Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
QUASICRYSTALS
Editors
Fran~oiseAxel, Universire Paris VII-Denis Diderot
Denis Gratias, CECN-CNRS, Vitry
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.
To his wife, who was with him in Les Houches and to their young daughter,
we present the expression of our heartfelt sympathy.
Foreword
Quasicrystals
COURSE 1
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
COURSE 4
Non-commutative models for quasicrystals
by P. Kramer and J. Garcia-Escudero
COURSE 5
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
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
COURSE 14
Singular words, invertible substitutions and local isomorphisms
by Wen Zhi-Ying
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
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
Annexes
Current deterministic sequences......................................................... 585
Glossary......................................................................................... 595
Quasicrystals,
Diophantine approximation
and algebraic numbers
Yves Meyer
CEREMADE
Universite Paris-Dauphine
Place de Lattre de Tassigny
75775 Paris Cedex 16 - FRANCE
1. Introduction.
This was the problem. Then I discovered some remarkable patterns A for
4 Y. Meyer
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.
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.
Once for all, the Fourier transform of a function / E Ll(JRn ) will be defined
by
(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
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,
wES
This implies
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* .
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
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
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.
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).
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
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.
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.
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.
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 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
and
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
and (A *)* = A.
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.
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
(Y) = (21r)n A( )
(5.2) W vol D cp -Y , Y =P2(d'") , d'" E D*.
(5.3)
This rapid decay explains why only finitely many layers A:n are observed in
the diffraction pattern of a quasicrystal.
6. Conclusion.
References.
The pentacrystals
J. Patera
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
T + a = 1, T2 = 1 + T, a2 = 1 + a, aT = -1, (1)
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
(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
al a2 a3 a4 al a2 a3 ra4
~ - 7r1l
~
a7 a6 as as ral ra2 ra3 a4
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.
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 ,
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
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
4. DEFINITION OF QUASICRYSTALS
For any fixed convex polytope P E lllI we define a penta crystal as the set EP of
points x,
(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
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
5. PHASONS
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)
6. QUASIADDITION
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)
7. EXAMPLES
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
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
. ...... ..
\
.. ....
· ..... ......
....
... ....... . ..... .
. : .....: ....
•••
-........ /
•••••• y
-.!~~. ~!.--
.. .... .... ..
/ ..
"" ...... .
. .. .... .. . .. .
..............
.. ....
....... .... .. ..... .. .... ......
. . .. ....
.......
. .... ...... .. ... · ... ... ... ....
.. .. ... ..
.... ... ... ..
.. •• + ••
. .. .... .... .
.... .. * • *.
• •..
... ... •......
· ... *.
• • * ••••••• ••
• . ••
.. ... ... .
• •• • •
• + •
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
ABSTRACT
1. INTRODUCTION
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
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
(4)
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
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
where tilde denotes transposition. Indeed, on the orthonormal basis one has
4 -~ )
R(e) = ( 1
'2
'
L(e) = ( J32 J323) (11)
implying
R(a) =( ~ -1 )
o ' L(a) = ( ; ~). (14)
(15)
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 :
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)
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
(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)
The norm N orm( 7') and t.he t.race Trace( r) are defined by
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)
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.
with gik = g/d elements of the positive definite metric tensor 9 of the ternary
quadratic form left invariant by the automorph A
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
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):
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):
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:
(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
Order 2 3 4 6 00 00
Trace -1 0 1 2 3 n>3
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
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:
(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
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
(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.
so that disregarding the common factor 7:i t.he basis transformation is given
U
by
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
D, C D'
1. Euclideal1 :
(l
-1 1
Rz(a) = 0 mAa) = ~l 1 (63)
0 0
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
6
with hyperbolic (L., Ly), p(ar~~Oli~ (P a)l1d ellipt.ic of order t.wo (E 1 , E 2 )
3. Indefinite:
U
8
16 )
L",(a) = N:(a) = 17 32 (69)
9 17
(70)
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
Lx X 2 ""
( =-:4~
3
)
(72)
so that
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
°°
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)
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
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
[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
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.
(3)
(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.
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:
(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:
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.
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.
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
(13)
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
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
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.
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)
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.
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:
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 :
Then if we use the multiplication law given in [17], it can be proved by induction
on n that for k=1,2:
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)
All the words belonging to the language generated by these grammars represent
parts of the patterns.
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
T. Janssen
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].
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.
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.
-
1)001)
•
• t .~
-- • •
.
t
•
•·
--
t I'
•
•·
•• .1
•~ • ~ (!)~
•, :::=>
•
•• • •
•
.•
4,
.'
•
•
•
••
•
.1
Fig. 1. - Diffraction pattern for a quasiperiodic structure. The peaks are arranged
in lines. .
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
(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].
.•.
0 0 0 0 0 0 0 0
.. .
. ~., ~ - ~- .~ ~ ~ ~ ~ .
~ ~
. ..
.. ~ . ~ .; ...$ . ~ ...
. . . . .
... '.'. . 't,
·· . ~.... ~... , ~.- .. ~ "' ~"" 0 0 0 0 0 0 0 0 0 0
.. . ~ .. :.... . ®.. 0 0 0 0
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
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.
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.
~. = Frac[i.ii;]. (1.13)
ii + !cos(27rif.ii). (1.16)
iE = ii + !cos(27r(if.ii+t)), (1.17)
,, ,,
,, ,, ,
,, ,,
,, ,,
,, ,, ,, ,
,, ,, ,,
, ,, ,,
,, , ,,
,, ,, ,
,, ,, ,,
,, ,
,
,, ,,
,, ,, ,, ,
,, ,, ,,
, ,, ,
,, , ,,
,, ,, ,,
,, ,,
,, ,,
, ,, ,,
,, , ,,
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.
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
(1.18)
(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.
(1.22)
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).
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.
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).
{(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.
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
m + nr ---+ n + (m+n)r
(1.31)
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)
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.
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.
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.
(1.35)
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.
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].
2. DIFFRACTION
(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.
called respectively the point, absolute continuous and singular continuous part.
The integrated structure factor is the function
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 )
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.
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.
~ = 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.
(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.
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
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.
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.
Le.:.-·
~...... . ...:..•.. .-.
"11111"
50 100 150 10
X Q-X
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
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
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.
J
tJ
L0]
I.JJ ~
7rr,]
dB sin B27rr;J exp( iH r '] cos B) (2.55)
1.2
"fort.15" - -
1
0.8
0.6
0.4
0.2
o
o 5 10 15 20 25 30 35 40
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
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.
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)
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).
(3.63)
n
(3.66)
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:
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.
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
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
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-
(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)
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
where the sum runs over 1st, 2nd, ... , n-th neighbour pairs. The coupling
constant Q mn depends on the separation of the neighbours.
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
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
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.
~~----------~----------~~--------------~
r x M r
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
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.
-,
:0-0
o
:-~
o
o
I
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
(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.
o
o
•
''""
"'.
'"
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.
4. SUBSTITUTIONAL CHAINS
4.1. Introduction
( 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.
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.
(4.92)
For each finite piece of the chain one has a maximum and a minimum of the
modulation.
If A is the second eigenvalue one has the following properties for this variation
function.
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
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).
Each subset Pi of the path P is projected on VI. One has for the atomic surface
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
---------',-
____ •. ___ ._____ I
--------r--------------
,
________
,
------- - - - - - ________ L ___________ _
, - - - -"T- - - - ----
-- -- -
,, _ _ _ .L. _ _ _ _
, - '"""T"'---
- --'-- ,,
-.,-
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
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)
oo 1)
0 ,
1 1
which has two equivalent substitutions. The substitutions have the Pisot prop-
erty, because the eigenvalues are
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
0. 8r---~--~---.---r---'----r---~--~---r--~
" fo rt.19 " -
311132
0.6
0. '
0.2
o ..... .
- 0.2
- 0.4
-0.6
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 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
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)
T(l)
= ( ~-El o
-1 ) = A,
T(2) AB with B = ( ~- E2
-1
0 ),
T(3) ABA.
FROM QUASIPERIODIC TO MORE COMPLEX SYSTEMS 135
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.
(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].
(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.
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
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
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.
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.
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.
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.
3.1. Self-similarity
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 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
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
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.
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
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.
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
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.
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.
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).
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.
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
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.
For this, we first need more information about the position of the forbidden
planes with respect to the lattice, and about their intersections.
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.
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.
==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
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- -
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
/
'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.
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
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
7.1. Definitions
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.
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.
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.
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
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'.
__ ~~ __ ~ ________________ ~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.
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.
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°.
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
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:
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
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
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
@@
@ \(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.)
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
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
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
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
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:
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] .
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).
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.
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].
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
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].
2.5,--------------,. 2.5-r-- - - - -- - - - . . , .
H HJ.
2.0
..: .. ", .
:.
...
, ','
1.5
', ,' : , ...... .
', _ -,
. -,. - ":
" , ,'
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.
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.
LS L
l
.1, J 1 J ,I.
(a)
J ,I,
r
r
~
r
-
-
l $L -
-
J
•
(b )
; r
l'S l
(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
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'
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]
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
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;
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).
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
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.
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
(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
[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
1. Introduction
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.
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.
(1)
(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.
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
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.
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)
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
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
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
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
SB(k) (12)
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.
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).
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
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
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
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.
(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.
(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.
-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
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
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
>.
.;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
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.
5. Gratings
function of the wires Pd(X) =Ln b(x - xn) . The density of the grating is then
(35)
(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
Acknowledgments
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
Nicolas Rivier
1. INTRODUCTION
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.
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
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.
1 A single picture to launch two Nobel prizes (N6el's and Shull's), quoted by a third [63].
252 N. Rivier
- 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.
n=d-dd-l (1)
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
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
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
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.
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
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).
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
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
... -+ 1tn+ 1(O/H) -+ 1tn(H) -+ 1tn(O) -+ 1tn(O/H) -+ 1tn-l (H) -+ ... (2)
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
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)
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 ).
··.........
....
....... ...............
.'... ......................
. ........... . .. .....
............
............. '.
'
·. ......... ..............
~
~ ' ,
'.'
' ' '.
:.::............... 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·:
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
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.
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.
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.
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
,, ,
X,
,, ,
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
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.
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.
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.
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.
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.
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
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.
("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).
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.
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.
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.
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
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.
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.
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.
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)
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
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
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.
(8)
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.
(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.
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
(10)
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
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.
5. CONCLUSIONS
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.
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
[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
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.
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:
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 .. ·
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.
o
01
0110
01101001
0110100110010110
- 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
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
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:
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:
Hence, \In ~ 0:
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)
2K_l
- 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
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
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.
7. MISCELLANEA
• 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
1021201020···
2102012101···.
This is the sequence of lengths of the strings of 1's between two consecutive O's
in the Morse sequence
011010011001011010010
1 3 4 5 7 9 11 12 13 15 16 17 19
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:
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
Ocw. 1l~o
1) A reverse automaton generating the Prouhet-Thue-Morse sequence:
1 o
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?
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
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.
Consider the two automata A' and A" where A' (resp. A") is the initial
state.
1 o
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.
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
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.
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_ )
• •
L r--_ _---,L
L-.._ _ =R L
L R
so that
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).
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
2c:
1 ~ dim (r) ~ 2.
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
so that
dim(r) ~ lim log(2eN) - 2
N -00 log a.,fN - .
320 J.-P. Allouche and M. Mendes France
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.
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
then
AUTOMATIC SEQUENCES 321
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=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
References
[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.
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:
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].
Going back to the Penrose tiling, remember that the underlying morphism
is the Fibonacci morphism:
0-01, 1-0.
2.1. An eXaIllple
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
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
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].
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
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}
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
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
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
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
f(n) '" o.
f(n) '" O.
f(n) '" O.
2a+ 1
if \ - 2v +2 '
,,- ">
v_ 0, 0 ~ a < 2 +
v 1
if not
and
L
00
f(n - 1) '"
v=o
( N-1 )1/2
11111 = l~..:~p ~ ~ I/(nW
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)
In other words
We then write
L uk(n).
00
s(n) ~
k=O
~ I<A)exp2i1rAn ~ I(n)
)..ES(J)
if and only if
~ 11(AW = 11/112.
)..ES(J)
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
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
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])
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)
pseudo-random: e 2i ,..an'
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
limsuplrl(n)1 > 0,
n-oo
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)
!
m<~
L f(2m)f(2m + 2k + 1)
m<!:l.
8(N,2k+1)= ~
+ m~~ J(2m + 1)f(2m + 2k + 2) + 0(1)
Hence
(1) 8(N,2k+1)=-8(~,k)-8(~,k+1)+0(1).
Similarly
Then
-y*(1) = -~ - ~-Y*(1),
{
-y*(1) = -~ - ~-Y*(1).
We now assume -y(k) exists for all k < 2K. Then equation (2) implies that
-y(2K) exists and
-y(2K) = -y(K).
1
-y(2K + 1) = -2(-y(K) + -y(K + 1)).
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
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)
limsup!,(n)! ~ ~.
n--+oo 3
and
'l(n) = 10 1 e 2imx d(11(X),
The following equalities hold
= 12 e2i7rntd(1 (~)
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
(5)
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).
f(2n) = (-It
{
f(2n + 1) = f(n).
ip..) = lim
N-oo
~
N
L: f(n)exp(-2i7rn~).
n<N
Define
S(N,~) = L: f(n) exp( -2i7rn~)
n<N
+e- 2i
m</f
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
k-1 1
S(N, A) =N L 2i+1 X(2i A)e«l - 2i)A) + O(log N).
j=O
j(A) = f X2~2~;)
i=O
e«l - 2i)A).
X(2i2a+l) =X(2a+.l)
2l 2l -)
= {I if j=f-2,
0 if not.
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
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 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
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,
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
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
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
automatic sequence. The simple fact that this last sequence takes only finitely
many values is by no means trivial.
References
[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
The inf is to be taken over all maps cp: S ->~. The opacity of A is by
definition
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
(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
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
+~ +~-
Consider for example the input sequence
e: +--++
The automaton sends it on
= 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.
+()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
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
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
References
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].
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
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
lim PT(a)
T/+oo
=2 :+1'
All configurations have equal probability: the system IS III a state of total
disorder.
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
Z-«(J,N) = L exp(-(J1ig(a)).
O'E{±}N
O'N=-l
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
Then
V(N) = M(€N_dV(N - 1),
where the "transfer matrix" is defined by
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
Hence the sign of ON (notice the subscript!) imposes the value of the spin UN
at equilibrium:
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.
(2)
As an easy exercise, the reader may verify that for a = 0, (H = 0), these
equations reduce to
8n E {-2, +2}
and if a> 0
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,
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
1 N-1 1
liminf N
N->oo
2::)A.€)n ~ -20:·
n=O
Proof. Recall
and
Hence
QED.
1 N-1
liminfN "(Aa€)n = {3,
N->oo L..J
n=O
N-1
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
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.
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
Department of Mathematics
Universite des Sciences et Technologies de Lille
F-59655 Villeneuve d'Asq Cedex, France
INTRODUCTION
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.
L t]tk8-(j - k)
l~],k~m
is nothing but
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]
N
lim
N--+oo 2 + 1 L 18-(n)12 = 0
N1
-N
1 N
lim 2N ~18-(nW = ~(a{x})2
N--+oo +1 ~ ~
-N xET
372 M. Queffelec
(T = (Te + (Tpp
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].
lim &(n) = o.
Inl ...... oo
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.
whose expansion is
PN(t) = ~ PN(n)e tnt
Inl<3 N
Every n admits a unique decomposition in base 3.
PN(n) = II (~ )1<,1 if n =L cJ 3J
J<N J<N
p(O)=1
p(n) = II(~ )1<,1 if n =L cJ 3J
p = N-+oo
lim PN · m
with
AUTOMATIC AND SUBSTITUTIVE SEQUENCES 375
II(1+cos2nt)
n=O n=O
e,Z"tjZ =
n=O n=O
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
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
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
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
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
u{qn) = u{n) n E 7L
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
We are looking for ergodic properties which induce purity laws (or dichotomy)
for those measures f.L enjoying them: for example
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
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:
au =weak- lim
N->oo
~I
N ~
'" u n etnt l 2 .m
n<N
3.2. Classification
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.
3.2.5. 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
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]
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
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.
4.2.4. 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.
v = h(u) where h: A -+ CD
4.3. Examples
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
x = closure(Tnu, n ~ 0)
Obviously, given c: > 0, d(x,y) < c: if and only if X[O,n] = Y[O,n] for some
n=n(c:). Now
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
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
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
• Fibonacci on {a, b}
((a) = ab
((b) = a
M(()=(~ ~) and e = T = 1 + J5
2
AUTOMATIC AND SUBSTITUTIVE SEQUENCES 387
((a) = ab
((b) = aa
(2(W) is by definition the following word, of length Iwol, whose letters belong
to [22:
• 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
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
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)
O-a,b(k) = N--+oo
lim N1 #{n < N, unH = a, Un = b}
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.
Uf = f T = e't f
0
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)
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
¢(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].
1(a)1 = q (a E A)
6.1. Examples
(a) = ab
(b) = ba
(a) = ab
(b) = aa
(a) = ae
(b) = de
(e) = ab
(d) = db
392 M. Queffelec
(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 -+ ab, b -+ aa
b
- - ab
aa
abaa
abab
- abaaabab
abaaabaa
- ...
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
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
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
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
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
(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]
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]
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}
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].
(1((0)1,1((1)1,"',1((8 - 1)1)
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.
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.
a ---- ababa
b ---- abb
M = (~ ~ ), T = 5, d = 4, (j = 4, (j' =1.
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)
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
But the order of letters is relevant for the description of the continuous part:
~ 0~ ~ D
while
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
(p, q)M = (1, I)M2 = (1, I)(T M - dl) by the Cayley-Hamilton equation
= T(p, q) - d(l, 1)
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
f(u) = 0, f(Tnu) = tn
f(Tx) =f(x) + 0: if Xo = 0
f(x) + 0: - 1 if Xo = 1
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)
~
P(X) = X2 - (n + 2)X + 1 = 0
satisfy
(J> 1 > (J' > 0
(0) = 001
(1) = 10
AUTOMATIC AND SUBSTITUTIVE SEQUENCES 407
7.3. Miscellaneous
((1) = 00
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.
tends to cp(a)cp(b)cp(c) = 1
and e d = 1.
1 1 0 0
1 0 1
o
1 0 1
1 0 o
2---+1. .. 13
'-v-'
a2
m-l---+1. .. 1m
'-v-'
am-l
m ---+ 1. .. 1
'-v-'
7.4. Conclusion
8. References
[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
1. INTRODUCTION
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
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
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
If u is the infinite sequence which is the fixed point of (J" starting with a, where
(J" is the substitution in (2), then
heal = 1, h(b) = -1
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
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.,
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
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
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
1«
_ vI n) ... v (n)
k
< 2" ,
_ _2n ~ w~n) ... w~n) ~ 0 for all 1 ~ k ~ 4n ,
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 (~ ~)).
A = {SI! ( u) : n ~ O}
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
,d ::::
• llr'tlr
!(a.,.,k) :::: f(a .. k) :::: f(b., f(vA:) :::: e"TO k:::: 0, ... ,9.
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'( bo,o,o)
(J'(vo) V9 V I,
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
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
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
1 n)
-(R
71
= -1 71
J U,,(x)dIJ(X) -+ Il(W), a.nd
1
-Un -+ II(W) I" - almost surely.
71
(St) = f f(Xt)d/l(X) = O.
In case (S.. ) is associat.ed to a primit.ive substit.ut.ion (T we have
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.
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)
o
(II 5 n 112)! /
as Il -+ 00 (6)
P
n '" 00 /3<n
RANDOM AND AUTOMATIC WALKS 429
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.
2
(52") =:34 for all 11 (7)
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 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
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 + v2)VN+lVii.
It follows that. for any x E X t7 one has (uniform in J:)
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
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
1. INTRODUCTION
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.
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.
(2)
This immediately shows that Wn is a palindrome for all n (a word v = VI ••• Vk
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".
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
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
Then
a1l"( a) d(b) = a
a1l"(b) a(b-Ia) = a(b- a(0) = (a(b))-lab = a-lab = b.
1
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,
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
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]).
References
1. Introduction
2. Thermodynamical entropy
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
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:
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.
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
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
Theorem 1 We have:
2L (3
H(rd::; log(-C ) + - ( 3 - '
£ e-1
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.
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
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
·
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
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.
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.
It is easy to check that this limit exists because of the subadditivity of the
function log(p( n)) :
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.
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.
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
x, Tx,"') Tn-I(x).
pn =p V T- I P V··· V T-(n-I) P.
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.
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.
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:
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.
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]):
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.
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 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.
Therefore, in these two extreme cases, the sequence (Hn)nEIN gives a charac-
terization of the ultimately periodic and the "random" 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:
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.
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:
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].
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 :
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 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
[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.
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.
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
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
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.
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)
To prove assertion 2°, write [AI (A2A3) All A3A2 = AIA2 (A3AIA3) A2 and
use formula (3) twice. We obtain
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.
_,_
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 .
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.
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:
(3) Strictly speaking, elements of this group are not words, but equivalence classes
of words.
TRACE MAPS 471
3.4. Endomorphisms
a(aba- I ) = a(a)a(b)a(a)-I
= aba(ab)-l = abab-Ia- I
(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
In this section, the free group on two generators a and b will be denoted simply
by r.
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 .
(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)
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"..
00
<I>' = ( 0 0
yz - 2x xz - 2y
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.
Proof We have
- 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.
and
(as previously, X stands for the collection of x's, and similarly for Y).
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).
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.
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.
6. COMMENTS
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.
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].
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.
[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.
[8] M. KOLAR & M. K. ALI Trace maps associated with general two-letter
substitution rules. Phys. Rev. A 42 (1990), 7112-7124.
[16] J. PEYRIERE On the trace map Eor products oE matrices associated witb
substitutive sequences. J. Stat. Phys. 62 (1991), 411-414.
[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.
[21] C. TRAINA Trace polynomial Eor two generator subgroups of 8L(2. <c).
Proc. Amer. Math. Soc. 79 (1980), 369-372.
[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
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
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
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
(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
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)
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
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
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
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{)
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
where
(3.1)
By iterating,
and similarly,
'ljJt = 1 0 = 'ljJf
'ljJJ = 0 1 = 'ljJ5
T m--+n = A = (~ ~) , ad - be = 1 (3.3)
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
( c~sx -sinx)
smx cos x
(3.4)
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
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,
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
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,
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*
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.
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-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}.
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,
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 )
while with
we get
.T.o2
'i.'n = /\\-k(n)T1..... m(n) S ( 01 )
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
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
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.
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
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
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,
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.
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
°
(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,
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:
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
Llnxn >-00
n=O
For Cantor's 'middle thirds' set Xn = 2/3 and thus the Lebesgue measure is
zero.
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.
H = L EP({E})
EEO'pp(H}
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
(5.5)
P(0) = 0 , P(lR) = I
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:
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
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
H= J EdPE
This equation has the following meaning. For any 'ljJ E Ji and any function f,
continuous on a(H),
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
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])
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
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
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
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:
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
{L 3- x
00
C= n n : Xn = 0 or 2}
n=l
= 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
J e'txda(x) = e,t/2 IT
n=l
cos t/3 n
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:
where the two projections commute. Let 'Ij; E H cont . Then for any E E lR
Problems.
and
J.L1/> = J.L1/>PP + J.L1/>ac + J.L1/>.c
What about the decomposition of Pcont ? Because
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
Choose for U the Fourier transformation which maps £2(1.) onto L~er.[O, 1]: For
any 'IjJ E £2(1.)
L
00
4 - E2
-2
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)
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). Notice that a complex solution for real E is the linear
combination of two real solutions with complex coefficients. Let
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.)
(6.1)
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.
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
. 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
SE = {W: (E,w) E S}
SW = {E: (E,w) E S} (6.3)
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
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].
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 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
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.
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
(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
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).
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.
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
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
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.
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
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]).
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).
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.
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.
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
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
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
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
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
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,
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
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,
and so on.
5.8.4
(i) Let B be any Borel set, then
(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
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!).
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
[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
[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
1. INTRODUCTION
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.
m = ± 20rder(m) (2 I + 1) (3)
I being an integer.
1
These auth ors fiIn d anomLOUSI d ' fuSlon
f' I'f R < _1
2 (when the
hierarchical constraint is important)
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.
2. HIERARCmCAL POTENTIALS
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)
den, m) = In - m I s (12)
A well known construction of the real number out of the rational ones is
the following.
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
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
1 co
For example: -1 = 1-2 =.~ 2i
J=O
defined as
v(n) = U{ I n I s) (16)
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"
v(l 2N + k) =v(k)
(18)
k = 1, ... , 2N -1 VN~O, leZ
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
which suggest, as it is indeed the case, that they correspond to fixed points
of a renormalisation group transformation (RG).
Theorem
1) The spectrum of H, a(H), is a Cantor set, for all R
2) Ga,p labelJiDll
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).
k
N(e) = 2N (21)
4. LIAPUNOV EXPONENTS
The following limit exist and defines the Liapunov exponent 'Y
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
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.
we have (25)
Indeed
2 2
"'n+l = 'tn - 2 + R'tn ('tn - '"n- 1 + 2) n~l
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
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
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.
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)
(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.
3.1. Definitions
and
(3)
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
jEN jEN
Therefore
. 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
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)
{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)
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).
4.1. Definitions
(12)
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
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).
[tmr]n,
.=0
(18)
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
The argument when a > 7'(0) is similar, but one has to consider the sets
(22)
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)
But
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
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 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
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
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].
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
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
1
"pn(t) = -logE(exptY
n
n ),
Of course, to deal with the probability P (Yn :s: na) one has to consider non-
positive t's.
lim 'ljJn(t)
n-+oo
= 'IjJ(t)
exists and is not identically -00.
Consider the Legendre transform of 'IjJ
More on large deviations can be found in [13, 18, 22, 23, 84, 89].
Bibliography
[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.
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.
1 • Substitution:
(J (a --+ ab 1 1 )
l...b --+ a 1 0
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
lim ~ __ I+{5
F -'t- 2
p--+oo p
2 • Other definition:
''GeDe.l~
_l':--..:II'FiboDaCCl•
1 • Substitution:
o(a-+ am bn
b-+a
lim 1+1
p-+ p
00
= l (m + ..Jm + 4n) ='t(m, n) =largest eigenvalue of M
2
2 • Other definition
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
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
2 • Other definitions:
{ u 2n =1
u2n+l = 1 - un with Uo = 1
2.2 2-automaton
3· Property
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.
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
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")
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
2 • Other definitions
with Uo = 0
CURRENT DETERMINISTIC SEQUENCES 593
2.2 2-automaton
Rudin-Shapiro
1 - Substitution
a -+ ab
( b -+ ac 11 01
o 1
10
0
00
1
J
a c -+ db
d -+ dc
o 0 1 1
N-l
sup L un eiqn s CW
q n=O
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
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
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
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
•
lim
1 n 1-+-
1\ J
sen) = 0 oft 1\sen) = eint ds(t)
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)
~ _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
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
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
A measure (1 on T is continuous ~
lim ~+1
N-+-
L
Inl sN
I :(n)2 =0
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