Algorithms and Computation

in Mathematics • Volume 26
Editors
Arjeh M. Cohen Henri Cohen
David Eisenbud Michael F. Singer
Bernd Sturmfels

For further volumes:

http://www.springer.com/series/3339
Oleg Karpenkov

Geometry
of Continued
Fractions
Oleg Karpenkov
Dept. of Mathematical Sciences
University of Liverpool
Liverpool, UK

ISSN 1431-1550 Algorithms and Computation in Mathematics

ISBN 978-3-642-39367-9 ISBN 978-3-642-39368-6 (eBook)
DOI 10.1007/978-3-642-39368-6
Springer Heidelberg New York Dordrecht London

Library of Congress Control Number: 2013946250

Mathematics Subject Classification (2010): 11J70, 11H06, 11P21, 52C05, 52C07

© Springer-Verlag Berlin Heidelberg 2013

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of pub-
lication, neither the authors nor the editors nor the publisher can accept any legal responsibility for any
errors or omissions that may be made. The publisher makes no warranty, express or implied, with respect
to the material contained herein.

Printed on acid-free paper

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

Preface

Continued fractions appear in many different branches of mathematics: the theory of

Diophantine approximations, algebraic number theory, coding theory, toric geome-
try, dynamical systems, ergodic theory, topology, etc. One of the metamathematical
explanations of this phenomenon is based on an interesting structure of the set of
real numbers endowed with two operations: addition a + b and inversion 1/b. This
structure appeared for the first time in the Euclidean algorithm, which was known
several thousand years ago. Similarly to the structures of fields and rings (with op-
erations of addition a + b and multiplication a ∗ b), structures with addition and
inversion can be found in many branches of mathematics. That is the reason why
continued fractions can be encountered far away from number theory. In particular,
continued fractions have a geometric interpretation in terms of integer geometry,
which we place as a cornerstone for this book.
The main goal of the first part of the book is to explore geometric ideas behind
regular continued fractions. On the one hand, we present geometrical interpretation
of classical theorems, such as the Gauss–Kuzmin theorem on the distribution of el-
ements of continued fractions, Lagrange’s theorem on the periodicity of continued
fractions, and the algorithm of Gaussian reduction. On the other hand, we present
some recent results related to toric geometry and the first steps of integer trigonom-
etry of lattices. The first part is rather elementary and will be interesting for both
students in mathematics and researchers. This part is a result of a series of lecture
courses at the Graz University of Technology (Austria). The material is appropriate
for master’s and doctoral students who already have basic knowledge of linear alge-
bra, algebraic number theory, and measure theory. Several chapters demand certain
experience in differential and algebraic geometry. Nevertheless, I believe that it is
possible for strong bachelor’s students as well to understand this material.
In the second part of the book we study an integer geometric generalization of
continued fractions to the multidimensional case. Such a generalization was first
considered by F. Klein in 1895. Later, this subject was almost completely abandoned
due to the computational complexity of the structure involved in the calculation
of the generalized continued fractions. The interest in Klein’s generalization was
revived by V.I. Arnold approximately one hundred years after its invention, when

v
vi Preface

computers became strong enough to overcome the computational complexity. After

a brief introduction to multidimensional integer geometry, we study essentially new
questions for the multidimensional cases and questions arising as extensions of the
classical ones (such as Lagrange’s theorem and Gauss–Kuzmin statistics). This part
is an exposition of recent results in this area. We emphasize that the majority of
examples and even certain statements of this part are on two-dimensional continued
fractions. The situation in higher dimensions is more technical and less studied,
and in many cases we formulate the corresponding problems and conjectures. The
second part is intended mostly for researchers in the fields of algebraic number
theory, Diophantine equations and approximations, and algebraic geometry. Several
chapters of this part can be added to a course for master’s or doctoral students.
Finally, I should mention many other interesting generalizations of continued
fractions, coming from algorithmic, dynamical, and approximation properties of
continued fractions. These generalizations are all distinct in higher dimensions. We
briefly describe the most famous of them in Chap. 23.
Liverpool, UK Oleg Karpenkov
February 2013
Acknowledgements

First, I would like to thank Vladimir Arnold, who introduced the subject of con-
tinued fractions to me and who provided me with all necessary remarks and dis-
cussions for many years. Second, I am grateful to many people who helped me with
remarks and corrections related to particular subjects discussed in this book. Among
them are F. Aicardi, T. Garrity, V. Goryunov, I. Pak, E.I. Pavlovskaya, C.M. Series,
M. Skopenkov, A.B. Sossinski, A.V. Ustinov, A.M. Vershik, and J. Wallner. Espe-
cially I would like to express my gratitude to Thomas Garrity and David Kramer
for exhaustively reading through the manuscript and giving some suggestions to
improve the book. Finally, I am grateful to my wife, Tanya, who encouraged and
inspired me during the years of working on this book.
The major part of this book was written at the Technische Universität Graz. The
work was completed at the University of Liverpool. I am grateful to the Technische
Universität Graz for hospitality and excellent working conditions. Work on this book
was supported by the Austrian Science Fund (FWF), grant M 1273-N18.

vii
Contents

Part I Regular Continued Fractions

1 Classical Notions and Definitions . . . . . . . . . . . . . . . . . . . . 3
1.1 Continued Fractions . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Definition of a Continued Fraction . . . . . . . . . . . . . 3
1.1.2 Regular Continued Fractions for Rational Numbers . . . . 4
1.1.3 Regular Continued Fractions and the Euclidean Algorithm . 5
1.1.4 Continued Fractions with Arbitrary Elements . . . . . . . . 6
1.2 Convergence of Infinite Regular Continued Fractions . . . . . . . . 6
1.3 Existence and Uniqueness of a Regular Continued Fraction for a
Given Real Number . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Monotone Behavior of Convergents . . . . . . . . . . . . . . . . . 12
1.5 Approximation Rates of Regular Continued Fractions . . . . . . . 14
1.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2 On Integer Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1 Basic Notions and Definitions . . . . . . . . . . . . . . . . . . . . 20
2.1.1 Objects and Congruence Relation of Integer Geometry . . . 20
2.1.2 Invariants of Integer Geometry . . . . . . . . . . . . . . . 20
2.1.3 Index of Sublattices . . . . . . . . . . . . . . . . . . . . . 21
2.1.4 Integer Length of Integer Segments . . . . . . . . . . . . . 22
2.1.5 Integer Distance to Integer Lines . . . . . . . . . . . . . . 23
2.1.6 Integer Area of Integer Triangles . . . . . . . . . . . . . . 24
2.1.7 Index of Rational Angles . . . . . . . . . . . . . . . . . . 24
2.2 Empty Triangles: Their Integer and Euclidean Areas . . . . . . . . 25
2.3 Integer Area of Polygons . . . . . . . . . . . . . . . . . . . . . . 26
2.4 Pick’s Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5 The Twelve-Point Theorem . . . . . . . . . . . . . . . . . . . . . 30
2.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3 Geometry of Regular Continued Fractions . . . . . . . . . . . . . . . 33
3.1 Classical Construction . . . . . . . . . . . . . . . . . . . . . . . . 33

ix
x Contents

3.2 Geometric Interpretation of the Elements of Continued Fractions . 37

3.3 Index of an Angle, Duality of Sails . . . . . . . . . . . . . . . . . 38
3.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4 Complete Invariant of Integer Angles . . . . . . . . . . . . . . . . . . 41
4.1 Integer Sines of Rational Angles . . . . . . . . . . . . . . . . . . 41
4.2 Sails for Arbitrary Angles and Their LLS Sequences . . . . . . . . 43
4.3 On Complete Invariants of Angles with Integer Vertex . . . . . . . 43
4.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5 Integer Trigonometry for Integer Angles . . . . . . . . . . . . . . . . 47
5.1 Definition of Trigonometric Functions . . . . . . . . . . . . . . . 47
5.2 Basic Properties of Integer Trigonometry . . . . . . . . . . . . . . 48
5.3 Transpose Integer Angles . . . . . . . . . . . . . . . . . . . . . . 49
5.4 Adjacent Integer Angles . . . . . . . . . . . . . . . . . . . . . . . 51
5.5 Right Integer Angles . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.6 Opposite Interior Angles . . . . . . . . . . . . . . . . . . . . . . . 55
5.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6 Integer Angles of Integer Triangles . . . . . . . . . . . . . . . . . . . 57
6.1 Integer Sine Formula . . . . . . . . . . . . . . . . . . . . . . . . . 57
6.2 On Integer Congruence Criteria for Triangles . . . . . . . . . . . . 58
6.3 On Sums of Angles in Triangles . . . . . . . . . . . . . . . . . . . 59
6.4 Angles and Segments of Integer Triangles . . . . . . . . . . . . . 61
6.5 Examples of Integer Triangles . . . . . . . . . . . . . . . . . . . . 62
6.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7 Continued Fractions and SL(2, Z) Conjugacy Classes. Elements of
Gauss’s Reduction Theory. Markov Spectrum . . . . . . . . . . . . . 67
7.1 Geometric Continued Fractions . . . . . . . . . . . . . . . . . . . 67
7.1.1 Definition of a Geometric Continued Fraction . . . . . . . 67
7.1.2 Geometric Continued Fractions of Real Spectrum
SL(2, R) Matrices . . . . . . . . . . . . . . . . . . . . . . 68
7.1.3 Duality of Sails . . . . . . . . . . . . . . . . . . . . . . . 69
7.1.4 LLS Sequences for Real Spectrum Matrices . . . . . . . . 69
7.1.5 Algebraic Sails . . . . . . . . . . . . . . . . . . . . . . . . 70
7.1.6 LLS Periods of Real Spectrum Matrices . . . . . . . . . . 70
7.2 Geometry of Gauss’s Reduction Theory . . . . . . . . . . . . . . . 71
7.2.1 Cases of Matrices with Complex, Real, and Coinciding
Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . 71
7.2.2 Reduced Matrices . . . . . . . . . . . . . . . . . . . . . . 72
7.2.3 Reduced Matrices and Integer Conjugacy Classes . . . . . 73
7.2.4 Complete Invariant of Integer Conjugacy Classes . . . . . . 73
7.2.5 Algebraicity of Matrices with Periodic LLS Sequences . . . 74
7.3 Some Technical Details and Open Questions Related to Gauss’s
Reduction Theory . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7.3.1 Proof of Theorem 7.14 . . . . . . . . . . . . . . . . . . . . 75
Contents xi

7.3.2 Calculation of Periods of the LLS Sequence . . . . . . . . 77

7.3.3 Complexity of Reduced Operators . . . . . . . . . . . . . 78
7.3.4 Frequencies of Reduced Matrices . . . . . . . . . . . . . . 79
7.4 Minima of Quadratic Forms and the Markov Spectrum . . . . . . . 81
7.4.1 Calculation of Minima of Quadratic Forms . . . . . . . . . 81
7.4.2 Some Properties of Markov Spectrum . . . . . . . . . . . . 82
7.4.3 Markov Numbers . . . . . . . . . . . . . . . . . . . . . . 83
7.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
8 Lagrange’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
8.1 The Dirichlet Group . . . . . . . . . . . . . . . . . . . . . . . . . 87
8.2 Construction of the Integer nth Root of a GL(2, Z) Matrix . . . . . 90
8.3 Pell’s Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
8.4 Periodic Continued Fractions and Quadratic Irrationalities . . . . . 93
8.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
9 Gauss–Kuzmin Statistics . . . . . . . . . . . . . . . . . . . . . . . . . 99
9.1 Some Information from Ergodic Theory . . . . . . . . . . . . . . 100
9.2 The Measure Space Related to Continued Fractions . . . . . . . . 101
9.2.1 Definition of the Measure Space Related to Continued
Fractions . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
9.2.2 Theorems on Density Points of Measurable Subsets . . . . 101
9.3 On the Gauss Map . . . . . . . . . . . . . . . . . . . . . . . . . . 102
9.3.1 The Gauss Map and Corresponding Invariant Measure . . . 102
9.3.2 An Example of an Invariant Set for the Gauss Map . . . . . 104
9.3.3 Ergodicity of the Gauss Map . . . . . . . . . . . . . . . . 105
9.4 Pointwise Gauss–Kuzmin Theorem . . . . . . . . . . . . . . . . . 107
9.5 Original Gauss–Kuzmin Theorem . . . . . . . . . . . . . . . . . . 108
9.6 Cross-Ratio in Projective Geometry . . . . . . . . . . . . . . . . . 108
9.6.1 Projective Linear Group . . . . . . . . . . . . . . . . . . . 109
9.6.2 Cross-Ratio, Infinitesimal Cross-Ratio . . . . . . . . . . . 109
9.7 Smooth Manifold of Geometric Continued Fractions . . . . . . . . 110
9.8 Möbius Measure on the Manifolds of Continued Fractions . . . . . 110
9.9 Explicit Formulas for the Möbius Form . . . . . . . . . . . . . . . 111
9.10 Relative Frequencies of Edges of One-Dimensional Continued
Fractions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
9.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
10 Geometric Aspects of Approximation . . . . . . . . . . . . . . . . . . 115
10.1 Two Types of Best Approximations of Rational Numbers . . . . . 115
10.1.1 Best Diophantine Approximations . . . . . . . . . . . . . 115
10.1.2 Strong Best Diophantine Approximations . . . . . . . . . . 120
10.2 Rational Approximations of Arrangements of Two Lines . . . . . . 122
10.2.1 Regular Angles and Related Markov–Davenport Forms . . 123
10.2.2 Integer Arrangements and Their Sizes . . . . . . . . . . . 124
10.2.3 Discrepancy Functional and Approximation Model . . . . . 124
xii Contents

10.2.4 Lagrange Estimates for the Case of Continued Fractions

with Bounded Elements . . . . . . . . . . . . . . . . . . . 125
10.2.5 Periodic Sails and Best Approximations in the Algebraic
Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
10.2.6 Finding Best Approximations of Line Arrangements . . . . 132
10.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
11 Geometry of Continued Fractions with Real Elements and Kepler’s
Second Law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
11.1 Continued Fractions with Integer Coefficients . . . . . . . . . . . 137
11.2 Continued Fractions with Real Coefficients . . . . . . . . . . . . . 139
11.2.1 Broken Lines Related to Sequences of Arbitrary Real
Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
11.2.2 Continued Fractions Related to Broken Lines . . . . . . . . 142
11.2.3 Geometry of Continued Fractions for Broken Lines . . . . 143
11.3 Areal and Angular Densities for Differentiable Curves . . . . . . . 146
11.3.1 Notions of Real and Angular Densities . . . . . . . . . . . 146
11.3.2 Curves and Broken Lines . . . . . . . . . . . . . . . . . . 148
11.3.3 Some Examples . . . . . . . . . . . . . . . . . . . . . . . 149
11.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
12 Extended Integer Angles and Their Summation . . . . . . . . . . . . 153
12.1 Extension of Integer Angles. Notion of Sums of Integer Angles . . 153
12.1.1 Extended Integer Angles and Revolution Number . . . . . 154
12.1.2 On Normal Forms of Extended Angles . . . . . . . . . . . 157
12.1.3 Trigonometry of Extended Angles. Associated Integer
Angles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
12.1.4 Opposite Extended Angles . . . . . . . . . . . . . . . . . 162
12.1.5 Sums of Extended Angles . . . . . . . . . . . . . . . . . . 162
12.1.6 Sums of Integer Angles . . . . . . . . . . . . . . . . . . . 164
12.2 Relations Between Extended and Integer Angles . . . . . . . . . . 164
12.3 Proof of Theorem 6.8(i) . . . . . . . . . . . . . . . . . . . . . . . 165
12.3.1 Two Preliminary Lemmas . . . . . . . . . . . . . . . . . . 166
12.3.2 Conclusion of the Proof of Theorem 6.8(i) . . . . . . . . . 170
12.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
13 Integer Angles of Polygons and Global Relations for Toric
Singularities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
13.1 Theorem on Angles of Integer Convex Polygons . . . . . . . . . . 173
13.2 Toric Surfaces and Their Singularities . . . . . . . . . . . . . . . . 175
13.2.1 Definition of Toric Surfaces . . . . . . . . . . . . . . . . . 175
13.2.2 Singularities of Toric Surfaces . . . . . . . . . . . . . . . . 176
13.3 Relations on Toric Singularities of Surfaces . . . . . . . . . . . . . 178
13.3.1 Toric Singularities of n-gons with Fixed Parameter n . . . . 178
13.3.2 Realizability of a Prescribed Set of Toric Singularities . . . 179
13.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
Contents xiii

Part II Multidimensional Continued Fractions

14 Basic Notions and Definitions of Multidimensional Integer Geometry 185
14.1 Basic Integer Invariants in Integer Geometry . . . . . . . . . . . . 185
14.1.1 Objects and the Congruence Relation . . . . . . . . . . . . 185
14.1.2 Integer Invariants and Indices of Sublattices . . . . . . . . 186
14.1.3 Integer Volume of Simplices . . . . . . . . . . . . . . . . . 187
14.1.4 Integer Angle Between Two Planes . . . . . . . . . . . . . 187
14.1.5 Integer Distance Between Two Disjoint Planes . . . . . . . 188
14.2 Integer and Euclidean Volumes of Basis Simplices . . . . . . . . . 189
14.3 Integer Volumes of Polyhedra . . . . . . . . . . . . . . . . . . . . 191
14.3.1 Interpretation of Integer Volumes of Simplices via
Euclidean Volumes . . . . . . . . . . . . . . . . . . . . . 192
14.3.2 Integer Volume of Polyhedra . . . . . . . . . . . . . . . . 192
14.3.3 Decomposition into Empty Simplices . . . . . . . . . . . . 193
14.4 Lattice Plücker Coordinates and Calculation of Integer Volumes
of Simplices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
14.4.1 Grassmann Algebra on Rn and k-Forms . . . . . . . . . . 194
14.4.2 Plücker Coordinates . . . . . . . . . . . . . . . . . . . . . 195
14.4.3 Oriented Lattices in Rn and Their Lattice Plücker
Embedding . . . . . . . . . . . . . . . . . . . . . . . . . . 196
14.4.4 Lattice Plücker Coordinates and Integer Volumes of
Simplices . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
14.5 Ehrhart Polynomials as Generalized Pick’s Formula . . . . . . . . 198
14.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
15 On Empty Simplices, Pyramids, Parallelepipeds . . . . . . . . . . . . 203
15.1 Classification of Empty Integer Tetrahedra . . . . . . . . . . . . . 203
15.2 Classification of Completely Empty Lattice Pyramids . . . . . . . 204
15.3 Two Open Problems Related to the Notion of Emptiness . . . . . . 206
15.3.1 Problem on Empty Simplices . . . . . . . . . . . . . . . . 206
15.3.2 Lonely Runner Conjecture . . . . . . . . . . . . . . . . . . 207
15.4 Proof of White’s Theorem and the Empty Tetrahedra Classification
Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
15.4.1 IDC-System . . . . . . . . . . . . . . . . . . . . . . . . . 208
15.4.2 A Lemma on Sections of an Integer Parallelepiped . . . . . 209
15.4.3 A Corollary on Integer Distances Between the Vertices
and the Opposite Faces of a Tetrahedron with Empty Faces 209
15.4.4 Lemma on One Integer Node . . . . . . . . . . . . . . . . 210
15.4.5 Proof of White’s Theorem . . . . . . . . . . . . . . . . . . 211
15.4.6 Deduction of Corollary 15.3 from White’s Theorem . . . . 213
15.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
16 Multidimensional Continued Fractions in the Sense of Klein . . . . . 215
16.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
16.2 Some Notation and Definitions . . . . . . . . . . . . . . . . . . . 216
16.2.1 A-Hulls and Their Boundaries . . . . . . . . . . . . . . . . 216
xiv Contents

16.2.2 Definition of Multidimensional Continued Fraction in the

Sense of Klein . . . . . . . . . . . . . . . . . . . . . . . . 217
16.2.3 Face Structure of Sails . . . . . . . . . . . . . . . . . . . . 217
16.3 Finite Continued Fractions . . . . . . . . . . . . . . . . . . . . . . 218
16.4 On a Generalized Kronecker’s Approximation Theorem . . . . . . 219
16.4.1 Addition of Sets in Rn . . . . . . . . . . . . . . . . . . . . 219
16.4.2 Integer Approximation Spaces and Affine Irrational
Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
16.4.3 Formulation of the Theorem . . . . . . . . . . . . . . . . . 221
16.4.4 Proof of the Multidimensional Kronecker’s Approximation
Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
16.5 Polyhedral Structure of Sails . . . . . . . . . . . . . . . . . . . . 224
16.5.1 The Intersection of the Closures of A-Hulls with Faces of
Corresponding Cones . . . . . . . . . . . . . . . . . . . . 224
16.5.2 Homeomorphic Types of Sails . . . . . . . . . . . . . . . . 226
16.5.3 Combinatorial Structure of Sails for Cones in General
Position . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
16.5.4 A-Hulls and Quasipolyhedra . . . . . . . . . . . . . . . . 231
16.6 Two-Dimensional Faces of Sails . . . . . . . . . . . . . . . . . . . 232
16.6.1 Faces with Integer Distance to the Origin Equal One . . . . 232
16.6.2 Faces with Integer Distance to the Origin Greater than One 234
16.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
17 Dirichlet Groups and Lattice Reduction . . . . . . . . . . . . . . . . 237
17.1 Orders, Units, and Dirichlet’s Unit Theorem . . . . . . . . . . . . 237
17.2 Dirichlet Groups and Groups of Units in Orders . . . . . . . . . . 238
17.2.1 Notion of a Dirichlet Group . . . . . . . . . . . . . . . . . 238
17.2.2 On Isomorphisms of Dirichlet Groups and Certain Groups
of Units . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
17.2.3 Dirichlet Groups Related to Orders That Do not Have
Complex Roots of Unity . . . . . . . . . . . . . . . . . . . 240
17.3 Calculation of a Basis of the Additive Group Γ (A) . . . . . . . . . 241
17.3.1 Step 1: Preliminary Statements . . . . . . . . . . . . . . . 241
17.3.2 Step 2: Application of the LLL-Algorithm . . . . . . . . . 242
17.3.3 Step 3: Calculation of an Integer Basis Having a Basis of
an Integer Sublattice . . . . . . . . . . . . . . . . . . . . . 242
17.4 Calculation of a Basis of the Positive Dirichlet Group Ξ+ (A) . . . 243
17.5 Lattice Reduction and the LLL-Algorithm . . . . . . . . . . . . . 243
17.5.1 Reduced Bases . . . . . . . . . . . . . . . . . . . . . . . . 244
17.5.2 The LLL-Algorithm . . . . . . . . . . . . . . . . . . . . . 245
17.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
18 Periodicity of Klein Polyhedra. Generalization of Lagrange’s
Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
18.1 Continued Fractions Associated to Matrices . . . . . . . . . . . . 249
18.2 Algebraic Periodic Multidimensional Continued Fractions . . . . . 250
Contents xv

18.3 Torus Decompositions of Periodic Sails in R3 . . . . . . . . . . . 251

18.4 Three Single Examples of Torus Decompositions in R3 . . . . . . 253
18.5 Examples of Infinite Series of Torus Decomposition . . . . . . . 257
18.6 Two-Dimensional Continued Fractions Associated to Transpose
Frobenius Normal Forms . . . . . . . . . . . . . . . . . . . . . . 261
18.7 Some Problems and Conjectures on Periodic Geometry of
Algebraic Sails . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
18.8 Generalized Lagrange’s Theorem . . . . . . . . . . . . . . . . . 265
18.9 Littlewood and Oppenheim Conjectures in the Framework of
Multidimensional Continued Fractions . . . . . . . . . . . . . . . 269
18.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
19 Multidimensional Gauss–Kuzmin Statistics . . . . . . . . . . . . . . 271
19.1 Möbius Measure on the Manifold of Continued Fractions . . . . . 271
19.1.1 Smooth Manifold of n-Dimensional Continued Fractions . 271
19.1.2 Möbius Measure on the Manifolds of Continued Fractions 272
19.2 Explicit Formulae for the Möbius Form . . . . . . . . . . . . . . 273
19.3 Relative Frequencies of Faces of Multidimensional Continued
Fractions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
19.4 Some Calculations of Frequencies for Faces in the
Two-Dimensional Case . . . . . . . . . . . . . . . . . . . . . . . 276
19.4.1 Some Hints for Computation of Approximate Values of
Relative Frequencies . . . . . . . . . . . . . . . . . . . . 276
19.4.2 Numeric Calculations of Relative Frequencies . . . . . . . 277
19.4.3 Two Particular Results on Relative Frequencies . . . . . . 279
19.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
20 On Construction of Multidimensional Continued Fractions . . . . . 281
20.1 Inductive Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 281
20.1.1 Some Background . . . . . . . . . . . . . . . . . . . . . 281
20.1.2 Description of the Algorithm . . . . . . . . . . . . . . . . 282
20.1.3 Step 1a: Construction of the First Hyperface . . . . . . . . 283
20.1.4 Step 1b, 4: How Decompose the Polytope into Its Faces . 284
20.1.5 Step 2: Construction of the Adjacent Hyperface . . . . . . 284
20.1.6 Step 2: Test of the Equivalence Class for the Hyperface
F  to Have Representatives in the Set of Hyperfaces D . . 285
20.2 Deductive Algorithms to Construct Sails . . . . . . . . . . . . . . 285
20.2.1 General Idea of Deductive Algorithms . . . . . . . . . . . 285
20.2.2 The First Deductive Algorithm . . . . . . . . . . . . . . . 286
20.2.3 The Second Deductive Algorithm . . . . . . . . . . . . . 287
20.2.4 Test of the Conjectures Produced in the Two-Dimensional
Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
20.2.5 On the Verification of a Conjecture for the
Multidimensional Case . . . . . . . . . . . . . . . . . . . 296
20.3 An Example of the Calculation of a Fundamental Domain . . . . 297
20.4 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
xvi Contents

21 Gauss Reduction in Higher Dimensions . . . . . . . . . . . . . . . . 301

21.1 Organization of This Chapter . . . . . . . . . . . . . . . . . . . . 301
21.2 Hessenberg Matrices and Conjugacy Classes . . . . . . . . . . . . 302
21.2.1 Notions and Definitions . . . . . . . . . . . . . . . . . . . 303
21.2.2 Construction of Perfect Hessenberg Matrices Conjugate to
a Given One . . . . . . . . . . . . . . . . . . . . . . . . . 305
21.2.3 Existence and Finiteness of ς -Reduced Hessenberg
Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
21.2.4 Families of Hessenberg Matrices with Given Hessenberg
Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
21.2.5 ς -Reduced Matrices in the 2-Dimensional Case . . . . . . 311
21.3 Complete Geometric Invariant of Conjugacy Classes . . . . . . . . 312
21.3.1 Continued Fractions in the Sense of Klein–Voronoi . . . . 312
21.3.2 Geometric Complete Invariants of Dirichlet Groups . . . . 316
21.3.3 Geometric Invariants of Conjugacy Classes . . . . . . . . . 317
21.4 Algorithmic Aspects of Reduction to ς -Reduced Matrices . . . . . 318
21.4.1 Markov–Davenport Characteristics . . . . . . . . . . . . . 318
21.4.2 Klein–Voronoi Continued Fractions and Minima of
MD-Characteristics . . . . . . . . . . . . . . . . . . . . . 321
21.4.3 Construction of ς -Reduced Matrices by Klein–Voronoi
Continued Fractions . . . . . . . . . . . . . . . . . . . . . 322
21.5 Diophantine Equations Related to the Markov–Davenport
Characteristic . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
21.5.1 Multidimensional w-Sails and w-Continued Fractions . . . 324
21.5.2 Solution of Eq. (21.1) . . . . . . . . . . . . . . . . . . . . 326
21.6 On Reduced Matrices in SL(3, Z) with Two Complex Conjugate
Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
21.6.1 Perfect Hessenberg Matrices of a Given Hessenberg Type . 327
21.6.2 Parabolic Structure of the Set of NRS-Matrices . . . . . . . 328
21.6.3 Theorem on Asymptotic Uniqueness of ς -Reduced
NRS-Matrices . . . . . . . . . . . . . . . . . . . . . . . . 329
21.6.4 Examples of NRS-Matrices for a Given Hessenberg Type . 331
21.6.5 Proof of Theorem 21.43 . . . . . . . . . . . . . . . . . . . 333
21.6.6 Proof of Theorem 21.48 . . . . . . . . . . . . . . . . . . . 336
21.7 Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
21.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
22 Approximation of Maximal Commutative Subgroups . . . . . . . . . 347
22.1 Rational Approximations of MCRS-Groups . . . . . . . . . . . . 347
22.1.1 Maximal Commutative Subgroups and Corresponding
Simplicial Cones . . . . . . . . . . . . . . . . . . . . . . . 348
22.1.2 Regular Subgroups and Markov–Davenport Forms . . . . . 349
22.1.3 Rational Subgroups and Their Sizes . . . . . . . . . . . . . 350
22.1.4 Discrepancy Functional . . . . . . . . . . . . . . . . . . . 351
22.1.5 Approximation Model . . . . . . . . . . . . . . . . . . . . 351
Contents xvii

22.1.6 Diophantine Approximation and MCRS-Group

Approximation . . . . . . . . . . . . . . . . . . . . . . . . 352
22.2 Simultaneous Approximation in R3 and MCRS-Group
Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
22.2.1 General Construction . . . . . . . . . . . . . . . . . . . . 353
22.2.2 A Ray of a Nonreal Spectrum Operator . . . . . . . . . . . 354
22.2.3 Two-Dimensional Golden Ratio . . . . . . . . . . . . . . . 355
22.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
23 Other Generalizations of Continued Fractions . . . . . . . . . . . . . 357
23.1 Relative Minima . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
23.1.1 Relative Minima and the Minkowski–Voronoi Complex . . 358
23.1.2 Minkowski–Voronoi Tessellations of the Plane . . . . . . . 360
23.1.3 Minkowski–Voronoi Continued Fractions in R3 . . . . . . 361
23.1.4 Combinatorial Properties of the Minkowski–Voronoi
Tessellation for Integer Sublattices . . . . . . . . . . . . . 362
23.2 Farey Addition, Farey Tessellation, Triangle Sequences . . . . . . 364
23.2.1 Farey Addition of Rational Numbers . . . . . . . . . . . . 364
23.2.2 Farey Tessellation . . . . . . . . . . . . . . . . . . . . . . 365
23.2.3 Descent Toward the Absolute . . . . . . . . . . . . . . . . 366
23.2.4 Triangle Sequences . . . . . . . . . . . . . . . . . . . . . 368
23.3 Decompositions of Coordinate Rectangular Bricks and O’Hara’s
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
23.3.1 Π -Congruence of Coordinate Rectangular Bricks . . . . . 374
23.3.2 Criterion of Π -Congruence of Coordinate Bricks . . . . . . 375
23.3.3 Geometric Version of O’Hara’s Algorithm for Partitions . . 375
23.4 Algorithmic Generalized Continued Fractions . . . . . . . . . . . 378
23.4.1 General Algorithmic Scheme . . . . . . . . . . . . . . . . 378
23.4.2 Examples of Algorithms . . . . . . . . . . . . . . . . . . . 379
23.4.3 Algebraic Periodicity . . . . . . . . . . . . . . . . . . . . 381
23.4.4 A Few Words About Convergents . . . . . . . . . . . . . . 381
23.5 Branching Continued Fractions . . . . . . . . . . . . . . . . . . . 382
23.6 Continued Fractions and Rational Knots and Links . . . . . . . . . 386
23.6.1 Necessary Definitions . . . . . . . . . . . . . . . . . . . . 386
23.6.2 Rational Tangles and Operations on Them . . . . . . . . . 387
23.6.3 Main Results on Rational Knots and Tangles . . . . . . . . 388
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 391
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
 Part I
Regular Continued Fractions

In the first part of this book we study geometry of continued fractions in the plane.
As usually happens for low dimensions, this case is the richest, and many in-
teresting results from different areas of mathematics show up here. While trying
to prove analogous multidimensional statements, several research groups have in-
vented completely different multidimensional analogues of continued fractions (see,
e.g., Chap. 23). This is due to the fact that many relations and regularities that are
similar in the planar case become completely different in higher dimensions. So we
start with the classical planar case.
The first two chapters are preliminary. Nevertheless, we strongly recommend that
the reader look through them to get the necessary notation. We start in Chap. 1 with
a brief introduction of the notions and definitions of the classical theory of continued
fractions. Further, in Chap. 2, we shed light on the geometry of the planar integer
lattice. We present this geometry from the classical point of view, where one has a
set of objects and a group of congruences acting on the set of objects. This approach
leads to natural definitions of various integer invariants, such as integer lengths and
integer areas.
In Chap. 3 we present the main geometric construction of continued fractions.
It is based on the notion of sails, which are attributes of integer angles in integer
geometry.
In the next three chapters we come to the problem of the description of integer
angles in terms of integer parameters of their sails. We start in Chap. 4 with a con-
struction of a complete invariant of integer angles (it is written in terms of integer
characteristics of the corresponding sails). Further, in Chap. 5, we use these invari-
ants to define trigonometric functions for integer angles in integer geometry. These
functions have many nice properties similar to Euclidean trigonometric functions,
while the others are totally different. For instance, there is an integer analogue of
the formula α + β + γ = π for the angles in a triangle in the Euclidean plane, which
we prove in Chap. 6.
In Chap. 7 we introduce a notion of geometric continued fractions related to
arrangements of pairs of lines passing through the origin in the plane. One can con-
sider these arrangements to be invariant lines of certain real two-dimensional matri-
2

ces. So geometric continued fractions give rise to powerful invariants of conjugacy

classes of matrices. In particular, we derive from them Gauss’s reduction theory for
SL(2, Z)-matrices. We conclude this chapter with a short discussion of the Markov
spectrum.
In Chap. 8 we focus on the classical question related to Lagrange’s theorem on
the periodicity of continued fractions for quadratic irrationalities, which is based on
the structure of the set of solutions of Pell’s equations. In this chapter we also discuss
a few questions related to Dirichlet groups of matrices, in particular a problem of
solving the equation X n = A in GL(n, Z).
We continue in Chap. 9 with a classical question of Gauss–Kuzmin statistics for
the distribution of elements of continued fractions for arbitrary integers. After the
introduction of the classical approach via ergodic theory, we present a geometric
meaning of the Gauss–Kuzmin statistics in terms of cross ratios in projective geom-
etry.
Regular continued fractions are very good approximations to real numbers. Since
there is a vast number of publications on this subject, we do not attempt to cover all
interesting problems on continued fractions related to approximations. In Chap. 10
we discuss two questions related to best approximations of real numbers and of
arrangements of lines in the plane.
In the last three chapters of the first part we introduce applications of contin-
ued fractions to differential and algebraic geometry. In Chap. 11 we introduce an
“infinitesimal” version of continued fractions arising as sails of planar curves. We
show the relation of these fractions to the motion of a body along this curve satis-
fying Kepler’s second law. Further, in Chap. 12, we give a supplementary definition
of extended integer angles, which we use to prove the formula on the sum of inte-
ger angles in a triangle. In addition, we introduce a definition of the sum of integer
angles, which is a good candidate for the correct addition operation on continued
fractions. Finally, in Chap. 13, we describe the global relations on singular points
of toric surfaces via integer tangents of the integer angles of polygons associated to
these surfaces.
Chapter 1
Classical Notions and Definitions

In this chapter we bring together some basic definitions and facts on continued frac-
tions. After a small introduction we prove a convergence theorem for infinite regular
continued fractions. Further, we prove existence and uniqueness of continued frac-
tions for a given number (odd and even continued fractions in the rational case).
Finally, we discuss approximation properties of continued fractions. For more de-
tails on the classical theory of continued fractions we refer the reader to [105].

1.1 Continued Fractions

1.1.1 Definition of a Continued Fraction

We start with basic definitions in the theory of regular continued fractions.

Definition 1.1 Let α, a0 , . . . , an be real numbers satisfying

1
α = a0 + (1.1)
1
a1 +
1
a2 +
.. 1
.+
an

The expression in the right-hand side of the equality is called a continued fraction
expansion of a given number α (or just a continued fraction of α), and denoted by
[a0 ; a1 : · · · : an ]. The numbers a0 , . . . , an are the elements of this continued fraction.

Definition 1.2 A continued fraction is odd (even) if it has an odd (even) number of
elements.

O. Karpenkov, Geometry of Continued Fractions, 3

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_1, © Springer-Verlag Berlin Heidelberg 2013
4 1 Classical Notions and Definitions

Definition 1.3 An infinite continued fraction with an infinite sequence of elements

a0 , a1 , . . . is the following limit if it exists

lim [a0 ; a1 : · · · : ak ].
k→∞

We denote it by [a0 ; a1 : · · · ].

Definition 1.4 The number [a0 ; a1 : · · · : ak−1 ] is called the k-convergent (or just
convergent) to the finite or infinite continued fraction [a0 ; a1 : · · · ].

Definition 1.5 When the first element a0 of a continued fraction is an integer and all
other elements are positive integers, we say that this continued fraction is regular.

1.1.2 Regular Continued Fractions for Rational Numbers

Let us show how to construct a regular continued fraction for a rational number α.

Algorithm to Construct a Regular Continued Fraction

Input data. Consider a rational number α.
Goal of the algorithm. We are constructing a continued fraction for α. At each
step we define a pair of integers (ak−1 , rk ).
Step 1. If α is an integer, then α = [α], and the algorithm terminates. Suppose now
that α is not an integer. Let us subtract the integer part α and invert the re-
mainder, i.e.,
1
α = α + .
1/(α − α)
Write a0 = α, and continue with the remaining part r1 = 1/(α − a0 ).
Inductive Step k. Suppose we have completed k − 1 steps and get the numbers
ak−2 and rk−1 . Let us find ak−1 and rk :

1
rk−1 = rk−1  + .
1/(rk−1 − rk−1 )

Then we set ak−1 = rk−1  and rk = 1/(rk−1 − ak−1 ).

Output. The algorithm for noninteger numbers stops at Step n when rn is an inte-
ger. We get

α = [a0 ; a1 : · · · : an−1 : an ],

where an = rn .
1.1 Continued Fractions 5

Remark 1.6 Let ri = pi /qi with positive relatively prime integers pi , qi for i ≥ 1.
Then for every k ≥ 1 we have

pk+1 1 1 qk
= rk+1 = = = .
qk+1 rk − rk  pk /qk − pk /qk  pk − qk pk /qk 

Since rk > 1, its denominator is less than its numerator (i.e., qk+1 < pk+1 = qk ).
Hence the sequence of denominators qk decreases with the growth of k. Therefore,
the algorithm stops in a finite number of steps.

Example 1.7 Let us study one example:

9 1 1
=1+ =1+ .
7 (7/2) 3 + 1/2

So we get the continued fraction [1; 3 : 2].

1.1.3 Regular Continued Fractions and the Euclidean Algorithm

The term continued fraction was used for the first time by J. Wallis in 1695, but the
story of continued fractions starts much earlier with the Euclidean algorithm, named
after the Greek mathematician Euclid, who described it in his Elements (Books
VII and X). Actually the algorithm was known before Euclid; it was mentioned
in the Topics of Aristotle. For further historical details we refer to the book [57] by
D.H. Fowler. The goal of the algorithm is to find the greatest common divisor of a
pair of integers. Let us briefly describe the algorithm.

Euclidean Algorithm
Input data. Consider two positive integers p and q.
Goal of the algorithm. To find the greatest common divisor of p and q (usually it
is denoted by gcd(p, q)). We do this in several iterative steps.
Step 1. Set q0 = q. Find integers a0 and q1 with q > r1 ≥ 0 such that

p = a0 q + q1 .

Inductive Step k. Suppose that we have completed k − 1 steps and get the integers
ak−2 and qk−1 . Find integers ak−1 and qk where qk−1 > qk ≥ 0 such that

qk−2 = ak−1 qk−1 + qk .

Output. The algorithm stops at Step n when rn = 0. It is clear that

gcd(p, q) = gcd(q, q1 ) = gcd(q1 , q2 ) = · · · = gcd(qn−2 , qn−1 ) = qn−1 .

6 1 Classical Notions and Definitions

Remark 1.8 Since the sequence (qk ) is a decreasing sequence of positive integers,
the algorithm always stops.

Remark 1.9 The integers ak−1 and rk at each step are uniquely defined. The Eu-
clidean algorithm actually generates the elements of a continued fraction. We always
have
p
= [a0 ; a1 : · · · : an ].
q

1.1.4 Continued Fractions with Arbitrary Elements

In general one can consider continued fractions with arbitrary real elements. In
Chap. 6 and in Chap. 13 we use continued fractions with arbitrary integer elements
to describe integer triangles and later to generalize to the case of toric varieties. In
Chap. 11 we present some of the geometric ideas behind continued fractions with
arbitrary integer and real coefficients. In the rest of the book we shall deal with
regular continued fractions.
In the case of continued fractions with arbitrary elements, we could encounter
infinity while calculating the value of a continued fraction. For instance, this is the
case for the continued fraction [1; 1 : 0]. To avoid the annoying consideration of
different cases of infinities, we propose to add an element ∞ to the field of real
numbers R and define the following operations:

1 1
∞ + a = a + ∞ = ∞, = ∞, = 0.
0 ∞

Denote the resulting set by R. Notice that we do not add two elements: +∞ and
−∞, but only one: ∞. In some sense we are considering the projectivization of the
real line.
In this notation, for the continued fraction [1 : 1; 0] we have
1 1 1
[1; 1 : 0] = 1 + =1+ =1+ = 1 + 0 = 1.
1 + 1/0 1+∞ ∞

1.2 Convergence of Infinite Regular Continued Fractions

In this section we show that a sequence of k-convergents of any infinite regular

continued fraction converges to some real number. First, we give a formula for nu-
merators and denominators of rational numbers in terms of elements of continued
fractions of these numbers. Second, we introduce partial numerators and denomi-
nators and prove some relations between them. Finally, we formulate and prove a
theorem on convergence of infinite continued fractions.
1.2 Convergence of Infinite Regular Continued Fractions 7

Let us say a few words about the numerators and denominators of rational num-
bers and the elements of their continued fractions. It turns out that the numerator
and the denominator can be expressed as polynomial functions of the elements of
the continued fraction. Namely, there exists a unique pair of polynomials Pk and Qk
in the variables x0 , . . . , xk with nonnegative integer coefficients such that
Pk (x0 , . . . , xk )
= [x0 ; x1 : · · · : xk ], and Pk (0, . . . , 0) + Qk (0, . . . , 0) = 1.
Qk (x0 , . . . , xk )
The first condition defines the polynomials only up to a multiplicative constant, so
the second condition is a necessary normalization condition.

Example 1.10 Let us calculate the above polynomials for the continued fractions
[x0 ], [x0 ; x1 ], [x0 ; x1 : x2 ], . . . . The polynomials are as follows:

P0 (x0 ) = x0 , Q0 (x0 ) = 1;
P1 (x0 , x1 ) = x0 x1 + 1, Q1 (x0 , x1 ) = x1 ;
P2 (x0 , x1 , x2 ) = x0 x1 x2 + x2 + x0 , Q2 (x0 , x1 , x2 ) = x1 x2 + 1;
...

Now the numerator and the denominator are described via polynomials P and Q
as follows. Consider a finite (or infinite) continued fraction [a0 ; a1 : · · · : an ] with
n ≥ k (or [a0 ; a1 : · · · ] respectively). We set

pk = Pk (a0 , . . . , ak ) and qk = Qk (a0 , . . . , ak ).

In some sense the numbers pk and qk may be considered partial numerators and
partial denominators of rational numbers. Namely (by Proposition 1.13 below)
these numbers are the numerators and denominators of k-convergents. If k = n,
we get the numerator and the denominator of [a0 ; a1 : · · · : an ].
Let us prove several interesting facts and relations on the numbers pk and qk (and
hence that they are true numerators and denominators for k = n). In Proposition 1.12
below we prove that the integers pk and qk are relatively prime for every k. Then in
Proposition 1.13 we prove an important recurrence relation on both numerators and
denominators, which we will use in several further chapters. After that we deduce
one more useful relation (see Proposition 1.15) and prove Theorem 1.16 on the
convergence of infinite continued fractions.
For the next several propositions we use some additional notation:

p̂k = Pk−1 (a1 , . . . , ak ) and q̂k = Qk−1 (a1 , . . . , ak ).

We start with the following lemma.

Lemma 1.11 The following holds:


pk = a0 p̂k + q̂k ,
qk = p̂k .
8 1 Classical Notions and Definitions

Proof Since
p̂k
= [a1 ; a2 : · · · : ak ]
q̂k
we get
pk 1 a0 p̂k + q̂k
= a0 + = .
qk p̂k /q̂k p̂k
Therefore,

pk = λ(a0 p̂k + q̂k ),
qk = λp̂k .
Now we rewrite the second condition for polynomials Pk and Qk :

1 = Pk (0, . . . , 0) + Qk (0, . . . , 0)
 
= λ a0 Pk−1 (0, . . . , 0) + Qk−1 (0, . . . , 0) + λPk−1 (0, . . . , 0).

It is clear that the last expression represents a positive integer that is divisible by λ.
Hence 1 is divisible by λ, and therefore λ = 1. 

Proposition 1.12 Let [a0 ; a1 : · · · : an ] be a continued fraction with integer ele-

ments. Then the corresponding integers pk and qk are relatively prime.

Proof We prove the statement by induction on k.

It is clear that p0 = a0 and q0 = 1 are relatively prime.
Suppose that the statement holds for k − 1. Then p̂k and q̂k are relatively prime
by the induction assumption. Now the statement holds directly from the equalities
of Lemma 1.11. 

Let us now present a recursive definition of partial numerators and denominators.

Proposition 1.13 For every integer k we get


pk = ak pk−1 + pk−2 ,
qk = ak qk−1 + qk−2 .

Proof We prove the statement by induction on k.

For k = 2 the statement holds, since
p0 a 0 p1 a 0 a 1 + 1
= and = ,
q0 1 q1 a1

and therefore
p2 a 2 + a 0 a 1 a 2 + a 0 a 2 p 1 + p 0
= = .
q2 a1 a2 + 1 a2 q1 + q0
1.2 Convergence of Infinite Regular Continued Fractions 9

Suppose the statement holds for k − 1. Let us prove it for k:

pk 1 1
= a0 + = a0 +
qk p̂k /q̂k (ak p̂k−1 + p̂k−2 )/(ak q̂k−1 + q̂k−2 )
a0 (ak p̂k−1 + p̂k−2 ) + ak q̂k−1 + q̂k−2
=
ak p̂k−1 + p̂k−2
ak (a0 p̂k−1 + q̂k−1 ) + (a0 p̂k−2 + q̂k−2 ) ak pk−1 + pk−2
= = .
ak p̂k−1 + p̂k−2 ak qk−1 + qk−2

(The last equality holds by Lemma 1.11.) Therefore, the relations of the system
hold. 

From Proposition 1.13 it follows that the partial numerators and denominators are
bounded by the Fibonacci numbers Fi (defined inductively as follows: F1 = F2 = 1,
and Fn = Fn−1 + Fn−2 ).

Corollary 1.14 For regular continued fractions the following estimates hold:

|pk | ≥ Fk and qk ≥ Fk+1 .

Proof We prove the statement by induction on k. Direct calculations show that

|p0 | ≥ 0, |p1 | ≥ 1, and |p2 | ≥ 1,

|q0 | ≥ 1, |q1 | ≥ 1, and |q2 | ≥ 2.

Suppose that the statement holds for k − 2 and k − 1, we prove it for k. Notice
that the qk are all positive and the pk are either all negative or all nonnegative. We
apply Proposition 1.13:

|pk | = ak |pk−1 | + |pk−2 | ≥ 1 · Fk−1 + Fk−2 = Fk

and
qk = ak qk−1 + qk−2 ≥ 1 · Fk + Fk−1 = Fk+1 .
This concludes the proof. 

In the next proposition we present another useful expression for partial numera-
tors and denominators.

Proposition 1.15 For every k ≥ 1, the following holds:

pk−1 pk (−1)k
− = .
qk−1 qk qk−1 qk
10 1 Classical Notions and Definitions

Proof Let us multiply both sides by qk−1 qk . We get

pk−1 qk − pk qk−1 = (−1)k .

We prove this by induction on k.

For k = 1 we have

p0 q1 − p1 q0 = a1 a0 − (a1 a0 + 1) = −1.

Suppose the statement holds for k − 1. Let us prove it for k. By Proposition 1.13
we get

pk−1 qk − pk qk−1 = pk−1 (ak qk−1 + qk−2 ) − (ak pk−1 + pk−2 )qk−1
= −(pk−2 qk−1 − pk−1 qk−2 ) = (−1)k .

Therefore, the statement holds. 

Now we are ready to prove the following fundamental theorem.

Theorem 1.16 The sequence of k-convergents of an arbitrary regular continued

fraction converges.

Notice that the theorem does not always hold for continued fractions with arbi-
trary real elements.

Proof of Theorem 1.16 Let [a0 ; a1 : · · · ] be an infinite regular continued fraction.

From Proposition 1.15 and Corollary 1.14 we have
 
 pk−1 pk  1
 −  ≤ .
q qk Fk Fk+1
k−1

Since the sum

∞
 1
Fk Fk+1
k=1

converges (we leave this statement as an exercise for the reader), the sequence ( pqkk )
is a Cauchy sequence. Therefore, ( pqkk ) converges. 

1.3 Existence and Uniqueness of a Regular Continued Fraction

for a Given Real Number
In the next theorem we show that for every real number there exists a regular con-
tinued fraction representing it. For an irrational number, the corresponding regular
continued fraction is unique and infinite. For a rational number there are exactly two
1.3 Existence and Uniqueness of a Regular Continued Fraction 11

continued fractions,

[a0 ; a1 : · · · : an ] = [a0 ; a1 : · · · : an − 1 : 1],

with one of them odd, the other even.

Theorem 1.17
(i) For every rational number there exist a unique odd and a unique even regular
continued fraction.
(ii) For every irrational number α there exists a unique infinite regular continued
fraction whose k-convergents converge to α.

For instance, 9
7 = [1; 3 : 2] = [1; 3 : 1 : 1] and π = [3; 7 : 15 : 1 : 292 : 1 : 1 : 1 :
2 : · · · ].

Proof of Theorem 1.17 Existence. In Sect. 1.1 we have shown how to construct a
regular continued fraction [a0 ; a1 : · · · : an ] for a rational number α. Notice that if α
is rational but not an integer, then an > 1, and therefore,

α = [a0 ; a1 : · · · : an ] = [a0 ; a1 : · · · : an − 1 : 1].

One of these continued fractions is odd and the other is even. For an integer α we
always get α = [α] = [α − 1; 1].
In the case of an irrational number α, the algorithm never terminates, generating
the regular continued fraction α  = [a0 ; a1 : a2 : · · · ] and the sequence of remainders
rk > 1 such that
α = [a0 ; a1 : · · · : ak−1 : rk ].
Let us show that α = α  . From Proposition 1.13 for [a0 ; a1 : · · · : ak−1 : rk ] and
[a0 ; a1 : · · · : ak−1 : · · · ] we have
pn−1 rn + pn−2 pn pn−1 an + pn−2
α= and = .
qn−1 rn + qn−2 qn qn−1 an + qn−2

Using these expressions and the fact that an = rn , we have

   
   
α − pn  =  (pn−1 qn−2 − pn−2 qn−1 )(rn − an ) 
 qn   (qn−1 rn + qn−2 )(qn−1 an + qn−2 ) 
 
 1 
< 
(qn−1 rn + qn−2 )(qn−1 an + qn−2 ) 
1 1
< ≤ .
qn−1 qn Fn Fn+1

The last inequality follows from Corollary 1.14.

12 1 Classical Notions and Definitions

Therefore, the sequence ( pqnn ) converges to α, and hence α = α  .

Uniqueness. Consider a rational number α. Let us prove the uniqueness of the
finite regular continued fraction α = [a0 ; a1 : · · · : an ], where an
= 1. We prove this
by reductio ad absurdum.
Suppose
  

α = [a0 ; a1 : · · · : ak : ak+1 : · · · : an ] = a0 ; a1 : · · · : ak : ak+1 : · · · : am ,

where ak+1
= ak+1 . Then we have

pk rk+1 + pk−1 pk rk+1

 
+ pk−1 
pk rk+1 + pk−1
α= =    =  .
qk rk+1 + qk−1 qk rk+1 + qk−1 qk rk+1 + qk−1

Therefore, rk+1 = rk+1 , and thus a  

k+1 = rk+1  = rk+1  = ak+1 . We arrived at a
contradiction.
The proof of uniqueness for continued fractions of irrational numbers is the same
as in the case of rational numbers. 

1.4 Monotone Behavior of Convergents

In this section we prove two statements on the monotone behavior of convergents.
We start with a statement on the sequences of odd and even convergents.

Proposition 1.18
(i) The sequence of even convergents (p2k /q2k ) is increasing, and the sequence of
odd convergents (p2k+1 /q2k+1 ) is decreasing.
(ii) For every real α and a nonnegative integer k we have
p2k p2k+1
≤α and ≥ α.
q2k q2k+1
Equality holds only for the last convergent in case of rational α.

Proof
(i) By Proposition 1.13 we have

pm−2 pm pm−2 pm−1 pm−1 pm

− = − − −
qm−2 qm qm−2 qm−1 qm−1 qm
m−1

(−1) 1 1
= − .
qm−1 qm−2 qm
Since the sequence of the denominators (qk ) is increasing (by Proposition 1.15),
we have that pm−2 /qm−2 is greater than pm /qm for all even m, and it is less for
all odd m. This concludes the proof of (i).
1.4 Monotone Behavior of Convergents 13

(ii) The sequence of even (odd) convergents is increasing (decreasing) and tends to
α in the irrational case, and we end up with some pn /qn = α in the rational
case. This implies the second statement of the proposition. 

In the second statement we show that the larger k is, the better a k-convergent
approximates α.

Proposition 1.19 The sequence of real numbers

 
 
α − pk , k = 0, 1, 2, . . . ,
 qk 

is strongly decreasing, except for the case of α = [a0 , 1, 1], where this sequence
consists of the following three equivalent elements: (1/2, 1/2, 0).

Proof Recall that rk denotes the reminder [ak+1 ; ak+2 : · · · ].

Let us first prove that
 
   
α − α > α − p1 .
 q1 
This inequality is equivalent to
   
 1   1 1 
a 0 + − a  >  a + − a + ,
 a1 + 1/r2
0  0 a + 1/r 0
a1 
1 2

which is equivalent to
1 1/r2
> ,
a1 + 1/r2 a1 (a1 + 1/r2 )
and further to
a1 r2 > 1.
The only case in which a1 r2 ≤ 1 is a1 = r2 = 1. This is exactly the exceptional case
mentioned in the formulation of the theorem.
Now we proceed with the general case of k ≥ 2. From Proposition 1.13 we have
 
 
α − pk−1  = |pk−1 qk−2 − qk−1 pk−2 |
 qk−1  qk−1 (qk−1 rk + qk−2 )
1
= . (1.2)
qk−1 (qk−1 rk + qk−2 )

Similarly, we have
 
  1
 α − pk  = . (1.3)
 qk  qk (qk rk+1 + qk−1 )
14 1 Classical Notions and Definitions

Let us estimate the expression on the right side of the last equality:
1 1 1
≤ =
qk (qk rk+1 + qk−1 ) qk (qk + qk−1 ) qk (qk−1 ak + qk−2 + qk−1 )
1 1
= ≤
qk (qk−1 (ak + 1) + qk−2 ) qk (qk−1 rk + qk−2 )
1
< . (1.4)
qk−1 (qk−1 rk + qk−2 )
In the last inequality we used the fact that qk > qk−1 , which follows directly from
Proposition 1.13 for k ≥ 2:

qk = qk−1 ak + qk−2 > qk−1 .

Let us substitute the first and the last expressions of inequality (1.4) by equalities
(1.2) and (1.3). We get
   
   
α − pk  > α − pk−1 .
 qk   qk−1 
Therefore, the sequence is strictly decreasing. 

1.5 Approximation Rates of Regular Continued Fractions

It is interesting to compare advantages and disadvantages of continued fractions
with respect to decimal expansions. It is clear that it is very easy algorithmically
to add and to multiply two numbers if you know their decimal expansions, while
for continued fractions these operations are hard. For instance, there is no algebraic
expression for the elements of sums and products via elements of summands or
factors. Probably that is the reason why in real life, decimal expansions are widely
known and continued fractions are not. On the other hand, quadratic irrationalities
have periodic continued fractions, while their decimal expansions are not distin-
guishable from an arbitrary transcendental number (see Chap. 8). In addition, con-
tinued fractions give exact expressions for certain expressions involving exponents
and logarithms (see Exercise 1.4 below). They play a key role in Gauss’s reduction
theory for describing SL(2, Z) matrices (see Chap. 7). It turns out that convergents
of continued fractions are good as rational approximations of real numbers. In this
section we say a few words about the approximation rates of convergents. Later,
in Chap. 10, we study questions related to best approximations of real numbers by
rational numbers.

Theorem 1.20 Consider the inequality

 
 
α − p  < c . (1.5)
 q  q2
1.5 Approximation Rates of Regular Continued Fractions 15

Let c ≥ √1 . Then for every α, the inequality has an infinite number of integer solu-
5
tions (p, q).

Let us reformulate the essence of Theorem 1.20.

Lemma 1.21 Theorem 1.20 is true if for every irrational α, there are infinitely many
integer solutions (p, q) of
 
 
α − p  ≤ √ 1 .
 q 5q 2

Proof It is clear that for rational α = p/q the integer pairs (np, nq) are solutions.
In the irrational case it is enough to prove the theorem for c = √1 . Let us prove that
5
the equality can occur at most twice. Suppose we have
p 1 p̂ 1
α− = ±√ and α − = ±√
q 5q 2 q̂ 5q̂ 2
for some choice of signs. Then

p p̂ 1 1 1
− = √ ± 2 −± 2 .
q q̂ 5 q̂ q

Since 1 and √1 are linearly independent over Q, we have

5

p p̂
= and |q| = |q̂|.
q q̂
Hence equality holds at most for two pairs of integers (p, q) and (−p, −q). 

Our next step in the proof of Theorem 1.20 is to give an estimate on the growth
of denominators.

Lemma 1.22 Consider a real number α with regular continued fraction [a0 ; a1 :
· · · ]. Let pk /qk be its k-convergents. Then for an infinite sequence of integers we
have
√
qk 1+ 5
≥ .
qk−1 2

Proof By Proposition 1.13 we have

qk qk−2
= ak + .
qk−1 qk−1
Hence if ak ≥ 2, then
√
qk 1+ 5
≥ ak ≥ 2 > .
qk−1 2
16 1 Classical Notions and Definitions

Suppose now that ak = 1 and let

√
qk 1+ 5
< .
qk−1 2
Then
√
qk−1 1 1 1+ 5
= > √ = .
qk−2 qk /qk−1 − 1 (1 + 5)/2 − 1 2
Therefore, at least one of every two sequential numbers qk−1 /qk−2 and qk /qk−1
satisfies the condition of the lemma, and hence there are infinitely many qk /qk−1
satisfying the condition. 

Proof of Theorem 1.20 From Lemma 1.21 it is enough to show that for every irra-
tional α there exist infinitely many pairs of rational numbers satisfying
 
 
α − p  ≤ √ 1 .
 q 5q 2
Consider two consecutive convergents. From Proposition 1.15 we have
 
 pk−1 pk  1
 −  = .
q q q
k−1 k k−1 qk

In addition, from Proposition 1.18 it follows that α is contained in the segment with
endpoints pk−1 /qk−1 and pk /qk . Hence if the distance between endpoints is small
enough, namely
1 1 1
≤ √ 2 + √ 2,
qk−1 qk 5qk−1 5qk
then either (pk−1 , qk−1 ) or (pk , qk ) is a solution of the above equation. The last
inequality is equivalent to the following:

qk−1 2 √ qk−1
− 5 + 1 > 0.
qk qk
√
±1+ 5
The quadratic polynomial on the left side has two roots: 2 . Hence the inequal-
ity holds when
√
qk−1 1 + 5
≥ .
qk 2
From Lemma 1.22 we have infinitely many k satisfying the last inequality. There-
fore, the number of solutions of the above inequality is also infinite. This concludes
the proof of Theorem 1.20. 

It turns out that for certain numbers the estimate of the approximation rate cannot
be essentially improved.
1.5 Approximation Rates of Regular Continued Fractions 17

Proposition 1.23 Let α be the golden ratio, i.e.,

√
1+ 5
α= = [1; 1 : 1 : 1 : 1 : · · · ]
2

If c < √1 , then inequality (1.5) has only finitely many solutions.

5

√
Denote the golden ratio by θ and its conjugate (1 − 5)/2 by θ .

Proof of Proposition 1.23 First of all, let us show that it is enough to check only all
the convergents pk /qk . From Theorem 1.16 it follows that best approximations are
convergents to a number. Let p/q be a rational number such that qk ≤ q < qk+1 .
Therefore,
     
2
 p  
2 pk  
2 pk 
q α −  ≥ q α −  ≥ qk α − .
q qk qk
Hence if p/q is a solution of inequality (1.5), then pk /qk is a solution of inequality
(1.5) as well. Therefore, if there are infinitely many solutions of inequality (1.5)
then there are infinitely many convergents to the golden ratio among them.
Second, we prove the statement for the convergents. From Proposition 1.13 it
follows that the k-convergent to the golden ratio equals Fk+1 /Fk . Recall Binet’s
formula for Fibonacci numbers via the golden ratio and its conjugate:

k
θk − θ
Fk = √
5

(for more details see [206]). We have

     k+1 − θ k+1 
  k 
      
θ − pk−1  = θ − Fk+1  = θ − θ = θ 
 qk−1   Fk   k   k 
θk − θ θk − θ
 
 1 − θ 2k   
=   = √ 1 1 − θ 2k .
k  5Fk2
(θ k − θ )2

2k
Since |θ | < 1, we have |1 − θ | = 1 + o(1) and hence
 

 
θ − pk−1  = √ 1 +o 2
1
.
 qk−1  2
5qk−1 qk−1

This implies the statement for convergents and concludes the proof of Proposi-
tion 1.23. 
18 1 Classical Notions and Definitions

1.6 Exercises

Exercise 1.1 Prove that for every k ≥ 2 we get

pk−2 pk (−1)k−1 ak
− = .
qk−2 qk qk qk−2

Exercise 1.2 Prove that for every k ≥ 1 we get

qk
= [ak ; ak−1 : · · · : a1 ].
qk−1

Exercise 1.3 Consider the convergents of a regular continued fraction. Prove that
for every k ≥ 0 we get
 
1  pk  1

≥ α− > .
qk qk+1  qk qk (qk+1 + qk )

Exercise 1.4 Prove that

√
(a) 2 = [1; (2)];
(b) exp(1) = [2; 1 : 2 : 1 : 1 : 4 : 1 : 1 : 6 : 1 : 1 : 8 : 1 : 1 : 10 : · · · ].

Exercise 1.5 Consider an irrational number α.

(a) Suppose that we know that α ≈ 4,17. Is it true that 417
100 is its best approximation?
(b) Find the set of all real numbers for which the rational number [1; 2 : 3 : 4] is one
of the best approximations.

Exercise 1.6 Construct an infinite continued fraction with real coefficients that has
exactly two limit points: 1 and −1.
Chapter 2
On Integer Geometry

In many questions, the geometric approach gives an intuitive visualization that leads
to a better understanding of a problem and sometimes even to its solution. In the next
chapters we give an interpretation of the elements of continued fractions in terms
of integer geometry, with the continued fractions being associated to certain invari-
ants of integer angles. The geometric viewpoint on continued fractions also gives
key ideas for generalizing Gauss–Kuzmin statistics to studying multidimensional
Gauss’s reduction theory, leading to several results in toric geometry.
The notion of geometry in general can be interpreted in many different ways. In
this book we think of a geometry as of a set of objects and a congruence relation,
which is normally defined by some group of transformations. For instance, in Eu-
clidean geometry in the plane, we study points, lines, segments, polygons, circles,
etc., with the congruence relation being defined by the group of all length-preserving
transformations O(2, R) (the orthogonal group). The aim of this chapter is to intro-
duce basic ideas of integer geometry. In order to have a general impression we start
with the simplest, two-dimensional, case. We will need multidimensional integer
geometry only in the second part of the book, so we describe it later, in Chaps. 14
and 15.
We start in Sect. 2.1 with general definitions of integer geometry, and in par-
ticular, define integer lengths, distances, areas of triangles, and indexes of angles.
Further, in Sects. 2.2 and 2.3 we extend the notion of integer area to the case of
arbitrary polygons whose vertices have integer coordinates. In Sect. 2.4 we formu-
late and prove the famous Pick’s formula that shows how to find areas of polytopes
simply by counting points with integer coordinates contained in them. Finally, in
Sect. 2.5 we formulate one theorem in the spirit of Pick’s theorem: it is the so-called
twelve-point theorem.

O. Karpenkov, Geometry of Continued Fractions, 19

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_2, © Springer-Verlag Berlin Heidelberg 2013
20 2 On Integer Geometry

Fig. 2.1 The following triangles are integer congruent

2.1 Basic Notions and Definitions

2.1.1 Objects and Congruence Relation of Integer Geometry

We consider the integer lattice Z2 in R2 . The objects of integer geometry are integer
points with integer coordinates, integer segments and polygons having all vertices
in the integer lattice, integer lines passing through pairs of integer points, integer
angles with vertices at integer points. The congruence relation is defined by the
group of affine transformations preserving the integer lattice, i.e. Aff(2, Z). This
group is a semidirect product of GL(2, Z) and the group of translations on integer
vectors. We use ∼ = to indicate that two objects are integer congruent.
The congruence relation is slightly different from that in the Euclidean case. We
illustrate this with integer congruent triangles in Fig. 2.1.

Remark 2.1 To avoid confusion we add prefixes for triangles and angles, writing
ABC and ∠ABC respectively.

2.1.2 Invariants of Integer Geometry

As long as we have objects and a group of transformations that acts on the objects,
we get invariants—quantities that are preserved by transformations of the objects.
Usually the study of these invariants is the main subject of the corresponding ge-
ometry. For instance, in Euclidean geometry, the invariants are lengths of segments,
areas of polyhedra, measures of angles, etc.
So what are the invariants in integer geometry? There are two different types of
invariants in integer geometry: affine invariants and lattice invariants. Affine invari-
ants are intrinsic to affine geometries in general. Affine transformations preserve
– the set of lines;
– the property that a point on a line is between two other points;
– the property that two points are in one half-plane with respect to a line.
Lattice invariants are induced by the group structure of the vectors in the lattice. The
majority of lattice invariants are indices of certain subgroups (i.e., sublattices).
2.1 Basic Notions and Definitions 21

2.1.3 Index of Sublattices

Recall several notions from group theory. Let H be a subgroup of a group G. Con-
sider an element g ∈ G. The sets

gH = {gh | h ∈ H } and Hg = {hg | h ∈ H }

are called left and right cosets of H in G.

The index of the subgroup H in the group G is the number of left cosets of H in
G, denoted by |G : H |. (For example, for a positive integer n, we have |Z : nZ| = n.)
Note that within this book we work with abelian groups Zk , where for every g left
cosets of g coincide with right cosets, i.e., gH = Hg.
In integer geometry we consider the additive group of integer vectors Zn , which
is called the integer lattice. A subgroup of an integer lattice is called an integer
sublattice of the integer lattice. Actually, any integer sublattice is isomorphic to
Zk for k ≤ n, with the sublattice contained in some k-dimensional linear subspace
of Rn . The number k is the dimension of the sublattice.
Let A be an integer point and G be an integer sublattice. We say that the set of
points
L = {A + g | g ∈ G}
is the integer affine sublattice. Basically, the group G does not depend on the choice
of the point A in L. The invariant definition for G is as follows: the lattice G is the
set of all vectors with endpoints in L.
Consider two integer affine sublattices L1 ⊂ L2 whose integer sublattices G1
and G2 have the same dimension. It is clear that G1 ⊂ G2 , so the index |G2 : G1 | is
well defined. Since |G2 : G1 | is an invariant of the pair (G1 , G2 ) under the action of
GL(n, Z), it is also an invariant of the pair (L1 , L2 ) under the action of Aff(n, Z).
One of the simple ways to calculate the index of a two-dimensional sublattice in
Z2 is by counting integer points in a basis parallelogram for the sublattice.

Proposition 2.2 The index of a sublattice generated by a pair of integer vectors v

and w in Z2 equals the number of all integer points P satisfying

AP = αv + βw with 0 ≤ α < 1; 0 ≤ β < 1,

where A is an arbitrary integer point.

Proof Let H be the subgroup of Z2 generated by v and w. Define

Par(v, w) = {αv + βw | 0 ≤ α, β < 1}.

First, we show that for every integer vector g there exists an integer point P ∈
Par(v, w) such that AP ∈ gH . The point P is constructed as follows. Let

g = λ1 v + λ2 w.
22 2 On Integer Geometry

Consider
   
P = λ1 − λ1  v + λ2 − λ2  w + A.
Since
0 ≤ λ1 − λ1 , λ2 − λ2  < 1,
the point P is inside the parallelogram and AP ∈ gH .
Second, we prove the uniqueness of P . Suppose that for two integer points
P1 , P2 ∈ Par(v, w) we have AP1 ∈ gH and AP2 ∈ gH . Hence the vector P1 P2 is
an element in H . The only element of H of type

αv + βw with 0 ≤ α, β < 1

is the zero vector. Hence P1 = P2 .

Therefore, the integer points of Par(v, w) are in one-to-one correspondence to
the cosets of H in Z2 . 

Example 2.3 Consider the points A, B, and C as follows:

Then we have six points satisfying the above condition. Therefore, the index of the
sublattice generated by AB and AC is 6.

2.1.4 Integer Length of Integer Segments

Now we are ready to define several integer invariants. We start with integer length.

Definition 2.4 The integer length of an integer segment AB is the number of integer
points in the interior of AB plus one. We denote it by l(AB).

For example, the integer lengths of the segments AB and CD are both equal to 2.

An alternative invariant definition is as follows: the integer length of an integer

segment AB is the index of the sublattice generated by the vector B − A in the lattice
of integer points in the line containing AB.
2.1 Basic Notions and Definitions 23

In Euclidean geometry the length is a complete invariant of congruence classes

of segments. The same is true in lattice geometry.

Proposition 2.5 Two integer segments are congruent if and only if they have the
same integer length.

Proof Let AB be an integer segment of length k. Let us prove that it is integer

congruent to the integer segment with endpoints O(0, 0) and K(k, 0). Consider an
integer translation sending A to the origin O. Let this translation send B to an
integer point B  (x, y).
Since l(OB ) = l(AB) = k, we have x = kx  and y = ky  , where x  and y  are
relatively prime. Hence there exists an integer pair (a, b) such that

bx  − ay  = 1.

Hence the matrix



x y
a b
is invertible, and therefore, it is in SL(2, Z). This matrix takes the segment OK
to OB , and thus OK ∼= OB ∼ = OB. Hence all the segments of integer length k are
congruent to the segment OK. Therefore, they are all integer congruent. 

2.1.5 Integer Distance to Integer Lines

There is no natural simple definition of orthogonal vectors in integer geometry. Still,

it is possible to give a natural definition of an integer distance from a point to a line.

Definition 2.6 Consider integer points A, B, and C that do not lie in one line. An
integer distance from the point A to the integer segment (line) BC is the index of
the sublattice generated by all integer vectors AV, where V is an integer point of the
line BC in the integer lattice. We denote it by ld(A, BC).
The points A, B, and C are collinear, we agree to say that the integer distance
from A to BC is zero.

One of the geometric interpretations of integer distance from a point A to BC

is as follows. Draw all integer lines parallel to BC. One of them contains the point
A (let us call it AA ). The integer distance ld(A, BC) is the number of lines in the
region bounded by the lines AB and AA plus one. For the example of Fig. 2.2, we
have the following:
24 2 On Integer Geometry

Fig. 2.2 A triangle ABC and the sublattices for calculation of ld(A, BC), lS(ABC), and
lα(∠BAC)

There are two integer lines parallel to AB and lying in the region with boundary
lines AB and AA . Hence, ld(A, BC) = 2 + 1 = 3.

2.1.6 Integer Area of Integer Triangles

Let us start with the notion of integer area for integer triangles. We will define the
integer areas for polygons later, in the next subsection.

Definition 2.7 The integer area, denoted by lS(ABC), of an integer triangle

ABC is the index of the sublattice generated by the vectors AB and AC in the
integer lattice.
For the points O, A, and B in one line we say that lS(ABC) = 0.

For instance, the triangle in Fig. 2.2 has integer area equal to 6.
As in Euclidean geometry, integer area does not uniquely determine the congru-
ence class of triangles. Nevertheless, all integer triangles of unit integer area are
congruent, since the vectors of the edges of such triangles generate the integer lat-
tice.

2.1.7 Index of Rational Angles

An angle is called rational if its vertex is an integer point and both its edges contain
integer points other than the vertex.
2.2 Empty Triangles: Their Integer and Euclidean Areas 25

Definition 2.8 The index of a rational angle ∠BAC, denoted by lα(∠AOB), is the
index of the sublattice generated by all integer vectors of the lines AB and AC in the
integer lattice.
In addition, if the points A, O, and B are collinear, we say that lα(∠AOB) = 0.

Geometrically, the index of an angle ∠BAC is the integer area of the smallest tri-
angle AB C whose edges AB and AC generate the sublattices of lines containing
the corresponding edges. For instance, the angle ∠BAC in Fig. 2.2 has index equal
to 2:

Let us conclude this subsection with a general remark.

Remark 2.9 We write l, ld, lS, and lα (starting with the letter “l”) to indicate that
all these notions are well defined for any lattice, and not just for the integer lattice.

2.2 Empty Triangles: Their Integer and Euclidean Areas

Let us completely study the case of the following “smallest” triangles.

Definition 2.10 An integer triangle is called empty if it does not contain integer
points other than its vertices.

We show that empty triangles are exactly the triangles of integer area 1 and Eu-
clidean area 1/2. In particular, this means that all empty triangles are integer con-
gruent (as we show later, this is not true in the multidimensional case for empty
tetrahedra).
Denote by S(ABC) the Euclidean area of the triangle ABC. Recall that
1 
S(ABC) = det(AB, AC).
2
(Here by det(v, w) we denote the determinant of the (2 × 2) matrix whose first and
second columns contain the coordinates of the vectors v and w respectively.)

Proposition 2.11 Consider an integer triangle ABC. Then the following state-
ments are equivalent:
(a) ABC is empty;
(b) lS(ABC) = 1;
(c) S(ABC) = 1/2.
26 2 On Integer Geometry

Proof (a) ⇒ (b). Let an integer triangle ABC be empty. Then by symmetry, a
parallelogram with edges AB and AC does not contain integer points except for
its vertices as well. Therefore, by Proposition 2.2 there is only one coset for the
subgroup generated by AB and AC. Hence, AB and AC generate the integer lattice,
and lS(ABC) = 1.
(b) ⇒ (c). Let lS(ABC) = 1. Hence the vectors AB and AC generate the integer
lattice. Thus any integer point is an integer combination of them, so

(1, 0) = λ1 AB + λ2 AC and (0, 1) = μ1 AB + μ2 AC,

with integers λ1 , λ2 , μ1 , μ2 . Let also

AB = b1 (1, 0) + b2 (0, 1) and AC = c1 (1, 0) + c2 (0, 1)

for some integers a1 , a2 , b1 , b2 . Therefore,

b1 c1 λ1 μ 1 1 0
= .
b2 c2 λ2 μ2 0 1
Since these matrices are integer, both their determinants equal either 1 or −1. Hence
the Euclidean area of ABC coincides with the Euclidean area of the triangle with
vertices (0, 0), (1, 0), and (0, 1) and equals 1/2.
(c) ⇒ (a). Consider an integer triangle of Euclidean area 1/2. Suppose that it has
an integer point in the interior or on its sides. Then there exists an integer triangle
with Euclidean area smaller than 1/2, which is impossible (since the determinant of
two integer vectors is an integer). 

Remark 2.12 Proposition 2.11 implies that all empty integer triangles are congru-
ent.

2.3 Integer Area of Polygons

In this section we show that every integer polygon is decomposable into empty
triangles. Then we define the integer area of polygons to be the number of empty
triangles in the decomposition. Additionally, we show that this definition of integer
area is well defined and coincides with the definition of the integer area of triangles.
Consider a closed broken line A0 A1 . . . An−1 An (An = A0 ) with finitely many
vertices and without self-intersections in R2 . By the polygonal Jordan curve theo-
rem, this broken line separates the plane into two regions: one of them is homeomor-
phic to a disk and the other to an annulus. An n-gon (or just a polygon) A1 . . . An is
the closure of the region homeomorphic to a disk.

Proposition 2.13
(i) Every integer polygon admits a decomposition into empty triangles.
(ii) Two decompositions of the same polygon into empty triangles have the same
number of empty triangles.
2.3 Integer Area of Polygons 27

(iii) The number of empty triangles in any decomposition of a triangle ABC

equals lS(ABC).

Proposition 2.13 introduces a natural extension of the integer area to the case of
integer polygons.

Definition 2.14 The integer area of an integer polygon is the number of empty
triangles in its decomposition into empty triangles.

Proof of Proposition 2.13

(i) We prove the statement for integer n-gons by induction on n.
Base of induction. We start with triangles. To prove the statement for trian-
gles we use another induction on the number of integer points in the closure
triangle.
If there are exactly three integer points in the closure of a triangle, then they
are the vertices of the triangle (since the triangle is integer). Hence this triangle
is empty; the decomposition consists of one empty triangle.
Suppose that every triangle with k  < k integer points (k ≥ 3) is decompos-
able into empty triangles. Consider an arbitrary integer triangle ABC with k
integer points. Since k ≥ 3, there exists an integer point P in the triangle dis-
tinct from the vertices. Decompose the triangle into ABP, BCP, and CAP
(excluding triangles of zero area). This decomposition consists of at least two
triangles, and each of these triangles has at most k − 1 integer points. By the in-
duction assumption each of these triangles admits a decomposition into empty
triangles. Hence ABC is decomposable as well.
Therefore, every integer triangle admits a decomposition into integer trian-
gles.
Induction step. Suppose that every integer k  -gon for k  < k is decompos-
able into empty triangles. Let us find a decomposition for an arbitrary integer
k-gon A1 . . . Ak . If there exists an integer index s such that As , As+1 , As+2 are
collinear then we just remove one of the points and reduce the number of ver-
tices. Suppose that this is not the case. Consider the ray with vertex at A1 and
containing A1 A2 and start to rotate it in the direction of the vertex A3 . We stop
when we reach A3 (then As = A3 ) or at the moment when the ray contains
some vertex As at the same half-plane as A1 with respect to the line A2 A3
for the first time. Now we decompose the polygon A1 . . . Ak into polygons
A1 . . . As and A1 As . . . Ak . Both of these polygons have fewer than k vertices.
By the induction assumption each of these polygons admits a decomposition
into empty triangles. Hence A1 . . . Ak is decomposable as well.
This concludes the proof of (i).
(ii) Let a polygon P have two decompositions, one into n1 empty triangles and the
other into n2 empty triangles. Then by Proposition 2.11 we have
n1 n2
= S(P ) = ,
2 2
28 2 On Integer Geometry

Fig. 2.3 Integer area is

additive

since Euclidean area is additive. Hence, n1 = n2 .

(iii) Consider an integer triangle ABC. Define

Par(AB, AC) = {A + αv + βw|0 ≤ α, β < 1}.

Denote by #(ABC) the number of empty triangles in the decomposition (by

(i) and (ii) this is well defined). Let us prove that if lS(ABC) = n, then
#(ABC) = n, by induction on n.
Base of induction. Suppose that lS(ABC) = 1. Then by Proposition 2.11
we have #(ABC) = 1.
Induction step. Suppose that the statement holds for every k  < k, for k > 1.
Let us prove it for k. Consider a triangle ABC such that lS(ABC) = k.
By Proposition 2.2, there exists an integer point P in the parallelogram
Par(AB, AC) distinct from the vertices. Without loss of generality, we suppose
that P is not on the edge AB. Set Q = P + AB and D = C + AB.
The triangle BQD is obtained from APC by shifting by the integer
vector AB (see Fig. 2.3). Hence the total number of orbits of integer points
with respect to the shift on vector multiples of AB in the parallelograms
Par(AB, AP) and Par(PQ, PC) equals the number of such orbits in the par-
allelogram Par(AB, AC). Therefore, from Proposition 2.2, we get

lS(ABC) = lS(APB) + lS(CPD).

By Proposition 2.11 we have

#(ABC) = 2S(ABC) = S(ABCD) = S(ABQP) + S(CPQD)

= 2S(ABP) + 2S(CPD) = #(APB) + #(CPD).

Since P is not on AB, we have #(APB) < k and #(CPD) < k. Therefore,
by the inductive hypothesis, we get

lS(ABC) = lS(APB) + lS(CPD) = #(APB) + #(CPD)

= #(ABC).

The proof is completed by induction. 

Corollary 2.15 The integer area of polygons in the plane is twice the Euclidean
area.
2.4 Pick’s Formula 29

Proof The statement holds for empty triangles by Proposition 2.11. Now the corol-
lary follows directly from Proposition 2.13 and the definition of integer area. 

Remark 2.16 Corollary 2.15 implies that the index of an integer sublattice generated
by vectors AB and AC in the integer lattice equals
 
det(AB, AC).

2.4 Pick’s Formula

We conclude this section with a nice formula that describes a relation between inte-
ger points in a polygon and its Euclidean area.

Theorem 2.17 (Pick’s formula) The Euclidean area S of an integer polygon satis-
fies the following relation:
S = I + E/2 − 1,
where I is the number of integer points in the interior of the polygon and E is the
number of integer points in the edges.

For the integer area one should multiply the right part of the formula by two.

Example 2.18 For instance, for the following pentagon

the number of inner integer points equals 4, the number of integer points on the
edges equals 6, and the area equals 6. So,

6 = 4 + 6/2 − 1.

Proof of Theorem 2.17 We prove Pick’s formula by the induction on the area.
Base of induction. If the Euclidean area of an integer polygon is 1/2, then by
Proposition 2.11 the polygon is an empty triangle. For the empty triangle we have

S = 0 + 3/2 − 1 = 1/2.

Induction step. Fix k. Suppose that the statement holds for every integer polygon
of area k  /2, where k  < k. Let us prove the statement for polygons of area k/2. Let
P be an integer polygon of area k/2. Then it can be decomposed into two integer
polygons P1 and P2 (for instance as in the proof of Proposition 2.13(i)) intersecting
30 2 On Integer Geometry

in an integer segment Ai Aj . Suppose that Pi has Ii interior points and Ei boundary

points. Let Ai Aj contain Ê integer points. Since the areas of P1 and P2 are less than
k/2, by the induction assumption induction we have

S(P ) = S(P1 ) + S(P2 ) = I1 + E1 /2 − 1 + I2 + E2 /2 − 1

= (I1 + I2 + Ê) + (E1 + E2 − 2Ê − 2)/2 − 1.

Since the number of inner integer points of P is I1 + I2 + Ê and the number of

boundary integer points is E1 + E2 − 2Ê − 2, Pick’s formula holds for P . This
concludes the induction step. 

We have formulated Pick’s theorem traditionally in Euclidean geometry. It is

clear that this theorem is not true for all lattices. Still, the lattice analogue of the
theorem holds for all lattices.

2.5 The Twelve-Point Theorem

Let us also mention here an interesting theorem coming from the theory of toric
varieties (see [59] and [106]; several interesting proofs of this theorem can be found
in [168] and [171]).
First we give a definition of a dual polygon.

Definition 2.19 Let P = A1 . . . An be a convex integer polygon. Suppose that O is

the only integer point in the interior of P . Draw the integer vectors OB1 , . . . , OBn
of unit integer length and parallel to A1 A2 , . . . , An A1 respectively. The polygon
P ∗ = B1 . . . Bn is said to be dual to P .

The following statement holds.

Theorem 2.20 (The Twelve-Point Theorem) Suppose that P is a convex integer

polygon containing a single integer point in its interior. Denote by E and E ∗ the
number of integer points in the boundary of P and P ∗ . Then we have

E + E ∗ = 12.

We leave the proof of this theorem as an exercise for the reader.

2.6 Exercises
Exercise 2.1 For any triangle ABC, show that

lS(ABC) = l(AB) ld(C, AB).

2.6 Exercises 31

Exercise 2.2 Let 1 , 2 , and 3 be three parallel integer lines and let Ai ∈ li be
integer points. Suppose also that the lines 1 and 3 are in different half-planes with
respect to the line 2 . Prove that

ld(A1 , 3 ) = ld(A1 , 2 ) + ld(A2 , 3 ).

Exercise 2.3 (Geometric interpretation of integer distance) Let ld(O, AB) = k.

Prove that there are exactly k − 1 integer lines parallel to AB such that the point
O and the segment AB are in different half-planes with respect to these lines.

Exercise 2.4 Find an example of two integer noncongruent triangles that have the
same integer area.

Exercise 2.5 Consider an empty tetrahedron in R3 with vertices in Z3 . Suppose

that it does not contain other points of the lattice Z3 . Is it true that the edges of the
tetrahedron generate the whole lattice Z3 ?

Exercise 2.6 For any triangle ABC, show that

lS(ABC) = l(AB) l(AC) lα(∠ABC).

Exercise 2.7 Prove that the index of a subgroup generated by v and w in the integer
lattice is the absolute value of det(v, w).

Exercise 2.8 Find a proof of the twelve-point theorem.

Chapter 3
Geometry of Regular Continued Fractions

Continued fractions play an important role in the geometry of numbers. In this chap-
ter we describe a classical geometric interpretation of regular continued fractions in
terms of integer lengths of edges and indices of angles for the boundaries of convex
hulls of all integer points inside certain angles. In the next chapter we will extend
this construction to construct a complete invariant of integer angles. For the geom-
etry of continued fractions with arbitrary elements see Chap. 11.

3.1 Classical Construction

Let us start with the classical geometric construction of continued fractions.
Consider an arbitrary number α ≥ 1. Denote a ray {y = αx | x ≥ 0} by rα . This
ray divides the first octant {(x, y)|x, y ≥ 0} into two angles. Consider one of them
and take the convex hull of all the integer points except the origin inside this angle.
The boundary of the convex hull is a broken line, which is called the sail of the
angle. The same construction is applied to the second angle. See Fig. 3.1 for an
example when α = 7/5.
If α is a rational number, then both sails of the angles consist of finitely many
segments and two rays, as in the example of α = 7/5. Denote the vertices of the
broken line in the sail of the angle containing (1, 0) by A0 , . . . , An starting with
A0 = (1, 0). In the same way we denote the vertices of the broken line in the sail of
the angle containing (0, 1) by B0 , . . . , Bm , starting with B0 = (0, 1). We call these
broken lines the principal part of the sails. As we will show later, m = n − 1 or
m = n.
In the case of an irrational α, each of the sails is a union of one ray and an infinite
broken line. Denote the broken line starting from (1, 0) by A0 A1 . . . , and the broken
line starting from (0, 1) by B0 B1 . . . respectively, and call them principal parts of
the sail.

O. Karpenkov, Geometry of Continued Fractions, 33

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_3, © Springer-Verlag Berlin Heidelberg 2013
34 3 Geometry of Regular Continued Fractions

Fig. 3.1 The sails for 7/5

It is interesting to note that in the case of α = 7/5, shown in Fig. 3.1, the ratios
of the coordinates of extremal points of the convex hulls

A1 = (1, 1), B1 = (3, 2), B2 = A2 = (7, 5)

are convergents to the number 7/5. This does not happen by chance for 7/5; such
a situation is typical. In next theorem we show a general relation between Ai and
Bi and the convergents for an arbitrary real number α ≥ 1 (we discuss the case of
0 ≤ α < 1 later, in Theorem 3.4).

Theorem 3.1 Consider α ≥ 1. Let A0 A1 A2 . . . and B0 B1 B2 . . . be the principal

parts of the corresponding sails (finite or infinite). Then

Ai = (p2i−2 , q2i−2 ) and Bi = (p2i−1 , q2i−1 ), i = 1, 2, . . . ,

where pk /qk are convergents. The only exception is for the rational case for the
following reason: the last vertices of the principal parts for both sails coincide with
(pn , qn ), where pn /qn is the last convergent, i.e., α = pn /qn .

We start the proof with the following lemma.

Lemma 3.2
(a) The segment with endpoints (p2k−2 , q2k−2 ) and (p2k , q2k ) is in the sail
A0 A1 A2 . . . .
(b) The segment with endpoints (p2k−1 , q2k−1 ) and (p2k+1 , q2k+1 ) is in the sail
B0 B1 B2 . . . .
3.1 Classical Construction 35

Proof We start with item (a). First, notice that both points (p2k−2 , q2k−2 ) and
(p2k , q2k ) are in the convex hull of the sail including A0 A1 A2 . . . , since
p2k−2 p2k
α> and α > .
q2k−2 q2k

Consider the line l passing through the points (p2k−2 , q2k−2 ) and (p2k , q2k ).
From Proposition 1.13 we know that all integer points on l are as follows:

(p2k−2 , q2k−2 ) + λ(p2k−1 , q2k−1 ), λ ∈ Z.

Let us prove that the line l is at unit integer distance to the origin. The inte-
ger distance is equivalent to the index of the sublattice generated by the vectors
(p2k−2 , q2k−2 ) and (p2k−1 , q2k−1 ). From Proposition 1.15 it follows that

|p2k−2 q2k−1 − p2k−1 q2k−2 | = 1,

i.e., the Euclidean area of the corresponding triangle is 1/2. Hence by Proposi-
tion 2.11 these vectors generate the lattice, and ld((0, 0), l) = 1.
Therefore, there is no integer point in the interior of the band between l and the
line parallel to l and passing through the origin.
It is clear that the point (p2k−2 , q2k−2 ) − (p2k−1 , q2k−1 ) has nonpositive coordi-
nates, and therefore it is not in the cone.
Let us show that the point

(p2k−2 , q2k−2 ) + (a2k + 1)(p2k−1 , q2k−1 ) = (p2k , q2k ) + (p2k−1 , q2k−1 )

is not in the cone. Notice that the point (p2k , q2k ) + (p2k−1 , q2k−1 ) belongs to the
segment with endpoints

(p2k−1 , q2k−1 ) and (p2k+1 , q2k+1 ) = (p2k−1 , q2k−1 ) + a2k+1 (p2k , q2k )

which is contained in the other cone of the octant.

Therefore, the segment with endpoints (p2k−2 , q2k−2 ) and (p2k , q2k ) is contained
in the sail A0 A1 A2 . . . (see Fig. 3.2).
The proof for the second item is similar. 

Proof of Theorem 3.1 Actually, Lemma 3.2 almost proves the theorem. We have to
check only the endpoints of broken lines.
First, note that A1 = (α, 1), i.e., A1 = (p0 , q0 ).
Second, (if α ∈/ Z), we have
 
B1 = (0, 1) + a2 α, 1 = (p1 , q1 ).

Finally, we have to check in the case of rational α whether the segment with
endpoints (pn−1 , qn−1 ) and (pn , qn ) for the last two convergents is in one of the
two sails. The point (pn−1 , qn−1 ) is in one of the two sails by Lemma 3.2. The
36 3 Geometry of Regular Continued Fractions

Fig. 3.2 The segment with

endpoints (p2k−2 , q2k−2 ) and
(p2k , q2k ) is in the sail

point (pn , qn ) is in the intersection of the sails. The last thing we have to show
is that the triangle with vertices (0, 0), (pn−1 , qn−1 ), and (pn , qn ) is empty. From
Proposition 1.15 it follows that the Euclidean area of the triangle is 1/2; hence
by Proposition 2.11 the triangle is empty. Therefore, the segment with endpoints
(pn−1 , qn−1 ) and (pn , qn ) is in the sail.
We have found all the segments of the principal parts of both sails, concluding
the proof. 

Remark 3.3 In the case of rational α with the principle parts of the corresponding
angles A1 . . . An and B1 . . . Bm , we have the following. If the last element of the odd
continued fraction is 1, then n = m + 1, otherwise n = m.

Let us formulate a similar theorem for the case 0 < α < 1.

Theorem 3.4 Consider 0 < α < 1. Let A0 A1 A2 . . . and B0 B1 B2 . . . be the principal

parts of the corresponding sails (finite or infinite). Then

Ai = (p2i , q2i ) and Bi = (p2i−1 , q2i−1 ), i = 1, 2, . . .

where pk /qk are convergents. The only exception is when α is rational, in which
case the last vertex of the principal parts for both sails coincide with (pn , qn ), where
pn /qn is the last convergent, i.e. α = pn /qn .

Proof The proof repeats the proof of Theorem 3.1 except for the following differ-
ence. Since a0 = 0, the point A0 should coincide with A1 . This is the explanation
for the shift in indices. 
3.2 Geometric Interpretation of the Elements of Continued Fractions 37

3.2 Geometric Interpretation of the Elements of Continued

Fractions

Let us first formulate a corollary to Theorem 3.1.

Corollary 3.5 Consider α ≥ 1. Let A0 A1 A2 . . . and B0 B1 B2 . . . be the principal

parts of the corresponding sails (finite or infinite). Then

l(Ai Ai+1 ) = a2i and l(Bi Bi+1 ) = a2i+1 , i = 0, 1, 2, . . . ,

where α = [a0 ; a1 : a2 : · · · ]. The only exception occurs when α is rational, in which

case the last edges of the principal parts for both sails coincide with (pn , qn ), where
pn /qn is the last convergent, i.e., α = pn /qn . The integer length of this edge is 1.

In the case of 0 < α < 1, the corollary also holds, the only difference being that
l(Ai Ai+1 ) = a2i+2 instead a2i .

Proof of Corollary 3.5 The corollary follows directly from the explicit formulas for
Ak and Bk of Theorem 3.1, after applying Proposition 1.13. 

Theorem 3.1 and Corollary 3.5 lead to an interesting algorithm for constructing
sails for regular continued fractions.

Algorithm to Construct Sails of Real Numbers

Input data. Given a regular continued fraction α = [a0 ; a1 : a2 : · · · ] with a0 > 0.
Set O = (0, 0).
Goal of the algorithm. To construct the sails for α.
Step 1. Assign A0 = (1, 0), B0 = (0, 1) and construct

A1 = A0 + a0 OB0 and further B1 = B0 + a1 OA1 .

Inductive Step k. Suppose that we have already constructed A0 . . . Ak and

B0 . . . Bk . Then put

Ak+1 = Ak + ak OBk and Bk+1 = Bk + ak+1 OAk+1 .

Output. If α is rational, then the algorithm constructs both sails in finite time. If α
is irrational, then the algorithm also calculates both sails but in infinitely many
steps.
We leave the case 0 < α < 1 as an easy exercise for the reader.
In Fig. 3.3 we show the steps of the algorithm for the particular case α = [1; 2 : 2].
38 3 Geometry of Regular Continued Fractions

Fig. 3.3 Construction of the continued fraction for 7/5

3.3 Index of an Angle, Duality of Sails

In this subsection we show that there exists a duality between edges and angles. An
important consequence of this duality is that all the elements of the regular continued
fraction can be read from one of the sails. We restrict ourselves to the case α ≥ 1.
Reduction to the case 0 < α < 1 is straightforward, so we omit it.

Theorem 3.6 Consider α ≥ 1. Let A0 A1 A2 . . . and B0 B1 B2 . . . be the principal

parts of the corresponding sails (finite or infinite). Then
lα(∠Ai Ai+1 Ai+2 ) = a2i+1 and lα(∠Bi Bi+1 Bi+2 ) = a2i+2
for all admissible indices, where α = [a0 ; a1 : a2 : · · · ].

Proof Let us calculate the index of an integer angle at some vertex of the principal
part of one of the two sails. By Theorem 3.1 it is equivalent to calculate the index of
the angle between a pair of vectors (pi−1 , qi−1 ) and (pi+1 , qi+1 ):
 
lα ∠(pi−1 , qi−1 )(0, 0)(pi+1 , qi+1 )
 
= |pi−1 qi+1 − pi+1 qi−1 | = pi−1 (ai qi + qi−1 ) − (ai pi + pi−1 )qi−1 
= ai |pi−1 qi − pi qi−1 | = ai .
The second equality follows from Proposition 1.13, while the last follows from
Proposition 1.15. 

Edge–Angle Duality From Corollary 3.5 and Theorem 3.6, we get that
lα(∠Ai Ai+1 Ai+2 ) = l(Bi Bi+1 ) and lα(∠Bi Bi+1 Bi+2 ) = l(Ai+1 Ai+2 ).
For the example of α = 7/5 (see Figs. 3.1 and 3.4), we get
lα(∠A0 A1 A2 ) = l(B0 B1 ) = a1 = 2;
lα(∠B0 B1 B2 ) = l(A1 A2 ) = a2 = 2.
3.4 Exercises 39

Fig. 3.4 The edge–angle

duality for 7/5

3.4 Exercises

Exercise 3.1 Prove that every sail is homeomorphic to R1 .

Exercise 3.2 Describe the edge–angle duality for the case 0 < α < 1.

Exercise 3.3 Construct the sails of continued fractions for rational numbers 3/2,
10/7, 18/13, and 13/18.

Exercise 3.4 Find geometrically all the convergents to the rational numbers 8/5,
9/4, and 17/8.

Exercise 3.5 Is it sufficient to know all indices for the angles in the principal parts
of the sail defined by α > 1 and the indices of angles of its dual sail to reconstruct
α in a unique way? The question is interesting for both of the following cases: if α
is rational and if α is irrational.
Chapter 4
Complete Invariant of Integer Angles

In this chapter we generalize the classical geometric interpretation of regular con-

tinued fractions presented in Chap. 3 to the case of arbitrary integer angles, con-
structing a certain integer broken line called the sail of an angle. We combine the
integer invariants of a sail into a sequence of positive integers called an LLS se-
quence. From one side, the notion of LLS sequence extends the notion of contin-
ued fraction (see Remark 4.7), about which we will say more in the next chap-
ter. From another side, LLS sequences distinguish the integer angles. Sails and
LLS sequences of angles play a central role in the geometry of numbers. In par-
ticular, we use LLS sequences in integer trigonometry and its relations to toric
singularities and in Gauss’s reduction theory. F. Klein generalized the notion of
sail to the multidimensional case to study integer solutions of homogeneous de-
composable forms. We will study this generalization in the second part of this
book.
In Sect. 4.1 we give a preliminary definition of an integer sine. Further, in
Sect. 4.2, we define sails of integer angles and corresponding LLS sequences. Fi-
nally, in Sect. 4.3 we prove that an LLS sequence is a complete invariant of an
integer angle.

4.1 Integer Sines of Rational Angles

Recall the following general definitions. An angle is integer if its vertex is an integer
point. If the integer angle ∠ABC has integer points distinct from B on both edges
AB and BC, we call it rational.
Let us introduce the notation of the integer sine function for rational angles. We
put in the definition an integer analogue of the sine formula of the Euclidean case.
In this section we consider an integer angle ∠ABC with integer vertex B.

O. Karpenkov, Geometry of Continued Fractions, 41

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_4, © Springer-Verlag Berlin Heidelberg 2013
42 4 Complete Invariant of Integer Angles

Definition 4.1 Let ∠ABC be a rational angle, where A and C are lattice points
distinct from B. The integer sine of the angle is the following number:

lS(∠ABC)
,
l(AB) l(BC)

denoted by lsin ∠ABC.

Remark 4.2 Notice that the integer sine is well defined, i.e., it does not depend on
the choice of points B and C. We leave the reader to check this as an exercise.

Let us briefly compare sin and lsin. One difference is in the multiplicative con-
stant 1/2. This is due to the fact that an empty triangle has integer area 1 but Eu-
clidean area 1/2, as shown in Proposition 2.11. Yet the difference between sin and
lsin is much stronger, as illustrated in the following proposition.

Proposition 4.3 The integer sine of a rational angle coincides with the index of the
angle.

In particular, this implies that the integer sine takes all possible nonnegative in-
teger values.

Proof of Proposition 4.3 Consider a rational angle ∠ABC with A, B, C not con-
tained on one line. Let A and C  be the nearest integer points to B in the open rays
BA and BC, respectively. Then from the definition of integer length, we get
   
l BA = l BC = 1.

Hence,
   
lsin ∠AB C = lS ∠AB C .

The integer area of AB C is the index of the sublattice generated by BA and BC
in Z2 . Since the vectors BA and BC generate all integer points of the lines AB and
AC, the integer area of AB C is equivalent to the index of the angle. Therefore,
we get

lsin(∠ABC) = lα(∠ABC).

If A, B, and C are on one line, then

lsin(∠ABC) = lα(∠ABC) = 0.

This concludes the proof. 

4.2 Sails for Arbitrary Angles and Their LLS Sequences 43

4.2 Sails for Arbitrary Angles and Their LLS Sequences

In Chap. 3 we studied sails for a certain subset of integer angles. Let us extend the
notion of sails to the case of arbitrary integer angles.

Definition 4.4 Let α be an arbitrary integer angle. Consider the convex hull of the
set of all integer points in α except the origin. The boundary of this convex hull is
called the sail of this angle.

Remark 4.5 Notice that the sail of a lattice angle is a broken line homeomorphic
to R. A sail is either a two-sided infinite broken line or a one-sided infinite broken
line including also a ray at one side or a finite broken line that includes two rays
at both sides. These rays appear when edges of angles contain integer points other
than the vertex of the corresponding angle.

Let us now define a very important characteristic of the sail that we will use in
the future.

Definition 4.6 Consider an arbitrary angle with integer vertex. Let the sail for this
angle be a broken line with sequence of vertices (Ai ). Define:

a2k = l Ak Ak+1 ,
a2k−1 = lsin(∠Ak−1 Ak Ak+1 )

for all admissible indices. The lattice length sine sequence (LLS sequence for short)
for the sail is the sequence (an ).

The LLS sequence can be either finite or infinite on one or both sides.

Remark 4.7 Notice that LLS sequences contain only positive integer elements.
Consider a finite to the left LLS sequence. Regular continued fractions define a
one-to-one correspondence between such LLS sequences and real numbers not
smaller than 1; namely, the sequence (a0 , a1 , . . .) corresponds to the real number
[a0 ; a1 : · · · ]. This continued fraction is an important invariant of integer angles (we
will call it the integer tangent), and we will study this invariant in the next chapter.

4.3 On Complete Invariants of Angles with Integer Vertex

Proposition 4.8 Integer length, integer area, index and integer sine of a rational
angle are invariant under the action of the group of integer affine transformations
Aff(2, Z).
44 4 Complete Invariant of Integer Angles

Proof Every integer linear transformation sends subgroups to subgroups and the
corresponding cosets to the corresponding cosets; hence all integer indices are in-
variant under integer linear transformations. Thus, every integer affine transforma-
tion preserves indexes of the corresponding integer lattice subgroups in Z2 as well.
Therefore, the integer area and the index (and hence the integer sine) of a rational
angle are also preserved. The integer length is preserved, since all the inner integer
points map to inner integer points. 

Corollary 4.9 The LLS sequence is an invariant of lattice angles with respect to
Aff(2, Z).

Proof First, note that convex hulls are preserved by the elements of Aff(2, Z) (we
leave this as an easy exercise for the reader). Then the statement follows directly
from the fact that the integer length and the index are invariants of Aff(2, Z). 

It is interesting to note that the angles ∠ABC and ∠CBA are not necessarily inte-
ger congruent (i.e., there is no integer affine transformation taking one to another).
For example, if we consider

A = (3, −2), B = (0, 0), and C = (2, 1),

then ∠ABC and ∠CBA are not integer congruent. We will say more about such pairs
of angles in the next section.

Theorem 4.10 Two angles with vertices at integer points are lattice congruent if
and only if they have the same LLS sequences.

Proof From Corollary 4.9 we know that the LLS sequence is an integer invariant of
angles. It remains to prove that if two LLS sequences coincide, then the correspond-
ing angles are integer congruent.
Consider two angles α and β with sails (Ai ) and (Bi ). Let the corresponding
LLS sequences coincide and be equivalent to (ai ). Let the vertices of the angles α
and β be Oα and Oβ respectively. Consider an affine transformation taking Oβ to
Oα , B0 to A0 , and B1 to A1 . This affine transformation is integer, since l(A0 A1 ) =
l(B0 B1 ) = a0 and ld(Oα , A0 A1 ) = ld(Oβ , B0 B1 ). Suppose the angle β is taken to
some angle γ with sail (Ci ) (we already know that the vertex of γ is Oα , C0 = A0 ,
and C1 = A1 ).
Let us prove that the broken lines A0 A1 A2 . . . and C0 C1 C2 . . . coincide by in-
duction. Suppose A0 A1 . . . Ak−1 coincides with C0 C1 . . . Ck−1 . Let us prove that
Ak = Ck .
First, we know that

lsin(∠Ak−2 Ak−1 Ak ) = lsin(∠Ck−2 Ck−1 Ck ) = a2k−3

and
l(Ak−1 Ak ) = l(Ck−1 Ck ) = a2k−2 .
4.3 On Complete Invariants of Angles with Integer Vertex 45

Hence we have

ld(Ak , Ak−2 Ak−1 ) = ld(Ck , Ck−2 Ck−1 ) = a2k−2 a2k−3 .

This follows from the simple fact (see Exercise 4.6) that
lS(∠ABC)
ld(A, BC) = .
l(BC)
Notice that ∠Ak−2 Ak−1 Ak and ∠Ck−2 Ck−1 Ck are two parts of convex sails, and
hence the points Ak and Ck are on the other side of the point Oα with respect to
the line Ak−2 Ak−1 = Ck−2 Ck−1 . Hence Ak and Ck are both on a line l1 parallel to
the line Ak−2 Ak−1 containing points at integer distance a2k−2 a2k−3 from the line
Ak−2 Ak−1 .
Second, we have

ld(Ak , Oα Ak−1 ) = lS(∠Oα Ak−1 Ak ) = a2k−1

= lS(∠Oα Ck−1 Ck ) = ld(Ck , Oα Ck−1 ),

where Oα Ak−1 = Oα Ck−1 . Again, by convexity we know that the points Ak and
Ck are in a different half space from the point Ak−2 = Ck−2 with respect to the line
Oα Ak−1 . Therefore, Ck and Ak are on a line l2 parallel to Oα Ak−1 .
Since Oα Ak−1 and Ak−2 Ak−1 are not parallel, the intersection of lines l1 and l2
is a point coinciding with both Ak and Ck .
Therefore, the broken lines A0 A1 A2 . . . and C0 C1 C2 . . . coincide.
The fact that . . . A−1 A0 A1 and . . . C−1 C0 C1 coincide follows from the consid-
ered case after performing a GL(2, Z) transformation taking A0 to A1 and A1 to
A0 , namely the transformation

1 0
.
a0 −1

We apply this transformation to both sails (Ai ) and (Ci ) and get that the images of
Ck and Ak coincide for every negative k. Therefore, the whole sails (Ai ) and (Bi )
coincide. 

Theorem 4.11 For every sequence of positive integers (odd finite or infinite on one
or both sides) there exists an angle with vertex at the origin whose LLS sequence is
this sequence.

Proof First we consider the case of a sequence a0 , a1 , a2 , . . . (odd finite or infinite

to the right). Let A = (1, 0), B = (0, 0), and C = (1, α), where [a0 ; a1 : a2 : · · · ] is
the continued fraction for α. The angle ∠ABC has the desired LLS sequence. This
follows directly from Corollary 3.5 and Theorem 3.6.
When the sequence is infinite to the left, we construct the angle ∠ABC for the
inverse sequence. Then the angle ∠CBA is the angle with the prescribed LLS se-
quence.
46 4 Complete Invariant of Integer Angles

The remaining case is the case of both-side infinite sequences. Here we construct
the angle for the part with nonnegative indices. Then apply to this angle the trans-
formation we used in the proof of Theorem 4.10 and construct the angle for the
remaining part of the sail and get the angle with the prescribed LLS sequence. 

4.4 Exercises

Exercise 4.1 Show that the integer sine of a rational angle ∠ABC does not depend
on the choice of integer points B and C on the edges.

Exercise 4.2 Prove that convex sets are taken to convex sets and their boundaries
to boundaries under affine transformations.

Exercise 4.3 Let A = (3, −2), B = (0, 0), and C = (2, 1). Prove that ∠ABC and
∠CBA are not integer congruent.

Exercise 4.4 Prove that if the principal parts of the sails of two angles are integer
congruent, then the angles themselves are integer congruent.

Exercise 4.5 Let d be a positive integer and l an integer line. Prove that all lattice
points lying at integer distance are contained in two lines parallel to l at the same
Euclidean distance from l.

Exercise 4.6 Prove that

lS(∠ABC)
ld(A, BC) = .
l(BC)
Chapter 5
Integer Trigonometry for Integer Angles

In this chapter we explain how to interpret regular continued fractions related to

LLS sequences in terms of integer trigonometric functions. Integer trigonometry
has many similarities to Euclidean trigonometry (for instance, integer arctangents
coincide with real arctangents; the formulas for adjacent angles are similar). From
another point of view they are totally different, since integer sines and cosines are
positive integers; there are two right angles in integer trigonometry, etc. In this chap-
ter we discuss basic properties of integer trigonometry. In Chap. 6 we use integer
trigonometric functions to describe angles of integer triangles, which will further
result in global relations for toric singularities on toric surfaces (see Chap. 13).
For rational angles we introduce definitions of integer sines, cosines, and tan-
gents. In addition to rational integer angles, there are three types of irrational inte-
ger angles. If an integer angle ∠ABC has an integer point distinct from the vertex B
in AB but not in BC (in BC but not in AB), we call the angle R-irrational (or respec-
tively L-irrational). In case both edges of an angle do not contain lattice points other
than B, the angle is called LR-irrational. It is only for R-irrational angles that we
have a definition of integer tangents. The trigonometric functions are not defined for
L-irrational and LR-irrational angles. For more information on integer trigonometry
we refer to [94] and [96].

5.1 Definition of Trigonometric Functions

We start with the definition of the integer tangent for rational and R-irrational angles.

Definition 5.1 Consider a rational or R-irrational angle ∠ABC. Let A, B, and C

be noncollinear. Suppose the LLS sequence of ∠ABC is (a0 , a1 , a2 , . . .) (finite or
infinite). Then the integer tangent of the angle ∠ABC equals

ltan(∠ABC) = [a0 ; a1 : a2 : · · · ].

When the points A, B, and C are collinear (but the points A and C are distinct from
B) we say that ltan(∠ABC) = 0.

O. Karpenkov, Geometry of Continued Fractions, 47

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_5, © Springer-Verlag Berlin Heidelberg 2013
48 5 Integer Trigonometry for Integer Angles

Consider a rational angle ∠ABC and recall that

lS(ABC)
lsin(∠ABC) = .
l(AB) l(BC)

Definition 5.2 For a rational angle α we define

lsin α
lcos α = .
ltan α

It is clear that the integer sine, integer tangent, and therefore integer cosine are
invariants of Aff(2, Z).
Now we give the inverse function to the integer tangent.

Definition 5.3 Consider a real α ≥ 1 or α = 0. The integer arctangent of α is the

angle with vertex at the origin and edges

{y = 0 | x ≥ 0} and {y = αx | x ≥ 0}.

We define the zero angle as larctan 0. The angle π is the angle ABC, where A =
(1, 0), B = (0, 0), and C = (−1, 0).

5.2 Basic Properties of Integer Trigonometry

First we show that the integer tangent and arctangent are in fact inverse to each
other.

Proposition 5.4
(i) For every real s ≥ 1, we have ltan(larctan s) = s.
(ii) For every rational or R-irrational angle α, the following holds:

larctan(ltan α) ∼
= α.

Proof
(i) From Corollary 3.5 and Theorem 3.6 we have that the elements of the LLS se-
quence for larctan s coincide with the elements of the regular continued fraction
for s. Hence the statement holds by definition of the integer tangent.
(ii) Both angles larctan(ltan α) and α have the same LLS sequences. Therefore,
they are congruent by Theorem 4.10. 

In the following proposition we collect several trigonometric properties.

5.3 Transpose Integer Angles 49

Proposition 5.5
(i) The values of integer trigonometric functions for integer congruent angles co-
incide.
(ii) The lattice sine and cosine of any rational angle are relatively prime positive
integers.
(iii) For every angle α the following inequalities hold:

lsin α ≥ lcos α and ltan α ≥ 1.

Equality holds if and only if the lattice vectors of the angle rays generate the
whole lattice.
(iv) (Description of lattice angles.) Two integer angles α and β are congruent if
and only if ltan α = ltan β.

Proof
(i) This is a direct corollary of the fact that integer sine and tangent are defined
only by values of certain indices.
(ii) By Proposition 5.4 it is enough to prove the assertion for angles larctan α for
rational α ≥ 1.
Consider m n ≥ 1, where m and n are relatively prime integers. Then the
sail of the angle will contain the point (n, m). It is also clear that the lat-
tice distance between the point (n, m) and the line y = 0 is m, and hence
n ) = m.
lsin(larctan m
Since ltan(larctan m n) = n n ) = m, we have
m
and lsin(larctan m
lcos(larctan n ) = n.
m

(iii) This is true for all integer arctangent angles. Therefore, it is true for all angles.
(iv) The LLS sequence is uniquely defined by the integer tangent. Therefore, the
statement follows from Theorem 4.10. 

5.3 Transpose Integer Angles

Let us give a definition of the integer angle transpose to a given one.

Definition 5.6 The integer angle ∠BOA is said to be transpose to the integer angle
∠AOB. We denote it by (∠AOB)t .

It immediately follows from the definition that for any integer angle α, we have
 t t
α ∼ = α.

Further, we will use the following notion. Suppose that some arbitrary integers
a, b, and c, where c ≥ 1, satisfy ab ≡ 1(mod c). Then we write
 −1
a ≡ b(mod c) .
50 5 Integer Trigonometry for Integer Angles

For the trigonometric functions of transpose integer angles the following rela-
tions hold.

Theorem 5.7 (Trigonometric relations for transpose angles) Let an integer angle α
be not contained in a line. Then
(1) If α ∼
= larctan(1), then α t ∼
= larctan(1).
(2) If α  larctan(1), then
     −1
lsin α t = lsin α, lcos α t ≡ lcos α(mod lsin α) .

Proof Consider an integer angle α. Let ltan α = p/q, where gcd(p, q) = 1. By

Proposition 5.4(ii), we have α ∼ = larctan(p/q). The case p/q = 1 is trivial. Con-
sider the case p/q > 1.
Let A = (1, 0), B = (p, q), and O = (0, 0). Suppose that an integer point
C = (q  , p  ) of the sail of the angle larctan(p/q) is the closest integer point to the
endpoint B. Both coordinates of C are positive integers, since p/q > 1. Since the
triangle BOC is empty and the orientation of the pair of vectors (OB, OC) does
not coincide with the orientation of the pair of vectors (OA, OB), we have

p p
det = −1.
q q

Consider a linear transformation ξ of the two-dimensional plane,

p − p q  − q
ξ= .
p −q

Since det(ξ ) = −1, the transformation ξ is integer-linear and changes the ori-
entation. Direct calculations show that the transformation ξ takes the angle
larctant (p/q) to the angle larctan(p/(p − p  )). (See the example in Fig. 5.1.)
Since gcd(p, p  ) = 1 and p > p − p  , the following holds:

lsin(α t ) = p,
lcos(α t ) = p − p  .

Since pq  − qp  = −1, we have qp  ≡ −1(mod p). Therefore,

   
lcos α lcos α t = q p − p  ≡ 1(mod p).

From that we have


lsin(α t ) = lsin α,
lcos(α t ) ≡ (lcos α(mod lsin α))−1 .

This concludes the proof of Theorem 5.7. 

5.4 Adjacent Integer Angles 51

Fig. 5.1 If an integer angle α is integer-congruent to larctan(7/5), then α t is integer-congruent to

larctan(7/3)

5.4 Adjacent Integer Angles

Now we define an integer angle adjacent to a given one.

Definition 5.8 An integer angle ∠BOA is said to be adjacent to an integer angle

∠AOB if the points A, O, and A are contained in the same straight line. We denote
the angle ∠BOA by π − ∠AOB.

For the trigonometric functions of adjacent integer angles the following relations
hold.

Theorem 5.9 (Trigonometric relations for adjacent angles) Let α be some integer
angle. Then one of the following holds:
(1) If α is the zero angle, then π − α = ∼ π.
(2) If α is the straight angle, then π − α ∼ = 0.
(3) If α ∼
= larctan(1), then π − α ∼ = larctan(1).
(4) If α is neither zero nor straight nor integer-congruent to larctan(1), then

ltan α
π −α∼ = larctant
,
ltan(α) − 1
 −1
lsin(π − α) = lsin α, lcos(π − α) ≡ − lcos α(mod lsin α) .
52 5 Integer Trigonometry for Integer Angles

Remark 5.10 Suppose that an integer angle α is neither zero nor straight. Then the
conditions

lcos(π − α) ≡ (− lcos α(mod lsin α))−1 ,
0 < lcos(π − α) ≤ lsin α,
uniquely determine the value lcos(π − α).

Proof of Theorem 5.9 Consider an integer angle α. Directly from the definitions it
follows that if α ∼
= 0, then π − α ∼
= π , and, if α ∼
= π , then π − α ∼
= 0.
Suppose that ltan α = p/q > 0, where gcd(p, q) = 1. Then by Proposition 5.4(ii)
we have α ∼= larctan(p/q). Therefore,

π −α∼
= π − larctan(p/q).

It follows immediately that if p/q = 1, then π − α ∼

= larctan(1).
Let now α  larctan(1), and hence p/q > 1. Consider a linear transformation ξ1
of the two-dimensional plane,

−1 1
ξ1 = .
0 1

Since det(ξ1 ) = −1, the transformation ξ1 is integer-linear and changes the ori-
entation. Direct calculations show that the transformation ξ1 takes the cone cor-
responding to the angle π − larctan(p/q) to the cone corresponding to the an-
gle larctant (p/(p − q)). Since ξ1 changes the orientation, we have to transpose
larctant (p/(p − q)). (See the example in Fig. 5.2.) Therefore,

ltan α
π −α∼ = larctant
.
ltan(α) − 1

Now we show that


lsin(π − α) = lsin α,
lcos(π − α) ≡ (− lcos α(mod lsin α))−1 .

Let A = (1, 0), B = (q, p), and O = (0, 0). Consider the sail of the integer angle
π − larctan(p/q), which is integer-congruent to π − α. Suppose that an integer
point C = (q  , p  ) of the sail for π − larctan(p/q) is the closest integer point to the
endpoint of the sail B (or equivalently, the segment BC is in the sail and has the
unit integer length). The coordinate p  is positive, since p/q > 1. Since the triangle
BOC is empty and the vectors OB and OC define the same orientation as OA
and OB,

p p
det = 1.
q q
Consider a linear transformation ξ2 of the two-dimensional plane,
5.4 Adjacent Integer Angles 53

Fig. 5.2 If an integer angle α is integer-congruent to larctan(7/5), then π − α is integer-congruent

to larctan(7/4)

p − p q − q
ξ2 = .
−p q

Since det(ξ2 ) = 1, the transformation ξ2 is integer-linear and orientation-preserving.

Direct calculations show that the transformation ξ2 takes the angle π − larctan(p/q)
to the angle larctan(p/(p − p  )). Since gcd(p, p  ) = 1, we have lsin(π − α) = p.
Since gcd(p, p  ) = 1 and p > p − p  , the following holds:

p
lcos(π − α) = lcos larctan = p − p .
p − p

Since pq  − qp  = 1, we have qp  ≡ 1(mod p). Therefore,

 
lcos α lcos(π − α) = q p − p  ≡ −1(mod p).

From that we have


lsin(π − α) = lsin α,
lcos(π − α) ≡ (− lcos α(mod lsin α))−1 .

This concludes the proof of Theorem 5.9. 

54 5 Integer Trigonometry for Integer Angles

Remark 5.11 The statements of Theorem 5.7 and Theorem 5.9 were first intro-
duced in terms of regular continued fractions and so-called zig-zags by P. Popescu-
Pampu [169].

The following statement is an easy corollary of Theorem 5.9.

Corollary 5.12 For every integer angle α, the following holds:

π − (π − α) ∼
= α.

5.5 Right Integer Angles

We define right integer angles by analogy with Euclidean angles, using their sym-
metric properties.

Definition 5.13 An integer angle is said to be right if it is self-dual and integer-

congruent to the adjacent angle.

It turns out that in integer geometry there exist exactly two integer inequivalent
right integer angles.

Proposition 5.14 Every integer right angle is integer-congruent to exactly one of

the following two angles: larctan(1), larctan(2).

Proof Let α be an integer right angle.

Since (π − 0)  0 and (π − π)  π , we have ltan α > 0.
By the definition of integer right angles and Theorem 5.7, we obtain
   −1
lcos(α) ≡ lcos α t ≡ lcos α(mod lsin α) .

By the definition of integer right angles and Theorem 5.9, we obtain

 −1
lcos(α) ≡ lcos(π − α) ≡ − lcos α(mod lsin α) .

Hence,
 −1  −1
lcos α(mod lsin α) ≡ − lcos α(mod lsin α) .

Therefore, lsin α = 1 or lsin α = 2.

The integer angles with lsin α = 1 are integer-congruent to larctan(1). The integer
angles with lsin α = 2 are integer-congruent to larctan(2). The proof is complete. 
5.6 Opposite Interior Angles 55

5.6 Opposite Interior Angles

First we give the definition of parallel lines. Two integer lines are said to be parallel
if there exists an integer shift of the plane by an integer vector taking the first line to
the second.

Definition 5.15 Consider two distinct integer parallel lines AB and CD, where A,
B, C, and D are integer points. Let the points A and D be in different open half-
planes with respect to the line BC. Then the integer angle ∠ABC is called opposite
interior to the integer angle ∠DCB.

We have the following proposition on opposite interior integer angles.

Proposition 5.16 Two integer angles opposite interior to each other are integer-
congruent.

Proof Consider two distinct integer parallel lines AB and CD. Let the points A and
D be in distinct open half-planes with respect to the line BC. Let us prove that
∠ABC ∼ = ∠DCB.
Consider the central symmetry S of the two-dimensional plane at the midpoint
of the segment BC. Let P be an arbitrary integer point. Note that S(P ) = C + PB.
Since the point C and the vector PB are both integer, S(P ) is also integer. Since the
central symmetry is self-inverse, the inverse map S −1 also takes the integer lattice
to itself.
Therefore, the central symmetry S is an integer-affine transformation. Since S
takes the angle ∠ABC to the angle ∠DCB, we obtain ∠ABC ∼ = ∠DCB. 

5.7 Exercises
Exercise 5.1 Find both algebraically and geometrically the integer trigonometric
functions for the adjacent and transpose angles to the angles 73 , larctan 94 , larctan 13
5 .

Exercise 5.2 Let α be an arbitrary integer angle. Find expressions in the trigono-
metric functions for the angles π − α t and (π − α)t . Are these two angles integer
congruent?

Exercise 5.3 Can an integer triangle have two integer right angles? What about
three integer right angles?
Chapter 6
Integer Angles of Integer Triangles

In this chapter we study integer triangles, based on the results from previous chap-
ters. We start with the sine formula for integer triangles. Then we introduce integer
analogues of classical Euclidean criteria for congruence for triangles and present
several examples. Further, we verify which triples of angles can be taken as angles
of an integer triangle; this generalizes the Euclidean condition α + β + γ = π for
the angles of a triangle (this formula will be used in Chap. 13 to study toric sin-
gularities). Then we exhibit trigonometric relations for angles of integer triangles.
Finally, we give examples of integer triangles with small area.

6.1 Integer Sine Formula

In this section we give the integer sine formula for angles and edges of integer
triangles.
Let A, B, C be three distinct and noncollinear integer points. We denote the
integer triangle with vertices A, B, and C by ABC.

Proposition 6.1 (The sine formula for integer triangles) For any integer triangle
ABC the following holds:
l(AB) l(BC) l(CA) l(AB) l(BC) l(CA)
= = = .
lsin(∠BCA) lsin(∠CAB) lsin(∠ABC) lS(ABC)

Proof We have
lS(ABC) = l(AB) l(AC) lsin(∠CAB) = l(BA) l(BC) lsin(∠BCA)
= l(CB) l(CA) lsin(∠ABC).
After inverting the expressions and multiplying by all three integer lengths, we get
the statement of the proposition. 

The following is an open problem.

O. Karpenkov, Geometry of Continued Fractions, 57

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_6, © Springer-Verlag Berlin Heidelberg 2013
58 6 Integer Angles of Integer Triangles

Fig. 6.1 A counterexample to the second criterion of integer congruence for triangles

Problem 1 Find an integer analogue of the cosine formula for triangles.

Recall that the cosine formula for Euclidean triangle ABC with a = |BC|, b =
|AC|, c = |AB|, and α = ∠BAC is
a 2 = b2 + c2 − 2bc cos α.
It seems that to write an integer cosine formula is a hard task. Currently there are no
theorems related to addition in integer trigonometry.

6.2 On Integer Congruence Criteria for Triangles

We start with integer analogues for the three Euclidean criteria of triangle congru-
ence. It turns out that only the first criterion is valid in integer geometry. Further, we
present an additional criterion of congruence: on three angles and the area.

Proposition 6.2 (The first criterion of integer triangle integer congruence) Con-
sider integer triangles ABC and A B C . Suppose that
AB ∼= A B , AC ∼= A C , and ∠CAB ∼ = ∠C A B .
Then A B C ∼
= ABC.

Proof By definition there exists an integer affine transformation taking ∠CAB to

∠C A B . Since the integer lengths of the corresponding segments are the same, the
transformation takes B and C to B  and C  . 

It turns out that the literal generalizations of the second and third criteria from
Euclidean geometry to integer geometry do not hold. The following two examples
illustrate these phenomena.

Example 6.3 (The second criterion of triangle integer congruence does not hold in
integer geometry) In Fig. 6.1 we show two integer triangles ABC and A B C .
We have
AB ∼
= A B , ∠ABC ∼ = ∠A B C ∼
= larctan(1), and
= ∠C A B ∼
∠CAB ∼ = larctan(1).
The triangle A B C is not integer congruent to the triangle ABC, since
 
lS(ABC) = 4 and lS A B C = 8.
6.3 On Sums of Angles in Triangles 59

Fig. 6.2 A counterexample to the third criterion of integer congruence for triangles

Fig. 6.3 The additional criterion of integer congruence is not improvable

Example 6.4 (The third criterion of triangle integer congruence does not hold in
integer geometry) In Fig. 6.2 we show two integer triangles ABC and A B C .
All edges of both triangles are integer congruent (of length one), but the triangles
are not integer congruent, since lS(ABC) = 1 and lS(A B C ) = 3.

Instead of the second and the third criteria, we have the following additional
criterion.

Proposition 6.5 (An additional criterion of integer triangle integer congruence)

Consider two integer triangles ABC and A B C of the same integer area. Sup-
pose that
∠ABC ∼= ∠A B C , ∠CAB ∼ = ∠C A B , ∠BCA ∼ = ∠B C A .
Then A B C ∼
= ABC.

Proof All the integer lengths of the corresponding edges are the same by the sine
formula. Hence the triangles are integer congruent by the first criterion. 

In the following example we show that the additional criterion of integer triangle
integer congruence is not improvable.

Example 6.6 In Fig. 6.3 we give an example of two integer inequivalent triangles
ABC and A B C of the same integer area 4 and congruent angles ∠ABC, ∠CAB,
and ∠A B C , ∠C A B all integer-equivalent to the angle larctan(1), but ABC 
A B C .

6.3 On Sums of Angles in Triangles

In Euclidean geometry a triple of angles is a triple of angles in some triangle if and
only if their sum equals π . Let us introduce a generalization of this statement to the
case of integer geometry.
60 6 Integer Angles of Integer Triangles

First, we reformulate the Euclidean criterion in the form of tan functions. A tri-
angle with angles α, β, and γ exists if and only if

tan(α + β + γ ) = 0,
tan(α + β) ∈/ [0; tan α]
(without loss of generality, we suppose that α is acute). The next theorem is a trans-
lation of this condition into the integer case.
Second we give several preliminary definitions.
Let n be an arbitrary positive integer, and let A = (x, y) be an arbitrary integer
point. Denote by nA the point (nx, ny).

Definition 6.7 Consider an integer polygon or broken line with vertices A0 , . . . , Ak .

The polygon or broken line nA0 . . . nAk is called n-multiple (or multiple) to
A0 , . . . , Ak .

Let pi (for i = 1, . . . , k) be rational numbers and let [a1,i ; a2,i : · · · : ani ,i ] be

their odd continued fractions. Define
]p1 , . . . , pk [
= [a1,1 ;a2,1 : · · · : an1 ,1 : a1,2 : a2,2 : · · · : an2 ,2 : · · · : a1,k : a2,k : · · · : ank ,k ].
For instance,

3 7 31
, = [1; 1 : 1 : 1 : 2 : 2] = .
2 5 19
Now we are ready to formulate the generalization of the Euclidean theorem on
sum of angles in triangles.

Theorem 6.8 (On sums of integer tangents of angles in integer triangles)

(i) Let (α1 , α2 , α3 ) be an ordered triple of angles. There exists a triangle with con-
secutive angles integer congruent to α1 , α2 , and α3 if and only if there exists
i ∈ {1, 2, 3} such that the angles α = αi , β = αi+1(mod 3) , γ = αi+2(mod 3) satisfy
the following conditions:
(a) for ξ = ] ltan α, −1, ltan β[, the following holds: ξ < 0 or ξ > ltan α or
ξ = ∞;
(b) ] ltan α, −1, ltan β, −1, ltan γ [ = 0.
(ii) Let α, β, and γ be the consecutive angles of some integer triangle. Then
this triangle is multiple to the triangle with vertices A0 = (0, 0), B0 =
(λ2 lcos α, λ2 lsin α), and C0 = (λ1 , 0), where
lcm(lsin α, lsin β, lsin γ ) lcm(lsin α, lsin β, lsin γ )
λ1 = and λ2 = .
gcd(lsin α, lsin γ ) gcd(lsin α, lsin β)

We are not yet ready to prove the first statement of this theorem; we shall do
it later, in Chap. 12. We prove the second statement of the theorem after a small
remark.
6.4 Angles and Segments of Integer Triangles 61

Remark 6.9 The statement of Theorem 6.8(i) does not necessarily hold for even
continued fractions of the tangents. For instance, consider an integer triangle with
integer area equaling 7 and all angles integer congruent to larctan 7/3. For the odd
continued fractions 7/3 = [2; 2 : 1] of all angles we have
[2; 2 : 1 : −1 : 2 : 2 : 1 : −1 : 2 : 2 : 1] = 0.
If instead we take the even continued fractions 7/3 = [2; 3], then we have
35
[2; 3 : −1 : 2 : 3 : −1 : 2 : 3] =
= 0.
13
Remark 6.10 In Chap. 13 we discuss the relation on integer angles of integer poly-
gons similar to the case of triangles.

Proof of the second statement of Theorem 6.8 Consider a triangle ABC with ra-
tional angles α, β, and γ (at vertices at A, B, and C respectively). Suppose that
for every k > 1 and every integer triangle KLM, the triangle ABC is not integer
congruent to the k-multiple of KLM. In other words, we have
 
gcd l(AB), l(BC), l(CA) = 1.
Suppose that S is the integer area of ABC. Then by the definition of integer
sine, the following holds:
⎧
⎨ l(AB) l(AC) = S/ lsin α,
l(BC) l(BA) = S/ lsin β,
⎩
l(CA) l(CB) = S/ lsin γ .
Since gcd(l(AB), l(BC), l(CA)) = 1, we have l(AB) = λ1 and l(AC) = λ2 .
Therefore, the triangle ABC is integer congruent to the triangle A0 B0 C0 of
the theorem. 

6.4 Angles and Segments of Integer Triangles

Let us find the integer tangents of all angles and the integer lengths of all edges
of any integer triangle, knowing the integer lengths of two edges and the integer
tangent of the angle between them. Suppose that we know the integer lengths of the
edges AB, AC and the integer tangent of the angle ∠BAC in the triangle ABC. Let
us show how to restore the integer length and the integer tangents for the remaining
edge and the rational angles of the triangle.
For simplicity we fix some integer basis and use the system of coordinates OXY
corresponding to this basis (denoted by (∗, ∗)).

Theorem 6.11 Consider some triangle ABC. Let

l(AB) = c, l(AC) = b, and ∠CAB ∼
= α.
Then the angles ∠BCA and ∠ABC are defined in the following way:
62 6 Integer Angles of Integer Triangles

Fig. 6.4 Three possible configuration of points O, D, and Q

⎧
⎨ π − larctan( c lcos α−b ) if c lcos α > b,
c lsin α

∠BCA ∼
= larctan(1) if c lcos α = b,
⎩ t c lsin α
larctan ( b−c lcos α ) if c lcos α < b,
⎧ t)
⎪ b lsin(α
⎨ π − larctan ( b lcos(α t )−c ) if b lcos(α ) > c,
t t

∠ABC ∼
= larctan(1) if b lcos(α t ) = c,
⎪
⎩ larctan( b lsin(α t ) )
c−b lcos(α t ) if b lcos(α t ) < c.

Proof We start with proving the formula for the angle ∠BCA. Let α ∼ = larctan(p/
q), where gcd(p, q) = 1. Then CAB ∼ = DOE, where D = (b, 0), O = (0, 0), and
E = (qc, pc). Let us express the angle ∠DEO. Denote by Q the point (qc, 0). If
qc − b = 0, then ∠BCA = ∠DEO ∼ = larctan 1. If qc − b
= 0, then we have

∼ cp ∼ c lsin α
∠QDE = larctan = larctan .
cq − b c lcos α − b
The expression for the angle ∠BCA follows directly from the above expression for
∠QDE, since ∠BCA ∼ = ∠QDE. (See Fig. 6.4; here l(OD) = b, l(OQ) = c lcos α,
and therefore l(DQ) = |c lcos α − b|.)
To obtain an expression for the angle ∠ABC, we consider the triangle BAC.
Calculate the angle ∠CBA and then transpose all angles in the expression. Finally,
the integer length of CB is defined from the integer sine formula. 

6.5 Examples of Integer Triangles

Let us define certain types of triangles occurring in integer geometry. Since dual
angles are not necessarily congruent, we have more different types than in the Eu-
clidean case.

Definition 6.12 An integer triangle ACB is said to be dual to the triangle ABC.
6.5 Examples of Integer Triangles 63

An integer triangle is said to be self-dual if it is integer congruent to the dual

triangle.
An integer triangle is said to be pseudo-isosceles if it has at least two integer
congruent angles.
An integer triangle is said to be integer isosceles if it is pseudo-isosceles and
self-dual.
An integer triangle is said to be pseudo-regular if all its angles and all its edges
are integer congruent.
An integer triangle is said to be integer regular if it is pseudo-regular and self-
dual.

By the first criterion of integer congruence for integer triangles, the number of
integer congruence classes for integer triangles with bounded integer area is always
finite. In Fig. 6.5 we show the complete list of 33 triangles representing all integer
congruence classes of integer triangles with integer areas not greater than 10. We
enumerate the vertices of the triangle clockwise. Near each vertex of a triangle we
write the tangent of the corresponding rational angle. Inside any triangle we write its
area. We draw dual triangles on the same light gray region (if they are not self-dual).
Integer regular triangles are colored in dark gray, integer isosceles but not integer
regular triangles are white, and the others are light gray.

Integer Triangles of Small Area The above criteria allow us to enumerate all
integer triangles of small integer area up to integer equivalence. In the following
table we write down the numbers N (d) of noncongruent integer triangles of in-
teger area d for d ≤ 20 (here the dual noncongruent triangles are counted as two).

d 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
N(d) 1 1 2 3 2 4 4 5 5 6 4 10 6 8 8 11 6 13 8 14

Let us prove an easy statement on the statistics of N (d).

Proposition 6.13 We have

d d(d + 1)
≤ N (d) ≤ .
3 2
Proof First, let us show that N (d) ≤ d(d + 1)/2. From Theorem 6.8(ii) we know
that every integer triangle is equivalent to some A0 B0 C0 where
A0 = (0, 0), B0 = (λ2 q, λ2 p), C0 = (λ1 , 0),
where 0 < q < p. The area of the triangle is exactly λ1 λ2 q. Hence λ2 q ≤ d. There
are exactly d(d + 1)/2 integer points satisfying all listed conditions. Each such point
can be chosen to construct B0 .
If d is divisible by λ2 q, we have an integer triangle with C0 constructed uniquely
by setting

d
C0 = ,0 .
λ2 q
64 6 Integer Angles of Integer Triangles

Fig. 6.5 List of integer triangles of integer area less than or equal to 10
6.6 Exercises 65

Hence there are at most d(d + 1)/2 triangles of area d.

Let us show that N (d) ≥ d/3. There are exactly d noncongruent angles that
appear in triangles of area d:
k
larctan , k = 1, 2, . . . , d.
d
Each triangle contains at most three noncongruent angles. Hence N (d) ≥ d/3. 

It appears that the growth rate is linear. For instance, for a prime number d, we
always have N(d) ≤ d. Still, in some exceptional cases, when d has many divisors,
it can happen that N (d) > d. For instance, N (240) = 248.

6.6 Exercises

Exercise 6.1 Find all triangles of area 11.

Exercise 6.2 Let ltan(ABC) = 7/5; l(AB) = 3; l(BC) = 5. Find the remaining
angles and the remaining edge.

Exercise 6.3 Prove that in Euclidean geometry there exists a triangle with pre-
scribed angles α, β, and γ if and only if

tan(α + β + γ ) = 0,
tan(α + β) ∈
/ [0; tan α]
(here without loss of generality we suppose that α is acute).

Exercise 6.4 Is there a triangle with angles congruent to larctan(24/7), larctan(24/

13), and larctan 4?

Exercise 6.5 Suppose q has n divisors. Find an upper estimate on N (q) linear in d
and n.
Chapter 7
Continued Fractions and SL(2, Z) Conjugacy
Classes. Elements of Gauss’s Reduction Theory.
Markov Spectrum

In this chapter we study the structure of the conjugacy classes of SL(2, Z). Recall
that SL(2, Z) is the group of all invertible matrices with integer coefficients and unit
determinant. We say that the matrices A and B from SL(2, Z) are integer conjugate
if there exists an SL(2, Z) matrix C such that B = CAC −1 .
Notice that in the case of algebraically closed fields (say in C), every matrix
is conjugate to its Jordan normal form. The situation with SL(n, Z) is much more
complicated, since Z is not a field. A description of integer conjugacy classes in the
two-dimensional case is the subject of Gauss’s reduction theory, where conjugacy
classes are classified by periods of certain periodic continued fractions (for addi-
tional information we refer to [102, 132], and [134]). We present first steps in the
study of multidimensional Gauss reduction theory in Chap. 21, in the second part of
this book.
In this chapter we discuss the main elements of classical Gauss reduction theory
based on the theory of geometric continued fractions studied in previous chapters
(see also [97]). In particular, we formulate several open problems and show relations
to the Markov spectrum.

7.1 Geometric Continued Fractions

In this section we start with a definition for geometric continued fractions and show
how such continued fractions are related to SL(2, R) matrices. Further, we prove
that all LLS sequences of dual sails coincide. We conclude this section with a discus-
sion of periodic sails and the notion of LLS periods of the corresponding SL(2, Z)
matrices.

7.1.1 Definition of a Geometric Continued Fraction

In 1895, F. Klein introduced a generalization of the geometric interpretation of reg-

ular continued fractions to the multidimensional case. We will discuss the general

O. Karpenkov, Geometry of Continued Fractions, 67

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_7, © Springer-Verlag Berlin Heidelberg 2013
68 7 Continued Fractions and SL(2, Z) Conjugacy Classes

7 18
Fig. 7.1 Geometric continued fraction of the matrix 5 13

multidimensional case in the second part of the book. Now we give a definition from
the one-dimensional Klein geometric continued fraction. It is slightly different from
the definition of regular continued fractions.

Definition 7.1 Consider a pair of distinct lines l1 and l2 passing through an integer
point O. The lines l1 and l2 divide the plane into four angles. A geometric continued
fraction is the union of the sails of these angles.

Notice that the LLS sequences of the angles between the lines are lattice invari-
ants of a geometric continued fraction. Since the LLS sequence of an angle is its
integer complete invariant (see Theorem 4.10), we have the following corollary.

Corollary 7.2 The LLS sequence is a complete invariant of lattice angles under the
group of lattice affine transformations Aff(2, Z).

7.1.2 Geometric Continued Fractions of Real Spectrum SL(2, R)

Matrices

Definition 7.3 Consider a real spectrum matrix A. The matrix A has exactly two
distinct eigenlines. The geometric continued fraction defined by these two straight
lines is said to be associated to A.

In Fig. 7.1 we show the geometric continued fraction for the matrix

7 18
.
5 13

Integer lengths of edges are denoted by black digits, and integer sines, by white. The
LLS sequences of all four sails are equivalent to or inverse to

(. . . , 2, 1, 1, 3, 2, 1, 1, . . .).
7.1 Geometric Continued Fractions 69

7.1.3 Duality of Sails

As we have already seen in Fig. 7.1, the continued fractions of adjacent angles are
equivalent up to a reversal of order. This happens not by chance, such situation is
general. Before showing this, we discuss the notion of sail duality.

Definition 7.4 Two sails are dual with respect to each other if the LLS sequence
of one sail coincides (up to an index shift if the LLS sequences are infinite on both
sides) with the inverted LLS sequence of the second sail.

Proposition 7.5 Let A be a real spectrum matrix with no integer eigenvectors. Then
the sails of the opposite octants are congruent. The sails of the adjacent octants are
dual.

Proof The sails of opposite octants are congruent, since they are taken one to an-
other by the symmetry about the origin.
Let us prove the duality. Let one of the sails for the geometric continued fraction
be a broken line (Ai ). Without loss of generality we may fix coordinates such that
A0 = (1, 0) and A1 = (1, a0 ). Set B0 = (0, 1). Notice that B0 is on the dual sail
(since the sail is defined by the lines y = αx and y = βx, where a0 < α < a0 + 1
and −1 < β < 0 in the chosen coordinates). Denote the remaining points of the dual
sail by Bi , starting from B0 .
Let (ai ) and (bi ) be the LLS sequences for the sails (Ai ) and (Bi ). In Chap. 3
we have already shown the edge–angle duality in the orthant of points with positive
coordinates. In other words, we know that ai = b−i−1 for i ≥ 1.
Let Bi be the point symmetric to Bi with respect to the origin (i ∈ Z). Recall that

B−1 = (a1 , a0 a1 + 1) and A2 = (a1 a2 + 1, a0 a1 a2 + a0 + a2 ).

Consider the linear transformation taking A2 to (1, 0) and A1 to (1, a2 ), namely

a0 a1 + 1 −a1
.
a0 a1 a2 + a0 + a2 −a1 a2 − 1

This transformation takes the point B−1  to the point (0, 1). Now we have edge–angle
 
duality of A2 A1 A0 . . . and B−1 B0 . . . . Therefore, ai = b−i−1 for i ≤ 1. Hence the
LLS sequences are inverse to each other. 

7.1.4 LLS Sequences for Real Spectrum Matrices

Let us define LLS sequences for real spectrum matrices.

Definition 7.6 The LLS sequence of a matrix A in GL(2, R) is the LLS sequence
for any of its sails up to the choice of a direction of a sequence and zero element.
70 7 Continued Fractions and SL(2, Z) Conjugacy Classes

In the next proposition we list some basic properties of LLS sequences for real
spectrum matrices.

Proposition 7.7
(i) For any sequence of integers infinite in two sides there exists a real spectrum
matrix whose LLS sequence coincides with the given sequence.
(ii) Let real spectrum matrices A and B have the same LLS sequence. Then there
exists a matrix C commuting with A and integer conjugate to B.

7.1.5 Algebraic Sails

Consider now the case of real spectrum matrices in SL(2, Z). It turns out that all
such matrices have characteristic polynomials irreducible over the rational numbers.
The sails of such matrices are called algebraic.

Proposition 7.8 Every algebraic sail has a periodic LLS sequence.

Remark 7.9 Later, in Corollary 7.20, we show that a sail with periodic LLS se-
quence is algebraic.

Proof of Proposition 7.8 Let A be a real spectrum matrix in SL(2, Z). The ma-
trix A preserves its invariant lines and the lattice Z2 . Hence it acts on each of the
sails by shifting the vertices of the corresponding broken line along the broken line.
Therefore, the LLS sequences of algebraic sails are periodic. 

In Fig. 7.1, we show the sails of a real spectrum algebraic matrix with the period
of the LLS sequence equal to (2, 1, 1, 3).

7.1.6 LLS Periods of Real Spectrum Matrices

Definition 7.10 Let M be a real spectrum SL(2, Z) matrix. Then M acts on the sail
for M as a shift. Hence M also defines a shift of the LLS sequence. The factor of
the LLS sequence with respect to this shift is called the LLS period of M.

So the LLS period is a cyclically ordered sequence of an even number of integer

elements. Notice that the LLS periods (1, 2, 1, 2) and (2, 1, 2, 1) are the same, and
the LLS periods (1, 2, 1, 2) and (1, 2) are distinct.

Proposition 7.11 Two real spectrum SL(2, Z) matrices M1 and M2 with positive
eigenvalues are integer conjugate if and only if their LLS periods coincide.
7.2 Geometry of Gauss’s Reduction Theory 71

Proof Let M1 and M2 be integer conjugate. Then they have integer congruent ge-
ometric continued fractions, and they define the same shift of these continued frac-
tions. Therefore the LLS periods of M1 and M2 coincide.
Suppose now that the LLS periods of the matrices M1 and M2 coincide. This
means that the corresponding continued fractions have the same LLS sequences,
and therefore they are integer congruent. Therefore, there exists a matrix M2 integer
congruent to M2 whose geometric continued fraction coincides with the geometric
continued fraction of M1 . Suppose v is a vertex of the geometric continued fraction
of the operator M1 . Then
   
M1 (v) = M2 (v) and M1 M1 (v) = M2 M1 (v) ,

since the matrices M1 and M2 define the same shift of the geometric continued
fraction. Since the vectors v and M1 (v) form a basis of R2 , we have

M1 = M2 .

Therefore, M1 is integer conjugate to M2 . 

7.2 Geometry of Gauss’s Reduction Theory

In this section we briefly discuss a geometric approach to the enumeration of integer
conjugacy classes of SL(2, Z) operators by means of periods of geometric continued
fractions.

7.2.1 Cases of Matrices with Complex, Real, and Coinciding

Eigenvalues

It is natural to split SL(2, Z) into the following three cases.

Case of Complex Spectra Consider SL(2, Z) matrices whose characteristic poly-

nomials have a pair of complex conjugate roots (we call such a matrix a complex
spectrum matrix). There are exactly three integer conjugacy classes of such matri-
ces, represented by

1 1 0 1 0 1
, , and .
−1 0 −1 0 −1 −1

Degenerate Case Let us now study matrices whose characteristic polynomial

has a double root (which actually equal to 1). Such matrices are integer conjugate
to exactly one matrix of the following family:

1 n
for n ≥ 0.
0 1
72 7 Continued Fractions and SL(2, Z) Conjugacy Classes

Case of Real Spectra The case of matrices with real distinct eigenvalues is the
most complicated (we call such a matrix a real spectrum matrix). We use geometric
continued fractions to construct a complete invariant of conjugacy classes of such
matrices. Let us collect some general definitions and theorems related to this case.

7.2.2 Reduced Matrices

There are two similar subfamilies of real spectrum SL(2, Z) matrices: those with
positive eigenvalues and those with negative eigenvalues. The one-to-one corre-
spondence between them is given by associating to a matrix M the matrix −M.
For simplicity, we consider only matrices with positive eigenvalues.

Definition 7.12 A matrix

a c
M=
b d
in SL(2, Z) is reduced if d > b ≥ a ≥ 0.

Remark The definition of a reduced matrix is slightly different from the classical
one: a matrix in SL(2, Z) is reduced if and only if it has nonnegative entries that are
nondecreasing downward and to the right. In the framework of geometric continued
fractions it is more convenient to use our definition, which we further generalize to
the multidimensional case in Chap. 21.

Theorem 7.13 For every real spectrum matrix A in SL(2, Z) either A or −A is

integer conjugate to a reduced matrix.

We show how to construct a reduced matrix from a given one in Sect. 7.3.2.
It is interesting to note that the LLS period of a reduced matrix (and in particular
its LLS sequence) can be written directly from the coefficients of the matrix.
a c 
Theorem 7.14 Consider a reduced matrix M = bd . Suppose b
a = [a1 ; a2 : · · · :
a2n−1 ] and λ =  d−1
b . Then its LLS period is

(a1 , a2 , . . . , a2n−1 , λ).

0 −1 
If a = 0, we have M = 1 λ+2 for λ ≥ 2. Then the LLS period for M is

(1, λ).

Remark 7.15 For an arbitrary reduced matrix M we associate the corresponding

sequence (a1 , a2 , . . . , a2n−1 , λ) (or (1, λ) respectively) as in Theorem 7.14.

We prove Theorem 7.14 later in Sect. 7.3.1.

7.2 Geometry of Gauss’s Reduction Theory 73

7.2.3 Reduced Matrices and Integer Conjugacy Classes

Let us continue building on Theorem 7.14.

Corollary 7.16 (On enumeration of reduced matrices) The set of real spectrum re-
duced matrices is in one-to-one correspondence (defined in Remark 7.15) with the
set of finite sequences consisting of an even number of positive integer elements.

Proof It is clear that two distinct reduced matrices have distinct sequences as de-
fined in Remark 7.15. Let us show that the sequence

(a1 , a2 , . . . , a2n−1 , λ)

is realizable for some

 −1 reduced
 matrix. If this sequence is (1, α), then the corre-
sponding matrix is 10 λ+2 . For all the other sequences we define the integers b and
a from
b
= [a1 ; a2 : · · · : a2n−1 ]
a
Now c and d are uniquely recovered from the facts that the matrix has a unit deter-
minant and λ =  d−1b . 

Remark One can ask, what happens if we consider a matrix constructed from a
sequence with an odd number of elements? It turns out that in this case, we have a
real spectrum GL(2, Z) matrix and negative determinant. The LLS sequence of the
associated geometric continued fraction has this sequence as one of the periods.

So the reduced matrices are “almost” normal forms, since each matrix could have
more than one normal form.

Corollary 7.17 (On almost normal forms) The number of reduced matrices in an
integer conjugacy class coincides with the number of elements in the minimal period
of the corresponding LLS sequence.

Proof In Proposition 7.11 we showed that two matrices are integer conjugate if their
LLS periods are conjugate. From Corollary 7.16 it follows that reduced matrices are
in one-to-one correspondence with finite sequences of even lengths. The number
of sequences that are cyclic permutations of each other exactly coincides with the
number of elements in the minimal period of the sequences, which is also a minimal
period of the corresponding LLS sequence. 

7.2.4 Complete Invariant of Integer Conjugacy Classes

We have the fundamental result:

74 7 Continued Fractions and SL(2, Z) Conjugacy Classes

Theorem 7.18 (On complete invariant of integer conjugacy classes)

(i) Every LLS period is a complete invariant of an integer conjugacy class for
SL(2, Z) matrices with distinct positive real eigenvalues.
(ii) An arbitrary finite sequence of an even length with positive integer elements is
realizable as a period of some LLS sequence.

Proof Proposition 7.11 is a reformulation of the first part of the theorem. Corol-
lary 7.16 implies that all sequences are realized. 

We conclude this subsection with the following example.

Example 7.19 Consider

1519 1164
M= .
−1964 −1505

The LLS period of M is (1, 2, 1, 2). Hence there are exactly two reduced matrices
represented by the sequences (1, 2, 1, 2) and (2, 1, 2, 1). The coefficients a and b
for the corresponding reduced matrices are

b 4 b 8
= [1; 2 : 1] = and = [2; 1 : 2] = .
a 3 a 3

We find the elements c and d of the reduced matrices from conditions λ =  d−1b 
and ad − bc = 1. Finally, we get both reduced matrices integer conjugate to M:

3 8 3 4
and .
4 11 8 11

7.2.5 Algebraicity of Matrices with Periodic LLS Sequences

Corollary 7.20 A sail with a periodic LLS sequence is algebraic (i.e., a sail of some
algebraic real spectrum matrix).

Proof In Theorem 7.14 we constructed the algebraic matrices for all finite se-
quences as periods. Then in Proposition 7.7 we showed that sails with equivalent
LLS sequences are either integer congruent or dual. Therefore, every sail with a
periodic LLS sequence is algebraic. 

Remark If an LLS sequence of a geometric continued fraction has an odd period,

then there exists a GL(2, Z) matrix B with negative discriminant whose square is
a shift of the sail along the sail. The existence of B implies that all sails of the
continued fraction are integer congruent dual.
7.3 Some Technical Details and Open Questions 75

7.3 Some Technical Details and Open Questions Related to

Gauss’s Reduction Theory
In this section we give a proof of Theorem 7.14, show an algorithm to construct a re-
duced matrix integer conjugate to a given one, and formulate several open questions
on frequencies of occurrences of reduced matrices and their complexities.

7.3.1 Proof of Theorem 7.14

 −1 
Case a = 0 Consider a matrix M = 10 λ+2 for λ ≥ 2. Direct calculations show
that both eigenvalues of the corresponding matrix are positive; therefore every sail
is invariant under the action of M.
Consider the sail containing the point P = (1, 0). Notice first that the triangle
with vertices O, P , and M(P ) = (0, 1) is empty. Hence for every integer k, the tri-
angle with vertices O, M k (P ), and M k+1 (P ) is empty. Secondly, the point M(P )
is in the triangle with vertices O, P , and M 2 (P ) = (−1, λ + 2). Hence for every
integer k, the triangle with vertices O, M k (P ), and M k+2 (P ) contains the point
M k+1 (P ). Therefore, the boundary of the convex hull of all points M n (P ) coin-
cides with the broken line with edges M k (P )M k+1 (P ) for k ∈ Z. Therefore, all the
vertices of this sail are of type M n (P ). We have
   
l P , M(P ) = 1 and lsin P , M(P ), M 2 (P ) = λ.

Therefore, the LLS sequence has period (1, λ).

Case a > 0 We start with the following notation. Let

c = r + λa and d = s + λb.

b , we have 1 ≤ s ≤ b. In the new notation we have

Since λ =  d−1

a r + λa
M= .
b s + λb

The discriminant of the characteristic polynomial of the matrix A equals (a +

λb + s)2 − 4. Since a > 0, we have b > 1. In addition, we have λ ≥ 1 and s ≥ 1.
Therefore, the discriminant of M is positive. Therefore, the matrix M is a real spec-
trum matrix. For every integer t > 2, the integer t 2 − 4 is not the square of some
integer, and hence the sails of the matrix M are infinite.
Our next goal is to construct a period for the LLS sequence. The matrix M has
the following eigenvalues:

a + λb + s ± (a + λb + s)2 − 4
.
2
76 7 Continued Fractions and SL(2, Z) Conjugacy Classes

They are both positive, and hence the matrix M takes each sail to itself.
Let us consider the sail S whose convex hull contains the point P = (1, 0). De-
note by α the closed integer angle with vertex at the origin O whose edges are
integer rays OP and OM(P ). Consider the sail of α. Denote the broken line of this
sail by Sα .
Now we show that the sail of α is completely contained in the sail of the matrix
M containing P . Denote by Sα∞ the following infinite broken line:

+∞
  i 
M (Sα ) .
i=−∞

By construction, the convex hull of Sα∞ coincides with the convex hull of the sail S
of the matrix M. Let us prove that Sα∞ coincides with S, i.e., with the boundary of
its convex hull. The broken line Sα∞ is the boundary of the convex hull if the convex
angles generated by every pair of its adjacent edges do not contain the origin. Due
to linear periodicity it is sufficient to check all the angles of one of the periods of
the broken line Sα∞ . For instance, we will examine all the angles whose vertices are
in Sα except for the angle with vertex at M(P ) (the situation is clear for the angles
at vertices M(P ) and P ). All convex angles generated by pairs of adjacent edges
at inner vertices of the sail Sα do not contain the origin by definition. It remains to
check the angle with vertex at P = (1, 0).
Since b/a > 1, the direction of the first edge is (0, 1). Let us choose an integer
point Q on the second edge of this angle such that l(PQ) = 1. By construction, the
triangle OPQ is empty, and hence lS(OPQ) = 1. Thus ld(Q, OP) = 1. Therefore,
Q = (x, −1) for some integer x. Since the set M −1 (Sα ) is convex, the value of x is
determined by the eigendirection


−a + λb + s + (a + λb + s)2 − 4
, −1
2b

as follows:
  
−a + λb + s + (a + λb + s)2 − 4
x= + 1.
2b
This implies that x is contained in the open interval (λ + 1 + (s − 1)/b, λ + 1 + s/b).
From b ≥ s > 0 we get
x = λ + 1.
Hence for λ ≥ 1, the convex angle at vertex P does not contain the origin. Therefore,
the broken line Sα∞ coincides with the sail.
Let b/a = [a0 ; · · · : a2n−1 ]. In Chap. 3 it was shown that the sail Sα consists of
exactly n+1 integer segments; the integer lengths of the consecutive segments equal
a0 , a2 , . . . , a2n ; the integer sines of the corresponding angles equal a1 , a3 , . . . , a2n−1
(see Corollary 3.5 and Theorem 3.6). The integer sine of the angle at point P equals
7.3 Some Technical Details and Open Questions 77

x − 1 = λ. Hence the LLS sequence of the sail S has period

(a0 , a1 , . . . , a2n , λ).
This concludes the proof of Theorem 7.14.

7.3.2 Calculation of Periods of the LLS Sequence

In this subsection we are interested in calculating the LLS periods of matrices. We

do it in two steps. First, we present an algorithm to construct a reduced matrix
integer congruent to a given one. Secondly, we use Theorem 7.14 to find explicitly
the LLS period.
For the remainder of this section we denote by [[p, r][q, s]] the matrix

p r
.
q s

Algorithm to Construct Reduced Matrices

Input data. We are given an SL(2, Z) matrix M = [[p, r][q, s]]. In addition, we
suppose that the characteristic polynomial of the matrix is irreducible over Q
(or, equivalently, that it does not have ±1 as roots) and has two positive real
roots.
Goal of the algorithm. To construct one of the periods of the LLS sequence for M.
Step 1. If q < 0, then we multiply the matrix [[p, r][q, s]] by −Id. Go to Step 2.
Step 2. We have q ≥ 0. After conjugation of the matrix [[p, r][q, s]] by the matrix
[[1, −q/p][0, 1]] we get the matrix [[p  , r  ][q  , s  ]], where 0 ≤ p  ≤ q  . Go to
Step 3.1.
Step 3.1. Suppose that q  = 1. Then p  = 0, r  = −1. In addition we have |s  | > 2,
otherwise, the matrix has either complex roots or rational roots. The algorithm
stops, and the output of the algorithm is the matrix [[0, −1][1, s  ]].
Step 3.2.1. Suppose that q  > 1 and s  > q  . Then the algorithm stops, and the out-
put of the algorithm is the matrix [[p  , r  ][q  , s  ]].
Step 3.2.2. Suppose that q  > 1 and s  < −q  . Conjugate the matrix [[p  , r  ][q  , s  ]]
by the matrix [[−1, 1][0, 1]]. As a result we have the matrix [[p  , r  ][q  , s  ]]
with q  = q  , p  = q  − p  , and s  = −q  − s  > 0. Go to Step 3.2.1 or to
Step 3.2.3 depending on s  and q  .
Step 3.2.3. Suppose that q  > 1 and |s  | ≤ |q  |. Notice that the absolute values of q 
and s  do not coincide, since the matrix has unit determinant. Hence |s| < |q|.
Then we have
 
    p  s  − 1  (q  − 1)2 + 1
r  =   ≤ q  − 1.
 q  ≤ q
Go to Step 1 with the matrix [[s  , q  ][r  , p  ]] with |r  | < |q  |. This is the matrix
obtained from [[p  , r  ][q  , s  ]] by conjugation with [[0, −1][−1, 0]].
Output. The reduced integer matrix that is conjugate to M.
78 7 Continued Fractions and SL(2, Z) Conjugacy Classes

Remark 7.21 During the algorithm we only use conjugations with GL(2, Z) ma-
trices and multiply by −Id. So in the output one should expect a reduced integer
matrix that is conjugate either to M or to −M. Nevertheless, since M, as well as
any reduced matrix, has all positive eigenvalues, the resulting matrix is conjugate
exactly to M.

Remark 7.22 The algorithm works in a finite number of steps, since after each iter-
ation the integer absolute value of q is decreases.

Remark 7.23 To find the LLS period of a matrix M, one should construct the re-
duced matrix M  conjugate to M using the above algorithm and then apply Theo-
rem 7.14 to M  .

7.3.3 Complexity of Reduced Operators

As we have already seen, in the majority of conjugacy classes there is more than one
reduced matrix. Then it is natural to introduce some additional notion of complexity
to reduce the number of the reduced matrices. Let us introduce one important no-
tion of complexity, whose extended definition we will use later in Chap. 21 in the
multidimensional case.

Definition 7.24 Let M = [[p, r][q, s]] be a reduced matrix. A ς -complexity of M

is the number q. We denote it by ς(M).

A geometric interpretation of ς(M) is as the integer area of the triangle with

vertices O(0, 0), v = (1, 0), and M(v).

Problem 2 Find a reduced matrix with minimal ς -complexity from a given conju-
gacy class.

Remark The minimal ς -complexity coincides with the minimal value of lsin POQ,
where P is an arbitrary integer point distinct from O, and Q = A(P ).

If the LLS period of M is (a1 , . . . , a2n ), then the above problem is equivalent to
finding the minimal numerator among the numerators of the rational numbers:
[a1 ; · · · : a2n−1 ], [a2 ; · · · : a2n ], [a3 ; · · · : a2n : a1 ], ...,
[a2n ; a1 : · · · : a2n−2 ].

Example 7.25 For instance, if the LLS period of M is (a, b), where a < b, then the
minimum for the numerators is a.

Example 7.26 Let the minimal period of the LLS sequence consists of three ele-
ments: (a, b, c, a, b, c). The numerator of the fraction [a; b : c : a : b] is minimal if
and only if c ≥ a, b.
7.3 Some Technical Details and Open Questions 79

The following example shows that it is not correct to skip the maximal entry of
the sequence.

Example 7.27 Consider the LLS period (1, 4, 5, 4, 1, 4). The minimum of the nu-
merators is attained at the fraction [1; 4 : 5 : 4 : 1], and not at the fraction [4; 1 : 4 :
1 : 4].

7.3.4 Frequencies of Reduced Matrices

We start with a proper probabilistic space. Let P = (a1 , a2 , . . . , a2n−1 , a2n ) be an

arbitrary sequence of positive integer elements. Denote by #N (P ) the number of
matrices M satisfying the following:
(i) The absolute value of every entry of M does not exceed N .
(ii) The sequence P is the LLS period of M.
(iii) The algorithm of the previous subsection with M as an input produces the
reduced matrix [[p, r][q, s]] with (p, q) = (0, 1) for the case of P = (1, a2 ),
and
q/p = [a1 ; a2 : · · · : a2n−1 ]
for all the remaining cases.
Let us formulate several pertinent open questions.

Problem 3
(i) Which reduced matrix with a given trace (or with a given LLS sequence) is the
most frequent as N tends to infinity?
(ii) What is the relative probabilities of two sequences representing the same LLS
period as N tends to infinity?
(iii) Is it true that the reduced matrix with minimal ς -complexity is the most fre-
quent among the reduced matrices of a given LLS period?

In Table 7.1 we show some values of #25000 (P ) for matrices with absolute value
of the trace less than 11.
Observe that SL(2, Z) matrices corresponding to the sequence (1, 2) are more
frequent than those corresponding to (1, 1). This is due to the fact that matrices
with LLS periods (1, 1) have integer congruent dual sails. One may enumerate the
matrices with multiplicities equivalent to the number of integer congruent sails and
obtain
4#25000 (1, 1) > 2#25000 (1, 2) + 2#25000 (2, 1).
We conclude with the following question. Denote by GK(P ) the frequency of a
sequence P = (a1 , a2 , . . . , a2n−1 ) in the sense of Gauss–Kuzmin:
80 7 Continued Fractions and SL(2, Z) Conjugacy Classes

Table 7.1 Values of #25000 (P ) for matrices with small absolute value of the trace
Absolute value Notation for classes Period Matrix Value of
of the trace of equivalent matrices P [[p, r][q, s]] #25000 (P )

3 L3 (1, 1) [[0, 1][−1, 3]] 663160

4 L4 (1, 2) [[0, 1][−1, 4]] 834328
(2, 1) [[1, 2][1, 3]] 304776
5 L5 (1, 3) [[0, 1][−1, 5]] 818200
(3, 1) [[1, 3][1, 4]] 194528
6 L6,1 (1, 4) [[0, 1][−1, 6]] 777128
(4, 1) [[1, 4][1, 5]] 141784
L6,2 (2, 2) [[1, 2][2, 5]] 446432
7 L7,1 (1, 5) [[0, 1][−1, 7]] 734904
(5, 1) [[1, 5][1, 6]] 110848
L7,2 (1, 1, 1, 1) [[2, 3][3, 5]] 201744
8 L8,1 (1, 6) [[0, 1][−1, 8]] 695560
(6, 1) [[1, 6][1, 7]] 90688
L8,2 (2, 3) [[1, 2][3, 7]] 435472
(3, 2) [[1, 3][2, 7]] 310872
9 L9 (1, 7) [[0, 1][−1, 9]] 660984
(7, 1) [[1, 7][1, 8]] 76552
10 L10,1 (1, 8) [[0, 1][−1, 10]] 630592
(8, 1) [[1, 8][1, 9]] 66064
L10,2 (2, 4) [[1, 2][4, 9]] 408216
(4, 2) [[1, 4][2, 9]] 239712
L10,3 (1, 1, 1, 2) [[2, 3][5, 8]] 260872
(2, 1, 1, 1) [[2, 5][3, 8]] 114084
(1, 2, 1, 1) [[3, 4][5, 7]] 149832
(1, 1, 2, 1) [[3, 5][4, 7]] 114084

1 (α1 + 1)α2
GK(P ) = ln ,
ln(2) α1 (α2 + 1)

where

α1 = [a1 ; a2 : · · · : a2n−2 : a2n−1 ] and α2 = [a1 ; a2 : · · · : a2n−2 : a2n−1 + 1].

Problem 4 Let

P1 = (a1 , a2 , . . . , a2n−1 , a2n ), P1 = (a1 , a2 , . . . , a2n−1 ),

P2 = (a2 , a3 , . . . , a2n , a1 ), P2 = (a2 , a3 , . . . , a2n ).
7.4 Minima of Quadratic Forms and the Markov Spectrum 81

Is the following true:

#n (P1 ) GK(P1 )
lim = ?
n→∞ #n (P2 ) GK(P2 )

For further references to Gauss–Kuzmin statistics, see Chap. 9.

7.4 Minima of Quadratic Forms and the Markov Spectrum

We conclude this chapter with a small observation of a classical subject that in some
sense generalizes the question of complexity of minimal periods.

7.4.1 Calculation of Minima of Quadratic Forms

Consider a quadratic form

f (x, y) = ax 2 + bxy + cy 2

with real coefficients and positive discriminant Δ(f ) = b2 − 4ac. Set

 
m(f ) = inf f (x, y).
(x,y)∈Z2 \{(0,0)}

√
The set of all possible values of Δ(f )/m(f ) is called the Markov spectrum. For
a general reference on the subject we recommend the book of T.W. Cusick [41].

Proposition 7.28
(i) Let f be a quadratic form with positive discriminant. Consider the geometric
continued fraction defined by two lines, all of whose points satisfy the equation

f (x, y) = 0.

Let S be the set of all vertices in all four sails of this continued fraction. Then
we have
 
m(f ) = inf f (x, y).
(x,y)∈S

(ii) Let the LLS sequence of the continued fraction be [. . . , a−1 , a0 , a1 , . . .]. Sup-
pose that vi = (xi , yi ) is a vertex of the sail corresponding to the element ai of
the sail (i.e., the integer sine at vertex vi in the sail equals ai ). Then

Δ(f )/m(f ) = ai + [0; ai+1 : ai+2 : · · · ] + [0; ai−1 : ai−2 : · · · ].
82 7 Continued Fractions and SL(2, Z) Conjugacy Classes

Remark 7.29 If the LLS sequence is periodic, the expression of Proposition 7.28(ii)
is


2
 1 1 4
Δ(f )/m(f ) = a0 + + − ,
ltan α ltan α t lsin2 α
where the integer angle α is integer congruent to larctan([a1 ; · · · : a2n ]), and where
(a0 , a1 , . . . , a2n ) is some minimal even period of the LLS sequence.

If f has integer coefficients, the corresponding continued fraction is either finite

or periodic. The minimum is then attained at some vertex of the sail of the geometric
continued fraction. So the calculation of this minimum is similar to the calculation
of the complexity of minimal periods studied above in this chapter.

Remark 7.30 Instead of a quadratic form in two variables one can take a form of
degree n in n variables corresponding to the product of n linear real forms. Any
such form defines a point in the multidimensional Markov spectrum. Namely, for
a form f one takes the infimum of the set of absolute values of f for all nonzero
integer points divided by the n-th root of the discriminant of f . There is not much
known about the multidimensional Markov spectrum. We refer the interested reader
to [49–51], and [194] (see also in [72]).

7.4.2 Some Properties of Markov Spectrum

The Markov spectrum is √a closed set that has√a complicated structure. It does not
contain points less than 5. On the segment [ 5, 3] the spectrum has a unique limit
point, which is 3. It was described by A. Markov (see in [135] and [136]). The first
elements in the spectrum in the increasing order are as follows:
√
√   √   221  
5 = (1) , 8 = (2) , = (2, 2, 1, 1) ,
5
√ √
1517   7565  
= (2, 2, 1, 1, 1, 1) , = (2, 2, 2, 2, 1, 1) ,
13 29
√
2600  
= (2, 2, 1, 1, 1, 1, 1, 1) .
17
All these points are enumerated in Theorem 7.31. All real numbers greater than the
so-called Freiman’s constant
√
221564096 + 283748 462
F= = 4.5278295661 . . .
491993569
are in the Markov spectrum. The segment [3, F ] has not been completely studied.
It has many gaps, i.e., open segments that belong to the complement to the Markov
spectrum.
√ √ The largest gap in the Markov spectrum below 3 is the open segment
( 12, 13). The first gaps, including the maximal one, were calculated by O. Per-
7.4 Minima of Quadratic Forms and the Markov Spectrum 83

ron in [166] and [167]. Nowadays many other gaps are known, but a complete de-
scription of the gaps is unknown. For further information we refer the interested
reader to [38] and [41] (see also [84]).
It is also known that in some neighborhood of 3 the Markov spectrum has
Lebesgue measure zero (for instance, in the paper [31] by R.T. Bumby, it is shown
that the spectrum is of measure zero for the ray x < 3.33440).

Theorem 7.31 (A. Markov [136])

√
(i) The Markov spectrum below 3 consists of the numbers 9m2 − 4/m, where m
is a positive integer such that

m2 + m21 + m22 = 3mm1 m2 , m2 ≤ m1 ≤ m,

for some positive integers m1 and m2 .

(ii) Let the triple (m, m1 , m2 ) fulfill the conditions of item (i). Suppose that u is the
least positive residue satisfying

m2 v ≡ ±m1 mod m

and v is defined from

u2 + 1 = vm.
Then the form

fm (x, y) = mx 2 + (3m − 2u)xy + (v − 3u)y 2

√
represents the value 9m2 − 4/m in the Markov spectrum.

Remark 7.32 It is interesting to observe that the theory of the Markov spectrum is
based on the study of continued fractions with elements 1, 2, and 3. This is due to
the fact that LLS sequences containing greater elements correspond to the quadratic
forms that contribute to the Markov spectrum above Freiman’s constant.

7.4.3 Markov Numbers

Finally, let us say a few words about the integer solutions of the equation mentioned
in Theorem 7.31. The equation

x 2 + y 2 + z2 = 3xyz

is called the Markov Diophantine equation. A Markov number is a positive integer

x for which there exist positive integers y and z such that the triple (x, y, z) is a
solution of the Markov Diophantine equation. The first few Markov numbers are

1, 2, 5, 13, 29, 34, 89, 169, 194, 233, 433, 610, 985, 1325.
84 7 Continued Fractions and SL(2, Z) Conjugacy Classes

Fig. 7.2 The Markov number tree

The corresponding solutions are Markov triples

(1, 1, 1), (1, 1, 2), (1, 2, 5), (1, 5, 13), (2, 5, 29),
(1, 13, 34), (1, 34, 89), (2, 29, 169), . . .

It turns out that Markov triples possess the following regularity.

Proposition 7.33 Every Markov triple is obtained from (1, 1, 1) by applying a se-
quence operations of the following two types:
– permute the numbers x, y, and z in the triple (x, y, z);
– if (x, y, z) is a Markov triple, then (x, y, 3xy − z) is also a Markov triple.
7.5 Exercises 85

Without loss of generality we consider only solutions with x ≤ y ≤ z. The struc-

ture of such solutions forms a tree, whose vertices are almost all of valence 3 (except
for two vertices). An edge between two of vertices corresponds to the operation
described in Proposition 7.33. We show several of the first vertices of the tree in
Fig. 7.2.

7.5 Exercises

Exercise 7.1 Draw a geometric continued fraction defined by the lines x − 2y = 1

and 3x + 4y = 3 and calculate the LLS sequences for all the sails.

Exercise 7.2 Is it true that two SL(2, R) matrices are integer conjugate if and only
if their geometric continued fractions are integer congruent?

Exercise 7.3 Consider a real spectrum SL(2, Z) matrix. Let its eigen lines be y =
α1,2 x. Then α1 and α2 are conjugate quadratic irrationalities.

Exercise 7.4 Let A be a real spectrum (2 × 2) matrix. Find all the matrices com-
muting with A.

Exercise 7.5 Is the statement of Proposition 7.5 true for rational angles (R-angles,
L-angles)? Find a correct analogue in these cases.

Exercise 7.6 Prove Proposition 7.7.

Exercise 7.7 Classify the integer conjugacy classes in the complex spectra case.

Exercise 7.8 Find a reduced form for the matrix

103 69
.
100 67
Chapter 8
Lagrange’s Theorem

The aim of this chapter is to study questions related to the periodicity of geomet-
ric and regular continued fractions. The main object here is to prove Lagrange’s
theorem stating that every quadratic irrationality has a periodic continued fraction,
conversely that every periodic continued fraction is a quadratic irrationality. One of
the ingredients to the proof of Lagrange theorem is the classical theorem on integer
solutions of Pell’s equation
m2 − dn2 = 1.
So, there is a strong relation between periodic fractions and quadratic irrationalities.
It is natural to ask what happens in cases of cubic, quartic, etc., irrationalities? How
can one generalize Lagrange’s theorem to the multidimensional case? We give a
partial answer to such questions in Chap. 18.
We start in Sect. 8.1 with the study of so-called Dirichlet groups, which are the
subgroups of GL(2, Z) preserving certain pairs of lines. These groups are closely
related to the periodicity of sails. The structure of a Dirichlet group is induced by
the structure of the group of units in orders (we will discuss this later in more detail
for the multidimensional case in Chap. 17); here we restrict ourselves to the simplest
two-dimensional case. In Sect. 8.2 we take a break and show how to take nth roots
of matrices using Gauss’s reduction theory. In Sect. 8.3 we study the solutions of
Pell’s equation. And finally, in Sect. 8.4 we prove Lagrange’s theorem.

8.1 The Dirichlet Group

Let A be a GL(2, Z) matrix with two distinct eigenvalues. Denote the set of all
GL(2, Z) matrices commuting with A by Ξ (A), i.e.,
 
Ξ (A) = B ∈ GL(n, Z) | AB = BA .

The set Ξ (A) is a group under the operation of matrix multiplication. We call it
the Dirichlet group in dimension 2. We study the multidimensional case later, in
Chap. 17.

O. Karpenkov, Geometry of Continued Fractions, 87

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_8, © Springer-Verlag Berlin Heidelberg 2013
88 8 Lagrange’s Theorem

In case of A with only real eigenvalues, we are interested in the subgroup of the
Dirichlet group Ξ (A) containing all matrices with positive eigenvalues. This group
is called the positive Dirichlet group and denoted by Ξ+ (A).

Definition 8.1 We say that two Dirichlet groups (or positive Dirichlet groups) Ξ1
and Ξ2 are integer congruent if there exists B ∈ GL(2, Z) such that

A → B −1 AB

is an isomorphism of Ξ1 and Ξ2 .

Proposition 8.2 Two Dirichlet groups (or positive Dirichlet groups) Ξ1 and Ξ2 are
integer congruent if there exist A ∈ Ξ1 with distinct eigenvalues and B ∈ GL(2, Z)
such that B −1 AB ∈ Ξ2 .

Let us study the structure of Dirichlet groups for the matrices of GL(2, Z). There
are three essentially different cases here. The first case is for complex spectrum
matrices. In the second and the third cases we have real spectrum matrices, i.e.,
matrices all of whose eigenvalues are real. In the second case we consider matrices
with rational eigenvalues: the eigenvalues of such GL(2, Z) matrices are ±1. In the
third case we study matrices with irrational eigenvalues:
√
the eigenvalues of such
a±b c
matrices are quadratic irrationalities (i.e., d , where a, b, d are integers and
c > 1 is a square-free integer).

Remark 8.3 Since conjugate matrices have isomorphic Dirichlet groups, it is

enough to study only one representative of each conjugacy class of matrices in
GL(2, Z).

Recall also that the Dirichlet groups are defined only for matrices with distinct
eigenvalues, so we will consider only matrices with distinct eigenvalues.

Complex Spectra Case There are exactly two Dirichlet groups of such type up
to conjugation. The representative matrices of these groups can be chosen from the
group SL(2, Z). We have already seen them in Chap. 7:

0 1 1 1
; .
−1 0 −1 0

The first matrix is the matrix of rotation by the angle π/2. Only its powers commute
with it, and hence the Dirichlet group is isomorphic to Z/4Z. The second matrix
represents the 6-symmetry of a lattice. The related Dirichlet group is isomorphic to
Z/6Z.
8.1 The Dirichlet Group 89

Real Spectra Case I: Rational Eigenvalues Here we have two conjugate classes
of matrices represented by

1 0 1 −1
A1 = ; A2 = .
0 −1 0 −1
The Dirichlet group for Ξ (Ai ) is isomorphic to Z/2Z ⊕ Z/2Z, and it is generated
by ±Ai (for i = 1, 2 respectively).

Real Spectra Case II: Irrational Eigenvalues This is the largest set of irre-
ducible matrices whose conjugacy classes are described via Gauss’s reduction the-
ory. We start with a GL(2, Z) matrix A with distinct irrational eigenvalues (as we
have already mentioned, the eigenvalues are conjugate quadratic irrationalities).
First, note that A2 is in SL(2, Z) and has distinct eigenvalues. Hence without loss
of generality we restrict ourselves to SL(2, Z) matrices. Such matrices have peri-
odic geometric continued fractions with periodic LLS sequences. Due to Gauss’s
reduction theory and Proposition 7.7 there exists a one-to-one correspondence be-
tween integer conjugacy classes of Dirichlet groups and minimal periods of LLS
sequences.

Theorem 8.4 Consider an irreducible real spectrum matrix A ∈ SL(2, Z). Then its
Dirichlet group Ξ (A) is homeomorphic to Z ⊕ Z/2Z.

Proof The proof is straightforward. Any matrix commuting with A has the same
eigenlines, so it either acts as a shift on the sail, or exchanges the sails. The group
of shifts is homeomorphic to Z. Suppose it is generated by Â. If the period is even,
then the dual sails are not congruent, and the group is Z ⊕ Z/2Z generated by Â
and −E. If the period of the LLS sequence of the sail is odd, then one can extract
a square root of  in SL(2, Z). This square root will exchange dual sails, which are
congruent in this case. The Dirichlet group is again Z ⊕ Z/2Z, and it is generated
by Â1/2 and −E. 

Remark 8.5 Let us explain how to find generators of a two-dimensional Dirichlet

group Ξ (A) for a real spectrum matrix with irrational eigenvalues. Take A2 . This
matrix has all positive eigenvalues, and therefore, it preserves every sail. Construct
a sail of A2 . The matrix A2 defines a shift of the LLS sequence and hence one of its
periods. Let this period be n times the minimal period. Then the second generator
of Ξ (A) is (A2 )1/n (which is easily constructed using a Jordan basis for A in which
the matrix is diagonal).

Corollary 8.6 Consider an irreducible real spectrum matrix A ∈ SL(2, Z). Then its
positive Dirichlet group Ξ+ (A) is homeomorphic to Z.

Example 8.7 Consider the reduced matrix

11 30 11 8 + 2·11
A= = .
15 41 15 11 + 2·15
90 8 Lagrange’s Theorem

The period of the LLS sequence related to this continued fraction is (1, 2, 1, 2, 1, 2).
So there exists a cube root of A in SL(2, Z). This cube root is also reduced and has
period of LLS sequence (1, 2). Hence by Theorem 7.14 it is

1 0 + 2·1 1 2
= .
1 1 + 2·1 1 3

Denote this matrix by Â. The group Ξ (A) is generated by  and −E.

8.2 Construction of the Integer nth Root of a GL(2, Z) Matrix

In this section we study how to check whether a certain matrix has roots that are
also integer matrices. We start with SL(2, Z) matrices and further say a few words
about GL(2, Z) case.

Algorithm to Construct an Integer nth Root of a GL(2, Z) Matrix

Data. Suppose that we are given an SL(2, Z) matrix A = [[p, r][q, s]] with positive
discriminant (this matrix has real spectrum with irrational eigenvalues) and a
positive integer n.
Algorithm goals. To verify whether there exists and if so to find a matrix B ∈
GL(2, Z) such that A = B n .
Step 1. First we find one of the reduced matrices A2 = [[p2 , r2 ][q2 , s2 ]] that is con-
gruent to A and remember the congruence matrix C (i.e., A2 = CAC −1 ).
Step 2. By Theorem 7.14 we have the formulas to write a certain period of the LLS
sequence for A2 . There exists A3 = (A2 )1/n that belongs to GL(2, Z) if and only
if the constructed period of the LLS sequence has a subperiod whose length is
1/n that of A2 .
Step 3. Suppose that the period repeats the sequence (a1 , . . . , ak ) n times. Then the
matrix A3 is constructed by the formula A3 = [[p3 , r3 ][q3 , s3 ]], where p3 and
q3 are relatively prime positive numbers defined by
q3
= [a1 ; · · · : ak−1 ].
p3

The remaining two coefficients are

q3 r2 q2 p3 + (s2 − p2 )q3
r3 = , s3 = .
q2 q2

Since q2 > 0, all the coefficients are well defined.

Output. We have verified that an nth root exists. If it exists, the resulting matrix B
of this root is as follows:
B = C −1 A3 C.
8.3 Pell’s Equation 91

Proposition 8.8 The matrix A3 constructed in Step 3 is an nth root of the matrix A2 .

Proof There exists a matrix M ∈ Ξ (A2 ) mapping (1, 0) to (a3 , b3 ). This matrix
corresponds to the sail integer self-congruence related to the shift for the period
(a1 , . . . , ak ) if k is even. If k is odd, we have an integer congruence of dual sails. In
both cases the matrix M is the nth root of A2 contained in Ξ (A2 ).
Since M(1, 0) = (a3 , b3 ), the first row of M coincides with the first row of A3 .
The coefficients in the second rows of A3 and M are uniquely defined by the fact
that A2 M = MA2 and A3 M = MA3 . The formulas are obtained from these matrix
equations (which are four equations in the elements). Hence A3 = M ∈ Ξ (A2 ) ⊂
GL(2, Z), and it is constructed by the formulas in the algorithm. 

Remark 8.9 Let now A ∈ GL(2, Z) and det(A) = −1. It is clear that A does not
have integer even roots. Suppose n is odd. Then to find the nth root one should
check the (2n)th root for the matrix A2 in SL(2, Z). The matrix (A2 )1/2n (if it is in
GL(2, Z)) is one of the nth roots of A.

8.3 Pell’s Equation

Before studying Lagrange’s theorem, we focus on the classical Pell’s equation,

which we will use in the proof of Lagrange theorem. Pell’s equation is a Diophan-
tine equation of the form

m2 − dn2 = 1,

where n and m are integer variables and d is not the square of an integer. Pell’s
equation has trivial solutions (±1, 0). The problem of finding nontrivial solutions
of Pell’s equation was posed by P. Fermat in 1657. The solution of Pell’s equation
was mistakenly attributed to J. Pell. In fact, W. Brouncker found a method to solve
this equation. The first to publish a strict proof was J.-L. Lagrange, in 1768 (see
[129]).

Theorem 8.10 Let d be a nonsquare integer. Then Pell’s equation

m2 − dn2 = 1

has positive solutions. In addition, there exists a positive solution (m1 , n1 ) such that
the set of all positive solutions coincides with the set of pairs (mk , nk ) with positive
integer parameter k defined as follows:


k

mk m1 dn1 1
= .
nk n1 m1 0
92 8 Lagrange’s Theorem

Proof Existence of a nonnegative solution. By Theorem 1.20 there are infinitely

many integer solutions of the inequality
 
√ 
 d − p  < √1 .
 q 5q 2
Letting (p, q) be any of them, then
 
 √
 2  √ √ √  √ √ p  2 2 d + 1
q d −p 2  = |q d −p||q d +p| =  d − p 2 d + d− q < √ .
 q  q  5
Therefore, there exists an integer k such that the equation

q 2 d − p2 = c

has infinitely many integer solutions. Choose among these solutions (m1 , n1 ) and
(m2 , n2 ) such that

m1 ≡ m2 (mod c) and n1 ≡ n2 (mod c).

(This is possible, since the number of distinct remainder pairs is finite.)

Now take
m1 m2 − dn1 n2 m2 n1 − dm1 n2
m̂ = and n̂ = .
c c
Notice that

m̂ ≡ m21 − dn21 = c ≡ 0(mod c) and n̂ ≡ m1 n1 − m1 n1 = 0(mod c).

Hence the point (m̂, n̂) is integer. Now let us consider

√ √
m̂2 − d n̂2 = (m̂ − d n̂)(m̂ + d n̂)
√ √
m1 − dn1 m1 + dn1 m21 − dn21 c
= √ · √ = = = 1.
m2 − dn2 m2 + dn2 m22 − dn22 c
(The second equality is established by direct calculations.) This concludes the proof
of the existence of nonnegative solutions.
Structure of the set of solutions. Let (m, n) be a positive solution of Pell’s equa-
tion. Then

m 1 m dn
= Am,n , where Am,n = .
n 0 n m
Notice that Am,n is an integer matrix with characteristic polynomial t 2 − 2mt + 1 in
variable t. Hence, Am,n is in SL(n, Z), and all its eigenvalues are distinct positive
real numbers. √
Consider an angle α√d with vertex at the origin and edges y = ±x/ d for x > 0.
Let us show that Am,n preserves α√d .
8.4 Periodic Continued Fractions and Quadratic Irrationalities 93

The union of lines containing α√d is defined by the equation

x 2 − dy 2 = 0.

Let us find the image of this union with respect to the action of Am,n . The equation
for the image is
(mx + dny)2 − d(nx + my)2 = 0,
which is equivalent (after simplification) to the initial equation

x 2 − dy 2 = 0.

Hence Am,n preserves the union of two lines. There are two possible situations here:
either these lines are invariant lines of Am,n , or Am,n interchanges the lines. Since
points with positive coordinates are taken to points with positive coordinates, the
lines are eigenlines. Finally, since the eigenvalues are positive real numbers, the
rays defining the angle α√d are preserved.
So the matrix Am,n is in the positive Dirichlet group Ξ+ preserving the angle
α√d . It is easy to see that the point (1, 0) is on the sail of α√d (it is the only point
with x ≤ 1 in the angle).
We have shown that every positive solution of Pell’s equation is defined by a shift
of a sail by a matrix of the positive Dirichlet group Ξ+ . From the other point of view,
every matrix of the Dirichlet group preserves the value x 2 − dy 2 (which is 1 in our
case). Recall that Ξ+ is isomorphic to Z. Letting A generate this group (we choose
A such that it takes (1, 0) to the half-plane y > 0), then all the solutions with positive
x-coordinate are of the form Ak (1, 0). We choose only a positive parameter k, in
order to have positive y-coordinates of the solutions. Finally, denoting by (m1 , n1 )
the point A(1, 0), then the matrix

m1 dn1
n1 m1

is in Ξ+ and defines the same shift of the sail (as A), hence equaling A. This con-
cludes the proof of the second statement of the theorem. 

8.4 Periodic Continued Fractions and Quadratic Irrationalities

A continued fraction [a0 ; a1 : · · · ] is called periodic if there exist positive integers

k0 and h such that for every k > k0 ,

ak+h = ak .

We denote it by [a0 ; a1 : · · · : ak0 : (ak0 +1 : · · · : ak0 +h )].

94 8 Lagrange’s Theorem

Theorem 8.11 (Lagrange)√

Every periodic regular continued fraction is a quadratic
a+b c
irrationality (i.e., d for some integers a, b, c, and d, where b
= 0, c > 1, d > 0,
and c is square-free). The converse is also true: every quadratic irrationality has a
periodic regular continued fraction.

Lemma 8.12 For every quadratic irrationality ξ there exists an SL(2, Z) matrix
such that one of its eigenvectors is (1, ξ ).

Proof Let ξ be a root of the equation c2 x 2 + c1 x + c0 = 0 with integer coefficients

c0 , c1 , c2 . Consider an arbitrary matrix

a11 a12
A= .
a21 a22

Its eigenvectors are



a22 − a11 ± (a22 − a11 )2 + 4a12 a21

1, .
2a12
Notice that ξ and its conjugate root are of the form

−c1 ± c12 − 4c0 c2
.
2c2
So the matrix A has a root (1, ξ ) if the following system is satisfied:
⎧
⎨ nc0 = −a21 ,
nc1 = a11 − a22 ,
⎩
nc2 = a12 ,

for some n
= 0.
Let us find a matrix of SL(2, Z) satisfying this system for some integer n. Since
n is an integer, the coefficients a12 and a21 are integers as well. Since det A = 1, we
have
det A = a11 a22 − a12 a21 = a11 (na11 − c1 ) + n2 c0 c2 = 1.
Therefore,

nc1 ± n2 (c12 − 4c0 c2 ) − 4
a11 = .
2
The coefficient a11 is an integer if and only if there exists an integer m satisfying
 
m2 = n2 c12 − 4c0 c2 − 4.

Set D = c12 − 4c0 c2 , m = m/2, and n = n/2 and rewrite the equation

m2 − Dn2 = 1.
8.4 Periodic Continued Fractions and Quadratic Irrationalities 95

Fig. 8.1 The sails for α and

β coincide starting from some
vertex

We end up with Pell’s equation. Since ξ is irrational, the discriminant D = c12 −

4c0 c2 is not the square of some integer (since ξ is real, D ≥ 0). Hence, by Theo-
rem 8.10, it has an integer solution (m0 , n0 ) with n0
= 0. Hence the matrix

m0 − n0 c1 2n0 c2

−2n0 c0 m0 + n0 c1


has integer elements and unit determinant. Therefore, it is in SL(2, Z). 

Proposition 8.13 Let α1 > α2 > α3 be distinct numbers and let α1 be irrational.
Consider two angles α and β defined by pairs of rays in the lines (y = α1 x, y =
α2 x) and (y = α1 x, y = α3 x), where x > 0. The LLS sequences of these two angles
coincide from some element (up to a sequence shift).

Proof Denote by γ the angle defined by rays (y = α2 x, y = α3 x), x > 0. It is clear

that every integer point of the angle β is either in α or in γ . Then the convex hull
of all integer points of β is the convex hull of the union of the convex hulls of all
integer points for α and γ . Hence the sail for β is contained in the sails for α and γ
except for one edge, which represents the common support line to the convex hulls
for α and β (see Fig. 8.1). Therefore, the sails of α and β coincide starting from
some vertex. Hence, the LLS sequences coincide as well. 

Proof of Lagrange’s theorem First, let us show that every periodic continued frac-
tion represents a quadratic irrationality. Since the continued fraction is infinite, the
corresponding number is irrational.
Suppose that the periodic continued fraction for ξ does not have a preperiod, i.e.,
 
ξ = (a0 ; a1 : · · · : an ) .

Then
pn ξ + pn−1
ξ = [a0 ; a1 : · · · : an : ξ ] = .
qn ξ + qn−1
96 8 Lagrange’s Theorem

(The last equality holds by Theorem 1.13.) Notice that the denominator qn−1 ξ +
qn−2 is nonzero, since ξ is irrational. Therefore ξ satisfies

qn−1 ξ 2 + (qn−1 − pn−1 )ξ − pn−2 = 0.

Hence ξ is a quadratic irrationality.

Suppose now that
 
ξ = a0 ; a1 : · · · : an : (an+1 : an+2 : · · · : an+m ) .

Set
 
ξ̂ = (an+1 ; an+2 : · · · : an+m ) .
Then by Theorem 1.13 we have

pn ξ̂ + pn−1
ξ = [a0 ; a1 : · · · : an : ξ̂ ] = .
qn ξ̂ + qn−1

We have already shown that ξ̂ is a quadratic irrationality. Hence ξ is a quadratic

irrationality as well.
Secondly, we prove that every quadratic irrationality has a periodic continued
fraction. Let ξ > 1 be a quadratic irrationality. By Lemma 8.12 there exists an
SL(2, Z) matrix A with an eigenvector (1, ξ ). By Theorem 7.18 the geometric con-
tinued fraction of A has periodic LLS sequence. The LLS sequence for the angle α
generated by the two vectors (1, 0) and (1, ξ ) is one-side infinite, and by Proposi-
tion 8.13, from some element it coincides (up to a shift) with the LLS sequence for
the matrix A. Hence, the LLS sequence for α is periodic from some point on. Hence
from Corollary 3.5 and Theorem 3.6, the continued fraction for ξ is periodic.
Suppose now that ξ < 1. The number ξ̂ = ξ − ξ  + 1 is quadratic and greater
than 1, and hence it is periodic by the above. The continued fractions for ξ and ξ̂
are distinct only in the first element. Hence the continued fraction for ξ is periodic
as well. 

8.5 Exercises

Exercise 8.1 Prove Proposition 8.2.

Exercise 8.2 Prove that the two Dirichlet groups of real spectrum matrices with
rational eigenvalues

1 0 1 −1
;
0 −1 0 −1
are not integer congruent.
8.5 Exercises 97

Exercise 8.3 Let A be a (2 × 2) matrix with distinct eigenvalues. Then every matrix
commuting with A has the same eigenlines.

Exercise 8.4 Do the following matrices have a cube root in GL(2, Z):

25 32 89 144
; ?
32 41 144 233

Exercise 8.5 Solve explicitly Pell’s equation

m2 − dn2 = 1

for d = 2, 3, 5.
√
Exercise 8.6 Find an SL(2, Z) matrix with eigenvector (1, 11).
Chapter 9
Gauss–Kuzmin Statistics

It turns out that the frequency of a positive integer k in a continued fraction for
almost all real numbers is equal to

1 1
ln 1 + ,
ln 2 k(k + 2)

i.e., for a general real x we have 42 % of 1, 17 % of 2, 9 % of 3, etc. This distribution

is traditionally called the Gauss–Kuzmin distribution. The statistics of the elements
in continued fractions first appeared in the letters of K.F. Gauss to P.S. Laplace at
the beginning of the nineteenth century (see [63]). The first proof with additional
estimates was developed by R.O. Kuzmin in [121] in 1928 (see also [122]), and a
little later, another proof with new estimates was given by P. Lévy in [131]. Further
investigations in this directions were made by E. Wirsing in [209].
In this chapter we describe two strategies to study distributions of elements in
continued fractions. A classical approach to the Gauss–Kuzmin distribution is based
on the ergodicity of the Gauss map. The second approach is related to the geometry
of continued fractions and its projective invariance. It is interesting to note that the
frequencies of elements has an unexpected interpretation in terms of cross-ratios
(see Remark 9.31). Unfortunately, the classical approach does not have a general-
ization to the case of multidimensional sails, since it is not clear what map is the
multidimensional analogue of the Gauss map. We avoid this problem by using a
geometric approach to define and investigate multidimensional statistical questions
for multidimensional sails. We describe the multidimensional case in Chap. 19.
In the first five sections of this chapter we discuss the classical ergodic approach
to the Gauss–Kuzmin distribution. In Sect. 9.1 we give some basic notions and defi-
nitions of ergodic theory. Further, in Sects. 9.2 and 9.3, we present a measure related
to continued fractions and the Gauss map. We prove the pointwise Gauss–Kuzmin
theorem and formulate the original Gauss–Kuzmin theorem in Sects. 9.4 and 9.5
respectively.
In the last five sections we study the statistic of edges of geometric one-
dimensional continued fractions. After a brief discussion of cross-ratios (Sect. 9.6)

O. Karpenkov, Geometry of Continued Fractions, 99

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_9, © Springer-Verlag Berlin Heidelberg 2013
100 9 Gauss–Kuzmin Statistics

we define a structure of a smooth manifold on the set of geometric continued frac-

tions CF 1 in Sect. 9.7. Further, in Sect. 9.8, we define the Möbius measure on CF 1
which is invariant under the group PGL(R, 2) acting on CF 1 ; we write the Möbius
form explicitly in Sect. 9.9. Finally, in Sect. 9.10, we define related frequencies of
edges of continued fractions and show that they coincide with the Gauss–Kuzmin
statistics of the elements of continued fractions.

9.1 Some Information from Ergodic Theory

Let X be a set, Σ a σ -algebra on X, and μ a measure on the elements of Σ . The

collection (X, Σ, μ) is called a measure space. If μ(X) = 1, the measure space is
called a probability measure space.
Given a transformation T of a set X to itself, for any μ-integrable function f on
X one can define the time average for f at the point x to be

1  k 
n−1
lim f T x .
n→∞ n
k=0

The space average If is


1
If = f dμ.
μ(X)
The space average always exists. The time average does not exist for all x. Neverthe-
less, in the case in which we are interested, it exists for almost all x. We formulate
the related theorem after one important definition.

Definition 9.1 Let (X, Σ, μ) be a measure space. A transformation T : X → X is

measure-preserving if it is measurable and
 
μ T −1 (A) = μ(A)

for every set A of Σ .

For measure-preserving transformations we have the following theorem.

Theorem 9.2 (Birkhoff’s Pointwise Ergodic Theorem) Consider an arbitrary mea-

sure space (X, Σ, μ) and a measure preserving transformation T on X. Let f be a
μ-integrable function on X. Then the time average converges almost everywhere to
an invariant function f .

Definition 9.3 Consider a probability measure space (X, Σ, μ). A measure-

preserving transformation T on X is ergodic if for every X  ∈ Σ satisfying
T −1 (X  ) = X  , either μ(X  ) = 0 or μ(X  ) = 1.
9.2 The Measure Space Related to Continued Fractions 101

Theorem 9.4 (Birkhoff–Khinchin’s Ergodic Theorem) Consider a probability

measure space (X, Σ, μ) and a measure preserving transformation T . Suppose
that T is ergodic. Then the values of the time average function are equal to the
space average (i.e., f (x) = If ) almost everywhere.

9.2 The Measure Space Related to Continued Fractions

In this section we define a measure space that is closely related to distributions of

the elements of continued fractions. For this measure we formulate a statement on
the density of points for measurable subsets, which we use in the essential way in
the proofs below.

9.2.1 Definition of the Measure Space Related to Continued

Fractions

Consider the measure space of the segment I = {x | 0 ≤ x < 1} with the Borel σ -
algebra Σ and the measure μ̂ defined on a measurable set S as

1 dx
μ̂(S) = .
ln 2 S 1 + x

The coefficient 1/ ln 2 is taken such that the measure of the segment I equals 1.

9.2.2 Theorems on Density Points of Measurable Subsets

We start with a classical theorem on Lebesgue measure space. Denote by B(x, ε)

the standard ball of radius ε centered at x.

Theorem 9.5 (Lebesgue density) Let λ be the n-dimensional Lebesgue measure

on Rn . If A ⊂ Rn is a Borel measurable set, then almost every point x ∈ A is a
Lebesgue density point:
λ(A ∩ B(x, ε))
lim = 1.
ε→0 λ(B(x, ε))

Here “almost every point” means “except for a subset of zero measure”.
The measure μ̂ is equivalent to the one-dimensional Lebesgue measure λ on the
segment [0, 1] (for more information on measure theory, see [137]). Hence we have
a similar statement in the case of the measure space (X, Σ, μ̂).
102 9 Gauss–Kuzmin Statistics

Fig. 9.1 The Gauss map

Corollary 9.6 (μ̂-density) Let X = [0, 1] and let μ̂ be as above. If A ⊂ X is a μ̂-

measurable set with positive measure μ̂(A), then almost every point in A satisfies

μ̂(A ∩ B(x, ε))

lim = 1.
ε→0 μ̂(B(x, ε))

9.3 On the Gauss Map

Let us introduce a transformation whose ergodic properties will form the basis for
the proof of the Gauss–Kuzmin theorem.

9.3.1 The Gauss Map and Corresponding Invariant Measure

We consider the measure space (X, Σ, μ̂) defined in the previous section. Define
the Gauss map T of a segment [0, 1] to itself as follows:

T (x) = {1/x},

where {r} denotes the fractional part r − r (see Fig. 9.1).

Proposition 9.7 The Gauss map T is measure-preserving for the measure space
(X, Σ, μ̂).

We start with the following lemma.

Lemma 9.8 Let x = [0; a1 : a2 : · · · ]. Then

 
−1
  1 
T (x) = [0; k : a1 : a2 : · · · ]|k ∈ Z+ = k ∈ Z + .
x +k
9.3 On the Gauss Map 103

Proof The first equality follows directly from the fact that
 
T [0; b1 : b2 : · · · ] = [0; b2 : b3 · · · ]

and the fact that every real number has a unique regular continued fraction expansion
with the last element not equal to 1.
The second equality is straightforward. 

Proof of Proposition 9.7 Consider a measurable set S. From Lemma 9.8 it follows
that

  1 dx
μ̂ T −1 (S) =
ln 2 T −1 (S) 1 + x
∞ 

1  dx
= .
ln 2 T −1 (S)∩[1/(k+1),1/k] 1 + x
k=1

Notice that on each open (i.e., without the boundary points) segment ]1/k,
1/(k + 1)[ the operator T is in one-to-one correspondence with the open segment
]0, 1[. Let us denote the inverse function to T on the segment ]1/k, 1/(k + 1)[ by
−1
T(k) . Therefore,


−1 1 1  −1 
T T (S) ∩ , = T T(k) (S) = S,
k+1 k

and we can apply the rule of differentiation of a composite function. From

Lemma 9.8 we know, that

−1 1
T(k) (x) = .
x +k
Then we have
   −1
dT(k) (x)
dx dx
= = −1
T −1 (S)∩[1/(k+1),1/k] 1+x −1
T(k) (S) 1+x −1
T (T(k) (S)) 1 + T(k) (x)
 
−d(1/(x + k)) dx
= =
S 1 + 1/(x + k) S (x + k)(x + k + 1)

−1
(the negative sign is taken, since the map T(k) : x → x+k
1
changes the orientation).
So we have
∞ 
 −1  1  dx
μ̂ T (S) = .
ln 2 S (x + k)(x + k + 1)
k=1
104 9 Gauss–Kuzmin Statistics

Since the integrated functions are nonnegative, we can change the order of the sum-
mation and the integration operations. We get
  ∞

 −1  1 1
μ̂ T (S) = dx
ln 2 S (x + k)(x + k + 1)
k=1
  ∞
 
1 1 1 1 dx
= − dx = = μ̂(S).
ln 2 S x +k x +k+1 ln 2 S x + 1
k=1

So for every measurable set S we have

 
μ̂ T −1 (S) = μ̂(S).

Therefore, the Gauss map T preserves the measure μ̂. 

Remark 9.9 (On the Euler–Mascheroni constant) By definition, the Euler–

Mascheroni constant (traditionally denoted by γ ) is the following infinite sum
 n 
1
γ = lim − ln n .
n→∞ k
k=1

It was first studied by L. Euler in 1734. It is not known whether γ is irrational. It

turns out that the Euler–Mascheroni constant can be expressed as an integral of the
Gauss map with respect to Lebesgue measure:
 1
γ =1− T (x)dx.
0

9.3.2 An Example of an Invariant Set for the Gauss Map

Let us consider one example of a measurable set that is invariant under the Gauss
map.
Denote by Ψ the set of all irrational numbers in the segment [0, 1] whose con-
tinued fractions contain only finitely many 1’s. It is clear that

T −1 (Ψ ) = Ψ,

since the operation T −1 shifts elements of continued fractions by one and inserts
the first element.

Proposition 9.10 The set Ψ is measurable (i.e., Ψ ∈ Σ ).

Proof Denote by Υn the set of all irrational numbers that contain the element 1
exactly at place n. Notice that

Υ1 = [1/2, 1],
9.3 On the Gauss Map 105

and therefore, it is measurable. Hence for every n the set

Υn+1 = T n (Υ1 )

is measurable.
Denote by Ψ0 the set of all irrational numbers that do not contain an element ‘1’.
Since
∞

Ψ0 = X \ Υn ,
n=1
the set Ψ0 is also measurable. Then

T −n (Ψ0 )

is measurable for any positive integer n. Hence

∞

Ψ= T −n (Ψ0 )
n=1

is measurable. 

We will prove later that the Gauss map is ergodic, and therefore, Ψ is either of
zero measure or full measure in X.

9.3.3 Ergodicity of the Gauss Map

In this subsection we prove the ergodicity of the Gauss map.

Proposition 9.11 The Gauss map is ergodic.

Before proving Proposition 9.11 we introduce some supplementary notation and

prove two lemmas.
For a sequence of positive integers (a1 , . . . , an ) denote by I(a1 ,...,an ) the segment
with endpoints [0; a1 : · · · : an−1 : an ] and [0; a1 : · · · : an−1 : an + 1]. It is clear that
the map
T n : I(a1 ,...,an ) → [0, 1]
is one-to-one on the segment I(a1 ,...,an ) , and the inverse to T n is

T(a−11 ,...,an ) : x → [0; a1 : · · · : an : 1/x].

In terms of k-convergents pk /qk = [0; a1 : · · · : ak ], the expression for T(a−11 ,...,an ) (x)
is as follows (see Proposition 1.13):
106 9 Gauss–Kuzmin Statistics

pn /x + pn−1 pn + pn−1 x
T(a−11 ,...,an ) (x) = = .
qn /x + qn−1 qn + qn−1 x

Lemma 9.12 The measure of a segment I(a1 ,...,an ) satisfies the following inequality:
1
μ̂(I(a1 ,...,an ) ) < .
ln 2(qn + qn−1 )(pn + qn )

Proof We have
  
1 dx 1  [0;a1 :···:an :1] dx 
μ̂(I(a1 ,...,an ) ) = =
ln 2 I(a1 ,...,an ) 1 + x ln 2  [0;a1 :···:an ] 1 + x 



1  pn + pn−1  pn 

=  ln 1 + 1+ 
ln 2 qn + qn−1 qn


1  1 

= ln 1 +
ln 2  (qn + qn−1 )(pn + qn ) 
1
< .
ln 2(qn + qn−1 )(pn + qn )
The last inequality follows from the concavity of the natural logarithm function. 

Lemma 9.13 For any invariant set S of positive measure and any interval I(a1 ,...,an ) ,
ln 2
μ̂(S ∩ I(a1 ,...,an ) ) ≥ μ̂(S)μ̂(I(a1 ,...,an ) ).
2
Proof Since the map T is surjective, we also have

T (S) = S.

Let μ̂(S) = c > 0. Let us prove that c = 1. We have

 

1 dx 1 pn + pn−1 x  pn + pn−1 x
= d 1+
ln 2 S∩I(a ,...,an ) 1 + x ln 2 S qn + qn−1 x qn + qn−1 x
1

1 dx
=
ln 2 S (qn + qn−1 x)(qn + qn−1 x + pn + pn−1 x)

1 dx
≥
ln 2 · qn (qn + qn−1 + pn + pn−1 ) S 1 + x
1
= μ̂(S)
qn (qn + qn−1 + pn + pn−1 )
1 ln 2
≥ μ̂(S) ≥ μ̂(S)μ̂(I(a1 ,...,an ) ).
(qn + qn−1 )(2pn + 2qn ) 2
The last inequality follows from Lemma 9.12. 
9.4 Pointwise Gauss–Kuzmin Theorem 107

Proof of Proposition 9.11 Let S be a measurable subset of the unit interval such that
T −1 (S) = S. Suppose also μ̂(S) > 0.
For any irrational number y = [0; a1 : a2 : · · · ], from Lemma 9.13 we have

μ̂(X \ S) ∩ B(y, μ̂(I(a1 ,...,an ) )) ln 2 μ̂(S)μ̂(I(a1 ,...,an ) ) ln 2

≤1− =1− μ̂(S).
μ̂(B(y, μ̂(I(a1 ,...,an ) ))) 2 μ̂(B(y, μ̂(I(a1 ,...,an ) ))) 4

Hence y is not a μ̂-density point of X \ S. Therefore, by Corollary 9.6 almost every

point of [0, 1] \ Q is not in X \ S, and therefore, it is in S. Hence
 
μ̂(S) ≥ μ̂ [0, 1] \ Q = 1

Hence μ̂(S) = 1, concluding the proof of ergodicity of T . 

9.4 Pointwise Gauss–Kuzmin Theorem

Consider x in the segment [0, 1]. Let the regular continued fraction for x be [0; a1 :
· · · : an ] (odd or infinite). For a positive integer k, set

#(k, n)
P̂n,k (x) = ,
n
where #(k, n) is the number of integer elements ai equal to k for i = 1, . . . , n. Define

P̂k (x) = lim P̂n,k (x).

n→∞

Theorem 9.14 For every positive integer k and almost every x (i.e., in the comple-
ment of a set of zero measure) the following holds:

1 1
P̂k (x) = ln 1 + .
ln 2 k(k + 2)

We consider this theorem a pointwise Gauss–Kuzmin theorem. To prove this

pointwise Gauss–Kuzmin theorem we use Birkhoff’s ergodic theorems.

Proof of Theorem 9.14 Consider a subset S ∈ I . Let χS be the characteristic func-

tion of S, i.e.,

1, if x ∈ S,
χS (x) =
0, otherwise.
Then
1
n−1
 
P̂n,k (x) = χ]1/(k+1),1/k] T s x .
n
s=0
108 9 Gauss–Kuzmin Statistics

Hence, by Birkhoff’s pointwise ergodic theorem, the limit P̂k (x) exists almost ev-
erywhere. Since the transformation T is ergodic, we apply the Birkhoff–Khinchin
ergodic theorem and get
 1  1/k

1 dx 1 1
P̂k (x) = χ]1/(k+1),1/k] d μ̂ = = ln 1 + .
0 ln 2 1/(k+1) 1 + x ln 2 k(k + 2) 

9.5 Original Gauss–Kuzmin Theorem

Let α be some irrational number between zero and one, and let [0; a1 : a2 : a3 : · · · ]
be its regular continued fraction.
Let mn (x) denote the measure of the set of real numbers α contained in the
segment [0, 1] such that T n (α) < x (here T is the Gauss map). In his letters to
P.S. Laplace C.F. Gauss formulated without proofs the following theorem.

Theorem 9.15 (Gauss–Kuzmin) For 0 ≤ x ≤ 1 the following holds:

ln(1 + x)
lim mn (x) = .
n→∞ ln 2

This theorem is technically complicated. For the proof we refer to the original
manuscripts of R.O. Kuzmin [121] and [122] (see also A.Ya. Khinchin [105]).
Denote by Pn (k), for an arbitrary integer k > 0, the measure of the set of all real
numbers α of the segment [0, 1] such that each of them has the number k at the nth
position. The limit limn→∞ Pn (k) is called the frequency of k for regular continued
fractions and is denoted by P (k).

Corollary 9.16 For every positive integer k, the following holds:

1 1
P (k) = ln 1 + .
ln 2 k(k + 2)

Proof Notice that Pn (k) = mn ( k1 ) − mn ( k+1

1
). Now the statement of the corollary
follows from the Gauss–Kuzmin theorem. 

9.6 Cross-Ratio in Projective Geometry

In this section we switch to the multidimensional case for a while in order to give
some definitions that are similar to the one-dimensional case (we will use these later,
in Chap. 19).
9.6 Cross-Ratio in Projective Geometry 109

9.6.1 Projective Linear Group

The projective linear group (or the group of projective transformations) is the quo-
tient group
PGL(R, n) = GL(R, n)/Z(R, n),
where Z(R, n) is the one-dimensional subgroup of all nonzero scalar transforma-
tions of Rn . The group PGL(R, n) acts on the equivalence classes of vectors in Rn
with respect to Z(R, n). We have

Rn /Z(R, n) = RP n−1 .

Consider the affine part Rn−1 ⊂ RP n−1 . The stabilizer for the affine part is exactly
the group Aff(R, n − 1).

9.6.2 Cross-Ratio, Infinitesimal Cross-Ratio

Consider an arbitrary line in Rn−1 with a Euclidean coordinate on it.

Definition 9.17 Consider a 4-tuple of points on a line with coordinates z1 , z2 , z3 ,

and z4 . The value
(z1 − z3 )(z2 − z4 )
(z2 − z3 )(z1 − z4 )
is called the cross-ratio of the 4-tuple.

It is clear that the cross-ratio does not depend on the choice of the Euclidean coor-
dinate on the line, and therefore, it a function on the space of ordered 4-tuples of
distinct points in a line.
As we have already noted above, the space Rn−1 can be considered an affine
chart Rn−1 ⊂ RP n−1 . Hence the action of PGL(R, n) is well defined on the closure
of Rn−1 (which is actually RP n−1 ). The projective transformations take planes to
planes and, in particular, lines to lines. So it is natural to ask what happens to the
cross-ratios of four points on a line.

Proposition 9.18 The cross-ratio of four points is an invariant of projective trans-

formations of Rn .

We are interested in the infinitesimal cross-ratio, which is the following 2-form:

dxdy
.
(x − y)2
110 9 Gauss–Kuzmin Statistics

Notice that in the denominator we have

  
(x − y)2 = lim x − (y + εdy) · (x + εdx) − y .
ε→0

Corollary 9.19 The infinitesimal cross-ratio is an invariant of projective transfor-

mations of Rn−1 .

Proof The density of the infinitesimal cross-ratio coincides with the asymptotic co-
efficient at ε 2 of the cross-ratios of 4-tuples of points:

x, y, x + εdx, y + εdy,

as ε tends to 0. Therefore, the infinitesimal cross-ratio is a projective invariant. 

9.7 Smooth Manifold of Geometric Continued Fractions

Denote the set of all geometric continued fractions by CF 1 . Consider an arbitrary

element of CF 1 . It is a continued fraction defined by an (unordered) pair of nonpar-
allel lines (1 , 2 ) passing through integer points.
Denote the sets of all ordered collections of two independent and dependent
straight lines by FCF 1 and Δ1 respectively. We say that FCF 1 is a space of geo-
metric framed continued fractions. We have
 
FCF 1 = RP 1 × RP 1 \ Δ1 = T 2 \ Δ1 and CF 1 = FCF 1 /(Z/2Z),

where Z/2Z is the group transposing the lines in geometric continued fractions.
Note that FCF 1 is a 2-fold covering of CF 1 . We call the map of “forgetting” the
order in the ordered collections the natural projection of the manifold FCF 1 to the
manifold CF 1 and denote it by p (i.e., p : FCF 1 → CF 1 ).
Notice that FCF 1 is homeomorphic to the annulus and CF 1 is homeomorphic to
the Möbius band.

9.8 Möbius Measure on the Manifolds of Continued Fractions

The group PGL(2, R) of transformations of RP 1 takes the set of all straight lines
passing through the origin in the plane into itself. Hence, PGL(2, R) naturally acts
on CF 1 and FCF 1 . It is clear that the action of PGL(2, R) is transitive, i.e., it takes
any (framed) continued fraction to any other. Notice that a stabilizer of any geomet-
ric continued fraction is one-dimensional.

Definition 9.20 A form on the manifold CF 1 (respectively FCF 1 ) is said to be a

Möbius form if it is invariant under the action of PGL(2, R).
9.9 Explicit Formulas for the Möbius Form 111

Remark 9.21 The name for the invariant forms comes from theory of energies of
knots and graphs in low dimensional topology, where these forms are used as den-
sities for Möbius energies that are invariant under the group of Möbius transforma-
tions in R3 (we refer the interested reader to the book by J. O’Hara [150]).

Proposition 9.22 All Möbius forms of the manifolds CF 1 and FCF 1 are propor-
tional.

Proof Transitivity of the action of PGL(2, R) implies that all Möbius forms of the
manifolds CF 1 and FCF 1 are proportional. 

Let ω be some volume form of the manifold M. Denote by μω a measure of the

manifold M that for every open measurable set S contained in the same piecewise
connected component of M is defined by the equality
 
 
μω (S) =  ω.
S

Definition 9.23 A measure μ of the manifold CF 1 (FCF 1 ) is said to be a Möbius

measure if there exists a Möbius form ω of CF 1 (FCF 1 ) such that μ = μω .

From Proposition 9.22 we have the following.

Corollary 9.24 Any two Möbius measures are proportional.

Remark 9.25 The projection p takes the Möbius measures of the manifold FCF 1
to the Möbius measures of the manifold CF 1 . This establishes an isomorphism
between the spaces of Möbius measures for CF 1 and FCF 1 . Since the manifold
of framed continued fractions possesses simpler chart systems, all formulas of the
work are given for the case of the framed continued fraction manifold. To calculate
a measure of some set F of the unframed continued fraction manifold, one should:
take p −1 (F ); calculate the Möbius measure of the obtained set of the manifold of
framed continued fractions, and divide the result by 2.

9.9 Explicit Formulas for the Möbius Form

Let us write down Möbius forms of the framed one-dimensional continued fraction
manifold FCF 1 explicitly in special charts.
Consider a vector space R2 equipped with standard metrics on it. Letting l be
an arbitrary straight line in R2 that does not pass through the origin, choose some
Euclidean coordinates Ol Xl on it. Denote by FCF 1,l a chart of the manifold FCF 1
that consists of all ordered pairs of straight lines both intersecting l. Let us associate
to any point of FCF 1,l (i.e. to a collection of two straight lines) coordinates (xl , yl ),
where xl and yl are the coordinates on l for the intersections of l with the first and
112 9 Gauss–Kuzmin Statistics

the second straight lines of the collection respectively. Denote by |v|l the Euclidean
length of a vector v in the coordinates Ol Xl Yl of the chart FCF 1,l . Note that the
chart FCF 1,l is the space R × R minus its diagonal.
Consider the following form in the chart FCF 1,l :
dxl ∧ dyl
ωl (xl , yl ) = .
|xl − yl |2l

Proposition 9.26 The measure μωl coincides with the restriction of some Möbius
measures to FCF 1,l .

Proof Notice that the form ωl (xl , yl ) coincides with the infinitesimal cross-ratio on
the line l. Hence it is invariant under projective transformations of l (on an every-
where dense subset) in the chart FCF 1,l . Therefore, the measure μωl coincides with
the restriction of some Möbius measures to FCF 1,l . 

Corollary 9.27 A restriction of an arbitrary Möbius measure to the chart FCF 1,l is
proportional to μωl .

Proof The statement follows from the proportionality of any two Möbius mea-
sures. 

Consider now the manifold FCF 1 as a set of ordered pairs of distinct points on
a circle R/πZ (this circle is a one-dimensional projective space obtained from the
unit circle by identifying antipodal points). The doubled angular coordinate ϕ of the
circle R/πZ induced by the coordinate x of the straight line R naturally defines the
coordinates (ϕ1 , ϕ2 ) of the manifold FCF 1 .

Proposition 9.28 The form ωl (xl , yl ) is extendible to some form ω1 of FCF 1 . In

coordinates (ϕ1 , ϕ2 ), the form ω1 can be written as follows:

1 ϕ1 − ϕ2
ω1 = cot2 dϕ1 ∧ dϕ2 .
4 2

The proof of Proposition 9.28 is left as an exercise for the reader.

9.10 Relative Frequencies of Edges of One-Dimensional

Continued Fractions
Without loss of generality, in this section we consider only the Möbius form ω1 of
Proposition 9.28. Denote the natural projection of the form μω1 to the manifold of
one-dimensional continued fractions CF 1 by μ1 .
Consider an arbitrary segment F with vertices at integer points. Denote by
CF 1 (F ) the set of continued fractions that contain the segment F as an edge.
9.10 Relative Frequencies of Edges of One-Dimensional Continued Fractions 113

Fig. 9.2 Rays defining a

continued fraction should lie
in the domain shaded in gray

Definition 9.29 The quantity μ1 (CF 1 (F )) is called the relative frequency of the
edge F .

Note that the relative frequencies of edges of the same integer-linear type are
equivalent. Every edge of a one-dimensional continued fraction is at unit integer
distance from the origin. Thus, the integer-linear type of a segment is defined by its
integer length (the number of inner integer points plus unity). Denote the relative
frequency of the edge of integer length k by μ1 ( k  ).

Proposition 9.30 For every positive integer k, the following holds:

   1
μ1 k = ln 1 + .
k(k + 2)

Proof Consider a particular representative of an integer-linear type of a length-k

segment: the segment with vertices (0, 1) and (k, 1). The one-dimensional continued
fraction contains the segment as an edge if and only if one of the straight lines
defining the fraction intersects the interval with vertices (−1, 1) and (0, 1) while
the other straight line intersects the interval with vertices (k, 1) and (k + 1, 1) (see
Fig. 9.2).
For the straight line l defined by the equation y = 1, we calculate the Möbius
measure of the Cartesian product of the described pair of intervals. By the last sec-
tion it follows that this quantity coincides with the relative frequency μ1 ( k  ). So,
 0  k+1  k+1

   dxl dyl 1 1

μ1 k = = − dyl
−1 k (xl − yl )2 k yl yl + 1

(k + 1)(k + 1) 1
= ln = ln 1 + .
k(k + 2) k(k + 2)
This proves the proposition. 
(k+1)(k+1)
Remark 9.31 Note that the argument of the logarithm k(k+2) is the cross-ratio
of points (−1, 1), (0, 1), (k, 1), and (k + 1, 1).

Corollary 9.32 The relative frequency μ1 ( k  ), up to the factor

 0  +∞
dxl dyl
ln 2 = ,
−1 1 (xl − yl )2
coincides with the Gauss–Kuzmin frequency P (k).
114 9 Gauss–Kuzmin Statistics

9.11 Exercises

Exercise 9.1
(a) Prove that the measure

1 dx
μ(S) =
ln 2 S 1+x
is a probability measure on the segment [0, 1], i.e., μ([0, 1]) = 1.
(b) Find μ([a, b]) for 0 ≤ a < b ≤ 1, where μ is as above.

Exercise 9.2 Ergodicity of the doubling map. Consider the space (S 1 , Σ, λ), where
X is the unit circle, Σ is the Borel σ -algebra, and λ is the Lebesgue measure. Con-
sider the doubling map T : S 1 → S 1 such that

T (ϕ) = 2ϕ.

Prove that T is measure-preserving and ergodic.

Exercise 9.3 Define the frequencies of subsequences in continued fractions. What

is the frequency of the sequence (1, 2, 3)?

Exercise 9.4 Prove the μ̂-density theorem from the Lebesgue density theorem.

Exercise 9.5 Recall that Ψ0 is the subset irrational numbers in [0, 1] whose contin-
ued fractions do not contain 1 as an element. Prove by elementary means (without
using ergodic theorems) that
μ̂(Ψ0 ) = 0.

Exercise 9.6 Prove the projective invariance of the cross-ratio.

Exercise 9.7 Prove that any two triples of points on a line are projectively equiva-
lent. Find a criterion for two 4-tuples of points on a line to be projectively equivalent.

Exercise 9.8 Prove that

(a) RP 1 is homeomorphic to a circle;
(b) FCF 1 is homeomorphic to an annulus;
(c) CF 1 is homeomorphic to the Möbius band.

Exercise 9.9 Prove Proposition 9.28.

Chapter 10
Geometric Aspects of Approximation

The approximation properties of continued fractions have attracted researchers for

centuries. There are many different directions of investigation in this important sub-
ject (the study of best approximations, badly approximable numbers, etc.). In this
chapter we consider two geometric questions of approximations by continued frac-
tions. In Sect. 10.1 we prove two classical results on best approximations of real
numbers by rational numbers. Further, in Sect. 10.2, we describe a rather new branch
of generalized Diophantine approximations concerning arrangements of two lines in
the plane passing through the origin. In this chapter we use some material related
to basic properties of continued fractions of Chap. 1, to geometry of numbers of
Chap. 3, and to Markov–Davenport forms of Chap. 7.

10.1 Two Types of Best Approximations of Rational Numbers

In this section we study the problem of best approximations of real numbers by

rational numbers. We prove here a classical result that all best approximations p/q
are represented by the points (p, q) of the corresponding sails. Conversely, the point
(p, q) of the sails is a best approximation if the denominator q satisfies the so-
called half-rule. Further, we discuss strong best approximations, which are exactly
convergents of the corresponding continued fractions.

10.1.1 Best Diophantine Approximations

We say that a rational number p/q (where q > 0) is a best approximation of a real
number α if for any other fraction p  /q  with 0 < q  < q we get
   
   
α − p  > α − p .
 q   q

O. Karpenkov, Geometry of Continued Fractions, 115

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_10, © Springer-Verlag Berlin Heidelberg 2013
116 10 Geometric Aspects of Approximation

Remark 10.1 By the definition of best approximations, every integer is a best ap-
proximation of every real number α. To be more consistent we should say that the
best approximations with integer denominator are α and α + 1.

Remark 10.2 From the definition of best approximations it is not immediate that
there are no two best approximations with the same denominator (greater than 1).
This is a direct corollary of Theorem 10.3 below.

Theorem 10.3 Let α be a positive real number with regular continued fraction
[a0 ; a1 : · · · ]. Denote by pk /qk its k-convergent and by rk the remainder part
[ak ; ak+1 : · · · ]. A rational number p/q (with q > 1) is a best approximation of
α if and only if there exist k ≥ 0 and an integer x such that

p/q = [a0 ; a1 : · · · : ak : x],

where x satisfies either

 
ak+1
< x ≤ ak+1
2
or alternatively x = ak+1 /2 in the specific case in which ak+1 is divisible by 2 and
α satisfies the inequality
qk+1
rk+1 <
qk
(which is sometimes called the half-rule condition).

Remark 10.4 The condition

qk+1
rk+1 <
qk
is equivalent to the condition

[ak+1 ; ak+2 : · · · ] < [ak+1 ; ak : · · · : a0 ].

Before proving this theorem, we need to prove several preparatory statements.

Proposition 10.5 Consider an arbitrary nonzero real number α. Then all its con-
vergents are best approximations for α.

Proof Let pk /qk be the k-convergent to α (we assume that the integers pk and qk
are relatively prime). We prove the proposition by induction on k.
Base of induction. If the denominator qk equals 1, then we have either α for
k = 0 or sometimes α + 1 for k = 1 when a1 = 1. As we have agreed in Re-
mark 10.1, these two numbers are best approximations.
Step of induction. Suppose that the (k − 1)-convergent pk−1 /qk−1 is a best ap-
proximation. Let us prove that the k-convergent pk /qk is a best approximation (we
10.1 Two Types of Best Approximations of Rational Numbers 117

Fig. 10.1 There is no integer

points in the strip between the
lines l1 and l2

suppose also that qk > 1). Set

A = (pk−1 , qk−1 ) and B = (pk , qk ).

From Theorem 3.1 it follows that A and B are vertices of two different sails for two
angles defined by the ray y = αx together with two positive coordinate rays.
Denote by l1 the line passing through the points O and A, and by l2 the line
passing through the points B and parallel to l1 . Draw the line through B parallel
to the line x = 0 and denote the intersection of this line with the ray OA by C (see
Fig. 10.1).
From Proposition 1.15 it follows that the triangle AOB is empty. Hence, there
are no integer points in the strip between the lines l1 and l2 (which is gray in
Fig. 10.1). Thus every integer point of the triangle BOC is either in the line OC or
coincides with B. Therefore, every integer point (p, q) with qk−1 ≤ q < qk satisfies
     
 
p       
 − α  > min  pk−1 − α ,  pk − α  =  pk − α 
q  q  q  q 
k−1 k k

(the last equality follows from Proposition 1.19). By induction, pk−1 /qk−1 is a best
approximation. Hence if q < qk−1 , then
     
p     
 − α  >  pk−1 − α  >  pk − α 
q  q  q 
k−1 k

(the second inequality follows from Proposition 1.19). Therefore, the k-convergent
pk /qk is a best approximation. The induction step is complete. 

Lemma 10.6 Consider a nonzero real number α > 0 with regular continued frac-
tion [a0 ; a1 : · · · ]. Let pi /qi be its i-convergent for i = 0, 1, 2, . . . . Then for every
k ≥ 1 every best approximation p/q of α satisfying

qk−1 < q ≤ qk
118 10 Geometric Aspects of Approximation

lies on the line passing through the point (pk , qk ) and parallel to the vector
(pk−1 , qk−1 ) (it is the line l2 in Fig. 10.1).

Proof Let A = (pk−1 , qk−1 ) and B = (pk , qk ). Suppose that the lines l1 and l2 are
as above (see Fig. 10.1).
Consider E(p, q). If ld(E, OA) > 1, then the triangle EOA contains at least one
integer point E  (p  , q  ) incident with EOA distinct from E and not contained in
AO. First, it is clear that q  < q, since E has the greatest second coordinate among
all the points of the triangle OAE. Secondly, E and E  are in the same half-plane
with respect to the line y = αx. Hence,
    
p   
 − α  >  p − α .
q   q 

Therefore, p/q is not a best approximation of α.

Suppose now that the point E is in the same half-plane with the point A with
respect to the line y = αx. Then the angle between OA and y = αx is not greater
than the angle between OE and y = αx. Therefore, E is not a best approximation.
So if E is a best approximation satisfying the conditions of the lemma, then
ld(E, OA) = 1 and E is in the same half-plane with the point B. Therefore, E is on
the line l2 . 

Corollary 10.7 All best approximations of α = [a0 ; a1 : · · · ] satisfying

qk < q ≤ qk+1

are of the form [a0 ; · · · : ak : m], where m is a positive integer not greater than ak .

Proof From Lemma 10.6 we know that all best approximations satisfying the con-
dition of the corollary are on a line l2 . All integer points of this line are of the form
B + mOA for an integer m, or in other words,

(pk , qk ) + m(pk−1 , qk−1 ).

From Proposition 1.13 it follows that

pk + mpk−1
= [a0 ; · · · : ak : m].
qk + mqk−1

For m ≤ 0 the point B + mOA is not in the first octant. For m > ak the second coor-
dinate of the point B + mOA is greater than qk . Therefore, we have 0 < m ≤ ak . 

Proof of Theorem 10.3 By Corollary 10.7 we know that all best approximations are
of the form
[a0 ; · · · : ak : x],
10.1 Two Types of Best Approximations of Rational Numbers 119

where 0 < x ≤ ak+1 . By Proposition 10.5 we get that all the convergents (i.e., the
case x = ak+1 ) are best approximations. It is easy to see that

[a0 ; · · · : ak : x] > [a0 ; · · · : ak : x + 1]

for x = 1, 2, . . . , ak − 2. Therefore, if [a0 ; · · · : ak : x] is a best approximation, then

[a0 ; · · · : ak : x + 1] is also a best approximation.
Now it remains only to find the smallest number x in the set {1, 2, . . . , ak − 2}
such that the number [a0 ; · · · : ak : x] is a best approximation. By Proposition 10.5
we know that the previous best approximation is exactly pk /qk . So we should check
when the following inequality holds:
   
[a0 ; · · · : ak : x] − α  < [a0 ; · · · : ak−1 : ak ] − α . (10.1)

Recall that
pk rk+1 + pk−1
α= ;
qk rk+1 + qk−1
pk
[a0 ; · · · : ak ] = ;
qk
pk x + pk−1
[a0 ; · · · : ak : x] = .
qk x + qk−1

Therefore,
 
[a0 ; · · · : ak−1 : ak ] − α  = |pk qk−1 − qk pk−1 | = 1
qk (qk rk+1 + qk−1 ) qk (qk rk+1 + qk−1 )

and
 
[a0 ; · · · : ak : x] − α  = |pk qk−1 − qk pk−1 |(rk+1 − x)
(qk x + qk−1 )(qk rk+1 + qk−1 )
rk+1 − x
= .
(qk x + qk−1 )(qk rk+1 + qk−1 )

Here |pk qk−1 − qk pk−1 | = 1 holds by Proposition 1.15. Notice that we do not con-
sider the case x ≥ rk+1 , since then the denominator of the corresponding fraction is
greater than qk+1 . Hence inequality (10.1) is equivalent to the inequality

rk+1 − x 1
< ,
(qk x + qk−1 ) qk

which is equivalent to the following one:

rk+1 − (qk−1 /qk )

x> .
2
120 10 Geometric Aspects of Approximation

Suppose that ak+1 is odd. Then

   
ak+1 rk+1 − (qk−1 /qk ) ak+1
< < + 1.
2 2 2
Therefore, for an integer x, inequality (10.1) is equivalent to
ak+1
x> .
2
Now suppose that ak+1 is even. Then
   
ak+1 rk+1 − (qk−1 /qk ) ak+1
−1 < < + 1.
2 2 2
Therefore, for an integer x, inequality (10.1) is satisfied for
ak+1
x> .
2
Finally, it remains to study the case of x = ak+1 /2. We should check whether the
following inequality holds:
rk+1 − (qk−1 /qk )
ak+1 /2 > .
2
It is equivalent to
qk−1
ak+1 > rk+1 − ,
qk
and hence
ak+1 qk − qk−1
> rk+1 .
qk
According to Proposition 1.13, the numerator of the left part equals to qk+1 . So for
x = ak+1 /2, inequality (10.1) is equivalent to
qk+1
rk+1 < .
qk
We have described all solutions of inequality (10.1). This concludes the proof of
Theorem 10.3. 

10.1.2 Strong Best Diophantine Approximations

We say that a rational number p/q (where q > 0) is a strong best approximation of
a real number α if for every fraction p  /q  with 0 < q < q  we get
  
q α − p   > |qα − p|.
10.1 Two Types of Best Approximations of Rational Numbers 121

In this subsection we show that all strong best approximations are convergents of
the corresponding number. In particular, every strong best approximation is a best
approximation.

Proposition 10.8 Every strong best approximation of a real number α is a best

approximation of α.

Proof Consider a strong best approximation p/q of α. Then for every rational num-
ber p  /q  with 0 < q < q  we get
  
q α − p   > |qα − p|.

Since 0 < q  < q, we get

1    1

q α − p   > |qα − p|,
q q
which is equivalent to
   
   
α − p  > α − p .
 q    q
Hence p/q is a best approximation. 

Theorem 10.9 The set of all strong best approximations of a real number α coin-
cides with the set of all convergents of α.

Remark 10.10 In the case of rational α = [a0 ; a1 : · · · : an ] (where an

= 1) the num-
ber p  /q  = [a0 ; a1 : · · · : an − 1] is not a strong best approximation of α. In this case
there is a unique number pn−1 /qn−1 = [a0 ; a1 : · · · : an−1 ] with smaller denomina-
tor that is as good as p  /q  :
  
q α − p   = |qn−1 α − pn−1 |.

Proof Let pi /qi denote the i-convergent of α = [a0 ; a1 : a2 : · · · ].

From Proposition 10.8 it follows that every strong best approximation is a best
approximation. By Theorem 10.3 every best approximation is on a sail of the cor-
responding angles. Therefore, by Theorem 3.1 and Proposition 1.13, every best ap-
proximation can be written for some k ≥ 1 as
hpk + pk−1
, (10.2)
hqk + qk−1
where h varies from 0 to ak+1 . For h = 0 and h = ak+1 we get the k-convergent and
the (k + 1)-convergent respectively. If 0 < h < ak+1 , then on the one hand,
 
(hqk + qk−1 )α − (hpk + pk−1 )
    
=  (h + 1)qk + qk−1 α − (h + 1)pk + pk−1 − (qk α − pk )
122 10 Geometric Aspects of Approximation
     
=  (h + 1)qk + qk−1 α − (h + 1)pk + pk−1  + (qk α − pk )
> |qk α − pk |;

but on the other hand, hqk +qk−1 > qk . Hence for 0 < h < ak+1 , the rational number
of expression (10.2) is not a strong best approximation. Therefore, every strong best
approximation of α is a convergent of α.
For the case of convergents we have
 
|qk+1 α − pk+1 | = (qk+1 + qk )α − (pk+1 + pk ) − (qk α − pk )
 
< |qk α − pk | − (qk+1 + qk )α − (pk+1 + pk ) < |qk α − pk |.

The first inequality holds since the integer points (pk+1 , qk+1 ) and (pk+1 +
pk , qk+1 + qk ) are in different half-planes with respect to the line y = αx. Therefore,
all the convergents are strong best approximations. 

Example 10.11 Consider the example of α = 47 21 = [2; 4 : 5]. All best approxima-
tions and strong best approximations are shown in the following table.

Approximations [2] [2; 3] [2; 4] [2; 4 : 3] [2; 4 : 4] [2; 4 : 5]

Best + + + + + +
Strong best + − + − − +

10.2 Rational Approximations of Arrangements of Two Lines

In the previous section we studied the classical problem of an approximation of real

numbers by rational numbers. In geometric language this means that we approx-
imate some real ray with vertex at the origin by integer rays with vertices at the
origin. Now we say a few words about the approximation theory of arrangements of
two lines intersecting at the origin. On the one hand, the situation here reminds us
of the situation of simultaneous approximation, while on the other hand, algebraic
aspects of arrangement approximation are not visible for simultaneous approxima-
tion. We briefly touch on the multidimensional case in Chap. 22. (For extra details
we refer to [101].)
In Sect. 10.2.1 we associate to any arrangement of two lines a Markov–Davenport
form. (Actually, we have already seen these formulas while studying the Markov
spectrum in Sect. 7.4.) Further, in Sect. 10.2.2, we say a few words about a natural
class of integer approximations for arrangements and introduce the notion of sizes
for integer approximations. We define the distance function (discrepancy) between
two arrangements in Sect. 10.2.3.
10.2 Rational Approximations of Arrangements of Two Lines 123

In Sect. 10.2.4 we prove a general estimate for the discrepancy of best approxi-
mations for the arrangements defined by pairs of continued fractions with bounded
elements: it turns out that in this case, the discrepancy is proportional to N −2 , where
N is the size of the corresponding best approximation. Further, we give an example
of an arrangement one of whose continued fractions has a rapid growing sequence
of elements. In this case the discrepancy is asymptotically not better than N −1−ε .
We study best approximations of arrangements whose lines are eigenlines of cer-
tain matrices in SL(2, Z) in Sect. 10.2.5. Finally, in Sect. 10.2.6 we show how to
simplify the calculation of best approximations in practice.

10.2.1 Regular Angles and Related Markov–Davenport Forms

In this section we study arrangements of two distinct lines passing through the ori-
gin.
We start with the notion of Markov–Davenport forms.

Definition 10.12 Consider an arbitrary quadratic form

f (x, y) = ax 2 + bxy + cy 2

with real coefficients and positive discriminant Δ(f ) = b2 − 4ac. Then the form
f (x, y)
√
Δ(f )
is called the Markov–Davenport form.

Consider an arbitrary arrangement R of lines l1 and l2 . Let the linear forms a1 x +

b1 y and a2 x + b2 y attain zero values at all vectors of l1 and l2 respectively. There
is a unique, up to a sign, Markov–Davenport form associated to R whose zero set
coincides with the union l1 ∪ l2 . It is written as
(a1 x + b1 y)(a2 x + b2 y)
.
a1 b2 − a2 b1
We say that this form is associated to the arrangement R and denote it by ΦR .

Example 10.13 Let us write the Markov–Davenport form for the arrangement of
eigenlines of the Fibonacci operator

1 1
.
1 0

The Fibonacci operator has two eigenlines,

y = −θ x and y = θ −1 x,
124 10 Geometric Aspects of Approximation
√
where θ is the golden ratio, i.e., θ = 1+ 5
2 . The Markov–Davenport form of the
Fibonacci operator is

(y + θ x)(y − θ −1 x) 1  
= √ x 2 − xy − y 2 .
−θ − θ −1 5

10.2.2 Integer Arrangements and Their Sizes

To construct an approximation theory for arrangements one should choose the set
of all approximations and define the quality of each of the approximations of this
subset. We consider the set of all integer arrangements as the set of approximations:
an arrangement of two lines passing through the origin is called integer if both its
lines contain integer points distinct from the origin. Let us define the quality of the
approximation (which we call the size).
For a vector v = (a, b) denote by |v| the norm max(|a|, |b|).

Definition 10.14 Consider an integer arrangement R of lines l1 and l2 passing

through the origin. Let v1 and v2 be two integer vectors of integer length one con-
tained in l1 and l2 respectively. The size of R is the real number
 
max |v1 |, |v2 | ,

which we denote by ν(R).

Remark 10.15 For a generic irrational arrangement the best approximation of any
size greater than 1 is unique. Nevertheless, for particular cases there could be more
than one best approximation of the same size.

10.2.3 Discrepancy Functional and Approximation Model

Let us now define a natural distance between two arrangements in the plane passing
through the origin.

Definition 10.16 Let R1 , R2 be two arrangements of pairs of lines in the plane.

Consider two symmetric bilinear forms

ΦR1 (v) + ΦR2 (v) and ΦR1 (v) − ΦR2 (v).

Take the maximal absolute values of the coefficients of these forms. Suppose that
they equal c1 and c2 . The discrepancy between R1 and R2 is min(c1 , c2 ). We denote
it by ρ(R1 , R2 ).
10.2 Rational Approximations of Arrangements of Two Lines 125

Remark 10.17 In case of arrangements R1 and R2 of ordered lines one should con-
sider the maximal absolute values of the coefficients only of one form

ΦR1 (v) − ΦR2 (v).

The problem of best approximations of an arbitrary arrangement by integer ar-

rangements is the following.

Problem of Best Approximations of Arrangements For a given arrangement R

and an integer N > 0, find an integer arrangement RN such that ν(RN ) ≤ N and
     
ρ(R, RN ) = min ρ R, R  | R is integer, ν R  ≤ N .

Remark 10.18 The space of all arrangements of two lines passing through the ori-
gin is two-dimensional. Nevertheless, the complexity of approximation by integer
arrangements is more complicated in comparison with simultaneous approximation
of a line in R3 (whose configuration space is also two-dimensional). This happens
because for the approximation of the arrangement of two lines one needs two pairs
of relatively prime integers, while for simultaneous approximation of a vector in R3 ,
a triple of relatively prime integers suffices. Nevertheless, geometric properties of
continued fractions essentially simplify the approximations of arrangements. This is
relevant especially for so-called algebraic arrangements, whose lines are eigenlines
of some SL(2, Z) matrix. In this case all the approximations are constructed from
some periodic sets that are easy to construct.

10.2.4 Lagrange Estimates for the Case of Continued Fractions

with Bounded Elements

In this subsection we prove an analogue of Lagrange’s theorem on the approxima-

tion rate of arrangements of two lines y = α1 x and y = α2 x whose coefficients α1
and α2 have continued fractions with bounded elements. In particular, this includes
all algebraic arrangements (as in Remark 10.18). For the convenience of the reader
we do not consider the degenerate case in which one of the lines is x = 0, which is
not so important, since it is equivalent to the approximation of a vector in the plane.

Theorem 10.19 Consider an arrangement R of the lines y = α1 x and y = α2 x.

Let α1 and α2 be real numbers having infinite continued fractions with bounded
elements. Then there exist constants C1 , C2 > 0 such that for every integer N > 0,
the best approximation RN satisfies

C1 C2
< ρ(R, RN ) < 2 .
N2 N
126 10 Geometric Aspects of Approximation

This theorem is a reformulation of one of the results of [101]. We begin the proof
with the following two technical lemmas.
Let Rδ1 ,δ2 denote the arrangement of lines y = (αi + δi )x for i = 1, 2.

Lemma 10.20 Consider a real number ε satisfying

0 < ε < 1/|α1 − α2 |.

Supposing that ρ(R, Rδ1 ,δ2 ) < ε, then

(1 + |α1 |)(α1 − α2 )2 (1 + |α2 |)(α1 − α2 )2

|δ1 | < ε and |δ2 | < ε.
|α2 |(1 − ε|α1 − α2 |) |α1 |(1 − ε|α1 − α2 |)

Proof The Markov–Davenport form of Rδ1 ,δ2 is

(y − (α1 + δ1 )x)(y − (α2 + δ2 )x)

ΦRδ1 ,δ2 (x, y) = .
(α2 + δ2 ) − (α1 + δ1 )
Then the form ΦR − ΦRδ1 ,δ2 is as follows:

α12 δ2 − α22 δ1 + α1 δ1 δ2 − α2 δ1 δ2 2
x
(α1 − α2 )(α1 − α2 + δ1 − δ2 )
α2 δ1 − α1 δ2
+ xy
(α1 − α2 )(α1 − α2 + δ1 − δ2 )
δ2 − δ1
+ y2. (10.3)
(α1 − α2 )(α1 − α2 + δ1 − δ2 )

Consider the absolute values of the coefficients at y 2 and at xy for the form
ΦR − ΦRδ1 ,δ2 . By the conditions of the lemma these coefficients are less then ε. We
have
 
 δ2 − δ1 
  (10.4)
 (α − α )(α − α + δ − δ )  < ε;
1 2 1 2 1 2
 
 α2 δ1 − α1 δ2 
  (10.5)
 (α − α )(α − α + δ − δ )  < ε.
1 2 1 2 1 2

Inequality (10.4) implies

(α1 − α2 )2
|δ1 − δ2 | < ε.
1 − ε|α1 − α2 |

From Inequality (10.5) we have

|(α1 − α2 )(α1 − α2 + δ1 − δ2 )|ε + |α1 (δ1 − δ2 )|

|δ1 | < ,
|α2 |
10.2 Rational Approximations of Arrangements of Two Lines 127

and hence we have


(α1 − α2 )2 (α1 − α2 )2
|δ1 | < |α1 − α2 | |α1 − α2 | + ε ε + |α1 | ε |α2 |
1 − ε|α1 − α2 | 1 − ε|α1 − α2 |
(1 + |α1 |)(α1 − α2 )2
< ε.
|α2 |(1 − ε|α1 − α2 |)
The inequality for δ2 is obtained in a similar way. 

Lemma 10.21 Let ε be a positive real number. Suppose that real numbers δ1 and
δ2 satisfy |δ1 | < ε and |δ2 | < ε. Then the following estimate holds:

max(2, 2(|α1 | + |α2 |), α12 + α22 + |α1 − α2 |ε)

ρ(R, Rδ1 ,δ2 ) < ε.
(|α1 − α2 |)(|α1 − α2 | + 2ε)

Proof The statement of the lemma follows directly from the estimate for R − Rδ1 ,δ2 ,
shown in (10.3). For the coefficient at x 2 we have
 2 
 α1 δ2 − α22 δ1 + α1 δ1 δ2 − α2 δ1 δ2  α12 + α22 + |α1 − α2 |ε
 
 (α − α )(α − α + δ − δ )  < (|α − α |)(|α − α | + 2ε) ε.
1 2 1 2 1 2 1 2 1 2

For the coefficient at xy we have

α2 δ1 − α1 δ2 2(|α1 | + |α2 |)
xy < ε.
(α1 − α2 )(α1 − α2 + δ1 − δ2 ) (|α1 − α2 |)(|α1 − α2 | + 2ε)

For the coefficient at y 2 we have

δ2 − δ1 2
y2 < ε.
(α1 − α2 )(α1 − α2 + δ1 − δ2 ) (|α1 − α2 |)(|α1 − α2 | + 2ε)
Therefore, the statement of the lemma holds by the definition of discrepancy. 

Proof of Theorem 10.19 Let us start with the proof of the inequality

C1
< ρ(R, RN ).
N2
Let α1 = [a0 ; a1 : · · · ] and pi /qi = [a0 ; a1 : · · · : ai ] for i = 0, 1, . . . . Without loss
of generality we assume that N > a0 . Assume that k is the greatest integer for which
pk ≤ N and qk ≤ N . Then we have the following estimates:
 


 p  
min α1 −   |p|≤N, |q|≤N
q
 
 pk+1  1

≥ α1 − ≥
qk+1  qk+1 (qk+1 + qk+2 )
128 10 Geometric Aspects of Approximation

1
≥
(ak+1 + 1)qk ((ak+1 + 1)qk + (ak+1 + 1)(ak+2 + 1)qk )
1 1
≥ · .
(ak+1 + 1)2 (ak+2 + 2) N 2

The second inequality follows from Exercise 1.3 on page 18 (we leave the proof for
the reader). For the third inequality we use the following inequality twice:

qi+1 = ai+1 qi + qi−1 < (ai + 1)qi , (10.6)

which follows directly from Proposition 1.13.

The same calculations are valid for α2 . After that, the constant C1 is obtained by
Lemma 10.20.
Now we prove the inequality

C2
ρ(R, RN ) > .
N2
Again we assume that k is the greatest integer for which |pk | ≤ N and qk ≤ N .
For α1 we have the following:
   
   
α1 − pk  <  pk+1 − pk  = 1 < ak+1 + 1 < (ak+1 + 1)(|α1 | + 1) .
2
 qk   qk+1 qk  qk qk+1 2
qk+1 N2

The first inequality follows from the fact that if the k-convergent is greater than α,
then the (k + 1)-convergent is less then α, and conversely. Proposition 1.15 implies
the second equality. The third inequality follows from inequality (10.6). In the last
inequality we use the assumption that either qk+1 > N or |pk+1 | > N , in which case

|pk+1 | N
qk+1 > >
|α1 | + 1 |α1 | + 1

(here the first inequality holds since |α − pk+1 /qk+1 | < 1). From conditions of the
theorem the set of all elements ai of the continued fraction for α is bounded. There-
 such that for every N there exists an approximation
fore, there exists a constant C2,1
of α1 of magnitude smaller than C2,1  /N 2 .
 such that for every N
A similar estimate holds for α2 . There exists a constant C2,2
there exists an approximation of α1 of magnitude smaller than C2,2  /N 2 . Therefore,

we can apply Lemma 10.21 in order to obtain the constant C2 . 

Remark 10.22 Now we say a few words about the case of unbounded elements of
continued fractions. Let ε be a small positive real number. If the elements of a con-
tinued fraction (say for α1 ) grow fast enough, then there exists a sequence Ni for
which the approximations RNi are of magnitude (N C)1+ε . Let us illustrate this by the
i
following example.
10.2 Rational Approximations of Arrangements of Two Lines 129

Example 10.23 Let M be an arbitrary positive integer. Consider an arrangement R

of the lines y = 0 and y = αx, where the continued fraction [a0 ; a1 : · · · ] for α is
defined inductively:

– a0 = 1;
pk
– let a0 , . . . , ak be defined and [a0 ; · · · : ak ] = qk ; then ak+1 = (qk )M−1 .

Let Nk =  qk+1
2  + 1 for k = 1, 2, . . . . Then there exists a positive constant C such
that for every integer i we have

C
ρ(R, RNi ) ≥ 1+1/M
.
Ni

Proof For every i we have

qi+1 ≥ ai qi = qiM−1 qi = qiM .

Therefore, the magnitude of best approximations for k-convergents are as follows:

 
  1 1 1 1
α1 − pk  ≥ ≥ 1/M = 1+1/M < .
 qk  qk (qk+1 + qk ) q 1+1/M
k+1 (2qk+1 ) 2qk+1 22+1/M Nk+1

From Theorem 10.3 it follows that there is no best approximation whose denomina-
tors q satisfies

qk < q ≤ N k .

Therefore, RNk = Rpk /qk . Hence

C
ρ(R, RNk ) = ρ(R, Rpk /qk ) ≥ 1+1/M
.
Nk

The last inequality follows from Lemma 10.20. This completes the proof. 

We suspect the existence of a badly approximable arrangement R and a constant

C such that there are only finitely many solutions N of the inequality

C
ρ(R, RN ) ≤ ,
N

as in the case of simultaneous approximation of vectors in R3 (see for instance

[127]).
130 10 Geometric Aspects of Approximation

Fig. 10.2 The arrangement of two lines and its geometric continued 1-, 2-, and 3-fractions

10.2.5 Periodic Sails and Best Approximations in the Algebraic

Case

10.2.5.1 Geometric w-Continued Fractions

Let us prove a relation between classical geometry of numbers and best simultane-
ous approximations.

Definition 10.24 Define inductively the w-sails for an arbitrary angle C with vertex
at the origin.
– let the 1-sail be the sail of C.
– suppose that all w-sails for w < w0 are defined. Then let the w0 -sail be
  
0 −1
  w
∂ conv C ∩ Z \{O} \
2
w-sail ,
w=1

where conv(M) denotes the convex hull of M.

Consider an arbitrary arrangement R of two lines l1 and l2 . The union of all
four w-sails for the cones defined by the lines l1 and l2 is the geometric continued
w-fraction of R.

In Fig. 10.2 we show the geometric continued 1-, 2-, and 3-fractions for a partic-
ular arrangement of two lines. It is interesting to notice that the geometric continued
3-fraction in the figure is homothetic to the geometric continued 2-fraction and to
the geometric continued fraction (which is a 1-fraction). It turns out that it is a gen-
eral fact that the w-continued fraction of any general arrangement R (i.e., whose
lines do not contain integer points except the origin) is homothetic to the geometric
continued fraction of R; the coefficient of homothety is w. We leave this statement
as an exercise for the reader.
10.2 Rational Approximations of Arrangements of Two Lines 131

10.2.5.2 Algebraic Arrangements

An arrangement of two lines l1 , l2 is said to be algebraic if there exists an oper-

ator A ∈ GL(2, Z) with irreducible, over Q, characteristic polynomial having two
real roots whose eigenlines are l1 and l2 . In this case the operator A acts on a w-
geometric continued fraction (for any w) as a transitive shift.

Definition 10.25 The minimal absolute value of the form ΦR on the integer lattice
minus the origin is called the Markov minimum of ΦR . We denote it by αR .

Recall that Markov minima are studied in the Markov spectrum theory discussed
in Sect. 7.4.

Lemma 10.26 Let v be an integer point in the w-continued fraction of an algebraic

arrangement R. Then
 
ΦR (v) ≥ wαR .

Proof The statement follows directly from the fact that the w-continued fraction of
R is homothetic to the continued fraction of R. 

Theorem 10.27 Let R be an algebraic arrangement. Then there exists C > 0 such
that for every N ∈ Z+ the following holds: each of the two lines of the best approx-
imation arrangement RN contains an integer point of some w-sail for w < C.

Proof Let R be an arrangement of lines y = α1 x and y = α2 x. Let also v1,N (x1,N ,

y1,N ) and v2,N (x2,N , y2,N ) be two integer vectors of unit integer length in the lines
defining RN .
Since R is algebraic, the continued fractions for α1 and α2 are periodic and there-
fore uniformly bounded. Now we can apply Theorem 10.19 to write the estimate
C2
ρ(R, RN ) < .
N2
Hence by Lemma 10.20 we have
 
 x1,N  C̃2
 − α 
1 < .
y N2
1,N

Notice that
 
   
ΦR (x1,N , y1,N ) =  (x1,N − α1 y1,N )(x1,N − α2 y1,N ) 
 α1 − α2 
   
 x1,N   x1,N − α2 y1,N 
=  
− α1  ·  y1,N .
y1,N α1 − α2

The first term above is bounded by C̃/N 2 for some constant C̃ that does not depend
on N . Hence,
132 10 Geometric Aspects of Approximation
 2   
 y1,N (x1,N /y1,N ) − α2   (x1,N /y1,N ) − α2 

|ΦR (x1,N , y1,N )| ≤ C̃  2 ·   .
N α1 − α2  ≤ C̃  α1 − α2 

Finally, the last expression is uniformly bounded in N . The same property holds for
the values |ΦR (v2,N )|.
Therefore, by Lemma 10.26 the vectors of v1,N and v2,N for all N are in the
union of a finite number of w-continued fractions for R. 

Conjecture 5 We conjecture that for all algebraic arrangements for almost all N ,
the vectors v1,N and v2,N defining RN are in a 1-geometric continued fraction.

10.2.6 Finding Best Approximations of Line Arrangements

In this subsection we show a general technique for calculating best approximations

for an arbitrary arrangement of lines R with eigenspaces y = α1 x and y = α2 x for
distinct real numbers α1 and α2 .

Proposition 10.28 Let m and n be two integers. Supposing that |α1 − pq | < δ (or
|α2 − pq | < δ, respectively), then the following holds:
 
 
α1 − p  > |α1 − α2 | |ΦR (p, q)|
 q  |α1 − α2 | + δ q2
 

 
α2 − p  > |α1 − α2 | |ΦR (p, q)| .
 q  |α1 − α2 | + δ q2

Proof We have
 
 
α1 − p  = 1 |p − α1 q| = 1 |p − α1 q|(p − α2 q)
 q q q p − α2 q
|ΦR (p, q)| |α1 − α2 | |α1 − α2 | |ΦR (p, q)|
= > .
q2 |α1 − α2 + (p/q − α1 )| |α1 − α2 | + δ q2

The same holds for the case of the approximations of α2 . 

Procedure of Best Approximation Calculation

Input data. We start with an arrangement R of two lines y = α1 x and y = α2 x and
a positive integer N . The task is to construct the best approximation RN .
Step 1. Find pi /qi , which is the i-convergent to α1 with maximal possible denom-
inator satisfying
10.2 Rational Approximations of Arrangements of Two Lines 133

1 ≤ pi ≤ N and 1 ≤ qi ≤ N.
Also find pj /qj , the convergent satisfying the same property for α2 .
Step 2. Consider the arrangement R of lines

pi pj
y= x and y = x.
qi qj

We know that
   
   pj 
 α − pi  ≤ 1 1
and α −   ≤   .

 qi  qi qi+1 qj qj qj +1

By Lemma 10.21 (for N > 1) we have

1 1
ρ(R, R) < C max ,   ,
qi qi+1 qj qj +1

where
max(2, 2(|α1 | + |α2 |), α12 + α22 + |α1 − α2 |)
C= .
(|α1 − α2 |)(|α1 − α2 |)
Step 3. From the definition of best approximations and the result of Step 1 we have

1 1
ρ(R, RN ) ≤ ρ(R, R) < C max , .
qi qi+1 qj qj +1

In most cases (except for a small denominator N ) this estimate satisfies

1
ρ(R, RN ) < .
2|α1 − α2 |
Choose δ1 and δ2 such that the lines of the best approximation RN are y =
(α1 + δ1 )x and y = (α2 + δ2 )x. In the notation of Lemma 10.20 this means that
RN = Rδ1 ,δ2 . Now we apply Lemma 10.20 to write the estimate for δ1 and δ2 :

2(1 + |α1 |)(α1 − α2 )2 1 1

|δ1 | < · C max ,   ;
|α2 | qi qi+1 qj qj +1

2(1 + |α2 |)(α1 − α2 )2 1 1

|δ2 | < · C max , ,
|α1 | qi qi+1 qj qj +1

where C is as in Step 2.
Step 4. Suppose now that the lines of the best approximation RN contains integer
vectors (n1 , m1 ) and (n2 , m2 ) of unit integer length. By Proposition 10.28 we
write the following estimates on the values of the Markov–Davenport form ΦR
at these vectors:
134 10 Geometric Aspects of Approximation
 
ΦR (m1 , n1 ) |α1 − α2 | + δ1  m1  |α1 − α2 | + δ1
< α1 − < δ1 ;
n21 |α1 − α2 |  n1  |α1 − α2 |
 
ΦR (m2 , n2 ) |α1 − α2 | + δ2  m2  |α1 − α2 | + δ2
< α1 − < δ2 .
n22 |α1 − α2 |  n2  |α1 − α2 |

Step 5. This is the last step. Here we should consider all integer arrangements that
contains integer points (m1 , n1 ) and (m2 , n2 ) satisfying the inequalities obtained
in Step 4 for ΦR (mi , ni ) for i = 1, 2. Then one should choose the arrangement
with the smallest discrepancy, which is RN .
Output. In the output we have the best approximation RN .

Remark 10.29 The above procedure of best approximation calculation seriously de-
creases the number of integer points in the square N × N that should be tested to
find RN . This is of a great importance especially for the case of algebraic arrange-
ments. The procedure gives an explicit estimate on the number of w-sails whose
points should be checked to construct best approximations. Since all w-sails are pe-
riodic, all its points are contained in the finite number of orbits with respect to the
corresponding Dirichlet group action.

Example 10.30
Input data. Consider an arrangement of eigenlines of the Fibonacci operator:

0 1
.
1 1

Denote by Fn the nth Fibonacci number.

Consider any integer N ≥ 100.
Step 1. Consider a positive integer k such that Fk ≤ N < Fk+1 and choose an ap-
proximation R with eigenspaces Fk−1 y − Fk x = 0 and Fk y + Fk−1 x = 0. Then
   
  1   1
α1 − Fk  ≤ , α1 + Fk−1  ≤ .
 Fk−1  Fk−1 Fk  Fk  Fk Fk+1
Step 2. Since 1/(Fk−1 Fk ) < 1/(55 · 89), we have
√ √
max(2, 2 5, 3 + 5/4895) 1
ρ(R, RN ) < √
5 + 2 5/4895 Fk−1 Fk
√
2 5 (89/55)3 3.79
< √ < 2 .
5 + 2 5/4895 N 2 N

Step 3. Now, following Lemma 10.20 we write the estimates on δ1 and δ2 of the
best approximation RN = Rδ1 ,δ2 :

80.35 18.97
|δ1 | < and |δ2 | < .
N2 N2
10.3 Exercises 135

Step 4. Suppose that the best approximation RN is defined by the two integer vec-
tors (m1 , n1 ) and (m2 , n2 ) of unit integer length. Then we have the following
two estimates:
|ΦR (m1 , n1 )| 80.65 |ΦR (m2 , n2 )| 18.99
< , < .
n21 N2 n22 N2

Step 5 and Output. We have completed the computations for N = 106 , and the an-
swer in this case is the matrix with eigenspaces F29 y − F30 x = 0 and F30 y +
F29 x = 0. We conjecture that for the Fibonacci matrix all the best approxima-
tions are the arrangements of lines Fk−1 y − Fk x = 0 and Fk y + Fk−1 x = 0 for
k = 0, 1, . . . .

10.3 Exercises

Exercise 10.1 Find all best approximations and all strong best approximations for
the numbers 433/186 and 575/247.

Exercise 10.2 Let ABC be an integer empty triangle. Consider a strip bounded by
two lines parallel to AB and passing through the points A and C respectively. Prove
that all integer points on this strip are at its boundary lines.

Exercise 10.3 Consider a positive real number α with a regular continued fraction
[a0 ; a1 : · · · ]. Let x be an integer satisfying 1 ≤ x ≤ ak+1 − 1. Prove that
   
[a0 ; a1 : · · · : ak : x] − α  > [a0 ; a1 : · · · : ak : x + 1] − α .

Exercise 10.4 Let α = [a0 ; a1 : · · · ] be a positive real number. Denote by pk /qk its
k-convergent and by rk the remainder part [ak ; ak+1 : · · · ]. Then the condition
qk+1
rk+1 <
qk

is equivalent to the condition

[ak+1 ; ak+2 : · · · ] < [ak+1 ; ak : · · · : a0 ].

Exercise 10.5 Write Markov–Davenport forms for the arrangements R1 and R2 of

eigenlines of the operators

3 5 7 −2
and
2 3 −3 1

respectively. Write the discrepancy between R1 and R2 .

136 10 Geometric Aspects of Approximation

Exercise 10.6 Consider an arbitrary angle C with vertex at the origin whose edges
do not contain other integer points. Prove that the w-sail of C is homothetic to the 1-
sail of C and find the coefficient of homothety. What happens for the angles having
integer points on their edges?

Exercise 10.7 Let R be an algebraic arrangement of two lines. Then the following
statements hold.
(a) The Markov–Davenport form ΦR attains only a finite number of values at the
integer points of the corresponding continued fraction. Denote this set by S.
(b) The form ΦR attains the value x at some vertex of the w-sail if and only if
x/w ∈ S.

Exercise 10.8 Find all best approximations of sizes 1, 2, 3, 4, 5 for the arrangement
of the eigenlines of the Fibonacci operator.
Chapter 11
Geometry of Continued Fractions with Real
Elements and Kepler’s Second Law

In the beginning of this book we discussed the geometric interpretation of regular

continued fractions in terms of LLS sequences of sails. Is there a natural extension
of this interpretation to the case of continued fractions with arbitrary elements? The
aim of this chapter is to answer this question.
In Sects. 11.1 and 11.2 we introduce a geometric interpretation of odd or infinite
continued fractions with arbitrary elements in terms of broken lines in the plane
having a selected point (say the origin). Further, in Sect. 11.3 we consider differen-
tiable curves as infinitesimal broken lines to define analogues of continued fractions
for curves. The resulting analogues possess an interesting interpretation in terms of
a motion of a body according to Kepler’s second law.

11.1 Continued Fractions with Integer Coefficients

In this section we give a formal extension of the notion of the LLS sequence for
sails to the case of certain integer broken lines. Let us specify the class of broken
lines we are dealing with.
A broken line is called integer if all its vertices are integer points. We restrict
ourselves to the case of broken lines all of whose triples of consecutive vertices are
not in a line.

Definition 11.1 Consider an integer broken line L in the complement to an integer

point V . We say that L is at unit integer distance from V (or V -broken line, for
short) if all edges of L are at unit integer distance from V .

Remark Without loss of generality, we mostly use only O-broken lines, where O is
the origin.

O. Karpenkov, Geometry of Continued Fractions, 137

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_11, © Springer-Verlag Berlin Heidelberg 2013
138 11 Geometry of Continued Fractions and Kepler’s Second Law

For arbitrary integer points A, B, and C, we consider the function of orientation

sgn, namely

sgn(ABC)
⎧
⎨ 1, if the pair of vectors (BA, BC) defines the positive orientation,
= 0, if the points A, B, and C are in a straight line,
⎩
−1, if BA, BC defines the negative orientation.

Let us generalize the notion of LLS sequences for sails to an arbitrary O-broken
line.

Definition 11.2 Consider an O-broken line A0 A1 . . . An . Put by definition

a0 = sgn(A0 VA1 ) l(A0 A1 ),

a1 = sgn(A0 VA1 ) sgn(A1 VA2 ) sgn(A0 A1 A2 ) lsin(∠A0 A1 A2 ),
a2 = sgn(A1 VA2 ) l(A1 A2 ),
...
a2n−3 = sgn(An−2 VAn−1 ) sgn(An−1 VAn ) sgn(An−2 An−1 An )
× lsin(∠An−2 An−1 An ),
a2n−2 = sgn(An−1 VAn ) l(An−1 An ).

The sequence (a0 , . . . , a2n−2 ) is called a lattice signed length-sine sequence for the
broken line A0 A1 . . . An , or LSLS sequence, for short.
The continued fraction for the O-broken line A0 A1 . . . An is the element [a0 ; a1 :
· · · : a2n−2 ] of R.

Notice that the LSLS sequence for the sail of some angle coincides with the LLS
sequence of this sail. So we consider LSLS sequences as a natural combinatoric–
geometric generalization of LLS sequences. Note also that an O-broken line is
uniquely defined by its LSLS sequence, the point A0 , and the direction of the vector
A0 A1 .
In Fig. 11.1 we show how to get signs of elements of the LSLS sequence from
the local geometry of a broken line. As an example we consider the O-broken line
with four vertices in Fig. 11.2.
In the next theorem we state that the LSLS sequence is the complete invariant
for the action of the group of proper integer congruences (i.e., integer congruences
preserving the orientation of the plane).

Theorem 11.3
(i) Two O-broken lines are proper integer congruent if and only if their LSLS se-
quences coincide.
11.2 Continued Fractions with Real Coefficients 139

Fig. 11.1 How to identify signs of elements in LSLS sequences

Fig. 11.2 An O-broken line

and the corresponding LSLS
sequence

(ii) Every finite sequence of nonzero integers is realizable as an LSLS sequence for
some O-broken line.

We skip the proof of this theorem, since later, in the next section, we introduce a
more general construction that generalize LSLS sequences for O-broken lines. For
further information on O-broken lines and their LSLS sequences we refer to [94]
and [96].

11.2 Continued Fractions with Real Coefficients

Let us describe a recent generalization of lattice properties of regular continued

fractions to the case of arbitrary continued fractions (see also [98]).
140 11 Geometry of Continued Fractions and Kepler’s Second Law

Further in this chapter we work in the plane with a marked point O (say the
origin). For two vectors v and w in the plane we denote the oriented volume of the
parallelogram spanned by the vectors v and w by |v × w|.

11.2.1 Broken Lines Related to Sequences of Arbitrary Real

Numbers

In Chap. 3 (see page 37) we gave a geometric algorithm to construct sails for real
numbers given the regular continued fractions of the corresponding numbers. The
algorithm has a sequence of positive integers (a0 , a1 , . . .) as input and its output
is the sails for the number [a0 , a1 , . . .]. In this subsection we slightly modify this
algorithm such that it will construct a broken line starting from an arbitrary sequence
of nonzero real numbers. This algorithm will give us a base to define continued
fractions related to broken lines (we do this later, in the next subsection).

Algorithm to Construct Broken Lines Related to Sequences of Real Numbers

Input data. Given a sequence (a0 , a1 , . . . , a2n ) of nonzero real numbers; the first
vertex A0 ; and the direction v of the first edge.
Goal of the algorithm. To construct a broken line A0 A1 . . . An .
Step 1. Assign
A1 = A0 + λv,
where λ is the solution of the equation |OA0 × OA1 | = a0 .
Inductive Step k. From previous steps we have the vertices A0 , . . . , Ak , for k ≥ 1.
Now let us find the vertex Ak+1 . First, we consider the point
1
P = Ak + Ak−1 Ak .
a2k−2
Notice that the point P is on the line Ak−1 Ak and the area of OAk P equals
1/2. Secondly, we put
Q = P + a2k−1 OAk .
Thirdly, we define the point Ak+1 (see Fig. 11.3):

Ak+1 = Ak + a2k Ak Q.

This concludes Step k.

Output. The broken line A0 . . . An .

Remark 11.4 The constructed broken line in some sense generalizes geometric con-
struction from the case of sails related to regular continued fractions (see Theo-
rem 3.1 above) to the case of broken lines related to arbitrary continued fractions
(see Theorem 11.10 and Corollary 11.11 below).
11.2 Continued Fractions with Real Coefficients 141

Fig. 11.3 Construction of the point Ak+1

Let us give a geometric meaning of a0 , a1 , . . . in terms of oriented volumes for

certain vectors defined by the above broken line.

Proposition 11.5 Let (a0 , a1 , . . . , a2n ) be a sequence of arbitrary nonzero ele-

ments. Consider any broken line A0 . . . An constructed by this sequence as above.
Then the following holds:

a2k = |OAk × OAk+1 |, k = 0, . . . , n,

|Ak Ak−1 × Ak Ak+1 |
a2k−1 = , k = 1, . . . , n.
a2k−2 a2k

Proof Let us prove this proposition by induction on k.

Base of induction. Directly from the first step of the construction we get

|OA0 × OA1 | = a0 .

Inductive Step. Supposing that the statement holds for k − 1, let us prove it for k.
We start with a2k :

|OAk × OAk+1 | = a2k |OAk × OQ|

a2k
= a2k |OAk × OP| = |OAk−1 × OAk | = a2k .
a2k−2
The last equality follows from the induction assumption.
Now we check the expression for a2k−1 .
|Ak Ak−1 × Ak Ak+1 | |Ak Ak−1 × Ak Q|
= = |PAk × PQ| = a2k−1 |OAk × OP|
a2k−2 a2k a2k−2
a2k−1
= |OAk−1 × OAk | = a2k−1 .
a2k−2
The inductive step is complete. This concludes the proof. 

Example 11.6 Let us consider the first interesting case of broken lines with three
vertices. Without loss of generality we fix A0 = (1, 0) and the direction v = (0, 1).
Letting the sequence be (a0 , a1 , a2 ), we have

A1 = (1, a0 ).
142 11 Geometry of Continued Fractions and Kepler’s Second Law

The corresponding points P and Q are as follows:

P = (1, 1 + a0 ), Q = (1 + a1 , 1 + a0 + a0 a1 ).

Finally, we have
A2 = (1 + a1 a2 , a0 + a2 + a0 a1 a2 ).
Notice that the coordinates of the point A2 coincide with the denominator and the
numerator of the continued fraction [a0 ; a1 : a2 ].

11.2.2 Continued Fractions Related to Broken Lines

Let us extend the definition of LLS sequences for sails to the case of broken lines,
basing the definition on the expressions of Proposition 11.5. Consider an arbitrary
broken line A0 . . . An satisfying the following condition: any of its edge is not con-
tained in a line passing through the origin O.

Definition 11.7 Define

a2k = |OAk × OAk+1 |, k = 0, . . . , n;

|Ak Ak−1 × Ak Ak+1 |
a2k−1 = , k = 1, . . . , n.
a2k−2 a2k
The sequence (a0 , . . . , a2n ) is called the LLS sequence for the broken line. The num-
ber [a0 ; · · · : a2n ] is said to be the continued fraction for the broken line A0 . . . An .

This definition extends the definition of LSLS sequences to O-broken lines.

Proposition 11.8 Let a broken line be an O-broken line. Then its LLS sequence
coincides with its LSLS sequence.

Since the proof of this proposition is a straightforward calculation, we leave it to

the reader.
Finally, we show how the GL(2, R)-transformations act on the LLS sequences.

Proposition 11.9 Consider two broken lines A0 . . . An and B0 . . . Bn whose LLS se-
quences are (a0 , . . . , a2n ) and (b0 , . . . , b2n ) respectively. Suppose that there exists a
GL(2, R)-operator taking the first broken line to the second one. Let the determinant
of this operator equal λ. Then we have

a2k = λb2k , k = 0, . . . , n,
a2k−1 = λ1 b2k−1 , k = 1, . . . , n.

Proof The statement follows directly from formulas of Proposition 11.5, since the
area of any parallelogram in the definition is multiplied by λ. 
11.2 Continued Fractions with Real Coefficients 143

11.2.3 Geometry of Continued Fractions for Broken Lines

Let us recall the notion of polynomials Pk and Qk from Chap. 1. The polynomials
Pk and Qk are uniquely defined by the following three conditions: first, they are
relatively prime; secondly,

Pk (x0 , . . . , xk )
= [x0 ; x1 : · · · : xk ];
Qk (x0 , . . . , xk )

thirdly,
Pk (0, . . . , 0) + Qk (0, . . . , 0) = 1.

Theorem 11.10 Consider a broken line A0 . . . An with LLS sequence (a0 , a1 ,

. . . , a2n ). Let also A0 = (1, 0) and A1 = (1, a0 ). Then
 
An = Q2n+1 (a0 , a1 , . . . , a2n ), P2n+1 (a0 , a1 , . . . , a2n ) .

Proof We prove this theorem by induction on the number of vertices in the broken
line, i.e., in n.
Base of induction. If the broken line has two vertices A0 , A1 and its LLS se-
quence is (a0 ), then A1 = (1, a0 ).
Step of induction. Assume that the statement holds for all broken lines with k
vertices. Let us give a proof for an arbitrary broken line with k + 1 vertices. Consider
a broken line A0 . . . Ak . Let its LLS sequence be (a0 , . . . , a2k ).
Let

a a + 1 −a1
T= 0 1 .
−a0 1
The transformation T takes A1 to (1, 0) and the line A2 A1 to the line x = 1. Under
the transformation T the broken line A0 . . . Ak is taken to another broken line, which
we denote by B0 B1 . . . Bk . Since the determinant of T equals 1, then by Proposi-
tion 11.9, the LLS sequence for B0 B1 . . . Bk coincides with (a0 , . . . , a2k ). By the
induction assumption, the statement holds for the broken line B1 . . . Bk . Thus we
have
 
Bk = Q2k−1 (a2 , . . . , a2k ), P2k−1 (a2 , . . . , a2k ) .
Letting Bk = (q, p), then
 
Ak = T −1 (Bk ) = p + a1 q, a0 p + (a0 a1 + 1)q .

Therefore, we have

a0 p + (a0 a1 + 1)q 1 P2k+1 (a0 , a1 , . . . , a2k )

= a0 + = .
p + a1 q a1 + p/q Q2k+1 (a0 , a1 , . . . , a2k )
144 11 Geometry of Continued Fractions and Kepler’s Second Law

Now we substitute p = P2k−1 (a2 , . . . , a2k ) and q = Q2k−1 (a2 , . . . , a2k ). Let us cal-
culate the value of the sum of the numerator and denominator in the leftmost fraction
at the point O = (0, . . . , 0):
  
a0 P2k−1 + (a0 a1 + 1)Q2k−1 + (P2k−1 + a1 Q2k−1 ) (O)
= (P2k−1 + Q2k−1 )(O) = 1.

The first equality holds since we take a0 = a1 = 0. The second equation holds by
definition of P2k−1 and Q2k−1 . Therefore,

P2k+1 (a0 , a1 , . . . , a2k ) = a0 P2k−1 (a2 , . . . , a2k ) + (a0 a1 + 1)Q2k−1 (a2 , . . . , a2k ),
Q2k+1 (a0 , a1 , . . . , a2k ) = P2k−1 (a2 , . . . , a2k ) + a1 Q2k−1 (a2 , . . . , a2k ).

This concludes the proof for the broken line A0 . . . Ak . The induction step is com-
plete. 

Theorem 11.10 extends the geometric interpretation of regular continued frac-

tions to the case of arbitrary continued fractions in the following way.

Corollary 11.11 Consider a broken line A0 . . . An with A0 = (1, 0), A1 = (1, a0 ),

and An = (x, y). Let α = [a0 ; a1 : · · · : a2n ] be the corresponding continued fraction
for this broken line. Then
y
= α.
x
For the case of an infinite value for α = [a0 ; a1 : · · · : a2n ],
x
= 0.
y

Remark 11.12 Supposing that the sequence of numbers (a0 , a1 , . . . , a2n ) in Corol-
lary 11.11 contains only positive integers, the broken line A0 . . . An coincides with
the sail of the angle defined by rays y = 0 and y = αx in the first quadrant of the
plane. In particular, A0 . . . An is in the boundary of a convex set and its LLS se-
quence coincides with the LLS sequence of the sail for the angle ∠A0 OAn .

From Corollary 11.11 we get the following statement.

Corollary 11.13 Consider two broken lines A0 . . . An and B0 . . . Bm with LLS se-
quences (a0 , a1 , . . . , a2n ) and (b0 , b1 , . . . , b2m ) respectively. Let B0 = A0 and the
vector A0 A1 coincide with the vector ab00 B0 B1 . Suppose that

[a0 ; · · · : a2n ] = [b0 ; · · · : b2m ].

Then the points An , Bm , and the origin O are contained in one line.
11.2 Continued Fractions with Real Coefficients 145

Fig. 11.4 Two examples of broken lines with three edges

Proof Consider a transformation in the group SL(2, R) taking the point A0 to (1, 0)
and A1 to some point on the line x = 1. It takes both broken lines to some broken
lines with B0 = A0 = (1, 0) and the points A1 and B1 to some points on the line
x = 1. By Proposition 11.9 this operator does not change the continued fraction.
Now we are in a position to apply Corollary 11.10: the images of the points An , Bm ,
and the origin are on a line. Therefore, the points An , Bm , and O are on a line. 

Knowing the continued fraction for a broken line, it is possible to say whether
the corresponding broken line is closed.

Corollary 11.14 (On necessary and sufficient conditions for broken lines to be
closed) Consider a broken line A0 A1 . . . An with the LLS sequence (a0 , a1 ,
. . . , a2n ). Let A0 = (1, 0) and A0 A1 be on the line x = 1. Then the broken line
is closed if and only if

Q2n+1 (a0 , a1 , . . . , a2n ) = 1 and P2n+1 (a0 , a1 , . . . , a2n ) = 0.

The condition P2n+1 = 0 is equivalent to the following one:

[a0 ; a1 : · · · : a2n ] = 0.

Notice also that for an arbitrary broken line the conditions are almost the same, as
seen from Proposition 11.9.

Example 11.15 Let us write the conditions for a broken line consisting of three
edges to form a triangle. Suppose that the LLS sequence of this broken line is
(a0 , a1 , a2 , a3 , a4 ). Then the conditions are introduced by the following system:

a0 a1 a2 a3 a4 + a0 a1 a2 + a0 a1 a4 + a0 a3 a4 + a2 a3 a4 + a0 + a2 + a4 = 0,
a1 a2 a3 a4 + a1 a2 + a1 a4 + a3 a4 + 1 = 1.

In Fig. 11.4 we show two examples of broken lines with three edges.

We conclude this section with the following open problem.

146 11 Geometry of Continued Fractions and Kepler’s Second Law

Problem 6 Consider a broken line L and a proper Euclidean transformation T (the

origin may not be preserved by T ). Find the relation between the elements of the
LLS sequences of the broken lines L and T (L).

Due to Proposition 11.9 it is sufficient to solve this problem only for translations
by a vector.

11.3 Areal and Angular Densities for Differentiable Curves

Let us go even further and consider a differentiable curve instead of a broken line.
In this situation it helps to think of curves as broken lines with infinitesimally small
segments. In some sense the LLS sequence “splits” into a pair of functions related to
the odd and even elements of the LLS sequence. Further, we call these two functions
areal and angular densities. In this section we discuss some basic properties of areal
and angular densities, in particular we shell show that the areal density is the inverse
to the velocity of a body that moves according to the second Kepler law.

11.3.1 Notions of Real and Angular Densities

Throughout this section we suppose that all curves are of class C 2 and have the
arc-length parameterization.

Definition 11.16 Let γ be an arc-length parameterized C 2 -class curve. By defini-

tion the areal density and the angular density at some point t are

|Oγ (t) × Oγ (t + ε)|  

A(t) = lim = Oγ (t) × γ̇ (t)
ε→0 ε
and
|γ (t)γ (t − ε) × γ (t)γ (t + ε)|
B(t) = lim .
ε→0 ε|Oγ (t − ε) × Oγ (t)||Oγ (t) × Oγ (t + ε)|

We continue with the following geometric interpretations for the function A.

Proposition 11.17 (Areal density and the Kepler’s second law) Suppose that a body
moves along the curve γ with velocity 1/A. Then the sector area velocity of the body
is constant and equals 1.

The proposition follows directly from the definition.

Let us show that the value of A2 B at a point t coincides with the signed curvature
at t (which we denote by κ(t)).
11.3 Areal and Angular Densities for Differentiable Curves 147

Proposition 11.18 Let the vectors Oγ (t) and γ̇ (t) be noncollinear at some mo-
ment t. Then the following holds:

A2 (t)B(t) = κ(t).

Proof By definition we have

A2 (t)B(t)

|Oγ (t) × Oγ (t + ε)| 2

= lim
ε→0 ε

|γ (t)γ (t − ε) × γ (t)γ (t + ε)|

×
ε|Oγ (t − ε) × Oγ (t)||Oγ (t) × Oγ (t + ε)|
|γ (t)γ (t − ε) × γ (t)γ (t + ε)|
= lim .
ε→0 ε3
Let us rewrite the curve γ in coordinates, i.e., γ (t) = (x(t), y(t)). Since the param-
eter t is arc-length we have
      
γ (t) = x  (t) 2 + y  (t) 2 = 1 and κ(t) = x  (t)y  (t) − y  (t)x  (t).

Let us consider the Taylor series expansions

ε 2 
γ (t + ε) = γ (t) + εγ  (t) + γ (t) + terms of higher order,
2
ε 2 
γ (t − ε) = γ (t) − εγ  (t) + γ (t) + terms of higher order.
2
Let us substitute

ε 2 
  ε 2 
γ (t)γ (t + ε) = εx + x , εy + y + terms of higher order,
2 2
2

 ε   ε 2 
γ (t)γ (t − ε) = −εx + x , −εy + y + terms of higher order.
2 2

Hence we have
|γ (t)γ (t − ε) × γ (t)γ (t + ε)| ε 3 (x  y  − y  x  ) + 4th order terms
lim 3
= lim
ε→0 ε ε→0 ε3
ε 3 κ(t) + 4th order terms
= lim
ε→0 ε3
= κ(t).

Therefore, A2 (t)B(t) = κ(t). 

148 11 Geometry of Continued Fractions and Kepler’s Second Law

Let us prove the theorem on existence of a curve with a given angular density.
In some sense this theorem is a smooth analogue of the algorithm to construct a
broken line by the elements of the corresponding continued fraction of Sect. 11.2.1.
It is interesting to observe that to reconstruct a curve it is necessary to know the areal
density (which corresponds only to odd elements in the LLS sequences for broken
lines).

Theorem 11.19 Suppose that we know the areal density A(t) smoothly depending
on a parameter t in some neighborhood of t0 , the starting position γ (t0 ), and the
origin O.
– If |A(t0 )| > |Oγ (t0 )|, then there is no finite curve satisfying the above conditions
– If |Oγ (t0 )| > |A(t0 )| > 0, then the curve γ is uniquely defined in some neighbor-
hood of the point γ (t0 ).

Proof In polar coordinates (r, ϕ) with center at O, the curve γ is defined by the
following system of differential equations:
 2
r ϕ̇ = A,
ṙ 2 + r 2 ϕ̇ 2 = 1.
This system is equivalent to the union of the following two systems:
 
ϕ̇ = rA2 , ϕ̇ = rA2 ,
 and 
2 2
ṙ = 1 − Ar 2 , ṙ = − 1 − Ar 2 .

By the main theorem of ordinary differential equations (e.g., see [8]), each of these
two systems has a finite solution if and only if |r| > |A| > 0. 

11.3.2 Curves and Broken Lines

There are many interesting questions related to the convergence of broken lines to
smooth curves in the context of areal and angular densities. Let us first show one
particular result in this direction.
Let γ (t) be a curve with arc-length parameter t ∈ [0, T ] and densities A(t) and
B(t). In addition we suppose that the vectors Oγ (t) and γ̇ (t) are linearly indepen-
dent for all admissible values of t. For a positive integer n we denote by γn the
broken line V0,n . . . Vn,n with vertices

i
Vi,n = γ T .
n
Suppose that the LLS sequence of γn is (a0,n , . . . , a2n,n ). Consider the following
two functions An and Bn :

An (t) = a2nt/T +1,n and Bn (t) = a2nt/T ,n .

11.3 Areal and Angular Densities for Differentiable Curves 149

These functions satisfy the following convergence theorem.

Theorem 11.20 Let γ be a C 2 -curve with arc-length parameterization. Then the

sequence of functions (An ) converges pointwise to the function A, and the sequence
of functions (Bn ) converges pointwise to the function B.

Proof This follows directly from the definition of density functions and the proper-
ties of LLS sequences shown in Proposition 11.5. 

We continue with two open problems. The first one is in some sense an inverse
problem to the statement of Theorem 1.16.

Problem 7 Let a sequence of broken lines Li converge pointwise to a C 2 -curve γ .

Study the relations between LLS sequences of the broken lines Li and their “limits”
from one side and the density functions of γ from the other side.

Notice that it is not clear what kind of limits for the elements of LLS sequence
one should use here.
Finally, it is interesting to have some criterion for the curve to be closed, as was
done for broken lines in Corollary 11.14.

Problem 8 What are the conditions on the functions A(t) and B(t) for the resulting
curve γ to be closed?

11.3.3 Some Examples

Let us write down the areal and angular densities for straight lines, ellipses, and
logarithmic spirals.

Lines Let O be the origin and γ the line x = a. Then we have

A(t) = a and B(t) = 0.

2 2
Ellipses and Their Centers Consider the ellipse xa 2 + yb2 = 1 with a ≥ b > 0 and
let O be the origin, i.e., at the symmetry center of the ellipse. Then we have the
following expressions for the areal and angular densities:
ab 1
A(t) =  and B(t) =  .
a 2 sin2 t + b2 cos2 t ab a 2 sin2 t + b2 cos2 t
Notice that the ratio of the functions A(t) and B(t) is constant:
A(t)
= a 2 b2 .
B(t)
150 11 Geometry of Continued Fractions and Kepler’s Second Law

2 2
Ellipses and Their Foci Consider again the ellipse xa 2 + yb2 = 1 with a ≥ b > 0.
√
Let now O be the point (− a 2 − b2 , 0), which is one of the foci of the ellipse. Then
the areal and the angular densities are as follows:
√
ab + b a 2 − b2 cos t
A(t) = 
a 2 sin2 t + b2 cos2 t
and
a
B(t) =  √ .
b a 2 sin2 t + b2 cos2 t(a + cos t a 2 − b2 )2

Remark on Keplerian planetary motion Consider the Sun and some planet that
moves around the Sun. Then the planet moves according to Kepler’s laws:
I. The orbit of the planet is an ellipse with the Sun at one of the two foci of the
ellipse.
II. The motion has constant sector area velocity.
III. The square of the orbital period of the planet is proportional to the cube of the
semimajor axis of its orbit.
2 2
The trajectory of the planet is an ellipse defined by the equation xa 2 + yb2 = 1
with a ≥ b > 0 in some Euclidean coordinates (according to Kepler’s first law), in
addition, the Sun is one of the foci. Then according to Kepler’s second law, the
planet moves with velocity λ/A(t) for all t. The constant λ is defined from Kepler’s
third law:
T 2 Te2
= 3,
a3 ae
where by T we denote the orbital period of our planet, and by Te and ae , respec-
tively, the orbital period and the semimajor axis for Earth. Denote by L the length
of the ellipse for the planet, i.e.,
 π/2 

b2
L = 4a 1 − 1 − 2 cos2 tdt.
0 a
L
Since T = |λ| 0 |1/A(t)|dt, we have

3/2
Te a
λ = ± L .
0 |1/A(t)|dt ae

Similar situations hold for parabolic and hyperbolic motions.

Logarithmic Spirals Consider the logarithmic spiral

 bt  
ae cos t, aebt sin t | t ∈ R .
11.4 Exercises 151

Then the areal and angular densities are expressed as follows:

√
aebt e−3bt b2 + 1
A(t) = √ and B(t) = .
b2 + 1 a3

Observe that
1
A3 (t)B(t) = ,
b2 + 1
meaning that for logarithmic spirals the function A3 B is constant.

Remark 11.21 When we consider lines, ellipses, and spirals we have certain rela-
tions between densities A and B: the functions A, A/B, and A3 B respectively are
constant. Notice that if A2 B is a constant function, then the curvature is constant,
and hence we get circles. It is natural to study what kind of curves are defined by
certain relations between the densities. For instance, the following question remains
open: what curves do we get if AB (or simply B) is constant?

11.4 Exercises

Exercise 11.1 Find a proof of Proposition 11.8.

Exercise 11.2 Draw a broken line on five vertices whose LLS sequence satisfies
(a) all elements are positive;
(b) all elements are negative;
(c) even elements are positive and odd elements are negative;
(d) odd elements are positive and even elements are negative.

Exercise 11.3 Calculate the areal and angular densities for

(a) straight lines;
(b) ellipses;
(c) logarithmic spirals.
Chapter 12
Extended Integer Angles and Their Summation

Let us start with the following question. Suppose that we have arbitrary numbers a,
b, and c satisfying
a + b = c.
How do we calculate the continued fraction for c knowing the continued fractions
for a and b? It turns out that this question is not a natural question within the theory
of continued fractions. The simplest algorithm works as follows:
– first calculate the rational number representations for a and b;
– add them according to basic school arithmetic;
– write the continued fraction for the sum.
One can hardly imagine any law to write the continued fraction for the sum directly.
The main obstacle here is that the summation of rational numbers does not have
a geometric explanation in terms of the integer lattice. In this chapter we propose
to consider a “geometric summation” of continued fractions, which we consider a
summation of integer angles.
In Sect. 12.1 we start with the notion of extended integer angles. These angles
are the integer analogues of Euclidean angles of the type kπ + ϕ for arbitrary in-
tegers k. We classify extended angles by writing normal forms representing all of
them. Finally, we define the M-sums of extended angles and integer angles. Further,
in Sect. 12.2, we show how the continued fractions of extended angles are expressed
in terms of the corresponding normal forms. Finally, in Sect. 12.3 we give a proof
of Theorem 6.8(i) on the sum of integer angles in integer triangles, which is based
on the techniques introduced in this chapter.

12.1 Extension of Integer Angles. Notion of Sums of Integer

Angles
We start with a few general definitions. An integer affine transformation of the plane
is said to be proper if it preserves the orientation of the plane. We say that two

O. Karpenkov, Geometry of Continued Fractions, 153

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_12, © Springer-Verlag Berlin Heidelberg 2013
154 12 Extended Integer Angles and Their Summation

sets S1 and S2 are proper integer congruent if there exists a proper integer affine
transformation of R2 taking the set S1 to the set S2 , which we write as S1 ∼
ˆ S2 .
=

12.1.1 Extended Integer Angles and Revolution Number

12.1.1.1 Equivalence Classes of Integer Oriented Broken Lines and the

Corresponding Extended Angles

We say that an integer oriented broken line An An−1 An−2 . . . A0 is inverse to the
integer oriented broken line A0 A1 A2 . . . An .

Definition 12.1 Consider an integer point V . Two integer oriented V -broken lines
l1 and l2 (see Definition 11.1) are said to be equivalent if their first and their last
vertices coincide respectively and the closed broken line generated by l1 and the
inverse of l2 is contractible in R2 \ {V }.

Definition 12.2 The equivalence class of integer oriented V -broken lines contain-
ing a broken line A0 A1 . . . An is said to be the extended integer angle for the bro-
ken line A0 A1 . . . An with vertex V (or for short, extended angle). We denote it by
∠(V , A0 A1 . . . An ).

The set of extended angles is invariant with respect to the action of (proper)
integer affine transformations. Therefore, (proper) integer congruence for extended
angles is well defined.

12.1.1.2 Revolution Numbers for Extended Angles

Let v be an arbitrary vector, and V an integer point. Consider the ray

r = {V + λv | λ ≥ 0}.

Let AB be an arbitrary (oriented from A to B) segment not contained in r. Suppose

also that V is not in AB. We denote by #(r, AB) the number
⎧
⎪
⎨0, AB does not intersect r,
#(r, AB) = 12 sgn(A(A + v)B), AB ∩ r ∈ {A, B},
⎪
⎩
sgn(A(A + v)B), AB ∩ r ∈ AB \ {A, B},

and call it the intersection number of the ray r and the segment AB.
12.1 Extension of Integer Angles. Notion of Sums of Integer Angles 155

Definition 12.3 Let A0 A1 . . . An be a V -broken line, and let r be a ray {V + λv |

λ ≥ 0}. We call the number

n
#(r, Ai−1 Ai )
i=1

the intersection number of the ray r the V -broken line A0 A1 . . . An and denote it by
#(r, A0 A1 . . . An ).

Definition 12.4 Consider an arbitrary extended angle ∠(V , A0 A1 . . . An ). Define

r+ = {V + λVA0 | λ ≥ 0} and r− = {V − λVA0 | λ ≥ 0}.

The number
1 
#(r+ , A0 A1 . . . An ) + #(r− , A0 A1 . . . An )
2
is called the revolution number for the extended angle ∠(V , A0 A1 . . . An ).
We denote it by #(∠(V , A0 A1 . . . An )). Additionally, we put by definition
#(∠(V , A0 )) = 0.

Example 12.5 Let O = (0, 0), A = (1, 0), B = (0, 1), and C = (−1, −1). Then
    1
# ∠(O, A) = 0, # ∠(O, AB) = ,
4
    3
# ∠(O, ABCA) = 1, # ∠(O, ACB) = − .
4
Let us show that the definition of the revolution number is well defined.

Proposition 12.6 The revolution number of any extended angle is well defined.

Proof Consider an arbitrary V -broken line and the corresponding extended angle
∠(V , A0 A1 . . . An ). Let

r+ = {V + λVA0 | λ ≥ 0} and r− = {V − λVA0 | λ ≥ 0}.

By definition, every segment of the broken line A0 A1 . . . An is at unit integer dis-

tance from V . Hence this broken line does not contain V , and the rays r+ and r− do
not contain edges of the broken line.
Suppose that
 
∠(V , A0 A1 . . . An ) = ∠ V  , A0 A1 . . . Am .

Let us show that

    
# ∠(V , A0 A1 . . . An ) = # ∠ V  , A0 A1 . . . Am .
156 12 Extended Integer Angles and Their Summation

From the definition we have V = V  , A0 = A0 , An = Am , and the broken line
A0 A1 . . . An Am−1 . . . A1 A0 is contractible in R2 \ {V }. This implies that
    
# ∠(V , A0 A1 . . . An ) − # ∠ V , A0 A1 . . . Am
1    
= # r+ , A0 A1 . . . An Am−1 . . . A1 A0 + # r− , A0 A1 . . . An Am−1 . . . A1 A0
2
=0+0=0

(we leave to the reader the proof of the second equality; see Exercise 12.2). Hence,
    
# ∠(V , A0 A1 . . . An ) = # ∠ V  , A0 A1 . . . Am .

Therefore, the revolution number of any extended angle is well defined. 

Remark Notice that the revolution number of a closed V -broken line coincides with
the degree of a point V with respect to the V -broken line.

Let us formulate one important property of the revolution number.

Proposition 12.7 The revolution number of extended angles is invariant under

proper integer congruences.

Finally, we formulate a nice expression for the rotation number of a closed in-
teger broken line with integer vertices not passing through the origin introduced in
a recent preprint [79] by A. Higashitani and M. Masuda. Recall that the rotation
number about the integer point V of a closed broken line L not passing through
V is the degree of the projection of L to the unit circle along the rays with vertex
at V . We denote the rotation number by RotV (L). Notice that a rotation number for
V -broken lines coincides with the revolution number of the corresponding extended
angles.

Proposition 12.8 Consider a closed integer V -broken line L and enumerate all
of its integer points A1 , . . . , Ad (not only vertices). In addition, set A0 = Ad and
Ad+1 = A1 . Then we have

1 1 
d d
RotV (L) = εi + ai ,
4 12
i=1 i=1

where εi = det(VAi , VAi+1 ) and ai is defined from the equation

εi−1 VAi−1 + εi VAi+1 + ai VAi = 0.

For the proof of this proposition we refer to [79] and [211].

12.1 Extension of Integer Angles. Notion of Sums of Integer Angles 157

12.1.1.3 Zero Integer Angles

In the next theorem we use zero integer angles and their trigonometric functions.
Let A, B, and C be three integer points on a straight line. Suppose that B is distinct
from A and C and that it is not between A and C. We say that the integer angle
∠ABC with vertex at B is zero. Further, we put by definition

lsin(∠ABC) = 0, lcos(∠ABC) = 1, ltan(∠ABC) = 0.

Denote by larctan(0) the angle ∠AOA, where A = (1, 0) and O is the origin.

12.1.2 On Normal Forms of Extended Angles

While working with an abstract definition of extended angles it is useful to keep in

mind particular broken lines that characterize the corresponding equivalence classes,
which we shall call the normal forms of these broken lines.
Denote a sequence

(a0 , . . . , an , a0 , . . . , an , . . . , a0 , . . . , an , b0 , . . . , bm ).
!" #
k times

by
 
(a0 , . . . , an )k , b0 , . . . , bm .

Definition 12.9 (Normal forms of extended angles) Consider an integer oriented

O-broken line A0 A1 . . . As , where O is the origin. Let A0 = (1, 0) and (if s > 0) let
A1 be on the straight line x = 1.
(I) We say that the extended angle ∠(O, A0 ) is of Type I and denote it by 0π +
larctan(0) (or 0, for short). The empty sequence is said to be characteristic for
0π + larctan(0).
If the LSLS sequence of the broken line A0 A1 . . . As coincides with one of the
following sequences (we call it the characteristic sequence for the corresponding
angle), then:
(IIk ) If ((1, −2, 1, −2)k−1 , 1, −2, 1), where k ≥ 1, then we denote the angle Φ0
by kπ + larctan(0) (or kπ , for short) and say that Φ0 is of Type IIk ;
(IIIk ) If ((−1, 2, −1, 2)k−1 , −1, 2, −1), where k ≥ 1, then we denote the angle Φ0
by −kπ + larctan(0) (or −kπ , for short) and say that Φ0 is of Type IIIk ;
(IVk ) If ((1, −2, 1, −2)k , a0 , . . . , a2n ), where k ≥ 0, n ≥ 0, ai > 0, for i =
0, . . . , 2n, then we denote the angle Φ0 by kπ + larctan([a0 ; a1 : · · · : a2n ])
and say that Φ0 is of Type IVk ;
158 12 Extended Integer Angles and Their Summation

Fig. 12.1 Examples of normal forms

(Vk ) If ((−1, 2, −1, 2)k , a0 , . . . , a2n ), where k > 0, n ≥ 0, ai > 0, for i =

0, . . . , 2n, then we denote the angle Φ0 by −kπ + larctan([a0 ; a1 : · · · : a2n ])
and say that Φ0 is of Type Vk .

See examples of several normal forms in Fig. 12.1.

Theorem 12.10 For every extended angle Φ there exists a unique normal angle
of Definition 12.9 that is proper integer congruent to Φ. (This angle is called the
normal form of Φ.)

We start the proof with the following lemma.

Lemma 12.11 Consider integers m, k ≥ 1, and ai > 0 for i = 0, . . . , 2n.

(i) Suppose that the LSLS sequences for the extended angles Φ1 and Φ2 are re-
spectively
 
(1, −2, 1, −2)k−1 , 1, −2, 1, −2, a0 , . . . , a2n and
 
(1, −2, 1, −2)k−1 , 1, −2, 1, m, a0 , . . . , a2n .

Then Φ1 is proper integer congruent to Φ2 .

12.1 Extension of Integer Angles. Notion of Sums of Integer Angles 159

(ii) Suppose that the LSLS sequences for the extended angles Φ1 and Φ2 are re-
spectively
 
(−1, 2, −1, 2)k−1 , −1, 2, −1, m, a0 , . . . , a2n and
 
(−1, 2, −1, 2)k−1 , −1, 2, −1, 2, a0 , . . . , a2n .

Then Φ1 is proper integer congruent to Φ2 .

Proof We start the proof with the first statement of the lemma. Without loss of
generality we assume that the vertices of the extended angles Φ1 and Φ2 are at the
origin, say

Φ1 = ∠(O, A0 . . . A2k+n+1 ),
Φ2 = ∠(O, B0 . . . B2k+n+1 ).

Additionally we assume that A0 = B0 = (1, 0) and the points A1 and B2 are on the
lines x = 1 and (m + 1)y = x respectively. These conditions together with the LSLS
sequences uniquely identify the vertices
⎧
⎪
⎪ A2l = ((−1)l , 0), for l < k − 1,
⎪
⎨ A2l+1 = ((−1) , (−1) ),
l l for l < k − 1,
⎪
⎪ A = ((−1) k , 0),
⎪
⎩
2k
A2k+1 = ((−1)k , (−1)k a0 ),

and
⎧
⎪
⎪B2l = ((−1)l , 0), for l < k − 1,
⎪
⎨B
2l+1 = ((−1) (−m − 1), (−1) ), for l < k − 1,
l l

⎪
⎪B2k = ((−1)k , 0),
⎪
⎩
B2k+1 = ((−1)k , (−1)k a0 ).
Hence A2k = B2k and A2k+1 = B2k+1 . Notice that the remaining parts of both LSLS
sequences (i.e., (a0 , . . . , a2n )) coincide. Therefore, the point Al coincides with the
point Bl for l > 2k.
On the one hand, it is clear that the integer oriented broken line

A0 . . . A2k (= Bk )Bk−1 Bk−2 . . . B0

is contractible. On the other hand, Al = Bl for l > 2k. Therefore, the angle Φ1 is
proper integer congruent to Φ2 . This concludes the proof of Lemma 12.11(i).
The proof of Lemma 12.11(ii) is similar, and we leave it as an exercise for the
reader. 

Proof of Theorem 12.10 Uniqueness of normal forms. Let us first show that the
extended angles listed in Definition 12.9 are pairwise proper integer noncongruent.
160 12 Extended Integer Angles and Their Summation

Consider the revolution numbers of the extended angles listed in Definition 12.9:

Types
I IIk (k ≥ 0) IIIk (k ≥ 0) IVk (k ≥ 0) Vk (k > 0)

Revolution numbers 0 1/2(k + 1) −1/2(k + 1) 1/4 + 1/2k 1/4 − 1/2k

Therefore, the revolution numbers of extended angles distinguish the types of the
angles.
For Types I, IIk , and IIIk the proof of uniqueness is complete, since any such
type contains exactly one extended angle.
Now we prove that any two normal forms of Type IVk are not proper integer
congruent for any integer k ≥ 0. Consider the extended angle
 
Φ = kπ + larctan [a0 ; a1 : · · · : a2n ]

of Definition 12.9. Assume that the O-broken line for the angle Φ is A0 A1 . . . Am ,
where m = 2|k| + n + 1. Notice that the LSLS sequence for this broken line is
characteristic for Φ. We consider the two cases of odd and even k separately.
If k is even, then the integer angle ∠A0 OAm is proper integer congruent to the
integer angle larctan([a0 ; a1 : · · · : a2n ]). The integer arctangent is a proper inte-
ger affine invariant for Φ. This invariant distinguishes all the extended angles of
Type IVk with even k.
Suppose now that k is odd. Set B = O + A0 O. The integer angle ∠BVAm is
proper integer congruent to larctan([a0 ; a1 : · · · : a2n ]). The integer arctangent is a
proper integer affine invariant for the extended angle Φ, distinguishing extended
angles of Type IVk with odd k.
The proof of uniqueness of normal forms of Type Vk repeats the proof for normal
forms of Type IVk .
Therefore, the extended angles listed in Definition 12.9 are not proper integer
congruent.
Existence of normal forms. Now we prove that every extended angle is proper
integer congruent to one of the normal forms.
Consider an arbitrary extended angle Φ = ∠(V , A0 A1 . . . An ). If #(Φ) = k/2
for some integer k, then Φ is proper integer congruent to an angle of one of the
Types I–III. If #(Φ) = 1/4, then the extended angle Φ is proper integer congruent
to the extended angle defined by the sail of the integer angle ∠A0 VAn of Type IV0 .
Suppose now that #(Φ) = 1/4 + k/2 for some positive integer k. Choose the
broken line defining Φ whose LSLS sequence is of the following form:
 
(1, −2, 1, −2)k−1 , 1, −2, 1, m, a0 , . . . , a2n , (12.1)

where ai > 0, for i = 0, . . . , 2n. Notice that it is always possible to construct such a
broken line: first, construct the broken line for
 
(1, −2, 1, −2)k−1 , 1, −2, 1 .
12.1 Extension of Integer Angles. Notion of Sums of Integer Angles 161

Secondly, take the sail of the corresponding angle. The union of the broken line
and the sail is a broken line whose LSLS sequence is exactly as in (12.1). By
Lemma 12.11 the extended angle Φ defined by such an LSLS sequence is proper
integer congruent to the extended angle of Type IVk defined by the sequence
 
(1, −2, 1, −2)k−1 , 1, −2, 1, −2, a0 , . . . , a2n ,

which is in the list of normal forms.

Finally, we consider the case #(Φ) = 1/4 − k/2 for some positive integer k. As
in the previous case there exists a broken line defining Φ whose LSLS sequence is
 
(−1, 2, −1, 2)k−1 , −1, 2, −1, m, a0 , . . . , a2n ,

where ai > 0, for i = 0, . . . , 2n. By Lemma 12.11 the extended angle defined by
this sequence is proper integer congruent to the extended angle of Type Vk defined
by the sequence
 
(−1, 2, −1, 2)k−1 , −1, 2, −1, 2, a0 , . . . , a2n ,

which is again in the list of normal forms.

This completes the proof of Theorem 12.10. 

Let us reformulate Theorem 12.10 in the following way.

Corollary 12.12 Two extended angles are proper integer congruent if and only if
they have the same normal form.

12.1.3 Trigonometry of Extended Angles. Associated Integer

Angles

Having the list of normal forms, we can extend the trigonometric functions to the
case of extended integer angles.

Definition 12.13 Consider an extended angle Φ with the normal form kπ + ϕ for
some integer (possible zero) angle ϕ and for an integer k.
(i) The integer angle ϕ is said to be associated with the extended angle Φ.
(ii) The numbers ltan(ϕ), lsin(ϕ), and lcos(ϕ) are called the integer tangent, the
integer sine, and the integer cosine of the extended angle Φ.

There exists a canonical embedding of the set of all integer angles into the set of
all extended angles, since every sail is an integer broken line at distance one from
the vertex of the angle.
162 12 Extended Integer Angles and Their Summation

Definition 12.14 Consider an arbitrary integer angle ϕ. The angle

0π + larctan(ltan ϕ)

is said to be corresponding to the angle ϕ, which we denote by ϕ.

From Theorem 12.10 it follows that for every integer angle ϕ there exists a unique
extended angle ϕ corresponding to ϕ. Therefore, two integer angles ϕ1 and ϕ2 are
proper integer congruent if and only if the corresponding extended angles ϕ 1 and
ϕ 2 are proper integer congruent.

12.1.4 Opposite Extended Angles

As in Euclidean geometry, in integer geometry every extended angle possesses an

opposite angle.

Definition 12.15 Consider an extended angle

Φ = ∠(V , A0 A1 . . . An ).

The angle
∠(V , An An−1 . . . A0 )
is said to be opposite to Φ, which we denote by −Φ.

Let us write the normal form for the opposite extended angle.

Proposition 12.16 For every extended angle Φ ∼

ˆ kπ + ϕ we have
=

−Φ ∼
ˆ (−k − 1)π + (π − ϕ).
=

12.1.5 Sums of Extended Angles

In the next definition we introduce sums of integer and extended angles. A sum of
two angles is not uniquely defined, in contrast to Euclidean geometry.

Definition 12.17 Consider arbitrary extended angles Φi , where i = 1, . . . , l for

l > 1. Let the normal forms of the Φi have the characteristic sequence

(a0,i , a1,i , . . . , a2ni ,i ) for i = 1, . . . , l.

12.1 Extension of Integer Angles. Notion of Sums of Integer Angles 163

Let M = (m1 , . . . , ml−1 ) be an arbitrary (l − 1)-tuple of integers. Construct an ex-

tended angle Φ with characteristic sequence

(a0,1 , a1,1 , . . . , a2n1 ,1 , m1 , a0,2 , . . . , a2n2 ,2 , m2 , . . . , ml−1 , a0,l , . . . , a2nl ,l ).

The normal form of Φ is called the M-sum of extended angles Φi (i = 1, . . . , l) and

is denoted by


l
Φi , or equivalently by Φ1 +m1 Φ2 +m2 · · · +ml−1 Φl .
M,i=1

It is clear that M-sums of extended angles are defined in a unique way up to the
set choice of the set M.

Example 12.18 Let Φ = 0π + larctan 1. Then

Φ +−3 Φ = π + larctan 1,
Φ +−2 Φ = π + larctan 0,
Φ +−1 Φ = 0π + larctan 1,
Φ +0 Φ = 0π + larctan 2,
3
Φ +1 Φ = 0π + larctan .
2

It is interesting to observe that the M-sum of extended angles is neither associa-

tive nor commutative.

Proposition 12.19 The M-sum of extended angles is nonassociative.

Proof For example, let

Φ1 = 0π + larctan 2,
3
Φ2 = 0π + larctan ,
2
Φ3 = 0π + larctan 5.

Then

Φ1 +−1 Φ2 +−1 Φ3 = π + larctan 4,

Φ1 +−1 (Φ2 +−1 Φ3 ) = 2π + larctan 0,
(Φ1 +−1 Φ2 ) +−1 Φ3 = 0π + larctan 1. 

Proposition 12.20 The M-sum of extended angles is noncommutative.

164 12 Extended Integer Angles and Their Summation

Proof For example, let

5
Φ1 = 0π + larctan 1 and Φ2 = 0π + larctan .
2
Then
12 13
Φ1 +1 Φ2 = 0π + larctan
= 0π + larctan = Φ2 + 1 Φ1 . 
7 5
Remark 12.21 The definition of M-sums of extended angles is canonically lifted to
the case of classes of proper integer congruences of extended angles.

12.1.6 Sums of Integer Angles

We conclude this section with the definition of M-sums of integer angles.

Definition 12.22 Consider integer angles αi , where i = 1, . . . , l for some l ≥ 2. Let

α i be the corresponding extended angles for αi , and let M = (m1 , . . . , ml−1 ) be an
arbitrary (l − 1)-tuple of integers. The integer angle ϕ associated with the extended
angle
Φ = α 1 +m1 α 2 +m2 · · · +ml−1 α l
is called the M-sum of integer angles αi (i = 1, . . . , l) and denoted by


l
αi , or equivalently by α1 +m1 α2 +m2 · · · +ml−1 αl .
M,i=1

Remark 12.23 The definition of M-sums of integer angles is also canonically lifted
to the case of classes of proper integer congruences of integer angles.

12.2 Relations Between Extended and Integer Angles

Recall that r denotes the maximal integer not greater than r.

Theorem 12.24 Consider an extended angle Φ = ∠(V , A0 A1 . . . An ). Suppose that

the normal form for Φ is kπ + ϕ for some pair (k, ϕ). Let (a0 , a1 , . . . , a2n−2 ) be the
LSLS sequence for the integer oriented broken line A0 A1 . . . An . Suppose that
p
[a0 ; a1 : · · · : a2n−2 ] = .
q
12.3 Proof of Theorem 6.8(i) 165

Then the following hold:

⎧
⎪
⎪ if p/q = ∞,
⎪larctan 1,
⎪
⎪
⎪ larctan pq , if p/q ≥ 1,
⎪
⎪
⎪
⎨larctan |p|
|q|−(|q|−1)/|p||p| , if 0 < p/q < 1,
ϕ∼=
⎪
⎪ 0, if p/q = 0,
⎪
⎪
⎪
⎪ |p|
π − larctan |q|−(|q|−1)/|p||p| if −1 < p/q < 0,
⎪
⎪
,
⎪
⎩π − larctan(− p ),
q if p/q ≤ −1.

Proof Without loss of generality we assume that V is the origin O, A0 = (1, 0), and
1  
A0 + sgn A0 O A1 A0 A1 = (1, 1).
a0
(One can get this after a certain integer affine transformation of the plane.)
By Theorem 3.1 the coordinates of the point An are (q, p). This directly implies
the statement of the theorem for the cases p > q > 0, p/q = 0, and p/q = ∞.
Suppose now that q > p > 0. Consider the integer angle ϕ = ∠A0 PAn . Let
B0 . . . Bm be the sail for ϕ. Direct calculation shows that the point
B0 B1
D = B0 +
l(B0 B1 )
coincides with the point (1 + (q − 1)/p, 1). Consider the proper integer linear
transformation T satisfying the following two conditions: first, T takes the point
A0 = B0 to itself, and second, T takes the point D to (1, 1) . These two conditions
uniquely identify T as

1 −(q − 1)/p
T= .
0 1
The transformation T takes the point An = Bm , with coordinates (q, p), to the point
with coordinates (q − (p − 1)/pp, p) . Therefore,
p
ϕ = larctan .
q − (q − 1)/pp
The proof for the case p > 0 and q < 0 repeats the proof described above after
switching to the corresponding adjacent angles.
Finally, the case with p < 0 is centrally symmetric with respect to all the cases
considered above. This concludes the proof of Theorem 12.24. 

12.3 Proof of Theorem 6.8(i)

Now we return to the proof of the first statement of the theorem on sums of integer
tangents for integer angles in integer triangles:
166 12 Extended Integer Angles and Their Summation

Fig. 12.2 For the given

triangle ABC we have
l1 (AB; C) = 5

Theorem 6.8(i) (On sums of integer tangents of angles in integer triangles) Let
(α1 , α2 , α3 ) be an ordered triple of angles. There exists a triangle with consecutive
angles integer congruent to α1 , α2 , and α3 if and only if there exists i ∈ {1, 2, 3}
such that the angles α = αi , β = αi+1(mod 3) , γ = αi+2(mod 3) satisfy the following
conditions:
(a) for ξ = ] ltan α, −1, ltan β[ the following holds ξ < 0, or ξ > ltan α, or ξ = ∞;
(b) ] ltan α, −1, ltan β, −1, ltan γ [ = 0.

12.3.1 Two Preliminary Lemmas

We begin the proof of Theorem 6.8(i) with two preliminary lemmas.

Definition 12.25 Let ABC be an integer triangle. Consider all integer points in
the interior of ABC and on its sides lying at unit integer distance from the side AB.
We denote their number by l1 (AB; C) (see Fig. 12.2).

Recall that all integer points at unit integer distance from AB described in the
definition are contained in one straight line parallel to AB. Besides that,

0 < l1 (AB; C) < l(AB).

Now we prove the following lemma.

Lemma 12.26 For every integer triangle ABC the following holds:

∠CAB +l(AB)−l1 (AB;C)−1 ∠ABC +l(BC)−l1 (BC;A)−1 ∠BCA = π

(where by ϕ we denote the extended angle corresponding to ϕ).

Proof Consider an arbitrary integer triangle ABC. Suppose that the pair of vectors
(BA, BC) defines the positive orientation of the plane (otherwise, we consider the
triangle ACB). Set

D = A + CB and E = A + CA

(see Fig. 12.4 on page 169).

12.3 Proof of Theorem 6.8(i) 167

Since CADB is a parallelogram, we have

BAD ∼
ˆ ABC,
=

and hence
∠BAD ∼
ˆ ∠ABC
= and l1 (BA; D) = l1 (AB; C).
Since EABD is a parallelogram, we get

AED ∼
ˆ BAD ∼
= ˆ ABC.
=

Therefore,
∠DAE ∼
ˆ ∠BCA
= and l1 (DA; E) = l1 (BC; A).
Let A0 . . . An be the sail of the angle ∠CAB with the corresponding LLS se-
quence
(a0 , . . . , a2n−2 ).
Further, let B0 B1 . . . Bm be the sail of ∠BAD (where B0 = An ) with the correspond-
ing LLS sequence
(b0 , . . . , b2m−2 ).
Finally, let C0 C1 . . . Cl be the sail of ∠DAE (where C0 = Bm ) with the correspond-
ing LLS sequence
(c0 , . . . , c2l−2 ).
Consider now the A-broken line

A0 . . . An B1 B2 . . . Bm C1 C2 . . . Cl .

The LSLS sequence for this broken line is

(a0 , . . . , a2n−2 , t, b0 , . . . , b2m−2 , u, c0 , . . . , c2l−2 ),

where t and u are certain integers. By the definition of the M-sum of extended
angles this sequence defines the extended angle

∠CAB +t ∠BAD +u ∠DAE,

which is equal to π by construction.

Now let us compute t. Denote by An the integer point in the segment An−1 An
that is next to An . Consider the set of integer points at unit integer distance from
AB and lying in the half-plane with AB at the boundary and containing the point D.
This set coincides with the following set (see an example in Fig. 12.3):
 
An,k = An + An An + kAAn | k ∈ Z .
168 12 Extended Integer Angles and Their Summation

Fig. 12.3 Integer points An,t

Denote by π+ the open half-plane bounded by the straight line AC and containing
the point B. Let also π− denote the complement of π+ in the plane. Since

An,−2 = A + An A,

the integer points An,k for k less than or equal to −2 are in the closed half-plane π− .
Since
An,−1 = A + An An ,
the integer points An,k for k greater than or equal to −1 are in the open half-
plane π+ .
The intersection of the parallelogram AEDB and the open half-plane π+ contains
exactly l(AB) points of the described set: these points are the points An,k with
−1 ≤ k ≤ l(AB) − 2.
Since
BAD ∼
ˆ ABC,
=
the number of points An,k in the closed triangle BAD is l1 (AB; C): these points
are exactly the points An,k with k satisfying

l(AB) − l1 (AB; C) − 1 ≤ k ≤ l(AB) − 2.

Denote by k0 the integer

l(AB) − l1 (AB; C) − 1.

The point An,k0 is contained in the segment B0 B1 of the sail of the integer angle
∠BAD (see Fig. 12.4). Since

∠BAD ∼
ˆ ∠ABC,
=
12.3 Proof of Theorem 6.8(i) 169

Fig. 12.4 The point An,k0

we have

t = sgn(An−1 AAn ) sgn(An AB1 ) sgn(An−1 An B1 ) lsin ∠An−1 An B1

= 1 · 1 · sgn(An−1 An An,k0 ) lsin ∠An−1 An An,k0
= sign(k0 )|k0 | = k0 = l(AB) − l1 (AB; C) − 1.

For the same reason,

u = l(DA) − l1 (DA; E) − 1 = l(BC) − l1 (BC; A) − 1.

Therefore,

∠CAB +l(AB)−l1 (AB;C)−1 ∠ABC +l(BC)−l1 (BC;A)−1 ∠BCA = π.

The proof is complete. 

Lemma 12.27 Let α, β, and γ be nonzero integer angles. Suppose that

α +u β +v γ = π.

Then there exists a triangle with three consecutive integer angles proper integer
congruent to α, β, and γ respectively.

Proof Let
O = (0, 0), A = (1, 0), and D = (−1, 0).
Choose the integer points

B = (q1 , p1 ) and C = (q2 , p2 )

with integers p1 , p2 and positive integers q1 , q2 such that

∠AOB = larctan(ltan α) and ∠AOC = ∠AOB +u β.

From the construction, the pair of vectors (OB, OC) defines the positive orientation,
and ∠BOC ∼ ˆ β. Since
=
α +u β +v γ = π and α +u β ∼
ˆ ∠AOC,
=
we have ∠COD ∼
ˆ γ.
=
170 12 Extended Integer Angles and Their Summation

Define
B  = (q1 p2 , p1 p2 ) and C  = (q2 p1 , p1 p2 )
and consider the triangle B OC . The coincidence of angles

∠B OC = ∠BOC

implies
∠B OC ∼
ˆ β.
=
Since β
= 0, the points B  and C  are distinct and the straight line B C does not
coincide with the straight line OA. Since the second coordinates of both points B 
and C  equal q1 q2 , the straight line B C is parallel to the straight line OA. Therefore,
by Proposition 5.16 we have

∠C B O ∼
= ∠AOB = ∠AOB ∼
=α = ∠C OD = ∠COD ∼
and ∠OC B ∼ = γ.

So, the consecutive integer angles of the triangle B OC are integer congruent
to α, β, and γ . 

12.3.2 Conclusion of the Proof of Theorem 6.8(i)

Let α, β, and γ be nonzero integer angles satisfying the conditions (a) and (b) of
Theorem 6.8(i). From the second condition
 
ltan(α), −1, ltan(β), −1, ltan(γ ) = 0,

it follows that
α +−1 β +−1 γ = kπ.
Since all three tangents are positive, we have either k = 1 or k = 2.
By the first condition,

] ltan α, −1, ltan β[ < 0 or ] ltan α, −1, ltan β[ > ltan α.

Hence
α +−1 β = 0π + ϕ,
for some integer angle ϕ, and hence k ≤ 1. Together with the above, this implies
that k = 1. Therefore, by Lemma 12.27 there exists a triangle with three consecutive
integer angles integer congruent to α, β, and γ , respectively.
Let us prove the converse statement. First we prove condition (b) of Theo-
rem 6.8(i). We do it by reductio ad absurdum. Suppose that there exists a triangle
ABC with consecutive integer angles

α = ∠CAB, β = ∠ABC, and γ = ∠BCA,

12.3 Proof of Theorem 6.8(i) 171

such that
⎧
⎪
⎨] ltan α, −1, ltan β, −1, ltan γ [
= 0,
] ltan β, −1, ltan γ , −1, ltan α[
= 0,
⎪
⎩
] ltan γ , −1, ltan α, −1, ltan β[
= 0.

This system of inequalities and Lemma 12.26 imply that at least two of the integers

l(AB) − l1 (AB; C) − 1, l(BC) − l1 (BC; A) − 1, and

l(CA) − l1 (CA; B) − 1

are nonnegative.
Without loss of generality we suppose that

l(AB) − l1 (AB; C) − 1 ≥ 0,
l(BC) − l1 (BC; A) − 1 ≥ 0.

On the one hand, since all integers of the continued fraction

 
r = ltan(α), l(AB) − l1 (AB; C) − 1, ltan(β), l(BC) − l1 (BC; A) − 1, ltan(γ )

are nonnegative and the last one is positive, either r > 0 or r = ∞. On the other
hand, by Lemma 12.26 and by Theorem 12.24, we have that r = 0/−1 = 0. We
arrive at a contradiction. Hence the second condition always holds.
Now we prove that condition (a) holds. Consider a triangle ABC with consec-
utive integer angles

α = ∠CAB, β = ∠ABC, and γ = ∠BCA.

We have proven the second condition, so without loss of generality we assume that
 
ltan(α), −1, ltan(β), −1, ltan(γ ) = 0.

Since

α +−1 β +−1 γ = π,

we have

α +−1 β = 0π + ϕ

for some integer angle ϕ. The last expression for ϕ implies the first condition of the
theorem.
This concludes the proof of Theorem 6.8(i).
172 12 Extended Integer Angles and Their Summation

12.4 Exercises

Exercise 12.1 Find the revolution number of the O-broken line with vertices

A0 = (1, 0), A1 (−1, 2), A2 = (0, −1),

A3 = (1, 1), A4 = (1, −2).

Find the normal form of the extended angle ∠(O, A0 A1 A2 A3 A4 ).

Exercise 12.2 Let A0 A1 . . . An be a closed V -broken line. Suppose that it is con-

tractible in R2 \ {V }. Prove that for every ray r with vertex at V we have

#(r, A0 A1 . . . An ) = 0.

Exercise 12.3 Find a proof of Lemma 12.11(ii).

Exercise 12.4
(a) Prove that every extended angle possesses a unique opposite extended angle.
(b) Prove that the M-sum is uniquely defined by the ordered sequence of extended
angles and the set of integers M.

Exercise 12.5 Calculate the normal forms for the following angles

5
(2π + larctan 1) +−1 3π + larctan ;
2

4 7
4π + larctan +0 3π + larctan ;
3 3

5
(−5π + larctan 119) +2 3π + larctan .
3

Exercise 12.6 Find an example of a triple of integer angles that are not the angles
of an integer triangle.

Exercise 12.7 Is it true that for an arbitrary pair of integer angles (α, β) there exists
an integer angle γ such that these three angles are angles of some integer triangle?
Chapter 13
Integer Angles of Polygons and Global Relations
for Toric Singularities

In Chap. 6 we proved a necessary and sufficient criterion for a triple of integer an-
gles to be the angles of some integer triangle. In this chapter we prove the analogous
statement for the integer angles of convex polygons. Further, we discuss an applica-
tion of these two statements to the theory of complex projective toric surfaces. We
refer the reader to the general theory of toric surfaces in the works of V.I. Danilov
[43], G. Ewald [55], W. Fulton [59], T. Oda [149], and A. Trevisan [199]. In this
book we do not consider the multidimensional case (we refer the interested reader
to the paper of H. Tsuchihashi [200]).
In Sect. 13.1 we formulate and prove a theorem on integer angles of convex
polygons. After a brief introduction of the main notions and definitions of complex
projective toric surfaces (Sect. 13.2) we discuss two problems related to singular
points of toric varieties using integer geometry techniques.

13.1 Theorem on Angles of Integer Convex Polygons

We start with a theorem on necessary and sufficient conditions for integer angles to
be the angles of some convex integer polygon.

Theorem 13.1 Let (α1 , . . . , αn ) be an arbitrary ordered n-tuple of nonzero integer

angles. Then the following two conditions are equivalent:
(i) there exists a convex n-gon with consecutive integer angles α1 , . . . , αn ;
(ii) there exists a sequence of integers (m1 , . . . , mn−1 ) such that

π − αi +m1 · · · +mn−1 π − αn = 2π.

Proof Condition (i) ⇒ Condition (ii). Suppose that there exists an integer convex
polygon A1 A2 . . . An with prescribed consecutive integer angles α1 , . . . , αn . We as-
sume that the pair of vectors (A2 A3 , A2 A1 ) defines the positive orientation of the

O. Karpenkov, Geometry of Continued Fractions, 173

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_13, © Springer-Verlag Berlin Heidelberg 2013
174 13 Integer Angles of Polygons and Global Relations for Toric Singularities

plane (otherwise, we consider the polygon An An−1 . . . A1 ). Let

B1 = O + An A1 and Bi = O + Ai−1 Ai for i = 2, . . . , n.

Define

∠Bi OBi+1 , if i = 1, . . . , n − 1,
βi =
∠Bn OB1 , if i = n.
Consider the broken line that is the union of all the consecutive sails for angles
β1 , . . . , βn . This broken line defines an extended angle proper integer congruent
to 2π + 0. The LSLS sequence for this broken line contains exactly n − 1 new
elements (together with all the elements of the LLS sequences for the sails of the
angles β1 , . . . , βn ). Denoting these numbers by m1 , . . . , mn−1 , we get

β1 +m1 · · · +mn−1 βn = 2π.

From the definition of βi (i = 1, . . . , n) it follows that βi ∼

ˆ π − αi . Hence
=
π − αi +m1 · · · +mn−1 π − αn = 2π.

Therefore, the first condition implies the second condition.

Condition (ii) ⇒ Condition (i). We assume that there exists M = (m1 , . . . , mn−1 )
such that
π − αi +m1 · · · +mn−1 π − αn = 2π.
This equation implies the existence of integer points B1 = (1, 0), Bi = (xi , yi ), for
i = 2, . . . , n − 1, and Bn = (−1, 0) such that

∠Bi OBi−1 ∼
ˆ π − αi−1 ,
= for i = 2, . . . , n, and ∠B1 OBn ∼
ˆ π − αn .
=
Denote by P the integer point

n
O+ OBi .
i=1

Since αi
= 0, the angle π − αi is not contained in a line (i = 1, . . . , n). Hence,
the origin O is an interior point of the convex hull of the points Bi for i = 1, . . . , k.
Therefore, there exists an integer s such that the integer triangle Bs PBs+1 contains
the origin O and in addition, O is not in the edge Bs Bs+1 (here Bn+1 = B1 and
B0 = Bn ). This implies that

O = λ1 OP + λ2 OBi + λ3 OBi+1 ,

where λ1 > 0, λ2 ≥ 0, and λ3 ≥ 0 are rational numbers. So there exist positive

integers ci (i = 1, . . . , n) such that

n
(ci OBi ) = 0. (13.1)
i=1
13.2 Toric Surfaces and Their Singularities 175

Fig. 13.1 Two integer noncongruent polygons with proper integer congruent integer angles

Consider A0 = O and Ai = Ai−1 + ci OBi for i = 2, . . . , n. Since the numbers

c1 , . . . , ci are integers, the broken line A0 A1 . . . An is integer. From Eq. (13.1) the
broken line is closed (i.e., A0 = An ). By construction, the integer angle at Ai of
this broken line is proper integer congruent to αi (i = 1, . . . , n). Since ci > 0 for
i = 1, . . . , n and since all the vectors OBi are in counterclockwise order, the broken
line is a convex polygon. Therefore, condition (ii) holds. 

Remark 13.2 On the one hand, Theorem 13.1 is an extension of Theorem 6.8(i).
On the other hand, the direct generalization of Theorem 6.8(ii) is false: the integer
angles do not uniquely determine the proper integer affine homothety types of integer
convex polygons. See an example in Fig. 13.1.

A completely satisfactory description of integer congruence classes of lattice

convex polygons has not yet been found. It is known only that the number of convex
polygons with lattice area bounded from above by n grows exponentially in n1/3 , as
n tends to infinity (see [6] and [15]).

13.2 Toric Surfaces and Their Singularities

13.2.1 Definition of Toric Surfaces

We start with the definition of complex projective toric surfaces associated to integer
convex polygons.

Definition 13.3 Consider an integer convex polygon P = A0 A1 . . . An . Let the in-

tersection of this (closed) polygon with the integer lattice Z2 consist of the points
Bi = (xi , yi ) for i = 0, . . . , m. We enumerate the points in such a way that Bi = Ai
for i = 0, . . . , n. Denote by Ω the following set in the complex projective space
CP m :
 x0 y0 −x0 −y0 x1 y1 −x1 −y1 y −x −y  
t1 t2 t3 : t1 t2 t3 : · · · : t1xm t2 m t3 m m |t1 , t2 , t3 ∈ C \ {0} .

The closure of the set Ω in the natural topology of CP m is called the complex
projective toric surface associated with the polygon P and denoted by XP .

Example 13.4 For instance, consider the following two polygons:

P= and Q =
176 13 Integer Angles of Polygons and Global Relations for Toric Singularities

Then XP = CP 2 , and XQ is the conic in CP 3 defined by the equation

x1 x3 = x0 x4 .

Notice that the integer area coincides with the degree of the resulting surface. This
is a general fact of toric geometry.

Remark 13.5 One can generalize the described construction to the case of polyhedra
of arbitrary dimensions. Notice that in the multidimensional case this construction
may have nonisolated singular points. In this situation there is a standard procedure
to resolve them (via certain ς -processes). The resulting variety is the toric variety
corresponding to the polyhedron.

Remark 13.6 One of the main properties of toric surfaces (and toric varieties in
general) is as follows. Every toric surface admits a huge group of isomorphisms.
While here we discuss only toric surfaces, similar statements hold for toric varieties
of arbitrary dimension. Consider a toric surface XP , parameterized as XP (t0 , t1 , t2 ).
Consider the multiplicative group of complex numbers C∗ = C \ {0}. Every toric
surface XP admits a natural action of the group C∗ × C∗ . An arbitrary element
(a, b) of the group C∗ × C∗ acts on the toric surface XP as follows:
 
f(a,b) : XP → XP , f(a,b) XP (t0 , t1 , t2 ) = XP (at0 , bt1 , t2 ).

We use this action below to find singular points of toric surfaces.

13.2.2 Singularities of Toric Surfaces

We begin with the following definition.

Definition 13.7 Algebraic singularities of complex projective toric surfaces are

called toric singularities.

Let us first collect general properties of complex projective toric surfaces.

For i = 0, . . . , m, set

Ãi = (0 : · · · : 0 : 1 : 0 : · · · : 0),

where 1 stands in the ith place.

Theorem 13.8
(i) The set XP is a complex projective complex-two-dimensional surface with iso-
lated algebraic singularities;
(ii) the complex projective toric surface contains the points Ãi for i = 0, . . . , n
(where n + 1 is the number of vertices of the convex polygon);
13.2 Toric Surfaces and Their Singularities 177

(iii) the points of XP \ {Ã0 , Ã1 , . . . , Ãn } are nonsingular;

(iv) the point Ãi is singular if and only if αi  larctan 1, where αi is the angle of
the polygon P at vertex Ai ;
(v) the algebraic singularity at Ãi (0 ≤ i ≤ n) is uniquely determined by the inte-
ger affine type of the nonoriented sail of αi .

Example 13.9 Consider the polygon

P=

and the corresponding toric surface XP . The surface XP ⊂ CP 3 is defined by the

equation
x0 x1 x2 = x33 .
Its singularities are the points

(1 : 0 : 0 : 0), (0 : 1 : 0 : 0), and (0 : 0 : 1 : 0).

In the appropriate affine charts these three singularities are defined by the equation

xy = z3 .

For the proof of Theorem 13.8 we refer to classical textbooks on toric geome-
try ([43, 55, 59, 149], etc.). Nevertheless, we would like to outline some ideas and
remarks related to the proof. The first item is a classical statement of algebraic ge-
ometry.

Outline of the proof of item (ii) Consider an arbitrary vertex Ai of the polygon P .
Let li be some strictly supporting line for the polygon at Ai (i.e., li ∩ P = {Ai }).
Consider a vector v orthogonal to li and directed away from the polygon. Assume
that v = (λ1 , λ2 ). Consider a curve in RP m ⊂ CP m with real parameter t:
 λ x +λ y λ x +λ y  
t 1 0 2 0 : t 1 1 2 1 : · · · : t λ1 xm +λ2 ym |t ∈ R .

It is clear that this curve is a subset of XP . By construction of v, the largest value of

the exponent will be at the place i corresponding to Ai . Thus the limit point of this
curve is exactly the point Ãi as t tends to infinity. Therefore, the point Ãi belongs to
the toric surface XP . For a similar reason, the points Ãn+1 , . . . , Ãm are not in XP .

Outline of the proof of item (iii) To detect all singular points of XP we use the action
of the group C∗ × C∗ introduced in Remark 13.6. Notice that the group takes sin-
gular points to singular points and nonsingular points to nonsingular points. There
are two-dimensional, one-dimensional, and zero-dimensional orbits of the action of
178 13 Integer Angles of Polygons and Global Relations for Toric Singularities

C∗ × C∗ on XP . Since all the singular points of XP are isolated, they are all con-
tained in the zero-dimensional orbits, i.e., at points where C∗ × C∗ acts trivially. It
can be only the points with a single nonzero coordinate

(0 : · · · : 0 : 1 : 0 : · · · : 0).

The toric surface XP contains only singular points with ones at places 0, . . . , n. That
is exactly the points Ã0 , . . . , Ãn .

Remarks on items (iv) and (v) As stated in Theorem 13.8(v), singularities of toric
surfaces are in one-to-one correspondence with integer affine types of nonoriented
sails of integer angles. Therefore, toric singularities are classified by the LLS se-
quences of corresponding integer angles up to switching the order of the sequence.
We call the corresponding LLS sequence the continued fractions of the toric singu-
larity. Therefore, the pair of rational numbers
 
ltan α, ltan α t

for an arbitrary integer angle α is a complete invariant of the corresponding toric

singularities, which we call the sail pair of the singularity.
The continued fractions for toric singularities are slightly different from the
Hirzebruch–Jung continued fractions for toric singularities (see [86] and [81]). The
latter ones have all negative integer elements except for the first one, which is a
positive integer. It is interesting to note that Hirzebruch–Jung continued fractions
generate a sequence of σ -processes that resolve the corresponding singularity. The
relations between regular continued fractions and Hirzebruch–Jung continued frac-
tions are described in [169].

13.3 Relations on Toric Singularities of Surfaces

In this subsection we study global conditions on complex projective toric surface
singularities associated to n-gons (in algebraic language, n is understood as the
Euler characteristic of the corresponding toric surface; see, for instance, [199]).
In Sect. 13.3.1 we fix an arbitrary n and study whether a certain n-tuple of toric
singularities is realizable as a set of singular points of a toric surface defined by
an n-gon. Further, in Sect. 13.3.2, we fix an arbitrary m-tuple of toric singularities
and construct an n-gon (with large enough n) whose singularities are exactly the
singularities of the m-tuple.

13.3.1 Toric Singularities of n-gons with Fixed Parameter n

We begin this subsection with triangles. The corresponding toric surfaces have at
most three toric singularities. The relations on these singularities are described by a
reformulation of Theorem 6.8(i).
13.3 Relations on Toric Singularities of Surfaces 179

Corollary 13.10 Consider three arbitrary complex-two-dimensional toric singular-

ities with sail pairs (ai , bi ), where i = 1, 2, 3. Then the following two statements are
equivalent:
(i) There exists a complex projective toric surface of Euler characteristic equal to 3
whose singular points have sail pairs (ai , bi ) for i = 1, 2, 3.
(ii) There exist a permutation σ ∈ S3 and rational numbers ci ∈ {ai , bi } for i =
1, 2, 3 such that the following conditions hold:
– the continued fraction ]cσ (1) , −1, cσ (2) [ is either negative, greater than cσ (1) ,
or equal to ∞;
– ]cσ (1) , −1, cσ (2) , −1, cσ (3) [ = 0.

Remark Let us say one more time that one should use only odd continued fractions
for c1 , c2 , and c3 in the statement of the above proposition.

When n > 3 we have the following weaker statement. It follows directly from
Theorem 13.1.

Corollary 13.11 Consider n arbitrary complex-two-dimensional toric singularities

with sail pairs (ai , bi ), where i = 1, . . . , n. Then (i) ⇒ (ii), where
(i) There exists a complex projective toric surface of Euler characteristic n whose
singular points have sail pairs (ai , bi ), for i = 1, . . . , n.
(ii) There exist rational numbers ci ∈ {ai , bi } for i = 1, . . . , n and a sequence of
integers (m1 , . . . , mn−1 ) such that the following condition holds:

]cσ (1) , m1 , cσ (2) , m2 , . . . , mn−1 , cσ (n) [ = 0.

The following problem is still open.

Problem 9 Find a necessary and sufficient condition for existence of a complex

projective toric surface of Euler characteristic n in terms of sail pairs of its singular-
ities and a sequence of integers M (as in Corollary 13.11).

Notice that all integers of M are greater than −2. In addition, at least one of the
elements of M is equal to −1.

13.3.2 Realizability of a Prescribed Set of Toric Singularities

Finally, we discuss the realizability of toric surfaces with a prescribed set of toric
singularities.

Proposition 13.12 For every collection (with multiplicities) of complex-two-dimen-

sional toric algebraic singularities there exists a complex projective toric surface
with this collection of singularities.
180 13 Integer Angles of Polygons and Global Relations for Toric Singularities

In the proof of Proposition 13.12 we use the following lemma.

Lemma 13.13 For every collection of integer angles αi (i = 1, . . . , n) there exist

an integer k ≥ n − 1 and a k-tuple of integers M = (m1 , . . . , mk ) such that

α1 +m1 · · · +mn−1 αn +mn larctan 1 +mn+1 · · · +mk larctan 1 = 2π.

Proof Consider any collection of integer angles αi (i = 1, . . . , n) and set

Φ = α1 +1 α2 +1 · · · +1 αn .

It is clear that one of the LSLS sequences for Φ is obtained by adding the LLS
sequences of the angles αi and the number 1 taken n − 1 times. All the elements
of this LSLS sequence are positive integers, and hence Φ is of the form ϕ + 0π for
some integer angle ϕ.
If ϕ ∼
= larctan 1, then we have

Φ +−2 larctan 1 +−2 larctan 1 +−2 larctan 1 = 2π.

Then k = n + 2, and M = (1, . . . , 1, −2, −2, −2).

Suppose now that ϕ  larctan 1. Then the following holds:

ϕ +−1 π − ϕ +−2 larctan 1 +−2 larctan 1 = 2π.

Consider the sail for the angle π −ϕ. Suppose that the sequence of all its consecutive
integer points (not only vertices) is (B0 , . . . , Bs ), where the order coincides with the
order of the sail. Then we have

∠Bi OBi+1 ∼
ˆ larctan 1
= for every i = 1, . . . , s.

Put bi = lsin ∠Bi OBi+1 for i = 1, . . . , s. We get

ϕ +−2 larctan 1 +−2 larctan 1 +−2 larctan 1

= α1 +1 α2 +1 · · · +1 αn
+−1 larctan 1 +b1 larctan 1 +b2 · · · +bs larctan 1
+−2 larctan 1 +−2 larctan 1 +−2 larctan 1
= 2π.

Therefore, k = n + s + 3, and

M = (1, 1, . . . , 1, 1, −1, b1 , . . . , bs , −2, −2, −2).

!" #
(n−1) times

The proof of Lemma 13.13 is complete. 

13.3 Relations on Toric Singularities of Surfaces 181

Proof of Proposition 13.12 Consider an arbitrary collection of surface toric alge-

braic singularities. Suppose that they are enumerated by integer angles αi (i =
1, . . . , n). By Lemma 13.13 there exist k ≥ n − 1 and M = (m1 , . . . , mk ) such that

(π − α1 ) +m1 · · · +mn−1 (π − αn ) +mn larctan 1 +mn+1 · · · +mk larctan 1 = 2π.

Then by Theorem 13.1 there exists a convex polygon P = A0 . . . Ak with integer

angles proper integer congruent to the angles αi (i = 1, . . . , n), and the last k − n + 1
angles equal larctan 1.
By Theorem 13.8(c) all singularities of the toric surface XP are contained in the
set {Ã0 , Ã1 , . . . , Ãk }. By Theorem 13.8(d) the points Ãi corresponding to larctan 1
are also nonsingular. The collection of toric singularities at the remaining points co-
incides with the given collection. This concludes the proof of Proposition 13.12. 

Example 13.14 Let us construct a projective toric surface having a unique toric sin-
gularity with the sail pair (7/5, 7/3). Consider α ∼
=ˆ larctan(7/5). First of all, we
draw the angle π − α and its adjacent angle π − (π − α) (which is α). Further,
we subdivide the half-plane in the complement to the union of α and π − α in a
standard way into two angles β and γ (such that the resulting LSLS sequence ends
in (−2, 1, −2, 1)).

Secondly, we subdivide the angle α = π − (π − α) into angles integer congruent

to larctan 1, shown here:

Finally, we construct the hexagon P all of whose angles are adjacent to the angles
of the obtained decomposition.

The toric surface XP has a unique singular point; its sail pair is (7/5, 7/3).
182 13 Integer Angles of Polygons and Global Relations for Toric Singularities

13.4 Exercises

Exercise 13.1 Consider an arbitrary toric surface XP . Characterize all points of

XP where the action of C∗ × C∗ on XP is (a) constant; (b) one-dimensional; (c)
two-dimensional.

Exercise 13.2 Write the toric surfaces for the following polygons

P= , Q= , and R = .

Enumerate all their singularities and find the sail pairs of the corresponding singu-
larities.

Exercise 13.3 Find a toric surface with a unique toric singularity whose sail pair is
(5/3, 5/2).

Exercise 13.4 Prove that the sequence M in Corollary 13.11 has at least one nega-
tive integer. Show that all negative numbers are equal to −1.

Exercise 13.5 Find an example of an integer quadrangle whose sequence M has

exactly one negative integer.
 Part II
Multidimensional Continued Fractions

In this second part we study multidimensional versions of several classical theo-

rems on continued fractions. For this purpose we use multidimensional continued
fractions in sense of Klein–Voronoi. The material of this part is relatively recent, so
we accompany it with related open problems and conjectures. In addition, we pro-
vide links to classical problems and conjectures, such as the Littlewood conjecture,
the lonely runner problem, White’s problem, etc. In conclusion, we briefly discuss
other generalizations of continued fractions.
As in the planar case, we begin with a discussion of basic features of multidi-
mensional integer geometry. In Chap. 14 we introduce the main integer objects and
present the corresponding invariants. Further, in Chap. 15 we study a more specific
question related to emptiness of integer simplices, pyramids, and tetrahedra. In par-
ticular, we formulate and proof White’s theorem on three-dimensional empty integer
triangles.
In Chap. 16 we introduce multidimensional continued fractions in the sense
of Klein, which are special polyhedral surfaces related to simplicial cones in real
spaces. We study both finite and infinite Klein’s continued fractions. We prove that
for a generic cone, the corresponding continued fraction has a good polyhedral
structure (using the multidimensional Kronecker’s approximation theorem). Finally,
we study the structure of two-dimensional compact faces of continued fractions in
the sense of Klein.
In the next two chapters, we investigate the periodic structure of algebraic mul-
tidimensional continued fractions. The structure of such continued fractions is pe-
riodic due to the action of the Dirichlet groups. We study some classical aspects
concerning Dirichlet groups in Chap. 17. In particular, we describe several techni-
cal details related to the calculation of the bases of those groups, and the questions of
lattice reduction (e.g., the simplest LLL-algorithm). Further, in Chap. 18 we discuss
periodic multidimensional continued fractions. Finally, we prove a generalized ver-
sion of Lagrange’s theorem and explain the relation of multidimensional continued
fractions to the Oppenheim conjecture.
We discuss the multidimensional Gauss–Kuzmin statistics in Chap. 19. Firstly,
we extend the Möbius measure of the projective plane to higher dimensions. Sec-
184

ondly, we formulate general results regarding the generalized Gauss–Kuzmin statis-

tics on relative frequencies of polygonal faces in continued fractions in the sense
of Klein. Finally, we give examples of explicit calculation for certain integer types
of faces. We conclude with several open problems on the distribution of faces in
multidimensional continued fractions.
In Chap. 20 we bring together several techniques to construct multidimensional
continued fractions. There are two different algorithmic approaches to this issue.
The first approach deals with a face-by-face construction of the Klein’s polyhedral
continued fraction. In the second approach one makes an approximation of a sail
and conjectures which faces of the approximation are the faces of the continued
fraction. We discuss the corresponding algorithm and study one particular example.
Chapter 21 is dedicated to multidimensional Gauss’s reduction theory. We show
that the Hessenberg matrices play the role of reduced matrices. Further we discuss
a complete geometric invariant of conjugacy classes of integer invertible matrices
in the simplest case that all eigenvalues are distinct. If a matrix has a real spectrum,
then the corresponding invariant is its Klein’s continued fraction. For the remaining
cases we use multidimensional continued fractions in the sense of Klein–Voronoi as
invariants, which naturally extends Klein’s continued fractions to the case of matri-
ces with arbitrary spectra. After that we present the technique of multidimensional
Gauss reduction. Our next goal is to show how to solve Diophantine equations re-
lated to Markov–Davenport characteristics, the main ingredient here is algebraic
periodicity induced by the Dirichlet group. We conclude this chapter with an ex-
haustive study of the new case of the family of three-dimensional operators having
two complex conjugate eigenvalues. This part of material has never been published
before.
We investigate questions concerning best approximation of multidimensional
simplicial cones in Chap. 22. In particular we discuss the relation to the maximal
commutative subgroup approximation. Further we show that simultaneous approx-
imation can be also considered in the context of maximal commutative subgroup
approximation.
Finally, in Chap. 23 we expose a collection of different generalizations of mul-
tidimensional continued fractions mentioning interesting theorems related to them.
We do not pretend that this collection is complete. The aim of this chapter is to show
the diversity of methods that are used to study the extensions of classical problems
of continued fraction theory.
Chapter 14
Basic Notions and Definitions
of Multidimensional Integer Geometry

In this chapter we generalize integer two-dimensional notions and definitions of

Chap. 2. As in the planar case, our approach is based on the study of integer invari-
ants. Further, we use them to study the properties of multidimensional continued
fractions. First, we introduce integer volumes of polytopes, distances, and angles.
Then we express volumes of polytopes, integer distances, and integer angles in terms
of integer volumes of simplices. Finally, we show how to compute integer volumes
of simplices via certain Plücker coordinates in the Grassmann algebra (this formula
is extremely useful for the computation of multidimensional integer invariants of
integer objects contained in integer planes). We conclude this chapter with a dis-
cussion of the Ehrhart polynomials, which one can consider a multidimensional
generalization of Pick’s formula in the plane.

14.1 Basic Integer Invariants in Integer Geometry

Let us expand integer geometry to the multidimensional case. The main object of
integer geometry is the lattice Zn in Rn , i.e., the lattice of vectors all of whose
coordinates are integers.

14.1.1 Objects and the Congruence Relation

We say that a vector (point) is integer if all its coordinates are integers. A plane in
Rn is called integer if its integer vectors form a sublattice whose rank equals the
dimension of the plane. We say that a polyhedron is integer if all its vertices are
integer. A cone in Rn is said to be integer if its vertex is integer and each of its
edges contains integer points distinct from the vertex.
The integer congruence relation is defined as in the two-dimensional case by the
group of affine transformations preserving the set of integer points, i.e., Aff(n, Z).
This group is a semidirect product of GL(n, Z) and the group of translations on
integer vectors. We use “∼=” to indicate that two objects are integer congruent.

O. Karpenkov, Geometry of Continued Fractions, 185

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_14, © Springer-Verlag Berlin Heidelberg 2013
186 14 Basic Notions and Definitions of Multidimensional Integer Geometry

14.1.2 Integer Invariants and Indices of Sublattices

As in the planar case described in Chap. 2, the integer invariants in the multidimen-
sional case are easily defined in terms of indices of certain sublattices. We first give
a geometric description of the index of sublattices in terms of the number of integer
points in certain parallelepipeds.

Proposition 14.1 The index of a sublattice generated by integer vectors v1 , . . . , vk

in an integer k-dimensional plane equals the number of all integer points P satisfy-
ing

n
AP = λi v i with 0 ≤ λi < 1, i ∈ {1, . . . , n},
i=1

where A is an arbitrary integer point.

Proof The proof of this statement is similar to the proof of the planar one. Let H be
a subgroup of Z2 generated by v1 , . . . , vk . Define
 n $
 

Par(v1 , . . . , vk ) = λi vi  0 ≤ λi < 1, i = 1, . . . , n .
i=1

First, we prove that for every integer vector g there exists an integer point P
lying in the set Par(v1 , . . . , vk ) such that AP ∈ gH . Let


n
g= μi v i .
i=1

Then consider

n
 
P =A+ μi − μi  vi .
i=1

Since 0 ≤ μi − μi  < 1 for i = 1, . . . , n, the point P is inside the parallelogram

and AP ∈ gH .
Second, we show the uniqueness of such a point P . Suppose that for two integer
points P1 , P2 ∈ Par(v1 , . . . , vk ) we have AP1 ∈ gH and AP2 ∈ gH . Hence the vector
P1 P2 is an element in H . The only element of H of type

αv + βw with 0 ≤ α, β < 1

is the zero vector. Hence P1 = P2 .

Therefore, the integer points of Par(v1 , . . . , vk ) are in one-to-one correspondence
with the cosets of H in Z2 . 
14.1 Basic Integer Invariants in Integer Geometry 187

14.1.3 Integer Volume of Simplices

Let us start with the notion of integer volume for integer simplices. We will define
the integer volumes for polyhedra in the next subsection.

Definition 14.2 An integer volume of an integer simplex A1 . . . An is the index of

the sublattice generated by the vectors A2 A1 , . . . , An A1 in the integer lattice of the
space Span(A2 A1 , . . . , An A1 ). Denote it by lV(A1 . . . An ).
For the points A1 . . . An that are contained in some (n − 2)-dimensional plane we
say that lV(A1 . . . An ) = 0.

Example 14.3 Consider a three-dimensional simplex S ⊂ R4 with vertices

s1 = (2, 3, 0, 1), s2 = (1, 4, 2, 4), s3 = (1, 0, 0, 4), s4 = (1, 0, 0, 1).

The integer volume of S is 6.

We postpone the calculation of the volume of simplices for a while (see Exam-
ple 14.32 below).

14.1.4 Integer Angle Between Two Planes

First, we give the definition of an integer angle for two linear subspaces.

Definition 14.4 Consider two integer linear spaces L1 and L2 that are not con-
tained one in another. The integer angle between L1 and L2 is the index of the
sublattice generated by all integer vectors of L1 ∪ L2 in the integer lattice of the
space Span(L1 , L2 ). Denote it by lα(L1 , L2 ).
In case one space is a subspace of the other we agree to say that the integer angle
between them is zero.

In order to calculate integer angles via integer volumes one can use the following
proposition.

Proposition 14.5 Consider two integer linear spaces L1 and L2 that are not con-
tained one in another. Let the sets of independent integer vectors

{u1 , . . . , uk , w1 , . . . , wm }, {v1 , . . . , vl , w1 , . . . , wm }, and {w1 , . . . , wm }

form bases in L1 , L2 , and L1 ∩ L2 respectively. Then we have

lV(u1 , . . . , uk , v1 , . . . , vl , w1 , . . . , wm ) lV(w1 , . . . , wm )
lα(L1 , L2 ) = .
lV(u1 , . . . , uk , w1 , . . . , wm ) lV(v1 , . . . , vl , w1 , . . . , wm )
188 14 Basic Notions and Definitions of Multidimensional Integer Geometry

Proof First, let us change the basis (wi ) of L1 ∩ L2 to the basis (w i ) that generates
the integer sublattice in L1 ∩ L2 . The value of the formula stays unchanged, since
the numerator and the denominator are both divided by lV2 (w1 , . . . , wm ).
Second, we extend the basis (wi ) with integer vectors ui to the basis of (ui , wi )
that generates the integer lattice of L1 . The value of the formula stays unchanged,
since the numerator and the denominator are both divided by lV(u1 , . . . , uk , w 1 ,
. . . , w m ).
Third, we extend the basis (w i ) with integer vectors v i to the basis of (v i , wi )
that generates the integer lattice of L2 . For similar reasons the value of the formula
stays unchanged.
From the definition we have
⎧
⎪
⎪lV(u1 , . . . , uk , v 1 , . . . , v k , w1 , . . . , wm ) = lα(L1 , L2 ),
⎪
⎨lV(u , . . . , u , w , . . . , w ) = 1,
1 k 1 m
⎪
⎪lV(v , . . . , v , w , . . . , w ) = 1,
⎪
⎩
1 k 1 m
lV(w 1 , . . . , wm ) = 1.

For these vectors the statement of the proposition is clearly holds. This concludes
the proof. 

Finally, we give a definition of the integer angle between arbitrary (affine) planes.

Definition 14.6 Consider two integer planes π1 and π2 . Let L1 and L2 be the spaces
of vectors corresponding to π1 and π2 . The integer angle between π1 and π2 is the
index lα(L1 , L2 ). We denote it by lα(π1 , π2 ).

14.1.5 Integer Distance Between Two Disjoint Planes

Definition 14.7 Consider two disjoint integer planes π1 and π2 . Let L1 and L2 be
the spaces of vectors corresponding to π1 and π2 , and let a = A1 A2 be a vector such
that A1 ∈ π1 and A2 ∈ π2 . The integer distance between π1 and π2 is the index of
the sublattice generated by all integer vectors of L1 ∪ L2 ∪ Span(a) in the integer
lattice of the space Span(L1 , L2 , a). Denote it by ld(π1 , π2 ).
In the case of an intersecting subspace of another we agree to say that the integer
distance between them is zero.

Remark 14.8 One of the simplest useful invariants in three-dimensional integer ge-
ometry is the distance between an integer point p and an integer plane π , i.e., the
index of the lattice generated by the integer vectors joining p and all integer points
of π in the whole integer lattice of Z3 . Integer distance has a nice geometric de-
scription. Draw all integer planes parallel to π . One of them contains the point p
(let us call it πp ). The integer distance ld(p, π) is the number of integer planes in
14.2 Integer and Euclidean Volumes of Basis Simplices 189

the region bounded by the lines πp and π plus one. Let us illustrate by the following
example:

There are three integer planes parallel to π and lying between πp and π . Hence,
ld(p, π) = 3 + 1 = 4.

Let us present a formula relating integer distances and integer volumes.

Proposition 14.9 Consider two disjoint integer planes π1 and π2 . Let L1 and L2
be the spaces of vectors corresponding to π1 and π2 , and let a be a vector with
one integer endpoint in π1 and one integer endpoint in π2 . Suppose that the sets of
independent integer vectors

{u1 , . . . , uk , w1 , . . . , wm }, {v1 , . . . , vl , w1 , . . . , wm }, and {w1 , . . . , wm }

form bases in L1 , L2 , and L1 ∩ L2 respectively. Then we have

lV(a, u1 , . . . , uk , v1 , . . . , vl , w1 , . . . , wm ) lV(w1 , . . . , wm )
ld(π1 , π2 ) = .
lV(u1 , . . . , uk , w1 , . . . , wm ) lV(v1 , . . . , vl , w1 , . . . , wm )

The proof of Proposition 14.9 is similar to the proof of Proposition 14.5, so we

skip it here.
In this section we have defined several invariants of planes and pairs of planes.
In general, many other nice invariants for different arrangements of integer planes
(for instance, for triples of integer lines whose span is 5-dimensional) are defined in
a similar way.

14.2 Integer and Euclidean Volumes of Basis Simplices

Here we study simplices of maximal dimension whose edges generate an integer
lattice. We prove Proposition 14.11, which is analogues to Proposition 2.11.

Definition 14.10 An integer polyhedron is called empty if it does not contain any
integer points other than its vertices.
190 14 Basic Notions and Definitions of Multidimensional Integer Geometry

Denote by V (A0 A1 . . . An ) the Euclidean volume of a simplex A0 A1 . . . An . Re-

call that
1 
V (A0 A1 . . . An ) = det(A0 A1 , A0 A2 , . . . , A0 An ).
n!
We also denote by P (A0 ; A1 , . . . , An ) the parallelepiped
 $
n 

A0 + λi Ai A0  0 ≤ λi ≤ 1 .
i=1

Proposition 14.11 Consider an integer simplex A0 A1 . . . An in Rn . Then the fol-

lowing statements are equivalent:
(a) P (A0 ; A1 , . . . , An ) is empty;
(b) lV(A0 A1 . . . An ) = 1;
(c) V (A0 A1 . . . An ) = n! 1
.

Remark 14.12 In Chap. 2 we showed that every empty triangle has integer area 1
(and Euclidean area 1/2). This is no longer true in the multidimensional case; for
instance, see Example 14.33. We describe the situation in the multidimensional case
in the next chapter.

Proof of Proposition 14.11 (a) ⇒ (b). Let a parallelepiped P (A0 ; A1 , . . . , An ) be

empty. Therefore, by Proposition 14.1 there is only one coset for the subgroup gen-
erated by vectors A1 A0 , . . . , An A0 . Hence, the vectors A1 A0 , . . . , An A0 generate
the integer lattice, and lV(A0 A1 . . . An ) = 1.
(b) ⇒ (c). Let lV(A0 A1 . . . An ) = 1. Hence the vectors A1 A0 , . . . , An A0 gener-
ate the integer lattice. Thus every integer point is an integer combination of them,
so for the j th coordinate vector ej we have


n
ej = λij Ai A0 ,
j =1

with integers λij for 1 ≤ i, j ≤ n. Denote by M1 the matrix where at the j th place
of the ith column is λij . Let also M2 be the matrix whose ith column is filled with
the coordinates of the vector A0 Ai for i = 1, . . . , n. Then

M1 · M2 = Id,

where Id is the identity matrix. Since these matrices are integer, both their determi-
nants equal either 1 or −1. Hence the Euclidean volume of the simplex A0 A1 . . . An
1
coincides with the Euclidean volume of the coordinate simplex and equals n! .
(c) ⇒ (a). Consider an integer simplex A0 A1 . . . An of Euclidean volume n! . Sup-
1

pose that the corresponding parallelepiped P (A0 ; A1 , . . . , An ) has an integer point

distinct from its vertices. Then there exists an integer simplex such that one of its
14.3 Integer Volumes of Polyhedra 191

facets (of codimension 1) coincides with a facet of A0 A1 . . . An , and the height of

this facet is smaller than the corresponding height in A0 A1 . . . An . Hence there exists
1
an integer simplex whose Euclidean volume is smaller than n! , which is impossible
(since the determinant of n integer vectors is an integer). 

Proposition 14.13 For an arbitrary simplex S of full dimension we have the follow-
ing formula:
 
lV(S) = V P (S) .

Proof Let us prove this statement by induction on the integer volume of simplices.
Base of induction. The statement for simplices of integer volume 1 follows di-
rectly from Proposition 14.11.
Step of induction. Let the statement hold for all simplices of integer volume less
then N (N > 1). Let us prove it for an arbitrary simplex S = v0 v1 . . . vn of lattice
volume N .
Consider a polyhedron P (S). By Proposition 14.11, since lV(S) > 1, there is
an integer point in P (S) distinct from the vertices. Without loss of generality we
suppose that this point is not on any facet parallel to v1 , . . . , vn . Hence

d(v0 , v1 . . . vn ) > 1.

Denote this value by d0 .

Consider an integer point v 0 at unit integer distance to the plane of the facet
v1 . . . vn . By Proposition 14.9 we get
lV(v0 v1 . . . vn )
= d0 .
lV(v 0 v1 . . . vn )
It is clear that for Euclidean volumes we have the same relation:
V (P (v0 v1 . . . vn ))
= d0 .
V (P (v 0 v1 . . . vn ))
Since d0 > 0, we have lV(v 0 v1 . . . vn ) < N , and hence we are in a position to apply
the induction assumption:
   
lV(v0 v1 . . . vn ) = d0 lV(v 0 v1 . . . vn ) = d0 V P (v 0 v1 . . . vn ) = V P (v0 v1 . . . vn ) .

This concludes the proof. 

14.3 Integer Volumes of Polyhedra

In this subsection we write down an expression for the integer volume of simplices
via certain Euclidean volumes. Further, we use the additivity of Euclidean volume
to get a natural extension of lattice volume to the case of integer polyhedra. Finally,
we discuss how to decompose an arbitrary convex integer polyhedron into integer
empty simplices.
192 14 Basic Notions and Definitions of Multidimensional Integer Geometry

14.3.1 Interpretation of Integer Volumes of Simplices via

Euclidean Volumes

We begin with a definition of determinants of sublattices.

Definition 14.14 Consider a sublattice L of the integer lattice Zn . The Euclidean

volume of the simplex whose edges generate L is called the determinant of L and
denoted by det L.

Let us say a few words about the determinant being well defined. Recall that L
can be mapped by a linear (not necessarily integer) mapping to Zk ⊂ Rk . In Propo-
sition 14.11 we showed that all simplices whose edges generate Zk have the same
Euclidean volume. Hence all simplices whose edges generate L also have the same
Euclidean volume.
In the following theorem we establish a relation between the integer volume of a
simplex S and the Euclidean volume of the corresponding parallelepiped P (S).

Theorem 14.15 Let S be a k-dimensional integer simplex in Rn and let L(S) denote
the integer lattice of the k-dimensional integer plane containing S. Then we have
V (P (S))
lV(S) = .
det(L(S))

Proof Consider a linear map T sending L to Zk ⊂ Rk . We have

     V (P (S))   
lV T (S) = lV(S), V T P (S) = , and det T L(S) = 1.
det(L(S))
Hence the statement of Theorem 14.15 follows directly from a similar statement for
simplices of full dimension:
    
lV T (S) = V T P (S) ,

which appears in Proposition 14.13. 

14.3.2 Integer Volume of Polyhedra

Using the result of Theorem 14.15 we extend the definition of integer area to general
polyhedra.

Definition 14.16 Consider an integer polyhedron P of dimension n. The integer

volume of P equals
V (P )
n! ,
det(L(P ))
which we denote by lV(P ).
14.3 Integer Volumes of Polyhedra 193

As in the two-dimensional case, we have the additivity property for lattice area.

Proposition 14.17 The integer volume of polyhedra is additive, i.e., if an integer

polyhedron P is a disjoint union of integer polyhedra P1 , . . . , Pk then


k
lV(P ) = lV(Pk ).
i=1

Proof By definition the integer volumes of polyhedra of dimension n are propor-

tional to their Euclidean volumes. Therefore, the additivity of Euclidean volume
implies additivity of integer volumes. 

14.3.3 Decomposition into Empty Simplices

Let us formulate two theorems on the existence of decompositions into simplices.

The first holds in arbitrary dimensions, while the second is specific to the three-
dimensional case.

Theorem 14.18 For every convex integer polyhedron P (not contained in a hyper-
plane) there exists a decomposition into integer empty simplices.

Proof Let us give a sketch of the proof. The proof is based on induction on the
number of integer points in P ∈ Rn .
Base of induction. If P has only n + 1 integer points inside, then it is an empty
simplex.
Step of induction. Suppose now that the statement holds for every k < m. Let us
prove it for m. Consider an integer polyhedron P . Suppose that it contain points
p1 , . . . , pm . Let f1 , . . . , fs be all the (n − 2)-dimensional faces of P . We fix one
integer point of P , say some pi , and consider all pyramids with vertex at this point
and bases in faces f1 , . . . , fs . These pyramids subdivide P into integer polyhedra.
If P contains at least n + 2 points, then it is possible to choose pi such that this
subdivision contains at least two smaller polyhedra (each of them contains a smaller
number of integer points). By induction all the smaller polyhedra are decomposable;
hence P is decomposable itself. 

The decomposition of Theorem 14.18 usually contains triangles of distinct inte-

ger types that are not necessarily integer congruent to the coordinate simplex. In the
paper [88] of J.-M. Kantor and K.S. Sarkaria, a stronger result for three-dimensional
polyhedra is given.

Theorem 14.19 For every convex integer polyhedron P in R3 there exists a decom-
position of 4P into integer tetrahedra congruent to the basis tetrahedron.
194 14 Basic Notions and Definitions of Multidimensional Integer Geometry

(Recall that nP is the polyhedron whose vertices have all the coordinates n-times
greater than the coordinates of P .)

Idea of the proof By Theorem 14.18 it is enough to find a proof for the list of empty
tetrahedra, which is known due to White’s theorem (see Corollary 15.3 below). It
turns out that for every empty tetrahedron T ∈ R3 , the tetrahedron 4T admits a
decomposition into integer tetrahedra congruent to the basis tetrahedron. We skip
here the technical details related to explicit constructions of these decompositions. 

14.4 Lattice Plücker Coordinates and Calculation of Integer

Volumes of Simplices
In order to calculate integer volumes of integer simplices we embed the space of all
k-dimensional sublattices into the Grassmann algebra. Further, we express the inte-
ger volume of a simplex in terms of certain Plücker coordinates of this embedding.

14.4.1 Grassmann Algebra on Rn and k-Forms

The Grassmann algebra (or exterior algebra) of the vector space Rn is the alge-
bra over R with multiplication (usually denoted by ∧), generated by elements 1,
e1 , . . . , en , where e1 , . . . , en is the basis of Rn . The relations defining the algebra
are
ei ∧ ej = −ej ∧ ei ;
e i ∧ 1 = 1 ∧ ei = e i ;
1 ∧ 1 = 1.

This algebra is denoted by Λ(Rn ), with the operation ∧ called the wedge product.
Let r be a positive integer. Denote by Λk (Rn ) the vector space generated by the
elements ei1 ∧ · · · ∧ eik , where i1 , . . . , ik ∈ {1, . . . , n}. In addition we put Λ0 Rn = R.
The space Λk (Rn ) is called the kth exterior power of R and its elements are called
k-forms on R.

Proposition 14.20
(i) The dimension of the kth exterior power of Rn is as follows:
 
 n n
, if k ≤ n,
dim Λ R = k
k
0, if k > n.

(ii) Any element of the exterior algebra can be written as a sum of k-forms, so we
have
       
Λ Rn = Λ 0 R n ⊕ Λ 1 R n ⊕ · · · ⊕ Λ n R n .
14.4 Lattice Plücker Coordinates and Calculation of Integer Volumes 195

Let us define a natural basis of the space Λ(Rn ) for an arbitrary k ≤ n. Let I and
J be two strictly increasing sequences of k elements of the set {1, . . . , n}. We say
that I > J if there exists s ≤ k such that it = jt for t < s and is > js . We call this
ordering lexicographic.

Definition 14.21 For a strictly increasing sequence I of k elements of the set

{1, . . . , n} we define
ω I = ei 1 ∧ · · · ∧ ei k .
The basis of Λk (Rn ) consisting of all k-forms of type ωI whose indices are ordered
lexicographically is called the lexicographic basis of Λ(Rn ).

Example 14.22 The dimension of the space Λ2 (R4 ) is 6. The lexicographic basis
of Λ2 (R4 ) is as follows:

(e1 ∧ e2 , e1 ∧ e3 , e1 ∧ e4 , e2 ∧ e3 , e2 ∧ e4 , e3 ∧ e4 ).

14.4.2 Plücker Coordinates

The set of all k-dimensional subspaces of Rn is called the linear Grassmannian, and
denoted by Gr(k, Rn ). Denote by P (Λk (Rn )) the projectivization of the rth exte-
rior power Λk (Rn ), i.e., we consider all forms up to a multiplication by a nonzero
constant.

Definition 14.23 The Plücker embedding of a Grassmannian into projective space

is the map
    
φ : Gr k, Rn → P Λk Rn ,
 
Span(v1 , . . . , vk ) → P {v1 ∧ . . . ∧ vk } .

It is clear that the form v1 ∧ · · · ∧ vk up to a nonzero multiplicative factor does

not depend on the choice of basis in Span(v1 , . . . , vk ).
Let e1 , . . . , en be a basis of Rn and let ω1 , . . . , ωN be the corresponding lexico-
graphic basis in Λk (Rn ).

Definition 14.24 Consider a k-dimensional linear subspace L in Rn with basis

v1 , . . . , vk . Let

N
v1 ∧ · · · ∧ vk = λi ω i .
i=1

The Plücker coordinates of L are the projective coordinates (λ1 : · · · : λN ).

196 14 Basic Notions and Definitions of Multidimensional Integer Geometry

To compute the Plücker coordinates one should calculate the determinants of

all (k×k) minors of the (k × n) matrix whose elements are coordinates of vectors
v1 , . . . , vk .

Example 14.25 Consider a two-dimensional linear subspace of R3 generated by

vectors v1 = (1, 1, 1) and v2 = (1, 2, 5). We have

v1 ∧ v2 = (e1 + e2 + e3 ) ∧ (e1 + 2e2 + 5e3 )

1 1 1 1 1 1
= det e ∧ e2 + det e ∧ e3 + det e ∧ e3 .
1 2 1 1 5 1 2 5 2

Hence the Plücker coordinates of this linear subspace are

1 1 1 1 1 1
det : det : det = (1 : 4 : 3).
1 2 1 5 2 5

14.4.3 Oriented Lattices in Rn and Their Lattice Plücker

Embedding

In this subsection we define a modification of the Plücker embedding that is useful

in the study of lattices. We say that a lattice is a k-lattice if it has k generators and
they span a k-dimensional subspace in Rn . A k-lattice is said to be oriented if we fix
an orientation of the span. Denote the space of all oriented k-lattices by Lat(k, Rn ),
which we call the oriented lattice Grassmannian.

Definition 14.26 The lattice Plücker embedding of an oriented lattice Grassmann-

ian in the space Λk (Rn ) is the map
   
φ : Lat k, Rn → Λk Rn ,
v1 , . . . , vk  → v1 ∧ · · · ∧ vk ,

where v1 , . . . , vk is a basis of the lattice v1 , . . . , vk  defining the positive orienta-

tion.

Remark 14.27 It is clear that the lattice Plücker embedding is well defined (since
different generators of the lattice define the same k-form up to sign, which is also
preserved for generators of positive orientation).

We define lattice Plücker coordinates in a similar way.

Definition 14.28 Consider a k-lattice L in Rn with generators v1 , . . . , vk defining

the basis of positive orientation. Let
14.4 Lattice Plücker Coordinates and Calculation of Integer Volumes 197


N
v1 ∧ · · · ∧ vk = λi ω i .
i=1

The lattice Plücker coordinates of L are the coordinates (λ1 , . . . , λN ).

Remark Notice that the lattice Plücker coordinates of lattices are affine and the
arbitrary Plücker coordinates of subspaces are projective.

Example 14.29 As in Example 14.25, we consider a two-dimensional linear sub-

space of R3 generated by vectors v1 = (1, 1, 1) and v2 = (1, 2, 5). Its lattice Plücker
coordinates are (1, 4, 3).

14.4.4 Lattice Plücker Coordinates and Integer Volumes of

Simplices

If L is a sublattice of the integer lattice Zn , then all its lattice Plücker coordinates
are integers. Let us now show how to write the integer volume of integer simplices
in terms of lattice Plücker coordinates.

Theorem 14.30 Let s1 s2 . . . sk+1 be a simplex in Rn and let (p1 , . . . , pN ) be the lat-
tice Plücker coordinates for the lattice generated by vectors si sk+1 for i = 1, . . . , k.
Then
lV(s1 s2 . . . sk+1 ) = gcd(p1 , . . . , pN ).

Denote by vi the vector si sk+1 for i = 1, . . . , k. We begin with the following

lemma.

Lemma 14.31 The number gcd(p1 , . . . , pN ) is a GL(n, Z)-invariant.

Proof Let

N
v1 ∧ · · · ∧ vk = pi ω i
i=1
(as usual, (ω1 , . . . , ωN ) is the lexicographic basis associated to a chosen basis
(e1 , . . . , en )). Consider an arbitrary A ∈ GL(n, Z). We have
 

N 
N
A(v1 ∧ · · · ∧ vk ) = A pi ω i = pi A(ωi ).
i=1 i=1

Since all k-forms A(ωi ) have integer coefficients and all the numbers pi are divis-
ible by gcd(p1 , . . . , pN ), the greatest common divisor for the lattice Plücker coor-
198 14 Basic Notions and Definitions of Multidimensional Integer Geometry

dinates for the form A(v1 ∧ · · · ∧ vk ) (denote them by (q1 , . . . , qN )) is divisible by

gcd(p1 , . . . , pN ) as well.
For the same reason, applying the inverse operator A−1 to the k-form A(ωi ), we
have that gcd(q1 , . . . , qN ) is divisible by gcd(p1 , . . . , pN ). Hence

gcd(q1 , . . . , qN ) = gcd(p1 , . . . , pN ).

Therefore, gcd(p1 , . . . , pN ) is a GL(n, Z)-invariant. 

Proof of Theorem 14.30 The statement holds for all lattices in the plane spanned by
e1 , . . . , ek . In this case the first lattice Plücker coordinate is the Euclidean volume
of the parallelogram P (s1 , s2 , . . . , sk+1 ), which coincides with the integer volume
of the simplex s1 s2 . . . sk+1 (see Corollary 14.15). All the other lattice Plücker coor-
dinates are zero.
An arbitrary k-dimensional plane can be taken to the coordinate k-plane by some
GL(n, Z) transformation. By Lemma 14.31, the gcd(p1 , . . . , pN ) is a GL(n, Z) in-
variant. So in this case it is also equal to the integer volume of the simplex. 

Example 14.32 Let us calculate the volume of the three-dimensional tetrahedron

S = s1 s2 s3 s4 of Example 14.3. Recall that

s1 = (2, 3, 0, 1), s2 = (1, 4, 2, 4), s3 = (1, 0, 0, 4), s4 = (1, 0, 0, 1).

This tetrahedron defines a three-dimensional sublattice in Z4 generated by vectors

v1 = s1 s4 , v2 = s2 s4 , and v3 = s3 s4 . So, we get

v1 ∧ v2 ∧ v3 = (e1 + 3e2 ) ∧ (4e2 + 2e3 + 3e4 ) ∧ 3e4

= 0 · e1 ∧ e2 ∧ e3 + 12 · e1 ∧ e2 ∧ e4
+ 6 · e1 ∧ e3 ∧ e4 + 18 · e2 ∧ e3 ∧ e4 .

The lattice Plücker coordinates are (0, 12, 6, 18), and their greater common divisor
is 6. Hence the integer volume of S equals 6.

14.5 Ehrhart Polynomials as Generalized Pick’s Formula

In this section we discuss one of the generalizations of Pick’s formula.We start with
the following example, showing that there is no straightforward combinatoric gen-
eralization of Pick’s formula.

Example 14.33 Consider the following two tetrahedra (see in Fig. 14.1):
 
T1 = conv (0, 0, 0), (1, 0, 0), (0, 1, 0), (0, 0, 1) ;
 
T2 = conv (0, 0, 0), (1, 0, 0), (0, 1, 0), (1, 1, 2) .
14.5 Ehrhart Polynomials as Generalized Pick’s Formula 199

Fig. 14.1 The empty

tetrahedra T1 and T2

These tetrahedra are both empty, but they have distinct areas:

V (T1 ) = 1/6 and V (T2 ) = 1/3.

The emptiness of the tetrahedra means that they have only four integer points as ver-
tices, and hence they have the same number of integer points in all the corresponding
faces.

The following estimate relates the number of integer vertices in the polyhedron
and its Euclidean volume.

Theorem 14.34 (S.V. Konyagin, K.A. Sevastyanov [115]) The number of ver-
tices of an n-dimensional convex integer polyhedron P is bounded from above by
c(n)V (P )(n−1)/(n+1) , where c(n) is a constant not depending on P .

So the generalization should have a different nature. Let us say a few words on
Ehrhart polynomials, which to some extent generalize Pick’s formula.

Definition 14.35 Let P be a d-dimensional convex integer polytope in Rn , and let

t be a real variable. The function
 
L(P , t) = # tP ∩ Zn

is called the Ehrhart polynomial.

Actually, the function L(P , t) is a polynomial only to a certain extent.

Theorem 14.36 (E. Ehrhart [54]) Consider an arbitrary convex polyhedron P in

Rn . Let us restrict the variable t to integers. Then the Ehrhart polynomial of P is a
polynomial of degree d.

So for the integer variable t there exist rational numbers a0 , . . . , ad such that

L(P , t) = ad t d + ad−1 t d−1 + · · · + a0 .

Let us discuss the lattice-geometric meaning of several coefficients.

200 14 Basic Notions and Definitions of Multidimensional Integer Geometry

Proposition 14.37 Let P be an integer polyhedron. Then

a0 = 1;
1 
ad−1 = lV(F );
2 · (d − 1)!
F
lV(P )
ad = .
d!
The sum in the formula for ad−1 is taken over all faces of P of codimension 1.

The other coefficients of the Ehrhart polynomial do not have a simple combina-
torial interpretation now. We refer the interested reader to a recent book by M. Beck
and S. Robins [18].

Remark 14.38 Let us study the Ehrhart polynomials for integer polygons in the
plane. Consider a polygon P , let I be the number of integer points in the interior of
P , and let E be the number of integer points in the union of edges of P . Then we
have
L(P , t) = A(P ) · t 2 + E/2 · t + 1.
By Definition 14.35 we have
 
L(P , 1) = # P ∩ Z2 = I + E.

The last two equations (for t = 1) imply Pick’s formula:

A(P ) = I + E/2 − 1.

Example 14.39 Let us write the Ehrhart polynomials for the tetrahedra T1 and T2 in
Example 14.33:
1 11
L(T1 , t) = x 3 + x 2 + x + 1;
6 6
1 3 5
L(T2 , t) = x + x 2 + x + 1.
3 3

14.6 Exercises

Exercise 14.1 Calculate the integer distance and the integer angle between the fol-
lowing two planes in R4 :
 
x = 0, x + y = 0,
and
y = 0, z + t = 4.
14.6 Exercises 201

Exercise 14.2 Find some integer invariants of arrangements of three lines in Rn .

Exercise 14.3 Prove the statements of Proposition 14.20.

Exercise 14.4 Prove that any two integer k-dimensional planes in Rn are integer
congruent.

Exercise 14.5 Find the Plücker coordinates of the sublattice generated by the fol-
lowing vectors:

v1 = (1, 2, 3, 0, 0), v2 = (1, 1, 1, 1, 1), and v3 = (0, 0, 3, 2, 1).

Exercise 14.6 Find and prove the expression for the leading coefficient of the
Ehrhart polynomial.

Exercise 14.7 Calculate the Ehrhart polynomial for the tetrahedron

 
T = conv (0, 0, 0), (2, 0, 0), (0, 3, 0), (1, 2, 3) .
Chapter 15
On Empty Simplices, Pyramids, Parallelepipeds

Recall that an integer polyhedron is called empty if it does not contain integer points
other than its vertices. In this chapter we give the classification of empty tetrahedra
and the classification of pyramids whose integer points are contained in the base
of pyramids in R3 . Later in the book we essentially use the classification of the
mentioned pyramids for studying faces of multidimensional continued fractions.
In particular, the describing of such pyramids simplifies the deductive algorithm
of Chap. 20 in the three-dimensional case. We continue with two open problems
related to empty objects in lattices. The first one is a problem of description of empty
simplices in dimensions greater than 3. The second is the lonely runner conjecture.
We conclude this chapter with a proof of a theorem on the classification of empty
tetrahedra.

15.1 Classification of Empty Integer Tetrahedra

Recall that in the two-dimensional case there is a unique class of empty triangles
(see Proposition 2.11 and Remark 2.12). In this section we examine the situation in
the three-dimensional case. We start with White’s theorem on geometric properties
of empty tetrahedra.
Let ABCD be a tetrahedron with enumerated vertices. Denote by P (ABCD) the
following parallelepiped:

{A + αAB + βAC + γ AD | 0 ≤ α ≤ 1, 0 ≤ β ≤ 1, 0 ≤ γ ≤ 1}.

Theorem 15.1 (White’s theorem [207]) Let ADBA be an empty tetrahedron. Then
all integer points in P (ADBA ) except the vertices are contained in one of the planes
passing through centrally symmetric edges distinct from the edges of the tetrahe-
dron.

Remark 15.2 For every parallelepiped there exist exactly three planes as in the the-
orem. See an example with four inner integer points in Fig. 15.1.

O. Karpenkov, Geometry of Continued Fractions, 203

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_15, © Springer-Verlag Berlin Heidelberg 2013
204 15 On Empty Simplices, Pyramids, Parallelepipeds

Fig. 15.1 If ADBA is empty, then all integer points are in one of the dark gray three planes. We
give an example with four points inside

Theorem 15.1 leads to a complete classification of empty integer tetrahedra with

one marked vertex.

Corollary 15.3
(i) Classification of marked empty tetrahedra. An integer empty tetrahedron with
a marked vertex is integer congruent to exactly one of the following marked
tetrahedra:
– a tetrahedron with vertices (0, 0, 0), (1, 0, 0), (1, 0, 1), and (1, 1, 0);
ξ
– T1,r of the list “T-W,” where r ≥ 2, 0 < ξ ≤ r/2, and gcd(ξ, r) = 1 (see
Fig. 15.2).
The point (0, 0, 0) is a marked vertex for all the tetrahedra in the list.
(ii) Classification of empty tetrahedra. Any empty tetrahedron is integer congruent
ξ ν (without marked
to an empty marked tetrahedron. The tetrahedra T1,r1 and T1,r 2
vertex) are integer congruent if and only if r1 = r2 and for r1 > 1, at least one
of the following four equations holds:

(ξ ) ≡ (±ν mod r1 )±1 .

Remark 15.4 Classifications of empty tetrahedra and empty marked tetrahedra co-
incide only for r = 1, 2, 3, 4, 5, 6, 8, 10, 12, 24. (This is the case when the square
of every element of the multiplicative group (Z/mZ)∗ equals ±1. Hence all the
summands in the multiplicative group are either Z/4Z or Z/2Z. There are 15 such
groups. Only nine of them satisfy the condition. In addition we have the case r = 1.)

We give all the proofs for the statements of this section in Sect. 15.4.

15.2 Classification of Completely Empty Lattice Pyramids

We say that a pyramid is marked if the vertex of the pyramid is specified. (So every
tetrahedron corresponds to four marked pyramids, while all the other pyramids are
in one-to-one correspondence with marked pyramids.) A pyramid in R3 is called
15.2 Classification of Completely Empty Lattice Pyramids 205

Fig. 15.2 The T-W list of empty pyramids. The origin (0, 0, 0) is the marked vertex for all the
pyramids

integer if its vertex is an integer point and its base is an integer polygon. An integer
pyramid is called completely empty if every integer point inside the pyramid is either
contained in the base or coincides with the vertex of the pyramid.
An integer pyramid is called one-story (multistory) if the integer distance from
the vertex to the base is 1 (≥ 1). One-story integer empty pyramids can have an arbi-
trary polygon as a base, since there is no geometric obstacle to constructing arbitrary
pyramids. The multistory case is more interesting, as described in the following the-
orem.

Theorem 15.5 ([91]) Every multistory completely empty convex three-dimensional

marked pyramid is either triangular or quadrangular. Every triangular multistory
completely empty pyramid is integer congruent to exactly one of the marked pyra-
mids from the list “T-W” (see Fig. 15.2). Every quadrangular multistory completely
empty pyramid is integer congruent to exactly one of the following pyramids:

We skip the complete technical proof Theorem 15.5 (for a complete proof we
refer to [91]).
206 15 On Empty Simplices, Pyramids, Parallelepipeds

Fig. 15.3 All empty

three-dimensional tetrahedra
are of width 1

15.3 Two Open Problems Related to the Notion of Emptiness

Let us briefly mention two open problems related to arrangements of integer points
in the interior of base parallelepipeds of integer sublattices.

15.3.1 Problem on Empty Simplices

In this subsection we say a few words about the following problem.

Problem 10 Classify empty integer simplices in dimension n.

From Proposition 2.11 it follows that all empty triangles are integer congru-
ent to the coordinate triangle (n = 2). In Corollary 15.3 all distinct empty three-
dimensional tetrahedra are listed (n = 3). For the rest of the cases (n ≥ 4), the prob-
lem is open. In the case n = 4, we would like to mention one interesting recent
result. To formulate it we introduce a notion of the width of integer polyhedra.

Definition 15.6 The width of an integer polyhedron P in Rn is the minimal integer

distance between two parallel integer hyperplanes containing P in the middle.

Let us illustrate this definition with several examples.

Example 15.7 The widths of the following triangles,

are respectively 1, 2, and 3. These triangles have minimal possible integer area
among the triangles of the given width: the areas are 1, 3, and 7 respectively.

Example 15.8 The width of every empty three-dimensional tetrahedron is 1.

The last example means that every empty tetrahedron is contained between two
neighboring parallel integer planes (see Fig. 15.3).
Almost all four-dimensional empty integer simplices have the following property.
15.3 Two Open Problems Related to the Notion of Emptiness 207

Theorem 15.9 (M. Barile et al. [16]) Up to possibly a finite number of exceptions,
every empty simplex in dimension 4 has width 1 or 2.

To conclude this subsection we observe once more that the complete classifi-
cation of empty simplices in R4 is still not known, while the situation in higher
dimensions is entirely open.

15.3.2 Lonely Runner Conjecture

In this subsection we discuss the lonely runner conjecture, which is also known as
view-obstruction problem for n-dimensional cubes. We start with a general formu-
lation of the conjecture.

Conjecture 11 (Lonely runner conjecture) Suppose k runners having distinct con-

stant speeds start at a common point and run laps on a circular track with circum-
ference 1. Then for any given runner, there is a time at which that runner is at arc
distance at least 1/k away from every other runner.

This conjecture was introduced by J.M. Wills [208] in 1967. Later it was consid-
ered by T.W. Cusick (in [37, 39], and [40]) as the geometric view-obstacle problem.
The cases k = 2, 3 are easy exercises. The conjecture holds for k = 4 (U. Betke and
J.M. Wills [19], T.W. Cusick [39]), k = 5 (T.W. Cusick and C. Pomerance [42], see
also in [20]), k = 6 (T. Bohman et al. [21]; see also in [20]), and k = 7 (J. Barajas
and O. Serra [13] and [14]). For k ≥ 8 it is not known whether this conjecture is
true.
Let us describe the question in geometric terms. Without loss of generality we
study the conjecture only for the first runner. We say that the starting point is the
zero point of a circle. It is usual to put the speed of the first runner equal to zero.
We associate each runner that travels with nonzero speed with a point on a circle.
The configuration space of all (k − 1)-tuples on the circle is the (k − 1)-dimensional
torus T k−1 (which is the direct product of k − 1 copies of circles S 1 = R/Z). We
mark the point (0, . . . , 0) on this torus as the first runner (who is actually not running
but standing). While the time goes from zero to infinity, the configuration of (k − 1)-
runners travels linearly in the torus T k−1 . The question is whether this trajectory is
always far from the middle point of the torus (1/2, 1/2, . . . , 1/2).
To simplify the picture it is good to consider the universal covering space of
T k−1 , which is Rk−1 . Here the standing runner is associated with an integer lattice
Zk−1 . Now we reformulate the lonely runner conjecture as a view-obstacle problem.

Lonely Runner Conjecture as a View-Obstacle Problem A hunter is standing

in the middle (origin) of an n-dimensional garden. At each vertex of the shifted
lattice Zk−1 + (1/2, 1/2, . . . , 1/2) a tree grows. All the trees are hypercubes of size
1 − 2/k centered at the vertex of the shifted lattice and with sides parallel to the
208 15 On Empty Simplices, Pyramids, Parallelepipeds

coordinate axes. Suppose the hunter is not willing to shoot in the directions having
at least one zero coordinate. Then he will shoot into some tree. It is supposed that
the bullet has zero size.
Finally, we give a discrete lattice version of the lonely runner conjecture.

Discrete Version of Lonely Runner Conjecture Does there exist a full-rank

sublattice of Zk−1 (suppose it is generated by v1 , . . . , vk−1 ) such that the hypercube
k−1  $
 1 k−1

λi v i  ≤ λi ≤ , i = 1, 2, . . . , k − 1
k k
i=1

does not contain integer points.

Remark 15.10 Let us say a few words about the inverse question of meeting of all
the runners. Suppose k runners having distinct constant speeds start at a common
point and run laps on a circular track. Then for every ε > 0 and T > 0, there is a
time t > T at which all the runners are at arc distance at most ε away from each
other. We leave this as an exercise to the reader.

15.4 Proof of White’s Theorem and the Empty Tetrahedra

Classification Theorems
We start with a proof of White’s theorem. Further, we deduce from it the classi-
fication of empty tetrahedra. Before proving White’s theorem, we introduce some
necessary definitions and prove several preliminary statements.

15.4.1 IDC-System

We begin with some necessary definitions.

Definition 15.11 Let A, B, C, and D be integer points that span the entire space
R3 (the collection of points is ordered). Let us construct a system of coordinates
related to these points. Let b, c, and d be the integer distances from the points B, C,
and D to the planes of the faces ACD, ABD, and ACD, respectively. Consider the
point A as the origin. Set the coordinates of B, C, and D equal to (b, 0, 0), (0, c, 0),
and (0, 0, d) respectively. The coordinates of the other points in R3 are defined by
linearity. We say that this system of coordinates is an integer-distance coordinate
system with respect to ABCD or (IDC-system for short).

We call the points with integer coordinates in an IDC-system the IDC-nodes. We

say that an IDC-node is integer if it is an integer point of R3 . Notice that every
15.4 White’s Theorem and the Empty Tetrahedra Classification Theorems 209

integer point in the old system is an IDC-node. The converse is true if and only if
the vectors AB, AC, and AD generate the integer lattice. The coordinates of integer
IDC-nodes coincide with the integer distances to the coordinate planes in the IDC-
system.

15.4.2 A Lemma on Sections of an Integer Parallelepiped

If a and b are two integer vectors, denote by La,b the integer sublattice generated by
a and b. Consider the set of all cosets in Z3 /La,b . Every left coset of this quotient
group is contained in a plane parallel to the plane spanned by the vectors a and b. In
addition, every such plane contains only finitely many left cosets, their number being
equal to the index of the sublattice La,b in the integer sublattice in the subspace
Span(a, b).

Lemma 15.12 Consider an integer parallelepiped P = ABCDA B C D and a

plane π parallel to ABCD. Let π intersect the parallelepiped (by some parallel-
ogram). Then the following two statements hold.
(i) The section of the parallelepiped P by the plane π contains representatives of
all left cosets of Z3 /La,b that are in π .
(ii) Every left coset of Z3 /La,b contained in π intersects the parallelepiped P , and
the intersection is described by one of the following cases:
(a) a point inside the parallelogram π ∩ P ;
(b) two points on opposite edges of π ∩ P ;
(c) four points coincide with the vertices of π ∩ P .

We leave the proof of this lemma as a simple exercise for the reader.

15.4.3 A Corollary on Integer Distances Between the Vertices and

the Opposite Faces of a Tetrahedron with Empty Faces

Corollary 15.13 All integer distances from the vertices of an integer (three-
dimensional) tetrahedron with empty faces to the opposite faces are equal.

Proof Consider an integer tetrahedron OABC with empty faces. Suppose that

ld(A, OBC) = n.

Let us show that ld(B, OAC) = ld(C, OAB) = n.

Consider the parallelepiped P (OABC). Since the triangles OBC, OAB, and OAC
are empty, the corresponding faces of P (OABC) are empty as well. Hence all faces
210 15 On Empty Simplices, Pyramids, Parallelepipeds

of the parallelepiped are empty. Consider all the integer planes parallel to OBC
that divide the parallelepiped into two nonempty parts. Since ld(A, OBC) = n, the
number of these planes equals n − 1.
By Lemma 15.12 every such integer plane contains exactly one left coset of the
lattice Z3 /LOB,OC . Since the faces of the parallelepiped are empty, the intersec-
tion of this class with a parallelepiped is one point in the interior. Hence the paral-
lelepiped P (OABC) contains exactly n − 1 integer points.
Suppose that ld(B, OAC) = m, and ld(B, OAC) = k. Then for similar reasons
the parallelepiped contains m − 1 and k − 1 interior integer points, respectively.
Therefore, we have k = m = n. 

15.4.4 Lemma on One Integer Node

For the proof of Theorem 15.1 we need the following lemma. Let ld(B, ACD) = r
for some r > 1. Consider the IDC-system with respect to ADBA (with integer-
distance coordinates (x, y, z)). Denote by B  , C, C  , and D  the vertices with coor-
dinates (0, r, r), (r, r, 0), (r, r, r), and (r, 0, r), respectively.

Lemma 15.14 There is a unique integer node in the interior of the intersection of
the plane x + y + z = r + 1 and the parallelepiped (here we restrict ourselves to the
case r > 1).

Proof Since ld(A, A B CD) = r and the parallelogram A B CD is contained in the

plane x + y + z = r, for every integer n, the plane x + y + z = n is integer. So the
plane x + y + z = r + 1 is also integer.
Consider a parallelogram obtained from A B CD by the shifting on the vector
(0, 0, 1) in the IDC-system. This parallelogram is contained in the plane x + y +
z = r + 1. The edges of the shifted parallelogram are on the planes of faces of the
parallelepiped and they do not contain vertices. Hence there exists a unique point
in the interior of the shifted parallelogram, denoted by K(x0 , y0 , z0 ). Therefore, we
have at most one point satisfying the conditions of the lemma.
Let us show that K(x0 , y0 , z0 ) is inside the parallelepiped. The shifted parallelo-
gram is defined by the following four inequalities and one equation:

0 ≤ x ≤ r, 0 ≤ y ≤ r, and x + y + z = r + 1.

Since the point (0, 0, r + 1) is not integer (r + 1 is not divisible by r), we have
z0 ≤ r.
Suppose that z0 ≤ 0. In this case the vector KC is integer; hence the point
K  = A + KC is an integer node (see Fig. 15.4). Notice that the point K  is in the
tetrahedron ADBA and it is distinct from the vertices of the tetrahedron. So ADBA is
not empty, contradicting the conditions of the lemma. Therefore, we obtain z0 > 0.
Hence we get 0 ≤ x0 ≤ r, 0 ≤ y0 ≤ r, and 0 ≤ z0 ≤ r. Since the edges of the
shifted parallelogram do not pass through the vertices of the parallelepiped, the
point K is in the interior of the parallelepiped. 
15.4 White’s Theorem and the Empty Tetrahedra Classification Theorems 211

Fig. 15.4 The points K and

K  are the integer nodes

15.4.5 Proof of White’s Theorem

We prove the statement of the theorem by induction on the number of points inside
the parallelepiped (or equivalently on the value of ld(A , ABD)).
Base of induction. Suppose that ld(A , ABD) = 2. Then the unique integer node
inside the parallelepiped is the center of symmetry of the parallelepiped. It is in the
intersection of all planes passing through opposite edges.
Step of induction. Suppose that the statement of the theorem holds for all empty
tetrahedra with integer distance from the vertices to the opposite edges less than r.
Let us prove the theorem for an arbitrary empty tetrahedron A ABD with
 
ld A , ABD = r.

By Lemma 15.14 there exists a unique integer node in the plane x + y + z = r + 1

inside the parallelepiped P (A ABD), denoted by A (x0 , y0 , z0 ) in the IDC-system
with respect to the tetrahedron ABDA . Without loss of generality, we suppose
that z0 ≥ x0 and z0 ≥ y0 . The tetrahedron ADBA is an empty integer tetrahe-
dron, since all the nodes inside ADBA except A are inside ADCA as well. We
get ld(A , ADB) < r, and hence by induction we have that all integer nodes of
P (ADBA ) are either in the plane A CD, in the plane A BC, or in the plane B BD
(where B  = A + AB). Consider the intersections of these planes with the plane
z = 1. One of these intersections should contain an integer node inside the paral-
lelepiped P (ADBA ).

Case 1. The Integer Nodes Are in the Plane A CD The plane A CD is defined
by the equation
r − x0
x+ z = r.
z0
212 15 On Empty Simplices, Pyramids, Parallelepipeds

The intersection of the plane A CD with the plane z = 1 is the line (x1 , t, 1), where t
is a linear parameter of the line and x1 is a constant equal to r − r−x0
z0 . This plane con-
tains integer nodes only if x1 is integer. Let us estimate it (assuming that y0 ≤ z0 ).
On the one hand,

r − x0 r − (r + 1 − z0 − y0 ) 1 y0
x1 = r − =r − =r + −1−
z0 z0 z0 z0
1
≥r −2+ > r − 2.
z0

On the other hand, x1 < r. Since x1 is an integer and r − 2 < x1 < r, we have
x1 = r − 1. Therefore, there exist integer nodes with coordinates (r − 1, t, 1), and
hence by Lemma 15.12 one of them (with coordinates (r − 1, t0 , 1), for some integer
t0 ) is contained in the parallelepiped P (ADBA ). So the rest of the integer nodes of
the parallelepiped P (ADBA ) have the coordinates (r − k, (k · t0 mod r), k), and
hence they are all in the parallelogram A B CD.

Case 2. The Integer Nodes Are in the Plane A BC The plane A BC is defined
by the equation
r − y0
y+ z = r.
z0
The intersection of the plane A BC with the plane z = 1 is the line (t, y1 , 1), where t
is a linear parameter of the line and y1 is a constant equal to r − r−y0
z0 . This plane con-
tains integer nodes only if x1 is integer. Let us estimate it (assuming that x0 ≤ z0 ).
On the one hand,

r − y0 r − (r + 1 − z0 − x0 ) 1 x0
y1 = r − =r − =r + −1−
z0 z0 z0 z0
1
≥r −2+ > r − 2.
z0

On the other hand, y1 < r. Since y1 and r − 2 < y1 < r, we have y1 = r − 1.

Therefore, there exist integer nodes with coordinates (t, r − 1, 1), and hence by
Lemma 15.12 one of them (with coordinates (t0 , r − 1, 1), for some integer t0 ) is
contained in the parallelepiped P (ADBA ). So the rest of the integer nodes of the
parallelepiped P (ADBA ) have the coordinates ((k · t0 mod r), r − k, k), and hence
they are all in the parallelogram A B CD.

The Integer Nodes Are not in the Plane B BD Let us show that the case of the
plane B BD is an empty case. This plane is defined by the equation

r − y0 − x0
x +y + z = r.
z0
15.4 White’s Theorem and the Empty Tetrahedra Classification Theorems 213

The intersection of the plane B BD with the plane z = 1 is the line ( a+t a−t
2 , 2 , 1),
where t is a linear parameter for the line and a is a constant equivalent to r −
r−x0 −y0
z0 . Such a line contains integer nodes only if a is integer. On the one hand,

r − y0 − x0 z0 − 1
a=r − =r − > r − 1.
z0 z0
On the other hand, we have a < r. Therefore, there are no integer nodes in the
intersection of the parallelepiped P (ADBA ) (and of the parallelepiped P (ADBA )
as well) with the plane z = 1, which is impossible.
Therefore, the first two cases enumerate all realizable positions of integer nodes.
In both cases the integer nodes are in one of the planes mentioned in the formulation
of the theorem. Hence, the statement holds for an arbitrary pyramid ABCDA with
ld(A , ABD) = r. The proof of theorem is completed by induction. 

Remark 15.15 Notice that the third case in the proof is empty, since we assume that
the third coordinate of A is not less than the first and the second coordinates.

15.4.6 Deduction of Corollary 15.3 from White’s Theorem

Proof of Corollary 15.3(i) We prove three statements: on completeness of the list,

on emptiness of the tetrahedra in the list, and on pairwise noncongruence of the
listed tetrahedra.
Completeness of the list. Considering an integer empty tetrahedron A ABD with
a marked vertex A , let ld(A , ABD) = r. If r = 1, then we get a tetrahedron integer
congruent to (0, 0, 0), (1, 0, 0), (1, 0, 1), and (1, 1, 0).
Suppose that r > 1. Then consider a point K a unit integer distance from the
plane ABD lying in the parallelepiped P (ABDA ). By White’s theorem, K is in one
of the three diagonal planes. Without loss of generality, we assume that K is in
A B CD. So K has coordinates (ξ, r − 1, 1) in the IDC-system related to ABDA .
Let us rewrite the coordinates of the point in the basis AB, AC, and AK (notice that
it is a basis of the integer lattice). We have

A = (0, 0, 0), B = (1, 0, 0), C = (0, 1, 0), and A = (−ξ, 1 − r, r).

Hence ABDA is integer congruent to T1,r .

ξ

The parameter ξ is relatively prime to r, since there would otherwise exist integer
ξ r−ξ
points on the edges distinct from the vertices. The tetrahedra T1,r and T1,r are
integer congruent. Therefore, we restrict the consideration of the tetrahedra to the
case 0 < ξ ≤ r/2.
Emptiness of the tetrahedra. All the tetrahedra are empty, since all integer points
in P (ABDA ) are contained in one of the diagonal planes that does not intersect the
tetrahedron in the interior of the parallelepiped.
214 15 On Empty Simplices, Pyramids, Parallelepipeds

Any two marked tetrahedra in the list are not integer congruent. Let us introduce
the following integer invariant to distinguish the tetrahedra. Consider an arbitrary
marked tetrahedron ABDA with vertex A and the related three-sided angle with
vertex at A and a section ABD. By Lemma 15.14 this plane contains a unique
integer point (denoted by K) on the plane parallel to ABD at integer distance r + 1
from A . Integer distances to the faces of the three-sided angle are 1, ξ , and r − ξ for
some ξ . The unordered collection [1, ξ, r − ξ ] is an integer invariant of the marked
tetrahedron A ABD. This invariant (together with ld(A , ABD) = r) distinguishes all
distinct tetrahedra in the list of the corollary. 

Proof of Corollary 15.3(ii) We prove the same three statements as in the proof of
Corollary 15.3(i).
Completeness of the list. Emptiness of the tetrahedra. By Corollary 15.3(i) every
empty tetrahedron is integer congruent to at least one of the marked tetrahedra in
the list of Corollary 15.3(i). We have already shown that all the tetrahedra in this list
are empty.
Integer congruence of tetrahedra in the list. The affine group of symmetries of
tetrahedra S4 includes the affine group of symmetries of marked tetrahedra S3 . For
each of the four angles of a tetrahedron we construct the integer point uniquely
defined by the conditions of Lemma 15.14. Direct calculation shows that the inte-
ger 4-tuples of distances from these points to four planes of the faces of an empty
tetrahedron are as follows

(1, 1, ξ, r − ξ ), (1, 1, ξ, r − ξ ), (ν, r − ν, 1, 1), and (ν, r − ν, 1, 1),

where r is the volume of the tetrahedron and the pair (ξ, ν) satisfies ξ · ν ≡
(1 mod r). So this unordered 4-tuple of unordered 4-tuples is an invariant of empty
ξ ν , T r−ξ , and T r−ν in the list of Corollary 15.3(i)
tetrahedra. The tetrahedra T1,r , T1,r 1,r r
are integer congruent; all the other pairs of empty tetrahedra in the list have distinct
4-tuples of unordered 4-tuples of distances, and hence they are not congruent. 

15.5 Exercises
Exercise 15.1 Find triangles of width 4 with the smallest possible integer area.

Exercise 15.2 Prove that the width of all empty 3-dimensional tetrahedra is 1.

Exercise 15.3 Find proofs for the lonely runner conjecture for 2 and 3 runners.

Exercise 15.4 Prove the statement of Remark 15.10.

Exercise 15.5 Write the Ehrhart polynomial for all empty tetrahedra.

Exercise 15.6 Prove the equivalence of the lonely runner conjectures in the classi-
cal, view-obstacle, and discrete forms.
Chapter 16
Multidimensional Continued Fractions
in the Sense of Klein

In the first part of this book we studied continued fractions from the geometric point
of view of sails and their integer invariants. In this chapter we present a natural
generalization of sails to the multidimensional case by F. Klein.

16.1 Background

In 1839, C. Hermite [77] posed the problem of generalizing ordinary continued

fractions to the higher-dimensional case. Since then, there have been many differ-
ent definitions generalizing different properties of ordinary continued fractions. In
this book we focus on the geometric generalization proposed by F. Klein in [108]
and [109]. Multidimensional continued fractions in the sense of Klein have many
relations with other branches of mathematics. For example, in [200], H. Tsuchi-
hashi described the relationship between periodic multidimensional continued frac-
tions and multidimensional cusp singularities. M.L. Kontsevich and Yu.M. Suhov
studied the statistical properties of random multidimensional continued fractions in
[113], and later they were studied by the author in [93]. O.N. German [67] and
J.-O. Moussafir [141] discussed the connection between the sails of multidimen-
sional continued fractions and Hilbert bases. The relations to approximation theory
of maximal commutative subgroups is discussed by A. Vershik and the author in
[101]. The combinatorial topological generalization of Lagrange’s theorem was ob-
tained by E.I. Korkina in [116] and its algebraic generalization by G. Lachaud [123].
The book [11] of V.I. Arnold is a good survey of geometric problems and theorems
associated with one-dimensional and multidimensional continued fractions in the
sense of Klein (see also his articles [7, 9], and [10]).
Originally, F. Klein introduced his multidimensional continued fractions in the
context of studying decomposable forms with integer coefficients, which we dis-
cuss later in more general settings in Chap. 21. Periodicity of multidimensional
sails plays an important role in the study of algebraic irrationalities, since the pe-
riodic structure is defined by the induced action of the group of units in the orders

O. Karpenkov, Geometry of Continued Fractions, 215

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_16, © Springer-Verlag Berlin Heidelberg 2013
216 16 Multidimensional Continued Fractions in the Sense of Klein

of algebraic fields. Later, we give a multidimensional analogue of Lagrange’s the-

orem, in Chap. 18. Periods of sails are complete invariants of conjugacy classes of
SL(n, Z) matrices in multidimensional Gauss’s reduction theory (see Chap. 21). In
Chap. 19 we present statistical properties of multidimensional sails. Most of the
interesting questions related to multidimensional sails appear already in the three-
dimensional case, so within this book we pay the most attention to sails in R3 . We
observe that some three-dimensional theorems have quite simple generalizations to
the case of higher dimensions, while some others become hard open problems even
in the four-dimensional case.
We say a few words about the other generalizations of multidimensional contin-
ued fractions in Chap. 23.

16.2 Some Notation and Definitions

16.2.1 A-Hulls and Their Boundaries

We begin with several preliminary definitions.

Consider a set S ⊂ Rn . We denote by S the closure of S. A point p ∈ S is called
interior if there exists a ball in Rn with center at p that is completely contained in S.
The union of all interior points of S is called the interior of S. The boundary of S
is the set of points in S not belonging to the interior of S. We denote it by ∂S. An
element of the boundary of S is called a boundary point of S.
A simplicial cone in Rn is the convex hull of k rays with the same vertex and
linearly independent directions. (It immediately follows that k ≤ n.) The vertex of
a simplicial cone is the common vertex of the rays defining the cone. The rays are
called the edges of the cone. The convex hull of any subset of rays defining the cone
C is called a face of C. A simplicial cone is called integer if its vertex is integer.
A simplicial cone is called rational if its vertex is integer and all its edges contain
integer points distinct from the vertex.

Definition 16.1 Consider an integer simplicial cone C of dimension d with center

at the origin in Rn . An A-hull of the cone C is the convex hull of all integer points
lying in the closure C except the origin. We denote this set by A-hull(C).

Now we formulate the classical concept of the sail, which is the cornerstone of
the theory of multidimensional continued fractions in the sense of Klein.

Definition 16.2 The boundary of the set A-hull(C) is called the sail of this cone.
16.2 Some Notation and Definitions 217

16.2.2 Definition of Multidimensional Continued Fraction in the

Sense of Klein

Let us give the classical definition of multidimensional continued fractions. Con-

sider a set of n + 1 hyperplanes of Rn+1 passing through the origin in general posi-
tion (i.e., whose union is homeomorphic of the union of all coordinate hyperplanes).
The complement to the union of these hyperplanes consists of N = 2n+1 open sim-
plicial cones C1 , . . . , CN .

Definition 16.3 The set of all sails for all the cones C1 , . . . , CN is called the n-
dimensional continued fraction associated to the given n + 1 hyperplanes in Rn+1
passing through O in general position.

In this chapter we investigate combinatorial and topological properties of sails.

First we study the case of finite multidimensional continued fractions. Further we
prove a generalized Kronecker’s approximation theorem (which is actually inter-
esting by itself) and deduce from it general theorems on the polyhedral structure of
sails. Finally, we discuss the situation around two-dimensional faces of sails in more
detail.

16.2.3 Face Structure of Sails

We begin with several preliminary general definitions.

Definition 16.4 Let S be a convex set in Rn . A hyperplane π is said to be supporting

if the following two conditions hold:
(i) the set S is entirely contained in one of the two half-spaces defined by π ;
(ii) π ∩ S
= ∅.

To define faces of sails we use the following rather abstract definition.

Definition 16.5 Assume that π is a supporting hyperplane for a convex set S and
that S is not a subset of π . Then we say that the set Fπ = π ∩ S is a face of S.
The dimension of Fπ is the dimension of a face. If Fπ is a point or a segment
(including rays), we say that the face is either a vertex or an edge.

Remark 16.6 Notice that it is not necessary that the faces of A-hulls coincide with
the faces of their closures. One of the illustrations of this is the cone of Exam-
ple 16.20 below, shown in Fig. 16.4. The shaded area is not a face of A-hull(C), but
it is a face of its closure. The face of the A-hull(C) is the dashed ray.

Let us now introduce the notion of faces for sails.

218 16 Multidimensional Continued Fractions in the Sense of Klein

Fig. 16.1 A sail of a finite

continued fraction

Definition 16.7 Consider a simplicial n-dimensional cone C. A subset F in the sail

of C is called a face of the sail if F is a face of A-hull(C).

In the most interesting cases the set A-hull(C) is closed (see Theorem 16.24). So
the following rather tautological proposition is of interest in studying the faces of
sails.

Proposition 16.8 Consider a simplicial n-dimensional cone C in Rn . Every face of

the set A-hull(C) is a convex hull of a subset of Zn .

16.3 Finite Continued Fractions

Consider a rational simplicial cone C in Rn+1 with vertex at the origin. The sail of
C has finitely many compact faces of all dimensions. In fact, all compact faces are
contained in the pyramid OA1 A2 . . . An+1 , where Ai is the first integer point on the
ith edge distinct from O. We call sails of rational cones finite. If a multidimensional
continued fraction has a finite sail, then all the other sails of this continued fraction
are also finite; we call this continued fraction finite.

Example 16.9 In Fig. 16.1 we give an example of a sail for a simplicial cone defined
by the three vectors (3, 2, 3), (1, −4, 3), and (−2, 1, 3).

We recall that the classification of two-dimensional cones (i.e., planar angles) and
therefore the corresponding geometric one-dimensional sails is provided by LLS
sequences for both the finite and infinite cases in Chap. 4. The problem of describing
multidimensional (in particular two-dimensional) finite sails is relatively new and
currently has no consistent answer. Let us study the two-dimensional sails that have
a unique compact face. It is clear that this face is triangular, since the cone has three
edges.

Theorem 16.10 Let the sail for a simplicial cone C in R3 have a unique compact
face. Then this face falls into one of the following two cases.
16.4 On a Generalized Kronecker’s Approximation Theorem 219

(i) If the integer distance from the origin to the triangular compact face is one,
then every integer triangle is realizable as this face. Two sails whose unique
compact faces are at integer distance one are integer congruent if and only if
their triangular faces are integer congruent.
(ii) Suppose that the face is at an integer distance to the origin greater than 1. Then
the integer congruence classes of corresponding sails are in one-to-one corre-
spondence with the list “T-W” of completely empty pyramids shown in Fig. 15.2
(here each pyramid is associated to a simplicial cone centered at a marked ver-
tex of the pyramid and having edges containing edges of the pyramid). The
compact faces of the sails are exactly the bases of such pyramids.

Proof If the integer distance from the face to the vertex is one, then the statement
is straightforward. If the distance is greater than one, then the statement follows
directly from Theorem 15.5 on the classification of integer multistory completely
empty marked pyramids. 

Later we formulate a more general theorem for two-dimensional faces at distance

greater than 1 (see Corollary 16.36).
For sails with the number of compact faces greater than 1 almost nothing is
known. The following general problem remains open.

Problem 12 Describe all finite two-dimensional sails (and the corresponding con-
tinued fractions).

The first tasks here are to investigate sails having two, three, four, etc., compact
faces.

16.4 On a Generalized Kronecker’s Approximation Theorem

Kronecker’s approximation theorem states the following: if θ is an arbitrary irra-
tional number, then the sequence of numbers {kθ } (for k > 0) is dense in the unit
interval. In this subsection we formulate and prove a multidimensional generaliza-
tion of Kronecker’s approximation theorem.
We begin with a preliminary definition of the operation of addition for sets.

16.4.1 Addition of Sets in Rn

Let S and T be two subsets of Rn . We define the sum of S and T as follows:

S ⊕ T = {p + q | p ∈ S, q ∈ T }.
We show an example of the sum of a triangle and a square in Fig. 16.2.
The sum operation is invariant with respect to affine transformations.
220 16 Multidimensional Continued Fractions in the Sense of Klein

Fig. 16.2 Addition of two convex polygons

Proposition 16.11 The sum of two convex sets is convex.

Proof The proof is straightforward. Consider two convex sets S and T . Let u, v ∈
S ⊕ T . Then u = s1 + t1 and v = s2 + t2 , where s1 , s2 ∈ S and t1 , t2 ∈ T . Hence for
every λ ∈ [0, 1] we have
λu + (1 − λ)v = λ(s1 + t1 ) + (1 − λ)(s2 + t2 )
   
= λs1 + (1 − λ)s2 + λt1 + (1 − λ)t2 .
The two summands in the last expression are in S and T respectively. Hence the
sum is in S ⊕ T . Hence for any two points of S ⊕ T the segment connecting these
points is also in S ⊕ T . Therefore, the set S ⊕ T is convex. 

16.4.2 Integer Approximation Spaces and Affine Irrational Vectors

Let us give some preliminary definitions. We call a subspace L integer if it contains

an integer sublattice of full rank in L.

Definition 16.12 Let u ∈ Rn . We call the intersection of all integer vector subspaces
of Rn containing u the integer approximation space of u. It is denoted by Ru .

Definition 16.13 A vector u ∈ Rn is called affinely irrational if its endpoint is not

contained in any integer affine hyperplane of Ru .

Example 16.14
√ √
(i) Consider the vector v = [ 2, 3]. The integer approximation space of v is
two-dimensional; the vector v is affinely√irrational.
(ii) The integer approximation space of [1, 2] is also two-dimensional (i.e., it is
R2 ). Nevertheless, this vector is not affinely irrational, since its end is contained
√ x√= 1.
in the integer line
(iii) Let now u = [ 2, 2]. Then Ru is the line y = x, and u is affinely irrational.
(iv) Finally, for the vector w = [1, 1] we have that Rw is the line y = x. The integer
affine hyperplanes in Rw are just integer points. The vector w is integer itself,
so w is not affinely irrational.

Further, we use the following proposition.

16.4 On a Generalized Kronecker’s Approximation Theorem 221

Proposition 16.15 Let u ∈ Rn . Then for every λ ∈ R \ S, where S is a countable

set, the vector λu is affinely irrational.

Proof The set S is formed by the intersections of the integer affine hyperplanes in
Ru with the line  = {λu | u ∈ R}. No such integer hyperplane π contains the line
, since otherwise, π would contain Ru ; hence π intersects  in at most one point.
Since there are countably many integer planes, S is countable. 

16.4.3 Formulation of the Theorem

We are interested in the following generalization of the Kronecker’s approximation

theorem (see [142] and [125]).

Theorem 16.16 (Multidimensional Kronecker’s Approximation Theorem) Con-

sider an arbitrary vector u in Rn .
(i) Let Lu = Ru ∩ Zn . Then Lu ⊕ {ku | k ∈ R} is dense in Ru .
(ii) For every x ∈ Ru and every ε > 0, the set B(x, ε) ⊕ {ku | k ∈ R} contains
infinitely many integer points.
(iii) For every x ∈ Ru and every ε > 0, the set B(x, ε) ⊕ {ku | k ∈ R+ } contains
infinitely many integer points.
(iv) There exists an integer sublattice T of rank n − dim Ru such that
Zn ⊕ {ku | k ∈ Z} = Ru ⊕ T .
(v) If the vector u is affinely irrational, we have that for every x ∈ Ru and every
ε > 0, the set B(x, ε) ⊕ {ku | k ∈ Z+ } contains infinitely many integer points.
√ √ √ √
Example 16.17 Consider a vector u = [ 2, 3, 2 + 3]. Then the space Ru is a
plane x1 + x2 = x3 and the vector u is affinely irrational. Hence all five statements
of the theorem hold.

16.4.4 Proof of the Multidimensional Kronecker’s Approximation

Theorem

In the proof of Theorem 16.16 we use the following lemma.

Lemma 16.18 Consider u ∈ Rn and let dim Ru = d. Suppose that the integer lattice
Lu = Ru ∩ Zn has an integer basis (e1 , . . . , ed ).
(i) For any nonzero λ, we have Ru = Rλu .
222 16 Multidimensional Continued Fractions in the Sense of Klein

(ii) Consider an affinely irrational vector u. Let


d
w = λ0 u + λi ei ,
i=1
where the numbers λ0 , λ1 , . . . , λn are rational, and additionally λ0
= 0. Then
Ru = Rw and the vector w is affinely irrational.
(iii) Let u = (a1 , . . . , ad−1 , 1), in coordinates related to the basis e1 , . . . , ed , de-
note ũ = (a1 , . . . , ad−1 , 0). Then Rũ = Span(e1 , . . . , ed−1 ) and ũ is affinely
irrational.

Proof Lemma 16.18(i) holds, since the linear spaces defined by u and λu coincide.
Let us prove Lemma 16.18(ii). Let u be affinely irrational and let w be as in
the statement of the lemma. Let us first prove that Rw = Ru . A hyperplane of Ru
containing w is integer if and only if a hyperplane shifted by u − w containing w is
integer. Hence Rw = Ru and w is affinely irrational.
Finally we prove Lemma 16.18(iii). Consider an arbitrary integer subspace R
containing ũ. Then the span of R and ed contains u. Hence Ru is in the span of Rũ
and ed . Therefore, Rũ coincides with the span of e1 , . . . , ed−1 .
Now we prove that ũ is affinely irrational by reductio ad absurdum. Suppose
that it is contained in some integer hyperplane of Span(e1 , . . . , ed−1 ) defined by
a linear equation f = 0. Then u satisfies two independent linear equations f = 0
and xd = 0. Hence u is contained in some integer plane π of codimension 2 in Ru .
Hence u is contained in the integer hyperplane Span(π, O) (where O is the origin)
of Ru , and hence Ru is not the integer approximation space of u. We arrive at a
contradiction. The proof is complete. 

Proof of Theorem 16.16 Let us prove Theorem 16.16 by induction in dimension of

Ru .
Base of induction. The statement for the case dim Ru = 1 follows directly from
the classical Kronecker’s approximation theorem (recall that here u is an irrational
number).
Step of induction. Let all the statements of the theorem hold for all vectors whose
integer approximation space is (d − 1)-dimensional. We first prove statement (v)
for an arbitrary affinely integer vector u with dim Ru = d by reductio ad absurdum.
Suppose that for some x ∈ Ru , there exists ε > 0 such that the set
B(x, ε) ⊕ {ku | k ∈ Z+ }
contains only finitely many integer points. Consider x0 = x + N u for a sufficiently
large N such that the set
B(x0 , ε) ⊕ {ku | k ∈ Z+ }
does not contain any integer point. For a pair of positive integers k1 > k2 > 0, denote
by Δ(k1 , k2 ) the vector
{x0 + k1 u} − {x0 + k2 u},
16.4 On a Generalized Kronecker’s Approximation Theorem 223

Fig. 16.3 Reduction of the dimension

where for a vector r = (r1 , . . . , rd ), its quotient part is denoted by {r}, i.e.,
 
{r} = r1 − r1 , . . . , rd − rd  .
It is clear that Δ(k1 , k2 ) belongs to the unit coordinate cube. Since this cube is
compact, there exist integers k̂1 > k̂2 > 0 such that
√
 
Δ(k̂1 , k̂2 ) < 3 ε.
2
Notice that Δ(k̂1 , k̂2 )
= 0, since (k̂1 − k̂2 )u ∈
/ Zn . For every nonnegative m, the ball
 
B(x0 , ε) ⊕ mΔ(k̂1 , k̂2 )
does not contain integer points, since this ball is obtained from the ball
 
B(x0 , ε) ⊕ (k̂1 − k̂2 )u
by a shift on the integer vector
x0 + k̂1 u − x0 + k̂2 u,
where r is the vector floor function, which is equal to r − {r}. Hence the set
 
B(x0 , ε) ⊕ mΔ(k̂1 , k̂2 ) | m ∈ Z+
does not contain integer points. By construction, this set contains an (ε/2)-tubular
neighborhood of a ray with vertex at x0 and direction Δ(k̂1 , k̂2 ), i.e., the set
 
T = B(x0 , ε/2) ⊕ tΔ(k̂1 , k̂2 ) | t ∈ R+ ;
see Fig. 16.3 (left). Now we are interested in the points of intersections of the ray
{x0 + tΔ(k̂1 , k̂2 ) | k ∈ R+ } with the integer planes parallel to the span of the vec-
tors e1 , . . . , ed−1 . Let the first and second intersections with these planes be y0 and
y0 + v. Then all the intersections are in the set {y0 + kv | k ∈ Z+ }. By construction,
the set
B(y0 , ε/2) ⊕ {kv | k ∈ Z+ }
224 16 Multidimensional Continued Fractions in the Sense of Klein

does not contain any integer point. Let y 0 and v be the point and the vector whose
first d − 1 coordinates coincide with the coordinates of y0 and v, respectively, and
the last coordinate is zero. Since the last coordinate of y0 and v are integers, the set
B(y 0 , ε/2) ⊕ {kv | k ∈ Z+ }
does not contain any integer point.
By Lemma 16.18(ii) it follows that Ru = RΔ(k̂1 ,k̂2 ) . Further, by Lemma 16.18(i),
we get RΔ(k̂1 ,k̂2 ) = Rv . Since the last coordinate of Rv equals 1, we are in the situa-
tion of Lemma 16.18(iii). Hence the vector v is affinely irrational and
Rv = Span(e1 , . . . , ed−1 ).
Hence y0 is in Rv . Then by the induction assumption for (d −1)-dimensional integer
approximation spaces, the set
B(y 0 , ε/2) ⊕ {kv | k ∈ Z+ }
contains infinitely many integer points. We arrive at a contradiction. Therefore, the
statement of Theorem 16.16(v) is true for an arbitrary vector u with dim Ru = d.
Theorem 16.16(iii) is an easy corollary of Theorem 16.16(v): namely, for an ar-
bitrary u take a scalar λ such that λu is affinely irrational (due to Proposition 16.15),
and hence Theorem 16.16(iii) for u follows from Theorem 16.16(v) for λu.
Theorem 16.16(i) and (ii) follow directly from Theorem 16.16(iii), and Theo-
rem 16.16(iv) follows from Theorem 16.16(i).
All these prove the correctness of the step of induction and conclude the proof of
the multidimensional Kronecker’s approximation theorem. 

16.5 Polyhedral Structure of Sails

In this section we study topological and combinatorial structures of sails. We first
show that all sails are homeomorphic to real spaces. Then we discuss the integer
polyhedral structure of sails.

16.5.1 The Intersection of the Closures of A-Hulls with Faces of

Corresponding Cones

We start with the following proposition for arbitrary simplicial cones.

Proposition 16.19 Consider a simplicial n-dimensional cone C in Rn with vertex

at the origin. Let F be a face of C or else F = C. Then we have
A-hull(C) ∩ F = A-hull(F ) ⊕ F.
If the dimension of the lattice contained in Span F equals the dimension of F , we
have
A-hull(F ) ⊕ F = A-hull(F ).
16.5 Polyhedral Structure of Sails 225

Fig. 16.4 The intersection of

a face F with the closure of
the A-hull of C

Example 16.20 In Fig. 16.4, we give an example of a three-dimensional simpli-

cial cone C in R3 with a two-dimensional face F containing a one-dimensional
integer sublattice. The integer points of the face F are shown by black dots. The
set A-hull(F ) is the integer ray containing all the integer points of F . It is shown
dashed in the figure. The intersection of the set A-hull(C) with the face F is colored
in gray.

Proof of Proposition 16.19 First, we prove that

A-hull(F ) ⊕ F ⊂ A-hull(C) ∩ F.
Let x ∈ A-hull(F ) ⊕ F . This means that x = p + v, where p is a point of the
closure of the A-hull of F and v is a vector contained in the face F . If v = 0, then
x coincides with p, which is in A-hull(F ) ⊂ A-hull(C). Suppose now that v
= 0.
For every positive integer N we choose an integer point pN = (a1 (N ), . . . , an (N ))
in C \ F such that
 
 v pN − p  1

 |v| − |p − p|  < N .
N
(It is always possible to choose such point, to be sufficiently far from the origin,
which we leave as an exercise for the reader.) Then the segment with endpoints p
and pN is contained in the A-hull of C. Set
|v|
ṽN = (pN − p).
|pN − p|
By construction we have
p + vN ∈ A-hull(C).
Then
v = lim ṽN .
N →∞
Hence
p + v = p + lim ṽN ∈ A-hull(C) ∩ F.
N →∞

Therefore, we get
A-hull(F ) ⊕ F ⊂ A-hull(C) ∩ F.
226 16 Multidimensional Continued Fractions in the Sense of Klein

Second, we prove that

A-hull(C) ∩ F ⊂ A-hull(F ) ⊕ F.
Let x be not in A-hull(F ) ⊕ F . Then there exists a supporting plane P d−1 ⊂
Span(F ) of the sum A-hull(F ) ⊕ F that separates x and A-hull(F ) ⊕ F (by Propo-
sition 16.11 the sum A-hull(F ) ⊕ F is convex), where d is the dimension of the
plane Span(F ). In addition, this hyperplane can be chosen such that it does not
contain any vector of F . Then the intersection P d−1 ∩ F is a (d − 1)-dimensional
simplex. Denote its vertices by v1 , . . . , vd . Then, first, the pyramid Ov1 . . . vd inter-
sects A-hull(F ) ⊕ F only in its base v1 . . . vd , and second, the pyramid contains x
in the interior. Let us choose the vertices vd+1 . . . vn on the edges of the cone C that
are not adjacent to the face F sufficiently close to the origin such that all the integer
points other than the origin (if any) of the pyramid Ov1 . . . vn are contained in the
base v1 . . . vn . This implies that the set A-hull(C) intersects the pyramid only in its
base. Since x is not in the base v1 . . . vn , it is not contained in A-hull(C) ∩ F .
So we have proved the first equality of the proposition.
Suppose now that the dimension of the lattice contained in F equals the dimen-
sion of F itself. Let us prove
A-hull(F ) ⊕ F = A-hull(F ).
It is sufficient to prove that for an arbitrary point x in the interior of A-hull(F ) and
vector v ∈ F the ray with vertex at x and direction v is in A-hull(F ). Since the
integer lattice is of full rank in F , for any edge ek of F there exist a sufficiently
large constant Ck and a sequence xm (k) of distinct integer points inside F with
distances to this edge not greater than Ck whose norm tends to infinity. Then, start-
ing from some m, the ray with vertex at x and direction v intersects all simplices
(xm (1), . . . , xm (d)), where d = dim F . Hence this ray is in the convex hull of x
and all the elements xm (1), . . . , xm (d) for m > 0. Therefore, this ray is a subset of
A-hull(F ). 

16.5.2 Homeomorphic Types of Sails

The homeomorphic types of sails are described by the following theorem.

Theorem 16.21 Every sail of an n-dimensional simplicial cone in Rn is homeo-

morphic to Rn−1 .

Let us prove the following lemma.

Lemma 16.22 Let C be an n-dimensional simplicial cone in Rn and let x be an

interior point of its A-hull. Then the shifted cone {x} ⊕ C is contained in the interior
of A-hull(C).
16.5 Polyhedral Structure of Sails 227

Fig. 16.5 The set Sx \ Cx

and a point xr

Proof Since x is an interior point of A-hull(C), there exists a ball with center at x
contained in A-hull(C). Consider a point x1
= x in this ball such that x − x1 ∈ C.
Then the cone {x} ⊕ C is in the interior of the cone {x1 } ⊕ C. By Proposition 16.19
in the case of F = C we have
A-hull(C) = A-hull(C) ∩ C = A-hull(C) ⊕ C.
Hence,
{x1 } ⊕ C ⊂ A-hull(C).
Therefore, the cone {x} ⊕ C is contained in the interior of the A-hull of the
cone C. 

Proof of Theorem 16.21 Let C be a simplicial n-dimensional cone in Rn . Consider

an arbitrary interior point x of the A-hull of C and take an (n − 1)-dimensional
sphere Sx with center at x that is completely contained in the A-hull. Denote by
Cx the closed cone C ⊕ {x}. Project the sail to the sphere Sx along the radial rays
with vertex x. Let us show that the projection is bijective with the set Sx \ Cx (see
Fig. 16.5).
First, by Lemma 16.22 the cone Cx is in the interior of the A-hull of C, and hence
there are no boundary points of A-hull (i.e., of the sail) that project to Sx ∩ Cx .
Second, we show the surjectivity. Each ray with vertex at x and intersecting
Sx \ Cx intersects the complement to the cone, which is not in the closure of the
A-hull. Hence this ray contains at least one point of the boundary of the A-hull (i.e.,
of the sail). Therefore, the projection of the sail to the set Sx \ Cx is surjective.
Finally, we prove the injectivity. Consider a ray r with center at x and intersecting
the set Sx \ Cx . Let xr be the last boundary point of the closure of the A-hull, i.e., all
the other points of the closure of the A-hull are in between x and xr (see Fig. 16.5).
Denote by Pr the convex hull of the set {xr } ∪ Sx The set A-hull(C) is convex and
contains the point xr and the sphere Sx . Therefore,
Pr ⊂ A-hull(C).
228 16 Multidimensional Continued Fractions in the Sense of Klein

Hence all points of the ray r between x and xr are interior points of the set
A-hull(C), and hence they are not on the sail (which is the boundary of A-hull(C)).
Therefore,
 
∂ A-hull(C) ∩ r = {xr }.
Therefore, the projection is injective.
We have shown that there is a one-to-one projection sending the sail to Sx \ Cx .
On the one hand, the projection of the sail is continuous. On the other hand, the
inverse to the projection is also continuous (this follows directly from the convexity
of the sail; we leave this as an exercise for the reader). The set Sx \ Cx is homeo-
morphic to an open (d − 1)-dimensional disk (and hence homeomorphic to Rd−1 ),
and therefore, the sail is homeomorphic to Rd−1 . 

16.5.3 Combinatorial Structure of Sails for Cones in General

Position

In this book we mostly discuss sails for cones that are either integer (i.e., each edge
has an integer point distinct from the origin) or in general position. The combina-
toric structure of the sails for integer cones is rather simple and was discussed in
Sect. 16.3 above. Let us examine the situation for sails in general position.

16.5.3.1 Theorem on Closeness of Sails for Cones in General Position

First, we give the definition of a cone in general position.

Definition 16.23 A simplicial n-dimensional cone in Rn is said to be in general

position if
– none of its faces contains integer points;
– each of its edges contains a vector whose integer approximation space coincides
with Rn .

It turns out that for cones in general position the set A-hull(C) is closed.

Theorem 16.24 Let a simplicial n-dimensional cone C be in general position. Then

the set A-hull(C) is closed.

We begin the proof of Theorem 16.24 with two lemmas.

Lemma 16.25 Let a simplicial n-dimensional cone C be in general position. Then

A-hull(C) ∩ ∂C = ∅.
16.5 Polyhedral Structure of Sails 229

Proof Consider an arbitrary face F of the cone C. From Proposition 16.19 we have
A-hull(C) ∩ F = A-hull(F ) ⊕ F = ∅ ⊕ F = ∅.
Hence the intersection of A-hull(C) with the union of all faces is empty as well. 

Lemma 16.26 Let C be a simplicial n-dimensional cone and let F be a face of

the sail of C. Consider an interior point p of C contained in F . Suppose that π
is a supporting plane of the set A-hull(C) passing through p. Then we have the
following:
(i) If π ∩ C is not compact, then C is not in general position.
(ii) If π ∩ C is compact, we have p ∈ A-hull(F ).

Proof
(i) We prove the statement by contradiction. Assume that C is in general posi-
tion. The hyperplane π divides the cone into two connected components, one
of which, say C1 , does not contain integer points in the interior (since it is a
supporting plane of the of A-hull(C)). By construction, the set C1 is not com-
pact, and hence its intersection with one of the edges of the cone C is a ray.
We assume that this edge contains a vector u whose approximation space is Rn .
Now consider any interior point x of the set C1 . On the one hand, there exists a
positive ε such that the tubular neighborhood of the edge shifted to x:
B(x, ε) ⊕ {ku | k ∈ R+ },
is completely contained in C1 . Therefore, by construction,
B(x, ε) ⊕ {ku | k ∈ R+ } ∩ Zn = ∅.
On the other hand, by the multidimensional Kronecker’s approximation theo-
rem (Theorem 16.16(iii)), this tubular neighborhood contains infinitely many
integer points. We arrive at a contradiction. Hence Lemma 16.26(i) holds.
(ii) Let π ∩ C be compact. Consider a sequence of points pi ∈ A-hull(C) that con-
verges to p. By definition, A-hull(C) is the convex hull of a subset of inte-
ger points. Hence by Carathéodory’s theorem, for every i there exists an inte-
ger simplex Δi ⊂ A-hull(C) containing the point pi . Since π is a supporting
plane, the distance from one of the vertices of Δi to the plane π is smaller than
|p − pi |. Since Zn is discrete and π ∩ C is compact, there exists N such that
for every i > N every simplex Δi has an integer vertex in π ∩ C. Since π ∩ C
is compact, there are only finitely many integer points in π ∩ C. Hence we can
choose an infinite subsequence p̃i in the sequence pi all of whose simplices Δ̃i
have the same integer vertex v1 in the plane π .
If v1 = p, then we are done. In case v1
= p, for every Δ̃i we can choose a
vertex wi
= v1 such that the distance from wi to the plane π tends to zero as
i tends to infinity. This is due to the compactness of π ∩ C. Therefore, there
exists Ñ such that all wi belong to π ∩ C for i > N . Since the set
(π ∩ C) ∩ Zn
230 16 Multidimensional Continued Fractions in the Sense of Klein

is finite, we can choose an infinite subsequence Δ̂i of Δ̃i whose vertices wj

coincide with an integer point of π , say with v2 .
So each simplex Δ̂i contains an integer edge v1 v2 in the plane π . We con-
tinue iteratively constructing an infinite subsequence of Δi with a common face
v1 . . . vk ⊂ π . Either at some step k we have p ∈ v1 . . . vk , or we reach k = n,
and therefore, there exists an infinite subsequence of (Δi ) all of whose elements
contain the simplex v1 . . . vn in π as a subset, and the last vertex is at positive
integer distance from π . In the latter case,
p ∈ v1 . . . vn = v1 . . . vn .
Therefore, the point p is in A-hull(C). 

Proof of Proposition 16.24 Consider an arbitrary point p of the sail. By Lem-

ma 16.25, p is an interior point of the cone C. Let π be a supporting plane contain-
ing p. By Lemma 16.26(i), since C is in general position, the set π ∩ C is compact.
Then by Lemma 16.26(ii) we have p ∈ A-hull(F ). Therefore, the set A-hull(F )
contains its boundary. 

16.5.3.2 Structural Theorem on Cones in General Position

Let us give a classical definition of extremal points and extremal rays.

Definition 16.27 A vertex in a convex set is said to be extremal if it is not contained

in the interior of any segment in the set.
An edge (or a ray) contained in a convex set is said to be extremal if every seg-
ment of this convex set either does not intersect the ray (edge) or at least one of its
endpoints is in the ray.

Remark 16.28 Notice that not every extremal vertex of a convex set S is a vertex
of S. For example, consider the union of the half-disk {x 2 + y 2 ≤ 1 | y ≥ 0} and the
square with vertices (1, 0), (1, −2), (−1, −2), and (−1, 0). Here the points (±1, 0)
are extremal vertices and yet they are not vertices (i.e., 0-dimensional faces).

Now we formulate the main structural theorem on cones in general position (see
also [125]).

Theorem 16.29 Let a simplicial n-dimensional cone C be in general position. Then

we have
(i) all faces of the sail for C are compact integer polyhedra;
(ii) the set of all vertices of the sail is discrete;
(iii) the sail does not contain rays;
(iv) each vertex of the sail is adjacent to only finitely many extremal edges.
16.5 Polyhedral Structure of Sails 231

Proof We begin with Theorem 16.29(iii), proving it by contradiction. Let a cone C

satisfy the conditions of the theorem and let its sail contain a ray r. Consider any
support hyperplane of A-hull(C) containing this ray. This hyperplane divides the
cone into two connected components, one of which, say C1 , does not have integer
points in the interior (since it is a supporting plane of A-hull(C)). By construction,
C1 is not compact, and hence by Lemma 16.26(i) the cone C is not in general
position. We have arrived at a contradiction. Hence Theorem 16.29(iii) holds.
Let us prove Theorem 16.29(i). Suppose, to the contrary a face F is not compact.
Hence there exists a sequence of points pi in F , that increase in norm to infinity.
Consider the sequence of points pi where
p1 p i
p i = p1 + .
|p1 pi |
All the elements of this sequence are points in the unit sphere centered at p1 . Since
the sphere is compact, there exists at least one accumulation point p in the sphere.
Hence the ray with vertex at p1 and direction p1 p is entirely contained in F =
F . Therefore, by Theorem 16.29(iii) the cone C is not in general position. This
contradicts the condition of the theorem. Hence the face F is compact.
By Proposition 16.8 the face F is the convex hull of a subset of Zn . Compactness
of F implies the finiteness of Zn ∩ F . Therefore, F is a compact integer polyhedron.
This concludes the proof of Theorem 16.29(i).
Let us prove Theorem 16.29(iv). Suppose the statement is false, and a vertex p
is adjacent to an infinite number of extremal edges. Since Zn is discrete, the length
of such edges is not universally bounded. Hence the closure of the union of all these
extremal edges contains a ray. Since all edges are extremal, this ray is in the sail.
This contradicts Theorem 16.29(iii).
Finally, Theorem 16.29(ii) holds, since the integer lattice is discrete. 

16.5.4 A-Hulls and Quasipolyhedra

To conclude this section we mention just one result on the quasipolyhedral structure
of sails.

Definition 16.30 A convex set in Rd is called quasipolyhedral if its intersection

with every polytope is polyhedral.

In some of the literature, quasipolyhedral sets are called generalized polyhedra.

They were studied for the first time by V.L. Klee in [107] in the framework of
separation theory. Let us list several properties describing quasipolyhedral sets.

Theorem 16.31 A convex set P is quasipolyhedral if and only if the following three
conditions are satisfied:
– P is closed;
232 16 Multidimensional Continued Fractions in the Sense of Klein

– the set of extremal points is locally finite;

– for every extremal point p, the set of all edges and extremal rays meeting at p is
finite.

Let us apply this theorem to sails.

Corollary 16.32 Consider a simplicial n-dimensional cone C in Rn . The closure

of the A-hull of C is a quasipolyhedral set if every only if any face of C containing
a nonzero integer point spans an integer affine space (i.e., a space with full-rank
integer sublattice in it).

For more information we refer to [125].

16.6 Two-Dimensional Faces of Sails

We conclude this chapter with a collection of some results and questions on two-
dimensional faces of continued fractions. We distinguish two essentially different
situations: questions related to faces that are a unit integer distance from the origin,
and questions related to faces that are integer distance from the origin greater than
one.

16.6.1 Faces with Integer Distance to the Origin Equal One

All two-dimensional faces of the sails of continued fractions are convex integer
polygons. If the integer distance to the origin is 1, then the only restriction for a
polygon P to be a face of an n-dimensional continued fraction is whether it is pos-
sible to inscribe P in some (n + 1)-gon Q (not necessarily integer!) such that the
convex hull of all integer points in Q coincides with P .

Proposition 16.33 Every n-gon is realizable as a face of an m-dimensional contin-

ued fraction if m ≥ n − 1.

Proof It is clear that every integer n-gon P can be inscribed in an m-gon Q, for
m ≥ n, such that the convex hull of all integer points in Q coincides with P . We
leave the details of the proof to the reader. 

So the main question arises when m < n − 1.

Problem 13 Which n-gons are not realizable as faces of an m-dimensional contin-

ued fraction?
16.6 Two-Dimensional Faces of Sails 233

Fig. 16.6 One of the vertices

of T is in the shaded (open)
angle

The answer to this question is not understood even in the case m = 2.

Let us introduce a family of examples of quadrangles that are not realizable as
faces of two-dimensional continued fractions.

Proposition 16.34 For arbitrary integers b ≥ a ≥ 1, the quadrangle with vertices

(−1, 0), (−a − 1, 1), (−1, 2), (b − 1, 1) cannot be a face of a two-dimensional
continued fraction.

Proof Suppose that the statement is false. Let there exist a two-dimensional con-
tinued fraction that has one of the compact faces (say F ) integer equivalent to a
quadrangle with vertices (−1, 0), (−a − 1, 1), (−1, 2), (b − 1, 1) for some integers
b ≥ a ≥ 1. Let π be the plane of the face. Let us introduce integer coordinates on
the plane π such that the coordinates of the vertices of the face F are exactly (a, 0),
(0, 1), (−b, 0), and (0, −1). Notice that a point in this plane is integer (as a point in
R3 ) if and only if it is integer in the new coordinates in π .
Every two-dimensional simplicial cone defining a two-dimensional continued
fraction has exactly three faces of dimension 2. So the intersection of the edges of
the cone with the plane π is a triangle, which we denote by T . Since F is a face of
the corresponding continued fraction we have, first, that the face F is contained in
the interior of the triangle T and, second, that the set T \ F does not contain any
integer points. Notice that the points (1, 1) and (1, −1) are not in F , and the point
(1, 0) is in F . In addition, the points (1, 0), (1, 1), and (1, −1) are contained in a
straight line. This implies that the open angle (denoted by ϕ) with vertex O(0, 0)
and edges passing through the points (1, 1) and (1, −1) respectively, contains at
least one vertex of the triangle T (see Fig. 16.6).
For two adjacent angles to ϕ and for the opposite angle ϕ, a similar statement
holds. Therefore, each of these four nonintersecting angles has a vertex of T , which
is impossible, since T is a triangle. We have arrived at a contradiction. 

All the questions discussed in this subsection are related to the two-dimensional
case. Similar questions are also interesting for the multidimensional case as well,
though it is clear that the complexity rises with the dimension.
234 16 Multidimensional Continued Fractions in the Sense of Klein

16.6.2 Faces with Integer Distance to the Origin Greater than One

We showed in the previous section that every two-dimensional face is realizable as

a face of some continued fraction (probably of a large dimension). It turns out that
the majority of such faces are a unit integer distance from the origin. There is a
complete description of all integer faces that occur at integer distance greater than
one.

Theorem 16.35 Every two-dimensional bounded face an integer distance greater

than one from the origin forms a multistory completely empty pyramid.

Therefore, the list of such faces is a sublist of all multistory two-dimensional

continued fractions (see Theorem 15.5).

Corollary 16.36
(i) The list of faces at integer distance greater than 1 for k-dimensional contin-
ued fractions for k > 2 coincides with the list of bases of multistory marked
pyramids.
(ii) If k = 2, the list of all admissible faces coincides with the list of bases of trian-
gular multistory marked pyramids.

Proof Existence of all admissible faces follows directly from Theorem 15.5 and
Proposition 16.33.
Quadrangular faces are not realizable in the two-dimensional case due to Propo-
sition 16.34. 

We conclude this chapter with the following open question.

Problem 14 Which three-dimensional polytopes are realizable as three-

dimensional faces of sails at integer distance greater than 1.

16.7 Exercises

Exercise 16.1 Prove the one-dimensional Kronecker’s approximation theorem: if θ

is an arbitrary irrational number, then the sequence of numbers {kθ } (for k > 0) is
dense in the unit interval.

Exercise 16.2 Construct the sail for the cone with vertex at the origin and generated
by integer vectors (1, 1, 1), (−1, 2, 4), and (2, 1, 4).

Exercise 16.3
(a) What is the sum of a round disk and a quadrangle?
16.7 Exercises 235

(b) Suppose the sum of a convex k-gon and a convex l-gon is an m-gon. For which
triples (k, l, m) is this possible?
(c) Suppose the sum of a convex k-dimensional polyhedron and an l-dimensional
polyhedron is an m-dimensional polyhedron. For which triples (k, l, m) is this
possible?

Exercise 16.4 Let C be a simplicial cone with the vertex at the origin and let F be
one of its faces. Consider a point p in F and a vector v of F . Prove that for every
positive integer N there exists an integer point pn = (a1 (N ), . . . , an (N )) in C \ F
such that
 
 v pN − p  1
 −
 |v| |p − p|  < N .
N

Exercise 16.5 Every n-gon is realizable as a face of an m-dimensional continued

fraction if m ≥ n − 1.

Exercise 16.6 For an arbitrary n construct an n-gon that is realizable as a face of a

two-dimensional continued fraction.
Chapter 17
Dirichlet Groups and Lattice Reduction

Let us study the structure of Dirichlet groups, which is actually the main reason for
the periodicity of algebraic multidimensional sails (see Chap. 18). The simplest case
of two-dimensional Dirichlet groups was studied in Chap. 8 in the first part of this
book. In this chapter we study the general multidimensional case.
We start with Dirichlet’s unity theorem on the structure of the group of units in
orders in Sect. 17.1. Further, in Sect. 17.2 we describe the relation between Dirich-
let groups and groups of units. In Sects. 17.3 and 17.4 we show how to calculate
bases of the Dirichlet groups and positive Dirichlet groups respectively. Finally, in
Sect. 17.5 we briefly discuss the LLL-algorithm on lattice reduction which helps to
decrease the computational complexity in many problems related to lattices (includ-
ing the calculation of Dirichlet group bases).

17.1 Orders, Units, and Dirichlet’s Unit Theorem

In this subsection we give some preliminary material on the field theory related to
the structure of the units in orders.

Definition 17.1 Let K be a field of algebraic numbers and μ1 , . . . , μm an arbitrary

finite sequence of numbers in K. The set M of all linear combinations c1 μ1 + · · · +
cm μm with integer coefficients c1 , . . . , cm is called a module of the field K. The
numbers μ1 , . . . , μm are called the generators of the module M.

Every module M in an algebraic field contains a basis, i.e., a sequence

μ1 , . . . , μm that generates M as linear combinations with integer coefficients such
that only the trivial combination represents zero. It is clear that the number of ele-
ments in the basis coincides with the number of linearly independent elements (over
Q) in M.

O. Karpenkov, Geometry of Continued Fractions, 237

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_17, © Springer-Verlag Berlin Heidelberg 2013
238 17 Dirichlet Groups and Lattice Reduction

Definition 17.2 Let K be a field of algebraic numbers of degree n and let M be a

module in K. If M contains n linearly independent (over Q) numbers, then it is said
to be complete, otherwise it is called incomplete.

Definition 17.3 A complete module in a field of algebraic numbers K is an order

in K if it is a ring containing 1.

Example 17.4 Let K be an algebraic number field. The set of all numbers in K
whose minimal polynomial has integer coefficients is an order in K. This order is
often called maximal, since every other order of K is contained in this order.

Definition 17.5 Let D ⊂ K be an arbitrary order. A number ε in D is a unit if ε−1

is contained in D.

Theorem 17.6 (Dirichlet’s unit theorem) Let K be a field of algebraic numbers of

degree n = s + 2t. Consider an arbitrary order D in K. Then D contains units
ε1 , . . . , εr for r = s + t − 1 such that every unit ε in D has a unique decomposition
of the form
ε = ξ ε1a1 · · · εrar ,
where a1 , . . . , ar are integers and ξ is a root of 1 contained in D.

We refer to [22] for a proof of this theorem.

17.2 Dirichlet Groups and Groups of Units in Orders

Let us describe the relation between Dirichlet groups and groups of units in orders
in algebraic number fields.

17.2.1 Notion of a Dirichlet Group

Consider an arbitrary integer matrix A with characteristic polynomial irreducible

over Q.

Definition 17.7 Denote by Γ (A) the set of all integer matrices commuting with A.
(i) The Dirichlet group Ξ (A) is the subset of invertible matrices in Γ (A).
(ii) The positive Dirichlet group Ξ+ (A) is the subset of Ξ (A) that consists of all
matrices with positive real eigenvalues.

The matrices of Γ (A) form a ring with standard matrix addition and multiplica-
tion. (As a group, Γ (A) is isomorphic to Zn+1 .)
17.2 Dirichlet Groups and Groups of Units in Orders 239

17.2.2 On Isomorphisms of Dirichlet Groups and Certain Groups

of Units

Consider an arbitrary matrix A with characteristic polynomial χA irreducible

over Q. Denote by Q[A] the set of all matrices that are polynomials in A with ratio-
nal coefficients. It is clear that Q[A] is a ring contained in Span(Id, A, . . . , An−1 ).
This ring is naturally isomorphic to quotient ring Q[x]/χA (x).
Consider an eigenvalue ξ of A. It is clear that the field Q[ξ ] is also isomorphic
to Q[x]/χA (x). Hence there exists a natural isomorphism

hA,ξ : Q[A] → Q[ξ ],

sending p(A) to p(ξ ) for every element p of the quotient ring Q[x]/χA (x). Since
Ξ (A) is contained in Q[A], we have an embedding:

hA,ξ : Ξ (A) → Q[ξ ].

In the proofs of Proposition 17.8 and Theorem 17.9 we use the following notation:
 
PA = p ∈ Q[x]/χA (x) | p(A) ∈ Mat(n, Z) .

From the definitions it is clear that

 
hA,ξ Γ (A) = p(ξ ) | p ∈ PA .

Proposition 17.8 Consider an arbitrary integer matrix A with characteristic poly-

nomial irreducible over Q. Let ξ be one of eigenvalues of A. Then the set
hA,ξ (Γ (A)) is an order in Q(ξ ).

Proof First, the set PA described above is closed under addition and multiplication,
since if p1 (A) and p2 (A) are integer matrices, then (p1 + p2 )(A) and (p1 p2 )(A)
are also integer matrices. Hence hA,ξ (Γ (A)) is closed under addition and multi-
plication. Since Q[A] is a ring (even a field), the addition and multiplication in
hA,ξ (Γ (A)) satisfy all ring axioms. Hence hA,ξ (Γ (A)) is a ring.
Second, as an additive group, hA,ξ (Γ (A)) is a subgroup of Mat(n, Z) isomorphic
to the free abelian group Zn×n , which implies that it is a free abelian group with at
most n2 generators, and hence it is a finite module.
Third, this module contains the identity matrix Id.
And finally, the elements Id, ξ, ξ 2 , . . . , ξ n−1 are linearly independent over Q,
and hence the set hA,ξ (Γ (A)) is a complete module. Therefore, by definition,
hA,ξ (Γ (A)) is an order. 

Theorem 17.9 Consider an arbitrary integer matrix A with characteristic polyno-

mial irreducible over Q. Let ξ be one of the eigenvalues of A. Then the Dirich-
let group Ξ (A) is isomorphic to the multiplicative group of units in the order
hA,ξ (Γ (A)) of the field Q(ξ ). The isomorphism is realized by the map hA,ξ as
above.
240 17 Dirichlet Groups and Lattice Reduction

Proof Recall that hA,ξ is a one-to-one map between Ξ (A) and hA,ξ (Ξ (A)).
Letting p(ξ ) ∈ hA,ξ (Γ (A)) be a unit, i.e., it is invertible, then there exists an
element q ∈ P (A) such that p(ξ ) · q(ξ ) = 1. Hence q = p −1 in the quotient ring
Q[x]/χA (x), and
p(A) · q(A) = Id.
Therefore, the matrix p(A) is invertible in Mat(n, Z), and hence it is in Ξ (A).
Suppose now that p(A) ∈ Ξ (A) has an inverse in Ξ (A) that is q(A). Let us
restrict these matrices to the eigenspace corresponding to ξ . We have directly p(ξ ) ·
q(ξ ) = 1. Thus, p(ξ ) is a unit.
Hence, the Dirichlet group Ξ (A) is isomorphic to the multiplicative group of
units in the order hA,ξ (Γ (A)), and the isomorphism is given by hA,ξ . 

From Dirichlet’s unit theorem (Theorem 17.6) and Theorem 17.9 we have the
following description of the Dirichlet group Ξ (A).

Corollary 17.10 Consider an arbitrary integer matrix A with characteristic poly-

nomial irreducible over Q. Suppose it has s real and 2t complex conjugate eigen-
values. Then there exists a finite abelian group G such that

Ξ (A) = G ⊕ Zs+t−1 .

17.2.3 Dirichlet Groups Related to Orders That Do not Have

Complex Roots of Unity

A polynomial is called cyclotomic if its roots are all distinct primitive nth roots of
unity for some integer n.
A cyclotomic polynomial is an irreducible polynomial. It is usually denoted by
Φn . These polynomials form only a very small subset of all polynomials. For in-
stance, for all n > 2, the degrees of Φn are even, hence if we study matrices with
irreducible characteristic polynomials in odd dimensions, then we do not have any
orders containing roots of unity other than ±1. In fact, the degree equals the Euler
totient φ(n) (the number of positive integers less than or equal to n that are coprime
to n). The sequence of Euler totients (φ(n)) converges to infinity as n tends to in-
finity. Hence for each even d there are only finitely many cyclotomic polynomials
of any fixed degree d.
If the Dirichlet group Ξ (A) related to orders does not have complex roots of
unity, then we directly have

Ξ (A) = (Z/2Z)k ⊕ Zs+t−1

for some integer k. Since the matrix −Id is in Ξ (A), we have k ≥ 1.

17.3 Calculation of a Basis of the Additive Group Γ (A) 241

In these cases, for the positive Dirichlet group, we have Ξ+ (A) = Zs+t−1 , since
all symmetries about the plane have negative eigenvalues and every subgroup of a
free abelian group is free abelian.
In the special case that all the eigenvalues of A are real we have the following
statement (we call such a matrix a real spectrum matrix).

Proposition 17.11 Consider an arbitrary integer real spectrum matrix A ∈

SL(n, Z) with characteristic polynomial irreducible over Q. Then Ξ+ (A) = Zn−1 .

Proof Every generator of Ξ+ (A) is a matrix with positive eigenvalues. Therefore,

the operator is not cyclic. Hence Ξ+ (A) is a free abelian group. Since for every
B ∈ Ξ (A) we have B 2 ∈ Ξ+ (A), the rank of Ξ+ (A) is n − 1. 

17.3 Calculation of a Basis of the Additive Group Γ (A)

In this subsection we show how to calculate the basis of an additive group Γ (A)
for an arbitrary integer matrix A with irreducible characteristic polynomial. Let us
2
identify the space Mat(n, R) with the space R(n) and consider the standard metrics
for this space (the distance between two matrices is the Euclidean distance between
2
the corresponding points in Rn ). Then every integer matrix corresponds to some
2
integer point of Rn . In Proposition 17.12 below we show that the group Γ (A) is
isomorphic to Z . Further, we construct a basis of a maximal-rank subgroup in
n+1

Γ (A). Finally, we choose a basis in the parallelepiped generated by this basis. So

the algorithm can be performed as follows.

Algorithm to Calculate a Basis of Γ (A)

Input data. In the input we have an arbitrary integer matrix A with irreducible
characteristic polynomial.
Goal of the algorithm. To calculate a basis of the group Γ (A).
Step 1. Since matrices Id, A, A2 , . . . , An are linearly independent over Q, they
form a basis of some sublattice of Γ (A).
Step 2. To reduce all the calculations we apply the LLL-algorithm to the basis of
Step 1 to get a reduced basis with much smaller coefficients.
Step 3. Calculate a basis of Γ (A) starting with the basis of a sublattice of Γ (A)
constructed in Step 2.
Output. The basis of Γ (A).

17.3.1 Step 1: Preliminary Statements

Proposition 17.12 For every integer irreducible real spectrum matrix A the set
Γ (A) forms an additive group isomorphic to Zn+1 .
242 17 Dirichlet Groups and Lattice Reduction

Proof In the diagonal basis, the group of all matrices commuting with A is iso-
morphic to Rn+1 by addition. Hence the set of all integer matrices forms an integer
lattice in this (n + 1)-dimensional subspace. So the group is isomorphic to Zk with
k ≤ n + 1.
The matrices Id, A, A2 , . . . , An are linearly independent over Q since the char-
acteristic polynomial of A is irreducible over Q. Hence, k = n + 1. 

Corollary 17.13 The group Γ (A) is the intersection of the integer lattice Zn×n ⊂
Mat(n, R) with the space Span(Id, A, A2 , . . . , An ).

17.3.2 Step 2: Application of the LLL-Algorithm

Applying the LLL-algorithm described below to the sublattice generated by the ma-
trices Id, A, A2 , . . . , An , one constructs a reduced basis. This will decrease the Eu-
clidean lengths of basis vectors generating the sublattice (see Sect. 17.5 for the full-
rank lattice algorithm and Remark 17.18 for the non-full rank lattice algorithm). In
general it is not necessary to use the LLL-algorithm here, so that one can proceed
directly to Step 3.

17.3.3 Step 3: Calculation of an Integer Basis Having a Basis of

an Integer Sublattice

In Step 2 we end up with a basis M1 , . . . , Mn+1 of a full-rank sublattice in the space

Span(Id, A, A2 , . . . , An ). Let us show how to calculate effectively a basis of the
integer lattice in the span.
We do this inductively.
Base of induction. We take M = λ1 M1 as the first vector of the basis, where λ is
the greatest common divisor of all the element in M1 .
Step of induction. Suppose we have calculated the integer basis M 1 , . . . , M k−1
in Span(M1 , . . . , Mk−1 ). Let us extend it with M k to a basis of an integer lattice in
Span(M1 , . . . , Mk ). As M k we choose the only integer vector of the following form:

1  λi
k−1
Mk = Mk + Mi,
dk di
i=1

where di is the integer distance from the ith point of the set M 1 , . . . , M k−1 , Mk to
the span of the remaining points, and for integers λi we have 0 ≤ λi < di . Since
 
ld M k , Span(M 1 , . . . , M k−1 ) = 1,

the matrices M 1 , . . . , M k form a basis of the lattice in Span(M1 , . . . , Mk ).

17.4 Calculation of a Basis of the Positive Dirichlet Group Ξ+ (A) 243

The distances di are calculated by the formula of Theorem 14.9, and the inte-
ger volumes in coordinates are the greatest common divisors of the corresponding
modified Plücker coordinates (see Theorem 14.30). The numbers λi are found by
exhaustive search.
So in n + 1 steps we obtain a basis of Γ (A) that is the integer lattice in the space
Span(Id, A, A2 , . . . , An ) according to Corollary 17.13.

17.4 Calculation of a Basis of the Positive Dirichlet

Group Ξ+ (A)
From the algorithmic point of view this step is the most complicated. We describe
only the idea for one of the simplest algorithms here and give the corresponding
references.
Let χ(x) be the characteristic polynomial of the matrix A and let ξ be one of the
roots of χ(x). Consider the map

hA,ξ : Ξ (A) → Q[ξ ]

as in Sect. 17.2.2. By Theorem 17.9 this map is an isomorphism between the ring
Ξ (A) and image hA,ξ (Ξ (A)), which is the multiplicative group of units. In the book
[22] the authors show how to construct a basis for the units of an order and a number
ρ such that the norms of all its elements are bounded from above by ρ (here the norm
is the standard norm on the order considered as Rs+t ). The method of constructing
the constant ρ is standard and quite technical, so we omit it. Now according to [22]
we construct a basis by enumeration of all vectors of the set h(Ξ (A)) inside the ball
Bρ (O), where Bρ (O) is a ρ-neighborhood of the origin. The preimage (i.e., h−1 )
of this basis gives us a basis of the group of invertible elements in the ring Ξ (A),
and hence it gives a basis of the subgroup Ξ+ (A).

Remark 17.14 The constant ρ is extremely large (it equals the exponent of some
polynomial of the coefficients of the matrix A). An effective algorithm for this step
can be found in the book [35] by H. Cohen. Using this algorithm one finds the basis
of units in polynomial time (with respect to the coefficients of the matrix A).

17.5 Lattice Reduction and the LLL-Algorithm

Whenever we are facing computational aspects in terms of the coefficients in some
lattice basis, it is useful first to choose the basis in a way that it has vectors that
(in some standard norm) as small as possible and then to perform the calculation.
In 1982, A.K. Lenstra, H.W. Lenstra, and L. Lovász in [130] proposed a notion of
a reduced basis and an algorithm to calculate it. Later the algorithm was named
the LLL-algorithm in honor of its inventors. We present the original LLL-algorithm
244 17 Dirichlet Groups and Lattice Reduction

below and refer to [23] and [35] for further information. In this subsection we denote
the Euclidean length of a vector v by |v|.

17.5.1 Reduced Bases

Let b1 , . . . , bn be a basis for L. Define inductively


i−1 bi , bj∗ 
bi∗ = bi − μij bj∗ , where μij =
bj∗ , bj∗ 
j =1

(by v, w we denote the inner product of two vectors). Actually, the vectors
b1∗ , . . . , bn∗ are the resulting vectors in the Gram–Schmidt orthogonalization process
(and μi,j are the coefficients that arise in the orthogonalization process).

Definition 17.15 A basis b1 , b2 , . . . , bn for a lattice L is said to be reduced if the

following two conditions hold:
(i) Size reduced conditions: for 1 ≤ j ≤ i ≤ n, |μi,j | ≤ 12 .
(ii) Lovász conditions: for 1 < i ≤ n, |bi∗ + μi,i−1 bi−1
∗ |2 ≥ 3 b∗ .
4 k−1

Letting L be a lattice generated by b1 , . . . , bn , denote by d(L) the determinant

of the lattice:
d(L) = | det(b1 , . . . , bn )|.
Let us estimate the lengths of the vectors in a reduced basis via the orthogonal basis
bi∗ and the determinant of L.

Theorem 17.16 (A.K. Lenstra, H.W. Lenstra, and L. Lovász [130]) Let b1 , b2 , . . . ,
bn be a reduced basis for a lattice L in Rn . Then we have
(i) |bj |2 ≤ 2i−1 |bi∗ |2 for 1 ≤ j ≤ i ≤ n,
%
(ii) d(L) ≤ ni=1 |bi | ≤ 2n(n−1)/4 d(L),
(iii) |b1 | ≤ 2n(n−1)/4 d(L)1/n .

Proof From the size reduced conditions and the Lovász conditions, we have

 ∗ 2  ∗  1  ∗ 2
b  ≥ 3 − μ 2   b 
i i,i−1 bi−1 ≥
4 2 i−1

for all admissible i, and therefore,

 ∗ 2  
b  ≤ 2i−j b∗ 2 , for 1 ≤ j ≤ i ≤ n.
j i
17.5 Lattice Reduction and the LLL-Algorithm 245

From the size reduced conditions, it follows that

 2  i−1
 2  2  i−1
 2    2
|bi |2 = bi∗  + μ2i,j bj∗  ≤ bi∗  + 2i−j −2 bi∗  = 2i−2 +1/2 ≤ 2i−1 bi∗  .
j =1 j =1

This implies the first item of Theorem 17.16:

 2  2
|bj |2 ≤ 2j −1 bj∗  ≤ 2i−1 bi∗  .

By definition of the basis b1∗ , b2∗ , . . . , bn∗ , it follows that

   &n
 
d(L) = det b1∗ , b2∗ , . . . , bn∗  = bi∗ ,
i=1

the last equality holding since the vectors bi∗ are mutually orthogonal. Now the
inequalities of the second and third items of Theorem 17.16 follow directly from
the inequalities of the first item and the fact that |bi∗ | < |bi |. 

17.5.2 The LLL-Algorithm

One can think of this algorithm as a type of generalization of Euclid’s algorithm.

We use the following notation. Letting r be the nearest integer to |μk,l |, denote
by [bk ]l the vector bk − rbl .

LLL-Algorithm
Input data. As input we have a basis of the lattice bi . The additional orthogonal-
ization data (bi∗ , μi,j ) is defined by bi as above. The algorithm is iterative. At
each iteration we replace the basis bi by a new one. Whenever we change the
basis bi we should update the orthogonalization data (bi∗ , μi,j ).
Goal of the algorithm. To calculate a reduced basis of the lattice generated by the
basis bi .
Parameters and conditions of the iteration process. An iteration has one param-
eter k ∈ {1, 2, . . . , n + 1}. For the first iteration we put k = 2. The algorithm
terminates when k = n + 1. After each iteration we have the value of k and the
following conditions satisfied:
(ik ) For 1 ≤ j ≤ i < k, |μi,j | ≤ 12 .
(iik ) For 1 < i < k, |bi∗ + μi,i−1 bi−1
∗ |2 ≥ 3 b∗ .
4 k−1
Description of the iteration. We have the basis bi and the parameter k ∈
{1, 2, . . . , n} (if k = n + 1, the algorithm terminates). First, if k > 1, we replace
bk by [bk ]k−1 and respectively update (bi∗ , μi,j ). Now |μk,k−1 | ≤ 12 . Second, we
choose one of two cases to continue.
246 17 Dirichlet Groups and Lattice Reduction

Case 1. We choose this case if

 ∗ 
b + μi,i−1 b∗ 2 < 3 b∗ and k ≥ 2.
i i−1
4 k−1
We interchange bk−1 and bk and update (bi∗ , μi,j ). Finally, we reduce k by one
(k := k − 1) and exit the iteration.
Case 2. We choose this case if
 ∗ 
b + μi,i−1 b∗ 2 ≥ 3 b∗ or k = 1.
i i−1
4 k−1
Then we consequently replace bk by [bk ]l and respectively update (bi∗ , μi,j ) for
l = k−1, k−2, . . . , 1. Finally, we increase k by one (k := k + 1) and exit the
iteration.
Output. The algorithm terminates once k = n + 1. As output, we have a reduced
basis of the lattice.
It is important to notice that the algorithm terminates in finitely many iterations.
The running time is bounded as follows.

Theorem 17.17 (A.K. Lenstra, H.W. Lenstra, and L. Lovász [130]) Let L ⊂ Zn be
a lattice with basis b1 , b2 , . . . , bn , and let B ∈ R, B ≥ 2, be such that |bi |2 ≤ B for
1 ≤ i ≤ n. Then the number of arithmetic operations needed by the basis reduction
in the described LLL-algorithm is O(n4 log B), and the integers on which these
operations are performed each have binary length O(n log B).

We are not going to give a proof here, since it is quite technical. The interested
reader is referred to the original manuscript [130].

Remark 17.18 The algorithm works also for the case of lattices of rank l < n. In
this case it terminates when k reaches l + 1.

17.6 Exercises
Exercise 17.1 Prove that Γ (A), considered as a group with matrix addition, is iso-
morphic to Zn+1 .

Exercise 17.2 Find all isomorphism types of Dirichlet groups for matrices in
SL(n, Z) whose characteristic polynomials are irreducible over Q, where
(a) n = 2;
(b) n = 3;
(c) n = 4.

Exercise 17.3 Prove that cyclotomic polynomials are irreducible over Q and of
even degree (except for (x ± 1)).
17.6 Exercises 247

Exercise 17.4 Find the generators and relation of the Dirichlet group Ξ (A), where
⎛ ⎞
0 1 0
(a) A = ⎝0 0 1 ⎠ ;
1 2 −4
⎛ ⎞
0 1 0
(b) A = ⎝0 0 1⎠ .
1 −4 0
Chapter 18
Periodicity of Klein Polyhedra. Generalization
of Lagrange’s Theorem

The sails of algebraic multidimensional continued fractions possess combinatorial

periodicity due to the action of the positive Dirichlet group on the sails. In case
of one-dimensional geometric continued fractions this periodicity is completely
described by the periodicity of the corresponding LLS sequences (see Chap. 7).
Around twenty years ago, V.I. Arnold posed a series of questions to study peri-
odicity of multidimensional continued fractions in the sense of Klein. At this mo-
ment some of his questions have answers, while others remain open. In this chapter
we discuss current progress in this direction. The questions related to periodicity
of algebraic sails are important in algebraic number theory, since they are in cor-
respondence with algebraic irrationalities. In particular, periods of algebraic sails
characterize the groups of units in the corresponding orders.
First, we associate to any matrix with real distinct eigenvalues a multidimen-
sional continued fraction. Second, we discuss periodicity of associated sails in the
algebraic case. Further, we give examples of periods of two-dimensional algebraic
continued fractions and formulate several questions arising in that context. Then
we state and prove a version of Lagrange’s theorem for multidimensional contin-
ued fractions. Finally, we say a few words about the relation of the Littlewood and
Oppenheim conjectures to periodic sails.

18.1 Continued Fractions Associated to Matrices

We begin with the following constructive definition.

Definition 18.1 Consider an (n + 1) × (n + 1) matrix A whose eigenvalues are all

distinct and real. Take the n-dimensional hyperplanes passing through the origin that
are spanned by n linearly independent eigenvectors of A. There are exactly n + 1
such hyperplanes. These hyperplanes define the n-dimensional continued fraction
called the multidimensional continued fraction associated to A.

O. Karpenkov, Geometry of Continued Fractions, 249

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_18, © Springer-Verlag Berlin Heidelberg 2013
250 18 Periodicity of Klein Polyhedra. Generalization of Lagrange’s Theorem

Let us formulate an algebraic criterion of integer congruence of associated mul-

tidimensional continued fractions.

Proposition 18.2 Let A and B be matrices of GL(n + 1, R) with distinct real eigen-
values. The continued fractions associated to A and B are integer congruent if and
only if there exists a matrix X ∈ GL(n + 1, Z) such that XAX −1 commutes with B.

Proof Let A and B be matrices of GL(n + 1, R) with distinct real irrational eigen-
values and suppose that their continued fractions are integer congruent. Since
the continued fractions are integer congruent, there exists a linear integer lattice-
preserving transformation of the space that maps the continued fraction of A to the
continued fraction of B. Under such a transformation the matrix A is conjugated by
some integer matrix X with unit determinant. All eigenvalues of the matrix XAX −1
are distinct and real (since conjugations preserve the characteristic polynomial of
the matrix). Since the invariant cones of the first continued fraction map to the in-
variant cones of the second one, the sets of eigendirections for the matrices B and
XAX −1 coincide. Hence these matrices are simultaneously diagonalizable in some
basis, and therefore, they commute.
Let us prove the converse. Suppose that there exists X ∈ GL(n + 1, Z) such that

à = XAX −1

commutes with B. Note that the eigenvalues of the matrices A and à coincide.
Therefore, all eigenvalues of the matrix à (just as for the matrix B) are real and
distinct. Let us consider a basis in which the matrix à is diagonal. Simple verifi-
cation shows that the matrix B is also diagonal in this basis. Hence the matrices Ã
and B define the same orthant decomposition of Rn+1 , and thus the same continued
fraction as well. 

18.2 Algebraic Periodic Multidimensional Continued Fractions

Now we formulate the notion of periodicity of continued fractions associated to
algebraic irrationalities. Recall the following general definition.

Definition 18.3 An integer matrix A is called irreducible if its characteristic poly-

nomial is irreducible over Q.

We write RS-matrix for real spectrum matrices.

It is clear that every irreducible RS-matrix A has distinct eigenvalues. Hence
there exists an associated multidimensional continued fraction to A. We work now
only with integer matrices of the group SL(n + 1, Z).

Definition 18.4 An n-dimensional continued fraction associated to an irreducible

RS-matrix A ∈ SL(n + 1, Z) is called an n-dimensional continued fraction of an
18.3 Torus Decompositions of Periodic Sails in R3 251

(n + 1)-algebraic irrationality. The case n = 1 (2) corresponds to one- (two-) di-

mensional continued fractions of quadratic (cubic) irrationalities.

Consider an irreducible RS-matrix A ∈ SL(n + 1, Z). From Proposition 17.11

we know that the positive Dirichlet group Ξ+ (A) is isomorphic to Zn . Each matrix
of Ξ+ (A) preserves the integer lattice and the union of all n + 1 hyperplanes, and
hence it preserves the n-dimensional continued fraction. Since all eigenvalues are
positive, the sails are mapped to themselves as well. The group Ξ+ (A) acts freely
on any sail. The factor of a sail under such a group action is isomorphic to the
n-dimensional torus.

Definition 18.5 By a fundamental domain of the sail we mean the union of some
faces that contains exactly one face from each equivalence class.

In Fig. 18.1 we give an example of the periodic continued fraction associated to

the matrix of the two-dimensional golden ratio (later we study this example in more
detail; see Example 18.9 below):
⎛ ⎞
1 1 1
⎝1 2 2⎠ .
1 2 3

In the top picture we have a fragment of the sail. In the middle there are two triangles
of two different orbits (white ones and dark ones). Finally at the bottom we show
one of the possible fundamental domains with respect to the action of the positive
Dirichlet group.

18.3 Torus Decompositions of Periodic Sails in R3

As we have already seen in the first part of the book (Chap. 7), the LLS sequence
is a complete invariant of one-dimensional sails. The situation in the multidimen-
sional case is much more complicated, since the combinatorial structure of sails
is no longer as simple as in the one-dimensional case. A complete invariant is not
known even for two-dimensional sails in R3 . Further, in this section we discuss
only periodic two-dimensional continued fractions in R3 , recalling that such sails
are homeomorphic to R2 by Theorem 16.21.

Definition 18.6 A class of integer congruences of certain objects in Rn is called an

integer affine type of these objects.

Let us discuss the case of R3 . It is not very hard to get an invariant that distin-
guishes all sails. For instance, if we know
– integer affine types of all faces in the sail;
252 18 Periodicity of Klein Polyhedra. Generalization of Lagrange’s Theorem

Fig. 18.1 The sail of an

algebraic continued fraction
(top); its periodic structure
(middle); a fundamental
domain (bottom)

– integer affine types of the pyramids at all vertices of the sail;

– additional data about one vertex v0 : integer affine type of the union of the pyramid
and the origin,
then we can uniquely reconstruct the sail, if it exists.
Let us briefly show how to do this.

Algorithm of Sail Reconstruction

Input data. As mentioned above.
18.4 Three Single Examples of Torus Decompositions in R3 253

Goal of the algorithm. To construct the sail (actually some of its finite part, say a
certain fundamental domain for periodic sails).
Step 1. From the third condition we construct the pyramid at v0 of the sail.
Step 2. From the first condition we find all polygonal faces adjacent to v0 .
Inductive step. In previous steps we have found several faces of the sail. By the
first condition identify all integer affine types of pyramids that are adjacent to
the constructed set of faces (at least in two edges each). By the second condition
construct all the faces in new pyramids.
Output. In a finite or countable number of steps we build the whole sail (up to
integer congruence).

Remark 18.7 In the case of algebraic irrationalities all the integer affine types of the
faces and pyramids would repeat periodically. Hence we need to store only a finite
amount of data regarding a single period.

This data set is far from forming a complete invariant of sails. It has many mon-
odromy conditions, so the majority of cases are nonrealizable. So the natural ques-
tion here is as follows: how to find whether a certain data set is realizable?
Recently, V.I. Arnold proposed a weaker invariant to distinguish the periodic
sails.

Definition 18.8 A torus decomposition corresponding to a periodic two-

dimensional sail is the factor-torus face decomposition of the sail equipped with
integer affine types of faces and integer distances to them.

V.I. Arnold conjectured that torus decompositions of integer noncongruent sails

are distinct. For all checked integer noncongruent sails this conjecture is true. (See
Conjecture 16.) Here we should mention that for torus decompositions, the realiz-
ability problem is not yet studied.
In the multidimensional case one can define similar torus decompositions. The
important question is then to find which decompositions are realizable and which
are not realizable.

18.4 Three Single Examples of Torus Decompositions in R3

In this and the next section we study several examples. We construct one of the sails
for each example.

Example 18.9 We start with the simplest example that

√
generalizes the regular con-
tinued fraction corresponding to the golden ratio 1+2 5 . Recall that the geometric
continued fraction corresponding to the golden ratio is associated to the matrix

1 1
.
1 2
254 18 Periodicity of Klein Polyhedra. Generalization of Lagrange’s Theorem

All four one-dimensional sails of the one-dimensional continued fraction corre-

sponding to the golden ratio are integer congruent. The associated continued fraction
has sails whose LLS sequences are (. . . , 1, 1, 1, . . .). This sequence has period (1),
which is the simplest possible for one-dimensional sails. The corresponding circle
decompositions consist of one vertex and one edge.
The generalization of this one-dimensional continued fraction to the multidimen-
sional case was given by E.I. Korkina in the work [118]. The multidimensional con-
tinued fraction associated to the matrix
⎛ ⎞
1 1 1 ··· 1
⎜1 2 2 · · · 2⎟
⎜ ⎟
⎜1 2 3 · · · 3⎟
⎜. . . . ⎟
⎝ .. .. .. . . ... ⎠
1 2 3 ··· k
is called the generalized golden ratio.

Remark 18.10 Unfortunately, the generalized golden ratios do not always give the
simplest periodic continued fractions in all dimensions, since the characteristic poly-
nomials are not always irreducible. For instance, there are nonconstant factors if the
dimension is k = 4, 7, 10, 12, 13, 16, 17, 19 for k ≤ 20. Further, we study the case
k = 3, which corresponds to an irreducible matrix. The case of the simplest matrices
for k = 4 was studied in [92].

Consider the continued fraction associated to the generalized golden ratio matrix
⎛ ⎞
1 1 1
M = ⎝1 2 2⎠ .
1 2 3
The torus decomposition corresponding to this matrix is homeomorphic to the fol-
lowing one:

Here the segment AB is identified with the segment DC and the segment AD with the
segment BC. In the picture we show only homeomorphic types of the faces (without
lattice structure). The integer affine types of the corresponding faces are given in the
next picture; both triangles have integer affine types of the simplest triangle:
18.4 Three Single Examples of Torus Decompositions in R3 255

The positive Dirichlet group Ξ+ (M) is generated by the two matrices

⎛ ⎞ ⎛ ⎞
1 1 1 1 0 1
⎝1 2 2⎠ and ⎝0 2 1⎠ .
1 2 3 1 1 2

The integer distance from the triangle ABD to the origin equals 2, and from the
triangle BCD it equals 1.
This torus decomposition was found by E.I. Korkina [118], G. Lachaud [125],
and A.D. Bryuno and V.I. Parusnikov [27] approximately at the same time.

Example 18.11 The second example was studied by A.D. Bryuno and V.I. Parus-
nikov [27]. They constructed the continued fraction that is associated to the follow-
ing matrix:
⎛ ⎞
1 1 1
M = ⎝1 −1 0⎠ .
1 0 0

The positive Dirichlet group Ξ+ (M) is generated by the following two matrices:

X = M 2, Y = 2Id − M 2 .

The torus decomposition corresponding to this matrix is also homeomorphic to the

following one:

Here the segment AB is identified with the segment DC and the segment AD with the
segment BC. The triangle ABD has the integer affine type of the simplest triangle.
The triangle BCD has the integer affine type of the triangle with vertices (−1, −1),
(0, 1), and (1, 0).

The integer distance from the triangle ABD to the origin equals 2, and from the
triangle BCD it equals 1.
256 18 Periodicity of Klein Polyhedra. Generalization of Lagrange’s Theorem

Example 18.12 The third example was given by V.I. Parusnikov [159]. This contin-
ued fraction is associated to the following matrix:
⎛ ⎞
0 1 0
M = ⎝0 0 1 ⎠.
1 1 −3

The positive Dirichlet group Ξ+ (M) is generated by the following two matrices:

X = M 2, Y = 3Id − 2M −1 .

The torus decomposition corresponding to this matrix is also homeomorphic to the

following one:

Here the segment AB is identified with the segment DC, and the polygonal line
AGD with the polygonal line BEC (the point G is identified with the point E). All
triangles have the integer affine type of the simplest triangle. The pentagon BEFDG
has the integer affine type of the pentagon with vertices (−1, 0), (−1, 1), (0, 1)
(1, 0), and (1, −1):

The integer distances from the triangle ABC and the pentagon BEFDG to the origin
equal 1, from the triangle CDF to the origin equals 2, and from the triangle CFE
equals 3.

The continued fractions constructed in Examples 18.9, 18.11, and 18.12 are
also known as the continued fractions corresponding to the first, the second, and
the third Davenport forms. Several other single torus decompositions for two-
dimensional continued fractions were investigated in the works [160, 162], and
[161] by V.I. Parusnikov.
For a matrix A we denote by A the sum of absolute values of all the coefficients
for the matrix. Then we have the following theorem (for more information see [90]).

Theorem 18.13 Let A ∈ SL(3, Z) be an irreducible RS-matrix.

18.5 Examples of Infinite Series of Torus Decomposition 257

(i) A ≥ 4.
(ii) If A = 5 (there are 48 possible matrices), then the corresponding continued
fraction is integer congruent to the generalization of the regular fraction for
the golden ratio. This is shown in Example 18.9.
(iii) For the case A = 6 there are 912 different possible matrices. Consider the
associated two-dimensional continued fractions:
– 480 continued fractions are integer congruent to the one in Example 18.9;
– 240 of them are integer congruent to the continued fraction of Example
18.12;
– 192 of them are integer congruent to the continued fraction of Example
18.11.

The classification of two-dimensional continued fractions with norm greater than

or equal to seven is unknown.

18.5 Examples of Infinite Series of Torus Decomposition

We continue with several infinite series of torus decompositions. The first two in-
finite series were calculated by E.I. Korkina in [118]. The first of them is shown
below.

Example 18.14 The continued fractions of this series are associated to the following
matrices for a ≥ 0:
⎛ ⎞
0 0 1
Ma = ⎝1 0 −a − 5⎠ .
0 1 a+6
The positive Dirichlet group Ξ+ (Ma ) is generated by the following two matrices:

X a = Ma , Ya = (Ma − Id)2 .

The torus decomposition corresponding to Ma is homeomorphic to the following

one:

Here the segment AB is identified with the segment DC and the segment AD with
the segment BC. The integer distance from the triangle ABD to the origin equals
a + 2, and from the triangle BCD it equals 1.
258 18 Periodicity of Klein Polyhedra. Generalization of Lagrange’s Theorem

Both triangles have integer affine types of the simplest triangle:

Several other examples of infinite series of continued fractions were studied in

[89], some of which we now present. The following series generalizes the one from
the previous example.

Example 18.15 This series depends on two integer parameters a, b ≥ 0. Consider

the continued fractions associated to the following matrices:
⎛ ⎞
0 1 0
Ma,b = ⎝0 0 1 ⎠.
1 1+a−b −(a + 2)(b + 1)

The positive Dirichlet group Ξ+ (Ma,b ) is generated by the following two matrices:

−2 −1  −1 
Xa,b = Ma,b , Ya,b = Ma,b Ma,b − (b + 1)Id .

The torus decomposition corresponding to Ma,b is as follows:

Here the segment AB is identified with the segment DC and the segment AD with
the segment BC. Both triangles have the same integer affine type of the triangle with
vertices (0, 0), (0, 1), and (b + 1, 0) (b = 5 in the picture):

The integer distance from the triangle ABD to the origin equals a + 2, and from the
triangle BCD it equals 1.

Remark If we substitute b = 0 in Example 18.15, then we have a series of matrices

whose associated continued fractions are integer congruent to the continued frac-
tions associated to the series of matrices of Example 18.14. We formulate this more
precisely in the next proposition.
18.5 Examples of Infinite Series of Torus Decomposition 259

Proposition 18.16 The continued fractions associated to the following two matrices
are integer congruent (for integers a ≥ 0):
⎛ ⎞ ⎛ ⎞
0 1 0 0 0 1
Ma,0 = ⎝0 0 1 ⎠, Ma = ⎝1 0 −a − 5⎠ .
1 1+a −a − 2 0 1 a+6

Proof The matrices (Id − Ma,0 )−1 and Ma are conjugate by the matrix X in the
group SL(3, Z):
⎛ ⎞
−1 −1 −2
X=⎝ 0 0 −1⎠
1 0 −1

(i.e., Ma = X −1 (Id − Ma,0 )−1 X). Therefore, the corresponding continued fractions
are integer congruent by Proposition 18.2. 

Example 18.17 The series of this example depends on an integer parameter a ≥ 1.

The continued fractions of the series are associated to the following matrices:
⎛ ⎞
0 1 0
Ma,b = ⎝0 0 1 ⎠.
1 a −2a − 3

The positive Dirichlet group Ξ+ (Ma ) is generated by the following two matrices:
 −1
Xa = Ma−2 , Ya = 2Id − Ma−2 .

The torus decomposition corresponding to Ma is as follows:

Here the segment AB is identified with the segment DC and the segment AD with
the segment BC. All four triangles have integer affine types of the simplest triangle:
260 18 Periodicity of Klein Polyhedra. Generalization of Lagrange’s Theorem

The integer distance from the triangles ABD, BDE, and BCE to the origin are equal
to a + 2, a + 1, and 1, respectively.

Example 18.18 The series of this example depend on two integer parameters a > 0
and b ≥ 0. The continued fractions of the series are associated to matrices
⎛ ⎞
0 1 0
Ma,b = ⎝0 0 1 ⎠.
1 (a + 2)(b + 2) − 3 3 − (a + 2)(b + 3)

The positive Dirichlet group Ξ+ (Ma,b ) is generated by

 −1  −2 −2
Xa,b = (b + 3)Id − (b + 2)Ma,b Ma,b and Ya,b = Ma,b .

The torus decomposition corresponding to this matrix is homeomorphic to the fol-

lowing one:

Here the segment AB is identified with the segment DC and the polygonal line AFD
with the polygonal line BEC (the point F is identified with the point E). The integer
affine types of the faces ABF, CGE, CDG, BDF, and DBEC are as follows:

The integer distance from the triangles ABF, BFD, CDG, CEG to the origin equals
1, 1, 2 + 2a + 2b + ab, and 3 + 2a + 2b + ab respectively. The quadrangle DBEC
is at unit distance from the origin.

For further details related to such series we refer to [89].

18.6 Two-Dimensional Continued Fractions 261

18.6 Two-Dimensional Continued Fractions Associated to

Transpose Frobenius Normal Forms

Let us consider the family of transpose Frobenius matrices (or transpose companion
matrices)
⎛ ⎞
0 1 0
Am,n := ⎝0 0 1 ⎠,
1 −m −n
where m and n are integers; see Fig. 18.2.

Remark 18.19 Suppose that the characteristic polynomial χAm,n (x) is irreducible
over Q. Then the operator of multiplication by the element x in the field
Q[x]/(χAm,n (x) ) in the natural basis {1, x, x 2 } has the matrix Am,n .

Let us observe some basic properties. Denote by Ω the subset of irreducible RS-
matrices of this family.

Proposition 18.20 The set Ω is defined by the inequality

n2 m2 − 4m3 + 4n3 − 18mn − 27 ≤ 0.

In addition it is necessary to subtract all the matrices of types: Aa,−a and Aa,a+2 ,
a ∈ Z (these are the matrices with reducible characteristic polynomials).

Notice that the set Ω has the following symmetry.

Proposition 18.21 The two-dimensional continued fractions for the cubic irra-
tionalities constructed by the matrices Am,n and A−n,−m are integer congruent.

Remark 18.22 There are some other integer congruences of continued fractions
in Ω. For instance, the matrices A0,−a and A−2a,−a 2 have integer congruent contin-
ued fractions (since the matrices A20,−a and A−2a,−a 2 are integer conjugate).

Observe that continued fractions of transpose Frobenius matrices do not cover all
possible integer congruence classes of continued fractions for arbitrary RS-matrices
in SL(3, Z). For instance, the continued fraction of the following matrix is not inte-
ger congruent to that of a Frobenius matrix:
⎛ ⎞
1 2 0
A=⎝ 0 1 2⎠
−7 0 29

(we discuss similar questions later, in Chap. 21). So the following question is of
interest.
262 18 Periodicity of Klein Polyhedra. Generalization of Lagrange’s Theorem

Problem 15 How often are continued fractions integer congruent to those of the
set Ω?

Finally, in Fig. 18.2 we give a table (introduced in [90]) whose squares are filled
with torus decompositions of the sails for continued fractions associated to the ma-
trices of Ω. Integer affine types of all the faces of the torus decomposition for one of
the sails for Am,n is shown in the square at the intersection of the row with number n
and the column with number m. If one of the roots of the characteristic polynomial
for the matrix equals 1 or −1, then the corresponding square is marked with the sign
∗ or # respectively. The squares that correspond to the matrices whose characteristic
polynomial has two complex conjugate roots are shaded in light gray.

Remark Notice that some of the series constructed above are clearly seen in
Fig. 18.2.

18.7 Some Problems and Conjectures on Periodic Geometry of

Algebraic Sails
We begin with a problem on a complete invariant of sails. In the one-dimensional
case the answer is known, since the LLS sequence is a complete invariant of sails. In
dimensions two and higher the problem is open (in both the periodic and the general
cases). For the periodic case we have the following conjecture and open problem. If
the conjecture is true and the problem is solved, then we have a complete invariant
for two-dimensional periodic sails.

Conjecture 16 (V.I. Arnold) Torus decompositions of integer noncongruent sails

are distinct.

Problem 17 (V.I. Arnold) Describe all torus decompositions that are realizable by
periodic two-dimensional continued fractions.

Only a few examples are known in this direction. Let us give several trivial ex-
amples of torus decompositions that do not correspond to sails.

Example 18.23 The following torus decomposition is not realizable by sails of pe-
riodic continued fractions. There is one integer point in the interval AD, which we
denote by E:
18.7 Some Problems and Conjectures 263

Fig. 18.2 Torus decompositions for matrices Am,n

Here the polygonal line AEB maps to the polygonal line DFC under one of the
matrices of the group SL(3, Z) that preserves the sail. Since AEB is a segment, DFC
is also a segment. Therefore, the points B, C, F , and D lie in the same plane. And
hence BF is not an edge of some face.
Now we describe the first nontrivial example of torus decomposition, which was
given by E.I. Korkina in the work [118].

Example 18.24 Consider the simplest torus triangulation. It consists of two triangles
with the simplest integer affine type of the triangle (0, 0), (0, 1), and (1, 0).
264 18 Periodicity of Klein Polyhedra. Generalization of Lagrange’s Theorem

The integer distances to both faces are 1. This decomposition is not realizable for
periodic sails.

This example (together with many other ones) allows us to formulate the follow-
ing conjecture.

Conjecture 18 For every torus decomposition for the continued fraction of a cubic
irrationality there exists some face with integer distance to the origin greater than
one.

On the other hand, every known (periodic) sail has a face with the integer distance
to the origin equalling one.

Conjecture 19 For every torus decomposition for the continued fraction of a cubic
irrationality there exists some face whose integer distance to the origin equals one.

The following example is a torus decomposition that consists of a single face.

Example 18.25 Consider the torus decomposition consisting of one face with inte-
ger affine type of the simplest parallelogram with the vertices (0, 0), (0, 1), (1, 1),
and (1, 0).

The integer distances to this face can be chosen arbitrarily. This decomposition is
not realizable for periodic sails of cubic irrationalities.

It seems that a torus decomposition with a single rectangular face is not realizable
in general. Here we conjecture a stronger statement.

Conjecture 20 The torus decomposition for every sail of every cubic irrationality
contains a triangular face.

The next question is related to continued fractions of the same cubic extension.
18.8 Generalized Lagrange’s Theorem 265

Problem 21 (V.I. Arnold) Classify continued fractions that corresponds to the same
cubic extensions of the field of rational numbers.

Almost nothing is known here (even in the one-dimensional case). For example,
even the finiteness of the number of possible continued fractions associated to the
same extension is unknown. (For properties of cubic extensions of rational numbers
see in the work of B.N. Delone and D.K. Faddeev [52].) Here is a nice example
showing that a characteristic polynomial does not distinguish two-dimensional con-
tinued fractions.

Example 18.26 The following two matrices having the same characteristic polyno-
mial x 3 + 11x 2 − 4x − 1 (and hence the same cubic extension of Q) define integer
noncongruent continued fractions:
⎛ ⎞3 ⎛ ⎞
0 1 0 0 1 0
⎝0 0 1 ⎠ , ⎝0 0 1 ⎠.
1 1 −2 1 4 −11

18.8 Generalized Lagrange’s Theorem

The combinatorial topological generalization of Lagrange’s theorem was announced
by E.I. Korkina in [116]; a complete proof of it was given in 2008 by O.N. German
and E.L. Lakshtanov in [70]. A slightly weaker algebraic generalization was pro-
posed in 1993 by G. Lachaud; see [123, 124], and [125]. In this chapter we present
a version of the combinatorial topological generalization of Lagrange’s theorem.

Definition 18.27 Let C be a cone and v a vertex of the sail for C. An edge star of
v is the union of all edges of the sail adjacent to v; we denote it by Stv .
We say that a star is regular if v is in the interior of the cone C.

Consider a star Stv and a point p (not necessary integer) in the complement
to Stv . We say that Stv is a regular star with respect to p if
– v is in the interior of conv(Stv , p).
– the set conv(Stv , p) \ (conv(Stv ) ∪ {p}) does not contain integer points.
Define
Γ (Stv ) = {p|Stv is regular star with respect to p}.

Theorem 18.28 Let Stv be a regular star. Then the set Γ (Stv ) is bounded.

Proof We prove the statement by induction on the dimension of the star.

Base of induction. The statement clearly holds for every star (i.e., for one point)
in R1 .
Step of induction. Let the statement hold for every star in Rd−1 ; we prove the
statement in Rd .
266 18 Periodicity of Klein Polyhedra. Generalization of Lagrange’s Theorem

Consider an arbitrary regular star Stv in Rd . Suppose Γ (Stv ) is not closed. Then
there exists a (unit) direction ξ that is an accumulation points of unbounded direc-
tions, i.e., for every neighborhood U ⊂ S d−1 of ξ and every N > 0 there exists a
point p ∈ Γ (Stv ) with direction in U and |vp| > N . Hence the half-prism

conv(Stv ) ⊕ {λξ |λ ≥ 0}

does not contain integer points in the interior except for the points of conv(Stv ).
There are two different cases to consider here.
The first case is v being inside the half-prism. Then either ξ is rational and the
cone contains integer points, or ξ is irrational and then the half-prism contains in-
teger points by the multidimensional Kronecker’s approximation theorem (Theo-
rem 16.16). We have arrived at a contradiction.
In the second case we have v on the boundary of the half-prism. This mean that
ξ is in the plane of some face Std−1 v of Stv . Let L be the integer hyperplane contain-
ing Std−1
v . The direction ξ is also an accumulation points of unbounded directions
in Γ (Std−1
v ), since the intersections of pyramids with direction close to ξ with L are
also unbounded. This contradicts with induction assumption.
Hence the set Γ (Stv ) is bounded. 

Corollary 18.29 The set Γ (Stv ) contains finitely many integer points.

Definition 18.30 Let C be a cone in Rn and let SC be its sail. A two-sided infinite
sequence (vi ) of vertices in the sail SC in Rn is called a chain of SC if for every
integer k we have
– vk and vk+1 are connected by an edge of the sail;
– the points vk , . . . , vk+n−1 are affinely independent (i.e., they span the (n − 1)-
dimensional affine plane).

Definition 18.31 Let C be a cone in Rn , SC its sail, and (vi ) a chain in SC . Consider
an arbitrary subset U of Aff(n, Z). We say that the chain (vi ) is U -periodic if there
exists a period t such that for every i there exists Ai ∈ U satisfying

Ai (Stvj ) = (Stvj +t ) for j = i, i + 1, . . . , i + n.

Denote by Hn the set of affine transformations Ax + b, where A is an integer

irreducible RS-matrix in GL(n, Z) (and hence it has distinct eigenvalues). Let Hn0
denote the subset of Hn with b = 0. Since we have a fixed basis in Rn , we identify
linear transformations with matrices in the fixed basis.

Theorem 18.32 (Geometric generalization of Lagrange theorem) Consider an ir-

rational cone C ⊂ Rn and let SC be its sail. The following two statements are equiv-
alent:
– there exists a GL(n, Z)-transformation A ∈ Hn0 such that A(C) = C.
– there is an Hn -periodic chain of vertices of the sail SC .
18.8 Generalized Lagrange’s Theorem 267

Remark 18.33 A stronger similar statement holds for the class of all affine transfor-
mations whose corresponding linear transformations satisfy:
– all eigenvalues are distinct from 1;
– all pairs of complex conjugate eigenvalues are distinct and distinct from the ab-
solute values of real eigenvalues.
To avoid technical details we restrict ourselves to the class of affine transforma-
tions with irreducible RS-matrix in GL(3, Z), i.e., to Hn . For more information on
the general case we refer to [70].

Lemma 18.34 Let (vi ) be an Hn -periodic chain of vertices for some sail S in Rn .
Then there exists a transformation A in Hn0 establishing a nontrivial shift of stars,
i.e., there exists a positive integer T such that for every k we have

AT (Stvk ) = Stvk+T .

Proof First, we prove that there exists an affine transformation in Hn establishing a

nontrivial shift of the union of all the stars along itself.
Set Bi = A−1i+1 Ai . Notice that for k = 1, . . . , n we have Bi (Stvk ) = Stvk , in partic-
ular Bi (vk ) = vk . Though it is not clear whether Bi acts on the stars as the identity
map or as a symmetry, nevertheless we know that it preserves the set of all ele-
mentary vectors corresponding to the edges of the stars. Denoting by r1 the sum of
all primitive vectors corresponding to the edges of the star (Stv1 ), let s1 = v1 + r1 .
Notice that, first, Bi (s1 ) = s1 , second, s1 is not contained in F .
Consider now the (n − 1)-dimensional plane spanning the face F and the point
v, denoted by π . The complement to the plane π contains at least one edge of the
star Stv1 . Denote by r2 the sum of all elementary vectors for edges of the star Stv1
in the half-space containing this edge, which we denote by π+ . Since Bi is a proper
affine transformation preserving the plane π , it preserves the half-space π+ . Thus it
preserves the set of all elementary vectors of the star Stv1 that are in this half-space,
and in particular, the vector r2 . Let s2 = v1 + r2 . We have Bi (s2 ) = s2 .
From all of the above, the transformation Bi preserves the n + 1 points v1 , . . . ,
vn−1 , r1 , and r2 . By the condition on vi and construction of r1 and r2 , these points
span Rn . Therefore, Bi is the identity transformation and Ai = Ai+1 . So the trans-
formation A1 (in Hn ) acts on the stars as a shift.
Let us find now a linear transformation acting on the stars as a shift. For an ar-
bitrary integer k, we consider a transformation A−k 1 . This transformation sends Stvk
to Stv1 and the set of all admissible vertices Γ (Stvk ) to Γ (Stv1 ). By Theorem 18.28,
Γ (Stv1 ) is bounded, and since it contains only integer points, Γ (Stv1 ) is finite. Let
this set have N points. Consider the set
 k 
A1 (O) | k = 0, . . . , N .

All these points are in Γ (StvN ), since Ak1 (StvN −k ) = Stk . Since the stars Stv1 and
StvN are integer congruent, the set Γ (StvN ) contains exactly N points. Therefore,
268 18 Periodicity of Klein Polyhedra. Generalization of Lagrange’s Theorem

j
there exist distinct i and j such that Ai1 (O) = A1 (O). Hence Ai−j (O) = O, mean-
ing that Ai−j is a linear transformation in Hn and thus in Hn0 .
The transformation Ai−j ∈ Hn0 establishes a nontrivial shift of the union of all
the stars (Stvi ) along itself. 

Proof of Theorem 18.32 Suppose that there exists an SL(n, Z)-transformation

A ∈ Hn0 such that A(C) = C. Then for an arbitrary vertex of the sail v the chain
(An (v)) with integer parameter n is Hn -periodic.
The converse statement is more complicated. Suppose that there exists an un-
bounded, in both directions (as an ordered set in Rn ), Hn -periodic chain of ver-
tices (vi ) of the sail SC . Then by Lemma 18.34 there is a transformation A in Hn0
establishing a nontrivial shift in the sequence of the stars Stvi . Let us prove that
A(C) = C.
Let us consider an (n − 1)-dimensional tetrahedron T = v1 . . . vn . Notice that T
is in the sail, and hence the hyperspace containing it does not contain the origin O.
Since Am (T ) = vm+1 . . . vm+n is in C, the tetrahedron T is in the cone A−m (C).
Let (ei ) be an eigenbasis of A such that the corresponding eigenvalues (λi ) form a
decreasing sequence. Denote also by (wi ) a basis corresponding to the cone C.
Let v = αk ek + · · · + αl el , where αk
= 0. Then asymptotically, as m → +∞ we
have

Am (v) = αk λm
k ek + smaller terms;

m
1
A−m (v) = αl el + smaller terms.
λl

Therefore, the sequences of rays defined by the directions of Am (v) and A−m as
m → +∞ converge to the rays with direction ek and el respectively. (Here to define
the convergence we consider the topology of the unit sphere. A ray is represented
by a point of intersection of this ray with the unit sphere.)
By the above, the set (Am (wi )) should contain the tetrahedron T for every inte-
ger m. Therefore, the limiting cone at +∞ also contains T . Since the tetrahedron T
is (n − 1)-dimensional and the hyperspace of T does not contain the origin, all the
rays in the limit has linearly independent directions. Therefore, the rays are distinct
and (possibly after some permutations of vectors ei ) the matrix of vectors wi in the
basis (ei ) is lower triangular:
⎛ ⎞
α1,1 0 ··· 0 0
⎜ α2,1 α1,2 ··· 0 0 ⎟
⎜ ⎟
⎜ .. .. . . .. ⎟
⎜ . . . . .
. . ⎟
⎜ ⎟
⎝αn−1,1 αn−1,2 · · · αn−1,n−1 0 ⎠
αn,1 αn,2 ··· αn,n−1 αn,n

For the same reasons, all the rays of the limiting cone at −∞ are independent.
Hence for every i there exists a vector wj that is contained in the span of e1 , . . . , ei .
18.9 Littlewood and Oppenheim Conjectures 269

This implies that all the coefficients under the diagonal are zeros. Thus, the matrix
is diagonal, and all vectors wi are eigenvectors of A. Therefore, A(C) = C (recall
that by Lemma 18.34 the transformation A is in Hn0 ). 

18.9 Littlewood and Oppenheim Conjectures in the Framework

of Multidimensional Continued Fractions
In this section we briefly show a link between the Littlewood and Oppenheim con-
jectures and a certain conjecture in the theory of multidimensional continued frac-
tions. Letting v be an arbitrary real number, denote by v the distance from this
number to the set of integers, i.e.,

v = min |v − k|.

k∈Z

Let us formulate the original conjectures.

Conjecture 22 (Littlewood conjecture) For an arbitrary pair of real numbers

(α, β) the following holds:

inf mmαmβ = 0.
m∈Z+

Conjecture 23 (Oppenheim conjecture) Let n ≥ 3. Consider n linearly indepen-

dent linear forms L1 , . . . , Ln in Rn . Suppose that
 
inf L1 (p) · L2 (p) · · · Ln (p) > 0.
p∈Z \{0}
n

Then there exists an irreducible RS-matrix A ∈ SL(n, Z) such that the planes
Li (p) = 0 for i = 1, . . . , n are the invariant planes of A.

Remark It is known that for n = 3 the Oppenheim conjecture implies the Littlewood
conjecture (see, for instance, [34]).

Now we reformulate the Oppenheim conjecture in terms of the geometry of sails.

We start with the following definition.

Definition 18.35 Let F be an arbitrary (n − 1)-dimensional face of a sail in Rn and

let v1 , . . . , vm be the vertices of F . Then the determinant of F is
  
det F = det(vi , . . . , vi ).
1 n
1≤i1 <···<in ≤m

We also say that the determinant of an edge star Stv is the Euclidean volume of
the convex hull conv(Stv ).
270 18 Periodicity of Klein Polyhedra. Generalization of Lagrange’s Theorem

Theorem 18.36 (O. German [70]) Let n ≥ 3. Consider n linearly independent lin-
ear forms L1 , . . . , Ln in Rn . Denote by C the cone
 
C = p ∈ Rn | Li (p) = 0, i = 1, . . . , n .

Then the following two statements are equivalent:

(i) infp∈Zn \{0} |L1 (p) · L2 (p) · · · Ln (p)| > 0;
(ii) the faces of the sail SC have uniformly bounded determinants.

This theorem reduces the Oppenheim conjecture to the following conjecture.

Conjecture 24 Let the faces and edge stars of a cone C have uniformly bounded
determinants. Then the sail SC is algebraically periodic (i.e., it is a sail of some
irreducible RS-matrix of SL(n, Z)).

Theorem 18.32 implies that to check periodicity it is enough to find an Hn -

periodic chain of vertices of the sail.
Finally, we would like to mention here one unsuccessful attempt (which is never-
theless worthy mentioning) to solve these conjectures in [191] and [192]. For further
details we refer the interested reader to [68, 69], and [70].

18.10 Exercises
Exercise 18.1 Suppose that a polynomial p with integer coefficients is irreducible
over Q. Prove that all the roots of p are distinct.

Exercise 18.2 Check that the Dirichlet groups for Examples 18.9, 18.11, and 18.12
are correctly calculated.

Exercise 18.3 Let A ∈ SL(3, Z) be an irreducible RS-matrix. Prove that A ≥ 5.

Exercise 18.4 Prove that the torus decomposition consisting of a single square of
integer area 2 is not realizable as a sail of a periodic two-dimensional continued
fraction.

Exercise 18.5 Consider arbitrary integers m and n. Suppose that the matrix Am,n is
an irreducible RS-matrix. Prove that the matrix A−n,−m is also an irreducible RS-
matrix and that the continued fraction associated to A−n,−m is integer congruent to
the continued fraction associated to Am,n .

Exercise 18.6 Consider a cone C and its sail SC in Rn . Let (vi ) be some chain of
the sail SC . Prove that each n − 1 consecutive edges of the chain are edges of a face
in SC .
Chapter 19
Multidimensional Gauss–Kuzmin Statistics

In this chapter we study the distribution of faces in multidimensional continued frac-

tions. The one-dimensional Gauss–Kuzmin distribution was described in Chap. 9,
where we discussed the classical approach via ergodic theory and a new geometric
approach. Currently, an ergodic approach to the distribution of faces in continued
fractions has not been developed. In fact, it is a hard open problem to find an ap-
propriate generalization of the Gauss map suitable to the study of ergodic properties
of faces of multidimensional sails. This problem can be avoided, and in fact, the
information on the distribution of faces is found via the generalization of the geo-
metric approach via Möbius measures described in the second part of Chap. 9 for the
one-dimensional case. We discuss the multidimensional analogue of the geometric
approach in this chapter.

19.1 Möbius Measure on the Manifold of Continued Fractions

We begin with a description of the structure of a smooth manifold on the set of
multidimensional continued fractions. Further, we extend the notions of Möbius
forms and measures to the multidimensional case.

19.1.1 Smooth Manifold of n-Dimensional Continued Fractions

Denote the set of all continued fractions of dimension n by CF n . An arbitrary con-

tinued fraction is defined by an unordered collection of hyperplanes (π1 , . . . , πn+1 ).
Denote by li , for i = 1, . . . , n + 1, the intersection of all the above hyperplanes ex-
cept for the hyperplane πi . Obviously, l1 , . . . , ln+1 are independent straight lines
(i.e., they are not contained in a hyperplane) passing through the origin. These
straight lines form an unordered collection of independent straight lines. Conversely,
every unordered collection of n + 1 independent straight lines uniquely determines
some continued fraction.

O. Karpenkov, Geometry of Continued Fractions, 271

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_19, © Springer-Verlag Berlin Heidelberg 2013
272 19 Multidimensional Gauss–Kuzmin Statistics

Denote the sets of all ordered collections of n + 1 independent and dependent

straight lines by FCF n and Δn , respectively. We say that FCF n is the space of n-
dimensional framed continued fractions. Also denote by Sn+1 the permutation group
acting on ordered collections of n + 1 straight lines. In this notation we have
 
FCF n = RP n × RP n!"× · · · × RP n# \ Δn and CF n = FCF n /Sn+1 .
n+1 times

This implies that the sets FCF n and CF n admit natural structures of smooth mani-
folds induced by the Cartesian product of n + 1 copies of RP n . Note also that FCF n
is an ((n + 1)!)-fold covering of CF n . We call the map p : FCF n → CF n of “for-
getting” the order in the ordered collections the natural projection of the manifold
FCF n to the manifold CF n .

19.1.2 Möbius Measure on the Manifolds of Continued Fractions

The group PGL(n + 1, R) takes the set of all straight lines passing through the ori-
gin of (n + 1)-dimensional space into itself (for the definition of PGL(n + 1, R) and
cross-ratios we refer to Chap. 9). Hence, PGL(n + 1, R) naturally acts on the mani-
folds CF n and FCF n . Furthermore, the action of PGL(n + 1, R) is transitive, i.e., it
takes any (framed) continued fraction to any other. Note that for any n-dimensional
(framed) continued fraction, the subgroup of PGL(n + 1, R) taking this continued
fraction to itself is of dimension n.

Definition 19.1 A form of the manifold CF n (respectively FCF n ) is said to be a

Möbius form if it is invariant under the action of PGL(n + 1, R).

Transitivity of the action of PGL(n + 1, R) implies that every two n-dimensional

Möbius forms of the manifolds CF n and FCF n are proportional.

Definition 19.2 Let M be a smooth enough arcwise connected manifold and let ω
be a volume form on it. Denote by μω the measure on M defined as follows: for
every open measurable set S ⊂ M,
 
 
μω (S) =  ω.
S

Definition 19.3 A measure μ of the manifold CF n (FCF n ) is said to be a Möbius

measure if there exists a Möbius form ω of CF n (FCF n ) such that μ = μω .

Every two Möbius measures of CF n (FCF n ) are proportional. The projection p

projects the Möbius measures of the manifold FCF n to the Möbius measures of the
manifold CF n .
19.2 Explicit Formulae for the Möbius Form 273

19.2 Explicit Formulae for the Möbius Form

Let us write down explicitly the Möbius forms for the manifold of framed n-
dimensional continued fractions FCF n for arbitrary n.
Consider Rn+1 with standard metrics on it. Let π be an arbitrary hyperplane of
the space Rn+1 with chosen Euclidean coordinates OX 1 . . . Xn , and assume that
π does not pass through the origin. We call the set of all collections of n + 1
ordered straight lines intersecting the plane π the chart FCF n,π of the manifold
FCF n . Let the intersection of π with the ith plane be a point with coordinates
(x1,i , . . . , xn,i ) on π . For an arbitrary tetrahedron A1 . . . An+1 in the plane π , we
denote by Vπ (A1 , . . . , An+1 ) its oriented Euclidean volume (with respect to the
orientation induced by the coordinates OX 1,1 . . . Xn,1 X1,2 . . . Xn,n+1 of the chart
FCF n,π ).

Remark 19.4 The chart FCF n,π is everywhere dense in (Rn )n+1 .

Consider the following form in the chart FCF n,π :

-n+1 -n
i=1 ( j =1 dxj,i )
ωπ (x1,1 , . . . , xn,n+1 ) = .
Vπ (A1 , . . . , An+1 )n+1

Proposition 19.5 The measure μωπ is a restriction of a Möbius measure to FCF n,π .

Proof Every transformation of the group PGL(n + 1, R) is in one-to-one correspon-

dence with the set of all projective transformations of the plane π . Let us show that
the form ωπ is invariant under the action of the transformations (of the everywhere
dense set) of the chart FCF n,π that are induced by projective transformations of the
hyperplane π .
Denote by |v|π the Euclidean length of a vector v in the coordinates
OX 1,1 . . . Xn,n+1 of the chart FCF n,π . Consider an arbitrary point (x1,1 , . . . , xn,n+1 )
of the chart FCF n,π . Denote by Ai = Ai (x1,i , . . . , xn,i ) the point depending on the
coordinates of the plane π with coordinates (x1,i , . . . , xn,i ), i = 1, . . . , n + 1. So
for the chosen point, we have a naturally defined tetrahedron A1 . . . An+1 in the
hyperplane π . Set

Aj Ai
f ij = , i, j = 1, . . . , n + 1; i
= j.
|Aj Ai |π

It is clear that the vectors f ij and f j i are parallel to the vector Aj Ai . The collection
(f ij ) is a basis in the tangent space to FCF n,π , which depends continuously on the
point in the chart. By dvij we denote the 1-form corresponding to the coordinate
along the vector f ij of FCF n,π .
Let us rewrite the form ωπ in new coordinates:
274 19 Multidimensional Gauss–Kuzmin Statistics

ωπ (x1,1 , . . . , xn,n+1 )
&
n+1
Vπ (Ai , A1 , . . . , Ai−1 , Ai+1 , . . . An+1 )

dv21 ∧ dv31 ∧ · · · ∧ dvn,n+1

= %n+1 ·
k=1,k
=i |A A |
k i π Vπ (A1 , . . . , An+1 )n+1
i=1
dv21 ∧ dv12 dv32 ∧ dv23 dvn+1,n ∧ dvn,n+1
= (−1)(n+3)/4 · ∧ ∧ ··· ∧ .
|A1 A2 |2π |A2 A3 |2π |An+1 An |2π

As in the one-dimensional case, the expression

Δvij Δvj i
|Ai Aj |2

for the infinitesimally small Δvij and Δvj i is the infinitesimal cross-ratio of the
four points Ai , Aj , Ai + Δvj i f j i , and Aj + Δvij f ij on the straight line Ai Aj .
Therefore, the form ωπ is invariant under the action of the transformations (of the
everywhere dense set) of the chart FCF n,π that are induced by the projective trans-
formations of the hyperplane π . Hence the measure μωπ coincides with the restric-
tion of some Möbius measure to FCF n,π . 

Corollary 19.6 A restriction of an arbitrary Möbius measure to the chart FCF n,π
is proportional to μωπ .

Proof The statement follows from the proportionality of any two Möbius mea-
sures. 

Let us fix an origin Oij for the straight line Ai Aj . The integral of the form dvij
(respectively dvj i ) for the segment Oij P defines the coordinate vij (vij ) of the point
P contained in the straight line Ai Aj . As in the one-dimensional case, consider a
projectivization of the straight line Ai Aj . Denote the angular coordinates by ϕij and
ϕj i , respectively. In these coordinates,

dvij ∧ dvj i 1 2 ϕj i − ϕij

= cot dϕij ∧ dϕj i .
|Ai Aj |2π 4 2

Then the following is true.

Corollary 19.7 The form ωπ is extendible to some form ωn of FCF n . In coordinates

vij , the form ωn is as follows:
n+1 n+1
 n+1 
(−1)(n+3)/4 & & 2 ϕij − ϕj i
. n+1.
ωn = cot · dϕij ∧ dϕj i .
2n(n+1) 2
i=1 j =i+1 i=1 j =i+1
19.3 Relative Frequencies of Faces of Multidimensional Continued Fractions 275

19.3 Relative Frequencies of Faces of Multidimensional

Continued Fractions
Without loss of generality we choose precisely the form ωn of Corollary 19.7. De-
note by μn the projection of the measure μωn to the manifold of multidimensional
continued fractions CF n .
Consider an arbitrary polytope F with vertices at integer points. Denote by
CF n (F ) the set of n-dimensional continued fractions that contain the polytope F
as a face.

Definition 19.8 The value μ(CF n (F )) is called the relative frequency of a face F .

Let us specify the integer-linear congruence of faces.

Definition 19.9 We say that two faces of a sail are integer-linear congruent if there
exists an SL(n, Z)-transformation sending one face to the other.

Remark 19.10 It some sense the structure of a face includes the origin and the poly-
hedron of this face. Of course, if two faces are integer-linear congruent, the corre-
sponding polyhedra are integer congruent. The converse is not true. For instance,
two empty triangles (that are integer congruent) can be integer distance one, and
integer distance two to the origin. Such faces are not integer-linear congruent.

Proposition 19.11 The frequencies of integer-linear congruent faces coincide.

The following problem is open for n ≥ 2.

Problem 25 Find faces of n-dimensional continued fractions with the highest rel-
ative frequencies. Is it true that for every constant C there exist only finitely many
pairwise integer noncongruent faces whose relative frequencies do not exceed C?
Find the corresponding asymptotics (with respect to C tending to infinity).

One of the first sets of problems on statistics of faces of multidimensional con-

tinued fractions was proposed by V.I. Arnold. Let us consider the set of three-
dimensional integer matrices with rational eigenvalues, denoted by A3 . The contin-
ued fraction associated to a matrix of A3 consists of finitely many faces. Denote by
A3 (m) the set of all the matrices of A3 for which the sums of absolute values of all
their elements are not greater than m. The number of such matrices is finite. Let us
calculate the number of triangles, quadrangles, etc., among the continued fractions,
constructed for the matrices of A3 (m). As m tends to infinity we have a general
distribution of the frequencies for triangles, quadrangles, etc. We call the resulting
frequencies of integer faces the frequencies of faces in the sense of Arnold. The
problems of V.I. Arnold include the study of certain properties of such distributions
(for instance, what is more frequent, triangles or quadrangles, what is the average
number of integer points inside the faces, etc.). The majority of these problems are
open. For further information we refer to [9] and [10].
276 19 Multidimensional Gauss–Kuzmin Statistics

Remark 19.12 Let us say a few words about the following particular question of
V.I. Arnold: which faces in two-dimensional sails are more frequent, triangular or
quadrangular? For instance, in the case of R3 , there are two natural models of ran-
dom polyhedra. The first one is to choose k random points in space and to construct
their convex hull. For the second method, one should consider the intersection of k
random half-spaces. In the first method, a random convex hull has only triangular
faces, while in the second the average number of vertices of faces is four (Schläfli–
Saharov theorem). So it is interesting to understand whether two-dimensional sails
are “closer” to random polyhedra in the first sense (in this case they have more
triangles) or in the second sense (then they should have more quadrangles).

The following conjecture connects the notions of relative frequencies and fre-
quencies in the sense of Arnold.

Conjecture 26 The relative frequencies of faces are proportional to the frequencies

of faces in the sense of Arnold.

This conjecture is open in the n-dimensional cases for n ≥ 2.

19.4 Some Calculations of Frequencies for Faces in the

Two-Dimensional Case

19.4.1 Some Hints for Computation of Approximate Values of

Relative Frequencies

Consider the space R3 with the standard metric on it. Let π be an arbitrary plane in
R3 not passing through the origin and with fixed system of Euclidean coordinates
Oπ Xπ Yπ . Let FCF 2,π be the corresponding chart of the manifold FCF 2 . For an
arbitrary triangle ABC of the plane π , we denote by Sπ (ABC) its oriented Euclidean
area in the coordinates Oπ X1 Y1 X2 Y2 X3 Y3 of the chart FCF 2,π . Denote by |v|π the
Euclidean length of a vector v in the coordinates Oπ X1 Y1 X2 Y2 X3 Y3 of the chart
FCF 2,π . Consider the following form in the chart FCF 2,π :

dx1 ∧ dy1 ∧ dx2 ∧ dy2 ∧ dx3 ∧ dy3

ωπ (x1 , y1 , x2 , y2 , x3 , y3 ) = .
Sπ ((x1 , y1 )(x2 , y2 )(x3 , y3 ))3
Note that the oriented area Sπ of the triangle (x1 , y1 )(x2 , y2 )(x3 , y3 ) can be ex-
pressed in the coordinates xi , yi as follows:
  1
Sπ (x1 , y1 )(x2 , y2 )(x3 , y3 ) = (x3 y2 − x2 y3 + x1 y3 − x3 y1 + x2 y1 − x1 y2 ).
2
For the approximate computations of relative frequencies of faces it is useful
to rewrite the form ωπ in the dual coordinates (see Remark 19.14 below). Define
19.4 Some Calculations of Frequencies for Faces 277

a triangle ABC in the plane π by three straight lines l1 , l2 , and l3 , where l1 passes
through B and C, l2 passes through A and C, and l3 passes through A, and B. Define
the straight line li (i = 1, 2, 3) in π by the equation (after first making a translation
of π in such a way that the origin is taken to some inner point of the triangle)

ai x + b i y = 1

in the variables x and y. Then if we know the 6-tuple of numbers (a1 , b1 , a2 , b2 ,

a3 , b3 ), we can restore the triangle in a unique way.

Proposition 19.13 In coordinates a1 , b1 , a2 , b2 , a3 , b3 the form ωπ can be written

as follows:
8da1 ∧ db1 ∧ da2 ∧ db2 ∧ da3 ∧ db3
− .
(a3 b2 − a2 b3 + a1 b3 − a3 b1 + a2 b1 − a1 b2 )3

So, we reduce the computation of relative frequency for the face F (i.e., the value
of μ2 (CF 2 (F ))) to the computation of the measure μω2 (p −1 (CF 2 (F ))). Consider
some plane π in R3 not passing through the origin. By Corollary 19.7 we have
     
μω2 p −1 CF 2 (F ) = μωπ p −1 CF 2 (F ) ∩ (FCF 2,π ).

Finally, the computation should be made for the set μωπ (p −1 (CF 2 (F ))) ∩ (FCF 2,π )
in dual coordinates ai , bi (see Proposition 19.13).

Remark 19.14 In coordinates ai , bi the computation of the value of the relative fre-
quency often reduces to the estimation of the integral on the disjoint union of a finite
number of six-dimensional Cartesian products of three triangles in coordinates ai ,
bi (see Proposition 19.13). The integration over such a simple domain significantly
increases the speed of approximate computations. In particular, the integration can
be reduced to the integration over some 4-dimensional domain.

19.4.2 Numeric Calculations of Relative Frequencies

19.4.2.1 Remark on Complexity of Numeric Calculations

Explicit calculation of relative frequencies for the faces seems not to be realizable.
Nevertheless, it is possible to make approximations of the corresponding integrals.
Usually, the greater the area of the polygon, the smaller its relative frequency. The
most complicated approximation calculations correspond to the simplest faces, such
as an empty triangle.
In Fig. 19.1 we examine examples of the following faces:
– triangular faces (0, 0, 1), (0, 1, 1), (1, 0, 1), and (0, 0, 1);
– triangular face (0, 2, 1), (2, 0, 1);
278 19 Multidimensional Gauss–Kuzmin Statistics

Fig. 19.1 The points painted in light gray correspond to the points at which the rays defining the
two-dimensional continued fraction can intersect the plane of the chosen face

– quadrangular face (0, 0, 1), (0, 1, 1), (1, 1, 1), (1, 0, 1).

For each face we show the plane containing the face.

A point is shaded gray if there exists a continued fraction such that one of the
rays defining it intersects this plane. The light gray points correspond to the points
at which the rays defining the two-dimensional continued fraction intersect the plane
of the chosen face.

19.4.2.2 Some Numeric Results

In Table 19.1 we show the numerical approximations of relative frequencies for

12 faces. In the column “N◦ ” we write a special sign for integer-linear congruence
classes of a face. The index denotes the integer distance from the corresponding
face to the origin. In the column “face” we draw a picture of the integer-linear con-
gruence class of the face. Further, in the columns “lS” and “ld” we write down the
integer areas of faces and integer distances from the planes of faces to the origin
respectively. Finally, in the column “μ2 ” we show approximate relative frequencies
for the corresponding face.
Notice that in the given examples the integer congruence classes of polygons and
integer distances to the origin determine the integer-linear congruence type of the
face. This is not always the case if we consider greater distances to the origin.

Remark 19.15 The majority of faces of two-dimensional continued fractions are

at unit integer distance from the origin. Only three infinite series and three partial
examples of faces lie at integer distances greater than one from the origin, as seen
in Chap. 15. If the distance to the face is increasing, then the frequency of faces is
decreasing on average. The average rate of decrease of the frequency is unknown to
the author.
19.5 Exercises 279

Table 19.1 Some numeric results of calculations of relative frequencies

N◦ Face lS ld μ2 N◦ Face lS ld μ2

I1 3 1 1.3990 · 10−2 VI1 7 1 3.1558 · 10−4

I3 3 3 1.0923 · 10−3 VI2 7 2 3.1558 · 10−4

II1 5 1 1.5001 · 10−3 VII1 11 1 3.4440 · 10−5

III1 7 1 3.0782 · 10−4 VIII1 7 1 5.6828 · 10−4

IV1 9 1 9.4173 · 10−5 IX1 7 1 1.1865 · 10−3

V1 11 1 3.6391 · 10−5 X1 6 1 9.9275 · 10−4

19.4.3 Two Particular Results on Relative Frequencies

In conclusion, we give two simple statements on relative frequencies of faces.

Proposition 19.16 Integer congruent polygons P1 and P2 at integer distances 1

and 2 to the origin have the same relative frequencies (see, for example, VI1 and
VI2 of Table 19.1).

Denote by An the triangle with vertices (0, 0, 1), (n, 0, 1), and (0, n, 1). Denote
by Bn the square with vertices (0, 0, 1), (n, 0, 1), (n, n, 1), and (0, n, 1).

Proposition 19.17 The following holds:

μ(CF 2 (An ))
lim = 8.
n→∞ μ(CF 2 (Bn ))

For further information related to multidimensional Gauss–Kuzmin statistics we

refer to the work [113] by M.L. Kontsevich and Yu.M. Suhov and [93] by the author.

19.5 Exercises
Exercise 19.1 Show that relative frequencies of integer-linear congruent faces are
equivalent.
280 19 Multidimensional Gauss–Kuzmin Statistics

Exercise 19.2 Prove Propositions 19.16 and 19.17.

Exercise 19.3 For every face F of Table 19.1 draw the intersection set of the plane
spanned by F with all the rays that define two-dimensional continued fractions hav-
ing F as a face (see examples in Fig. 19.1).

Exercise 19.4 Calculate up to the second digit the relative frequencies for the faces
of types VII1 and X1 in Table 19.1.
Chapter 20
On Construction of Multidimensional Continued
Fractions

In the first part of this book we saw that the LLS sequences completely determine
all possible sails of integer angles in the one-dimensional case. The situation in the
multidimensional case is much more complicated. Of course, the convex hull algo-
rithms can compute all the vertices and faces of sails for finite continued fractions,
but it is not clear how to construct (or to describe) vertices of infinite sails of multi-
dimensional continued fractions in general. What integer-combinatorial structures
could the infinite sails have? There is no single example in the case of aperiodic in-
finite continued fractions of dimension greater than one. The situation is better with
periodic algebraic sails, where each sail is characterized by its fundamental domain
and the group of period shifts (i.e., the positive Dirichlet group).
In this chapter we show the main algorithms that are used to construct examples
of multidimensional continued fractions (finite, periodic, or finite parts of arbitrary
sails). We begin with some definitions and background. Further, we discuss one
inductive and two deductive algorithms to construct continued fractions. Finally, we
demonstrate one of the deductive algorithms on a particular example.

20.1 Inductive Algorithm

20.1.1 Some Background

A multidimensional periodic algebraic continued fraction is a set of infinite polyhe-

dral surfaces (i.e., sails) that contain an infinite number of faces. As we have already
mentioned, the quotient of every sail under the positive Dirichlet group action is
isomorphic to the n-dimensional torus. The algebraic periodicity of the polyhedron
allows us to reconstruct the whole continued fraction knowing only the fundamental
domain (i.e., the union of faces that contains exactly one face from each equivalence
class with respect to the action of the corresponding group Ξ+ ). Moreover, every
fundamental domain contains only a finite number of faces of the whole algebraic

O. Karpenkov, Geometry of Continued Fractions, 281

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_20, © Springer-Verlag Berlin Heidelberg 2013
282 20 On Construction of Multidimensional Continued Fractions

periodic continued fraction. Hence we are faced with the problem of finding a good
algorithm that enumerates all the faces for this domain.
There was no algorithm for constructing multidimensional continued fractions
until T. Shintani’s work [185] in 1976. The algorithm was further developed by
R. Okazaki [153] in 1993. The authors showed the so-called inductive method for
constructing fundamental domains of multidimensional continued fractions. The al-
gorithm produces the fundamental domain face by face, verifying that each new face
does not lie in the same orbit with some face constructed previously. Applying the
algorithm, one finds the fundamental domain in finitely many steps. In some sense
this is similar to the classical construction of one-dimensional continued fractions.
Using inductive algorithms, E. Thomas and A.T. Vasques obtained several fun-
damental domains for the two-dimensional case in [196]. Later, E. Korkina in [117–
119] and G. Lachaud in [123] produced an infinite number of fundamental domains
for periodic algebraic two-dimensional continued fractions.

20.1.2 Description of the Algorithm

For a polytope F and its hyperface E we say that the collection (E, F ) is a hyper-
flag. Every (n − 1)-dimensional face of a sail for an (n − 1)-dimensional continued
fraction is called a hyperface of the sail.

Inductive Algorithm for Constructing the Fundamental Domains for Sails of

Multidimensional Continued Fractions
Input data. We have a matrix A in SL(n, Z) and one of its invariant cones. We start
with an empty list of hyperflags L and empty fundamental domain D.
Goal of the algorithm. To construct one of the fundamental domains of the con-
tinued fraction for A in the given invariant cone.
Step 1a. Construct a hyperface F0 of the sail of the cone.
Step 1b. Calculate the face-structure of this face. Add all the hyperflags containing
F0 to the list L as unmarked.
Step 2. Take an unmarked hyperflag (F, E) of the list L and
– construct the hyperface F  adjacent to the hyperface F along E  ;
– check whether F  is not an image of another face F of the list D under the
action of some element of the positive Dirichlet group Ξ+ (A). If F  is not an
image, then add F  to the list D;
– mark (F, E) as marked; add (F  , E) as marked.
Step 3. Until the list L contains at least one unmarked hyperflag, go to Step 2.
Step 4. Find the face decomposition of all hyperfaces in the list D. It is the closure
of one of the fundamental domains.
Output. The face decomposition of the hyperfaces in the list D is the closure of
one of the fundamental domains of an (n − 1)-dimensional continued fraction.
20.1 Inductive Algorithm 283

Remark 20.1 The algorithm is relatively slow in Step 2, but the advantage is that
one does not need to construct the basis of the positive Dirichlet group. If we work
with nonalgebraic sails, then this algorithm constructs a finite domain of a sail (here
in Step 2, we only construct an adjacent face).

20.1.3 Step 1a: Construction of the First Hyperface

20.1.3.1 How to Calculate One Vertex of the Sail

Consider an arbitrary invariant cone C. Let us find some integer point in C. Shift the
basis unit parallelepiped completely into the cone C. There exists an integer point
lying inside the shifted parallelepiped. The coordinates of this point coincide with
integer parts of coordinates for one of the 2n vertices of this parallelepiped.
Having found some integer point P of the invariant cone C, let us construct some
vertex of the sail of C. Consider some integer plane π passing through the origin
such that
π ∩ C = O,
where O is the origin. Suppose that the integer distance from the point P to this
plane is equal to d. Now we look through all the simplices obtained as intersections
of the cone C with planes parallel to π at integer distances to the origin equal to
1, . . . , d. Suppose that the first simplex containing integer points lies in the plane at
integer distance d  ≤ d. The convex hull of all points of this simplex coincides with
some faces of the sail. All vertices of this face are vertices of the sail. Choose one
of them.

20.1.3.2 Inductive Construction of a Hyperface

Suppose that we have already constructed a face Fk of the sail of dimension k − 1

(where k < n). Let us construct a face of the sail of dimension k. Let π1 be an integer
plane of dimension n−2 that contains Fk , intersects the cone C in a compact set, and
does not pass through the origin O. Consider a two-dimensional plane π2 orthogonal
to π1 . Let us project the cone and all its inner integer points along π1 to π2 , denoting
this projection by f . The image of the cone f (C) is a two-dimensional cone. Letting
S be the set of all integer points inside the cone, then f (S) is contained in the cone
f (C). In addition, f (S) is contained in f (Zn ), which is a two-dimensional discrete
lattice (since π1 is integer).
Consider a line l in π2 passing through f (Fk ) and assume that this line cuts from
the cone f (C) a triangle that contains points of f (S). Take the convex hull of these
points in the triangle. One of the two neighboring points of f (S) to f (Fk ) in the
boundary of this convex hull is in the boundary of the convex hull for f (S) (we call
this point P ). Let s ∈ S and f (s) = P . Then the span of s and the face Fk is in some
284 20 On Construction of Multidimensional Continued Fractions

face of the sail of dimension k. The face Fk+1 consists of all integer points in the
tetrahedral intersection of the plane spanned by Fk and s with the cone C.
So we get the face Fk+1 . Proceed further until we get a hyperface Fn .

20.1.4 Step 1b, 4: How Decompose the Polytope into Its Faces

Let π be an oriented plane of dimension r, and let p0 , p1 , . . . , pr be points in this

plane. Define
⎧
⎪
⎨1 if p0 p1 , . . . , p0 pr defines a positive orientation,
signπ (p0 , p1 , . . . , pr ) = 0 if p0 p1 , . . . , p0 pr are dependent,
⎪
⎩
−1 if p0 p1 , . . . , p0 pr defines a negative orientation.

Proposition 20.2 Consider an r-dimensional convex polytope p1 . . . ps , and sup-

pose it spans the plane π of dimension r. Consider a subset of indices i1 < · · · <
ir ≤ s. Let the tetrahedron T = pi1 . . . pir have nonzero Euclidean (or integer)
(r − 1)-dimensional volume. Then we have:
(i) The tetrahedron T is in a hyperface if and only if the function

signπ (pj , pi1 , . . . , pir )

is either nonnegative or nonpositive simultaneously for all j ∈ {1, . . . , s}.

(ii) If T is in some hyperface, then this hyperface consists of points pj satisfying
the equation
signπ (pj , pi1 , . . . , pir ) = 0.

Proof The condition of the first item means exactly that all points are in one half-
plane with respect to the hyperplane containing the tetrahedron pi1 . . . pir .
The condition of the second item enumerates the points contained in the plane
of T . 

Remark 20.3 Using Proposition 20.2 one constructs all hyperfaces of a polyhe-
dron P . Further decomposition into faces is done iteratively, applying the hyperface
search to already constructed faces.

20.1.5 Step 2: Construction of the Adjacent Hyperface

Suppose that we know a hyperface F and one of its hyperfaces E. Let us construct
the hyperface FE adjacent to F via E.
20.2 Deductive Algorithms to Construct Sails 285

Let π1 be an integer plane of dimension n − 2 containing E and a two-

dimensional plane π2 orthogonal to π1 . Let us project the cone and all its inner
integer points along π1 to π2 , again denoting this projection by f . The image of
the cone f (C) is a two-dimensional cone. Letting S be the set of all integer points
inside the cone, then f (S) is contained in the cone f (C), and in addition, f (S) is
contained in f (Zn ), which is a two-dimensional discrete lattice (since π1 is integer).
The image of F is the segment AB, and the image of the vertex O of the cone is the
point denoted by O  . Consider a point in f (S) \ O AB, denoting it by P .
Without loss of generality, we suppose that the points A and P are in one half-
plane with respect to the line O B. The line AP cuts from the cone f (P ) a triangle
containing points of f (S). Take the convex hull of these points in the triangle. One
of the two neighboring points of f (S) to B in the boundary of this convex hull is
in the boundary of the convex hull for f (S) (we call this point Q). Let s ∈ S and
f (s) = Q. Then the span of s and the face E is in a hyperface of the sail adjacent
to E. The hyperface FE consists of all integer points in the tetrahedral intersection
of the plane spanned by E and s with the cone C.

20.1.6 Step 2: Test of the Equivalence Class for the Hyperface F 

to Have Representatives in the Set of Hyperfaces D

Consider two ordered bases (ei1 ) and (ei2 ) with matrices M1 and M2 in the standard
coordinate basis. It is clear that (ei1 ) is integer congruent to (ei2 ) if and only if

M1 (M2 )−1 ∈ SL(n, Z).

Here one needs to check that the elements of M1 (M2 )−1 are all integer and that the
determinant equals 1.
If in addition we are interested whether these two bases are in the same class
with respect to some Dirichlet group Ξ (A), then we should check that the matrix
M1 (M2 )−1 commutes with the matrix A.
Suppose that we are willing to test whether the equivalence class of a hyperface
F  has representatives in the set of faces D. Take an arbitrary n-tuple of independent
integer points in F  and the basis (ei ) of Rn corresponding to it. The test is positive if
there exist a face in D and an n-tuple of independent integer points in this face such
that the corresponding basis (gi ) is integer congruent to (ei ) and the transformation
matrix commutes with the matrix A defining the cone.

20.2 Deductive Algorithms to Construct Sails

20.2.1 General Idea of Deductive Algorithms

In inductive algorithms one constructs faces of the sail one by one inductively. There
is another approach to constructing faces of the sail for the cone C. First, one should
286 20 On Construction of Multidimensional Continued Fractions

construct an approximation of the sail itself, and, second, choose the faces of the
approximation that are the faces of the sail. This method is especially useful in
the algebraic periodic case. Suppose that we are given an integer irreducible real
spectrum matrix A ∈ SL(n + 1, Z). To compute some fundamental domain of a sail
of the continued fraction associated to A it is sufficient to do the following:
1. Compute a convex hull approximation of the sail. Namely, take a large enough
convenient set of integer points and find its convex hull.
2. Make a conjecture on some fundamental domain. Here we need to guess a set of
faces that might form a fundamental domain. We do this by finding a repeatable
pattern in the geometry of the faces.
3. Prove the conjecture if possible.
4. If you cannot prove the conjecture, start with 1, but with a larger convenient set
of points.
In this situation the following two questions are relevant:
How can one find a convenient set of integer points for the approximation of the
sail?
How can one test whether the conjecture of a fundamental domain of the sail is
true?
Let us give answers to these questions.

20.2.2 The First Deductive Algorithm

In [142] J.O. Moussafir developed an essentially different approach to constructing

continued fractions. It works for an arbitrary (not necessarily periodic) continued
fraction and computes any bounded part of an infinite polyhedron. The approach is
based on deduction. One produces a conjecture on the face structure for a large part
of the continued fraction. Then it remains to prove that every conjectured face is
indeed a face of the part. This method can be also applied to the case of periodic
continued fractions.

The Scheme of the First Deductive Algorithm

Input data. A cone C.
Goal of the algorithm. Construct a large enough set of faces in C.
Step 1. Find an approximation of the cone and make a face decomposition (better
if inside the approximated cone).
Step 2. Determine which faces are the real faces of the sail. Go to Step 1 until there
are enough faces.
Step 3. In the algebraic case, find the fundamental domain of the sail.
Output. A large enough set of faces in C.
20.2 Deductive Algorithms to Construct Sails 287

In the first step we approximate the cone in the following way. The cone is de-
fined by its edges, so we take a good approximation of all the edges, and get a cone
approximating our cone.
In the second step we should check that certain polyhedra are faces of the sail
for the cone. So we are solving the following problem: starting with a cone C with
vertex at the origin and an integer polyhedron F , find out whether F is a face of the
sail for C. Denote by πF an integer hyperplane containing F .
(i) Check that the intersection of the cone C with πF is a simplex, find all the
integer points inside it, and check that their convex hull coincides with F .
(ii) Find the integer distance d from the origin to F . Check that all integer planes
parallel to πF and having the origin and F in distinct half-planes (there are
exactly d − 1 such planes) do not have integer points in the intersection with
the cone C.
In the third step we find out which faces of the constructed part are shifted by
matrices of the group Ξ+ (A) to each other. This was discussed in Sect. 20.1.5.
As an example, we discuss one particular case of two-dimensional sails later, in
Sect. 20.2.4.

20.2.3 The Second Deductive Algorithm

In this section we describe another deductive construction adapted especially to

fundamental domains of periodic continued fractions. The construction involves a
method for conjecturing the structure of the fundamental domain and an algorithm
testing whether the conjectured domain is indeed fundamental. Usually the number
of “false” vertices of this approximation is much smaller than the number of “false”
vertices of the approximation produced by the first deductive algorithm.
Note that this algorithm substantially uses the periodicity of a multidimensional
continued fraction, and hence it is impossible to apply it to aperiodic continued
fractions.
Almost all known simple examples and series of examples of fundamental do-
mains were constructed using this algorithm (see [89]). In particular, the complete
list of all two-dimensional periodic continued fractions constructed by matrices of
small norm (| ∗ | ≤ 6) was found in [90] (see also Theorem 18.13 above).
Let us briefly itemize the main steps of the algorithm.

The Second Deductive Algorithm to Construct Fundamental Domains

Input data. A matrix A ∈ SL(n, Z) and one of its invariant cones.
Goal of the algorithm. Construct a fundamental domain for the sail of the given
matrix A for the given invariant cone.
Step 1. Calculate the basis of the additive group of the ring Γ (A).
Step 2. Calculate the basis of the group Ξ+ (A) (using the result of Step 1).
Step 3. Find some vertex of the sail (see Sect. 20.1.3).
288 20 On Construction of Multidimensional Continued Fractions

Step 4. Make a conjecture on a fundamental domain of the sail (using the results of
Step 2 and Step 3).
Step 5. Test the resulting conjecture from Step 4.
Output. A fundamental domain of an operator A in the given invariant cone.

Remark 20.4 It is assumed that the fundamental domain conjectured in Step 4 and
the basis A1 , . . . , An of the group Ξ+ (A) satisfy the following conditions:
(i) the closure of the fundamental domain is homeomorphic to the disk;
(ii) the operators with matrices A1 , . . . , An define the gluing of this disk to the n-
dimensional torus.

We show how to test conjectures in the case of two-dimensional continued frac-

tions in the description of Step 5. The result is partially based on Corollary 16.36
on the classification of two-dimensional faces at the integer distances to the origin
greater than one (see also in [91]). For the case of n-dimensional continued frac-
tions for n ≥ 3, the last step is quite complicated, since the classification of three-
dimensional faces at integer distances to the origin greater than one is unknown.

Remark 20.5 Note that all deductive algorithms are not algorithms in the strict
sense. One needs to choose some basis of Ξ+ (A) in the right way, produce a good
conjecture, and then test it. Even the algorithmic recognition of the period for the
given picture of a sail approximation is thought to be a hard problem. That is why
this “algorithm” cannot be completed with certainty by some computer program.
But on the other hand, this deductive algorithm is effective in practice. All of the ex-
amples listed in the article [89] were produced using this algorithm. The examples
of this paper generalize and expand almost all known periods of the sails calculated
before.

Remark 20.6 Deductive methods can be naturally generalized to the calculations

of periods of certain other periodic multidimensional continued fractions related
to cones (for more information we refer to the works by H. Minkowski [139] and
G.F. Voronoi [205]).

General questions concerning the lattice bases (Steps 1 and 2) were discussed in
Chap. 17 for the calculation of a vertex of the sail (Step 3); see Sect. 20.1.3. Step 5
is general for both deductive algorithms. It will be discussed in the next subsection.
So it remains to describe Step 4.

20.2.3.1 Step 4. How to Produce a Conjecture on the Fundamental Domain of

a Sail

Suppose that we know some vertex V of the sail in the invariant cone (from Step 3)
and a basis A1 , . . . , An for the group Ξ+ (A) (from Step 2). Let us briefly discuss
how to produce a conjecture on a fundamental domain of the sail. First, we compute
20.2 Deductive Algorithms to Construct Sails 289

the set of integer points that contains all vertices of some fundamental domain of
the sail. Second, we show how to choose an infinite sequence of special polyhedron
approximations for the sail. Finally, using a picture of these approximations, we
formulate a conjecture on a fundamental domain of the sail.

Proposition 20.7 Let V be a vertex of the sail of the n-dimensional continued frac-
tion of an (n + 1)-algebraic irrationality. Then there exists a fundamental domain
of the sail such that all vertices of this domain are contained in the convex hull H
of the origin and of the 2n distinct points of the form
 n 
& ε
Vε1 ,...,εn = Ai i (V ),
i=1

where εi ∈ {0, 1} for 1 ≤ i ≤ n.

Proof Consider the polyhedral cone C with vertex at the origin and base at the
convex polyhedron with vertices Vε1 ,...,εn . We take the union of all images of this
polyhedral cone under the actions of the operators with matrices

&
n
Am1 ,...,mn = Am i
i ,
i=1

for 1 ≤ i ≤ n, where mi ∈ Z. Obviously this union is equivalent to the union of

the whole open invariant cone and the origin. Therefore, every vertex of the sail
is obtained from a vertex contained in the cone C by applying an operator with
matrix Am1 ,...,mn for some integers mi . Moreover, the convex hull of all integer
points of the given invariant cone contains the convex hull of the vertices of the
form Am1 ,...,mn (V ). Hence the sail (i.e., the boundary of the convex hull of integer
points) is contained in the closure of the complement in the invariant cone to the
convex hull of all integer points of the form Am1 ,...,mn (V ). This complement is a
subset of the union of polyhedra obtained from H by an action of some operator
with matrix Am1 ,...,mn (for some integers mi , 1 ≤ i ≤ n). 

We skip the classical description of the computation of the convex hull for the
integer points contained in the polyhedron H . Denote the vertices of this convex
hull by Vr for 0 < r ≤ N . Here N is the total number of such points.

Definition 20.8 The convex hull of the finite set of points

 
Am1 ,...,mn (Vr ) | 1 ≤ mi ≤ m, ∀i : 0 ≤ i ≤ n

is called the nth special polyhedron approximation for the sail.

The defined set contains approximately, but fewer than, mn N points. (Since we
calculated some image points for the boundary of H several times, we do not
290 20 On Construction of Multidimensional Continued Fractions

know the exact number of points.) The number N is fixed for the given genera-
tors A1 , . . . , An while m varies. We should try to make a conjecture with the least
possible m.

Remark 20.9 For all the examples listed in the paper [89] (and for the example
of the last section) it was sufficient to take m = 2 to produce the corresponding
conjectures.

Remark 20.10 The “quality” of the approximation strongly depends on the choice
of the basis of Ξ+ (A).

Remark 20.11 Note that though the algorithm works for n + 1 arbitrary linearly
independent matrices of Ξ+ (A), it is not necessary to find generators of Ξ+ (A).
Suppose that we know matrices A1 , . . . , An that generate only some full-rank sub-
group of the group Ξ+ (A). Let the index of this subgroup be equal to k. Then we
are faced with the following two problems. Firstly, the number N will be approx-
imately k times greater than in the previous case. Secondly, one should also find a
conjecture on generators of the group Ξ+ (A).

20.2.4 Test of the Conjectures Produced in the Two-Dimensional

Case

In this section we explain how to test conjectures for the case of two-dimensional
periodic continued fractions (for more information we refer to [95]). The test con-
sists of seven stages. We prove here that these seven stages are sufficient for the
verification of whether the conjecture produced is true. The complexity of these
stages is polynomial in the number of all faces.

20.2.4.1 Brief Description of the Test Stages and Formulation of the Main
Results

Suppose that we have a conjecture on some fundamental domain D for some sail of a
two-dimensional periodic continued fraction associated to some integer irreducible
real spectrum matrix A. Let B1 , B2 be the basis of the group Ξ+ (A). Let pk (for k =
0, 1, 2) be the number of k-dimensional faces of the fundamental domain D. Denote
by Fi (i = 1, . . . , p2 ) the two-dimensional faces, i.e., polygons. All vertices and
edges adjacent to each face are known. It is also conjectured that the fundamental
domain D and the basis B1 , B2 satisfy the following conditions:
(i) the closure of the fundamental domain is homeomorphic to the two-dimensional
disk;
(ii) B1 and B2 define the gluing of this disk to the n-dimensional torus T 2 (the
fundamental domain D is in one-to-one correspondence with this torus).
20.2 Deductive Algorithms to Construct Sails 291

Test of the Conjecture Our test consists of the following seven stages:
1. test of condition (i);
2. test of condition (ii);
3. calculation of all integer distances from the origin to the two-dimensional planes
containing faces Fi and verification of their positivity;
4. test the nonexistence of integer points inside the pyramids with vertices at the
origin and bases at Fi (here the integer points of faces Fi are permitted);
5. test the convexity of dihedral angles (for all edges of the fundamental domain);
6. verification that all stars of the vertices are regular;
7. test whether all vertices of D are in the same invariant cone.
Let us give several technical definitions related to conditions 5 and 6.

Definition 20.12
(i) A dihedral angle is called well-placed if the origin is contained in the corre-
sponding opposite dihedral angle.
(ii) Consider an arbitrary polyhedral surface P and its vertex v. The n-star of the
vertex v in P is the union of all faces of P of dimension not greater than n
containing v.

Let us now define the regular stars at vertex v of a fundamental domain. First, we
construct a 2-star Rv of the conjectured sail using the algorithm of sail reconstruc-
tion (see page 252). Second, let p : W → T 2 be the universal covering of the torus
obtained by gluing the fundamental domain via the shift operators. Take any vertex
w ∈ W corresponding to the vertex v. Denote the 2-star at w ∈ W by Rw . Denote
the canonical projection of Rw to Rv by ξ .
If v
= (a, 0, 0) for some positive a, then we set v m = (1/m, 0, 0) for m ∈ Z+ . (If
v = (a, 0, 0), then we set v m = (0, 1/m, 0).) A 2-star at v is called regular if for the
sequence of rays lm with vertex at the origin and passing through the point v + v m
there exists a positive k such that for every m ≥ k the following holds: the set

ξ −1 (lm ∩ Ru )

consists of exactly one point.

Let us first formulate a theorem on complexity of the conjecture test.

Theorem 20.13 The described conjecture test for the fundamental domain D re-
quires every than
C(p0 + p1 + p2 )4
additions, multiplications, and comparisons of two integers, where C is a universal
constant that does not depend on the pi .

We skip the proof of this technical theorem here and refer the interested reader
to [95].
292 20 On Construction of Multidimensional Continued Fractions

Remark 20.14 Note that here we do not take into account the complexity of addi-
tions, multiplications, and comparison of two large integers. We think of any such
operation as a single operation (as a unit of time). There are some known bounds
for the number of digits of the integers that are linear with respect to the coefficients
of the matrix A. So the complexity should be multiplied by some polynomial in the
coefficients of A.

Theorem 20.15 Let the set of faces D satisfy the following conditions:
(1) condition (i);
(2) condition (ii);
(3) positivity of all integer distances from the origin to the two-dimensional planes
containing faces Fi ;
(4) there are no integer points inside the pyramids with vertices at the origin and
bases at Fi (here integer points on the faces Fi are permitted);
(5) all dihedral angles are convex;
(6) all stars of the vertices are regular;
(7) all vertices of D are contained in the same invariant cone.
Then D is a fundamental domain of some sail of the continued fraction associated
to the matrix A.

Let us prove that these seven stages are sufficient for the test.

20.2.4.2 Lemma on the Injectivity of the Face Projection

We prove Theorem 20.15 in four lemmas.

First let us give the necessary notation. Let the matrices B1 and B2 generate
Ξ+ (A). For any integers n and m, we write Bn,m for the matrix B1n B2m . We suppose
that our domain D satisfies Conditions 1–7 of Theorem 20.15. Let

U= Bn,m (D).
n,m∈Z

Consider the two-dimensional unit sphere S 2 centered at the origin O. We denote

by π the following map:
π : R3 \ O → S,
where every point x ∈ R3 \ O maps to the point at the intersection of S 2 and the ray
with vertex at the origin and containing x.

Lemma 20.16 For any face of the polygonal surface U , the map π is welldefined
and injective on it.
20.2 Deductive Algorithms to Construct Sails 293

Proof Consider any two-dimensional face F of the surface U . By condition 3, the

distance from the origin to the plane containing F is greater than zero. Hence this
plane does not contain the origin. Then π is welldefined and injective on F .
Let now E be some edge of U . By conditions 1 and 2, this edge is adjacent to
some two-dimensional face and therefore is contained in some plane that does not
pass through the origin. Thus the line containing E does not pass through the origin.
Hence π is well defined and injective on E.
The injectivity for the vertices is obvious. 

20.2.4.3 Lemma on the Finite Covering of the Fundamental Domain

Let x ∈ R3 \ O. Denote by Nx the tetrahedral angle with vertex at the origin and
base with vertices x, B1 (x), B1 B2 (x), and B2 (x). Notice that


Bn,m (Nx ) \ O
n,m∈Z

is one of eight invariant cones of the continued fraction associated to A that con-
tains x. Note that from conditions 1, 2, and 6 it follows that all points of D are
contained in one open invariant cone, which we denote by C.

Lemma 20.17 Let x be some point of the open invariant cone C. Then the union of
all faces of D is contained in a finite union of solid angles of the type Bn,m (Nx ).

Proof By Dirichlet’s unit theorem it follows that for every interior point a of the
open invariant cone C there exists an open neighborhood satisfying the follow-
ing condition. The neighborhood can be covered by four solid angles of the type
Bn,m (Nx ) when a belongs to an edge of some Bk,l (Nx ), by two solid angles when a
belongs to the face of some Bk,l (Nx ), and by one solid angle in the remaining cases.
In every case, the neighborhood can be covered by some finite union of solid angles
of type Bn,m (Nx ).
Consider a covering of D by such neighborhoods that correspond to each point of
the closure of D. Since the closure of D is closed and bounded in R3 , it is compact.
Hence this covering contains some finite subcovering. Therefore, the union of all
faces of D is contained in the finite union of solid angles of type Bn,m (Nx ). 

Corollary 20.18 Let x be contained in the open invariant cone C. Then the solid
angle Nx contains only points from a finite number of fundamental domains of the
type Bn,m (D).

Proof From the last lemma it follows that D is contained in the finite union
/ l
k=1 Bnk ,mk (Nx ) (for some positive l). Then the solid angle Nx can contain only
points of the fundamental domains B−nk ,−mk (D) for 1 ≤ k ≤ l. 
294 20 On Construction of Multidimensional Continued Fractions

20.2.4.4 Lemma on the Bijectivity of the Projection

Lemma 20.19 The map π bijectively takes the polygonal surface U to the set
S 2 ∩ C.

Proof As was shown above, the surface U is contained in C and is taken to S 2 ∩ C

under the map π .
Let us introduce the following notation. By condition 2, the action of the opera-
tors with matrices B1 and B2 determines a gluing of the fundamental domain. After
gluing we obtain a torus, denoted by T 2 . Let W be the universal covering of T 2 . The
face decomposition on T 2 lifts to a face decomposition on W . There is a natural two-
parameter family (with two integer parameters) of projections pn,m : W → U that
map faces to faces (since the group of shifts Bk,l acts on U ). Let us choose one of
these projections and denote it by p (p : W → U ).
Consider the map π ◦ p : W → S 2 . This map does not have branch points at the
images of open faces of W , since every face of W bijectively maps to some face of
U , and the corresponding face of U injectively maps to S 2 ∩ C by Lemma 20.16.
Two faces with a common edge of the universal covering W map to some two
faces with common edge of the surface U . Such faces of U generate a well-placed
dihedral angle, and hence also injectively map to S 2 ∩ C. So the map π ◦ p does not
have branch points at the images of open edges.
We now consider some vertex v of W . The edges and faces of W with common
vertex v, by condition 6, form a regular 2-star. These edges also map to some edges
of U with common vertex. Thus there exists a sequence of points that tends to
π ◦ p(v) (contained in S 2 ) such that the preimage of every point of the sequence has
exactly one preimage in the 2-star of v. Hence π ◦ p does not have branch points
in the branch containing the star at π ◦ p(v). Therefore, π ◦ p does not have any
branch points at the vertices. So the map π ◦ p : W → S 2 ∩ C does not have branch
points.
Consider an arbitrary point x ∈ S 2 ∩ C and the solid angle Nx corresponding to
it. Let x1 and x2 be two points of S 2 ∩ Nx . We will now show that the preimages
(π ◦ p)−1 (x1 ) and (π ◦ p)−1 (x2 ) contain the same number of points. Let us join
the points x1 and x2 by some curve inside S 2 ∩ Nx . By Corollary 20.18, we know
that the preimage of this curve is contained in a finite number of faces of W . Since
there are no branch points in any face (and their number is finite) and there are no
boundary faces of W , the number of preimages for π ◦ p is some (finite) discrete
and continuous function on this curve. Therefore, the number of preimages for π ◦ p
of any two points of S 2 ∩ Nx is the same. Hence the number of preimages for π ◦ p
of any two points of S 2 ∩ C is the same.
From this we conclude that the projection π ◦ p of the universal covering W
(homeomorphic to an open disk) to S 2 ∩ C (i.e., homeomorphic to an open disk) is
a nonramified covering with finitely many branches. Since the covering (of an open
disk by an open disk) is piecewise connected, the number of branches equals one.
Since by definition, p : W → U is surjective and by all of the above it is injective,
the maps p : W → U and π : U → S 2 ∩ C are bijective. 
20.2 Deductive Algorithms to Construct Sails 295

20.2.4.5 Lemma on Convexity

Since every ball centered at the origin contains only a finite number of vertices of U
(and they do not form a sequence tending to the boundary of the invariant cone C),
the polyhedral surface U divides the space R3 into two connected components. De-
note by H the connected component of the complement to U that does not contain
the origin.

Lemma 20.20 The set H is convex.

Proof Suppose that some plane passing through the origin intersects the polygonal
surface U and does not contain any vertex of U . By Lemma 20.19 such a plane
intersects U at some piecewise-connected broken line with an infinite number of
edges. The complement of the plane of this broken line consists of two connected
components. By assumption all vertices of this broken line are contained in open
edges of U . By condition 5 all dihedral angles of U are well-placed. Thus the angle
at every vertex of intersection of H with our plane is less than a straight angle.
Hence by the previous lemma the intersection is convex.
Consider the set of all planes that pass through the origin, intersect U , and do not
contain vertices of U . This set is dense in the set of all planes passing through the
origin and intersecting U . Therefore, by continuity it follows that the intersection of
H with any plane passing through the origin (and intersecting U ) is convex.
Now we prove that the set H is convex. Let x1 and x2 be some points of H .
Consider the plane that spans x1 , x2 , and the origin. This plane intersects U , since
x1 is in H and the origin is not in H . By the above, the intersection of H with this
plane is convex. Hence the segment with endpoints x1 and x2 is contained in H .
Thus H is convex (by the definition of convexity). 

20.2.4.6 Conclusion of the Proof of Theorem 20.15: the Main Part

So the constructed polygonal surface U has the following properties:

– by Lemma 20.20, U bounds the convex set H ;
– by construction, all vertices of U are integer points;
– by Condition 4, the set C \ H does not contain integer points.
Therefore, the polygonal surface U is the boundary of the convex hull of all
integer points inside C. Thus by definition, U is one of the sails of the continued
fraction associated to the matrix A. 
Let us formulate one important conjecture here.

Conjecture 27 Conditions 1–6 imply condition 7.

296 20 On Construction of Multidimensional Continued Fractions

20.2.5 On the Verification of a Conjecture for the

Multidimensional Case

Here we briefly outline an idea of how to test a conjecture on a fundamental domain

of a multidimensional continued fraction.

Conjecture for the Multidimensional Case Suppose that we have a conjec-

ture on some fundamental domain D, and also some basis B1 , . . . , Bn of the
group Ξ+ (A). Also let the fundamental domain and the basis have the following
properties:

(i) the closure of the fundamental domain is homeomorphic to the disk;

(ii) the operators with matrices B1 , . . . , Bn define the gluing of this disk to the n-
dimensional torus.

How Does One Test the Conjecture for Fundamental Domains of Multidi-
mensional Continued Fractions? The verification of conditions (i) and (ii) is
straightforward and is omitted. If these conditions hold, we check whether all the
n-dimensional faces of the fundamental domain are faces of the sail. This can be
done in the following way.
Suppose that the integer distances from the origin to the planes of faces Fi are
equal to di (i = 1, . . . , p, where p is the number of all n-dimensional faces). Our
conjecture is true if and only if for all i = 1, . . . , p the following conditions hold:

(a) For every integer d < di consider the plane parallel to the face Fi with integer
distances to the origin equal to d. The intersection of our invariant cone with
this plane does not contain any integer point.
(b) For d = di the convex hull of all integer points in the intersection coincides with
face Fi .

The verification of conditions (a) and (b) is quite complicated from the algorithmic
point of view.
We conclude this section with the important inverse question of constructing pe-
riodic continued fractions.

Problem 28 (V.I. Arnold) Does there exist an algorithm to decide whether a given
type of fundamental domain is realizable by a periodic continued fraction?

The answer to this question is unknown even for the two-dimensional periodic
continued fractions.
20.3 An Example of the Calculation of a Fundamental Domain 297

Fig. 20.1 The closure of the

fundamental domain of a sail
of a fraction associated to the
matrix Ab−a−1,(a+2)(b+1)
(here b = 6, and a is
arbitrary)

20.3 An Example of the Calculation of a Fundamental Domain

Let us construct fundamental domains for two-dimensional continued fractions of
Example 18.15. Recall that
⎛ ⎞
0 1 0
Am,n = ⎝0 0 1 ⎠.
1 −m −n

Theorem 20.21 Let m = b − a − 1, n = (a + 2)(b + 1) (where a, b ≥ 0). Consider

the sail for the matrix Am,n containing the point (0, 0, 1). Let

A = (1, 0, a + 2), B = (0, 0, 1),

 
C = (b − a − 1, 1, 0), D = (b + 1)2 , b + 1, 1 .

Then the following set of faces forms one of the fundamental domains:
(1) the vertex A;
(2) the edges AB, AD, and BD;
(3) the triangular faces ABD and BDC.

The closure of the fundamental domain is homeomorphic to the square shown in

Fig. 20.1 (for the case of an arbitrary a, and b = 6).

Proof of Theorem 20.21 Steps 1 and 2. We omit the first and the second steps (these
steps are classical; see [35]) and here write down the result. The following two
matrices generate the group Ξ+ (A):
 −1 
Xa,b = A−2m,n , Ya,b = A−1
m,n Am,n − (b + 1)Id ,

where Id is the identity element in the group SL(3, Z).

Step 3. We prove that (0, 0, 1) is a vertex of the sail. Consider the plane passing
through A, B, and D:

(−1 − a)x + (ab + a + b + 1)y + z = 1.

When the equations (in variables x, y, and z)

(−1 − a)x + (ab + a + b + 1)y + z = α

298 20 On Construction of Multidimensional Continued Fractions

do not have any integer solution for 0 < α < 1, the integer distance from ABD to
the origin is equal to one. There are exactly three integer points (A, B, and D) in
the intersection of the plane and the invariant cone. We leave the proof of this as an
exercise for the reader.
Step 4. The conjecture of the fundamental domain was produced in the statement
of this theorem.
Step 5. It remains to test the conjectured fundamental domain. For the test we
need some extra points:
−1
E = Xa,b (B)

= 1, −ab − a − 2b − 2, a 2 b2 + 2a 2 b

+ 4ab2 + a 2 + 8ab + 4b2 + 5a + 7b + 5 ;
F = Ya,b (B) = (−a − 2, 1, 0);
−1  
H = Xa,b (F ) = 0, −b − 1, ab2 + 2ab + 2b2 + a + 4b + 3 .

1. (Test of condition (i)). It can be shown in the usual way that the faces have com-
mon edge BD, and the edges intersect only at vertices. This implies that all ad-
jacencies are correct, and that only one or two faces are adjacent to each edge.
The closure of the boundary is a closed broken line ABCDA, homeomorphic to
the circle.
2. (Test of condition (ii)). Direct calculations show that the operator with matrix
Xa,b takes the segment AB to DC (the point A goes to the point D and B to
C). The operator with matrix Ya,b takes AD to BC (A goes to B and D to C).
Obviously no other points glue together. The Euler characteristic of the obtained
surface equals 2 − 3 + 1, i.e., zero, and the surface is orientable.
3. (Calculation of all integer distances from the origin to the two-dimensional
planes containing faces.) The integer distance to the plane of ABD equals
⎛ ⎞
 1 0 b 2 + 2b + 1 
1   b+1
· ⎝ 0 0 b + 1 ⎠ = = 1.
b+1  a+1 1 1  b+1

The integer distance from the origin to the plane of BDC equals
⎛ ⎞
 0 b2 + 2b + 1 b − a − 1 
 
1
· ⎝0 b+1 1 ⎠ = ab + 2b + a + 2 = a + 2.
b+1  1  b+1
1 0 

4. (Test on nonexistence of integer points inside the pyramids with vertices at the
origin and bases at the faces.) Since the integer distance from the origin to the
plane containing ABD equals one, the pyramid corresponding to ABD does not
contain integer points different from O and the points of the face ABD.
20.3 An Example of the Calculation of a Fundamental Domain 299

The face BDC is integer-linear congruent to the face with vertices (1, a + 1,
−a − 2), (b + 2, a + 1, −a − 2), (1, a + 2, −a − 2) of the list “T-W” (see
Fig. 15.2). The corresponding transformation taking BCD to the face of the list
“T-W” is as follows:
⎛ ⎞
b+1 b−a−1 b−a
⎝ 1 1 1 ⎠.
0 −1 −1

By Theorem 15.5 the pyramid corresponding to BDC does not contain integer
points different from O and the points of the face BDC.
5. (Test on convexity of dihedral angles.) Let us first consider the edge BD. This
edge is adjacent to the faces ABD and BDC. The face ABD is contained
in the plane fABD (x, y, z) = 0, and the face BDC is contained in the plane
fBDC (x, y, z) = 0, where

fABD (x, y, z) = (−1 − a)x + (ab + a + b + 1)y + z − 1;

fBDC (x, y, z) = x + (b + 1)y − (a + 2)z + (a + 2).

To test that the dihedral angle corresponding to the edge BD is well-placed it is

sufficient to verify the following: the point C and the origin O lie in different
half-spaces with respect to the plane spanned by the points A, B, and D and the
points A and O lie in different half-spaces with respect to the plane spanned by
the points C, B, and D. So we need to solve the following system:

fABD (C) · fABD (O) < 0,
fBDC (A) · fBDC (O) < 0.

This system is equivalent to


(a 2 + 3a + 2) · (−1) < 0,
(−a 2 − 3a − 1) · (a + 2) < 0.

Since a ≥ 0, the inequalities hold. Thus the dihedral angle associated with
the edge BD is well-placed. Since the cases of dihedral angles associated to the
edges AB (and the faces ADB and AEB) and BC (and the faces BDC and CBF)
can be verified in the same way, we omit their descriptions. This concludes the
test of Condition 5.
6. (Verification that all 2-stars of the vertices are regular.) There is only one vertex
in the torus decomposition. Any lift of this point to the universal covering W is
adjacent to six edges and six faces. Consider a vertex of the universal covering
that maps to the point B. The corresponding 2-star maps to six edges BC, BD,
BA, BE, BH, and BF and to six faces BCD, BDA, BAE, BEH, BHF, and BFC,
where
−1  
H = Xa,b (F ) = 0, −b − 1, ab2 + 2ab + 2b2 + a + 4b + 3 .
300 20 On Construction of Multidimensional Continued Fractions

We will check that for every sufficiently small positive ε, a ray lε with vertex at
the origin and passing through the point Pε = (ε, 0, 1) intersects exactly one of
the faces of 2-stars.
First we will check that for every sufficiently small positive ε, the ray lε in-
tersects the triangle BCF, or equivalently, that the ray lε is contained in the tri-
hedral angle with vertex at the origin O and base in the triangle BCF. The two-
dimensional face of the trihedral angle containing B, C, and O can be defined
by fABO = 0, and the two-dimensional face of the trihedral angle containing B,
F , and O can be defined by fBFO = 0; the two-dimensional face of the trihedral
angle containing C, F , and O can be defined by fCFO = 0, where

fBCO (x, y, z) = x + (a + 1 − b)y;

fBFO (x, y, z) = x + (a + 2)y;
fCFO (x, y, z) = z.

For every sufficiently small positive ε, the ray lε is contained in the dihedral
angle defined above if the following conditions hold: the points Pε and F are
in the same closed half-space with respect to the plane fBCO = 0, the points Pε
and C are in the same closed half-space with respect to the plane fBFO = 0, and
the points Pε and B are in the same closed half-space with respect to the plane
fCFO = 0. Since the points Pε and B are close to each other for sufficiently small
ε, they are in the same closed half-space with respect to the plane fCFO = 0. We
now check the remaining two conditions:
 
fBCO (Pε ) · fBCO (F ) ≥ 0, (−b − 1)ε ≥ 0,
⇔
fBFO (Pε ) · fBFO (C) ≥ 0, (b + 1)ε ≥ 0.

Since b, ε ≥ 0, the first inequality does not hold. Thus for every sufficiently small
positive ε, the ray lε does not intersect the triangle BCF.
The cases of the triangles BCD, BDA, BAE, BEH, and BHF are similar to
those described above and are omitted here.
The ray lε (for any sufficiently small positive ε) intersects the bijective im-
age of the 2-star of the vertex at exactly one point contained in the edge AB.
Therefore, all 2-stars associated to the vertices are regular.
7. (Test that all the vertices of D are in the same invariant cone.) The test of the
seventh stage for this theorem is trivial, since D contains exactly one vertex. 

20.4 Exercise

Exercise 20.1 Construct fundamental domains of the matrices in Examples 18.9–

18.12 and prove that the corresponding triangulations are correct.
Chapter 21
Gauss Reduction in Higher Dimensions

In this chapter we continue to study integer conjugacy classes of integer matrices in

general dimension. Recall that matrices A and B in SL(n, Z) are integer conjugate
if there exists a matrix C in GL(n, Z) such that

B = CAC −1 .

The natural problem here is as follows: describe the set of integer conjugacy classes
in SL(n, Z). In this section we give an answer to this question for matrices whose
characteristic polynomials are irreducible over the field of rational numbers.
Gauss‘s reduction theory gives a complete geometric description of conjugacy
classes for the case n = 2, as we have already discussed in Chap. 7. In the multi-
dimensional case the situation is more complicated. It is relatively simple to check
whether two given matrices are integer conjugate (see, for instance, [4] and [73]),
but to distinguish conjugacy classes is a much harder task. In this chapter we de-
scribe a generalization of Gauss’s reduction theory to the multidimensional case
(see also [99]). We study questions related to the three-dimensional case in more
detail.
There exists an alternative algebraic approach to the study of SL(n, Z) conjugacy
classes, which we do not touch in this book. There one starts with GL(n, Q) con-
jugacy classes and splits them into GL(n, Z) conjugacy classes. This reduces the
problem to certain problems related to orders of algebraic fields that are defined by
the roots of the characteristic polynomial of the corresponding matrices.

21.1 Organization of This Chapter

In this chapter we investigate the following five aspects.
I. Generalized reduced matrices. In the multidimensional Gauss’s reduction the-
ory, Hessenberg matrices play the role of reduced matrices. A Hessenberg ma-
trix is a matrix that vanishes below the superdiagonal (see [193]); they were in-
troduced by K. Hessenberg in [78]. Hessenberg matrices were essentially used

O. Karpenkov, Geometry of Continued Fractions, 301

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_21, © Springer-Verlag Berlin Heidelberg 2013
302 21 Gauss Reduction in Higher Dimensions

in the QR-algorithm for the eigenvalue problem (see [53, 198], and [154]). Fur-
ther, in [99] they were considered in the framework of the multidimensional
Gauss reduction theory as a multidimensional analogue of reduced matrices.
In Sect. 21.2 we introduce a natural notion of Hessenberg complexity for
Hessenberg matrices. Further we show that each integer conjugacy class of
SL(n, Z) has only finitely many distinct Hessenberg matrices with minimal
complexity (Theorem 21.8).
II. Complete invariant of integer conjugacy classes. In Sect. 21.3 we introduce
a complete geometric invariant of a Dirichlet group. It is a periodic multidi-
mensional continued fractions in the sense of Klein–Voronoi (Theorem 21.23).
Klein–Voronoi continued fractions naturally generalize Klein continued frac-
tions. Further, we deduce the complete invariants of integer conjugacy classes
of GL(n, Z) matrices. We show that the conjugacy classes are represented by
periodic shifts of Klein–Voronoi periodic continued fractions (Theorem 21.24).
III. Calculation of reduced matrices. In Sect. 21.4 we introduce a technique to con-
struct reduced matrices in any integer conjugacy class that has minimal Hessen-
berg complexity in this class.
IV. Solution of Diophantine equations related to certain decomposable forms. In
Sect. 21.5 we study the integer points at the level sets of the decomposable
forms (we have already discussed this question in Chap. 7 in connection with
the Markov spectrum). As a technical detail we introduce a notion of multi-
dimensional w-continued fractions similar to its one-dimensional version of
Chap. 10.
V. Integer conjugacy classes of SL(3, Z). We examine the structure of integer con-
jugacy classes of SL(3, Z) via Klein–Voronoi continued fractions. The main re-
sults concern only the operators with a pair of complex conjugate eigenvectors.
It turns out that in this case, Hessenberg matrices distinguish corresponding
conjugacy classes asymptotically (Theorem 21.48). The analogous statement
does not hold for the case of operators with three real eigenvalues. In general it
is much less known for the case of three real eigenvalues.

21.2 Hessenberg Matrices and Conjugacy Classes

In Sect. 21.2.1 we begin with necessary definitions and notation related to Hessen-
berg matrices. Further, in Sect. 21.2.2 we discuss algorithmic aspects of construc-
tion of perfect Hessenberg matrices integer congruent to a given one. In Sect. 21.2.3
we show that every integer conjugacy class with irreducible characteristic polyno-
mial has a finite nonzero number of ς -reduced matrices. Further in Sect. 21.2.4 we
study the set of perfect Hessenberg matrices. We think of this set as a “book” whose
“pages” are enumerated by Hessenberg types. Since the matrices from the same
page are distinguished by their polynomials, any two matrices on one page are not
integer congruent. We conclude this section with a brief discussion of ς -reduced
2-dimensional matrices of the same Hessenberg type.
21.2 Hessenberg Matrices and Conjugacy Classes 303

21.2.1 Notions and Definitions

In this subsection we briefly introduce Hessenberg matrices, which we consider as

generalized reduced matrices in the multidimensional case.

21.2.1.1 Perfect Hessenberg Matrices

A matrix M of the form

⎛ ⎞
a1,1 a1,2 ··· a1,n−2 a1,n−1 a1,n
⎜a2,1 a2,2 ··· a2,n−2 a2,n−1 a2,n ⎟
⎜ ⎟
⎜ 0 a3,2 ··· a3,n−2 a3,n−1 a3,n ⎟
⎜ ⎟
⎜ .. .. .. .. .. .. ⎟
⎜ . . . . . . ⎟
⎜ ⎟
⎝ 0 0 ··· an−1,n−2 an−1,n−1 an−1,n ⎠
0 0 ··· 0 an,n−1 an,n

is called an (upper) Hessenberg matrix. We say that the matrix M is of Hessenberg

type
a1,1 , a2,1 |a1,2 , a2,2 , a3,2 | · · · |a1,n−1 , a2,n−1 , . . . , an,n−1 .

Definition 21.1 A Hessenberg matrix in SL(n, Z) is said to be perfect if for every

pair of integers (i, j ) satisfying 1 ≤ i < j + 1 ≤ n the following inequalities hold:
0 ≤ ai,j < aj +1,j .

In other words, all elements in the first n − 1 columns of a perfect Hessenberg

matrix are nonnegative integers, and in addition, the element aj +1,j is maximal in
each of these columns.

21.2.1.2 Reduced Hessenberg Matrices

We study only the simplest general case of SL(n, Z) matrices whose characteris-
tic polynomials are irreducible over Q. Every such matrix is integer conjugate to a
perfect Hessenberg matrix, and each conjugacy class of SL(n, Z) contains infinitely
many perfect Hessenberg matrices. To reduce the number of such matrices, we in-
troduce a natural notion of complexity.

Definition 21.2 Consider a Hessenberg matrix M = (ai,j ). The integer

&
n−1
|aj +1,j |n−j
j =1

is called the Hessenberg complexity of M. Denote it by ς(M).

304 21 Gauss Reduction in Higher Dimensions

The Hessenberg complexity is in fact equivalent to the volume of the paral-

lelepiped spanned by v = (1, 0, . . . , 0), M(v), M 2 (v), . . . , M n−1 (v), as we will dis-
cuss in Sect. 21.4.1.
An integer Hessenberg matrix has unit Hessenberg complexity if and only if

a2,1 = · · · = an,n−1 = 1.

In the literature such matrices are known as Frobenius matrices (or companion ma-
trices). The elements of the last column of any Frobenius matrix are exactly the
coefficients of the characteristic polynomial multiplied alternately by ±1.

Example 21.3 Consider the matrix

⎛ ⎞
1 2 3
⎝2 3 6 ⎠.
0 5 −1

This matrix is a perfect Hessenberg matrix of type 1, 2|2, 3, 5. Its Hessenberg
complexity is 22 · 5 = 20.

Definition 21.4 A perfect Hessenberg matrix M is said to be ς -nonreduced if there

exists an integer matrix M  of smaller Hessenberg complexity integer congruent
to M. Otherwise, we say that M is ς -reduced.

Remark Let us say a few words about the relation between reduced matrices in-
troduced in Chap. 7 and ς -reduced matrices in SL(2, Z). On the one hand, every
reduced real spectrum matrix is a perfect Hessenberg matrix. On the other hand,
every ς -reduced matrix is also reduced, although there are certain reduced matrices
that are not ς -reduced.

Later, in Theorem 21.8, we show that the number of ς -reduced matrices is finite
for every integer conjugacy class. Some classes contain more than one ς -reduced
perfect Hessenberg matrix.

Example 21.5 The ς -reduced Hessenberg matrices (with Hessenberg complexity

equivalent to 3)
⎛ ⎞ ⎛ ⎞
0 1 2 0 2 3
M1 = ⎝1 0 0⎠ and M2 = ⎝1 1 1⎠
0 3 5 0 3 4

are integer conjugate.

21.2 Hessenberg Matrices and Conjugacy Classes 305

21.2.2 Construction of Perfect Hessenberg Matrices Conjugate to

a Given One

In this subsection we show how to construct perfect Hessenberg matrices that are
integer congruent to a given one. The main ingredients of the algorithm are in the
proof of the following proposition.

Proposition 21.6 Consider an SL(n, Z) matrix M whose characteristic polynomial

is irreducible over Q. Then for every vector v of integer length 1 there exists a unique
matrix C satisfying the following conditions:
– C(1, 0, . . . , 0) = v;
– the matrix CMC −1 is perfect Hessenberg (we denote this matrix by (M|v)).

In other words, for a given integer operator A, every integer vector of unit integer
length is uniquely extended to the basis of the integer lattice in which the operator
A has a perfect Hessenberg matrix.

Proof Existence. Consider an SL(n, Z) matrix M with irreducible characteristic

polynomial. Let A denotes the linear operator with matrix M. Consider an integer
vector v of unit integer length. For i = 1, . . . , n − 1 we define
 
Vi = Span v, A(v), A2 (v), . . . , Ai−1 (v) .

Since the characteristic polynomial of M is irreducible, the set of all spaces Vi forms
a complete flag in Rn . All the vectors Aj (v) are integer vectors; hence every space
Vi contains a full-rank integer sublattice.
Let us inductively construct an integer basis {ẽi } of a vector space Rn such that:
– for every i the vectors ẽ1 , . . . , ẽi generate the sublattice Zn ∩ Vi ;
– the matrix of A is perfect Hessenberg in the basis {ẽi }.
Base of induction. Let ẽ1 = v. Since l(ẽ1 ) = 1, the vector ẽ1 generates Zn ∩ V1 .
Step of induction. Suppose that we have already constructed ẽ1 , . . . , ẽk forming
a basis of Zn ∩ Vk . Let us now calculate ẽk+1 . Choose an integer vector gk+1 of the
space Vk+1 a unit integer distance to the space Vk . Since ẽ1 , . . . , ẽk generate Zn ∩Vk ,
the vectors ẽ1 , . . . , ẽk , gk+1 generate the sublattice Zn ∩ Vk+1 . Since A(ẽk ) ∈ Vk+1 ,
we have

k
A(ẽk ) = qi,k ẽi + ak+1,k gk+1 .
i=1

For i = 1, . . . , k, we define bi,k and ai,k as integer quotients and remainders of the
following equations:

qi,k = bi,k · |ak+1,k | + ai,k , where 0 ≤ ai,k < ak+1,k .

306 21 Gauss Reduction in Higher Dimensions

Then we have
 

k 
k
A(ẽk ) = |ak+1,k | sign(ak+1,k )gk+1 + bi,k ẽi + ai,k ẽi
i=1 i=1

(where sign is the sing function over the real numbers). Finally, we get


k
ẽk+1 = sign(ak+1,k )gk+1 + bi,k ẽi .
i=1

The characteristic polynomial of A is irreducible over Q, and hence the integer

spaces Vi are not invariant subspaces of A. Therefore, the set {ẽ1 , . . . , ẽk+1 } is a
basis of Zn ∩ Vk+1 . The induction step is complete.
Denote by M̂ the matrix of A in the basis {ẽi }. By construction, the matrix M̂ is
a perfect Hessenberg matrix, and its Hessenberg type is
0    1
a1,1 , |a2,1 |a1,2 , a2,2 , |a3,2 | · · · a1,n−1 , . . . , an−1,n−1 , |an,n−1 | .

Both M and M̂ are matrices of A in two different integer bases, which means they
are integer conjugate. The integers ai+1,i are nonzero (i = 1, . . . , n − 1), since the
characteristic polynomial of A is irreducible. Therefore, ς(M̂) > 0. Denote by C
the transition matrix to the basis {ẽi }. Then C(ẽ1 ) = v and M̂ = CMC −1 (where M̂
is a perfect Hessenberg matrix).
Uniqueness. Let C1 and C2 satisfy the conditions of the theorem. Suppose that
C1 and C2 are the transition matrices to the bases {ẽi } and {êi } respectively. Then

ẽ1 = C1 (e1 ) = v = C2 (e1 ) = ê1 .

For i > 1 the equality ẽi = êi follows from the uniqueness of the choice of the coef-
ficients in the proof of existence. Hence the bases {ẽi } and {êi } coincide. Therefore,
C1 = C 2 . 

Now let us briefly summarize the algorithm.

Algorithm to Construct Perfect Hessenberg Matrices

Input data. (M, v). Here M is a matrix of a lattice-preserving operator A whose
characteristic polynomial is irreducible over q, and v is an integer vector of unit
integer length.
Step 1. Put ẽ1 = v.
Inductive Step k. We have already constructed ẽ1 , . . . , ẽk . Choose gk+1 ∈ Vk+1
satisfying ld(gk+1 , Vk ) = 1 and find integers qi,k for i = 1, . . . , k and ak+1,k
from the decomposition


k
A(ẽk ) = qi,k ẽi + ak+1,k gk+1 .
i=1
21.2 Hessenberg Matrices and Conjugacy Classes 307

Find bi,k and ai,k (i = 1, . . . , k) as integer quotients and reminders of the equa-
tions
qi,k = bi,k · |ak+1,k | + ai,k .
Put

k
ẽk+1 = sign(ak+1,k )gk+1 + bi,k ẽi .
i=1

Output data. The perfect Hessenberg matrix CMC −1 , where C is the transition
matrix to {ẽk }.
Note that the basis {ẽi } is constructed in n steps.
Later, in the proof of Theorem 21.31, we use the following corollary.

Corollary 21.7 Consider an SL(n, Z) operator A with matrix M and let B ∈

Ξ (A). Then for an arbitrary v we have (M|v) = (M|B(v)).

Proof Each step of the algorithm produces the same data for v and B(v), due to the
fact that A and B commute. Therefore, (M|v) = (M|B(v)). 

21.2.3 Existence and Finiteness of ς -Reduced Hessenberg

Matrices

It turns out that ς -reduced Hessenberg matrices are good for the classification of
conjugacy classes in SL(n, Z). Although they do not form complete invariants of
conjugacy classes, they are present in each class and their number is finite. We give
an explicit construction of all ς -reduced Hessenberg matrices conjugate to a given
one in Sect. 21.4.3 via Klein–Voronoi continued fractions.

Theorem 21.8 Let M be an SL(n, Z) matrix. Then the number of ς -reduced Hes-
senberg matrices integer conjugate to M is finite and greater than zero.

The proof of this theorem is based on the following proposition.

Proposition 21.9 Every Hessenberg matrix with positive Hessenberg complexity is

identified by its Hessenberg type and characteristic polynomial.

Proof Let M = (ai,j ) be a Hessenberg matrix with positive Hessenberg complexity.

The first n − 1 columns of M are entirely defined by the Hessenberg type of M. The
last column is uniquely defined from the characteristic polynomial of M,

x n + cn−1 x n−1 + · · · + c1 x + c0 ,
308 21 Gauss Reduction in Higher Dimensions

in the following way. For every k the coefficient ck is a polynomial in variables

ai,j that does not depend on a1,n , . . . , ak,n . This polynomial has a unique monomial
containing ak+1,n , namely
 n−1 
&
aj +1,j ak+1,n .
j =k+1

Since ς(M)
= 0, the product in the parentheses is nonzero. Hence ak+1,n is uniquely
defined by ck and the elements ai,j of the first n−1 columns. Therefore, all elements
of M are defined by the Hessenberg type of M and the characteristic polynomial
of M. 

Proof of Theorem 21.8 Existence. By Proposition 21.6 there exist perfect Hessen-
berg matrices integer conjugate to M. The set of values of Hessenberg complexity
is discrete and bounded from below; hence there exists a perfect Hessenberg matrix
M̃ integer conjugate to M and with minimal possible Hessenberg complexity. By
definition the matrix M̃ is a ς -reduced.
Finiteness. Let c be the Hessenberg complexity of M̃. By definition the Hessen-
berg complexity of all the other ς -reduced Hessenberg matrices integer conjugate to
M equals c. The number of Hessenberg types whose Hessenberg complexity equals
c is finite. Notice that the integer conjugate matrices have the same characteris-
tic polynomials. From Proposition 21.9 there exists at most one integer Hessenberg
matrix with a given Hessenberg type and a given polynomial. Therefore, the number
of ς -reduced Hessenberg matrices integer conjugate to M is finite. 

21.2.4 Families of Hessenberg Matrices with Given Hessenberg

Type

In this subsection we discuss the structure of Hessenberg matrices with given Hes-
senberg type. We begin with a particular example.

Example 21.10 Let us examine the Hessenberg type 0, 1|1, 0, 2. The set of all
matrices of this Hessenberg type is a two-parameter family with parameters m and
n:
⎧⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎫
  ⎨ 0 1 1 0 0 0 0 0 1  ⎬
H 0, 1|1, 0, 2 = ⎝1 0 0⎠ + m ⎝0 0 1⎠ + n ⎝0 0 0⎠  m, n ∈ Z .
⎩ ⎭
0 2 1 0 0 0 0 0 2

Define
⎛ ⎞
0 1 n+1
H0,1|1,0,2 (m, n) = ⎝1 0 m ⎠.
(1,0,1)

0 2 2n + 1
21.2 Hessenberg Matrices and Conjugacy Classes 309

Fig. 21.1 Matrices of Hessenberg type 0, 1|1, 0, 2

(1,0,1)
The discriminant of H0,1|1,0,2 (m, n) equals

−44 − 44n2 − 56mn − 32n3 + 32m3

+ 16m2 n2 + 16mn2 + 16m2 n − 56n − 8m + 52m2 .

The set of matrices with negative discriminant for the given family coincides with
the union of the sets of integer solutions of the following quadratic inequalities:

2m ≤ −n2 − n − 2 and 2n ≤ m2 + m.
(1,0,1)
In Fig. 21.1 we represent a matrix H0,1|1,0,2 (m, n) by the square in the inter-
section of the mth column and the nth row. Matrices with reducible characteris-
tic polynomials correspond to black squares. Light gray squares represent matrices
with three real eigenvalues. Matrices shown as white squares have a pair of complex
conjugate eigenvalues.

It turns out that the general case is similar. Consider a Hessenberg type

Ω = a1,1 , a2,1 |a1,2 , a2,2 , a3,2 | · · · |a1,n−1 , . . . , an−1,n−1 , an,n−1 .

Denote by H (Ω) the set of all Hessenberg matrices in SL(n, Z) of Hessenberg

type Ω.
For k = 1, . . . , n − 1 we put

vk (Ω) = (ak,1 , . . . , ak,k+1 , 0, . . . , 0)

310 21 Gauss Reduction in Higher Dimensions

and denote by Mk (Ω) the matrix whose first n − 1 columns are equal to zero and
the last one equals to vk (Ω). Denote also by σ (Ω) the (n − 1)-dimensional simplex
with vertices
O (the origin), v1 , . . . , vn−1 .

Theorem 21.11
(i) The set H (Ω) is not empty if and only if lV(σ (Ω)) = 1.
(ii) Let M0 ∈ H (Ω). Then H (Ω) is an integer affine (n − 1)-dimensional sublattice
in the lattice of all integer (n × n) matrices, i.e.,
 $

n−1 

H (Ω) = M0 + ci Mi (Ω)  c1 , . . . , cn−1 ∈ Z .
i=1

The proof of Theorem 21.11 is based on the following lemma.

Lemma 21.12 Consider an operator A with integer Hessenberg matrix M of

type Ω. Let v be the vector standing in the last column of M. Then M ∈ GL(n, Z)
if and only if the following conditions hold:
– lV(σ (Ω)) = 1;
– ld(v, Span(σ (Ω))) = 1.

Proof Necessary condition. Consider an operator A with Hessenberg GL(n, Z) ma-

trix M in some integer basis {gi }. Denote by Sgn−1 the (n − 1)-dimensional simplex
with vertices
O, g1 , . . . , gn−1 .
Since M ∈ GL(n, Z), the operator A preserves integer volumes and integer dis-
tances. Since lV(Sgn−1 ) = 1, we have the first condition:
      
lV σ (Ω) = lV A Sgn−1 = lV Sgn−1 = 1.

Notice that
    
A Span Sgn−1 = Span σ (Ω) and A(gn ) = v.

Therefore, we get the second condition:

        
ld v, Span σ (Ω) = ld gn , A Span Sgn−1 = ld gn , Span(g1 , . . . , gn−1 ) = 1.

Sufficient condition. Suppose that both conditions of the lemma hold. Then the
operator A preserves the integer lattice (generated by g1 , . . . , gn ). Therefore, M ∈
SL(n, Z). 

Proof of Theorem 21.11 (i) Suppose that lV(σ (Ω)) = 1. Choose an integer vector
v at unit integer distance to the plane Span(σ (Ω)). By Lemma 21.12 both matrices
21.2 Hessenberg Matrices and Conjugacy Classes 311

defined by Ω and vectors ±v are in GL(n, Z). One of them is in SL(n, Z). Con-
versely, if H (Ω) contains an SL(n, Z) matrix, then by Lemma 21.12 the integer
volume of σ (Ω) equals 1.
Statement (ii) is straightforward, since the determinant of a matrix is an additive
function with respect to the operation of vector addition in the last column. 

21.2.5 ς -Reduced Matrices in the 2-Dimensional Case

Let us say a few words about ς -reduced 2-dimensional matrices in families H (Ω).
We start with the following example.

Example 21.13 Consider the family of matrices H (3, 4):

3 3m + 2
.
4 4m + 3

Their Hessenberg complexity equals 4. Almost all matrices of this family have
irreducible, over Q, characteristic polynomials with two real roots (except for
m = −1, −2). For m = 0, 1, 2, 3, . . . the LLS-periods of the matrices are

(2, 2), (1, 2, 1, 1), (1, 2, 1, 2), (1, 2, 1, 3), ..., (1, 2, 1, m), ....

The matrices are ς -reduced for m ≥ 2. For m = −3, −4, −5, . . . the corresponding
LLS-periods are

(4, 1), (4, 2), (4, 3), (4, 4), ..., (4, −2 − m), ....

Starting from m ≤ −6 the matrices are ς -reduced.

The example of Hessenberg type 3, 4 shows that almost all matrices of this
Hessenberg type are ς -reduced. That is actually the case for all Hessenberg types in
SL(2, Z).

Theorem 21.14 Almost all matrices of a given Hessenberg type in SL(2, Z) are
ς -reduced.

This follows from Theorem 21.31 below and direct calculation of complexities
for all vertices of the period; we skip the proof here.
In Theorem 21.48 we prove similar behavior for SL(3, Z) matrices having two
complex conjugate eigenvalues (see also Conjecture 29).
312 21 Gauss Reduction in Higher Dimensions

21.3 Complete Geometric Invariant of Conjugacy Classes

In this section we introduce a geometric complete invariant of integer conjugacy

classes: multidimensional continued fractions in the sense of Klein–Voronoi. We
start with necessary definitions in Sect. 21.3.1. In particular, we give a general defi-
nition of Klein–Voronoi continued fractions in all dimensions in Sect. 21.3.1.1 and
discuss their algebraic periodicity. Further, in Sect. 21.3.2 we show that the Klein–
Voronoi continued fraction for an operator A is a geometric complete invariant of
the corresponding Dirichlet group Ξ (A). Finally, in Sect. 21.3.3 we discuss that
periodic shifts distinguish integer conjugacy classes of SL(n, Z) matrices.

21.3.1 Continued Fractions in the Sense of Klein–Voronoi

Approximately at the time of the work by F. Klein on continued fractions related

to real matrices, G. Voronoi in his dissertation [205] introduced a geometric algo-
rithmic definition that covers cases of matrices having complex conjugate eigenval-
ues. In [28] and [29] J.A. Buchmann generalized Voronoi’s algorithm, making it
more convenient for computation of fundamental units in orders. We use ideas of
J.A. Buchmann to define the multidimensional continued fraction in the sense of
Klein–Voronoi for all cases. Note that if all eigenvalues of an operator are real then
the Klein–Voronoi multidimensional continued fraction is a continued fraction in
the sense of Klein.

21.3.1.1 General Definitions

Consider an operator A in GL(n, R) with distinct eigenvalues. Suppose that it has k

real eigenvalues r1 , . . . , rk and 2l complex conjugate eigenvalues c1 , c1 , . . . , cl , cl ,
where k + 2l = n. Denote by LR (A) the space spanned by the real eigenvectors.
Denote by T l (A) the set of all real operators commuting with A such that their
real eigenvalues are all units and the absolute values for all complex eigenvalues
equal one. In other words:
 
TA = B ∈ GL(n, R) | AB = BA, spec(B) ⊂ S 1 , B|LR (A) = Id |LR (A) ,

where spec(B) is the spectrum of B and S 1 is the complex unit circle. In fact, T l (A)
is an abelian group with the operation of matrix multiplication.
For a vector v in Rn , we denote by TA (v) the orbit of v with respect to the
action of the group of operators T l (A). If v is in general position with respect to the
operator A (i.e., it does not lie in invariant planes of A), then TA (v) is homeomorphic
to the l-dimensional torus. For a vector of an invariant plane of A the orbit TA (v) is
also homeomorphic to a torus of positive dimension not greater than l, or to a point.
21.3 Complete Geometric Invariant of Conjugacy Classes 313

Example 21.15 Suppose that A is a real spectrum matrix, i.e., l = 0. Since all its
eigenvectors are real, T 0 (A) consists only of the unit operator and TA (v) = {v}.

Example 21.16 Now consider the case of a pair complex conjugate eigenvalues,
i.e., l = 1. The group T 1 (A) corresponds to elliptic rotations in the invariant plane
of A corresponding to complex eigenvalues. Such rotations are parameterized by
the angle of rotation. A general orbit of TA (v) is an ellipse around the (n − 2)-
dimensional invariant subspace corresponding to real eigenvalues. Every orbit in
the invariant subspace spanned by real eigenvalues consists of one point.

Let gi be a real eigenvector with eigenvalue ri for i = 1, . . . , k, and let gk+2j −1

and gk+2j be vectors corresponding to the real and imaginary parts of some complex
eigenvector with eigenvalue cj for j = 1, . . . , l. Consider the coordinate system
corresponding to the basis {gi }:
OX 1 X2 . . . Xk Y1 Z1 Y2 Z2 . . . Yl Zl .
Denote by π the (k + l)-dimensional plane OX 1 X2 . . . Xk Y1 Y2 . . . Yl . Let π+ be
the cone in the plane π defined by the equations yi ≥ 0 for i = 1, . . . , l. For every v
the orbit TA (v) intersects the cone π+ in a unique point.

Definition 21.17 A point p in the cone π+ is said to be π -integer if the orbit TA (p)
contains at least one integer point.

Consider all (real) hyperplanes invariant under the action of the operator A. There
are exactly k such hyperplanes. In the above coordinates, the ith of them is defined
by the equation xi = 0. The complement to the union of all invariant hyperplanes in
the cone π+ consists of 2k arcwise connected components C1 (A), . . . , C2k (A).

Definition 21.18 The convex hull of all π -integer points except the origin contained
in an arcwise connected component Ci (A) is called a factor-sail of Ci (A). The set
of all factor-sails is said to be the factor-continued fraction for the operator A. We
denote it by Ŝi (A).

Definition 21.19 The union of all orbits TA (∗) in Rn represented by the points in
the factor-sail of Ci (A) is called the sail of TA (Ci ), which we denote by Si (A). The
union of all sails is said to be the continued fraction for the operator A in the sense
of Klein–Voronoi (see Fig. 21.2 below). We denote it by KVCF(A).

In algebraic language we write the factor sail as

   
Ŝi (A) = ∂ conv q ∈ π+ | TA (q) ∩ Ci (A) ∩ Zn
= ∅ \ {O} ,
and hence
k
 
2
Si (A) = TA (p) and KVCF(A) = Si (A).
p∈Ŝi (A) i=1
314 21 Gauss Reduction in Higher Dimensions

Fig. 21.2 A three-dimensional example (Example 21.22): a the cone π+ and the eigenplane; b the
continued factor-fraction; c a sail of the continued fraction

Definition 21.20 The intersection of a factor-sail with a hyperplane in π is said

to be an m-dimensional face of the factor-sail if it is homeomorphic to the m-
dimensional disk.
The union of all orbits in Rn represented by points in some face of the factor-sail
is called the orbit-face of the operator A.
Integer points of the sail are said to be vertices of this sail.

In the simplest possible cases of k + l = 1, every factor-sail of A is a point. If

k + l > 1, then every factor-sail of A is an infinite polyhedral surface homeomorphic
to Rk+l−1 .
21.3 Complete Geometric Invariant of Conjugacy Classes 315

21.3.1.2 Algebraic Continued Fractions

Consider now an operator A in the group GL(n, Z) whose characteristic polynomial

is irreducible over Q. Suppose that it has k real roots and 2l complex conjugate
roots, where k + 2l = n.
Further, we consider the Dirichlet group Ξ (A) of all GL(n, Z) operators com-
muting with A and the corresponding positive Dirichlet group Ξ+ (A) (for more
details see Chap. 17). The Dirichlet group Ξ (A) takes a Klein–Voronoi continued
fraction to itself and permutes the sails. The positive Dirichlet group Ξ+ (A) consists
exactly of operators preserving every sail. By the Dirichlet unit theorem, the group
Ξ (A) is homomorphic to Zk+l−1 ⊕ G, where G is a finite abelian group. The group
Ξ+ (A) is homeomorphic to Zk+l−1 and its action on any sail is free. The quotient of
any sail by the action of Ξ+ (A) is homeomorphic to the (n − 1)-dimensional torus.

Definition 21.21 A fundamental domain of KVCF(A) is a collection of orbit-faces

of KVCF(A) representing exactly one of the different classes of KVCF(A)/Ξ (A).
A fundamental domain of a sail Si (A) is a collection of orbit-faces of Si (A)
representing exactly one of the different classes of Si (A)/Ξ+ (A).

Example 21.22 Let us study an operator A with a Frobenius matrix

⎛ ⎞
0 0 1
⎝1 0 1⎠ .
0 1 3

This operator has one real and two complex conjugate eigenvalues. Therefore, the
cone π+ for A is a two-dimensional half-plane. In Fig. 21.2a the half-plane π+ is
shaded in light gray and the invariant plane corresponding to the pair of complex
eigenvectors is in dark gray. The vector shown in Fig. 21.2a with endpoint at the
origin is an eigenvector of A.
In Fig. 21.2b we show the cone π+ . The invariant plane separates π+ into two
parts. The dots on π+ are the π -integer points. The boundaries of the convex hulls
in each part of π+ are two factor-sails. One factor-sail is taken to another by the
induced action of − Id.
Finally, in Fig. 21.2c we show one of the sails. Three orbit-vertices shown in
the figure correspond to the vectors (1, 0, 0), (0, 1, 0), and (0, 0, 1): the large dark
points (0, 1, 0) and (0, 0, 1) are visible at the corresponding orbit-vertices.
The positive Dirichlet group Ξ+ (A) in our example is homeomorphic to Z, and
is generated by A. The group Ξ (A) is homeomorphic to Z ⊕ Z/2Z with generators
A and − Id. The operator A takes the point (1, 0, 0) and its orbit-vertex to the point
(0, 1, 0) and the corresponding orbit-vertex. Therefore, every fundamental domain
of the continued fraction for the operator A contains one orbit-vertex and one ver-
tex edge. For instance, we can choose the orbit-vertex corresponding to the point
(1, 0, 0) and the orbit-edge corresponding to the “tube” connecting orbit-vectors for
the points (1, 0, 0) and (0, 1, 0).
316 21 Gauss Reduction in Higher Dimensions

21.3.2 Geometric Complete Invariants of Dirichlet Groups

One of the main properties of Klein–Voronoi continued fractions is that they classify
all Dirichlet groups.

Theorem 21.23 Let A, B ∈ GL(n, Z) have characteristic polynomials irreducible

over Q. Then Ξ (A) = Ξ (B) if and only if KVCF(A) = KVCF(B).

Remark If the characteristic polynomial of a matrix is irreducible over Q, then all its
eigenvectors are distinct, so all matrices of Proposition 21.23 possess Klein–Voronoi
continued fractions.

Proof of Theorem 21.23 Supposing that Ξ (A) = Ξ (B) then A and B commute.
Hence they have the same eigenvectors (since they do not have multiple eigenvalues)
and the same orbits: TA = TB . In addition, we choose the same cone π+ for both A
and B. Therefore, by definition, the Klein–Voronoi continued fractions for A and B
coincide.
Let us prove the converse statement. Assume that the Klein–Voronoi continued
fractions for A and B coincide. Suppose that A √ has real eigenvectors g1 , . . . , gk
and complex conjugate eigenvectors gk+2j −1 ± −1gk+2j for j = 1, . . . , l, where
k + 2l = n. Consider the coordinate system corresponding to the basis {gi }:

OX 1 X2 . . . Xk Y1 Z1 Y2 Z2 . . . Yl Zl .

In these coordinates define the form ΦA as follows:

 k 
& & l
 2 
ΦA (x1 , . . . , xn ) = xi yj + zj .
2

i=1 j =1

Similarly we define the form ΦB for the operator B. Note that A preserves the form
ΦA up to a multiplicative scalar: a simple calculation shows that for every v ∈ Rn
we have
 
ΦA A(v) = det(A)ΦA (v).
The same is true for B and ΦB .
The Klein–Voronoi continued fraction for A (which is also KVCF(B)) asymp-
totically coincides with the set ΦA = 0 (and ΦB = 0 respectively) at infinity, and
hence the matrices A and B have all the same invariant subspaces. In particular,
their one-dimensional real eigenspaces corresponding to real eigenvectors and their
two-dimensional eigenspaces (we denote them by π1 , . . . , πl ) defined by pairs of
complex conjugate roots coincide. This implies that A and B commute if and only
if they commute for the vectors of the invariant planes π1 , . . ., πl .
Let us show that operators A and B restricted to a plane πi (i = 1, . . . , l) com-
mute. For a vertex v of KVCF(A) we set
 
Tv = KVCF(A) ∩ v + π1 , . . . , πl  .
21.3 Complete Geometric Invariant of Conjugacy Classes 317

From construction, Tv coincides with the orbit TA (v). Since the space spanned by
all planes π1 , . . . , πl is invariant for both A and B, the set Tv coincides with the
orbit TB (v), and hence TA (v) = TB (v).
Consider an arbitrary matrix C of the set TA . Since all eigenvalues of C have
unit modulus and C is diagonalizable in the eigenbasis of A, C preserves ΦA up to
a scalar det(C) = ±1. Therefore, |ΦA | is constant on TA (v). For the same reason,
|ΦB | is constant on TA (v) = TB (v). Therefore, by linearity, ΦA = c · ΦB for some
constant c
= 0. Hence, A preserves ΦB up to a multiplicative scalar.
Consider now the plane πj for some 1 ≤ j ≤ l and take coordinates (yj , zj ). It
is clear that if an operator preserves ΦB (v) up to a multiplicative scalar, then its
restriction to the plane πj preserves the form

yj2 + zj2

up to a multiplicative factor. There are two types of matrices that preserve the set of
level sets of this form:

a b −a b
and
−b a b a

with arbitrary real parameters a and b. The matrices of the second family have two
real eigenvalues in πj , which is by the above not the case for B. Therefore, both A
and B are from the first family. All matrices of the first family commute. Hence A
commutes with B in planes πj for 1 ≤ j ≤ l.
Therefore, A and B are both diagonalizable in a certain complex basis. Hence A
and B commute. Therefore, Ξ (A) = Ξ (B). 

21.3.3 Geometric Invariants of Conjugacy Classes

In this subsection we fix some integer basis in Rn and identify operators with
matrices in this basis. On the one hand, from Theorem 21.23 it follows that the
Klein–Voronoi continued fraction for A identifies the Dirichlet group Ξ (A) in a
unique way. On the other hand, the operator A defines a shift of KVCF(A) along
itself, which we denote by PA . It is clear that distinct operators of Ξ (A) de-
fine inequivalent shifts. So every A ∈ GL(n, Z) is uniquely identified with a pair
(KVCF(A), PA ). The group GL(n, Z) naturally acts on pairs (KVCF(A), PA ) by
left multiplication on the first factor and by conjugation on the second factor. We
have the following important tautological theorem.

Theorem 21.24 (On geometric complete invariant) Two matrices A1 , A2 of the

group GL(n, Z) whose characteristic polynomial is irreducible over Q are integer
conjugate if and only if the pairs (KVCF(A1 ), PA1 ) and (KVCF(A2 ), PA2 ) are in
the same GL(n, Z)-orbit.
318 21 Gauss Reduction in Higher Dimensions

Remark 21.25 The important consequence of this theorem is that all geometric
GL(n, Z)-invariants of the Klein–Voronoi continued fraction for A (such as integer
distances and integer volumes of certain integer point configurations of KVCF(A))
are invariants of the conjugacy class of A. Fundamental domains of Klein continued
fractions and their invariants in the real spectrum three-dimensional case (n = k = 3,
l = 0) were studied in works [89, 90, 118, 119, 123, 125], etc. Currently there is not
much known about the construction of Klein–Voronoi continued fractions in the
nonreal spectrum case.

21.4 Algorithmic Aspects of Reduction to ς -Reduced Matrices

In Sect. 21.2 above we showed the existence and finiteness of ς -reduced matrices
in each integer conjugacy class of SL(n, Z) matrices. The aim of this section is to
introduce techniques to construct ς -reduced matrices integer conjugate to a given
one. In Sect. 21.4.1 we give a geometric interpretation of Hessenberg complexity
as a volume of a certain simplex, which is called the MD-characteristic. Further, in
Sect. 21.4.2 we use the MD-characteristics to prove that the algorithm of Sect. 21.2.2
constructs all ς -reduced matrices starting from integer vertices of a Klein–Voronoi
continued fraction. The corresponding techniques are outlined in Sect. 21.4.3.

21.4.1 Markov–Davenport Characteristics

In this subsection we characterize the Hessenberg complexity in terms of the

Markov–Davenport characteristic.

21.4.1.1 Definition of the MD-Characteristic and Its Invariance Under the

Action of the Dirichlet Group

The study of Markov–Davenport characteristics is closely related to the theory of

minima of absolute values of homogeneous forms with integer coefficients in n-
variables of degree n. One of the first works in this area was written by A. Markov
[135] for decomposable quadratic forms in two variables. Further, H. Davenport, in
a series of works [46–49], and [50], made the first steps for the case of decomposable
forms for n = 3.
Consider A ∈ SL(n, Z). Denote by P (A, v) the parallelepiped spanned by vec-
tors v, A(v), . . . , An−1 (v), i.e.,
 $

n−1 

P (A, v) = O + λi A (v)  0 ≤ λi ≤ 1, i = 0, . . . , n − 1 ,
i

i=0

where O is the origin.

21.4 Algorithmic Aspects of Reduction to ς -Reduced Matrices 319

Definition 21.26 The Markov–Davenport characteristic (or the MD-characteristic,

for short) of an SL(n, Z) operator A is the functional
 
ΔA : Rn → R defined by ΔA (v) = V P (A, v) ,

where V (P (A, v)) is the nonoriented volume of P (A, v).

Proposition 21.27 Consider A ∈ SL(n, Z) and let B ∈ Ξ (A). Then for an arbitrary
point v we have
 
ΔA (v) = ΔA B(v) .

Remark Proposition 21.27 implies that the MD-characteristic naturally defines a

function over the set of all orbits of the Dirichlet group.

Proof of Proposition 21.27 Since B ∈ Ξ (A), we have the equality An B(v) =

BAn (v). Hence the parallelepiped P (A, B(v)) coincides with B(P (A, v)). Since
B ∈ SL(n, Z), the volume of the parallelepiped is preserved. Therefore,
 
ΔA (v) = ΔA B(v) . 

21.4.1.2 Homogeneous Forms Associated to SL(n, Z) Operators

Consider any SL(n, Z) operator A. Suppose that it has k real eigenvectors√ g1 , . . . , g k

with eigenvalues r1 , . . . , rk and 2l complex eigenvectors gk+2j −1 ± −1gk+2j with
complex conjugate eigenvalues c1 , c1 , . . . , cl , cl , where k + 2l = n. Consider the
system of coordinates

OX 1 X2 . . . Xk Y1 Z1 Y2 Z2 . . . Yl Zl

corresponding to the basis {gi }. A form

 k 
& & l
 2 
α xi yj + zj
2

i=1 j =1

with nonzero α is said to be associated to the operator A.

Proposition 21.28 Let A be an SL(n, Z) operator whose characteristic polynomial

has distinct roots. Then the MD-characteristic of A coincides with the absolute
value of a form associated to A for a certain nonzero α.

Proof Let us consider the formulas of the MD-characteristic of A in the eigenbasis.

We assume that the coordinates in this eigenbasis are (t1 , . . . , tn ). Then for any
vector v = (t1 , . . . , tn ) we have
 j j j j j j 
Aj (x) = r1 t1 , . . . , rk tk , c1 tk+1 , c1 tk+2 , . . . , cl tk+2l−1 , cl tk+2l .
320 21 Gauss Reduction in Higher Dimensions

Therefore,
 k   k 
& & l  α & &
l
 2 
   
ΔA (t1 , . . . , tn ) = α  ti (tk+2j −1 tt+2j ) = l  xi yj + zj2 .
  4  
i=1 j =1 i=1 j =1

Simple calculations show that α

= 0. 

21.4.1.3 Hessenberg Complexity in Terms of the MD-Characteristic

Proposition 21.29 Consider an operator A with Hessenberg matrix M = (ai,j ) in

some integer basis {ẽi }. The Hessenberg complexity ς(M) equals the value of the
MD-characteristic ΔA (ẽ1 ).

Proof Denote by Vk the plane spanned by vectors v, A(v), A2 (v), . . . , Ak−1 (v).
Let us inductively show that
 k 
&
A (ẽ1 ) =
k
ai,i+1 ẽk+1 + vk , where vk ∈ Vk .
i=1

Base of induction. We have A(ẽ1 ) = a1,2 ẽ2 + a1,1 ẽ1 .

Step of induction. Suppose that the statement holds for k = m, i.e.,
m 
&
A (ẽ1 ) =
m
ai,i+1 ẽm+1 + vm , and vm ∈ Vm .
i=1

Let us prove the statement for m + 1. Since M is Hessenberg, A(vm ) is in Vm+1 .

Therefore, we have
 m  
&
A m+1
(ẽ1 ) = A ai,i+1 ẽm+1 + A(vm )
i=1
m+1 
&
= ai,i+1 ẽm+1
i=1
   
&
m
 
+ A(vm ) + ai,i+1 A(ẽm+1 ) − am+1,m+2 ẽm+2 .
i=1

The second summand in the last expression is in Vm+1 . We have completed the
induction step.
Therefore,
&
n−1
ΔA (ẽ1 ) = |ai+1,i |n−i = ς(M).

i=1
21.4 Algorithmic Aspects of Reduction to ς -Reduced Matrices 321

In the proof of the next theorem we use the following corollary.

Corollary 21.30 Consider an operator A with matrix M. Let v be any primitive in-
teger vector. Consider the matrix (M|v) constructed by the algorithm of Sect. 21.2.2.
Then we have
ς(M|v) = ΔA (v).

21.4.2 Klein–Voronoi Continued Fractions and Minima of

MD-Characteristics

In the following theorem we use Klein–Voronoi continued fractions to find minima

of MD-characteristics.

Theorem 21.31 Let A be an SL(n, Z) operator with matrix M in some integer

basis. Suppose that the characteristic polynomial of A is irreducible over Q. Let
U be a fundamental domain of the Klein–Voronoi continued fractions for A (see
Definition 21.21). Then we have:
(i) For every ς -reduced matrix M̂ integer conjugate to M, there exists v ∈ U such
that M̂ = (M|v).
(ii) Let v ∈ U . The matrix (M|v) is ς -reduced if and only if the MD-characteristic
ΔA attains its minimal value on the set Zn \ {O} at the point v.

Proof Theorem 21.31(ii) follows directly from Corollary 21.30.

Let us prove Theorem 21.31(i). By Proposition 21.28 there exists a nonzero con-
stant α such that in the system of coordinates OX 1 X2 . . . Xk Y1 Z1 Y2 Z2 . . . Yl Zl the
MD-characteristic ΦA is expressed as follows:
 k 
& & l
 2 
 2 
α  xi yi + zi .
 
i=1 i=1

Suppose that the minimal absolute value of ΦA on the set of integer points except
the origin equals m0 .
Consider a cone π+ in the plane z1 = · · · = zl = 0 and choose the coordinates
(x1 , . . . , xk , y1 , . . . , yl ) in it. Consider a projection of Rn to the cone along the orbits
TA (∗). Since the MD-characteristic is constant on TA (u) for every u, the projection
of the MD-characteristic is well defined. Denote it by Φ̃A . In the chosen coordinates
of π+ , the function Φ̃A is written as follows:
 k 
& & l 
 2
α xj y j .
 
j =1 j =1
322 21 Gauss Reduction in Higher Dimensions

This function is convex in every orthant of the cone π+ . Since every factor-sail is
the boundary of a certain convex hull in each orthant, all minima of the convex
function Φ̃A restricted to the convex hulls are attained at the boundary, i.e., at π -
integer points of factor-sails. Therefore, all integer minima of ΦA on Zn \ {O} are
at vertices of the Klein–Voronoi continued fraction.
By Corollary 21.30, the Hessenberg complexity ς(M|v) coincides with the MD-
characteristic ΔA (v). Since every matrix integer conjugate to M has a presentation
in the form (M|v) and all integer minima of the MD-characteristic are attained at
vertices of the Klein–Voronoi continued fraction, every ς -reduced operator M̃ is
represented as (M|v0 ) for some vertex v0 ∈ KVCF(A). By Corollary 21.7, for every
B ∈ Ξ (A) we have
 
M|B(v0 ) = (M|v0 ).
Hence a vector v0 can be chosen in the fundamental domain U . This concludes the
proof. 

Remark 21.32 In Example 21.5 the ς -reduced Hessenberg matrices

⎛ ⎞ ⎛ ⎞
0 1 2 0 2 3
M1 = ⎝1 0 0⎠ and M2 = ⎝1 1 1⎠
0 3 5 0 3 4

are integer conjugate but do not coincide. The reason for this is as follows. Con-
sider the Klein–Voronoi continued fraction of A with matrix M1 . It contains in-
teger vertices p1 = (1, 0, 0) and p2 = (0, 1, −1). It turns out that p1 and p2 are
not in the same orbit of the Dirichlet group but have the same MD-characteristic,
namely 3. Hence we get two distinct integer conjugate ς -reduced Hessenberg ma-
trices: M1 = (M1 |(1, 0, 0)) and M2 = (M1 |(0, 1, −1)).

21.4.3 Construction of ς -Reduced Matrices by Klein–Voronoi

Continued Fractions

As we have already proved, the ς -reduced Hessenberg matrix for the operator A
is constructed starting from some vertex in a fundamental domain of the Klein–
Voronoi multidimensional continued fraction. We use this property to find all ς -
reduced Hessenberg matrices in a given integer conjugacy class of matrices.

Algorithm (Techniques) to Find ς -Reduced Matrices in Integer Conjugacy

Classes
Input data. An integer matrix M for A.
Goal of the algorithm. To construct all ς -reduced matrices integer conjugate
to M.
21.4 Algorithmic Aspects of Reduction to ς -Reduced Matrices 323

Step 1. Find a fundamental domain of the Klein–Voronoi continued fraction for the
operator A (see Remark 21.33 below).
Step 2. Take all vertices of the fundamental domain constructed in Step 1 and
find among them all vertices with minimal value of the MD-characteristic (say
v1 , . . . , vk ).
Step 3. By Theorem 21.31(i) and (ii) all ς -reduced matrices integer conjugate to
M are (M|v1 ), . . . , (M|vk ). They are all constructed by the algorithm described
in Sect. 21.2.2.
Output. A list of all ς -reduced matrices integer conjugate to M.

Remark 21.33 Currently, Step 1 is the most complicated. The real spectrum case of
matrices with all eigenvalues being real has been well studied. For the algorithms
of constructing multidimensional continued fractions in this case, see Chap. 20 and
the papers by R. Okazaki [153], T. Shintani [185], J.-O. Moussafir [142], and the
author [95]. E. Korkina in [119] and [118], G. Lachaud in [123, 125], A.D. Bruno
and V.I. Parusnikov in [27, 161, 162], and [163], and the author in [89] and [90]
have produced a large number of fundamental domains for periodic algebraic two-
dimensional continued fractions (see also the site [25] by K. Briggs). Some funda-
mental domains in the three-dimensional case are found, for instance, in [92]. The
case with complex conjugate eigenvalues which we study in the next section, is new.

Example 21.34 Let us consider the example of the operator A defined by the matrix
⎛ ⎞
−2 −4 −3
⎝1 2 2 ⎠.
−1 −1 3

The characteristic polynomial of this operator has three distinct real roots. There-
fore, the Klein–Voronoi continued fraction consists of eight sails. The compositions
of operators − Id, A, and 2 Id +A−1 define integer congruences for all these sails,
and hence all ς -reduced operators are written from vertices of one sail. Consider
a sail containing the point (1, 0, 0). There are exactly three distinct orbits of the
Dirichlet group containing the vertices in this sail. They are defined by the follow-
ing points:
(0, 0, 1), (1, 0, 0), and (3, −1, 1).
(We skip all calculations of convex hulls corresponding to the sail.) The MD-
characteristic of these vectors are respectively 1, 2, and 4. So the minimum of the
MD-characteristic (which is 1 in this case) is attained on the vertices of the orbit of
the Dirichlet group containing (0, 0, 1). Therefore, there exists a unique ς -reduced
Hessenberg matrix, which is
⎛ ⎞
0 0 1
⎝1 0 1 ⎠ .
0 1 −3
324 21 Gauss Reduction in Higher Dimensions

The perfect Hessenberg matrices for the vertices (1, 0, 0) and (3, −1, 1) are respec-
tively
⎛ ⎞ ⎛ ⎞
0 1 −1 1 0 −1
⎝1 0 0 ⎠ and ⎝2 0 3 ⎠ ,
0 2 −3 0 1 −4
and their ς -complexities are 2 and 4.

21.5 Diophantine Equations Related to the Markov–Davenport

Characteristic

Let A be an arbitrary SL(n, Z) operator whose characteristic polynomial is irre-

ducible over Q, and let N be an arbitrary integer. In this section we discuss how to
solve the Diophantine equation

ΔA (v) = N. (21.1)

We begin with a general definition of w-continued fractions.

21.5.1 Multidimensional w-Sails and w-Continued Fractions

We have already introduced the w-sails of continued fractions in the one-

dimensional case in Sect. 10.2.5. Let us generalize this notion to the multidimen-
sional case.

21.5.1.1 Definition

Consider an operator A in SL(n, Z) whose characteristic polynomial is irreducible

over Q. As above, we suppose that A has k real eigenvalues and l pairs of com-
plex conjugate eigenvalues, where k + 2l = n. Recall that the complement to the
union of all invariant hyperplanes in the cone π+ consists of 2k arcwise connected
components denoted by C1 (A), . . . , C2k (A).

Definition 21.35 We define the n-sails for an arbitrary simplicial cone C with ver-
tex at the origin inductively.
– let the 1-sail be the sail of C.
– suppose that all w-sails for w < w0 are defined. The convex hull of all π -integer
points except the origin and except for the π -integer points for all w-factor-sails
with w < w0 contained in an arcwise connected component Ci (A) is called the
w0 -factor-sail of Ci (A). We denote it by Ŝw0 ,i (A).
21.5 Diophantine Equations Related to the Markov–Davenport Characteristic 325

Let us write the w0 -factor-sail of Ci in one formula:

  
0 −1
w
  
Ŝw0 ,i (A) = ∂ conv q ∈ π+ | TA (q) ∩ Ci (A) ∩ Zn
= ∅ \ {O} \ Ŝw,i (A) .
w=1

The set of all w-factor-sails for a given operator is called the w-factor-continued
fraction for the operator A.

Definition 21.36 The union of all orbits TA (∗) in Rn represented by the points in
the w-factor-sail of Ci (A) is called the w-sail of TA (Ci ); we denote it by Sw,i (A).
The union of all sails is said to be the w-continued fraction for the operator A in
the sense of Klein–Voronoi. We denote it by KVCF w (A).

In other words, we have

k
 
2
Sw,i = TA (p) and KVCF w (A) = Sw,i (A).
p∈Ŝw,i (A) i=1

21.5.1.2 Minima of Markov–Davenport Characteristics on w-Sails

Denote by αw the minimal absolute value of the Markov–Davenport characteristic

|ΔA | on KVCF w (A).

Proposition 21.37 The sequence α1 , α2 , α3 , . . . is strictly increasing.

Remark 21.38 The w-sails for arbitrary w > 0 are not necessarily homothetic to the
1-sail, as it was in the one-dimensional case of Sect. 10.2.5.

Proof of Proposition 21.37 Consider the coordinate system

OX 1 X2 . . . Xk Y1 Z1 Y2 Z2 . . . Yl Zl

introduced in Sect. 21.3.1.1 above. In these coordinates the form ΔA is written as

follows:
 k 
& & l
 2 
ΔA (x1 , . . . , xn ) = c xi yj + zj ,
2

i=1 j =1

for some nonzero constant c. Consider an arbitrary ray r with vertex at the ori-
gin and not contained in any invariant subspace of A. Assume that its direction is
(λ1 , . . . , λn ). Then at points of this ray the function ΔA is as follows:
 k 
& & l
 2  n
ΔA (λ1 t, . . . , λn t) = c λi λk+2j −1 + λk+2j t .
2
(21.2)
i=1 j =1
326 21 Gauss Reduction in Higher Dimensions

It is clear that the ray r intersects the w-sail first and only thereafter intersects the
(w + 1)-sail. Notice that the expression on the right side of Eq. (21.2) is proportional
to t n (with respect to t), which increases as t increases. Therefore, the value of ΔA
at the point of intersection of the ray r with the (w + 1)-sail is strictly greater than
the value at the point of intersection with the corresponding w-sail.
Notice that every possible limit value of ΔA on the (w + 1)-sail is attained at
some point of its fundamental domain (since the closure of a fundamental domain
is compact). In particular, the infimum of |ΔA | at the (w + 1)-sail is attained as a
minimum at some point. Therefore, αw+1 > αw . 

Corollary 21.39 For an arbitrary integer w, αw ≥ w.

Proof The convexity of a Markov–Davenport form on the intersection of cones Ci

with the plane π+ implies that the minimum αw is attained at an integer vertex of the
KVCF w (A). Therefore, αw is a positive integer for every w ≥ 0. Now the statement
of the corollary follows directly from Proposition 21.37. 

21.5.2 Solution of Eq. (21.1)

The following general techniques helps to write all the solutions of Eq. (21.1).

Techniques to Solve Eq. (21.1)

Input data. Given a pair (A, N ), where A is a GL(n, Z) matrix whose characteris-
tic polynomial is irreducible over Q, and N is an integer.
Goal of the algorithm. Find all integer solutions of ΔA (v) = N .
Step 1. Write the MD-characteristic ΔA according to Definition 21.26.
Step 2. Find the fundamental domains of all w-sails Cw,i for the operator A (for all
admissible i) and w ≤ N ; denote their union by U . Notice that U ∩ Zn is finite.
Step 3. Check all integer points of U whether they satisfy the equation

ΔA (v) = N.

Denote the set of all integer solutions in U by ZU .

Output data. The solution of Δ(v) = N is the set

B(ZU ).
B∈Ξ (A)

Remark 21.40 By Corollary 21.39 it is enough to check only the vertices of the
w-sails for w ≤ N .
21.6 On Reduced NRS-Matrices in SL(3, Z) 327

21.6 On Reduced Matrices in SL(3, Z) with Two Complex

Conjugate Eigenvalues

In this section we study the case of Hessenberg SL(3, Z) matrices with two complex
conjugate and one real eigenvalue. It turns out that the majority of such matrices
of the same Hessenberg type are ς -reduced. We start in Sect. 21.6.1 with some
notation and definitions. In Sect. 21.6.2 we formulate a supplementary theorem on
the parabolic structure of the set of nonreal spectrum matrices, whose proof we give
in Sect. 21.6.5. Further, in Sect. 21.6.3 we formulate the main result on asymptotic
uniqueness of ς -reduced matrices, whose proof is in Sect. 21.6.6. In Sect. 21.6.4 we
examine some particular examples of families of matrices with fixed Hessenberg
type. All the results of this section are new.

21.6.1 Perfect Hessenberg Matrices of a Given Hessenberg Type

For simplicity, we study only SL(3, Z) matrices whose characteristic polynomials

are irreducible over Q. There are two main geometrically essentially different cases
of such matrices: the real spectrum matrices (or RS-matrices for short), whose char-
acteristic polynomials have only real eigenvalues, and the nonreal spectrum ma-
trices (or NRS-matrices) whose characteristic polynomials have a pair of complex
conjugate and one real eigenvalue.
For a Hessenberg type Ω we denote the subset of all NRS-matrices in H (Ω) by
NRS(Ω).

Definition 21.41 Let Ω = a11 , a21 |a12 , a22 , a32 . Consider v = (a13 , a23 , a33 )
such that the determinant of the matrix (aij ) equals 1. Define
⎛ ⎞
a11 a12 a11 m + a12 n + a13
HΩv (m, n) = ⎝a21 a22 a21 m + a22 n + a23 ⎠ .
0 a32 a32 n + a33

It is clear that
 
H (Ω) = HΩv (m, n) | m ∈ Z, n ∈ Z .
In this context, to choose a vector v means to choose the origin O in the plane
H (Ω). So the set H (Ω) has the structure of a two-dimensional plane (see also
Sect. 21.2.4 above). We denote by OMN the coordinate system corresponding to the
2 a .
parameters (m, n). Notice that the Hessenberg complexity of this family is a12 23
v
Let DiscrΩ (m, n) denote the discriminant of the characteristic polynomial of
HΩv (m, n). Then the set NRS(Ω) is defined by the following inequality in variables
n and m:
DiscrvΩ (m, n) < 0.
328 21 Gauss Reduction in Higher Dimensions

Fig. 21.3 The family of

matrices of Hessenberg type
0, 1|0, 0, 1

Example 21.42 In Fig. 21.3 we show the subset NRS(0, 1|0, 0, 1). For this exam-
ple we choose v = (0, 0, 1).

21.6.2 Parabolic Structure of the Set of NRS-Matrices

The set NRS(0, 1|0, 0, 1) in Fig. 21.3 “looks like” the intersection of Z2 with the
union of the convex hulls of two parabolas. Let us formalize this in a general state-
ment.
Consider the matrix HΩv (0, 0) = (aij ) and define b1 , b2 , and b3 such that the
characteristic polynomial of this matrix in the variable t is

−t 3 + b1 t 2 − b2 t + b3 .

Even though in the case of SL(3, Z) we have b3 = 1, nevertheless we write b3 for

generality reasons. For the family HΩv (m, n) we define the following two quadratic
functions:

p1,Ω (m, n) = m − α1 n2 − β1 n − γ1 ;

n a21 m − a11 n 2 a21 m − a11 n

p2,Ω (m, n) = − α2 − β2 − γ2 ,
a21 a21 a21
21.6 On Reduced NRS-Matrices in SL(3, Z) 329

where
⎧ a32 ⎧ a32 a21
⎪ ⎪
⎪ α1 = − 4a21 ,
⎪
⎪ ⎪ α2 = 4b3 ,
⎪
⎪
⎪
⎪ ⎪
⎪
⎨ a11 − a22 − a33 ⎨ b2
β1 = , β2 = − ,
⎪
⎪
2a21 ⎪
⎪ 2b3
⎪
⎪ ⎪
⎪
⎪
⎪ 4b2 − b12 ⎪
⎪ b2 − 4b1 b3
⎩ γ1 = , ⎩ γ2 = 2 .
4a21 a32 4a21 a32 b3
In the real plane OMN of the family HΩv (m, n) we consider the interior of the
circle of radius R with center at the origin (0, 0) and denote it by BR (O). For a real
number t we set
    
Λt = (m, n) | p1,Ω (m, n) − t p2,Ω (m, n) − t < 0 .

Theorem 21.43 For every positive ε there exists R > 0 such that the following in-
clusions hold:

Λε \BR (O) ⊂ NRS(Ω)\BR (O) ⊂ Λ−ε \BR (O).

We give the proof of this theorem in Sect. 21.6.5 below.

21.6.3 Theorem on Asymptotic Uniqueness of ς -Reduced

NRS-Matrices

Recall that a ray is said to be integer if its vertex is integer and it contains integer
points distinct from the vertex.

Definition 21.44 An integer ray in H (Ω) is said to be an NRS-ray if all its integer
points correspond to NRS-matrices. A direction is said to be asymptotic for the set
NRS(Ω) if there exists an NRS-ray with this direction.

As stated in Theorem 21.43, for every Hessenberg type Ω the set NRS(Ω) almost
coincides with the union of the convex hulls of two parabolas. This implies the
following statement.

Proposition 21.45 Let Ω = a11 , a21 |a12 , a22 , a32 . There are exactly two asymp-
totic directions for the set NRS(Ω), defined by the vectors (−1, 0) and (a11 , a21 ).

Consider a family of Hessenberg matrices HΩv for an appropriate integer vector v.

Definition 21.46 Define

m,n  
R1,Ω,v = HΩv (m − t, n) | t ∈ Z≥0 ;
m,n  
R2,Ω,v = HΩv (m + a11 t, n + a21 t) | t ∈ Z≥0 .
330 21 Gauss Reduction in Higher Dimensions

Fig. 21.4 Every NRS-ray contains finitely many ς -nonreduced matrices

m,n m,n
By R1,Ω,v (t), or respectively by R2,Ω,v (t), we denote the tth element in the corre-
sponding family.

m,n m,n
Remark 21.47 The families R1,Ω,v and R2,Ω,v coincide with the sets of all integer
points of certain rays with directions (−1, 0) and (a11 , a21 ) respectively. Conversely,
from Proposition 21.45 it follows that the set of integer points of every NRS-ray
m,n m,n
coincides either with R1,Ω,v or with R2,Ω,v for some integers m and n.

In Fig. 21.4 we show in dark gray two NRS-rays:

−9,5
– R1,1,2|1,1,3,(0,0,−1) from the left;
−2,−1
– R2,1,2|1,1,3,(0,0,−1) from the right.
Now we are ready to formulate the main result on asymptotic behavior of NRS-
matrices, which we prove later in Sect. 21.6.6.
21.6 On Reduced NRS-Matrices in SL(3, Z) 331

(1,0,0)
Fig. 21.5 The family of Hessenberg matrices H0,1|1,0,2 (m, n)

Theorem 21.48 (On asymptotic ς -reducibility and uniqueness)

(i) Every NRS-ray (as in Fig. 21.4) contains only finitely many ς -nonreduced ma-
trices.
(ii) Every NRS-ray contains only finitely many matrices that are integer conjugate
to some other ς -reduced matrix.

Example 21.49 Every NRS-ray for the Hessenberg type 0, 1|0, 0, 1 contains only
ς -reduced perfect matrices. Experiments show that every NRS-ray for 0, 1|1, 0, 2
contains at most one ς -nonreduced matrix (see Fig. 21.5).

21.6.4 Examples of NRS-Matrices for a Given Hessenberg Type

In this subsection we study several examples of families NRS(Ω) for the Hessenberg
types:

0, 1|0, 0, 1, 0, 1|1, 0, 2, 0, 1|1, 1, 2, and 1, 2|1, 1, 3.

In Figs. 21.5, 21.6, and 21.7 the dark gray squares correspond to ς -nonreduced
matrices. We also fill with gray the squares corresponding to ς -reduced Hessenberg
matrices that are the nth powers (n ≥ 2) of some integer matrices.
(1,0,0)
Hessenberg Perfect NRS-Matrices H0,1|0,0,1 (m, n) The Hessenberg com-
plexity of all these matrices is 1, and therefore, they are all ς -reduced (see the family
in Fig. 21.3 on page 328).
332 21 Gauss Reduction in Higher Dimensions

(1,0,1)
Fig. 21.6 The family of Hessenberg matrices H0,1|1,1,2 (m, n)

(0,0,−1)
Fig. 21.7 The family of Hessenberg matrices H1,2|1,1,3

(1,0,0)
Hessenberg Perfect NRS-Matrices H0,1|1,0,2 (m, n) The Hessenberg com-
plexity of these matrices equals 2. Experiments show that 12 such matrices are ς -
nonreduced (see the family in Fig. 21.5). It is conjectured that all other Hessenberg
matrices of NRS(0, 1|1, 0, 2) are ς -reduced.
(1,0,1)
Hessenberg Perfect NRS-Matrices H0,1|1,1,2 (m, n) The Hessenberg com-
plexity of these matrices equals 2. We have found 12 ς -nonreduced matrices in
21.6 On Reduced NRS-Matrices in SL(3, Z) 333

the family. It is conjectured that all other Hessenberg matrices of NRS(0, 1|1, 1, 2)
are ς -reduced. See Fig. 21.6.

(0,0,−1)
Hessenberg Perfect NRS-Matrices H1,2|1,1,3 (m, n) This is a more compli-
cated example of a family of Hessenberg perfect NRS-matrices, with their com-
plexity equaling 12. We have found 27 ς -nonreduced matrices in the family. It is
conjectured that all other Hessenberg matrices of NRS(1, 2|1, 1, 3) are ς -reduced.
See Fig. 21.7.

21.6.5 Proof of Theorem 21.43

We begin the proof with several lemmas, but first a small remark.

Remark The set NRS(Ω) is defined by the inequality

DiscrvΩ (m, n) < 0.

In the left part of the inequality there is a polynomial of degree 4 in variables m

and n. Notice that the product
2 2
 
16a21 a32 b3 p1,Ω (m, n)p2,Ω (m, n)

is a good approximation to DiscrvΩ (m, n) at infinity; i.e., the polynomial

 
DiscrvΩ (m, n) − 16a21
2 2
a32 b3 p1,Ω (m, n)p2,Ω (m, n)

is a polynomial of degree only 2 in the variables m and n.

(1,0,0)
Lemma 21.50 The planar curve Discr0,1|0,0,1 (m, n) = 0 is contained in the do-
main defined by the inequalities
 2
(m − 4n + 3)(n2 + 4m + 3) ≥ 0,
(m2 − 4n − 3)(n2 + 4m − 3) − 72 ≤ 0.

(1,0,0)
Remark Lemma 21.50 implies that the curve Discr0,1|0,0,1 (m, n) = 0 is contained
in some tubular neighborhood of the curve
 2  
m − 4n n2 + 4m = 0.

Proof of Lemma 21.50 Note that

(1,0,0)   
Discr0,1|0,0,1 (m, n) = m2 − 4n n2 + 4m − 2mn − 27.
334 21 Gauss Reduction in Higher Dimensions

Thus, we have
(1,0,0)   
Discr0,1|0,0,1 (m, n) − m2 − 4n + 3 n2 + 4m + 3

= −2(n − 3)2 − 2(m + 3)2 − (n + m)2 ≤ 0,

and
(1,0,0)   
Discr0,1|0,0,1 (m, n) − m2 − 4n − 3 n2 + 4m − 3 + 72

= 2(n − 3)2 + 2(m + 3)2 + (n − m)2 ≥ 0.

(1,0,0)
Therefore, the curve Discr0,1|0,0,1 (m, n) = 0 is contained in the domain defined in
the lemma. 

Lemma 21.51 For every Ω = a11 , a21 |a12 , a22 , a32  there exists an affine (not nec-
essarily integer) transformation of the plane OMN taking the curve DiscrvΩ (m, n) =
(1,0,0)
0 to the curve Discr0,1|0,0,1 (m, n) = 0.

Proof Let HΩv (0, 0) = (ai,j ). Note that every matrix HΩv (m, n) is rational conjugate
to the matrix
(1,0,0)   
H0,1|0,0,1 m ,n ,
where
 
m = a23 a32 − a11 a33 + a12 a21 − a22 a33 − a11 a22 + a21 a32 m − a11 a32 n,
n = a11 + a22 + a33 + a32 n,

by the matrix
⎛ 2 +a a
⎞
1 a11 a11 12 21
v
XΩ = ⎝0 a21 a11 a21 + a21 a22 ⎠ .
0 0 a21 a32
(1,0,0)
Therefore, the curve DiscrvΩ (m, n) = 0 is mapped to the curve Discr0,1|0,0,1 (m,
n) = 0 bijectively. In OMN coordinates this map corresponds to the following affine
transformation:

m a a m − a11 a32 n a a − a11 a33 + a12 a21 − a22 a33 − a11 a22
"→ 21 32 + 23 32 .
n a32 n a11 + a22 + a33

This completes the proof of the lemma. 

Proof of Theorem 21.43 Consider a family of matrices HΩv (−p1,Ω (0, t) + ε, t) with
real parameter t. Direct calculations show that for ε
= 0 the discriminant of the
matrices for this family is a polynomial of the fourth degree in the variable t, and
  1  
DiscrvΩ −p1,Ω (0, t) + ε, t = a21 a32
5
εt 4 + O t 3 .
4
21.6 On Reduced NRS-Matrices in SL(3, Z) 335

Therefore, there exists a neighborhood of infinity with respect to the variable t such
that the function DiscrvΩ (−p1,Ω (0, t) + ε, t) is positive for positive ε in this neigh-
borhood, and negative for negative ε. 

Hence for a given ε there exists a sufficiently large N1 = N1 (ε) such that for ev-
ery t > N1 , there exists a solution of the equation DiscrvΩ (m, n) = 0 at the segment
with endpoints
   
−p1,Ω (0, t) + ε, t and −p1,Ω (0, t) − ε, t

of the plane OMN.

Now we examine the family in the variable t for the second parabola:

a11 a21
HΩv t − a11 p2,Ω (t, 0) −  ε, −a21 p2,Ω (t, 0) −  ε .
2 + a2
a11 2 + a2
a11
21 21

For the same reasons, for a given ε there exists a sufficiently large N2 = N2 (ε)
such that for every t > N2 there exists a solution of the equation DiscrvΩ (m, n) = 0
on the segment with endpoints

a11 a21
t − a11 p2,Ω (t, 0) −  ε, −a21 p2,Ω (t, 0) −  ε and
2 + a2
a11 a 2 + a2
21 11 21

a11 a21
t − a11 p2,Ω (t, 0) +  ε, −a21 p2,Ω (t, 0) +  ε
2 + a2
a11 2 + a2
a11
21 21

of the plane OMN.

The set
p1,Ω (m, n)p2,Ω (m, n) = 0
is a union of two parabolas and therefore has four parabolic branches “approaching
infinity.” We have shown that for each of these four parabolic branches there ex-
ists a branch of the curve DiscrvΩ (m, n) = 0 contained in the ε-tube of the chosen
parabolic branch if we are far enough from the origin.
(1,0,0)
From Lemma 21.50 we know that Discr0,1|0,0,1 (m, n) = 0 is contained in some
tubular neighborhood of

p1,0,1|0,0,1 (m, n)p2,0,1|0,0,1 (m, n) = 0.

Then by Lemma 21.51, the curve DiscrvΩ (m, n) = 0 is contained in some tubular
neighborhood of the curve

p1,Ω (m, n)p2,Ω (m, n) = 0

336 21 Gauss Reduction in Higher Dimensions

outside some ball centered at the origin. Finally, by Viète’s Theorem, the intersec-
tion of the curve DiscrvΩ (m, n) = 0 with each of the parallel lines
a11 + a21
t : n−m=t
a21
contains at most four points. Therefore, there exists a sufficiently large T such that
for any t ≥ T , the intersection of the curve DiscrvΩ (m, n) = 0 and t contains ex-
actly 4 points corresponding to the branches of the parabolas p1,Ω (m, n) = 0 and
p2,Ω (m, n) = 0 lying in Λ−ε \ Λε .
From Lemma 21.50 and Lemma 21.51 it follows that the set
 
a11 + a21  
n − m < T ∩ DiscrvΩ (m, n) = 0
a21
is compact.
Therefore, there exists R = R(ε, N1 , N2 , T ) such that in the complement to the
ball BR (O) we have
Λε ⊂ NRS(Ω) ⊂ Λ−ε .
The proof of Theorem 21.43 is complete.

21.6.6 Proof of Theorem 21.48

First we prove that a fundamental domain of a general Klein–Voronoi continued

fraction can be chosen from a certain bounded subset of R3 . Then we describe
asymptotically the geometric structure of Klein–Voronoi continued fractions. Fi-
nally, we prove Theorem 21.48.

21.6.6.1 One General Fact on Fundamental Domains of Klein–Voronoi

Continued Fractions for NRS-Matrices of SL(3, Z)

Consider an NRS-operator A and any integer point p distinct from the origin O.
Denote by ΓA0 (p) the convex hull of the union of two orbits corresponding to the
points p and A(p). For every integer k we denote by ΓAk (p) the set Ak (ΓA0 (p)).

Proposition 21.52 Let p be a nonzero integer point. Then there exists a fundamen-
tal domain of the Klein–Voronoi continued fraction for A with all (integer) orbit-
vertices contained in the set ΓA0 (p).

The proof is based on the following lemma. Let


ΓA (p) = ΓAk (p).
k∈Z
21.6 On Reduced NRS-Matrices in SL(3, Z) 337

Lemma 21.53 Let p be a nonzero integer point. Then one of the Klein–Voronoi
sails for A is contained in the set ΓA (p).

Proof Notice that the set ΓA (p) is a union of orbits. Let us project ΓA (p) to the half-
plane π+ . The set ΓA (p) projects to the closure of the complement of the convex
hull for the points ΓA (Ak (p)) ∩ π+ for all integer k in the angle defined by the
eigenspaces. Since all points Ak (p) are integer, their convex hull is contained in the
convex hull of all points corresponding to integer orbits in the angle. Hence the set
ΓA (p) ∩ π+ contains the projection of the sail. Therefore, the set ΓA (p) contains
one of the sails. 

Proof of Proposition 21.52 Since − Id exchanges the sails, one can choose a fun-
damental domain entirely contained in one sail. Let this sail contain a point p.
By Lemma 21.53 the set ΓA (p) contains this sail. Therefore, all orbit-vertices of
a fundamental domain for the Klein–Voronoi continued fraction can be chosen in
ΓA0 (p). 

21.6.6.2 Geometry of Klein–Voronoi Continued Fractions for Matrices of

m,n
R1,Ω,v

Let us prove the following statement.

m,n
Proposition 21.54 Consider an NRS-ray R1,Ω,v . Then there exists a constant C >
0 such that for every t > C, all orbit-vertices of one of the fundamental domains
m,n
for the Klein–Voronoi continued fraction of the matrix R1,Ω,v (t) are contained in
the set of all integer orbits corresponding to the integer points in the convex hull of
(1, 0, 0), (a11 , a21 , 0), and (−a11 , −a21 , 0).

We begin with the case of matrices of Hessenberg type Ω0 = 0, 1|0, 0, 1. Such
matrices form a family H (Ω0 ) with real parameters m and n:
⎛ ⎞
0 0 1
H0,1|0,0,1 (m, n) = ⎝1 0 m⎠ .
(1,0,0)

0 1 n

Put v0 = (1, 0, 0).

m,n
Lemma 21.55 Let R1,Ω 0 ,v0
be an NRS-ray. Then for every ε > 0 there exists C > 0
such that for every t > C the convex hull of the union of two orbit-vertices

TR m,n (1, 0, 0) and TR m,n (0, 1, 0)

1,Ω0 ,v0 (t) 1,Ω0 ,v0 (t)

is contained in the ε-tubular neighborhood of the convex hull of three points

(1, 0, 0), (0, 1, 0), (0, −1, 0).
338 21 Gauss Reduction in Higher Dimensions

Remark Lemma 21.55 means that the corresponding domain tends to be flat as t
m,n
tends to infinity. Recall that R1,Ω 0 ,v0
(t)(1, 0, 0) = (0, 1, 0) for every t.

Proof Let us find the asymptotics of eigenvectors and eigenplanes for operators
m,n m,n
R1,Ω 0 ,v0
(t) as t tends to +∞. Denote a real eigenvector of R1,Ω 0 ,v0
(t) by e(t).
Direct calculations show that there exists μ
= 0 such that
  
e(t) = μ (1, 0, 0) + O t −1 .
m,n
Consider the invariant real plane of the operator R1,Ω 0 ,v0
(t) (it corresponds to a pair
of complex conjugate eigenvalues). Note that this plane is a union of a certain family
m,n
of closed orbits of R1,Ω 0 ,v0
(t). Direct calculations show that any orbit in this family
is an ellipse with semiaxes λgmax (t) and λgmin (t) for some positive real number λ,
where

gmax (t) = (0, t, 0) + O(1),

   
gmin (t) = 0, 0, t 1/2 + O t −1/2 .
√
The vectors gmax (t) ± −1gmin (t) are two complex eigenvectors of the operator
m,n
R1,Ω 0 ,v0
(t). For the ratio of the lengths of maximal and minimal semiaxes of any
orbit we have the following asymptotic estimate:

λ|gmax (t)|  
= |t|1/2 + O |t|−1/2 .
λ|gmin (t)|

First, since
1  
(1, 0, 0) − e(t) = O |t|−1 ,
μ
the minimal semiaxis of the orbit-vertex TR m,n (1, 0, 0) is asymptotically not
1,Ω0 ,v0 (t)
greater than O(t −1 ). Therefore, the length of the maximal semiaxis is asymptot-
ically not greater than some function of type O(|t|−1/2 ). Hence, the orbit of the
point (1, 0, 0) is contained in the (C1 |t|−1/2 )-ball of the point (1, 0, 0), where C1 is
a constant that does not depend on t.
Second, we have
1  
(0, 1, 0) − gmax (t) = O |t|−1 .
t
Thus, the length of the maximal semiaxis of the orbit-vertex TR m,n (1, 0, 0) is
1,Ω0 ,v0 (t)
asymptotically not greater than some function 1+O(t −1/2 ). Hence, the length of the
minimal semiaxis is asymptotically not greater than some function O(|t|−1/2 ). This
implies that the orbit of the point (0, 1, 0) is contained in the (C2 |t|−1/2 )-tubular
neighborhood of the segment with vertices (0, 1, 0) and (0, −1, 0), where C2 is a
constant that does not depend on t.
21.6 On Reduced NRS-Matrices in SL(3, Z) 339

Therefore, the convex hull of the union of two orbit-vertices

TR m,n (1, 0, 0) and TR m,n (0, 1, 0)

1,Ω0 ,v0 (t) 1,Ω0 ,v0 (t)

is contained in the C-tubular neighborhood of the triangle with vertices (1, 0, 0),
(0, 1, 0), (0, −1, 0), where C = |t|−1/2 max(C1 , C2 ). This concludes the proof of
the lemma. 

Let us now extend the statement of Proposition 21.52 to the general case of Hes-
senberg matrices.
m,n
Corollary 21.56 Let Ω = a11 , a21 |a12 , a22 , a32  and let R1,Ω,v be an NRS-ray.
Then for every ε > 0 there exists C > 0 such that for every t > C the convex hull of
the union of two orbit-vertices

TR m,n (1, 0, 0) and TR m,n (a11 , a21 , 0)

1,Ω,v (t) 1,Ω,v (t)

is contained in the ε-tubular neighborhood of the convex hull of three points

(1, 0, 0), (a11 , a21 , 0), (−a11 , −a21 , 0).
m,n
Remark Recall that R1,Ω,v (t)(1, 0, 0) = (a11 , a21 , 0) for every t.

Proof of Corollary 21.56 Define

⎛ ⎞
a21 a32 −a32 a11 a11 a22 − a21 a12
X=⎝ 0 a32 −a11 − a22 ⎠ .
0 0 1

The operator X defines two linear functions l1 and l2 on two variables m and n with
coefficients depending only on a11 , a21 , a12 , a22 , and a32 satisfying

t
, l2 (m, n) = XHΩ (m − t, n)X −1 .
(1,0,0)
H0,1|0,0,1 l1 (m, n) −
a21 a32
m,n
Therefore, the ray R1,Ω,v after a change of coordinates and a rescaling is taken to
m̃,ñ
the ray R1,Ω 0 ,v0
of matrices with Hessenberg type Ω0 = 0, 1|0, 0, 1 for certain m̃
and ñ.
Lemma 21.55 implies the following. For every ε > 0 there exists a positive con-
stant such that for every t greater than this constant the convex hull of the union of
two orbit-vertices

TR m̃,ñ (1, 0, 0) and TR m̃,ñ (0, 1, 0)

1,Ω0 ,v0 (t) 1,Ω0 ,v0 (t)

is contained in the ε-tubular neighborhood of the triangle with vertices (1, 0, 0),
(0, 1, 0), (0, −1, 0).
340 21 Gauss Reduction in Higher Dimensions

If we reformulate the last statement for the family of operators in the old coordi-
nates, then we get the statement of the corollary. 
m,n
Proof of Proposition 21.54 Notice that the operator R1,Ω,v (t) takes the point
(1, 0, 0) to the point (a11 , a21 , 0). Therefore, the convex hull of the union of two
orbit-vertices
TR m,n (1, 0, 0) and TR m,n (a11 , a21 , 0)
1,Ω,v (t) 1,Ω,v (t)

(we denote it by W (t)) coincides with the set ΓR0m,n (1, 0, 0). 
1,Ω,v (t)

From Proposition 21.52 it follows that there exists a fundamental domain for
a Klein–Voronoi continued fraction with all its orbit-vertices contained in W (t).
Choose a sufficiently small ε0 such that the ε0 -tubular neighborhood of the triangle
with vertices

(1, 0, 0), (a1,1 , a2,1 , 0), and (−a11 , −a21 , 0)

does not contain integer points distinct from the points of the triangle. From Corol-
lary 21.56 it follows that for a sufficiently large t the set W (t) is contained in
the ε0 -tubular neighborhood of the triangle. This implies the statement of Propo-
sition 21.54. 

21.6.6.3 Geometry of Klein–Voronoi Continued Fractions for Matrices of

m,n
R2,Ω,v

Now let us study the remaining case of the rays of matrices with asymptotic direc-
tion (a11 , a21 ). We recall that Ω0 = 0, 1|0, 0, 1.
m,n
Proposition 21.57 Consider an NRS-ray R2,Ω,v . Then there exists a constant
C > 0 such that for every t > C all orbit-vertices of one of the fundamental domains
m,n
for the Klein–Voronoi continued fraction of the matrix R2,Ω,v (t) are contained in
the set of all integer orbits corresponding to the integer points in the convex hull of
(1, 0, 0), (−1, 0, 0), and (a11 , a21 , 0).

The proof of this proposition is based on a corollary of the following lemma.

m,n
Lemma 21.58 Let R2,Ω 0 ,v0
be an NRS-ray. Then for every ε > 0 there exists C > 0
such that for every t > C the convex hull of the union of two orbit-vertices

TR m,n (1, 0, 0) and TR m,n (0, 1, 0)

2,Ω0 ,v0 (t) 2,Ω0 ,v0 (t)

is contained in the ε-tubular neighborhood of the convex hull of three points

(1, 0, 0), (−1, 0, 0), (0, 1, 0).
21.6 On Reduced NRS-Matrices in SL(3, Z) 341

Proof First, notice that the Klein–Voronoi continued fractions for the operators A
and A−1 coincide.
Second

H0,1|0,0,1 (m, n + t) = XH0,1|0,0,1 (−n − t, −m)X −1 ,

(1,0,0) (1,0,0)

where
⎛ ⎞
0 −1 −n − t
X = ⎝−1 0 −m ⎠ .
0 0 −1
Thus, in the new coordinates we obtain the equivalent statement for the ray
−n,−m
R1,Ω 0 ,v0
(t). Now Lemma 21.58 follows directly from Lemma 21.55. 

m,n
Corollary 21.59 Let Ω = a11 , a21 |a12 , a22 , a32  and let R2,Ω,v be an NRS-ray.
Then for every ε > 0 there exists C > 0 such that for every t > C the convex hull of
the union of two orbit-vertices

THΩv (m+a11 t,n+a21 t) (1, 0, 0) and THΩv (m+a11 t,n+a21 t) (a11 , a21 , 0)

is contained in the ε-tubular neighborhood of the triangle with vertices (1, 0, 0),
(−1, 0, 0), and (a11 , a21 , 0).

Remark We omit the proofs of Corollary 21.59 and Proposition 21.57, since they
repeat the proofs of Corollary 21.56 and Proposition 21.54.

21.6.6.4 Proof of Theorem 21.48

Step 1. Let A be an operator with Hessenberg matrix M in SL(3, Z). By Propo-

sition 21.29 the Hessenberg complexity of the Hessenberg matrix M coincides
with the MD-characteristic ΔA (1, 0, 0). Therefore, the Hessenberg matrix M is ς -
reduced if and only if the MD-characteristic of A attains the minimal possible ab-
solute value on the integer lattice, except at the origin, exactly at the point (1, 0, 0).
Step 2. By Theorem 21.31(ii) we know that all minima of the set of absolute
values for the MD-characteristic of A are attained at integer points of the Klein–
Voronoi sails for A.
Step 3. By Theorem 21.31(i) we can restrict the search of the minimal absolute
value of the MD-characteristic to an arbitrary fundamental domains of the Klein–
Voronoi continued fraction.
Step 4.1. The case of NRS-rays with asymptotic direction (−1, 0). By Proposi-
tion 21.54 there exists T > 0 such that for every integer t > T all integer points of
m,n
one of the fundamental domains for R1,Ω,v (t) are contained in the convex hull of
three points (1, 0, 0), (a11 , a21 , 0), and (−a11 , −a21 , 0).
342 21 Gauss Reduction in Higher Dimensions

This triangle contains only finitely many integer points, all of which have the last
coordinate equal to zero. The value of the MD-characteristic for the points of type
(x, y, 0) equals
(a21 x − a11 y)a32
2 2
y t + C,
where the constant C does not depend on t, but only on x, y, and the elements ai,j
of the matrix HΩv (m, n).
So, for any point (x, y, 0), the MD-characteristics linearly increase with re-
spect to the parameter t. The only exceptions are the points of type λ(1, 0, 0) and
μ(a11 , a21 , 0) (for integers λ and μ). The values of the MD-characteristics are con-
stant in these points with respect to t.
Since there are finitely many integer points in the triangle (1, 0, 0), (a11 , a21 , 0),
and (−a11 , −a21 , 0), for sufficiently large t the MD-characteristic at points of the
m,n
triangle attains its minimum at (1, 0, 0) or at (a11 , a21 , 0). Since R1,Ω,v (t) takes
the point (1, 0, 0) to the point (a11 , a21 , 0), the values of the MD-characteristics at
(1, 0, 0) and at (a11 , a21 , 0) coincide.
Therefore, for sufficiently large t the matrix
(1,0,0)
Ha11 ,a21 |a12 ,a22 ,a32  (m − t, n)

is always ς -reduced and there are no other ς -reduced matrices integer conjugate to
the given one. This implies both statements of Theorem 21.48 for every NRS-ray
with asymptotic direction (−1, 0).
Step 4.2. The case of NRS-rays with asymptotic direction (a11 , a21 ). This case is
similar to the case of NRS-rays with asymptotic direction (−1, 0), so we omit the
proof here.
Proof of Theorem 21.48 is complete. 

21.7 Open Problems

In this section we formulate open questions on the structure of the sets of NRS-
matrices and briefly describe the situation for RS-matrices.

NRS-Matrices As we have shown in Theorem 21.48, the number of ς -

nonreduced matrices in any NRS-ray is finite. Here we conjecture a stronger state-
ment.

Conjecture 29 Let Ω be an arbitrary Hessenberg type. All but a finite number of

NRS-matrices of type Ω are ς -reduced.

If the answer to this conjecture is positive, we immediately have the following

general question.

Problem 30 Study the asymptotics of the number of ς -nonreduced NRS-matrices

with respect to the growth of Hessenberg complexity.
21.7 Open Problems 343

Denote the conjectured number of ς -nonreduced NRS-matrices of Hessenberg

type Ω by #(Ω). Numerous calculations give rise to the following table for all types
with Hessenberg complexity less than 5.

Ω
0, 1|0, 0, 1 0, 1|1, 0, 2 0, 1|1, 1, 2 0, 1|1, 0, 3 0, 1|1, 1, 3 0, 1|1, 2, 3

ς(Ω) 1 2 2 3 3 3
#(Ω) 0 12 12 6 10 10

Ω
0, 1|2, 0, 3 0, 1|2, 1, 3 0, 1|2, 2, 3 1, 2|0, 0, 1 0, 1|1, 0, 4 0, 1|1, 1, 4

ς(Ω) 3 3 3 4 4 4
#(Ω) 14 10 10 94 6 8

Ω
0, 1|1, 2, 4 0, 1|1, 3, 4 0, 1|3, 0, 4 0, 1|3, 1, 4 0, 1|3, 2, 4 0, 1|3, 3, 4

ς(Ω) 4 4 4 4 4 4
#(Ω) 10 8 10 12 8 8

RS-Matrices Now we say a few words about real spectrum matrices (i.e., about
SL(3, Z)-matrices with three distinct real roots). Mostly we consider the family
H (0, 1|1, 0, 2); the situation with the other Hessenberg types is similar.
Recall that
⎛ ⎞
0 1 n+1
H0,1|1,0,2 (m, n) = ⎝1 0 m ⎠.
(1,0,1)

0 2 2n + 1
This matrix is of Hessenberg type 0, 1|1, 0, 2, and its Hessenberg complexity
(1,0,1)
equals 2. Hence H0,1|1,0,2 (m, n) is ς -reduced if and only if it is not integer conju-
gate to some matrix of unit Hessenberg complexity. All such matrices are of Hes-
senberg type 0, 1|0, 0, 1.
(1,0,1)
In Fig. 21.8 we show all matrices H0,1|1,0,2 (m, n) within the square
−20 ≤ m, n ≤ 20.
The square in the intersection of the mth column with the nth row corresponds to
(1,0,1)
the matrix H0,1|1,0,2 (m, n). It is colored in black if the characteristic polynomial
has rational roots. The square is colored in gray if the characteristic polynomial is
irreducible and there exists an integer vector (x, y, z) with the coordinates satisfying
−1000 ≤ x, y, z ≤ 1000,
such that the MD-characteristic of H0,1|1,0,2 (1, 0, 1)(m, n) equals 1 at (x, y, z).
All the rest of the squares are white.
344 21 Gauss Reduction in Higher Dimensions

Fig. 21.8 The family of matrices of Hessenberg type 0, 1|1, 0, 2

If a square is gray, then the corresponding matrix is ς -nonreduced (see Re-

mark 21.29). If a square is white, then we cannot conclude whether the matrix is
ς -reduced (since the integer vector with unit MD-characteristics may have coordi-
nates with absolute values greater than 1000).
It is most probable that white squares in Fig. 21.8 represent ς -reduced matrices.
We have checked explicitly all white squares for

−10 ≤ m, n ≤ 10,

and found that all of them are ς -reduced (such squares are contained inside the big
black square shown on the figure).
We show a boundary broken line between the NRS- and RS-squares in gray.

Remark In Fig. 21.8 the NRS-domain is easily visualized. It almost completely con-
sists of white squares. On the other hand, the RS-domain contains a relatively large
number of black squares. This indicates a significant difference between the RS and
NRS cases.
21.8 Exercises 345

Direct calculations of the corresponding MD-characteristics show that the fol-

lowing proposition holds.
(1,0,1)
Proposition 21.60 If an integer m + n is odd, then H0,1|1,0,2 (m, n) is ς -reduced.

Remark On the one hand, Proposition 21.60 implies the existence of rays en-
tirely consisting of ς -reduced matrices. On the other hand, in contrast to the NRS-
matrices, there are some rays consisting entirely of ς -nonreduced RS-matrices. For
instance, all matrices corresponding to the integer points of the lines
(1) m = n;
(2) m = n + 2;
(3) m = −n;
(4) m = −n − 2;
(5) n = 3m − 4;
(6) m = 3n + 6
are not ς -reduced (we do note state that the list of such lines is complete).

So Theorem 21.48 does not have a direct generalization to the RS-case and we
end up with the following problem.

Problem 31 What is the percentage of ς -reduced matrices among matrices of a

given Hessenberg type Ω?

It is likely that almost all Hessenberg matrices are ς -reduced (except for some
measure-zero subset).

21.8 Exercises
Exercise 21.1 Let π be an integer k-dimensional plane in Rn . Describe the locus
of all integer points a unit integer distance to π .

Exercise 21.2 Prove that the Hessenberg matrices

⎛ ⎞ ⎛ ⎞
0 1 3 0 2 5
⎝1 0 0⎠ and ⎝1 1 2⎠
0 3 8 0 3 7
are not integer conjugate but have the same Hessenberg complexity and the same
characteristic polynomial.

Exercise 21.3 Let A ∈ GL(n, R) preserve the level sets of the form x 2 + y 2 . Prove
that there exists α such that either

cos α sin α − cos α sin α

A= or A = .
− sin α cos α sin α cos α
346 21 Gauss Reduction in Higher Dimensions

Exercise 21.4 Consider A ∈ GL(n, Z) whose characteristic polynomial is irre-

ducible over Q. Find the generators of the abelian group T l (A) in the eigenbasis
of A.

Exercise 21.5 Consider A, B ∈ GL(n, Z) whose characteristic polynomial is irre-

ducible over Q. Then A and B commute if and only if the Dirichlet groups of A and
B coincide.
(1,0,1)
Exercise 21.6 If an integer m + n is odd, then H0,1|1,0,2 (m, n) is ς -reduced.

(1,0,1)
Exercise 21.7 Prove that all matrices H0,1|1,0,2 (m, n) corresponding to integer
points of the lines
(1) m = n;
(2) m = n + 2;
(3) m = −n;
(4) m = −n − 2;
(5) n = 3m − 4;
(6) m = 3n + 6
are ς -reduced. Are there some other similar lines in the plane (m, n)?

Exercise 21.8 How many different nonempty Hessenberg types of complexity n

exist for n = 1, 2, . . . , 10?

Exercise 21.9 Prove that the minimum of the Markov–Davenport characteristic on

the corresponding w-sail is attained at some vertex of this w-sail.

Exercise 21.10 Find an example of an invariant cone for an irreducible operator

whose w-sails are not homothetic to each other.
Chapter 22
Approximation of Maximal Commutative
Subgroups

We have already discussed some geometric approximation aspects in the plane in

Chap. 10: we have studied approximations, first, of an arbitrary ray with vertex at the
origin, and second, of an arrangement of two lines passing through the origin. In this
chapter we briefly discuss an approximation problem of maximal commutative sub-
groups of GL(n, R) by rational subgroups (introduced recently in [101]). In general
this problem touches the theory of simultaneous approximation and both subjects
of Chap. 10. The problem of approximation of real spectrum maximal commutative
subgroups has much in common with the problem of approximation of nondegener-
ate simplicial cones. So it is clear that multidimensional continued fractions should
be a useful tool here. Also we would like to mention that the approximation problem
is linked to the so-called limit shape problems (see for instance [202]).
In Sect. 22.1 we give general definitions and formulate the problem of best ap-
proximations of maximal commutative subgroups. Further, in Sect. 22.2 we discuss
the connection of three-dimensional maximal commutative subgroup approximation
to the classical case of simultaneous
√ approximation of vectors in R3 .
In this chapter we denote −1 by I .

22.1 Rational Approximations of MCRS-Groups

In this section we give general definitions and formulate basic concepts of approxi-
mation of maximal commutative subgroups. We begin in Sect. 22.1.1 with the def-
inition of MCRS-groups and say a few words about how they relate to simplicial
cones. Further, in Sect. 22.1.2 we extend the definition of two-dimensional Markov–
Davenport forms to the multidimensional case. We define rational subgroups and
their “size” in Sect. 22.1.3. The distance function (discrepancy) between two sub-
groups is introduced in Sect. 22.1.4. Then in Sect. 22.1.5 we give the definition of
best approximations. Finally, we conclude this section with a link to the classical
problem of approximating real numbers by rational numbers.

O. Karpenkov, Geometry of Continued Fractions, 347

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_22, © Springer-Verlag Berlin Heidelberg 2013
348 22 Approximation of Maximal Commutative Subgroups

22.1.1 Maximal Commutative Subgroups and Corresponding

Simplicial Cones

22.1.1.1 Maximal Commutative Subgroups

For an arbitrary matrix A of GL(n, R), we denote by CGL(n,R) (A) its centralizer,
i.e., the set of all the elements of GL(n, R) commuting with A.
We say that a matrix is regular if all its eigenvalues are distinct (but not nec-
essarily real). In this chapter we discuss only the most studied case of centralizers
defined by regular matrices. Notice that if a matrix A is regular, then the centralizer
of CGL(n,R) (A) is commutative.

Definition 22.1 Let A be a regular matrix in GL(n, R). We call the centralizer
CGL(n,R) (A) the maximal commutative subgroup of GL(n, R) (or MCRS-group for
short).

We say that a subspace is invariant for a centralizer CGL(n,R) (A) if it is invari-

ant for every element of CGL(n,R) (A). Let a regular operator A have k real and 2l
complex eigenvalues (k + 2l = n). Then the MCRS-group CGL(n,R) (A) has k one-
dimensional and l two-dimensional minimal invariant subspaces. In case l = 0, we
say that the MCRS-group is real spectrum.

Remark 22.2 Notice that if A is a regular matrix in GL(n, Z), then the Dirichlet
group Ξ (A) is a subgroup of the centralizer CGL(n,R) (A). We would like just to
mention that Dirichlet groups usually play an important role in the study of best
approximations of the corresponding MCRS-groups.

22.1.1.2 The Space of Simplicial Cones

An n-sided simplicial cone in Rn is the convex hull of the union of n unordered

rays with vertex at the origin. We consider only the cones that are not contained in
a hyperplane. Denote by Simpln the space of all n-sided simplicial cones.
Let us describe a relation between real spectrum maximal commutative sub-
groups from one side and n-sided simplicial cones from the other. There exists a nat-
ural 2n -fold covering of the space of all real spectrum MCRS-groups by the space
Simpln : the cones map to MCRS-groups whose eigendirections are the extremal
rays of the cones. Therefore, approximation problems within the space Simpln can
be studied in terms of the MCRS-groups.

Remark 22.3 The space Simpln is not compact for n > 1. It admits a transitive right
action of GL(n, R) and possesses an invariant measure μn . This measure is usually
called the Möbius measure, and is similar to the Möbius measure for the space of
all multidimensional continued fractions discussed in Chap. 19.
22.1 Rational Approximations of MCRS-Groups 349

22.1.2 Regular Subgroups and Markov–Davenport Forms

Consider an arbitrary MCRS-group A in GL(n, R) and denote its eigenlines (both

real and conjugate) by l1 , . . . , ln . For i = 1, . . . , n, denote by Li a nonzero linear
form over the space Cn that attains zero values at all vectors of the complex lines
lj for j
= i. Let det(L1 , . . . , Ln ) be the determinant of the matrix having in the kth
column the coefficients of the form Lk for k = 1, . . . , n in the dual basis.

Definition 22.4 We say that the form

%n
k=1 (Lk (x1 , . . . , xn ))
det(L1 , . . . , Ln )
is the Markov–Davenport form for the MCRS-group A and denote it by ΦA .

Remark 22.5 Notice that Markov–Davenport forms are closely related to MD-
characteristics studied in the previous section. Namely, if A ∈ A is an operator with
distinct eigenvalues, then the Markov–Davenport form is proportional to the MD-
characteristic ΔA .

Let us continue with a particular example.

Example 22.6 Consider the MCRS-group corresponding to the Fibonacci operator

1 1
.
1 0
The Fibonacci operator has two eigenlines,

y = −θ x and y = θ −1 x,
√
1+ 5
where θ is the golden ration 2 . So the Markov–Davenport form of the Fibonacci
operator is
(y + θ x)(y − θ −1 x) 1  
−1
= √ −x 2 + xy + y 2 .
θ −θ 5

Notice that we have already studied two-dimensional Markov–Davenport forms

in Chap. 7 and Chap. 10. Let us formulate and prove the following two simple
properties of Markov–Davenport forms.

Proposition 22.7 The Markov–Davenport form of an MCRS-group A is defined by

A up to a sign.

Proof All regular operators of A have the same set of eigenlines. Hence the forms
Li are uniquely defined by the MCRS-group up to a multiplication by a scalar and
permutations. Every scalar multiplication is normalized by the determinant in the
denominator, while a sign of a permutation is not caught by the denominator. 
350 22 Approximation of Maximal Commutative Subgroups

Proposition 22.8 All the coefficients of the Markov–Davenport form are either si-
multaneously real, if there is an even number of minimal invariant planes of the
corresponding MCRS-group, or simultaneously complex.

Proof By definition, every MCRS-group contains a real operator with distinct

eigenvalues. Linear forms related to real eigenvalues have real coefficients. Pairs
of linear forms corresponding to pairs of complex conjugate eigenvalues are com-
plex conjugate up to a multiplicative complex factor. Without loss of generality we
choose this factor to be equal to 1. This implies that the product of two such lin-
ear forms is real. Due to the determinant in the denominator √each pair of complex
conjugate roots brings an additional multiplicative factor I = −1. 

Further details of the proof we leave to the reader.

22.1.3 Rational Subgroups and Their Sizes

Recall some definitions related to complex geometry. A complex number is called

a Gaussian integer if it is of the form a + I b with integers a and b. The set of all
Gaussian integers forms a lattice with the operation of vector addition. We say that
a vector is Gaussian if all its coordinates are Gaussian integers. A one-dimensional
subspace of Cn is called Gaussian if it contains a nonzero Gaussian vector. A Gaus-
sian vector is said to be primitive if all its coordinates are relatively prime with
respect to the multiplication of Gaussian integers.

Proposition 22.9 Every Gaussian one-dimensional subspace contains exactly four

primitive Gaussian vectors. In addition, a vector v is primitive if and only if the
vectors −v, I v, and −I v are primitive as well.

Notice that in the case of C1 there are four primitive vectors (or units): they are
±1 and ±I .
For a complex vector v = (a1 + I b1 , . . . , an + I bn ), denote by |v| the norm
5 6
max ai2 + bi2 .
i=1,...,n

It is clear that the minimum of the norm | ∗ | among the Gaussian vectors in a Gauss-
ian one-dimensional subspace is attained at a primitive Gaussian vector.
Now we are ready to define rational MCRS-groups.

Definition 22.10 An MCRS-group A is called rational if every of its eigenlines

contains a Gaussian nonzero vector.

Definition 22.11 Consider a rational MCRS-group A. Let l1 , . . . , ln be the eigen-

lines of A. Let vi be a primitive Gaussian vector in li for i = 1, . . . , n. The size of A
22.1 Rational Approximations of MCRS-Groups 351

is the real number

 
min |v1 |, . . . , |vn | ,
i=1,...,n

which we denote by ν(A).

22.1.4 Discrepancy Functional

Let us define a natural distance between MCRS-groups. Let A1 and A2 be two

MCRS-groups. Consider the following two symmetric homogeneous forms of de-
gree n each, the combination of two Davenport forms:

ΦA 1 + ΦA 2 and ΦA1 − ΦA2 .

Take two maximal absolute values of the coefficients of these forms (separately).
The minimum of these two maximal values is considered the distance between A1
and A2 , which we call the discrepancy and denote by ρ(A1 , A2 ).

Example 22.12 Consider the following two operators

0 −1
with eigenvectors (I, 1) and (−I, 1),
1 0

1 1
with eigenvectors (1, 2) and (1, −2).
4 1

These operators have distinct eigenvalues, and therefore, they define MCRS-groups
(denote them by A1 and A2 respectively). Both MCRS-groups are rational, where
the operators of the first one have real eigenvalues, and where the operators of the
second have pairs of complex conjugate eigenvalues. Direct calculations show that

ν(A1 ) = 1 and ν(A2 ) = 2.

Now let us calculate the discrepancy ρ(A1 , A2 ). We have

 
   2 2
ΦA (v) ± ΦA (v) = I x + y ± y − 4x 
2 2
1 2  2 4 
√
and therefore ρ(A1 , A2 ) = 3
2 .

22.1.5 Approximation Model

In the above notation the problem of approximation of MCRS-groups by rational

MCRS-groups can be formulated as follows.
352 22 Approximation of Maximal Commutative Subgroups

Definition 22.13 (The problem of best approximation) For a given MCRS-group A

and a positive integer N , find a rational MCRS-group AN of size not exceeding N
such that for every other rational MCRS-group A of size not exceeding N we have
 
ρ(A, AN ) ≤ ρ A, A .

Remark 22.14 There is another important class of MCRS-groups that would also be
interesting to consider in this context. An MCRS-group is said to be algebraic if it
contains regular operators of GL(n, Z). It would be interesting to develop approxi-
mation techniques of MCRS-groups by algebraic MCRS-groups.

22.1.6 Diophantine Approximation and MCRS-Group

Approximation

The classical problem of approximating real numbers by rational numbers is a par-

ticular case of the problem of best approximations of MCRS-groups in the following
sense.
For a nonzero real number α, we denote by A[α] the MCRS-group of GL(2, R)
R
with invariant lines x = 0 and y = αx. Denote by Ω[0,1] the set of A[α] satisfying
Q R
1 > α > 0. Denote also by Ω[0,1] the subset of Ω[0,1] defined by rational numbers α.
For every pair of relatively prime integers (p, q) satisfying 0 ≤ pq ≤ 1 we have


p
ν A = q.
q
Let us calculate the discrepancy between two MCRS-groups A[α1 ] and A[α2 ]
for arbitrary positive real numbers α1 and α2 ,
x(y − α1 x) x(y − α2 x)
ΦA[α1 ] − ΦA[α2 ] = − = (α2 − α1 )x 2 ,
1 1
x(y − α1 x) x(y − α2 x)
ΦA[α1 ] + ΦA[α2 ] = + = 2xy − (α2 + α1 )x 2 .
1 1
Since α1 > 0 and α2 > 0, we have
 
ρ A[α1 ], A[α2 ] = |α1 − α2 |.

Now the classical problem of approximation of real numbers by rational num-

bers is reformulated in terms of MCRS-groups as follows: for a given MCRS-group
R Q
A[α] ∈ Ω[0,1] and a positive N , find an approximation A[p/q] ∈ Ω[0,1] of size not
Q
exceeding N such that for every other MCRS-group A[p  /q  ] ∈ Ω[0,1] of size not
exceeding N ,
    
ρ A[α], A[p/q] ≤ ρ A[α], A p  /q  .
22.2 Simultaneous Approximation in R3 and MCRS-Group Approximation 353

22.2 Simultaneous Approximation in R3 and MCRS-Group

Approximation
The theory of simultaneous approximation of a real vector by vectors with rational
coefficients is also to some extent a special case of MCRS-group approximation.
We say a few words about this similarity in this section. After a brief description
of the approximation model, we study two particular examples of simultaneous ap-
proximation in the framework of MCRS-group approximation.

22.2.1 General Construction

Let (a, b, c) be a vector in R3 . Denote by A[a, b, c] the maximal commutative sub-

group defined by an operator with the following three eigenvectors

(a, b, c), (0, 1, I ), (0, 1, −I ).

We state the problem of best approximation as follows: for a given MCRS-group

A[a, b, c] and a positive number N , find an approximation A[a  , b , c ] of size not
exceeding N such that for every other MCRS-group A[a  , b , c ] of size not exceed-
ing N , we have
     
ρ A[a, b, c], A a  , b , c ≤ ρ A[a, b, c], A a  , b , c .

Notice that
 
ν A[a, b, c] = max(a, b, c)
and

b2 + c2 3 b 2 c 2 1 2 1 2
ΦA[a,b,c] (x, y, z) = I − x + x y + x z − xy − xz .
2a 2 a a 2 2

The discrepancy between the MCRS-groups A[a, b, c] and A[a  , b , c ] is as fol-

lows:
  
ρ A[a, b, c], A a  , b , c
     

 b b   c c   b2 + c2 b 2 + c 2 
= min max  −  ,  −  ,  − ,
a a a a 2a 2 2a  2 
     

 b b   c c   b2 + c2 b 2 + c 2 
  
max  +  ,  +  ,  + .
a a a a 2a 2 2a  2 
In the case of positive a, b, c, a  , b , and c , we simply have
     

       b b   c c   b2 + c2 b 2 + c 2 

ρ A[a, b, c], A a , b , c = max  −  ,  −  ,    − .
a a a a 2a 2 2a  2 
354 22 Approximation of Maximal Commutative Subgroups

Remark 22.15 For the case of classical simultaneous approximation one takes a
slightly different distance between the vectors:
   

       b b   c c 
  
ρ A[a, b, c], A a , b , c = max  −  ,  −   .
a a a a

So the described approximation model is not exactly the classical model of simulta-
neous approximation.

Let us study two particular examples.

22.2.2 A Ray of a Nonreal Spectrum Operator

Consider the nonreal spectrum algebraic operator

⎛ ⎞
0 1 1
B = ⎝0 0 1⎠ .
1 0 0

This operator can be thought of as the simplest nonreal spectrum operator from a
geometric point of view. Denote the real eigenvalue of the operator B by ξ1 and its
complex conjugate eigenvalues by ξ2 and ξ3 . Notice that

|ξ1 | > |ξ2 | = |ξ3 |.

Let us approximate the eigenline corresponding to ξ1 . Let vξ1 be the vector in this
eigenline having the first coordinate equal to 1. In this setting we have

ξ1 ≈ 1.3247179573 and vξ1 ≈ (1, 0.5698402911, 0.7548776662).

The set of best approximations with size not exceeding 106 contains 48 elements.
All these elements are of type B ni (1, 0, 0), where (ni ) is the following sequence:

(4, 6, 7, 8, 9, 10, . . . , 50, 51, 52).

We conjecture that the set of best approximations coincides with the set of points
B k (1, 0, 0), where k = 4 or k ≥ 6; the approximation rate in this case is CN −3/2 .
The structure of the set of best approximations is usually closely related to the cor-
responding Dirichlet group. In this case the Dirichlet group Ξ (B) is homeomorphic
to Z ⊕ Z/2Z; the generators are B and −Id.
22.2 Simultaneous Approximation in R3 and MCRS-Group Approximation 355

22.2.3 Two-Dimensional Golden Ratio

Consider now the algebraic real spectrum operator of the two-dimensional golden
ratio
⎛ ⎞
3 2 1
G = ⎝2 2 1⎠ .
1 1 1
As we have already seen in Chap. 18, this operator has the simplest continued frac-
tion in the sense of Klein. The Dirichlet group Ξ (G) is generated by the following
two operators:
⎛ ⎞ ⎛ ⎞
1 1 1 0 1 0
H1 = ⎝1 1 0⎠ and H2 = ⎝1 −1 1 ⎠ .
1 0 0 0 1 −1

Notice that G = H12 and H2 = (H1 − Id)−1 . The operator H1 here is the three-
dimensional Fibonacci operator. Denote the eigenvalues of H1 by ξ1 , ξ2 , and ξ3
such that
|ξ1 | > |ξ2 | > |ξ3 |.
Let us approximate the eigenline corresponding to ξ1 . Denote by vξ1 the vector
of this eigenline having the last coordinate equal to 1. Notice that

ξ1 ≈ 2.2469796037 and vξ1 ≈ (2.2469796037, 1.8019377358, 1).

The set of best approximations of size not exceeding 106 contains 41 elements;
40 of them (except for the third best approximation) are contained in the set
 m n 
H1 H2 (1, 0, 0) | m, n ∈ Z .

In the following table we show the consecutive pairs of powers (mi , ni ) for the these
41 best approximations except for the third one. So in the column i we get m = mi ,
n = ni for the approximation H1mi H2ni (1, 0, 0).

i
1 2 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

mi 1 2 3 3 4 4 5 5 6 6 6 7 7 8 8 9 9 10 10 11 11
ni 1 1 2 1 2 1 3 2 3 2 1 3 2 3 2 4 3 4 3 5 4

i
23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41

mi 11 12 12 13 13 14 14 15 15 15 16 16 17 17 18 18 19 19 19
ni 3 4 3 5 4 5 4 6 5 4 5 4 6 5 6 5 7 6 5
356 22 Approximation of Maximal Commutative Subgroups

The third best approximation is A[3, 2, 1]; it is not of the form H1m H2n (1, 0, 0).
We conjecture that the set of all best approximations except A[3, 2, 1] is contained
in the set of all points of type H1m H2n (1, 0, 0), with the approximation rate in this
case being CN −3/2 .
We conclude this chapter with the following general remark.

Remark 22.16 In the above two examples we have several algebraic artifacts (such
as missing B 5 (1, 0, 0) as a best approximation for the first example and an addi-
tional best approximation A[3, 2, 1] for the second example). This is not a surprise,
since we do not approximate the triples of eigenvectors simultaneously but a certain
eigenvector together with two vectors

(0, 1, I ), (0, 1, −I ).

We mix vectors of different natures, and as a result we have irregularities. It is

probable that in most common situations such artifacts may occur infinitely many
times.

22.3 Exercises
Exercise 22.1
(a) Prove that if an operator A is regular, then the centralizer of A ∈ GL(n, R) is
commutative.
(b) Describe all matrices whose centralizers are commutative.

Exercise 22.2 For an arbitrary A ∈ GL(n, R) describe all invariant subspaces of

CGL(n,R) (A).

Exercise 22.3 Let G be a commutative subgroup of GL(n, R). Suppose that for
every matrix A ∈ GL(n, R) \ G there exists a matrix à ∈ G such that A and à do
not commute. Is it true that there exists a matrix B ∈ G such that G = CGL(n,R) (B)?

Exercise 22.4 Prove that the space Simpl2 is homeomorphic to the Möbius band.

Exercise 22.5 Find the first 10 best approximations A[a, b, c] for the minimal and
the middle eigenvectors of the three-dimensional golden ratio G.

Exercise 22.6 Using the theory of MCRS-groups formulate a problem of approxi-

mation of planes in R4 by integer planes.
Chapter 23
Other Generalizations of Continued Fractions

In this chapter we present some other generalizations of regular continued fractions

to the multidimensional case. The main goal for us here is to give different geomet-
ric constructions related to such continued fractions (whenever possible). We say a
few words about Minkowski–Voronoi continued fractions (Sect. 23.1), triangle se-
quences related to Farey addition (Sect. 23.2), O’Hara’s algorithm related to decom-
position of rectangular parallelepipeds (Sect. 23.3), geometric continued fractions
(Sect. 23.4), and determinant generalizations of continued fractions (Sect. 23.5). Fi-
nally, in Sect. 23.6 we describe the relation of regular continued fractions to rational
knots and links.
We do not pretend to give a complete list of generalizations of continued frac-
tions. The idea is to show the diversity of generalizations. Let us give several ref-
erences to some subjects that are beyond the scope of this book but that may be
interesting and useful in the framework of geometry of continued fractions: p-adic
continued fractions [74, 80, 133, 173], complex continued fractions [174], Rosen
continued fractions and their expansions [120, 172], geodesic continued fraction
approach based on Minkowski reduction [128]. Finally, we would like to mention a
relation of continued fractions to tiling of a square by rectangles [58, 126, 170].

23.1 Relative Minima

In this section we consider a geometric generalization of continued fractions in

terms of local minima. It appears for the first time in the works of H. Minkowski
[139] and G.F. Voronoi [204] (see also in [205]) and is used to study units in al-
gebraic fields. Several properties of local minima were later studied by G. Bullig
in [30]. The main object of Minkowski–Voronoi continued fractions is an arbitrary
complete lattice Γ in R3 with respect to the fixed coordinate lattice. Statistical prop-
erties of relative minima were studied by A.A. Illarionov in [82] and [83]. See also
[33] and [71] for more information about three-dimensional relative minima. A de-
scription of the multidimensional case can be found in paper [32] by V.A. Bykovskii.

O. Karpenkov, Geometry of Continued Fractions, 357

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6_23, © Springer-Verlag Berlin Heidelberg 2013
358 23 Other Generalizations of Continued Fractions

The main idea of the Minkowski approach is to define extremal nodes of a lattice
Γ with respect to a family of convex sets that can be chosen with certain freedom
(for instance, vertices of sails considered in previous chapters of this book can be
considered tetrahedral local minima). In this section we describe the classical ap-
proach that works with coordinate parallelepipeds.
For simplicity, in this section we consider only sets in R3 in general position (in
the sense that no two vertices of the lattice are in a plane parallel to some coordinate
plane).

23.1.1 Relative Minima and the Minkowski–Voronoi Complex

We begin with a rather algebraic definition of the Minkowski–Voronoi complex. In

later subsections, we will construct a tessellation of the plane showing the geometric
nature of the complex.
Let T be an arbitrary subset of Rn≥0 = (R≥0 )n . For i = 1, . . . , n we set
 
Ti = max xi | (x1 , . . . , xn ) ∈ T .

We associate to T the following parallelepiped:

 
Π(T ) = (x1 , . . . , xn ) | 0 ≤ xi ≤ Txi , i = 1, . . . , n .

Definition 23.1 Let S be an arbitrary subset of Rn≥0 . An element s ∈ S is called a

Voronoi relative minimum with respect to S if the parallelepiped Π({s}) contains no
points of S except for the origin.

Definition 23.2 Let S be an arbitrary subset of Rn≥0 and let Ŝ be the set of all
its Voronoi relative minima. A finite subset T ⊂ Ŝ is called minimal if the paral-
lelepiped Π(T ) contains no points of Ŝ \ T except for the origin. We denote the set
of all minimal k-element subsets of Ŝ ⊂ S by Mk (S).

It is clear that any minimal subset of a minimal subset is also minimal.

Definition 23.3 A Minkowski–Voronoi complex MV(S) is an (n − 1)-dimensional

complex such that
(i) the k-dimensional faces of MV(S) are enumerated by its minimal (n − k)-
element subsets (i.e., by the elements of Mn−k (S));
(ii) a face with a minimal subset T1 is adjacent to a face with a minimal subset
T2
= T1 if and only if T1 ⊂ T2 .

Remark 23.4 In the three-dimensional case it is also common to consider the

Voronoi and Minkowski graphs that are subcomplexes of the Minkowski–Voronoi
complex. They are defined as follows.
23.1 Relative Minima 359

The Voronoi graph is the graph whose vertices and edges are respectively vertices
and edges of the Minkowski–Voronoi complex.
The Minkowski graph is the graph whose vertices are edges are respectively faces
and edges of the Minkowski–Voronoi complex (two vertices in the Minkowski graph
are connected by an edge if and only if the corresponding faces in the Minkowski–
Voronoi complex have a common edge).
So in some sense Minkowski and Voronoi graphs are dual to each other.

Example 23.5 Let us consider an example of a 6-element set S0 ⊂ R3 defined as

follows
S0 = {s1 , s2 , s3 , s4 , s5 , s6 },
where

s1 = (3, 0, 0), s2 = (0, 3, 0), s3 = (0, 0, 3),

s4 = (2, 1, 2), s5 = (1, 2, 1), s6 = (2, 3, 4).

There are only five Voronoi relative minima for the set S0 , namely the vectors
s1 , . . . , s5 . The Minkowski–Voronoi complex contains 5 vertices, 6 edges, and 5
faces. Its vertices are

v1 = {s1 , s3 , s4 }, v2 = {s3 , s4 , s5 }, v3 = {s1 , s4 , s5 },

v4 = {s2 , s3 , s5 }, v5 = {s1 , s2 , s5 }.

Its edges are

e1 = {s1 , s3 }, e2 = {s3 , s2 }, e3 = {s1 , s2 },

e4 = {s3 , s4 }, e5 = {s1 , s4 }, e6 = {s4 , s5 },
e7 = {s3 , s5 }, e8 = {s1 , s5 }, e9 = {s2 , s5 }.

Its faces are

f1 = {s1 }, f2 = {s2 }, f3 = {s3 }, f4 = {s4 }, f5 = {s5 }.

Finally, we describe the complex MV(S) as a tessellation of an open two-

dimensional disk. We show vertices (on the left), edges (in the middle), and faces
(on the right) separately:
360 23 Other Generalizations of Continued Fractions

23.1.2 Minkowski–Voronoi Tessellations of the Plane

In this subsection we discuss some geometric images standing behind the

Minkowski–Voronoi complex in the three-dimensional case.

Definition 23.6 Let S be an arbitrary subset of R3≥0 . The Minkowski polyhedron for
S is the boundary of the set
 
S ⊕ R3≥0 = s + r | s ∈ S, r ∈ R3≥0 .

In other words, the Minkowski polyhedron is the boundary of the union of copies
of the positive octant shifted by vertices of the set S.

Definition 23.7 The Minkowski–Voronoi tessellation for a set S ⊂ R3≥0 is a tessel-

lation of the plane x + y + z = 0 obtained by the following three steps.
Step 1. Consider the Minkowski polyhedron for the set S and project it orthogo-
nally to the plane x + y + z = 0. This projection induces a tessellation of the plane
by edges of the Minkowski polyhedron.
Step 2. Remove from the tessellation of Step 1 all vertices corresponding to local
minima of the function x + y + z on the Minkowski polyhedron (this is exactly the
relative minima of S). Remove also all edges coming from all the removed vertices.
Step 3. After Step 2 some of the vertices are of valence 1. For each vertex v of
valence 1 and the only remaining edge wv with endpoint at v we replace the edge
wv by the ray wv with vertex at w and passing through v.

Proposition 23.8 The Minkowski–Voronoi tessellation for a nice set S has the com-
binatorial structure of the Minkowski–Voronoi complex MV(S).

Remark 23.9 Here we do not specify what the word “nice” means. We say only that
it includes finite sets and complete lattices considered below.

Example 23.10 Consider the set S0 as in Example 23.5. In Fig. 23.1 we show the
Minkowski polyhedron (on the left) and the corresponding Minkowski–Voronoi tes-
sellation (on the right). The local minima of the function x + y + z on the Minkowski
polyhedron for S0 are the relative minima f1 , . . . , f5 . They identify the faces of the
complex MV(S0 ). The local maxima of the function x + y + z on the Minkowski
polyhedron for S0 are v1 , . . . , v5 , corresponding to vertices of the complex MV(S0 ).
The vertices v1 , . . . , v5 are as follows:

v1 = (3, 1, 3), v2 = (2, 2, 3), v3 = (3, 2, 2),

v4 = (1, 3, 3), v5 = (3, 3, 1).
23.1 Relative Minima 361

Fig. 23.1 The Minkowski polyhedron (on the left) and the corresponding Minkowski–Voronoi
tessellation (on the right)

23.1.3 Minkowski–Voronoi Continued Fractions in R3

Now let us give a definition of Minkowski–Voronoi continued fractions for complete

lattices in R3 . Consider a lattice Γ defined as follows:

Γ = {m1 v1 + m2 v2 + m3 v3 : m1 , m2 , m3 ∈ Z},

where v1 , v2 , v3 are linearly independent vectors in R3 . Define

    
|Γ | = |x|, |y|, |z| | (x, y, z) ∈ Γ \ (0, 0, 0) .

Definition 23.11 We say that an edge {s1 , s2 } of the complex MV(|Γ |) with vertices
{s1 , s2 , s3 } and {s1 , s2 , s4 } is nondegenerate if at least one of the triples (s1 , s2 , s3 )
and (s1 , s2 , s4 ) generates the lattice Γ .
Suppose that {s1 , s2 , s3 } generates the lattice. Then we equip the edge {s1 , s2 }
with a matrix that sends s1 and s2 to themselves and takes s3 to s4 .
So if both triples (s1 , s2 , s3 ) and (s1 , s2 , s4 ) generate the lattice Γ , then the cor-
responding edge is equipped with a pair of matrices inverse to each other.

Finally, we have all the tools that are needed to give a definition of the
Minkowski–Voronoi continued fraction.

Definition 23.12 Let Γ be an arbitrary complete lattice in R3 . The Minkowski–

Voronoi continued fraction for Γ is the Minkowski–Voronoi complex MV(|Γ |) all
of whose nondegenerate edges are equipped.

We refer to [139] and [75, 76] for more details related to general minimal systems
and to the papers [71] and [201] concerning the arbitrary case.
Let us give a small remark on the periodic algebraic case.
362 23 Other Generalizations of Continued Fractions

Remark 23.13 Let α, β, and γ be three roots of a cubic polynomial x 3 + ax 2 +

bx + 1 with integer coefficients a and b. Consider the lattice generated by vectors
 
(1, 1, 1), (α, β, γ ), and α 2 , β 2 , γ 2 .

From Dirichlet’s unit theorem it follows that the multiplicative subgroup of di-
agonal matrices with positive diagonal elements preserving the lattice is isomorphic
to Z2 . The diagonal matrices of this group preserve both the Minkowski–Voronoi
complex and all translation matrices for the corresponding equipped edges. There-
fore, the Minkowski–Voronoi continued fraction in this case is doubly periodic.

Remark 23.14 Let us say a few words about the two-dimensional case. Consider a
complete lattice Γ ⊂ R2 . The Minkowski–Voronoi complex for Γ is a broken line.
Each edge ei is equipped with a matrix

0 ±1
.
1 ai

Consider now sails of lattices in each of the four coordinate octants for the lattice Γ
(i.e., we consider the boundaries of the convex hulls for points of Γ but not of Z3 ).
Then the integers ai correspond to the integer lengths of these sails.

23.1.4 Combinatorial Properties of the Minkowski–Voronoi

Tessellation for Integer Sublattices

If Γ is a full-rank sublattice of Zn , then the Minkowski–Voronoi complex MV(|Γ |)

contains finitely many faces. Not much is known about its combinatorial structure.
The first open problem here is as follows.

Problem 32 Give a combinatorial description (a complete invariant) of all realiz-

able finite Minkowski–Voronoi complexes for integer sublattices.

In particular, it is interesting to consider the following problem.

Problem 33 How many different N -vertex Minkowski–Voronoi complexes for in-

teger lattices do there exist for a fixed N ?

Notice that the problem of a combinatorial description of the Minkowski–

Voronoi complex is also open in the case of infinite complexes as well. Nevertheless
it seems almost impossible to have a nice complete invariant for it, since there is
no testing tool to check whether an infinite complex is realizable as MV(|Γ |) for
some lattice Γ . As in the case of sails in multidimensional continued fractions, the
only infinite case in which it is possible to calculate some examples is the case
of periodic Minkowski–Voronoi complexes corresponding to algebraic lattices (see
23.1 Relative Minima 363

Fig. 23.2 The

Minkowski–Voronoi
tessellation for |Γ171 |

Remark 23.13 above). So for the infinite case, we propose a simpler version of Prob-
lem 32.

Problem 34 Find criteria of nonrealizability of the Minkowski–Voronoi complexes

for arbitrary lattices.

In [210] C. Yannopoulos gives several examples of representations using only

three directions for the edges of the Voronoi graph in the plane (which is the 1-
skeleton of the Minkowski–Voronoi complex). In a recent preprint [100] the authors
extend such representations to the general case of integer sublattices. We announce
this result here.

Theorem 23.15 Let Γ be a sublattice of Zn . The Minkowski–Voronoi tessellation

can be chosen in such a way that the following conditions hold:

3 (for k ∈ Z).
– All finite edges are straight segments of one of the directions kπ
– Each vertex that does not contain infinite edges is a vertex of the following 8
types:

Example 23.16 Consider the lattice Γ171 generated by the following three vectors:

(171, 0, 0), (0, 171, 0), and (2, 32, −1).

Notice that the first integer point on the z-axis is (0, 0, 171). The diagram of
the Minkowski–Voronoi tessellation for |Γ171 | satisfying the conditions of Theo-
rem 23.15 is shown in Fig. 23.2.
364 23 Other Generalizations of Continued Fractions

23.2 Farey Addition, Farey Tessellation, Triangle Sequences

23.2.1 Farey Addition of Rational Numbers

We begin with a definition of Farey addition of rational numbers.

Definition 23.17 Let pq11 and pq22 be two rational numbers. We assume that p1 is
relatively prime to q1 and p2 is relatively prime to q2 . The rational number

p1 p 2 p 1 + p 2
⊕ =
q1 q2 q1 + q2
p1 p2
is called the Farey sum of q1 and q2 . (Note that for every integer m we consider the
fraction m1 .)

There is a nice way to construct the set of rational numbers using the operation
of Farey addition. We do it inductively in countably many steps.

Construction of the Rational Numbers

Base of construction. We start with the sequence of integer numbers (ai,1 ), where
ai,1 = i for i ∈ Z.
The kth step of the construction. Suppose that we have constructed a sequence
(ai,k ) (which is infinite on both sides). Let us extend this sequence to the se-
quence (ai,k+1 ). Set

a2i,k+1 = ai,k , for i ∈ Z;

a2i+1,k+1 = ai,k ⊕ ai+1,k , for i ∈ Z.

In other words, we put between every two numbers of the sequence (ai,k ) their
Farey sum.

Example 23.18 The sequences for the first three steps are as follows:

Base: . . . , −4, −3, −2, −1, 0, 1, 2, 3, 4, . . .

3 1 1 3
Step 1: . . . , −2, − , −1, − , 0, , 1, , 2, . . .
2 2 2 2
2 1 1 1 1 2
Step 2: . . . , −1, − , − , − , 0, , , , 1, . . .
3 2 3 3 2 3
...

Remark 23.19 The part of the sequence (ai,k ) contained between the elements 0
and 1 is sometimes called the kth Farey sequence.
23.2 Farey Addition, Farey Tessellation, Triangle Sequences 365

Remark 23.20 The denominators appearing in the Farey sequences form Stern’s
diatomic sequence, which is defined as follows:

a1 = 1;
a2i = ai , for i ∈ Z+ ;
a2i+1 = ai + ai+1 , for i ∈ Z+ .

The first few elements of this sequence are

1, 1, 2, 1, 3, 2, 3, 1, 4, 3, 5, 2, 5, 3, 4, 1, . . .

Further properties of Stern’s diatomic sequence can be found in a recent survey

[148] by S. Northshield. In this relation we would like to mention the work [112]
by A. Knauf on ferromagnetic spin chains, where this sequence appears as Pascal’s
triangle with memory (see also [110] and [111]).

23.2.2 Farey Tessellation

The above construction of rational numbers appears in hyperbolic geometry as Farey

tessellation. In this section we work with the model of the hyperbolic plane in the
upper half-plane of the real plane R2 with coordinates (x, y).
We add the point at infinity to the line y = 0 and call it the absolute. Considering
the coordinate x as the coordinate on the absolute, we say that the coordinate of
infinity is ∞. For every pair of points (a, b) in the absolute there exists a unique
hyperbolic line passing through these points, denoted by l(a, b). For every line l
in the hyperbolic plane there exists a unique hyperbolic reflection with the axis l.
A triangle all three of whose vertices are on the absolute is called an ideal hyperbolic
triangle.

Definition 23.21 The Farey tessellation is the minimal possible decomposition of

the hyperbolic plane into ideal triangles such that the following two conditions hold:
– it contains the ideal triangle with vertices 0, 1, and ∞;
– the tessellation is preserved by the group of isometries of the hyperbolic plane
generated by all reflections whose axes are sides of ideal triangles in the tessella-
tion.

There is a simple way to construct this tessellation. We start with the ideal tri-
angle with vertices 0, 1, and ∞. Reflecting this triangle with respect to the axes
l(0, 1), l(0, ∞), and l(1, ∞), we get three new triangles with vertex sets {1, 2, ∞},
{−1, 0, ∞}, and {0, 12 , 1} respectively. Add all of them to the tessellation. Continue
iteratively to reflect the obtained picture with respect to the edges of new triangles.
366 23 Other Generalizations of Continued Fractions

Fig. 23.3 Farey tessellation in the hyperbolic plane

In countably many steps we construct all ideal triangles of the tessellation. We show
the Farey tessellation in Fig. 23.3.

Proposition 23.22
(i) The set of vertices of all ideal triangles in the Farey tessellation coincides with
the subset of the points on the absolute whose coordinates are rational or ∞.
(ii) An ideal triangle with vertices p < q < r is in the Farey tessellation if and
only if there exist i and k such that p = ai,k , r = ai+1,k , and q = p ⊕ r (i.e.,
q = a2i+1,k+1 ).

The continued fractions for the vertices of ideal triangles of the Farey tessellation
has the following surprising regularity.

Proposition 23.23 For every ideal triangle T in the Farey tessellation there exists
a sequence of integers a0 , a1 , . . . , an+1 such that the vertices of T have the coordi-
nates p, q, and p ⊕ q, where

p = [0; a0 : a1 : · · · : an ], q = [0; a0 : a1 : · · · : an : an+1 ],

and p ⊕ q = [0; a0 : a1 : · · · : an : an+1 + 1].

23.2.3 Descent Toward the Absolute

It turns out that the continued fraction of a real number itself can be interpreted in
terms of the Farey tessellation. Let us explain this.
It is clear that the Farey tessellation is invariant with respect to the shift on the
vector (1, 0), so for simplicity we restrict ourselves to the case of real numbers α
satisfying 0 ≤ α < 1.
In the upper half-plane model one can imagine the Farey tessellation as a pyra-
mid. We say one more time that we consider now only the part of the tessellation
contained in the band 0 ≤ x < 1. On the top of the pyramid there is the ideal triangle
with vertices 0, 1, and ∞. Suppose that we are at this triangle. It is permitted either
23.2 Farey Addition, Farey Tessellation, Triangle Sequences 367

Fig. 23.4 Descent in the Farey pyramid. The descent sequence is LLRR, and hence the exit point
is 37 = [0; 2 : 3]

to exit the pyramid at vertex 1 or to descend to the adjacent triangle with vertices 0,
1
2 , and 1. To fix notation, we say that we then descend left.
Suppose now that we are at some ideal triangle of the tessellation with vertices
x1 < x2 < x3 . Then it is permitted either to exit the pyramid at vertex x2 or to
descend to one of the neighboring triangles in the tessellation: either to descend left
to the triangle with vertices x1 < x4 < x2 or to descend right to the triangle with
vertices x2 < x5 < x3 for the appropriate point x4 or x5 respectively (see Fig. 23.4).
In a finite or infinite number of steps we descend to some point α of the absolute.
At each step of the descent we go either to the right or to the left, so it is natural to
define the descent sequence whose elements are letters L (if we descend to the left)
and R (if we descend to the right). If the sequence is nonempty, the first is always L.
Let us denote by a1 the number of letters L standing before the first letter R.
Denote by a2 the number of letters R standing before the next letter L, etc.

Example 23.24 For instance, for the sequence (L, L, L, R, R, L, L, L, L, R, R,

L, L) we get

a1 = 3; a2 = 2; a3 = 4; a4 = 2; a5 = 2.

We have the following interesting theorem.

Theorem 23.25 In the above notation the traveler will either exit the pyramid at
the rational number
[0; a1 : a2 : · · · : an−1 : an + 1]
368 23 Other Generalizations of Continued Fractions

in a finite number of steps or descend to the irrational number

[0; a1 : a2 : · · · ]

on the absolute.

For instance, in Example 23.24 we exit at point 238 69

= [0; 3 : 2 : 4 : 2 : 3] in
finitely many steps. In the example considered in Fig. 23.4 we have the descent
sequence LLRR and the exit point 37 = [0; 2 : 3].
Currently it is not clear what the analogue of the Farey tessellation in multidi-
mensional hyperbolic geometry is, so we conclude this section with the following
general open problem.

Problem 35 Find a natural generalization of the Farey tessellation to higher-

dimensional hyperbolic geometry.

In this regard we would like to mention the works [143–147] by V.V. Nikulin on
discrete reflection groups in hyperbolic spaces.

Remark 23.26 It is also interesting to consider Farey tessellation in the de Sitter

geometry, which is in some sense dual to the projective geometry. This is a new ge-
ometric approach to the classic theory of quadratic forms, developed by V.I. Arnold
[5] and F. Aicardi [1] and [2].

23.2.4 Triangle Sequences

In this subsection we briefly observe several results related to triangle sequences.

Triangle sequences generalize classical Farey addition to the two-dimensional case.
They were introduced by T. Garrity in [60]. For multidimensional triangle sequences
we refer to [45].

23.2.4.1 Definition of Triangle Sequence and Its Algebraic Properties

We work with the triangle

 
 = (x, y) | 1 ≥ x ≥ y > 0 .

Consider the following partition of the triangle  into smaller triangles 1 , 2 , . . . ,

where
23.2 Farey Addition, Farey Tessellation, Triangle Sequences 369
 
k = (x, y) ∈  | 1 − x − ky ≥ 0 > 1 − x − (k + 1)y ≥ 0

for k ∈ Z+ . Now let us define the map T that maps each triangle n to the trian-
gle . For every point (x, y) of the triangle n , define

y 1 − x − ky
T (x, y) = , .
x x

The point T (x, y) is in the closure  =  ∪ ∂.

Definition 23.27 Consider an arbitrary point (x, y) of the triangle . Let T k (x, y) ∈
n . Then we put ak = n, for k = 1, 2, . . . . The sequence

(a0 , a1 , a2 , . . .)

is called the triangle sequence for the pair (x, y).

Note that the triangle sequence is said to terminate at step k if T k (x, y) ∈  \ .

In most cases the triangle sequence is infinite.

Remark 23.28 In [138] the authors proved that the triangle map T :  →  is er-
godic with respect to the Lebesgue measure.

23.2.4.2 Multidimensional Farey Addition

The definition of Farey addition is straightforward.

Definition 23.29 Consider two vectors in Rn whose coordinates are represented as

ratios of numbers:

p1 pn r1 rn
T= ,..., and S = ,..., .
q1 qn s1 sn

Then the Farey sum of T and S is the rational vector

p1 + r1 pn + rn
T ⊕S = ,..., .
q1 + s1 qn + sn

We now use the following notation. Let T = (p1 , . . . , pn ) be an integer vector in

Rn . We denote by X̂ the following representation of a rational vector in Rn−1 :

p2 p 3 pn
T̂ = , ,..., .
p1 p1 q1
370 23 Other Generalizations of Continued Fractions

Finally, we put by definition

T +̂ S = T̂ ⊕ Ŝ.

23.2.4.3 Explicit Formula for Vertices of Triangles

The map T defines a natural nested partition of the triangle  into smaller triangles.
This partition contains triangles of the following type.

Definition 23.30 Let (a0 , . . . , an ) be a sequence of nonnegative integers. Define

 
(a0 , . . . , an ) = (x, y) | T k (x, y) ∈ ak , for all k ≤ n .

Our next goal is to write an explicit formula for such triangles. We begin a brief
geometric description of triangle sequences. One can consider the triangle sequence
for (α, β) as a method to construct a sequence of integer vectors (Ck ) whose el-
ements are “almost orthogonal” to the vector (1, α, β), i.e., the dot product of the
vectors in the sequence with the point (1, α, β) in R3 is small. Set
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
1 0 0
C−3 = ⎝0⎠ , C−2 = ⎝1⎠ , C−1 = ⎝0⎠ .
0 0 1

Suppose that the triangle sequence for (α, β) is (ai ). Then we set by definition

Ck = Ck−3 − Ck−2 − ak Ck−1 .

Define also
Xk = Ck × Ck+1 ,
where v × w denotes the cross product of two vectors v and w in R3 .

Theorem 23.31 The vertices of the triangle (a0 , . . . , an ) are X̂n−1 , X̂n , and
Xn +̂ Xn−2 .

As is usual for continued fraction algorithms, the proof is obtained by induction

(see the proof in [12]).

23.2.4.4 Convergence of Triangle Sequences

A triangle sequence does not always converge to a unique number. However, it does
so under certain conditions.
23.2 Farey Addition, Farey Tessellation, Triangle Sequences 371

Consider an infinite triangle sequence (a0 , a1 , a2 , . . .) of nonzero elements. Let

the triangle (a0 , . . . , an ) have vertices X̂n−1 , X̂n , and Xn +̂ Xn−2 . Set

d(X̂n−1 , X̂n+1 )
λn = ,
d(X̂n−1 , Xn +̂ Xn−2 )

where d(Y1 , Y2 ) is the Euclidean distance between the points Y1 and Y2 .

We have the following theorem.

Theorem 23.32 Assume that (an ) is a triangle sequence of nonzero elements. Then
the triangle sequence describes a unique pair (α, β) when
∞
&
(1 − λn ) = 0.
n=N

Let us now give a criterion of nonconvergence of a triangle sequence.

Theorem 23.33 Consider a sequence (an ). Suppose that there exists N such that
an > 0 for any n > N and
∞
&
(1 − λn ) > 0.
n=N

Then the triangle sequence (an ) does not correspond to a unique point.

We refer to [12] for the proof of the above two theorems. Weak convergence is
discussed in [138].

23.2.4.5 Algebraic Periodicity

One of the most challenging properties that one thinks of while generalizing con-
tinued fractions to the multidimensional case is algebraic periodicity. If a triangle
sequence is periodic, then the corresponding pair of real numbers correspond to cer-
tain solutions of a cubic equation with integer coefficients. Although the converse
statement is not true, there is the following nice result announced in [44], which
seems to improve the situation.
For simplicity we embed the triangle  in R3 , simply by adding the first coordi-
˜ The vertices of 
nate 1 and denoting the resulting triangle by . ˜ are
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
1 1 1
v1 = ⎝0⎠ , v2 = ⎝1⎠ , v3 = ⎝1⎠ .
0 0 1
372 23 Other Generalizations of Continued Fractions

Now we define three important matrices:

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
0 0 1 1 0 1 1 1 1
A0 = ⎝1 0 0⎠ , A1 = ⎝0 1 0⎠ , B = ⎝0 1 1⎠ .
0 1 1 0 0 1 0 0 1

We embed the permutation group on three elements into GL(3, Z) as a subgroup

of matrices permuting the columns.

Definition 23.34 For an arbitrary triple of column permutations (σ, τ0 , τ1 ) we de-

fine the following two matrices:

F0,σ,τ0 = σ A0 τ0 and F1,σ,τ1 = σ A1 τ1 .

Denote by k (σ, τ0 , τ1 ) the image of  under the map F1,σ,τ

k
1
F0,σ,τ0 .

We use M t to denote the transpose matrix to M.

Definition 23.35 Let us define the map Tσ,τ0 ,τ1 for an arbitrary triple of column per-
mutations σ, τ0 , τ1 . Consider a pair (x, y) ∈ . Suppose that (x, y) ∈ k (σ, τ0 , τ1 ).
Then we define
 −1 t
(a, b, c) = (1, x, y) · BF0,σ,τ F −k B −1 .
0 1,σ,τ1

We define Tσ,τ0 ,τ1 (x, y) as follows:

b c
Tσ,τ0 ,τ1 (x, y) = , .
a a

Remark 23.36 We have actually defined 216 different maps corresponding to dif-
−1
ferent triples of permutations. Notice that the matrix (BF0,σ,τ F −k B −1 )t is in
0 1,σ,τ1
SL(3, Z). The triangle partition maps Tσ,τ0 ,τ1 are called TRIP maps.

Finally, we describe the following three classes of maps, where n ∈ Z≥0 :

Class 1. Te,e,e ◦ (Te,(123),(123)
1 )n ;
Class 2. Te,(23),e ◦ (Te,(123),(123)
1 )n ;
Class 3. T(23),(23),e ◦ (T(13),(12),e
0 )n .
Notice that for each choice of a class and a parameter n we get a different al-
gorithm defined by the composition T̂ of the corresponding n + 1 TRIP maps.
Similar to Definition 23.35, the composed map T̂ sends the triangle  to the set
(m1 ,m2 ,...,mn ,me ) , which is encoded by n + 1 integer parameters appearing in the
composition T̂ .
23.3 Coordinate Rectangular Bricks and O’Hara’s Algorithm 373

Definition 23.37 Let (x, y) ∈ . Fix a composed map T̂ of one of Classes 1, 2, 3

with a fixed parameter n. For an arbitrary integer k we denote by a k the (n + 1)-
dimensional vector of nonnegative integers such that T̂ k (x, y) is in a n . We say that
(a i ) is the triangle sequence for (x, y) with respect to T̂ .

Now we are ready to formulate an interesting theorem announced in [44] in

which the authors in some sense cover all cubic irrationalities by periodic triangle
sequences of several continued fraction algorithms in the following way.

Theorem 23.38 Let K be a cubic extension of Q. Suppose that the element u ∈ OK

satisfies 0 < u < 1. Then either (u, u2 ), (u2 , u4 ), (u, u2 − u), (u2 , u2 − u4 ),
(uu , (uu )2 − uu ), or ((uu )2 , (uu )2 − (uu )4 ), has a periodic triangle sequence
under some composed map T in Class 1, 2, or 3 (where u is one of the conjugates
of u).

The following conjecture is given for the multidimensional case.

Conjecture 36 (T. Garrity) For each positive integer d, there is a finite num-
ber of multidimensional continued fraction algorithms, so that for every d-tuple
(ξ1 , . . . , ξd ) with Q(ξ1 , . . . , ξd ) a degree-d algebraic number field, there is a multi-
dimensional continued fraction algorithm in the family spanned by the initial algo-
rithms such that it generates a periodic output.

Remark 23.39 Farey fractions give rise to a thermodynamic approach to real num-
bers (for more information we refer to [56, 110, 111], and [61]). Triangle sequences
are used to generalize this approach to the case of pairs of real numbers [62].

Remark 23.40 There are several other generalizations in the spirit of triangle se-
quences. For instance, the classical Minkowski ?(x) function (see [75, 76, 140, 197,
203], etc.) was generalized to the two-dimensional case by O.R. Beaver and T. Gar-
rity in [17] (see also [158]).

23.3 Decompositions of Coordinate Rectangular Bricks and

O’Hara’s Algorithm

Surprisingly, the Euclidean algorithm arises in the theory of partitions related to

O’Hara’s algorithms, introduced by K.M. O’Hara in [151] (see also [152]). In this
section we briefly describe the geometric approach to O’Hara’s algorithm via special
decompositions of parallelepipeds. In particular we describe their relation to regular
continued fractions and their generalizations. For simplicity we consider here only
the finite-dimensional case. For the general infinite-dimensional case for permuta-
tions of Z+ and further information on the finite-dimensional case, we refer to [114]
and [156].
374 23 Other Generalizations of Continued Fractions

Fig. 23.5 A decomposition

of bricks R(6, 11) and
R(22, 3)

23.3.1 Π-Congruence of Coordinate Rectangular Bricks

We begin with several general definitions. If a polyhedron P is a disjoint union

of a finite or countable number of convex polyhedra P1 , P2 , . . . , then we say that
{P1 , P2 , . . .} is a decomposition of P .

Definition 23.41 Let π be a linear subspace of codimension one in Rn . We say that

two convex polyhedra P and Q in Rn are π -congruent if there exist decompositions
{P1 , . . . , Pk } and {Q1 , . . . , Qk } of P and Q respectively such that for every i ∈
{1, . . . , k} there exists a translation Ti along a vector in π taking Pi to Qi (i.e.,
Ti (Pi ) = Qi ).
Two convex polyhedra are said to be Π -congruent if there exists a space π such
that these polyhedra are π -congruent.

Definition 23.42 If instead of finite decompositions of polyhedra we have count-

able decompositions, then the corresponding polyhedra are called asymptotically
π -congruent (respectively asymptotically Π -congruent).

In this section we are interested only in the coordinate parallelepipeds defined as

R(a1 , . . . , an ) = [0, a1 ) × · · · × [0, an ).

We call them coordinate bricks.

Example 23.43 Let us consider two bricks R(6, 11) and R(22, 3). We show their
decompositions in Fig. 23.5 simultaneously. Each of the bricks is subdivided into
six smaller rectangles labeled by Roman numerals. The rectangle with label n is
shifted to the rectangle with label n . For instance, the gray rectangle labeled by V
is taken to the gray rectangle labeled by V  . Here all the translation vectors are in
the linear space x + 2y = 0.

In the next two subsections we answer the following two natural questions:
(a) Which coordinate bricks are (asymptotically) Π -congruent?
23.3 Coordinate Rectangular Bricks and O’Hara’s Algorithm 375

(b) How does one construct the decompositions and translations showing the
(asymptotic) Π -congruence of such boxes?

23.3.2 Criterion of Π-Congruence of Coordinate Bricks

We begin with the following definition.

Definition 23.44 Two coordinate bricks P = R(a1 , . . . , an ) and Q = R(b1 , . . . , bn )

are called relatively rational if there exists λ
= 0 such that λP and λQ are integer
regular parallelepipeds.

In an arbitrary dimension n the following general theorem holds.

Theorem 23.45
(i) Two coordinate bricks of the same dimension are asymptotically Π -congruent
if and only if they have the same volume.
(ii) Two relatively rational rectangular coordinate boxes are Π -congruent if and
only if they have the same volume.

In addition we have the following simple formula to calculate the equation of the
plane π .

Proposition 23.46 Let R(a1 , . . . , an ) and R(b1 , . . . , bn ) be π -congruent. Then the

hyperspace π is uniquely defined by the equation
a1 a1 a2 a1 . . . an−1
x1 + x2 + x3 + · · · + xn = 0.
b2 b2 b3 b 2 . . . bn

For a proof of Theorem 23.45 we refer to [114].

23.3.3 Geometric Version of O’Hara’s Algorithm for Partitions

The answer to the second question (on construction of decompositions and transla-
tions) is given by O’Hara’s algorithm, which comes from the theory of general par-
titions (we refer the interested reader to [3, 151, 152, 155, 156]). In this subsection
we consider only the geometric representation of the algorithm related to continued
fractions. In some sense the idea of this algorithm is an algorithm-definition, giv-
ing a constructive description of all the translations Ti : for an arbitrary point of the
first coordinate brick the algorithm constructs the image of this point in the second
coordinate brick. The resulting decomposition is a decomposition into boxes, all of
whose edges are parallel to coordinate vectors.
376 23 Other Generalizations of Continued Fractions

O’Hara’s Algorithm in a Geometric Form

Predefined parameters of the algorithm. We are given two coordinate bricks
R(a1 , . . . , an ) and R(b1 , . . . , bn ).
Input data. A point v ∈ R(a1 , . . . , an ).
Aim of the algorithm. To construct the point w ∈ R(b1 , . . . , bn ) corresponding
to v.
Step 1. Put v1 = v.
Recursive Step k. We have constructed a point vk−1 . If vk−1 is contained in the
brick R(b1 , . . . , bn ), then the algorithm terminates and we put w = vk−1 . Sup-
pose now that vk−1 is not in the brick R(b1 , . . . , bn ). Then there exists an index
j such that the j th coordinate of vk−1 is not less than bj , allowing us to recur-
sively define
vk = vk−1 − bj ej + aj −1 ej −1
(here ej and ej −1 are the corresponding coordinate vectors whose indices are in
the group Z/nZ).
Output. The algorithm returns a vector w ∈ R(b1 , . . . , bn ).

Remark 23.47 It turns out that the result w does not depend on the choice of the
coordinate j in any recursive step, but entirely on the input data v (see [151] and
[152] for more details).

Now we are ready to define a mapping between the coordinate bricks.

Definition 23.48 Let P and Q be two coordinate bricks of the same volume in Rn .
We define a bijection ϕP ,Q : P → Q at every v ∈ P as follows:

ϕP ,Q (v) = w,

where w is the output of the algorithm with the input data v.

Observe the following property of ϕP ,Q .

Proposition 23.49 The function ϕP ,Q is a piecewise linear bijective function be-

tween the coordinate bricks P and Q.

Due to piecewise linearity we have the following natural decomposition of the

coordinate bricks.

Definition 23.50 Let P and Q be two coordinate bricks of the same volume in
Rn . Consider a natural decomposition of the coordinate brick P into regions on
which the function ϕP ,Q is linear. Define a decomposition of Q as the image of the
decomposition of P via the map ϕP ,Q . We say that these decompositions of P and
Q are associated to the bijection ϕP ,Q .
23.3 Coordinate Rectangular Bricks and O’Hara’s Algorithm 377

Fig. 23.6 A decomposition

of bricks R(9, 25) and
R(25, 9)

In the general case the described decompositions have a countable number of

boxes. However, when P and Q are relatively rational, the decompositions of P
and Q are finite.
The following proposition gives an answer to the second question of this section:
how does one construct the decompositions and translations showing the (asymp-
totic) Π -congruence of such boxes?

Proposition 23.51 Let P = R(a1 , . . . , an ) and Q = R(a1 , . . . , an ) be asymptot-

ically Π -congruent coordinate bricks. Then P and Q are asymptotically π -
congruent, where the hyperplane π is defined by the equation
a1 a1 a2 a1 . . . an−1
x1 + x2 + x3 + · · · + xn = 0.
b2 b2 b3 b 2 . . . bn
In addition, the decompositions of P and Q associated to the bijection ϕP ,Q estab-
lish the asymptotic π -congruence. The restriction of ϕP ,Q to parallelepipeds in the
decomposition of P identifies all the translation vectors.

Here we notice again that in the case of relatively rational coordinate bricks P
and Q, one has π -congruence instead of asymptotic π -congruence.

Remark 23.52 In the special case of rectangular bricks R(a, b) and R(b, a) with
relatively prime positive integers a and b and the space π = {x + y = 0}, O’Hara’s
algorithm is a version of the Euclidean algorithm. Geometrically the elements of the
continued fraction for b/a correspond to the numbers of equivalent squares in the
corresponding layers of the decomposition of R(a, b).

Example 23.53 We illustrate this with the example of a = 9 and b = 25. The as-
sociated decompositions of R(9, 25) and of R(25, 9) consist of two large squares,
one average square, three small squares, and two very small squares. This can be
read from the regular continued fraction: 25 9 = [2; 1 : 3 : 2]. In Fig. 23.6 we show
the layers of equivalent squares as bold rectangles.

Remark 23.54 Notice that after a coordinate rescaling the decomposition remains
combinatorially the same. Therefore, we can always choose the coordinate scale to
work with R(1, b) and R(b, 1), where π = {x + y = 0} for some b > 0.
378 23 Other Generalizations of Continued Fractions

From the above two remarks it follows that the notion of partitions in the two-
dimensional case correlates with the notion of regular continued fractions. So we
conclude with the following natural problem.

Problem 37 (I. Pak [156]) Find a relation between multidimensional continued

fractions and the geometry of the bijection constructed by O’Hara’s algorithm.

23.4 Algorithmic Generalized Continued Fractions

We have already considered two algorithms that generalize the Euclidean algorithm:
O’Hara’s algorithm and triangle sequences. There are in fact many other different
generalizations of the Euclidean algorithm in the spirit of the two mentioned above.
No geometric interpretation is known for most of them. To give some impression
of the algorithmic approaches to continued fractions we mention here several other
algorithms of this nature. We are not planning to go into details of these approaches,
since it is far away from the scope of this book. The best reference to the algorithmic
approach to continued fractions is the book [180] by F. Schweiger (see also [24] and
[195]). First we give a general algorithmic scheme in which such algorithms can be
realized (for additional details we refer to the preprint [181] of F. Schweiger). We
then show some examples of known algorithms. Finally, we briefly discuss some
problems related to algorithmic generalizations of continued fractions.

23.4.1 General Algorithmic Scheme

Most of the generalized Euclidean algorithms work according to the following

scheme.

Scheme of a Generalized Euclidean Algorithm

Predefined parameters of the algorithm. Consider the space Rn for some inte-
ger n. The following data specifies each algorithm:
(i) the domain Δ ⊂ Rn on which the algorithm is defined;
(ii) the operation of subtraction S : Rn → Rn ;
(iii) the operation of inversion T : Rn → Rn ;
(iv) the termination domain Δ̂ ⊂ Δ, at which the algorithm terminates.
We give examples of 4-tuples (Δ, S, T , Δ̂) below. Notice that it is usually re-
quired that
T ◦ S(Δ) ⊂ Δ.
Input data. A real vector v ∈ Δ.
Aim of the algorithm. To construct the generalized continued fraction.
23.4 Algorithmic Generalized Continued Fractions 379

Step 0. Put v0 = v.
Recursive Step k. We have constructed vk−1 . If vk ∈ Δ̂, then the algorithm termi-
nates. If vk ∈
/ Δ̂, the recursive step consists of three substeps:
Substep k.1: perform an operation of subtraction S;
Substep k.2: invert the result via the operation of inversion T ;
Substep k.3: write the (k − 1)th element of the continued fraction based on
Substep k.1 and Substep k.2.
We conclude the step by putting

vk = T ◦ S(vk−1 ).

Output. If the algorithm stops (say at step m + 1), then the last element vm is
considered a generalized greatest common divisor. The corresponding contin-
ued fraction is generated in the iterative steps; it is calculated by parameters of
subtractions S and inversions T .

23.4.2 Examples of Algorithms

First we write the classical Euclidean algorithm in the framework of the general
algorithmic scheme. We will then give several examples of generalizations.

Example 23.55 (Euclidean algorithm) Predefined parameters of the algorithm are

as follows:
 
Δ = (x, y) | x ≥ y > 0; x, y ∈ Z ;
 

x
S : (x, y) "→ x − y, y ;
y
T : (x, y) "→ (y, x);
 
Δ̂ = (x, 0) | x ∈ Z+ .

In Substep k.3 we remember the element ak−1 =  pqkk , where vk−1 = (pk−1 , qk−1 )
is the vector defined in Step k − 1. The resulting continued fraction is

[a0 ; a1 : a2 : · · · ].

Example 23.56 (Jacobi–Perron algorithm ([85, 164, 165])) Historically, the first al-
gorithmic generalization of regular continued fractions is the Jacobi–Perron algo-
rithm (see also [176]). We put
 
Δ = (x0 , x1 , . . . , xn ) | x0 ≥ xi ≥ 0, 1 ≤ i ≤ n ;
     

x0 x2 xn
S : (x0 , x1 , . . . , xn ) "→ x0 − , x1 , x2 − , . . . , xn − ;
x1 x1 x1
380 23 Other Generalizations of Continued Fractions

T : (x0 , x1 , . . . , xn ) "→ (x1 , . . . , xn , x0 );

 
Δ̂ = (x0 , x1 , . . . , xn ) | x1 = 0 ∩ Δ.

In Substep k.3 we remember the n-dimensional vector

     

x0,k−1 x2,k−1 xn,k−1

ak−1 = , ,..., ,
x1,k−1 x1,k−1 x1,k−1

where vk−1 = (x0,k−1 , x1,k−1 , . . . xn,k−1 ) is the vector in Rn+1 defined in Step k − 1.
The resulting generalized continued fraction here is the sequence of n-dimensional
vectors
(a0 , a1 , a2 , . . .).

Remark 23.57 If the Jacobi–Perron algorithm terminates, one actually can remove
the zero coordinate x1 = 0 and continue the algorithm in the real space whose di-
mension is one less (finally we end up at n = 1).

Example 23.58 (Generalized subtractive algorithm ([177])) The generalized sub-

tractive algorithm is defined by the following data. First, p is fixed for the algorithm.
Second, we put
 
Δ = (x0 , x1 , . . . , xn ) | x0 ≥ x1 ≥ · · · ≥ xn ≥ 0 ;
S : (x0 , x1 , . . . , xn ) "→ (x0 − xp , x1 , . . . , xn );
T is a permutation of coordinates putting them in nonincreasing order;
 
Δ̂ = (x0 , x1 , . . . , xn ) | xp = 0 ∩ Δ.

We remember the permutation that was used in Substep k.2. Indeed, we always ex-
change the first element with some element ak−1 ∈ {0, 1, . . . , n}. So we can choose
a sequence of integers
(a0 , a1 , a2 , . . .)
as the generalized p-subtractive continued fraction.

Remark 23.59 The first subtractive algorithm was introduced in [26] by V. Brun,
where he considered the case p = 1. Later, E.S. Selmer in [183, 184] studied the
subtractive algorithm for the case p = n.

Example 23.60 (Fully subtractive algorithm ([178, 179])) In the fully subtractive
algorithm we also fix p and consider
 
Δ = (x0 , x1 , . . . , xn ) | x0 ≥ x1 ≥ · · · ≥ xn ≥ 0 ;
S : (x0 , x1 , . . . , xn ) "→ (x0 − xp , x1 − xp , . . . , xp−1 − xp , xp , . . . , xn );
23.4 Algorithmic Generalized Continued Fractions 381

T is a permutation of coordinates putting them in nonincreasing order;

 
Δ̂ = (x0 , x1 , . . . , xn ) | xp = 0 ∩ Δ.

We remember the permutation that was used in Substep k.2, denoting it by ak−1 .
The generalized continued fraction is the sequence of permutations (a0 , a1 , a2 , . . .).

23.4.3 Algebraic Periodicity

Let us mention here a famous conjecture known as Jacobi’s last theorem.

Problem 38 (Jacobi’s last theorem) Let K be a totally real cubic number field. Con-
sider arbitrary elements y and z of K satisfying 0 < y, z < 1 such that 1, y, and z
are independent over Q. Is it true that the Jacobi–Perron algorithm generates an
eventually periodic continued fraction with starting data v = (1, y, z)?

This problem is open not only for the Jacobi–Perron algorithm but also for other
similar algorithms. The converse statement is true.

Proposition 23.61 Periodic algorithmic continued fractions correspond to certain

algebraic irrationalities.

23.4.4 A Few Words About Convergents

Let us briefly explain how to generate convergents for a vector v with a generalized
continued fraction (a0 , a1 , . . .).
This is usually done by inverting the map T ◦ S. First of all, notice that the values
in the image of the map (T ◦ S)−1 are enumerated by the possible values of the
elements of the corresponding continued fractions, so we can write (T ◦ S)−1 a to
indicate which of the images should be taken.
One starts with some vector w0 . It is natural to choose w0 either (1, 0, . . . , 0) or
the generalized greatest common divisor (in case of existence). Then the vectors

w1 = (T ◦ S)−1
a0 (w0 ), w2 = (T ◦ S)−1
a1 (w1 ), w3 = (T ◦ S)−1
a2 (w2 ), ...

are called the convergents of v.

Remark 23.62 In questions of approximation it is natural to reduce projectively the

first coordinates:

x1 xn
p = (x0 , . . . , xn ) "→ p̃ = ,..., .
x0 x0
Now arises the question of the quality of approximation of the vector ṽ by the vectors
w̃i . For further information on approximation aspects we refer to [180].
382 23 Other Generalizations of Continued Fractions

23.5 Branching Continued Fractions

In this subsection we observe some interesting representations of real numbers as
ratios of certain determinants of infinite matrices related to continued fractions or
their special generalizations. It is nice to admit that for every algebraic number there
exists a periodic representation defining the corresponding number. Matrix repre-
sentations do not give a complete invariant of algebraic numbers in the following
sense: any real number has many different representations, and it is quite hard to de-
termine whether two of them define the same number. For further information and
references we refer to the works of V.Ya. Skorobogatko [189, 190], V.I. Shmoı̆lov
[186–188], I. Gelfand and V. Retakh [64, 66].
We start with a representation of real numbers by tridiagonal determinants whose
elements are the elements of the following continued fractions (for further details
we refer to [87]):
 
a11 a12 0 0 0 · · ·

−1 a22 a23 0 0 · · ·

 0 −1 a33 a34 0 · · ·
 
 0 0 −1 a44 a45 · · ·

 0 0 0 −1 a55 · · ·

a12  · · · · · · · ·
a11 + =   . (23.1)
a23 a22 a23 0 0 · · ·
a22 + a34 
a33 + −1 a33 a34 0 · · ·
a45  
a44 +  0 −1 a44 a45 · · ·
a55 + · · ·  
 0 0 −1 a55 · · ·

 · · · · · · ·

Equation (23.1) holds for any continued fraction whose sequence of convergents
converges to a real number. Here we use dots in matrices to denote the limit of the
ratios of determinants of the corresponding (n + 1) × (n + 1) matrix and n × n
matrix as n tends to infinity. We will also do the same in the above formula. In the
case of Eq. (23.1) we have an even stronger statement for n-convergents:
 
a11 a12 0 · · · 0 0 

−1 a22 a23 · · · 0 0 

 0 −1 a33 · · · 0 0 

 . 
 · · · .. · · 

 0 0 0 · · · an−1,n−1 an−1,n 

a12  0 0 0 ··· −1 an,n 
a11 + =   .
a23 a22 a23 · · · 0 0 
a22 + an−1,n−1 
··· + −1 a33 · · · 0 0 

an,n  . 
 · · .. · · 

 0 0 · · · an−1,n−1 an−1,n 

 0 0 ··· −1 an,n 
23.5 Branching Continued Fractions 383

Let us give the definition of Hessenberg continued fractions.

Definition 23.63 Consider a real number x and a sequence of real numbers ai . Sup-
pose that
 
−a1 −a2 −a3 −a4 · · ·
 
 −1 −a1 −a2 −a3 · · ·
 
 0 −1 −a1 −a2 · · ·

 0 0 −1 −a1 · · ·

 · · · · · ·
x=   ,
−a1 −a2 −a3 · · ·
 
 −1 −a1 −a2 · · ·
 
 0 −1 −a1 · · ·

 · · · · · ·
i.e., the limit from the right exists and equals x. Then the expression from the right
is called the Hessenberg continued fraction for x.

The most interesting case here occurs when a sequence (ai ) has only finitely
many nonzero entries.

Proposition 23.64 Let (ai ) be a sequence of real numbers with finitely many
nonzero elements, assuming that the element an is the last nonzero element. Suppose
that the Hessenberg continued fraction for the sequence (ai ) converges to some real
number x. Then x satisfies the equation

x n + a1 x n−1 + a2 x n−2 + · · · + an−1 x + an = 0.

In the particular case of cubic equations we have an additional nice formula,

which could be considered a generalized two-dimensional regular continued frac-
tion.

Proposition 23.65 Suppose that a Hessenberg continued fraction for a real number
x is defined by a sequence (a1 , a2 , a3 , 0, 0, . . .). Then we have

a2 − a3
a3
a2 −
a1 − · · ·
a1 − a2 − · · ·
a1 −
a1 − · · ·
x = −a1 + a3 .
a2 − a2 − · · ·
a1 −
a1 − · · ·
a1 − a3
a2 −
a1 − · · ·
a1 − a2 − · · ·
a1 −
a1 − · · ·
384 23 Other Generalizations of Continued Fractions

Later A.Z. Nikiporetz gave the following generalization of Hessenberg continued

fractions.

Definition 23.66 Consider a real number xk and a two-sided sequence of real num-
bers (ai ). Suppose that
 
 −ak −ak+1 −ak+2 −ak+3 · · ·

−ak−1 −ak −ak+1 −ak+2 · · ·

−ak−2 −ak−1 −ak −ak+1 · · ·

−ak−3 −ak−2 −ak−1 −ak · · ·

 · · · · . . .
xk =  
 −ak −ak+1 −ak+2 · · ·

−ak−1 −ak −ak+1 · · ·

−ak−2 −ak−1 −ak · · ·

−ak−3 −ak−2 −ak−1 · · ·
 
 · · · . . .
 
−ak−1 −ak −ak+1 −ak+2 · · ·

−ak−2 −ak−1 −ak −ak+1 · · ·

−ak−3 −ak−2 −ak−1 −ak · · ·

−ak−4 −ak−3 −ak−2 −ak−1 · · ·
 
 · · · · . . .
:   ,
−ak−1 −ak −ak+1 · · ·

−ak−2 −ak−1 −ak · · ·

−ak−3 −ak−2 −ak−1 · · ·
 
 · · · . . .

i.e., the limit from the right exists and equals xk . Then the expression from the right
is called the Nikiporetz continued fraction for x.

In analogy to Proposition 23.64 for Hessenberg continued fractions we get the

following statement for Nikiporetz continued fractions.

Proposition 23.67 Let (ai ) be a two-sided sequence of real numbers with finitely
many nonzero elements. We assume that all elements with negative indices are zeros,
a0 = 1, and the element an is the last nonzero element. Suppose that the Nikiporetz
continued fraction for the sequence (ai ) converges to some real number xk . Then xk
satisfies the equation

xkn + a1 xkn−1 + a2 xkn−2 + · · · + an−1 xk + an = 0.

Notice that in the particular case k = 1 we have the statement of Proposi-

tion 23.64.
We continue with Skorobogatko continued fractions.
23.5 Branching Continued Fractions 385

Definition 23.68 The expression

a1 a2
α = b0 + a + a6
b1 + b +a3 4 a5
3 · · · + b4 + · · · b2 + b +
5 · · · + b6 + · · ·

is called the Skorobogatko continued fraction for α.

Proposition 23.69 The Skorobogatko continued fraction for α as in the previous

definition is written in determinant form as follows:
 
 b 0 a 1 a 2 0 0 0 0 · · ·
 
−1 b1 0 a3 a4 0 0 · · ·

−1 0 b2 0 0 a5 a6 · · ·

 0 −1 0 b3 0 0 0 · · ·
 
 0 −1 0 0 b4 0 0 · · ·

0 0 −1 0 0 b5 0 · · ·

0 0 −1 0 0 0 b6 · · ·

 · · · · · · · · · ·
α=   .
 b1 0 a3 a4 0 0 · · ·

0 b2 0 0 a5 a6 · · ·

−1 0 b3 0 0 0 · · ·
 
−1 0 0 b4 0 0 · · ·

 0 −1 0 0 b5 0 · · ·
 
 0 −1 0 0 0 b6 · · ·
 
 · · · · · · · · ·

Finally, we would like to mention another generalization of continued fractions

for the case of noncommuting elements and its expression in terms of quasidetermi-
nants studied by I. Gelfand and V. Retakh in [64, 66] (see also in [157]). Here we
formulate only a consequence for the case of commutative elements (for the non-
commutative case we refer the interested reader to the original papers mentioned
above).

Definition 23.70 Consider a real number x and a collection of real numbers ai,j .
Suppose that
 
a11 a12 a13 a14 · · ·
 
−1 a22 a23 a24 · · ·
 
 0 −1 a33 a34 · · ·
 
 0 0 −1 a44 · · ·

 · · · · · · ·
x=   ,
a22 a23 a24 · · ·
 
−1 a33 a34 · · ·
 
 0 −1 a44 · · ·
 
 · · · · · ·
386 23 Other Generalizations of Continued Fractions

i.e., the limit from the right exists and equals x. Then the expression from the right
is called the quasideterminant for x.

We mention once more that the notion of quasideterminant in the case of non-
commutative elements ai,j is more complicated; see [65] and [66].

Proposition 23.71 In the setting of the above definition, we have

 1
x = a11 + a1j1  .
1
j1
=1 a2j1 + a2j2
a3j2 + · · ·
j2
=1,j1

23.6 Continued Fractions and Rational Knots and Links

In this section we explain how continued fractions are used in topology for the
classification of rational knots and links. For additional information and references
we refer to [103] and [104].

23.6.1 Necessary Definitions

Recall that a knot is a smooth embedding of a circle R/Z into R3 ; a link is a smooth
embedding of several circles into R3 .

Definition 23.72 We say that a knot K has an n-bridge representation if K is iso-

topic to some knot having only n maxima and n minima as critical points of the
natural height function on K given by the z-coordinate in R3 .
The bridge number of a knot K is the minimal number n such that K has an
n-bridge representation.
We say that a knot of bridge number n is an n-bridge knot.
A 2-bridge knot is said to be a rational knot.

For the study of n-bridge knots it is natural to consider n-tangles.

Definition 23.73 An n-tangle is a proper embedding of the disjoint union of n arcs

into a three-dimensional ball with 2n marked points such that the endpoints of the
arcs map to distinct marked boundary points.

An n-tangle is called trivial if its endpoints are connected by straight lines.

23.6 Continued Fractions and Rational Knots and Links 387

23.6.2 Rational Tangles and Operations on Them

Definition 23.74 We say that a 2-tangle is a rational tangle if there exists a smooth
deformation of the three-dimensional ball under which the 2-tangle evolves to the
trivial 2-tangle.

Graphically, tangles are represented as special regular mappings of the arcs to

the plane (i.e., having only finitely many singular points, and such that all these
singularities are double crossings). Two pairs of endpoints are mapped to the lines
y = 0 and y = 1 respectively, and the rest are mapped to the band bounded by these
lines. At each crossing we indicate which of the two branches of the double crossing
is above and which is below. (See an example of a tangle in Fig. 23.7.)

Basic Tangles First we define the following three tangles:

T (0) = , T (1) = , T (∞) = .

Let us now define the addition, multiplication by −1, and inversion operations
on 2-tangles.

The Sum Operation The sum of two tangles T1 and T2 is the tangle T1 + T2
defined as follows:

T 1 + T2 = .

Multiplication by −1 If we change all the crossings in the tangle T , then we get

the tangle called opposite to T and denoted by −T .

The Inversion Operation Let T be an arbitrary 2-tangle. The inverse tangle T i

is defined as follows:

Ti = .

Definition 23.75 Let m be a positive integer. We define

T (m) = T (1) + · · · + T (1) .

!" #
m

Letting [a0 ; a1 : a2 : · · · : an ] be a continued fraction with integer coefficients,

then set
     i i i
T [a0 ; a1 : a2 : · · · : an ] = T (a0 ) + T (a1 ) + T (a2 ) + · · · + T i (an ) .
388 23 Other Generalizations of Continued Fractions

Fig. 23.7 A rational tangle

T ([4; 2 : 3])

As we will see later, in Theorem 23.76, the isotopy class of a tangle T ( pq ) does
not depend on the choice of the integer elements of the continued fractions repre-
senting the fraction pq .

23.6.3 Main Results on Rational Knots and Tangles

The next theorem shows that the set of rational tangles is in a natural one-to-one
correspondence with the rational numbers. This theorem is a reformulation of the
Conway theorem from [36].

Theorem 23.76
(i) Two tangles T ([a0 ; a1 : a2 : · · · : an ]) and T ([b0 ; b1 : b2 : · · · : bm ]) for contin-
ued fractions with integer coefficients are isotopic if and only if

[a0 ; a1 : a2 : · · · : an ] = [b0 ; b1 : b2 : · · · : bm ].
p
(ii) For every rational tangle T there exists a rational number q such that T is
isotopic to T ( pq ).

To get a knot from a tangle one should use the following closing operation.
Each tangle diagram has two endpoints on the line y = 0. Connect them by some
curve that does not intersect the diagram. Do the same for the two endpoints on the
line y = 1. See an example of a tangle T ( 32 ) and the corresponding trefoil knot in
Fig. 23.8.

Theorem 23.77 (H. Schubert [175]) Let pq11 and pq22 be two rational numbers satis-
fying gcd(p1 , q1 ) = gcd(p2 , q2 ) = 1. Then the corresponding knots or links K( pq11 )
and K( pq22 ) are isotopic if and only if the following two conditions hold:
(a) p1 = p2 ;
(b) either q1 ≡ q2 (mod p1 ) or q1 q2 ≡ 1(mod p1 ).
23.6 Continued Fractions and Rational Knots and Links 389

Fig. 23.8 A rational tangle and its closure

Remark 23.78 The classification of rational knots arises in the topology of three-
dimensional manifolds. Namely, the 2-fold branch coverings spaces of S 3 along the
rational knots and links represent lens spaces; see the book [182] for more details.

So the problem of classification of rational knots is completely solved. The next

step in this direction would be the classification of 3-bridge knots.

Problem 39 Describe all 3-bridge knots and links.

It is natural to suppose that one should use an appropriate generalization of con-

tinued fractions to investigate this question. So we have an additional question on a
natural extension of continued fractions.

Problem 40 Find a generalization of continued fractions that classifies rational 3-

tangles.

Rational 3-tangles can be defined similarly to rational 2-tangles. Nevertheless,

we leave some freedom in the definition in order to try to achieve the goals of Prob-
lem 39.
References

1. F. Aicardi, Symmetries of quadratic form classes and of quadratic surd continued fractions. I.
A Poincaré tiling of the de Sitter world. Bull. Braz. Math. Soc. (N.S.) 40(3), 301–340 (2009)
2. F. Aicardi, Symmetries of quadratic form classes and of quadratic surd continued fractions.
II. Classification of the periods’ palindromes. Bull. Braz. Math. Soc. (N.S.) 41(1), 83–124
(2010)
3. G.E. Andrews, The Theory of Partitions. Cambridge Mathematical Library (Cambridge Uni-
versity Press, Cambridge, 1998). Reprint of the 1976 original
4. H. Appelgate, H. Onishi, The similarity problem for 3 × 3 integer matrices. Linear Algebra
Appl. 42, 159–174 (1982)
5. V. Arnold, Arithmetics of binary quadratic forms, symmetry of their continued fractions
and geometry of their de Sitter world. Bull. Braz. Math. Soc. (N.S.) 34(1), 1–42 (2003).
Dedicated to the 50th anniversary of IMPA
6. V.I. Arnold, Statistics of integral convex polygons. Funct. Anal. Appl. 14(2), 79–81 (1980).
Russian version: Funkc. Anal. Prilozh. 14(2), 1–3 (1980)
7. V.I. Arnold, A-graded algebras and continued fractions. Commun. Pure Appl. Math. 42(7),
993–1000 (1989)
8. V.I. Arnold, Ordinary Differential Equations. Springer Textbook (Springer, Berlin, 1992).
Translated from the third Russian edition by Roger Cooke
9. V.I. Arnold, Higher-dimensional continued fractions. Regul. Chaotic Dyn. 3(3), 10–17
(1998). J. Moser at 70 (in Russian)
10. V.I. Arnold, Preface, in Pseudoperiodic Topology. Am. Math. Soc. Transl. Ser. 2, vol. 197
(Am. Math. Soc., Providence, 1999), pp. ix–xii
11. V.I. Arnold, Continued Fractions (Moscow Center of Continuous Mathematical Education,
Moscow, 2002) (in Russian)
12. S. Assaf, L.-C. Chen, T. Cheslack-Postava, B. Cooper, A. Diesl, T. Garrity, M. Lepinski,
A. Schuyler, A dual approach to triangle sequences: a multidimensional continued fraction
algorithm. Integers 5(1), A8 (2005)
13. J. Barajas, O. Serra, The lonely runner problem with seven runners, in Fifth Conference
on Discrete Mathematics and Computer Science, University of Valladolid. Ciencias (Val-
ladolid), vol. 23 (Secr. Publ. Intercamb. Ed., Valladolid, 2006), pp. 111–116 (in Spanish)
14. J. Barajas, O. Serra, The lonely runner with seven runners. Electron. J. Comb. 15(1), R48
(2008)
15. I. Bárány, A.M. Vershik, On the number of convex lattice polytopes. Geom. Funct. Anal.
2(4), 381–393 (1992)
16. M. Barile, D. Bernardi, A. Borisov, J.-M. Kantor, On empty lattice simplices in dimension 4.
Proc. Am. Math. Soc. 139(12), 4247–4253 (2011)

O. Karpenkov, Geometry of Continued Fractions, 391

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6, © Springer-Verlag Berlin Heidelberg 2013
392 References

17. O.R. Beaver, T. Garrity, A two-dimensional Minkowski ?(x) function. J. Number Theory
107(1), 105–134 (2004)
18. M. Beck, S. Robins, Computing the Continuous Discretely: Integer-Point Enumeration in
Polyhedra. Undergraduate Texts in Mathematics (Springer, New York, 2007)
19. U. Betke, J.M. Wills, Untere Schranken für zwei diophantische Approximations-Funktionen.
Monatshefte Math. 76, 214–217 (1972)
20. W. Bienia, L. Goddyn, P. Gvozdjak, A. Sebő, M. Tarsi, Flows, view obstructions, and the
lonely runner. J. Comb. Theory, Ser. B 72(1), 1–9 (1998)
21. T. Bohman, R. Holzman, D. Kleitman, Six lonely runners. Electron. J. Comb. 8(2), R3
(2001). In honor of Aviezri Fraenkel on the occasion of his 70th birthday
22. A.I. Borevich, I.R. Shafarevich, Number Theory. Pure and Applied Mathematics, vol. 20
(Academic Press, New York, 1966). Translated from the Russian by Newcomb Greenleaf
23. P. Borwein, Computational Excursions in Analysis and Number Theory. CMS Books in
Mathematics, vol. 10 (Springer, New York, 2002)
24. A.J. Brentjes, Multidimensional Continued Fraction Algorithms. Mathematical Centre
Tracts, vol. 145 (Mathematisch Centrum, Amsterdam, 1981)
25. K. Briggs, Klein polyhedra. http://keithbriggs.info/klein-polyhedra.html
26. V. Brun, Algorithmes euclidiens pour trois et quatre nombres, in Treizième congrès des math-
ématiciens scandinaves, tenu à Helsinki 18–23 août 1957 (Mercators Tryckeri, Helsinki,
1958), pp. 45–64
27. A.D. Bryuno, V.I. Parusnikov, Klein polyhedra for two Davenport cubic forms. Math. Notes
- Ross. Akad. 56(3–4), 994–1007 (1995). Russian version: Mat. Zametki 56(4), 9–27 (1994)
28. J. Buchmann, A generalization of Voronoı̆’s unit algorithm. I. J. Number Theory 20(2), 177–
191 (1985)
29. J. Buchmann, A generalization of Voronoı̆’s unit algorithm. II. J. Number Theory 20(2),
192–209 (1985)
30. G. Bullig, Zur Kettenbruchtheorie im Dreidimensionalen (Z 1). Abh. Math. Semin. Hansis-
che Univ. 13, 321–343 (1940)
31. R.T. Bumby, Hausdorff dimensions of Cantor sets. J. Reine Angew. Math. 331, 192–206
(1982)
32. V.A. Bykovskiı̆, Connectivity of Minkowskii Graph for Multidimensional Complete Lattices
(Dalnauka, Vladivostok, 2000)
33. V.A. Bykovskiı̆, O.A. Gorkusha, Minimal bases of three-dimensional complete lattices. Sb.
Math. 192(1–2), 215–223 (2001). Russian version: Mat. Sb. 192(2), 57–66 (2001)
34. J.W.S. Cassels, H.P.F. Swinnerton-Dyer, On the product of three homogeneous linear forms
and the indefinite ternary quadratic forms. Philos. Trans. R. Soc. Lond. Ser. A 248, 73–96
(1955)
35. H. Cohen, A Course in Computational Algebraic Number Theory. Graduate Texts in Mathe-
matics, vol. 138 (Springer, Berlin, 1993)
36. J.H. Conway, An enumeration of knots and links, and some of their algebraic properties,
in Computational Problems in Abstract Algebra, Oxford, 1967 (Pergamon, Oxford, 1970),
pp. 329–358
37. T.W. Cusick, View-obstruction problems. Aequ. Math. 9, 165–170 (1973)
38. T.W. Cusick, The largest gaps in the lower Markoff spectrum. Duke Math. J. 41, 453–463
(1974)
39. T.W. Cusick, View-obstruction problems in n-dimensional geometry. J. Comb. Theory, Ser.
A 16, 1–11 (1974)
40. T.W. Cusick, View-obstruction problems. II. Proc. Am. Math. Soc. 84(1), 25–28 (1982)
41. T.W. Cusick, M.E. Flahive, The Markoff and Lagrange Spectra. Mathematical Surveys and
Monographs, vol. 30 (Am. Math. Soc., Providence, 1989)
42. T.W. Cusick, C. Pomerance, View-obstruction problems. III. J. Number Theory 19(2), 131–
139 (1984)
43. V.I. Danilov, The geometry of toric varieties. Russ. Math. Surv. 33(2), 97–154 (1978). Rus-
sian version: Usp. Mat. Nauk 33(2), 85–134 (1978)
References 393

44. K. Dasaratha, L. Flapan, T. Garrity, Ch. Lee, C. Mihaila, N. Neumann-Chun, S. Peluse, M.

Stroffregen, Cubic irrationals and periodicity via a family of multi-dimensional continued
fraction algorithms (2012). arXiv:1208.4244
45. K. Dasaratha, L. Flapan, T. Garrity, Ch. Lee, C. Mihaila, N. Neumann-Chun, S. Peluse,
M. Stroffregen, A generalized family of multidimensional continued fractions: TRIP maps
(2012). arXiv:1206.7077
46. H. Davenport, On the product of three homogeneous linear forms. I. Proc. Lond. Math. Soc.
13, 139–145 (1938)
47. H. Davenport, On the product of three homogeneous linear forms. II. Proc. Lond. Math. Soc.
(2) 44, 412–431 (1938)
48. H. Davenport, On the product of three homogeneous linear forms. III. Proc. Lond. Math.
Soc. (2) 45, 98–125 (1939)
49. H. Davenport, Note on the product of three homogeneous linear forms. J. Lond. Math. Soc.
16, 98–101 (1941)
50. H. Davenport, On the product of three homogeneous linear forms. IV. Proc. Camb. Philos.
Soc. 39, 1–21 (1943)
51. H. Davenport, On the product of three non-homogeneous linear forms. Proc. Camb. Philos.
Soc. 43, 137–152 (1947)
52. B.N. Delone, D.K. Faddeev, The Theory of Irrationalities of the Third Degree. Translations
of Mathematical Monographs, vol. 10 (Am. Math. Soc., Providence, 1964)
53. J.W. Demmel, Applied Numerical Linear Algebra (Society for Industrial and Applied Math-
ematics, Philadelphia, 1997)
54. E. Ehrhart, Sur les polyèdres rationnels homothétiques à n dimensions. C. R. Math. Acad.
Sci. Paris 254, 616–618 (1962)
55. G. Ewald, Combinatorial Convexity and Algebraic Geometry. Graduate Texts in Mathemat-
ics, vol. 168 (Springer, New York, 1996)
56. J. Fiala, P. Kleban, Generalized number theoretic spin chain-connections to dynamical sys-
tems and expectation values. J. Stat. Phys. 121(3–4), 553–577 (2005)
57. D. Fowler, The Mathematics of Plato’s Academy: A New Reconstruction, 2nd edn. (Oxford
University Press, New York, 1999)
58. C. Freiling, M. Laczkovich, D. Rinne, Rectangling a rectangle. Discrete Comput. Geom.
17(2), 217–225 (1997)
59. W. Fulton, Introduction to Toric Varieties. Annals of Mathematics Studies, vol. 131 (Prince-
ton University Press, Princeton, 1993). The William H. Roever Lectures in Geometry
60. T. Garrity, On periodic sequences for algebraic numbers. J. Number Theory 88(1), 86–103
(2001)
61. T. Garrity, A thermodynamic classification of real numbers. J. Number Theory 130(7), 1537–
1559 (2010)
62. T. Garrity, A thermodynamic classification of pairs of real numbers via the triangle multi-
dimensional continued fraction (2012). arXiv:1205.5663
63. C.F. Gauss, Recherches Arithmétiques (Blanchard, Paris, 1807)
64. I. Gelfand, V. Retakh, Quasideterminants. I. Sel. Math. New Ser. 3(4), 517–546 (1997)
65. I.M. Gelfand, V.S. Retakh, Determinants of matrices over noncommutative rings. Funct.
Anal. Appl. 25(2), 91–102 (1991). Russian version: Funkc. Anal. Prilozh. 25(2), 13–25
(1991)
66. I.M. Gelfand, V.S. Retakh, Theory of noncommutative determinants, and characteristic func-
tions of graphs. Funct. Anal. Appl. 26(4), 231–246 (1993). Russian version: Funkc. Anal.
Prilozh. 26(4), 1–20 (1992)
67. O.N. German, Sails and Hilbert bases. Proc. Steklov Inst. Math. 239(4), 88–95 (2002). Rus-
sian version: Tr. Mat. Inst. Steklova 239(4), 98–105 (2002)
68. O.N. German, Sails and norm minima of lattices. Sb. Math. 196(3), 337–367 (2005). Russian
version: Mat. Sb. 196(3), 31–60 (2005)
69. O.N. German, Klein polyhedra and norm minima of lattices. Dokl. Akad. Nauk, Ross. Akad.
Nauk 406(3), 298–302 (2006)
394 References

70. O.N. German, E.L. Lakshtanov, On a multidimensional generalization of Lagrange’s theorem

for continued fractions. Izv. Math. 72(1), 47–61 (2008). Russian version: Izv. Ross. Akad.
Nauk, Ser. Mat. 72(1), 51–66 (2008)
71. O.A. Gorkusha, Minimal bases of three-dimensional complete lattices. Math. Notes - Ross.
Akad. 69(3–4), 320–328 (2001). Russian version: Mat. Zametki 69(3), 353–362 (2001)
72. P.M. Gruber, C.G. Lekkerkerker, Geometry of Numbers, 2nd edn. North-Holland Mathemat-
ical Library, vol. 37 (North-Holland, Amsterdam, 1987)
73. F.J. Grunewald, Solution of the conjugacy problem in certain arithmetic groups, in Word
Problems II, Oxford, 1976. Stud. Logic Foundations Math., vol. 95 (North-Holland, Amster-
dam, 1980), pp. 101–139
74. J. Hančl, A. Jaššová, P. Lertchoosakul, R. Nair, On the metric theory of p-adic continued
fractions. Indag. Math. (N.S.) 24(1), 42–56 (2013)
75. H. Hancock, Development of the Minkowski Geometry of Numbers, vol. 1 (Dover, New York,
1964)
76. H. Hancock, Development of the Minkowski Geometry of Numbers, vol. 2 (Dover, New York,
1964)
77. C. Hermite, Letter to C. D. J. Jacobi. J. Reine Angew. Math. 40, 286 (1839)
78. K. Hessenberg, Thesis, Technische Hochschule, Darmstadt, Germany, 1942
79. A. Higashitani, M. Masuda, Lattice multi-polygons (2012). arXiv:1204.0088
80. J. Hirsh, L.C. Washington, p-adic continued fractions. Ramanujan J. 25(3), 389–403 (2011)
81. F. Hirzebruch, Über vierdimensionale Riemannsche Flächen mehrdeutiger analytischer
Funktionen von zwei komplexen Veränderlichen. Math. Ann. 126, 1–22 (1953)
82. A.A. Illarionov, Estimates for the number of relative minima of lattices. Math. Notes - Ross.
Akad. 89(1–2), 245–254 (2011). Russian version: Mat. Zametki 89(2), 249–259 (2011)
83. A.A. Illarionov, The average number of relative minima of three-dimensional integer lattices.
St. Petersburg Math. J. 23(3), 551–570 (2012). Russian version: Algebra Anal. 23(3), 189–
215 (2011)
84. T.H. Jackson, Note on the minimum of an indefinite binary quadratic form. J. Lond. Math.
Soc. (2) 5, 209–214 (1972)
85. C.G.J. Jacobi, Allgemeine Theorie der Kettenbruchähnlichen Algorithmen, in welchen jede
Zahl aus drei vorhergehenden gebildet wird (Aus den hinterlassenen Papieren von C. G. J.
Jacobi mitgetheilt durch Herrn E. Heine). J. Reine Angew. Math. 69(1), 29–64 (1868)
86. H.W.E. Jung, Darstellung der Funktionen eines algebraischen Körpers zweier unabhängigen
Veränderlichen x, y in der Umgebung einer Stelle x = a, y = b. J. Reine Angew. Math. 133,
289–314 (1908)
87. V.F. Kagan, Origins of Determinant Theory (Gos. Izd. Ukraine, Odessa, 1922) (in Russian)
88. J.-M. Kantor, K.S. Sarkaria, On primitive subdivisions of an elementary tetrahedron. Pac. J.
Math. 211(1), 123–155 (2003)
89. O. Karpenkov, On the triangulations of tori associated with two-dimensional continued frac-
tions of cubic irrationalities. Funct. Anal. Appl. 38(2), 102–110 (2004). Russian version:
Funkc. Anal. Prilozh. 38(2), 28–37 (2004)
90. O. Karpenkov, On two-dimensional continued fractions of hyperbolic integer matrices with
small norm. Russ. Math. Surv. 59(5), 959–960 (2004). Russian version: Usp. Mat. Nauk
59(5), 149–150 (2004)
91. O. Karpenkov, Classification of three-dimensional multistoried completely hollow convex
marked pyramids. Russ. Math. Surv. 60, 165–166 (2005). Russian version: Usp. Mat. Nauk
60(1), 169–170 (2005)
92. O. Karpenkov, Three examples of three-dimensional continued fractions in the sense of
Klein. C. R. Math. Acad. Sci. Paris 343(1), 5–7 (2006)
93. O. Karpenkov, On an invariant Möbius measure and the Gauss-Kuz’min face distribution.
Proc. Steklov Inst. Math. 258, 74–86 (2007). Russian version: Tr. Mat. Inst. Steklova 258,
79–92 (2007)
94. O. Karpenkov, Elementary notions of lattice trigonometry. Math. Scand. 102(2), 161–205
(2008)
References 395

95. O. Karpenkov, Constructing multidimensional periodic continued fractions in the sense of

Klein. Math. Comput. 78(267), 1687–1711 (2009)
96. O. Karpenkov, On irrational lattice angles. Funct. Anal. Other Math. 2(2–4), 221–239 (2009)
97. O. Karpenkov, On determination of periods of geometric continued fractions for two-
dimensional algebraic hyperbolic operators. Math. Notes - Ross. Akad. 88(1–2), 28–38
(2010). Russian version: Mat. Zametki 88(1), 30–42 (2010)
98. O. Karpenkov, Continued fractions and the second Kepler law. Manuscr. Math. 134(1–2),
157–169 (2011)
99. O. Karpenkov, Multidimensional Gauss reduction theory for conjugacy classes of SL(n, Z).
Preprint (2012)
100. O.N. Karpenkov, A.V. Ustinov, On construction of the Minkowski-Voronoi complex.
Preprint (2013)
101. O.N. Karpenkov, A.M. Vershik, Rational approximation of maximal commutative subgroups
of GL(n, R). J. Fixed Point Theory Appl. 7(1), 241–263 (2010)
102. S. Katok, Continued fractions, hyperbolic geometry and quadratic forms, in MASS Selecta
(Am. Math. Soc., Providence, 2003), pp. 121–160
103. L.H. Kauffman, S. Lambropoulou, On the classification of rational knots. Enseign. Math. (2)
49(3–4), 357–410 (2003)
104. L.H. Kauffman, S. Lambropoulou, On the classification of rational tangles. Adv. Appl. Math.
33(2), 199–237 (2004)
105. A.Ya. Khinchin, Continued Fractions (FISMATGIS, Moscow, 1961)
106. A.G. Khovanskiı̆, Newton polytopes, curves on toric surfaces, and inversion of Weil’s the-
orem. Russ. Math. Surv. 52(6), 1251–1279 (1997). Russian version: Usp. Mat. Nauk 52(6),
113–142 (1997)
107. V. Klee, Some characterizations of convex polyhedra. Acta Math. 102, 79–107 (1959)
108. F. Klein, Ueber eine geometrische Auffassung der gewöhnliche Kettenbruchentwicklung.
Nachr. Ges. Wiss. Göttingen Math-Phys. Kl. 3, 352–357 (1895)
109. F. Klein, Sur une représentation géométrique de développement en fraction continue ordi-
naire. Nouv. Ann. Math. 15(3), 327–331 (1896)
110. A. Knauf, On a ferromagnetic spin chain. Commun. Math. Phys. 153(1), 77–115 (1993)
111. A. Knauf, On a ferromagnetic spin chain. II. Thermodynamic limit. J. Math. Phys. 35(1),
228–236 (1994)
112. A. Knauf, Number theory, dynamical systems and statistical mechanics. Rev. Math. Phys.
11(8), 1027–1060 (1999)
113. M.L. Kontsevich, Yu.M. Suhov, Statistics of Klein polyhedra and multidimensional contin-
ued fractions, in Pseudoperiodic Topology. Am. Math. Soc. Transl. Ser. 2, vol. 197 (Am.
Math. Soc., Providence, 1999), pp. 9–27
114. M. Konvalinka, I. Pak, Geometry and complexity of O’Hara’s algorithm. Adv. Appl. Math.
42(2), 157–175 (2009)
115. S.V. Konyagin, K.A. Sevastyanov, Estimate of the number of vertices of a convex integral
polyhedron in terms of its volume. Funkc. Anal. Prilozh. 18(1), 13–15 (1984)
116. E. Korkina, La périodicité des fractions continues multidimensionnelles. C. R. Acad. Sci.,
Sér. 1 Math. 319(8), 777–780 (1994)
117. E.I. Korkina, The simplest 2-dimensional continued fraction, in International Geometrical
Colloquium, Moscow (1993)
118. E.I. Korkina, Two-dimensional continued fractions. The simplest examples. Tr. Mat. Inst.
Steklova 209, 143–166 (1995)
119. E.I. Korkina, The simplest 2-dimensional continued fraction. J. Math. Sci. 82(5), 3680–3685
(1996)
120. C. Kraaikamp, I. Smeets, Approximation results for α-Rosen fractions. Unif. Distrib. Theory
5(2), 15–53 (2010)
121. R.O. Kuzmin, On one problem of Gauss. Dokl. Akad. Nauk SSSR, Ser. A, 375–380 (1928)
122. R.O. Kuzmin, On a problem of Gauss. Atti del Congresso Internazionale dei Matematici,
Bologna 6, 83–89 (1932)
396 References

123. G. Lachaud, Polyèdre d’Arnol’d et voile d’un cône simplicial: analogues du théorème de
Lagrange. C. R. Acad. Sci., Sér. 1 Math. 317(8), 711–716 (1993)
124. G. Lachaud, Sails and Klein polyhedra, in Number Theory, Tiruchirapalli, 1996. Contemp.
Math., vol. 210 (Am. Math. Soc., Providence, 1998), pp. 373–385
125. G. Lachaud, Voiles et Polyhedres de Klein (Act. Sci. Ind., Hermann, 2002)
126. M. Laczkovich, G. Szekeres, Tilings of the square with similar rectangles. Discrete Comput.
Geom. 13(3–4), 569–572 (1995)
127. J.C. Lagarias, Best simultaneous Diophantine approximations. I. Growth rates of best ap-
proximation denominators. Trans. Am. Math. Soc. 272(2), 545–554 (1982)
128. J.C. Lagarias, Geodesic multidimensional continued fractions. Proc. Lond. Math. Soc. (3)
69(3), 464–488 (1994)
129. J.-L. Lagrange, Solution d’un problème d’arithmétique, in Œuvres de Lagrange, vol. 1
(Gauthier-Villars, Paris, 1867–1892), pp. 671–732
130. A.K. Lenstra, H.W. Lenstra Jr., L. Lovász, Factoring polynomials with rational coefficients.
Math. Ann. 261(4), 515–534 (1982)
131. P. Levy, Sur les lois de probabilité dont dependent les quotients complets et incomplets d’une
fraction continue. Bull. Soc. Math. Fr. 57, 178–194 (1929)
132. J. Lewis, D. Zagier, Period functions and the Selberg zeta function for the modular group,
in The Mathematical Beauty of Physics, Saclay, 1996. Adv. Ser. Math. Phys., vol. 24 (World
Scientific, River Edge, 1997), pp. 83–97
133. K. Mahler, Lectures on Diophantine Approximations. Part I: g-adic Numbers and Roth’s
Theorem (University of Notre Dame Press, Notre Dame, 1961). Prepared from the notes by
R.P. Bambah of my lectures given at the University of Notre Dame in the Fall of 1957
134. Y.I. Manin, M. Marcolli, Continued fractions, modular symbols, and noncommutative geom-
etry. Sel. Math. New Ser. 8(3), 475–521 (2002)
135. A. Markoff, Sur les formes quadratiques binaires indéfinies. Math. Ann. 15(3–4), 381–406
(1879)
136. A. Markoff, Sur les formes quadratiques binaires indéfinies. Math. Ann. 17(3), 379–399
(1879)
137. P. Mattila, Geometry of Sets and Measures in Euclidean Spaces: Fractals and Rectifiability.
Cambridge Studies in Advanced Mathematics, vol. 44 (Cambridge University Press, Cam-
bridge, 1995)
138. A. Messaoudi, A. Nogueira, F. Schweiger, Ergodic properties of triangle partitions. Monat-
shefte Math. 157(3), 283–299 (2009)
139. H. Minkowski, Généralisation de la théorie des fractions continues. Ann. Sci. Éc. Norm. Sup.
(3) 13, 41–60 (1896)
140. H. Minkowski, Gesammelte Abhandlungen (Am. Math. Soc., Chelsea, 1967), pp. 293–315
141. J.-O. Moussafir, Sails and Hilbert bases. Funct. Anal. Appl. 34(2), 114–118 (2000). Russian
version: Funkc. Anal. Prilozh. 34, (2) 2000, 43–49
142. J.-O. Moussafir, Voiles et Polyédres de Klein: Geometrie, Algorithmes et Statistiques. Doc-
teur en sciences thèse, Université Paris IX—Dauphine, 2000
143. V.V. Nikulin, On the classification of arithmetic groups generated by reflections in
Lobachevskiı̆ spaces. Izv. Akad. Nauk SSSR, Ser. Mat. 45(1), 113–142 (1981)
144. V.V. Nikulin, On the classification of arithmetic groups generated by reflections in
Lobachevskiı̆ spaces. Izv. Akad. Nauk SSSR, Ser. Mat. 45(1), 240 (1981)
145. V.V. Nikulin, Quotient-groups of groups of automorphisms of hyperbolic forms by subgroups
generated by 2-reflections. Algebro-geometric applications, in Current Problems in Mathe-
matics, vol. 18 (Akad. Nauk SSSR, Vsesoyuz. Inst. Nauchn. i Tekhn. Informatsii, Moscow,
1981), pp. 3–114
146. V.V. Nikulin, Discrete reflection groups in Lobachevsky spaces and algebraic surfaces, in
Proceedings of the International Congress of Mathematicians, Berkeley, CA, 1986, vol. 1
(Am. Math. Soc., Providence, 1987), pp. 654–671
147. V.V. Nikulin, Finiteness of the number of arithmetic groups generated by reflections in
Lobachevskiı̆ spaces. Izv. Math. 71(1), 53–56 (2007). Russian version: Izv. Ross. Akad.
References 397

Nauk, Ser. Mat. 71(1), 55–60 (2007)

148. S. Northshield, Stern’s diatomic sequence 0, 1, 1, 2, 1, 3, 2, 3, 1, 4, . . . . Am. Math. Mon.
117(7), 581–598 (2010)
149. T. Oda, Convex Bodies and Algebraic Geometry: An Introduction to the Theory of Toric
Varieties. Ergebnisse der Mathematik und ihrer Grenzgebiete (3), vol. 15 (Springer, Berlin,
1988). Translated from the Japanese
150. J. O’Hara, Energy of Knots and Conformal Geometry. Series on Knots and Everything,
vol. 33 (World Scientific, River Edge, 2003)
151. K.M. O’Hara, Structure and Complexity of the Involution Principle for Partitions (ProQuest
LLC, Ann Arbor, 1984). Ph.D. thesis, University of California, Berkeley
152. K.M. O’Hara, Bijections for partition identities. J. Comb. Theory, Ser. A 49(1), 13–25 (1988)
153. R. Okazaki, On an effective determination of a Shintani’s decomposition of the cone Rn+ .
J. Math. Kyoto Univ. 33(4), 1057–1070 (1993)
154. J.M. Ortega, H.F. Kaiser, The LLT and QR methods for symmetric tridiagonal matrices.
Comput. J. 6, 99–101 (1963/1964)
155. I. Pak, Partition identities and geometric bijections. Proc. Am. Math. Soc. 132(12), 3457–
3462 (2004) (electronic)
156. I. Pak, Partition bijections, a survey. Ramanujan J. 12(1), 5–75 (2006)
157. I. Pak, A. Postnikov, V. Retakh, Noncommutative Lagrange theorem and inversion polyno-
mials. Preprint (1995). www.math.ucla.edu/~pak/papers
158. G. Panti, Multidimensional continued fractions and a Minkowski function. Monatshefte
Math. 154(3), 247–264 (2008)
159. V.I. Parusnikov, Klein’s polyhedra for the third extremal ternary cubic form. Technical report,
preprint 137, Keldysh Institute of the RAS, Moscow (1995)
160. V.I. Parusnikov, Klein’s polyhedra for the fifth extremal cubic form. Technical report, preprint
69, Keldysh Institute of the RAS, Moscow (1998)
161. V.I. Parusnikov, Klein’s polyhedra for the seventh extremal cubic form. Technical report,
preprint 79, Keldysh Institute of the RAS, Moscow (1999)
162. V.I. Parusnikov, Klein polyhedra for the fourth extremal cubic form. Math. Notes - Ross.
Akad. 67(1–2), 87–102 (2000). Russian version: Mat. Zametki 67(1), 110–128 (2000)
163. V.I. Parusnikov, Klein polyhedra for three extremal cubic forms. Math. Notes - Ross. Akad.
77(4), 566–583 (2005). Russian version: Mat. Zametki 77(3–4), 523–538 (2000)
164. O. Perron, Grundlagen für eine Theorie des Jacobischen Kettenbruchalgorithmus. Math.
Ann. 64(1), 1–76 (1907)
165. O. Perron, Erweiterung eines Markoffschen Satzes über die Konvergenz gewisser Ketten-
brüche. Math. Ann. 74(4), 545–554 (1913)
166. O. Perron, Über die Approximation irrationaler Zahlen durch rationale. I. Essay 4, S.-B.
Heidelberg Akad. Wiss., 1921
167. O. Perron, Über die Approximation irrationaler Zahlen durch rationale. II. Essay 8, S.-B.
Heidelberg Akad. Wiss., 1921
168. B. Poonen, F. Rodriguez-Villegas, Lattice polygons and the number 12. Am. Math. Mon.
107(3), 238–250 (2000)
169. P. Popescu-Pampu, The geometry of continued fractions and the topology of surface singu-
larities, in Singularities in Geometry and Topology 2004 (2007), pp. 119–195
170. M. Prasolov, M. Skopenkov, Tiling by rectangles and alternating current. J. Comb. Theory,
Ser. A 118(3), 920–937 (2011)
171. D. Repovsh, M. Skopenkov, M. Tsentsel, An elementary proof of the twelve lattice point the-
orem. Math. Notes - Ross. Akad. 77(1–2), 108–111 (2005). Russian version: Mat. Zametki
77(1), 117–120 (2005)
172. D. Rosen, A class of continued fractions associated with certain properly discontinuous
groups. Duke Math. J. 21, 549–563 (1954)
173. A.A. Ruban, Certain metric properties of the p-adic numbers. Sib. Mat. Zh. 11, 222–227
(1970)
398 References

174. A.L. Schmidt, Ergodic theory of complex continued fractions, in Number Theory with an
Emphasis on the Markoff Spectrum, Provo, UT, 1991. Lecture Notes in Pure and Appl. Math.,
vol. 147 (Dekker, New York, 1993), pp. 215–226
175. H. Schubert, Knoten mit zwei Brücken. Math. Z. 65, 133–170 (1956)
176. F. Schweiger, The Metrical Theory of Jacobi-Perron Algorithm. Lecture Notes in Mathemat-
ics, vol. 334 (Springer, Berlin, 1973)
177. F. Schweiger, Ergodic properties of multi-dimensional subtractive algorithms, in New Trends
in Probability and Statistics, Palanga, 1991, vol. 2 (VSP, Utrecht, 1992), pp. 91–100
178. F. Schweiger, Invariant measures for fully subtractive algorithms. Anz. Österreich. Akad.
Wiss. Math.-Natur. Kl. 131, 25–30 (1995)
179. F. Schweiger, Fully subtractive algorithms. Österreich. Akad. Wiss. Math.-Natur. Kl.
Sitzungsber. II 204, 23–32 (1996)
180. F. Schweiger, Multidimensional Continued Fractions. Oxford Science Publications (Oxford
University Press, Oxford, 2000)
181. F. Schweiger, Multidimensional continued fractions—new results and old problems. Preprint
(2008). http://www.cirm.univ-mrs.fr/videos/2008/exposes/287w2/Schweiger.pdf
182. H. Seifert, W. Threlfall, Seifert and Threlfall: A Textbook of Topology. Pure and Applied
Mathematics, vol. 89 (Academic Press, New York, 1980). Translated from the German edi-
tion of 1934 by Michael A. Goldman, with a preface by Joan S. Birman, with “Topology of
3-dimensional fibered spaces” by Seifert, translated from the German by Wolfgang Heil
183. E.S. Selmer, Continued fractions in several dimensions. Nord. Mat. Tidskr. 9, 37–43 (1961)
184. E.S. Selmer, Continued fractions in several dimensions. Nord. Mat. Tidskr. 9, 95 (1961)
185. T. Shintani, On evaluation of zeta functions of totally real algebraic number fields at non-
positive integers. J. Fac. Sci., Univ. Tokyo, Sect. 1A, Math. 23(2), 393–417 (1976)
186. V.I. Shmoı̆lov, Nepreryvnye Drobi. Tom I. Periodicheskie Nepreryvnye Drobi (Merkator,
Lviv, 2004)
187. V.I. Shmoı̆lov, Nepreryvnye Drobi. Tom II. Raskhodyashchiesya Nepreryvnye Drobi (Merka-
tor, Lviv, 2004)
188. V.I. Shmoı̆lov, Nepreryvnye Drobi. Tom III. Iz Istorii Nepreryvnykh Drobei (Merkator, Lviv,
2004)
189. V.Ya. Skorobogatko, Ideas and results of the theory of branching continued fractions and
their application to the solution of differential equations, in General Theory of Boundary
Value Problems (Naukova Dumka, Kiev, 1983), pp. 187–197
190. V.Ya. Skorobogatko, Teoriya Vetvyashchikhsya Tsepnykh Drobei i Ee Primenenie v Vychis-
litelnoi Matematike (Nauka, Moscow, 1983)
191. B.F. Skubenko, Minima of a decomposable cubic form of three variables. Zapiski nauch.
sem. LOMI (1988)
192. B.F. Skubenko, Minima of decomposable forms of degree n of n variables for n ≥ 3. Zapiski
nauch. sem. LOMI (1990)
193. J. Stoer, R. Bulirsch, Introduction to Numerical Analysis, 3rd edn. Texts in Applied Mathe-
matics, vol. 12 (Springer, New York, 2002). Translated from the German by R. Bartels, W.
Gautschi and C. Witzgall
194. H.P.F. Swinnerton-Dyer, On the product of three homogeneous linear forms. Acta Arith. 18,
371–385 (1971)
195. G. Szekeres, Multidimensional continued fractions. Ann. Univ. Sci. Bp. Rolando Eötvös
Nomin., Sect. Math. 13, 113–140 (1970)
196. E. Thomas, A.T. Vasquez, On the resolution of cusp singularities and the Shintani decompo-
sition in totally real cubic number fields. Math. Ann. 247(1), 1–20 (1980)
197. R.F. Tichy, J. Uitz, An extension of Minkowski’s singular function. Appl. Math. Lett. 8(5),
39–46 (1995)
198. L.N. Trefethen, D. Bau, Numerical Linear Algebra (Society for Industrial and Applied Math-
ematics, Philadelphia, 1997)
199. A. Trevisan, Lattice polytopes and toric varieties. Master’s thesis, Leiden University, 2007.
http://www.math.leidenuniv.nl/scripties/trevisan.pdf
References 399

200. H. Tsuchihashi, Higher-dimensional analogues of periodic continued fractions and cusp sin-
gularities. Tohoku Math. J. 35(4), 607–639 (1983)
201. A.V. Ustinov, Minimal vector systems in 3-dimensional lattices and analog of Vahlen’s the-
orem for 3-dimensional Minkowski’s continued fractions. Sovrem. Probl. Mat. 16, 103–128
(2012)
202. A.M. Vershik, Statistical mechanics of combinatorial partitions, and their limit configura-
tions. Funct. Anal. Appl. 30(2), 90–105 (1996)
203. P. Viader, J. Paradís, L. Bibiloni, A new light on Minkowski’s ?(x) function. J. Number
Theory 73(2), 212–227 (1998)
204. G. Voronoı̆, A generalization of the algorithm of continued fractions. Ph.D. thesis, Warsaw,
1896 (in Russian)
205. G.F. Voronoı̆, On a Generalization of the Algorithm of Continued Fraction. Collected Works
in Three Volumes (USSR Ac. Sci, Kiev, 1952) (in Russian)
206. D. Wells, The Penguin Dictionary of Curious and Interesting Numbers (Penguin, Middlesex,
1986)
207. G.K. White, Lattice tetrahedra. Can. J. Math. 16, 389–396 (1964)
208. J.M. Wills, Zwei Sätze über inhomogene diophantische Approximation von Irrationalzahlen.
Monatshefte Math. 71, 263–269 (1967)
209. E. Wirsing, On the theorem of Gauss-Kusmin-Lévy and a Frobenius-type theorem for func-
tion spaces. Acta Arith. 24, 507–528 (1973/1974). Collection of articles dedicated to Carl
Ludwig Siegel on the occasion of his seventy-fifth birthday, V
210. C. Yannopoulos, Zur Kettenbruchtheorie im Dreidimensionalen. Math. Z. 47, 105–110
(1940)
211. R.T. Živaljević, Rotation number of a unimodular cycle: an elementary approach (2012).
arXiv:1209.4981
Index

Symbols to find ς -reduced matrices in integer

π -congruent polyhedra, 374 conjugacy classes, 322
asymptotically, 374 Angle, rational, 24, 41
π -integer point of a sail, 313 Approximation
Π -congruent polyhedra, 374 best Diophantine, 115
asymptotically, 374 strong, 120
ς -complexity, 78 nth special polyhedron, 289
size of, 124
A Arrangement
A-hull, 216 algebraic, 131
badly approximable, 129
Absolute, 365
Asymptotic direction, 329
Affinely irrational vector, 220
Average
Algorithm
space, 100
Euclidean, 5, 379
time, 100
generalized Euclidean, general scheme, 378
Jacobi–Perron, 379
B
LLL-, 245 Binet’s formula, 17
of sail reconstruction, 252 Boundary of a set, 216
O’Hara’s, 375
subtractive C
fully, 380 Centralizer, 348
generalized, 380 Characteristic function of a set, 107
to construct a regular continued fraction, 4 Complex projective toric surface, 175
to construct an integer nth root of a Conjecture
GL(2, Z) matrix, 90 Arnold, 262
to construct broken lines related to Garrity, 373
sequences of real numbers, 140 Littlewood, 269
to construct faces of a continued fraction lonely runner, 207
first deductive, 286 Oppenheim, 269
second deductive, 287 Constant
to construct fundamental domains, Euler–Mascheroni, 104
inductive, 282 Freiman’s, 82
to construct perfect Hessenberg matrices, Continued fraction, 3
306 element of, 3
to construct reduced matrices, 77 even, 3
to construct sails of real numbers, 37 for an arbitrary broken line, 142

O. Karpenkov, Geometry of Continued Fractions, 401

Algorithms and Computation in Mathematics 26,
DOI 10.1007/978-3-642-39368-6, © Springer-Verlag Berlin Heidelberg 2013
402 Index

Continued fraction (cont.) F

for an O-broken line, 138 Face of a convex set, 217
geometric, 68 Face of a sail, 218
associated to a matrix, 68 Factor-continued fraction, 313
Hessenberg, 383 Factor-sail, 313
infinite, 4 Farey
Minkowski–Voronoi, 361 sequence, kth, 364
multidimensional sum, 364
associated to an operator, 249 of vectors, 369
associated with given planes, 217 tessellation, 365
finite, 218 Form
in the sense of Klein–Voronoi, 313 associated to an arrangement, 123
of algebraic irrationality, 251 associated to an operator, 319
torus decomposition, 253 Markov–Davenport, 123
Nikiporetz, 384 Frequency
odd, 3 of a face in the sense of Arnold, 275
of toric singularity, 178 of an element, 108
periodic, 93 relative
regular, 4 of a face, 275
Skorobogatko, 385 of an edge, 113
Convergent, 4
Coordinate brick, 374 G
relatively rational, 375 Gauss map, 102
Coset, left and right, 21 Gauss–Kuzmin distribution, 99
Cross-ratio, 109 Gaussian
infinitesimal, 109 integer, 350
Cyclotomic polynomial, 240 one-dimensional subspace, 350
vector, 350
D primitive, 350
Decomposition of a polyhedron, 374 Golden ratio, 17
Density generalized, 254
angular, 146 Graph
areal, 146 Minkowski, 359
Descent sequence, 367 Voronoi, 359
Determinant Grassmann (exterior) algebra, 194
of a face, 269 Grassmannian
of an edge star, 269 linear, 195
of the lattice, 244 oriented lattice, 196
Dirichlet group, 87, 238
positive, 88, 238 H
Discrepancy, 124 Hessenberg
Dual polygon, 30 complexity, 303
type, 303
E Hyperflag, 282
Edge star, 265
regular, 265 I
with respect to a point, 265 IDC
Ehrhart polynomial, 199 -nodes, 208
Empty -system, 208
integer triangle, 25 Ideal hyperbolic triangle, 365
polyhedron, 189, 203 Index
Euler totient, 240 of a rational angle, 25
Extremal vertex, edge, ray of a convex set, 230 of a subgroup, 21
Index 403

Integer pseudo-regular, 63
affine type, 251 self-dual, 63
angle, 20 vector, 185
adjacent, 51 volume, 187, 192
associated to an extended angle, 161 Integer-linear congruence of faces, 275
between planes, 188 Inverse broken line, 154
between spaces, 187
irrational, 47 K
L-irrational, 47 k-form, 194
LR-irrational, 47 Kepler planetary motion, 150
opposite interior, 55 Knot, 386
R-irrational, 47 n-bridge, 386
right, 54 n-bridge representation of, 386
transpose, 49 rational, 386
approximation space, 220
arctangent, 48 L
area, 24, 27 Lexicographic
arrangement, 124 basis, 195
broken line, 137 ordering, 195
cone, 185
LLS period, 70
congruence, 20, 185
LLS sequence, 43
proper, 138, 154
for a broken line, 142
conjugate matrices, 67, 301
of a matrix, 69
cosine, 161
LSLS sequence, 138
distance, 23
between planes, 188
extended angle, 154 M
corresponding to an integer angle, 162 M-sum
normal form, 157 of extended integer angles, 163
opposite, 162 of integer angles, 164
geometry, 20 Marked pyramid, 204
lattice, 21 Markov
length, 22 Diophantine equation, 83
line, 20 minimum, 131
parallel lines, 55 number, 83
plane, 185 spectrum, 81
point, 20, 185 multidimensional, 82
polygon, 20 tripple, 84
polyhedron, 185 Markov–Davenport
pyramid, 205 characteristic, 319
completely empty, 205 form for an MCRS-group, 349
one-story, multistory, 205 Matrix
ray, 329 complex spectrum, 71
segment, 20 Frobenius (companion), 304
sine, 42, 161 transpose, 261
sublattice, 21 Hessenberg, 301, 303
affine, 21 ς -reduced, ς -nonreduced, 304
subspace, 220 perfect, 303
tangent, 43, 47, 161 irreducible, 250
triangle NRS-, 327
dual, 62 real spectrum, 72, 241
integer isosceles, 63 reduced, 72
integer regular, 63 regular, 348
pseudo-isosceles, 63 RS-, 250, 327
404 Index

Maximal commutative subgroup Q

(MCRS-group), 348 Quadratic irrationality, 94
algebraic, 352 Quasideterminant, 386
discrepancy between, 351 Quasipolyhedral convex set, 231
rational, 350
real spectrum, 348 R
size of, 351 Reduced basis, 244
Minimal subset, 358 Regular 2-star of a vertex, 291
Minkowski–Voronoi Revolution number, 155
complex, 358 Rotation number, 156
tessellation, 360
Möbius S
form, 110 Sail
multidimensional, 272 algebraic, 70
measure, 111 chain of, 266
multidimensional, 272, 348 U -periodic, 266
Module of a field, 237 duality, 38, 69
complete, incomplete, 238 finite, 218
Multiple polygon, broken line, 60 fundamental domain of, 251, 315
hyperface of, 282
N multidimensional, 216, 313
n-sided simplicial cone, 348 of an angle, 33, 43
n-star of a vertex, 291 orbit face of, 314
pair of singularity, 178
Necessary and sufficient conditions of broken
principal part of, 33
lines to be closed, 145
Simplicial cone, 216
NRS-ray, 329
integer, 216
rational, 216
O
Space
Order of a field, 238
measure, 100
maximal, 238
probability, 100
Oriented lattice, 196 Space of framed continued fractions
geometric one-dimensional, 110
P multidimensional, 272
Pell’s equation, 91 Stern’s diatomic sequence, 365
Pick’s formula, 29 Sum
Plücker of sets, 219
coordinates, 195 of tangles, 387
lattice, 197 Supporting hyperplane, 217
embedding, 195
lattice, 196 T
Polyhedron Tangle, 386
generalized, 231 closing operation, 388
Minkowski, 360 inverse, 387
Problem opposite, 387
Arnold, 262, 265, 296 rational, 387
Jacobi’s last theorem, 381 trivial, 386
of best approximations Theorem
classical, 115 Birkhoff–Khinchin’s ergodic, 101
for MCRS-groups, 353 Birkhoff’s pointwise ergodic, 100
half-rule condition, 116 Dirichlet’s unit, 238
Pak, 378 Gauss–Kuzmin
view-obstruction, 207 original, 108
Projective linear group, 109 pointwise, 107
Index 405

Theorem (cont.) Triangle sequence, 369

Kronecker’s approximation, 219 terminates, 369
multidimensional, 221 TRIP map, 372
Lagrange, 94
geometric generalization, 266 U
Lebesgue density, 101 Unit in an order, 238
on geometric complete invariant of matrix
integer congruence, 317 V
on sum of integer tangents in integer V -broken line, 137
triangles, 60 equivalent, 154
Schubert, 388 Voronoi relative minimum, 358
twelve-point, 30
White’s, 203 W
Toric singularities, 176 w-sail, w-continued fraction
Transformation geometric, 130
ergodic, 100 multidimensional, 325
measure-preserving, 100 Well-placed dihedral angle, 291