cantor set (because a(H) contains
no isolated point).
5. The spectrum of H(w = 0) is purely singular continuous.
6. For all E E a(H) all solutions of H(w = 0)'ljJ = E'ljJ are polynomially
bounded (Iochum, Raymond, Testard [66), [67]).

It is an open question whether the spectrum is purely singular continuous for all
w. Let us recall that TH(w)T- 1 = H(Tw) = H(w+a) is unitarily equivalent to
H(w), therefore the singular continuity holds true for w = ka (mod 1), k E Z,
which is a countable dense set in [0,1]. Due to [64), this result has recently been
extended to an uncountable dense set of w, still of zero measure (cf. Section
6.8).

7.2. General Sturmian potentials

The case of Sturmian potentials for arbitrary irrational a was investigated by

Bellissard, Iochum, Scoppola and Testard [17]. The methods applied to the
Fibonacci case can be extended to treat the general problem, even though this
is technically more involved. All the results found for the Fibonacci potential
remain valid and the same questions are unanswered.
 SCHRODINGER DIFFERENCE EQUATION ... 529

4/5
3/4

2/3

3/5

112

~i~~
r..".,
2/5

113
m
114
115 ,
'A'
",
IU

o
-2 -1 o +1 +2
E = energy
Fig. 2. - Spectrum versus a for Sturmian sequences potentials.

There is an interesting numerical work by Ostlund and Kim [105] and a

rigorous study by Bellissard, Iochum and Testard [18] on the a-dependence of
the spectrum of H = Ho + v(a), where

v~o:) = X[O,a) (na) ,

cf. Section 2. For rational a the potential is periodic and the spectrum is the
union of a finite number of intervals. When approaching a rational a from above
530 A. Siito

and from below, the numerical plot of the energies belonging to the spectrum
clearly reveals a discontinuity, see Figure 2. [18] show that this reflects the
discontinuity of the characteristic function which generates the potential. The
gap edges vary continuously on irrational a's. For a rational value r,
lim v(a) =1= lim v(a)
atr a!r

and none of them equals v(r). In fact, the two limits are not periodic, only
ultimately periodic (as if we took a periodic sequence, cut off a finite segment
which is not a period and glue the two infinite pieces together). Both yield
the same essential spectrum as Ho + v(r) and both create (different) isolated
eigenvalues in the gaps.

7.3. Period doubling potential

The potential is generated by the period doubling substitution ~(a) = ab,

~(b) = aa. The corresponding Schrodinger equation was studied by Bellissard,
Bovier and Ghez [15], [27]. The trace map is a fundamental tool of the analysis.
Results 1.-5. valid for the Fibonacci potential hold true: The spectrum is a
Cantor set of zero Lebesgue measure, it is purely singular continuous (again,
not known for all elements of the hull) and the Lyapunov exponent vanishes
in the spectrum. There is a rather detailed knowledge about the gaps, their
behaviour as a function of the potential strength. Gap labelling is complete:
all admissible values are taken by the IDS in the gaps.

7.4. Thue-Morse potential

Historically, the first studies were done by Axel et al. on the phonon frequency
spectrum of the harmonic chain with masses generated by the Thue-Morse
substitution ~(a) = ab, ~(b) = ba. In [9], [10], [11] we find the proof that
the phonon spectrum is a Cantor set, a numerical work on its box dimension
suggesting zero Lebesgue measure, gap labelling and a study of generalized
eigenfunctions. This latter established the existence of extended states for a
dense set in the phonon spectrum.
The Schrodinger equation with the potential generated by this substitution
has also been widely studied, namely by Bellissard [12], Delyon and Peyriere
[43] and Bovier and Ghez [26], [27]. Valuable numerical work was done by
Riklund et al. [121]. Bellissard [12] identified the set RangN (d. (6.8)) and
studied the dependence of the gap widths on the potential strength. Delyon
and Peyriere [43] proved the absence of decaying solutions and, hence, the
continuity of the spectrum. They showed that the generalized eigenfunctions
are not 'too small' on a geometric progression. The continuity of the spectrum
was obtained as a byproduct also in [15], [27]. Bovier and Ghez [26], [27]
proved, in a more general context (see next section), that the spectrum is of
zero Lebesgue measure. With the continuity this implies that the spectrum is
purely singular continuous.
 SCHRODINGER DIFFERENCE EQUATION ... 531

7.5. Systematic study of substitutional potentials

Bovier and Ghez [26], [27] succeeded to find a general condition on prImI-
tive substitutions, assuring that the corresponding Schr6dinger operator has a
spectrum of zero Lebesgue measure. Their work deeply exploits the Kohir-Nori
results on trace maps ([84], see also [111]).
The action of a substitution ~ can be defined on transfer matrices and on
traces of transfer matrices. ~ acting on traces is called a trace map. ~ carries
a letter a of the alphabet A into a word, the transfer matrix corresponding to
a into a product, i.e. a monomial, of transfer matrices in reversed order, and
the trace of the transfer matrix into a polynomial of traces of transfer matrices
belonging to letters of an enlarged alphabet B. These generalized letters are
special words over the original alphabet A, not containing repeated letters of
A. Therefore B is also finite, and one can see that the trace map is closed in
the following sense:
~ carries the trace belonging to any letter of B into a polynomial of traces
belonging to letters of B.
Let b1 , ... , bq be the letters of B and x, a real variable associated with bi . The
idea of Bovier and Ghez is to retain the highest degree monomial
q

j,(X1, ... Xq ) = II x~"

)=1

of the polynomial corresponding to bi and to define a substitution ¢ on B such

that ¢(b,) contains b) k,) times. Based on earlier experience, namely with the
Fibonacci substitution, one may expect that, by imposing some conditions on
¢, it is possible to control the high iterates of the trace map and, hopefully, the
spectrum. The right notion is semi-primitivity, a property which is somewhat
weaker than primitivity (see [26J for the definition) and is straightforward to
verify. The authors prove that, if ~ is primitive, ¢ is semi-primitive and the
substitutional sequence contains the word bb for some b E B, CT(H) is a set
of zero Lebesgue measure. Although the proof is more complicated than in
the cases discussed earlier, the main idea is again to show that the Lyapunov
exponent vanishes in the spectrum (the existence of the word bb is used here)
and, then, to confront the Ishii-Pastur-Kotani theorem with the Kotani theorem
for finite-ranged potentials. If the substitutional sequence starts with a square,
the two-interval version of the Gordon theorem yields also that the spectrum
is purely singular continuous.
The conditions of the theorem are fulfilled by many named substitutional
sequences as, for example, the Fibonacci, Thue-Morse, period doubling, binary
non-Pisot, ternary non-Pisot and circle sequences. Apart from the binary non-
Pisot sequence, singular continuity is also verified in the cases listed above (but
it is not proved for the other potentials in the hull). A notable exception is
the Rudin-Shapiro substitution which is primitive and therefore the spectrum
is purely singular ([64]), but for which the substitution ¢ is not semi-primitive
and, hence, it is not known whether u(H) is of zero Lebesgue measure.
532 A. SiitO

8. Solutions of the problems

2.1
The proof is based on the observation that for bounded sequences pointwise
convergence is equivalent to convergence in the metric d.
(i) 1. implies 2.
Indeed, s E O(s) = O(t).
(ii) 2. implies 3.
O(s) is bounded: for all t E O(s), d(s, t) ::; 6sup ISnl.
Suppose that s is not almost-periodic. Then there exists an € > 0 and an
infinite sequence of real intervals (ak' bk ) such that bk -ak -+ 00 and d(Tn s, s) >
€ for every integer n E U(ak' bk). Choose an integer nk in (ak' bk ) such that
nk -ak -+ 00 and bk -nk -+ 00. Since O(s) is closed and bounded (compact in
the topology generated by d), Tnks has at least one limit point tin O(s). Fix
an arbitrary integer n. We have the triangle inequality
d(s, Tnt) ~ d(s, Tnd n s ) - d(Tnd n s , Tnt) .
For k sufficiently large, nk + n E (ak,bk) and thus d(s, Tnk+n s ) > €. On the
other hand, d(Tnk s, t) -+ 0 implies that d(Tnk+n s, Tnt) -+ 0 as k -+ 00. It
follows that d(s, Tnt) ~ € for all integers n. Hence, s cannot be in O(t), which
contradicts 2.
(iii) 3. implies 2.
Let t E O( s), t = lim J ..... oo TiJ s, where lim means pointwise limit. Fix an
€ > 0 and let {nd be the almost-periods with gaps smaller than £E:. Given j,
choose k such that nk-l < i j < nk· Then m J == nk - i J < £0 and
d(s,TmJt) < d(s,Tnks) + d(Tnks, TmJ t)
< €+d(TmJT'Js,TmJt)
Take the limit j -+ 00. Since 0 < mj < £0' there will be an m E (0, £E:) which
occurs infinitely many times among the mJ's. Taking the limit only on the
subsequence of j values for which mj = m, we obtain d(s, Tmt) ::; €. Since €
was arbitrary, this proves that s E O(t).
(iv) 2. implies 1.
Suppose t E O(s). Then Tnt E O(s) for every integer n, because O(s) is
shift-invariant. O(s) being closed, this implies O(t) C O(s). On the other
hand, according to 2., s E O(t). Interchanging sand t, the above argument
yields O(s) C O(t).
2.2
Let A C O(s) be a T-invariant set. One has to show that p(A) = 0 or 1.
Suppose the opposite, i.e., 0 < p(A) < 1. Define a probability measure J.L by
setting
J.L(B) = p(A n B)/ p(A)
It is easy to verify that J.L is a T-invariant probability; on the other hand, it
differs from p (e.g., on O(s) \ A). This contradicts the uniqueness of p.
 SCHRODINGER DIFFERENCE EQUATION ... 533

2.3
Let wEn. If {Tnw}:=o is an infinite set, the sequence contains no repetition.
Therefore
00 00

n=O n=O
and hence p( {w}) = 0.
3.2.1
The proof goes by induction. tr AO tr I = = 2. Multiplying the Caley-

°
Hamilton equation .
A2 - (tr A)A + I =
with A k - 2 and taking the trace one finds

tr Ak = tr A tr A k- 1 - tr A k- 2

in which one may recognize the recurrence relation for the Chebyshev polyno-
mials. Setting tr A = 2 cos a, the solution of the recurrence is

tr Ak = 2 cos ka
5.1.1
Let H1/J = E1/J where 1/J is subexponential: for any a °
> there exists a b > °
such that
l1/Jk I < be a1kl for all k,
but 1/J is not in (2(Z). It follows that either E~=o l1/Jkl 2 = 00 or E~=-oo l1/Jkl 2 =
00. Suppose, for instance, the first. Define 1/Jn = 1/J/(E"!:o l1/Jd 2 )1/2 for n > 0.
Then H1/Jn = E1/Jn, and rpn, given by

rpk - ° °
n _ {1/J1: if ~ k ~ n
otherwise

is a normalized vector for all n. We show that a suitable subsequence of

°
{rpn}:=l is a Weyl sequence. It is clear that for all k, rpk --+ with increasing
n, so rpn goes to zero weakly. On the other hand,

(Hrpnlk = Erpk if k =F -l,O,n,n + 1

This implies that

ek = ° if k =F -l,O,n,n + 1
(cf. Eq.(5.3)). It is easy to check that

Therefore
534 A. Siito

where the vectors on the right correspond to the definition (3.1). The first
term goes to zero as n increases, the second, in general, not. However, there
exists a subsequence nk such that IIw~: II --+ 0 with increasing k. Indeed, if the
opposite were true, one could find a positive constant c such that IIw~1I2 > c
for all n > o. Now
n n-l

C < Ilw~1I2 = IIw n ll 2/ L l"pkl 2 ~ 211w n 1l 2/ L II wkll 2

k=O

and therefore

which contradicts the supposed subexponential nature of"p. The Weyl sequence
we were looking for is {'P nk H"=l.
5.2.1
In general, let E be any accumulation point of the eigenvalues of H, say,
En --+ E as n --+ 00. Let"pn be the corresponding orthonormal eigenvectors.
Then "pn form a Weyl sequence. First, they go to zero weakly: for any vector
'I' E 1i, the Bessel inequality

L
00

11'1'1122: 1("pn,'P}1 2
n=l

implies that ("pn, 'P) --+ 0 with increasing n. Second,

H"pn = En"pn = E"pn + (En - E}'¢Jn

and II(En - E}'¢Jnll = lEn - EI --+ o.

5.2.2
Choose a sequence nk such that Vnk --+ E as k --+ 00. If {jn is the unit vector
concentrated on n,

so the inverse of E - V is unbounded. By definition, this means that E is in

the spectrum.
One can also construct a vector "p such that the solution of (E - V}'P = "p
for '1', given by

is not square-summable. Write E = cos 27rB, then trivially

IE - Vnl ~ 27rmin Ina

pEz
- B- pi
 SCHRODINGER DIFFERENCE EQUATION ... 535

On the other hand, by the Kronecker theorem (Theorem 440 in [57]), there ex-
ists an increasing sequence nk of positive integers and a sequence Pk of integers
such that
InkQ - () - Pkl < 3/nk
The example is obtained by choosing 1/Jn = E - Vn for n = nk and zero
otherwise: 111/J1I2 < 611"4 but t.pnk = 1 for all k, so t.p is not square-summable.
5.6.1
If x is not in supp JL then for sufficiently small c > 0, JL( x + c) - JL( x - c) =
JL«x - c, x + c» = 0 and, hence, x is not a point of increase of the distribution
function. If x E supp JL then for every c > 0,

JL(x + c) - JL(x - c) = JL«x - c, x + c]) ~ JL«x - c,x + c» > 0 ,

that is, x is a point of increase.

5.6.2
Let JL, v be ac measures, A a common essential support. Suppose that JL(B) =
O. v(BnAC) = 0 because A is an essential support of v. m(BnA) = 0, because
B n A c A, an essential support of JL, and JL(B n A) = o. But v is ac, so
v(B n A) = o. We found v «: JL. The opposite is obtained by interchanging JL
and v.
5.6.3
(i) Let C be a 'thick' Cantor set (Le., m(C) > 0),

n=O CCn -- U n=O uk=l J nk

C c -- U OO OO kn

where Jnk are the open intervals appearing in the construction of C (cf. Sec.5.3).
Let fnk(X) be a continuous function which is strictly positive on Jnk and van-
ishes outside J nk , and

f: t J
n=Ok=l
fnk(x)dx < 00

Then Ln Lk f nk is the density of an absolutely continuous measure JL. supp JL =

lR (because the closure of CC is lR), and an essential support of JL is CC. So
JL( C) = 0 but m( C) > 0, therefore supp JL is not an essential support of JL.
(ii) For 0 < P < q, P and q integers which are relatively primes, let JL( {p/q}) =
2- p - Q • supp JL = [0,1] while the rational numbers in (0,1) form an essential
support. JL([O, 1]\«:11) = 0 but m([O, 1]\«:11) = 1, so the support is not an essential
support.
(iii) Let JL be concentrated on a discrete set (whose only accumulation points
can be ±oo), for instance, JL(lR) = L:'=l JL( {n}). Then supp JL is the smallest
essential support.
5.6.4
(i) If JL( x) is absolutely continuous (ac) on any finite interval then the measure
JL is ac.
Indeed, let A be a Borel set, m(A) = 0 and suppose first that A is covered by
536 A. Siito

a finite open interval J. Fix c > 0 and let 8 > 0 correspond to c (cf. Eq.(5.7)).
According to the definition of a set of zero Lebesgue measure, one can find an
open set 0 such that A C 0 and m(O) < 8. 0 is the union of disjoint open
intervals (ai, b1), (a2' ~), ... , all in J, so that A C U(ai' b,) and I:(b, - ai) < 8.
Thus,

Since c is arbitrary, JL(A) = O.

If A is not covered by a finite interval, it is covered by count ably many finite
intervals Jk. JL(A n Jk) = 0 for all k, thus JL(A) = O.
(ii) If JL is an ae measure, i.e., JL « m, then JL(x) is an ae function on any
finite interval.
We show this by proving that, given a finite interval J, for any c > 0 there
exists a 8 > 0 such that for Borel sets A C J, m(A) < 8 implies JL(A) < c.
Suppose this does not hold true. Then there exists an c > 0 such that for
all n ~ 1 and for suitably chosen An C J, m(An) < 1/2n but JL(An) ~ c.
Let Bn = Uk'=n+1Ak. Now m(Bn) ~ 1/2n and JL(Bn) ~ JL(An+d ~ c. Since
Bn C J for all nand Bn is a decreasing sequence, it has a limit B C J for
which m(B) = 0 and JL(B) ~ f. This contradicts JL « m.
5.6.5
Suppose JL is a singular measure. Then there is some set A of zero Lebesgue
measure such that JL is concentrated on A. The intersection of any essential
support with A yields an essential support of zero Lebesgue measure. If B is
an essential support and m(B) > 0 then JL(B \ A) = 0 and m(B \ A) > 0, a
contradiction.
5.6.6
The definition of JLpp is constructive, see Eq.(5.6), therefore the pp part is
unique. Let
JL -_ JLpp +1+1_
JLsc JL ac - JLpp +2+2
JLsc JL ac
If JL!c i= JL~c' there exists a set Ll such that m( Ll) = 0 and JL!c (Ll) i= JL~c (Ll).
But then JL!c(Ll) i= JL;c(Ll) which contradicts JL!c(Ll) = JL;c(Ll) = o.
5.7.1
Let C be the middle-thirds Cantor set, a the se measure whose distribution
function is the Cantor function. Let JL be a pp measure whose smallest essential
support is the set of boundary points {em} of the complement of C, cf. Sec.
5.3. Then supp JL = supp a = C and C is also an essential support for both
measures.
5.8.1
The following statements are equivalent: JL", is concentrated on A, JL",(IR) =
JL",(A), 111jJ11 2 = IIP(A)1jJ11 2 , II (I - P(A) )1jJ1l = 0, (I - P(A))1jJ = 0, P(A)1jJ = 1jJ.
If 1jJ E P(A)1i then P(A)1jJ = 1jJ, so JL",(AC) = 0 and sUPP JL", C A.
5.8.2
Let A be any Borel set such that JL", is concentrated on A. Then from
Problem 5.8.1, P(A)1jJ = 1jJ. H commutes with P(A), therefore P(A)H1jJ = H1jJ.
 SCHRODINGER DIFFERENCE EQUATION ... 537

From Problem 5.B.1 it follows that /LH..p is concentrated on A. By definition,

we get /LH..p « /L..p.
If /L..p is singular, it is concentrated on a set A such that meA) = o. There-
fore /LH..p is also singular. If /L..p is continuous, for any E E JR, /L..p( {E}) =
IIP({E})1jJ1I 2 = 0, therefore P({E})1jJ = O. Hence,

/LH..p({E}) = IIP({E})H1jJ1I 2 = IIHP({E})1jJ1I2 = 0,

that is, /LH..p is continuous. SO /LH..p is sc if /L..p is sc. If /L..p is ac then /LH..p is ac
because meA) = 0 implies /L..p(A) = 0, which implies /LH..p(A) = o.
5.8.3 .
Let A, B, C be disjoint essential supports respectively for (/L..p )pp, (/L..p )ac, (/L..p )sc.
/L..p is concentrated on Au B U C. From Problem 5.B.1,

1jJ = peA U B U C)1jJ = P(A)1jJ + P(B)1jJ + P(C)1jJ == 1jJPP + 1jJac + 1jJsc

Now
P(A)P(B) = P(A)P(C) = P(B)P(C) = 0
therefore 1jJPP, 1jJac ,1jJsc are pairwise orthogonal, and for any Borel set D.

/L..p(D.) = IIP(D.)1jJ1I 2 = IIP(D.)P(A)1jJ1I 2 + IIP(D.)P(B)1jJ1I 2 + IIP(D.)P(C)1jJ1I2

= /L..ppp (D.) + /L..pac (D.) + /L..pBC (D.)

Using the relations P(D.)P(A) = P(D. n A), etc., we find

and so on.
5.8.4
(i) Let B be any Borel set, then

(1jJ, P(B)1jJ) + (1jJ' - 1jJ, P(B)1jJ') + (1jJ, p(B)(1jJi - 1jJ))

/L..p(B) + O(II1jJi - 1jJ1I) -+ /L..p(B)

(ii) 1jJi are singular, so there exist sets A, such that /L..p.(A~) = 0 and
meA,) = o. Now m(uA,) = 0, for all i /L..p.«UAj)C) ::; /L..p.(A~) = 0 and
by (i), /L..p«UAj)C) = o. So /L..p is concentrated on a set of zero Lebesgue mea-
sure and, hence, is singular. If all the /L..p. are pp or all are sc then /L..p is pp or
sc, respectively, because Hpp and Hcont are closed subspaces.
5.8.5
(i) [HacJP is ac: Let B be a Borel set, m(B) = o. Take any 1jJ E H, then

II [Hac]P(B)1jJ1I2 = IIP(B)[Hac]1jJ1I2 = /L..pac(B) = (/L..p)ac(B) = 0

therefore [Hac]P(B) = o.
(ii) [H.cJP ::; [7tcontlP = Pcont , therefore [7t sc ]P is a continuous measure.
We have to prove that it is singular. Choose an orthonormal basis {1jJd. Let
538 A. Siito

A. be sets of zero Lebesgue measure, /Lt/J8C concentrated on Ai. Take A = UA •.

m(A) = 0 and we show that [1£sc]P is c~ncentrated on A.

IIP(A C)[1£sc]'I/J,1I = IIP(AC)'I/J~clI = IIP(nJAj)'I/J:clI = II II P(A;)p(AD'l/Jrll = 0

jf:.,

The last equality holds because P(Ai)'l/Jic = 0, according to Problem 5.8.1. It

follows that [1£sc]P(AC)'I/J, = 0 for all i, and therefore [1£sc]P(AC) = O.
5.8.6
If /L(B) = 0 then /Lt/J(B) = 0 for all 'I/J E 1£. Accordingly, IIP(B)'l/J1l = 0 for
all 'I/J E 1£ which means that P(B) = O.
5.8.7
Let B be a Borel set, /Lt/Jl(B)+/Lt/J2(B) = O. Then IIP(B)'l/J 111 = IIP(B)'l/J 211 =
o and hence p(B)'l/J1 = p(B)'l/J2 = O. This implies
/Lat/J1+13t/J2(B) laI 2I1P(B)'l/J111 2 + 1,B1 2 IIP(B)'l/J2 112
+ a*,B('l/Jl,P(B)'l/J2) +a,B*('l/J2,P(B)'l/J1) =0
5.8.8
According to Problem 5.8.6, it suffices to show that /Lt/J « L: c,/Lt/J' for all
'I/J E 1£. Writing 'I/J = L: ai'I/Ji, one proves

/LEa, t/J' « L C,/Lt/J'

exactly as in Problem 5.8.7.
5.8.9
(i) The set is linearly independent: Taking any finite linear combination with
not all coefficients vanishing and expanding the sum in the canonical basis {{in},
at least one {in appears with nonzero coefficient.
(ii) The set generates the canonical basis {{in}~=_OCl' For n = 0 one obtains
{i0, {i1 and by induction one finds that {i-n, {i-n+! , ... , {in, {in+! can be expanded
with the vectors {Hk{iO, Hk{i1}k=0'
5.8.10
Let 'l/J1 E 1£1 and 'l/J2 E 1£2 where 1£i are subspaces of 1£.

/Lt/Jl+t/J2(B) = IIP(B)('l/J1 + 'l/J 2)1I2

= /Lt/Jl(B) + /Lt/J2(B) + (P(B)'l/J1,'l/J2) + ('l/J2,P(B)'l/J1)
and
(p(B)'l/J1, 'l/J2) = (P(B)[1£1]'l/J 1, [1£2]'l/J2) = 0
for every Borel set B if and only if the two subspaces are orthogonal and one
of them, say 1£1, is H-invariant (and, hence, [1£1] commutes with P(B) for all
B). Then, however, 1£2 C 1£[ which is H -invariant and contains 'l/J 2.
6.2.1
Choose a sequence {nd such that V k == V(Tnkw) --+ V(w') pointwise. The
corresponding operators Hk = Ho + Vk all have the same spectrum, a(H(w)),
 SCHRODINGER DIFFERENCE EQUATION ... 539

and converge strongly to H(w'). Equation (6.1) holds in the form a(H(w)) :J
a(H(w')). The opposite is also true because V(w') is also minimal.
6.2.2
Any point in the spectrum which is not an eigenvalue is in the essential
spectrum. So we have to prove that all the eigenvalues are in the essential
spectrum. Let E be an eigenvalue, '¢ the corresponding normalized eigenvector.
Because V is recurrent, there exists an increasing sequence {nd such that
Vk = TnkVT-n k (V: = Vn+nk ) tends to V pointwise. Let <pk = T-nk'¢.
«
This goes to zero weakly <p, <pk) - 0 for any fixed vector <p), so it is a Weyl
sequence if c;k == (Ho + V - E)<pk _ 0 (in norm!).

There exists a sequence of integers mk - 00 and a sequence 'T/k - 0 such that

IV: - Vn l 2 < 'T/k for Inl < mk, and l:lnl~mk l'¢nl 2 < 'T/k. Therefore
II(V k - V),¢1I2 =( L + L )(V; - Vn)21'¢nI 2 < 'T/k(l + 4sup IVm1 2 )
Inl<mk Inl~mk

so c;k goes to zero, indeed.

6.9.1
According to Section 6.3, the IDS can be obtained as

N(E) = lim NnL{E)

n->oo

where NnL{E) is l/nL times the number of eigenvalues less than E of H nL , the
restriction of H to an interval of length nL. A convenient choice of boundary
condition is

The matrix of HnL has V1 , ... , VnL in the diagonal, 1 above and below it and
e-'''' and e'" in the upper right and lower left corner, respectively. Finding the
eigenvalues of HnL is equivalent to finding the E values for which the transfer
matrix Tl->ndE) = T1'-.L (E) has eigenvalues e±'''' or trace 2 cos a: this can
be seen from
'l1nL = Tl->nL 'l1o
(cf. Section 3.2) and the boundary condition. Choose a = /2, then one has
7l'

to look at the zeros of trTl~L(E). l.From Figure 1 and Problem 3.2.1 one
learns that this function has exactly n zeros in each band of the spectrum of
H; therefore the IDS in the gap just above the kth band is kn/nL = k/ L,
independently of n, which proves the assertion.
540 A. Siito

References

[1] J.P. Allouche: This volume.

[2] J.P. Allouche, J. Peyri(~re: "Sur une formule de recurrence sur les traces
de produits de matrices associees a certaines substitutions".
C.R. Acad. Sci. Paris 302 serie II, 1135-1136 (1986).
[3] AI-Naggar, D.B. Pearson: "A New Asymptotic Condition for Absolutely
Continuous Spectrum of the Sturm-Liouville Operator on the Half-Line".
Helv. Phys. Acta. 67, 144-166 (1994).
[4] S. Aubry, G. Andre: "Analycity Breaking and Anderson Localisation in
Incommensurate Lattices".
Ann. Israel Phys. Soc 3,'133-140 (1980).
[5] Y. Avishai, D. Berend: "1Tace Maps for Arbitrary Substitution Se-
quences".
J. Phys. A: Math. Gen. 26, 2437-2443 (1993).
[6] J. Avron, P.H.M. v. Mouche, B. Simon: "On the Measure of the Spectrum
for the Almost Mathieu Operator".
Commun. Math. Phys. 132, 103-118 (1990).
[7] J. Avron, B. Simon: "Almost Periodic Schrodinger Operators I: Limit
Periodic
Potentials" .
Commun. Math. Phys. 82, 101-120 (1981).
[8] J. Avron, B. Simon: "Almost Periodic Schrodinger Operators II: The
Integrated Density of States".
Duke Math. J. 50, 369-391 (1983).
[9] F. Axel, J.P. Allouche, M. Kleman, M. Mendes-France, J. Peyriere: "Vi-
brational Modes in a One Dimensional "Quasi-Alloy": the Morse Case".
J. de Physique (Paris) C3 47,181-186 (1986).
[10] F. Axel, J. Peyriere: "Spectrum and Extended States in a Harmonic
Chain with Controlled Disorder: Effects of the Thue-Morse Symmetry".
J. Stat. Phys. 57, 1013-1047 (1989).
[11] F. Axel, J. Peyriere: "Etats etendus dans une chaine adesordre controJe".
C.R. Accad. Sci. Paris 306 serie II, 179-182 (1988).
[12] J. Bellissard: "Spectral Properties ofSchrodinger's Operator with a Thue-
Morse Potential".
in "Number Theory and Physics", ed. J.M. Luck, P. Moussa, M. Wald-
schmidt, Springer Verlag: Springer Proceedings in Physics 47, 140-150
(1990).
[13] J. Bellissard: "Gap Labelling Theorems for Schrodinger Operators".
in "From Number Theory to Physics",
ed. M. Waldschmidt, P. Moussa, J .M. Luck, C. Itzykson, chapter 12,
538-630 Springer Verlag, (1992).
 SCHRODINGER DIFFERENCE EQUATION ... 541

[14] J. Bellissard: "K-Theory ofC*-Algebras in Solid State Physics".

in "Statistical Mechanics and Field Theory: Mathematical Aspects" ,
ed. T.C. Dorlas, N.M. Hungenholtz, M. Winnink, Springer Verlag: Lec-
ture Notes in Physics 257, 99-156 (1986).
[15] J. Bellissard, A. Bovier, J.-M. Ghez: "Spectral Properties of a Tight
Binding Hamiltonian with Period Doubling Potential".
Commun. Math. Phys. 135, 379-399 (1991).
[16] J. Bellissard, A. Bovier, J.-M. Ghez: "Gap Labelling Theorems for One
Dimensional Discrete Schrodinger Operators".
Rev. Math. Phys. 4, 1-37 (1992).
[17] J. Bellissard, B. Iochum, E. Scoppola, D. Testard: "Spectral Properties
of One Dimensional Quasi-Crystals".
Commun. Math. Phys. 125, 327-345 (1986).
[18] J. Bellissard, B. Iochum, D. Testard: "Continuity Properties of the Elec-
tronic Spectrum of ID Quasicrystals".
Commun. Math. Phys. 141, 353-380 (1991).
[19] J. Bellissard, R. Lima, D. Testard: "A Metal-Insulator Transition for the
Almost Mathieu Model".
Commun. Math. Phys. 88, 207-234 (1983).
[20] J. Bellissard, E. Scoppola: "The Density of States for Almost Periodic
Schrodinger Operators and the Frequency Module: A Counter-Example".
Commun. Math. Phys. 85, 301-308 (1982).
[21] J. Bellissard, B. Simon: "Cantor Spectrum for the Almost Mathieu Equa-
tion".
J. Func. Anal. 48, 408-419 (1982).
[22] Ju.M. Berezanskii: "Expansions in Eigenfunctions of Selfadjoint Opera-
tors".
A.M.S. Translations of Mathematical Monographs 17, (1968).
[23] V. Berthe: This volume.
[24] A.S. Besicovitch: "Almost Periodic Functions".
Cambridge University Press, (1932).
[25] Ph. Bougerol, J. Lacroix: "Products of Random matrices with Applica-
tion to Schrodinger Operators".
Birkhauser: Progress in Probability and Statistics, (1985).
[26] A. Bovier, J.-M. Ghez: "Spectral Properties of One-Dimensional Schrodinger
Operators with Potentials Generated by Substitutions".
Commun. Math. Phys. 158, 45-66 (1993).
[27] A. Bovier, J.-M. Ghez: Erratum to "Spectral Properties of One-Dimensional
Schrodinger Operators with Potentials Generated by Substitutions",
Commun. Math. Phys. 158, 45-66 (1993).
Commun. Math. Phys. 166, 431-432 (1994).
542 A. Siito

[28] R. Carmona, A. Klein, F. Martinelli: "Anderson Localization for Bernoulli

and Other Singular Potentials".
Commun. Math. Phys. 108, 41-66 (1987).
[29] R. Carmona, J. Lacroix: "Spectral Theory of Random Schrodinger Op-
erators".
Birkhauser: Probability and its Applications, (1990).
[30] M. Casdagli: "Symbolic Dynamics for the Renormalization Map of a
Quasiperiodic Schrodinger Equation".
Commun. Math. Phys. 107, 295-318 (1986).
[31] M.-D. Choi, G.A. Elliot, N. Yui: "Gauss Polynomials and the Rotation
Algebra". Invent. Math 99, 225-246 (1990).
[32] V. Chulaevsky, F. Delyon: "Purely Absolutely Continous Spectrum for
Almost Mathieu Operators".
J. Stat. Phys. 55, 1279-1284 (1989).
[33] V.A. Chulaevsky, Ya.G. Sinai: "Anderson Localisation for 1-D Discrete
Schrodinger Operator with Two-Frequency Potential".
Commun. Math. Phys. 125, 91-112 (1989).
[34] V.A. Chulaevsky, Ya.G. Sinai: "The Exponential Localization and Struc-
ture of the Spectrum for 1D Quasi-Periodic Discrete Schrodinger Opera-
tors".
Rev. Math. Phys. 3, 241-284 (1991).
[35] W. Craig, B. Simon: "Subharmonicity of the Lyapunov Index".
Duke Math. J. 50, 551-560 (1983).
[36] W. Craig, B. Simon: "Log Holder continuity of the Integrated Density of
States for Stochastic Jacobi Matrices".
Commun. Math. Phys. 90, 207-218 (1983).
[37] H.L. Cyeon, R.G. Froese, W. Kirsch, B. Simon: "Schrodinger Operators
with Application to Quantum Mechanics and Global Geometry".
Springer Verlag, (1987).
[38] P. Deift, B. Simon: "Almost Periodic Schrodinger Operators III: The
Absolutely Continuous Spectrum in One Dimension".
Commun. Math. Phys. 90, 389-411 (1983).
[39] F.M. Dekking: "The Spectrum of Dynamical Systems Arising from Sub-
stitutions of Constant Length".
Z. Wahrscheinlichkeitstheorie verw. Gebiete 41, 221-239 (1978).
[40] F .M. Dekking: This volume.
[41] F. Delyon: "Abscence of Localization for the Almost Mathieu Equation".
J. Phys. A: Math. Gen. 20, L21-L23 (1987).
[42] F. Delyon, D. Petritis "Absence of Localization in a Class of Schrodinger
Operators with Quasiperiodic Potential".
Commun. Math. Phys. 103, 441-444 (1986).
 SCHRODINGER DIFFERENCE EQUATION ... 543

[43] F. Delyon, J. Peyriere: "Recurrence of the Eigenstates of a Schrodinger

Operator with Automatic Potential".
J. Stat. Phys. 64, 363-368 (1991).
[44] F. Delyon, B. Souillard: "The Rotation Number for Finite Difference
Operators and its Properties".
Commun. Math. Phys. 89, 415-426 (1983).
[45] M.S.P. Eastham: "The spectral Theory of Periodic Differential Equa-
tions".
Scottish Academic Press, (1973).
[46] P. Erdos, RC. Herndon: "Theories of Electrons in One-Dimensional Dis-
ordered Systems".
Adv. Phys. 31, 65-163 (1982).
[47] H. Furstenberg, H. Kesten: "Products of Random Matrices".
Ann. Math. Stat. 31, 457-469 (1960).
[48] S. Fishman, D.R Grempel, RE. Prange: "Localization in an Incommen-
surate Potential: An Exactly Solvable Model".
Phys. Rev. Lett. 49, 833-836 (1982).
[49] S. Fishman, D.R Grempel, RE. Prange: "Localization in ad-Dimensional
Incommensurate Structure".
Phys. Rev. B 29, 4272-4276 (1984).
[50] J. Frohlich, T. Spencer, P. Wittwer: "Localization for a Class of One
Dimensional Quasi-Periodic Schrodinger Operators".
Commun. Math. Phys. 132, 5-25 (1990).
[51] D.J. Gilbert: "On Subordinacy and Analysis of the Spectrum ofSchrodinger
Operators with Two Singular Endpoints".
Proc. Royal. Soc. Edinburgh 112A, 213-229 (1989).
[52] D.J. Gilbert, D.B. Pearson: "On Subordinacy and Analysis of the Spec-
trum of One-Dimensional Schrodinger Operators".
J. Math. Anal. and Appl. 128, 30-56 (1987).
[53] I.Ya. Goldsheid: "Asymptotic Properties of the Product of Random Ma-
trices Depending on a Parameter".
in "Muiticomponent Random Systems" edited by RL. Dobrushin
and Ya.G. Sinai, Marcel Dekker Inc., 239-283 (1980).
[54] A.Ya. Gordon: "On the Point Spectrum of the One Dimensional Schrodinger
Operator".
Usp. Math. Nauk. 31, 257 (1976).
[55] A.Ya. Gordon: "Pure Point Spectrum Under I-Parameter Perturbations
and Instability of Anderson Localization".
Commun. Math. Phys. 164, 489-505 (1994).
[56J W.H. Gottschalk: "Substitution Minimal Sets".
Transactions Amer. Math. Soc. 109, 467-491 (1963).
544 A. Siito

[57] G.H. Hardy, E.M. Wright: "An Introduction to the Theory of Numbers".
Fourth Edition, Oxford University Press, (1971).
[58] D. Herbert, R. Jones: "Localized States in Disordered Systems".
J. Phys. C 4, 1145-1161 (1971).
[59] M.R. Herman: "Une methode pour minorer les exposants de Lyapounov
et quelques exemples montrant Ie caractere local d'un theoreme d'Arnold
et de Moser sur Ie tore de dimension 2".
Comment. Math. Helvetici 58, 453-502 (1983).
[60] H. Hiramoto, M. Kohmoto: "New Localization in a Quasiperiodic Sys-
tem".
Phys. Rev. Lett. 62, 2714-2717 (1989).
[61] H. Hiramoto, M. Kohmoto: "Electronic Spectral and Wavefunction Prop-
erties of One-Dimensional Quasiperiodic Systems: A Scaling Approach".
Int. J. of Mod. Phys. B 6, 281-320 (1992).
[62] H. Hiramoto, M. Kohmoto: "Scaling Analysis of Quasiperiodic Systems:
Generalized Harper Model".
Phys. Rev. B 40, 8225-8234 (1989).
[63] A. Hof: "Some Remarks on Discrete Aperiodic Schr&linger Operators".
J. Stat. Phys. 72, 1353-1374 (1993).
[64] A. Hof, O. Knill, B. Simon: "Singular Continuous Spectrum for Palin-
dromic Schrodinger Operators".
Preprint (1994).
[65] K. Iguchi: "Equivalence Between the Nielsen and the Scaling Transfor-
mations in One-Dimensional Quasiperodic Systems".
J. Math. Phys. 34, 3481-3490 (1993).
[66] B. Iochum, D. Testard: "Power Law Growth for the Resistance in the
Fibonacci Model".
J. Stat. Phys. 65, 715-723 (1991).
[67] B. Iochum, L. Raymond, D. Testard: "Resistance of One-Dimensional
Quasicrystals" .
Physica A 187, 353-368 (1992).
[68] K. Ishii: "Localization of Eigenstates and Transport Phenomena in the
One Dimensional Disordered System".
Suppl. Prog. Theor. Phys .. 53, 77-138 (1973).
[69] S.Ya. Jitomirskaya: "Anderson Localization for the Almost Mathieu Equa-
tion I: A Nonperturbative Proof".
Commun. Math. Phys. 165, 49-57 (1994).
[70] S. Ya. Jitomirskaya: "Anderson Localization for the Almost Mathieu Equa-
tion II: Point Spectrum for A> 2".
Preprint (1994).
 SCHRODINGER DIFFERENCE EQUATION ... 545

[71] S. Jitomirskaya, B. Simon: "Operators with Singular Continuous Spec-

trum:
III: Almost Periodic Schrodinger Operators".
Commun. Math. Phys. 165, 201-205 (1994).
[72] R. Johnson: "A Review of Recent Works on Almost Periodic Differential
and Difference Operators".
Acta Appl. Math. 1, 54-78 (1983).
[73] R. Johnson, J. Moser: "The Rotation Number for Almost Periodic Po-
tentials".
Commun. Math. Phys. 84, 403-438 (1982).
Erratum: Commun. Math. Phys. 90, 317-318 (1983).
[74] G. Jona-Lasinio, F. Martinelli, E. Scoppola: "Multiple Tunneling in d-
Dimension of a Quantum Particle in a Hierarchical System".
Ann. Inst. Henri Poincare 42, 73-108 (1985).
[75] I.S. Kac: "On the Multiplicity of the Spectrum of a Second Order Differ-
ential Operator".
Soviet Math. 3, 1035-1039 (1962).
[76] I.S. Kac: "On the Multiplicity of the Spectrum of a Second Order Differ-
ential Operator".
Izv. Akad. Nauk SSSR Ser. Mat. 27, 1081-1112 (1963) (in Russian).
[77] T. Kato: "Perturbation Theory for Linear Operators".
Springer Verlag: Grund. der math. Wissen. 132 2nd ed., 2nd print,
(1984).
[78] S. Khan, D.B. Pearson: "Subordinacy and Spectral Theory for Infinite
Matrices".
Helv. Phys. Acta 65, 505-527 (1992).
[79] M. Kohmoto, L.P. Kadanoff, C. Tang: "Localization Problem in One
Dimension: Mapping and Escape".
Phys. Rev. Lett. 50, 1870-1872 (1983).
[80] S. Ostlund, R. Pandit, D. Rand, H.J. Schnellnhuber, E.D. Siggia: "One-
Dimension Schrodinger Equation with an Almost Periodic Potential".
Phys. Rev. Lett. 50, 1873-1876 (1983).
[81] M. Kohmoto, Y. Oono: "Cantor Spectrum for an Almost Periodic Schrodinger
Equation and a Dynamical Map".
Phys. Lett. 102 A, 145-148 (1984).
[82] M. Kohmoto, B. Sutherland, K. Iguchi: "Localisation in Optics: Quasiperi-
odic Media".
Phys. Rev. Lett. 58, 2436-2438 (1987).
[83] M. Kolar, M.K. Ali: "Dace Maps Associated with General Two-Letter
Substitution Rules".
Phys. Rev. A 42, 7112-7124 (1990).
546 A. Siito

[84] M. Kolar, F. Nori: "nace Maps of General Substitutional Sequences".

Phys. Rev. B 42, 1062-1065 (1990).
[85] J. Kollar, A. Siito: "The Kronig-Penney Model on a Fibonacci Lattice".
Phys. Lett. 117 A, 203-209 (1986).
[86] S. Kotani: "Lyapunov Indices Determine Absolute Continuous Spectra
of Stationary One Dimensional Schrodinger Operators".
Proc. Kyoto Stoch. Conf. (1983).
[87] S. Kotani: "Lyapunov Exponents and Spectra for One Dimensional Ran-
dom Schrodinger Operators".
Proceedings of the A.M.S. meeting on 'Random Matrices' Brunswick (1984).
[88] S. Kotani: "Jacobi Matrices with Random Potential taking Finitely Many
Values".
Rev. Math. Phys. 1, 129-133 (1989).
[89] H. Kunz: This volume.
[90] H. Kunz, R. Livi, A. Siito: "Cantor Spectrum and Singular Continuity
for a Hierarchical Hamiltonian".
Commun. Math. Phys. 122, 643-679 (1989).
[91] H. Kunz, B. Souillard: "Sur le spectre des operateurs aux differences
finies aleatoires".
Commun. Math. Phys. 78, 201-246 (1980).
[92] Y. Last: "On the Measure of Gaps and Spectra for Discrete ID Schrodinger
Operators" .
Commun. Math. Phys. 149, 347-360 (1992).
[93] Y. Last: "A Relation Between a.c. Spectrum of Ergodic Jacobi Matrices
and the Spectra of Periodic Approximants".
Commun. Math. Phys. 151, 183-192 (1993).
[94] Y. Last: "Zero Measure Spectrum for the Almost Mathieu Operator".
Commun. Math. Phys. 164,421-432 (1994).
[95] P. Liardet: "Some metric properties of subsequences".
Acta Arith. 55, 119-135 (1990).
[96] R. Livi, A. Maritan, S. Ruffo: "The Spectrum of a I-D Hierarchical
Model".
J. Stat. Phys. 52, 595-608 (1988).
[97] R. Livi, A. Politi, S. Ruffo: "Repeller Structure in a Hierarchical Model:
1. Topological Properties".
J. Stat. Phys. 65, 53-72 (1991).
[98] R. Livi, A. Politi, S. Ruffo: "Repeller Structure in a Hierarchical Model:
II. Metric Properties".
J. Stat. Phys. 65, 73-95 (1991).
 SCHRODINGER DIFFERENCE EQUATION ... 547

[99] J.M. Luck: "Cantor Spectra and Scaling of Gap Widths in Deterministic
Aperiodic Systems".
Phys. Rev. B 39, 5834-5849 (1989).
[100] V.A. Mandelshtam, S.Ya. Zhitomirskaya: "lD-Quasiperiodic Operators.
Latent Symmetries".
Commun. Math. Phys. 139,589-604 (1991).
[101] F. Martinelli, E. Scoppola: "Introduction to the Mathematical Theory of
Anderson Localization".
Rivista del nuovo cimento 10, 1-90 (1987).
[102] P. Michel: "Stricte ergodicite d'ensembles minimaux de substitutions".
C. R. Acad. Sci. Paris Serie A-B 278, 811-813 (1974).
[103] J. Moser: "An Example of a Schrodinger Equation with Almost Periodic
Potential and Nowhere Dense Spectrum".
Comment. Math. Helvetici 56, 198-224 (1981).
[104] V.I. Oseledec "A Multiplicative Ergodic Theorem, Ljapunov Character-
istic Numbers for Dynamical Systems".
Trans. Moscow Math. Soc. 19, 197-231 (1968).
[105] S. Ostlund, S. Kim: "Renormalisation of Quasiperiodic Mappings".
Physica Scripta T 9, 193-198 (1985).
[106] L.A. Pastur: "Spectral Properties of Disordered Systems in the One Body
Approximation" .
Commun. Math. Phys. 75, 179-196 (1980).
[107] L. Pastur, A. Figotin: "Spectra of Random and Almost-Periodic Opera-
tors".
Springer Verlag: Grund. der math. Wissen. 297, (1992).
[108] D.B. Pearson: "Quantum Scattering and Spectral Theory".
Academic Press, (1988).
[109] D.B. Pearson: "Singular Continuous Measures in Scattering Theory."
Commun. Math. Phys. 60, 13-36 (1978).
[110] J. Peyriere: "On the 1}ace Map for Products of Matrices Associated with
Substitutive Sequences".
J. Stat. Phys. 62, 411--414 (1991).
[111] J. Peyriere: This volume.
[112] J. Peyriere, Z.-Y. Wen, Z.-X. Wen: "Polynomes associes aux endomor-
phismes de groupes libres".
L'Enseignement Mathematique 39, 153-175 (1993).
[113] M. Queffelec: "Substitution Dynamical Systems. Spectral Analysis".
Springer Verlag: Lecture Notes in Math. 1294, (1987).
[114] M. Queffelec: This volume.
548 A. SlitO

[115] M. Reed, B. Simon: "Methods of Modern Mathematical Physics.

I : Functional Analysis.
IV: Analysis of Operators".
Academic Press, (1980).
[116] N. Riedel: "Point Spectrum for the Almost Mathieu Equation".
C.R. Math. Rep. Acad. Sci. Canada VIII, 399-403 (1986).
[117] N. Riedel: "Almost Mathieu Operators and Rotation C*-Algebras".
Proc. London Math. Soc. 56, 281-302 (1988).
[118] N. Riedel: "Absence of Cantor Spectrum for a Class of Schrodinger Op-
erators".
Bull. of the A.M.S. 29, 85-87 (1993).
[119] N. Riedel: "The Spectrum of a Class of Almost Periodic Operators".
Preprint (1993).
[120] N. Riedel: "Regularity of the Spectrum for the Almost Mathieu Opera-
tor".
Preprint (1993).
[121] R. Riklund, M. Severin, Y. Liu: "The Thue-Morse Aperiodic Crystal, a
Link Between the Fibonacci Quasicrystal and the Periodic Crystal".
Int. J. Mod. Phys. B 1, 121-132 (1987).
[122] R. del Rio, N. Makarov, B. Simon: "Operators with Singular Continuous
Spectrum: II. Rank One Operators".
Commun. Math. Phys. 165, 59-67 (1994).
[123] D. Ruelle "Ergodic Theory of Differentiable Dynamical Systems".
Publ. Math. IRES 50, 275-306 (1979).
[124] S. Saks: "Theory of the Integral".
Dover Pub. Inc., (1964).
[125] T. Schneider, D. Wurtz, A. Politi, M. Zannetti: "Schrodinger Problem
for Hierarchical Heterostructures".
Phys. Rev. B 36, 1789-1792 (1987).
[126] M.A. Shubin: "Discrete Magnetic Laplacian".
Commun. Math. Phys. 164, 259-275 (1994).
[127] B. Simon: "Almost Periodic Schrodinger Operators: A Review".
Adv. in Appl. Math. 3, 463-490 (1982).
[128] B. Simon: "Schrodinger Semigroups".
Bull. Amer. Math. Soc. 7, 447-526 (1982).
[129] B. Simon: "Kotani Theory for One Dimensional Stochastic Jacobi Ma-
trices".
Commun. Math. Phys. 89, 227-234 (1983).
[130] B. Simon: "Almost Periodic Schrodinger Operators IV: The Maryland
Model".
Annals of Phys. 159, 157-183 (1985).
 SCHRODINGER DIFFERENCE EQUATION ... 549

[131] Ya.G. Sinai: "Anderson Localization for One-Dimensional Difference Schrodin·

ger Operator with Quasiperiodic Potential".
J. Stat. Phys. 46, 861-909 (1987).
[132] E. Sorets, T. Spencer: "Positive Lyapunov Exponents for Schrodinger
Operators with Quasi-Periodic Potentials".
Commun. Math. Phys. 142, 543-566 (1991).
[133] A. Soshnikov: "Difff!rence Almost-Periodic Schrodinger Operators: Corol-
laries of Localization".
Commun. Math. Phys. 153, 465-477 (1983).
[134] B. Sutherland, M. Kohmoto: "Resistance of a One-Dimensional Qua-
sicrystal: Power Law Growth".
Phys. Rev. B 36, 5877-5886 (1987).
[135] A. Siito: "The Spectrum of a Quasiperiodic Schrodinger Operator".
Commun. Math. Phys. 111, 409-415 (1987).
[136] A. Siito: "Singular Continuous Spectrum on a Cantor Set of Zero Lebesgue
Measure for the Fibonacci Hamiltonian".
J. Stat. Phys. 56, 525-531 (1989).
[137] D. Thouless: "A relation Between the Density of States and Range of
Localization for One-Dimensional Random Systems".
J. Phys. C 5, 77-81 (1972).
[138] D. Thouless: "Bandwidths for a Quasiperiodic Tight Binding Model".
Phys. Rev. B 28, 4272-4276 (1983).
[139] D. Thouless: "Scaling for the Discrete Mathieu Equation".
Commun. Math. Phys. 127, 187-193 (1990).
[140] M. Toda: "Theory of Nonlinear Lattices".
Springer Verlag: Springer Series in Solid-State Sciences 20, (1981).
 COURSE 18

SchrOdinger equation in a hierarchical potential

Herve Kunz

Institut de Physique Theorique EPF-Lausanne

PHB-Ecublens CH-1015 Lausanne
Switzerland

After a brief summary of the work done on dynamics in a hierarchical

potential, we describe the rigorous results obtained on the spectral
properties of a one dimensional Schr~dinger equation with a hierarchical
potential.

1. INTRODUCTION

Recently there has been interest in the relaxation properties of hierarchical

structures like spin glasses, macromolecules and computing structures, for
example. As our experience with human hierarchical constraints teaches us,
one should expect a considerable slowing down of the motion under such
constraints. In this context, it appeared also natural to study the motion of
a quantum particle in such a medium. An experimental realisation of such
motion, when it is essentially constrained to be one-dimensional is possible
in artificially grown superlattices.

As it is well explained in a nice review article by R. Rammal, G.

Toulouse and M. Virasaro [lJ a characteristic feature of such systems is that
the natural distance between states is ultrametric, i.e. all triangles are
isoceless in this distance. Ultrametric spaces have been studied by
552 H. Kunz

mathematicians since one century and are at the basis of the so-called p-
adric analysis, which was used in some models of field theory.

One of the first problem studied was that of a random walk in a

hierarchical environment, which was apparentely first introduced to analyse
long-time behaviour of economic systems. Its one-dimensional version was
analysed by the physicists Huberman and Kerszberg [2], Domany and Teitel
[3] Maritan and Stella [4].

The model is defined as follows :

At each site m of a one-dimensional lattice a random walker can make a

jump to its neighbours with a probability w(m ± 11 m), so that the probability
Pm of finding the walker at m obeys the usual equation

Pm = w(m + 1) I m) (Pm +l - Pm) + w(m-ll m) (Pm -l -Pm) (1)

The hierarchical structure is introduced by defining the jump probability

to be :

w(m + 1) I m) = w(m 1m + 1) = ). Rorder(m) (2)

where the order of the integer m is defined by the equation

m = ± 20rder(m) (2 I + 1) (3)

I being an integer.

)., and R represent parameters of the problem. In other words the

probability to jump from the site m + 1 to the neighbouring site m depends
on the position of m in the hierarchy measured by order m. If m is odd its
order is zero, and the probability is equal to A.. If m is even, but when
divided by two gives an odd number, its order is one, and the probability is
)'R. If m is four times an odd number, its order is two and the probability is
)'R2, and so on. Starting at the origin, Le. Pm(O) = am one is interested by
the diffusion properties of the walker, characterized by the long-time
behaviour of

(m 2 (t» =L m 2 Pm(t) (4)

m
 SCHRODINGER EQUATION IN A HIERARCHICAL POTENTIAL 553

1
These auth ors fiIn d anomLOUSI d ' fuSlon
f' I'f R < _1
2 (when the
hierarchical constraint is important)

(m 2(t» ... tx(R) (5)

and normal diffusion if R > ~

(m 2 (t» ... D(R)t (6)

One can say that the system exhibits a dynamical phase transition.

A surprising feature is however the fact that the same result would be
obtained if one were considering the jump probabilities w(m+ll m) as
independent random variables with a singular probability distribution f(w) =
w-b , with b = 1 + \fUn2,lnR). But in this case one knows that all states are
localised, despite the fact that the configuration average (m 2 (t» "" tx(R).
This poses the interesting question of understanding the nature of the
spectrum of the operator appearing in the right-hand side of equation (1),
when the jump probabilities are hierarchically distributed.

A related problem appears in the study of the one-dimensional

Schr6dinger equation with a hierarchical potential, whose hamiltonian is
given by

(H",>m ="'m+l + "'m-l + v(m) "'m (8)

the potential v(m) being hierarchical, i.e.

v(m) = A. f(order (m» (9)

Such a model was introduced by Schneider, Wurtz, Politi, Zannetti [5]

and Livi, Maritan, Ruffo [6]. It was numerically studied and further
analysed by Livi, Politi and Ruffo [7]. Related models were also introduced
by Roman [8] and Keirstead, Ceccatto and Huberman [9]. A continuous
version of a qualitatively similar multidimensional model was rigorously
analysed by Jona-Lasino, Martinelli and Scoppola [10].
554 H. Kunz

We have rigorously studied the spectral properties of the one-

dimensional Schrodinger equation with a hierarchical potential in a joint
work with R. Livi and A Silto [11].

2. HIERARCmCAL POTENTIALS

In order to better understand the properties of hierarchical potentials, some

mathematical concepts appear useful.

On the integers n E Z one can define a valuation

I n I s= s-order (n) (10)

s being any real number larger than one. This valuation possesses the
following properties :

1) I n I s ~,and I n I s = 0 ~ n = 0
2) I-nls = Inls
3) In+m Is ~ max [ I n Is, Im Is]
4) Inmls = Inls Imls (11)

Compared to the familiar absolute value, the unusual property of this

valuation is the third one, which implies: 12n I s ~ I n Is.

It is then possible to introduce a distance between integers

den, m) = In - m I s (12)

Such a distance possesses the ultrametric property

den, m) ~ max (d(n, k), d(k, m» (13)

for any triple n, m, k.

This means that in this metric, all triangles are isoceles.

A well known construction of the real number out of the rational ones is
the following.

A sequence of rational numbers {rn} is a Cauchy sequence with respect

to the usual distance if
 SCHRODINGER EQUATION IN A HIERARCHICAL POTENTIAL 555

V e >0 I rn - rm I < e when n, m ~ N (e).

Another Cauchy sequence {r~} of rational numbers is said to be

equivalent to {rn} if

I rn - r' I when n, m ~ M (e ).
m

Consider now the set of equivalent Cauchy sequences. One says that
they define a real number x, which is identified with the rational number r,
when these sequences are equivalent to the constant sequence rn = r. In all
other cases x is said to be an irrational number. In all cases one says that
the sequences rn converges to the real number x. Cauchy sequences of real
numbers always converge to a real number. One says that the real numbers
have been obtained from the rational ones by completing them with respect
to the usual distance.

One can follow exactly the same strategy for the integers if one consider
Cauchy sequences of integers Ink} , with respect to the ultrametric distance,
i.e. such that

V e 1~-r}ls s e when k,I~N(e).

Completing the integers in this way gets a new ring, the so-called 2-adic
integers, 12.
Any number Ol e 12 can be written as

Olj e {O, I} (14)

A finite number of 1 in the sequence of Olj means that Ol is a positive

integer, whereas a finite number of 0 means that Ol is a negative integer.

1 co
For example: -1 = 1-2 =.~ 2i
J=O

The "rest" corresponds to 2-adic integers. Addition of two numbers Ol and

Ol' is made by adding Olj and Ol'j modulo 2. The order of a number Ol is then
556 H. Kunz

defined as

order (CD) = min {k ICDk = 1} (15)

We could of course have replaced 2 in this construction by any prime

number p. We would define in this way p-adic numbers.

An ultrametric potential (hierarchical) can then be simply defined as

v(n) = U{ I n I s) (16)

and can be extended to 12.

If f is continuous and bounded

lim v(n+mk) =v(n+CD) CD e 12 (17)

k -+ 00

This property shows that the introduction of 2-adic numbers is natural.

In fact the sequence n + CD (n e Z ) can be considered as a trajectory in

an ergodic system.

The map is here simply the translation by one unit and the invariant
measure on this space is the Bemouilli one, i.e. like in coin tossing the {CDj}
can be considered as independent random variables with :

1
p(CDj = 0) = p(CDj = 1) = 2"

We see therefore that an ultrametric potential is an ergodic one.

Moreover, when f is continuous and bounded it is limit-periodic.

Indeed, we have the basic property

v(l 2N + k) =v(k)
(18)
k = 1, ... , 2N -1 VN~O, leZ

Therefore, if vN(n) is the periodic potential of period 2N such that

vN(n) = v(n) when n = 1, ... , 2N then we see that VN and v differ only at
the sites I 2N I e Z therefore
 SCHRODINGER EQUATION IN A HIERARCHICAL POTENTIAL 557

lim sup 1v(n) - VN(n) 1 = A- lim

N ~ 00 n N ~ 00 n~

which is the definition of a limit-periodic function.

It turns out that for an understanding of the qualitative properties of

the spectrum of the Schrodinger equation, only three type of potentials
have to be considered. Amusingly, they correspond to Coulomb potentials in
the ultrametric distance.

A-
R < 1 I _ R (1 - 1n 1R-1) Id Coulomb
v(n) = R =1 - A- log 1n 12 2d Coulomb (19)
A- 1
R>1 l-R (1- lnIR) 3d Coulomb

These potentials posses a dilatation symmetry

v(2n) = R v(n) + A- (20)

which suggest, as it is indeed the case, that they correspond to fixed points
of a renormalisation group transformation (RG).

3. SPECTRAL PROPERTIES OF THE SCHR<>nINGER EQUATION

Our main results can be summarised in the following theorem, which is

obtained for the specific potential given by equation (19).

Theorem
1) The spectrum of H, a(H), is a Cantor set, for all R
2) Ga,p labelJiDll

Consider the eigenvalues of the Dirichlet problem

H\jI = e\jl on [1, 2N - 1]

Call them eN (k) k = 1, ... 2N - 1

558 H. Kunz

Then the set {eN (k.»)N ,k forms a dense set in the spectrum. Points of this
set are at the border of the spectral gaps of H. (Remember that the gaps are
dense everywhere).

When e = eN (k), the integrated density of states of the hamiltonian H,

N(e), has the property

k
N(e) = 2N (21)

This property is completely analogous to the corresponding one for the

standard Cantor triadic set.

4. LIAPUNOV EXPONENTS

The following limit exist and defines the Liapunov exponent 'Y

'Y = lim sup x1 In II Ax Ax-I.·. AlII (22)

x-+ oo

where Ax is the transfer matrix

Ax = G 1
-v(x) -1
0
)
(23)

It can be proven that the Liapunov exponent 'Y vanishes for all energies
in the spectrum when R ~ 1 and for all energies except those in a residual
set S when R < 1. But in this last case we have 'Y = 0 when e E [ - 2, 2] for
example.

5. SPECTRAL MEASURES

When R ~ 1 (v(n) unbounded) the spectrum is purely singular continuous.

When R < 1 (v(n) bounded)

the spectrum is purely continuous in a(H) I S, for example in [- 2, 2].

 SCHRODINGER EQUATION IN A HIERARCHICAL POTENTIAL 559

The set S corresponds to possible non isolated eigenvalues near the

border of the spectrum, i.e. to localised states.

A natural conjecture is that when R < 1 the spectrum is absolutely

continuous, with the possibility of localised states at its upper boundary.
When R ~ 1, one expects that the spectrum is a Cantor set with zero
Lebesgue measure.

Numerical studies [7] have revealed a very interesting multifractal

structure of the spectrum. There exist unfortunately no rigorous results on
this type of properties. This is related to our poor understanding of gap
openings in periodic potentials.

The method used to obtain these results is to combine information

obtained from characteristic properties of one-dimensional Schr6dinger
equations with those of a dynamical system obtained by looking at the trace
map existing in these models, which can also be understood as a
renormalisation group transformation.

The trace map in our case is the following. Let Aj be the transfer matrix
(23). Then we have A;; = ~N +j j = 1, ... 2N - 1 because of property (18) of
the potential.

Therefore if Mn = ~n ... Al (24)

we have (25)

Indeed

Since det Aj = 1, equation (25) induces the following trace map,

if

"'n = trMn (26)

2 2
"'n+l = 'tn - 2 + R'tn ('tn - '"n- 1 + 2) n~l

'to =e , "'1 = e2 - A. e - 2 (27)

560 H. Kunz

The problem is to determine the behaviour of the "tn when n ~ and to DO

relate this behaviour to the position of the energy e in the spectrum.

The result is that e E (J (H) if and only if {"tn} is bounded when R ~ 1 and
{"tn} is bounded or {"tn} is unbounded, but

"tn
~ ~ R , when R < 1.
n-1

The last possibility, when R < 1, appear when the energy is in the
exceptional set S.

Another way to look at the problem is to decimate over the odd sites
recursively (RG transformation). Then after a length scale reduction of 2m ,
the equation H'I' = e'l' can be shown to become

(28)

We see that the energies e have been renormalised to "tm , and the
coupling constant A. to A.m. More explicitely one finds :

2
110+1 ="tm - 2 - A.m "tm
(29)

with "to = e A.o = A.. In fact A.m = A. Rm "tm-1 "tm-2 ... "to. (30)

In this way one can obtain, for example an "explicit" formula for the
Green function

fo
21t n-1
Gz (0, 0) =
dO
21t
A. 1
- [2cosO - z + 1-R + 1-R l: A.nTI ("tm + 2cos 2m O)-l 1 (31)
n=l
DO

m::::O
r

but now "to = z (complex)

so that in order to get more information from this formula, one would need
to analyse the trace map for complex values of "tm , a rather difficult task.
 SCHRODINGER EQUATION IN A HIERARCHICAL POTENTIAL 561

One sees therefore that the results on this type of problem are by no
means complete and further work would be needed, to understand the
multifractal structure of the spectrum for example.

Finally, let us note that nothing has been done on the multidimensional
case, except on the diffusion properties of some continuous model [10].

References

[1] Rammal R., Toulouse G., Virasoro M., Rev. Mod. Phys. 58, 765 (1986)
[2] Huberman B.A., Kerszberg M., J. Phys. A : Math. Gen. 18, L331 (1985)
[3] Teitel S., Domany E., Phys. Rev. Lett. 55, 2176 (1985)
[4] Maritan A, Stella A., J. Phys. A : Math. Gen. 19, L269 (1986)
[5] Schneider T., Wurtz D., Politi A., Zanetti M., Phys. Rev. B 36, 1789
(1987)
[6] Livi R., Maritan A, Ruffo S., J. Stat. Phys. 52, 595 (1988)
[7] Livi R., Politi A., Ruffo S., J. Stat. Phys. 65, 53 (1991), J. Stat. Phys.
66, 73 (1991)
[8] Roman H., Phys. Rev. B 36, 7163 (1987)
[9] Keirstead W. P., Ceccatto H. A., Huberman B. A., J. Stat. Phys. 53,
733 (1988)
[10] Jona-Lasino G., Martinelli F., Scoppola E., Ann. Inst. Henri Poincare
42, 733 (1985)
[11] Kunz H., Livi R., SOW A Comm. Math. Phys. 122, 643 (1989)
 COURSE 19

Introduction to multifractal analysis

J. Peyriere

Universite Paris-Sud
MatMmatiques, Bat. 425
Unite associee au CNRS 757
91405 Orsay cedex, France

1. INTRODUCTION

The term "multifractal" and the use of the thermodynamical formalism in this
context appear in [44, 32, 43J. But the description of multifractal behaviour
goes back to [61-63]. Since then, multifractal analysis has been of wide use
and it is impossible to give an exhaustive set of references. Originally, one
analyzed measures (though it was not always clearly stated). But nowadays
one analyzes also functions and sequences of Choquet capacities. There has
been a tremendous development of the literature on theoretical considerations
on multifractality as well as on its applications.
The aim of this course is limited. We simply seek to provide the reader with
the necessary concepts. To do so, we consider the simplest situation of the
analysis of measures on the real line.

2. SHORT RECALLS ON MEASURES

2.1. Borel sets and measures

Let n be an integer larger than 1. We denote by B the smallest set of subsets

of IRn which contains the open sets and which is closed with respect to the
following operations: taking the complement of a set in IRn or taking the
union of a countable family. The elements of B are called Borel sets.
564 J. PeyrU~re

A positive Borel measure on ntn is a function J.L from B to [0, +00] such that,
for any countable (Le. either finite or denumerably infinite) family {An} of
disjoint Borel sets, one has

(1)

One could think of J.L(A) as the mass of A. Then the additivity requirement
becomes natural. Nevertheless, there are several difficulties. The first one is
that, in general, it is impossible to assign a mass to any subset of ntn in such a
way that (1) be still valid. The sets for which this is possible are called measur-
able. Borel sets are measurable. The second difficulty is that it is not easy to
describe Borel sets. This means that functions on B are even more difficult to
define and describe. Fortunately, there is a theorem by Caratheodory asserting
that a function defined on parallelepipeds and satisfying (1) can be extended
to B so as to become a measure. For instance, the Lebesgue measure is the one
which assigns to a parallelepiped the product of the lengths of its sides. From
now on, instead of positive Borel measure we shall simply say measure.
A measure J.L is said to be bounded if its total mass ",(ntn) is finite (when it
equals 1, we have a probability measure).
If J.L is a measure on ntn and P(x) a property depending on the point x in
ntn , the locution "P(x) is true for ",-almost every x" means that the set of
points x such P(x) is false is contained in a set of J.L-measure 0.

2.2. An example: the trinomial measures

We are going to define measures on [0, 1[ (this means measures on nt such that
any Borel set not intersecting [0, 1[ has zero measure). Let us divide the interval
I = [0, 1[ into three equal intervals 10 = [0,1/3[, h = [1/3,2/3[, and [2/3,1[,
and perform the same operation on each of them. The three subintervals of I J
will be denoted by Ijo, Ijl' and I J2 • And so on. In other terms, if Cl, C2,· .. ,Ck
is a sequence of O's, 1's, and 2's,

Now, let us consider m = (mo,ml,m2) a triple of non-negative numbers

such that mo + ml + m2 = 1, and set

(2)

(one can remark that this last quantity is simply mtomf 1 m: 2 where #j stands
for the number of i's such that 1 ~ i ~ k and ei = j).
The Caratheodory theorem shows that the above equation defines a proba-
bility measure J.Lm. When mo = ml = m2 = 1/3, we get the Lebesgue measure
 MULTIFRACTALS 565

on [0,1], and when rno = rn2 = 1/2, we get the natural measure on the triadic
Cantor set.
Of course, what we have just done using triadic expansions and successive
trisections of intervals can be done with another base. In this case, we call the
resulting measure J.Lm a multinomial measure.

3. SEVERAL NOTIONS OF DIMENSIONS

3.1. Definitions

3.1.1. Hausdorff measures and dimension

Let E be a subset of IRn. If p is a positive real number, we call p-covering

of E by balls (or simply p-covering) any countable collection of balls B J the
diameters of which (denoted by IBj I) are less than p and such that E C U EJ'
If Q and p are positive real numbers, we set

and
(3)

As a function of E, when restricted to Borel sets, 1-{0I. is a measure called

the Hausdorff measure in dimension Q. As a matter of fact, in the present
course, we are mostly interested in its behaviour as a function of Q. We have
the following facts

- if Q < f3 and 1-{OI.(E) < 00 then 1-{f3(E) = 0,

and
- if Q < f3 and 1-{f3(E) > 0 then 1-{Q(E) = 00
from which we deduce that the number dim E = sup {Q I1-{Q(E) = oo}, called
the Hausdorff dimension of E, has the following property: 1-{Q(E) equals 00 for
o :::; Q < dim E, and 0 for Q > dim E (nothing general can be said on 1-{Q(E)
for Q = dim E).
The Hausdorff dimension has the following properties:

E C F ==> dimE:::; dimF,

dimUEn = sup dim En (countable union),
dim (countable set) = O.
566 J. Peyriere

3.1.2. The Bouligand-Minkowski dimension (or "box dimension")

Let E be a bounded subset of JRn . Consider NE{E) the minimum number of
elements of coverings of E by balls of diameter e. The Bouligand-Minkowski
dimension of E is the number

u
A{E) = 1·Imsup log NE
E--+O
1
-
(E) .
oge
Here is an other way (from which the name box dimension comes) of comput-
ing~. Consider a hypercube containing E, divide it into 2n equal hypercubes,
and let wn{E) be the number of those small hypercubes which meet E. It is
then easy to show that

u
A{E) = 1·Imsup logwn{E) .
n--+oo n log 2
Obviously, if E c F we have ~(E) :::; ~(F), and if E is the closure of E,
~(E) = ~(E).
There are several other ways of defining ~. Here is one which is relevant to
our topic. If E is a bounded subset of JRn and p a positive number, we call
p-packing of E any collection of disjoint balls of diameter less than p such that
any of them intersects E. If a is a positive number, define

P:{E)=su p {2:IB j l'" I {B j } is ap-packing ofE}

and
P"'{E) = lim P"'{E).
p--+O p
(4)

It can be shown [92] that

~(E) = in£{ a I
~ 0 P"'{E) = o} = sup {a ~ 0 IP'" = oo}. (5)
But this apparent symmetry with the definition of Hausdorff dimension is
misleading: P'" restricted to Borel sets is not a measure.
3.1.3. Packing dimension
The Bouligand-Minkowski dimension behaves differently than the Hausdorff
dimension: it may happen the box dimension of a countable union U En is not
equal to the supremum of ~En. Indeed, the box dimension of a single point is
o and the box dimension of the set of rational points in the interval [0,1] is l.
To obviate this inconvenient, C. Tricot [93] introduced the following variant of
the Bouligand-Minkowski dimension

DimE=inf sUP{~{En)IECUEn}. (6)

n
This is the packing dimension. It is well behaved with respect to countable
unions, and therefore assigns the value 0 to countable sets. More on packing
dimension and packing measures can be found in [94].
 MULTIFRACTALS 567

3.2. Estimating the Hausdorff dimension

3.2.1. Comparison of these dimensions

For a bounded set E, one has the relation 1{~(E) ~ pC>. Np(E) for which it
follows that dim E ~ Ll(E). Because dim and Dim are well behaved with
respect to countable unions, we conclude that dim E ~ Dim E.

3.2.2. Lower bound for dim

Proposition. Let JL be a positive and bounded Borel measure on IR n such

that, for any ball B, one has JL(B) ~ C IBIC>. (where, of course, C does not
depend on B). Then, any Borel set E such that JL(E) > 0 has a Hausdorff
dimension not less a.
Proof. Consider an arbitrary covering E C UjEN B j of E by balls. One has

jEN jEN

Therefore

1{C>.(E) 2: ~ JL(E) > 0

and
dimE 2: a.

Indeed, if we are interested in evaluating the Hausdorff dimension of a single

set E and willing to use the preceding proposition, we have to consider a
suitable measure JL. Moreover, the asumption on JL is too stringent in the sense
that we need to ask for the inequality JL(B) ~ C IBr only for balls which meet
E. But the assumption on JL can be further weakened.
Billingsley lemma. Let JL be a positive and bounded Borel measure, and E
a Borel set such that JL(E) > 0 and

. 10gJL(B) }
E C {x I B'\.{x}
hm Iog IBI = a

(where B "" {x} means that the ball B shrinks to {x}}. Then dimE = a.
The proof will not be given here. It can be found in [8], with a little more
general a statement. The following corollary is of common use.
Corollary. Let JL be a positive and bounded Borel measure such that

. log JL(B)
hm
B'\.{x}
I IBI
og
=a for JL-almost every x.
568 J. PeyrU~re

Then any Borel set E such that IL( E} >

than Q.
°has a Hausdorff dimension not less
For short, if a measure fulfills the condition in the above corollary, we shall
say it satisfies an approximate HOlder condition of order Q.

3.2.3. Example
We are going to compute some Hausdorff dimensions by using the trinomial
measures introduced in section 2.2.
If x is a point in [0,1[, consider its triadic expansion

X=
""'
L-~ c· (with Cj E {O, 1, 2} },
J:2:1

(this expansion is unique if x is not of the form k 3-1; in this case, we consider
the expansion ending with 000 ... ), and set, for i = 0, 1 or 2,

(7)

Let us now consider a trinomial measures ILm. It results from formula (2)
that
_ [ 'Po(x,n) 'Pl(x,n) 'P2(X,n)] n
ILm (1n ( x )} - mo m1 m2 , (8)

from which it follows

{X I n-+oo
lim CPi(X, n} = mi for i = 0, 1, 2} C
{x
. loglLm{In(X}) ~ }
Il~~ log IIn(x}1 = - ~ m.log3 m. , (9)

where log3 stands for the base 3 logarithm.

Set
N m = {x n-+oo I
lim CPi(X, n} = m. for i = 0, 1, 2} . (1O)

The cJ's, considered as functions on [0, 1[ endowed with the probability ILm,
are Li.d. (independent identically distributed) random variables. Then it re-
sults from the law of large numbers that, for i = 0, 1, and 2, CP.(x, n} goes to
m, as n goes to 00 for ILm-almost every x. This means that ILm (Nm ) = 1.
It then results from (9) and Billingsley lemma(i} the following formula due
to Eggleston [21]
2
dimNm = - L mi log3 mi· (11)
i=O

e) To tell the truth, a little more subtle form of Billingsley lemma is needed:
instead of considering all the intervals shrinking to {x}, it is enough to use the triadic
intervals only.
 MULTIFRACTALS 569

We can even say a bit more: any set of positive JLm-measure has a Hausdorff
dimension not less than - L~=o mi log3 mi.
Two facts are worth to be noticed. First, if mOml m2 #- 0, any subinterval of
[0,1] has a positive JLm-measure, and therefore meets N m . In particular, N m
is dense in [0,1]. Secondly, if m #- m ' , we have N m n N m, = 0, JLm (Nm ) = 1,
and JLm' (Nm,) = 1. This means that JLm and JLm' are mutually singular (see
M. Queffelec's contribution to this volume).

4. THE MULTIFRACTAL FORMALISM

4.1. Definitions

We are given an integer v 2: 2 and a probability measure JL on [0,1[.

The interval [fn, ~ [ (for 0 S j < v n ) is denoted by In,j, or sometimes
let ... en (with €J E {O, 1,···,v -I} for 1 S j S n) as explained in section 2.2.
If x E [0,1[, then x belongs to a unique In,j which will be denoted by In(x).
log" will stand for the base v logarithm.

4.1.1. The pointwise Holder exponent

The Holder exponent of JL at x is

(12)

provided that the limit exists.

Roughly speaking, the meaning of this exponent is that when n is large one
has JL (In(x)) :::oJ IIn(x)I"'h(x)

4.1.2. The T function

For q E JR, we set
(13)

where the dash means that the summation runs over the intervals of positive
measure. The function Tn is non-increasing and concave. Clearly, Tn(l) = 0 and
limsup(-Tn(O)) is the Bouligand-Minkowski dimension of the closed support
Supp JL of JL.
One has

Therefore one has Tn(q) S q - 1. Moreover, since the intervals In,j are nested,
and due to the inequality (al + ... + a,,)q 2: ai + ... + a~, valid for q 2: 1 and
positive aJ's, we have Tn(q) S Tn+1(q) for q 2: 1.
570 J. PeyrU~re

Set
T(q) = liminfTn(q).
n--+oo
(14)

It results from the above discussion that T(q) = lim Tn(q) for q > 1, and
T(O) = -~(SuPPJL). In particular, T(I) = 0 and T(q) ~ q - 1.
If lim..~oo Tn(q) exists for any q, then T = lim Tn.
4.1.3. The multifractal spectrum
For any nonnegative number 0, we consider the set
(15)
and its Hausdorff dimension

1(0) = dimE". (16)

The function 1 is called the multifractal spectrum of the measure 1'.
4.1.4. The multifractal formula
The multifractal formalism is said to hold for I' at 0 if one has
1(0) = T*(O), (17)
where T* is the Legendre transform of T (see appendix 1).
Up to now, there exist no satisfactory conditions implying the multifractal
formalism to hold. The existing ones are either too complicated or do not
emcompass sufficiently many cases.
One could expect that that the existence, for any q, of the limit lim Tn(q)
would suffice. But, in the next section, it will be shown that it is not so. In
section 4.3, we take as an example the trinomial measure, and, in section 4.4,
we discuss a setting where the multifractal formalism works.

4.2. A counterexample to the multifractal formalism

We take II = 5. Then the intervals In" are the 5-adic sub-intervals of [0,1[,
and it is more convenient to shift to notations close to the ones of section 2.2:
the intervals of the k-th generation are denoted by I~1~2 ... ok where the e,'S are
in the set {0,I,2,3,4}.
We define a measure I' in the following way

I' (I~lE2···~k) =
0 if el E {I, 2, 3},
0 if el = 0 and it exists j E {2, ... , k} such that eJ E {O, 2, 4},
0 if el = 4 and it exists j E {2, ... , k} such that ej E {I,3},
2- k if el = 0 and eJ E {I, 3} for 1 < j ~ k,
3 - /2 if
1 k el = 4 and ej E {O, 2, 4} for 1 < j ~ k.
 MULTIFRACTALS 571

This measure is the sum of two pieces which both are multinomial measures.
Let us compute r(q). We have

Therefore
(q-l)IOg52 if q~1
r(q) = {
(q - 1)log5 3 if q:$ 1,
and
-co if a tt [log5 2, log5 3]
r*(a) ={ a
if a E [log5 2, log5 3].
On the other hand, the only values of a for which Ea is non-empty are log5 2
and log53. It follows that, for a E ]log5 2, log5 3[, one has f(a) f:. r*(a).

4.3. The trinomial measures

For multinomial measures, the computation of the various multifractal quanti-

ties can be carried out. This is why we first treat this example.
Let J-Lm be a measure of the kind defined in section 2.2 and take lJ = 3. It is
easy to check the following formula

[tmr]n,
.=0
(18)

from which we get an explicit expression for r:

2
r(q) = -log3 L m;. (19)
.=0

In other terms
(20)

Now we are going to prove that the multifractal formalism holds for /-Lm.
First, we have to prove the inequality DimEa :$ r*(a). We are going to
prove it in the case a < r'(O), and give clues for the other case.
Let f3 be a number such that 0 < f3 :$ T'(O). Set

(21)

On the other hand, it results from (20) that, for q ~ 0, the number ~J(f3) of
intervals 101 "'0, such that /-L (I0 1"' 0 ,) ~ 3- J ,6 is less than 3J (q,6-T(q)). It follows
572 J. Peyrii~re

that N 3 - J (BnU3)) ::; 3j (q/3-T(q)) for q ~ 0 and j ~ n. From this, we deduce that
~ (BnU3)) ::; qf3 - 7(q) for any q ~ 0, and, since f3 ::; 7'(0), ~ (BnU3)) ::; 7*(f3).
But, we have

where ko > (7'(0) - a)-I. Therefore,

Dim E,,::; lim Dim

k_oo
UBn(a + -k1 )::; lim sup ~ (Bn(a
k-oo n>1
+ -k1 )) = 7*(a).
n~1 -

The argument when a > 7'(0) is similar, but one has to consider the sets

(22)

We now have to show the opposite inequality. Let f = (fo, h, h) be a triple

of non-negative numbers subject to conditions

fo + h + 12 = 1 (23)
and
(24)

It results from the law of large numbers that <P.(x) converges towards f.
for ILf-almost every x for i = 0, 1 ou 2. Therefore, due to formula (8) and
" ( ) logJLm (In(x)) d f
cond ItlOn 24, logIIn(x)1 converges towar s a as n goes to 00 or ILralmost
every x. This means that ILf(E,,) = 1. It results (see section 3.2.3) that
dim E" 2: - 2::;=0 f. log3 j.. Therefore we have

dim Eo: ~ sup { - t. f.log 3 f.1 f satisfies (23) and (24) } (25)

By using the Lagrange multipliers theory, this supremum can be shown to

be 7*(a) and reached for fJ = m~ / ~ m~, where q is determined by the
09<3
equation 7'(q) = a. In other terms, there exists a measure ILf satisfying an
approximate Holder condition of order a and such that ILf(Eo:) > O.
It is worth noticing that it is not the box dimension of Eo: which is
involved. Indeed, the set Eo:, which is dense in [0,1]' has a box dimension
equal to 1.
 MULTIFRACTALS 573

4.4. A few results

Theorem. One has

DimE", ::; T*(a).
Sketch of proof. Let us consider the map Yn : j f---t log// It (1n,j). It can be
considered as a random variable if the set {O, 1", . , v n - I} is endowed with
the uniform probability. Its moment generating function is

E (etYn ) = v- n L'It (I n ,j ) t/log //,

O::;J<//n
therefore

But

So, by the Chernoff inequality (see the appendix on large deviations),

lim sup .!.log// #

n ...... oo n
{j IIt (In'J) ~ v-n",} ::; inf 1ta - T
t2:0 og v
(-1
t )
og v
= T*(a), (26)

the last equality being valid if a is less than the value for which T is maximum.
As in the preceding section, this inequality yields an upper bound for N 3 -n (B",)
which allows to conclude.

As it was said already, there are only few simple conditions insuring that the
converse inequality holds. Here is one [11,85].
Proposition. If there exists a constant c > 0 such that, for any sequences
and 7]1, ... , 7]n of D's, 1 's, and 2's, one has
C1, ... , Cm

C-11t(I'1, .. ·,,~)It(IT/l'''',T/n)::; 1t(I'l""'~,T/l''',T/n)::; CIt(I'l, .. ,'~)It(IT/l''',T/n)

then the multifractal formalism holds for It at any a of the form a = T'(q).
The condition in this proposition means that the measure It is somewhat
self-similar. The multinomial measures, which are strictly self-similar, satisfy
this condition with c = 1.
Sketch of proof. For a fixed q, one can construct a measure Itq such that there
exist a constant, such that, for any sequence C1, ... , Cn, one has

,v
-1 nr(q) It (1'l,"','n )q < (1 )<
_ Itq 'l,"','n _ ,v nr(q) It (1'l,",'n )q . (27)

The construction of such a measure, called a Gibbs measure, in this non-

dynamical context can be found in [78, 79, 85]. It can be shown, by using
574 J. Peyriere

the large deviations theory, that, if 0 = 1"(q) exists, the measure ILq satisfy
an approximate Holder condition and that IL(Ea) > O. Then the conclusion
follows from Billinsley lemma.
One can notice that, in the case of trinomial measures, the measure ILl
obtained by an optimization process is a Gibbs measure. Indeed, instead of
having (27), one has

ILl (I€1,. ... £n ) -- V

n7'(q)
IL (1£l,. ... £n )q •

4.5. Negative dimensions

The preceding sections raise the following intriguing question: what is the
meaning of negative values of 1'*(o)? If one goes through the proof of the
theorem in section 4.4, one can show that 1'*(0) < 0 implies Ea = 0. But this
is not a satisfactory answer. Indeed, (26) is still valid. Moreover, if lim n -+ oo 1'n
exists, the Gartner-Ellis theorem (see the appendix on large deviations) yields
a more precise form of (26): 1'* = j, where

j(o) = lim lim sup .!:.logll

£-+0 n-+oon
# {k I 0 :::; k< vn and 1.!:.IOgll IL{In.k)
n
+ 01 :::; e} .

This means that the previous formula gives a direct definition ofthe spectrum.
This is why B. Mandelbrot [64-72] prefers to start from j instead of 1'. Besides,
the consideration of / leads him [72] to define a new notion of dimension (the
ELNA dimension) which can assume negative values. Negative values of /(0)
correspond to extremely rare events. When using T, these negative values
correspond, via Legendre transform, to values of 1'(q) for large Iql's. This
is why, when dealing with experimental data, it is difficult to get a precise
estimate of negative values of j(o), and accurately evaluating these negative
values necessitates oversampling. For a complete discussion of negative - or
latent - dimensions see [70, 72].

4.6. Concluding remarks

The setting for the multifractal analysis of measures can be broadened. For
instance, one can define the l' function when the intervals 1n.i are not nested
or when the intervals of the same generation do not have the same length (see
e.g. [43, 17, 11]). Also it can be extended further [84].
Multifractal analysis of functions and of sequences of Choquet capacities
leads to interesting developements and has important applications, but is not
within the scope of this course, which is only meant as an introduction to this
subject.
 MULTIFRACTALS 575

Appendix 1

Legendre transforms

If 'IjJ is a concave function from lR to [-00, +00[, its Legendre transform is

the following function 'IjJ* from lR to [-00, +00[:

"p*(a) = inf [aq - "p(q)]. (L)

qER

Indeed the only interesting case is when "p is not identically -00, and we
assume it is so. In this case, the set on which it is finite is an interval on which
the infimum in formula (L) can be computed.
As an infimum of affine functions, "p* is concave. If"p is differentiable, its
derivative is non-increasing and "p*(a) = aqcx - "p(qcx) where qcx is such that
a = "p'(qcx).
The function "p* is finite on an certain interval. If"p is finite on the whole of
lR, then the limits amin and a max of 'IjJ(q)/q as q goes to +00 or to -00 exist
and are finite. Then the interval on which 'IjJ* is finite has amin and a max for
endpoints.
One can show the equality "p** = "p.

Appendix 2

Large deviations

Let {Yn } be a sequence of random variables, consider the logarithms of their

moment generating functions

1
"pn(t) = -logE(exptY
n
n ),

and set "p(t) = limsuPn--+oo "pn(t).

If a E lR and t 2: 0, one has

P (Yn 2: na) = P (e tYn 2: enta )

< expn'IjJn(t)
- expnta
~ expn("pn(t) - ta) ,

from which it follows

lim sup .!.logP (Yn 2: na) ~ "p(t) - ta,

n~oo n
576 J. Peyriere

where 'IjJ(q) = limsup'IjJn(q).

n-+oo
Minimizing on t 2: 0 yields the Chernoff inequality :

lim sup ..!.log P (Yn 2: na) :s: - sup [ta - 'IjJ(t)] .

n-+oo n t::::o

Of course, to deal with the probability P (Yn :s: na) one has to consider non-
positive t's.

Indeed the Chernoff inequality is sharp in the following situation. Assume

that, for any real number t, the limit

lim 'ljJn(t)
n-+oo
= 'IjJ(t)
exists and is not identically -00.
Consider the Legendre transform of 'IjJ

lex) = sup [tx - 'IjJ(t)].

tEIR

We then have the following theorem.

The Gartner-Ellis theorem.

- inf l(x)::; lim inf ..!.log P

xEja,b[ n-+oo n
(Yn n E]a, b[)

:s: lim sup ..!.logP

n-+oo n
(Yn n E [a, b]) :s: - inf lex).
xE[a,bj

In case the function I is continuous and minimum at a, one has

lim ~ logP C: 2: a) = -lea) for a 2: a

lim ~ logP (:n :s: a) = -lea) for a:S: a.

More on large deviations can be found in [13, 18, 22, 23, 84, 89].

Bibliography

[1] Arbeiter M. and Patzshke M., Random self-similar multifractals. Advances

in Math., submitted.
[2] Arneodo A., Bacry E., and Muzy J.F., Wavelet analysis offractal signals.
Direct determination of the singularity spectrum of fully developped turbulence
data. Preprint (1991).
 MULTIFRACTALS 577

[3] Arm~odo A., Muzy J.-F., and Bacry E., Multifractal formalism for fractal
signals. Preprint, Paul-Pascal Research Center, 1992.
[3] Bedford T., Hausdorff dimension and box dimension in self-similar sets.
In Proceedings Conference on Topology and Measure (V. Binz, GDR, 1987).
[4] Ben Nasr F., Analyse multifractale de mesures. Comptes Rendus Acad.
Sc. Paris 319, Serie I (1994), 807-810.
[5] Berroir J.-P., Analyse multifractale d'images. These, Universite Paris-
Dauphine, 1993.
[6] Berroir J.-P. and Levy Vehel J., Medical images segmentation with mul-
tifractals. In IFCS, 1993.
[7] Berroir J.-P. and Levy Vehel J., Multifractal tools for image processing.
In SCIA, Tromsl/l, 1993, IAPR
[8] Billingsley P., Ergodic theory and information. J. Wiley 1965.
[9] Bacry E., Muzy J.-F., and Arneodo A., Singularity spectrum of fractal
signal from wavelet analysis: exact results. J. Stat. Phys., to appear.
[10] Bohr T. and Rand D., The entropy function for characteristic exponents.
Physica 25D (1987), 387-398.
[11] Brown G., Michon G., and Peyriere J., On the multifractal analysis of
measures. J. Stat. Phys. 66 (1992), 775-790.
[12] Cawley R and Mauldin RD., Multifractal decompositions of Moran frac-
tals. Advances in Math. 92 (1992), 196-236.
[13] Chernoff H., A measure of asymptotic efficiency for tests of a hypothesis
based on the sum of observations.. Ann. Math. Stat. 23 (1952), 493-507.
[14] Chhabra A. and Jensen RV., Direct determination of the f(a) singularity
spectrum. Phys. Rev., March 1989.
[15] Chhabra A., Jensen, R, and Sreenivasan K., Extraction of underlying
multiplicative processes from multifractals via the thermodynamic formalism.
Phys. Rev. A 40 (1989), 4593-4611.
[16] Collet P. and Koukiou F., Large Deviations for Multiplicative Chaos. Com-
mun. Math. Phys. 147 (1992), 329-342.
[17] Collet P., Lebowitz J., and Porzio A., The dimension spectrum of some
dynamical systems. J. Stat. Phys. 47 (1987), 609-644.
[18] Cramer H., Sur un nouveau tMoreme-limite de la tMorie des probabilites.
Actualites Scientifiques et Industrielles 3, 1938.
[19] Daubechies 1. and Lagarias J., On a thermodynamic formalism for multi-
fractal functions. In The State of Matter, (dedicated to Elliott Lieb), eds. M.
Aizenman et H. Araki, World Scientific (Singapore, 1994), 213-265.
578 J. Peyriere

[20] Dubois J., Levy Vehel J., Sailhac P., and Gilbert F., MuItifractal Analysis
of Time Series : Application to Piton de la Fournaise Volcano. Journal of
Geophysical Research, submitted, 1994.
[21] Eggleston H.G., The fractional dimension of a set defined by decimal prop-
erties. Q. J. Math. Oxford, Ser. 20 (1949), 31-46.
[22] Ellis R., Large deviations for a general class of random vectors. Ann.
Prob. 12(1) (1984), 1-12.
[23] Ellis R., Entropy, Large Deviations and Statistical Mechanics. Springer,
1985.
[24] Evertsz C.J.G. and Berkner K., Large deviation and self-similarity analysis
of graphs: DAX stock prices. Chaos, Solitons & Fractals 6 (1995), 121-130.
[25] Evertsz C.J.G., Hayakawa Y., and Mandelbrot B.B, Exactly self similar
"left-sided" muItifractal measures. Phys. Rev. A42 (1990), 4528-4536.
[26] Evertsz C.J.G and Mandelbrot B.B., Chaos and Fractals. Springer 1992,
922-952.
[27] Evertsz C.J.G and Mandelbrot B.B., Self-similarity of harmonic measure
on DLA. Physica A 185 (1992), 77-86.
[28] Evertsz C.J.G and Mandelbrot B.B., Harmonic measure around a linearly
self-similar tree. J. Phys. A 25 (1992), 1781-1797.
[29] Falconer K., The Geometry of Fractal Sets. Cambridge University Press
1985.
[30] Falconer K., Fractal Geometry: Mathematical Foundation and Applica-
tions. J. Wiley & Sons (1990).
[31] Falconer K., The MuItifractal Spectrum of Statistically Self-Similar Mea-
sures. Preprint, 1993.
[32] Frisch U. and Parisi G., Fully developed turbulence and intermittency in
turbulence, and predictability in geophysical fluid dynamics and climate dy-
namics. In International School of Physics "Enrico Fermi" , course 88. M. Ghil,
ed. (North-Holland, Amsterdam, 1985, p. 84).
[33] Ghez J.-M. and Vaienti S., On the wavelet analysis for muItifractal sets.
J. Stat. Phys. 57 (1989), 415-419.
[34] Ghez J.-M. and Vaienti S., Rigourous wavelet analysis for multifractal
sets. In Large Scale Structures in Nonlinear Physics, eds. J.-L Fournier and
P.-L. Sulem. Lecture Notes in Physics 392, 320-326. Springer.
[35] Ghez J.-M. and Vaienti S., A Frostman-like theorem for the wavelet tran-
form on fractal sets. Nonlinearity 4 (1991),1241-1249.
[36] Ghez J.-M. and Vaienti S., Integrated wavelets on fractal sets: 1. The
correlation dimension. Nonlinearity 5 (1992), 777-790.
 MULTIFRACTALS 579

[37] Ghez J.-M. and Vaienti S., Integrated wavelets on fractal sets: II. The
generalized dimensions. Nonlinearity 5 (1992), 791-804.
[38] Ghez J.-M. Orlandini E., Tesi M.-C., and Vaienti S., Dynamical integral
transform on fractal sets and the computation of entropy. Physica D 63 (1993),
282-298.
[39] Ghez J.-M., Siboni S., and Vaienti S., Wavelet tranform and multientropy
analysis of strange sets. In Progress in Wavelet Analysis and Applications (Pro-
ceedings ofthe international conference "Wavelets and Applications"). Eds.
Y. Meyer and S. Roques. Frontieres.
[40] Godreche C. and Luck J.-M., Multifractal analysis in reciprocal space and
the nature of the Fourier transform of self-similar structures. J. Phys. A: Math.
Gen. 23 (1990), 3769-3797.
[41] Gutzwiller M.C. and Mandelbrot B.B., Invariant multifractal measures in
chaotic Hamiltonian systems, and related structures. Phys. Rev. Lett. 60
(1988), 673-676.
[42] Gupta V.K. and Waymire E., A statistical analysis of mesoscale rainfall
as a random cascade. J. Appld. Meteor 32 (1993), 251-267.
[43] Halsey T.C, Jensen M.H., KadanoffL.P., Procaccia I., and Shraiman B.J.,
Fractal measures and their singularities: The characterization of strange sets.
Phys. Rev. A 33 (1986), 1141-1151.
[44] Hentschel H. and Procaccia I., The infinite number of generalized dimen-
sions of fractals and strange attractors. Physica 8D (1983), 435-444.
[46] Holley R. and Waymire E., Multifractal dimensions and scaling exponents
for strongly bounded random cascades. Ann. of App. Prob. 2(4) (1992),
819-845.
[47] Jaffard S., Sur la dimension de Hausdorff des points singuliers d'une fonc-
tion. Comptes Rendus Acad. Sc. Paris 314 Serie I (1991), 31-36.
[48] Jaffard S., Construction de fonctions multifractales ayant un spectre de
singularites prescrit. Comptes Rendus Acad. Sc. Paris, Serie I, 315 (1992),
19-24.
[49] Jaffard S., Formalisme multifractal pour les fonctions. Comptes Rendus
Acad. Sc. Paris 317 Serie I (1993), 745-750.
[50] Jaffard S., Multifractal' Formalism for functions Part I: Results valid for
all functions et Part II: Selfsimilar functions. Preprint (1993).
[51] Jaffard S., Determination du spectre de singularites de la fonction
En~l n- 2 sin 7rn 2 x. Preprint du CERMA.
[52] Kahane J.P. and Peyriere J., Sur certaines martingales de Benoit Mandel-
brot. Advances in Math. 22 (1976), 131-145.
580 J. Peyriere

[53] Levy Vehel J., Medical Image Segmentation with Multifractals. In New
Approaches in Classification and Data Analysis, serie: Studies in Classification,
Data Analysis and Knowledge Organization. Springer 1994.
[54] Levy Vehel J., Multifractal Segmentation of Remotely Sensed Images. In
Fractals in Geoscience and Remote Sensing, Office for Official, Publications of
the European Community 1994.
[55] Levy Vehel J. and Berroir J.-P., Image analysis through multifractal de-
scription. In IFIP 2nd International Working Conference on Fractals in the
Natural and Applied Sciences, ed. M.M. Novak, 1993.
[56] Levy Vehel J. and Mignot P., Multifractal segmentation of images. Frac-
tals 2(3) (1994),371-377.
[57] Levy Vehel J., Mignot P., and Berroir J.P., Multifractals, texture, and
image analysis. In CVPR, 1992.
[58] Levy Vehel J. and Vojak R., Multifractal description of road traffic struc-
ture. Technical Report 1943, INRIA, June 1993.
[59] Levy Vehel J. and Vojak R., Mutual Multifractal Analysis of Sequences of
Choquet Capacities: Preliminary Results. Preprint 1995.
[60] Mandelbrot B.B, Possible refinement of the lognormal hypothesis concern-
ing the distribution of energy dissipation in intermittent turbulence. In Sta-
tistical Models and Turbulence. Edited by M. Rosenblatt and C. Van Atta.
Lectures Notes in Physics 12, Springer 1972, 331-35l.
[61] Mandelbrot B.B., Intermittent turbulence in self-similar cascades: diver-
gence of high moments and dimension of the carrier. J. Fluid. Mech. 62 (1974),
331-358.
[62] Mandelbrot B.B., Multiplications aleatoires iterees et distributions invari-
antes par moyenne ponderee aleatoire, I & II. Comptes Rendus Acad. Sc. Paris
278A (1974), 289-292 & 355-358.
[63] Mandelbrot B.B., Fractal measures (their infinite moment sequences and
dimensions) and multiplicative chaos. Technical report, Physics Department,
IBM Research Center, Mathematics Department, Harvard University, 1986.
[64] Mandelbrot B.B., An introduction to multifractal distribution functions.
In Fluctuations and Pattern Formation (Cargese 1988). Edited by H.E. Stanley
and N. Ostrowsky. Kluwer 1988, 345-360.
[65] Mandelbrot B.B., Multifractal measures, especially for the geophysicists.
Pure and Applied Geophysics 131 (1989), 5-42.
[66] Mandelbrot B.B., A class of multifractal measures with negative (latent)
value for the dimension f(a). Fractals: physical origin and properties (Erice,
1988). Edited by Luciano Pietronero, Plenum 1989, 3-29.
 MULTIFRACTALS 581

[67] Mandelbrot B.B., Two meanings of multifractality, and the notion of neg-
ative fractal dimension. Soviet-American chaos meeting (Woods Hole, 1989).
Edited by Kenneth Ford and David Campbell, American Institute of Physics
1990.
[68] Mandelbrot B.B., Limit lognormal multifractal measures. Frontiers of
Physics: Landau memorial conference (Tel Aviv, 1988). Edited by E. Gostman.
Pergamon 1990,309-340.
[69] Mandelbrot B.B., New "anomalous" multiplicative multifractals: left sided
/(0:) and the modeling of DLA. Condensed matter Physics, in honour of Cyrill
Domb (Bar Ilan, 1990). Physica A168 (1990), 95-111.
[70] Mandelbrot B.B, Random multifractals: negative dimensions and the re-
sulting limitations of the thermodymamic formalism. Proc. London Math.
Soc. A434 (1991), 79-88.
[71] Mandelbrot B.B, The Minkowski measure and multifractal anomalies in
invariant measures of parabolic dynamical systems. In Chaos in Australia
(Sydney, 1990). Edited by G. Brown and A. Opie. World Publishing 1993,
83-94.
[72] Mandelbrot B.B., Negative ELNA dimensions and ELNA Holders, mul-
tifractals and their Holder spectra, and the role of lateral pre-asymptotics in
science. In J.P. Kahane meeting (Orsay, 1993). Journal of Fourier Analysis
and Applications 1995.
[73] Mandelbrot B.B. and Evertsz C.J.G., Multifractality of the harmonic mea-
sure on DLA, and extended self-similarity. Physica A 177 (1991), 386-393.
[74] Mandelbrot B.B., Evertsz C.J.G., and Hayakawa Y., Exactly self-similar
"left-sided" multifractal measures. Phys. Rev. A 42 (1990), 4528-4536.
[75] Mandelbrot B.B. and Riedi R., Multifractal formalism for infinite multi-
nomial measures. Advances in Applied Mathematics.
[76] Meneveau C. and Sreenivasan K.R., The multifractal spectrum of the dis-
sipation field in turbulent flows. Nuclear Physics B (Proc. Suppl.) 2 (1987),
49-76.
[77] Meneveau C. and Sreenivasan K.R., Measurement of /(0:) from scaling of
histograms and applications to dynamical systems and fully developed turbu-
lence. Phys. Lett. A 137 (1989), 103-112.
[78] Michon G., Une construction des mesures de Gibbs sur certains ensembles
de Cantor. Comptes Rendus Acad. Sc. Paris 308 (1989), 315-318.
[79] Michon G., Mesures de Gibbs sur les Cantor reguliers. Ann. IHP, Phys.
Th. 58 n° 3 (1993), 267-285.
(80] Michon M. and Peyriere J., Thermodynamique des ensembles de Cantor
autosimilaires. Chinese Ann. of Math. 15B (1994), 253-272.
582 J. Peyriere

[81] Muzy J.F., Arneodo A., and Bacry E., Wavelet analysis offractal signals:
direct determination of the singularity spectrum of fully developed turbulence
data. Springer, 1991.
[82] Muzy J.-F., Bacry E., and Arneodo A., Wavelets and multifractal formal-
ism for singular signals: application to turbulence data. Preprint
[83] Olsen L., A multifractal formalism. Advances in Math. to appear.
[84] Oono Y., Large deviation and statistical Physics. Progress of Theoretical
Phys. 99 (1989), 165-205.
[85] Peyriere J., Multifractal measures. In Probabilistic and Stochastic Meth-
ods in Analysis (Proceedings of the NATO ASI, II Ciocco 1991). Ed. J. Byrnes,
Kluwer Academic Publishers (1992).
[86] Procaccia 1., The characterization offractal measures as interwoven sets of
singularities: Global universality at the transition to chaos. Technical report,
department of Chemical Physics, The Weizmann Institute of Science, Rehovot
76100, Israel.
[87] Rand D., The singularity spectrum f(a) for cookie-cutter. Ergodic theory
and dynamical sys. 9 (1989), 527-541.
[88] Riedi R., An improved multifractal formalism and self-similar measures.
J. Math. Anal. Appli., to appear
[89] Sanov 1., On the probability of large deviations ofrandom variables. Mat.
Sb. 42 (1957), 11-44 (in Russian); English translation in Selected Translations
in Mathematical Statistics, 1961.
[90] Tessier Y., Lovejoy S., and Schertzer D., Universal Multifractals: Theory
and Observation for Rain and Clouds. J. Appld. Meteor. 32 (1993), 223-250.
[91] Tricot C., Sur la classification des ensembles boreliens de mesure de
Lebesgue nulle. These (nO 1921), Universite de Geneve (1980).
[92] Tricot C., Two definitions of fractional dimension. Math. Proc. Cam-
bridge Philos. Soc. 91 (1982), 57-74.
[93] Tricot C. and Taylor S.J., Packing measure and its evaluation for a brow-
nian path. Trans. Amer. Math. Soc. 288 n02 (1985), 679-699.
[94] Vojak R., Levy Vehel J., and Danech Pajou M., Multifractal description
of road traffic structure. In 7th IFAC/IFORS Symposium on Transportation
Systems: Theory and Application of Advanced Technology, 942-947, Tianjin
1994.
Annexes
 CmTent deterministic sequences

The elements collected below may enable the reader either to shape a
deterministic sequence having predefined properties, or to decipher from its
characteristics the principal features of a given disordered system.

Details, proofs and references are to be found in the Courses.

To use the 2- and 4-automata presented here, "feed" them with the
binary decomposition of natural integers starting with the initial state
indicated by an arrow, so as to generate the corresponding terms of the
sequence.

uo, Ub U2 ••• Un ••• always denotes the elements of the sequence

which is the fixed point of the substitution, dependini on u.o..
586

2 letter alphabet (0, 1) or (a, b)

Non-automatic, Sturmian sequence

1 • Substitution:
(J (a --+ ab 1 1 )
l...b --+ a 1 0

Eigenvalues : A. = 't or - rl with 't = ({5 + 1)/2.

Invertible
Pisot substitution, bounded atomic surface.
I Det(M) I = 1 : quasiperiodic.
Fourier module of rank 2.
Fourier Transform of sequence in the infinite length limit atomic
(countable number)
Number of factors of length n in infinite sequence ("complexity") :
p(n) = n + 1 for all n ~ 1

Construction :
Fo= 1 b
Fl = 1 a
F2 = 2 a b
Fa =3 a b a
F4 = 5 a b a a b
F5= 8 a b a a b a b a
F6 = 13 a b a a b a b a a b a a b

Sequence length at iteration p = Fp = Fp-I + F p -2

(Fo= Fl = 1)

lim ~ __ I+{5
F -'t- 2
p--+oo p

2 • Other definition:

Local concatenation rule for building blocks:

An+2 = An+l An (means "glue An+l and An")·
with Ao = b, Al = a .
 CURRENT DETERMINISTIC SEQUENCES 587

''GeDe.l~
_l':--..:II'FiboDaCCl•

2 letter alphabet (0, 1) or (a, b)

In general, non automatic sequences

1 • Substitution:
o(a-+ am bn
b-+a

EigenfJolueB : A. =(m ± ...J m 2 + 4n)l2.

Pi80t if n < m + 1 : bounded atomic surface.
Quasiperiodic if n = 1, Fourier module of rank two.
Unbounded atomic surface if n ~ m + 1.
Invertible if I Det(M) I = 1 ~ n = 1.
Fourier Transform of sequence in the infinite length limit atomic or
singular continuous depending on m and n.

Sequence length at iteration p Fp =mFp.1 + n Fp-2

with Fo = Fl = 1

lim 1+1
p-+ p
00
= l (m + ..Jm + 4n) ='t(m, n) =largest eigenvalue of M
2

2 • Other definition

Local concatenation rule for building blocks : (See Fibonacci)

A = A:+l A; (means "glue m times A with n times A")
~ ~1 p
with Ao =b, Al = a .
588

Regular paper-folding
4 letter alphabet (a, b, c, d) then 2 letter (0, 1)

(::
1 - Substitution

;~ 1
1 01 10 00 J
(J c~ad M= ( 0 1 0 1
d ~ cd
o 0 1 1

Followed by the projection: a, b ~ 0, c, d ~ 1 yields the sequence.

Eigenvectors (0, 0, 0, 0) (1, -1, -1, 1) (0, - 1, 0, 1) (1, 1, 1, 1)

Eigenvalues o o 1 1
Projection (0, 0) (0,0) (- 1, 1) (1, 1)

Substitution non-Pi sot : unbounded variation.

Rank of Fourier module is not finite
Projection : Point spectrum, not quasiperiodic, unbounded atomic surface.
Non invertible
Fourier Transform of projection in the infinite length limit: atomic

Number of factors of length n in infinite sequence ("complexity") :

p(n) = 4n for n ~ 7

Construction
a
a b
abc b
abcbadab
abc bad a b abc dab c b

then by projection 0 0 1 0 0 1 0 0 0 0 1 1 0 0 1 0
Sequence length at iteration p = 2P

2 - Other definitions:
2.1 recursive on the alphabet (0, 1) :

=0
U4 n
{ U4n+2 = 1 with Uo = 0
u2n+l = un
the sequence being denoted
 CURRENT DETERMINISTIC SEQUENCES 589

2.2 4-automaton:

~1

Period doubHng
2 letter alphabet (0, 1) or (a, b)

1 . Substitution

(J
(
1-+10
0-+11 or
(:: :: 1
1 0
2 )

Eigenvalues : A. = 2, - 1.
Substitution non-Pisot, unbounded atomic surface.
IDet(M) I = 2 : not quasiperiodic, infinite rank Fourier module.
Non invertible
Fourier Transform of sequence in the infinite length limit: atomic
Construction
1
1 0
101 1
10111010

Sequence length at iteration p = 2P

590

2 • Other definitions:

2.1 recursive on the alphabet (0, 1)

{ u 2n =1
u2n+l = 1 - un with Uo = 1

the sequence being Uo u 1 u2 ... un ...

2.2 2-automaton

3· Property

summing up the terms modulo two gives the Prouhet-Thue-Morse sequence

(up to its first term)
1 0 1 1 1 0 1 0
1 1 0 100 1 1
 CURRENT DETERMINISTIC SEQUENCES 591

Circle map
3 letter alphabet (a, b, c), then 2 letter (0, 1)

1 . Substitution
a-+cac
(
ab-+acc a c
c-+abc a c
M=( 122
001
232
)

followed by projection a, b -+ 0, C -+ 1.

Eigenvector (0, - 1, 1) (4 + 2't, - 3 - 2't, 1) (2 - 2't, 2't - 1, 1)

Eigenvalue -1 2-{5=-'t3 2+{5=-c3
Projection (0,0) (1 + 't, - 1) ('t, 1)

Substitution non-Pisot
Projection: bounded atomic surface. Quasiperiodic with Fourier module of
rank two.
Non automatic
Invertible
Fourier Transform of sequence in the infinite length limit: atomic

2 • Other definition

un =X (na)
A

WI
'th "!
L> = 2 ' a= 't
-2 + {5 ,
't = 1--2- and X is the characteristic
A
function of the interval [0, A] on the circle.
The sequence obtained is identical with the fixed point of a with Uo = a.
592

Prouhet-Thue-Morse

2 letter alphabet (0, 1) or (1, - 1), or (a, b)

1 . Substitution

0(: : : ~ 1
1
1
1 )

Eigenvalues 0 and 2.
Substitution non Pisot
No Fourier module.
Fourier Transform of sequence in the infinite length limit singular
continuous may have atomic component
Number of factors of length n in infinite sequence ("complexity")

3.2 m + 4r if O:s; r < 2 m • l

p(n)= { 4.2m + 2r if 2m-l s r < 2 m
with n = 2m + r + 1, r integer

Construction
a
a b
a b b a
abbabaab
a b b a b a a b b a a b a b b a

Length of sequence at iteration p = 2P.

2 • Other definitions

2.1 recursive on the alphabet (0, 1) :

with Uo = 0
 CURRENT DETERMINISTIC SEQUENCES 593

2.2 2-automaton

2.3 local concatenation rule for building blocks

An+l =An An (where 0 = 1, 1 = 0)

with Ao = 0, hence Al = 0 1, etc.

2.4 Denote 0 1 10 1 001 1 00 1... = Uo ul u2 ... un ... :

un is the sum modulo 2 of the binary digits of n

Rudin-Shapiro

4 letter alphabet (a, b, c, d) then 2 letter (0, 1) or (1, -1)

1 - Substitution
a -+ ab
( b -+ ac 11 01
o 1
10
0
00
1
J
a c -+ db
d -+ dc
o 0 1 1

followed by the projection a, b -+ 0, c, d, -+ 1,

or a, b = - 1, c,d = + 1
using the alphabet (- 1, + 1)

Eigenvectors (1,-1,-1,1) (1,1,1,1) (-1, 1+12, -1-12,1) (-1,1-12. -1+12,1)

Eigenvalues o 2 12 -12
Projection (0, 0) (1+1) (1, -1) (1, -1)

Substitution non-Pi sot :

Projection with unbounded variation, no point spectrum, no Fourier module
Number of factors of length n in infinite sequence ("complexity")
594

p(n) = 8n - 8 for all n ~ 8

Fourier Transform of sequence in the infinite length limit absolutely
continuous with bounded intensity
The square root N property

N-l
sup L un eiqn s CW
q n=O

with C = 2(1 + f2),

where N is the sequence total length and Un is
on the alphabet (- 1, + 1)
Construction
a
a b
a b a c
a b a cab d b then -+ 0 0 0 1 0 0 1 0 or -1, -1 -1, 1 -1 -1 1 -1
with the notation 110 Ul U2 ... Un .....

Length of sequence at iteration p = 2P.

2 • Other definitions

2.1 recursive on the alphabet (0, 1)

U2n =un
{ U4n+l =un with Uo =0
U4n+3 = 1 - U2n+l
2.2 4-automaton

o 1 1 0
_-4-=-_ --*-- --+--~O
~a-~-~---'--ft-~-~
O~ 1~ 10
7'
2.3 Un is the number of (possibly overlapping) occurrence
of 11 in the binary expansion of n, mod. 2
U29? 29 =( 1 1 1 0 1) base two
1 1 1 0 1 two occurrences of 11, hence u29 = 0
 Glossary

Tlut following-elementll - they are necessarily incomplete, for which we

apologize - have been collected for non-specialist readers, in order to facilitate
tlutir access to tlut Courses.

Abelian group : its elements commute. Example Z (the integers under

addition)
Abelianization: Kramer
Absolutely continuous: Mendes France 4.3, Queffelec, Berthe, Sdto.
See Walters P., Ergodic theory - Introductory lectures, Springer.
Lecture Notes, Math. 458 (1975)
Acceptance domain: Patera
Almost periodic: Meyer 2 definition, Janssen 1,7, Berthe, Sdto,
Almost periodic in the sense of Besicovitch : Mendes France 4.2
Almost periodic Radon measure : Meyer 2 definition
£-AImost periodic : Meyer 2 definition
Almost-Mathieu equation: Sdto
Alphabet: Kramer, Mendes France 1.1, Queffelec, Dekking 2, Peyriere
Trace Maps, Sdto. A finite set J.I, the elements of which are called
"letters". From the letters are formed words (or blocks) by
juxtaposition just as in the current language
Alternating sequence: Mendes France
Ammann decoration : Katz 5.2
Amorphous: Rivier
Aperiodic crystal: Steurer 2
Approldmants : Janssen 3.3" -Steurer 3, Sdto. Lattice periodic structure
obtained from a quasiperiodic one by a (small) phason strain.
Arrhenius law : Kalugin
Atomic measure : Berthe, Queffelec. See measure
Atomic surface : Janssen 4.2, Kalugin, Katz 2.2, Rivier. The closure of the
set of points in the unit cell in superspace which are equivalent by
lattice translations with atomic positions in physical space
Automatic sequence : Mendes France 1.4, Queffelec 4, Berthe
Automatic walk : Dekking 2, page 4
Automaton (finite) : Mendes France 1.5
596

Automorph (of an integral quadratic form) : Janner 4. It is defined e.g. in

Dickson, Introduction to the Theory of Numbers, Dover, 1957 on
p.72 and in Buell, Binary Quadratic Forms, Springer, 1989 on p.31.
It is obtained as an integral solution of the Pell (plus) equation and
corresponds to a positive unit of the associated quadratic field

Bernoulli measure : Queffelec 1.4.3. The simplest example of a Bernoulli

measure is the measure associated to the infinitely lasting game of

tossing a unbiased coin, i. e.,

-(620 + ~6)
~= *
1
More generally, one calls Bernoulli measure any convergent infinite
product of discrete convolution measures having finite support
Binary integral quadratic forms : Janner 4.6. Binary quadratic forms ;
Quadratic fields ; Pell equation(s) : These concepts are discussed in
practically every textbook on number theory. See in particular :
1. D.A. Buell, Binary Quadratic Forms, Springer, 1989.
2. L.E. Dickson, Introduction to the Theory of Numbers, Dover,
1957.
3. H. Cohn, A Second Course in Number Theory, Wiley, 1963.
4. H. Hasse, Vorlesungen uber Zahlentheorie, Springer 1964.
Block frequency : Queffelec, Berthe 6.3,
Consider an infinite sequence u taking its values in a finite
alphabet JI. Let w be a word appearing in u. For all N, define N
w
the number of appearances of w in the first terms of the sequence u.
N
If lim ¥ exists, it represents the frequency of appearance of w
N-+-
in u.
Caution : the limit may not exist. For instance, if JI = {0,1} and
2n 2n+1
u =0 11 0000 ... 1 0 ..... the frequency of 0 or 1 does
not exist
Borel set : Queffelec, SQto. The class of "Borel sets'" on R is somehow the
class of sets that can be obtained from repeated denumerable unions
or denumerable intersections of intervals. There is no description of
Borel sets. Note that their cardinality and the cardinality of R are
identical, therefore they are relatively rare.
A '1ength" can be defined for these sets. See measure
Borel support : Queffelec. A Borelian measure ~ on R admits B for its
Borel support if ~(BC) = O. It is clear that it is not defined in a
unique way since, if N is negligible for J1, BUN and BrlNc are still
Borelian supports of ~.
 GLOSSARY 597

Bounded measure : Queffelec, Soto. See measure

Cantor function : Suto

Cantor measure : Queffelec. It is the Bernoulli measure

*1 2" (~o+ ~2.3 on)

00
1

or the measure defined by the following self similarity property:

fh(x) d~(x) = ~ fh ( ~) c4L(X) 1

+ -
2
fh(X;2) d~(x)
[0,1] [0,1] [0,1]
for any positive measurable h on [0,1]
Cantor set : Queffelec, Kunz, Suto
Cartan matrix : Patera
Cauchy sequences : Kunz
Cayley-Hamilton theorem : Queffelec, Peyriere Trace Maps, Soto
CBED Convergent Beam Electron Diffraction : Steurer
Cellular automaton : Mendes France 2.3, 3.3
Cesaro mean : Berthe

A sequence (un) n E N is said to converge by Cesaro mean if

_ 1 N-1
u =NI,un converges. If a sequence converges then it converges by
2
Cesaro mean. The converse is not true.
Chebyschev polynomials : Peyriere Trace Maps
Cholesteric liquid crystal: Rivier. See P.G. de Gennes, The physics of
liquid crystals, Clarendon, Oxford 1974 and The structure and
energetics of defects in liquid crystals, S. Chandrasekhar and
G.S. Ranjaneth, Adv. Phys. 35 (1986) 507
Circle potentials : SOto
Closure : Queffelec, Soto
Coboundary : Queffelec
Cocycle : Queffelec
Complexity : Mendes France 5, Soto
Composite structure : Janssen 1.2. A crystal phase consisting of two or
more subsystems which are displacively modulated incommensurate
crystal phases, whereas the basic structures are mutually
incommensurate. See: T. Janssen and A. Janner, Adv.Phys. 36, 519
(1987)
Concatenation : Peyriere Trace Maps
Continued fraction: Mendes France 2.5
598

Continuous measure: Queffelec. See measure

Continuous singular: Men~s France 4.3. See measure
Continuous spectrum; Suto
Correlation measures: Queffelec 3. See measure
Correlation function: Mendes France 4.3, Queffelec 3.1.1
Co:s:eter group: Kramer 4, Patera. A group generated by reflections
Curvature: Rivier
Cut and projection: Patera, Katz 2.3

Decoration: Katz 5
Decoration algorithm: Rivier
Decurving: Rivier
Defects: Rivier
Deflation: Katz 3.1
Delaunay property: Dekking 4
Delaunay set : Meyer 1 definition
Dense point spectrum: Suto
Dichotomy property: Queffelec
Dimension of a curve: Mendes France 2.4
Dirac mass (or Bragg peak): Mendes France 4.3
Disclination: Rivier
Discrete maz:imal spectral type: Queffelec
Discrete measure: Queffelec 1.2. See measure
Discrete point spectrum: Suto
Discrete substitution: Queffelec
Dislocation: Rivier
Distribution function: suto
Dragon curve: Mendes France 2.3
Drawing rule: Mendes France 1.1
Dynamic structure factor : Janssen 3.4. Differential cross section for
inelastic scattering as a function of energy and momentum transfer.
(See: S.W. Lovesey : Neutron Scattering, Oxford University Press,
Oxford)
Dynamical Bystem : Dekking 5. See page 19 of P. Walters, An introduction
to ergodic theory, Grad. Texts in Math. 79, Springer, New York,
1981
Dynamical theory of X-ray diffraction: Steurer 3.1, Mikulic 2.2.
Diffration theory taking into account multiple scattering processes.
It is employed for the calculation of diffraction curves of perfect
crystals (semi-infinite substrates, superlattices)
 GLOSSARY 599

Economic systems: Kunz

Elliptic matrix: SUto
Energy spectrum: Wen
Entropy: Berth'
Ergodic, Ergodicity: Mendes France 7.4, Queffelec, Dekking, Berth',
Suto. See page 27 of P. Walters, An introduction to ergodic theory,
Grad. Texts in Math. 79, Springer, New York, 1981. A measure JL is
called T-er,odic or er,odic if the T-invariant sets in A have either JL
measure 0 or 1. See measure
Essential spectrum: SUto
Essential support: suto
EXAFS Extended X Ray Absorption Fine Structure: Steurer

Fibonacci word: Wen

Fibre bundle: Rivier
Fized point: Mendes France 1.1, 3.2
Flips: Katz
Floquet theory: suto
Formal grammar: Kramer 6, Mendes France 1.1. A term of algebraic
linguistics. See Eilenberg S., Automata, Languages and Machines,
vol. A and B (Academic Press, New York, 1976).
Fourier coefficients: Meyer, suto
Fourier module: Janner, Janssen 1.1. An infinite set of vectors in
reciprocal space that are linear combinations with integer
coefficients of a discrete set of basis vectors
Fourier transform: Meyer 2, Mendes France 1.1, Suto
Fourier-Bohr coefficient: Mendes France 4.1
Fourier Bohr spectrum: Mendes France 4.3
Fractal atomic surfaces: Janssen 4.3. Atomic surface with a Hausdorff
dimension that is not an integer, and smaller than the dimension of
the internal space
Frank-Kasper phases: Rivier
Fraunhofer approdmation: Mikulic 2.2.1. Usually employed in the
kinematical theory to solve the diffraction integral. It replaces the
spherical waves of incoming and diffracted waves propagating in a
sample by plane waves. This leads to a formal expression of the
amplitude of the diffracted wave proportional to the Fourier
transform of the electron density.
Free group automorphism: Kramer 2
Free group: Kramer 2
Free group endomorphism: Dekking 2. A map from a free group to itself
which preserves the group operation.
600

Free monoid: Mendes France 1.1 definition, Wen, Peyriere Trace Maps
Fricke's lemma: Peyriere Trace Maps
Fricke-Klein geometry : Kramer 8. A geometric approach to pairs of
unimodular two by two matrices and their products due to R Fricke
and F Klein, Vorlesungen uber die Theorie der automorphen
Funktionen, I & II, Teubner, Reprint, 1965
Frustration: Rivier. An excellent introduction by G. Toulouse, in ref. [6]
of Rivier's Course, p. 195

Gap labelling: Sdto

Gaps: Sdto
Gauge invariance: Rivier
Generalized Cantor set: Sdto
Generalized random walk: Dekking 6
Glass: Rivier
Gordon theorem: Sdto
Gram matrix: Patera
Graph: Dekking 6
Gratings: Mikulic
Growth rules: Kalugin, Mendes France 1.1

Harmonious: Meyer 4
Harper equation: Sdto
Helly theorem: Katz
Hierarchical potential : Kunz. A potential which is a function of the
ultrametric distance from a point to the origin
Homogeneous automaton: Mendes France, 5, 1.2
Homology : Rivier
Homomorphism : Dekking. Any map which preserves given algebraic
structures
Homotopic path: Rivier
Homotopy, Homotopy cia.. : Katz, Rivier
Hopf tenure: Rivier
HRTEM High Resolution Transmission Electron Diffraction: Steurer
Hull: Sdto
Hyperbolic matrix: Sdto

Icosahedral group: Patera

Ideal: Peyriere Trace Maps
 GLOSSARY 601

Incommensurate crystal phases: Janssen 1.2. A quasiperiodic system of

particles obtained from a lattice periodic system by a (quasi) periodic
modulation. The latter can be displacive, occupational or a
combination of both
Indefinite ternary integral quadratic forms: Janner 5, 6. An
historically important book is Fricke/K.lein, Vorlesungen uber die
Theorie der automorphen Funktionen, I & II, Teubner, Reprint,
1965, I pp. 502-535. In a more modern exposition, the same topics
are discussed in : W. Magnus, Noneuclidean Tesselations and Their
Groups, Academic Press, 1974.
Infinite word: Wen
Inflation: Katz 3.1, 3.2
Inflation symmetry: Rivier
Integrated density of states: S\1to
Internal space : Janssen 1.3, 1.4, Rivier. The orthogonal component of the
physical space in the n-dimensional superspace. Also called:
perpendicular space
Invariant measure : Queffelec. A measure defined on (X, A) is called
T-inyariant or invariant with respect to the transformation T from X
to X if ~(T"lA) =~(A) for each A E .9l or (equivalently) if \i(x; ; t{x)

d~x» = Jt{Tx) d~(x) for each measurable function f ~ 0 on X.)

x

Invertible substitution: Wen

Ising automaton: Mendes France 7

Kinematical theory of X.ray diffraction: Mikulic 2.1, Steurer 3.1.

Diffraction theory, neglecting all multiple scattering processes. It is
mostly employed in connection with the Fraunhofer approximation.
Klfitze decomposition: Katz
Kotani theorem: S\1to
Kronig Penney model: Janssen 5.1. One-dimensional model for electrons
in solids, the quantum mechanical behaviour of an electron in a
potential consisting of delta peaks on atomic positions, or of
constant potentials between the atoms. See : S. Gasiorowicz :
Quantum Physics, Wiley, New York

Landauer resistance: S\1to

Laves phase: Rivier
602

Lebesgue decomposition theorem: Mendes France 4.3, Queffelec. Any

positive bounded measure on R can be uniquely decomposed into
the sum of a discrete (or atomic or pure point) measure, a singular
continuous measure and an absolutely continuous measure
Lebesgue measure: Mendes France, Queffelec, SUto. See measure
Lie algebra: Patera
Limit-periodic: Kunz. A function which is the uniform limit of periodic
functions.
Litteral map (or projection): Mendes France 1.3
Ljapunov exponent: SUto, Kunz
Local isomorphism: Wen

Manifold: Peyriere Trace Maps

Matching rules: Katz 1 definition, Kalugin, Mendes France 1.1
Mathieu equation: Suto
Maximal spectral type: Queffelec. A measure (or any measure equivalent
to it) that dominates all the spectral measures associated to a
unitary operator
ME Mossbauer Effect: Steurer
Mean almost periodic: Mendes France 4.3, Berthe
Mean-periodic sequence: Mendes France
Mean-square displacement: Dekking 7
Measurable function : Queffelec. A function f : X -+ R is called measurable
iff"l(B) E ~ for all Borel sets B E ~R).
Note that in case of X = R, any continuous function is measurable
Measurable mapping: Queffelec. A map T : X -+ X is called measurable if
T-l(A) E ~ for all A E ~
=
Measures on T [0, 2n) : Queffelec
- A measure s is called a Rajchman measure if

•
lim
1 n 1-+-
1\ J
sen) = 0 oft 1\sen) = eint ds(t)

- Riesz product: F. Riesz constructed this type of measure on T to

have an example of measure the Fourier cmfficients of which are
exactly known. The construction, explained in Queffelec 2., is not
simple.
- correlation measure the term comes from probability theory
Wiener introduced the class of complex sequences (un) for which:
 GLOSSARY 603

g(k) = lim
N.-+oo N
1 L u n +k u n exists for all k E Z.
n<N

(g(-k) = g(k»
One can then check that g(k) is the k th Fourier cmfficient of
a positive measure on T called correlation measure of u.
Measure: Queffelec, Suto
1) Borel measure on (R, B(R» (positive). It is the extension of the
notion of length to B(R). IJ. is such a measure if
IJ.«(/j) = 0 and IJ.(U Bn) = when the Bn are disjoint Borel sets.
n
IJ. is bounded if IJ.(R) < + 00, i. e., if its total mass is finite ..
2) IJ. ~ 0 having mass 1 is a probability measure
3) A IJml. measure is the difference of two positive measures IJ.I - 1J.2
(lJ.i ~ 0).
4) A complex measure is a measure of the form IJ.I + iIJ.2 with the lJ.i
real.
5) If X is a set with as-algebra .9L, we define m to be a positive
measure on ( Xs .9L ) if IJ. is an application from .9L to IR +
IJ.«(/j) = 0
{
such that ~
IJ.(U An) = .LJIJ.(U An)

when the An are disjoint elements of .9l .

6)A measure on R is discrete (or atomic or pure point) if it can be

written as Lan 6xn ' where xn E IR and 6x is a Dirac mass in x.

~ _lin
Example : Poisson measure I" 6n with parameter I.
~ e
n. n~0
7) A measure on R is continuous if it charges no point of R, i. e.
1J.({x} = 0 when x E IR
8) A measure on IR is sjn~ular if it is perpendicular to the Lebesgue
measure m of 1R, i. e. if it is carried by a Borel set B of zero m-
measure.
So m(B) = 0 while IJ.{B c ) = o.
A discrete measure is always singular in this sense. A
continuous measure can be singular or not.
Example : The Cantor measure IJ. is carried by the Cantor
604

triadic set K of zero m-measure. It is continuous and singular. JI. +m

is not singular.
9)A measure is absolutely continuous ( with respect to the
Lebesgue measure) if it annihilates the same Borel sets as m, that is
m(B) =0 if Jl.{B) = o.
By the Radon Nikodym theorem, JI. can be written f.m, where
f =~ is the Radon Nikodym derivative. It is a ~ 0 function of

L l(m). It is clear that the converse is true.

10) A measure JI. defined on (X, JI) is T-invariant or invariant with
respect to the transformatioon T from X to X, if JI.(T-IA) = JI.(A) for
all A E JI, or

i fl:x) dp.(x) i fl:Tx) dJl.{x) for all f measurable ~ 0 on X

11) A measure JI. is T-ergodic if the T-invariant sets of JI, i. e., A

-1
such that T A = A are of JI.-measure 0 ou 1.
12) A measure JI. is T-mixing if = lim JI.( T-n An B) = JI.(A) JI.(B)
n-+ DO
for all A, B E JI
Metallic glass: Rivier. See Glassy metals I, H. J. Guntherodt and H.
Beck, Springer 1981
MEM (Mu::imum Entropy Method): Steurer 4.1
Metric: Janner 1, 2. In a given real n-dimensional vector space V a
quadratic form f(x l,···,:x,J .. I.7,k=l gik Xi x k of rank n is given
for any X E V, where n is the rank of the symmetric matrix 8 =
(8ik) . A metric is defined in terms of a scalar product :

:xy = I.~k-l g·k X. Xk for any X, Y E V, so that gik is the

I, - I I
metric tensor. This metric is positive definite if {(x) = 0 for any
X E V, it is negative definite if - {(x) is positive definite and
indefinite otherwise. Note that this definition implies the extension
of the concept of metric space (normally requiring a positive definite
form) to include the pseudometric case as well. See for more details
the Encyclopedic Dictionary of Mathematics, Kiyosi Its, Ed. (The
MIT Press, 1987) p. 1014 (Metric Spaces) and p. 1292 (Quadratic
Fonns).
Metric entropy : Berthe. See Walters P •. Ergodic theory - Introductory
lectures, Springer. Lecture Notes, Math. 458 (1975)
Minimality: Sato
Model sets : Meyer 3 definition
Module: Let A be a commutative ring with unit e. An A -module is a
 GLOSSARY 605

commutative group E endowed with an operation A x E -+ E

denoted (A, x) -+ Ax such that
ex = x
(A + ~)x = Ax + J.I.X
i..(x + y) = Ax + AY
Monoid: Kramer
Monopole: Rivier
Morphisms: Mendes France 1.1, Wen
Multimetrical : Janner 1. It is an approach allowing to express structural
relations in crystals (considered as a 3-dimensional Euclidean matter
distribution having Fourier wave vectors which are integral linear
combinations of n basic ones spanning the Fourier Z-module) in
terms of geometrical transformations which satisfy certain
compatibility conditions, but which need not to be 3-dimensional,
nor Euclidean.
A similar approach is that of the superspace crystallography, based
on ndimensional space groups (which induce non-Euclidean
transformations in 3-dimensions) where the Fourier Z-module is
described as a n-dimensional lattice with positive definite metric. It
is also possible to consider more than one metric for a given Fourier
Z -module. One then obtains a multimetrical description of the
crystal. In this case, however, a point-atom approximation is needed
for allowing compatibility among the symmetry transformations
leaving different metric tensors invariant.
Mutually singular: Queffelec, S\1to. See measure

Negautomorph: Janner 4. Negautomorph of an integral quadratic form :

it is obtained as an integral solution of the Pell (minus) equation,
which when expressed in matrix form has a negative determinant. It
occurs only when the associated quadratic field has negative units,
which in fact correspond to the negautomorphs of the corresponding
integral quadratic form.
Nematic liquid crystal : Rivier. See P.G. de Gennes, The physics of liquid
crystals, Clarendon, Oxford 1974, The structure and energetics of
defects in liquid crystals, S. Chandrasekhar and G. S. Ranjaneth,
Adv. Phys.aQ(1986) 507
NMR Nuclear Magnetic Resonance: Steurer
Non commutative crystallography : Kramer 3. An extension of
crystallography where translations are replaced by non-commuting
shift operations
Non-abelian group: Kramer, Rivier. Its element do W2i commute. Example
: rotations
606

Normal sequence: Berthe. A normal sequence is such that all the blocks
of given length have the same frequency

2-0dometer : Queffelec
Only ergodic: Silto
Opacity: Mendbs France 6, Berthe
Orbit: Silto
Order of an integral valuation: Kunz. A good introduction can be found
in the book by Mahler K.: p-adic numbers and their functions,
(Cambridge University Press 1981)
Orientation preserving transformations: Janner. In the case
considered here (of automorphs of indefinite ternary integral
quadratic forms) relevant is the orientation of a circle, as defined in
the book by W. Magnus, Noneuclidean Tesselation and Their
Groups (.Academic Press, New York, 1974) on page 5. Orientation
preserving transformations for discontinuous groups of the
compactified complex plane are discussed in chapter II of the same
book.
Osedelet's theorem: Silto

p-adic integers: Kunz

Palindrome : Wen
Paperfolding walk: Dekking 1.2
Parabolic matrix: Silto
Pauli matrices: Rivier. A representation of quaternions, see any textbook
in quantum mechanics
Peano curve: Mendes France. See B. B. Mandelbrot, The Fractal
Geometry of Nature; W. H. Freeman, publ. 1982
Pell equation: Janner 4
Penrose patterns: Kramer, Dekking 4. A non-periodic tiling of the plane
with 5-fold point symmetry
Perron-Frobenius eigenvalue: Dekking 2, Rivier. See E. Seneta, Non-
negative matrices and Markov chains, 2nd ed., Springer, New York,
1981
Perron-Frobenius Theorem: Queffelec, Dekking. See E. Seneta, Non-
negative matrices and Markov chains, 2nd ed., Springer, New York,
1981
Phason: Janssen, Kalugin, Rivier
Phason Dip: Rivier
Phonons : Janssen 3. Elementary vibrational excitations of a crystal, the
quanta of the harmonic oscillations around the equilibrium points.
 GLOSSARY 607

(See : C. Kittel: Introduction to solid state physics, Wiley, New

York)
Pisot number: Meyer 2 definition, Janssen 4.2, Queffelec 1.4.3. Solution
of a polynomial equation with integer coefficients, which is larger
than one, whereas the moduli of its conjugates are smaller than
unity
Pisot substitution: Queffelec
Poisson summation formula: Meyer 2.4
Polya's theorem: Dekking 1. See page 9 of B.D. Hughes and S. Prager.
Random processes and random systems : an introduction. The
Mathematics and Physics of Disordered Media. Eds. B.D. Hughes
and B.W. Ninham, SLN Math. 1035, Springer, Berlin, 1983, p.l - 86
Polytope: Rivier, see H. S. M. Coxeter. Regular Polytopes, Dover 1973
Positive definite sequence : Queffelec. A sequence (an) of complex numbers

is positive definite if ~ t.
£.J 1
t J. a I·J
.. ~ 0 for all sequences (t , ... , t )
1 n
l~i~n
lsj~n
of complex numbers.
"-
If s is a measure on T, (s(n>ne :l)} is a positive definite
sequence. Bochner's theorem then says that these are the only
ones.
Positive measure: Queffelec. See measure
Powder diffraction: Steurer 3.1
Primitive substitution: Queffelec, Wen, Berthe, Stito.
A substitution a is primitive if its matrix Ma is primitive, i. e., for
k
some k ~ 1 all entries of Ma are > o. This means that there exists k ~
1 such that each letter b from the alphabet JiI appears in all the
iterates ak(a).)
Primitive word : Wen
Probability measure: Queffelec, Dekking, Stito
Procesi-Razmyslov algebra: Peyriere Trace Maps
Pseudo-random sequence: Queffelec. Pseudo random sequence : it is a
sequence the correlation measure of which is continuous.
Primitive substitution: It is a substitution the matrix M of which is
primitive, such that the crefficients of Mk are all > 0 for a certain k
~ 1. This means there exists k ~ 1 such that any letter v of the
alphabet JiI appears in all the iterates z\a) e e JI..
Pure point measure: Queffelec. See measure
Push-down automata: Kramer 7. A class of automata which are specific to
context-free languages
608

Q [+ I] : Patera
Q-invariant: Queffelec
Q-odometer: Queffelec
Quadratic fields: Janner
Quadratic Pisot substitution: Queffelec
Quasi-lattice walk: Dekking
Quasiperiodic: Janssen 1.1, 1.7, Katz 2.1. A function is quasiperiodic if its
Fourier transform is only different from zero on a Fourier module of
finite rank.
Quaternions : Patera, Rivier. Generalisation of complex numbers, useful
representation of rotations

Radon-Nikodym theorem: Queffelec

Rajchman measures: Queffelec. See measure
Random sequence: Queffelec. One definition can be the following : for a
sequence which takes the values 0 and I, a definition could be given
in analogy with an infinite sequence obtained by the outcomes
(head or tails) of independent throws with an unbiased coin : each
word w =wl w2 ... Wn E {O, U* of length n appears with frequency 2~
in the sequence.
Caution I : if the coin is biased (say heads has probability 113) then
mathematicians will still call the sequence random (words w = WI •..
wn now have frequency (:) (~k) (~n-k), where k = Iwlo is the
number of zeroes in the word w.
Caution II : even if the throws are no longer independent
mathematiciens will often call the sequence random. )
Random walk: Dekking I, Kunz. See B.D. Hughes and S. Prager,
Random processes and random systems : an introduction. The
Mathematics and Physics of Disordered Media. Eds. B.D. Hughes
and B.W. Ninham, SLN Math. 1035, Springer, Berlin, 1983, p.l - 86
Rank of a Fourier module : Janssen. Minimal number of basis vectors that
generates the module if one uses integer coefficients
Recurrent walk: Dekking I, Suto. See page 9 of B.D. Hughes and S.
Prager. Random processes and random systems: an introductionand
The Mathematics and Physics of Disordered Media. Eds. B.D.
Hughes and B.W. Ninham, SLN Math. 1035, Springer, Berlin, 1983,
p.l - 86
Reducible: Kramer
Reliability factor: Steurer
=
RG renormalization group: Kunz
 GLOSSARY 609

Rien product: Queffelec. See measure

Robinson tiling: Mendbs France, Robinson R. M.. Undecidability and non-
periodicity of tilings of the plane, Inventiones Math. 12 (1971) 177-
209
Root lattice: Patera
Root sublattice: Patera
Root system of Lie Algebra : Patera

SAED Selected Area Electron Diffraction: Steurer

Salem number : Meyer 2
Scale symmetric point set: Janssen 1.6. Here is a factor I and for each
radius R an origin in physical space such that multiplication by I of
all distances to the origin maps the point set within a sphere of
radius R into itself
Scaling: Dekking 3, Application of a non-degenerate linear map
Schwartz cIass S(Rn): Meyer
Self similarity : Dekking. See page 117 of K. Falconer, Fractal geometry ;
mathematical foundations and applications, Wiley, Chichester, 1990,
Katz
Self-affine: Dekking. See page 126-127 of K. Falconer, Fractal geometry;
mathematical foundations and applications, Wiley, Chichester, 1990
Semi-group: Wen
Semi-kinematical approximation: Mikulic 2.2.1. all approximation
theory of X-ray diffraction in the case of weak intensities
Semi-norm: Mendes France
Sequences:
Automatic : Mendes France 1.4
Circle map : Mendes France 1.4
Fibonacci : Mendbs France 1.1, Dekking
Generalized Fibonacci : Mendes France 1.4
Generalized Rudin-Shapiro : Mendes France 5.2.2
Paperfolding : Mendbs France 1.3, 2, Dekking
Period doubling : Mendes France 1.4
Prouhet- Thue-Morse : Mendbs France 1.2, Dekking, SUto
Rudin-Shapiro: Mendes France 1.4, Dekking
Sturmian : Mendes France 5.1, Bertbe, SUto
Thue-Morse : Mendes France 1.2, Dekking, Bertha, SUto
Toeplitz : Mendes France, Wen
Shelling problem: Patera
Shift operator: Dekking, Suto. See page 21 of P. Walters. An
introduction to ergodic theory, Grad. Texts in Math. 79, Springer,
New York, 1981
610

Shift-invariant orbit : Queffelec

Shift-invariant probability measure: Suto
Shuftle product: Mendes France 2.3
Simplices: Rivier
Singular continuous : Queffelec 1.3, Bertha, Suto. See measure
Singular decomposition of infinite word: Wen
Singular probability distribution: Kunz
Singular word: Wen
SL(2,C): Peyriere Trace Maps
Smectic liquid crystal: Rivier. See P.G. de Gennes, The physics of liquid
crystals, Clarendon, Oxford 1974, The structure and energetics of
defects in liquid crystals, S. Chandrasekhar and G. S. Ranjaneth,
Adv. Phys . .3..5. (1986) 507
SMF Symmetry-minimum function: Steurer
Spectral measure : Mendes France 4.3, Queffelec, Suto. This term is used
in a particular context. Let U be a unitary operator on a separable
Hilbert space H, and f an element of H, the sequence
( <ukf, f>H)k E Z
is the sequence of Fourier cmfficients of a positive bounded measure
a called the spectral measure of f.
f
Spectral projection: Suto
Spectral properties: Kunz
Spectral theory : Suto
Spectral type: Queffelec, Suto
Spectrum :
Fourier-Bohr : Mendes France
Energy : Mendes France, Suto
STEM Scanning Tunneleing Microscopy: Steurer
Strict ergodicity: Suto
Strongly connected automaton: Mendes France 6.2
Structure factor: Janssen 2.1. Fourier transform of a density consisting of
delta functions on the atomic positions. See : J.M. Cowley :
Diffraction physics, North-Holland, Amsterdam
Sturmian sequence: Mendes France 5.1, Wen, Bertha 6.2.. A sturmian
sequence is a sequence of complexity p(n) = n+1, for all n. It is the
itinerary of the orbit of a point of the unit circle under a rotation of
irrational angle a with respect to complementary intervals of length
a and 1- a.
Sub-manifold: Peyriere Trace Maps
Subordinacy: Suto
Substitutional chain: Janssen 4. One-dimensional chain of atoms, for
which the distances, hopping frequencies, site energies et cetera
 GLOSSARY 611

follow the order of letters in a word obtained by a substitution rule.

Subword: Wen
Superspace group : Janssen 1.4. n-dimensional space group with a point
group that leaves the physical space invariant.
Superspace : Janner 1, Janssen 1.4, Rivier. A vector space for a
quasiperiodic system with Fourier module of rank n, of dimension n
and the direct sum of the physical space and an additional (internal
or perpendicular) space. See : International Tables for
Crystallography, Vol. C, ed. A.J.C. Wilson, Kluwer, Dordrecht,
Section 9.8, p.797
Support: SOto, Queffelec
Surface gratings : Mikulic 5. Lateral grating structures grown or etched
on a surface of a superlattice
Systematic extinction: Janssen 1.5. Vanishing of the intensity of a
diffraction peak as a consequence of symmetry, e.g. a glide plane or a
screw axis. See: J.M. Cowley: Diffraction Physics, North-Holland,
Amsterdam

T-ergodic measure: Queffelec. See measure

T·invariant or T q invariant measure: Queffelec. See measure
T.mixing measure: Queffelec. See measure
Textures: Rivier
Thuring machine: Mendes France 3.3
Tight.binding : SOto
Tilings:
Canonical or Penrose-like: Katz 2.4
Oblique : Katz 2.4.2,
Octogonal or Ammann : Katz 2.4.3, Kalugin,
Topological entanglement: Rivier
Topological entropy: Mendes France 5, Berthe. See Walters P., Ergodic
theory - Introductory lectures, Springer. Lecture Notes, Math. 458
(1975). The topological entropy of a sequence of complexity p(n) is
equal to lim log p(n)
n
n=+oo
Topological singularity: Rivier. See Volovik G.E. and Mineev V.P.,
Zh.E.T.F.Pis'ma 23 (1976) 647, Mermin D., Rev. Mod.Phys.51
(1979) 591 and Michel L., Rev.Mod.Phys. 52 (1980) 617
Topological stability: Rivier
Topological theory of defects: Rivier
Trace map: Wen, Peyriere Trace Maps, SOto
Transfer matrices: Mendes France 7.2, SOto, Kunz
Transient : Dekking 1, See page 9 of B.D. Hughes and S. Prager. Random
612

processes and random systems : an introduction and The

Mathematics and Physics of Disordered Media. Eds. B.D. Hughes
and B.W. Ninham, SLN Math. 1035, Springer, Berlin, 1983, p.l - 86
Triangle patterns: Kramer. A non-periodic tiling of the plane with 5-fold
point symmetry
Two-sided sequences: Queffelec. A two-sided sequence: (an) is a
neZ
sequence which is indexed by Z .

Ultimately periodic: Mendlls France, Bertha 6.1, Sl1to. Ultimately

periodic sequence, a sequence which is periodic from some index on
Ultrametric: Kunz. A distance in which all triangles are isoceles.
Unique ergodicity: Berthe, Sl1to. See Walters P., Ergodic theory -
Introductory lectures, Springer. Lecture Notes, Math. 458 (1975)

Variety: Peyrillre Trace Maps

VQ: Patera

Weakly periodic: Janssen 1.7 definition

Weierstrass curve: Mendlls France. See B. B. Mandelbrot, The Fractal
Geometry of Nature; W. H. Freeman, publ. 1982
Weyl group: Patera
Weyl sequence: Sl1to
Wiener spectrum: Mendlls France 4.3
Wiener's criterion: Queffelec.

A measure (1 on T is continuous ~
lim ~+1
N-+-
L
Inl sN
I :(n)2 =0

Words: Kramer Mendlls France 1.1, Wen, Peyrillre Trace Maps

Worms: Katz
Wyckoff positions: Janner 6. This crystallographic concept is discussed by
H. Wondratchek in the International Tables for Crystallography,
Vol. A. Space Group Symmetry, Th. Hahn Ed., Kluwer, 1992, pp.
724-726.

Z-module: Janner 1, Janssen 1.1, Steurer. It is a free Abelian group. The

rank of the Z-module is the number of free generators. Considered
is the case where these elements are vectors. The dimension is then
that of the vector space spanned by the Z-module.
 Extra references

The following references have been collected by J.P. Allouche and M. Mendes
France. Quite a few have already been quoted in one or the other of the Courses.
They are meant to help the reader get in touch with the literature of the field.
The list is of course not complete.

References

- Ali M. K., Gumbs G., Quasiperiodic dynamics for a generalized third

order Fibonacci series, Phys. Rev. B (Brief Reports) 39 (1988) 7091-
7093.
- Allouche J.-P., Finite automata in I-D and 2-D Physics, in: Number
Theory and Physics, Proceedings of the Winter School, Les Houches,
France, March 7-16 1989, Luck J.-M., Moussa P., Waldschmidt M. Eds,
Springer Proceedings in Physics 47 (Springer Verlag, 1990) p. 177-184.
- Allouche J.-P., Complexity ofinfinite sequences and the Ising transducer,
in: Cellular Automata and Cooperative Systems, Boccara N., Goles E.,
Martinez S., Picco P. Eds, NATO ASI series, Series C: Mathematical
and Physical Sciences, vol. 396 (Kluwer Academic Publishers, Dordrecht,
1993) p. 1-9.
- Allouche J.-P., Mendes France M., Suite de Rudin-Shapiro et modele
d'Ising, Bull. Soc. math. Prance 113 (1985) 273-283.
- Allouche J.-P., Mendes France M., Quasi-crystal Ising chain and au-
tomata theory, J. Stat. Phys. 42 (1986) 809-821.
- Allouche J.-P., Peyriere J., Sur une formule de recurrence sur les traces
de produits de matrices associes a certaines substitutions, C. R. Acad.
Sci. Paris, Serie II 302 (1986) 1135-1136.
- Allouche J.-P., Salon 0., Robinson tHings and 2-dimensional automata,
in: Quasicrystals, networks and molecules of fivefold symmetry, Hargittai
1. Ed (VCH Publishers Inc., 1990) p. 97-105.
- Aubry S., Godreche C., Incommensurate structure with no average lat-
tice: an example of a one-dimensional quasicrystal, (Gratias D., Michel
L. Eds), J. de Phys. Colloq. 47 C3 (1986) 187-196.
- Aubry S., Godreche C., Luck J.-M., A structure intermediate between
quasiperiodic and random, Europhys. Lett. 4 (1987) 639-.
614

- Aubry S., Godreche C., Luck J.-M., Scaling properties of a structure

intermediate between quasiperiodic and random, J. Stat. Phys. 51 (1988)
1033-1075.
- Aubry S., Godreche C., Vallet F., Incommensurate structure with no
average lattice: an example of a one-dimensional quasicrystal, J. Phys.
48 (1987) 327-334.
- Avishai Y., Berend D., Transmission through Fibonacci chain, Phys.
Rev. B 43 (1991) 6873-6880.
- Avishai Y., Berend D., Transmission through a Thue-Morse chain, Phys.
Rev. B 45 (1992) 2717-2724.
- Avishai Y., Berend D., Trace maps for arbitrary substitution sequences,
J. Phys. A: Math. Gen. 26 (1993) 2437-2443.
- Avishai Y., Berend D., Glaubman D., Minimum-dimension trace maps
for substitution sequences, Phys. Rev. Lett. 72 (1994) 1842-1845.
- Axel F., Allouche J.-P., KIernan M., Mendes France M., Peyriere, J.,
Vibrational modes in a one-dimensional "quasi alloy", the Morse case,
(Gratias D., Michel L. Eds), J. de Phys. Colloq. 47 C3 (1986) 181-187.
- Axel F., Allouche J.-P., Wen Z.-Y., On certain properties of high-reso-
lution X-ray diffraction spectra of finite-size generalized Rudin-Shapiro
multilayer heterostructures, J. Phys.: Condens. Matter 4 (1992) 8713-
8728.
- Axel F., Peyriere J., Spectrum and extended states in a harmonic chain
with controlled disorder: effects of the Thue-Morse symmetry, J. Stat.
Phys. 57 (1989) 1013-1047.
- Axel F., Terauchi H., High-resolution X-ray diffraction spectra of Thue-
Morse GaAs-AIAs heterostructures: towards a novel description of dis-
order, Phys. Rev. Lett. 66 (1991) 2223-2226.
- Baake M., Grimm V., Joseph D., Trace maps, invariants, and some of
their applications, Int. J. Mod. Phys. B 7 (1993) 1527-1550.
- Baake M., Grimm V., Pisani C., Partition function zeros for aperiodic
systems, J. Stat. Phys. (1995) to appear.
- Baake M., Joseph D., Kramer P., Periodic clustering in the spectrum of
quasiperiodic Kronig-Penney models, Phys. Lett. A 168 (1992) 199-208.
- Baake M., Roberts J. A. G., Symmetries and reversing symmetries of
trace maps, in: Proceedings 3rd International Wigner Symposium, Boyle
L. L., Solomon A. 1. Eds (Oxford, 1993).
- Bajema K., Merlin R., Raman scattering by acoustic phonons in Fi-
bonacci GaAs-AIAs superlattices, Phys. Rev. B 36 (1987) 4555-4557.
- Barache D., Luck J.-M., Electronic spectra of strongly modulated ape-
riodic sequences, Phys. Rev. B 49 (1994) 15004-15016.
- Bellissard J., Spectral properties of Schrodinger's operator with a Thue-
Morse potential, in: Number theory and physics, Luck J.-M., Moussa
P., Waldschmidt M. Eds, Springer Proceedings in Physics 47 (Springer,
Berlin, Heidelberg, New York, 1990) p. 140-150.
 EXTRA REFERENCES 615

- Bellissard J., Gap labelling theorems for Schr6dinger's operators, in:

From number theory to physics, Waldschmidt M., Moussa P., Luck J.-
M., Itzykson C. Eds (Springer, Berlin, Heidelberg, New York, 1992) p.
538-630.
- Bellissiud J., Bovier A., Ghez J.-M., Spectral properties of a tight bind-
ing hamiltonian with period doubling potential, Comm. Math. Phys. 135
(1991) 379-399.
- Bellissard J., Bovier A., Ghez J.-M., Gap labelling theorems for one-
dimensional discrete Schr6dinger operators, Rev. Math. Phys. 4 (1992)
1-37.
- Bellissard J., Bovier A., Ghez J.-M., Discrete Schr6dinger operators with
potentials generated by substitutions, in: Differential equations with ap-
plications to mathematical physics, Ames W. F., Harrell E. M., Herod
J. V. Eds (Academic Press, Boston, 1993) p. 13-23.
- Bellissard J., Iochum B., Scoppola E., Testard D., Spectral properties of
one-dimensional quasi-crystals, Comm. Math. Phys. 125 (1989) 527-543.
- Benza V. G., Kohli M., Ali M. K., Phase transition in the generalized Fi-
bonacci quantum Ising models, Phys. Rev. B (Rapid Comm.) 41 (1990)
9578-.
- Bovier A., Ghez J.-M., Schr6dinger operators with substitution poten-
tials, in: Cellular Automata and Cooperative Systems, Boccara N., Goles
E., Martinez S., Picco P. Eds, NATO ASI Series, Series C: Mathematical
and Physical Sciences, vol. 396 (Kluwer Academic Publishers, Dordrecht,
1993) p. 67-83.
- Bovier A., Ghez J.-M., Spectral properties of one-dimensional Schr6din-
ger operators with potentials generated by substitutions, Comm. Math.
Phys. 158 (1993) 45-66 and erratum, Comm. Math. Phys. to appear.
- Bovier A., Ghez J.-M., Remarks on the spectral properties of tight bind-
ing and Kronig-Penney models with substitution sequences, J. Phys. A:
Math. Gen. (1995) to appear.
- Casdagli M., Symbolic dynamics for the renormalization map of a quasi-
periodic Schr6dinger equation, Comm. Math. Phys. 107 (1986) 295-318.
- Cheng Z., Savit R., Merlin R., Structure and electronic properties of
Thue-Morse lattices, Phys. Rev. B 37 (1988) 4375-4382.
- Combescure M., Recurrent versus diffusive dynamics for a kicked quan-
tum sytem, J. Stat. Phys. 62 (1991) 779-79l.
- Combes cure M., Recurrent versus diffusive dynamics for a kicked quan-
tum oscillator, Ann. Inst. Henri Poincare 57 (1992) 67-87.
- Doria M. M., Nori F., Satija 1. 1., Thue-Morse quantum Ising model,
Phys. Rev. B 39 (1989) 6802--6806.
- Dulea M., Physical properties of one-dimensional deterministic aperiodic
systems (Thesis, Link6ping, 1992).
- Dulea M., Johansson M., Riklund R., Unusual scaling of the spectrum in
a deterministic aperiodic tight-binding model, Phys. Rev. B 47 (1993)
8547-8551.
616

- Ghosh P. K., The Kronig-Penney model on a generalized Fibonacci lat-

tice, Phys. Lett. A 161 (1991) 153-157.
- Godreche C., Beyond quasiperiodicity: scaling properties of a Fourier
spectrum, in: Universalities in condensed matter, Jullien R. et al. Eds,
Springer Proceedings in Physics 32 (Springer Verlag, 1988).
- Godreche C., The sphinx: a limit-periodic tiling of the plane, J. Phys.
A: Math. Gen. 22 (1989) 1163-1166.
- Godreche C., Types of order and diffraction spectra for tilings of the line,
in: Number theory and physics, Luck J.-M., Moussa P., Waldschmidt M.
Eds, Springer Proceedings in Physics 47 (Springer, Berlin, Heidelberg,
New York, 1990) p. 86-99.
- Godreche C., Non-Pisot tilings and singular scattering, Phase Transi-
tions 32 (1991) 45-.
- Godreche C., Lan~on F., A simple example of a non-Pisot tiling with
fivefold symmetry, J. Physique 2 (1992) 207-220.
- Godreche C., Luck J.-M., Tilings of the plane and their diffraction spec-
tra, Proceedings of the Anniversary Adriatico Research Conference on
Quasicrystals, Trieste, Italy, July 4-7 1989, Jaric M., Lundqvist S. Eds
(World Scientific, 1989).
- Godreche C., Luck J.-M., Quasiperiodicity and randomness in tilings of
the plane, J. Stat. Phys. 55 (1989) 1-:-28.
- Godreche C., Luck J.-M., Multifractal analysis in reciprocal space and
the nature of the Fourier transform of self-similar structures, J. Phys. A:
Math. Gen. 23 (1990) 3769-3797.
- Godreche C., Luck J.-M., Indexing the diffraction spectrum of a non-
Pisot self-similar structure, Phys. Rev. B 45 (1992) 176-.
- Godreche C., Luck J.-M., Janner A., Janssen T., Fractal atomic surfaces
of self-similar quasiperiodic tHings of the plane, J. Physique 13 (1993)
1921-.
- Godreche C., Luck J.-M., Orland H., Magnetic phase structure on the
Penrose lattice, .J. Stat. Phys. 45 (1986) 777-800.
- Godreche C., Luck J.-M., Vallet F., Quasiperiodicity and types of order:
a study in one dimension, J. Phys. A: Math. Gen. 20 (1987) 4483--4499.
- Godreche C., Oguey C., Construction of average lattices for quasiperiodic
structure by the section method, J. Physique 51 (1990) 21-37.
- Godreche C., Orland H., Renormalisation on the Penrose lattice, (Gra-
tias D., Michel L. Eds), J. de Phys. Colloq. 47 C3 (1986) 197-203.
- Grimm U., Baake M., Non-periodic Ising quantum chains and conformal
invariance, J. Stat. Phys. 74 (1994) 1233--1245.
- Gumbs G., Ali M. K., Dynamical maps, Cantor spectra and localization
for Fibonacci and related quasiperiodic lattices, Phys. Rev. Lett. 60
(1988) 1081-1084.
- Gumbs G., Ali M. K., Scaling and eigenstates for a class of one-dimensio-
nal quasiperiodic lattices, J. Phys. A: Math. Gen. 21 (1988) 517-521.
 EXTRA REFERENCES 617

- Gumbs G., Ali M. K, Electronic properties of the tight-binding Fi-

bonacci hamiltonian, J. Phys. A: Math. Gen. 22 (1989) 951-970.
- Hof A., Quasi-crystals, aperiodicity and lattice systems (Thesis, Gronin-
gen, 1992).
- Hof A., Knill 0., Simon B., Singular continuous spectrum for palindromic
Schrodinger operators, Preprint Caltech (1994).
- Holzer M., Three classes of one-dimensional, two-tile Penrose tilings and
the Fibonacci Kronig-Penney model as a generic case, Phys. Rev. B 38
(1988) 1709-1720.
- Hornquist M., Johansson M., Singular continuous electron spectrum for
a class of circle sequences, Preprint Linkoping University (1994) and J.
Phys. A: Math. Gen. to appear.
- Hu P., Ting C. S., Electron transmission in a one-dimensional quasicrys-
tal, Phys. Rev. B 34 (1986) 8331-8334.
- Iochum B., Raymond L., Testard D., Resistance of one-dimensional qua-
sicrystals, Physica A 187 (1992) 353-368.
- Iochum B., Testard D., Power law growth for the resistance in the Fi-
bonacci model, J. Stat. Phys. 65 (1991) 715-723.
- Kalugin P. A., Kitaev A. Y., Levitov L. S., Electron spectrum of a one-
dimensional quasicrystal, Sov. Phys. JETP 64 (1986) 410-416.
- Kim Y.-J., Oh G. Y., Lee M. H., Electronic structure of a one-dimensio-
nal quasicrystal, J. Korean Phys. Soc. 22 (1989) 136-140.
- Kohmoto M., Kadanoff L. P., Tang C., Localization problem in one di-
mension: mapping and escape, Phys. Rev. Lett. 50 (1983) 1870-1872.
- Kohmoto M., Oono Y., Cantor spectrum for an almost periodic SchrO-
dinger equation and a dynamical map, Phys. Lett. 102 A (1984) 145-148.
- Kolar M., Ali M. K, Generalized Fibonacci superlattices dynamical trace
maps and magnetic excitations, Phys. Rev. B 39 (1989) 426-432.
- Kolar M., Ali M. K., Attractors of some volume non-preserving trace
maps, Phys. Rev. A 39 (1989) 6538-6543.
- Kolar M., Ali M. K, Attractors in quantum Ising models, Phys. Rev. B
40 (1989) 11083-11089.
- Kolar M., Ali M. K, One-dimensional generalized Fibonacci tHings,
Phys. Rev. B 41 (1990) 7108-.
- Kolar M., Ali M. K, Trace maps associated with general two-letter sub-
stitution rules, Phys. Rev. A 42 (1990) 7112-7124.
- Kolar M., Ali M. K, Nori F., Generalized Thue-Morse chains and their
physical properties, Phys. Rev. B 43 (1991) 1034-.
- Kolar M., Iochum B., Raymond L., Structure factor of ID systems (su-
perlattices) based on two-letter substitution rules. I. D (Bragg) peaks, J.
Phys. A: Math. Gen. 26 (1993) 7343-7366.
- Kolar M., Nori F., Trace maps of general substitutional sequences, Phys.
Rev. B 42 (1990) 1062-1065.
- Kolar M., SlitO A., The Kronig-Penney model on a Fibonacci lattice,
Phys. Lett. A 117 (1986) 203-209.
618

- Luck J.-M., Cantor spectra and scaling of gap widths in deterministic

aperiodic systems, Phys. Rev. B 39 (1989) 5834-5849.
- Luck J.-M., Critical behavior of the aperiodic quantum Ising chain in a
transverse magnetic field, J. Stat. Phys. 72 (1993) 417-458.
- Luck J.-M., Godreche C., Janner A., Janssen T., The nature of the
atomic surfaces of quasiperiodic self-similar structures, J. Phys. A: Math.
Gen. 26 (1993) 1951-1999.
- Luck J.-M., Orland H., Smilanski U., On the response of a two-level
quantum system to a class of time dependent quasiperiodic perturba-
tions, J. Stat. Phys. 53 (1988) 551-564.
- MacDonald A. H., Aers G. C., Continuum-model acoustic and electronic
properties for a Fibonacci superlattice, Phys. Rev. B 36 (1987) 9142-
9145.
- Macia E., Dominguez-Adame F., Exciton optical absorption in self-simi-
lar aperiodic lattices, Preprint Madrid University (1994).
- Merlin R, Bajema K, Raman scattering by acoustic phonons and struc-
tural properties of Fibonacci, Thue-Morse and random superlattices, J.
de Phys. Colloq. 48 C5 (1987) 503-506.
- Merlin R, Bajema K, Clarke R, Juang F.-Y., Bhattacharya P. K,
Quasi-periodic GaAs-AIAs heterostructures, Phys. Rev. Lett. 55 (1985)
1768-1770.
- Ostlund S., Pandit R, Rand D., Schellnhuber H. J., Siggia E. D., SchrO-
dinger equation with an almost periodic potential, Phys. Rev. Lett. 50
(1983) 1873-1876.
- Peyriere J., On the trace map for products of matrices associated with
substitutive sequences, J. Stat. Phys. 62 (1991) 411-414.
- Peyriere J., Wen Z.-Y., Wen Z.-X., Polynomes associes aux endomor-
phismes de groupes libres, Enseign. Math. 39 (1993) 153-175.
- Rivier N., A botanical quasicrystal, (Gratias D., Michel L. Eds) , J. de
Phys. Colloq. 47 C3 (1986) 299-309.
- Roberts J. A. G., Baake M., The dynamics oft race maps, in: Hamiltonian
mechanics: integrability and chaotic behaviour, Seimenis J. Ed, NATO
ASI Series, Series B: Physics (Plenum Press, New York, 1994) p. 275-
285.
- Roberts J. A. G., Baake M., Trace maps as 3D reversible dynamical
systems with an invariant, J. Stat. Phys. 74 (1994) 829-888.
- Shechtman D., Blech I., Gratias D., Cahn J. V., Metallic phase with
long-range orientational order and no translational symmetry, Phys. Rev.
Lett. 53 (1984) 1951-1953.
- Steinhardt P. J., Ostlund S., The physics of quasicrystals (World Scien-
tific, Singapore, Philadelphia, 1987).
- Sutherland B., Kohmoto M., Resistance of a one-dimensional quasicrys-
tal, Phys. Rev. B 36 (1987) 5877-5886.
- Siito A., The spectrum of a quasiperiodic Schrodinger operator, Comm.
Math. Phys. 111 (1987) 409-415.
 EXTRA REFERENCES 619

- Siito A., Singular continuous spectrum on a Cantor set of zero Lebesgue

measure for the Fibonacci hamiltonian, J. Stat. Phys. 56 (1989) 525-53l.
- Thomas U., Kramer P., The Fibonacci quasicrystal reconsidered: vari-
ety of energy spectra for continuous Schrodinger equations with simple
potentials, Int. J. Mod. Phys. B 3 (1989) 1205-1235.
- Todd J., Merlin R., Clarke R., Mohanty K. M., Axe J. D., Synchrotron
X-ray study of a Fibonacci superlattice, Phys. Rev. Lett. 57 (1986) 1157-
1160.
- Tracy C. A., Universality classes of some aperiodic Ising models, J. Phys.
A: Math. Gen. 21 (1988) 603-605.
- Wen Z.-X., Wen Z.-Y., On the leading term and the degree of the poly-
nomial trace mapping associated with a substitution, J. Stat. Phys. 75
(1994) 627-64l.
- Wiirtz D., Soerensen M. P., Schneider T., Quasiperiodic Kronig-Penney
model on a Fibonacci superlattice, Helv. Phys. Acta 61 (1988) 345-362.