Handbook of Geometric

Constraint Systems Principles

Discrete Mathematics and Its Applications
Series Editors
Miklos Bona
Donald L. Kreher
Douglas West
Patrice Ossona de Mendez

Handbook of Product Graphs, Second Edition

Richard Hammack, Wilfried Imrich, and Sandi Klavžar
Handbook of Mathematical Induction:
Theory and Applications
David S. Gunderson
Graph Polynomials
Yongtang Shi, Matthias Dehmer, Xueliang Li, and Ivan Gutman
Introduction to Combinatorics, Second Edition
W. D. Wallis and J. C. George
Representation Theory of Symmetric Groups
Pierre-Loïc Méliot
Advanced Number Theory with Applications
Richard A. Mollin
A Multidisciplinary Introduction to Information Security
Stig F. Mjølsnes
Combinatorics of Compositions and Words
Silvia Heubach and Toufik Mansour
Handbook of Linear Algebra, Second Edition
Leslie Hogben
Combinatorics, Second Edition
Nicholas A. Loehr
Handbook of Discrete and Computational Geometry, Third Edition
C. Toth, Jacob E. Goodman and Joseph O’Rourke
Handbook of Discrete and Combinatorial Mathematics, Second Edition
Kenneth H. Rosen
Crossing Numbers of Graphs
Marcus Schaefer
Graph Searching Games and Probabilistic Methods
Anthony Bonato and Paweł Prałat
Handbook of Geometric Constraint Systems Principles
Meera Sitharam, Audrey St. John, and Jessica Sidman,
https://www.crcpress.com/Discrete-Mathematics-and-Its-Applications/book-series/CHDISMTHAPP?page=1
&order=dtitle&size=12&view=list&status=published,forthcoming
 Handbook of Geometric
Constraint Systems Principles

Meera Sitharam
Audrey St. John
Jessica Sidman
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2019 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed on acid-free paper

Version Date: 20180528

International Standard Book Number-13: 978-1-4987-3891-0 (Hardback)

This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been
made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity
of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright
holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this
form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may
rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized
in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying,
microfilming, and recording, or in any information storage or retrieval system, without written permission from the
publishers.

For permission to photocopy or use material electronically from this work, please access www.copyright.com
(http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers,
MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of
users. For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been
arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for
identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
 To the works of:

Hilda Polaczek Geiringer 1893–1973

Neil White 1945–2014

Michel Marie Deza 1939–2016

Wenjun Wu 1919–2017
There is no royal road to geometry.
– Euclid

Geometry is not true, it is advantageous.

– Henri Poincare

Geometry, inasmuch as it is concerned with real space, is no longer considered a part of pure
mathematics; like mechanics and physics, it belongs among the applications of mathematics.
– Hermann Weyl

Think geometrically, prove algebraically.

– John Tate
Contents

Foreword xxi

Preface xxiii

Contributors xxv

1 Overview and Preliminaries 1

Meera Sitharam and Troy Baker
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Specifying a GCS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.2 Fundamental GCS Questions . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.3 Tractability and Computational Complexity . . . . . . . . . . . . . . . . . 3
1.2 Parts and Chapters of the Handbook . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Part I: Geometric Reasoning Techniques . . . . . . . . . . . . . . . . . . . 4
1.2.2 Part II: Distance Geometry, Configuration Space, and Real Algebraic Ge-
ometry Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.3 Part III: Geometric Rigidity Techniques . . . . . . . . . . . . . . . . . . . 6
1.2.4 Part IV: Combinatorial Rigidity Techniques . . . . . . . . . . . . . . . . . 7
1.2.4.1 Inductive Constructions . . . . . . . . . . . . . . . . . . . . . . 8
1.2.4.2 Body Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.4.3 Body-Cad, and Point-Line Frameworks . . . . . . . . . . . . . . 9
1.2.4.4 Symmetric and Periodic Frameworks and
Frameworks under Polyhedral Norms . . . . . . . . . . . . . . . 9
1.2.5 Missing Topics and Chapters . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Terminology Reconciliation and Basic Concepts . . . . . . . . . . . . . . . . . . 10
1.3.1 Constrainedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.2 Rigidity of Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.3 Generic Rigidity of Frameworks . . . . . . . . . . . . . . . . . . . . . . . 12
1.3.4 Approximate Degree-of-Freedom and Sparsity . . . . . . . . . . . . . . . 13
1.4 Alternative Pathway through the Book . . . . . . . . . . . . . . . . . . . . . . . . 15

I Geometric Reasoning, Factorization and Decomposition 19

2 Computer-Assisted Theorem Proving in Synthetic Geometry 21
Julien Narboux, Predrag Janičić, and Jacques Fleuriot
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2 Automated Theorem Proving . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.1 Foundations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.2 Nondegenerate Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.3 Purely Synthetic Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.3.1 Early Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2.3.2 Deductive Database Method, GRAMY, and
iGeoTutor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

ix
x Contents

2.2.3.3 Logic-Based Approaches . . . . . . . . . . . . . . . . . . . . . 26

2.2.4 Semisynthetic Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2.4.1 Area Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2.4.2 Full-Angle Method . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2.4.3 Vector-Based Method . . . . . . . . . . . . . . . . . . . . . . . 31
2.2.4.4 Mass-Point Method . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2.5 Provers Implementations and Repositories of Theorems . . . . . . . . . . 33
2.3 Interactive Theorem Proving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.1 Formalization of Foundations of Geometry . . . . . . . . . . . . . . . . . 34
2.3.1.1 Hilbert’s Geometry . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.1.2 Tarski’s Geometry . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.3.1.3 Axiom Systems and Continuity Properties . . . . . . . . . . . . 38
2.3.1.4 Other Axiom Systems and Geometries . . . . . . . . . . . . . . 40
2.3.1.5 Meta-Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.2 Higher Level Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.3 Other Formalizations Related to Geometry . . . . . . . . . . . . . . . . . 42
2.3.4 Verified Automated Reasoning . . . . . . . . . . . . . . . . . . . . . . . . 44

3 Coordinate-Free Theorem Proving in Incidence Geometry 61

Jürgen Richter-Gebert and Hongbo Li
3.1 Incidence Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.1.1 Incidence Geometry in the Plane . . . . . . . . . . . . . . . . . . . . . . . 62
3.1.2 Other Primitive Operations . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.1.3 Projective Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.2 Bracket Algebra: Straightening, Division, and Final Polynomials . . . . . . . . . . 67
3.2.1 Bracket Algebra and Straightening . . . . . . . . . . . . . . . . . . . . . . 67
3.2.2 Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.2.3 Final Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.3 Cayley Expansion and Factorization . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.3.1 Cayley Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.3.2 Cayley Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.3.3 Cayley Expansion and Factorization in Geometric Theorem Proving . . . . 76
3.3.4 Rational Invariants and Antisymmetrization . . . . . . . . . . . . . . . . . 77
3.4 Bracket Algebra for Euclidean Geometry . . . . . . . . . . . . . . . . . . . . . . 79
3.4.1 The Points I and J . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.4.2 Proving Euclidean Theorems . . . . . . . . . . . . . . . . . . . . . . . . . 80

4 Special Positions of Frameworks and the Grassmann-Cayley Algebra 85

Jessica Sidman and William Traves
4.1 Introduction: the Grassmann-Cayley Algebra and Frameworks . . . . . . . . . . . 85
4.2 Projective Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.2.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.2.2 Homogeneous Coordinates and Points at Infinity . . . . . . . . . . . . . . 88
4.2.3 Equations on Projective Space . . . . . . . . . . . . . . . . . . . . . . . . 88
4.2.4 Duality Between Lines and Points in P2 . . . . . . . . . . . . . . . . . . . 89
4.2.5 Grassmannians and Plücker Coordinates . . . . . . . . . . . . . . . . . . . 89
4.2.6 More About Lines in 3-space . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.3 The Bracket Algebra and Rings of Invariants . . . . . . . . . . . . . . . . . . . . 91
4.3.1 Group Actions and Invariant Polynomials . . . . . . . . . . . . . . . . . . 93
4.3.2 Relations Among the Brackets . . . . . . . . . . . . . . . . . . . . . . . . 94
4.4 The Grassmann-Cayley Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Contents xi

4.4.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.4.2 The cross product as a Join . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.4.3 Properties of the Exterior Product . . . . . . . . . . . . . . . . . . . . . . 99
4.4.4 Brackets and the Grassmann-Cayley algebra . . . . . . . . . . . . . . . . . 99
4.4.5 The Join and Meet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.5 Cayley Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.5.2 The Pure Condition as a Bracket Monomial . . . . . . . . . . . . . . . . . 101
4.5.3 White’s Algorithm for Multilinear Grassmann-Cayley
Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

5 From Molecular Distance Geometry to Conformal Geometric Algebra 107

Timothy F. Havel and Hongbo Li
5.1 Euclidean Distance Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.2 The Distance Geometry Theory of Molecular Conformation . . . . . . . . . . . . 110
5.3 Inductive Geometric Reasoning by Random Sampling . . . . . . . . . . . . . . . 115
5.4 From Distances to Advanced Euclidean Invariants . . . . . . . . . . . . . . . . . 122
5.5 Geometric Reasoning in Euclidean Conformal Geometry . . . . . . . . . . . . . . 129

6 Tree-Decomposable and Underconstrained Geometric Constraint

Problems 139
Ioannis Fudos, Christoph M. Hoffmann, and Robert Joan-Arinyo
6.1 Introduction, Concepts, and Scope . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.1.1 Geometric Constraint Systems (GCS) . . . . . . . . . . . . . . . . . . . . 141
6.1.2 Constraint Graph, Deficit, and Generic Solvability . . . . . . . . . . . . . 142
6.1.3 Instance Solvability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.1.4 Root Identification and Valid Parameter Ranges . . . . . . . . . . . . . . . 144
6.1.5 Variational and Serializable Constraint Problems . . . . . . . . . . . . . . 145
6.1.6 Triangle-Decomposing Solvers . . . . . . . . . . . . . . . . . . . . . . . . 145
6.1.7 Scope and Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.2 Geometric Constraint Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.3 Constraint Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.3.1 Geometric Elements and Degrees of Freedom . . . . . . . . . . . . . . . . 149
6.3.2 Geometric Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.3.3 Compound Geometric Elements . . . . . . . . . . . . . . . . . . . . . . . 150
6.3.4 Serializable Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.3.5 Variational Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.3.6 Triangle Decomposability . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.3.7 Generic Solvability and the Church-Rosser Property . . . . . . . . . . . . 155
6.3.8 2D and 3D Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.4 Solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
6.4.1 2D Triangle-Decomposable Constraint Problems . . . . . . . . . . . . . . 159
6.4.2 Root Identification and Order Type . . . . . . . . . . . . . . . . . . . . . 160
6.4.3 Extended Geometric Vocabulary . . . . . . . . . . . . . . . . . . . . . . . 166
6.5 Spatial Geometric Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
6.6 Under-Constrained Geometric Constraint Problems . . . . . . . . . . . . . . . . . 171
xii Contents

7 Geometric Constraint Decomposition: The General Case 181

Troy Baker, Meera Sitharam, and Rahul Prabhu
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
7.1.1 Terminology and Basic Concepts . . . . . . . . . . . . . . . . . . . . . . 182
7.1.2 Triangle-Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
7.1.3 Dulmage-Mendelsohn Decomposition . . . . . . . . . . . . . . . . . . . . 185
7.1.4 Assur Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
7.1.5 The Frontier Vertex Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 187
7.1.6 Canonical Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . 188
7.2 Efficient Realization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
7.2.1 Numerical Instability of Rigid Body Incidence and Seam Matroid . . . . . 191
7.2.2 Optimal Parameterization in Recombination . . . . . . . . . . . . . . . . . 192
7.2.3 Reconciling Conflicting Combinatorial Preprocessors . . . . . . . . . . . . 194
7.3 Handling Under- and Over-Constrained Systems . . . . . . . . . . . . . . . . . . 195
7.4 User Intervention in DR-Planning . . . . . . . . . . . . . . . . . . . . . . . . . . 195
7.4.1 Root Selection and Navigation . . . . . . . . . . . . . . . . . . . . . . . . 195
7.4.2 Changing Constraint Parameters . . . . . . . . . . . . . . . . . . . . . . . 196
7.4.3 Dynamic Maintenance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
7.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

II Distance Geometry, Real Algebraic Geometry, and

Configuration Spaces 199
8 Dimensional and Universal Rigidities of Bar Frameworks 201
A. Y. Alfakih
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
8.2 Stress Matrices and Gale Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 202
8.3 Dimensional Rigidity Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
8.4 Universal Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
8.4.1 Affine Motions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
8.4.2 Universal Rigidity Basic Results . . . . . . . . . . . . . . . . . . . . . . . 207
8.4.3 Universal Rigidity for Special Graphs . . . . . . . . . . . . . . . . . . . . 208
8.5 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

9 Computations of Metric/Cut Polyhedra and Their Relatives 213

Mathieu Dutour Sikirić, Michel-Marie Deza, and Elena I. Deza
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
9.2 Metric and Cut Cones and Polytopes . . . . . . . . . . . . . . . . . . . . . . . . . 215
9.3 Hypermetric Cone and Hypermetric Polytope . . . . . . . . . . . . . . . . . . . . 216
9.4 Cut and Metric Polytopes of Graphs . . . . . . . . . . . . . . . . . . . . . . . . . 218
9.5 Quasi-Semimetric Polyhedra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
9.6 Partial Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
9.7 Supermetric and Hemimetric Cones . . . . . . . . . . . . . . . . . . . . . . . . . 225
9.8 Software Computations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228

10 Cayley Configuration Spaces 233

Meera Sitharam, Menghan Wang, Joel Willoughby, and Rahul Prabhu
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
10.1.1 Euclidean Distance Cone . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
10.2 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
10.3 Related Chapters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Contents xiii

10.4 Characterizing 2D Graphs with Convex Cayley Configuration Spaces . . . . . . . 236

10.5 Extension to other Norms, Higher Dimensions, and Flattenability . . . . . . . . . 238
10.5.1 Computing Bounds of Convex Cayley Configuration Spaces in 3D for Par-
tial 3-Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
10.5.2 Some Background on the Distance Cone . . . . . . . . . . . . . . . . . . . 240
10.5.3 Genericity and Independence in the Context of Flattenability . . . . . . . . 242
10.6 Efficient Realization through Optimal Cayley Modification . . . . . . . . . . . . . 245
10.7 Cayley Configuration Spaces of 1-Dof Tree-Decomposable Linkages in 2D . . . . 246
10.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250

11 Constraint Varieties in Mechanism Science 253

Hans-Peter Schröcker, Martin Pfurner, and Josef Schadlbauer
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
11.1.1 Linkages and Joints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
11.1.2 Base and End-Effector Frame . . . . . . . . . . . . . . . . . . . . . . . . 255
11.2 Mechanisms and Algebraic Varieties . . . . . . . . . . . . . . . . . . . . . . . . . 255
11.2.1 Geometric Constraint Equations . . . . . . . . . . . . . . . . . . . . . . . 256
11.2.2 Study Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
11.2.3 Dual Quaternions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
11.2.4 Analyzing Mechanisms via Algebraic Varieties . . . . . . . . . . . . . . . 260
11.3 Serial Manipulators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
11.3.1 Direct and Inverse Kinematics . . . . . . . . . . . . . . . . . . . . . . . . 264
11.3.2 Synthesis of Open and Closed Serial Chains . . . . . . . . . . . . . . . . . 264
11.3.3 Singularities and Path Planning . . . . . . . . . . . . . . . . . . . . . . . 266
11.3.4 Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
11.4 Parallel Manipulators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
11.4.1 Direct and Inverse Kinematics . . . . . . . . . . . . . . . . . . . . . . . . 267
11.4.2 Singularities and Self-Motions . . . . . . . . . . . . . . . . . . . . . . . . 268
11.4.3 Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269

12 Real Algebraic Geometry for Geometric Constraints 273

Frank Sottile
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
12.2 Ideals and Varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
12.3 ... and Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
12.4 Structure of Algebraic Varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
12.4.1 Zariski Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
12.4.2 Smooth and Singular Points . . . . . . . . . . . . . . . . . . . . . . . . . 279
12.4.3 Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
12.5 Real Algebraic Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
12.5.1 Algebraic Relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
12.5.2 Semi-Algebraic Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
12.5.3 Certificates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283

III Geometric Rigidty 287

13 Polyhedra in 3-Space 289
Brigitte Servatius
13.1 Euler’s Conjecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
13.2 Cauchy’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
13.3 Co-Dimension 2 Results – Bricard Octahedra . . . . . . . . . . . . . . . . . . . . 290
13.4 Polyhedral Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
xiv Contents

14 Tensegrity 299
Robert Connelly and Anthony Nixon
14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
14.2 Tensegrity Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
14.2.1 Combinatorics of Tensegrities . . . . . . . . . . . . . . . . . . . . . . . . 304
14.2.2 Geometric Interpretations . . . . . . . . . . . . . . . . . . . . . . . . . . 304
14.2.3 Packings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
14.3 Types of Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
14.3.1 Global Rigidity and Stress Matrices . . . . . . . . . . . . . . . . . . . . . 306
14.3.2 Universal and Dimensional Rigidity . . . . . . . . . . . . . . . . . . . . . 306
14.3.3 Operations on Tensegrities . . . . . . . . . . . . . . . . . . . . . . . . . . 308
14.4 Examples and Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
14.4.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
14.4.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311

15 Geometric Conditions of Rigidity in Nongeneric Settings 317

Oleg Karpenkov
15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
15.2 Configuration Space of Tensegrities and its Stratification . . . . . . . . . . . . . . 319
15.2.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
15.2.2 Definition of a Tensegrity . . . . . . . . . . . . . . . . . . . . . . . . . . 320
15.2.3 Stratification of the Space of Tensegrities . . . . . . . . . . . . . . . . . . 321
15.2.4 Tensegrities on 4 Points in the Plane . . . . . . . . . . . . . . . . . . . . . 321
15.3 Extended Cayley Algebra and the Corresponding Geometric
Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
15.3.1 Extended Cayley Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . 322
15.3.2 Geometric Relations on Configuration Spaces of Points and Lines . . . . . 324
15.4 Geometric Conditions of Infinitesimal Flexibility in Terms of
Extended Cayley Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
15.4.1 Examples in the Plane . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
15.4.2 Frameworks in General Position . . . . . . . . . . . . . . . . . . . . . . . 326
15.4.3 Non-parallelizable Tensegrities . . . . . . . . . . . . . . . . . . . . . . . . 326
15.4.4 Geometric Conditions for Existence Non-parallelizable Tensegrities . . . . 327
15.4.5 Conjecture on Strong Geometric Conditions for Tensegrities . . . . . . . . 327
15.5 Surgeries on Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
15.6 Algorithm to Write Geometric Conditions of Realizability of Generic Tensegrities 330
15.6.1 Framed Cycles in General Gosition . . . . . . . . . . . . . . . . . . . . . 330
15.6.1.1 Basic Definitions . . . . . . . . . . . . . . . . . . . . . . . . . 330
15.6.1.2 Geometric Conditions for Framed Cycles . . . . . . . . . . . . . 331
15.6.1.3 Geometric Conditions for Trivalent Graphs . . . . . . . . . . . . 332
15.6.2 Resolution Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
15.6.2.1 Definition of Resolution Schemes . . . . . . . . . . . . . . . . . 332
15.6.2.2 Resolution of a Framework . . . . . . . . . . . . . . . . . . . . 333
15.6.2.3 HΦ-Surgeries on Completely Generic Resolution Schemes . . . 333
15.6.3 Construction of Framing for Pairs of Leaves in Completely Generic Reso-
lution Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
15.6.4 Framed Cycles Associated to Generic Resolutions of a Graph . . . . . . . 335
15.6.5 Natural Correspondences Between ΞG (P) and the Set of all Resolutions for
G(P) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
15.6.6 Techniques to Construct Geometric Conditions Defining Tensegrities . . . 336
Contents xv

16 Generic Global Rigidity in General Dimension 341

Steven J. Gortler
16.1 Basic Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
16.2 Connelly’s Sufficiency Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
16.3 Hendrickson’s Necessary Conditions . . . . . . . . . . . . . . . . . . . . . . . . 345
16.3.1 Nonsufficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
16.4 Necessity of Connelly’s Condition . . . . . . . . . . . . . . . . . . . . . . . . . . 347
16.5 Randomized Algorithm for Testing Generic Global Rigidity . . . . . . . . . . . . 347
16.6 Surgery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
16.7 Other Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349

17 Change of Metrics in Rigidity Theory 351

Anthony Nixon and Walter Whiteley
17.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
17.2 Projective Transfer of Infinitesimal Rigidity . . . . . . . . . . . . . . . . . . . . . 352
17.2.1 Coning and Spherical Frameworks . . . . . . . . . . . . . . . . . . . . . . 353
17.2.2 Rigidity Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
17.2.2.1 Spherical to Affine Transfer . . . . . . . . . . . . . . . . . . . . 357
17.2.3 Equilibrium Stresses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
17.2.4 Point-Hyperplane Frameworks . . . . . . . . . . . . . . . . . . . . . . . . 359
17.2.5 Tensegrity Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
17.3 Projective Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
17.4 Pseudo-Euclidean Geometries . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
17.4.1 Hyperbolic and Minkowski Spaces . . . . . . . . . . . . . . . . . . . . . . 364
17.5 Transfer of Symmetric Infinitesimal Rigidity . . . . . . . . . . . . . . . . . . . . 364
17.5.1 Symmetric Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
17.6 Global Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
17.6.1 Universal Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
17.6.2 Projective Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . 369
17.6.3 Pseudo-Euclidean Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . 370
17.7 Summary and Related Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370

IV Combinatorial Rigidity 375

18 Planar Rigidity 377
Brigitte Servatius and Herman Servatius
18.1 Rigidity of Bar and Joint Frameworks . . . . . . . . . . . . . . . . . . . . . . . . 378
18.1.1 Rigidity Matrix and Augmented Rigidity Matrices . . . . . . . . . . . . . 380
18.1.2 Rigidity Matrix as a Transformation . . . . . . . . . . . . . . . . . . . . . 382
18.1.3 The Infinitesimal Rigidity Matroid of a Framework . . . . . . . . . . . . . 384
18.2 Abstract Rigidity Matroids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
18.2.1 Characterizations of A2 and (A2 )⊥ . . . . . . . . . . . . . . . . . . . . . 387
18.2.2 The 2-Dimensional Generic Rigidity Matroid . . . . . . . . . . . . . . . . 389
18.2.3 Cycles in G2 (n) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
18.2.4 Rigid Components of G2 (G) . . . . . . . . . . . . . . . . . . . . . . . . . 391
18.2.5 Representability of G2 (n) . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
18.3 Rigidity and Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
18.3.1 Birigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
18.3.2 Tree Decomposition Theorems . . . . . . . . . . . . . . . . . . . . . . . . 397
18.3.2.1 Computation of Independence in G2 (n) . . . . . . . . . . . . . . 398
18.3.3 Pinned Frameworks and Assur Decomposition . . . . . . . . . . . . . . . 400
xvi Contents

18.3.3.1 Isostatic Pinned Framework . . . . . . . . . . . . . . . . . . . . 400

18.3.4 Body and Pin Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
18.3.5 Rigidity of Random Graphs . . . . . . . . . . . . . . . . . . . . . . . . . 407

19 Inductive Constructions for Combinatorial Local and Global Rigidity 413

Anthony Nixon and Elissa Ross
19.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
19.2 Rigidity in Rd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
19.2.1 Inductive Operations on Frameworks . . . . . . . . . . . . . . . . . . . . 416
19.2.2 Recursive Characterizations of Graphs . . . . . . . . . . . . . . . . . . . . 418
19.2.3 Combinatorial Characterizations of Rigidity . . . . . . . . . . . . . . . . . 421
19.3 Body-Bar, Body-Hinge, Molecular, etc. . . . . . . . . . . . . . . . . . . . . . . . 422
19.3.1 Geometry and Combinatorics . . . . . . . . . . . . . . . . . . . . . . . . 423
19.3.2 Characterizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
19.4 Further Rigidity Contexts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
19.4.1 Frameworks with Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . 425
19.4.2 Infinite Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
19.4.3 Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
19.4.4 Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
19.4.5 Applications of Rigidity Techniques . . . . . . . . . . . . . . . . . . . . . 428
19.4.6 Direction-Length Frameworks and CAD . . . . . . . . . . . . . . . . . . . 428
19.4.7 Nearly Generic Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . 428
19.5 Summary Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429

20 Rigidity of Body-Bar-Hinge Frameworks 435

Csaba Király and Shin-ichi Tanigawa
20.1 Rigidity of Body-Bar-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . . 436
20.1.1 Body-Bar Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
20.1.2 Body-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
20.1.3 Body-Bar-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . . . . 442
20.2 Generic Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
20.2.1 Body-Bar Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
20.2.2 Body-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
20.3 Other Related Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
20.3.1 Plate-Bar Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
20.3.2 Identified Body-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . 445
20.3.3 Panel-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
20.3.4 Molecular Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
20.3.5 Body-Pin Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
20.3.6 Body-Bar Frameworks with Boundaries . . . . . . . . . . . . . . . . . . . 449
20.3.7 Other Variants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
20.4 Generic Global Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
20.4.1 Body-Bar Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
20.4.2 Body-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
20.4.3 Counterexamples to Hendrickson’s Conjecture . . . . . . . . . . . . . . . 452
20.5 Graph Theoretical Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
20.5.1 Tree Packing and Connectivity . . . . . . . . . . . . . . . . . . . . . . . . 453
20.5.2 Brick Partitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
20.5.3 Constructive Characterizations . . . . . . . . . . . . . . . . . . . . . . . . 455
20.5.4 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
Contents xvii

21 Global Rigidity of Two-Dimensional Frameworks 461

Bill Jackson, Tibor Jordán, and Shin-Ichi Tanigawa
21.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
21.2 Conditions for Global Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
21.2.1 Stress Matrix Characterization in Rd . . . . . . . . . . . . . . . . . . . . . 463
21.2.2 Hendrickson’s Necessary Conditions for Global Rigidity . . . . . . . . . . 464
21.3 Graph Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
21.4 Characterization of Global Rigidity in R1 and R2 . . . . . . . . . . . . . . . . . . 466
21.5 The Rigidity Matroid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
21.5.1 Rd -Independent Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
21.5.2 Rd -Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
21.5.3 Rd -Connected Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
21.6 Special Families of Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
21.6.1 Highly Connected Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 470
21.6.2 Vertex-Redundantly Rigid Graphs . . . . . . . . . . . . . . . . . . . . . . 471
21.6.3 Vertex Transitive Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
21.6.4 Graphs of Large Minimum Degree . . . . . . . . . . . . . . . . . . . . . . 472
21.6.5 Random Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
21.6.6 Unit Disk Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
21.6.7 Squares of Gaphs, Line Graphs, and Zeolites . . . . . . . . . . . . . . . . 473
21.7 Related Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474
21.7.1 Globally Linked Pairs of Vertices . . . . . . . . . . . . . . . . . . . . . . 474
21.7.2 Globally Rigid Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
21.7.3 Globally Loose Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
21.7.4 Uniquely Localizable Vertices . . . . . . . . . . . . . . . . . . . . . . . . 477
21.7.5 The Number of Non-Equivalent Realizations . . . . . . . . . . . . . . . . 477
21.7.6 Stability Lemma and Neighborhood Results . . . . . . . . . . . . . . . . . 478
21.8 Direction Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479
21.8.1 Parallel Drawings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 480
21.8.2 Direction-Length Global Rigidity . . . . . . . . . . . . . . . . . . . . . . 481
21.9 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482
21.10Optimization Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482

22 Point-Line Frameworks 487

Bill Jackson and J.C. Owen
22.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
22.1.1 Motivation from CAD . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
22.1.2 Motivation from Automated Deduction in Geometry (ADG) and Theorem
Proving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
22.1.3 Constraint Graphs and Frameworks . . . . . . . . . . . . . . . . . . . . . 488
22.2 Point-Line Graphs and Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . 490
22.2.1 Point-Line Frameworks and the Rigidity Map . . . . . . . . . . . . . . . . 491
22.2.2 The Rigidity Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492
22.2.3 The Rigidity Matroid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 494
22.2.4 Affine Properties of the Point-Line Rigidity Matrix . . . . . . . . . . . . . 494
22.2.5 Fixed-Slope Point-Line Frameworks . . . . . . . . . . . . . . . . . . . . . 495
22.3 Characterization of the Generic Rigidity Matroid for Point-Line Frameworks in R2 496
22.3.1 A Count Matroid for Point-Line Graphs . . . . . . . . . . . . . . . . . . . 496
22.3.2 A Characterization of Independence in MPL (G) when G is Naturally Bi-
partite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
22.3.3 A Characterization of Independence in MPL (G) . . . . . . . . . . . . . . 501
xviii Contents

22.3.4 The Rank Function for MPL (G) . . . . . . . . . . . . . . . . . . . . . . . 501

22.4 Extensions to Rd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
22.4.1 Point-Line Frameworks in Rd . . . . . . . . . . . . . . . . . . . . . . . . 502
22.4.2 Point-Hyperplane Frameworks in Rd . . . . . . . . . . . . . . . . . . . . 502
22.5 Direction-Length Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503

23 Generic Rigidity of Body-and-Cad Frameworks 505

Audrey St. John
23.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
23.2 Algebraic Body-and-Cad Rigidity Theory . . . . . . . . . . . . . . . . . . . . . . 506
23.2.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506
23.2.2 Getting to Know Body-and-Cad Frameworks . . . . . . . . . . . . . . . . 507
23.2.3 Formalization of the Algebraic Setting . . . . . . . . . . . . . . . . . . . . 509
23.2.4 Building a 3D Body-and-Cad Framework . . . . . . . . . . . . . . . . . . 512
23.3 Infinitesimal Body-and-Cad Rigidity Theory . . . . . . . . . . . . . . . . . . . . 513
23.3.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
23.3.2 The Pattern of the Rigidity Matrix . . . . . . . . . . . . . . . . . . . . . . 515
23.3.2.1 Primitive Angular and Blind Constraints . . . . . . . . . . . . . 515
23.3.3 Generic Rigidty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
23.4 Combinatorial Body-and-Cad Rigidity Theory . . . . . . . . . . . . . . . . . . . 517
23.4.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
23.4.2 The Rigidity Matroid and Sparsity . . . . . . . . . . . . . . . . . . . . . . 518
23.4.3 Characterizing Generic Body-and-Cad Rigidity . . . . . . . . . . . . . . . 518
23.4.4 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519
23.5 Open Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522

24 Rigidity with Polyhedral Norms 525

Derek Kitson
24.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
24.1.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
24.1.2 Statement of the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
24.2 Rigidity of Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
24.2.1 Points of Differentiability . . . . . . . . . . . . . . . . . . . . . . . . . . 527
24.2.2 The Rigidity Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528
24.2.3 Framework Colors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529
24.2.4 Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530
24.2.5 Path Chasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530
24.2.6 Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531
24.3 Rigidity of Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535
24.3.1 Sparsity Counts and Tree Decompositions . . . . . . . . . . . . . . . . . . 536
24.3.2 Regular Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
24.3.3 Symmetric Isostatic Placements . . . . . . . . . . . . . . . . . . . . . . . 537
24.3.4 Symmetric Tree Decompositions . . . . . . . . . . . . . . . . . . . . . . . 539

25 Combinatorial Rigidity of Symmetric and Periodic Frameworks 543

Bernd Schulze
25.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 543
25.2 Incidentally Symmetric Isostatic Frameworks . . . . . . . . . . . . . . . . . . . . 544
25.2.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 544
25.2.2 Symmetry-Adapted Maxwell Counts . . . . . . . . . . . . . . . . . . . . 545
25.2.3 Characterizations of Symmetric Isostatic Graphs . . . . . . . . . . . . . . 548
Contents xix

25.3 Forced-Symmetric Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . 549

25.3.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 549
25.3.2 Symmetric Motions and the Orbit Rigidity Matrix . . . . . . . . . . . . . 551
25.3.3 Characterizations of Forced-Symmetric Rigid Graphs . . . . . . . . . . . . 553
25.4 Incidentally Symmetric Infinitesimally Rigid
Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554
25.4.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554
25.4.2 Phase-Symmetric Orbit Rigidity Matrices . . . . . . . . . . . . . . . . . . 556
25.4.3 Characterizations of Symmetric Infinitesimally Rigid Graphs . . . . . . . . 557
25.5 Periodic Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559
25.5.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559
25.5.2 Maxwell Counts for Periodic Rigidity . . . . . . . . . . . . . . . . . . . . 560
25.5.3 Characterizations of Periodic Rigid Graphs . . . . . . . . . . . . . . . . . 561

Index 567
Foreword

Geometric constraint systems arise in a diverse range of applications including: computer aided
engineering and architectural design, molecular and materials modeling, robotics and animation,
sensor networks, machine learning, and dimension reduction. Broadly, a geometric constraint sys-
tem (GCS) is defined on a set of geometric primitives (e.g., points, lines, rigid bodies) by specifying
geometric relationships (such as distances, angles, or incidences). The core GCS foundations come
from at least four interwoven topic areas and research communities: (i) combinatorial and geomet-
ric rigidity, (ii) automated geometric theorem proving, (iii) geometrically constrained configuration
spaces and, (iv) distance geometry. Indeed, the principles, tools, and techniques rely on invariant
theory, combinatorial and discrete geometry, algebraic geometry and topology, convex/semidefinite
analysis, with algorithmic foundations and complexity going back to Cauchy, Cayley, Hilbert, Klein,
and Maxwell.
With such a rich array of communities working on GCS research, this handbook is intended as
an entry point to the principal mathematical and computational tools, techniques and results cur-
rently in use. It was born out of continued requests for a single source containing the core principles
and results that would be accessible to beginners and experts alike (from the graduate student start-
ing research to the algebraic geometer interested in applications to the roboticist seeking to engineer
a swarm of autonomous agents). Recognizing that readers may come from a wide variety of back-
grounds, we hope that this book will be a useful tool for navigating the concepts, approaches, and
results found in GCS research.
We are grateful to the authors of the chapters that follow; their expertise provides the roadmap
for developing a unified view of the varied perspectives. We pledge any royalties toward supporting
the activities of the four research communities represented by the four parts of this handbook, es-
pecially the activities of young researchers. We would like to thank Louis Theran for his feedback.
We thank Rahul Prabhu for his kind and timely expert help with LATEX. And, finally, we cannot put
into words the debt of gratitude owed to our families for their unconditional support and patience
during this process.

Meera Sitharam
Audrey St. John
Jessica Sidman

xxi
Preface

The goal of this book is to provide a resource for those aiming to become acquainted with the
fundamentals as well as experts looking to pinpoint specific results or approaches in the broad
landscape. The flow of the handbook is intended to take readers from the general algebro-geometric
approaches to more specialized contexts permitting combinatorial analysis and efficient algorithms.
Chapters are grouped by the main techniques being deployed, in the hopes that readers can find
the material best-suited to their expertise. Of course, the overlapping nature of the material being
presented prevents a neat partitioning of chapters by topic area, but we hope the juxtapositioning of
the chapters helps the reader to see how the subject is connected.
Chapter 1 provides an overview of the book as a more detailed starting point and is expected to
help the reader navigate the book effectively. It includes a basic introduction, some preliminaries,
and an overview of the various topics and methods. We also give an alternative pathway through the
book, intended to help a newcomer become acquainted with the domain. We hope this is a first step
toward a unifying foundation for the rich set of GCS problems.

xxiii
Contributors

Abdo Y. Alfakih Queen Mary University of London

University of Windsor London, England
Windsor Ontario, Canada
Predrag Janičić
Troy Baker University of Belgrade
University of Florida Belgrade, Serbia
Gainesville, Florida
Robert Joan-Arinyo
Robert Connelly Universitat Politècnica de Catalunya
Cornell University Barcelona, Catalonia
Ithaca, New York
Tibor Jordán
Michel-Marie Deza ELTE Eötvös Loránd University and
École Normale Supérieure MTA-ELTE Egerváry Research Group on
Paris, France Combinatorial Optimization
Budapest, Hungary
Elena I. Deza
Moscow State Pedagogical University Oleg Karpenkov
Moscow, Russia University of Liverpool
Liverpool, United Kingdom
Mathieu Dutour Sikirić
Rudjer Bosković Institute Csaba Király
Zagreb, Croatia ELTE Eötvös Loránd University and
MTA-ELTE Egerváry Research Group on
Jacques Fleuriot Combinatorial Optimization
University of Edinburgh Budapest, Hungary
Scotland, United Kingdom
Derek Kitson
Ioannis Fudos Lancaster University
University of Ioannina Lancaster, United Kingdom
Ioannina, Greece
Hongbo Li
Steven J. Gortler Chinese Academy of Sciences
Harvard University Beijing, China
Cambridge, Massachusetts
Julien Narboux
Timothy F. Havel University of Strasbourg
Energy Compression Inc. Strasbourg, France
Boston, Massachusetts
Anthony Nixon
Christoph M. Hoffmann Lancaster University
Purdue University Lancaster, United Kingdom
West Lafayette, Indiana
John C. Owen
Bill Jackson Siemens

xxv
xxvi Handbook of Geometric Constraint Systems Principles

Cambridge, England Jessica Sidman

Mount Holyoke College
Martin Pfurner South Hadley, Massachusetts
University of Innsbruck
Innsbruck, Austria Meera Sitharam
University of Florida
Rahul Prabhu
Gainesville, Florida
University of Florida
Gainesville, Florida
Frank Sottile
Jürgen Richter-Gebert Texas A&M University
Technical University of Munich College Station, Texas
Munich, Germany
Audrey St. John
Elissa Ross Mount Holyoke College
MESH Consultants Inc., Fields Institute South Hadley, Massachusetts
Toronto, Canada
Shin-ichi Tanigawa
Josef Schadlbauer The University of Tokyo
University of Innsbruck Tokyo, Japan
Innsbruck, Austria
William Traves
Hans-Peter Schröcker US Naval Academy
University of Innsbruck Annapolis, Maryland
Innsbruck, Austria
Menghan Wang
Bernd Schulze
University of Florida
Lancaster University
Gainesville, Florida
Lancaster, United Kingdom

Brigitte Servatius Walter Whiteley

Worcester Polytechnic Institute York University
Worcester, Massachusetts Toronto, Canada

Herman Servatius Joel Willoughby

Worcester Polytechnic Institute University of Florida
Worcester, Massachusetts Gainesville, Florida
Chapter 1
Overview and Preliminaries
Meera Sitharam
University of Florida

Troy Baker
University of Florida

CONTENTS
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Specifying a GCS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.2 Fundamental GCS Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.3 Tractability and Computational Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Parts and Chapters of the Handbook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Part I: Geometric Reasoning Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.2 Part II: Distance Geometry, Configuration Space, and Real Algebraic
Geometry Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.3 Part III: Geometric Rigidity Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.4 Part IV: Combinatorial Rigidity Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.4.1 Inductive Constructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.4.2 Body Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.4.3 Body-Cad, and Point-Line Frameworks . . . . . . . . . . . . . . . . . . . . . . 9
1.2.4.4 Symmetric and Periodic Frameworks and Frameworks under
Polyhedral Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.5 Missing Topics and Chapters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Terminology Reconciliation and Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.1 Constrainedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.2 Rigidity of Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.3 Generic Rigidity of Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3.4 Approximate Degree-of-Freedom and Sparsity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4 Alternative Pathway through the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

In this chapter, we begin with a generalized introduction to geometric constraint systems before
giving an overview of the book’s contents. We conclude with a section reconciling terminology and
concepts that have arisen in different communities and as well as an alternative pathway through the
book especially for the novice reader.

1
2 Handbook of Geometric Constraint Systems Principles

1.1 Introduction
A geometric constraint system is generally defined on a finite collection of geometric primitives
(e.g., 0-dimensional points, 1-dimensional lines, general d-dimensional hyperplanes, d-dimensional
rigid bodies, conics, cylinders). The setting is a given Euclidean or non-Euclidean geometry over the
reals, generally of fixed dimension, and constraints specify geometric n-ary relationships among the
primitives. These constraints can be logical (e.g., incidence, perpendicularity, tangency) or metric
(e.g., distance, angle, orientation) and may be either equalities or inequalities. Typically, a con-
straint can be expressed as a set of quadratic polynomials with real (often rational or even integer)
coefficients. The combinatorics of a GCS are usually captured separately in a constraint graph: a
(hyper)graph where each vertex represents a geometric primitive and each (hyper)edge represents a
constraint on the corresponding primitives.
A realization (or solution) of a GCS is a placement (or configuration) of the geometric primitives
that satisfies the constraints. The realizations of a GCS can be found algebraically by solving a
system of polynomial equations corresponding to the GCS, where the variables are the coordinates
of the geometric primitives. Thus, the set of realizations of such a system consists of the solutions
to a finite collection of polynomial equations and is hence a variety. Typically, our primary interest
is in the real points of this variety, but if we consider solutions over C, then the full power of
algebro-geometric methods may be brought to bear.
In the geometric setting it is generally implied (as it is implied throughout this chapter, unless
explicitly stated otherwise) that we are concerned with the realization space modulo some group of
trivial motions that is designated a priori. For example, in Euclidean space the trivial motions are
comprised of translation and
rotation; in d-dimensional Euclidean space, there
is a d-dimensional
space of translations and d2 -dimensional space of rotations giving a d+1 2 -dimensional space of
trivial motions. Embedding Rd into projective space can help us to see unifying principles in in-
cidence and other constraint systems. Sometimes, the realization is pinned (or grounded), i.e., the
trivial group is chosen to be the empty group.

1.1.1 Specifying a GCS

To illustrate these core concepts, consider specifying the most common GCS for classical rigid-
ity theory: the Euclidean bar-and-joint, or Euclidean distance constraint system. The geometric
primitives are 0-dimensional points (called “joints”), the constraints are specified distances between
points (called “bars”) and the ambient space is Rd . G = (V, E) associates a vertex to each joint and
an edge (u, v) to each bar constraining the joints represented by vertices u and v. Then a bar-and-
joint constraint system of G can be defined as a tuple (G, δ ) where δ : E → R assigns distance
values to the bars. A bar-and-joint constraint system is also called a linkage. A configuration of the
joints in Rd is given by a map p : V → Rd and is a realization of (G, δ ) if the distance between
p(u) and p(v) is δ (u, v) for all (u, v) ∈ E. For example, let (G = ({1, 2}, {(1, 2)}), δ ) be a Euclidean
bar-and-joint system consisting of two joints with one bar between them specifying a distance of 4;
i.e., δ ((1, 2)) = 4. Then, if (x1 , y1 ) and (x2 , y2 ) are the variables for the coordinates of joints 1 and
2, respectively, realizations of this linkage are the solutions to the single constraint equation

(x1 − x2 )2 + (y1 − y2 )2 = δ 2 .

Notice that some geometric constraints may lead to multiple equations; if we were to place an
incidence constraint between two points, it would give the equations:

x1 = x2 and y1 = y2 .
 Overview and Preliminaries 3

Now consider specifying a GCS whose geometric primitives are rigid bodies in Euclidean space;
a “bar” constraint can be placed between two bodies by picking a point on each and constraining
the distance between them. Such a system describes a Euclidean body-and-bar constraint system.
Realizations are solutions to the quadratic system of distance equations, i.e., placements of the
bodies (e.g., by assigning elements of the special Euclidean group SE(d)) that satisfy the bar lengths.

1.1.2 Fundamental GCS Questions

Given a GCS C with n equality constraints, we seek approaches for finding realizations and/or
giving structural characterization of C based on properties of the resulting set S of geometrically
constrained configurations or realizations. That is, when the solution space S has co-dimension
n (independent C); is finite (locally rigid C); has the singleton property (globally rigid C). Other
properties of S such as dimension (degrees of freedom), connectedness, singularities (deformation
paths and extreme configurations) are also of interest. More generally, many of these properties
can be deduced by deriving dependent (often inconsistent) constraints that are locally or globally
implied by the given GCS, or by ascertaining its independence.

1.1.3 Tractability and Computational Complexity

In its full generality, the fundamental questions encompass the first order theory of the real closed
fields. This theory is complete, and automated theorem proving over the reals (RCF) is decidable
as shown by Tarski [36] (i.e., does not suffer from Gödel’s incompleteness of Peano’s first order
theory of natural numbers). That said, its algorithmic complexity is essentially that of polynomial
ideal membership, commonly using Gröbner bases or cylindrical algebraic decomposition [4, 2],
which is prohibitive, being complete for the class EXPSPACE [25]. Even the existential theory of
the reals is NP-hard, with the best-known algorithms requiring doubly exponential time [25].

1.2 Parts and Chapters of the Handbook

This handbook contains a sampling of a wide range of theories and methods that attempt to cir-
cumvent the above intractability by taking advantage of properties specific to various types of geo-
metric constraint systems. We have organized it into parts based on the main techniques underlying
each chapter, starting with the algebro-geometric techniques and concluding with combinatorial
approaches. This essentially provides a flow from approaches that address the general (generic and
non-generic) GCS setting to those that work under restricted (generic) settings. Indeed, assumptions
of genericity appear throughout Part IV, while Part I contains approaches to address non-generic sit-
uations.

Part I Geometric Reasoning Techniques: Chapters 2–7 address more general geometric constraint
systems, with many of the approaches based on algebraic methods.
Part II Distance Geometry, Configuration Space, and Real Algebraic Geometry Techniques:
Chapters 8–12 span the underlying topics of distance geometry, configuration spaces, and
(real) algebraic geometry.
Part III Geometric Rigidity Techniques: Chapters 13–17 (mostly based in Rigidity Theory), while
often restricted to generic assumptions, require geometric analyses.
 4 Handbook of Geometric Constraint Systems Principles

Part IV Combinatorial Rigidity Techniques: Chapters 18–25 conclude the book with the Rigidity
Theory settings that permit combinatorial approaches.

1.2.1 Part I: Geometric Reasoning Techniques

For the specific case of deriving dependent geometric constraints from a GCS, i.e., automated geom-
etry theorem proving, an algorithm that is significantly more efficient than either of the previously
mentioned (exponential time and polynomial space) approaches is the Wu-Ritt characteristic set or
triangle decomposition method [9, 35]. This and other techniques for automated geometric theorem
proving are discussed in Chapter 2.
Since we are restricting ourselves to geometric constraints, the relevant polynomials are typi-
cally invariants of transformation or trivial motion groups that define the underlying geometry (Eu-
clidean, Projective, etc.). Invariant polynomials permit synthetic, coordinate-free, and even metric-
free computational approaches to deriving dependent constraints, e.g., using bracket algebras. For
example, the Grassman-Cayley algebra [40] yields a synthetic computational approach to deriving
dependent constraints in projective and incidence geometry. In fact, such invariant theoretic methods
even extend to finite geometries [33, 32].
Chapter 5 introduces the bracket algebra and Grassmann-Cayley algebra for the plane with
a view toward proving theorems in Projective and Euclidean geometry. The bracket algebra and
Grassmann-Cayley algebra appear again in Chapter 4, this time in the context of nongeneric or
special realizations of a GCS. The goal is to provide a geometric meaning to the algebraic condi-
tion (so-called pure condition) that makes the realization special, and this involves the technique
of multilinear Cayley factorization. The chapter includes introductions to projective space, homo-
geneous coordinates, the Grassmannian, and Plücker coordinates and ends with examples applying
the theory to body-and-bar GCS.
As another example of a similar approach, Chapter 3 develops the theory and applications of
a variety of types of Euclidean invariants in deriving constraint dependences. The n2 pairwise

distance polynomials between n points in real Euclidean space are Euclidean invariants that are
related by the Cayley-Menger syzygies. These are described in the following classical theorem on
Euclidean distance matrices.
A Euclidean distance matrix (EDM) for Rd is an n × n square matrix of pairwise (squared)
distances between n points in Rd . It is denoted ∆[n] with distance entries δi j for 1 ≤ i, j ≤ n. For
S ⊆ [n], the submatrix ∆S has entries δi j for i, j ∈ S. The volume matrix ∆ˆ S is the |S| + 1 × |S| + 1
matrix obtained from ∆S by bordering ∆S with a top row (0, 1, . . . , 1) and a left column (0, 1, . . . , 1)T .
Now det(∆ˆ S ) computes the volume of the simplex with points in S, and is called a Cayley-Menger
determinant [34]. The next theorem effectively says that the volumes of simplices formed by d + 2
points in Rd is 0.

Theorem 1.1 (Cayley-Menger Relations) A real symmetric matrix ∆[n] with 0 diagonal and posi-
tive entries is a Euclidean distance matrix for Rd only if det(∆ˆ S ) = 0 for all S ⊆ [n] with |S| ≥ d + 2.

A more direct approach for both Euclidean distance or Projective constraint systems is to simply
solve for the set of realizations of the constraint system to determine (in)dependent constraints. For
example, some GCS permit a ruler and compass type construction of the (finite set of) solutions,
which is equivalent to solving a recursive triangular block decomposition of the constraint system.
A broad class of GCS occurring in computer aided design are of this type, i.e., their underlying
combinatorial structure, or constraint graph is triangle-decomposable or tree-decomposable [7].
It is known that such GCS are Quadratically Radically Solvable (QRS), i.e., the coordinates
of the solutions are in nested quadratic extensions of the coefficient field. As to whether all generic
Overview and Preliminaries 5

QRS systems have triangle-decomposable constraint graphs is still an open problem, with the equiv-
alence being shown only when the constraint graph is topologically planar [26]. Solving these sys-
tems entails two stages: the recursive triangle decomposition stage, a combinatorial procedure on
the constraint graph, and a solution or realization stage, that obtains the solution of the correspond-
ing recursively decomposed system through a bottom-up process of assembling or recombining the
(generically finitely many) solutions of subsystems.
Chapter 6 gives many natural examples of triangle-decomposable constraint systems and for-
mal algorithms for recursive decomposition and recombination. In general, the time complexity
of solving or realization is bottle-necked by the largest subsystem that must be solved simultane-
ously and a polynomial time preprocessing algorithm to identify the subsystems can be beneficial.
For non-triangle-decomposable GCS, the above method can be generalized to obtain recursive de-
compositions into subgraphs that approximate generically rigid subsystems that have finitely many
isolated solutions or realizations (see Section 1.3 for Terminology and Basic Concepts).
The process of identifying these subsystems and finding a partial order in which to solve them is
called decomposition-recombination (DR-) planning, the topic of Chapter 7. Such algorithms vary
in their generality but often leverage geometric properties of specific constraint and primitive types
or use a priori knowledge of patterns in their arrangement. However, in general, at some point an
algebraic system must be solved for recombination of the decomposed subsystem solutions. Chapter
7 surveys several such methods for decomposition and recombination of more general constraint
systems.
However, most of the above-mentioned approaches (including those that rely on generic rigidity
or finiteness of the solution set for decomposition) do not differentiate between real and complex
solution spaces. In particular to specialize the invariant-theoretic approach to the reals requires im-
posing additional inequality constraints beyond the Cayley-Menger conditions in the previous theo-
rem, by asserting that all simplices have positive volumes. For 1- and 2-dimensional simplices (line
segments and triangles), this gives exactly the metric condition on real Euclidean distances.

Theorem 1.2 (Cayley-Menger Inequalities) A real symmetric matrix with 0 diagonal and positive
entries ∆[n] is a Euclidean distance matrix for Rd if and only if det(∆ˆ S ) ≥ 0 for all S ⊆ [n], |S| ≥ 2;
and det(∆ˆ S ) = 0 for all S ⊆ [n] with |S| ≥ d + 2.

Note that the Euclidean invariant approaches described in Chapter 5 use all the Cayley-Menger
conditions including the above inequalities.

1.2.2 Part II: Distance Geometry, Configuration Space, and Real Algebraic Geom-
etry Techniques
The inequalitites in Theorem 1.2 can be viewed as partly arising from the metric property of real
Euclidean space. This leads to another tool for dealing with geometric constraints with an underlying
metric, namely distance or metric geometry. The classical theorem of Schoenberg [30, 31] (which
generalizes to infinite dimensional Hilbert spaces) is stated for finite dimensional real Euclidean
distance matrices below. It is equivalent to the conjunction of the two Cayley-Menger theorems
above.

Theorem 1.3 (Schoenberg’s Theorem) A real symmetric matrix with 0 diagonal and positive en-
tries ∆[n] is a Euclidean distance matrix for Rd if and only if it is negative semidefinite on the
subspace of all vectors orthogonal to the all 1’s vector and rank(∆[n] ) ≤ d + 1.

The convexity and face structure of the Euclidean distance cone yield powerful techniques for
understanding distance constraint systems, including implied or dependent constraints, and differ-
ent types of rigidity. Chapter 8 surveys some of these techniques. Chapter 12 introduces the tools
6 Handbook of Geometric Constraint Systems Principles

of real algebraic geometry, specifically semialgebraic sets that involve polynomial equalities and
inequalities (such as the Cayley-Menger determinantal equalities and inequalities above) and the
positivenstellensatz as tools for defining generic realizations and dealing with distance constraint
systems. It starts with a brief introduction to the correspondence between ideals and varieties over
C, and then turns to a discussion of varieties defined by polynomials with real coefficients, which
may be viewed as either real or complex varieties. It culminates with a view of the projection of
the d-dimensional stratum of the Euclidean (squared) distance cone as a semialgebraic set and its
application to rigidity.
Chapter 9 explores the structure of general metric cones. Chapter 10 employs properties of pro-
jections and fibers of rank d strata of the Euclidean distance cones to characterize distance constraint
systems (their underlying graphs) whose configuration spaces generically map finitely-many-to-one
to a convex set and whose singular configurations have a simple description. These characterizations
extend to when the distance constraints are inequalities and the distances are l pp norms. The tech-
niques yield interesting configuration space properties of a common class of plane linkage mecha-
nisms (Euclidean distance constraint systems with one degree of freedom in R2 ) arising from the
QRS or triangle decomposable constraint graphs mentioned above.
However, questions about linkage mechanisms in general are inherently difficult: Kempe’s uni-
versality theorem [20] states that the space of configurations can trace out any desired algebraic
curve. Chapter 11 explores constraint varieties of mechanisms and describes how so-called study
parameters and dual quaternions are used in kinematics.

1.2.3 Part III: Geometric Rigidity Techniques

The question of local and global uniqueness of polyhedra whose faces have a given combinatorics
has been studied by a long line of researchers starting from Cauchy in the early 1800s to Alexandrov
in the 1950s. These results concern the geometric rigidity of polyhedral frameworks, or just poly-
hedra, namely 3-dimensional polytopes that are composed of planar rigid panel faces; face panels
can rotate about the edge or hinge on which they are incident with another panel. A triangulated
polyhedron’s faces are all triangular.
Although these results concern the rigidity-related properties of frameworks, i.e., specific re-
alizations of a combinatorial structure of constraints (from which the GCS can be extracted), as
opposed to rigidity-related properties of GCSs (as discussed so far), in the latter part of this chapter
we will reconcile the slight differences in these two ways of thinking.
Cauchy [5] showed that all convex polyhedra are rigid and in fact globally rigid if convexity
is stipulated. (Mistakes in Cauchy’s proof were fixed and the result extended by a series of subse-
quent researchers.) Despite a long standing conjecture that the result extended to all triangulated
polyhedra (convex or not), verified for many subclasses [10], the general statement was disproven
by counterexample [6]. These results and further developments are discussed in Chapter 13.
Let G = (V, E) be the graph associated to the edge skeleton (or 1-skeleton) of a triangulated
polyhedron and let p : V → R3 be the map that assigns the coordinates of each vertex of the poly-
hedron to a vertex of G. Then (G, p) is a bar-joint framework that is a realization of a bar-joint
constraint system in 3D, with a distance constraint graph G. Thus, Cauchy’s theorem shows that
bar-joint frameworks arising from convex triangulated polyhedra are rigid (and globally rigid if
convexity of the framework is stipulated).
Chapter 16 discusses geometric conditions for global rigidity of generic bar-joint frameworks in
arbitrary dimensions [11]. Recall that a framework is globally rigid if it is the unique realization of
its underlying GCS. A framework (G, p) is generic w.r.t. a property P (such as global rigidity) if for
some neighborhood N (p), for all frameworks (G, q) with q ∈ N (p) (G, q) satisfies P if and only if
(G, p) satisfies P. When the context, namely the property P is clear, we simply say the framework
is generic.
Overview and Preliminaries 7

The result employs a feature of the framework’s equilibrium self-stress (defined in Section 1.3)
and further shows that global rigidity is in fact a generic property, i.e., either all generic frameworks
(G, p) of a graph G are globally rigid or none are. In other words, the property of the framework
only depends on the graph G, but is given a geometric characterization in Chapter 16. Chapters that
additionally give combinatorial as opposed to geometric characterizations of such properties that
depend only on the constraint graph are described in the next section on combinatorial rigidity.
Chapter 14 considers tensegrity frameworks [28] in which the underlying GCS involves in-
equality as opposed to equality constraints. Some edges of the constraint graph, called struts, have
distance lower bounds and others, called ties have distance upper bounds which restrict the sign of
the equilibrium self-stress they can carry. Tensegrity frameworks that represent packed incompress-
ible spheres contain only struts. Bar-joint frameworks are special cases of tensegrity frameworks
where all edges have both distance upper and lower bounds (fixed distances).
Chapter 14 gives geometric, equilibrium self-stress based characterizations for rigidity and other
related properties of tensegrity frameworks both in general and generic settings, and connects them
to rigidity properties of bar frameworks.
Chapter 15 specifically considers nongeneric tensegrity frameworks and uses extended Cayley
algebra (discussed earlier under the Geometric Reasoning Section of the handbook) to give geomet-
ric conditions for rigidity related properties.
Chapter 17 deals with properties related to rigidity that are invariant under various transfor-
mations (beyond the trivial motion group of the underlying geometry). Using characterizations of
properties related to rigidity theory, techniques (such as Coning or Maxwell-Cremona diagrams for
understanding stresses) for GCSs and frameworks in one geometry can be extended to another. For
example, techniques and characterizations from Euclidean geometry can be extended to say Affine,
Projective, Spherical, Minkowski, and Hyperbolic geometries that are defined using Cayley-Klein
metrics or using the trivial motion groups under which the metrics are invariant.

1.2.4 Part IV: Combinatorial Rigidity Techniques

As mentioned earlier, (in)dependence and other properties related to rigidity are often generic (under
appropriate, careful definitions of genericity), i.e., they hold for all generic frameworks and/or GCSs
with a given constraint graph G, or for none of them. They depend only on the underlying constraint
graph G; (in)dependence is captured by a geometric or algebraic matroid such as the (generic)
rigidity matroid defined formally in the second part of this chapter, whose ground set is related to
the algebraic constraints represented by the edges (and nonedges) of G.
This section of the handbook deals with such (in)dependence properties that are equivalently
characterized by purely combinatorial sparsity or graphic matroids, which do not use the algebraic
structure of the constraint polynomials or brackets over the reals (or of the coefficient field of the
constraint polynomials). The book [12] has so far been the trusted source on combinatorial rigidity,
but significant progress has been made since it was published.
A classic example of a purely combinatorial rigidity characterization is a celebrated result of
Laman [21] published in 1970, though recently it has also been found in a forgotten work of Hilda
Pollaczek-Beiringer [27] from 1927.

Theorem 1.4 ([27, 21]) A 2D bar-joint graph G = (V, E) is rigid (all its generic frameworks/GCSs
are rigid) if and only if |E| = 2|V | − 3 and for any subsystem G0 = (V 0 , E 0 ) where |V 0 | > 1, |E 0 | ≤
2|V 0 | − 3.

Graphs satisfying such counting conditions – which keep track of the internal degrees of freedom
(dof) of the system – are often referred to as Laman graphs. This combinatorial characterization of
bar-joint rigidity in 2D led to a series of increasingly refined algorithms to detect rigidity, as well
as maximal rigid subgraphs in flexible constraint graphs. Chronologically, there was a network flow
8 Handbook of Geometric Constraint Systems Principles

based algorithm [16], a matroid sums algorithm [8], a bipartite matching algorithm [14], and finally
what is known as the pebble game [18]. A version of the pebble game (a special case of network
flow) is used for recursive decomposition into approximately rigid subgraphs in the DR-planning
algorithms mentioned earlier. The idea of pebble games has since been extended to the class (k, l)-
sparse graphs [22] where l < 2k, as discussed in Section 1.3.4.
It turns out that the latter condition in the above theorem defines independence in a sparsity
matroid on V ×V . The analogous condition d|V | − d+1


2 formulated by Maxwell in the nineteenth
century [24] is necessary but not sufficient for d ≥ 3 as discussed later in this chapter, see the famous
“double banana” graph, Figure 1.4. Typically, the so-called “Maxwell direction” is showing that
independence in the combinatorial matroid is necessary for independence in the algebraic rigidity
matroid, and is the easier one. The converse direction – that completes the equivalence of the two
types of matroids – is the challenging one.
However, there are combinatorial characterizations of independence and local rigidity of bar-
joint frameworks in 2D which extend to other types of 2D frameworks such as body-bar, body-
hyperpin, 2D bar-joint on the sphere (or 3D line-angle), 2D point-line-incidence-direction, etc. Of-
ten there are more than one equivalent characterization. For example, Lovász and Yemini [23] found
an alternate characterization of 2D bar-joint rigidity that is superficially quite different from Laman-
Pollaczek-Beiringer’s characterization. Chapter 18 and Chapter 21 discuss, respectively, such local
and global rigidity characterizations. The latter chapter begins with the first combinatorial charac-
terization of generic global rigidity, namely that of 2D bar-joint frameworks [17]. Chapter 22 gives
a combinatorial matroid that captures the rigidity matroid of more challenging frameworks in 2D
involving angles between lines with distances between points and lines.

1.2.4.1 Inductive Constructions

In 1911, Henneberg [15] gave the following constructive definition of the class of what later became
Laman-Pollaczek-Beiringer graphs.

Definition 1.1 Henneberg Construction A Henneberg construction of a graph G is a sequence

of the following operations which, beginning with a single edge, results in G.
(a) Add a new vertex and two edges connecting it to two existing vertices.
(b) Subdivide an existing edge and add an additional edge from the new vertex to another existing
vertex.

Laman [21] used this class in the proof of Theorem 1.4. The basic structure of the proof was to
show that the class described in the theorem was exactly the class of graphs with Henneberg con-
structions. Then he proved that for any Henneberg construction there is a positioning of the vertices
that will have no infinitesimal motions. The earlier proof of Pollaczek-Beiringer was stronger, point-
ing out that almost all (or generic positionings) that avoid certain algebraic conditions will have no
infinitesimal motions.
Inductive constructions are one of the mainstays of results that show equivalence between an
algebraic rigidity matroid and a combinatorial or graphic matroid, specifically for the difficult di-
rection, i.e., showing that independence in the combinatorial matroid implies independence in the
algebraic rigidity matroid. Chapter 19 systematically surveys such inductive constructions in many
proofs of combinatorial rigidity characterizations.

1.2.4.2 Body Frameworks

A body geometric primitive is a finite n-dimensional rigid object; this is rather general, including
points, line segments, plane segments, etc. but also any other rigid free-form shape of the same
dimension as the space. A body-hinge system has body primitives and hinge constraints, which are
Overview and Preliminaries 9

incidences between two primitives in d-dimensions where d is less than the dimension of either
object about which the primitives can rotate. In 2D, this must be a bar-joint system. A body-bar
system also has body primitives (as discussed in Section 1.2.4.2), but the constraints are distances
between generic points on the body. Unlike the higher dimensional bar-joint systems, body-bar-
hinge systems have a combinatorial characterization of independence and local rigidity in arbitrary
dimensions, first proved in [38, 39] who pioneered the use of so-called pure conditions (where
certain determinants vanish) to describe genericity. The characterizations extend to special classes
of bar-joint systems that can be cast as body-bar-hinge systems. Combinatorial characterizations of
global rigidity also exist.
Polyhedra (see Section 1.2.3) are a subclass of body-hinge structures. The bodies are panels
(polygonal faces) and the hinges connect the panels; moreover, the system must completely en-
close a volume. The so-called molecular conjecture [37] (referring to the ability to model pro-
tein backbones as body-hinge structures) stated that the rigidity of coplanar hinge and panel hinge
frameworks obeyed the same combinatorial characterization as generic body-hinge structures. It
was proven for general dimension in [19].
Chapter 20 surveys combinatorial characterizations of both local and global rigidity for body-
hinge structures in arbitrary dimensions.

1.2.4.3 Body-Cad, and Point-Line Frameworks

The set of body-cad frameworks is a catch-all category for 3D constraint systems [13]. Motivated
by CAD design software, it includes many of the common constraints and primitives seen in the
industry. The above-mentioned categories are in fact specializations of this class. Primitives include
points, lines, and planes, and constraints include coincidence, angular (parallel, perpendicular, or
arbitrary fixed angles), and distance (cad). Chapter 23 discusses a combinatorial characterization of
(infinitesimal) rigidity for such systems.

1.2.4.4 Symmetric and Periodic Frameworks and Frameworks under

Polyhedral Norms
Chapter 25 develops the set up and techniques for extending combinatorial rigidity characterizations
to symmetric and periodic frameworks (for different symmetry groups). Chapter 24 does the same
for extending from Euclidean distance to polyhedral norms.

1.2.5 Missing Topics and Chapters

The sampling of topics in this handbook would have been more comprehensive with chapters on (a)
Wu-Ritt’s characteristic set method for automated geometry theorem proving mentioned earlier, (b)
the topology (homology and cohomology) of linkage configuration spaces, related to Walker’s prob-
lem, (c) combinatorial and algorithmic studies on expansive motions of linkages and origami related
to the Carpenter’s rule problem, (d) sphere-packing rigidity with results arising from analytic per-
spectives besides the tensegrity perspective, (e) conjectures and progress on characterizing generic
bar-joint rigidity in 3D, and (f) the exploration of genericity, rigidity, and configuration spaces of
periodic and infinite frameworks. Of these, there is extensive expository literature on (a),(b),(c).
We hope that the next revision of the handbook may include chapters condensing the substantial
amount of work on (d), (e), and (f). Finally the area is rich in subtopics and directions arising from
numerous applications, from computer aided engineering and architectural design, molecular and
materials modeling, machine learning and complexity. These however would be outside the scope
of this handbook on geometric constraint principles, being more appropriate for a handbook on
geometric constraints applications.
10 Handbook of Geometric Constraint Systems Principles

1.3 Terminology Reconciliation and Basic Concepts

In this section we clarify slightly different types of terminology used to talk about GCS and frame-
works, and their relationship. We additionally introduce the overall program of combinatorial char-
acterizations of GCS and frameworks: define notions of genericity, introduce the concept of a rigid-
ity matrix and the notion of infinitesimal rigidity, which is a linearization of local rigidity that is
generically equivalent and is used to define the so-called (generic) rigidity matroid.

1.3.1 Constrainedness
The notion of constrainedness exists to discuss the characteristics of the solution space of a GCS.
An over-constrained system is one that has no solutions. A well-constrained system is one that has
a finite number of solutions. An under-constrained system is one that has infinitely many solutions.
Each variable in a system has an approximate degree-of-freedom (dof). Each constraint contributes
at least one equation.
The definitions above apply to a system of equations with sufficiently general parameters. How-
ever, some specific assignments of values to parameters (e.g., length and angle measures) in the
equations corresponding to the same underlying constraint graph may result in a different classifi-
cation. That is, these properties are not structural (or combinatorial) A system is called generic over
a ground field K (usually Q or R in our setting) if the designated constraint parameters are alge-
braically independent over K, i.e., they are not the solutions of any nontrivial polynomial equation
with coefficients in K. Weaker definitions of genericity of GCS exist, for example, stating that the
parameters are not the zeroes of a given set of polynomials, or some finite but unspecified set of poly-
nomials. In all these cases, the set of nongeneric parameter values is measure-zero, i.e., choosing
the parameters at random will result in a generic system with probability of one. Unless otherwise
specified, the strongest definition of algebraic independence of parameter values is implied.
A generic constraint system being under-, well-, or over-constrained implies that all generic
parameter assignments to the same underlying constraint graph result in an under-, well-, or over-
constrained system, respectively. Therefore, being generically *-constrained is a combinatorial
property. Thus, generic constrainedness terminology can be extended to the underlying constraint
graph. Generic *-constrainedness does not imply *-constrainedness of a nongeneric system, and
similarly the converse does not hold either. Therefore, classifying nongeneric constraint systems is
an intractable problem.
However, in the generically over-constrained setting, a specific class of non-generic systems
is of interest. Such a system with a generic assignment to the parameters will have no solution,
but those non-generic systems that do have a solution are called consistently over-constrained. The
sets of generically well-over-constrained and generically under-over-constrained systems are dis-
joint, complementary subsets of generically over-constrained. A generically well-over-constrained
system is one with a spanning generically well-constrained subsystem. A generically under-over-
constrained is any other over-constrained system.
Note that traditionally generically under-constrained was taken to mean the union of generically
under-constrained and generically under-over-constrained as defined here. The current definition is
cleaner since it ensures that the sets of generically under-, well-, and over-constrained are disjoint.

1.3.2 Rigidity of Frameworks

Two frameworks (G, p) and (G, q) are congruent if p and q are congruent modulo trivial motions.
They are equivalent if the underlying GCS of (G, p) is identical to that of (G, q).
A framework (G, p) is rigid if there exists a nonempty neighborhood N (p) (in the Euclidean
Overview and Preliminaries 11

Figure 1.1
A 2D bar-joint framework that is rigid but not infinitesimally rigid. A nontrivial infinitesimal motion
is indicated by the dashed arrow.

topology) such that for all p0 ∈ N (p), congruence of frameworks (G, p) and (G, p0 ) implies equiva-
lence. If the framework is not rigid, it is flexible. A rigid framework is minimally rigid (or isostatic)
if the removal of any hyperedge of G results in a flexible framework. The framework is globally
rigid (or strongly rigid) if, for all p0 for which the framework (G, p0 ) is equivalent to (G, p), it also
holds that they are congruent.
Given a constraint graph G, there is a system of polynomial equations F = { f1 , . . . , fm } and
variables X = {x1 , . . . , xn } (as explained in Section 1.1.) The rigidity matrix R(G) is the Jacobian of
this system with respect to X, i.e., the m × n matrix with element (i, j) equal to ∂ fi /∂ x j . The rigidity
matrix of a framework (G, p) (written as R(G, p)) is R(G) with all variables xi replaced by p(xi ).
When working with most constraint systems, the equations are quadratic and this process is often
referred to as linearization of the polynomial system.
As explained earlier, the solution space of a geometric constraint system is a variety in K n (where
we may let K = C to make full use of the algebraic theory); it is the set of zeros of the corresponding
polynomial system. The row span of the rigidity matrix R(G, p) is the space of normals to this variety
at point (p(x1 ), p(x2 ), . . . , p(xn )). Any infinitesimal movement on the tangent space, orthogonal to
the space of normals (i.e., along the variety), will give another solution. The tangent space is the
right nullspace (or kernel) of R(G, p). Geometrically, any infinitesimal vector in the right nullspace
represents an infinitesimal change to each primitive such that the resulting framework still satisfies
all of the constraints.
The left nullspace of the rigidity matrix also has a geometric interpretation. The left nullspace
only has nonzero vectors if there is a linear dependency in the rigidity matrix, which corresponds to
an equilibrium self-stress of the system.
The degree-of-freedom (dof ) of the framework is the nullity of R(G, p) (i.e., the rank of the right
nullspace.) The framework is infinitesimally rigid if the dof is equal to the number of trivial motions.
Geometrically, this means that all infinitesimal motions arise from trivial motions of the space. By
the rank-nullity theorem, a framework is also infinitesimally rigid exactly when the rank of R(G, p)
is n − k, where n is the number of variables and k is the number of trivial motions.
In the case of
d-dimensional Euclidean space, the space of trivial motions has dimension d+1 2 , which are the d
translations plus 2 rotations. If the right nullspace of R(G, p) has dimension greater than d+1
d


2 ,
then the framework (G, p) is said to have infinitesimal motions (or infinitesimal flexes).
It is clear that infinitesimal rigidity implies rigidity, and we give an example of a bar-joint frame-
work that is rigid but not infinitesimally rigid to show that the converse is false. A comprehensive
introductory treatment of the rigidity of graphs (bar-joint systems) can be found in Ref. [12].
To prove that rigidity does not imply infinitesimal rigidity, consider Figure 1.1. This is a rigid
bar-joint graph that has an infinitesimal motion that is zero at all places except at the single vertex
in the direction of the arrow. A rigid framework with a nontrivial infinitesimal motion is called
degenerate and must be in a nongeneric realization. A combinatorial characterization of rigidity is
12 Handbook of Geometric Constraint Systems Principles

Figure 1.2
Two different 2D frameworks of the same generically flexible bar-joint constraint graph. The de-
generate framework on the left is rigid, whereas the generic configuration on the right is flexible.

only guaranteed to hold for a certain generic class of realizations, and we formalize what we mean
by genericity in the next section.
The rigidity matrix has a natural notion of dependence, based on the linear dependence of the
rows (or columns) of the matrix. As such, the rigidity matroid of a framework [12] is simply the
linear matroid of the rows of its rigidity matrix. That is, the row vectors of the matrix comprise the
ground set, and the linearly independent subsets of rows comprise the family of independent sets.
Therefore, the matroid rank (the maximum cardinality of an element in the matroid) is exactly the
rank of the matrix. Since each row corresponds to some constraint, a dependent row corresponds to a
dependent constraint. The framework as a whole is independent if there are no dependent constraints
and is dependent otherwise.

1.3.3 Generic Rigidity of Frameworks

Determining rigidity of frameworks is difficult; deciding global [29] and local rigidity [1] are
both strongly NP-hard for bar-joint systems. However, if a certain measure-zero set of primitive
arrangements is excluded, determining rigidity can become much easier for certain constraints and
primitives. This is the set of degenerate frameworks, which will be discussed in this section.
In Section 1.3.1, the notion of genericity was discussed in the context of geometric constraint
systems. When considering frameworks, genericity has a different meaning. A framework (G, p) is
said to be generic with respect to property P (G, p) satisfies N P if and only if there exists some
neighborhood N (p) around p such that for all points p0 ∈ N (p) the framework (G, p0 ) satisfies P.
A property P is said to be generic when, for all constraint graphs G, either all generic frame-
works of G (w.r.t. P) satisfy P or none satisfy P. If a property of frameworks (G, p) is generic, then
it is a combinatorial property of the underlying constraint graph alone. Intuitively, a generic prop-
erty of a framework is one that is maintained if primitives were “wiggled” by small amounts in any
direction. For example, the independence of the rigidity matrix is a generic property of frameworks.
To illustrate the importance of considering generic frameworks, consider the following exam-
ples. See Figure 1.2 which depicts two frameworks of the same constraint graph. The nongeneric
framework on the left is rigid while being generically flexible. The three bar “chain” is taut, dis-
allowing finite flexes; with only slightly different lengths, this would no longer be rigid. See also
Figure 1.3 which also depicts two frameworks corresponding to a single constraint graph. The non-
generic framework on the left is flexible while being generically rigid. The three vertical bars are
the same length, permitting a vertical shear; with different lengths, there would be no infinitesimal
motions.
Whereas rigidity does not imply infinitesimal rigidity of a framework, in the generic case it
does. The Implicit Function Theorem from multivariate calculus, as in Ref. [3], shows that when
a framework is generic w.r.t. infinitesimal rigidity, every infinitesimal flex can be converted into a
Overview and Preliminaries 13

Figure 1.3
Two different 2D frameworks of the same generically rigid bar-joint constraint graph (known as
C2 ×C3 ). The degenerate framework on the left is flexible, whereas the generic configuration on the
right is rigid.

Table 1.1
Correspondence between constrainedness terminology when used in the context of generic systems
and generic frameworks.
Generic Systems Realizations Generic Framework
Under-constrained Infinite solutions Independent and flexible
Well-constrained Finite solutions Independent and rigid
Over-constrained No solutions Dependent
Under-over-constrained No solutions Dependent and flexible
Well-over-constrained No solutions Dependent and rigid

finite flex, i.e., rigidity in fact implies infinitesimal rigidity. Since infinitesimal rigidity is a generic
property, this shows that rigidity is also a generic property. As mentioned, this property can be
thought of as a combinatorial property of the underlying constraint graph. Therefore, a constraint
graph is called rigid if some generic framework of the graph is infinitesimally rigid. The notions of
flexibility and minimal rigidity have obvious meanings in the generic sense as well.
The generic rigidity matroid of a geometric constraint system with constraint graph G can be
defined in two ways: (1) the rigidity matroid of any generic framework (G, p), or (2) the rigidity ma-
troid formed by the rows of the rigidity matrix R(G) of indeterminates. This leads to combinatorial
notions of independence among constraints.
Table 1.1 establishes a rough correspondence between the genericity of a GCS, a framework, and
rigidity. Traditionally, generically under-constrained was used to describe all flexible frameworks
and was therefore a union of under- and under-over-constrained as defined here. The symmetry of
the new definitions displayed above is another argument for the new terminology.

1.3.4 Approximate Degree-of-Freedom and Sparsity

There is a real, algebraic notion of degree-of-freedom that is discussed in Section 1.3.2. This sec-
tion introduces a different, combinatorial idea of degree-of-freedom that can be used to approxi-
mate rigidity. This was briefly mentioned in Section 1.3.1, and has an obvious relationship with
constrainedness. The following methods are very general, working for any type of primitive or con-
straint. However, a positive result for rigidity by this characterization is often only a necessary con-
dition for true rigidity. Many systems require additional considerations, if there is a combinatorial
characterization at all.
In the constraint graph, each primitive has some degrees-of-freedom (dofs) and each constraint
eliminates some dofs between participating primitives. In some of the literature, particularly that
which uses network flow based algorithms, dof corresponds to the negation of the density of the
constraint graph. Given a constraint graph G = (P,C), with primitives P and constraints C, and a
14 Handbook of Geometric Constraint Systems Principles

Figure 1.4
A 3D bar-joint framework, known as the double-banana. It is flexible due to a rotation about the
dashed line, despite being dof-rigid (i.e., (3, 6)-tight).

weight function w on P and C, the density of G is

X X
d(G) = w(c) − w(p).
c∈C p∈P

Given some constant k equal to the number of trivial motions of the underlying geometry, a con-
straint graph is minimally dof-rigid if it has k dofs and every subgraph has at least k dofs. In terms
of density, the graph is minimally dof-rigid if d(G) = −k and for all subgraphs G0 , d(G0 ) ≤ −k. A
graph is dof-rigid if it contains some minimally dof-rigid subgraph.
For example, consider 2D bar-joint systems: points can be thought of as having 2 dofs (transla-
tion but not rotation) and an edge between two points destroys 1 dof, leaving a system with 3 dof
(translation and rotation.) In Euclidean 2D space, there are 3 trivial motions and therefore a sin-
gle edge would be dof-rigid. In fact, as mentioned earlier, this notion of dof-rigid exactly captures
generic rigidity of 2D bar-joint systems. We follow the convention of referring to methods using dof
analysis as Laman counts.
We can understand the rigidity of some systems combinatorially using only dof analysis. How-
ever, this combinatorial notion of dof-rigid does not usually imply generic rigidity. Consider 3D
bar-joint systems: points instead have 3 dofs, edges still eliminate 1 dof, and the space has 6 trivial
motions. By the counts, the famous “double banana” graph in Figure 1.4 is dof-rigid; however, the
“bananas” can clearly swivel about the dashed hinge.
As mentioned earlier, the Laman-Pollaczek-Beiringer theorem gives a purely combinatorial
property (no algebra, simply counting) to capture the properies of the rigidity matrix and there-
fore the matroid. These subsystems are exactly the independent sets of the 2D bar-joint rigidity
matroid.
This theorem motivated the notion of sparsity and sparsity counts [22]. This terminology is used
to discuss constraint graphs where all primitives have k dofs and all constraints eliminate one dof
and are binary; however, the theory does allow loops, effectively permitting unary constraints, and
allows for multiedges, so constraints that eliminate n dofs can be represented in as many edges. A
graph G = (V, E) is (k, l)-sparse, for every induced subgraph G0 = (V 0 , E 0 ), the inequality |E 0 | ≤
k|V 0 | − l. The graph is (k, l)-tight if it additionally satisfies |E| = k|V | − l.
For example, Laman graphs would be the set of (2, 3)-tight graphs. A 2D system of 2D rigid
bodies and distance constraints would use k = 3 and l = 3; in fact, (3, 3)-tight graphs are exactly
the class of 2D rigid body-bar graphs. For a fixed k and l there is often a natural interpretation of a
(k, l)-tight graph as a constraint system in which geometric primitives have k dof.
For all l < 2k, these counts define a sparsity matroid where the basis is the set of edges and the
independent sets are the edges in the (k, l)-sparse subgraphs. This allows for an efficient class of
algorithms, called pebble games, which can detect (k, l)-sparse graphs in polynomial time, if l < 2k.
 References 15

1.4 Alternative Pathway through the Book

Given the interconnected nature of the chapters of this book there are many logical ways of or-
dering the material. Here, we describe an alternative navigation that might be more accessible to a
newcomer.
In this suggested pathway, Group I contains chapters highlighting the different perspectives
from developing solvers for general geometric constraint systems to analyzing the rigidity of bar-
and-joint structures to constructing purely combinatorial objects arising in rigidity theory. Groups
II and III focus on the problems studied in Rigidity Theory, introducing concepts such as local and
global rigidity and considering models initially restricted to points before generalizing to other types
of geometric primitives (e.g., rigid bodies or lines). Groups IV and V shift the structure to partition
by the underlying approaches (metric geometry and algebraic methods).
Group I Getting Started: Chapters 1, 6, 7, 18, and 19
To get the lay of the land, start with two sets of chapters highlighting the perspec-
tives of historically distinct communities. Chapters 6 and 7 focus on decomposition-
recombination approaches used in computing realizations of a general GCS. In contrast,
Chapter 18 restricts its content to the classical structure studied in Rigidity Theory of
2-dimensional bar-and-joint frameworks (introduced in Section 1.1.1); the combinatorial
property characterizing generic bar-and-joint rigidity is studied in a generalized setting in
Chapter 19.
Group II Rigidity Theory for Point Primitives: Chapters 21, 16, 8, 14, 15, and 25
Building upon the fundamentals introduced in Group I, continue the Rigidity Theory per-
spective, with topics posed in the setting of GCS with points as the geometric primitives.
Group III Rigidity Theory for Other Primitives: Chapters 20, 13, 22, and 23
Next, move to the Rigidity Theory for systems defined on geometric primitives beyond
simply points (e.g., rigid bodies, lines).
Group IV Metric Geometry: Chapters 24, 17, 9, and 10
Shifting to an organization based on underlying techniques, start with approaches relying
on distance and metric geometry.
Group V Algebraic Methods: Chapters 12, 11, 3, 2, 5, and 4
Finally, conclude with chapters based on algebraic methods.
The intention of this pathway is to help to get readers started with the varying perspectives on
GCS formulation (Group I). Then, start with topics in Rigidity Theory (Groups II and III) before
shifting to an order partitioned more by the underlying machinery (Groups IV and V).

Acknowledgment
We thank Jessica Sidman and Rahul Prabhu for a careful reading, and Audrey St. John for providing
the alternative pathway.

References
[1] Timothy Good Abbott. Generalizations of Kempe’s universality theorem. PhD thesis, Mas-
sachusetts Institute of Technology, 2008.
16 References

[2] Dennis S Arnon, George E Collins, and Scott McCallum. Cylindrical algebraic decomposition
I: The basic algorithm. SIAM Journal on Computing, 13(4):865–877, 1984.
[3] L. Asimow and B. Roth. The rigidity of graphs. Transactions of the American Mathematical
Society, 245:279–289, November 1978.
[4] B. Buchberger. A note on the complexity of constructing Gröbner-bases, pages 137–145.
Springer Berlin Heidelberg, Berlin, Heidelberg, 1983.
[5] Augustin-Louis Cauchy. Sur les polygones et les polyèdres: second mémoire. Journal de
l’École Polytechnique, 9:87, 1813.
[6] Robert Connelly. The rigidity of polyhedral surfaces. Mathematics Magazine, 52(5):275–283,
November 1979.
[7] Ioannis Fudos and Christoph M. Hoffmann. Correctness proof of a geometric constraint
solver. International Journal of Computational Geometry & Applications, 06(04):405–420,
1996.
[8] Harold N Gabow and Herbert H Westermann. Forests, frames, and games: algorithms for
matroid sums and applications. Algorithmica, 7(1-6):465–497, 1992.
[9] Giovanni Gallo and Bud Mishra. Wu-Ritt characteristic sets and their complexity. Discrete
and Computational Geometry: Papers from the DIMACS Special Year, 6:111–136, 1991.
[10] Herman Gluck. Almost all simply connected closed surfaces are rigid. In Geometric topology,
volume 438 of Lecture Notes in Mathematics, pages 225–239. Springer, 1975.
[11] Steven J Gortler, Alexander D Healy, and Dylan P Thurston. Characterizing generic global
rigidity. American Journal of Mathematics, 132(4):897–939, 2010.
[12] Jack E. Graver, Brigitte Servatius, and Herman Servatius. Combinatorial Rigidity, volume 2.
Graduate Studies in Math., AMS, 1993.
[13] Kirk Haller, Audrey Lee-St.John, Meera Sitharam, Ileana Streinu, and Neil White. Body-
and-cad geometric constraint systems. Comput. Geom. Theory Appl., 45(8):385–405, October
2012.
[14] Bruce Hendrickson. Conditions for unique graph realizations. SIAM J. Comput., 21(1):65–84,
February 1992.
[15] Lebrecht Henneberg. Die graphische Statik der starren Systeme, volume 31. BG Teubner,
Leipzig, 1911.
[16] Hiroshi Imai. On combinatorial structures of line drawings of polyhedra. Discrete Applied
Mathematics, 10(1):79 – 92, 1985.
[17] B. Jackson and T. Jordán. Connected rigidity matroids and unique realizations of graphs. J.
Combinatorial Theory Ser B, 94:1–29, 2005.
[18] Donald J. Jacobs and Bruce Hendrickson. An algorithm for two-dimensional rigidity perco-
lation: The pebble game. Journal of Computational Physics, 137(2):346 – 365, 1997.
[19] Naoki Katoh and Shin-ichi Tanigawa. A proof of the molecular conjecture. Discrete &
Computational Geometry, 45(4):647–700, 2011.
[20] Alfred B Kempe. On a general method of describing plane curves of the nth degree by
linkwork. Proceedings of the London Mathematical Society, 1(1):213–216, 1875.
[21] G. Laman. On graphs and rigidity of plane skeletal structures. Journal of Engineering Math-
ematics, 4(4):331–340, October 1970.
References 17

[22] Audrey Lee and Ileana Streinu. Pebble game algorithms and sparse graphs. Discrete Mathe-
matics, 308(8):1425–1437, 2008. Third European Conference on Combinatorics, Graph The-
ory, and Applications.
[23] L. Lovász and Y. Yemini. On generic rigidity in the plane. SIAM Journal on Algebraic
Discrete Methods, 3(1):91–98, 1982.
[24] James Clerk Maxwell. On reciprocal figures and diagrams of forces. Philosophical Magazine,
27:250–261, 1864.
[25] Ernst Mayr and R. Cori. Membership in polynomial ideals over Q is exponential space com-
plete, pages 400–406. Springer Berlin Heidelberg, Berlin, Heidelberg, 1989.
[26] J Owen and S Power. The non-solvability by radicals of generic 3-connected planar laman
graphs. Transactions of the American Mathematical Society, 359(5):2269–2303, 2007.
[27] H. Pollaczek-Geiringer. über die gliederung ebener fachwerk. Zeitschrift fur Angewandte
Mathematik und Mechanik (ZAMM), 7:58–72, 1927.
[28] B. Roth and W. Whiteley. Tensegrity frameworks. Transactions of the American Mathematical
Society, 265(2):419–446, June 1981.
[29] James B Saxe. Embeddability of weighted graphs in k-space is strongly NP-hard. Carnegie-
Mellon University, Department of Computer Science, 1980.
[30] I. J. Schoenberg. On certain metric spaces arising from euclidean spaces by a change of metric
and their imbedding in hilbert space. Annals of Mathematics, 38(4):787–793, December 1936.
[31] I. J. Schoenberg. Metric spaces and positive definite functions. Transactions of the American
Mathematical Society, 44(3):522–526, November 1938.
[32] Neil JA Sloane. Error-correcting codes and invariant theory: new applications of a nineteenth-
century technique. American Mathematical Monthly, pages 82–107, 1977.
[33] Larry Smith. Polynomial invariants of finite groups. a survey of recent developments. Bulletin
of the American Mathematical Society, 34(3):211–250, 1997.
[34] D. M. Y. Sommerville. An Introduction to the Geometry of n Dimensions. 1958.
[35] Ágnes Szántó. Complexity of the Wu-Ritt decomposition. In Proceedings of the second
international symposium on parallel symbolic computation, pages 139–149. ACM, 1997.
[36] Alfred Tarski. A decision method for elementary algebra and geometry. In Quantifier elimi-
nation and cylindrical algebraic decomposition, pages 24–84. Springer, 1998.
[37] T.-S. Tay and W. Whiteley. Recent advances in the generic rigidity of structures. Structural
Topology, 9:31–38, 1984.
[38] Tiong-Seng Tay. Rigidity of multi-graphs. i. linking rigid bodies in n-space. Journal of
Combinatorial Theory, Series B, 36(1):95 – 112, 1984.
[39] Neil White and Walter Whiteley. The algebraic geometry of motions of bar-and-body frame-
works. SIAM J. Algebraic Discrete Methods, 8(1):1–32, 1987.
[40] Neil L. White. A Tutorial on Grassmann-Cayley Algebra, pages 93–106. Springer Nether-
lands, Dordrecht, 1995.
 Part I

Geometric Reasoning, Factorization

and Decomposition
Chapter 2
Computer-Assisted Theorem Proving in
Synthetic Geometry
Julien Narboux
University of Strasbourg

Predrag Janičić
University of Belgrade

Jacques Fleuriot
University of Edinburgh

CONTENTS
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2 Automated Theorem Proving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.1 Foundations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.2 Nondegenerate Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.3 Purely Synthetic Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.3.1 Early Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2.3.2 Deductive Database Method, GRAMY, and iGeoTutor . . . 24
2.2.3.3 Logic-Based Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.4 Semisynthetic Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2.4.1 Area Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2.4.2 Full-Angle Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2.4.3 Vector-Based Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2.4.4 Mass-Point Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2.5 Provers Implementations and Repositories of Theorems . . . . . . . . . . . . . . . . . . . . 33
2.3 Interactive Theorem Proving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.1 Formalization of Foundations of Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.1.1 Hilbert’s Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.1.2 Tarski’s Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.3.1.3 Axiom Systems and Continuity Properties . . . . . . . . . . . . . . . . . . . 38
2.3.1.4 Other Axiom Systems and Geometries . . . . . . . . . . . . . . . . . . . . . . . 40
2.3.1.5 Meta-Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.2 Higher Level Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.3 Other Formalizations Related to Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.3.4 Verified Automated Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

21
22 Handbook of Geometric Constraint Systems Principles

2.1 Introduction
Computer-assisted proof in mathematics has been underway since the pioneering times of comput-
ers in the 1950s. Starting from early systems with very limited capability, computer-assisted theorem
proving has evolved to demonstrate theorems never proved before by humans [156] and assist with
monumental efforts spanning several man-years [116]. In this endeavour, geometry plays an impor-
tant part, just as it has throughout the history of mathematics. This is due to its pervasive role: it
is a paradigmatic form of reasoning, with applications to education, mathematical and physical re-
search, but also to many applied areas such as robotics, computer vision, and CAD [48]. Moreover,
many of the search techniques and other algorithmic features developed for geometric reasoning
have influenced other areas of artificial intelligence.
As for other subfields of computer-assisted proof, the mechanization of geometry spans both
automated and interactive theorem proving. In the former, computers aim to prove theorems com-
pletely automatically, while in the latter, the role of the system is to act as a proof assistant that
verifies the reasoning steps of the user, guides the proving process, and provides some limited au-
tomation. These two branches are often connected through methods that can produce geometric
proofs automatically, where either the proofs or the methods themselves are fully verified.
Just as for its pen-and-paper counterpart, computer-assisted proof in geometry is subject to dif-
ferent foundations: algebraic or synthetic (axiomatic) ones. This chapter aims to provide a com-
prehensive survey of the latter and of its applications. In particular, it will mostly deal with pla-
nar Euclidean geometry [6], which in this case generally means a theory consisting of geometric
statements true in R2 . There are several formal systems that aim to axiomatize this theory or its
subtheories, including those due to Euclid, Hilbert, and Tarski.
In what follows, to ensure a coherent exposition, we will use a uniform notation (which may
differ at times from that used by the original authors). In particular, we will denote points by upper-
case letters, lines by lowercase ones, the strict notion of betweenness by A−B · −C· (i.e., B belongs to
segment AC and is different from A and C) while its nonstrict version will be denoted by A−B−C,
collinearity by Col A BC, perpendicularity by ⊥, cyclicity of points, i.e., points lying on the same cir-
cle, by cyclic(A, B,C, D), the angle between half-lines AB and AC by ∠ABC, the full-angle between
lines AB and CD by ∠[AB,CD], a triangle with vertices A, B, C by 4ABC, congruence between seg-
ments, triangles, or between angles by ∼ =, and equality over measures of angles or over full-angles
by =.

2.2 Automated Theorem Proving

Automated theorem proving in geometry is often considered a “classical AI domain.” Its methods
can be split into three major families or styles: algebraic, synthetic, and semisynthetic.
The algebraic methods deal with the algebraized formulation of geometric statements and usu-
ally involve dealing with the membership of polynomial ideals [45, 217, 224], quantifier elimina-
tions [64, 212], or use coordinate-free approaches based on bracket and Cayley algebras, described
in Chapter 3. Although they are powerful, they cannot produce human-readable proofs and, gener-
ally, consist of steps that do not have any obvious meaning in synthetic geometry. The second and
the third groups of methods, the subjects of this review, focus on proving theorems via geometric ax-
ioms (or higher-level geometric lemmas) and often try to automate the traditional theorem proving
approaches, while attempting to generate human-readable proofs.
Automated theorem proving is used in various contexts, e.g., for mathematical education [19],
Computer-Assisted Theorem Proving in Synthetic Geometry 23

in the simplification of geometric axiom systems [69] and in the study of incidence geometry using
term rewriting techniques [5].
Unless otherwise stated, the methods presented next only deal with planar geometries assuming
the parallel postulate.
Note also that the examples will be presented in a uniform way, although we have taken care to
preserve the essence of the methods being applied. Finally, we remark that there are other surveys
[48, 90, 97, 131, 218] that cover some of the approaches that we will discuss next.

2.2.1 Foundations
Aside from the algorithmic techniques used, automated theorem proving methods rely on various
choices with regard to the underlying logic and geometric knowledge. One issue relates to a set
of axioms to be used. Some methods use well-known geometric axioms sets, but most use custom
(finite sets of first-order) axioms. In the latter case, the axioms are actually simple theorems of Eu-
clidean geometry whose proofs are not considered and are asserted as facts belonging to common
geometric knowledge (hence they are often called “lemmas” or “rules”). The set of axioms is not
necessarily minimal as they are often selected to ease the automatic proof of more complex theo-
rems. For many methods, choosing an appropriate set and level of axioms is one of most critical
issues when it comes to power and efficiency.

2.2.2 Nondegenerate Conditions

The notion of nondegenerate (NDG) conditions arises for each style of automated geometric rea-
soning. Namely, it is often the case that the goal (denoted by G) is implied by the configuration
(denoted by C) plus some additional conditions [46, 48, 55]. For such conditions (denoted by ndg),
the following formulae (where ∀∗ and ∃∗ denote universal and existential closure) have to be valid:
∃ ∗ (C ∧ ndg) and ∀ ∗ (C ∧ ndg ⇒ G). In many cases, the methods can automatically produce such
NDG conditions, although not necessarily the “weakest” ones. There are, however, algorithms for
computing the weakest NDG conditions [44].

2.2.3 Purely Synthetic Methods

Purely synthetic methods, or simply synthetic methods do not use coordinates and algebraic forms
for the geometric statements. Many of these techniques add auxiliary elements to the geometric
configuration under consideration, so that certain axioms can be applied. This usually leads to a
combinatorial explosion in the search space. The challenge then rests in controlling this explosion
and in developing suitable heuristics in order to avoid unnecessary construction steps. Due to the na-
ture of these problems, such synthetic proof techniques are sometimes called Artificial Intelligence
(AI) methods for automated theorem proving in geometry.
The very first AI method was developed by Gelernter et al. Quoting the authors: “In early spring,
1959, an IBM 704 computer, with the assistance of a program comprising some 20000 individual in-
structions, proved its first theorem in elementary Euclidean plane geometry” [101]. Their program,
called the Geometry Machine and written in FORTRAN [99, 100], was not only the first automated
theorem prover for geometry, but also one of the very first automated reasoning systems for mathe-
matics. Although its power, from a modern point of view, was very limited, this system is important
both for historical reasons and for introducing a number of ideas and techniques that were used by
many subsequent reasoning systems. At the time, geometry was viewed as a typical, paradigmatic
AI domain but also as a potentially easy domain where simple ad-hoc rules, basic forward and back-
ward chaining ∗ applied exhaustively, and simple heuristics could be used to bear easy fruit e.g. such
∗ Forward and backward chaining are two important forms of inferences within reasoning systems. The former can be viewed

as a sequence of applications of modus ponens that derives new facts from existing premises in order to prove the goal, while
24 Handbook of Geometric Constraint Systems Principles

as cracking the whole area of high school geometry. In the description of the approaches, logical
representation, algorithms, implementation details, and even the description of features of the pro-
gramming languages were often intermixed. Over time, more sophisticated and mature approaches
have emerged, with many still using some of the early techniques and ideas though.

2.2.3.1 Early Systems

Gelernter’s Geometry Machine implemented several reasoning techniques. These included the use
of “diagrams” (concrete models in the Cartesian plane) in an attempt to reject false subgoals before
any attempt to prove them, dealing with symmetries, and the use of simplification rules (akin to
modern rewrite rules). The system worked by backward reasoning – starting with the given goal
and trying to decompose it to simpler provable subgoals. Its basic rules were based on axioms about
the congruence of triangles. Thus, proving that two segments were congruent could be done by
showing that these two segments were the corresponding edges of two congruent triangles.

Example. Proving the following theorem, in less than 20s, was one of the big triumphs of Geom-
etry Machine: if ∠ABD ∼
= ∠DBC, AD ⊥ AB, DC ⊥ BC, then AD ∼ =CD (Figure 2.1):
∠ABD ∼= ∠DBC (Premise)
∠DAB is right angle (by Definition of perpendicular)
∠DCB is right angle (by Definition of perpendicular)
∠BAD ∼= ∠BCD (by All right angles are congruent)
BD ∼
= BD (by Reflexivity of congruence)
4ADB ∼= 4CDB (by Congruence of triangles, rule Side-Angle-Angle)
AD ∼
=CD (by Corresponding elements of
congruent triangles are congruent)
There were several subsequent systems improving on Gelern- A
ter’s ideas, e.g., by combining backward and forward chaining,
by trying to model the human solving process more faithfully,
or by being designed to serve as support for tutoring systems D
[2, 109, 56, 57, 89, 103, 105, 136, 166]. However, despite all these
efforts, these early systems had a very limited scope and were only B C
able to prove geometric problems of small or moderate complexity.
They didn’t treat NDG conditions and were not able (or were able Figure 2.1
only to a limited extent) to add new, “auxiliary” points, necessary Geometry machine diagram.
in many proofs. So, they typically dealt only with axioms and conjectures of the following form
(universal closure is assumed): A1 (~x) ∧ . . . ∧ An (~x) ⇒ B(~x), where~x denotes a sequence of variables,
Ai and B are atomic formulae or their negations.

2.2.3.2 Deductive Database Method, GRAMY, and iGeoTutor

There are several theorem proving methods, including the deductive database (DD) [54] and those
used by the systems GRAMY [155] and iGeoTutor [220], that deal with “rules” and conjectures
of the form (universal closure is assumed): A1 (~x) ∧ . . . ∧ An (~x) ⇒ B1 (~x) and also rules of the form:
A1 (~x) ∧ . . . ∧ An (~x) ⇒ ∃~y(B1 (~x,~y) ∧ . . . ∧ Bm (~x,~y)) where ~x and ~y denote sequences of variables,
Bi are atomic formulae and A j are atomic formulae or their negations. There are no disjunctions
either in the rules or in the conjectures. Hence, these methods cannot prove conjectures involving
existential quantifiers, but can use new, auxiliary points (or segments), while searching for a proof. †
the latter can be viewed as applying modus ponens backward to refine the goal into subgoals that can hopefully be proven
from the premises.
† Using auxiliary points makes a substantial change compared to the early methods (Section 2.2.3.1), a change that enabled

the proof of a wider set of complex theorems.

Computer-Assisted Theorem Proving in Synthetic Geometry 25

One of the main challenges, though, lies in controlling the introduction of additional objects since
these can lead to a combinatorial explosion.
Unlike for algebraic methods, a common motivation here is the generation of human-readable
synthetic proofs that are as close as possible to those taught in schools. Moreover, in the case of
GRAMY, the generation of several proofs is attempted, making it suitable for some forms of tutor-
ing.

Scope. The methods deal with formulae containing no function symbols and with fixed sets of
predicates symbols. For instance, the DD method uses predicate symbols corresponding to geo-
metric relations (over points), such as Col , ⊥, cyclic(), and equality over full-angles (see Section
2.2.4.2). GRAMY deals not only with points, but also with segments, angles and triangles, and with
a set of predicate symbols that includes ∼
= , ⊥, k, membership, etc. Each system uses a fixed set of
axioms, for instance, the DD method uses around 75 axioms of the first form, including:
D41: cyclic(A, B, P, Q) ⇒ ∠[PA, PB] = ∠[QA, QB]
D42: ∠[PA, PB] = ∠[QA, QB] ∧ ¬ Col P Q A B ⇒ cyclic(A, B, P, Q)
D74: ∠[AB,CD] = ∠[PQ,UV ] ∧ PQ ⊥ UV ⇒ AB ⊥ CD
and around 20 rules of the second form, including, for example:
X1: OM ⊥ MA ∧ ∠[XO, MO] = ∠[MO, AO] ⇒ ∃B (Col B A M ∧ Col B O X)
The DD method was reported as being able to prove, in a matter of seconds and via hundreds of
derived facts, 160 out of the 600 theorems in the authors’ collection of results proved by Wu’s
method. GRAMY and iGeoTutor were applied on smaller benchmark sets gathered from different
sources.

Theorem proving mechanisms. For a given conjecture, atomic formulae from the hypotheses
are considered as “facts.” All three methods use forward chaining for deriving new facts using the
available axioms (the DD method takes some ideas from deductive database theory [96]). There is
a number of techniques (e.g., based on symmetries) used for keeping the number of stored derived
facts low and the proving process efficient.
Auxiliary points are introduced only in a controlled manner, determined by various strategies,
such as introducing new points only if new facts cannot be derived using forward chaining, or intro-
ducing points only through a very limited number of templates, specific geometry configurations.
GRAMY and iGeoTutor do not handle NDG conditions, while the DD method treats them using
a form of negation as failure augmented with some basic diagrammatic model checking.
iGeoTutor can deal with conjectures involving some arithmetical constraints and for that purpose
uses external provers that are specialized for theories like linear arithmetic and are available within
the SMT (Satisfiability Modulo Theory) solver Z3 [160].

Example. The system iGeoTutor can prove the following theo- A D

rem: Given a square ABCD, a point in its interior such that AP ∼
= PD P
◦
and ∠PAD = 15 , prove that the triangle 4PBC is equilateral. The
system, using the fact AB ∼
= AD, decides to use the congruent tri- Q
angles template, and adds a points Q such that ∠BAQ ∼ = ∠PAD
and AQ ∼= AP (such that 4AQB ∼ = 4APD holds). Later, the sys-
tem decides to use another template and introduces the segment
QP. By the initial forward chaining, the system deduces BP ∼ = CP, B C
◦
∠APD = 150 (among others facts). The rest of the proof is as fol-
Figure 2.2
lows:
iGeoTutor diagram.
26 Handbook of Geometric Constraint Systems Principles

4AQB ∼
= 4APD (by Congruence of triangles, rule Side-Angle-Side)
∠BAQ = 15◦ , ∠BQA = 150◦ (by Corresponding parts of
congruent triangles are congruent)
∠QAP = 60◦ (∠QAP = 90◦ − ∠BAQ − ∠PAD)
∠AQP = 60◦ , ∠APQ = 60◦ (by AQ ∼= AP, Isosceles triangle)
AQ ∼
= PQ (by 4AQP is equilateral)
∠BQP = 150◦ (∠BQP = 360◦ − ∠BQA − ∠AQP)
4AQB ∼= 4PQB (by BQ ∼= BQ, ∠BQA ∼ = ∠BQP
∼
AQ = PQ, Side-Angle-Side)
AB ∼
= BP (by Corresponding parts of
congruent triangles are congruent)
BC ∼
= BP, BC ∼
=CP (by AB ∼= BC, AB ∼= BP, BP ∼=CP,
Transitivity of congruence)

Properties. These methods are complete, with respect to the set of rules being used, for proofs
that do not require the introduction of auxiliary points since the search space is then finite. This
feature enables not only proof, but also the deduction of additional facts from a given configuration
and hence the discovery of new theorems. If a proof requires auxiliary elements, completeness can
be proved only under specific conditions.

2.2.3.3 Logic-Based Approaches

Geometric theorems can be proved not only by dedicated systems but also by general ones, typically
by theorem provers for first order logic or some of its fragments such as Coherent logic (CL). Such
automated provers usually cover rich sets of formulae that include existential quantification. We
review some of these approaches next.
Coherent logic consists of formulae of the following form (universal closure is assumed) [15,
85]: A1 (~x) ∧ . . . ∧ An (~x) ⇒ ∃~y(B1 (~x,~y) ∨ . . . ∨ Bm (~x,~y)), ~x and ~y denote sequences of variables, Ai
denotes an atomic formula, and B j denotes a conjunction of atomic formulae. There are no function
symbols with arity greater than zero and there is no negation. Resolution logic (RL) deals with
clauses, i.e., formulae of the following form (universal closure is assumed): A1 (~x) ∨ . . . ∨ An (~x),
where ~x denotes sequences of variables, and Ai denotes an atomic formula or a negation of an
atomic formula.
CL conveys a wider range of formulae compared to the DD method, for instance (Section
2.2.3.2) – there can be existential quantification over variables, not only in axioms but also in con-
jectures and, also, there can be disjunctions. CL can also be considered as an extension of RL, but
in contrast to the resolution-based proving, the CL conjecture can be proved directly and unchanged
(refutation, Skolemization and transformation to clausal form are not used). The domain of pro-
cedures for CL actually covers first-order logic because every first-order theory can be translated
into coherent logic, possibly with additional predicate symbols [185, 85]. Checking validity of an
arbitrary first-order formula can be replaced by checking unsatisfiability of a corresponding set of
clauses (after refutation, Skolemization and transformation to clausal form).
Provability in CL and unsatisfiability in RL are semi-decidable and there is a number of methods
and provers for coherent logic (some of them are based on simple forward reasoning and iterative
deepening, with a number of techniques for narrowing the search space [15, 208, 209], while some
use more advanced techniques, like lemma learning and back-jumping [167, 169]) and much more
for RL [196]. CL admits a simple sequent-calculus style proof system, and any corresponding CL
proof has a simple structure [208]. Readable proofs in a forward reasoning style can be easily
obtained in CL [15]. The existing theorem proving methods for CL and RL do not deal with NDG
conditions but can use the same heuristics as for the DD method (Section 2.2.3.2).
Computer-Assisted Theorem Proving in Synthetic Geometry 27

CL provers have been used in a variety of settings and domains. In particular, they have been
applied to Euclidean geometry using an axiom system similar to Borsuk’s [18, 129] and to Hilbert’s
and Tarski’s axiomatics, with proofs exported to Isabelle, Coq, and natural language [208, 209].
They have also been used to prove the correctness of solutions to ruler and compass construction
problems [153]. They have been used for projective plane geometry, where a proof of Hessenberg’s
theorem was carried out with Coq proof objects generated [16]. Recent work has also seen them
combined with resolution theorem provers such as Vampire (for filtering relevant axioms) [210].

Example. The following theorem from Tarski’s geometry [201] can be proved in CL (using the
theorem prover ArgoCLP [209]): Assuming that A−B−C, (A, B) ∼ = (A, D), and (C, B) ∼
= (C, D), show
that B = D. The presented proof is obtained by simplifying and transforming the generated proof,
so it reintroduces negation and uses reductio ad absurdum [152].
1. It holds that B−A−A (by th 3 1).
2. From the facts that A−B−C, it holds that Col C A B (by ax 4 10 3).
3. From the facts that (A, B) ∼= (A, D), it holds that (A, D) ∼
= (A, B) (by th 2 2).
4. It holds that A = B or A 6= B (by ax g1).
5. Assume that A = B.
6. From the facts that (A, D) ∼= (A, B) and A = B it holds that (A, D) ∼= (A, A).
7. From the facts that (A, D) ∼= (A, A), it holds that A = D (by ax 3).
8. From the facts that A = B and A = D it holds that B = D.
This proves the conjecture.
9. Assume that A 6= B.
Let us prove that A 6= C by reductio ad absurdum.
10. Assume that A = C.
11. From the facts that A−B−C and A = C it holds that A−B−A.
12. From the facts that A−B−A, and B−A−A, it holds that A = B (by th 3 4).
13. From the facts that A 6= B, and A = B we get a contradiction.
Contradiction.
Therefore, it holds that A 6= C.
14. From the fact that A 6= C, it holds that C 6= A (by the equality axioms).
15. From the facts that C 6= A, Col C A B, (C, B) ∼ = (C, D), and (A, B) ∼= (A, D), it holds that B = D (by th 4 18).
This proves the conjecture.
Quaife used the resolution theorem prover OTTER to prove several non-trivial theorems in
Tarski’s geometry [201], with a slightly modified axiom system [189]. During theorem proving, a
number of techniques were employed to guide resolution and, upon success, some post-processing
used to translate the resolution proofs into a more readable form. More recently, Beeson and Wos
carried out some similar work, but with a newer version of OTTER and with much more success
thanks to a number of techniques and strategies that have become available in the meantime [10, 11].
They proved around 200 theorems from the book by Schwabhäuser et al. [201]. Of these theorems,
76% were proved automatically using different custom heuristics and strategies, while for the others
heavy human support (in a form of lemmas and hints) was required. This latest work did not involve
the production of readable proofs from the resolution ones. Even more recently, other resolution
provers with state-of-the-art techniques have led to an even higher percentage of theorems from the
same corpus being proved completely automatically, without any guidance by humans [216].

Example. The following proof, slightly reformulated for the sake of uniformity, is generated by
Quaife’s approach: if C is between B and D, each of which is between A and E, then C is between
A and E.
28 Handbook of Geometric Constraint Systems Principles

37 U−V−W ⇒ W−V−U (by Axiom)

45 U−V−X,V−W−X ⇒ U−W−X (by Lemma)
46 U−V−W,U−W−X ⇒ U−V−X (by Lemma)
74 U−V−X,U−W−X ⇒ U−V−W,U−W−V (by Lemma)
77 A−B−E (by Hyp)
78 B−C−D (by Hyp)
79 A−D−E (by Hyp)
80 ¬A−C−E (by negated goal)
91 ¬A−C−D (by 80, 46, 79)
92 ¬A−C−B (by 80, 46, 77)
109 ¬A−B−D (by 91, 45, 78)
127 ¬B−C−A (by 92, 37)
184 A−D−B (by 109, 74, 79, 77)
253 ¬B−D−A (by 127, 46, 78)
309 Contradiction! (by 184, 37, 253)

2.2.4 Semisynthetic Methods

Semisynthetic methods, sometimes also called coordinate-free methods or geometric invariant
methods, do not use algebraic formulation of geometry problems, but express conjectures in terms
of certain geometric quantities and prove them by manipulating equalities over expressions in these
quantities. This approach can also lead to combinatorial explosion, but in many cases can give short
and readable proofs.

2.2.4.1 Area Method

The area method is a procedure for a fragment of Euclidean plane geometry [50, 51, 130, 229].
It uses suitably chosen geometry quantities, such as area of triangle, and can efficiently prove
many non-trivial theorems and produces proofs that are often very concise and human-readable.
The method had been extended to solid Euclidean geometry [52], to non-Euclidean geometries
[225, 226] and, in conjunction with Collins algorithm [63], to a system for proving geometry in-
equalities [197].

Scope. A conjecture consists of a construction and a goal, where the construction is expressed in
terms of (five basic) specific construction primitives (or constructions composed of the primitive
ones), and the goal is an equality over expressions given in terms of (three basic) specific primitive
geometry quantities. Both the construction and the goal are expressed only in terms of points (i.e.,
cannot involve lines or circles explicitly).
An example of a construction primitive is I NTER Y U V P Q, which indicates that point Y is the
intersection of lines UV and PQ. For a construction step to be well-defined, certain NDG conditions
may be required. The above construction step has a NDG condition U 6= V ∧ P 6= Q ∧ UV ∦ PQ.
Intersections of two circles and intersections of a line and a circle are supported by construction
primitives only in some special cases. Additional construction steps can be expressed in terms of
the basic ones.
The geometric quantities used are: the signed ratio of parallel directed segments, denoted
AB
CD
, the signed area for a triangle ABC; denoted SABC (negated for the triangle with the op-
posite orientation); the Pythagoras difference, denoted PABC (for the points A, B, C, defined as
2 2 2
PABC = AB +CB − AC ). Using these quantities, a number of geometric predicates can be simply
expressed, for instance: A = B iff PABA = 0; Col A BC iff SABC = 0; AB ⊥ CD iff PABA 6= 0 ∧ PCDC 6=
0 ∧ PACD = PBCD ; AB k CD iff PABA 6= 0 ∧ PCDC 6= 0 ∧ SACD = SBCD , etc.
The method implemented by its authors proved 500 theorems from their collection [51].
Computer-Assisted Theorem Proving in Synthetic Geometry 29

Theorem proving mechanism. The method works by the elimination of constructed points in
reverse order, using a set of specific elimination lemmas. ‡ All the lemmas used by the method can
be proved by an elegant, custom axiom system (Section 2.3.1.4).
There is an elimination lemma for each pair of construction step and geometric quantity. For
instance, the following lemma is used for eliminating a point constructed on a line at a given ratio
from a signed area:
PY
If Y is a point constructed on line PQ, such that PQ = λ then for any points A and B

SABY = λ SABQ + (1 − λ )SABP .

The combined NDG conditions of the conjecture is the conjunction of those for the correspond-
ing construction steps, of the conditions that the denominators of the ratios of parallel directed
segments in the goal equality are not equal to zero, and of the conditions that lines appearing in
ratios of segments in the goal are parallel. It is then proved that the goal equality follows from the
construction specification and the combined NDG conditions. §
Apart from the basic NDG conditions, there are also side con- D
K
ditions in some of the elimination lemmas having two cases — pos- C
itive (always of the form “A is on PQ”) and negative (always of the
form “A is not on PQ”). If one side condition can be proved, then
L J
that case is applied. Otherwise, in one variation of the method, the
proof process branches into two cases, and in another, the negative
case is assumed and added to the NDG conditions [130]. B
I
A
Example. As an example, we give the proof of Varignon’s theo-
Figure 2.3
rem: Given a quadrilateral ABCD, let I, J, K and L be the midpoints
Varignon’s theorem.
of AB, BC, CD, DA, then IJKL is a parallelogram. We give below
the proof that IJ k KL, the proof that JK k IL is similar. Note that
a synthetic proof within Coq is given in Figure 2.9.
SKIJ − SLIJ
SKIC SLIC
= SKIB 2 + 2 − 2 − 2
SLIB
J Eliminated
SBKA SBKB SCKA SCKB SBLA
= 2 + 2 + 2 + 2 − 2 − I Eliminated
SBLB SCLA SCLB
2 − 2 − 2
1
= 2 (SBKA + SCKA + SCKB − SBLA − SCLA − Simplification
SCLB )
= 12 ( SABC SABD SACC SACD
2 + 2 + 2 + 2 + 2 + K Eliminated
SBCC
SBCD
2 − SBLA − SCLA − SCLB )
= 12 ( SABC SABD SACC SACD
2 + 2 + 2 + 2 + 2 + L Eliminated
SBCC
SBCD
2 − SABA2 − SABD
2 − SACA
2 − SACD
2 −
SBCA SBCD
2 − 2 )
= 14 (SABC + SBCA ) Simplification
= 0 Simplification

Properties. The method is terminating, sound, and complete: for each geometric statement in its
scope, it can decide whether it is a theorem, i.e., it is a decision procedure for this fragment of
geometry. Its complexity is exponential in the number of points involved [229].
‡ A later variant of the method also deals with nonconstructive statements, described in terms of various geometric predicates

[53].
§ If the negation of some NDG condition of a geometric statement is implied by the remaining construction steps, the

left-hand side of the implication is inconsistent and the statement is trivially valid.
30 Handbook of Geometric Constraint Systems Principles

2.2.4.2 Full-Angle Method

The full-angle method [47] is, in spirit, closely related to the area method (Section 2.2.4.1) and can
also produce elegant proofs for a number of complex theorems. The idea of eliminating points is
extended to eliminating lines. The main motivation of the full-angle method is the fact that using
“traditional angles” in geometrical proofs typically leads to considering a number of cases. For
instance, for four distinct cyclic points A, B, C, D, one can claim that the angles ∠ABC and ∠ADC
are congruent if B and D are on the same side of line AC or complementary if they are on opposite
sides (Figure 2.4). On the other hand, with full-angles one can simply (without using order relation
or orientations of plane) state ∠[AD,CD] = ∠[AB,CB]. Namely, it holds ∠[AB, BC] = ∠[DE, EF]
iff ∠ABC ∼ = ∠DEF and the two angles have the same orientation or ∠ABC = 180◦ − ∠DEF and the
two angles have opposite orientations.
A full-angle is defined to be an ordered pair ∠[m, n] B
of two intersecting lines m and n, such that ∠[m, n] is D
equal to another full-angle ∠[u, v] if there is a rotation
R such that R(m) k u and R(n) k v (therefore, any full-
angle can be considered as an equivalence class) [51].
The sum of two full-angles is defined as follows: given A C
four lines m, n, u, and v, and a rotation R such that
R(u) k n, then ∠[m, n] + ∠[u, v] = ∠[m, R(v)]. For arbi- D
trary line m, ∠[m, m] is denoted by 0. For arbitrary lines
Figure 2.4
m and n, ∠[m, n] can be denoted also by −∠[n, m]. It can
Cyclic points and peripheral angles.
be proved that full-angles form an Abelian group with the
operation +, the neutral element 0, and with inverse element corresponding to the (unary) operator
−. In addition, ∠[m, n] + −∠[u, v] is abbreviated by ∠[m, n] − ∠[u, v], and for arbitrary perpendicular
lines m and n, ∠[m, n] is denoted by 1.
It can be proved that full-angles satisfy around 20 properties useful for transforming goals,
including the following ones (where, for each full-angle ∠[AB,CD], it is assumed that A 6= B and
C 6= D):
R4: 1 + 1 = 0
R6: if Col P Q X then ∠[AB, PX] = ∠[AB, PQ]
R10: if cyclic(A, B,C, D) then ∠[AD,CD] = ∠[AB,CB]
R13: ∠[AB,CD] = −∠[CD, AB]
R14: for any line UV , ∠[AB,CD] = ∠[AB,UV ] + ∠[UV,CD]

Scope. The full-angle method deals with conjectures consisting of hypotheses, expressed in terms
of relevant construction steps (cf. the area method in Section 2.2.4.1) or in terms of other geometric
predicates (as in the later variation of the area method), and of a goal that is an equality over full-
angles.

Theorem proving mechanism. The proof method uses forward chaining for exhaustively deduc-
ing new facts from the existing ones, using lemmas (rules) like:
F1: if m k n and m k l, then n k l
F5: if PA ⊥ PB then QA ⊥ QB iff cyclic(A, B, P, Q)
F8: if AB k AC, then Col A BC
K2: if m ⊥ n and u ⊥ v, then ∠[m, u] = ∠[n, v]
Some rules have NDG conditions attached and they can be treated as for the area method (see
Section 2.2.4.1).
Using derived facts, rules like the ones listed above are used for elimination of points (R6) or
lines (R10) from the goal – this is again analogous to the area method although the expressions are
now simpler since there are no multiplications or divisions over full-angles.
Computer-Assisted Theorem Proving in Synthetic Geometry 31

The configuration is not necessarily expressed in terms of constructive statements and so (unlike
in the basic version of the area method) there is no implicit order in which points can be eliminated.
So, an ordering has to be imposed over points, which extends to full-angles, in order to control
the application of the rules. For instance, the rule: For any line UV , ∠[AB,CD] = ∠[AB,UV ] +
∠[UV,CD] is used only if the two new full-angles can be further reduced to full-angles less than
∠[AB,CD] (in the ordering).

Example. For Simson’s theorem (Figure 2.5) the hypothe-

D E
ses are cyclic(A, B,C, D), E is the foot from D to BC (i.e., C
Col BC E and DE ⊥ BC), F is the foot from D to AC (i.e.,
F
Col AC F and DF ⊥ AC), G is the foot from D to AB (i.e.,
Col A B G and DG ⊥ AB), and the goal is that G, F, E are
collinear, i.e., ∠[GF, GE] = 0.
B
The proof assumes that for each full-angle ∠[XY,UV ] used A G
it holds that X 6= U and Y 6= V . The following point order
is used O, A, B,C, D, E, F, G, and the following facts (among
Figure 2.5
others) can be derived from the hypotheses: cyclic(F, A, D, G)
Simson’s theorem.
(because FA ⊥ FD, GA ⊥ GD), cyclic(E, B, D, G) (because
EB ⊥ ED, GB ⊥ GD). In the following proof outline, applications of rules related to the symmetry
properties of relations are not shown.

∠[GF, GE] =
= ∠[GF, GD] + ∠[GD, GE] (by R14)
= ∠[AF, AD] + ∠[GD, GE] (by R10, cyclic(F, A, D, G))
= ∠[AF, AD] + ∠[BD, BE] (by R10, cyclic(E, B, D, G))
= ∠[AF, AD] − ∠[BE, BD] (by R13)
= ∠[AC, AD] − ∠[BE, BD] (by R6, Col A F C)
= ∠[AC, AD] − ∠[BC, BD] (by R6, Col BC E)
= ∠[AC, AD] − ∠[AC, AD] (by R10, cyclic(A, B,C, D))
= 0

Properties. The method is not complete, but can be used as a complement to the area method.
When applied to a conjecture in its scope, if it succeeds, the generated proof is typically short and
readable. Otherwise, the goal is transformed into a goal for the area method: an equality α = β is
transformed ¶ into tan(α) = tan(β ) and then further using the following equations (where PABCD =
PABD − PCBD ):

tan(∠[AB,CD]) + tan(∠[PQ,UV ])
tan(∠[AB,CD] + ∠[PQ,UV ]) =
1 − tan(∠[AB,CD]) tan(∠[PQ,UV ])

4SACBD
tan(∠[AB,CD]) =
PADBC
Since the area method is complete, the above gives a decision procedure for formulae belonging
to the scope of the full-angle method [47].

2.2.4.3 Vector-Based Method

The idea of using vectors for automating geometric proofs has been proposed by several authors, but
probably the most important work in the area is due to Chou, Gao, and Zhang [49]. Their method
¶ The 4SACBD
function tan for the full-angle (corresponding to the usual trigonometric function), tan(∠[AB,CD]) = PADBC
, is
well-defined, thanks to the fact that ∠[AB,CD] = ∠[PQ,UV ] iff SACBD PPUQV = SPUQV PACBD .
32 Handbook of Geometric Constraint Systems Principles

is, in spirit, close to the area method (Section 2.2.4.1), in the way the hypotheses are described
constructively and the constructed points are eliminated from the goal one by one using appropriate
lemmas.

Scope. The hypotheses are expressed in terms of (four) specific construction primitives. For in-
−→ −→
stance, P RATIO A W U V r denotes the construction of a point A such that WA = rUV , where r
is a rational number, an expression over geometric quantities, or a parameter (the NDG condition
is U 6= V ). Additional construction steps can be suitably expressed in terms of the basic ones. For
instance, the construction of the midpoint, denoted M IDPOINT M A B, can be expressed as P RATIO
M A A B 1/2.
−
→ −→
A goal is either an equality over vectors or an equality involving the inner products (hAB, CDi)
−→ −→
and exterior products ([AB, CD]) of vectors over constructed points.
There are two kinds of NDG conditions: those induced by the construction steps and those
necessary for the goal to be defined (denominators are not zero). Then, the conjecture is changed by
augmenting the hypotheses with these NDG conditions.

Theorem proving mechanism. The method works by the elimination of constructed points in
reverse order, using a set of specific elimination lemmas, like for the area method. There are elim-
ination lemmas for each pair (construction step, geometry quantity). For instance, the following
lemma is used for the elimination of a point Y constructed by the P RATIO step from the linear
quantity G(Y ) satisfying G(αY1 + βY2 ) = αG(Y1 ) + β G(Y2 ), for any real numbers α and β :
If Y is introduced by P RATIO Y W U V r, then G(Y ) = G(W ) + r(G(V ) − G(U)).

Properties. The method is terminating, sound, and complete: for each geometry statement in its
domain, it can decide whether it is a theorem, i.e., the method is a decision procedure for its fragment
of geometry. The complexity of the method is exponential in the number of involved points [49].

2.2.4.4 Mass-Point Method

Barycentric coordinates and mass points have been used in geometry at least since 1969 [68], and
were introduced in automated theorem proving for geometry by Zou and Zhang [230]. In the non-
complex case, the method is similar to the method used by Kimberling for studying triangle centers
(Section 2.3.2). A mass point is mP, where m (“mass”) is a positive real number, and P is a point in
a plane. Two mass points mP and nQ are equal iff m = n and P = Q.

Scope. The conjecture consists of hypotheses (in the form of a construction) and a goal. Hypothe-
ses are expressed in terms of three free (arbitrary) points and subsequent points are obtained by
five basic geometric constructions and some compound ones (that enable a constructed point to be
expressed as a linear combination of the three basis points), including:
−→ −→
C3 L RATIO X A B r, that gives a point X on the line AB such that AX = rAB, where r is a rational
number, a rational expression, or a variable. Specially, M IDPOINT X A B denotes L RATIO X A B
−→ −→
1/2. Constructing a point X such that AX = rXB (or (1 + r)X = A + rB) is denoted by M RATIO X
A B r.
C5 I NTER X U V A B, that gives the intersection point X of lines UV and AB (the NDG condition
is that X is not equal to some of the points U, V , A, B, and that UV and AB are not parallel; otherwise,
the prover fails).
The goal is a predicate over constructed points, one from a set that includes, for instance,
Col A BC. For this predicate, it can be proved [228]: if P, Q, and R are points of the plane ABC (A, B,
C are noncollinear points), and P = a p A + b p B + c pC, Q = aq A + bq B + cqC, R = ar A + br B + crC,
Computer-Assisted Theorem Proving in Synthetic Geometry 33

then P, Q and R are collinear iff  

 ap bp cp 
 
 aq bq cq  = 0.
 
 ar br cr 
The method is extended to deal with additional constructions (such as a construction of a circle)
and uses complex numbers for convenience. This extended version has a scope strictly wider than
the basic version.
The authors of the implementation successfully used it for proving hundreds of nontrivial theo-
rems. Although the generated proofs are understandable, they are still not human-like proofs.

Theorem proving mechanism. The mass point method works by Q

expressing all constructed points as a linear combination, of three B1
C1
(or two, for some simple conjectures) free points, then reformulat- P
ing the goal the same way and finally proving it as a goal over real
numbers. C
A1

Example. Desargues’s theorem (see also Section 2.3.2) states B

that, given two triangles 4ABC and 4A1 B1C1 , if the lines AA1 , S A R
BB1 , CC1 intersect in a point S, then the intersection points P of BC
Figure 2.6
and B1C1 , Q of CA and C1 A1 , and R of AB and A1 B1 are collinear
Desargues’s theorem.
(Figure 2.6). The theorem has to be slightly reformulated in order
to use the mass point method. The construction is as follows:
Let A, B, and C be three free points.
Let S be an arbitrary point of the plane ABC, hence S = aA + bB + cC for some real numbers a,
b, c, such that a + b + c = 1.
1 x 1 x
M RATIO A1 S A x, for some x (then it holds A1 = 1+x S + 1+x A = 1+x (aA + bB + cC) + 1+x A=
a+x b c
1+x A + 1+x B + 1+x C).
M RATIO B1 S B y, for some y (then it holds B1 = 1+y a
A + b+y c
1+y B + 1+y C).
a b c+z
M RATIO C1 S C z, for some z (then it holds C1 = 1+z A + 1+z B + 1+z C).
y
I NTER P B C B1 C1 (then it holds yB − zC = (1 + y)B1 − (1 + z)C1 = (y − z)P, i.e., P = y−z B−
z
y−z C).
x z
I NTER Q A C A1 C1 (then it holds Q = x−z A − x−z C).
x y
I NTER R A B A1 B1 (then it holds R = x−y A − x−y B).
The goal is to prove that P, Q, R are collinear, which is done by showing that:
y z
− y−z
 
 0 y−z

 x z


 x−z 0 − x−z  = 0
 x y
x−y − x−y 0 

Properties. The mass point method provides a decision procedure for conjectures within its
scope [230].

2.2.5 Provers Implementations and Repositories of Theorems

There are a number of tools, typically providing dynamic geometry functionalities, that have support
for the automated proof of geometry theorems. We mention the most notable ones next.
GEX/jGEX/MMP/Geometer is a family of systems equipped with provers based on algebraic
approaches, the DD method, the area method, the vector method, and the full-angle method [98,
227]. GeoGebra [19], also equipped with several algebraic-based provers and tools based on the area
34 Handbook of Geometric Constraint Systems Principles

method, is aimed at education. It can also work with the Coq proof assistant to support interactive
proofs [180]. GeoProof is another tool linked to Coq that can use provers based on the area method,
Wu’s method, and the Gröbner basis method for generating machine verifiable proofs [163]. GCLC
is a system that supports two algebraic methods and the area method [128]. Theorema [36], built
on top of Mathematica, is a general mathematical tool with support for several theorem proving
approaches, including the area method. OpenGeoProver is a library with several algebraic provers
and one based on the area method [151]. Geometry Explorer uses the full-angle method [223] and
provides means of visualizing geometric proofs as graphs.
There are ongoing efforts toward linking dynamic geometry systems with automated theorem
proving and also with automated discovery, intelligent management of geometry knowledge, tutor-
ing, eLearning, and so on [43, 142, 191, 219].
Finally, aside from collections of theorems available within the above tools, we note the exis-
tence of a dedicated repository of geometry theorems known as TGTP [190].

2.3 Interactive Theorem Proving

A proof assistant is a piece of software that can check mathematical assertions interactively. The
main ones that have been used for the formalization of geometry are Coq [14, 66, 67], Isabelle [168,
175, 221], HOL4 [207], HOL-Light [119], and Mizar [215, 222]. They differ in their mathematical
foundations (e.g., type theory, higher order logic [HOL], or set theory) and their proof language. In
procedural style proof assistants (e.g., Coq and HOL Light), proofs are described as sequences of
commands that modify the proof state whereas in proof assistants that use a declarative language
(e.g., Mizar and Isabelle), the proofs are structured and contain the intermediate assertions that were
given by the user and justified by the system.

2.3.1 Formalization of Foundations of Geometry

There are several ways in which the foundations of geometry can be laid.
In the synthetic approach, the geometry theory is built from axioms, with non-logical symbols
corresponding to geometric predicates, and sorts corresponding to geometric objects.
The best-known modern axiomatic systems along these lines are those of Hilbert [124] and
Tarski [214], which we will examine in detail next.
In the analytic approach, a field F is assumed (usually R, the reals), the space is defined as Fn ,
and the geometric objects and predicates are defined.
In the mixed analytic/synthetic approaches, one assumes both the existence of a field and also
some geometric axioms. For example, the axiomatic systems for geometry proposed for education
in North America by the School Mathematics Study Group in the 1960s are based on Birkhoff’s ax-
iomatic system [17] in which the underlying field (R) serves to measure distances and angles. This
approach, known as the metric approach, is developed in a number of modern sources [154, 159].
A similar approach is used by Chou, Gao, and Zhang for the foundations of the area method [229]
(Section 2.2.4.1), where the underlying field is used to express ratios of signed distances and ar-
eas. The axioms and properties of the area method have been formalized in Coq [130]. Geometry
can also be defined as a space of objects and a group of transformations acting on it (Erlangen
program [135]), and several axiom systems based on this approach have been proposed [86, 170].
Axiom systems based on intuitionistic logic have also been proposed for geometry. Von Plato,
for instance, uses the concept of apartness of points and convergence of lines to study plane geome-
try [183, 184]. Beeson, for his part, introduces a constructive version of Tarski’s axiom system [7, 8].
Computer-Assisted Theorem Proving in Synthetic Geometry 35

Axiomatics exist for a variety of other geometries (hyperbolic, elliptic, projective, etc.) that will
not be examined in this chapter. Details about such axiomatizations can be found in surveys by
Pambuccian [171, 172, 173], for instance. In what follows, we mainly concentrate on Hilbert and
Tarski’s axiomatic systems and their formalizations.

2.3.1.1 Hilbert’s Geometry

Hilbert’s axiom system Hilbert stated his axioms in natural language, but left various aspects
implicit and imprecise [124]. One formulation of the first four groups of planar axioms (the fifth
group of axioms is discussed in Section 2.3.1.3) in terms of first-order logic (FOL) is given next,
although there are other possible formalizations [39, 123, 199, 200].
Hilbert’s axiom system can be formulated in a three-sorted first-order language, with individual
variables interpreted as points, lines, and planes (the third sort is only needed when dealing with
space geometry). It involves four nonlogical predicate symbols: one for the incidence of a point on
a line (denoted by P ∈ l in what follows), one for strict betweenness (denoted by A−B · −C),
· one for
∼
the congruence of segments (denoted by AB =CD), and one for the congruence of angles (denoted
by ∠ABC ∼ = ∠DEF).
The axioms are stated by Hilbert using notions from set theory, but the use of set theory is
inessential and the formulation of Hilbert’s axioms given next does not rely on it (the use of ∈ is
merely as a syntactically convenient predicate).
The first group of Hilbert’s axioms deals with incidence: for any two points there is one and
only one line through the two points; on every line there are two points; there is a line and point not
incident to it (or alternatively: there are three noncollinear points). The axioms, given in FOL form,
are as follows:
Incidence Axioms
I 1 A 6= B ⇒ ∃l A ∈ l ∧ B ∈ l
I 2 A 6= B ∧ A ∈ l ∧ B ∈ l ∧ A ∈ m ∧ B ∈ m ⇒ l = m
I 3 ∃AB B ∈ l ∧ A ∈ l ∧ A 6= B
I 4 ∃P0 l0 ¬ P0 ∈ l0
The second group of axioms provides properties of betweenness. The betweenness predicate of
Hilbert is strict (as opposed to that of Tarski) as ensured by the first axiom (II 1). The fourth axiom
is the axiom of Pasch: if a line intersects one side of a triangle, then it also intersects one of the two
other sides.
Order axioms
II 1 A−B· −C· ⇒ Col A BC
·
A−B−C · ⇒ A 6= C
· −C
A−B · ⇒ C−B · −A
·
II 2 A 6= B ⇒ ∃C A−B · −C
·
II 3 A−B· −C· ⇒ ¬B−C · −A
·
Def cut l A B := ¬A ∈ l ∧ ¬B ∈ l ∧ ∃I (I ∈ l ∧ A−I · −B)
·
II 4 ¬ Col A BC ∧ ¬C ∈ l ∧ cut lAB ⇒ (cut lAC ∨ cut lBC)
The third group of axioms provides congruence properties. The aspects that are implicit in
Hilbert’s text or are consequences of the set theoretical formulation of the axioms are suffixed with
lowercase letters. For example, in Hilbert’s text, the segment CD is the same as segment DC, a
fact captured by axiom IV-a in our case. Axiom IV 2 is the transitivity property of the congruence
predicate. Axiom IV 3 is the segment addition axiom. Axiom IV 4 is the angle construction axiom:
−→ −→
given an angle ∠BAC and a side of a half line OX, there is a unique half line OY such that the
−→ −→
angle between OX and OY is ∠BAC. Axiom IV 5 is the side-angle-side property (two triangles are
congruent if two pairs of their corresponding sides and angles determined by them are congruent).
36 Handbook of Geometric Constraint Systems Principles

Congruence Axioms
IV-a AB ∼=CD ⇒ AB ∼ = DC
IV 1 A 6= B ∧ M ∈ l ⇒ ∃A0 B0 (A0 ∈ l ∧ B0 ∈ l ∧ A0−M · 0 ∧ MA0 ∼
· −B =
0 ∼
AB ∧ MB = AB)
IV 2 AB ∼=CD ∧ AB ∼ = EF ⇒ CD ∼ = EF
Def disjoint A B C D := ¬∃P (A−P · −B
· ∧C−P · −D)
·
IV 3 Col A BC ∧Col A B C ∧disjoint ABBC ∧disjoint A0 B0 B0C0 ∧
0 0 0

AB ∼= A0 B0 ∧ BC ∼= B0C0 ⇒ AC ∼= A0C0
IV-b ∼
¬ Col A BC ⇒ ∠ABC = ∠ABC
IV-c ¬ Col A BC ⇒ ∠ABC ∼ = ∠CBA
IV-d ∠ABC ∼ = ∠DEF ⇒ ∠CBA ∼ = ∠FED
Def same side A B l := ∃P cut lAP ∧ cut lBP
Def same side’ A B X Y := X 6= Y ∧ ∀l X ∈ l ∧ Y ∈ l ⇒
same side ABl
−
→
Def B ∈ PA := P−A · −B
· ∨ P−B
· −A· ∨ (P 6= A ∧ A = B)
∼ −
→ 0 − → −→ −→
IV-e ∠ABC = ∠DEF ∧A ∈ BA∧C ∈ BC ∧D0 ∈ ED∧F 0 ∈ EF ⇒
0

∠A0 BC0 ∼= ∠D0 EF 0

IV 4.1 ¬ Col P O X ∧ ¬ Col A BC ⇒ ∃Y (∠ABC ∼ = ∠XOY ∧
same side’ PY OX)
IV 4.2 ¬ Col P O X ∧ ¬ Col A BC ∧ ∠ABC ∼ = ∠XOY ∧ ∠ABC−→ ∼
=
0 0
∠XOY ∧ same side’ PY OX ∧ same side’ PY OX ⇒ Y ∈ OY 0

IV 5 ¬ Col A BC ∧ ¬ Col A0 B0 C0 ∧ AB ∼ = A0 B0 ∧ AC ∼ = A0C0 ∧

∼ 0 0 0
∠BAC = ∠B A C ⇒ ∠ABC = ∠A B C∼ 0 0 0

The fourth group of axioms consists of a single axiom, Playfair’s version of Euclid’s fifth pos-
tulate which states that there is at most one parallel line to a given line through a given point.

Parallels Axiom
Def l k m := ¬ ∃X (X ∈ l ∧ X ∈ m)
IV I ∀lPm1 m2 ¬P ∈ l ∧ l k m1 ∧ P ∈ m1 ∧ l k m2 ∧ P ∈ m2 ⇒ m1 = m2

Formalization of Hilbert’s Gründlagen der Geometrie Dehlinger, Dufourd, and Schreck have
studied the formalization of Hilbert’s foundations of geometry in the intuitionist setting of Coq [71].
In this context, they highlighted that Hilbert’s proofs require the excluded middle axiom applied to
the equality between points. They focus on the first two groups of axioms and prove some be-
tweenness properties. Meikle and Fleuriot formalized the first three groups of axioms in their three-
dimensional version within the Isabelle/HOL proof assistant [157]. They went up to the twelfth
theorem of Hilbert’s book k . Scott continued that formalization and revised it [202]. He corrected
some “subtle errors in the formalization of Group III” such as the definition of a point being inside
an angle, which led to the set of points inside an angle being always empty. But Scott’s definition,
which states that a point P is inside an angle if there are two points A and B on the sides of the angle
such that P belongs to segment AB, is also flawed in the sense that it is only valid in Euclidean geom-
etry since this property is equivalent to the parallel postulate [23]. A more general definition is used
by Schwabäuser et al. [201] (see Section 2.3.1.1). In further work, this time using HOL Light, Scott
and Fleuriot developed a discovery system that attempted to fill gaps in Hilbert’s incidence-related
proofs automatically [203].
Independently, Richter also formalized a substantial number of results based on Hilbert’s axioms
and a metric axiom system within HOL Light [193] while Braun et al. have formalized the fragment
of [124] necessary to prove Tarski’s axioms [32].
k We use the numbering of theorems from the tenth edition.
Computer-Assisted Theorem Proving in Synthetic Geometry 37

2.3.1.2 Tarski’s Geometry

Szmielew provided a rigorous development of geometry based on Tarski’s axioms, which was later
extended by Schwabäuser et al. [201]. The proofs presented there include the important and non-
trivial proofs obtained by Gupta that the midpoint can be constructed without any continuity ax-
iom [113]. Tarski’s axiom system has evolved over time and a comprehensive description of its
history can be found elsewhere [214].

Tarski’s axiom system Tarski’s axiom system can be formulated in an unsorted first-order lan-
guage (with all individual variables interpreted as points). It involves only two nonlogical predicate
symbols: one for nonstrict betweenness (denoted by A−B−C) and one for congruence (denoted by
AB ∼
=CD). The axioms are formulated as follows:

A1 Symmetry AB ∼
= BA
A2 Pseudo-Transitivity AB ∼
=CD ∧ AB ∼ = EF ⇒ CD ∼ = EF
A3 Cong Identity ∼
AB =CC ⇒ A = B
A4 Segment construction ∃E A−B−E ∧ BE ∼ =CD
A5 Five-segment AB ∼
= A0 B0 ∧ BC ∼= B0C0 ∧
AD ∼
= A0 D0 ∧ BD ∼= B0 D0 ∧
A−B−C ∧ A −B −C0 ∧ A 6= B ⇒ CD ∼
0 0
=C0 D0
A6 Between Identity A−B−A ⇒ A = B
A7 Inner Pasch A−P−C ∧ B−Q−C ⇒ ∃X P−X−B ∧ Q−X−A
A8 Lower Dimension ∃ABC ¬A−B−C ∧ ¬B−C−A ∧ ¬C−A−B
A9 Upper Dimension AP ∼
= AQ ∧ BP ∼= BQ ∧CP ∼ =CQ ∧ P 6= Q
⇒ A−B−C ∨ B−C−A ∨C−A−B.
A10 Parallel postulate ∃XY (A−D−T ∧ B−D−C ∧ A 6= D ⇒
A−B−X ∧ A−C−Y ∧ X−T−Y )

The symmetry and transitivity axioms, i.e., A1 and A2 , imply that equidistance is an equiva-
lence relation. The identity axiom for equidistance (A3 ) ensures that only degenerate line segments
can be congruent to a degenerate line segment. The axiom of segment construction (A4 ) allows
the extension of a line segment by a given length. The five-segments axiom (A5 ) is similar to the
side-angle-side property, but expressed without mentioning angles, using only the betweenness and
congruence relations (Figure 2.7(a)). The lengths of AB, AD, and BD determine the angle ∠CBD.
The identity axiom for betweenness expresses that the only possibility to have B between A and A
is to have A and B equal. The inner form of Pasch’s axiom (A7 , Figure 2.7(b)) is a variant of the
axiom that Pasch introduced to repair the defects of Euclid [174]. This version of Pasch’s axiom was
introduced by Peano [176]. Inner Pasch’s axiom is a form of the axiom that holds even in space, i.e.,
it does not assume a dimension axiom. The lower 2-dimensional axiom (A8 ) asserts the existence
of three noncollinear points. The upper 2-dimensional axiom (A9 ) implies that all the points are
coplanar. The version of the parallel postulate (A10 ) is a statement which can be expressed easily
in the language of Tarski’s geometry (Figure 2.7(c)). It is equivalent to the uniqueness of parallels
(Playfair’s postulate) that Hilbert uses or to Euclid’s fifth postulate.

Formalization of Metamathematische Methoden in Der Geometrie Narboux formalized the

first eight chapters of Schwabhäuser et al.’s axiomatic development using the Coq proof assis-
tant [162, 164]. Braun and Narboux extended this formalization up to the twelfth chapter and pro-
vided the formal proof that every model of Tarski’s axiom A1 –A10 is a model of Hilbert’s first four
groups of axioms [32]. Makarios formally verified that the Cartesian plane over R is a model of
Tarski’s axioms using the Isabelle/HOL proof assistant. He also proved that the Klein-Beltrami
plane is a model of Tarski’s axioms (except for the parallel postulate), thereby demonstrating the
independence of the parallel postulate [149]. Furthermore, Makarios proposed a simplification to
38 Handbook of Geometric Constraint Systems Principles

D D′ C

P B C
Q D
X
A

A B C A′ B′ C′ X T Y
B
(a) Five segment (b) Inner Pasch (c) Tarski’s parallel postulate

Figure 2.7
Illustration for some of Tarski’s axioms.

Tarski’s axioms: by changing the order of the points in the five-segment axiom, the pseudo reflex-
ivity of congruence was formally derived from the other axioms [149]. In Coq, Braun and Narboux
formalized two synthetic proofs of Pappus’s theorem [33]: the usual version that assumes the par-
allel postulate and another one in neutral geometry (i.e., without assuming the parallel postulate).
The proofs are formalized versions of those presented by Hilbert [124] and described in detail by
Schwabhäuser et al. [201]. Having a synthetic proof is necessary because Pappus’s theorem is used
in the arithmetization of geometry that justifies analytic proofs. Boutry et al. have studied Tarski’s
geometry in the context of intuitionist logic (without the excluded middle axiom) but assuming the
decidability of equality of points. They showed that the decidability of equality of points is equiv-
alent to that of congruence and betweenness. They also proved that the decidability of the other
predicates, with the exception of line intersection, can be derived from that of equality [25]. Richter
et al. [108, 194] ported some of the earlier Coq proofs by Narboux to Mizar and then extended the
formalization. They proved that the real line is a model of Tarskis axioms A1 –A7 and formalized
about 50 theorems by Schwabhäuser et al. [201]. Also in Mizar, Coghetto and Grabowski proved
that the Cartesian plane over R is a model of Tarski’s axiom [62], as was done by Harrison using
HOL-Light.
In Isabelle/HOL, Marić and Petrović proved that the Cartesian plane over R is a model of both
Tarski’s axioms and Hilbert’s axioms [177, 151]. Boutry and Cohen demonstrated that Cartesian
planes over Pythagorean ordered fields (i.e., ones in which sums of squares are squares) are models
of Tarski’s axioms A1 –A10 [20]. Moreover, Boutry et al. formalized the arithmetization of Tarski’s
geometry, i.e., that from any model of its planar axioms, it is possible to define a Pythagorean
ordered field and associate pairs of coordinates to points [21]. This is the culminating result of both
the development by Schwabhäuser et al. [201] and Hilbert’s book [124]. Boutry et al. also proved the
usual algebraic characterization of geometric predicates, which allows the use of algebraic methods
for automatic theorem proving in a synthetic setting.

2.3.1.3 Axiom Systems and Continuity Properties

Although the synthetic and analytic approaches to geometry seem quite different, Descartes proved
that the second one can be derived from the first one. By defining addition, multiplication, and
square roots geometrically, he introduced the arithmetization and coordinatization of geometry [74].
However, despite this analytic recasting, there are geometries that are too weak for arithmetization
and, moreover, not all geometries can be reduced to algebra. Thus, having a synthetic approach to
both the foundations of geometry and to automated reasoning in geometry is still vital.
Hilbert’s axioms (Groups I–IV) as presented in Section 2.3.1.1 are bi-interpretable with Tarski’s
axioms A1 –A10 . The claim that there exists at least one line parallel to l through P is not part of
these axioms: it can be derived from the other axioms. This implies that a different set of axioms is
needed for the foundations of elliptic geometry.
Computer-Assisted Theorem Proving in Synthetic Geometry 39

Neutral geometry is the set of theorems which are valid both in hyperbolic and Euclidean ge-
ometry, it can defined by the axioms A1 –A9 of Tarski or Hilbert’s Groups I–III.
Our earlier presentation of Tarski and Hilbert’s axiomatics did not include the continuity axioms
as a large part of geometry can be built without them. We now describe some continuity proper-
ties starting with a class (Figure 2.8) that states the existence of a point of intersection between
circles/lines/segments. The following three properties are equivalent assuming Tarski’s axioms A1 –
A9 [122, 211]:

Circle–Circle continuity Given two circles C1 and C2 , if there is a point on C1 which is inside C2
and vice-versa, then there is a point of intersection of the two circles.

Line–Circle continuity Given a line l which contains a point inside the circle C , there is a point
on l which is also on C .
Segment–Circle continuity Given a segment PQ and a circle C , if P is inside C , and Q outside
then there is a point of the segment PQ which is on C .

(a) Circle–Circle (b) Line–Circle (c) Segment–Circle

Figure 2.8
Elementary continuity.
M is a model of A1 –A10 , if, as mentioned in the previous section, it is isomorphic to Cartesian
planes over Pythagorean fields. By adding the Circle–Circle continuity axiom, an axiomatization
of ruler and compass geometry is obtained. M is a model of A1 –A10 plus Circle–Circle if M is
isomorphic to the Cartesian plane over a Euclidean field (positives are squares).
The continuity property (A11 ) introduced by Tarski is a first-order axiom schema which restricts
the Dedekind cut axiom to first-order definable cuts:
First-Order Dedekind’s Schema A11

(∃a ∀xy (X ∧Y ⇒ a−x−y)) ⇒ (∃b ∀xy (X ∧Y ⇒ x−b−y))

where X and Y stand for any first-order formula in the language of Tarski’s axiom system
which does not contain any free occurrence of a, b, x in X and a, b, y in Y .
Tarski showed [213] that M is a model of A1 –A11 if and only if it is isomorphic to the Cartesian
plane over some real closed field.
The full-continuity axiom is the second-order axiom.
Dedekind’s Axiom A∗11

∀XY (∃a ∀xy (x ∈ X ∧ y ∈ Y ⇒ a−x−y)) ⇒

(∃b ∀xy (x ∈ X ∧ y ∈ Y ⇒ x−b−y))
40 Handbook of Geometric Constraint Systems Principles

The theory A1 –A10 , A∗11 is categorical. M is a model of A1 –A10 ,A∗11 if and only if it is isomorphic
to the real Cartesian plane.
Hilbert has two continuity axioms. The first one is:
V 1: Archimedes’s Axiom Given two line segments AB and CD, there exists a number n such that
n copies of segment CD is greater than AB.
This can be derived from A∗11 but not from A11 as it is a second order property. If M is a model
of A1 –A10 in which the geometric version of Archimedes’s axiom holds then M is isomorphic to
the Cartesian plane over an Archimedean and Pythagorean field.
Hilbert’s second continuity axiom is a property about the models of the other axioms:
V 2: Line completeness An extension of a set of points on a line with its order and congruence
relations that would preserve the relations existing among the original elements as well as the
fundamental properties of line order and congruence that follows from Axioms I–III and from
V 1 is impossible.
Giovanni argues that Hilbert was aware of the Dedekind continuity axiom A∗11 and chose Axiom
V 2 because it does not imply Archimedes’s axiom [104]. The role of this axiom and its relationship
with Archimedes’s axiom is discussed in [88].

Formalization of continuity properties and ruler and compass geometry Duprat has proposed
the formalization of several axiom systems for Euclidean plane geometry in Coq. His aim was to ax-
iomatize a geometry where objects correspond to ruler and compass constructible ones [81, 82, 83].
He proved the corresponding axioms of Hilbert (except continuity) from his set of 20 axioms [84]
with two nonlogical predicate symbols: one for ternary left-turn and one for quaternary congruence.
Tarski’s and Hilbert’s systems can also be used to axiomatize ruler and compass geometry. Gries et
al. have formalized (as part of the GeoCoq library) the equivalence between different versions of the
Circle–Circle intersection axiom and also the equivalence between Circle–Line and Circle–Segment
axioms. They have shown that the Circle–Line intersection can be derived from the Circle–Circle
intersection and that the Circle–Circle intersection can be derived from A11 . This proof relies on the
nontrivial fact that the triangular inequality can be derived without the circle continuity axiom.

2.3.1.4 Other Axiom Systems and Geometries

Euclid’s geometry Avigad et al. have proposed a formal system for Euclid’s Elements [4]. This
includes Tarski’s axioms, Circle/Line continuity, and some other axioms related to the relative posi-
tion of points in the plane that can be derived from Tarski’s axioms. Beeson et al. have given another
formalization of Euclid’s axioms and produced proofs in a forward reasoning style for all the propo-
sitions of the first book of Euclid’s Elements [9]. Their proofs are checked using both HOL Light
and Coq.

Foundations of the area method The lemmas used by the area method (Section 2.2.4.1) can be
proved within, say, Hilbert’s geometry (i.e., within its fragment for plane geometry), but the proofs
are involved because in order to define the signed area and Pythagoras difference, arithmetization
of geometry is needed. As an alternative, Chou et al. proposed an elegant higher-level axiomatiza-
tion [50], suitable for the area method, later further extended and improved by Narboux, yielding an
axiomatic system that includes axioms for fields and only 13 axioms (such as SABC = −SBAC ). Us-
ing this system, Narboux proved (formally, within Coq) all properties used by the area method, thus
demonstrating the correctness of the method [130, 161]. It was also proved in a machine verifiable
way that the axioms can be derived from either Guilhot’s formal development [179, 181], Hilbert’s
or Tarski’s axioms [22].
Computer-Assisted Theorem Proving in Synthetic Geometry 41

Projective geometry Magaud et al. have proposed a Coq formalization of projective plane geom-
etry based on the notion of rank [147]. The correspondence between the rank axiom system and tra-
ditional axiom system was then completed and extended by Braun et al. [30, 31]. In Mizar, Coghetto
has formalized a proof of Pascal’s theorem [61] that is based on several articles by Leończuk and
Prażmowskin [140], and by Skaba [206].

Complex plane Marić and Petrović have formalized the extended complex plane using the Is-
abelle/HOL proof assistant [150]. They defined homogeneous coordinates, stereographic projection,
Möbius transformations, generalized circles and proved their properties.

2.3.1.5 Meta-Theory
There has also been some mechanization efforts targeted at proving meta-theoretical results. Makar-
ios proved the independence of the parallel postulate and the equivalence between different variants
of Tarski’s axiom system in Isabelle/HOL [149]. Boutry et al. formalized the equivalence of 34 ver-
sions of the parallel postulate and studied the impact of intuitionist logic on it [23]. Romanos and
Paulson formalized a proof of the impossibility of trisecting the angle and doubling the cube [198]
in Isabelle/HOL as did Harrison in HOL Light.

2.3.2 Higher Level Results

In the section, we present the formalizations of some of the best known theorems in synthetic ge-
ometry.

Desargues’ theorem Desargues’ theorem is an important result in geometry. It states that two
triangles are in perspective axially if and only if they are in perspective centrally (see also Sec-
tion 2.2.4.4). The theorem can be proved without any assumption in space but there are non-
Desarguesian planes such as Moulton’s plane. Kusak formalized the proof of Desargues’ theo-
rem in the projective space using Mizar [138]. A formalization of the same result and of Moul-
ton’s plane using the concept of rank was obtained using the Coq proof assistant by Magaud et
al. [146, 147, 148]. Braun has formalized Hessenberg’s proof of Desargues’ theorem in the context
of Tarski’s geometry [201]. Desargues’ theorem can be also verified by the Coq’s implementation
of the area method [130]. Harrison mechanized the proof of Desargues’ theorem using a technique
based on the computation of determinants [195].

Morley’s theorem Morley’s theorem states that the three points of intersection of the adjacent
angle trisectors form an equilateral triangle. Coghetto formalized the proof of Letac [141] using
the Mizar proof assistant [58]. The proof is based on trigonometry, the law of sines and the law of
cosines. Harrison mechanized Connes’ proof [65] in HOL Light. Guilhot formalized a proof based
on a set of lemmas about trigonometric functions, while Ida et al. proposed a computational proof
using origami [126, 127].

Triangle centers Triangle centers such as the center of gravity, the circumcenter, the orthocenter
have been studied for centuries by geometers. The existence and uniqueness of the circumcenter,
incenter, orthocenter, and gravity center have been formalized in Coq by Boutry et al. In particular,
they examined the difference between the computer checked proof of the existence of the center
of gravity and the usual proof published in textbooks [26]. The equivalence of the existence of the
circumcenter with the parallel postulate was also formalized [23]. In Mizar, Coghetto has formalized
the existence and uniqueness of circumcenter, orthocenter, and centroid [59, 60].
Kimberling has curated an electronic encyclopedia of triangle centers (ETC) that contains over
12,000 centers and many of their properties [134]. Recently, Narboux developed a certified version
42 Handbook of Geometric Constraint Systems Principles

of (parts of) the ETC such that some of the properties are accompanied by a Coq-checked proof of
their validity [165].

Sum of angles of a triangle The fact that the sum of angles of a triangle is congruent to the
sum of two right angles has also been studied. Boutry et al. formalized in Coq its equivalence with
the parallel postulate [23]. This required the sum of angles to be defined geometrically without
using the Archimedean axiom, hence without a means of measuring angles as real numbers [111].
Guilhot proved in Coq that the measure of the sum of angles of a triangle is π using her axiom
system based on mass points and by assuming axioms similar to the protractor postulate [112].
Harrison formalized the same theorem in HOL Light using the law of sines and cosines. Eberl has
ported Harrison’s proof to Isabelle/HOL [87]. Mizar’s formalization by Chang et al. uses complex
numbers [42].

Pythagorean theorem The Pythagorean theorem is an important property in Euclidean geometry.

Guilhot has formalized an analytic proof using Coq and the ring tactic [112]. A formal proof can
also be obtained in one line using algebraic techniques such as Gröbner basis or Wu’s method as
implemented in Coq, Isabelle/HOL and HOL Light. The Mizar proof by Kubo and Nakamura uses
the complex numbers [42]. Braun et al. have formalized a synthetic proof starting from Tarski’s
axioms [22]. Having a synthetic proof of the Pythagorean theorem is necessary as the latter is used
in the justification of the analytic approach that relies on the arithmetization of geometry.

Heron’s formula The formalization in Mizar [192] uses a coordinate-based definition of the area
of a triangle. The HOL Light formalization by Harrison uses a more general definition of area based
on measure theory. There are (at least) two Coq’s formalization. The one by Narboux is based on
the axioms of the area method (see Section 2.2.4.1) while Guilhot has one derived from her axioms
about mass points.

Thales’ circle theorem Braun and Magaud have formalized several analytic and synthetic proofs
of this theorem and of its reciprocal in Coq [29]. Boutry et al. have formalized its equivalence to
the parallel postulate [23] and the proof can also be obtained using algebraic methods in HOL Light
and in Coq (see Section 2.3.4).

Intersecting chord and Ptolemy’s theorems Pham proposed a formalization in Coq based on
Guilhot’s axiom system [178]. Bulwahn formalized those theorems in the real Cartesian plane using
Isabelle/HOL [37, 38]. Pottier et al. proved a version of these theorems not involving betweenness
and square of distances using Gröbner basis [110]. Harrison verified these theorems using alge-
braic techniques in HOL Light and, as for the Pythagorean theorem, the Mizar formalization used
complex numbers [192].

2.3.3 Other Formalizations Related to Geometry

In this section, we briefly describe some important formalization efforts which are not in the field
of synthetic geometry, but are related to geometry.

π and the area of the circle The number π can be given by many different definitions. Bertot
and Allais prove formally using the Coq proof assistant that several definitions are equivalent [13].
The cosine function is defined as a power series and π is then defined as twice the unique value α
between 1 and 2 such that cos(α) = 0. They show that this definition is equivalent to a definition
based on the area of a unit circle using Riemann integral. They also proved that Archimedes’s
process to compute π is also equivalent. The perimeter of the circle can be seen as the limit of
a sequence of inscribed regular polygons. Finally, they also provide means to actually compute
Computer-Assisted Theorem Proving in Synthetic Geometry 43

approximations of π. The area of a circle has been derived by Eberl in Isabelle/HOL (via integration
by substitution) and Harrison in HOL Light. In both cases, this involves a definition of area based
on measure theory.

Euler’s formula Euler’s polyhedron formula asserts that V − E + F = 2 where V , E, and F are the
number of vertices, edges, and faces of polyhedron, respectively. This formula is interesting from
the formalization point of view because of its long history of proofs and refutations [139]. Alama
formalized Poincaré’s proof using a definition of polyhedrons based on incidence matrices [1]. Du-
fourd formalized an intuitionist proof in Coq involving polyhedron defined using hyper-maps [78].

The Jordan Curve Theorem The Jordan Curve Theorem is a well-known example of a statement
that looks trivial but is indeed very hard to prove rigorously. The Mizar formalization started in 1992
and was completed in September 2005 [137]. Hales completed the formalization of the theorem
in HOL Light in January 2005 [117]. Scott and Fleuriot formalized a version of the theorem for
polygons using only Hilbert’s axioms Group I and II, i.e., with only the weak notions of incidence
and order [205]. Dufourd formalized a discrete version using hyper-maps [79].

Four color theorem The incorrect proof by Kempe [133] and the fact that Appel and Haken’s
proof [3] involved ad-hoc computer programs made the four color theorem a good candidate for
formalization. Gonthier produced a formal proof using hyper-maps [106, 107]. He used a version
of Euler’s formula to define planar maps and proved a combinatorial version of the Jordan Curve
Theorem equivalent to his version the of Euler formula.

The Kepler conjecture The largest formalization effort related to geometry is that of Ferguson
and Hales’ proof of the Kepler conjecture. Their pen-and-paper proof was completed in 1998 but not
published in full until 2006, with the editor adding as a side note that the reviewers were only “99%
certain” of the correctness of the proof. This compelled Hales to start the Flyspeck formalization
project [116], with the aim of providing a fully verified proof in HOL Light. During the formaliza-
tion, which involved a large international collaboration and the formalization of some early results
in Isabelle/HOL, a few minor errors and incomplete arguments were uncovered [118]. Hales pub-
lished a book describing the formalized version of the proof [114], while the final, fully checked
result was finally achieved using HOL Light in August 2014 [115].

Nonstandard analysis Fleuriot carried out the formalization of geometric proofs from Newton’s
Principia using nonstandard analysis in Isabelle/HOL [92]. By combining synthetic proofs based on
the area method (see Section 2.2.4.1) and full-angles (see Section 2.2.4.2) with a rigorous treatment
of infinitesimals [91], he formalized Kepler’s law of Equal Areas and discovered a flaw in Newton’s
famous Kepler Proposition [94]. Magaud, Cholet, and Fuchs carried out formalization of the discrete
model of the continuum introduced by Harthong and Reeb in Coq [145] and Fleuriot performed a
similar study using Isabelle/HOL [93].

Origami Origami, the art of paper folding, provides geometrical tools allowing some construc-
tions that cannot be realized by ruler and compass. Ida et al. have studied the mathematical and
computational aspect of origami. Kaliszyk and Ida have also investigated how the Mathematica
proofs based on Gröbner basis and Cylindrical Algebraic Decomposition can be reproduced in Is-
abelle/HOL, HOL Light, and Coq [132]. Despite the limitations of proof assistants when it comes to
algebraic reasoning, such combinations of interactive and automatic theorem proving can be useful
for the mechanized verification of theorems about origami constructions [125].

Computational and combinatorial geometry Dufourd and Puitg were the first to use a proof
assistant to formalize geometric data structures [187, 188]. Pichardie and Bertot proved a con-
44 Handbook of Geometric Constraint Systems Principles

vex hull algorithm using the Coq proof assistant [182]. They show that Knuth’s counterclockwise
systems (CCS) hold in R2 and then verify an incremental algorithm for computing convex hulls.
Dehlinger and Dufourd defined a formal specification of generalized maps in Coq and proved the
trading theorem [70]. Dufourd also developed a Coq formalization of hyper-maps [77] and applied
it to the verification of an image segmentation algorithm [75, 76]. Meikle and Fleuriot verified in
Isabelle/HOL the Graham Scan convex hull algorithm using an embedded While language, Hoare
logic, and the properties of signed areas formalized via Knuth’s CCS [158]. Brun et al. formalized
the same algorithm using hyper-maps [35]. They also have derived an imperative program [34].
Dufourd and Bertot applied these formalizations to the proof of a plane Delaunay triangulation
algorithm [80]. Brandt proposed the use of three-valued logic to simplify the formalization of de-
generated conditions and apply its method to formalize a convex-hull algorithm [27, 28]. Fuchs
and Théry have formalized Grassmann-Cayley algebra in Coq [95]. Ma et al. have used geometric
algebra to formalize robotic manipulation algorithms [144]. Liu et al. have used Coq to verify an
aircraft proximity property [143].

2.3.4 Verified Automated Reasoning

Proof assistants are designed for checking proofs generated by the user, but it is also possible to
check proofs generated by automated theorem provers. The latter can be used to prove theorems
wholesale or, at a smaller scale, to automate parts of proofs which are produced interactively. There
are several ways to integrate an ATP in a proof assistant. It is possible to define the ATP inside the
proof assistant and prove its correctness; this is called the reflexive approach [12]. It is also possible
to rely on external tools that generate a proof which is then checked by the proof assistant. Instead
of exporting the whole proof it is also possible to export a certificate, i.e., a piece of information
that is sufficient to reconstruct a proof efficiently inside the proof assistant. For example, to prove
that a polynomial is not irreducible, an oracle can use a complex factorization algorithm to compute
the factors. The proof assistant can then use a simple algorithm to check that the factorization is
correct [73].
Narboux has used the tactic language of Coq [72] to implement the area method in Coq (see
Section 2.2.4.1) [130, 161]. Grégoire et al. have used this approach to bring the power of Gröbner
basis algorithms to the Coq proof assistant [110, 186]. Genevaux et al. then extended their work
to an implementation of Wu’s method [102]. Chaieb et al. integrated quantifier elimination and
Gröbner basis within Isabelle/HOL [40, 41] and applied these to some geometric examples. Petrović
proposed an Isabelle/HOL formalization of the concept of geometric construction, with an algebraic
semantics [177]. Harrison used algebraic methods to check some examples in geometry using HOL
Light [120, 121].
At a small scale level, Boutry et al. proposed a reflexive approach to prove some transitivity
properties and deals with some symmetries implicitly [24, 26] generalizing the theory EL that Bal-
biani et al. studied using term rewriting techniques [5]. The example shown in Figure 2.9 uses the
tactics described by Boutry et al. [26]. Scott and Fleuriot introduced a generic combinator language
that was then instantiated to give automatic forward proof strategies that could deal with incidence
reasoning in Hilbert’s proof, especially in cases involving Pasch’s axiom [204]. As pointed out in
Section 2.2.3.2, Stojanović et al. generated readable proofs (of theorems from Tarski’s geometry)
which are checked by Coq and Isabelle/HOL [208].

Acknowledgment The authors are grateful to Michael Beeson, Roland Cogheto, Pedro Quaresma,
Predrag Tanović, Dongming Wang, and the anonymous reviewer for their comments on preliminary
versions of this chapter. Janičić was partly supported by grant ON174021 of the Ministry of Science
of Serbia. Fleuriot was supported by EPSRC grants EP/L011794/1 and EP/N014758/1.
References 45
Lemma varignon : forall A B C D I J K L, A<>C -> B<>D -> ~ Col I J K ->
Midpoint I A B -> Midpoint J B C -> Midpoint K C D -> Midpoint L A D ->
Parallelogram I J K L.
Proof.
intros.
assert_diffs. (* Deduce some inequalities *)
assert (Par I L B D) by perm_apply (triangle_mid_par B D A L I).
(* Applying the midpoint theorem in the triangle BDA. *)
assert (Par J K B D) by perm_apply (triangle_mid_par B D C K J).
(* Applying the midpoint theorem in the triangle BDC. *)
assert (Par I L J K) by (apply par_trans with B D;finish).
(* Transitivity of parallelism *)
assert (Par I J A C) by perm_apply (triangle_mid_par A C B J I).
(* Applying the midpoint theorem in the triangle ACB. *)
assert (Par L K A C) by perm_apply (triangle_mid_par A C D K L).
(* Applying the midpoint theorem in the triangle ACD. *)
assert (Par I J K L) by (apply par_trans with A C;finish).
(* Transitivity of parallelism *)
apply par_2_plg;finish.
(* If in the opposite sides of a quadrilatral are parallel and
two opposite side are distinct then it is a parallelogram. *)
Qed.

Figure 2.9
Coq script of the proof of Varignon’s theorem based on some tactics for partial automation.

References
[1] Jesse Alama. Euler’s polyhedron formula. Formalized Mathematics, 16(1):7–17, 2008.
[2] J. R. Anderson, C. F. Boyle, and G. Yost. The geometry tutor. The Journal of Mathematical
Behavior, pages 5–20, 1986.
[3] K. Appel and W. Haken. Every planar map is four colorable. Part I: Discharging. Illinois
Journal of Mathematics, 21(3):429–490, 1977.
[4] Jeremy Avigad, Edward Dean, and John Mumma. A Formal System for Euclid’s Elements.
The Review of Symbolic Logic, 2:700–768, 2009.
[5] Philippe Balbiani and Luis del Cerro. Affine geometry of collinearity and conditional term
rewriting. In Hubert Comon and Jean-Pierre Jounnaud, editors, Term Rewriting, volume 909
of Lecture Notes in Computer Science, pages 196–213. Springer Berlin / Heidelberg, 1995.
10.1007/3-540-59340-3 14.
[6] Michael Beeson. Foreword to Metamathematische Methoden in der Geometrie. Ishi Press,
2011.
[7] Michael Beeson. A constructive version of Tarski’s geometry. Annals of Pure and Applied
Logic, 166(11):1199–1273, 2015.
[8] Michael Beeson. Constructive geometry and the parallel postulate. Bulletin of Symbolic
Logic, 22(1):1–104, 2016.
[9] Michael Beeson, Julien Narboux, and Freek Wiedijk. Proof-checking Euclid. 2017.
[10] Michael Beeson and Larry Wos. OTTER proofs of theorems in Tarskian geometry. In
Stéphane Demri, Deepak Kapur, and Christoph Weidenbach, editors, 7th International Joint
Conference, IJCAR 2014, Held as Part of the Vienna Summer of Logic, Vienna, Austria, July
19-22, 2014, Proceedings, volume 8562 of Lecture Notes in Computer Science, pages 495–
510. Springer, 2014.
46 References

[11] Michael Beeson and Larry Wos. Finding Proofs in Tarskian Geometry. Journal of Automated
Reasoning, submitted, 2015.
[12] Yves Bertot. Coq in a Hurry. May 2010.
[13] Yves Bertot and Guillaume Allais. Views of PI: Definition and computation. Journal of
Formalized Reasoning, 7(1):105–129, 2014.
[14] Yves Bertot and Pierre Castéran. Interactive Theorem Proving and Program Development,
Coq’Art: The Calculus of Inductive Constructions. Texts in Theoretical Computer Science.
An EATCS Series. Springer, 2004.
[15] Marc Bezem and Thierry Coquand. Automating Coherent Logic. In Geoff Sutcliffe and
Andrei Voronkov, editors, 12th International Conference on Logic for Programming, Artifi-
cial Intelligence, and Reasoning — LPAR 2005, volume 3835 of Lecture Notes in Computer
Science, pages 246–260. Springer-Verlag, 2005.
[16] Marc Bezem and Dimitri Hendriks. On the Mechanization of the Proof of Hessenberg’s
Theorem in Coherent Logic. Journal of Automated Reasoning, 40(1):61–85, 2008.
[17] George D. Birkhoff. A set of postulates for plane geometry, based on scale and protractor.
Annals of Mathematics, 33:329–345, 1932.
[18] Karol Borsuk and Wanda Szmielew. Foundations of Geometry: Euclidean and Bolyai-
Lobachevskian geometry, Projective Geometry. North-Holland Pub Co, 1960.
[19] Francisco Botana, Markus Hohenwarter, Predrag Janičić, Zoltán Kovács, Ivan Petrović,
Tomás Recio, and Simon Weitzhofer. Automated Theorem Proving in GeoGebra: Current
Achievements. Journal of Automated Reasoning, 55(1):39–59, 2015.
[20] Pierre Boutry. On the Formalization of Foundations of Geometry. PhD thesis, University of
Strasbourg. in preparation.
[21] Pierre Boutry, Gabriel Braun, and Julien Narboux. From Tarski to Descartes: Formalization
of the Arithmetization of Euclidean Geometry. In The 7th International Symposium on Sym-
bolic Computation in Software (SCSS 2016), EasyChair Proceedings in Computing, page 15,
Tokyo, Japan, March 2016.
[22] Pierre Boutry, Gabriel Braun, and Julien Narboux. Formalization of the Arithmetization of
Euclidean Plane Geometry and Applications. Journal of Symbolic Computation, page 23,
2017. In Press.
[23] Pierre Boutry, Charly Gries, Julien Narboux, and Pascal Schreck. Parallel postulates and con-
tinuity axioms: a mechanized study in intuitionistic logic using Coq. Journal of Automated
Reasoning, page 72, 2017. to appear.
[24] Pierre Boutry, Julien Narboux, and Pascal Schreck. A reflexive tactic for automated genera-
tion of proofs of incidence to an affine variety. October 2015.
[25] Pierre Boutry, Julien Narboux, Pascal Schreck, and Gabriel Braun. A short note about
case distinctions in Tarski’s geometry. In Francisco Botana and Pedro Quaresma, editors,
Proceedings of the 10th Int. Workshop on Automated Deduction in Geometry, volume TR
2014/01 of Proceedings of ADG 2014, pages 51–65, Coimbra, Portugal, July 2014. Univer-
sity of Coimbra.
[26] Pierre Boutry, Julien Narboux, Pascal Schreck, and Gabriel Braun. Using small scale au-
tomation to improve both accessibility and readability of formal proofs in geometry. In
Francisco Botana and Pedro Quaresma, editors, Proceedings of the 10th Int. Workshop on
References 47

Automated Deduction in Geometry, volume TR 2014/01 of Proceedings of ADG 2014, pages

31–49, Coimbra, Portugal, July 2014. University of Coimbra.
[27] Jens Brandt. A Layered Approach to Polygon Processing for Safety-Critical Embedded Sys-
tems. PhD thesis, Universität Kaiserslautern, 2007.
[28] Jens Brandt and Klaus Schneider. Dependable Polygon-Processing Algorithms for Safety-
Critical Embedded Systems. In Laurence T. Yang, Makoto Amamiya, Zhen Liu, Minyi Guo,
and Franz J. Rammig, editors, Embedded and Ubiquitous Computing – EUC 2005, number
3824 in Lecture Notes in Computer Science, pages 405–417. Springer Berlin Heidelberg,
December 2005. DOI: 10.1007/11596356 42.
[29] David Braun and Nicolas Magaud. Des preuves formelles en Coq du théorème de Thalès
pour les cercles. In David Baelde and Jade Alglave, editors, Vingt-sixièmes Journées Fran-
cophones des Langages Applicatifs (JFLA 2015), Le Val d’Ajol, France, January 2015.
[30] David Braun, Nicolas Magaud, and Pascal Schreck. An Equivalence Proof Between Rank
Theory and Incidence Projective Geometry. In Proceedings of Automated Deduction in Ge-
ometry, pages 62–77, Strasbourg, 2016.
[31] David Braun, Nicolas Magaud, and Pascal Schreck. Formalizing Finite Models of Projective
Geometry in Coq: a Benchmark for Proof Automation. 2017.
[32] Gabriel Braun and Julien Narboux. From Tarski to Hilbert. In Tetsuo Ida and Jacques
Fleuriot, editors, Post-Proceedings of Automated Deduction in Geometry 2012, volume 7993
of LNCS, pages 89–109, Edinburgh, United Kingdom, September 2012. Springer.
[33] Gabriel Braun and Julien Narboux. A synthetic proof of Pappus’ theorem in Tarski’s geom-
etry. Journal of Automated Reasoning, 58(2):23, February 2017.
[34] Christophe Brun, Jean-François Dufourd, and Nicolas Magaud. Designing and proving cor-
rect a convex hull algorithm with hypermaps in Coq. Computational Geometry, 45(8):436–
457, 2012.
[35] Christophe Brun, Jean-François Dufourd, and Nicolas Magaud. Formal Proof in Coq and
Derivation of an Imperative Program to Compute Convex Hulls. In Tetsuo Ida and Jacques D.
Fleuriot, editors, Automated Deduction in Geometry, number 7993 in Lecture Notes in Com-
puter Science, pages 71–88. Springer Berlin Heidelberg, 2012. DOI: 10.1007/978-3-642-
40672-0 6.
[36] Bruno Buchberger, Adrian Crǎciun, Tudor Jebelean, Laura Kovács, Temur Kutsia, Koji Nak-
agawa, Florina Piroi, Nikolaj Popov, Judit Robu, Markus Rosenkranz, and others. Theorema:
Towards Computer-Aided Mathematical Theory Exploration. Journal of Applied Logic,
4(4):470–504, 2006.
[37] Lukas Bulwahn. Intersecting Chords Theorem. Archive of Formal Proofs, October 2016.
[38] Lukas Bulwahn. Ptolemy’s Theorem. Archive of Formal Proofs, August 2016.
[39] Ugo Cassina. Ancora sui fondamenti della geometria secondo Hilbert. I, II, III. Istituto Lom-
bardo di Scienze e Lettere, Rendiconti, Classe di Scienze Matematiche e Naturali, 12(81):71–
94, 1948.
[40] Amine Chaieb and Tobias Nipkow. Verifying and reflecting quantifier elimination for Pres-
burger arithmetic. In G. Sutcliffe and A. Voronkov, editors, Logic for Programming, Artificial
Intelligence, and Reasoning (LPAR 2005), volume 3835 of LNCS, pages 367–380. Springer,
2005.
48 References

[41] Amine Chaieb and Makarius Wenzel. Context aware Calculation and Deduction — Ring
Equalities via Gröbner Bases in Isabelle. In M. Kauers, M. Kerber, R. Miner, and W. Wind-
steiger, editors, CALCULEMUS 2007, volume 4573 of Lecture Notes in Computer Science,
pages 27–39. Springer, 2007.
[42] Wenpai Chang, Yatsuka Nakamura, and Piotr Rudnicki. Inner products and angles of com-
plex numbers. Formalized Mathematics, 11(3):275–280, 2003.
[43] Xiaoyu Chen and Dongming Wang. Formalization and Specification of Geometric Knowl-
edge Objects. Mathematics in Computer Science, 7(4):439–454, December 2013.
[44] XueFeng Chen and DingKang Wang. The Projection of Quasi Variety and Its Application on
Geometric Theorem Proving and Formula Deduction. In Franz Winkler, editor, Automated
Deduction in Geometry, 4th International Workshop, ADG 2002, Hagenberg Castle, Austria,
September 4-6, 2002, Revised Papers, volume 2930 of Lecture Notes in Computer Science,
pages 21–30. Springer, 2004.
[45] Shang-Ching Chou. Proving and discovering geometry theorems using Wu’s method. PhD
thesis, The University of Texas, Austin, December 1985.
[46] Shang-Ching Chou and Xiao-Shan Gao. A class of geometry statements of constructive type
and geometry theorem proving. In Proceeding of CADE 92. Academia Sinica, 1992.
[47] Shang-Ching Chou and Xiao-Shan Gao. Automated Generation of Readable Proofs with
Geometric Invariants I. Multiple and Shortest Proof Generation. J. Autom. Reasoning,
17(3):325–347, 1996.
[48] Shang-Ching Chou and Xiao-Shan Gao. Automated Reasoning in Geometry. In John Alan
Robinson and Andrei Voronkov, editors, Handbook of Automated Reasoning, pages 707–749.
Elsevier and MIT Press, 2001.
[49] Shang-Ching Chou, Xiao-Shan Gao, and Jing-Zhong Zhang. Automated Geometry Theorem
Proving by Vector Calculation. In Manuel Bronstein, editor, Proceedings of the 1993 Inter-
national Symposium on Symbolic and Algebraic Computation, ISSAC ’93, pages 284–291.
ACM, 1993.
[50] Shang-Ching Chou, Xiao-Shan Gao, and Jing-Zhong Zhang. Automated production of tradi-
tional proofs for constructive geometry theorems. In Moshe Vardi, editor, Proceedings of the
Eighth Annual IEEE Symposium on Logic in Computer Science LICS, pages 48–56. IEEE
Computer Society Press, June 1993.
[51] Shang-Ching Chou, Xiao-Shan Gao, and Jing-Zhong Zhang. Machine Proofs in Geometry.
World Scientific, Singapore, 1994.
[52] Shang-Ching Chou, Xiao-Shan Gao, and Jing-Zhong Zhang. Automated Production of Tra-
ditional Proofs in Solid Geometry. Journal of Automated Reasoning, 14:257–291, 1995.
[53] Shang-Ching Chou, Xiao-Shan Gao, and Jing-Zhong Zhang. Automated Generation of read-
able proofs with geometric invariants, Theorem Proving with Full Angle. Journal of Auto-
mated Reasoning, 17:325–347, 1996.
[54] Shang-ching Chou, Xiao-shan Gao, and Jing-zhong Zhang. A Deductive Database Approach
to Automated Geometry Theorem Proving and Discovering. Journal of Automated Reason-
ing, 25:219–246, 2000.
[55] Chou Shang-Ching and Gao Xiao-Shan. A Survey of Geometric Reasoning Using Algebraic
Methods. pages 97–119. Birkhäuser Boston, Boston, MA, 1996. DOI: 10.1007/978-1-4612-
4088-4 5.
References 49

[56] H. Coelho and L. M. Pereira. GEOM: A Prolog Geometry Theorem Prover. Memórias
525, Laboratório Nacional de Engenharia Civil, Ministério de Habitação e Obras Públicas,
Portugal, 1979.
[57] H. Coelho and L. M. Pereira. Automated reasoning in geometry theorem proving with Pro-
log. Journal of Automated Reasoning, 2(4):329–390, 1986.
[58] Roland Coghetto. Morley’s Trisector Theorem. Formalized Mathematics, 23(2):75–79,
2015.
[59] Roland Coghetto. Altitude, Orthocenter of a Triangle and Triangulation. Formalized Mathe-
matics, 24(1):27–36, 2016.
[60] Roland Coghetto. Circumcenter, circumcircle and centroid of a triangle. Formalized Mathe-
matics, 24(1):17–26, 2016.
[61] Roland Coghetto. Pascal’s Theorem in Real Projective Plane. 25:110, 2017.
[62] Roland Coghetto and Adam Grabowski. Tarski Geometry Axioms — Part II. Formalized
Mathematics, 24:157–166, 2016.
[63] George E. Collins. Quantifier Elimination for Real Closed Fields by Cylindrical Algebraic
Decomposition. volume 33 of Lecture Notes In Computer Science, pages 134–183. Springer-
Verlag, 1975.
[64] George E. Collins. Quantifier Elimination for Real Closed Fields by Cylindrical Algebraic
Decomposition: a synopsis. SIGSAM Bull., 10(1):10–12, February 1976.
[65] Alain Connes. A new proof of Morley’s theorem. Publications Mathématiques de l’IHÉS,
88:43–46, 1998.
[66] Thierry Coquand and Gérard Huet. The Calculus of Constructions. In Information and
Computation, volume 76. 1988.
[67] Thierry Coquand and Christine Paulin-Mohring. Inductively defined types. In P. Martin-Löf
and G. Mints, editors, Proceedings of Colog’88, volume 417 of Lecture Notes in Computer
Science. Springer-Verlag, 1990.
[68] H. S. M. Coxeter. Introduction to Geometry. Wiley, 2nd edition, 1969.
[69] Li Dafa, Peifa Jia, and Xinxin Li. Simplifying von Plato’s Axiomatization of Constructive
Apartness Geometry. In Annals of Pure and Applied Logic, volume 102, pages 1–26. 2000.
[70] Christophe Dehlinger and Jean-François Dufourd. Formalizing generalized maps in Coq.
Theoretical Computer Science, 323(1-3):351–397, 2004.
[71] Christophe Dehlinger, Jean-François Dufourd, and Pascal Schreck. Higher-Order Intuition-
istic Formalization and Proofs in Hilbert’s Elementary Geometry. In Automated Deduction
in Geometry, volume 2061 of Lecture Notes in Computer Science, pages 306–324. Springer,
2001.
[72] David Delahaye. A Tactic Language for the System Coq. In Proceedings of Logic for
Programming and Automated Reasoning (LPAR), Reunion Island (France), volume 1955 of
Lecture Notes in Artificial Intelligence, pages 85–95. Springer-Verlag, November 2000.
[73] David Delahaye and Micaela Mayero. Dealing with Algebraic Expressions over a Field in
Coq using Maple. In Journal of Symbolic Computation: special issue on the integration of
automated reasoning and computer algebra systems, volume 39, pages 569–592, 2005.
[74] René Descartes. La géométrie. Open Court, Chicago, 1925.
50 References

[75] Jean-François Dufourd. Définition et preuve d’un algorithme fonctionnel de segmentation

d’image basé sur les hypercartes. In Proceedings of JFLA 2007, Aix-les-Bains, January
2007.
[76] Jean-François Dufourd. Design and formal proof of a new optimal image segmentation
program with hypermaps. Pattern Recognition, 2007.
[77] Jean-François Dufourd. A hypermap framework for computer assisted proofs in surface
subdivisions. In Symposium on Applied Computing, Seoul, March 2007.
[78] Jean-François Dufourd. Polyhedra Genus Theorem and Euler Formula: A Hypermap-
formalized Intuitionistic Proof. Theor. Comput. Sci., 403(2-3):133–159, August 2008.
[79] Jean-François Dufourd. An intuitionistic proof of a discrete form of the Jordan curve the-
orem formalized in Coq with combinatorial hypermaps. Journal of Automated Reasoning,
43(1):19–51, 2009.
[80] Jean-François Dufourd and Yves Bertot. Formal study of plane Delaunay triangulation. In
Lawrence Paulson and Matt Kaufmann, editors, Interactive Theorem Proving, volume 6172,
pages 211–226, Edinburgh, United Kingdom, July 2010. Springer.
[81] Jean Duprat. Proof assistant for plane geometry. In Workshop on Lambda Calculus, Type
Theory and Mathematical Logic, volume 12 of Schedea Informatica, Cracovie, June 2003.
[82] Jean Duprat. Une axiomatique de la géométrie plane en Coq. In JFLA (Journées Franco-
phones des Langages Applicatifs), pages 123–136, Etretat, France, January 2008. INRIA.
[83] Jean Duprat. The Euclid’s Plane: Formalization and Implementation in Coq. In ADG 2010,
2010.
[84] Jean Duprat. Fondements de géométrie euclidienne. 2010.
[85] Roy Dyckhoff and Sara Negri. Geometrization of first-order logic. The Bulletin of Symbolic
Logic, 21:123–163, 2015.
[86] Jerzy Dydak. Axiomatization of geometry employing group actions. January 2015.
[87] Manuel Eberl. Basic Geometric Properties of Triangles. Archive of Formal Proofs, December
2015.
[88] Philip Ehrlich. From Completeness to Archimedean Completeness: An Essay in the Foun-
dations of Euclidean Geometry. A Symposium on David Hilbert, 110:57–76, 1997.
[89] E. W. Elcock. Representation of knowledge in geometry machine. Machine Intelligence,
8:11–29, 1977.
[90] Alfredo Ferro and Giovanni Gallo. Automated theorem proving in elementary geometry. Le
Matematiche, 43(1,2):195–224, 1990.
[91] Jacques D. Fleuriot. Nonstandard Geometric Proofs. In Automated Deduction in Geome-
try, Third International Workshop, ADG 2000, Zurich, Switzerland, September 25-27, 2000,
Revised Papers, pages 246–267, 2000.
[92] Jacques D. Fleuriot. Theorem Proving in Infinitesimal Geometry. Logic Journal of the IGPL,
9(3):447–474, 2001.
[93] Jacques D. Fleuriot. Exploring the foundations of discrete analytical geometry in Is-
abelle/HOL. In International Workshop on Automated Deduction in Geometry, volume 6877
of LNCS, pages 34–50. Springer, 2010.
References 51

[94] Jacques D. Fleuriot and Lawrence C. Paulson. Proving Newton’s Propositio Kepleriana Us-
ing Geometry and Nonstandard Analysis in Isabelle. In Automated Deduction in Geometry,
Second International Workshop, ADG’98, Beijing, China, August 1-3, 1998, Proceedings,
pages 47–66, 1998.
[95] Laurent Fuchs and Laurent Théry. A Formalization of Grassmann-Cayley Algebra in COQ
and Its Application to Theorem Proving in Projective Geometry. In Pascal Schreck, Julien
Narboux, and Jürgen Richter-Gebert, editors, Automated Deduction in Geometry, volume
6877 of Lecture Notes in Computer Science, pages 51–67. Springer Berlin Heidelberg, 2011.
[96] Herve Gallaire, Jack Minker, and Jean-Marie Nicolas. Logic and Databases: A Deductive
Approach. ACM Comput. Surv., 16(2):153–185, June 1984.
[97] Xiao-Shan Gao. Chapter 10 — Search methods revisited. In Xiao-Shan Gao and Dong-
ming Wang, editors, Mathematics Mechanization and Applications, pages 253–271. Aca-
demic Press, London, 2000. DOI: 10.1016/B978-012734760-8/50011-9.
[98] Xiao-Shan Gao and Qiang Lin. MMP/Geometer — A Software Package for Automated Ge-
ometric Reasoning. In Proceedings of Automated Deduction in Geometry (ADG02), volume
2930 of Lecture Notes in Computer Science, pages 44–66. Springer-Verlag, 2004.
[99] Herbert Gelernter. Realization of a geometry theorem proving machine. In Proceedings of
the International Conference Information Processing, pages 273–282, Paris, 1959.
[100] Herbert Gelernter. Realisation of a Geometry-Proving Machine. In Automation of Reasoning.
Springer-Verlag, Berlin, 1983.
[101] Herbert Gelernter, J. R. Hansen, and Donald Loveland. Empirical explorations of the geom-
etry theorem machine. In Papers presented at the May 3-5, 1960, western joint IRE-AIEE-
ACM Computer Conference, IRE-AIEE-ACM ’60 (Western), pages 143–149, San Francisco,
California, 1960. ACM.
[102] Jean-David Genevaux, Julien Narboux, and Pascal Schreck. Formalization of Wu’s Simple
Method in Coq. In Jean-Pierre Jouannaud and Zhong Shao, editors, CPP 2011 First Inter-
national Conference on Certified Programs and Proofs, volume 7086 of Lecture Notes in
Computer Science, pages 71–86, Kenting, Taiwan, December 2011. Springer-Verlag.
[103] Paul C. Gilmore. An Examination of the Geometry Theorem Machine. Artif. Intell.,
1(3):171–187, 1970.
[104] Eduardo N. Giovannini. Completitud y Continuidad En Fundamentos de la Geometrı́a de
Hilbert (Completeness and Continuity in Hilbert’s Foundations of Geometry). Theoria: Re-
vista de Teorı́a, Historia y Fundamentos de la Ciencia, 28(1):139–163, 2013.
[105] Ira Goldstein. Elementary Geometry Theorem Proving. AI Lab memo 280, MIT, April 1973.
[106] Georges Gonthier. A Computer Checked Proof of the Four Colour Theorem. 2004.
[107] Georges Gonthier. Formal Proof — The Four-Color Theorem. Notices of the American
Mathematical Society, 55(11):1382–1393, 2008.
[108] Adam Grabowski. Tarski’s geometry modelled in Mizar computerized proof assistant. In
Computer Science and Information Systems (FedCSIS), 2016 Federated Conference on,
pages 373–381. IEEE, 2016.
[109] James G. Greeno, Maria E. Magone, and Seth Chaiklin. Theory of constructions and set in
problem solving. Memory and Cognition, 7(6):445–461, 1979.
52 References

[110] Benjamin Grégoire, Loı̈c Pottier, and Laurent Théry. Proof Certificates for Algebra and Their
Application to Automatic Geometry Theorem Proving. In Post-proceedings of Automated
Deduction in Geometry (ADG 2008), volume 6301 of Lecture Notes in Artificial Intelligence,
pages 42–59. Springer, 2011.
[111] Charly Gries, Pierre Boutry, and Julien Narboux. Somme des angles d’un triangle et unicité
de la parallèle : une preuve d’équivalence formalisée en Coq. In Les vingt-septièmes Journées
Francophones des Langages Applicatifs (JFLA 2016), Actes des Vingt-septièmes Journées
Francophones des Langages Applicatifs (JFLA 2016), page 15, Saint Malo, France, January
2016. Jade Algave and Julien Signoles.
[112] Frédérique Guilhot. Formalisation en Coq d’un cours de géométrie pour le lycée. In Journées
Francophones des Langages Applicatifs. INRIA, January 2004.
[113] Haragauri Narayan Gupta. Contributions to the Axiomatic Foundations of Geometry. PhD
thesis, University of California, Berkley, 1965.
[114] Thomas Hales. Dense Sphere Packings: A Blueprint for Formal Proofs. Cambridge Univer-
sity Press, New York, NY, USA, 2012.
[115] Thomas Hales, Mark Adams, Gertrud Bauer, Dang Tat Dat, John Harrison, Hoang Le Truong,
Cezary Kaliszyk, Victor Magron, Sean Mclaughlin, Nguyen Tat Thang, Nguyen Quang
Truong, Tobias Nipkow, Steven Obua, Joseph Pleso, Jason Rute, Alexey Solovyev,
Ta Thi Hoai An, Tran Nam Trung, Trieu Thi Diep, Josef Urban, Vu Khac Ky, and Roland
Zumkeller. A Formal Proof of the Kepler Conjecture. Forum of Mathematics, Pi, 5, 2017.
[116] Thomas C. Hales. Introduction to the Flyspeck Project. In Mathematics, Algorithms,
Proofs, volume 05021 of Dagstuhl Seminar Proceedings. Internationales Begegnungs- und
Forschungszentrum für Informatik (IBFI), Schloss Dagstuhl, Germany, 2006.
[117] Thomas C. Hales. The Jordan Curve Theorem, Formally and Informally. The American
Mathematical Monthly, 114(10):882–894, 2007.
[118] Thomas C. Hales, John Harrison, Sean McLaughlin, Tobias Nipkow, Steven Obua, and
Roland Zumkeller. A Revision of the Proof of the Kepler Conjecture. Discrete & Com-
putational Geometry, 44(1):1–34, July 2010.
[119] John Harrison. HOL Light: A Tutorial Introduction. In Mandayam K. Srivas and Albert John
Camilleri, editors, Formal Methods in Computer-Aided Design, volume 1166 of Lecture
Notes in Computer Science. Springer, 1996.
[120] John Harrison. A HOL Theory of Euclidean space. In Joe Hurd and Tom Melham, editors,
Theorem Proving in Higher Order Logics, 18th International Conference, TPHOLs 2005,
volume 3603 of Lecture Notes in Computer Science, pages 114–129, Oxford, UK, August
2005. Springer-Verlag.
[121] John Harrison. The HOL Light Theory of Euclidean Space. Journal of Automated Reasoning,
50(2):173–190, 2013.
[122] Robin Hartshorne. Geometry:Euclid and Beyond. Undergraduate texts in mathematics.
Springer, 2000.
[123] Olaf Helmer-Hirschberg. Axiomatischer Aufbau der Geometrie in formalisierter Darstel-
lung. Schriften des mathematische Seminars und des Instituts fur angewandte Mathematik
der Universität Berlin, 2:175–201, 1935.
[124] David Hilbert. Grundlagen der Geometrie. Leipzig, 1899.
References 53

[125] Tetsuo Ida. Interactive vs. automated proofs in computational origami. In 2012 14th Interna-
tional Symposium on Symbolic and Numeric Algorithms for Scientific Computing (SYNASC),
pages 7–7, September 2012.
[126] Tetsuo Ida, Asem Kasem, Fadoua Ghourabi, and Hidekazu Takahashi. Morley’s theorem
revisited: Origami construction and automated proof. Journal of Symbolic Computation,
46(5):571–583, May 2011.
[127] Tetsuo Ida, Hidekazu Takahashi, and Mircea Marin. Computational Origami of a Morley’s
Triangle. In Michael Kohlhase, editor, Mathematical Knowledge Management, number 3863
in Lecture Notes in Computer Science, pages 267–282. Springer Berlin Heidelberg, July
2005. DOI: 10.1007/11618027 18.
[128] Predrag Janičić. GCLC - A Tool for Constructive Euclidean Geometry and More Than That.
In Andrés Iglesias and Nobuki Takayama, editors, Mathematical Software - ICMS 2006, Sec-
ond International Congress on Mathematical Software, Castro Urdiales, Spain, September
1-3, 2006, Proceedings, volume 4151 of Lecture Notes in Computer Science, pages 58–73.
Springer, 2006.
[129] Predrag Janičić and Stevan Kordić. EUCLID — the Geometry Theorem Prover. FILOMAT,
9(3):723–732, 1995.
[130] Predrag Janičić, Julien Narboux, and Pedro Quaresma. The Area Method : A Recapitulation.
Journal of Automated Reasoning, 48(4):489–532, 2012.
[131] Jianguo Jiang and Jingzhong Zhang. A review and prospect of readable machine proofs for
geometry theorems. Journal of Systems Science and Complexity, 25(4):802–820, 2012.
[132] Cezary Kaliszyk and Tetsuo Ida. Proof Assistant Decision Procedures for Formalizing
Origami. In James H. Davenport, William M. Farmer, Josef Urban, and Florian Rabe, edi-
tors, Intelligent Computer Mathematics, number 6824 in Lecture Notes in Computer Science,
pages 45–57. Springer Berlin Heidelberg, July 2011. DOI: 10.1007/978-3-642-22673-1 4.
[133] A. B. Kempe. On the Geographical Problem of the Four Colours. American Journal of
Mathematics, 2(3):193–200, 1879.
[134] Clark Kimberling. Triangle Centers and Central Triangles. 2001.
[135] Felix C. Klein. A comparative review of recent researches in geometry. PhD thesis, 1872.
[136] Kenneth R. Koedinger and John R. Anderson. Abstract Planning and Perceptual Chunks:
Elements of Expertise in Geometry. Cognitive Science, 14(4):511–550, 1990.
[137] Artur Kornilowicz. Jordan curve theorem. Formalized Mathematics, 13(4):481–491, 2005.
[138] Eugeniusz Kusak. Desargues Theorem In Projective 3-Space. Formalized Mathematics, 2(1),
1991.
[139] Imre Lakatos, John Worrall, and Elie Zahar, editors. Proofs and Refutations. Cambridge
University Press, 1976.
[140] Wojciech Leonczuk and Krzysztof Prazmowski. Projective Spaces. Journal of Formalized
Mathematics, 1(4):767–776, 1990.
[141] A. Letac. Solutions (Morley’s triangle). Problem N 490. Sphinx: Revue Mensuelle des
Questions Récréatives, 9, 1939.
[142] Tielin Liang and Dongming Wang. Towards a Geometric-Object-Oriented Language. In
Automated Deduction in Geometry, pages 130–155, 2004.
54 References

[143] Dongxi Liu, NealeL. Fulton, John Zic, and Martin de Groot. Verifying an Aircraft Proximity
Characterization Method in Coq. In Lindsay Groves and Jing Sun, editors, Formal Methods
and Software Engineering, volume 8144 of Lecture Notes in Computer Science, pages 86–
101. Springer Berlin Heidelberg, 2013.
[144] Sha Ma, Zhiping Shi, Zhenzhou Shao, Yong Guan, Liming Li, and Yongdong Li. Higher-
Order Logic Formalization of Conformal Geometric Algebra and its Application in Verifying
a Robotic Manipulation Algorithm. Advances in Applied Clifford Algebras, 26(4):1305–
1330, 2016.
[145] Nicolas Magaud, Agathe Chollet, and Laurent Fuchs. Formalizing a discrete model of the
continuum in Coq from a discrete geometry perspective. Annals of Mathematics and Artifi-
cial Intelligence, 74(3-4):309–332, October 2014.
[146] Nicolas Magaud, Julien Narboux, and Pascal Schreck. Formalizing Desargues’ Theorem in
Coq using Ranks. In SAC, pages 1110–1115, 2009.
[147] Nicolas Magaud, Julien Narboux, and Pascal Schreck. Formalizing Projective Plane Geome-
try in Coq. In Thomas Sturm, editor, Post-proceedings of Automated Deduction in Geometry
(ADG) 2008, volume 6301 of LNAI, pages 141–162, Shanghai, China, 2011. Springer.
[148] Nicolas Magaud, Julien Narboux, and Pascal Schreck. A Case Study in Formalizing Pro-
jective Geometry in Coq: Desargues Theorem. Computational Geometry, 45(8):406–424,
2012.
[149] Timothy James McKenzie Makarios. A Mechanical Verification of the Independence of
Tarski’s Euclidean Axiom. Master Thesis, Victoria University of Wellington, 2012.
[150] Filip Marić and Danijela Petrović. Formalizing complex plane geometry. Annals of Mathe-
matics and Artificial Intelligence, 74(3-4):271–308, 2015.
[151] Filip Marić, Ivan Petrović, Danijela Petrović, and Predrag Janičić. Formalization and Imple-
mentation of Algebraic Methods in Geometry. In Pedro Quaresma and Ralph-Johan Back,
editors, Proceedings First Workshop on CTP Components for Educational Software, Wro-
claw, Poland, 31th July 2011, volume 79 of Electronic Proceedings in Theoretical Computer
Science, pages 63–81. Open Publishing Association, 2012.
[152] Vesna Marinković. Proof Simplification in the Framework of Coherent Logic. Computing
and Informatics, 34(2):337–366, 2015.
[153] Vesna Marinković, Predrag Janičić, and Pascal Schreck. Computer Theorem Proving for
Verifiable Solving of Geometric Construction Problems. In Francisco Botana and Pedro
Quaresma, editors, Automated Deduction in Geometry - 10th International Workshop, ADG
2014, Coimbra, Portugal, July 9-11, 2014, Revised Selected Papers, volume 9201 of Lecture
Notes in Computer Science, pages 72–93. Springer, 2015.
[154] G. E. Martin. The Foundations of Geometry and the Non-Euclidean Plane. Undergraduate
Texts in Mathematics. Springer, 1998.
[155] Noboru Matsuda and Kurt Vanlehn. GRAMY: A Geometry Theorem Prover Capable of
Construction. Journal of Automated Reasoning, 32:3–33, 2004.
[156] William McCune. Solution of the Robbins Problem. Journal of Automated Reasoning,
19(3):263–276, 1997.
[157] Laura Meikle and Jacques D. Fleuriot. Formalizing Hilbert’s Grundlagen in Isabelle/Isar.
In Theorem Proving in Higher Order Logics, volume 2758 of Lecture Notes in Computer
Science, pages 319–334. Springer, 2003.
References 55

[158] Laura Meikle and Jacques D. Fleuriot. Mechanical Theorem Proving in Computational Ge-
ometry. In Hoon Hong and Dongming Wang, editors, Proceedings of Automated Deduc-
tion in Geometry 2004, volume 3763 of Lecture Notes in Computer Science, pages 1–18.
Springer-Verlag, November 2005.
[159] Richard S Millman and George D Parker. Geometry, A Metric Approach with Models.
Springer Science & Business Media, 1991.
[160] Leonardo Mendonça de Moura and Nikolaj Bjørner. Z3: An Efficient SMT Solver. In
C. R. Ramakrishnan and Jakob Rehof, editors, Tools and Algorithms for the Construction
and Analysis of Systems, 14th International Conference, TACAS 2008, Held as Part of the
Joint European Conferences on Theory and Practice of Software, ETAPS 2008, Budapest,
Hungary, March 29-April 6, 2008. Proceedings, volume 4963 of Lecture Notes in Computer
Science, pages 337–340. Springer, 2008.
[161] Julien Narboux. A Decision Procedure for Geometry in Coq. In Slind Konrad, Bunker
Annett, and Gopalakrishnan Ganesh, editors, Proceedings of TPHOLs’2004, volume 3223
of Lecture Notes in Computer Science. Springer-Verlag, 2004.
[162] Julien Narboux. Formalisation et automatisation du raisonnement géométrique en Coq. PhD
thesis, Université Paris Sud, September 2006.
[163] Julien Narboux. A Graphical User Interface for Formal Proofs in Geometry. Journal of
Automated Reasoning, 39(2):161–180, 2007.
[164] Julien Narboux. Mechanical Theorem Proving in Tarski’s Geometry. In Francisco Botana
Eugenio Roanes Lozano, editor, Post-Proceedings of Automated Deduction in Geometry
2006, volume 4869 of Lecture Notes in Computer Science, pages 139–156, Pontevedra,
Spain, 2008. Springer.
[165] Julien Narboux and David Braun. Towards A Certified Version of the Encyclopedia of Tri-
angle Centers. In J. Rafael Sandra, Dongming Wang, and Jing Yang, editors, Special Issue
on Geometric Reasoning, pages 1–17. Springer, 2016.
[166] A.J. Nevis. Plane geometry theorem proving using forward chaining. Artificial Intelligence,
6(1):1–23, 1975.
[167] Mladen Nikolić and Predrag Janičić. CDCL-Based Abstract State Transition System for
Coherent Logic. In Jeuring J. et al., editor, Intelligent Computer Mathematics - CICM 2012,
volume 7362 of Lecture Notes in Computer Science. Springer, 2012.
[168] Tobias Nipkow, Lawrence C. Paulson, and Markus Wenzel. Isabelle HOL:
a Proof Assistant for Higher-Order Logic. Springer, 2005. \sc url: \tt
http://www.cl.cam.ac.uk/research/hvg/Isabelle/dist/Isabelle/doc.
[169] Hans de Nivelle and Jia Meng. Geometric Resolution: A Proof Procedure Based on Finite
Model Search. In Automated Reasoning, Third International Joint Conference, IJCAR, vol-
ume 4130 of Lecture Notes in Computer Science, pages 303–317. Springer, 2006.
[170] Victor Pambuccian. Groups and Plane Geometry. Studia Logica, 81(3):387–398, 2005.
[171] Victor Pambuccian. Axiomatizations of hyperbolic and absolute geometries. In Non-
Euclidean Geometries, volume 581, pages 119–153. Springer, 2006.
[172] Victor Pambuccian. Axiomatizing geometric constructions. Journal of Applied Logic,
6(1):24–46, 2008.
[173] Victor Pambuccian. The axiomatics of ordered geometry: I. Ordered incidence spaces. Ex-
positiones Mathematicae, 29(1):24–66, 2011.
56 References

[174] Moritz Pasch. Vorlesungen über neuere Geometrie. Teubner, Leipzig, 1882.
[175] Lawrence C. Paulson and Tobias Nipkow. Isabelle Tutorial and User’s Manual. Technical
Report 189, University of Cambridge, Computer Laboratory, January 1990.
[176] Giuseppe Peano. Principii de Geometria. Fratelli Bocca, Torino, 1889.
[177] Danijela Petrović and Filip Marić. Formalizing analytic geometries. In Proceedings of Au-
tomated Deduction in Geometry, September 2012.
[178] Tuan Minh Pham. Similar triangles and orientation in plane elementary geometry for Coq-
based proofs. In Proceedings of the 2010 ACM Symposium on Applied Computing, pages
1268–1269. ACM, 2010.
[179] Tuan Minh Pham. Formal Description of Geometrical Properties. Theses, Univeristé Nice
Sophia Antipolis, November 2011.
[180] Tuan Minh Pham and Yves Bertot. A Combination of a Dynamic Geometry Software With a
Proof Assistant for Interactive Formal Proofs. Electron. Notes Theor. Comput. Sci., 285:43–
55, September 2012.
[181] Tuan Minh Pham, Yves Bertot, and Julien Narboux. A Coq-based Library for Interactive and
Automated Theorem Proving in Plane Geometry. In Proceedings of the 11th International
Conference on Computational Science and Its Applications (ICCSA 2011), volume 6785 of
Lecture Notes in Computer Science, pages 368–383. Springer-Verlag, 2011.
[182] David Pichardie and Yves Bertot. Formalizing Convex Hulls Algorithms. In Proc. of 14th In-
ternational Conference on Theorem Proving in Higher Order Logics (TPHOLs’01), volume
2152 of Lecture Notes in Computer Science, pages 346–361. Springer-Verlag, 2001.
[183] Jan von Plato. The axioms of constructive geometry. Annals of Pure and Applied Logic,
76:169–200, 1995.
[184] Jan von Plato. A constructive theory of ordered affine geometry. In Indagationes Mathemat-
icae, volume 9, pages 549–562. 1998.
[185] Andrew Polonsky. Proofs, Types and Lambda Calculus. PhD thesis, University of Bergen,
2011.
[186] Loı̈c Pottier. Connecting Gröbner Bases Programs with Coq to do Proofs in Algebra, Geom-
etry and Arithmetics. In G. Sutcliffe, P. Rudnicki, R. Schmidt, B. Konev, and S. Schulz, edi-
tors, Knowledge Exchange: Automated Provers and Proof Assistants, volume 418 of CEUR
Workshop Proceedings, Doha, Qatar, 2008. CEUR-WS.org.
[187] François Puitg. Preuves en modélisation géométrique par le calcul des constructions induc-
tives. PhD thesis, Université de Strasbourg, 1999.
[188] FranÇois Puitg and Jean FranÇois Dufourd. Formal specification and theorem proving break-
throughs in geometric modeling. In Jim Grundy and Malcolm Newey, editors, Theorem
Proving in Higher Order Logics: 11th International Conference, TPHOLs’98 Canberra, Aus-
tralia September 27–October 1, 1998 Proceedings, pages 401–422, Berlin, Heidelberg, 1998.
Springer Berlin Heidelberg. DOI: 10.1007/BFb0055149.
[189] Art Quaife. Automated Development of Tarski’s Geometry. Journal of Automated Reason-
ing, 5(1):97–118, 1989.
[190] Pedro Quaresma. Thousands of Geometric Problems for Geometric Theorem Provers
(TGTP). In Pascal Schreck, Julien Narboux, and Jürgen Richter-Gebert, editors, Automated
Deduction in Geometry, volume 6877 of Lecture Notes in Computer Science, pages 169–181.
Springer, 2011.
References 57

[191] Pedro Quaresma. Towards an Intelligent and Dynamic Geometry Book. Mathematics in
Computer Science, pages 1–11, 2017.
[192] Marco Riccardi. Heron’s Formula and Ptolemy’s Theorem. Formalized Mathematics, 16(1-
4):97–101, 2008.
[193] William Richter. Formalizing Rigorous Hilbert Axiomatic Geometry Proofs in the Proof
Assistant Hol Light.
[194] William Richter, Adam Grabowski, and Jesse Alama. Tarski Geometry Axioms. Formalized
Mathematics, 22(2):167–176, 2014.
[195] Jürgen Richter-Gebert. Meditations on Ceva’s Theorem. In Ellers E. Davis, C., editor, The
Coxeter Legacy: Reflections and Projections, pages 227–254. American Mathematical Soci-
ety, 2006.
[196] John Alan Robinson and Andrei Voronkov, editors. Handbook of Automated Reasoning (Vol
2). Elsevier and MIT Press, 2001.
[197] Judit Robu. Automated Geometric Theorem Proving. PhD thesis, RISC, Johannes Kepler
University, Linz, Austria, 2002.
[198] Ralph Romanos and Lawrence Paulson. Proving the Impossibility of Trisecting an Angle and
Doubling the Cube. 2014.
[199] Karel Rössler. Géométrie abstraite mécanisée. Publications de la Faculté des Sciences de
l’Université Charles (Praha), (134):29, 1934.
[200] Wolfram Schwabhäuser. Uber dire Vollständigkeit der elementaren euklidischen Geometrie.
Zeitschrift für mathematische Logik und Grundlagen der Mathematik, 2:137–165, 1956.
[201] Wolfram Schwabhäuser, Wanda Szmielew, and Alfred Tarski. Metamathematische Methoden
in der Geometrie. Springer-Verlag, Berlin, 1983.
[202] Phil Scott. Mechanising Hilbert’s Foundations of Geometry in Isabelle. Master Thesis,
University of Edinburgh, 2008.
[203] Phil Scott and Jacques D. Fleuriot. An Investigation of Hilbert’s Implicit Reasoning through
Proof Discovery in Idle-Time. In Automated Deduction in Geometry - 8th International
Workshop, ADG 2010, Munich, Germany, July 22-24, 2010, Revised Selected Papers, pages
182–200, 2010.
[204] Phil Scott and Jacques D. Fleuriot. A Combinator Language for Theorem Discovery. In Intel-
ligent Computer Mathematics - 11th International Conference, AISC 2012, 19th Symposium,
Calculemus 2012, 5th International Workshop, DML 2012, 11th International Conference,
MKM 2012, Systems and Projects, Held as Part of CICM 2012, Bremen, Germany, July 8-13,
2012. Proceedings, pages 371–385, 2012.
[205] Phil Scott and Jacques D. Fleuriot. Compass-free Navigation of Mazes. In James H. Dav-
enport and Fadoua Ghourabi, editors, SCSS 2016. 7th International Symposium on Symbolic
Computation in Software Science, volume 39 of EPiC Series in Computing, pages 143–155.
EasyChair, 2016.
[206] Wojciech Skaba. The Collinearity Structure. Formalized Mathematics, 1(4):3, 1990.
[207] Konrad Slind and Michael Norrish. A Brief Overview of HOL4. In Otmane Ait Mohamed,
César Muñoz, and Sofiène Tahar, editors, Theorem Proving in Higher Order Logics: 21st
International Conference, TPHOLs 2008, Montreal, Canada, August 18-21, 2008. Proceed-
ings, pages 28–32. Springer Berlin Heidelberg, Berlin, Heidelberg, 2008. DOI: 10.1007/978-
3-540-71067-7 6.
58 References

[208] Sana Stojanović, Julien Narboux, Marc Bezem, and Predrag Janičić. A Vernacular for Co-
herent Logic. In Stephen M. Watt, James H. Davenport, Alan P. Sexton, Petr Sojka, and
Josef Urban, editors, Intelligent Computer Mathematics, volume 8543 of Lecture Notes in
Computer Science, pages 388–403. Springer International Publishing, 2014.
[209] Sana Stojanović, Vesna Pavlović, and Predrag Janičić. A Coherent Logic Based Geometry
Theorem Prover Capable of Producing Formal and Readable Proofs. In Automated Deduction
in Geometry, volume 6877 of Lecture Notes in Computer Science, pages 201–220. Springer,
2011.
[210] Sana Stojanović Durdević, Julien Narboux, and Predrag Janičić. Automated Generation of
Machine Verifiable and Readable Proofs: A Case Study of Tarski’s Geometry. Annals of
Mathematics and Artificial Intelligence, page 25, 2015.
[211] J. Strommer. Über die Kreisaxiome. Periodica Mathematica Hungarica, 4:3–16, 1973.
[212] Alfred Tarski. A Decision Method for Elementary Algebra and Geometry. University of
California Press, 1951.
[213] Alfred Tarski. What is Elementary Geometry? In P. Suppes L. Henkin and A. Tarski, edi-
tors, The axiomatic Method, with special reference to Geometry and Physics, pages 16–29,
Amsterdam, 1959. North-Holland.
[214] Alfred Tarski and Steven Givant. Tarski’s System of Geometry. The Bulletin of Symbolic
Logic, 5(2):175–214, June 1999.
[215] Andrzej Trybulec. Mizar. In Freek Wiedijk, editor, The Seventeen Provers of the World,
volume 3600 of Lecture Notes in Computer Science. Springer, 2006.
[216] Josef Urban and Robert Veroff. Experiments with State-of-the-art Automated Provers on
Problems in Tarskian Geometry. In Boris Konev, Stephan Schulz, and Laurent Simon, edi-
tors, IWIL-2015. 11th International Workshop on the Implementation of Logics, volume 40
of EPiC Series in Computing, pages 122–126. EasyChair, 2016.
[217] Dongming Wang. Elimination Procedures for Mechanical Theorem Proving in Geometry.
Ann. Math. Artif. Intell., 13(1-2):1–24, 1995.
[218] Dongming Wang. Geometry machines: From AI to SMC. In Jacques Calmet, John Campbell,
and Jochen Pfalzgraf, editors, Artificial Intelligence and Symbolic Mathematical Computa-
tion, volume 1138 of Lecture Notes in Computer Science, pages 213–239. Springer Berlin /
Heidelberg, 1996. DOI: 10.1007/3-540-61732-9 60.
[219] Dongming Wang, Xiaoyu Chen, Wenya An, Lei Jiang, and Dan Song. OpenGeo: An Open
Geometric Knowledge Base. In Hoon Hong and Chee Yap, editors, Mathematical Software
– ICMS 2014, volume 8592 of Lecture Notes in Computer Science, pages 240–245. Springer
Berlin Heidelberg, 2014. DOI: 10.1007/978-3-662-44199-2 38.
[220] Ke Wang and Zhendong Su. Automated Geometry Theorem Proving for Human-readable
Proofs. In Proceedings of the 24th International Conference on Artificial Intelligence, IJ-
CAI’15, pages 1193–1199, Buenos Aires, Argentina, 2015. AAAI Press.
[221] Markus Wenzel. Isabelle/Isar — A Versatile Environment for Human-Readable Formal Proof
Documents. PhD thesis, Institut für Informatik, Technische Universität München, 2002.
[222] Freek Wiedijk, editor. The Seventeen Provers of the World, volume 3600 of Lecture Notes in
Computer Science. Springer, 2006.
References 59

[223] Sean Wilson and Jacques D. Fleuriot. Combining Dynamic Geometry, Automated Geometry
Theorem Proving and Diagrammatic Proofs. In ETAPS Satellite Workshop on User Interfaces
for Theorem Provers (UITP), Edinburgh, 2005. Springer.
[224] Wen-Tsun Wu. On the Decision Problem and the Mechanization of Theorem-Proving in
Elementary Geometry. In Automated Theorem Proving: After 25 Years, volume 29 of Con-
temporary Mathematics, pages 213–234. American Mathematical Society, 1984.
[225] Lu Yang, Xiao-Shan Gao, Shang-Ching Chou, and Jing-Zhong Zhang. Automated Produc-
tion of Readable Proofs for Theorems in Non-Euclidian Geometries. In Dongming Wang,
editor, International Workshop on Automated Deduction in Geometry, Selected Papers, vol-
ume 1360 of Lecture Notes in Computer Science, pages 171–188, Toulouse, 1996. Springer.
[226] Lu Yang, Xiao-Shan Gao, Shang-Ching Chou, and Jing. Zhong. Zhang. Automated Proving
and Discovering of Theorems in Non-Euclidean Geometries. In Proceedings of Automated
Deduction in Geometry (ADG98), Lecture Notes in Artificial Intelligence, pages 171–188,
Berlin, Heidelberg, 1998. Springer-Verlag.
[227] Zheng Ye, Shang-Ching Chou, and Xiao-Shan Gao. Visually Dynamic Presentation of Proofs
in Plane Geometry. Journal of Automated Reasoning, 45(3):243–266, December 2009.
[228] Paul Yiu. Introduction to the Geometry of the Triangle. Department of Mathematics, Florida
Atlantic University, 2001.
[229] Jing-Zhong Zhang, Shang-Ching Chou, and Xiao-Shan Gao. Automated Production of Tra-
ditional Proofs for Theorems in Euclidean Geometry. Ann. Math. Artif. Intell., 13(1-2):109–
138, 1995.
[230] Yu Zou and Jingzhong Zhang. Automated Generation of Readable Proofs for Constructive
Geometry Statements with the Mass Point Method. In Pascal Schreck, Julien Narboux, and
Jürgen Richter-Gebert, editors, 8th International Workshop, ADG 2010, Revised Selected
Papers, volume 6877 of Lecture Notes In Computer Science, pages 221–258, Munich, 2011.
Springer.
Chapter 3
Coordinate-Free Theorem Proving in
Incidence Geometry
Jürgen Richter-Gebert
Faculty of Mathematics, Technical University of Munich, Germany

Hongbo Li
Academy of Mathematics and Systems Science, Chinese Academy of Sciences;
University of Chinese Academy of Sciences, China

CONTENTS
3.1 Incidence Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.1.1 Incidence Geometry in the Plane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.1.2 Other Primitive Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.1.3 Projective Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.2 Bracket Algebra: Straightening, Division, and Final Polynomials . . . . . . . . . . . . . . . . . . . 67
3.2.1 Bracket Algebra and Straightening . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.2.2 Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.2.3 Final Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.3 Cayley Expansion and Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.3.1 Cayley Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.3.2 Cayley Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.3.3 Cayley Expansion and Factorization in Geometric Theorem Proving . . . . . . . 76
3.3.4 Rational Invariants and Antisymmetrization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.4 Bracket Algebra for Euclidean Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.4.1 The Points I and J . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.4.2 Proving Euclidean Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

Introduction
The objects under consideration in this chapter are the classical objects of elementary geometry:
points, lines, planes, circles, conics, etc. Many geometric theorems deal with the interdependence
of properties that hold between such objects. A point may lie on a line or on a circle, a circle and
line may touch tangentially, two lines may enclose a certain angle, two points may be at a certain
distance, etc. A geometric theorem is typically stated in a way where certain such relations (the
hypotheses) imply another such relation (typically under the presence of certain non-degeneracy
assumptions). Incidence geometry exclusively deals with properties that do not refer to measure-
ments and are of purely qualitative kind. Thus, the incidence of a point and a line is a prototypical

61
62 Handbook of Geometric Constraint Systems Principles

elementary property that may occur in a theorem of incidence geometry, while the reference to a
certain angle typically is not of this type.
In what follows standard techniques based on bracket algebra and Grassmann-Cayley algebra
for the coordinate-free treatment of incidence statements and constructions will be explained. We
shall focus on the case of two-dimensional projective incidence geometry; the extension to high
dimensions is straightforward. See Section 2 of [30] for a general introduction to projective space.
For an elaborate introduction to this topic we recommend [18, 25].

3.1 Incidence Geometry

3.1.1 Incidence Geometry in the Plane

Glossary
Homogeneous coordinates: For p ∈ Kn let [v] := {λ v | λ ∈ K − {0}}. Vectors that are identified
by scalar multiples are called homogeneous coordinates. We abbreviate [(p1 , p2 , . . . , pn )] =:
(p1 : p2 : · · · : pn ).
Projective
 plane over K: For a field K let PK (points) and LK (lines) be two disjoint copies of the
set [p] | p ∈ K3 − {(0, 0, 0)} =: KP2 of equivalence classes of non-zero three-dimensional
vectors. An incidence relation IK ⊂ PK × LK is defined by [p]IK [l] ⇐⇒ hp, li = 0, where
the angular brackets denote the pairing between vector space K3 and its dual. The triple
(PK , LK , IK ) is the projective plane over K.

Standard embedding: For a point (x, y) ∈ K2 its homogenization is defined as (x : y : 1). For a
line l ⊂ K2 defined by the equation ax + by + c = 0 its homogeneous coordinates are defined
by (a : b : c). Incidences between points and lines in K2 translate to incidences in the projective
plane of the standard embedding.

We consider the projective plane (PK , LK , IK ) over a field K. Points in that plane are represented
by their homogeneous coordinates [p] ∈ PK . By this each point p is represented by a non-zero vector
(x, y, z) such that two vectors represent the same point if and only if they differ by a non-zero scalar
multiple λ . Equivalently the points are identified with the one-dimensional linear subspaces of K3 .
Since most operations and relations that we will consider are represented by linear operations, it is
often sufficient to perform the calculations in terms of the representants. When we speak of the point
p with p ∈ K3 we mean the projective point represented by the equivalence class [p]. Sometimes it
is necessary to refer to the vector p itself.
In the case of K = R it is instructive to consider the points as the intersection of the correspond-
ing one-dimensional linear subspaces with the unit sphere S2 := {(x, y, z) | x2 + y2 + z2 = 1}. In this
view each point in RP2 corresponds to an antipodal pair of points on the unit sphere. A line [l] ∈ LK
may be identified with the two-dimensional subspace of K3 normal to the vector l. Incidence be-
tween a point (x : y : z) and a line (a : b : c) corresponds to incidence between the corresponding
subspaces and is algebraically resembled by the vanishing of the scalar product ax + by + zc = 0. In
the case of K = R a subspace representing a line may be intersected with the unit sphere and by this
corresponds to a great circle on S2 .
In the standard embedding each point (x, y) ∈ R2 is represented by its homogenization (x : y : 1).
This covers almost all points of PR except for those of the form (x : y : 0). They correspond to
points that are infinitely far away and have no direct correspondence in R2 . For each direction
Coordinate-Free Theorem Proving in Incidence Geometry 63

(equivalence class of parallels in R2 ) there is exactly one such point at infinity incident to all lines
of a parallel class. All infinite points lie on a common line with parameter (0 : 0 : 1) which is the
line at infinity. Literally the same interpretation applies to all fields K. See Section 2.2 of [30] for
more on homogeneous coordinates and points at infinity.

Glossary
Join: For different points [p], [q] ∈ PK the unique line in LK incident to both. Can be calculated by
meet(p, q) = p × q, where the cross symbol denotes the vector product in vector algebra.
Meet: For different lines [l], [m] ∈ LK the unique point in PK incident to both. Can be calculated
by join(l, m) = l × m.
Grassmann-Cayley algebra: A non-associative algebra with two multilinear and graded-
antisymmetric products: the wedge and vee products. For example, the operations meet and
join are denoted by the wedge and vee products, respectively:

meet(p, q) = p ∧ q and join(l, m) = l ∨ m.

They satisfy the following graded-antisymmetric properties: for p of grade d p and q of grade
dq , if n is the dimension of the base vector space, then

p ∧ q = (−1)(n−d p )(n−dq ) q ∧ p; p ∨ q = (−1)d p dq q ∨ p.

As to the grade of an element, a scalar is of grade 0, a nonzero vector in the base vector space is
of grade 1, and the wedge product of r elements of grade d1 , d2 , . . . , dr respectively, if nonzero,
is of grade d1 + d2 + · · · + dr .
Collinear: Three points [p], [q], [r] ∈ PK are collinear if they are incident to the same line. This
happens if and only if det(p, q, r) = 0.
Concurrent: Three lines [l], [m], [g] ∈ PK are concurrent if they are incident to the same point. This
happens if and only if det(l, m, g) = 0.
Bracket: Shorthand for a determinant: [p, q, r] := det(p, q, r).
Non-degeneracy condition: A relation indicating that a certain dependency (typically an incidence
or a higher algebraic dependency) does not occur.
Incidence theorem: A true statement of the form (H1 , . . . Hk , N1 , . . . , Nr ) =⇒ C, where Hi (the
hypotheses) are statements of incidence, Ni are non-degeneracy conditions and C (the conclu-
sion) is an incidence statement.

If one considers incidence relations from a structural point of view, the projective plane is way
more natural than the usual plane K2 . While in K2 it may happen that two affine lines do not have an
intersection (in case they are parallel), in the projective plane over K two distinct lines will always
have exactly one projective point in common: their meet. Likewise any pair of distinct points in KP2
have exactly one line incident to both of them: their join. (Remark: in this way the points and lines
of KP2 satisfy the usual axioms of an abstract projective plane claiming that join and meet always
exist and are unique; see Section 4 of [30] for an algebraic definition of the meet). The fact that p is
incident to l if and only if the scalar product hp, li vanishes directly translates into the fact that the
join and meet operations can be algebraically expressed by the vector product in K3 . See Section
4.2 of [30] for an explicit algebraic computation relating the join and vector products.
64 Handbook of Geometric Constraint Systems Principles

If the vector product p×q turns out to be the zero vector, then p and q must be linearly dependent
and hence indicate an inadmissible (degenerate) operation of joining two identical points.
Collinearity of three points p1 , p2 , p3 corresponds to the existence of a line that is simultaneously
incident to all three points. In this case the points are linearly dependent. Hence collinearity can be
expressed by the vanishing of the determinant det(p1 , p2 , p3 ). Similarly, the fact that three lines
l1 , l2 , l3 meet at a common point corresponds to the vanishing of the determinant det(l1 , l2 , l3 ). For
better readability we will abbreviate a 3 × 3 determinant det(a, b, c) by the bracket notation [a, b, c].
A minimal instance of a theorem of projective incidence geometry is Pappus’ Theorem (Figure
3.1, left). In textbooks this theorem is often naively stated in the following way.
If a, b, c and A, B,C are two triples of collinear points, then the points Z = (a ∨ B) ∧ (b ∨ A), X =
(b ∨C) ∧ (c ∨ B), Y = (c ∨ A) ∧ (a ∨C) are collinear as well.

Figure 3.1
The theorems of Pappus and Pascal.

However, a bit of care is necessary here. It might happen that the initial six points are located in
a way that some of the join and meet operators are degenerate. In this case the above formulation
makes no sense. Thus, it is a wise choice to furthermore require that each join and meet operation is
indeed non-degenerate. (This can be achieved by imposing that a, b, c, A, B,C are mutually different.)
In principle the algebra of meet and join operations is very well suited to perform direct cal-
culations in which three-dimensional vectors represent geometric points and lines. One crucial in-
gredient of such calculations comes from the fact that nested vector products can be reformulated
as linear combinations of expressions involving geometric objects and determinants. The following
very useful formula can be easily proved by expansion of the left and right sides:

(p ∨ q) ∧ (r ∨ s) = [p, q, r] s − [p, q, s] r. (3.1)

A first approach to proving Pappus’ Theorem is by transforming the join and meet operations
into pure bracket expressions. We want to prove the collinearity of the points X,Y, Z under the hy-
potheses of the theorem. We check whether the determinant [X,Y, Z] vanishes. From the construction
we obtain:
 
[X,Y, Z] = [a, B, b]A − [a, B, A]b, [b,C, c]B − [b,C, B]c, [c, A, a]C − [c, A,C]a .
Coordinate-Free Theorem Proving in Incidence Geometry 65

Expanding this formula we obtain eight summands:

+[a, B, b][b,C, c][c, A, a][A, B,C] −[a, B, b][b,C, c][c, A,C][A, B, a]

| {z }
z }| {
−[a, B, A][b,C, c][c, A, a][b, B,C] −[a, B, b][b,C, B][c, A, a][A, c,C]
−[a, B, A][b,C, B][c, A,C][a, b, c] +[a, B, b][b,C, B][c, A,C][A, c, a]
z }| {
+[a, B, A][b,C, c][c, A,C][b, B, a] +[a, B, A][b,C, B][c, A, a][b, c,C].
| {z }

Observe that the six terms cancel pairwise by summands of the same underbrace, or lowerbrace, or
underline, and we are left with the expression:

+[a, B, b][b,C, c][c, A, a][A, B,C] − [a, B, A][b,C, B][c, A,C][a, b, c].

This expression vanishes if the hypotheses of Pappus’ Theorem are satisfied since in this case
the two underlined determinants must be zero.
A closer look at this proof of Pappus’ Theorem reveals two principal problems with this kind
of coordinate-free calculation. First of all the calculation heavily depends on the fact that there is
a nice cancellation pattern between the expanded bracket monomials. However, it might happen
that bracket expressions vanish for less obvious reasons than simply being equal. We will deal with
this phenomenon in depth in Section 3.2.1. The second issue that arises comes from the fact that
this technique heavily relied on the expansion of the multiplication of sums of terms. This may
easily lead to an exponential blowup of the calculations in the middle of the calculation (before
cancellation occurs). Already for theorems slightly larger than Pappus’ Theorem this may lead to
practical intractability of such calculations even with the aid of computers.

3.1.2 Other Primitive Operations

One might think that building a system that exclusively uses points, lines, and incidences as prim-
itive objects and relations leads to a relatively restricted expressiveness. Fortunately this is not the
case. More elaborate constructions (or equivalently algebraic expressions) may be used to express
more advanced geometric relations. To exemplify this let us consider the expression

+[a, B, b][b,C, c][c, A, a][A, B,C] − [a, B, A][b,C, B][c, A,C][a, b, c]

we obtained in our previous calculation. To get a more symmetric form of this formula we apply a
cyclic shift to the variable names a → c → b → a and change it into

+[c, B, a][a,C, b][b, A, c][A, B,C] − [c, B, A][a,C, B][b, A,C][c, a, b].

Reordering the indices in the brackets yields the following very symmetric form:

+[a, b, c][a, B,C][A, b,C][A, B, c] − [A, B,C][A, b, c][a, B, c][a, b,C] (∗)

We have seen that this expression vanishes if both triples (a, b, c) and (A, B,C) are collinear. How-
ever this is not the only situation in which this expression vanishes. In fact, vanishing of this expres-
sion is equivalent to the six points a, b, c, A, B,C being on a common conic.
To see this we first define a conic to be the solution of a quadratic equation in homogeneous
coordinates. A point (x, y, z)T lies on a conic given by parameters (α, β , γ, δ , ε, ζ ) if it satisfies the
equation
αx2 + β y2 + γz2 + δ xy + εxz + ζ yz = 0.
66 Handbook of Geometric Constraint Systems Principles

In the standard affine embedding, conics correspond to ellipses, hyperbolas, parabolas, or may de-
generate into pairs of lines (which may even coincide). For p = (x, y, z)T we define the quadratic
operator
sq(p) := (x2 , y2 , z2 , xy, xz, yz).
We first observe that if for five points p1 , . . . , p5 the vectors sq(p1 ), . . . , sq(p5 ) are linearly indepen-
dent then there is a unique conic passing through these five points, which is given by the solution
of a system of linear equations. A sixth point lying on this conic must satisfy a quadratic equation.
Now let us consider the above equation (∗). We first observe that each point occurs exactly twice
in each expression. Hence if we fix five of the points (say a, b, c, A, B) the expression (if suitably
non-degenerate) being zero gives a quadratic relation in the last point C. Here the non-degeneracy
condition is that the formula does not automatically vanish for all possible choices of C. This can
be achieved by making sure that no quadruple of the points a, b, c, A, B are collinear (we omit the
proof of this fact here). It turns out that the quadratic equation in C describes the unique conic that
passes through the other five points. For this it is sufficient to show that if C is any of the points
in the set {a, b, c, A, B} then the expression (∗) will vanish. This is trivial for C ∈ {a, b, A, B} since
in this case one determinant in each summand becomes zero. If C = c then it is easy to check that
both summands are identical and hence their difference will vanish as well. We can rephrase the last
result in the following way, that is essentially an algebraic formulation of Pascal’s Theorem [25]
(Figure 3.1, right):
Six points a, b, c, A, B,C are on a common conic if the expression (∗) vanishes. This is the case if
Z = (a ∨ B) ∧ (b ∨ A), X = (b ∨C) ∧ (c ∨ B), Y = (c ∨ A) ∧ (a ∨C) are collinear.

3.1.3 Projective Invariance

Glossary
Projective transformation: A map τ : PK → PK induced by the map p 7→ M · p for a non-singular
matrix M. The action of this map is well defined.
Collineation: A map τ : PK → PK is a collineation if three points a, b, c are collinear whenever so
are τ(a), τ(b), τ(c) and vice versa. Over R every collineation is a projective transformation. In
general a collinearity is a projective transformation composed with a field automorphism.
Projective property: Consider the space (PK )n of configurations of n points. A property among
these points is a subset X ⊆ (PK )n of this space with the interpretation that a point configuration
has the property X if it lies in this subset.
Projectively invariant property: A property X of n points with
([p1 ], . . . , [pn ]) ∈ X ⇔ ([M · p1 ], . . . , [M · pn ]) ∈ X,
for all projective transformations M.
Determinant map: Let E = {1, . . . , n} be the index set of a list of vectors P = (p1 , . . . , pn ) with
underlying point configuration ([p1 ], . . . , [pn ]). The determinant map associates to every triple
the corresponding determinant
ΞP : E 3 → K
(i, j, k) 7→ [pi , p j , pk ]

Proper configuration: A configuration P is proper if not all points are collinear. In this case the
map ΞP is not identically zero.
Coordinate-Free Theorem Proving in Incidence Geometry 67

It turns out that projectively invariant properties are in close relation to the determinants of a
point configuration.

Theorem 3.1 Up to projective equivalence a proper configuration ([p1 ], . . . , [pn ]) of P =

(p1 , . . . , pn ) is uniquely determined by the map ΞP .

A proof of this theorem may be found in [25]. The basic idea of the proof is that after fixing
a projective basis the positions of the points that satisfy the determinant values may be uniquely
reconstructed. As a consequence every projectively invariant property may be decided entirely by
the knowledge of the map Ξ. In view of this theorem it is not surprising that the fact that six points
lie on a conic can be decided by the evaluation of a polynomial in the determinants among the
vectors representing the points.

Remark 3.1 In fact there are much stronger versions of the above theorem that also preserve the
algebraic structure of such properties. In a stronger form (known as the First Fundamental Theorem
of Invariant Theory, compare [12, 32]) it states that every projectively invariant property that may
be encoded as the vanishing of a polynomial in the coordinate entries of the points may be rewritten
as the vanishing of a polynomial of the determinants among vectors representing these points. See
Section 3.1 of [30] for more on group actions and invariant polynomials.

3.2 Bracket Algebra: Straightening, Division, and Final Polyno-

mials
3.2.1 Bracket Algebra and Straightening

Glossary
Grassmann-Plücker relation in K3 : Let a, b, c, d, e, f be six vectors in K3 then

[a, b, c][d, e, f ] − [a, b, d][c, e, f ] + [a, b, e][c, d, f ] − [a, b, f ][c, d, e] = 0.

Similar expressions hold for all dimensions.

Formal bracket: A symbol [[i jk]] with i, j, k ∈ E = {1, 2, . . . , n} that will play the role of variables
mimicking the role of formal determinants. The comma punctuation among the indices will be
used only when an expression occurs as a vector variable in the bracket.
Bracket polynomial: A polynomial in the polynomial ring

R := K[ [[i jk]]; (i, j, k) ∈ E 3 ].

Alternating ideal: An ideal mimicking the alternating determinant rules

D    E
Alt := [[i jk]] + [[ jik]]  i, j, k ∈ E ∪ [[i jk]] + [[ik j]]  i, j, k ∈ E .
 

In the quotient ring R = R/Alt brackets that have the same index set are identified (with appro-
priate sign changes in accordance to the alternating determinant rules).
68 Handbook of Geometric Constraint Systems Principles

Grassmann-Plücker ideal: An ideal in R mimicking the Grassmann-Plücker relations.

D  E
GP := [[abc]][[de f ]] − [[abd]][[ce f ]] + [[abe]][[cd f ]] − [[ab f ]][[cde]]  a, b, c, d, e, f ∈ E .


Bracket ring (or algebra): The quotient [35]

BR := R/hAlt ∪ GPi = R/hGPi.

In the light of Remark 3.1 it is a natural step to express properties of projective geometry entirely
on the basis of determinants (brackets) among the vectors representing the points of a configuration.
For this a natural question to ask is what kind of maps Ξ : E 3 → K can be created as images from
point configurations.
In fact the determinant values satisfy polynomial relations, the so-called Grassmann-Plücker
relations that can be easily proved by expansion of the brackets. They impose natural restrictions on
the map ΞP that comes from a vector configuration P. If Ξ comes from a configuration then it must
satisfy all Grassmann-Plücker relations on subsets of points.
The Second Fundamental Theorem of Invariant Theory [22, 32] provides a close relation of the
possible maps Ξ that come from point configurations and Grassmann-Plücker relations.

Theorem 3.2 Let Ξ : E 3 → K be an alternating map that does not vanish identically, i.e., for all
indices a, b, c ∈ E, Ξ(a, b, c) = −Ξ(b, a, c) = −Ξ(c, b, a) = −Ξ(a, c, b), and there exist a, b, c ∈ E
such that Ξ(a, b, c) 6= 0. If for all indices a, b, c, d, e, f ∈ E,

Ξ(a, b, c)Ξ(d, e, f ) − Ξ(a, b, d)Ξ(c, e, f ) + Ξ(a, b, e)Ξ(c, d, f ) − Ξ(a, b, f )Ξ(c, d, e) = 0

then there is a vector configuration P with Ξ = ΞP .

The close relation of Grassmann-Plücker relations to possible maps Ξ provides us with the
possibility to describe a projective configuration entirely in terms of (formal) determinants. For this
we fix an index set E = {1, 2, . . . , n} and mimic the behavior of determinants by formal symbols.
Since determinants coming from a vector configuration are characterized by the fact that they
are alternating and satisfy the Grassmann-Plücker relations, the bracket ring consists of bracket
polynomials such that two polynomials are identified if they evaluate to the same value for every
vector configuration. To make this statement more precise we consider the map ΞP and reinterpret it
with the formal brackets as arguments by setting ΞP ([[i jk]]) := ΞP (i, j, k). We extend this map by the
natural homomorphism to polynomials in R by setting ΞP (x + y) = ΞP (x) + ΞP (y) and ΞP (x · y) =
ΞP (x) · ΞP (y). The bracket ring has the following crucial property:
Theorem 3.3 Let X and Y be polynomials in R. The following statements are equivalent:

(i) ΞP (X) = ΞP (Y ) for all vector configurations P.

(ii) X = Y in the bracket ring BR.
For example, consider the two bracket polynomials

X = [[abc]][[de f ]] − [[abd]][[ce f ]] and

Y = [[abe]][[cd f ]] − [[ab f ]][[cde]].

They are equal to each other in BR since their sum is a Grassmann-Plücker relation. If the corre-
sponding determinant polynomial ΞP (X) vanishes for a configuration P then so does ΞP (Y ). Both
polynomials encode the same projectively invariant property (if they vanish), namely that the lines
Coordinate-Free Theorem Proving in Incidence Geometry 69

join(a, b), join(c, d), join(e, f ) meet. See Section 3.2 of [30] for an explicit algebraic description of
the relations on brackets.
As there are algebraic relations among the formal brackets taken as arguments, two drastically
different polynomials in the same bracket ring may still be equal to each other. The normal form
of a bracket polynomial f refers to another bracket polynomial N( f ) such that any two bracket
polynomials f , g are equal if and only if N( f ) and N(g) are identical. The normal form N( f ) of f is
said to be straight, and any algorithm changing f into N( f ) for arbitrary input bracket polynomial
f is called a straightening algorithm [1, 2].
The term “straight” comes from the following unique property of bracket monomials in normal
form. A bracket monomial
f = [[a1 b1 c1 ]][[a2 b2 c2 ]] · · · [[an bn cn ]]
can be written in the following tableau form by piling up its bracket arguments:
 
a1 b1 c1
 a2 b2 c2 
f = . ..  .
 
..
 .. . . 
an bn cn

Given a total order among the vector variables (or indices of vector variables bearing the same
name) ai , b j , ck , then f is in normal form if and only if in its tableau form, along each row, the order
is increasing, and along each column, the order is non-decreasing.   
d1 d3 d5 d1 d4 d5
For example, when d1 ≺ d2 ≺ · · · ≺ d5 , then is straight, while
d2 d4 d5 d2 d3 d5
is not, because its second column is not “straight.”
The classical Young’s straightening algorithm [36, 37] is based on the so-called van der Waerden
relations. They include not only the Grassmann-Plücker relations, but also the following:

[[ab1 b2 ]][[b3 b4 c]] − [[ab1 b3 ]][[b2 b4 c]] + [[ab1 b4 ]][[b2 b3 c]]

(3.2)
+[[ab2 b3 ]][[b1 b4 c]] − [[ab2 b4 ]][[b1 b3 c]] + [[ab3 b4 ]][[b1 b2 c]] = 0.

By the Grassmann-Plücker relations, the two lines on the left side of the above equality are both
equal to [[ab1 c]][[b2 b3 b4 ]], so the van der Waerden relations are in the ideal hGPi.
In [33], it is shown that all van der Waerden relations form a Gröbner basis of hGPi, and the
classical Young’s straightening algorithm is a top reduction procedure with respect to the Gröbner
basis. When the order among bracket monomials is the row-deglex order (to be introduced below),
then Young’s straightening algorithm reduces the row-deglex order of the input bracket polynomial
every time a van der Waerden relation is employed.

Glossary
Straight bracket monomial: A bracket monomial in tableau form where along each row from left
to right, the order of vector variables is increasing, and along each column from top to bottom,
the order of vector variables is non-decreasing. The procedure changing a bracket polynomial
into its straight form (normal form) is called straightening.
van der Waerden relation: An identity in the bracket ring obtained from the fact that the antisym-
metric tensor product of any four vector variables b1 , b2 , b3 , b4 is zero. Given two other vector
variables a, c, there are six different ways of dividing the 4-tuple of vector variables into two
pairs, with the first pair allocated to a and the second pair allocated to c, such that each 3-tuple
form a bracket, and a bracket monomial of degree two is generated by each allocation. As the
antisymmetric tensor product of the original 4-tuple is zero, so is the signed sum (3.2) of the
degree-2 bracket monomials over all possible allocations.
70 Handbook of Geometric Constraint Systems Principles

Row-deglex order: A total order among such monic bracket monomials that within each bracket,
the order of vector variables is increasing. The row-deglex order is the degree lexico-
graphical order among the resulting sequences of vector variables obtained by removing
all bracket  symbols fromthe monicbracket monomials. For example, if a1 ≺ a2 ≺ . . . ≺
a1 a4 a1 a5
a6 , then  a2 a5  ≺  a2 a3  in row-deglex order, because when scanned by row,
a3 a6 a4 a6
a1 a4 a2 a5 a3 a6 ≺ a1 a5 a2 a3 a4 a6 degree lexicographically.
Admissible order: An admissible order among straight monic bracket monomials is a total order
that is preserved by multiplication and then straightening. For example, the row-deglex order is
not an admissible order among straight monic bracket monomials.

3.2.2 Division

Glossary
Negative column-deglex order: A total order among monic bracket monomials where within each
bracket, the order of vector variables is increasing. It is the negative of the degree lexicograph-
ical order among the resulting sequences of vector variables obtained by scanning the tableau
forms of the monic bracket monomials columnwise
  from the first column to the last. For exam-
a1 a4 a1 a5
ple, if a1 ≺ a2 ≺ . . . ≺ a6 , then  a2 a5    a2 a3  in negative column-deglex order,
a3 a6 a4 a6
because when scanned by column, a1 a2 a3 a4 a5 a6 ≺ a1 a2 a4 a5 a3 a6 degree lexicographically.

Columnwise straightening: Given a monic bracket monomial f , its columnwise straightening is

denoted by f  , and is a monic bracket monomial obtained from the tableau form of f by sorting
the entries of each column of f so that the entries become non-decreasing within each column.
The operator “ ” can be extended linearly to all bracket polynomials.

Columnwise reduction: Given two monic bracket monomials f , g, for any column of g, if all its
entries belong to the same column of f , then f is said to be columnwise reducible with respect
to g. Let h be the monic bracket monomial whose tableau form is obtained from that of f by
removing the entries of g columnwise from f , called the columnwise quotient of f with respect
to g. Then f = gh + r for some bracket polynomial r called the columnwise remainder, and the
procedure is called the columnwise reduction of f by g.

Homogeneous bracket polynomial: A bracket polynomial that is homogeneous in its every vector
variable.

   
d1 d4 d5 d1 d3 d5
For example, when f = , then f  = ; it is clearly straight.
  d2 d3 d5 d2 d4 d5
Given g = d2 d4 d5 , then
 f is columnwise reducible with respect to g, the columnwise
quotient is h = d1 d3 d5 , and the columnwise remainder is
     
d1 d4 d5 d1 d3 d5 d1 d2 d5
r = f − gh = − =− .
d2 d3 d5 d2 d4 d5 d3 d4 d5

The last step follows a three-summand Grassmann-Plücker relation.

Coordinate-Free Theorem Proving in Incidence Geometry 71

Theorem 3.4 [8] Given a homogeneous bracket polynomial f , the leading term of its normal form
N( f ) under the negative column-deglex order is the same with the leading term of f  . In particular
when f is a bracket monomial, then the leading term of N( f ) is f  , and the latter is always straight.

Theorem 3.5 [19] The negative column-deglex order is an admissible order among straight ho-
mogeneous bracket polynomials: let f , g be two straight homogeneous bracket polynomials and
f ≺ g in the negative column-deglex order, then for any nonzero homogeneous bracket polynomial
h, N( f h) ≺ N(gh).

Given two homogeneous bracket polynomials f , g, the division of f by g is defined as the re-
duction of f with respect to hg, GPi. Usually this can be done by first computing a Gröbner basis
GB [31, 41] of the ideal and then using it to make top reduction to f . The number of elements of
GB is usually greater than one, and the elements cannot be predicted from the explicit form of g,
so the division by Gröbner basis is dramatically different from the canonical division between two
polynomials by top reduction, where the divisor contains only one polynomial all the time.
Now that the negative column-deglex order “≺” is an admissible order among straight homoge-
neous bracket polynomials, for two homogeneous bracket polynomials f  g in straight form, the
columnwise reduction of the leading term lt( f ) of f by the leading term lt(g) of g results in a colum-
nwise quotient h that is a bracket monomial, such that lt( f ) = h · lt(g) + r, where the columnwise
remainder r satisfies N(r) ≺ lt( f ). Then f = h · g + r̃, where

r̃ = r + ( f − lt( f )) − h · (g − lt(g))

satisfies N(r̃) ≺ lt( f ). This is called an invariant top reduction of f by g. The invariant top reduction
of N(r̃) by g can be done similarly. By successive invariant top reductions with respect to g, we get
f = h0 · g + r0 , where h0 , r0 are bracket polynomials, N(r0 ) ≺ f , and lt(N(r0 )) is not columnwise
reducible with respect to lt(g). This procedure is called an invariant division of f by g, and h0 , r0 are
called the invariant quotient and invariant remainder, respectively.

3.2.3 Final Polynomials

Glossary
Non-degeneracy monomial: A bracket monomial N ∈ R such that N = b1 ·b2 · · · bl with bd = [[i jk]]
and (i, j, k) 6∈ T for all d ∈ {1, . . . , l}, where T is a finite set of collinear triples.
T -vanishing polynomial: A bracket monomial A ∈ R in the ideal
D E
{[[i jk]] | (i, j, k) ∈ T } .

Final polynomial: A polynomial X ∈ GP such that X = A + N for a T -vanishing polynomial A

and a non-degeneracy monomial N.

Being able to express projective geometry entirely on the level of determinants implies that
there must be ways to prove geometric theorems without ever referring to concrete points and their
specific coordinates. Usually such coordinate-free calculations tend to be considerably shorter and
more instructive than calculations that refer to a concrete embedding and to a particular choice
of a basis. The technique for this is called final polynomials and was introduced by Bokowski and
Sturmfels [6] and independently Whiteley [40] in the context of non-realizability proofs for matroids
and oriented matroids.
72 Handbook of Geometric Constraint Systems Principles

We will start by exemplifying the method of final polynomials with the example of a proof
of Pappus’ Theorem. For this we consider the following version of Pappus’ Theorem which in its
statement is slightly more symmetric than our original formulation. Our final polynomial will in
particular be of the type of bi-quadratic final polynomials which are algorithmically easier to find
[5, 24] and admit additional structural properties [26].
Let 1, 2, . . . , 9 be the indices of nine points in the projective plane. Consider the nine triples of points
(1, 2, 3), (1, 5, 9), (1, 6, 8), (4, 5, 6), (4, 3, 8), (4, 2, 9), (7, 2, 6), (7, 5, 3) and (7, 8, 9). Under the non-
degeneracy assumptions that all other triples are non-collinear, the collinearity of eight of the nine
triples implies the collinearity of the last one.
To give a proof that is entirely based on calculations with determinants consider the following
polynomial X in formal determinants. It is a linear combination of Grassmann-Plücker relations.
Hence it must be zero in the bracket ring BR. Equivalently we can state that the evaluation ΞP (X)
(where we replace each formal bracket with the value of the corresponding determinant) must vanish
for every point configuration P.


X = + [[714]][[735]] − [[713]][[745]] + [[715]][[743]] · [[148]][[127]][[149]][[467]]


+ [[471]][[438]] − [[473]][[418]] + [[478]][[413]] · [[157]][[127]][[149]][[467]]


+ [[147]][[132]] − [[143]][[172]] + [[142]][[173]] · [[157]][[478]][[149]][[467]]


− [[147]][[195]] − [[149]][[175]] + [[145]][[179]] · [[478]][[124]][[137]][[467]]


− [[714]][[798]] − [[719]][[748]] + [[718]][[749]] · [[145]][[124]][[137]][[467]]


− [[471]][[492]] − [[479]][[412]] + [[472]][[419]] · [[145]][[178]][[137]][[467]]


+ [[471]][[465]] − [[476]][[415]] + [[475]][[416]] · [[178]][[247]][[149]][[137]]


+ [[147]][[168]] − [[146]][[178]] + [[148]][[176]] · [[457]][[247]][[149]][[137]]


+ [[714]][[762]] − [[716]][[742]] + [[712]][[746]] · [[457]][[148]][[149]][[137]].

Within the ring R we can apply alternating determinant rules and calculations with polynomials
to rewrite the above polynomial in the following way (after expanding the summands, reordering
brackets and cancelling terms):

X = +[[147]][[735]][[148]][[127]][[149]][[467]] + [[147]][[438]][[157]][[127]][[149]][[467]]
+[[147]][[132]][[157]][[478]][[149]][[467]] − [[147]][[195]][[478]][[124]][[137]][[467]]
−[[147]][[798]][[145]][[124]][[137]][[467]] − [[147]][[492]][[145]][[178]][[137]][[467]]
+[[147]][[465]][[178]][[247]][[149]][[137]] + [[147]][[168]][[457]][[247]][[149]][[137]]
+[[147]][[762]][[457]][[148]][[149]][[137]].

We know that also the above expression must evaluate to zero if we replace formal determinants
by the values of real determinants coming from any vector configuration. The underlined brack-
ets refer to the triples that play a role in our formulation of Pappus’ Theorem. Now assume that
in contradiction to Pappus’ Theorem there is a point configuration P in which only eight of the
collinearities are satisfied but not the ninth. Our non-degeneracy assumptions imply that all other
brackets are non-zero. This immediately leads to a contradiction to the above expression being zero,
since in this case all summands except for one would vanish.
The decisive point is that this proving strategy works in general and provides us with a general
tool for carrying out geometric proofs on the level of determinants. To see this consider the following
setup. Assume that we want to prove the non-existence of a configuration P which has a certain set of
Coordinate-Free Theorem Proving in Incidence Geometry 73

collinear triples T and all other triples are non-collinear (like in our proof an assumed configuration
that satisfies only eight of the Pappus triples).
The existence of a final polynomial X proves the non-existence of a configuration that has ex-
clusively the collinearities indicated by T . X being in GP implies that this polynomial evaluates to
zero for any given configuration. The property X = A + N implies it to be non-zero if the triples in
T are realized as collinear triples. Over algebraically closed fields the converse is also true.

Theorem 3.6 Over an algebraically closed field, if a configuration is not realizable then it must
admit a final polynomial.

The non-existence of a configuration in the presence of a final polynomial can also be explored in
general fields. However the converse direction relies on Hilbert’s Nullstellensatz. There are versions
of final polynomials that apply to the real numbers field. Details may be found in [6, 32].

3.3 Cayley Expansion and Factorization

3.3.1 Cayley Expansion

Glossary
Cayley bracket: Any expression in Grassmann-Cayley algebra that contains only the meet and join
operations is called a monic Cayley monomial. If in addition the expression is scalar-valued,
then it is called a monic Cayley bracket. A Cayley bracket is the scaling of a monic Cayley
bracket by a factor of K.
Cayley expansion: By the First Fundamental Theorem of Invariant Theory, any Cayley bracket
equals a homogeneous bracket polynomial. Changing a Cayley bracket into an equal bracket
polynomial is called Cayley expansion.
Binomial Cayley expansion: It refers to changing a Cayley bracket into an equal homogeneous
bracket binomial. Formula (3.1) provides the following shuffle formula of join-splitting bino-
mial Cayley expansion:

(a1 ∨ a2 ) ∧ (b1 ∨ b2 ) ∧ (c1 ∨ c2 ) = [[a1 a2 b1 ]][[b2 c1 c2 ]] − [[a1 a2 b2 ]][[b1 c1 c2 ]]

= −[[a1 b1 b2 ]][[a2 c1 c2 ]] + [[a2 b1 b2 ]][[a1 c1 c2 ]] (3.3)
= −[[a1 a2 c1 ]][[b1 b2 c2 ]] + [[a1 a2 c2 ]][[b1 b2 c1 ]].

The three different binomial Cayley expansions are obtained by alternatively distributing the
two vector variables of b1 ∨ b2 , or a1 ∨ a2 , or c1 ∨ c2 , to the other two joins, respectively. The
equality of the three different results follows Grassmann-Plücker relations.
Monomial Cayley expansion: It refers to changing a Cayley bracket into another equal expression
that is the multiplication of more than one Cayley bracket. For example in (3.3), when a1 = b1 ,
then the first two expansion results are identically the following monomial:

(a1 ∨ a2 ) ∧ (a1 ∨ b2 ) ∧ (c1 ∨ c2 ) = −[[a1 a2 b2 ]][[a1 c1 c2 ]]. (3.4)

Changing a Cayley bracket into a bracket polynomial is a procedure of eliminating the meet
74 Handbook of Geometric Constraint Systems Principles

operation, and algebraically this is a procedure of simplification, as the meet and join operations are
not associative when both occur in the same Cayley monomial. Cayley expansion has a prominent
feature that the expansion result is generally not unique, and in some cases, even for monomial
Cayley expansion.
Although formula (3.3) gives three different expansion results, all of them are by distributing
entries of a join operation to the other join operations. In duality, for some Cayley monomials one
can distribute entries of a meet operation to other meet operations [3, 9]. For example, let

f = [[(a1 ∨ a2 ) ∧ (a01 ∨ a02 ), (a1 ∨ a3 ) ∧ (a01 ∨ a03 ), (a2 ∨ a3 ) ∧ (a02 ∨ a03 )]], (3.5)

then by distributing the two entries a1 ∨ a2 and a01 ∨ a02 of the first meet operation to the other two
meet operations, we get

f = −((a1 ∨ a2 ) ∧ (a1 ∨ a3 ) ∧ (a01 ∨ a03 )) · ((a01 ∨ a02 ) ∧ (a2 ∨ a3 ) ∧ (a02 ∨ a03 ))

+((a1 ∨ a2 ) ∧ (a2 ∨ a3 ) ∧ (a02 ∨ a03 )) · ((a01 ∨ a02 ) ∧ (a1 ∨ a3 ) ∧ (a01 ∨ a03 )) (3.6)
= [[a1 a2 a3 ]][[a01 a02 a03 ]] · ([[a1 a01 a03 ]][[a2 a3 a02 ]] − [[a1 a3 a01 ]][[a2 a02 a03 ]]).

Alternatively one can use the join-splitting shuffle formula to expand (a1 ∨ a2 ) ∧ (a01 ∨ a02 ) first,
as follows:
f = −[[a1 a01 a02 ]][[a2 , a1 a3 ∨ a01 a03 , a2 a3 ∨ a02 a03 ]]
+[[a2 a01 a02 ]][[a1 , a1 a3 ∨ a01 a03 , a2 a3 ∨ a02 a03 ]]
= [[a1 a01 a02 ]][[a2 a02 a03 ]][[a2 , a1 a3 ∨ a01 a03 , a3 ]] (3.7)
−[[a2 a01 a02 ]][[a1 a01 a03 ]][[a1 , a3 , a2 a3 ∨ a02 a03 ]]
= [[a1 a2 a3 ]](−[[a1 a01 a02 ]][[a2 a02 a03 ]][[a3 a01 a03 ]] + [[a2 a01 a02 ]][[a1 a01 a03 ]][[a3 a02 a03 ]]).

Besides (3.3) and (3.4), there are other formulas for binomial Cayley expansion and monomial
Cayley expansion. For example we have the following six formulas [16]:

i. ((a1 ∨ a2 ) ∧ (a3 ∨ a4 ) ∧ (a5 ∨ a6 )) · ((a1 ∨ a3 ) ∧ (a2 ∨ a4 ) ∧ (a5 ∨ a6 ))

= [[a1 a2 a4 ]][[a1 a3 a4 ]][[a2 a5 a6 ]][[a3 a5 a6 ]] − [[a1 a2 a3 ]][[a2 a3 a4 ]][[a1 a5 a6 ]][[a4 a5 a6 ]];

ii. [[a1 , (a01 ∨ a02 ) ∧ (a03 ∨ a04 ), (a01 ∨ a02 ) ∧ (a003 ∨ a004 )]]
= [[a1 a01 a02 ]] ((a01 ∨ a02 ) ∧ (a03 ∨ a04 ) ∧ (a003 ∨ a004 ));

iii. [[(a1 ∨ a2 ) ∧ (a3 ∨ a4 ), (a1 ∨ a2 ) ∧ (a03 ∨ a04 ), (a001 ∨ a002 ) ∧ (a003 ∨ a004 )]]
= −((a1 ∨ a2 ) ∧ (a3 ∨ a4 ) ∧ (a03 ∨ a04 )) · ((a1 ∨ a2 ) ∧ (a001 ∨ a002 ) ∧ (a003 ∨ a004 ));
(3.8)
iv. [[(a1 ∨ a2 ) ∧ (a3 ∨ a4 ), (a1 ∨ a02 ) ∧ (a03 ∨ a04 ), (a2 ∨ a02 ) ∧ (a003 ∨ a004 )]]
= −[[a1 a2 a02 ]] ([[a1 a3 a4 ]][[a2 a003 a004 ]][[a02 a03 a04 ]] − [[a1 a03 a04 ]][[a2 a3 a4 ]][[a02 a003 a004 ]]);

v. [[(a1 ∨ a2 ) ∧ (a3 ∨ a4 ), (a1 ∨ a3 ) ∧ (a2 ∨ a4 ), (a1 ∨ a4 ) ∧ (a03 ∨ a04 )]]

= [[a1 a2 a4 ]][[a1 a3 a4 ]] ([[a1 a2 a3 ]][[a4 a03 a04 ]] + [[a1 a03 a04 ]][[a2 a3 a4 ]]);

vi. [[(a1 ∨ a2 ) ∧ (a3 ∨ a4 ), (a1 ∨ a02 ) ∧ (a3 ∨ a04 ), (a2 ∨ a02 ) ∧ (a4 ∨ a04 )]]
= −[[a1 a2 a02 ]][[a3 a4 a04 ]] ((a1 ∨ a3 ) ∧ (a2 ∨ a4 ) ∧ (a02 ∨ a04 )).

As the result of Cayley expansion of a Cayley polynomial is generally not unique, the Cayley
expansions leading to factored and shortest results are often critical for effective symbolic com-
puting of incidence configurations. Their finding and classification are the main content of Cayley
expansion theory [16].
Coordinate-Free Theorem Proving in Incidence Geometry 75

How diverse can different Cayley expansion results of the same Cayley bracket be? For generic
vector variables, while (a1 ∨ a2 ) ∧ (a3 ∨ a4 ) ∧ (a5 ∨ a6 ) has only three different Cayley expansion
results in bracket polynomials,
• ((a1 ∨ a2 ) ∧ (a3 ∨ a4 ) ∧ (a5 ∨ a6 )) · ((a01 ∨ a02 ) ∧ (a03 ∨ a04 ) ∧ (a05 ∨ a06 )) has 45 different expan-
sion results in bracket polynomials;
• [[a1 , (a01 ∨a02 )∧(a03 ∨a04 ), (a001 ∨a002 )∧(a003 ∨a004 )]] has 46 different expansion results in bracket
polynomials;
• [[(a1 ∨ a2 ) ∧ (a3 ∨ a4 ), (a01 ∨ a02 ) ∧ (a03 ∨ a04 ), (a001 ∨ a002 ) ∧ (a003 ∨ a004 )]] has 16,847 different
expansion results in bracket polynomials.

3.3.2 Cayley Factorization

Glossary
Cayley factorization: The inverse of Cayley expansion: changing a homogeneous bracket polyno-
mial into an equal Cayley bracket.
Rational Cayley factorization: Changing a homogeneous bracket polynomial into an equal ra-
tional function where the numerator is a Cayley bracket, and the denominator is a bracket
monomial.

If a homogeneous bracket polynomial can be converted into a Cayley bracket, then the incidence
geometric interpretation of the bracket polynomial can be read directly from the expression of the
Cayley bracket. Cayley factorization is a procedure of generating geometric interpretation for a
projective invariant by incidence constructions. Unfortunately, for a general homogeneous bracket
polynomial, its Cayley factorization usually does not exist. See Section 5 of [30] for the special case
of multilinear Cayley factorization.
Besides geometric translation, Cayley factorization is also useful in Cayley expansion. Suppose
that a Cayley bracket is changed into different factored forms of bracket polynomials. Then Cayley
factorization may be used to unify different factors of the Cayley expansion results. This technique
is particularly useful in robust symbolic computing of projective conic theorems, in that the defect
of diverse Cayley expansion results may be significantly reduced or even eliminated.
For example, (3.6) and (3.7) are two different Cayley expansion results of the same input (3.5).
Although they are both binomial, the two results are from different shuffle formulas. To unify the
different factors of the results, equality (3.3) from right to left, and the fourth equality of (3.8) from
right to left, can be used to make Cayley factorization, and the results are unified to

[[a1 a2 a3 ]][[a01 a02 a03 ]] (a1 ∨ a01 ) ∧ (a2 ∨ a02 ) ∧ (a3 ∨ a03 ). (3.9)

Cayley factorization is a difficult task. Only the simplest case where a bracket polynomial is
linear with respect to every vector variable of it is solved [37]. Even the following seemingly simple
question remains open: is the following Crapo’s binomial

[[a1 a02 a03 ]][[a2 a03 a04 ]] · · · [[ak a01 a02 ]] + (−1)k−1 [[a1 a01 a02 ]][[a2 a02 a03 ]] · · · [[ak a0k a01 ]]

Cayley factorizable?
In [34], the problem of “rational Cayley factorizability” was investigated. The problem is as
follows: Given a bracket polynomial that is not Cayley factorizable, is it Cayley factorizable after
76 Handbook of Geometric Constraint Systems Principles

being multiplied with a suitable bracket monomial? An affirmative answer was given in [31]. The
technique rational Cayley factorization is very important in projective geometric computing. Still
there is no general algorithm for this factorization.
The following is a simple example of rational Cayley factorization. By equality (3.3) from right
to left,
[[a1 a3 a4 ]][[a2 a5 a6 ]] − [[a2 a3 a4 ]][[a1 a5 a6 ]] = (a1 ∨ a2 ) ∨ (a3 ∨ a4 ) ∨ (a5 ∨ a6 )
is a Cayley factorization. If the minus sign on the left side is changed to a plus sign, the bracket
binomial is no longer Cayley factorizable. Instead, it is rational Cayley factorizable, by the fifth
equality of (3.8) from right to left:
[[a1 a3 a4 ]][[a2 a5 a6 ]] + [[a2 a3 a4 ]][[a1 a5 a6 ]]
[[(a1 ∨ a2 ) ∨ (a5 ∨ a6 ), (a1 ∨ a3 ) ∨ (a2 ∨ a4 ), (a1 ∨ a4 ) ∨ (a2 ∨ a3 )]]
=
[[a1 a2 a3 ]][[a1 a2 a4 ]]
[[(a1 ∨ a2 ) ∨ (a3 ∨ a4 ), (a1 ∨ a5 ) ∨ (a2 ∨ a6 ), (a1 ∨ a6 ) ∨ (a2 ∨ a5 )]]
= .
[[a1 a2 a5 ]][[a1 a2 a6 ]]

3.3.3 Cayley Expansion and Factorization in Geometric Theorem Proving

Cayley expansion and factorization have various applications in combinatorics [10, 20], computer
vision [15, 17], robotics [38, 39], etc., and can be generalized to superalgebras [11, 13, 27, 28, 29].
Cayley expansion and factorization are particularly useful in proving incidence theorems having
linear construction sequences [18]. In a linear construction sequence of a geometric configuration,
any geometric entity is constructed by equations linear in the vector variable representing the ge-
ometric entity. A constrained geometric entity in a linear construction sequence is usually repre-
sented by a Cayley monomial. To prove a conclusion in the form of an equality in Cayley brackets,
we only need to substitute the Cayley monomial representations of the geometric entities into the
corresponding vector variables in the Cayley brackets in the reverse order of the linear construction
sequence, and simplify the result by Cayley expansion and factorization. This method works very
well, and for all projective incidence theorems tested so far, each theorem can be given a bino-
mial proof, i.e., the maximal number of terms of the conclusion expression during elimination and
simplification is two.
The benefit of the Cayley expansion and factorization method is not restricted to providing
short proofs. If one constraint of equality type is removed from the original geometric configura-
tion, the Cayley expansion and factorization approach usually provides quantitative description of
the relationship between a measurement of the discrepancy of the conclusion from being true and
a measurement of the discrepancy of the missing constraint from being satisfied. The following
Desargues’ Theorem [4] is a typical example.
For two triangles a1 a2 a3 and a01 a02 a03 in the plane, if the three lines a1 a01 , a2 a02 , a3 a03 concur, then the
three points of intersection a = a1 a2 ∩ a01 a02 , b = a1 a3 ∩ a01 a03 , and c = a2 a3 ∩ a02 a03 are collinear.
There is only one equality constraint in the hypothesis, which is the concurrence of the three
lines a1 a01 , a2 a02 , a3 a03 . In Grassmann-Cayley algebra, the hypothesis is represented by

f := (a1 ∨ a01 ) ∧ (a2 ∨ a02 ) ∧ (a3 ∨ a03 ) = 0,

and the conclusion is represented by

g := [[(a1 ∨ a2 ) ∧ (a01 ∨ a02 ), (a1 ∨ a3 ) ∧ (a01 ∨ a03 ), (a2 ∨ a3 ) ∧ (a02 ∨ a03 )]] = 0.

We disclose the relationship between the hypothesis expression and the conclusion expression by
Cayley expansion and factorization.
Coordinate-Free Theorem Proving in Incidence Geometry 77

a2
a2'

a1'
a1
a3'

a3 a

c
b

Figure 3.2
Desargues’ Theorem.

By Cayley expansion (3.6) or (3.7), then Cayley factorization (3.9), we get the following identity,
which is just the last equality of (3.8):
[[(a1 ∨ a2 ) ∧ (a01 ∨ a02 ), (a1 ∨ a3 ) ∧ (a01 ∨ a03 ), (a2 ∨ a3 ) ∧ (a02 ∨ a03 )]]
(3.10)
= [[a1 a2 a3 ]][[a01 a02 a03 ]] (a1 ∨ a01 ) ∧ (a2 ∨ a02 ) ∧ (a3 ∨ a03 ).
So Desargues’ Theorem and its converse are direct corollaries of (3.10). The two geometric
theorems when represented in this identity form can be directly used in symbolic manipulations as
term rewriting rules. This is a much higher level of algebraization of geometric theorems.
When changed into polynomials of homogeneous coordinates, f contains 48 terms, while g con-
tains as many as 1,290 terms. The effect of controlling input expression size and middle expression
swell by Cayley expansion and factorization is obvious.

3.3.4 Rational Invariants and Antisymmetrization

A projective space has various subspaces, and each subspace has its own invariants. An invariant of
a projective subspace is no longer an invariant of the whole space. For example, on a projective line,
the bracket of two points is an invariant, but no longer an invariant of the projective plane, instead it
becomes the covariant representing the line: the join operation of the two points.
Nevertheless, the ratio of two invariants of a projective subspace is always an invariant, called
an invariant ratio. For example, the ratio of two brackets [a1 , a2 ] and [a3 , a4 ] for collinear points
a1 , a2 , a3 , a4 is an invariant: for any point a5 not collinear with the four points,
[a1 , a2 ] [[a1 a2 a5 ]]
=
[a3 , a4 ] [[a3 a4 a5 ]]
is independent of the choice of dummy vector variable a5 . This is a ubiquitous property of all
invariant ratios in projective geometry.

Glossary
Invariant ratio: The signed ratio of two collinear line segments. Invariant ratios are as fundamental
as brackets in projective geometry. They are the direct heritage of low dimensional invariants.
Rational monomial invariant: A rational monomial function of brackets and invariant ratios.
Completion of rational monomial invariant: It refers to changing a rational monomial invariant
into an equal rational bracket monomial.
78 Handbook of Geometric Constraint Systems Principles

Antisymmetrization of rational monomial invariant: A method of completing rational mono-

mial invariants by constructing meet operations in the numerator and denominator of a rational
monomial invariant, respectively, then making monomial Cayley expansions.

The completion of a rational monomial invariant is always possible by adding dummy vector
variables to the components of invariant ratios, so the aim should be to find the simplest completion.
Since the numerator and denominator of an invariant ratio are both covariants, they can connect
with other covariants by the meet operation. A monomial of invariant ratios is then changed into
a monomial of ratios of meet operations. This transformation is called a partial antisymmetrization
of the rational monomial invariant. After the antisymmetrization, by monomial Cayley expansions,
a monomial of ratios of meet operations can be changed into a rational bracket monomial. This is
the antisymmetrization approach [18] to the completion of rational monomial invariants. It is very
useful in proving incidence theorems involving ratios. The following is an illustrative example.
Let a01 , a02 , a03 be points collinear with sides a2 a3 , a1 a3 , a1 a2 of triangle a1 a2 a3 , respectively.
(a) [Ceva’s Theorem and its converse (Fig 3.3, left)] Lines a1 a01 , a2 a02 , a3 a03 concur if and only if
a01 a2 a02 a3 a03 a1
= 1. (3.11)
a3 a01 a1 a02 a2 a03

(b) [Menelaus’ Theorem and its converse (Fig 3.3, right)] Points a01 , a02 , a03 are collinear if and
only if
a01 a2 a02 a3 a03 a1
= −1. (3.12)
a3 a01 a1 a02 a2 a03

a1 a1

a3' a2' a2'

a3'

a2 a1' a3 a1' a2 a3

Figure 3.3
Ceva’s Theorem (left) and Menelaus’ Theorem (right).

Denote the left side of (3.11) by p. It has a natural antisymmetrization

(a01 ∨ a2 ) ∧ (a02 ∨ a3 ) ∧ (a03 ∨ a1 )
p= .
(a3 ∨ a01 ) ∧ (a1 ∨ a02 ) ∧ (a2 ∨ a03 )
The following Cayley expansion of p, where the priorities in the expansion are described by the
parentheses, leads to Ceva’s Theorem:
((a01 ∨ a2 ) ∧ (a02 ∨ a3 )) ∧ (a03 ∨ a1 ) [[a2 a01 a02 ]][[a1 a3 a03 ]]
= .
((a2 ∨ a03 ) ∧ (a3 ∨ a01 )) ∧ (a1 ∨ a02 ) [[a3 a01 a03 ]][[a1 a2 a02 ]]
Coordinate-Free Theorem Proving in Incidence Geometry 79

It changes (3.11) into

[[a2 a01 a02 ]][[a1 a3 a03 ]] − [[a1 a2 a02 ]][[a3 a01 a03 ]] = −(a1 ∨ a01 ) ∨ (a2 ∨ a02 ) ∨ (a3 ∨ a03 ) = 0.

The following Cayley expansion of p leads to Menelaus’ Theorem:

((a01 ∨ a2 ) ∧ (a02 ∨ a3 )) ∧ (a03 ∨ a1 ) [[a2 a01 a02 ]][[a1 a3 a03 ]]
= − .
((a3 ∨ a01 ) ∧ (a1 ∨ a02 )) ∧ (a2 ∨ a03 ) [[a1 a01 a02 ]][[a2 a3 a03 ]]
It changes (3.12) into

[[a2 a01 a02 ]][[a1 a3 a03 ]] − [[a1 a01 a02 ]][[a2 a3 a03 ]] = (a1 ∨ a2 ) ∧ (a01 ∨ a02 ) ∧ (a3 ∨ a03 )
= −[[a1 a2 a3 ]][[a01 a02 a03 ]]
= 0.

3.4 Bracket Algebra for Euclidean Geometry

So far all our considerations apply to configurations and theorems in the projective plane. The only
properties that might be encoded are those that are invariant under the group of projective transfor-
mations. One might be misled to think that the algebraic expressiveness and the proving techniques
introduced in the previous sections cannot be applied to theorems that involve measurements and
refer to distances and angles.
In a sense rather the opposite is the case. In a suitable setup all properties and relations of Eu-
clidean (and even hyperbolic or elliptic) geometry can be expressed in terms of projective geometry.
The working horse behind this is the so-called Cayley-Klein geometries that can be traced back to
the works of Beltrami, Cayley, and Klein. A detailed account of them can be found in [14, 23, 25].
Here we will only roughly sketch how Euclidean properties can be expressed in projective terms.
The key to the embedding of Euclidean properties into a projective setup is the use of complex
numbers in a very controlled way. We consider the real projective plane RP2 and introduce two
extra points with complex homogeneous coordinates: I = (1, −i, 0)T and J = (1, i, 0)T .

3.4.1 The Points I and J

Glossary
I and J: Two points with special homogeneous coordinates
√ I = (1, −i, 0)T and J = (1, i, 0)T with
respect to the standard embedding, where i = −1. They are both on the line at infinity l∞ of
the plane.
Concyclicity: In the standard embedding, all conics passing simultaneously through I and J are
exactly the circles (including the limit situations of the union of a line and the line at infinity).
Euclidean similarity: A transformation that leaves absolute values of angles as well as ratios of
lengths invariant. Euclidean similarities are those projective transformations that stabilize the
pair {I, J}.
Cross ratio: A rational monomial invariant for four points A, B,C, D on a line l. If point X is not
[[XAC]][[XBD]]
incident to line l then cr(A, B;C, D) := [[XAD]][[XBC]] . The result is independent of the choice of
X.
80 Handbook of Geometric Constraint Systems Principles

Laguerre’s formula for angles: The angle from line l to line m in homogeneous coordinates can
be calculated by 2i1 ln(cr(L, M; I, J)) with L = l ∧ l∞ and M = m ∧ l∞ .

If we consider the standard embedding R2 → RP2 via (x, y) 7→ (x, y, 1) the two points I and J
are exactly those points that are obtained by intersecting an arbitrary circle with the line at infinity.
To see this consider a circle (x − mx )2 + (y − my )2 = r2 with center m and radius r. Written in
homogeneous coordinates this circle is a conic given by the equation

x2 + y2 − 2mx · xz − 2my · yz + (m2x + m2y − r2 ) · z2 = 0.

Intersecting this conic with the line at infinity l = (0, 0, 1)T asks for those solutions of the conic
equation that also satisfy z = 0. Thus, one has to solve x2 + y2 = 0. Up to scalar multiples the unique
two solutions are I = (1, −i, 0)T and J = (1, i, 0)T .
It is remarkable that the solution is independent of the particular choice of the initial circle. In
other words: All circles pass through I and J. Conversely it can be shown that any conic passing
through the two points either is a proper circle (in the standard embedding), or decomposes into the
union of a line l and the line at infinity. The latter situation can be considered as the RP2 equivalent
of a circle with infinite radius (compare [25]). In other words: circles are the conics through I and
J.
Transformations that map circles to circles in Euclidean geometry are those that leave ratios of
lengths invariant. They turn out to be rotations, reflections, translations, scalings, and arbitrary com-
positions of those (summarized under the term Euclidean similarity transformations). In particular,
these transformations preserve the absolute values of angles (which is an alternative characteri-
zation). With respect to the standard embedding of R2 into RP2 this group forms a subgroup of
the projective transformations. Combining these considerations with the above characterization of
circles we can say that Euclidean similarity transformations are exactly those transformations that
leave the pair {I, J} invariant.
In fact this is a special case of a more general construction. The pair {I, J} can be considered
as a special degenerate conic. The theory of Cayley-Klein geometries characterizes measurement of
angles and distances by referring the calculation of cross ratios to general conics. This construction
is remarkably rich, and within the theory of Cayley-Klein geometries Euclidean, hyperbolic, elliptic,
and relativistic geometries occur as special cases [14, 23, 25].
The above characterization of Euclidean similarities implies that every property that is invariant
under Euclidean similarity transformations can be expressed as a projective invariant in which I
and J play a special role. Rephrased in a different way Euclidean similarity geometry is projective
geometry with two points I and J distinguished. We exemplify this general principle by Laguerre’s
formula that allows us to calculate the angle between two lines.

Theorem 3.7 (Laguerre’s formula) Given two lines l and m in the Euclidean plane (in the stan-
dard embedding), the absolute value of the angle between them can be calculated by
1
| ln(cr(L, M; I, J))|, with L = l ∧ l∞ , M = l ∧ m∞ .
2i

3.4.2 Proving Euclidean Theorems

The considerations of the previous subsection show that it is possible to express properties of Eu-
clidean geometry by performing projective calculations and expressing Euclidean properties with
projective invariant measures in which I and J play a distinguished role. Thus the projective prov-
ing techniques introduced previously may also be applied to purely Euclidean situation. For this it
Coordinate-Free Theorem Proving in Incidence Geometry 81

is necessary to express every Euclidean relation by a projectively invariant property that involves I
and J. We are going to exemplify this philosophy by working one concrete example.
We consider Miquel’s Theorem about six circles and eight points (Figure 3.4):
Let 1, 2, . . . , 8 be the indices of eight distinct points in the Euclidean plane. Consider the six quadru-
ples of points (1, 2, 3, 4), (1, 2, 7, 8), (5, 2, 3, 8), (5, 6, 3, 4), (1, 6, 7, 4), (5, 6, 7, 8). If five of these
quadruples are concyclic so is the last one.

Figure 3.4
The Theorem of Miquel.

First we observe that we can characterize the concyclicity of four points 1, 2, 3, 4 by expressing
the fact that 1, 2, 3, 4, I, J are on a common conic. Using our characterization of conics by bracket
binomials we see that this can be algebraically expressed by
[[12I]][[34I]][[14J]][[32J]] = [[12J]][[34J]][[14I]][[32I]].
Observe that the following transformation property arises, that finally justifies this characterization.
Assume that (1, 2, 3, 4) are concyclic and by this satisfy the above equation. Now consider a pro-
jective transformation τ that leaves the pair {I, J} invariant (i.e., a Euclidean similarity). Without
loss of generality assume that τ(I) = I and τ(J) = J. Since the transformation is projective and the
above formula is stable under projective transformations, we get
[[10 20 I]][[30 40 I]][[10 40 J]][[30 20 J]] = [[10 20 J]][[30 40 J]][[10 40 I]][[30 20 I]],
with p0 = τ(p) for p ∈ {1, 2, 3, 4}. Thus as expected, the transformed points again satisfy our al-
gebraic concyclicity characterization. The above-mentioned Theorem of Miquel is now an easy
consequence.
Assume that the first five concyclicities are satisfied; this gives the following five equations:
[[12I]][[34I]][[14J]][[32J]] = [[12J]][[34J]][[14I]][[32I]],
[[12J]][[65J]][[15I]][[62I]] = [[12I]][[65I]][[15J]][[62J]],
[[62J]][[37J]][[67I]][[32I]] = [[62I]][[37I]][[67J]][[32J]],
[[87J]][[34J]][[84I]][[37I]] = [[87I]][[34I]][[84J]][[37J]],
[[15J]][[84J]][[14I]][[85I]] = [[15I]][[84I]][[14J]][[85J]].
82 References

The situation is similar to that of manipulating the conclusion part of a bi-quadratic final polynomial.
Multiplying all the left sides and all the right sides and cancelling everything that occurs on both
sides (we can do this since any line containing two distinct real points does contain neither I nor J),
we are left with
[[65J]][[87J]][[67I]][[85I]] = [[65I]][[87I]][[67J]][[85J]],
an equation expressing the concyclicity of 5, 6, 7, 8.

For more elaborate discussion of Euclidean theorems from a projective viewpoint including
many elaborated examples, we refer to [7, 25].
Acknowledgment. The first author is supported by the DFG Collaborative Research Center TRR
109, “Discretization in Geometry and Dynamics”; the second author is supported by the NSFC
Project 11671388 and the CAS Frontier Key Project QYZDJ-SSW-SYS022.

References
[1] Abhyankar, S.S. Invariant theory and enumerative combinatorics of young tableaux. In:
Mundy, J.L. and Zisserman, A. (eds.) Geometric Invariance in Computer Vision. MIT Press,
Cambridge, MA, pp. 45–76, 1992.
[2] Anick, D. & Rota, G.-C. Higher-order syzygies for the bracket algebra and for the ring of
coordinates of the Grassmannian. Proc. Nat. Acad. Sci. 88(18): 8087–8090, 1991.
[3] Barnabei, M., Brini, A., & Rota, G.-C. On the exterior calculus of invariant theory. J. of
Algebra 96: 120–160, 1985.
[4] Berger, M. Geometry I, II. Springer, 1987.
[5] Bokowski, J. & Richter-Gebert, J. On the finding of final polynomials. Europ. J. Combina-
torics 11: 21–34, 1990.
[6] Bokowski, J. & Sturmfels, B. Computational synthetic geometry. In: Lecture Notes in Math-
ematics 1355, Springer-Verlag, Berlin, Heidelberg, 1989.
[7] Crapo, H. & Richter-Gebert, J. Automatic proving of geometric theorems. In: Invariant Meth-
ods in Discrete and Computational Geometry, White, N. (ed.), Kluwer Academic Publishers,
Dodrecht, 107–139, 1995.
[8] Désarménien, J. An algorithm for the Rota Straightening Formula, Discrete Mathematics 30:
51–68, 1980.
[9] Doubilet, P., Rota, G.-C., & Stein, J. On the foundations of combinatorial theory IX: Combi-
natorial Methods in Invariant Theory. Stud. Appl. Math. 57: 185–216, 1974.
[10] Dress, A. & Wenzel, D. Grassmann-Plücker relations and matroids with Coefficients. Ad-
vances in Mathematics 86: 68–110, 1991.
[11] Grosshans, F.D., Rota, G.-C., & Stein, J. Invariant Theory and Superalgebras, AMS, 1987.
[12] Gurevich, G.B. Foundations of the Theory of Algebraic Invariants, P. Noordhoff, Groningen,
1964.
[13] Huang, R.Q., Rota, G.-C., & Stein, J. Supersymmetric bracket algebra and invariant theory,
Acta Appl. Math. 21: 193–246, 1990.
References 83

[14] Kowol, G. Projektive Geometrie und Cayley-Klein Geometrien der Ebene, Birkhäuser, 2009.
[15] Li, H. & Sommer, G. Coordinate-free projective geometry for computer vision. In: Sommer,
G. (ed.), Geometric Computing with Clifford Algebras, Springer, Heidelberg, pp. 415–454,
2001.
[16] Li, H. & Wu, Y. Automated short proof generation in projective geometry with Cayley and
bracket algebras I, II. J. Symb. Comput. 36(5): 717–762, 763–809, 2003.
[17] Li, H., Zhao, L., & Chen, Y. A symbolic approach to polyhedral scene analysis by parametric
calotte propagation. Robotica 26(4): 483-501, 2008.
[18] Li, H. Invariant Algebras and Geometric Reasoning. World Scientific, Singapore, 2008.
[19] Li, H., Shao, C., Huang, L., & Liu, Y. Reduction among bracket polynomials. In: Proc. IS-
SAC’14, ACM Press, pp. 304–311, 2014.
[20] Mainetti, M. & Yan, C.H. Arguesian Identities in Linear Lattices. Advances in Mathematics
144: 50–93, 1999.
[21] McMillan, T. & White, N. The dotted straightening algorithm. J. Symb. Comput. 11: 471–482,
1991.
[22] Olver, P.J. Classical Invariant Theory. Cambridge University Press, Cambridge, 1999.
[23] Onishchik, A.L. & Sulanke, R. Projective and Cayley-Klein Geometries, Springer Mono-
graphs in Mathematics, Springer, 2006.
[24] Richter-Gebert, J. Mechanical theorem proving in projective geometry. Annals of Mathemat-
ics and Artificial Intelligence 13: 139–172, 1995.
[25] Richter-Gebert, J. Perspectives on Projective Geometry, Springer, 2011.
[26] Richter-Gebert, J. Meditations on Ceva’s Theorem. In: The Coxeter Legacy: Reflections and
Projections, Davis, C. & Ellers, E. (Eds.), American Mathematical Society, Fields Institute,
227–254, 2006.
[27] Rota, G.-C. & Stein, J. Applications of Cayley Algebras. Academia Nazionale dei Lincei atti
dei Convegni Lincei 17, Colloquio Internazionale sulle Teorie Combinatoire, Tomo 2, Roma,
1976.
[28] Rota, G.-C. & Stein, J. Symbolic method in invariant theory, Proc. Nat. Acad. Sci. 83: 844–
847, 1986.
[29] Rota, G.-C. & Sturmfels, B. Introduction to invariant theory in superalgebras. In: Staton, D.
(ed.), Invariant Theory and Tableaux, Springer, New York, pp. 1–35, 1990.
[30] Sidman, J. & Traves, W. Cayley factorization and special positions. In: Sitharam, M., St. John,
A., and Sidman, J. (ed.), Handbook of Geometric Constraints Systems: Principles.
[31] Sturmfels, B. Computational algebraic geometry of projective configurations. J. Symbolic
Computation 11: 595–618, 1991.
[32] Sturmfels, B. Algorithms in Invariant Theory, Springer-Verlag, Wien, New York, 1993.
[33] Sturmfels, B. & White, N. Gröbner bases and invariant theory. Adv. Math. 76(2): 245–259,
1989.
[34] Sturmfels, B. & Whiteley, W. On the synthetic factorization of homogeneous invariants. J.
Symbolic Computation 11: 439–454, 1991.
84 References

[35] White, N. The bracket ring of combinatorial geometry I. Transactions of AMS 202: 79–95,
1975.
[36] White, N. Implementation of the straightening algorithm of classical invariant theory. In: Sta-
ton, D. (ed.), Invariant Theory and Tableaux, Springer, New York, pp. 36–45, 1990.
[37] White, N. Multilinear Cayley factorization. J. Symb. Comput. 11: 421–438, 1991.
[38] White, N. Geometric applications of the Grassmann-Cayley algebra. In: Goodman, J.E. and
O’Rourke, J. (eds.), Handbook of Discrete and Computational Geometry, CRC Press, Boca
Raton, FL, 1997.
[39] White, N. Grassmann-Cayley algebra and robotics applications. In: Bayro-Corrochano, E.
(ed.), Handbook of Geometric Computing, Springer, Heidelberg, pp. 629–656, 2005.
[40] Whiteley, W. Logic and invariant computation for analytic geometry. In: Symbolic Computa-
tions in Geometry, I.M.A. Preprint # 389, University of Minnesota, January 1988.
[41] Whiteley, W. Invariant computations for analytic projective geometry. J. Symb. Comput. 11:
549–578, 1991.
Chapter 4
Special Positions of Frameworks and the
Grassmann-Cayley Algebra
Jessica Sidman and William Traves
Mount Holyoke College and US Naval Academy

CONTENTS
4.1 Introduction: the Grassmann-Cayley Algebra and Frameworks . . . . . . . . . . . . . . . . . . . . . 85
4.2 Projective Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.2.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.2.2 Homogeneous Coordinates and Points at Infinity . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.2.3 Equations on Projective Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.2.4 Duality Between Lines and Points in P2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.2.5 Grassmannians and Plücker Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.2.6 More About Lines in 3-space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.3 The Bracket Algebra and Rings of Invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.3.1 Group Actions and Invariant Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.3.2 Relations Among the Brackets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.4 The Grassmann-Cayley Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.4.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.4.2 The cross product as a Join . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.4.3 Properties of the Exterior Product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.4.4 Brackets and the Grassmann-Cayley algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.4.5 The Join and Meet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.5 Cayley Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.5.2 The Pure Condition as a Bracket Monomial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.5.3 White’s Algorithm for Multilinear Grassmann-Cayley Factorization . . . 103
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

4.1 Introduction: the Grassmann-Cayley Algebra and Frame-

works
Given a framework determined by a system of geometric constraints, a fundamental question is to
determine whether the framework is rigid or flexible. If the framework is a bar-and-joint or body-
and-bar framework, the constraints are all given by fixing the distance between selected pairs of
points with bars. The theorems of Laman [2], (actually discovered first by Pollaczek-Geiringer [5]),

85
86 Handbook of Geometric Constraint Systems Principles

Tay [12], and White-Whiteley [13] characterize rigidity combinatorially in various settings, but
these combinatorial characterizations only determine the behavior of sufficiently generic realiza-
tions of a framework with given combinatorics.
The pure condition of a generically rigid framework, introduced by White and Whiteley in [13]
and [14], is a polynomial in brackets. The goal of this chapter is to give an introduction to the
Grassmann-Cayley algebra, which can be used to give a geometric interpretation of the vanishing
of bracket polynomials in order to better understand when a generically rigid framework admits
nontrival internal motions.
In order to motivate the development of the Grassmann-Cayley algebra we present the simplest
example of a generically rigid structure in the plane, two rigid bodies connected by three bars, which
we represent by a multigraph with a vertex for each body and an edge for each bar in Figure 4.1.

a
b
1 2
c

Figure 4.1
Vertices 1 and 2 correspond to rigid bodies in P2 . Three bars with generic attachment points create
a rigid framework.

Following [13], we embed the framework in the projective plane P2 and label each bar with a
vector corresponding to the line in the direction of the bar. This vector is obtained by taking the cross
product of the vectors corresponding to the endpoints of the bars. Equivalently, the bars are labeled
by their Plücker coordinates in a projective plane that is dual to the plane in which the framework is
embedded. (See Section 4.2 for a more detailed introduction to projective space, Section 4.2.5 for
an introduction to Plücker coordinates, and Section 4.2.4 for a discussion of duality.) The resulting
body-and-bar framework is generically infinitesimally rigid and has a nontrivial infinitesimal motion
exactly when the three bars are parallel or meet at a point (see Figure 4.2).

Figure 4.2
Three realizations of the graph from Figure 4.1. The first is generic. The middle figure depicts an
embedding in which the bars are parallel, and the lines along the bars meet at a point on the line at
infinity. In the third example the point of coincidence is in the finite plane.

In P2 lines are parallel exactly when they meet at a point on the line at infinity, hence the
framework has an infinitesimal motion exactly when the lines along the bars are coincident. Three
lines in P2 are coincident when their Plücker coordinates in the dual projective plane are collinear.
This is precisely the condition that the 3 × 3 matrix [abc] has determinant zero.
In Section 4.2 we give an introduction to projective space, the Grassmannian, and Plücker co-
ordinates. In Section 4.3 we discuss the bracket ring as the ring of invariant polynomials on Pn and
as a quotient ring with relations given by Plücker relations and Van der Waerden syzygies. We treat
Special Positions of Frameworks and the Grassmann-Cayley Algebra 87

the Grassmann-Cayley algebra in Section 4.4 and Cayley factorization in Section 4.5, illustrating
the theory with examples from the rigidity theory of body-and-bar frameworks. In the final section
we include a new result, Theorem 4.1 due to the first author, Cai, St. John, and Theran, character-
izing which body-and-bar frameworks have pure conditions which may be represented by a bracket
monomial. We use the notation and conventions of Sturmfels [9] throughout, so that the reader who
would like a fuller exposition may transition easily between this chapter and [9].

4.2 Projective Space

4.2.1 Motivation
The frameworks that we wish to consider live naturally in Rn , a real vector space of dimension n, but
it is often useful to work in the augmented n-dimensional projective space Pn . To construct Pn we
attach new points to Rn at infinity; these new points encode the ways that a sequence of points can
diverge to infinity. The resulting projective space has many properties not found in our Euclidean
space. For instance, Bézout’s Theorem for the projective plane says that any two curves defined by
polynomial equations of degrees d and e meet in de points, suitably interpreted. Indeed, our points
of intersection may lie at infinity and can even have complex coordinates; if the curves are tangent at
some of the intersection points, we count these points with multiplicity (for details, see Schenk [7]
or Smith et al. [8]). As a result of Bézout’s Theorem, any pair of lines in the plane, each described
by the vanishing of a polynomial of degree 1, must meet; when the lines are parallel, the point of
intersection lies at infinity. Indeed, each point at infinity can be thought of as the point of intersection
of a family of parallel lines. In the setting of body-and-bar frameworks this point of view unifies
how we think about translations and rotations, as an infinitesimal translation perpendicular to these
lines can then be thought of as an infinitesimal rotation about the point at infinity.

Projective space: Projective space Pn is obtained by adding a hyperplane of points at infinity to

an n-dimensional vector space V . The projective space Pn may be viewed as parameterizing
1-dimensional subspaces of V .
Homogeneous coordinates: Homogeneous coordinates give a way to describe points in projective
space. A polynomial P on V can be extended to all of Pn by homogenizing P with respect to
an additional variable. Though such polynomials are no longer functions on Pn , their vanishing
sets are well-defined.
Duality: There is an incidence-preserving duality between points in the projective plane P2 and
∗
lines in P2 . Lines in P2 are elements of a dual projective plane, P2 .
Grassmannian: The Grassmannian G(d +1, n+1) is a variety parameterizing (d +1)-dimensional
subspaces of an (n + 1)-dimensional vector space. This is equivalent to G(d, n), parameterizing
d-dimensional linear spaces of Pn . Both varieties are manifolds of dimension (d + 1)(n − d).
Plücker coordinates: The Plücker coordinates give an embedding of G(d + 1, n + 1) into a projec-
tive space. Each subspace is represented by a d + 1 by n + 1 matrix X whose rows form a basis
of the subspace. The Plücker coordinates are then just the determinants of the (d + 1) by (d + 1)
submatrices of X. The Plücker coordinates are also called brackets and the determinant of the
submatrix given by the columns i1 , i2 , . . . , id+1 is denoted [i1 i2 . . . id+1 ]. The vanishing sets of
homogeneous polynomials in the brackets form well-defined subsets of the Grassmannian.
88 Handbook of Geometric Constraint Systems Principles

4.2.2 Homogeneous Coordinates and Points at Infinity

August Ferdinand Möbius introduced coordinates on projective space that allow us to discuss finite
and infinite points using the same notation. For each nonzero point (a1 , . . . , an ) ∈ Rn we add a
(superfluous) coordinate [1 : a1 : . . . : an ]. More generally, points in projective space have the form
[a0 : a1 : . . . : an ] (with not all coordinates ai equal to zero), but we identify any two points that differ
by a nonzero scalar multiple,

[a0 : a1 : . . . : an ] = [a0 λ : a1 λ : . . . : an λ ] for λ 6= 0.

The notation uses the colon rather than the comma to suggest that the ratio of coordinates is essential
to the location of a point in Pn , rather than the actual coordinates themselves, which can be simul-
taneously scaled by a common nonzero multiple λ . With this convention, points in Pn whose initial
coordinate is nonzero (and hence can be scaled to 1) correspond to points in Rn . For an introduction
to the projective plane, P2 as a quotient of the sphere S2 , see Section 4.1.1 of Chapter 3.
Points in Pn whose initial coordinate is zero correspond to points at infinity. For example, let’s
look for the intersection of two parallel lines in the plane, e.g. the lines y = 2x + 3 and y = 2x + 5.
These lines don’t meet at all in R2 ; their intersection in P2 must lie among the points at infinity. To
see which point, we take the limit of the points [1 : x : 2x + 3] for x ∈ R as x goes to infinity:
 
1 3
lim [1 : x : 2x + 3] = lim : 1 : 2+ = [0 : 1 : 2].
x→∞ x→∞ x x

A similar limit computation shows that the points on the second line also approach [0 : 1 : 2] as x
approaches infinity. Indeed, the points on any line in R2 with slope equal to m approach the point
[0 : 1 : m] as x approaches infinity. Moreover, it does not matter whether x approaches ∞ or −∞,
the points on the lines approach the same limit point. Though one might first imagine that we need
to add a circle of limit points at infinity to R2 – one for each direction pointing from the origin – in
fact, the antipodal points on this circle are identified in P2 . Instead, we see that our homogeneous
coordinates represent points of R2 together with an additional point for each line through the origin
in R2 . As a point of interest, we remark that if you were to thicken this curve of points at infinity
to a band, the resulting surface would have just one edge (and one side) – the Möbius band. It is
curious that this portion of the projective plane bears Möbius’s name rather than the coordinates he
introduced.
In the following sections we focus on low-dimensional examples such as P2 and P3 . By special-
izing we are able to make our computations concrete. As well, our low-dimensional focus is well
suited to applications involving physical body-and-bar frameworks.

4.2.3 Equations on Projective Space

In the Euclidean plane R2 with coordinates (a1 , a2 ) a polynomial equation such as a21 + 2a2 − 3 = 0
implicitly describes a curve but the solutions of this equation don’t interact well with our new
coordinates on P2 ; after all, if we multiply [1 : a1 : a2 ] by an arbitrary nonzero scalar λ we don’t
change the projective point but the resulting coordinates no longer satisfy the equation. We can deal
with this issue by homogenizing the equation, multiplying each term by a power of a0 so that each
term has the same degree as the original polynomial. For example, the homogenization of a21 +
2a2 − 3 is a21 + 2a0 a2 − 3a20 . Scaling each coordinate by a nonzero constant λ causes the evaluation
of the homogeneous polynomial to be multiplied by a power of λ , so the homogeneous polynomial
evaluates to zero independently of the representative chosen for the point in projective space. For this
reason we use the vanishing of homogeneous polynomials to define objects in projective space and
the points in Pn satisfying the homogenization of a polynomial condition on Rn can be thought of as
the points on the extension of the original object to projective space. Indeed, the curves appearing
Special Positions of Frameworks and the Grassmann-Cayley Algebra 89

in Bézout’s Theorem must be projective plane curves defined by the vanishing of homogeneous
polynomials.

4.2.4 Duality Between Lines and Points in P2

The vanishing of a linear polynomial describes a line in P2 . Of particular interest is the curve defined
by a0 = 0; this consists of the points at infinity in P2 , which form a line in P2 rather than the circle
that one might naively picture from our discussion in Section 4.2.2. It is worth recomputing the
intersection of the parallel lines y = 2x + 3 and y = 2x + 5 using the homogenized versions of the
polynomials. Using the homogenizing variable w, the homogenized equations are y − 2x − 3w = 0
and y − 2x − 5w = 0 and their intersection consists of the nonzero points [w : x : y] = [0 : x : 2x], any
of which represent the intersection point [0 : 1 : 2] in P2 . This example illustrates yet another way of
looking at points and lines in P2 . Points in P2 correspond to 1-dimensional linear subspaces of R3
and lines in P2 correspond to 2-dimensional linear subspaces.
Lines in the projective plane have a peculiar property: if a0 , a1 , and a2 are the coordinates on P2
then the line
c0 a0 + c1 a1 + c2 a2 = 0 (4.1)
is completely determined by the tuple of constants (c0 , c1 , c2 ). Moreover, the line doesn’t change if
we multiply its equation by a nonzero scalar, so the line corresponds uniquely to the projective point
[c0 : c1 : c2 ]. That is, the lines in P2 can be identified with points in another copy of P2 , which we
∗
denote by P2 . Recalling our earlier model of the projective plane in which points in P2 correspond
to 1-dimensional linear subspaces and lines correspond to 2-dimensional linear subspaces and fixing
the usual inner product on 3-space, we see that the dual of a projective line given by equation (4.1)
∗
is the point [c0 : c1 : c2 ] ∈ P2 corresponding to the subspace orthogonal to the plane in 3-space
determined by equation (4.1).

4.2.5 Grassmannians and Plücker Coordinates

∗
The dual space P2 parameterizes lines in the projective plane. Describing lines in higher dimen-
sional projective spaces requires a new perspective. Each point in Pn corresponds to a 1-dimensional
subspace of Rn+1 . When we picked a nonzero (n + 1)-tuple to represent our point in Pn , we were
really picking a basis for the underlying subspace. And when we identified all nonzero scalar multi-
ples of this tuple we were really identifying all tuples that span the same subspace. Now each line in
Pn can be thought of as a 2-dimensional subspace of Rn+1 . Given a basis of this space (equivalently,
given two distinct points on the line), we can arrange the vectors in a 2 × (n + 1) matrix M of full
rank. Multiplying M on the left by any invertible 2 × 2 matrix A implements elementary row opera-
tions on the matrix M and gives a new basis for the subspace corresponding to the line. Just as in our
construction of projective space by identifying vectors that differ by a nonzero scalar multiple, we’d
like to break the set of 2 × (n + 1) matrices of full rank into equivalence classes, where two matrices
are identified if their rows span the same subspace (equivalently, if there is an invertible 2 × 2 ma-
trix A so that multiplication of one matrix on the left by A gives the other matrix). Each equivalence
class represents a 2-dimensional subspace in Rn+1 and the set of these equivalence classes is called
the Grassmannian G(2, n + 1). Alternatively, if we associate each equivalence class with a line in
Pn , the set of equivalence classes is then referred to as the Grassmannian G(1, n). We use the bold
font G to indicate that we are thinking about objects projectively.
The coordinates on the Grassmannian G(1, n) are called Plücker coordinates. There is one co-
ordinate for each of the n+1


2 2 × 2 submatrices of the 2 × (n + 1) matrix M. Each coordinate takes
a value given by the determinant of the corresponding 2 × 2 submatrix. For example, there are six
coordinates for lines in P3 ; each is denoted by a bracket of numbers representing the columns in a
90 Handbook of Geometric Constraint Systems Principles

2 × 2 submatrix. For instance, the coordinates of the matrix

 
1 2 1 2 [12] = 1 [13] = 1 [14] = 3
M= are
0 1 1 3 [23] = 1 [24] = 4 [34] = 1.
As with homogeneous coordinates on Pn , it is not the actual values of the coordinates that matter, but
rather their ratio relative to the other coordinates. Indeed, multiplying M by an invertible 2×2 matrix
A changes the representative matrix for our line but does not change the line itself. At the same time,
multiplying M by A multiplies each of the determinant coordinates by det(A) 6= 0. So
the Plücker
coordinates give an embedding of G(1, n) into the projective space PN , where N = n+1 2 − 1.
We’ve explained how to get coordinates from the line; now we explain how to recover the line
from the coordinates. Since the matrix M has full rank, one of the coordinates [i j] is nonzero. It fol-
lows that we can multiply M by an appropriate invertible matrix A so that the (i, j) submatrix of AM
is the identity matrix. The newly scaled values of the coordinates [ik] and [k j] (for k 6= i, j) record
the values of the remaining entries in AM,
determining a basis for the line corresponding to the
original matrix M. Though there are n+1 2 coordinates, the Grassmannian G(1, n) has much lower
dimension. If we change basis so that the (i, j) minor is the identity matrix, the remaining 2(n − 1)
coordinates can take arbitrary values. The space G(1, n) is a manifold of dimension 2(n − 1). In par-
ticular, the collection G(1, 3) of lines in 3-space is 4 dimensional. This seems intuitively reasonable:
imagine two parallel planes in 3-space; most lines in 3-space are determined by where they intersect
these two planes, and these two points of intersection are determined by their 4 coordinates.
To parameterize the set of linear spaces of Pn of dimension d (corresponding to linear sub-
spaces of Rn+1 of dimension d + 1) we consider full-rank (d + 1) × (n + 1) matrices modulo left
multiplication by (d + 1) × (d + 1) invertible matrices. The homogeneous coordinates on G(d, n)
n+1

or G(d + 1, n + 1) are the determinants of the (d + 1) × (d + 1) submatrices. There are d+1 such
n+1

submatrices so these determinants embed G(d, n) as a subset of PN , where N = d+1 − 1. As in the
construction of G(1, 3), since one of the subdeterminants is nonzero we can left multiply by an in-
vertible matrix A to force that submatrix to be the identity. The remaining (d + 1)(n + 1 − (d + 1)) =
(d + 1)(n − d) entries can take arbitrary values so the spaces G(d, n) and G(d + 1, n + 1) have di-
mension (d + 1)(n − d).
Just as we require our polynomials to be homogeneous in order to define curves in the pro-
jective plane, the polynomials that define algebraic subsets of the Grassmannian G(d, n) must be
homogeneous in the sense that each monomial contains the same number of Plücker coordinates.

4.2.6 More About Lines in 3-space

In this section we consider some bracket polynomials (polynomials in the Plücker coordinates) that
describe the behavior of lines in 3-space.
Example 1 Suppose that we want to find an algebraic characterization of when two lines in P3
intersect. In other words, given `1 ∈ G(1, 3) and `2 ∈ G(1, 3), we would like to find a bracket
polynomial that evaluates to zero when these two lines intersect. Thinking of each line as a 2-
dimensional subspace in R4 , we pick a basis v1 , v2 for `1 and w1 , w2 for `2 . Using these bases
we can construct the Plücker coordinates for each line. Let [i j]k stand for the [i j] coordinate for
line `k (k = 1, 2, respectively). The two lines meet if and only if the span of their basis vectors is a
3-dimensional subspace of R4 . This can be detected by seeing if the determinant of the 4 × 4 matrix
with rows v1 , v2 , w1 and w2 vanishes. When we expand this determinant via Laplace expansion along
the top two rows we obtain an expression for the determinant in terms of the Plücker coordinates
for the lines:
[12]1 [34]2 − [13]1 [24]2 + [14]1 [23]2 + [23]1 [14]2 − [24]1 [13]2 + [34]1 [12]2 .
This expression equals zero precisely when the two lines intersect or coincide.
Special Positions of Frameworks and the Grassmann-Cayley Algebra 91

Example 2 Now we present a similar example where we will see that the equations derived using
analogous geometric reasoning vanish on a set containing several components, only one of which
corresponds to the geometry that we are trying to characterize. We’ll consider three lines `1 , `2 , and
`3 and try to find equations in the brackets that guarantee that the three lines all meet at a single
point. First we’ll try to determine the dimension of the space parameterizing three lines that meet in
a common point. Each of the three lines sits in G(1, 3), a four-dimensional variety, so without any
restrictions, the three lines are parameterized by 12 parameters. To describe a single line ` ⊂ P3
that goes through a point P, it is enough to specify any other point Q on the line `. There is a
3-dimensional space of possible values for Q ∈ P3 but lots of choices for Q give the same line; Q
and Q0 give rise to the same line ` precisely when Q, Q0 and P are collinear. Identifying points Q
and Q0 that are collinear with P we obtain a 2-dimensional space parameterizing lines through P.
One way to parameterize the lines `1 , `2 , and `3 that meet at a point is to first choose the point P
(3 dimensions of choice) and then choose each of the three lines going through P (2 dimensions of
choice for each line). This leads to a parameter space of dimension 3 + 2 + 2 + 2 = 9.
We might expect the set of lines that meet in a point to be cut out by 12−9 = 3 bracket equations.
Three such equations spring to mind! Just take the equations that indicate that `1 meets `2 , `1 meets
`3 , and `2 meets `3 .
These do cut out a 9-dimensional set of lines in P3 but it isn’t quite the correct 9-dimensional
set. Indeed, sitting inside this set is the subset ∆ where all three lines are equal. That isn’t what we
wanted. More generally we don’t want lines in the set ∆i j , where `i = ` j either. We could write down
equations that characterize this situation and then require one of those equations to be nonzero.
That would eliminate these unwanted sets but it also makes the description of the set by equations
considerably more complicated. Moreover, the 3 equations that we used to cut out our 9-dimensional
set of lines actually contain two 9-dimensional components. One of them corresponds to the set of
lines that meet in a point and the other corresponds to the set of lines that are coplanar (such lines
must always meet each other in projective space). We could add more bracket equations that vanish
on the first component but not on the second. Adding enough equations intelligently will eventually
remove the second component from our set but again the description of the space of 3 coincident
lines becomes more complicated. This may be unavoidable but perhaps the three equations that we
started with were just a bad choice and some other set of three equations cuts out the required set
of lines. In algebraic geometry we say that a set of codimension-k (i.e., a set that has dimension k
less than its ambient space) is a complete intersection if it is cut out by k equations. However, not
all codimension-k sets are complete intersections. It can be very difficult to tell if a particular set is
a complete intersection.

4.3 The Bracket Algebra and Rings of Invariants

Many properties of geometric objects in Pd don’t depend on the coordinate system that we choose
for the ambient projective space. For instance, whether two lines in P3 intersect depends on the lines
but not the coordinate system in which we represent the lines. As another example, whether the
linearized distance constraints in a body-and-bar framework are satisfied should not depend on the
coordinate system we use to describe the ambient space. Such conditions are said to be invariant
under change of coordinates. In this section we’ll lay out the mathematical framework used to
describe invariant conditions, a topic that, perhaps surprisingly, links back up to Grassmannians and
relations among the Plücker coordinates. For an alternative introduction in the plane, see Section
3.1.1 of Chapter 3.
Automorphism Group: The automorphism group PGLd of Pd consists of all invertible linear
92 Handbook of Geometric Constraint Systems Principles

transformations on Cd+1 modulo the multiples of the identity matrix. Equivalently, it consists
of all linear transformations from Pd to itself.
Coordinates and brackets: Given n points in Pd we form the d + 1 by n matrix X of their coordi-
nates. In contrast to the earlier construction of Grassmannians, here we think of each point as
a column of X. Again we consider brackets, determinants of d + 1 by d + 1 submatrices of the
matrix X of coordinates. We label each point (and each column of X) by a symbol drawn from
a set Λ. At various times we use numbers, letters, or subscripted letters to denote the elements
of Λ, whichever is most convenient for the application at hand.
Homogeneity: A polynomial in the brackets is homogeneous of degree d if every term is the prod-
uct of d brackets. It is multilinear if every symbol in the polynomial appears in every monomial
exactly once. More generally, it is multihomogeneous if each symbol in the polynomial appears
in every monomial an equal number of times.
Algebraic and invariant properties: An algebraic property of the points is one that can be char-
acterized by the vanishing of a system of well-defined polynomials in the coordinates given by
X. Such a property is invariant if it holds independent of the coordinates chosen for the ambi-
ent space Pd . The set of collections of points where a given set of multihomogeneous bracket
polynomials evaluate to zero is an invariant set in the sense that membership in the set is an
invariant property.
Bracket algebra: Every invariant set can be written as the vanishing set of a collection of mul-
tihomogeneous bracket polynomials. The ring of invariants is also referred to as the bracket
algebra.

Plücker relations: The brackets are not independent. The relations on the brackets are gen-
erated by the alternating property: if σ is a permutation of 1, . . . , d + 1, [i1 · · · id+1 ] =
sgn(σ )[σ (i1 ) · · · σ (id+1 )] and the quadratic Plücker relations.
Each Plücker relation is determined by a choice α = (α1 , . . . , αd ) of d points and a second
choice β = (β1 , . . . , βd+2 ) of d + 2 points. The Plücker relation has the form
d+2
X
(−1)t−1 [α1 . . . αd βt ][β1 . . . β̂t . . . βd+2 ] = 0,
t=1

where the notation β̂t means to omit the term βt from the bracket.

Van der Waerden syzygies: To determine whether two polynomials in the brackets are the same,
we reduce them to a normal form using Van der Waerden syzygies. Given an integer s with
1 ≤ s ≤ d + 1 and tuples α = (α1 , . . . , αs−1 ), β = (β1 , . . . , βd+2 ), and γ = (γ1 , . . . , γd+1−s ) of
points, we define [[α β̇ γ]] to be
X
sgn(τ ∗ , τ) · [α1 . . . αs−1 βτ1∗ . . . βτd+2−s
∗ ] · [βτ1 . . . βτs γ1 . . . γd+1−s ],
τ = {τ1 < . . . < τs }
⊂ {1, . . . , d + 2}

where τ ∗ = {1, . . . , d + 2} \ τ is the complement of τ and sgn(τ ∗ , τ) is 1 if an even number of

transpositions suffices to permute the (d + 2)-tuple (τ ∗ , τ) to (1, . . . , d + 2) and −1 otherwise.
The Van der Waerden syzygy is the relation [[α β̇ γ]] = 0. The dotted notation helps identify the
symbols split among the two brackets. The Plücker relations are the Van der Waerden syzygies
with s = d + 1.
Special Positions of Frameworks and the Grassmann-Cayley Algebra 93

Tableau: Monomials in the brackets are ordered lexicographically. Given a monomial in the brack-
ets we write the brackets as rows in a tableau in lexicographic order. A tableau is standard if
its columns are sorted and nonstandard otherwise.
Normal form: Polynomials in the brackets whose monomials are all standard are said to be reduced
to normal form. Polynomials in the brackets can be iteratively reduced to normal form by
adding a multiple of an appropriate Van der Waerden syzygy to the lexicographically largest
nonstandard monomial.

4.3.1 Group Actions and Invariant Polynomials

Given a column vector P consisting of (d + 1) entries representing the coordinates of a point in pro-
jective space we can change basis in Pd , obtaining new coordinates P0 for the point, by multiplying
on the left by an invertible (d + 1) × (d + 1) matrix A:

P0 = AP.

Viewed in a slightly different way, we can imagine that the matrix A acts on the projective space,
moving the point with coordinates P to the point with coordinates P0 . However, not all invertible
matrices A actually move points: the matrices that are scalar multiples of the (d + 1) × (d + 1) iden-
tity matrix just scale the coordinates of the point P and so don’t move the point at all in projective
space. In order to recognize these matrices that stabilize all points in Pd , mathematicians speak of
the automorphism group PGLd of Pd ; it consists of all linear transformations of Pd to itself. PGLd
is the space obtained by taking the collection of invertible (d + 1) × (d + 1) matrices and identifying
those that are scalar multiples of one another.
Now suppose that we are given n points P1 , . . . , Pn in a projective space Pd . We organize these
points into a (d + 1) × n matrix

X = (xi j ) (0 ≤ i ≤ d; 1 ≤ j ≤ n),

where xi j represents the ith coordinate of point Pj ; that is, Pj = [x0 j : x1 j : . . . : xd j ]T . An algebraic
property of the n points is a property that can be expressed as the vanishing of a well-defined
polynomial function F(x01 , . . . , xdn ) of the xi j . The polynomial F defines a well-defined function
precisely if for each row i, the number of variables xi j (counted with multiplicity) from that row
appearing in each of the monomials in F is constant. The property is invariant if it does not depend
on the choice of coordinates for Pd ; that is, the property holds for P1 , . . . , Pn if and only if for each
invertible matrix A the property holds for all P10 , . . . , Pn0 , where Pj0 = APj .

Example 3 Given 3 or more points P1 , P2 , P3 , . . . in P2 , the polynomial function

 
x01 x02 x03
F = det x11 x12 x13 
x21 x22 x23

is an invariant polynomial that only depends on the first three points. If A ∈ GL3 and Pj0 = APj then
evaluating F at the transformed points Pj0 equals the product of det(A) with the evaluation of F at
the original points Pj ,
F(P10 , P20 , P30 , . . .) = det(A)F(P1 , P2 , P3 , . . .),
so the property defined by F = 0 holds independent of the coordinates. We often write [123] for the
determinant defining this polynomial F. The invariant property can be stated succinctly: the poly-
nomial [123] vanishes precisely when the three points P1 , P2 , and P3 are collinear. Stated in this way
94 Handbook of Geometric Constraint Systems Principles

it is clear that the property doesn’t depend on the coordinates on the ambient space. More generally,
for points in Pd , all determinants [ j0 j1 . . . jd ] of submatrices consisting of d + 1 of the columns of X
are invariant. Products of the brackets are invariant and any multihomogeneous polynomial (each
symbol jk appears the same number of times in every monomial) is invariant. For instance, the
polynomial [123][245] + [125][234] is invariant but the polynomial [123][245] + [124][345] is not.

In fact the First Fundamental Theorem of Invariant Theory states that the only invariant alge-
braic conditions on a collection of points in Pd are the multihomogeneous polynomials built out of
the brackets [ j0 j1 . . . jd ]. Even when an invariant algebraic condition is presented to us as a multiho-
mogeneous polynomial it can be extremely difficult to determine a corresponding condition on the
geometry of the points. This topic is treated in more detail in Section 4.5. One can also ask which
maps from a finite set of d + 1 elements to C (or any other field) arise from configurations of points,
and this point of view is discussed in Section 3.1.3 of Chapter 3.
The ring of invariants C[x01 , . . . , xdn ]PGLd consists of all sums and products of polynomials that
are invariant under a linear change of coordinates. Every element of the ring of invariants can be
expressed in terms of the brackets [ j0 j1 . . . jd ]. The brackets themselves are not algebraically inde-
pendent. The Second Fundamental Theorem of Invariant Theory describes the relations among the
brackets, a topic treated in the next section. However, the main result is that all the relations come
from identities that follow from the interpretation of the brackets as determinants. It follows that the
ring of invariants C[x01 , . . . , xdn ]PGLd can be identified with the coordinate ring of the Grassmannian
G(d, n − 1). We’ll refer to this ring as the bracket algebra, in honor of its generators.
The two different interpretations of the bracket algebra – as the invariant ring C[x01 , . . . , xdn ]PGLd
of n points in Pd and as the coordinate ring of the Grassmannian G(d, n − 1) of (d + 1)-dimensional
linear spaces of Pn−1 – arise from two different interpretations of the matrix X of coordinates. When
we view X as a collection of columns, we get the interpretation in terms of points in Pd and when
the view X as a collection of rows we get the interpretation in terms of linear spaces in Pn−1 .

4.3.2 Relations Among the Brackets

The relations among the brackets are generated by special relations called Plücker relations. The
Plücker relations all come from Cramer’s rule, a well-known result about determinants. For instance,
for 4 points in P1 , the only Plücker relation is

[12][34] − [13][24] + [14][23] = 0.

We explain how this relation arises, following the line of reasoning set out in Richter-Gebert [6].
Since the first three points are collinear we can express the first point as a linear combination of the
second and third points; that is, there is a nontrivial solution to the matrix equation
    
x02 x03 a2 x
= 01 . (4.2)
x12 x13 a3 x11

Using Cramer’s rule we find the solutions to be

[13] [21]
a2 = and a3 = .
[23] [23]

Substituting the solution into (4.2) gives a vector equation and multiplying by [23] gives a linear
dependence relation with brackets as coefficients
       
x01 x02 x 0
[23] − [13] − [21] 03 = .
x11 x12 x13 0
Special Positions of Frameworks and the Grassmann-Cayley Algebra 95

Writing [23] · 1 − [13] · 2 − [21] · 3 for the left-hand side of this equation, we bracket with the last
point and use the linearity of the determinant to compute the Plücker relation:

0 = [[23] · 1 − [13] · 2 − [21] · 3, 4] = [14][23] − [13][24] + [12][34].

For higher dimensional projective spaces there are multiple Plücker relations. Given a set of
points in Pd each Plücker relation is determined by a choice α = (α1 , . . . , αd ) of d points and a
second choice β = (β1 , . . . , βd+2 ) of d + 2 points. The Plücker relation has the form
d+2
X
(−1)t−1 [α1 . . . αd βt ][β1 . . . β̂t . . . βd+2 ] = 0,
t=1

where the notation β̂t means to omit the term βt from the bracket. The Plücker relations generate
the ideal of all relations among the brackets in the sense that every relation is obtained by summing
multiples of the Plücker relations by polynomials in the brackets. Some authors use the term Plücker
relation more generally to refer to any quadratic relation among the brackets, but we won’t follow
this convention.
The same argument that establishes the Plücker relation on P1 verifies the relation in Pd . Specif-
ically, consider the following matrix equation that expresses a linear dependence relation among
d + 2 points β1 , . . . , βd+2 in Pd ,
    
x0β2 . . . x0βd+2 a2 x0β1
 .. ..   ..  =  ..  .
 . .  .   . 
xdβ2 . . . xdβd+2 ad+2 xdβ1

Use Cramer’s rule to find the solution,

at = [β2 . . . (β1 β̂t ) . . . βd+2 ]/[β2 . . . βd+2 ] (2 ≤ t ≤ d + 2),

where the notation (β1 β̂t ) means replace the term βt in β2 . . . βd+2 with β1 . Since transposition of
columns multiplies a determinant by −1 we can rewrite this as

at = (−1)t [β1 . . . β̂t . . . βd+2 ]/[β2 . . . βd+2 ].

It follows that
d+2
X
(−1)t [β1 . . . β̂t . . . βd+2 ] · βt = [β2 . . . βd+2 ] · β1 .
t=2
Moving all terms to the right-hand side, we obtain the relation
d+2
X
(−1)t−1 [β1 . . . β̂t . . . βd+2 ] · βt = 0.
t=1

Now we place the vector in the left-hand side as the right-most vector in the (d + 1) × (d + 1) matrix
whose first d columns come from the coordinates of the points α1 , . . . , αd and take the determinant
(which must evaluate to zero). Using the linearity of the determinant we can write the determinant
as the Plücker relation
d+2
X
(−1)t−1 [β1 . . . β̂t . . . βd+2 ][α1 . . . αd βt ] = 0.
t=1

The presence of multiple Plücker relations makes it difficult to determine if two expressions
96 Handbook of Geometric Constraint Systems Principles

in the brackets are equal. One way to deal with such questions is to first reduce each expression
f to a normal form N( f ), a special representative for each collection of equivalent expressions.
(The normal form of the bracket of 3 points is also discussed in Section 3.2.1 of Chapter 3.) Two
expressions f and g are equivalent if and only if their normal forms are equal, N( f ) = N(g). The
reduction to normal form can be viewed as an implementation of the normal form algorithm from
Gröbner bases (see Sturmfels [9] or Sturmfels and White [10] for details). It can also be viewed as
an implementation of the straightening algorithm due to Alfred Young [15]. We’ll explain how to
reduce (or straighten) an arbitrary expression in the brackets to its normal form. The method uses a
more general relation that holds in the bracket algebra called a Van der Waerden syzygy. The word
syzygy is used by algebraic geometers to refer to relations among polynomial expressions such as
the Plücker relations.
A Van der Waerden syzygy on Pd is characterized in the following way. Let s be an inte-
ger between 1 and d and let α = (α1 , . . . , αs−1 ) be a tuple of s − 1 points in Pd . Further let
β = (β1 , . . . , βd+2 ) be a tuple of d + 2 points in Pd and let γ = (γ1 , . . . , γd+1−s ) be a tuple of d + 1 − s
points in Pd . Define [[α β̇ γ]] to be
X
sgn(τ ∗ , τ) · [α1 . . . αs−1 βτ1∗ . . . βτd+2−s
∗ ] · [βτ1 . . . βτs γ1 . . . γd+1−s ],
τ = {τ1 < . . . < τs }
⊂ {1, . . . , d + 2}

where τ ∗ = {1, . . . , d + 2} \ τ is the complement of τ and sgn(τ ∗ , τ) is 1 if an even number of

transpositions suffices to permute the (d + 2)-tuple (τ ∗ , τ) to (1, . . . , d + 2) and −1 otherwise. The
Van der Waerden syzygy is the relation [[α β̇ γ]] = 0. The notation is set up so that if s = d and γ = (),
then [[α β̇ γ]] is a Plücker relation. Hongbo Li gives a nice elementary proof (using a clever averaging
method combined with induction) that each Van der Waerden syzygy is in the ideal generated by
the Plücker relations (see Li [3, Proposition 2.9]).
To describe the straightening algorithm we first need to define an ordering on bracket monomials
in Pd and standard and nonstandard monomials. There is an ordering on brackets: [λ1 . . . λd+1 ] is
less than [ω1 . . . ωd+1 ] if, reading left-to-right, the first place they differ has smaller symbol in λ : that
is, there is an index t so that λi = ωi for 1 ≤ i < t and λt < ωt . Given two monomials in the brackets
we order them lexicographically. That is, we first compare their smallest brackets and if these are
different then the monomial with the smaller bracket is smaller. If these brackets are the same we
move on to compare their next smallest brackets and continue until a decision has been reached. This
is precisely how we decide on the order of two words with the same length in a dictionary, which is
why the ordering is called lexicographic. For example, [123][146][235] < [123][156][234]. Given a
monomial in the brackets, e.g., [145][234][156] we first arrange the brackets in lexicographic order
[145] < [156] < [234] and then write them as rows in a tableau T (really, a matrix, though this is the
common term in the mathematical literature), in increasing order:
 
1 4 5
T = 1 5 6 .
2 3 4

The tableau T is standard if its columns are all sorted, that is, reading down each column, the
numbers never decrease; otherwise T is nonstandard. In our example above, T is nonstandard; the
first column is sorted but the second is not and there is a violation in the second and third rows
since 5 > 3. If T is nonstandard it can be straightened by adding an appropriate Van der Waerden
polynomial. Just find the highest position with a violation in the left-most unsorted column. Suppose
that the violation occurs in column s and label the higher row involved in the violation as

[α1 . . . αs−1 β1 . . . βd+2−s ]

Special Positions of Frameworks and the Grassmann-Cayley Algebra 97

and the following row as

[βd+3−s . . . βd+2 γ1 . . . γd+1−s ].
In our example, the rows are [α1 β1 β2 ] = [156] and [β3 β4 γ1 ] = [234]. To straighten the nonstandard
monomial we subtract a multiple of the Van der Waerden polynomial [[α β̇ γ]]. In our example, the
Van der Waerden polynomial is [[15̇6̇2̇3̇4]], which equals

[156][234] − [152][634] + [162][534] + [153][624] − [163][524] + [123][564]

= [156][234] + [125][346] − [126][345] − [135][246] + [136][245] + [123][456].

Let’s take a moment to pause and observe the utility of the dotted notation; the dotted terms in
the notation [[α β̇ γ]] denote the β ’s, the terms that are split
in the expansion of the Van der Waerden
syzygy. There are 6 terms in the syzygy since there are 42 = 6 ways to pick 2 of the 4 β 0 s to put in
the first bracket.
To help straighten the monomial [145][234][156] we solve for the nonstandard monomial
[234][156] in the Van der Waerden syzygy and substitute the resulting expression into the mono-
mial; equivalently, in this instance, we subtract a multiple of the Van der Waerden syzygy (we need
to multiply by the brackets not appearing in the violation) from the monomial. The result is
[145][234][156] −[145][[15̇6̇2̇3̇4]]
= −[145][125][346] +[145][126][345] +[145][135][246] −[145][136][245] −[145][123][456]
= −[125][145][346] +[126][145][345] +[135][145][246] −[136][145][245] −[123][145][456].

The second and fourth monomials in the resulting polynomial are nonstandard. We straighten the
largest nonstandard monomial. In this case we straighten [136][145][245] by adding a multiple of
the appropriate Van der Waerden syzygy; the reader can check that this is [245][[136̇1̇4̇5̇]]. This has
the effect of producing a new polynomial in which some of the monomials might be nonstandard
but these new monomials can only be smaller than the one we’ve eliminated. Continuing in this way
we eventually reach a polynomial with only standard monomials. This is the normal or straightened
form of the original polynomial. The normal form for our example polynomial [145][234][156] is
[123][145][456] − [124][145][356] + [134][145][256].

Example 4 Consider four points a, b, c, and d in P1 . The second monomial in the bracket poly-
nomial P(a, b, c, d) = [ac][bd] − [ad][bc] is nonstandard with the violation occurring in the second
column. To straighten P we use the Plücker relation

[[ad˙ḃċ]] = [ad][bc] − [ab][dc] + [ac][db] = [ad][bc] + [ab][cd] − [ac][bd] = 0

to express the nonstandard monomial [ad][bc] in a different manner and substitute into P. We obtain
a standard monomial [ab][cd] equivalent to P.

4.4 The Grassmann-Cayley Algebra

4.4.1 Motivation
Let V be a vector space of dimension d over a field k. The Grassmann-Cayley algebra is generated
by elements that correspond to subspaces of V and has two operations, the join and meet, that cor-
respond to subspace sum and intersection under suitable conditions. We introduce the Grassmann-
Cayley algebra with d arbitrary here, and take an algebraic point of view. For a more geometric
discussion in the plane, please see Chapter 3 of this handbook by Li and Richter-Gebert.
98 Handbook of Geometric Constraint Systems Principles

Exterior product: We denote the exterior product of vectors v1 , . . . , vm ∈ V by v1 ∨ · · · ∨ vm . (Typi-

cally, the exterior product is denoted by “ ∧ ”, but we reserve that symbol for the meet operation
in the Grassmann-Cayley algebra.) In the literature on the Grassmann-Cayley algebra the ∨ is
frequently omitted, and we write v1 · · · vm for the exterior product.
Extensor: The exterior product of m vectors is an extensor of step m.
Join: The join of extensors α and β is α ∨ β . If α is an extensor of step m and β is an extensor of
step n, and the vector space spanned by the factors of α and β has dimension m + n, then their
join has step m + n; otherwise their join is zero.
Exterior algebra: The exterior algebra of V is the k-algebra generated by all of the extensors on
V ; that is, V = ⊕V ∨i , where V ∨i is the span of the extensors of step i. The exterior algebra is
the quotient of the tensor algebra obtained by setting v ∨ v = 0 for all v ∈ V, hence the name
“extensor” (standing for “exterior tensor”).
Meet: The meet of extensors v1 · · · vm and w1 · · · wn is defined if m + n ≥ d. In this, case v1 · · · vm ∧
w1 · · · wn is defined to be the following extensor of step m + n − d:
X
sgn(σ )[vσ (1) · · · vσ (d−n) w1 · · · wn ]vσ (d−n+1) · · · vσ (m) .
σ

The sum is taken over all permutations σ of {1, 2, . . . , m} such that σ (1) < σ (2) < · · · < σ (d −
n) and σ (d − n + 1) < σ (d − n + 2) < · · · < σ (m). Each such permutation σ is called a shuffle.
Note that when the step of the meet is zero, the resulting element is a bracket.
Dotted notation: As in Section 4.3.2, we use the dotted notation and write
[v̇1 · · · v̇d−n w1 · · · wn ]v̇d−n+1 · · · v̇m
for X
sgn(σ )[vσ (1) · · · vσ (d−n) w1 · · · wn ]vσ (d−n+1) · · · vσ (m) ,
σ
where the sum is over all shuffles σ .
Grassmann-Cayley algebra: The Grassmann-Cayley algebra of V is the exterior algebra of V
together with the meet operation.

4.4.2 The cross product as a Join

We show that the join of vectors in R3 is the familiar cross product, suitably interpreted.
Let V = R3 with standard basis vectors e1 , e2 , and e3 . Note that V ∨2 = V ∨ V has basis e2 ∨
e3 , e1 ∨ e3 , e1 ∨ e2 and the join of v = v1 e1 + v2 e2 + v3 e3 and w = w1 e1 + w2 e2 + w3 e3 is
(v1 e1 + v2 e2 + v3 e3 ) ∨ (w1 e1 + w2 e2 + w3 e3 )
= (v2 w3 − v3 w2 )e2 ∨ e3 − (v1 w3 − v3 w1 )e1 ∨ e3 + (v1 w2 − v2 w1 )e1 ∨ e2 .
Note that the coefficients of v ∨ w are the coefficients of the cross product of v and w. The set
{e∗i } is a dual basis for V ∗ , and if we define a linear map given by e∗1 7→ e2 ∨ e3 , e∗2 7→ e1 ∨ e3
and e∗3 7→ e1 ∨ e2 , we have an isomorphism from V ∗ to V ∨2 . Thus, v ∨ w gives a linear function
on V whose kernel is the plane spanned by v and w. So, here we can see that the join of v and w
corresponds to the subspace spanned by v and w in a natural way.
If we think of the points [v1 : v2 : v3 ] and [w1 : w2 : w3 ] in P2 , then the same computation above
shows that the extensor represents the line spanned by the points, and that the coordinates of the
extensor in this basis give the coefficients of the equation of the line. Thus, we see that lines in P2
are parameterized by a dual P2 whose coordinates correspond to step 2 extensors.
Special Positions of Frameworks and the Grassmann-Cayley Algebra 99

4.4.3 Properties of the Exterior Product

The exterior product is associative, alternating (or equivalently, antisymmetric or anticommutative)
and multilinear, and we can see that together these two properties imply that the exterior product of
a set of vectors is nonzero if and only if they are linearly independent.

Lemma 4.1
Let V be a vector space over a field k.
(a) If v ∈ V , then v ∨ v = 0.
(b) If v = r1 v1 + · · · rm vm , where ri ∈ k and vi ∈ V, then v ∨ v1 ∨ v2 ∨ · · · ∨ vm = 0.

Proof 4.1 For (1), note that the alternating property states that if we transpose the order of two
factors in an exterior product then its sign changes. As a consequence, v ∨ v = 0 for any v ∈ V, and
any exterior product with two factors the same must P be zero.
For (2), by linearity v ∨ v1 ∨ v2 ∨ · · · ∨ vm = ri (vi ∨ v1 ∨ · · · ∨ vm ), and we can see that each
summand has two factors that are the same.

From (2) of 4.1 we see that each extensor of step m gives rise to a unique m-dimensional linear
subspace of V, as the factors in the product are a basis for a vector space. In fact, given a vector
space V , the exterior product of the vectors in a basis for V is unique up to a scalar multiple.

Example 5 (Extensors and the Grassmannian) Suppose that we work in V = R4 with standard
basis e1 , e2 , e3 , e4 and U is the 2-dimensional subspace spanned by e1 and e2 . Any other basis for U
has the form ae1 + be2 , ce1 + de2 where ad − bc 6= 0. The extensor (ae1 + be2 ) ∨ (ce1 + de2 ) expands
to
(ac)e1 ∨ e1 + (ad)e1 ∨ e2 + (bc)e2 ∨ e1 + (bd)e2 ∨ e2 = (ad − bc)e1 ∨ e2 ,
and we see that U is indeed represented by a unique extensor up to a scalar multiple.
In fact, we see that for any dimension 2 vector subspace U in R4 , the exterior product of a
pair of basis vectors is an extensor in the span of the 6 extensors e1 ∨ e2 , . . . , e3 ∨ e4 obtained by
joining pairs of the four standard basis vectors. Since this extensor representing U is only unique
up to scalar multiple, we see that in fact each rank 2 vector space of R4 maps to a point in P5 .
The extensors ei ∨ e j are the Plücker coordinates of G(2, 4), and the sole Plücker relation arises by
looking at relations on the 2 × 2 minors of the 2 × 4 matrix whose rows are given by a basis for U.
This is exactly how the Plücker relation on 4 points in P1 was derived in Section 4.3.2 and indeed
the Grassmannian G(2, 4) is a degree 2 hypersurface in P5 defined by this same equation.

This example illustrates an important general phenomenon first observed in Section 4.3.1. The
bracket ring on n symbols with brackets consisting of d symbols is both the ring of invariants for n
points on Pd−1 and the coordinate ring of the Grassmannian G(d, n) parameterizing d-dimensional
subspaces of Cn (or d − 1-dimensional linear spaces in Pn−1 ).

4.4.4 Brackets and the Grassmann-Cayley algebra

When we write an extensor of step d in terms of the standard basis we can see how the alternating
and multilinear properties force brackets to arise. For example, suppose that we have d = 2, a =
a1 e1 + a2 e2 , and b = b1 e1 + b2 e2 . Then

a ∨ b = (a1 e1 + a2 e2 ) ∨ (b1 e1 + b2 e2 )
= a1 e1 ∨ (b1 e1 + b2 e2 ) + a2 e2 ∨ (b1 e1 + b2 e2 )
= (a1 b2 − a2 b1 )e1 ∨ e2
100 Handbook of Geometric Constraint Systems Principles

and we see that we have the determinant of the matrix with a and b as column vectors. In other
words, the coefficient of e1 ∨ e2 is the bracket [ab].
More generally we have

Lemma 4.2
If a1 , . . . , ad ∈ V, then a1 ∨ · · · ∨ ad = [a1 · · · ad ]e1 ∨ · · · ∨ ed .

Proof 4.2 As in the example above, if we write the ai in terms of the basis vectors ei , and use the
alternating and multilinear properties to write the resulting extensor in terms of the exterior product
of the standard basis vectors in order, the result follows.

4.4.5 The Join and Meet

The join and meet in the Grassmann-Cayley algebra were designed to capture the operations of
subspace sum and intersection. Let U and W be two subspaces of a vector space V with bases
u1 , . . . , um and w1 , . . . , wn . Note that u1 · · · um ∨ w1 · · · wn represents U +W if and only if U ∩W = 0,
otherwise the join is zero even though the subspace sum is not.
Since we have m + n ≥ d in the definition of the meet, the meet only captures the intersection of
subspaces whose span is all of V . If U +W is indeed all of V , then u1 · · · um ∧ w1 · · · wn is a sum of
extensors of step d − (m + n). In order to interpret the meet as the intersection of U and W we need
to show that the meet of extensors is again an extensor.
Example 6 Let V = R4 with the standard basis. Let u = e1 ∨ e2 ∨ e3 and w = e2 ∨ e3 ∨ e4 . Then

u ∧ w = [e1 e2 e3 e4 ]e2 ∨ e3 − [e2 e2 e3 e4 ]e1 ∨ e3 + [e3 e2 e3 e4 ]e1 ∨ e2 = [e1 e2 e3 e4 ]e2 ∨ e3 .

We see that u ∧ w is a scalar multiple of e2 ∨ e3 , and that e2 and e3 form a basis for the intersection
of the subspaces corresponding to the extensors u and w.

Lemma 4.3
Let V be a vector space with vectors u1 . . . , um forming a basis for a subspace U, and vectors
w1 . . . , wn forming a basis for a subspace W. Let u = u1 · · · um and w = w1 · · · wn . If u ∧ w 6= 0, then
u ∧ w is indeed an extensor.

Proof 4.3 Let C be a basis for U ∩W and expand C to a basis BU of U. We can also expand C to
a basis BW for W so that C ∪ (BU \C) ∪ (BW \C) is a basis for V. Let α denote the exterior product
of the elements of BU and β denote the exterior product of the elements in BW . Then because of
the alternating property, when we expand the sum in the definition of α ∧ β we find that either the
entire sum is zero or the only nonzero term is supported on the exterior product of the elements of
C. Since the exterior product of the elements of a basis for a vector space is unique up to a scalar
multiple, we see that α ∧ β is a scalar multiple of u ∧ w, which must then be an extensor and hence
corresponds to a subspace (the subspace U ∩W ).

4.5 Cayley Factorization

4.5.1 Motivation
The vanishing of a simple Grassman-Cayley expression has a geometric interpretation. For example,
in P2 , the join a ∨ b ∨ c is zero if and only if the three points a, b, and c are collinear. If we pick a
Special Positions of Frameworks and the Grassmann-Cayley Algebra 101

basis {ei } for the underlying vector space, then a∨b∨c = [abc](e1 ∨e2 ∨e3 ), and we can see that the
vanishing of the join is equivalent to the vanishing of a bracket monomial. In rigidity theory bracket
polynomials arise as the pure conditions of frameworks, and we want to understand when such a
bracket polynomial vanishes in terms of the geometry of the points in the brackets. In this section we
review an algorithm that implements Grassmann-Cayley factorization. It takes a multihomogeneous
bracket polynomial P as input and attempts to find a simple Grassmann-Cayley expression, which
expands to P if it exists. We follow the treatment in [9]. See Chapter 3.3 which contains a discussion
of Cayley factorization and Cayley expansion in the plane.

Bracket monomial: A bracket monomial is a product of brackets.

Simple Grassmann-Cayley expression: A simple Grassmann-Cayley expression is an element of
the Grassmann-Cayley algebra in which only meets and joins appear (i.e., sums are forbidden).
Atomic extensors: Given a multilinear bracket polynomial P(a1 , . . . , an ), we partition the set
{a1 , . . . , an } into equivalence classes which we call atomic extensors. Two points ai and a j
are in the same atomic extensor if P vanishes when ai and a j are equal.

4.5.2 The Pure Condition as a Bracket Monomial

The simplest generalization of the vanishing of a single bracket is the vanishing of a bracket mono-
mial. Since the bracket ring is a quotient ring, if P is a bracket polynomial there are (infinitely)
many different coset representatives of P, each of which may be obtained by adding a element of
the defining ideal of the quotient ring to P. In rigidity theory, bracket polynomials arise as the pure
conditions of body-and-bar frameworks. Via the k-fan diagrams of White and Whitely [13], we can
construct a bracket polynomial from a tie-down of a body-and-bar framework. The example below
shows that a judicious choice of tie-down may greatly simplify the analysis.
Example 7 The multigraph below is an isostatic planar body-and-bar framework in the plane.

2
a

1 d e f
b
c
3

If we tie down vertex 1 then there are three 3-fan diagrams, and the pure condition is represented by

2 2 2
a a a

1 d e f 1 d e f 1 d e f
b b b
c c c
3 3 3
102 Handbook of Geometric Constraint Systems Principles

[ae f ][bcd] − [ad f ][bce] + [ade][bc f ].

If we tie down vertex 2 there is a unique 3-fan diagram and we obtain the representative

[abc][de f ].

Of course these two representations are equivalent; one can add [[bcȧd˙ė f˙]] to the first representation
to get the second.

2
a

1 d e f
b
c
3

In fact, the first author, together with Ruimin Cai, Audrey St. John, and Louis Theran have
characterized the body-and-bar frameworks whose pure condition can be represented by a bracket
monomial. We digress briefly to give the proof of this result here, as we think it may be of interest
to those interested in rigidity theory, and it is not a lengthy proof.
Recall that if we fix a positive integer d, a graph G is a (d, d)-Henneberg graph if there exists
a sequence of graphs, G0 , . . . , Gm = G such that G0 is a graph consisting of a single vertex, and the
graph Gi is obtained by adding a new vertex vi to the vertex set of Gi−1 together with d edges from
vi to some vertices in Gi−1 (where multi-edges are allowed).
We define a multigraph G to be a (d, d)-recursively Henneberg graph if either G is a (d, d)-
Henneberg graph or if there exists a subgraph G0 such that G0 and G/G0 are (d, d)-recursively
Henneberg graphs. The notion of a recursively Henneberg graph is related to the construction of
quadratically solvable graphs due to Owen in Theorem 3.2 of [4].

a
b
1 c 2 12

e a
d f b d f
e
1 c 2
g g
4 3 4 3
h h
i i

Figure 4.3
A recursively (3, 3)-Henneberg graph G is depicted on the left. The middle graph G0 is a Henneberg
subgraph of G, and the graph on the right is G/G0 .

Theorem 4.1 (Cai, Sidman, St. John, Theran) If G is a multigraph representing an isostatic
body-and-bar framework labeled by d-vectors, then the pure condition of G can be represented
by a bracket monomial if and only if G is a recursively (d, d)-Henneberg graph.
Special Positions of Frameworks and the Grassmann-Cayley Algebra 103

Proof 4.4 Suppose that the pure condition P of G has a representation as a bracket monomial, and
proceed by induction on the number of vertices of G. If P is a single bracket, then we are in the base
case where G is a graph with two vertices joined by d edges, and G is a recursively (d, d)-Henneberg
graph. If P is a product of more than one bracket, then P factors, and by Proposition 4.4 in [13], G
has an isostatic subgraph G0 . The pure condition of G is the product of the pure conditions of G0 and
G/G0 , so these graphs both have pure conditions that are products of brackets. By induction, they
are recursively (d, d)-Henneberg graphs, so G is as well.
For the reverse direction, suppose that G is a recursively Henneberg (d, d)-graph. Then it either
consists of d edges between two vertices (in which case the pure condition is a single bracket) or
it has a proper subgraph G0 such that G0 and G/G0 are recursively Henneberg (d, d)-graphs. By
induction, the pure conditions of G0 and G/G0 are bracket monomials, and the product of these
bracket monomials is the pure condition of G.

4.5.3 White’s Algorithm for Multilinear Grassmann-Cayley Factorization

The bracket polynomials that occur as the pure condition of a body-and-bar framework are homo-
geneous and multilinear. White’s algorithm implements Grassmann-Cayley factorization for multi-
linear expressions in the bracket algebra. We follow the presentation in [9], using the same notation
for consistency.

White’s algorithm: Let P be a bracket polynomial.

(a) Partition the set of points in P into a set Q of atomic extensors. If there do not exist atomic
extensors whose step sizes sum to d, then the algorithm terminates.
(b) Choose an atomic extensor A of step d in Q, and straighten P with the variables in A first.
This yields a new coset representative of P = A · P0 . Repeat this step with atomic extensors
Q0 = Q\{A} and bracket polynomial P0 while atomic extensors of step d remain. If P0 is a
scalar, the algorithm terminates with the Cayley factorization of the original polynomial.
(c) If there exist atomic extensors E and F of steps j and k respectively, where j, k < d and
j + k > d, order the variables so that the variables in E come first, followed by the variables
in F. If the result of straightening has E first in each tableaux, followed by shuffles of F,
followed by the rest of the variables, then E ∧ F is part of the factorization. If we cannot find
such E and F, then the algorithm terminates and the polynomial cannot be factored.
(d) If E ∧ F is a factor, then delete E and F from Q. Define a new symbol G = E ∧ F, add G to
Q, and recompute the atomic extensors. Replace P by the bracket polynomial whose tableaux
have one fewer row by deleting the top row and replacing occurrences of F in the next row
by dummy variables. Start over with step 2 now.
Example 8 We illustrate the first two steps in White’s algorithm, finding atomic extensors and
straightening, on the polynomial P(a, b, c, d) = [ac][bd] − [ad][bc]. Computing
• P(a, a, c, d) = [ac][ad] − [ad][ac] = 0,
• P(a, b, c, c) = [ac][bc] − [ac][bc] = 0,
• P(a, b, a, d) = [aa][bd] − [ad][ba] = [ad][ab].
we see that the atomic extensors of P are ab and cd.
We put a > b > c > d and straighten using the syzygy S = [[aḃċd]] ˙ = [ab][cd] − [ac][bd] +
[ad][bc]. Since P + S = [ab][cd], we can see that the Cayley factorization is ab ∧ cd, and we are
done.
104 Handbook of Geometric Constraint Systems Principles

We illustrate the full algorithm on an isostatic body-and-bar framework with four bodies in the
plane.

Example 9 (Sidman, St. John, Theran)

1 h 2

i
a b f e

4 d 3
c

P = −[abc][d f i][egh] + [abd][c f i][egh] + [abc][dei][ f gh] − [abd][cei][ f gh].

(a) Atomic extensors are Q = {ab, cd, e f , gh, i}.
(b) There are no atomic extensors of step d = 3.
(c) If we put a > b > c > d > e > f > g > h > i, then P is already straightened with a > b > c > d.
Indeed, P = −[abċ][d˙ f i][egh] + [abċ][dei][
˙ f gh].

(d) Store ab ∧ cd and continue with bracket polynomial P0 = −[u f i][egh] + [uei][ f gh].
(e) The atomic extensors of P0 are Q0 = {ui, e f , gh}.
(f) Still no atomic extensors of step d = 3.
(g) Straighten with the order u > i > e > f > g > h. We have [ui f ][egh] − [uie][ f gh] = [ui f˙][ėgh].
(h) Store ui ∧ e f and continue with [vgh]. Since this is a bracket of step d = 3, we are done.
(i) Unwinding all of this, we have (((ab ∧ cd) ∨ i) ∧ e f ) ∨ gh.
(j) This can be interpreted as saying that the condition P is equivalent to the requirement that
three points must be collinear: the three points are given by intersecting lines e and f , lines g
and h, and line i with the line through the intersection of a with b and c with d.

Example 10 (Sidman, St. John, Theran) We give an isostatic planar body-and-bar framework
whose pure condition does not have a Grassmann-Cayley factorization. The pure condition of the
graph below is [ab f ][cdi][egh] − [abg][cdi][e f h] + [abi][cde][ f gh]. We checked that it is not fac-
torable using an implementation of White’s algorithm in Macaulay 2 [1].

4 g 3

i
a b e
h

1 d 2
c
References 105

It remains an open problem to develop an algorithm that produces a Grassmann-Cayley factor-

ization of an arbitrary bracket polynomial. One interesting result regarding the Grassmann-Cayley
factorization of arbitrary bracket polynomials is due to Sturmfels and Whiteley.
Theorem 4.2 (Sturmfels-Whiteley, [11]) If P is a multihomogeneous bracket polynomial there ex-
ists a bracket monomial M so that MP has a Grassmann-Cayley factorization.
As Sturmfels notes in [9], the proof is constructive but the degree of the monomial M may be very
large, so a factorization algorithm based on this theorem would be impractical in many cases. If
we are given M and then factor MP, we see that if M 6= 0 then P = 0 is equivalent to geometry
encoded in the Grassmann-Cayley factorization. That is, as long as the degeneracies encoded in
the vanishing of M do not arise, then the vanishing of the bracket polynomial is equivalent to the
geometric interpretation of the Grassmann-Cayley factorization of MP.
Acknowledgments. Ruimin Cai worked on Theorem 4.1 as part of her undergraduate research at
Mount Holyoke College, and the proof presented here is due to joint work with the first author,
Audrey St. John, and Louis Theran. The examples of body-and-bar frameworks appearing in the
text were produced jointly with Audrey St. John and Louis Theran. Code for Grassmann-Cayley
factorization was implemented by the first author and Josephine Yu at the August 2010 Macaulay 2
workshop funded by NSF grant number 0964128 and NSA grant number H98230-10-1-0218.

References
[1] Daniel R. Grayson and Michael E. Stillman. Macaulay2, a software system for research in
algebraic geometry. Available at http://www.math.uiuc.edu/Macaulay2/.
[2] Gerard Laman. On graphs and rigidity of plane skeletal structures. J. Engrg. Math., 4:331–
340, 1970.
[3] Hongbo Li. Invariant algebras and geometric reasoning. World Scientific, 2008.
[4] John C. Owen and Steve C. Power. The non-solvability by radicals of generic 3-connected
planar Laman graphs. Trans. Amer. Math. Soc., 359(5):2269–2303, 2007.
[5] Hilda Pollaczek-Geiringer. Über die gliederung ebener fachwerk. Zeitschrift fur Angewandte
Mathematik und Mechanik (ZAMM), 7:58–72, 1927.
[6] Jürgen Richter-Gebert. Perspectives on projective geometry.
[7] Hal Schenck. Computational algebraic geometry, volume 58 of London Mathematical Society
Student Texts. Cambridge University Press, Cambridge, 2003.
[8] Karen E. Smith, Lauri Kahanpää, Pekka Kekäläinen, and William Traves. An invitation to
algebraic geometry. Universitext. Springer-Verlag, New York, 2000.
[9] Bernd Sturmfels. Algorithms in invariant theory. Texts and Monographs in Symbolic Com-
putation. Springer, New York, second edition, 2008.
[10] Bernd Sturmfels and Neil White. Gröbner bases and invariant theory. Adv. Math., 76(2):245–
259, 1989.
[11] Bernd Sturmfels and Walter Whiteley. On the synthetic factorization of projectively invariant
polynomials.
106 References

[12] Tiong-Seng Tay. On the generic rigidity of bar-frameworks. Adv. in Appl. Math., 23(1):14–28,
1999.
[13] Neil White and Walter Whiteley. The algebraic geometry of motions of bar-and-body frame-
works. SIAM J. Algebraic Discrete Methods, 8(1):1–32, 1987.
[14] Neil L. White and Walter Whiteley. The algebraic geometry of stresses in frameworks. SIAM
J. Algebraic Discrete Methods, 4(4):481–511, 1983.
[15] Alfred Young. The Collected Papers of Alfred Young (1873–1940).
Chapter 5
From Molecular Distance Geometry to
Conformal Geometric Algebra
Timothy F. Havel
Energy Compression Inc., Boston, Massachusetts

Hongbo Li
Academy of Mathematics and Systems Science, Chinese Academy of Sciences;
School of Mathematics, University of Chinese Academy of Sciences, China

CONTENTS
5.1 Euclidean Distance Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.2 The Distance Geometry Theory of Molecular Conformation . . . . . . . . . . . . . . . . . . . . . . . . 110
5.3 Inductive Geometric Reasoning by Random Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.4 From Distances to Advanced Euclidean Invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.5 Geometric Reasoning in Euclidean Conformal
Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

5.1 Euclidean Distance Geometry

Glossary
Semimetric space: A set A with nonnegative symmetric function ρ : A × A → R≥0 that vanishes
only on the diagonal: ρ(a, b) = 0 ⇔ a = b for all a, b ∈ A.
Metric space: The domain A of a semimetric ρ that satisfies the triangle inequality: ρ(a, b) ≤
ρ(a, c) + ρ(b, c) for all a, b, c ∈ A.
Metric vector space: A real vector space V equipped with a quadratic form Q : V → R, or
equiv-
alently, a bilinear form R : V × V → R, where R(v, w) = 12 Q(v) + Q(w) − Q(v − w) for all
v, w ∈ V.
Metric affine space: A set P together with an antisymmetric translation operator →
−
·· : P × P → V
onto a metric vector space (V, Q) which satisfies pq = 0 ⇔ p = q and pq + −
−→ −→ → = −
qr → for
pr
all p, q, r ∈ P.

Euclidean space: A metric affine space E = (P, (V, Q), →

−
·· ) on which the quadratic form is positive-
− → −→
definite, meaning Q( pq ) ≥ 0 and Q( pq ) = 0 ⇔ p = q for all p, q ∈ P.

107
108 Handbook of Geometric Constraint Systems Principles

Euclidean embedding: An injective mapping ε : A ,→ P of a semimetric space (A, ρ) into a Eu-

clidean space E such that ρ ε −1 (p), ε −1 (q) = Q(− →) for all p, q ∈ ε(A); when such a map-

2
pq
ping exists, we say that ρ is embeddable in Euclidean space.
n-dimensional Euclidean distance function: The metric ρ : A×A → R≥0 of a metric space (A, ρ)
which can be embedded in a Euclidean space En = (P, (V, Q), →
−
·· ) of dimension n = dim(V)
but not n − 1.
n-dimensional Euclidean group: Given a Euclidean space En of dimension n, a fixed origin o ∈ P,
a translation t ∈ V and a rotation R ∈ O, where O is the group of isometries of (V, Q), the n-
dimensional Euclidean group En may be defined via its action ω on P, namely q = ω t,R (p) ⇔
−→ = t + R−
oq →; composition in E is given by ω 0 0 ◦ ω = ω 0 0 0 , which shows that
op n t ,R t,R t +R t,R R
this definition is origin-independent.

Distance geometry may be regarded as the study of “geometric” metric spaces [39]. These are metric
spaces (A, ρ) with a transitive group action such that the metric is invariant with respect to the action
induced on the unordered pairs of A. Early investigators including Karl Menger, Johan Seidel, and
Leonard Blumenthal demonstrated that the classical elliptic, Euclidean, and hyperbolic geometries
are completely characterized by their metrics alone [4, 40, 54]. Implicit in their work is the fact
that all the invariant geometric quantities and relations of these geometries can be expressed in a
coordinate-free fashion using the interpoint distances as the primitive notion.
For n-dimensional Euclidean geometry, the group in question is the n-dimensional Euclidean
group En of translations composed with proper and improper rotations. The n = 2 case dates back
at least to Heron of Alexandria, who in AD 60 gave us what is now known as Heron’s formula for
the area A of a triangle with vertices a, b, c in terms of the lengths of its sides da,b , db,c , da,c . Letting
2s = da,b + db,c + da,c be its perimeter, this is given by:

A2 = s(s − da,b )(s − da,c )(s − db,c ) (5.1)

This turns out to be the first in a sequence of determinants of matrices involving squared interpoint
distances, known as Cayley-Menger determinants. These were first shown by Arthur Cayley to be
proportional to the squared volume D of the simplex spanned by its N vertices a, b, c, . . . , d:

0 1 1 1 ... 1 
 
 2 2 2 

1
 0 da,b da,c . . . da,d 
1 d 2 0 2
db,c 2 
. . . db,d
 a,b 
−(−2)N−1 ((N − 1)!)2 D(a, b, c, . . . , d) :=  1 d 2 2
db,c 0 2
. . . dc,d 

 a,c 
. .. .. .. .. .. 
 .. . . . . . 

1 d 2 2
db,d 2
dc,d ... 0 
a,d
(5.2)
Note in particular that D(a, b) = da,b 2 , an alternative notation which we will frequently use.

Given a semimetric ρ : A × A → R≥0 , the Cayley-Menger determinant on a finite sequence of

N points [a, b, c, . . . , d] ∈ AN is obtained by setting dx,y ≡ ρ(x, y) for all subscripts x, y in Equation
(5.2). Menger’s intrinsic characterization of the Euclidean distance function may now be stated as
follows.
Theorem 5.1 – A semimetric ρ : A × A → R≥0 on a (not necessarily finite) set A is Euclidean of
dimension n if and only if there exists a sequence of n + 1 distinct points [a, b, c, . . . , d] ∈ An+1 such
that the Cayley-Menger determinants on those points satisfy
D(a, b), D(a, b, c), . . . , D(a, b, c, . . . , d) > 0 , (5.3)
From Molecular Distance Geometry to Conformal Geometric Algebra 109

and for every pair of points f, g ∈ A we also have

0 = D(a, b, c, . . . , d, f) = D(a, b, c, . . . , d, f, g) . (5.4)

Theorem 5.1 is actually a corollary of a stronger theorem, also due to Menger.

Theorem 5.2 – [Menger’s Theorem] A semimetric ρ : A × A → R≥0 is Euclidean of dimension
n if and only if all Cayley-Menger determinants on M ≤ n + 1 points are nonnegative, at least
one determinant on n + 1 points is strictly positive, all determinants on n + 2 points vanish, and a
Cayley-Menger determinant on at least one set of n + 3 points is nonnegative (in which case it is
necessarily zero).
Another way of stating this theorem is that if every subset of n + 2 points can be isometrically
embedded in an n- but not generally (n − 1)-dimensional Euclidean space, then the semimetric is
Euclidean of dimension n unless A consists of exactly n + 3 points and the Cayley-Menger determi-
nant on those n + 3 points is strictly negative. Such semimetrics are called pseudo-Euclidean, and
can be obtained by a general construction known as isogonal conjugation, starting from an arbitrary
full-dimensional (n + 2)-point subset of n-dimensional Euclidean space [4].
Note that for metric spaces the restriction ρ(x, y) = 0 ⇔ x = y in the foregoing is trivial since
an easy analysis of Equation (5.1) with A2 = D(x, y, z) shows that 0 = dx,y only if dx,z = dy,z for all
z ∈ A, so the points x and y may be identified. Henceforth, by a slight abuse of notation, whenever
a semimetric is Euclidean we will not distinguish between a set of points in its domain and the
embedded image thereof. Indeed we shall generally restrict ourselves to Euclidean spaces, since
what is of interest to us is the Euclidean interpretation of Cayley-Menger determinants and related
homogeneous polynomials in the interpoint distances squared D(x, y).
Another, at first glance rather different, intrinsic characterization of the Euclidean metric is the
following theorem due to I. J. Schoenberg [53]:

Theorem 5.3 – [Schoenberg’s Theorem] A semimetric ρ : A × A → R≥0 on a finite set A =

{a1 , . . . , aN } is Euclidean dimension n < N if and only if Schoenberg’s quadratic form
N N
1 XX
QD (x) := − D(ak , a` ) xk x` , (5.5)
2
k=1 `=1

with D(a, a0 ) = da,a 2 0 2

0 := ρ(a, a ) as before, is positive semidefinite of rank n on the subspace of

[x1 , . . . , xN ]> =: x ∈ RN defined by 1> x = Nk=1 xk = 0, where 1 is a column of 1’s.

P
 
Now let D be the matrix of squared distances D(a j , ak ) j,k=1,...,N , and D̃ be D augmented by a 0-th
row and column of 1’s with a 0 on the diagonal, as in Cayley-Menger determinants. Then on the
subspace 1> x = 0 the augmented form

Q̃D (x̃) := − 12 x̃> D̃ x̃ (5.6)

is numerically equal to Schoenberg’s quadratic form QD for all x̃ = [x0 ; x] ∈ RN+1 with x0 = 0.
Moreover, given any set of n + 1 independent points in n-dimensional Euclidean space, say
{p1 , . . . , pn , pn+1 }, the barycentric coordinates x ∈ Rn+1 (1> x = 1) of any point x versus this ba-
sis can be found by solving the linear system of equations:
    
0 1 1 ··· 1 x0 1
1
 0 D(p1 , p2 ) · · · D(p1 , pn+1 )
  x1 
   D(x, p1 ) 
 
 1 D(p1 , p2 )
 0 · · · D(p2 , pn+1 )  x2  = 
  
 D(x, p2 ) 

(5.7)
 .. .. .. .. ..   ..   .. 
. . . . .   .   . 
1 D(p1 , pn+1 ) D(p1 , p2 ) · · · 0 xn+1 D(x, pn+1 )
110 Handbook of Geometric Constraint Systems Principles

The 0-th coordinate x0 is equal to the determinant of the augmented matrix for the linear system
obtained by dropping the 0-th column from the matrix above.
In the special case that D(pk , pn+1 ) = 1 for k = 1, . . . , n and D(p j , pk ) = 2 for all j and k with
1 < j < k ≤ n, the basis constitutes an orthonormal frame with origin pn+1 , and the barycentric
coordinates x1 , . . . , xn are just the usual Cartesian coordinates of x with respect to this frame. From
these observations it may be seen that Schoenberg’s form is nothing but the squared Euclidean
lengths of the interpoint vectors obtained by taking the differences of their barycentric coordinates
[8]. It is striking nonetheless that these facts are naturally expressed in a space of n + 2 dimensions.
We will presently provide a more complete geometric interpretation of this (n + 2)-dimensional
space. Meanwhile, to further spur the reader’s curiosity, we note that the rank of the matrix of
squared distances D also has something to say about the geometry of the embedded points. For
example, the four-point determinant expands to:

2 2 2 
 
 0 dp,q dp,r dp,s
 
 2 2 2
 dp,q 0 dq,r dq,s 




 2 2 2
 = − dp,q dr,s +dp,r dq,s +dp,s dq,r −dp,q dr,s +dp,r dq,s +dp,s dq,r
d
 p,r dq,r 0 dr,s  


 2 2 2 dp,q dr,s − dp,r dq,s +dp,s dq,r dp,q dr,s +dp,r dq,s − dp,s dq,r
 dp,s dq,s dr,s 0 

(5.8)
in analogy to Heron’s formula (5.1). The non-negativity of the last three factors on the right is
known as Ptolemy’s inequality, which holds as an equality if and only if the four points lie on a
circle or a line in a plane of dimension 2. More generally, the rank of an N × N matrix of squared
Euclidean distances is m < N if and only if the N points lie on a (m − 2)-dimensional hypersphere
in Euclidean space (of possibly infinite radius, in which case it is a (m − 2)-dimensional Euclidean
subspace).
More recent results in Euclidean distance geometry, including the fact that the inverse of the
bordered matrix appearing in Cayley-Menger determinants is a matrix of inner products of outer
normals to the corresponding n-dimensional simplex bordered by the barycentric coordinates of its
orthocenter, and the applications of this fact to the classification of such simplices, may be found in
the late Miroslav Fiedler’s book [14].

5.2 The Distance Geometry Theory of Molecular Conformation

Glossary

Distance constraint: Lower and upper bounds `p,q ≤ up,q on the distance between a pair of points
p, q; the constraint is exact when `p,q = up,q .

Chirality constraint: A subset χ̃(p, q, . . . , r) of the power set 2{−1,0,+1} containing the possible
signs of the oriented volume of the n-dimensional simplex spanned by the ordered (n + 1)-
tuple of points p, q, . . . , r.
Conformation: Any element of the (nN − n(n + 1)/2)-dimensional manifold of orbits of a set of
N points in an n-dimensional Euclidean space, under the action of the Euclidean group En , that
satisfies the given distance and chirality constraints.
From Molecular Distance Geometry to Conformal Geometric Algebra 111
H H
H H
H H
C C C
x
C C C
H H
H H
H H
Figure 5.1
Standard chemical depictions of the three most common local structures in organic (carbon-based)
chemistry, wherein carbon atoms are labeled C, hydrogen H, chemical bonds are drawn as solid
lines and other fixed distances are drawn with light dashed lines: At a carbon-carbon single bond,
as in the ethane molecule shown on the left, the tetrahedra of atoms around each carbon can rotate
around the bond’s axis; at a double bond, as in the ethylene molecule in the center, the atoms bonded
to the carbons prefer to be coplanar with the carbons and one another; and at a triple bond, as in the
acetylene molecule on the right, all atoms are collinear with the carbon bond.

Conformation space: The set of orbits, with boundary and singularities, consisting of all confor-
mations as above; it may be obtained by letting the Euclidean group act on the real semialge-
braic variety defined by the distance and chirality constraints relative to an arbitrary orthonor-
mal coordinate frame.

Triangle inequality limits: The infimum and supremum of each individual distance over all N-
point metric spaces consistent with the distance constraints.
Euclidean distance limits: The infimum and supremum of each individual distance over all sets of
N points in Euclidean space the distances of which satisfy the distance constraints.
n-dimensional distance limits: The infimum and supremum of each individual distance over all
sets of N points in n-dimensional Euclidean space (n < N) which satisfy the distance and chi-
rality constraints.

Euclidean distance geometry is the natural foundation of a coordinate-free approach to the study
of molecular conformation [8]. A molecule is a set of N > 1 atoms which are held together in
three-dimensional Euclidean space by chemical bonds, despite being constantly jostled by random,
thermally driven collisions with other molecules. Over time, the atoms of the molecule undergo a
random, but highly biased, walk within a real but “fuzzy” variety, with boundaries and singularities,
embedded within the (3N − 6)-dimensional space of orbits (unless the conformation is collinear, in
which case it’s 3N − 5) under the action of the 3D (three-dimensional) Euclidean group of transla-
tions and rotations [11]. This fuzzy variety is often referred to as the “conformation space” of the
molecule. In the case of large molecules such as the proteins and nucleic acids commonly encoun-
tered in biological organisms, the number of atoms may be in the thousands or even millions, and
the dimensionality of conformation space by itself is daunting.
The usual approach to understanding this fuzzy variety is based on computational brute force.
Starting from a given (perhaps experimentally determined) conformation, Newton’s equations of
motion are solved in an iterative process known as molecular dynamics, using an analytic (semi-
empirical) or quantum mechanical (first principles) approximation to the internal energy of the
molecule [52]. This produces a time-ordered sequence of thousands or even millions of conforma-
tions, each represented with essentially meaningless precision by a set of e.g., Cartesian coordinates
for the atoms versus an arbitrary coordinate frame. In order to obtain an approximation to a possible
physical trajectory of an single molecule, a small time step of order femtoseconds must be used.
112 Handbook of Geometric Constraint Systems Principles

As a result, successive conformations in such sequences differ only slightly from one another, and
only in the (often unattainable, even with today’s supercomputers) limit of simulations spanning
milliseconds or more can one hope to have obtained a reasonably complete sample of the conforma-
tion space of large molecules. Although the trajectory itself can be viewed as a movie, the problem
of making sense of the massive amounts of data in a trajectory is also highly nontrivial.
The preceding paragraph hardly does justice to an immense area of applied computational sci-
ence, which recently earned the 2013 Nobel Prize in Chemistry [58]. Yet the fact remains that
despite many heroic attempts to sample conformation space exhaustively and to render this high-
dimensional set of chemically relevant conformations comprehensible, we have a long way to go.
In contrast to analytic approaches such as molecular dynamics, the distance geometry approach to
the study of molecular conformation is synthetic in nature: A “molecule” is simply defined by the
totality of invariant geometric properties associated with its conformation space. While many differ-
ent geometric properties could be (and are) used, considerable simplicity and generality is attained
by focusing on the most elementary yet fundamental, namely the interatomic distances. For exam-
ple, the distances da,b between bonded pairs of atoms {a, b} are generally the same to within a few
percent in all possible conformations of the molecule, and so may be treated as fixed constants to a
good approximation. Similarly, the angles between pairs of chemical bonds {a, b} and {b, c} with
a common atom b are often also nearly constant, and given that the bond lengths themselves are
constant may be determined by treating the “geminal” distance da,c as constant. More generally, all
the distances in any essentially rigid portion of the molecule may be treated as fixed constants (see
Figure 5.1).
In any molecule with a nontrivial conformation space there will also be distances that are not
even approximately constant, but that does not mean they can assume any combination of values
consistent with embedability in 3D Euclidean space. If nothing else, atoms repel one another at short
distances even when there are no chemical bonds between them, and this effectively imposes lower
bounds on all nonbonded and nongeminal pairs of atoms. These lower bounds are typically modeled
as a sum of hard-sphere radii associated with the chemical types of the atoms. While the attractive
forces between pairs of atoms become insignificant at large separations, the cumulative effects of
many such interactions, possibly combined with the molecule’s interactions with its surroundings,
may also lead to nontrivial upper bounds on nonbonded, nongeminal distances in any of its possible
conformations. The distance geometry approach to the study of molecular conformation is predi-
cated on the empirical observation that one can do a reasonable job of describing the conformation
space of just about any molecule by specifying lower and upper bounds on its interatomic distances
(together with the chirality constraints introduced in the next paragraph). This implicit characteriza-
tion of conformation space is much more compact, and also more comprehensible, than a molecular
dynamics trajectory. It may in fact be used to summarize the results of such a simulation, simply by
setting the bounds to the minimum and maximum values of the distances observed in the trajectory.
By themselves, however, neither the distances nor their bounds can distinguish a molecule from
its mirror image, since the distances are invariant under reflections. A chiral molecule is one that
cannot be superimposed on its mirror image by translation, rotation and intramolecular motions
such as bond rotations, in which case its chirality specifies which mirror image it is. Chirality is
most often encountered in molecules wherein four chemically distinct kinds of atoms are bonded to
a common carbon atom, which hence lie on the vertices of a tetrahedron with the carbon atom in
its interior. The chirality of this tetrahedron can be specified by the sign of its oriented volume V
relative to a given ordering of its vertex atoms a, b, c, d, namely


χ(a, b, c, d) := sign V (a, b, c, d) . (5.9)
From Molecular Distance Geometry to Conformal Geometric Algebra 113
H H H
H H H
C C C
HO CO2 H HO CO2 H HO2 C OH
C C C
HO2 C OH HO CO2 H HO CO2 H
dextro meso levo

Figure 5.2
The dextro and levo forms of tartaric acid are enantiomers, or mirror images, of one another and
hence have the same interatomic distances in at least one pair of their possible conformations,
whereas the meso form has an internal plane of symmetry in at least one of its possible confor-
mations and so is achiral.

The oriented volume of the tetrahedron, in turn, is given by the determinant

 
1 1 1 1
 
1  xa xb xc xd 
V (a, b, c, d) := , (5.10)
3!  ya yb yc yd 

 
 za zb zc zd 

where x, y, z are the Cartesian coordinates of the subscript atoms. Note this definition of chirality
applies to any rigid quadruple of chemically distinct atoms in a molecule, not merely those bonded
to a common carbon atom. It can be generalized to nonrigid quadruples whenever these have a
common chirality in all possible conformations of the molecule. The chirality of the molecule as a
whole is determined by the chirality of any one of its chiral quadruples, or chiral centers as they are
known. Note however that if any conformation of the molecule has an internal plane of symmetry,
then the molecule will be achiral even if it has chiral centers (see Figure 5.2).
As might be expected, the distances and chiralities are not wholly independent. Instead, the
relative chirality of any pair of chiral quadruples is given by


χ(a1 , a2 , a3 , a4 ) χ(b1 , b2 , b3 , b4 ) = sign D(a1 , a2 , a3 , a4 ; b1 , b2 , b3 , b4 ) , (5.11)

where D(a1 , . . . , aM ; b, . . . , bM ) is more generally a nonsymmetric M × M Cayley-Menger determi-

nant of the form:
 
0
 1 1 ··· 1 
2 2
 1 da1 ,b1 da1 ,b2 ··· da21 ,bM 
 
 
2 2
−(−2)M−1
((M−1)!) D(a1 , ..., aM ; b1 , ..., bM ) :=  1 da2 ,b1 da2 ,b2
2  ··· da22 ,bM  (5.12)

. .. .. .. .. 
 .. . . . . 

1 d 2 2 2
 
aM ,b1 daM ,b2 ··· daM ,bM 

These determinants may be interpreted geometrically as the products of oriented volumes whenever
the 2M points a1 , . . . , aM , b1 , . . . , bM are embeddable in a Euclidean space of dimension n = M − 1.
Surprisingly, they contain none of the distances D(a j , ak ) or D(b j , bk ) for j and k with 1 ≤ j, k ≤ M.
The following determinantal identity, which may be derived from Jacobi’s theorem on a minor of the
adjugate, shows conversely that the consistency of the distances as a whole with this interpretation,
together with the nonnegativity of all the symmetric Cayley-Menger determinants on n + 1 or fewer
114 Handbook of Geometric Constraint Systems Principles

d d
c

b ⇐⇒ b
a a

c
Figure 5.3
The two roots of the four-point Cayley-Menger determinant D(a, b, c, d), viewed as a quadratic in
2 , correspond to planar conformations in which the atoms c, d are on the same (left) or opposite
dc,d
(right) sides of the line of the bond connecting atoms a, b. Chemists refer to these as the cis and
trans conformations of the bond a − b.

points, implies that the semimetric ρ(a, b) := da,b is n-dimensional Euclidean:

n2
(n−1)2
D(a1 , ..., an−1 ) D(a1 , ..., an−1 , b, c) = D(a1 , ..., an−1 , b) D(a1 , ..., an−1 , c)
− D(a1 , ..., an−1 , b; a1 , ..., an−1 , c)2 (5.13)

Moreover the following “cross-swap” relation

D(E; F) D(G; H) = D(E; G) D(F; H) (5.14)

holds among any four (n + 1)-tuples of points E, F, G, H in an n-dimensional Euclidean space.

We now consider the implications of these relations among the distances for bounds on the
distances among the set of atoms A of a molecule. Given upper bounds ua,b , ub,c on two of the
distances among three atoms, the triangle inequality implies that the third distance can be no greater
than their sum. This implicit bound ūa,c := ua,b + ub,c is called the upper triangle inequality limit
on the third distance. Any explicit bound ua,c > ūa,c on the third distance is therefore redundant, in
that we can set ua,c := ūa,c without changing the set of conformations consistent with the bounds.
Similarly, the inverse triangle inequality imposes lower limits on da,c of the form `¯a,c := max(`a,b −
ub,c , `b,c − ua,b ), and if `a,c < `¯a,c we may set `a,c := `¯a,c without changing the conformation space
consistent with the bounds.
Iterating on these relations, we find that the “tightest” upper limits ū among the atoms A are
equal to the shortest paths in a graph with edge lengths equal to the given upper bounds u, whereas
the lower limits are given by

`¯a,b = max max(`a,c − ūb,c , `b,c − ūa,c ) ,

(5.15)
c ∈A

for all a, b ∈ A. Thus the lower and upper limits can be computed in polynomial time by finding
the shortest paths in a graph on two disjoint copies A, A0 of the atoms. The lengths of the edges
within each copy are equal to the given upper distance bounds, whereas the directed edge lengths
connecting each pair of atoms a ∈ A, b0 ∈ A0 are equal to the negatives of the corresponding lower
bounds, −`a,b0 [8].
The polynomial-time computability of the triangle inequality limits is a benefit of the factoriz-
ability of the three-point Cayley-Menger determinants into linear factors (Equation 5.1). The non-
negativity of the four-point Cayley-Menger determinants also impose limits on the distances for
any given set of lower and upper bounds, but no polynomial-time or even closed (provably finite)
algorithm for their computation is known. In the simplest case that all but one of the distances
among four atoms are known, the sixth distance must lie between the square roots of the roots
From Molecular Distance Geometry to Conformal Geometric Algebra 115
2 obtained by expanding the four-point Cayley-Menger
of the quadratic in the squared distance dc,d
determinant along its last row and column:

4 2
288 D(a, b, c, d) = −2 D(a, b) dc,d + 32 D(a, b, c; a, b, d)|d 2 =0 dc,d
c,d

+ 288 D(a, b, c, d)|d 2 =0 (5.16)

c,d

These correspond to the two conformations shown in Figure 5.3.

The implications of this “tetrangle inequality” for the lower and upper limits on one of the
distances among any four atoms, given lower and upper bounds on the other five distances, are
also rather more complicated than the triangle inequality limits, and are illustrated in Figure 5.4.
The extension of these results to include chirality constraints may be found in Ref. [51]. Similar
results can be derived for the pentangle and, in principle, even higher-order inequalities, but are
too complicated to be presented here. These higher-order inequalities, moreover, are valid in nD
Euclidean space for arbitrary n ≤ N − 1, and hence do not have much to say about the 3D case that
is of interest to chemists. In particular, they cannot derive the implications of the sphere packing
constraints mentioned above, due to the fact that nonbonded, nongeminal pairs of atoms can be
viewed as hard nonoverlapping spheres. In chemistry, the consequences of these ubiquitous lower
distance bounds are called excluded volume effects, and in large molecules they can eliminate the
vast majority of otherwise possible conformations. To some extent, they could be derived by taking
account of the fact that the pentangle inequality holds as an equality in three dimensions. In this
case, however, the active distance bounds depend on the relative chiralities of the five quadruples
of atoms in each quintuple, further complicating the combinatorics of a search for what we call
the three-dimensional Euclidean distance limits. Thus, despite its potential relative to an analytical
approach, there are many open problems that need to be solved before a purely distance theoretic
approach to molecular conformation can deal with complex large molecules. This has led to the
hybrid approach introduced in the next section.

5.3 Inductive Geometric Reasoning by Random

Sampling
Glossary
Conformational ensemble: A random sample from conformation space.
Inductive geometric reasoning: Imputing the common geometric features of a conformational en-
semble to be necessary consequences of the distance and chirality constraints that define it.
Metrization: The computation of a random metric space from between the upper and lower triangle
inequality limits.

EMBED algorithm: An approach (not a true “algorithm”) to computing the Cartesian coordinates
of a conformational ensemble based on diagonalization of an estimate of the metric matrix
derived from a distance matrix obtained by metrization.
Metric matrix: The matrix of inner products of all pairs of vectors between an arbitrary origin,
usually taken as the centroid of the atoms, and the atoms themselves.
Error function: A function of the Cartesian coordinates of the atoms that increases monotonically
116 Handbook of Geometric Constraint Systems Principles

c d d c a b a b
d c

a b a b c d

a b a

c d a b

b c d c d

Figure 5.4
The combinations of bounds among four atoms a, b, c, d that can be active constraints in the con-
formations wherein the lower (top row) and upper (bottom row) tetrangle inequality limits on dc,d
are attained. Distances at their lower bounds are indicated with a heavy line, while distances at their
upper bounds are indicated with wavey lines. If we interpret the lower bounds as the struts and the
upper bounds as cables in a tensegrity framework, then the frameworks of the two configurations on
the left and right sides of this figure are globally rigid [7].The combinations of bounds among four
atoms a, b, c, d that can be active constraints in the conformations wherein the lower (top row) and
upper (bottom row) tetrangle inequality limits on dc,d are attained. Distances at their lower bounds
are indicated with a heavy line, while distances at their upper bounds are indicated with wavey
lines. If we interpret the lower bounds as the struts and the upper bounds as cables in a tensegrity
framework, then the frameworks of the two configurations on the left and right sides of this figure
are globally rigid [7].

with increasing violations of the distance constraints as well as the oriented volume constraints
derived from the distance and chirality constraints.

In this section we introduce an inductive approach to geometric reasoning, which can be used
to gain insight into the distance geometry descriptions of even very large and complex molecules
such as proteins and nucleic acids. This is done by computing a random sample of conformation
space, called a conformational ensemble, and then analyzing it to determine the common geometric
features of all its members. Given an ensemble that is sufficiently large and random, those com-
mon geometric features will, with high probability, be necessary consequences of the distance and
chirality constraints. The geometric features of interest need not themselves be either distances or
oriented volumes, but can include aggregate properties such as the size and shape of the molecule,
which atoms are on its surface versus buried under surface atoms, or even kinematic features such
as correlations in atomic motions. Whereas one certainly could obtain a conformational ensemble
from a molecular dynamics simulation of sufficient duration, the techniques we shall now describe
can produce random ensembles of diverse conformations in far less time [8].
2 ]
Let A = [a, b, . . . , z] be a total ordering of the set of atoms A, and let L = [`a,b a,b∈A , U =
2
[ua,b ]a,b∈A be symmetric, nonnegative and zero-diagonal matrices containing the lower and upper
distance bounds squared (L ≤ U). Also let C = [χ̃a,b,c,d ]a,b,c,d∈A be the antisymmetric tensor of
possible chiralities for its quadruples of atoms, which has its values in the power set 2{−1,0,+1} .
Given this input, the EMBED algorithm proceeds through the following steps, which are described
in greater detail in the ensuing text:
¯ ū from the interatomic distance bounds `, u, as de-
(a) Compute the triangle inequality limits `,
From Molecular Distance Geometry to Conformal Geometric Algebra 117

scribed in the previous section; to the extent computationally practical, it is advantageous to

also compute selected tetrangle or higher-order distance limits;
(b) Generate a random metric space within the distance limits, i.e., a symmetric zero-diagonal
matrix R = [ra,b ]a,b∈A such that `¯a,b ≤ ra,b ≤ ra,c + rb,c ≤ ūa,c + ūb,c for all a, b, c ∈ A;
2 ] 1
(c) Let D := [ra,b a,b∈A be the corresponding matrix of squared metrics, and let M := − 2 PDP be
>
the projection on the subspace orthogonal to the vector 1 of 1’s, i.e. P := I − 11 /N;

(d) Let QΛΛQ> be an eigenvalue decomposition of M, with the eigenvalues λ1 ≥ λ2 ≥ · · · ≥ λN

arranged in nonincreasing order;
(e) If the third eigenvalue λ3 < 0, return to step 2; otherwise, let [q1 , q2 , . . . , qN ] be the columns
of the orthogonal matrix Q;
√ √ √
(f) The vectors x = λ1 q1 , y = λ2 q2 , z = λ3 q3 can now be taken as center-of-mass (as-
suming all atoms have equal masses), principal-axis (ditto), three-dimensional Cartesian co-
ordinates for the atoms, with the property that the squared distances computed from these
coordinates approximate those in D and hence do not seriously violate the distance bounds,
since L ≤ D ≤ U by construction;

(g) Compute the chirality error function C(x, y, z) (see below for its definition) for the embedded
coordinates and the mirror image thereof, and keep the mirror image that minimizes this
function;
(h) Using your favorite global optimization algorithm, minimize the violations of the distance
bounds and chirality constraints with respect to the Cartesian coordinates until the residual
violations are acceptably small;

(i) Repeat steps 2 through 8 until you have as many conformations as desired for your ensemble.
Following this overview, we now explain the individual steps in greater detail.
The random metric space given by the matrix R in step 2 is obtained by setting the lower and
upper bounds `a,b , ua,b on the distance between an arbitrary pair of atoms equal to some random
value `¯a,b ≤ ra,b ≤ ūa,b between the corresponding limits, and then computing the new triangle
inequality limits implied by the modified bounds `a,b := ra,b =: ua,b . Iterating in this fashion over
all pairs of atoms, one attains the desired metric space once all the lower and upper bounds have
become equal. Adopting a term from topology, this process has been dubbed metrization. Given that
the majority of triangle inequality limits are unchanged by equating the lower and upper bounds
between one pair of atoms, the new distance limits can be obtained in far less time than would be
needed to compute the limits from scratch in each iteration. Indeed, if one treats an arbitrary atom as
the root of a shortest-path tree, and chooses the distances to each of the other, as-yet untreated atoms
sequentially, the shortest-path tree can be updated in time linear in N on each iteration. This enables
the random metric spaces R to be computed in time proportional to at most N 3 , just as with a single
set of triangle inequality limits. Restricting the iteration to all distances from a single root atom to
all other as-yet untreated atoms does however introduce correlations in the values of the metrics and
hence a bias in the sampling, unless the order in which the atoms are taken as root atoms is chosen
randomly on each pass over steps 2 through 8.
The projection in step 3 may be shown to transform the entries of the matrix of Schönberg’s
quadratic form − 12 D = − 12 [da,b
2 ] 1 2
a,b∈A := − 2 [ra,b ]a,b∈A into a matrix M = [ma,b ]a,b∈A of inner prod-
ucts, called the metric matrix, given via the law of cosines as
2 2 2
ma,b = 12 d0,a


+ d0,b − da,b , (5.17)
118 Handbook of Geometric Constraint Systems Principles

where the squared distances to the new point 0 are given by

2 1 X 2 1 X 2
d0,a = da,f − df,g . (5.18)
N 2N 2
f ∈A f, g ∈ A

2 is the squared
In the three-dimensional Euclidean case, it was first shown by Lagrange that d0,a
distance between the point a and the centroid 0 of the points. This interpretation also holds for
any set of points embedded in a metric affine space of any dimension, and hence for an arbitrary
semimetric space (cf. Theorem 5.3).
In steps 4 through 6, it can be shown that the metric matrix M0 = xx> + yy> + zz> minimizes
the Frobenius norm kM − M00 kF over all rank 3 positive semidefinite matrices M00 . This is why the
matrix of three-dimensional squared Euclidean distances, obtained from the embedded coordinates
using the Hadamard (or entrywise) matrix product “” as D0 =

(x1> − 1x> )  (x1> − 1x> ) + (y1> − 1y> )  (y1> − 1y> ) + (z1> − 1z> )  (z1> − 1z> ), (5.19)

is a good approximation to the non-Euclidean estimate D = [ra,b 2 ]

a,b∈A thereof obtained from
metrization. As a result, when the input distance bounds are reasonably “tight,” the embedded
atomic coordinates x, y, z will be close to those of a conformation that satisfies the distance con-
straints, and the relative chiralities of the chiral centers will also be largely correct.
Step 7 attempts to also get the absolute chiralities χ largely correct, essentially by choosing
the mirror image of the embedded coordinates which minimizes the volume-weighted sum of the
chirality constraint violations over all chiral centers. Then in step 8, a real-valued error function,
which measures the violations of the distance and chirality constraints, is minimized to eliminate
the residual constraint violations. In the case of the distance constraints, the function

`¯a,b
2 2 2 !
da,b (x, y, z)2
X  
G(x, y, z) = max 0, − 1 + max 0, 2
−1 , (5.20)
da,b (x, y, z)2 ūa,b
a,b∈A

where da,b (x, y, z)2 = (xa − xb )2 + (ya − yb )2 + (za − zb )2 , has been found to be relatively free of
local minima and to give chemically reasonable structures even when (as is generally the case)
the residual distance constraint violations cannot be eliminated entirely. The total error function is
then obtained by adding the chirality error function C to G. Letting Va,b,c,d (x, y, z) be the oriented
volumes computed from the coordinates as in Equation (5.10), and L̄a,b,c,d , Ūa,b,c,d be the minimum
and maximum unsigned volumes consistent with the distance and chirality constraints among the
corresponding quadruple, this is defined as
X X   2
C(x, y, z) = min 0, χ ·Va,b,c,d (x, y, z) − L̄a,b,c,d +
a,b,c,d ∈ A χ ∈ χ̃a,b,c,d  2 
0∈/ χ̃a,b,c,d max 0, χ ·Va,b,c,d (x, y, z) − Ūa,b,c,d
X X  2
+ max 0, χ ·Va,b,c,d (x, y, z) − Ūa,b,c,d (5.21)
a,b,c,d ∈ A χ ∈ χ̃a,b,c,d
0 ∈ χ̃a,b,c,d χ 6= 0
X
+ Va,b,c,d (x, y, z)2 ,
a,b,c,d ∈ A
χ̃a,b,c,d = {0}

where a, b, c, d ∈ A4 implies a sum over all quadruples a, b, c, d ∈ A with a < b < c < d with respect
to its total ordering A. Note that in the last line of this equation, we are treating planarity as a
“chirality” constraint, enforced by penalizing any oriented volume other than zero.
From Molecular Distance Geometry to Conformal Geometric Algebra 119

Figure 5.5
A conformational ensemble for the protein E. coli Flavodoxin, obtained by homology modeling
using distance constraints derived from multiple sequence alignments with four homologous (evo-
lutionarily related) Flavodoxins, superimposed upon the crystal structure of the protein which was
subsequently determined (heavy lines). Since each conformation contains nearly 1000 atoms, only
the main-chain atoms, aromatic side-chains, and flavin cofactor are shown for simplicity. Repro-
duced with permission from [19].

The EMBED algorithm is best known as a means of determining “the” conformation of proteins
and other large biomolecules in aqueous solution from Nuclear Magnetic Resonance (NMR) spec-
troscopy [57, 59]. The most important geometric parameters that can be extracted from these spectra
are extensive lists of distances between pairs of hydrogen atoms that are less than 0.5 nm in all con-
formations. While many initially doubted that such imprecise constraints could suffice to determine
the relative positions of the polypeptide backbone atoms of the protein with a precision approaching
that of crystallographic methods, another Nobel prize in chemistry proved them wrong [21, 62]. In
the years since, other spectroscopic parameters have gained prominence which, like all Euclidean
invariants, can be expressed as functions of the interatomic distances, but these functions are rela-
tively complicated [48]. As a result, the propagation of experimental uncertainties makes it difficult
to derive reasonably tight bounds on individual distances from them. Distance-based conformational
sampling has however also been applied to many other problems in computational chemistry, includ-
ing ring closure [23], drug design [9], and protein structure prediction from sequence alignments
with homologous proteins of known structure [19]. Figure 1.5 dipicts a conformational ensemble
for the protein Flavodoxin obtained from the latter application, known as homology modeling.
Early implementations of the EMBED algorithm minimized the error function by gradient de-
scent methods such as conjugate gradients [17, 22]. Subsequently, considerably more robust con-
vergence was obtained by dynamical simulated annealing, wherein the error function is treated as
if it were the energy function in a molecular dynamics simulation [16, 18]. Starting from a high
temperature (T ∝ 2(C + G)/3R, where C and G are the distance and chirality error functions as
above and R is the gas constant), the temperature is gradually reduced to zero by scaling down the
velocities at each point along the trajectory. While it may seem ironic to be using molecular dy-
namics in a method designed to improve upon it for conformational= sampling purposes, the error
functions turn out to be much softer and to have far fewer local minima than physically realistic
energy functions. Moreover, since the intent is not to produce a physically meaningful trajectory
but merely to simulate the natural tendency of thermodynamic systems to fall into a low-energy
state upon annealing, a time step of order milliseconds can be utilized, thereby covering a far larger
region of conformation space in a given amount of real time. Although convergence to zero error
is seldom attained in large problems, the residual violations of the distance and chirality constraints
are often small enough to be of no chemical significance.
Over the last 25 years, a great many different global optimization algorithms have been applied
to the minimization of such error functions [38, 43]. Some are able to solve large problems much
120 Handbook of Geometric Constraint Systems Principles

faster than dynamical simulated annealing, even starting from random rather than embedded coordi-
nates. In many cases, unfortunately, these algorithms have made chemically unrealistic assumptions
about the problem to be solved, particularly regarding the precision of the distances. Indeed a great
deal of this work has focused on the distance matrix completion problem, wherein all distances
are either known exactly or are completely unknown, and the goal is to fill in the missing entries
of the distance matrix so as to make it embeddable in a low, if not three-dimensional, Euclidean
space. Even when reasonably general distance and chirality constraints are assumed, however, the
optimization problem was solved well enough for most chemical intents and purposes years ago.
In the hope of spurring future work on more important issues, we close this section by giving two
open problems, the solutions of which would considerably broaden the scope of distance geometry’s
chemical applications.

Problem I: Tight Outer Estimates of the 3D Euclidean Distance Limits. Typically, if one takes
the minimum and maximum values of the distances over a conformational ensemble as the bounds
from which to compute a new conformational ensemble using the EMBED algorithm, one finds that
the fit of the distances in the unoptimized conformations to the new, as well as the old, distance
bounds is so good that very little, and certainly no demanding global, optimization is needed. This
means that improved estimates of the 3D Euclidean distance limits would be an effective means of
reducing the EMBED algorithm’s dependence on global optimization. A computationally tractable
means of estimating these limits would moreover help one to discover errors in the distance bounds;
in the simple case of the triangle inequality, for example, contradictions of the form
ua,b + ub,c < la,b (5.22)
are always discovered when computing the triangle inequality limits. The ability to discover, and
pinpoint, such contradictions using the higher order inequalities and equalities of distance geome-
try would be very useful to chemists, who must gather large quantities distance information from
diverse sources in the course of studying the conformation of a given molecule. It also enables an
approach to geometric reasoning via contradiction. One could, for example, attempt to prove that
two atoms were necessarily in close proximity to one another by imposing a large lower bound on
the distance between them, and then testing the combined constraints for contradictions.
The estimate of the distance limits L̄, Ū obtained by taking the minimum and maximum values
of the distances over a large conformational ensemble is what we call an inner estimate: The true
3D distance limits L̄3D , Ū3D will always be looser than L̄, Ū, i.e. L̄3D ≤ L̄ and Ū3D ≥ Ū (because
it is impossible for a finite ensemble to completely sample the infinitude of possible combinations
of values for the distances). In order to prove the existence of contradictions (ū < `)¯ with no false
positives (erroneous contradictions), these inequalities need to point the other way, i.e. L̄3D ≥ L̄
and Ū3D ≤ Ū. Estimates of the 3D Euclidean distance limits with this property will be called outer
estimates. In principle they can be obtained from any suitable (computationally tractible) relaxation
of the system of inequalities and equalities which the distances and chiralities mutually satisfy
in three dimensions [41]. In practice, such a relaxation must not be overly simplistic, or at least
must provide a significant improvement over that obtained from the triangle inequality alone. Most
importantly, it must provide information about the lower limits on the distances which result from
the hard sphere packing constraints characteristic of atoms in 3D molecules, as described previously.
The fact that packing and covering problems have long been studied in mathematics and yet still
contain many open questions implies the problem of computing such tight outer estimates will be
challenging. The payoff for such an algorithmic approach will nevertheless be comparably large,
and could well impact mathematics as much as chemistry. Some preliminary work on the one-
dimensional version of this problem has already been performed [13].

Problem II: Computations in Rings of Symmetric Distance Polynomials. Many experimental

sources of information on molecules yield information on the distances among atoms but are am-
From Molecular Distance Geometry to Conformal Geometric Algebra 121

biguous concerning which atoms those distances stem from. For example, X-ray diffraction provides
a 3D map of the density of electrons in a molecule, which has peaks at the positions of its atoms.
With large molecules or low-resolution data, fitting a conformation with all the correct bond lengths
and angles to such an electron density map so that all the atoms are near the correct peaks can be
difficult or impossible. It is much easier to find the peaks and compute all the distances between
them, although these distances may contain significant errors from peak overlaps due to the limited
resolution of the map. A representation of the distance matrix that did not assume a total ordering
of the atoms, or equivalently, which was invariant to permutations of the underlying set, could help
solve such problems. This would be done by fitting a conformation of the molecule, represented by
a distance matrix versus a total ordering of its atoms, to the permutation-invariant representation
so as to make the permutation-invariant representation calculated by symmetrization of the former
match the latter to within the error bounds on its values. As usual in distance geometry, it will be
easier to do this using the squared distances rather than the distances themselves.
In the simplest nontrivial case of three vertices (points, atoms), the polynomials in question are:
2 +d2 +d2
P1 (a, b, c) = da,b b,c a,c
2 d2 +d2 d2 +d2 d2
P2 (a, b, c) = da,b b,c a,b a,c b,c a,c (5.23)
P3 (a, b, c) = 2 d2 d2
da,b b,c a,c

These are the same as the elementary symmetric polynomials among three indeterminates, so there’s
nothing new here. It is worth noting, nonetheless, that 4P1 (a, b, c) − P0 (a, b, c)2 = 16D(a, b, c) ≥ 0,
so these polynomials inherit dependencies from those of Euclidean distances.
The four-point elementary symmetric distance polynomials are something new:
2 2 2 2 2 2
P1 (a, b, c, d) = da,b + db,c + da,d + db,c + db,d + dc,d
2 2 2 2 2 2
P2 (a, b, c, d) = da,b dc,d + da,c db,d + da,d db,c
2 2 2 2 2 2 2 2 2 2 2 2
P3 (a, b, c, d) = da,b da,c + da,b da,d + da,c da,d + da,b db,c + da,b db,d + db,c db,d +
2 2 2 2 2 2 2 2 2 2 2 2
da,c db,c + da,c dc,d + db,c dc,d + da,d db,d + da,d dc,d + db,d dc,d
2 2 2 2 2 2 2 2 2 2 2 2
P4 (a, b, c, d) = da,b da,c da,d + da,b db,c db,d + da,c db,c dc,d + da,d db,d dc,d
2 2

2 2 2 2

2 2

2 2 2 2


P5 (a, b, c, d) = da,b + dc,d da,c db,d + da,d db,c + da,c + db,d da,b dc,d + da,d db,c
2 2

2 2 2 2


+ da,d + db,c da,b dc,d + da,c db,d
2 2 2 2 2 2 2 2 2 2 2 2
P6 (a, b, c, d) = da,b da,c db,c + da,b db,d db,d + da,c da,d dc,d + da,d db,c dc,d (5.24)
2 2 2

2 2 2 2 2 2

2 2 2
P7 (a, b, c, d) = da,b + da,c + da,d db,c db,d dc,d + da,b + db,c + db,d da,c da,d dc,d +
2 2 2

2 2 2 2 2 2

2 2 2
da,c + db,c + dc,d da,b da,d db,c + da,d + db,d + dc,d da,b da,c db,c
2 2 2 2 2 2 2 2 2 2 2 2
P8 (a, b, c, d) = da,c da,d db,c db,d + da,b da,d db,c dc,d + da,b da,c db,d dc,d
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
P9 (a, b, c, d) = da,c da,d db,c db,d dc,d + da,b da,d db,c db,d dc,d + da,b da,c db,c db,d dc,d +
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
da,b da,c da,d db,d dc,d + da,b da,c da,d db,c dc,d + da,b da,c da,d db,c db,d
2 2 2 2 2 2
P10 (a, b, c, d) = da,b da,c da,d db,c db,d dc,d

It may be observed that these polynomials are in a one-to-one correspondence with the loopless,
undirected and unlabeled graphs on four vertices, save for the empty graph. In this correspondence,
P9 is the complementary graph of P1 (paths of length 1), P8 is the complement of P2 (pairs of disjoint
edges), P7 the complement of P3 (paths of length 2), P6 the complement of P4 (stars), and P5 is self-
122 Handbook of Geometric Constraint Systems Principles

complementary (paths of length 3). It is readily shown that

−P1 (a, b, c, d)P2 (a, b, c, d) + P5 (a, b, c, d) − 2P6 (a, b, c, d) = 288D(a, b, c, d) ≥ 0 , (5.25)

and these polynomials inherit other dependencies from those of Euclidean distances as well, e.g.,

− P1 (a, b, c, d)2 + 2 P2 (a, b, c, d) + P3 (a, b, c, d)

(5.26)
= 4 D(b, c, d) + D(a, c, d) + D(a, b, d) + D(a, b, c) ≥ 0 .

Moreover, since we have 10 polynomials in only 6 indeterminates, it is clear they will satisfy 4 other
algebraically independent equality relations as well. These appear to be rather too complicated to
be found simply by inspection.
In order to work in the ring generated by these polynomials, a straightening (or normal form)
algorithm is needed that allows one to determine if one expression can be transformed into another
using the equality relations (syzygies). Ideally, this should be done in the ring of symmetric dis-
tance functions, obtained in the limit as the number of vertices N → ∞. In addition to the fully
symmetric case introduced above, it would also be of some chemical interest to be able to do the
same for subgroups of the full symmetric group, particularly those which constitute a restriction of
the symmetric group to one or more subsets of the vertices [10, 46]. Finally, there are several gener-
alizations of purely mathematical interest. These include the analogous systems of polynomials for
general symmetric matrices, which correspond to unlabeled graphs with loops, and nonsymmetric
matrices, which correspond to unlabeled digraphs with loops. The boolean case also promises a
novel approach to the graph isomorphism problem.

5.4 From Distances to Advanced Euclidean Invariants

Glossary
Null vector: A nonzero vector of a metric vector space whose inner product with itself is zero.
Positive/Negative vector: A vector of a real metric vector space whose inner product with itself is
positive/negative.
Conformal point at infinity: An extraneous point outside an n-dimensional Euclidean affine space
obtained by pinching the hyperplane at infinity (tangent space) of the affine space to a single
point, resulting in a compact topological space homeomorphic to an n-dimensional sphere.
Conformal model: A realization of n-dimensional Euclidean distance geometry in an (n + 2)-
dimensional Minkowski vector space, or more accurately, as an n-dimensional quadratic surface
of the Minkowski space.

In a Euclidean vector space, for two vectors a, b representing two points relative to an arbitrary
origin, their difference a − b represents the displacement from b to a in the Euclidean affine space.
The displacement has Euclidean distance dab , so
2
dab = (a − b)2 = (a − b) · (a − b) = a2 + b2 − 2a · b. (5.27)

What are these inner products geometrically: a2 , b2 , and a · b? They always depend on the reference
point (the origin of all vectors representing points), and are geometrically meaningless.
From Molecular Distance Geometry to Conformal Geometric Algebra 123

e0
e

Figure 5.6
Illustration of the map f from a 1-space (the x-axis) of Rn to the space Rn+1,1 spanned by Rn , e, e0 .
The x-axis is mapped to a parabola in an affine plane normal to the e0 -axis.

If there were a vector space model of Euclidean geometry in which a · b had geometric meaning,
what could it be? It would have to be some geometric relation between the two points. The only
possibility is the distance, the unique invariant between the two points. Then a · a has to be zero,
as it measures the distance between a point and itself. So “point” a should be represented by a null
vector. From (5.27) we get
(a − b)2 d2
a·b = − = − ab . (5.28)
2 2
1
 
..
 . 
Let e1 , e2 , . . . , en , e0 , e be a basis of Rn+1,1 with metric 
 1  . Then

 0 −1 
−1 0

x12 + x22 + · · · + xn2

 
f : (x1 , x2 , . . . , xn ) ∈ Rn 7→ x1 , x2 , . . . , xn , 1, − ∈ Rn+1,1 (5.29)
2

is an isometry from n-dimensional Euclidean space into (n+2)-dimensional Minkowski space, such
that (5.28) is satisfied (Figure 5.6). The basis vector e denotes the conformal point at infinity of the
Euclidean space, and e0 = f(0) denotes the origin of Rn . Denote the orthogonal complement of the
plane spanned by e and e0 in Rn+1,1 by Rn . Then

x2
f(x) := x + e0 + e, ∀x ∈ Rn . (5.30)
2
The image space of isometry f is

Ne := {y ∈ Rn+1,1 | y · y = 0, y · e = −1}. (5.31)

This unusual nonlinear realization of Euclidean distance geometry turns out to bring an amazing
124 Handbook of Geometric Constraint Systems Principles

amount of benefits in both geometric modeling and symbolic manipulation [30]. It is a classical
result [3] that any orthogonal transformation of Rn+1,1 induces a conformal transformation in Ne ∪
{e} = En ∪ {conformal point at infinity} via the conformal model, and conversely, any conformal
transformation can be generated in this way. This justifies the name “conformal” of the model.

Glossary
Minkowski representation: A conformal object such as line, circle, plane, sphere is represented
by an extensor of Minkowski signature in the Grassmann algebra over the conformal model.
Dual Minkowski representation: Hodge dual of the Minkowski representation.
Affine representation: A point “a” (null vector) in the conformal model has affine representation
e ∨ a, where “∨” denotes the outer product in Grassmann algebra. Suitable for affine geometry.
Dual affine representation: Hodge dual of the affine representation.
Homogeneous model: Discard the constraint a · e = −1 for null vector a representing an affine
point. Two null vectors represent the same point (including the conformal point at infinity) if
and only if they differ at most by scale [32].

The following are Minkowski representations of planes/lines/circles/spheres:

• Line ab: e ∨ a ∨ b. Point d on line C: d ∨ C = 0.
• Plane abc: e ∨ a ∨ b ∨ c.
• Circle abc (circum-circle of the triangle): a ∨ b ∨ c.
• Sphere abcd (circum-sphere of the tetrahedron): a ∨ b ∨ c ∨ d.
The following are dual representations:
• Plane normal to vector n ∈ R3 and passing through point a:
positive vector p := n + (a · n)e. Point c on plane p: c · p = 0.
• Sphere with center ø and radius R: positive vector ø − R2 e/2.
• Line of intersection of two planes p1 , p2 : p1 ∨ p2 .
• Circle as intersection of two spheres, or a sphere and a plane p1 , p2 : p1 ∨ p2 .
The following are Euclidean interpretations of the inner products of positive vectors and null
vectors in the conformal model [29]:
• For two points c1 and c2 (null vector representation),

dc21 c2
c1 · c2 = − . (5.32)
2

• For a point c and a hyperplane n + δ e (positive vector representation),

c · (n + δ e) = c · n − δ . (5.33)

The inner product is positive, zero, or negative, if and only if the vector from the hyperplane
to the point is along n, zero, or along −n, respectively. Its absolute value equals the distance
between the point and the hyperplane.
From Molecular Distance Geometry to Conformal Geometric Algebra 125

• For a point c1 and a sphere c2 − R2 e/2,

R2 R2 − dc21 c2
c1 · (c2 − e) = . (5.34)
2 2
This inner product, known as the power of the point with respect to the sphere, is positive,
zero, or negative, if and only if the point is inside, on, or outside the sphere respectively. Let
dmax and dmin be respectively the maximal distance and minimal distance between point c1
and the points on the sphere. Then the absolute value of (5.34) equals dmax dmin /2.
• For two hyperplanes n1 + δ1 e and n2 + δ2 e,
(n1 + δ1 e) · (n2 + δ2 e) = n1 · n2 . (5.35)

• For a hyperplane n + δ e and a sphere c − R2 e/2,

R2
(n + δ e) · (c − e) = n · c − δ . (5.36)
2
It is positive, zero, or negative, if and only if the vector from the hyperplane to the center c of
the sphere is along n, zero, or along −n respectively. Its absolute value equals the distance
between the center and the hyperplane.
• For two spheres c1 − R21 e/2 and c2 − R22 e/2,

R21 R2 R2 + R22 − dc21 c2

(c1 − e) · (c2 − 2 e) = 1 . (5.37)
2 2 2
It is zero if the two spheres are perpendicular to each other, i.e., they intersect at 90◦ .
More generally, when the two spheres intersect, (5.37) equals the cosine of the angle of
intersection multiplied by R1 R2 .
Any negative vector s ∈ Rn+1,1 after rescaling, can be written in the standard form s = c +
2
(R /2) e, where c2 = 0. The Euclidean geometric meaning of the inner product of a negative vector
with another vector in Rn+1,1 is as follows [34]:
• Let c1 , c2 be points represented by null vectors in Ne . Then

R2 dc2 c + R2
c1 · (c2 + e) = − 1 2 . (5.38)
2 2
Let y be any point on the sphere of center and radius (c2 , R) such that line c2 y ⊥ line c1 c2 ,
then (5.38) equals −dc21 y /2.
• Let n be a unit vector in Rn , then
R2
(n + δ e) · (c + e) = c · n − δ . (5.39)
2
Let y be any point on sphere (c2 , R) such that line c2 y ⊥ direction n, then (5.39) equals the
signed distance from hyperplane n + δ e to point y.
• Let c1 , c2 ∈ Ne , then
R21 R2 R2 − R22 − dc21 c2
(c1 − e) · (c2 + 2 e) = 1 . (5.40)
2 2 2
Let y be any point on sphere (c2 , R) such that line c2 y ⊥ line c1 c2 , and let dmax , dmin be
respectively the maximal signed distance and the minimal signed distance from point y to
the points on sphere c1 − R21 e/2, then (5.40) equals −dmax dmin /2.
126 Handbook of Geometric Constraint Systems Principles

• Let c1 , c2 ∈ Ne , then

R21 R2 R2 + R22 + dc21 c2

(c1 + e) · (c2 + 2 e) = − 1 . (5.41)
2 2 2
Let y be any point on sphere (c2 , R2 ) such that line c2 y ⊥ line c1 c2 , and let z be any point
2 /2.
on sphere (c1 , R1 ) satisfying c1 z ⊥ c1 y. Then (5.41) equals −dyz
While the conformal model is suitable for geometric interpretation, the homogeneous model is
convenient for symbolic computation.

Glossary
Second point of intersection: For two circles/lines ab1 c1 and ab2 c2 in the plane, “point” a (may
be either a finite point or the conformal point at infinity) is trivially a point of intersection;
besides this “point,” the other “point” of intersection is denoted by b1 c1 ∩a b2 c2 . In particular,
if b1 c1 ∩a b2 c2 = a, then the two circles/lines are tangent to each other at “point” a; in the
case of two lines, they are parallel to each other (so that they are “tangent” to each other at the
conformal point at infinity a = e).
Reduced meet product: For two extensors a ∨ b1 ∨ c1 and a ∨ b2 ∨ c2 in the Grassmann-Cayley
algebra (also known as Grassmann algebra, cf. the previous chapter and the notations therein,
or [60]) over the conformal model of the Euclidean plane,

(b1 ∨ c1 ) ∧a (b2 ∨ c2 ) := [ab1 c1 c2 ]b2 − [ab1 c1 b2 ]c2 = [ab1 b2 c2 ]c1 − [ac1 b2 c2 ]b1

is the reduced meet product [33] of the two extensors with respect to a. It can be generalized to
the conformal model of n-dimensional Euclidean space.
Clifford algebra: The Clifford algebra [6, 49] over a metric vector space V n is the quotient of
⊗(V n ) modulo the two-sided ideal generated by elements of the form a ⊗ a − a · a for all a ∈ V n .
The quotient of the tensor product is called the Clifford multiplication (or geometric product),
denoted by the juxtaposition of elements. The Clifford multiplication is associative, multilinear,
and satisfies ab = a · b + a ∨ b for all a, b ∈ V n .
Monic Clifford monomial: The Clifford multiplication of a sequence of vector variables of V n .
Geometric algebra: Clifford himself called the algebra now bearing his name “geometric algebra.”
This term is nowadays reserved specifically for the version of Clifford algebra [24, 25, 28]
where the geometric product is defined on the Grassmann-Cayley algebra over the base metric
vector space V n .
Conformal geometric algebra: It refers to the geometric algebra (Clifford algebra + Grassmann-
Cayley algebra) over Rn+1,1 , together with the n-dimensional conformal interpretations of alge-
braic elements in this algebra [1, 2, 12, 15, 26, 27, 30, 50, 56]. The first application of Clifford
algebra to the study of the conformal model dates back to 1892 in the work of E. Müller [44, 45].

In the Grassmann-Cayley algebra over the conformal model R3,1 of the Euclidean plane, the
meet product of two circles/lines ab1 c1 and ab2 c2 in the plane is

(a ∨ b1 ∨ c1 ) ∧ (a ∨ b2 ∨ c2 ) = a ∨ {(b1 ∨ c1 ) ∧a (b2 ∨ c2 )}.

The resulting outer product has two ingredients: the trivial point of intersection a (null vector), and
another vector (b1 ∨c1 )∧a (b2 ∨c2 ) that is generally not null. The second vector needs to be changed
From Molecular Distance Geometry to Conformal Geometric Algebra 127

into a null vector in order to represent the point b1 c1 ∩a b2 c2 in the conformal model. This can be
realized by reflecting null vector a with respect to invertible vector (b1 ∨ c1 ) ∧a (b2 ∨ c2 ).
In Clifford algebra, the reflection of a vector y with respect to an invertible vector x is defined
by
y 7→ Adx (y) := − xyx−1 .
In our mission of representing point b1 c1 ∩a b2 c2 in the conformal model, the following homoge-
neous reflection is more convenient to use, as it allows for (b1 ∨ c1 ) ∧a (b2 ∨ c2 ) to be non-invertible,
i.e., null:
1
Ny (x) := xyx.
2
Then b1 c1 ∩a b2 c2 is represented multiplicatively by Na ((b1 ∨ c1 ) ∧a (b2 ∨ c2 )).

Glossary
Basic algebraic invariant: An algebraic G-invariant is a polynomial of coordinate variables, such
that for any coordinate transformation in a group G, the polynomial is invariant [47]. In oriented
Euclidean geometry, the basic algebraic invariants are squared distances and signed volumes of
simplexes [63].
Advanced algebraic invariant: It refers to a scalar-valued monomial of an algebra such that if
the multiplications in the monomial are expanded relative to a coordinate system, then the
monomial becomes a polynomial of basic algebraic invariants [34].
Clifford bracket algebra: The Clifford bracket algebra of vector variables a1 , . . . , am in metric vec-
tor space V n is the commutative ring generated by two kinds of Clifford brackets [33], the
“hyper-determinant” and “hyper-inner product,” by prolonging the bracket of n vector vari-
ables and inner product of two vector variables respectively with the Clifford multiplication:
for all k, l ≥ 0,

a1 · a2 = ha1 a2 i prolonged to ha1 a2 . . . a2k i := ha1 a2 . . . a2k i0 ,

[a1 . . . an ] = (a1 ∨ · · · ∨ an )∼ prolonged to [a1 a2 . . . an+2l ] := ha1 a2 . . . an+2l i∼
n .

Here “∼ ” is the Hodge dual operator in Grassmann algebra, and h ii is the i-grading operator in
Grassmann algebra that extracts the ith-graded part of the argument [25].
Null bracket algebra: Clifford bracket algebra of null vector variables [34], i.e., aa = 0 for any
vector variable a.

For points a1 , · · · , aM ∈ Rn , let the dot symbol denote the contraction (also called the inner
product, or interior product) of the Clifford algebra inherited from the contraction of the tensor
algebra over Rn+1,1 . This may also be defined via the following Laplace expansion:

 
 0
 −1 −1 ... −1 

 −1 0 − 21 da21 a2 ... 1 2
− 2 da1 aM 
 
 
1 2
(e ∨ f(a1 ) ∨ . . . ∨ f(aM )) · (e ∨ f(a1 ) ∨ . . . ∨ f(aM )) =  −1 − 2 da2 a1 ... − 21 da22 aM 
 
 0

 .. .. .. .. ..
 
.

 . . . . 
 
 −1 − 1 da2 a − 1 da2 a
 
2 M 1 2 M 2 ... 0 
(5.42)
128 Handbook of Geometric Constraint Systems Principles

The right side is the Cayley-Menger determinant (5.2) up to a factor of ((M − 1)!)2 . When the
left side is replaced by (e∨f(a1 )∨. . .∨f(aM ))·(e∨f(b1 )∨. . .∨f(bM )), then its Laplace expansion is
the general nonsymmetric Cayley-Menger determinant (5.12). Some applications of the conformal
model in distance geometry can be found in [20].
When the conformal point at infinity in (5.42) is replaced by a point, we get

(f(a1 ) ∨ · · · ∨ f(aM )) · (f(a1 ) ∨ · · · ∨ f(aM )) = det(f(ai ) · f(a j ))M×M = det(−da2i a j /2)M×M ,

from which we get the following n-dimensional Ptolemy’s Theorem [3]:

Let a1 , · · · , an+2 be points in En , then they are on a hypersphere or hyperplane if and only if
det(da2i a j )(n+2)×(n+2) = 0.
Furthermore, if we replace some/all points on the left side of (5.42) with positive/negative vectors
of Rn+1,1 , we get various extensions of the Cayley-Menger determinant to conformal objects other
than points [42].
Cayley-Menger determinants and their various extensions are advanced algebraic invariants:
each is a monomial of the inner-product Grassmann algebra over the conformal model, and each
when expanded equals a polynomial in the squared distances, or in the inner products of non-
null vectors representing geometric objects other than points. The above proof of n-dimensional
Ptolemy’s Theorem shows the effectiveness of using advanced algebraic invariants in geometric
reasoning.
Still the Cayley-Menger determinant and all its generalizations are limited to monomials of
length (maximal number of inclusive vector variables) at most 2(n + 2), as the outer product of
any n + 3 vector variables in Rn+1,1 is zero. They are still inefficient in more complex geometric
reasoning, and advanced algebraic invariants of unlimited length are called for.
The two Clifford brackets extend the determinant of n + 2 vector variables of Rn+1,1 to a “hyper-
determinant” of n + 2l vector variables for arbitrary l ≥ 1, and extend the inner product of two
vector variables to a “hyper-inner product” of 2k vector variables for arbitrary k ≥ 1, respectively.
The respective anti-symmetry and symmetry of the determinant and inner product functions are
replaced by the following symmetries:
• Reversion:
ha1 a2 · · · a2k i = ha2k a2k−1 · · · a1 i,
n(n−1)
[a1 a2 · · · an+2l ] = (−1) 2 [an+2l an+2l−1 · · · a1 ].

• Shift:
ha1 a2 · · · a2k i = ha2k a1 a2 · · · a2k−1 i,
[a1 a2 · · · an+2l ] = (−1)n−1 [an+2l a1 · · · an+2l−1 ].

The null bracket algebra over Rn+1,1 is related to the Clifford bracket algebra over Rn as follows:
Let u, v and the ai be null vectors in Rn+1,1 . Let each ai represent a point in the orthogonal comple-
ment Rn of the Minkowski plane spanned by u, v, and let − a−→
i a j denote the displacement vector from
point ai to point a j . Then

1 −−→ −−→ −−−−−→ −−−→

ha1 a2 · · · a2k i = ha1 a2 a2 a3 · · · a2k−1 a2k a2k a1 i,
2
1  −→ −−→ −−−−−−−−→ −−−−→
[a1 a2 · · · an+2l ] = (−1)n − a1 a2 a2 a3 · · · an+2l−1 an+2l an+2l a1 ].
2
In 2D trigonometry, the Clifford brackets of the conformal model of Euclidean plane have the
From Molecular Distance Geometry to Conformal Geometric Algebra 129

a2

a1

a4

a3

Figure 5.7
Illustration of the notation ∠(a1 a2 a3 , a1 a3 a4 ): circle a1 a2 a3 in the figure has anti-clockwise orien-
tation (from a1 to a2 to a3 ), and the tangent direction of the circle at point a1 points from a1 to a2 .
Similarly, circle a1 a3 a4 in the figure has clockwise orientation, so the tangent direction of the circle
at point a1 points from a4 to a1 . The angle of rotation from the first tangent to the second one is
∠(a1 a2 a3 , a1 a3 a4 ).

following geometric interpretations [35]:

da1 a2 da2 a3 · · · da2l+1 a2l+2 da2l+2 a1
ha1 a2 · · · a2l+2 i = − cos(∠(a1 a2 a3 , a1 a3 a4 )
2
+∠(a1 a4 a5 , a1 a5 a6 ) + · · · + ∠(a1 a2l a2l+1 , a1 a2l+1 a2l+2 ));
da1 a2 da2 a3 · · · da2l+1 a2l+2 da2l+2 a1
[a1 a2 · · · a2l+2 ] = − sin(∠(a1 a2 a3 , a1 a3 a4 )
2
+∠(a1 a4 a5 , a1 a5 a6 ) + · · · + ∠(a1 a2l a2l+1 , a1 a2l+1 a2l+2 )),

where ∠(a1 a2 a3 , a1 a3 a4 ) denotes the angle from the tangent direction of oriented circle a1 a2 a3 at
point a1 to the tangent direction of oriented circle a1 a3 a4 at point a1 , as shown in Figure 5.7. It
should be remarked that the special role played by a1 in these formulas is not fundamental, in that
any point can be made to serve that role by cyclic shift and reversion.

5.5 Geometric Reasoning in Euclidean Conformal

Geometry

Glossary
Monomial proof: Throughout the process of verifying the conclusion equality of a theorem, the
intermediate expression remains a monomial.
Binomial proof: The intermediate expression remains 2-termed during algebraic manipulations.
Robust proof: In theorem proving by an algebraic method, if there is more than one algebraic
manipulation of the same quality available for the intermediate expression at one step (e.g.,
these algebraic manipulations all change the intermediate polynomial from k terms to l terms),
then no matter which algebraic manipulation of the indicated quality is selected at the current
step, it ultimately leads to an algebraic proof of the same size of the intermediate expression as
130 Handbook of Geometric Constraint Systems Principles

6 1

2 3 4
8

Figure 5.8
Miquel’s 4-Circle Theorem, with the points of intersection labeled as in the text.

others (e.g., the intermediate polynomial is always within the same number of terms for each
selection made).
Quantitative extension of geometric theorem: An algebraic description of a geometric theorem
is usually the following: given a set of equality constraints f1 = 0, . . . , fk = 0 and inequality
constraints g1 6= 0, . . . , gl 6= 0 as the hypothesis, the conclusion c = 0 is true. If it is derived
that when some equality constraints in the hypothesis are removed, say f1 = 0, . . . , fs = 0 are
missing where s < k, the conclusion expression and the missing constraints satisfy

tc = f1 h1 + · · · + fs hs (5.43)

for some geometrically meaningful expressions t, h1 , . . . , hk , then an explicit form of (5.43) is

called a quantitative extension of the geometric theorem [32], as it describes the quantitative
relationship between the conclusion expression and the missing constraints, and provides a
stronger and more general result than the original theorem.

Conformal geometric algebra and null bracket algebra together provide a powerful tool for geo-
metric reasoning in Euclidean conformal geometry. Most theorems in classical Euclidean geometry
involving lines, circles, planes, spheres are given robust monomial/binomial proofs. When one or
more equality constraints are removed from the hypothesis of a theorem, usually a quantitative
extension of the theorem can be found. We illustrate this by several examples.
Example 1. (Miquel’s 4-Circle Theorem [61], Figure 5.8) Four circles intersect at eight points cycli-
cally. If 1, 2, 3, 4 are con-cyclic, so are 5, 6, 7, 8.
There are five concyclicity constraints in the hypothesis, and the conclusion is the sixth concyclicity.
For symmetry, we remove the concyclicity of 1, 2, 3, 4, and see how the conclusion varies. The new
geometric configuration is constructed as follows:
Free points: 1, 2, 3, 4, 5, 7.
Intersections: 6 = 15 ∩2 37, 8 = 15 ∩4 37.
In conformal geometric algebra, their representations are, respectively,

6 = 2−1 {(1 ∨ 5) ∧2 (3 ∨ 7)}2{(1 ∨ 5) ∧2 (3 ∨ 7)},

8 = 2−1 {(1 ∨ 5) ∧4 (3 ∨ 7)}4{(1 ∨ 5) ∧4 (3 ∨ 7)}.
From Molecular Distance Geometry to Conformal Geometric Algebra 131
1
3’ 2’

4
2 3
1’

Figure 5.9
Miquel’s 3-Circle Theorem, with the points of intersection labeled as in the text.

Conclusion expression: [5678]. It equals zero if and only if 5, 6, 7, 8 are con-cyclic.

We compute the expression by eliminating 6, 8 and expanding the result. The following is a typical
monomial proof/computing:
6, 8
[5678] = 2−2 [5{(1 ∨ 5) ∧2 (3 ∨ 7)}2{(1 ∨ 5) ∧2 (3 ∨ 7)}
7{(1 ∨ 5) ∧4 (3 ∨ 7)}4{(1 ∨ 5) ∧4 (3 ∨ 7)}]
expand
= − 2−2 [1257][1457][2357][3457][51237341] (5.44)
monomial
= (1 · 5)(3 · 7)[1257][1457][2357][3457][1234].

In the second step, the Cayley expansion [31] is to eliminate the reduced meet product; in the
third step, the monomial factorizations are by the identity aba = 2(a · b)a for null vector a, where
aba = 151 and 373 in the rightmost factor.
The removed concyclicity expression [1234] occurs naturally in the final result, so Miquel’s
4-Circle Theorem is automatically discovered. In addition, the above computing together with sim-
ilar computing of (5 · 6)(7 · 8) yield the following homogeneous identity that is invariant under the
rescaling of any vector variable, which is a quantitative version of the theorem and discloses how
the conclusion depends on the concyclicity constraint of 1, 2, 3, 4:
[5678] [1234] [1257][3457]
= .
(5 · 6)(7 · 8) (1 · 2)(3 · 4) [1457][2357]
Example 2. (Miquel’s 3-Circle Theorem [5], Figure 5.9) Let 10 , 20 , 30 be points on lines 23, 13, 12
respectively. Then circles 120 30 , 10 230 , 10 20 3 meet at a common point 4.
Construct point 4 as the second intersection of circles 120 30 and 10 230 . The hypothesis contains
three collinearity constraints and two concyclicity constraints. The conclusion is the concyclicity of
10 , 20 , 3, 4.
We remove all the collinearity constraints to see how the conclusion expression [10 20 34] depends
on them. The following is a simple monomial computing:
[10 20 34] = [10 420 3]
4
= 2−1 [10 {(1 ∨ 20 ) ∧30 (10 ∨ 2)}30 {(1 ∨ 20 ) ∧30 (10 ∨ 2)}20 3] (5.45)
expand
= 2−1 [110 20 30 ][210 20 30 ][10 230 120 3].

In the above computing, after rearranging the input [10 20 34] to [10 420 3] and substituting the ex-
pression of 4 into it, each meet product allows one and only one monomial Cayley expansion. For
example, since the meet product (1 ∨ 20 ) ∧30 (10 ∨ 2) is a neighbor of null vector 10 , the Cayley ex-
pansion by separating 10 , 2 automatically deletes the term of 10 , leaving only a multiple of 2 left in
the expansion result. Finding a good neighbor makes good control of the size of Cayley expansion.
132 Handbook of Geometric Constraint Systems Principles
4

3’
3

1’ 2
2’ o123

Figure 5.10
Generalized Simson’s Theorem, with the points labeled as in the text and the triangle formed by the
constructed points shaded.

If the three collinearities are restored, it is easily seen that generically [110 20 30 ] 6= 0 and
[210 20 30 ]
6= 0. So it is factor [10 230 120 3] = 0 that leads to the original conclusion under the two
non-degeneracy conditions.
The removed collinearities can be reformulated as following: (i) three circles/lines 10 23, 120 3,
1230 concur; (ii) they concur at the conformal point at infinity (so that they are lines, not circles).
By a computing procedure similar to (5.45), we get that constraint (i) can be represented by
[120 310 230 ] = 0 under some additional non-degeneracy conditions, and (ii) is redundant under these
additional conditions. So the equivalence between the conclusion and the three collinearities is the
following shift symmetry of Clifford bracket:

[10 230 120 3] = − [120 310 230 ],

and Miquel’s 3-Circle Theorem is a special case of this symmetry.

Example 3. (Generalized Simson’s Theorem [3], Figure 5.10) In the plane, draw feet 10 , 20 , 30 from
point 4 to the three sides of triangle 123 respectively. Then

d4ø 2
1
S10 20 30 = (1 − 2123 )S123 , (5.46)
4 R123

where S123 , ø123 , and R123 are the signed area, circum-center, and radius of the circum-circle of
triangle 123 respectively. In particular, 1, 2, 3, 4 are con-cyclic if and only if 10 , 20 , 30 are collinear,
which is the original Simson’s Theorem.
The construction of the geometric configuration is as follows:
Free points: 1, 2, 3, 4.
Feet: 10 = ¶4,23 (projection of point 4 on line 23), 20 = ¶4,13 , 30 = ¶4,12 .
Their (reduced) representations in conformal geometric algebra are, respectively,

10 = (2 ∨ 3) ∧e (4 ∨ he23i∼
3 ) mod e,
20 = (3 ∨ 1) ∧e (4 ∨ he31i∼
3 ) mod e, (5.47)
30 = (1 ∨ 2) ∧e (4 ∨ he12i∼
3 ) mod e,

where “mod e” denotes that the two sides of the equality are equal up to an additional term λ e for
some scalar λ . Equation (5.47) is geometrically readable: e ∨ 2 ∨ 3 = he23i3 represents line 23, and
the Hodge dual he23i∼ 3 represents the normal direction of the line, so the intersection of line 23 with
the line passing through point 4 and the normal direction of line 23 is the foot 10 .
Conclusion expression: [e10 20 30 ] (twice the signed area of triangle 10 20 30 ).
 From Molecular Distance Geometry to Conformal Geometric Algebra 133

We compute the conclusion expression by eliminating the three feet and making simplification,
with the hope that [e123] and [1234] occur naturally as factors in the result, according to (5.46). The
following is a typical binomial proof/computing:

10 ,20 ,30
[e10 20 30 ] = [e{(2 ∨ 3) ∧e (4 ∨ he23i∼ ∼ ∼
3 )}{(3 ∨ 1) ∧e (4 ∨ he31i3 )} {(1 ∨ 2) ∧e (4 ∨ he12i3 )}]
expand
= [e123]([e34he31i∼ ∼ ∼ ∼ ∼ ∼
3 ][e124][ehe23i3 4he12i3 ] − [e24he12i3 ][e134][ehe23i3 4he31i3 ])
duality
= [e123](he34he31i3 i[e124][ehe23i3 4he12i3 ] − he24he12i3 i[e134][ehe23i3 4he31i3 ])
ungrading
= 2−3 [e123](he34e31i[e124][e23e4e12] − he24e12i[e134][e23e4e31])
monomial
= (e · 2)(e · 3)(e · 4)[e123]2 (he134i[e124] − he124i[e134])
contract
= 2(e · 1)(e · 2)(e · 3)(e · 4)2 [e123]2 [1234].

The second step is Cayley expansion to remove the reduced meet product. Notice that not all the
meet products are eliminated at the same time, instead they are removed one by one. For exam-
ple, if we expand the first meet product (2 ∨ 3) ∧e (4 ∨ he23i∼ 3 ) by separating 2, 3, the result is a
linear combination (binomial) of the two null vectors 2, 3, and in successive Cayley expansions,
any separation generating one of 2, 3 leads to a monomial result (null vector advantage), so that the
expansion result remains 2-termed.
The third step in the computing is to eliminate the Hodge dual operator; the fourth step is to
remove the 3-grading operator; the fifth step is monomial factorizations; the last step is contraction
to reduce the number of terms.
So factors [e123]2 [1234] come up in the result automatically. The following quantitative result
is also obtained by similar computing of (e · 10 )(e · 20 )(e · 30 ):

[e10 20 30 ] [1234][e123]2
= .
(e · 10 )(e · 20 )(e · 30 ) 4(e · 1)(e · 2)(e · 3)(e · 4)(1 · 2)(2 · 3)(1 · 3)

In more geometrical terms, the above identity can be written as

[1234]
[e10 20 30 ] = ,
2R2123

because R2123 = −2(1 · 2)(1 · 3)(2 · 3)[e123]−2 .

It is a common phenomenon that algebraic computing based on conformal geometric algebra
representations and null bracket algebra manipulations are extremely short, and the results are geo-
metrically meaningful. The simplicity may be partially attributed to the fact that many complicated
syzygies among basic invariants are integrated into simple symmetries within advanced invariants.
The above examples have demonstrated the effectiveness of expansion and factorization in null
bracket algebra. Besides these two symbolic manipulations, there are two other basic ones: nor-
malization and division. The normalization converts a Clifford bracket polynomial into a unique
normal form, and the division among Clifford bracket polynomials is drastically different from the
coordinate polynomial case. Some recent advances can be found in [36, 37].
Acknowledgment. The second author is supported by the NSFC Project 11671388 and the CAS
Frontier Key Project QYZDJ-SSW-SYS022.
134 References

References
[1] E. Bayro-Corrochano and C. López-Franco. Omnidirectional vision: Unified model using
conformal geometry. ECCV’04 pages 536–548, 2004.
[2] E. Bayro-Corrochano, L. Reyes-Lozano, J. Zamora-Esquivel. Conformal geometric algebra
for robotic vision. J. Mathematical Imaging and Vision 24:55–81, 2006.
[3] M. Berger Geometry I, II. Springer Verlag, Berlin, Heidelberg, 1987.
[4] L. M. Blumenthal. Theory and Applications of Distance Geometry. Cambridge Univ. Press,
Cambridge,1953. Reprinted by the Chelsea Publishing Co., 1970.
[5] S. C. Chou, X. S. Gao and J. Z. Zhang. Machine Proofs in Geometry. World Scientific, Sin-
gapore, 1994.
[6] W. K. Clifford. Application of Grassmann’s extensive algebra. Am. J. Math. I:350–358, 1878.
[7] R. Connelly. Rigidity and energy. Invent. Math. 66:11–33, 1982.
[8] G. M. Crippen and T. F. Havel. Distance Geometry and Molecular Conformation. Research
Studies Press, Taunton, UK, 1988.
[9] G. M. Crippen and S. A. Wildman. Quantitative structure-activity relationships (QSAR). In
A. K. Ghose and V. N. Viswanadhan, editors, Combinatorial Library Design and Evaluation,
pages 131–156. Marcel Dekker, New York, NY, 2001.
[10] G. M. Crippen. Cluster distance geometry of polypeptide chains. J. Comput. Chem., 25:1305–
1312, 2004.
[11] D. B. Dix. Polyspherical coordinate systems on orbit spaces with applications to biomolecular
shape. Acta Appl. Math., 90:247–306, 2006.
[12] L. Dorst, D. Fontijne and S. Mann. Geometric Algebra for Computer Science. Morgan Kauf-
mann Publishers, Elsevier Inc., 2007.
[13] A. W. M. Dress and T. F. Havel. Bound smoothing under chirality constraints. SIAM J. Disc.
Math., 4:535–549, 1990.
[14] M. Fiedler. Matrices and Graphs in Geometry. Cambridge University Press, 2011.
[15] D. Fontijne, L. Dorst. Modeling 3D Euclidean geometry–performance and elegance of five
models of 3D Euclidean geometry in a ray tracing application. Computer Graphics Appl.
23:68-78, 2003.
[16] M. Fuhrmans, A. G. Milbradt, and C. Renner. Comparison of protocols for calculation of
peptide structures from experimental NMR data. J. Chem. Theory Comput., 2:201–208, 2006.
[17] T. F. Havel and K. Wüthrich. A distance geometry program for determining the structures of
small proteins and other macromolecules from nuclear magnetic resonance measurements of
1 H − 1 H proximities in solution. Bull. Math. Biol., 46:673–698, 1984.

[18] T. F. Havel. An evaluation of computational strategies for use in the determination of protein
structure from distance constraints obtained by nuclear magnetic resonance. In D. Nobel and
T. L. Blundell, editors, volume 56 of Progress in Biophysics and Molecular Biology, pages
43–78. Permagon Press, Oxford, UK, 1991.
[19] T. F. Havel. Predicting the structure of the flavodoxin from Erchericia coli by homology
modeling, distance geometry and molecular dynamics. Molec. Simul., 10:175–210, 1993.
References 135

[20] T. F. Havel. Computational synthetic geometry with Clifford algebra. In: Automated Deduc-
tion in Geometry, D. Wang (ed.), Lect. Notes in Artif. Intellig. 1360, pp. 102–114, Springer-
Verlag, 1997.
[21] T. F. Havel. Metric matrix embedding in protein structure calculations, NMR spectra analysis,
and relaxation theory. Magn. Reson. Chem., 41:S37–S50, 2003.
[22] T. F. Havel, I. D. Kuntz, and G. M. Crippen. Theory and practice of distance geometry. Bull.
Math. Biol., 45:665–720, 1983.
[23] T. F. Havel and I. Najfeld. A new system of equations, based on geometric algebra, for ring
closure in cyclic molecules. In J. Fleischer, J. Grabmeier, F. W. Hehl, and W. Küchlin, editors,
Computer Algebra in Science and Engineering, pages 243–259. World Scientific, Singapore,
1995.
[24] D. Hestenes. Space-Time Algebra. Gordon and Breach, New York, 1966.
[25] D. Hestenes and G. Sobczyk. Clifford Algebra to Geometric Calculus, Kluwer Acad., Dor-
drecht NL, 1984.
[26] D. Hildenbrand. Geometric computing in computer graphics using conformal geometric alge-
bra. Computers and Graphics 29:795–803, 2005.
[27] A. Lasenby and J. Lasenby. Surface evolution and representation using geometric algebra. In
R. Cipolla and R. Martin, editors, The Mathematics of Surfaces IX, pages 144–168. Springer
Verlag, London, 2000.
[28] A. Lasenby and C. Doran. Geometric Algebra for Physicists. Cambridge University Press,
Cambridge, 2003.
[29] Li, H. Some applications of Clifford algebra to geometries. In X. S. Gao et al., editors, LNAI
1669, Automated Deduction in Geometry, pages 156–179. Springer Verlag, Heidelberg, 1998.
[30] H. Li, D. Hestenes and A. Rockwood. Generalized homogeneous coordinates for computa-
tional geometry. In G. Sommer, editor, Geometric Computing with Clifford Algebras, pages
27–60. Springer Verlag, Heidelberg, 2001.
[31] H. Li and Y. Wu. Automated short proof generation in projective geometry with Cayley and
bracket algebras I. Incidence geometry. J. of Symbolic Comput. 36:717–762, 2003.
[32] H. Li. Symbolic computation in the homogeneous geometric model with Clifford algebra. In
J. Gutierrez, editor, Proc. ISSAC 2004, pages 221–228. ACM Press, New York NY, 2004.
[33] H. Li. A recipe for symbolic geometric computing: Long geometric product, BREEFS and
Clifford factorization. In C. W. Brown, editor, Proc. ISSAC 2007, pages 261–268. ACM Press,
New York, NY, 2007.
[34] H. Li. Invariant Algebras and Geometric Reasoning. World Scientific, Singapore 2008.
[35] H. Li and L. Huang. Complex brackets, balanced complex differences, and applications in
symbolic geometric computing. In Proc. ISSAC 2008, pages 181–188. ACM Press, New York,
NY, 2008.
[36] H. Li. Normalization of polynomials in algebraic invariants of three-dimensional orthogonal
geometry. arXiv: 1302.7194v1 [cs.SC], 2013.
[37] H. Li, C. Shao, L. Huang and Y. Liu. Reduction among bracket polynomials. In Proc. IS-
SAC’14, pages 304–311. ACM Press, New York NY, 2014.
[38] L. Liberti, C. Lavor, N. Maculan and A. Mucherino. Euclidean distance geometry and appli-
cations. SIAM Rev., 56:3–69, 2014.
136 References

[39] K. Menger. Untersuchungen über allgemeine Metrik. Math. Annal., 100:75–163, 1928.
[40] K. Menger. New foundation of Euclidean geometry. Amer. J. Math., 53:721–745, 1931.
[41] D. S. Mitrinović, J. E. Pečarić and V. Volenec. Recent Advances in Geometric Inequalities.
Springer Verlag (originally Kluwer Academic), Dordrecht, NL, 1989.
[42] B. Mourrain and N. Stolfi. Computational symbolic geometry. In: N. L. White, editor, Invari-
ant Methods in Discrete and Computational Geometry, pages 107–139. D. Reidel, Dordrecht,
NL, 1995.
[43] A. Mucherino, C. Lavor, L. Liberti and N. Maculan, editors. Distance Geometry – Theory,
Methods and Applications. Springer Verlag, New York, NY, 2013.
[44] E. Müller. Die Kugelgeometrie nach den Principien der Grassmann’schen Ausdehnungslehre
(Teil I). Monatshefte Math. Phys. 3: 365–402, 1892.
[45] E. Müller. Die Kugelgeometrie nach den Principien der Grassmann’schen Ausdehnungslehre
(Teil II). Monatshefte Math. Phys. 4: 1–52, 1893.
[46] I. Najfeld and T. F. Havel. Embedding with a rigid substructure. J. Math. Chem., 21:223–260,
1997.
[47] P. J. Olver. Classical Invariant Theory. Cambridge University Press, Cambridge, 1999.
[48] J. H. Prestegard, C. M. Bougault and A. I. Kishore. Residual dipolar couplings in structure
determination of biomolecules. Chem. Rev., 104:3519–3540, 2004.
[49] M. Riesz. Clifford Numbers and Spinors, 1958. From lecture notes made in 1957-8, edited by
E. Bolinder and P. Lounesto. Kluwer Academic, Dordrecht NL, 1993.
[50] B. Rosenhahn and G. Sommer. Pose estimation in conformal geometric algebra I, II. J. Math-
ematical Imaging and Vision 22:27–48, 22:49–70, 2005.
[51] Aleix Rull, Josep M. Porta and Federico Thomas. Distance bound smoothing under orienta-
tion constraints. IEEE Intnl. Conf. Robotics & Automation, pages 1431–1436, 2015.
[52] Tamar Schlick. Molecular Modeling and Simulation, 2nd ed., Springer Media, New York,
NY, 2010.
[53] I. J. Schoenberg. On certain metric spaces arising from euclidean spaces by a change of metric
and their imbeddings in Hilbert space. Annals Math., 38:787–793, 1937.
[54] J. J. Seidel. Distance-geometric development of two-dimensional euclidean, hyperbolic and
spherical geometry. Simon Stevin, 29:32–50, 65–76, 1952.
[55] E. Snapper and R. J. Troyer. Metric Affine Geometry. Academic Press, 1971. Reprinted by
Dover Publications, 1989.
[56] G. Sommer. A geometric algebra approach to some problems of robot vision. In J. Byrnes, ed-
itor, NATO Science Series 136, Computational Non-commutative Algebra and Applications,
pages 309–338. Kluwer Acad., Dordrech,t NL, 2004.
[57] H. Takashima. High-resolution protein structure determination by NMR. Ann. Rep. NMR
Spect., 59:235–273, 2006.
[58] Walter Thiel and Genard Hummer. Nobel 2013 chemistry: Methods for computational chem-
istry. Nature, 504:96–97, 2013.
[59] G. Wagner, S. Hyberts, and T. F. Havel. NMR structure determination in solution: A critique
and comparison with X-ray crystallography. Ann. Rev. Biophys. Biomol. Struct., 21:167–198,
1992.
References 137

[60] N. White. Geometric applications of the Grassmann-Cayley algebra. In J. E. Goodman and J.

O’Rourke, editors, Handbook of Discrete and Computational Geometry. CRC Press, Chap-
man & Hall, Boca Raton, FL, 1997.
[61] W. T. Wu. Basic Principles of Mechanical Theorem Proving in Geometries I: Part of Elemen-
tary Geometries. Science Press, Beijing, 1984; Springer Verlag, Vienna AT, 1994.
[62] K. Wüthrich. NMR structures of biological macromolecules. Magn. Reson. Chem., 41:S89–
S98, 2003.
[63] H. Weyl. The Classical Groups. Princeton Univ. Press, Princeton, NJ, 1939.
Chapter 6
Tree-Decomposable and
Underconstrained Geometric Constraint
Problems
Ioannis Fudos
Department of Computer Science and Engineering, University of Ioannina, Ioannina, Greece

Christoph M. Hoffmann
Department of Computer Science, Purdue University, West Lafayette, IN

Robert Joan-Arinyo
Department of Computer Science, Universitat Politècnica de Catalunya, Barcelona, Catalonia

CONTENTS
6.1 Introduction, Concepts, and Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.1.1 Geometric Constraint Systems (GCS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.1.2 Constraint Graph, Deficit, and Generic Solvability . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.1.3 Instance Solvability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.1.4 Root Identification and Valid Parameter Ranges . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.1.5 Variational and Serializable Constraint Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.1.6 Triangle-Decomposing Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.1.7 Scope and Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.2 Geometric Constraint Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.3 Constraint Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.3.1 Geometric Elements and Degrees of Freedom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.3.2 Geometric Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.3.3 Compound Geometric Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.3.4 Serializable Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.3.5 Variational Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.3.6 Triangle Decomposability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.3.7 Generic Solvability and the Church-Rosser Property . . . . . . . . . . . . . . . . . . . . . . . 155
6.3.8 2D and 3D Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.4 Solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
6.4.1 2D Triangle-Decomposable Constraint Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 159
Operation 1: Minimal GCS Placement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
Operation 2: Constructing one Element from two Constraints . . . . . . . . . . . . . . . . . . . . . . . 159
Operation 3: Matching two Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6.4.2 Root Identification and Order Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Moving Selected Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

139
140 Handbook of Geometric Constraint Systems Principles

Adding Extra Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

Order-Based Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
Dialog with the Solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
Design Paradigm Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
6.4.3 Extended Geometric Vocabulary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
Variable-Radius Circles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
6.5 Spatial Geometric Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
Points and Planes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
Points, Lines, and Planes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
6.6 Under-Constrained Geometric Constraint Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176

6.1 Introduction, Concepts, and Scope

Geometric constraint system (GCS): Also known as a geometric constraint problem, is a pair
P = (U, F), where U is a set of geometric objects (or geometric elements) on which we impose
a set F of geometric constraints.
Constraint graph: Given a GCS P = (U, F), its Constraint Graph G = (V, E) is a labeled undi-
rected graph whose vertices V are the geometric objects in U, each labeled with its degrees of
freedom. An edge e ∈ E represents a binary constraint between two geometric objects and is
labeled by the number of equations formulating the geometric constraint.
Deficit: Deficit κ of a constraint graph is defined as the sum of the labels of its vertices minus the
sum of the labels of its edges. The deficit determines the degrees of freedom of the compound
geometric object defined by the corresponding GCS.
Quasi generically overconstrained: A GCS problem P is quasi generically over-constrained if
there is an induced subgraphPof the associated
P constraint graph, such that for the induced sub-
graph the following holds: l(v) − l(e) < κ and none of the constraints fixes the geometric
structure with respect to the global coordinate system, where κ = 3 in the plane and κ = 6 in
3-space.
Quasi generically underconstrained: A GCS problem P in thePEuclidean P plane/space is quasi
generically under-constrained if it is not overconstrained and l(v) − l(e) > κ.
Instance solvability: A GCS is well-constrained if the associated (algebraic) equation system has
one or more discrete real solutions. It is over-constrained if it has no solution, and is under-
constrained if it has a continuum of solutions.

A geometric constraint problem (we refer the reader to Chapter 1), also known as a geometric
constraint system, consists of a finite set of geometric objects, such as points, lines, circles, planes,
spheres, etc., and constraints upon them, such as incidence, distance, tangency, and so on. A solution
of a geometric constraint problem P is a coordinate assignment for each of the geometric objects of
P that places them in relation to each other such that all constraints of P are satisfied. A problem P
may have a unique solution, it may have more than one solution, or it may have no solution.
In this chapter, we are concerned with geometric constraint solvers, i.e., with programs that
find one or more solutions of a geometric constraint problem. If no solution exists, the solver is
expected to announce that no solution has been found. Owing to the complexity, type or difficulty
of a constraint problem, it is possible that the solver does not find a solution even though one may
Tree-Decomposable and Underconstrained Geometric Constraint Problems 141

exist. Thus, there may be false negatives, but there should never be false positives. Intuitively, the
ability to find solutions can be considered a measure of solver’s competence.
We consider static constraint problems and their solvers. We do not consider dynamic con-
straint solvers, also known as dynamic geometry programs, in which specific geometric elements
are moved, interactively or along prescribed trajectories, while continually maintaining all stipu-
lated constraints. However, if we have a solver for static constraint problems that is sufficiently fast
and competent, we can build a dynamic geometry program from it by solving the static problem for
a sufficiently dense sampling of the trajectory of the moving element(s).
The work we survey has its roots in applications, especially in mechanical computer-aided de-
sign (MCAD). The constraint solvers used in MCAD took a quantum leap with the work by Owen
[46]. Owen’s algorithm solves a geometric constraint problem by an initial, graph-based structural
analysis that extracts generic subproblems and determines how they would combine to form a com-
plete solution. These subproblems are then handed to an algebraic solver that solves the specific
instances of the generic subproblems and combines them. Owen’s graph analysis is top down. A
bottom-up analysis was proposed in [3]. Subsequent work expanded the knowledge of
(a) The structure and properties of the constraint graph, see Section 6.3;
(b) The geometric vocabulary, see Section 6.4;

(c) The understanding of spatial constraint systems, see Section 6.5.

We also look briefly at under-constrained problems, see Section 6.6.
Restricted to points and distances, the constraint graph analysis has deep roots in mathematics
and combinatorics; see, e.g., [17, 38, 43] for some of these connections. Here, we limit the dis-
cussion to constraint problems that are motivated by MCAD applications, and that means that the
geometric vocabulary has to be richer than points and distances between them. In the remainder of
this section we introduce informally major concepts and methods for solving geometric constraint
systems (GCS) with an application perspective in mind.
Note in particular that some definitions in this section may be provisional. Those definitions,
although formally incorrect, are so given nevertheless because they facilitate understanding the ma-
terial. They are clearly identified along with examples that show where they fall short — and what
can be done about it.

6.1.1 Geometric Constraint Systems (GCS)

Fix the space in which to consider a geometric constraint system (GCS). We again refer the reader
to Chapter 1. Typical examples are the Euclidean space in 2 or 3 dimensions. Each object to be
placed by a GCS instance in that space has a specific number of degrees of freedom (dof), i.e., a
specific number of independent coordinates. In Euclidean 2-space, points and lines each have 2 dof.
In Euclidean 3-space, a point has 3 dof and a line has 4. A constraint on such objects corresponds
to one or more equations expressing the constraints on the coordinates. So, requiring a distance d
between two points A = (Ax , Ay ) and B = (Bx , By ) in 2-space would be expressed by

(Ax − Bx )2 + (Ay − By )2 = d 2

Requiring that the two points are coincident would entail two equations:

Ax = Bx
Ay = By

Accordingly, a GCS can be viewed simply as a system of equations: A solution of the GCS, if
one exists, is a valuation of the variables, the coordinates, that satisfies all equations. Viewed in this
142 Handbook of Geometric Constraint Systems Principles

(a) (b)

Figure 6.1
Constraint problem and associated constraint graph.

foundational way, solving a GCS boils down to formulating a system of equations in the coordinates
of the geometric entities and solving the system by any means appropriate. The equations are almost
always algebraic.

6.1.2 Constraint Graph, Deficit, and Generic Solvability

The approach of treating a GCS as an (unstructured) system of equations is inefficient and almost
always unnecessary. Given a GCS, we analyze the structure of the corresponding equation system,
and seek to identify smaller subsystems that can be solved independently and that admit especially
efficient solution algorithms. Triangle decomposable systems, the main subject of this chapter, do
this by analyzing a constraint graph that mirrors the equation structure. The constraint graph is re-
cursively broken down into subgraphs that correspond to independently solvable subsystems. Once
subsystems have been solved, their solutions are combined and expose, in the aggregate, additional
subsystems that now can be solved separately. Given the GCS problem P = (U, F), with the set
of geometric objects U and constraints F upon them, we define the constraint graph G(P) as fol-
lows.We again refer the reader to Chapter 1.

Definition 6.1 Given the GCS P = (U, F), its constraint graph G = (V, E) is a labeled undirected
graph whose vertices are the geometric objects in U, each labeled with its degrees of freedom. There
is an edge (u, v) in E if there is a constraint between the geometric objects u0 and v0 , corresponding
to u and v, respectively. The edge is labeled by the number of independent equations corresponding
to the constraint between u0 and v0 .

Example 11 Consider the constraint problem of Figure 6.1 that comprises four points in the plane,
labeled A through D. A line between two points indicates a distance constraint on the points. Thus
there are 5 distance constraints, shown left. The corresponding constraint graph is shown on the
right.

Since a GCS naturally corresponds to a system of equations, we expect that the number of in-
dependent equations should equal the number of variables. The number of independent variables
equals the sum of degrees of freedom, i.e., the sum of vertex labels of the constraint graph. More-
over, the number of independent equations equals the sum of the edge labels, in most situations.
Thus, we investigate ho the structure of the constraint graph reflects the structure of the equation
system that corresponds to the GCS.
As stated, the GCS of Example 11 does not prescribe where on the plane to position and orient
the points after solving the GCS. Thus, while this GCS is quasi well-constrained, the sum of vertex
labels equals the sum of edge labels plus 3. Here, the deficit of 3 corresponds to 3 missing equations
Tree-Decomposable and Underconstrained Geometric Constraint Problems 143

v2 v4 v6

v1 v3 v5

Figure 6.2
A GCS with an overconstrained and an underconstrained subgraph.

that fix where, in the plane, to place the solution. If necessary, the remaining degrees of freedom
can be determined by placing the points with respect to a global coordinate system, for example by
adding three equations that place A at the origin and B on the (positive) x-axis.
So, if the position
P and orientation
P of the solution is undetermined, we are assured that a con-
straint graph where l(v) − l(e) < 3 in the plane corresponds to an equation system in which at
least one equation is not independent. Consequently, we conclude

Definition 6.2
A GCS problem P is quasi generically over-constrained if there is an induced subgraph
P of
Pthe asso-
ciated constraint graph, such that for the induced subgraph the following holds: l(v)− l(e) < κ
and none of the constraints fixes the geometric structure with respect to the global coordinate sys-
tem, where κ = 3 in the plane and κ = 6 in 3-space.

Extrapolating this line of reasoning, we might be led to the conclusion that we can use this
structural graph property to define

Definition 6.3
A GCS problem P in the Euclidean
P Pplane/space is generically quasi under-constrained if it is not
quasi overconstrained and l(v) − l(e) > κ.

Definition 6.4 (Provisional)

A GCS problem P in the Euclidean plane/space is quasi generically well-constrained if: (i) it is not
quasi over-constrained and (ii) the sum of vertex labels of the associated constraint graph equals the
sum of the edge labels plus the deficit κ.

Note, however, for a graph to be well-constrained Definition 6.4 is not sufficient and will be
refined in Section 6.3.7. The problem with Definition 6.4 is best illustrated by an example. The
following example exhibits a constraint graph for which the sum of the labels of the vertices equals
the sum of the labels of the edges plus κ but it contains an overconstrained subgraph and therefore
the GCS cannot be well-constrained.

Example 12 Consider the constraint graph of Figure 6.2. For specificity, let the graph vertices
be points and the edges distances. The deficit is 3, so it seems that a problem with this constraint
graph is quasi well-constrained. Now the subgraph induced by vertices v1 , ..., v4 is overconstrained;
drop any of the subgraph edges and you obtain a well-constrained subproblem with a deficit 3.
On the other hand, the subgraph induced by the vertices v3 ...v6 has a deficit of 4 and is in fact
underconstrained; vertices v5 and v6 cannot be placed. There is an extra constraint in the first
subproblem that is not available for the second subproblem.
144 Handbook of Geometric Constraint Systems Principles

Simple dof counting fails to identify all well-constrained graphs in 2D. Likewise, a GCS in 3D
with deficit 6 that is not overconstrained does not always have a well-constrained graph. Chapter 1
partially resolves the abovementioned issues using a notion of minimal dof rigidity, which captures
generic well constrained 2D distance constrained systems, as we will see in Section 6.3.7. Neverthe-
less, as pointed out in Chapter 1, in general, even this does not capture generically well constrained
GCS.
Note that the definitions and properties explained assume in particular that the GCS solution
does not have certain symmetries. For instance, consider placing concentrically two circles of given
radii. The constraint graph has two vertices, labeled 2 each, and
Pone edge Plabeled 2, because concen-
tricity implies that the two centers coincide. Here the deficit l(v) − l(e) is 2, less than κ = 3.
The system appears to be quasi over-constrained, but the constraint system is well-constrained. The
lower deficit reflects the rotational symmetry of the solution.
In the following, we exclude GCS that fix the structure with respect to the coordinate system, as
well as GCS that exhibit symmetries that reduce κ.

6.1.3 Instance Solvability

Consider constructing a triangle 4(A, B,C) with the three sides a, b, c through each of the vertex
pairs, in the Euclidean plane. The three points and three lines together have 12 degrees of freedom.
Each vertex is incident to two lines, so there are six incidence constraints, each contributing one
equation. We add three angle constraints, one for each line pair. The resulting constraint graph has
a deficit of three and thus appears to be quasi generically well-constrained. However, the problem
is not well-constrained: if the three angle constraints add up to π, then the GCS is actually under-
constrained. If they add up to π/4, say, the GCS has no solution.
The problem arises from the interdependence of the three angle constraints. Thus generic solv-
ability does not guarantee that problem instances are solvable. In particular, the constraint graph
does not record specific dependencies among the equations. Such dependencies often arise from
geometry theorems.

Definition 6.5 A GCS is well-constrained if the associated (algebraic) equation system has one
or more discrete real solutions. It is over-constrained if it has no solution, and is under-constrained
if it has a continuum of solutions.

The definition of generic solvability is fair in the sense that dependencies among some of the
constraints, and their equational form, can arise from geometry theorems. The example above is
based on a simple theorem, but more complex theorems can arise and may be as hard to detect as
solving the equations in the first place.

6.1.4 Root Identification and Valid Parameter Ranges

Consider Example 11. With the distance constraints as drawn, the GCS has multiple solutions, some
shown in Figure 6.3. From the equational perspective none is distinguished. From an application
perspective usually one is intended. Since the number of distinct solutions can grow exponentially
with the number of constrained geometric objects, it is not reasonable to determine all solutions
and let the application choose. This application-specific problem is the root identification problem.
Some authors use the term chirality.
Solvability generalizes to the determination of valid parameter ranges: It is clear that the dis-
tances between the points A, B,C, and the distances between the points B,C, D must satisfy the
triangle inequality for there to be a (nondegenerate) solution. Considering the metric constraints
(distance, angle) as coordinates of points in a configuration space, what is the manifold of points
Tree-Decomposable and Underconstrained Geometric Constraint Problems 145

D
A B

(a) (b) (c)

Figure 6.3
Different solutions of the constraint problem of Example 11.

Figure 6.4
A variational constraint problem.

that are associated with solvable GCS instances? Is the manifold connected or are there solutions
that cannot be reached from every starting configuration?

6.1.5 Variational and Serializable Constraint Problems

Some GCS have equation systems that are naturally triangular. That is, intuitively, the geometric
objects can be ordered such that they can be placed one-by-one. Such GCS are serializable or se-
quential constraint problems. GCS that are not serializable have been called variational constraint
problems in some of the application-oriented literature. The examples discussed so far are all se-
quential. The following example shows a variational problem.

Example 13 Place 9 points in the plane subject to the 15 distance constraints indicated by the
lines, in Figure 6.4. The following three groups of 4 points each can be placed with respect to each
other: (A, B,C, D), (D, E, F, G), and (G, H, I, A). Each of these subproblems is sequential in nature.
However, the overall placement problem of all points is not serializable; it is variational.

Serializable constraint problems are of limited expressiveness. Due to their potential efficiency,
however, they are used to great effect in some dynamic geometry packages such as Cinderella [6]
and GeoGebra [15].

6.1.6 Triangle-Decomposing Solvers

The overall strategy of triangle-decomposing solvers is to construct the constraint graph and, an-
alyzing the graph, to recursively isolate solvable subproblems. Then, the solved subproblems are
recombined. The process is recursive.
146 Handbook of Geometric Constraint Systems Principles

For planar constraint systems, triangle-decomposability is understood to be a recursive property

in which the base case consists of two vertices (2 dof each) and one edge between them eliminating
one of the four dof’s. Larger structures are then solved by isolating three solved structures that pair-
wise share a single geometric element and thus can be combined by placing the solved structures
relative to each other. When this decomposition is done top-down, the solved structures are tricon-
nected; [46]. When the decomposition is done bottom-up, triples of solved components are sought
that pairwise share a geometric primitive; [3]. The act of combining such triples has been called a
cluster merge.
For the problem of Example 13, illustrated in Figure 6.4, the graph is first decomposed into the
three subproblems discussed. Next, each of these subproblems is further broken down. For the sub-
problem A, B,C, D, for example, this might be done by placing, with respect to each other, A and B, B
and C, and C and A. Those three subsystems are then combined into the triangle A, B,C. This merg-
ing step combines three subsystems that pairwise share a point, placing the three shared geometric
elements in one geometric construct. The triangle can then be merged with the two subsystems C
and D, and B and D. The result so obtained is the solved subsystem with vertices A, B,C, D. Sim-
ilarly, we solve subproblems D, E, F, G and G, H, I, A, each in two merge steps. Finally, the entire
problem is solved by merging the three subsystem.
The recursion solving the problem of Figure 6.4 can be thought of either as a top-down decom-
position, or as a bottom-up reconstruction.
Note that the top-level merge step places the shared objects, here the points A, D and G. The
distances between these points are not given but they can be obtained from the three subproblem
solutions. We can think of this process as a plan that is formulated based on the constraint graph
analysis. Two questions arise:
(a) Since the plan so formulated is not unique, do different decompositions arrive at different
solutions? This question is settled by investigating the nature of the recursive decomposition
and whether it satisfies a Church-Rosser property.
(b) What specific subsystems of equations must be solved and how? This question is approached
by analyzing the basic subgraph configurations that can occur and are allowed by the geomet-
ric constraint solver, both when decomposing and when recombining.
Concerning the second question, in our example two key operations are needed: placing two points
at a prescribed distance, and placing a third point at a required distance from two fixed points. There
is also a third operation that places a rigid geometric configuration X such that two points of X
match two given points, using translation and rotation. This operation is used when recombining the
three subproblems.
Carrying out these operations entails solving equation systems of a fixed structure. Often, these
equation systems are small and can be solved very efficiently. The simplest cases restrict the geo-
metric repertoire to points, lines, and circles of fixed radius. Triangle decomposable systems then
require at most solving univariate quadratic polynomials.
More complicated primitives can be added. They include circles of a-priori unknown radius, e.g.,
[37]; conic sections, e.g., [9]; and Bézier curves, e.g., [20]. Additions impact both the constraint
graph analysis as well as the richness of algebraic equation systems that have to be solved. The
impact on the graph analysis can be limited to adding more patterns to cluster merging. So extended,
triangle decomposition becomes a more general analysis that can be called tree decomposition. The
decomposition remains a tree, but the tree may have different numbers of subtrees associated with
the growing repertoire of subgraph patterns. The number of possible patterns is infinite. Therefore,
this approach becomes self-limiting. More than that, rigidity no longer has a simple characterization.
The contributions to the equation solving algorithms is also considerable. Section 6.4.3 catalogs
what is known about the equation systems required to add just circles of unknown radius.
Tree-Decomposable and Underconstrained Geometric Constraint Problems 147

Because of these rigors, more general techniques have to be considered. For the extended graph
analysis, the DR-type algorithms solve the problem by dynamically identifying generically solv-
able subgraphs; [22, 23]. Likewise, the growth of irreducible algebraic equation systems increas-
ingly motivates searching general equation solvers, including numerical solving techniques, such as
Newton iteration, homotopy continuation, and other procedural techniques, for instance [50].

6.1.7 Scope and Organization

We begin with the graph analysis of triangle-decomposable constraint systems in 2D. These systems
play an important role in linkage analysis and graph rigidity, [53, 54, 55]. But even in the Euclidean
plane, applications of geometric constraint systems argue for more expressive systems. We can
increase expressiveness by enlarging the repertoire of geometric objects, as well as by admitting
more complex cluster merging operations.
In the plane, an extended repertoire includes foremost variable-radius circles; that is, circles
whose radius is not given explicitly but must be inferred based on the constraints. More advanced
objects, such as conics and certain parametric curves, have also been considered. Those additional
geometries impact the subsystems of equations that must be solved.
More complex cluster merging operations affect both the constraint graph structures encountered
as well as the equation systems that must be solved to position the geometric structures accordingly.
We will discuss some examples.
After discussing constraint graphs and quasi generic over-, well-, and under-constrained graphs,
we consider the equations that must be solved so as to obtain specific solutions. Here, we explain
the structure of those systems as well as some tools to transform the equation systems into simpler
ones. Questions of root identification, valid parameter ranges, and order type of solutions are to be
discussed. Much is known about these topics in the Euclidean plane. For Euclidean 3-space much
less is known. There the equation systems are in general much harder, and the number of cases
that should be considered is larger. Even simple sequential problems can require daunting equation
systems. A case in point is to find a line in 3-space that is at prescribed distances from four given
points, discussed in Section 6.5.

6.2 Geometric Constraint Systems

Solution of a GCS: A solution of a geometric constraint system (U, F) is an assignment of coor-
dinates instantiating the elements of U such that the constraints F are all satisfied.

Definition 6.6 A geometric constraint system (GCS) is a pair (U, F) consisting of a finite set U
of geometric elements in an ambient space and a finite set F of constraints upon them.

The ambient space typically is n-dimensional Euclidean space. The majority of applications
require n = 2 or n = 3; e.g., [3, 31, 46]. Spherical geometry may also be considered, for instance in
nautical applications.
The imposed constraints typically are binary relations. We do not consider higher-order con-
straints, such as “C be the midpoint between points A and B.” Note, however, that such constraints
can often be expressed by several binary constraints. This can be done in a variety of ways, with or
without variable-radius circles.
148 Handbook of Geometric Constraint Systems Principles

Definition 6.7 A solution of a geometric constraint system (U, F) is an assignment of coordinates

instantiating the elements of U such that the constraints F are all satisfied.

There may be several solutions [10]. Moreover, solutions may or may not be required to be in
prescribed position and orientation, in a global coordinate system.
As defined, a GCS is a static problem in that solutions fix the geometric elements with respect
to each other. The dynamic geometry problem asks to maintain constraints as some elements move
with respect to each other. We consider only static constraint problems and their solvers.
By geometric coverage we understand the diversity of geometric elements admissible in U.
Points, lines and circles of given radius are adequate for many applications in Euclidean 2-space
[46]. For GCS in Euclidean 3-space, an analogous geometric coverage could be points, lines, planes,
as well as spheres and cylinders of fixed radii. Here, the number of solutions even of simple GCS
can be very large [13].

6.3 Constraint Graph

Rigid body: A rigid body is a set of geometric objects whose relative placement is fixed. A rigid
body in 2D has in general 3 degrees of freedom while in 3D a rigid body has in general 6
degrees of freedom.

Rigidity: A GCS is rigid or has rigidity in 2D or 3D if the resulting set of geometric elements after
placement with respect to the constraints is rigid. In graph theory a graph is rigid if the structure
formed by replacing the edges by rigid rods and the vertices by flexible hinges is rigid.(links to
infinitesimal rigidity and structural regidity).
Serializable GCS: A GCS is serializable if the geometric objects can be ordered in such a way that
they can be placed sequentially one-by-one as function of preceding, already placed elements.
Variational GCS: A solvable GCS that is not serializable is variational.
Triangle decomposable GCS: A GCS is triangle decomposable if it can be solved by using as
construction operations the placement of two or three geometric elements with pairwise con-
straints.
Degrees of freedom: Degrees of freedom (dof) are the number of independent, one-dimensional
variables by which a geometric object can be instantiated and positioned. Geometric constraints
consume one or more dof expressed by a number of independent equations.
Compound (complex) geometric objects (elements): A cluster of geometric objects constitutes a
compound geometric object (element). The definition of a compound geometric object may
also include a set of geometric constraints imposed on the geometric objects. Therefore a rigid
body is a compound geometric object with deficit 3 in 2D or 6 in 3D.
Quadratically solvable GCS: A quadratically solvable GCS is a system of equations whose zeros
can be obtained successively solving a sequence of quadratic equations where the coefficients
in an equation may depend on the solution of preceding equations in the sequence.
Laman graph and Laman’s Theorem: Laman’s theorem and Laman graph provide a sufficient
and necessary condition for characterizing a well constrained (rigid) GCS in 2D that consists
of points and distances. A constraint graph that satisfies this condition is called Laman graph.
Tree-Decomposable and Underconstrained Geometric Constraint Problems 149

Table 6.1
Minimal GCS in the Euclidean plane; d(...) denotes distance, a(...) denotes angle.

Geoms Constraint Notes

Points A, B d(A, B) Distance not zero
Point A, line m d(m, A) Zero distance allowed
Lines m, n a(m, n) Lines not parallel

Table 6.2
Minimal GCS in Euclidean 3-space; d(...) denotes distance, a(...) denotes angle.

Geoms Constraint Notes

Points p1 , p2 , p3 d(pi , pk ) No zero distance
Point p, line L d(p, L) Distance not zero
Lines L1 , L2 d(L1 , L2 ), a(L1 , L2 ) Lines not parallel
Planes E1 , E2 , E3 a(Pi , Pk ) No parallel planes

The constraint graph of Definition 6.1 is an abstract representation of the equation system equiv-
alent to the geometric constraint problem. The analysis of the graph yields a set of operations used
to solve the equations. Ideally, those operations are simple, for instance univariate polynomials of
degree 2 [3]. To start the graph analysis, we find minimal constraint problems. That is, constraint
problems with a minimal number of geometric objects whose solution defines a local coordinate
frame. Such problems depend on ambient space. Consider the following:

Definition 6.8 Given a constraint problem in the Euclidean plane, consisting of two points A
and B and a nonzero distance constraint between them. Such a problem is minimal. The associated
constraint graph G = ({A, B}, {(A, B)}) is a minimal constraint graph.

This minimal constraint problem establishes a coordinate system of the Euclidean plane in which
−
→
A is the origin and the oriented line AB is the x-axis. This is not the only minimal constraint problem
in the plane. Table 6.1 shows the minimal problems involving the basic geometric objects with
2 degrees of freedom. Note that fixed-radius circles can be used in lieu of points as long as the
centers and points are not coincident. Two parallel lines, at prescribed distance in the plane, are not
considered minimal because they do not establish a coordinate frame.
Table 6.2 shows the main cases for Euclidean 3-space. In the case of two lines that are skew, a
third line L3 is constructed that connects the two points of closest approach. Here, L1 and L3 define
a plane that is oriented by L2 . If the lines L1 and L2 intersect, they lie on a common plane that
is coordinatized by the two lines and is oriented by defining the third coordinate direction using a
right-hand rule. If the two lines are parallel, they define a common plane but fail to coordinatize it.

6.3.1 Geometric Elements and Degrees of Freedom

A geometric element has a characteristic number of degrees of freedom that depends on the type of
geometry and the dimensionality of ambient space.

Definition 6.9 Degrees of freedom are the number of independent, one-dimensional variables by
which a geometric object can be instantiated and positioned.
150 Handbook of Geometric Constraint Systems Principles

Table 6.3
Degrees of freedom for elementary objects and rigid objects.

Geom Geometric Meaning of dof 2D 3D

Point Variables representing coordinates 2 3
Line 2D: distance from origin and direction
3D: distance from origin, direction in 3D 2 4
direction on the plane
Plane Distance from origin, direction in 3D 3
Circle, fixed-radius Coordinates of center, orientation in 3D 2 5
Circle, variable-radius Coordinates of center, radius, orientation in 3D 3 6
Rigid body 2D: 2 displacements, 1 orientation 3 6
3D: 3 displacements, 3 orientation
Sphere, fixed- or 3 displacements or 3/4
variable-radius 3 displacements, radius
Ellipse, variable axes Center, axes lengths, axes orientation 5 7
Ellipsoid, variable axes Center, axes lengths, axes orientation 9

Elementary objects are geometric objects that have a certain number of degrees of freedom and
cannot be decomposed further into more elementary objects. Compound objects can be charac-
terized by a GCS, and consist of one or several elementary objects placed in relation to each other
according to a GCS solution instance. A complex object is rigid if its shape cannot change or, equiv-
alently, if the relative position and orientation of the elementary objects that comprise the complex
object cannot change.
The degrees of freedom for elementary geometric objects and for rigid objects in two and three
dimensions are summarized in Table 6.3. Singularities of coordinatization can trigger robustness
issues in GCS. Therefore, the representation of elementary geometric objects should be uniform,
without singularities. For example, if we represent lines by the familiar y = mx + b formula, lines
parallel to the y-axis cannot be so represented. This problem can be avoided if we represent a line
by its distance from the origin and the direction of the normal vector of the line.

6.3.2 Geometric Constraints

Geometric constraints consume one or more degrees of freedom and are expressed by an equal
number of independent equations. A basic set of geometric constraints in the plane is summarized
in Table 6.4. Constraints can be expressed directly. Alternatively, some can be reduced equivalently
to other constraints. For example, suppose we stipulate that a circle of radius r be tangent to a line L
in the plane. Then we can require equivalently that the center of the circle be at distance r from L. For
a comprehensive list of constraints involving planes, points and lines in 3D see for example [40, 57].
Finally we should note that distance 0 between two points takes out two degrees of freedom since
this causes dimension reduction (from 1-dimensional to 0-dimensional). Similarly, requiring that
two lines be parallel at distance d in 3-space eliminates 3 dof, but if d = 0 the constraint eliminates
4 dof.

6.3.3 Compound Geometric Elements

Several geometric elements can be conceptually grouped into compound elements. A circular arc is
an example. Compound geometric elements are convenient concepts for graphical user interfaces of
GCS solvers. Internally, they are decomposed into geometric elements and implied constraints. We
illustrate this dual view with the example of a circular arc shown in Figure 6.5 and explain why it
has overall two intrinsic parameters.
Tree-Decomposable and Underconstrained Geometric Constraint Problems 151

Since the arc does not have a fixed radius, it lies on a variable-radius circle that has 3 dof. The
two end points contribute an additional 4 dof. The two end points are incident to the circle, reducing
the overall dof to 5. Position and orientation of the arc, in some coordinate system, requires 3 dof,
leaving two intrinsic arc parameters. They can be interpreted as 1 dof for the radius of the arc, and
1 dof for the distance between the end points.
Note that this constraint problem has four solutions in general. Orienting the end points, and
connecting them with an oriented line, or circle, the arc can be on one of two sides of the directed
line. Moreover, the two end points divide the circle into a shorter and a longer arc, in general.
Which one is chosen accounts for the other two solutions. Applications require selecting one of
these solutions. Several conventions can be followed to determine a unique segment: preservation of
the original configuration, preservation of the direction of the curve from the first endpoint towards
the second endpoint, the shortest segment, and so on.

6.3.4 Serializable Graphs

When a GCS is serializable (Section 6.1.5) geometric objects can be ordered in such a way that they
can be placed sequentially one-by-one as function of preceeding, already placed elements. This idea
can be formalized as follows.

Definition 6.10 Let G = (V, E) be a GCS graph and x0 , x1 , x2 . . . be elements in V . We say that
xi depends on xk , xr , . . ., written xi > xk , xr , . . ., if xi can be placed only after the xk , xr , . . . have been
placed.

Table 6.4
Types of constraints and number of dof eliminated.
Type Constraint 2D 3D
point-point distance One equation representing the distance be- 1 1
tween two points p1 , p2 under a metric || ||:
||p1 − p2 || = d with d > 0.
Angle between lines and planes Angle between two lines (can be repre- 1 1
sented by the angle between the normal vec-
tors).
Exceptions in 3D: Parallelism between
lines eliminates 2 dof. Line-plane orthogo-
nality in 3D eliminates 2 dof.
Point on point For any metric ||p1 − p2 || = 0 is equivalent 2 3
to p1 = p2 .
Point-line distance One equation expressing the point-line dis- 1 1
tance.
Line-line parallel distance Parallelism and distance. 2 3
Plane-plane parallel distance Parallelism and distance. 2 3
Point on line Same as point-line distance In 3D an addi- 1 2
tional dof is canceled.
Line on line Same as parallel distance between lines in 2 4
2D. In 3D an additional dof is canceled.
Point on plane Same as point-plane distance. - 1
Line on plane Plane-line parallelism and zero distance. - 2
Fixing elementary object Fixing all or some of the dof of an ele- dof dof
mentary object.
152 Handbook of Geometric Constraint Systems Principles

A (2) B (2)

on (1) on (1)

C (3)

(a) (b) (c) (d)

Figure 6.5
Building a complex object using one variable radius circle, two end points, and two incidence con-
straints.

H > A, I
G > I, H
D > A, G
C > A, D
B > A, D
E > D, G
F > D, G
(a) (b)

Figure 6.6
(a) DAG derived from a serializable graph, with (A, I) as starting pair. (b) A construction sequence
respecting the dependencies.

If the constraint graph is serializable, then the pair (G, >) is a directed acyclic graph (DAG) and
admits topological sorting [1]. See Example 13. More formally,

Definition 6.11 Let G(V, E) be the constraint graph associated with a GCS in Euclidean 2-space.
Without loss of generality assume that G is connected. We say that the GCS is serializable if (G, >)
describes a sequence of dependencies such that, under a suitable enumeration of V ,
(a) There are elements x0 and x1 ∈ V that induce a minimal constraint graph.

(b) Each subsequent element xi , 2 ≤ i ≤ |V | depends on elements x j and xk where j < i and
k < i.

In general, the enumeration is not unique and depends on the pair x0 , x1 ∈ V that is placed first.
However, as we will see later, different possible sequences derived from a given DAG are equivalent
in the sense that they lead to the same final placement for all the objects in V with respect to each
other; see Section 6.3.7.

Example 14 Consider the graph in Figure 6.6(a). Edges have been directed to show dependencies
of placement. Choosing (A, I) as starting pair, a valid dependence relation is obtained. The list in
Figure 6.6(b) gives a serial construction based on the graph (G, >).
Tree-Decomposable and Underconstrained Geometric Constraint Problems 153

C
H > A, I
G > I, H
A D

B E > G, F
H E
D > E, F
F

I B > A, C
G D > C, B
(a) (b)

Figure 6.7
(a) DAG derived from the variational graph in Figure 6.4 with starting pairs (A, I), (G, F) and (C, B).
(b) Three different subsequences of construction dependencies that can be identified in the DAG.

6.3.5 Variational Graphs

Definition 6.12 A GCS which is not serializable is called variational.

When the GCS is serializable, starting with a minimal GCS and applying the dependence rela-
tionship to the constraint graph G = (V, E) generates a sequence of dependencies that includes only
a subset of elements in V . We call it a subsequence and the corresponding subgraph a cluster.
Assuming that the variational GCS is solvable, one repeatedly selects a minimal GCS and ap-
plies the dependence relationship, using graph edges not yet used, resulting in a collection of sub-
sequences.
Intuitively, the situation described means that clusters corresponding to different sequences must
be merged, usually applying translations and rotations defined by elements shared by subsequences.
From an equational point of view, the existence of different subsequences reveals that there are
several underlying equations that must be solved simultaneously.

Example 15 Figure 6.7a shows a variational DAG corresponding to the variational constraint
problem in Figure 6.4. We choose three starting pairs (A, I), (G, F), and (C, B). Each allows us
to build a DAG from some of the graph vertices and edges. They are listed in Figure 6.7(b). Notice
that each subsequence identifies a serializable subgraph.

6.3.6 Triangle Decomposability

The strategy of triangle-decomposing solvers, as sketched in Section 6.1, is based on decomposing
the constraint graph recursively. Decomposition splits a (sub)graph into three (sub)subgraphs that
share one vertex pairwise. This is called a triangle decomposition step. More complex splitting
configurations can be considered and called tree decomposition. Figure 6.8 shows a triangle and a
more complex tree decomposition step.
From a practical point of view, in the case of GCS in Euclidean 2-space, triangle-decomposable
constraint graphs suffice for solving many problems that arise in applications.
We define a triangle decomposition step as follows.

Definition 6.13 Let G = (V, E) be a graph. We say that three subgraphs of G, Gi (Vi , Ei ), 1 ≤ i ≤ 3,
define a triangle decomposition step of G if
154 Handbook of Geometric Constraint Systems Principles

(a) (b)

Figure 6.8
(a) Triangle decomposition step. (b) More complex tree decomposition step. Split subgraphs are
shown as ovals, shared vertices as dots.
G1

A D

B
I E

F
G
H
G3
G2

Figure 6.9
A triangle-decomposition step for the graph shown in Figure 6.7(a).

(a) V1 ∪V2 ∪V3 = V , E1 ∪ E2 ∪ E3 = E and Ei ∩ E j = ∅, i 6= j, and

(b) There are three vertices, say u, v, w ∈ V , such that V1 ∩V2 = {u}, V2 ∩V3 = {v} and V3 ∩V1 =
{w}.

Example 16 Consider the graph G in Figure 6.7(a). As shown in Figure 6.9, the subgraphs G1 , G2
and G3 define a triangle decomposition step of G. Vertices pairwise shared by the subgraphs are
A, D and G.

Definition 6.14 We say that a ternary tree T is a triangle decomposition for the graph G if
(a) The root of T is the graph G.
(b) Each node of T is a subgraph G0 ⊂ G which is either the root of a ternary tree generated by a
triangle decomposition step of G0 or a leaf node with a minimal associated subgraph.

Definition 6.15 A graph for which there is a triangle decomposition is called triangle-
decomposable.

In general, a triangle decomposition of a graph is not unique. However, if the graph is triangle
decomposable by one sequence of decomposition steps, then any legal sequence will decompose
the graph [10].
Tree-Decomposable and Underconstrained Geometric Constraint Problems 155

(a) (b)

Figure 6.10
Two different triangle decompositions for the graph shown in Figure 6.7.

Example 17 Consider the graph and the triangle decomposition step shown in Figure 6.9. Now
recursively apply decomposition steps to each of the subgraphs G1 , G2 , G3 until reaching minimal
subgraphs. Figure 6.10 shows two triangle decompositions for the graph considered that differ in
the subtree rooted at node DEFG. Notice, however, that the set of terminal nodes is the same in
both triangle decompositions.

Assume that the three subgraphs G1 , G2 , G3 are (graphs of) solvable GCS subproblems. Let u
and v be the shared vertices of G2 and there is no constraint between them. Then u and v constitute a
virtual minimal GCS where the constraint relating the two elements is not given but can be deduced
from the solution of G2 . A similar statement can be made about G1 and G3 and their shared vertices.
The triangle pattern is not the only decomposition construct [3]. Others include the pattern
shown in Figure 6.8b. Intuitively, a decomposition pattern represents an equation system that must
be solved simultaneously. Decomposition patterns are infinite in number; see also SECTION 2.2

6.3.7 Generic Solvability and the Church-Rosser Property

Recall Definitions 6.13, 6.15, and 6.4.
Triangle-decomposable graphs, of GCS in the Euclidean plane, are generically well-constrained,
and can be solved either bottom-up [11] or top-down [36]. This assertion is based on the shared
geometric elements, of each triangle decomposition step, having 2 dof and being in general position
with respect to each other. Variable-radius circles have 3 dof and give rise to a special case illustrated
in Figure 6.11.
Triangle decomposable graphs are also called ternary-decomposable on account of the topology
of the decomposition tree. A (recursive) decomposition of a well-constrained triangle decompos-
able graph is not unique. However, it can be shown using the Church-Rosser property for reduction
systems that if one triangle decomposition sequence fully reduces the graph, then all such decom-
positions must succeed [10]. The advantage of restricting to triangle-decomposable problems is that
a fixed, finite repertoire of algebraic equation systems suffices to solve this class of problems.
Triangle decomposable graphs are a subset of the set of well-constrained constraint graphs of
planar GCS. The entire set has been characterized by Laman in [38].

Definition 6.16 Let G = (V, E) be a connected, undirected graph whose vertices represent points
in 2D and edges represent distances between points. G is a well-constrained constraint graph of a
GCS iff, the deficit of G is 3 and, for every subset U ⊂ V , the induced subgraph (U, F) has a deficit
of no less than 3.
156 Handbook of Geometric Constraint Systems Principles

B L3

L2 C

A L1

(a) (b)

Figure 6.11
Well-constrained graph with three lines, two points, and a variable-radius circle. (a) Constraint
problem. Dashed lines represent metric constraints. (b) Constraint graph. All vertices have 2 dof
except C which has three. Solid lines represent incidence constraints.

(a) (b)

Figure 6.12
Well-constrained graphs that are not triangle-decomposable. (a) Graph K3,3 . (b) Desargue’s graph.
All vertices have 2 dof, all edges consume 1 dof.

Two examples of well-constrained graphs that are not triangle-decomposable are shown in Fig-
ure 6.12. In triangle decomposition, the irreducible constituents of the constraint graph are the min-
imal constraint graphs, Definition 6.8. For the general case, the set of irreducible constraint graphs
has been characterized in [21] using a network flow approach. Conceptually, irreducible constraint
graphs must be solved as a single equation system. Since the graphs can be arbitrarily complex, so
can be the equation systems.
For the planar case, when we have only points and distances the set of triangle decomposable
graphs coincides with the set of quadratically solvable graphs. However if we extend to lines and
angles, there are quadratically solvable graphs which are not triangle decomposable. Consider the
problem of finding a triangle from its three altitudes shown in Figure 6.13(a). The corresponding
graph is shown in Figure 6.13(b) where the hexagon edges are point-on-line constraints and the
diagonals are point-line distance constraints. The geometric problem is quadratically solvable, [32],
but the graph, K3,3 , is not triangle decomposable.
Laman’s theorem holds even if we extend the repertoire of geometries to any geometry having 2
degrees of freedom and the constraints to virtually any constraint of Table 6.4. However, if we extend
the set of geometries to include for example variable radius circles, then the Laman condition is no
longer sufficient.

Example 18 Consider the GCS of Figure 6.14. We have two rigid clusters C1 = {V1 ,V2 ,V3 ,V4 } and
C2 = {V1 ,V20 ,V30 ,V40 }, where V1 is a variable-radius circle, V2 ,V3 ,V20 ,V30 are points, and V4 ,V40 are
lines. The constraints e1 , e2 , e01 , e02 are distances from the center of V1 , e6 , e06 are distances from the
Tree-Decomposable and Underconstrained Geometric Constraint Problems 157
A b

hb c
b C
c
ha hc

B a C
B a

(a) (b)

Figure 6.13
Deriving the geometry of a triangle given its three altitudes. (a) Formulating the three altitude prob-
lem. (b) The resulting constraint graph.

1
2
1 3

5
4
2 2 1
5
6
3
3
2
6 3
4 4
1 2
4

Figure 6.14
A Laman graph in 2D that is not rigid. V1 is a variable radius circle. V2 ,V3 ,V02 ,V03 are points. V4 ,
V04 are lines. e1 ,e2 ,e01 ,e02 are points. e3 ,e4 ,e03 ,e04 are on constraints. e5 , e05 are distances. e6 , e06 are
distances from cirlce ( circumference).

circumference of V1 , e5 , e05 are distance constraints, and e3 , e4 , e03 , e04 are incidence constraints. The
two clusters share the variable-radius circle V1 .
The graph is clearly a Laman graph but is not rigid , since it is underconstrained — C1 and C2
can move independently around circle V1 . The problem is also overconstrained since the radius of
V1 can be derived independently from cluster C1 and from cluster C2 .

6.3.8 2D and 3D Graphs

Graph analysis for spatial constraint problems is not nearly as mature as the planar case. In the
planar case, Definition 6.8 conceptualizes minimal GCS with which a bottom-up graph analysis
would start and a top-down analysis would terminate. Table 6.1 lists the configurations. In 3-space,
analogously, the corresponding minimal problem may consist of three points or three planes, with
three constraints between them. However, the major configurations shown in Table 6.2 also include
other cases, for example two lines.
In the plane, serially adding one additional point or line requires two constraints. In 3-space,
adding a point or plane requires three. The inclusion of lines in 3-space leads to serious compli-
cations. For example, since lines have 4 dof, see Table 6.3, we have to be able to construct lines
from given distances from 4 fixed points. Equivalently, we have to construct a common tangent to 4
158 Handbook of Geometric Constraint Systems Principles

Figure 6.15
Two hexahedra sharing two points.

fixed spheres, a problem known to have up to 12 real solutions, [24]. Work in [63] has explored the
optimization of the algebraic complexity of 3D subsystems.
In 3D, the Laman condition is not sufficient. Figure 6.15 illustrates two hexahedra sharing two
vertices. If the length of the edges is given the GCS that arises is also known as the double banana
problem. The graph is a Laman graph but the problem corresponds to two rigid bodies (each hex-
ahedron is a rigid body) sharing two vertices and is thus non rigid in the sense that the two rigid
bodies are free to rotate around the axis defined by the two shared points. The problem is also clearly
overconstrained since the distance of the two shared geometries can be derived independently by
each of the two rigid bodies.
Extending the theory in [56, 60], [16, 40] give a combinatorial condition that is necessary and
sufficient for rigidity for extended types of 3D constraints including between point-point, point-
line, plane etc. except for point-point coincidence. Note that this characterization does not capture
the cases of Figures 6.14 and 6.15. The condition holds for both cases, but rigidity is not guaranteed
since they involve point coincidences between rigid bodies directly or indirectly. In this approach, a
multigraph (V, (B, A)) is formulated, where vertices V represent rigid bodies and edges (B, A) stand
for primitive constraints that represent single equations between the two 6-vectors that describe the
rigid body motion of the two vertices. Primitive constraints intuitively affect at most one degree of
freedom. Each geometric constraint is translated to a number of primitive constraints (see Appendix
C of [40]). A distinction is made between primitive angular and blind constraints: a primitive angular
constraint may affect only a rotational degree of freedom. All other primitive constraints that may
affect either a rotational or translational degree of freedom are called blind constraints.
Therefore edges are of two types: angular (A) and blind (B). Such a scheme is minimally rigid
if and only if there is a subset B0 of the blind edges such that (i) B − B0 is an edge disjoint union of
3 spanning trees and (ii) A ∪ B0 is an edge disjoint union of 3 spanning trees.

6.4 Solver
Root Identification or Root Selection: A well constrained GCS may have a large number of so-
lutions. The problem of selecting the user intended solution is referred to as root identification
or root selection.

After the constraint graph has been analyzed, the implied underlying equations are to be solved.
We discuss now how to do that.
Tree-Decomposable and Underconstrained Geometric Constraint Problems 159

Table 6.5
Placing minimal GCS: p represents points, L represents lines, d distance, a angle.

d(p1 , p2 ) Set p1 = (0, 0), p2 = (d, 0)

d(p, L) Place L on the x-axis, p = (0, d)
a(L1 , L2 ) Place L1 on the x-axis, L2 through the origin at angle a

6.4.1 2D Triangle-Decomposable Constraint Problems

We restrict to points and lines in the Euclidean plane. As discussed in Section 6.3.6, triangle-
decomposable constraint systems in the plane require solvers that implement three operations:
(a) The two geometric elements of a minimal subgraph (Definition 6.8) are placed consistent with
the constraint between them.
(b) A third geometric element is placed by two constraints on two geometric elements already
placed.
(c) Given two geometric elements in fixed position, a rigid-body transformation is done that
repositions the two elements elsewhere.
These operations are applied to the decomposition tree, progressing from the leaves of the tree to the
root. The solving order is bottom-up regardless whether the decomposition tree was built top-down
or bottom-up [3, 10, 46].

Example 19 Consider the constraint system of Example 11. We choose the subgraph induced by A
and B as minimal and place A at the origin and B on the positive x-axis, at the stipulated distance
from A, so executing Operation 1.
We place C by the two constraints on A and B, solving at most quadratic, univariate equations,
executing Operation 2. The triangle A, B,C is thereby constructed.
Assume that we have solved the triangle B,C, D in like manner, separately and with B and
D as vertices of the minimal subgraph. We can now assemble the two triangles by a rigid-body
transformation that moves the triangle B,C, D such that the points B and C are matched, using
Operation 3.

Note that we can extend the geometric vocabulary of points and lines, adding circles of given radius
at no cost. A fixed-radius circle is replaced by its center. A point-on-circle constraint is replaced
by a distance constraint between the point and the center, and a tangency constraint by a distance
constraint between the tangent and the center.

Operation 1: Minimal GCS Placement

This operation chooses a default coordinate system. There are three pairs of geometric elements
that occur in minimal GCS: (point, point), (point, line), and (line, line). The chosen placements are
shown in Table 6.5. Other choices could be made.

Operation 2: Constructing one Element from two Constraints

Two elements A and B are given, a third element C is to be placed by constraints upon them. There
are six cases, with p denoting a point and L denoting a line.

(p, p) → p (p, L) → p (L, L) → p

(p, p) → L (p, L) → L (L, L) → L
160 Handbook of Geometric Constraint Systems Principles

(a) (b) (c)

Figure 6.16
Constructing a line at specific distances from two points. Equivalently, finding common tangents of
two circles with radius equal to the stipulated distances. The degenerate cases of 1 or 3 solutions are
not shown.

The sixth case, (L, L) → L is underconstrained. See also [3]. We illustrate with the case (p, p) → L.
Example 20 In the case (p, p) → L a line should be constructed by respective distance from two
points. Consider the two circles with the given points as center and radius equal to the stipulated
distance, as shown in Figure 6.16. Depending on the three distances, there will be 4, 2 or no real
solution in general.
Order the points A and B, and orient the circles centered at those points counter-clockwise.
Orient the line to be constructed so that the projection of A onto the line precedes the projection
of B. Then we can distinguish the up to four tangents in a coordinate-independent way: observe
whether the line orientation is consistent with the circle orientation (+), or whether the orientations
are opposite (–). See also [3].
The degenerate cases where the two circles are tangent to each other yield three solutions or
one. In theses cases there is one double solution that represents the coincidence of two solutions,
with orientations (+,–) and (–,+), or with orientations (+,+) and (–,–).

Operation 3: Matching two Elements

The operation requires that the geometric elements to be matched be congruent. The operation is a
rigid-body transformation and is routine.

6.4.2 Root Identification and Order Type

We noted that constructing one element from two constraints can have multiple solutions. As a
result, a well-constrained GCS has in general an exponential number of solution instances. We shall
illustrate this with the simple construction (p, p) → p.

Example 21 Consider placing n points, by 2n − 3 distance constraints between them, and assume
that the distance constraints are such that we can place the points by sequentially applying the
construction (p, p) → p. In general, each new point can be placed in two different locations: Let
Pi and Pj be known points from which the new point Q is to be placed, at distance di and d j ,
respectively. Draw two circles, one about Pi with radius di , the other about Pj with radius d j as
shown in Figure 6.17. The intersection of the two circles are the possible locations of Q. For n
points, therefore, we could have up to 2n−2 solution instances.

Note that not all construction paths derive real solutions. If, in Example 21, the distance between
Pi and Pj is larger than the sum of di and d j , then there is no real solution for placing Q and therefore
any subsequent construction using this instance of Q is not feasible. Therefore, one might argue that
this pruning may result in polynomial algorithms. However, this is unlikely since the problem of
Tree-Decomposable and Underconstrained Geometric Constraint Problems 161

Q
di
dj
Pi
Pj
Q’

Figure 6.17
Placing one point Q from two points Pi , Pj already known.

d3
P4

d4 P3

α P4
d2 P3
P1

P2
d1 P1
P1 P2 P

(a) (b) (c) (d)

Figure 6.18
(a) GCS consisting of four points, four straight segments, four point-point distances and an angle.
(b),(c) and (d) Three different solution instances to the GCS.

determining whether a well-constrained GCS has a real solution has been shown to be NP-complete
[11].
In general, an application will require one specific solution, usually known as the intended so-
lution. To identify it is not always a trivial undertaking. In [3] finding the intended solution is called
the Root Identification Problem. Notice that, on a technical level, selecting the intended solution
corresponds to selecting one among a number of different roots of a system of nonlinear algebraic
equations.
A well-constrained GCS would not necessarily include enough information to identify which
solution is the intended one. Consider the following example.

Example 22 The well-constrained GCS in Figure 6.18a consists of four points, four straight seg-
ments, four point-point distances and an angle. The solution includes four instances. Two corre-
spond to the one shown in Figure 6.18b and to a symmetric arrangement of the same shape. Solution
instances in Figures 6.18c and 6.18d are structurally different.

Clearly, the GCS sketch in Figure 6.18a does not include any hint on which solution instance
must be chosen to be displayed on the user’s screen. Thus, additional information must be sup-
plied to the solver. In [3], approaches applied to overcome this issue have been classified into five
categories: Selectively moving geometric elements, adding extra constraints to narrow down the
number of possible solution instances, placement heuristics, a dialogue with the constraint solver
that identifies interactively the intended solution, and applying a design paradigm approach based
on structuring the GCS hierarchically. Next we elaborate on each category.

Moving Selected Geometry

In this approach, a solution is presented graphically to the user. The user selects, again graphically,
certain geometric elements that are considered misplaced. The user then moves those elements
where they should be placed in relation to other elements. This approach is used by the DCM solver
[7, 46].
162 Handbook of Geometric Constraint Systems Principles

Adding Extra Constraints

Adding a set of extra constraints to narrow down which is the intended solution instance is an
intuitive and simple approach to solving the root identification problem. Extra constraints could
capture domain knowledge from the application or could be just geometric — and actually over-
specify the GCS. Unfortunately, both ideas result in NP-complete problems [3]. Nevertheless, extra
constraints along with genetic algorithms have been applied to solve the root identification problem
showing a promising potential. Authors in [61] argue that the approach is both effective and efficient
in search spaces with up to 2100 solution instances. A different application of genetic algorithms to
solve the root identification problem is described in [4]. The approach mixes a genetic algorithm
with a chaos optimization method.

Example 23 Genetic algorithms described in [33, 41, 48], use extra geometric and topological
constraints defined as logical predicates on oriented geometries. For example, assume that the
polygon in Figure 6.18a is oriented counterclockwise. The solution shown in Figure 6.18c would
be selected as the intended one by requiring the following two predicates to be fulfilled
−→ −→ −→
PointOnSide(P3 , P1 P2 , le f t), Chirality(P2 P3 , P3 P4 , cw)

Order-Based Heuristics
All solvers known to us derive information from the initial GCS sketch and use it to select a specific
solution. This is reasonable, because one can expect that a user sketch is similar to what is intended.
For instance, by observing on which side of an oriented line a specific point lies in the input sketch
it is often appropriate to select solutions that preserve this sidedness. The solver described in [3]
seeks to preserve the sidedness of the geometric elements in each construction step: The orientation
of three points with respect to each other, of two lines and one point, and of one line and two
points. The work described in [3] implements an additional heuristic for arc tangency which aims
at preserving the type of tangency present in the sketch. See Figure 6.20. When the rules fail, the
solver opens a dialogue to allow the user to amend the rules as the situation might require. These
heuristics are also applied in the solver described in [34].

Example 24 Consider placing three points, P1 , P2 , and P3 , relative to each other. The points have
been drawn in the initial sketch in the position shown in Figure 6.19a. The order defined by the
−→
points can be determined as follows. Determine where P3 lies with respect to the line P1 P2 . If P3
is on the line, then determine whether it lies between P1 and P2 , preceding P1 or following P2 . The
solver will preserve this orientation if possible. For this example, the solver will choose the point P3
as shown in Figure 6.19b.

Dialog with the Solver

A useful paradigm for user-solver interaction has to be intuitive and must account for the fact that
most application users will not be intimately knowledgeable about the technical working of the
solver. So, we need a simple but effective communication paradigm by which the user can interact
with the solver and direct it to a different solution, or even browse through a subset of solutions in
case the one shown in the user’s screen is not the intended one.

Example 25 SolBCN, a ruler-and-compass solver described in [34] offers a simple user-solver

interaction tool. Figures 6.21a and b respectively show the GCS sketch input by the user and the
solution instance selected by the heuristic rules implemented. Then, the user can trigger the solution
selector by clicking on a button and a list with the set of construction steps where a quadratic
Tree-Decomposable and Underconstrained Geometric Constraint Problems 163
P3

P3

P1 P2

P1 P2

P3

(a) (b)

Figure 6.19
Placing three points, P1 , P2 , and P3 , relative to each other. (a) Points placed in the initial sketch and
induced orientation. (b) P3 and P30 are two possible placements for the third point. Preserving the
orientation defined in the sketch leads to select P3 as the intended placement.

(a) (b)

Figure 6.20
Tangency types. (a) Arc and segment tangency. (b) Circle-circle tangency.

equation must be solved is displayed, Figure 6.21c. The user can then change the square root sign
for each of these construction steps by either selecting it directly or navigating with the next/previous
pair of buttons. Figure 6.21d shows a solution different from the first one so obtained.

Navigating the GCS solution space using the approach illustrated in the Example 25 is simple.
But it has obvious drawbacks. On the one hand, the number of items in the list selector grows
exponentially with the number of quadratic construction steps in the GCS. On the other hand it
is difficult to anticipate how choosing a root sign for a construction step will affect the solution
selected by the next sign chosen by the user.
These problems are avoided by considering that, conceptually, all possible solution instances of
a GCS can be arranged in a tree whose leaves are the different instances, and whose internal nodes
correspond to stages in the placement of individual geometric elements. The different branches from
a particular node are the different choices for placing the geometric element. Since the tree has depth
proportional to the number of elements in the GCS, stepping from one solution instance to another
is proportional to the tree depth only. Moreover, it is possible to define an incremental approach by
allowing the user to select at each construction step which tree branch should be used. In solvers
based on the DR-planning paradigm [28], this tree naturally is the construction plan generated by
the solver.

Example 26 The Equation and Solution Manager [51] features a scalable method to navigate the
solution space of GCS. The method incrementally assembles the desired solution to the GCS and
avoids combinatorial explosion, by offering the user a visual walk-through of the solution instances
to recursively constructed subsystems and by permitting the user to make gradual, adaptive solution
choices. Figure 6.22 illustrates the approach.

Design Paradigm Approach

One of the difficulties in selecting the intended solution of a GCS stems from the fact that geometric
elements in a problem sketch are not grouped into logical structures. Authors in [3] argue that
hierarchically structuring the constraint problem would alleviate the complexity of solving the root
164 Handbook of Geometric Constraint Systems Principles

(a) (b)

(c) (d)

Figure 6.21
User-solver dialog offered by the ruler-and-compass solver described in [34]. (a) GCS sketch.
(b) Solution instance selected by the heuristics implemented in the solver. (c) Solution instance
selector. (d) Solution instance displayed after changing the square root signs of some construction
steps.

identification problem, for example grouping geometric elements as design features. First a basic,
dimension-driven sketch would be given. Then, subsequent dimension-driven steps would modify
the basic sketch and add complexity. By doing so, the design intent would become more evident and
some of the technical problems would be simplified.

Example 27 Consider solving the GCS in Figure 6.23a. The role of the arc is clearly to round the
adjacent segments, and thus it is most likely that the solution shown in Figure 6.23b is the one the
user meant rather than the one in Figure 6.23c, when changing the angles to 30◦ . However, the
solver would be unaware of the intended meaning of the arc. Instead, the user could sketch first the
quadrilateral without the arc, and then add the arc to round a vertex. When changing some of the
dimensional constraints, the role of the arc would remain that of a round, so preserving the user
intent.
Tree-Decomposable and Underconstrained Geometric Constraint Problems 165

(a) (b)

(c) (d)

Figure 6.22
Incremental solution space navigation described in [51]. a) GCS problem including three circles. b)
Construction plan graph for the GCS solution. c) GCS solution instance after choosing one of the
possible solutions for each construction step. d) A different GCS solution instance after rebuilding
the partial construction corresponding to the construction step labeled G14 in the tree.

(a) (b) (c)

Figure 6.23
Solution selection by the design paradigm approach: Panel (a) shows the final GCS, panels (b) and
(c) two different solution instances. If the arc is introduced as a rounding feature of a constrained
quadrilateral, then selecting solution (b) over solution (c) is a logical choice.
166 Handbook of Geometric Constraint Systems Principles

6.4.3 Extended Geometric Vocabulary

So far we discussed 2D constraint solvers that use only points and lines, as well as circles of given
radii. Now we will add other geometric element types. This has implications for both the equation
solvers as well as for the constraint graph analysis.
The simplest addition allows GCS to include geometric elements of the new type, but only if
these elements can be constructed sequentially, from explicit constraints on a set of already placed
geometric elements. A more difficult addition would allow the use of elements of the added type
in the same way as the core vocabulary of points and lines. Note that a new element type can have
more than two degrees of freedom.

Variable-Radius Circles
Circles whose radii have not been given explicitly are arguably the most basic extension of the 2D
core solver. Variable-radius circles have three degrees of freedom. We consider two ways in which
they can arise:
(a) A variable-radius circle is to be constructed by a sequential step from three, already placed
geometric entities.

(b) A variable-radius circle is to be determined in a step analogous to merging three clusters

(Definition 6.13) and the circle acts as a cluster.
Note that a variable-radius circle cannot be a shared element in the sense of Definition 6.13. Shared
elements have already been constructed; therefore a shared variable-radius circle has already be-
come a fixed-radius circle.
In the following, we consider points to be circles of zero radius. Consequently, there are four
ways in which a variable-radius circle can be constructed sequentially. Table 6.6 summarizes the
equation systems.
It is advantageous to convert a line-distance constraint into two separate tangency constraints, so
simplifying the equations that must be solved. For example, if the sought circle C is to be at distance
d from line L, we work instead with two problems, one in which the circle C is to be tangent to a
parallel line L+ of L. Here L+ is at distance d from L, on one side of L. In the other problem, C is to
be tangent to a parallel line L− also at distance d, but on the other side of L. Analogously, perimeter
distance from a given circle reduces to tangency with a circle whose radius has been increased or
reduced by said distance.
When deriving the algebraic equations, we work with cyclographic maps, orienting both lines
and circles. Briefly, an oriented circle C is mapped to a normal cone µ(C), with an axis parallel to the
z-axis and a half angle of π/4. The cone intersects the xy-plane in C. Depending on the orientation,
the apex of the cone is above or below the xy-plane. Considered as zero-radius circle, the point P
in the xy-plane maps to the normal cone µ(P) with apex in the xy-plane. Oriented lines L in the
xy-plane are mapped to planes through L at an angle of π/4 with the xy-plane. This reformulates the
constraint problem as a spatial intersection problem of planes and cones. See [5, 19, 25, 26, 47] for
details and further reading.
The intersection of two normal cones is a conic, in affine space, plus a shared circle at infinity.
The conic lies in a plane whose equation is readily obtained by subtracting the two cone equations:
if K1 = 0 and K2 = 0 are the two normal cones, then K1 − K2 = 0 is the plane that contains the conic
in which the two cones intersect.

Example 28 Consider the sequential construction problem of finding a circle that is tangent to
three given circles. This is the classical Apollonius problem that has eight solutions in general.
We orient the circles and require that the sought circle be oriented consistently with the given
circles at the points of tangency. After orienting the given circles, we can map the problem to the
Tree-Decomposable and Underconstrained Geometric Constraint Problems 167

Table 6.6
Sequential construction of variable-radius circles. All constraints are tangent constraints. Equations
formulated using cyclographic maps. For the cases LCC and CCC the linear equation(s) are from
intersecting two cones.

Given Equation Notes

Elements System
LLL (1,1) Intersect two angle bisectors
LLC (1,2) Intersect two planes and a cone
LCC (1,1,2) Intersect two planes and a cone
CCC (1,1,2) Apollonius problem; Example 28

intersection, in 3-space, of three normal cones, C1 , C2 , and C3 , each arising from an oriented circle.
Intersect two cone pairs, say C1 ∩ C2 and C1 ∩ C3 , obtaining two planes that, in turn, intersect in
a line L in 3-space. Then intersect L with one of the cones, say C2 . Two points are obtained that,
understood as the apex of a normal cone, map each to one (oriented) circle in the plane that is
a solution; see [49]. Algebraically, the solution is obtained by solving linear equations plus one
univariate quadratic equation.
There are 8 ways to orient the three circles, but they correspond pairwise, so only four such
problems must be solved. If one or more circles are points, they must be considered oriented both
ways. So, for each zero-radius circle, the number of solutions reduces by a factor of 2. The special
cases of the Apollonius problem have been mapped out and solved in [49] using cyclographic maps.

Now consider the determination of variable-radius circles in a cluster merge. Here, there are
two constraints from each cluster to the variable-radius circle to be constructed, and the two clusters
share a geometric element. The situation is analogous to the triangle merge characterized in Defi-
nition 6.13. The various cases and how to solve them have been studied and solved in [26, 25, 5].
Specifically, [26, 25] map out the cases in which the constraints are on the perimeter of the variable-
radius circle; and [5] considers constraints on the center of the variable-radius circle as well.
Tables 6.7 and 6.8 summarize the results from those papers. The approach is conceptually as
follows. Let S1 = {E0 , E1 , E2 } and S2 = {E0 , E3 , E4 } be the clusters constraining the variable-radius
circle. The two clusters share element E0 , either a line denoted L, or a circle denoted C. The clusters
can move relative to each other, translating along the shared line if E0 = L, or rotating about the
center of the shared circle if E0 = C. Elements E1 and E2 belong to S1 and constrain the sought
circle. Likewise, E3 and E4 are the constraining elements of S2 . Proceed as follows:
(a) Fix the cluster S1 that has the more difficult constraining elements E1 and E2 ; i.e., the cluster
with the larger number of circles.
(b) Choose a convenient coordinate system: the shared line E0 = L as the x-axis, or the origin as
the center of the shared circle if E0 = C.
(c) Construct the cyclographic map of all constraining elements. The cones and planes of S2 are
parameterized by the distance d between S1 and S2 , or else by the angle θ between S1 and S2 .
(d) Construct three planes from the constraining elements E1 , . . . , E4 . They are either cyclo-
graphic maps of lines, or normal cone intersections. The elements of S2 give rise to parame-
terized coefficients, by distance d for translation along the x-axis, or by angle θ for rotation
around the origin, of the moving cluster S2 . Intersect the planes, so obtaining a point with
parameterized coordinates.
(e) Substitute the parameterized point into the equation of the element E1 of the fixed cluster, so
168 Handbook of Geometric Constraint Systems Principles

Table 6.7
Cluster cases; all constraints on circle perimeter.

Constraint type Planes Polynomial degree

E(LL, LL) [L2 ], [L3 ]t , [L4 ]t (1,2)
E(CL, LL) [L2 ], [L3 ]t , [L4 ]t (2,4)
E(CL,CL) [L2 ], [L4 ]t , ([C1 ], [C3 ]t ) (4,4)
E(CC, LL) ([C1 ], [C2 ]), [L3 ]t , [L4 ]t (2,4)
([C1 ], [C2 ])
E(CC,CL) ([C1 ], [C3 ]t ) (4,4)
[L4 ]
([C1 ], [C2 ])
E(CC,CC) ([C1 ], [C3 ]t ) (4,4)
([C3 ], [C4 ])t

Note: E(...) denotes whether clusters share a line L, the translational case, or share a circle C,
rotational case. [X] denotes the cyclographic map equation µ(X) of X; [X]t denotes the equation
with coefficients parameterized by distance t (translation case) or by angle θ (rotation case). (X,Y )
denotes the intersection plane equation of X and Y . The parameterized point is substituted into the
equation [C1 ], except for the first case where it is substituted into [L1 ]. (m, n) denotes the equation
degrees, namely m for the translation case E = L, and n for the rotation case E = C.

obtaining a univariate polynomial that finds the intersection point(s) of the four cyclographic
objects; a polynomial in d or θ .
(f) Solve the polynomial as described, each obtained by a particular configuration of orientations.
Some of the constraints can be on the center c of the variable-radius circle, and [5] considers
those cases. Note that there can be at most two constraints on the center of the variable-radius circle,
for otherwise the relative position of S2 to S1 would be determined and the role of the variable-radius
circle would be curtailed.
The problem is again solved in the same conceptual manner, but with a twist. When a constraint
is placed on the center c of the variable-radius circle, the constraint can be expressed by extending
cyclographic maps with a map τ(X). Here, τ(L) is a vertical plane through the line L. Moreover,
τ(C) is a cylinder through the circle C with axis parallel to the z-axis. The results so obtain are
summarized in Table 6.8.
A problem is denoted E0 (E1 E2 , E3 E4 ). E0 is the shared element by the two clusters, a line L or a
circle C. E1 and E2 are the two elements of the fixed cluster constraining the variable-radius circle.
E3 and E4 are the two elements of the moving cluster constraining the variable-radius circle. The
numbers (m,n) are the equation degrees when E0 = L (m), and E0 = C (n). An element Ek0 constrains
the center, an element E the circumference, of the variable-radius circle.

6.5 Spatial Geometric Constraints

Completable graph, completed graph and completion: An under-constrained graph G(V, E) is
Tree-Decomposable and Underconstrained Geometric Constraint Problems 169

Table 6.8
Cluster cases with constraints on the center of the variable-radius circle.

One center constraint Two center constraints

Problem Degree Problem Degree
E(LL,LL’) (1,2) E(LL,L’L’) (1,2)
E(LL’,LL’) (1,2)
E(CL,LL’) (2,4) E(CL,L’L’) (2,4)
E(CL’,LL) (2,4) E(CL’,LL’) (2,4)
E(C’L,LL) (2,4) E(C’L,LL’) (2,4)
E(C’L’,LL) (2,4)
E(CL,CL’) (4,4) E(CL,C’L’) (4,8)
E(CL,C’L) (4,16) E(CL’,CL’) (4,4)
E(C’L,CL’) (4,16)
E(C’L,C’L) (4,4)
E(CC,LL’) (2,4) E(CC,L’L’) (2,4)
E(CC’,LL) (4,4) E(CC’,LL’) (4,4)
E(C’C’,LL) (2,4)
E(CC,CL’) (4,4) E(CC,C’L’) (4,8)
E(CC,C’L) (4,32) E(CC’,CL’) (8,32)
E(CC’,CL) (8,32) E(CC’,C’L) (8,32)
E(C’C’,CL) (4,8)
E(CC,CC’) (16,64) E(CC,C’C’) (2,8)
E(CC’,CC’) (16,64)

Note: (m, n) denotes the equation degree for E = L and E = C, respectively. L0 denotes a constraint
between a line and the center of the variable-radius circle; C0 denotes a constraint between a circle
and the center of the variable-radius circle.

completable if there is a set of additional edges E 0 such that the graph G0 (V, E ∪ E 0 ) is well-
constrained. We say that a minimal set E 0 is a completion of G and that G0 is a completed graph
of G.

Compared to constraint problems in the plane, our knowledge of spatial constraint systems is
relatively modest. The constraint graph analysis applies with some notable caveats. For example,
Laman’s characterization of rigidity does not apply in 3-space, not even when restricting to points
only, and distances between them; see Section 6.3.8. Furthermore, the subsystems isolated by con-
straint graph decomposition can be complex, especially if lines are admitted to the geometric vo-
cabulary. We illustrate the latter point with a few examples.

Points and Planes

Points and planes comprise the most elementary vocabulary in spatial constraint solving. Both have
three degrees of freedom and are dual of each other. In analogy to the minimal constraint graph in
2D (Definition 6.8), a minimal constraint graph in 3D consists of three elements, points or planes,
and three constraints, forming a triangle. The initial placement for the four combinations places the
elements in canonical order, planes first, points second. Table 6.9 summarizes the method [8]. Note
that the constraint between two planes is an angle, and the constraint between two points or a point
170 Handbook of Geometric Constraint Systems Principles

Table 6.9
Placement of three entities that are mutually constrained. P denotes a plane p a point.
Canonical Placement of Three Points or Planes
P1 placed as the plane z = 0
P1 , P2 , P3 P2 placed to intersect P1 in the y-axis
P3 placed to contain the origin
P1 placed as the plane z = 0
P1 , P2 , p3 P2 placed to intersect P1 in the y-axis
p3 is placed on the xz-plane
P1 placed as the plane z = 0
P1 , p2 , p3 p2 is placed on the positive z-axis
p3 is placed on the xz-plane with z ≥ 0
p1 is placed at the origin
p1 , p2 , p3 p2 is placed on the positive x-axis
p3 is placed on the xz-plane with z ≥ 0

and a plane is a distance. The initial placement fails for the exceptional angles 0 and π, as well as
for distance 0 between two points.
Sequential constructions of points and planes are straightforward. The locus of a third point p,
at respective distances from two known points p1 and p2 , is the intersection of two spheres centered
at p1 and p2 . It is a circle that is contained in a plane perpendicular to the line through p1 ad p2 . As
before, this fact can be used to simplify the algebra.
The simplest, nonsequential constraint system is the octahedron, consisting of 6 elements and
12 constraints [29, 30]. The name derives from the constraint graph that has the topology of the
octahedron. There are 7 major configurations according to the number of planes. The configurations
with 5 and with 6 planes are structurally underconstrained. Solutions of the octahedron constraint
system have been proposed in [8, 29, 30, 44]. The number of distinct solutions is up to 16.

Points, Lines, and Planes

A line in 3-space has four degrees of freedom. Usually lines are represented with 6 coordinates,
using Plücker coordinates, or with 8, using dual quaternions; e.g., [2]. Consequently, the implicit
relationships between the coordinates have to be made explicit by additional equations when solving
constraint systems with lines in 3-space.
The sequential construction of lines can be trivial, for instance determining a line by distance
from two intersecting planes. But it can also be hard, for instance, when determining a line by
distance from four given points in space. The latter problem, in geometric terms, asks for common
tangents to four fixed spheres. This problem has up to 12 real solutions. The upper bound of 12 was
shown in [31] using elementary algebra. The lower bound of 12 was established in [42] with an
example. In contrast to the restricted problem of only points and planes, this sequential constraint
problem involving lines is therefore much more demanding.
We can constrain two lines by distance from each other, by angle, and also by both distance and
angle. Small constraint configurations that involve lines have been investigated [13, 14]. The papers
show that there are two variational configurations of four lines. They are shown in Figure 6.24. The
papers also show that the number of configurations with five geometric elements, including lines, is
17. Moreover, when six elements are considered, the number of distinct configurations grows to 683.
Because of this daunting growth pattern, it is natural to seek alternatives. One such alternative has
been proposed in [58, 59] where, instead of lines, only segments of lines are allowed. Consequently,
Tree-Decomposable and Underconstrained Geometric Constraint Problems 171

Figure 6.24
The two variational constraint problems with 4 lines. The 4 double lines represent both angle and
distance, the solid diagonals distance, and the dashed diagonal angle constraints.

many constraints can be formulated as constraints on end points. Note that in many applications this
is perfectly adequate.

6.6 Under-Constrained Geometric Constraint Problems

In general, existing geometric constraint solving techniques have been developed under the assump-
tion that problems are well-constrained, that is, they adhere to Definition 6.4 given in Section 6.1.
Put differently, the number of constraints and their placement on the geometric elements define a
problem with a finite number of solution instances for nondegenerate configurations. However, there
are a number of scenarios where the assumption of well-constrained does not apply. Examples are
early stages of the design process when only a few parameters are fixed or in cooperative design
systems where different activities in product design and manufacture examine different subsets of
the information in the design model [27]. The problem then is under-constrained, that is, there are
infinitely many solution instances for nondegenerate configurations.

Example 29 Consider the hook of a car trunk locker shown in Figure 6.25. Once distances d1
and d2 have placed the center of the exterior circle of the hook with respect to the hook’s axis of
rotation, the designer is mainly interested in finding a value for the angle α where the exterior circle
is tangent to the small circle transitioning to the inner circle of the hook. The angle α has to be such
that the hook smoothly rotates while closing and opening the hood. At this design stage, the stem
shape of the hook is irrelevant.

This, and other simple examples taken from computer-aided design, illustrate the need for ef-
ficient and reliable techniques to deal with under-constrained systems. The same need is found in
other fields where geometric constraint solving plays a central role, such as kinematics, dynamic
geometry, robotics, as well as molecular modeling applications.
Recent work on under-constrained GCS with one degree of freedom, has brought significant
progress in understanding and formalizing generically under-constrained systems; [52, 54, 55]. The
work focuses on GCS restricted to points and distances, also generically called linkages.
The goal of geometric constraint solving is to effectively determine realizations or embeddings
of geometric objects in the ambient space in which the GCS problem is formulated. Thus, the
current trend is that solving an under-constrained GCS should be understood as solving some well-
constrained GCS derived from the given one.
There are two ways to transform an under-constrained GCS into a well-constrained one: adding
to the GCS as many extra constraints as needed or removing from the GCS unconstrained geometric
entities. Note that removing constrained entities makes little sense. Accordingly, the literature on
172 Handbook of Geometric Constraint Systems Principles
R

r
α

d1

d2

Figure 6.25
Hook of a car trunk locker.

Over−constrained Well−constrained Under−constrained

Tree decomposable

Solvable Completable

Figure 6.26
A partition of the geometric constraint graphs set, G. The set of tree-decomposable graphs straddles
over the sets of well- and over-constraint graphs.

under-constrained GCS advocates to transform an under-constrained GCS into a well-constrained

one by adding new constraints. This technique was formally defined in [35] as follows.

Definition 6.17 Let G(V, E) be an under-constrained graph associated with a GCS problem. Let
E 0 be a set of additional edges each bounded by two distinct vertices in V such that the graph
G0 (V, E ∪ E 0 ) is well-constrained. We say that E 0 is a completion for G and that G0 is the completed
graph of G.

Let G denote the set of geometric constraint graphs. Definitions 6.4, 6.3 and, 6.2 given in Sec-
tion 6.1 induce in G a partition as shown in Figure 6.26. The set of tree-decomposable graphs
straddles over the sets of well- and over-constraint graphs. As described in Section 6.3.6, the set of
well-constrained, tree-decomposable graphs are solvable by the tree decomposition approach.
Within the set of under-constrained graphs we can distinguish two families: Those which are
not tree-decomposable and those which are. It is easy to see that there is no completion for a
graph in the first family that could transform the graph into a tree-decomposable one. Consider-
ing graphs in the second family, Definition 6.4 fixes the number of extra constraints that must be
added. However deciding which constraints should actually be added to the graph is not a straight-
forward matter because the resulting graph could be either over-constrained or well-constrained but
not tree-decomposable.

Example 30 Figure 6.27a shows an under-constrained graph. To see that it is tree-decomposable

just consider as the first decomposition step the subgraphs induced by the sets of vertices {E, F, G},
Tree-Decomposable and Underconstrained Geometric Constraint Problems 173

A C A C A C

B B B

G G
G
D D D

E F E F E F

(a) (b) (c)

Figure 6.27
(a) An under-constrained, tree-decomposable graph G. (b) A tree-decomposable completion of G.
(c) A non tree-decomposable completion of G.

{A, G} and {A, B,C, D, F}. Finding the steps needed to complete a tree-decomposition is rou-
tine. The completion E = {(A, G), (G, B), (G, D), (E, F)} generates the well-constrained graph
shown in Figure 6.27b. To see that the graph is tree-decomposable take as a first decomposi-
tion step the two minimal constraint graphs with edges {(E, D)} and {(E, F)} plus the sub-
graph induced by edges {(A, B), (A,C), (B,C), (B, D), (C, F), (D, F)}. However, completion E =
{(A, E), (E, G), (E, D), (G, D)} results in the graph depicted in Figure 6.27c which is no tree-
decomposable.

In what follows we restrict to the completion problem for triangle decomposable GCS.

Definition 6.18 An under-constrained graph G(V, E) is said to be completable if there is a set of

edges E 0 which is a completion for G.

Notice that completability does not require triangle-decomposability of the underconstrained

graph, nor does it imply that a well-constrained, completed graph be tree-decomposable.

Definition 6.19 Let G(V, E) be a triangle-decomposable, under-constrained graph and let E 0 be a

completion for G. We say that E 0 is a triangle-decomposable completion, (td-completion in short)
for G if G0 (V, E ∪ E 0 ) is triangle-decomposable.

Reported techniques dealing with under-constrained GCS differ mainly in the way they figure
out completions as well as whether they aim at figuring out td-completions or just completions. The
work in [39] describes an algorithm where the constraint graph is captured as a bipartite connectivity
graph whose nodes are either geometric entities or constraints. Each edge connects a constraint node
with the constrained geometric node. In analogy to sequential solvers the graph edges are directed
to indicate which constraints are used to fix (incident) geometric objects. The connectivity graph is
analyzed according to the degrees of freedom of under-constrained geometric nodes. Each under-
constrained geometric node is a candidate to support an additional edge to a new constraint, or if
there is an edge that heads a propagation path to an existing constraint node that has an unused
condition. When there are several candidates on which the new constraint can be established, the
selection is left to the user. A similar approach to solve under-constrained GCS based on degrees of
freedom analysis is described in [45].
Two different notions of td-completion were introduced in [35]. The first one is called free
completion and is computed in three steps. First a triangle decomposition for the given graph is
174 Handbook of Geometric Constraint Systems Principles
A C

ABCDEFG

CF ABCDG DEF

D
ABC AG DBG DE DF EF

AB BC AC DB BG DG E F

(a) (b)

Figure 6.28
(a) A triangle-decomposition for the under-constrained graph G(V, E) in Figure 6.27a. Pairs of ver-
tices {A, G}, {D, E}, {B, G} and {D, G} in leaf nodes do not define graph edges. (b) A set of addi-
tional edges defined on V (G). Edges (A, E), (A, G), (B, G), (D, E), (D, G) and (E, G) do not belong
to E.

figured out. Then the set of under-constrained leaf nodes in the decomposition is identified. Notice
that each node in this set stores a subgraph G(V, E) where |V | = 2 and E is the empty set. Thus
one edge is missing. Finally the completion is computed as the set of missing edges in the under-
constrained leaf nodes of the triangle-decomposition.

Example 31 Figure 6.28a shows a triangle-decomposition for the under-constrained graph

G(V, E) in Figure 6.27a. We have |V | = 7 and |E| = 7. For 2D problems, the general property
of a well-constrained graph described in Section 6.1 is the Laman theorem, [38], 2|V | − |E| = 3.
Hence the number of additional edges required to complete G to a well-constrained graph is
|E 0 | = 2|V | − |E| − 3 = 4. The set of leaves in the triangle-decomposition corresponding to
under-constrained minimal graphs includes exactly four elements: {A, G}, {B, G}, {D, E} and
{D, G}. Thus E 0 = {(A, G), (B, G), (D, E), (D, G)} is a free completion for G. The completed well-
constrained graph G0 (V, E ∪ E 0 ) is shown in Figure 6.27b.

The second td-completion is called conditional completion. The first and second steps are the
same as in the free completion. However, in the third step, edges to complete under-constrained leaf
nodes in the decomposition are drawn from an additional graph defined over a subset of vertices
of the given graph. If the number of those edges that are not in the original graph is smaller than
the number required by Definition 6.4, then the completed graph will remain under-constrained.
However a free completion can eventually be applied to get a well-constrained completion.

Example 32 Consider again the under-constrained graph G(V, E) in Figure 6.27a and its triangle-
decomposition shown in Figure 6.28a. The set of under-constrained pairs of vertices in the triangle-
decomposition is {(A, G), (B, G), (D, E), (D, G)}. Assume that the set of additional edges defined
on V (G) is
{(A, B), (A, E), (A, G), (B, G), (C, F), (D, F), (E, G), (E, D), (G, D)}
as shown in Figure 6.28b. Now, additional edges for the completion must be drawn from a set
E ∗ such that E ∩ E ∗ = ∅. In the case at hand, E ∗ = {(A, E), (A, G), (B, G), (D, E), (D, G), (E, G)}.
Thus, a completion for G(V, E) is E 0 = {(A, G), (B, G), (D, E), (D, G)} ⊂ E ∗ . Figure 6.27b shows
the completed graph G0 (V, E ∪ E 0 ).
Tree-Decomposable and Underconstrained Geometric Constraint Problems 175
P1
d1 P2 on L1 α L2
C0
P2
d0 d1 on on on on
L2
α P3
P0 P1 d0 P0 d2 P3
d2
L1 C1

(a) (b) (c)

Figure 6.29
(a) A crankshaft and connecting rod in a reciprocating piston engine. (b) The crankshaft and con-
necting rod abstracted as a GCS. (c) Geometric constraint graph.

A technique to complete general under-constrained graphs is described in [35]. The approach

is based on transforming the problem of computing a completion into a combinatorial optimiza-
tion problem. Edges in the graph and in the additional set are assigned different weights. Then a
greedy algorithm generates a well-constrained problem provided that there are enough edges in the
additional set. A variant of this approach is reported in [62].
There is a class of 2D, triangle-decomposable, under-constrained GCS that occur in a number
of fields like dynamic geometry or mechanical computer aided design. In these GCS the geometries
are points, the constraints are usually point-point distances, and exactly one edge is missing in the
associated constraint graph. Such GCS are known as linkages.

Example 33 Figure 6.29 a shows an illustration of a crankshaft and connecting rod in a recip-
rocating piston engine. The crankshaft and connecting rod can be abstracted as the GCS shown
in Figure 6.29b. The GCS includes four points P0 , P1 , P2 , P3 , two lines L1 , L2 , three point-point dis-
tances, d0 , d1 , d2 , and one line-line angle, α. Moreover, points P0 , P1 and P2 must be on the line L1 ,
and points P0 , P3 must be on the line L2 . If, for example, values of either the distance d0 or the angle
α are freely assigned, then the GCS can be considered a linkage.

Reachability is an important problem in fields such as dynamic geometry or conformational

molecular geometry. It can be formalized as follows:
Let Rs and Re be two realizations of a well defined geometric construction where Rs
is called the starting instance and Re the ending instance. Are there continuous trans-
formations that preserve the incidence relationships established in the geometric con-
struction and transform Rs to Re?
The well-defined geometric construction in the reachability problem can be understood as a linkage,
that is, a well-constrained GCS problem where values assigned to one specific constraint can freely
change. In [18] the reachability problem is solved assuming that the underlying GCS is triangle-
decomposable. The approach first computes the set of intervals of values that the free constraint
can take for which the linkage is realizable. This set of intervals is known as the linkage Cayley
configuration space [12]. When both Rs and Re are realizations with the free constraint taking
values within one Cayley interval, the path is an interval arc. When Rs and Re realizations belong
to different intervals, finding a path entails figuring out whether there are continuous transitions
between consecutive intervals that permit the linkage to reach Re when starting at Rs. If more than
one such path is found, one is chosen according to some predefined strategy. In [18] the minimal
arc length path is the one chosen.
176 References
− π2 π
2

σ0 (s1 ) = +1; σ0 (s2 ) = +1 A B A → [− π2 , 0]

σ1 (s1 ) = +1; σ1 (s2 ) = −1 A B C B → [0, π2 ]
σ2 (s1 ) = −1; σ2 (s2 ) = +1 C D C → [− π2 , π2 ]
σ3 (s1 ) = −1; σ3 (s2 ) = −1
0 2.2 10.6 17.4 25.8 30 d1

(a) (b)

Figure 6.30
(a) Cayley configuration for distance constraint d1 . (b) Cayley configuration for angle constraint α.

Example 34 The crankshaft and connecting rod GCS in Figure 6.29b is triangle-decomposable,
ruler-and-compass solvable. A construction plan that places each geometric element with respect
to each other is

1. P0 = origin() 5. P2 = intLC(L1 ,C0 , s1 )

2. P1 = distPP(P0 , d0 ) 6. L2 = linePA(P0 , α)
3. L1 = line2P(P0 , P1 ) 7. C1 = circleCR(P0 , d2 )
4. C0 = circleCR(P1 , d1 ) 8. P3 = intLC(L2 ,C1 , s2 )
When a construction step has more than one solution, an orientation parameter is needed to select
the desired solution instance. Parameters s1 and s2 in steps 5 and 8 respectively allow to select one
of the two points in a line-circle intersection. Figure 6.30a shows the Cayley configuration space
when the point-point distance constraint value d1 changes. Intervals labeled A, B, C, and D define
orientations and parameter values for which the GCS is realizable. Intervals A and C yield realiza-
tions consistent with the one depicted in Figure 6.29a. Intervals B and D yield realizations where
point P1 would be placed on the line L1 opposite to P2 with respect to P0 . The Cayley configuration
space when the varying parameter is the angle α is shown in Figure 6.30b. Now orientations are
represented along a radial axis and angle values for which the solution is realizable are represented
as circular intervals.

Linkages are extensively studied in [53, 54, 55]. The object of these works is to lay sound
theoretical foundations for a reliable and efficient computation of Cayley configuration spaces for
general tree decomposable linkages. New concepts like size and computational complexity are in-
troduced and efficient algorithms are developed to answer a number of questions on linkages like
effectively computing Cayley configuration spaces and solving the reachability problem. Methods
so far applied to compute Cayley configuration spaces, like the one used in [18], suffer from po-
tential combinatorial growth. The work in [54, 55] shows that for low Cayley complexity GCS
problems, computing the configuration space is polynomial in the number of geometric elements of
the problem.

References
[1] A. Aho, J. Hopcroft, and J. Ullman. The Design and Analysis of Computer Algorithms.
Addison-Wesley, 1974.
[2] W. Blaschke. Kinematik und Quaternionen. VEB Verlag der Wissenschaften, Berlin, Ger-
many, 1960.
References 177

[3] W. Bouma, I. Fudos, C. Hoffmann, J. Cai, and R. Paige. Geometric constraint solver.
Computer-Aided Design, 27(6):487–501, June 1995.
[4] C.H. Cao, W. Fu, and W. Li. The research of a new geometric constraint solver. In X.-T
Yan, C.-Y Jiang, and N.P. Juster, editors, Perspectives from Europe, and Asia on Engineering
Design and Manufacturing, A comparison of engineering design and manufacture in Europe
and Asia, pages 205–214. Springer Scinence+Business Media, LLC., 2004.
[5] C.-S. Chiang and R. Joan-Arinyo. Revisiting variable radius circles in constructive geometric
constraint solving. Computer-Aided Geometric Design, 221(4):371–399, 2004 2004.
[6] Cinderella. The interactive geometry software Cinderella, June 2007. http://cinderella.de/tiki-
index.php.
[7] D-Cubed. The Dimensional Constraint Manager. Cambridge, England, 2003. 2D DCM
Version 44.0, 3D DCM Version 28.0.
[8] C. Durand and C. Hoffmann. A systematic framework for solving geometric constraints ana-
lytically. JSC, 30:493–519, 2000.
[9] I. Fudos and C.M. Hoffmann. Constraint-based parametric conics for CAD. Computer-Aided
Design, 28(2):91–100, 1996.
[10] I. Fudos and C.M. Hoffmann. Correctness proof of a geometric constraint solver. Interna-
tional Journal of Computational Geometry and Applications, 6(4):405–420, 1996.
[11] I. Fudos and C.M. Hoffmann. A graph-constructive approach to solving systems of geometric
constraints. ACM Transactions on Graphics, 16(2):179–216, April 1997.
[12] H. Gao and M. Sitharam. Combinatorial classification of 2D underconstrained sytems. In
Proceedings of the Seventh Asian Symposium on Computer Mathematics (ASCM 2005), pages
118–127, September 6 2005.
[13] X.-S. Gao, C. M. Hoffmann, and W. Yang. Solving spatial basic geometric constraint config-
urations with locus intersection. CAD, 36:111–122, 2004.
[14] X.-S Gao, C.M. Hoffmann, and W.-Q Yang. Solving spatial basic geometric constraint con-
figurations with locus intersection. In Proceedings of Solid Modeling SM’02, Saarbrucken,
Germany, June, 17-21 2002. ACM Press.
[15] GeoGebra. http://www.geocebra.org/cms, July 2007.
[16] Kirk Haller, Audrey Lee-St.John, Meera Sitharam, Ileana Streinu, and Neil White. Body-
and-cad geometric constraint systems. Comput. Geom. Theory Appl., 45(8):385–405, October
2012.
[17] L. Henneberg. Die graphische Statik der starren Körper. Springer, 1908.
[18] M. Hidalgo and R. Joan-Arinyo. The reachability problem in constructive geometric con-
straint solving based dynamic geometry. Journal of Automated Reasoning, 44(7):709–720,
2013. DOI:10.1007/s10817-013-9280-y.
[19] C. M. Hoffmann. Computer vision, descriptive geometry and classical mechanics. In B. Fal-
cidieno, I. Hermann, and C. Pienovi, editors, Computer Graphics and Mathematics, pages
229–224. Springer Verlag, Eurographics Series, 1992.
[20] C. M. Hoffmann and J. Peters. Geometric constraints for CAGD. In M. Daehlen, T. Lyche,
and L. Schumaker, editors, Mathematical Methods for Curves and Surfaces, pages 237–254.
Vanderbilt University Press, 1995.
178 References

[21] Christoph M. Hoffmann, Andrew Lomonosov, and Meera Sitharam. Finding solvable subsets
of constraint graphs. In Principles and Practice of Constraint Programming - CP97, Third
International Conference, Linz, Austria, October 29 - November 1, 1997, Proceedings, pages
463–477, 1997.
[22] Christoph M. Hoffmann, Andrew Lomonosov, and Meera Sitharam. Decomposition plans for
geometric constraint problems, part ii: new algorithms. Journal of Symbolic Computation,
31(4):409–427, 2001.
[23] Christoph M. Hoffmann, Andrew Lomonosov, and Meera Sitharam. Decomposition plans
for geometric constraint systems, part i: Performance measures for cad. Journal of Symbolic
Computation, 31(4):367–408, 2001.
[24] Christoph M. Hoffmann and Bo Yuan. On spatial constraint solving approaches. In Revised
Papers from the Third International Workshop on Automated Deduction in Geometry, ADG
’00, pages 1–15, London, UK, 2001. Springer-Verlag.
[25] C.M. Hoffmann and C.-S. Chiang. Variable-radius circles of cluster merging in geometric
constraints. Part II: Rotational clusters. Computer-Aided Design, 34:799–805, October 2002.
[26] C.M. Hoffmann and C.-X. Chiang. Variable-radius circles of cluster merging in geomet-
ric constraints. Part I: Translational clusters. Computer-Aided Design, 34:787–797, October
2002.
[27] C.M. Hoffmann and R. Joan-Arinyo. Distributed maintenance of multiple product views.
Computer-Aided Design, 32(7):421–431, June 2000.
[28] C.M. Hoffmann, A. Lomonosov, and M. Sitharam. Decompostion Plans for Geometric Con-
straint Systems, Part I: Performance Measurements for CAD. Journal of Symbolic Computa-
tion, 31:367–408, 2001.
[29] C.M. Hoffmann and P.J. Vermeer. Geometric constraint solving in R2 and R3 . In D.-Z. Du
and F. Hwang, editors, Computing in Euclidean Geometry, pages 266–298. World Scientific
Publishing, 1995.
[30] C.M. Hoffmann and P.J. Vermeer. A spatial constraint problem. In J.P. Merlet and B. Ravani,
editors, Computational Kinematics’95, pages 83–92. Kluwer Academic Publ., 1995.
[31] C.M. Hoffmann and B. Yuan. On spatial constrint solving approaches. In J. Richter-Gebert
and D. Wand, editors, Proceedings of ADG’2000, Zurich, Switzerland, 2000.
[32] R. Joan-Arinyo. Triangles, ruler and compass. Technical Report LSI-95-6-R, Department
LiSI, Universitat Politècnica de Catalunya, 1995.
[33] R. Joan-Arinyo, M.V. Luzón, and A. Soto. Genetic algorithms for root multiselection in
constructive geometric constraint solving. Computer & Graphics, 27(1):51–60, 2003.
[34] R. Joan-Arinyo and A. Soto. A ruler-and-compass geometric constraint solver. In M.J. Pratt,
R.D. Sriram, and M.J. Wozny, editors, Product Modeling for Computer Integrated Design and
Manufacture, pages 384–393. Chapman and Hall, London, 1997.
[35] R. Joan-Arinyo, A. Soto-Riera, S. Vila-Marta, and J. Vilaplana. Transforming an undercon-
strained geometric constraint problem into a wellconstrained one. In G. Elber and V.Shapiro,
editors, Eight Symposium on Solid Modeling and Applications, pages 33–44, Seattle (WA),
June 16-20 2003. ACM Press.
[36] R. Joan-Arinyo, A. Soto-Riera, S. Vila-Marta, and J. Vilaplana-Pasto. Revisiting decomposi-
tion analysis of geometric constraint graphs. In Proceedings of the Seventh ACM Symposium
on Solid Modeling and Applications, SMA ’02, pages 105–115, New York, NY, 2002. ACM.
References 179

[37] R.R. Kavasseri. Variable radius circle computations in geometric constraint solving. Master’s
thesis, Computyer Sciences. Purdue University, August 1966.
[38] G Laman. On graphs and the rigidity of plane skeletal structures. J. Engineering Mathematics,
4:331–340, 1970.
[39] R.S. Latham and A.E. Middleditch. Connectivity analysis: a tool for processing geometric
constraints. Computer-Aided Design, 28(11):917–928, November 1996.
[40] Audrey Lee-St.John and Jessica Sidman. Combinatorics and the rigidity of CAD systems.
Computer-Aided Design, 45(2):473 – 482, 2013. Solid and Physical Modeling 2012.
[41] M.V. Luzón, A. Soto, J.F. Gálvez, and R. Joan-Arinyo. Searching the solution space in con-
structive geometric constraint solving with genetic algorithms. Applied Intelligence, 22:109–
124, 2005.
[42] I. Macdonald, J. Pach, and T. Theobald. Common tangents to four unit balls in R 3. Discrete
and Comp. Geometry, 26:1–17, 2001.
[43] James Clerk Maxwell. On reciprocal figures and diagrams of forces. Philosophical Magazine,
27:250–261, 1864.
[44] D. Michelucci and S. Foufou. Using the Cayley-Menger determinants for geometric constraint
solving. In Solid Modeling and Applications, pages 285–290. ACM, New York, 2004.
[45] A. Noort, M. Dohem, and W.F. Bronsvoort. Solving over and underconstrained geometric
models. In Geometric Constraint Solving. Springer-Verlag, Berlin, Heidelberg, 1998.
[46] J.C. Owen. Algebraic solution for geometry from dimensional constraints. In R. Rossignac
and J. Turner, editors, Symposium on Solid Modeling Foundations and CAD/CAM Applica-
tions, pages 397–407, Austin, TX, June 5–7 1991. ACM Press.
[47] H. Pottmann and M. Peternell. Applications of Laguerre geometry in cagd. CAGD, 15:165–
188, 1998.
[48] E. Yeguas R. Joan-Arinyo, M.V. Luzón. Search space pruning to solve the root identification
problem in geometric constraint solving,. Computer-Aided Design and Applications, 6(1):15–
25, 2009.
[49] K. Ramanathan. Variable radius circle computations in geometric constraint solving. Master’s
thesis, Department of Computer Science. Purdue University, 1996.
[50] E. C. Sherbrooke and N. M. Patrikalakis. Computation of the solutions of nonlinear polyno-
mial systems. CAGD, 10:379–405, 1993.
[51] M. Sitharam, A. Arbree, Y. Zhou, and N. Kohareswaran. Solution space navigation for geo-
metric constraint systems. ACM Transactions on Graphics, 25(2):194–213, April 2006.
[52] M. Sitharam and H. Gao. Characterizing graphs with convex and connected Cayley configu-
ration spaces. Discrete & Computational Geometry, 43(3):594–625, 2010.
[53] M. Sitharam and M. Wang. How the beast really moves: Cayley analysis of mechanisms
realization spaces using CayMos. Computer-Aided Design, 46:205–210, January 2014.
[54] M. Sitharam, M. Wang, and H. Gao. Cayley configuration spaces of 1-dof tree-decomposable
linkages, Part I: Structure and extreme points. arXiv:1112.6008v7[cs.CG], 26 Feb 2014.
[55] M. Sitharam, M. Wang, and H. Gao. Cayley configuration spaces of 1-dof tree-decomposable
linkages, Part II: Combinatorial characterization of complexity. arXiv:1112.6009v4[cs.CG],
7 Nov 2012.
180 References

[56] Tiong-Seng Tay. Rigidity of multi-graphs. i. linking rigid bodies in n-space. Journal of
Combinatorial Theory, Series B, 36(1):95 – 112, 1984.
[57] Gilles Trombettoni and Marta Wilczkowiak. Gpdof a fast algorithm to decompose under-
constrained geometric constraint systems: Application to 3d modeling. International Journal
of Computational Geometry and Applications, 16(05n06):479–511, 2006.
[58] H.A. van der Meiden. Semantics of Families of Objects. PhD thesis, Delft Technical Univer-
sity, 2008.
[59] H.A. van der Meiden and W.F. Bronsvoort. A constructive approach to calculate parameter
ranges for systems of geometric constraints. Computer-Aided Design, 38(4):275–283, 2006.
[60] Neil White and Walter Whiteley. The algebraic geometry of motions of bar-and-body frame-
works. SIAM J. Algebraic Discrete Methods, 8(1):1–32, January 1987.
[61] E. Yeguas, R. Joan-Arinyo, and M.V. Luzón. Modeling the performance of evolutionary
algorithms on the root identification problem: A case study with PBIL and CHC algorithms.
Evolutionary Computation, 19(1):107–135, 2011.
[62] G.-F. Zhang and X.-S Gao. Well-constrained completion and decomposition for under-
constrained geometric consraint problems. Int. Jour. of Computational Geometry & Appli-
cations, 16(5-6):461–478, 2006.
[63] Y. Zhou. Combinatorial decomposition, generic independence and algebraic complexity of
geometric constraint systems. Applications in biology and engineering. PhD thesis, University
of Florida, 2006.
Chapter 7
Geometric Constraint Decomposition:
The General Case
Troy Baker
University of Florida

Meera Sitharam
University of Florida

Rahul Prabhu
University of Florida

CONTENTS
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
7.1.1 Terminology and Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
7.1.2 Triangle-Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
7.1.3 Dulmage-Mendelsohn Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.1.4 Assur Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.1.5 The Frontier Vertex Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
7.1.6 Canonical Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
7.2 Efficient Realization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
7.2.1 Numerical Instability of Rigid Body Incidence and Seam Matroid . . . . . . . . . . 191
7.2.2 Optimal Parameterization in Recombination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
7.2.3 Reconciling Conflicting Combinatorial Preprocessors . . . . . . . . . . . . . . . . . . . . . . 194
7.3 Handling Under- and Over-Constrained Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
7.4 User Intervention in DR-Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
7.4.1 Root Selection and Navigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
7.4.2 Changing Constraint Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
7.4.3 Dynamic Maintenance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
7.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

7.1 Introduction
Solving or realizing geometric constraint systems requires finding real solutions to a large multi-
variate polynomial system (of equalities and inequalities representing the constraints); this requires
double exponential time in the number of variables, even if the type or orientation of the solution
is specified. Thus, to realize a geometric constraint system, it is crucial to perform recursive de-

181
182 Handbook of Geometric Constraint Systems Principles

composition into locally rigid subsystems (which have finitely many solutions), and then apply the
reverse process of recombining the subsystem solutions. The effectiveness of the decomposition is
determined by the size of the largest (indecomposable) subsystem that must be solved. This size
measures the complexity bottleneck. It is an NP-hard combinatorial optimization problem to find
a DR-plan that optimizes this measure. This chapter briefly surveys a series of recursive decom-
position algorithms and culminates in an algorithm to solve the optimal DR-planning problem in
polynomial time for a broad class of minimally rigid or isostatic systems.

7.1.1 Terminology and Basic Concepts

Decomposition-recombination (DR-) planning has been studied since the mid 1980’s. It was first
precisely formulated in [8]. Several useful surveys have been compiled regarding DR-planning [11,
12, 17], to which the interested reader is referred.
DR-planning is a divide-and-conquer technique primarily used for determining the largest sub-
system of polynomials that must be solved simultaneously, thereby reducing the overall time to
realize the system. Informally, a DR-plan is a recursive decomposition of a rigid constraint graph
into rigid subgraphs. These subgraphs can be solved independently of one another and then quickly
recombined into the original graph.
The two main goals of a DR-plan [8] are that (1) the solutions of the small subsystems can be
recombined into a solution for the entire system and (2) the subsystems should be geometrically
meaningful. Consideration (1) is developed further in Section 7.2.2 and consideration (2) should be
clear from the definitions (the subsystems are rigid graphs in the system). Different types of DR-
plans further satisfy (2) by imposing more conditions on the subsystem, as discussed later in this
section. The reader is referred to Chapter 1 for definitions of rigid graphs, independent graphs, dof
rigid graphs, over/under-constrained graphs, etc.

Definition 7.1 Decomposition-recombination (DR-) plan A DR-plan of constraint graph G is

a forest where:

• Each node is a rigid subgraph of G.

• A root node is a maximal rigid subgraph.
• An internal node is the union of its children.

• A leaf node is a single constraint and the involved primitives.

See Figure 7.1 for two simple examples of DR-plans of a 2D bar-joint constraint graph.
Note that this definition permits the same rigid subgraph to appear in multiple nodes of the
DR-plan. Not permitting such duplication would, in general, require the DR-plan to be defined as a
directed acyclic graph instead of a forest. In this way, a DR-plan can also be thought of as a partial
order on the nodes. The forest representation is used to simplify book keeping.
Observe also that the concept of a DR-plan is very general; it is usable for any type of constraint
graphwith an underlying combinatorial (abstract) rigidity matroid and potentially broader classes of
matroid-like structures with certain intersection properties. Furthermore, the DR-plan is combina-
torial, concerned only with the underlying constraint graph. Therefore, a change in parameters has
no impact on the DR-plan (see Section 7.4.2 for additional discussion.)
The size of a DR-plan is the total number of nodes in the forest. The fan-in of a node is the number
of children, and the fan-in of a DR-plan is the maximum number of children over all nodes in the
forest.
A DR-plan is optimal if it minimizes the fan-in over all possible DR-plans. This criterion for
Geometric Constraint Decomposition: The General Case 183

Figure 7.1
Two DR-plans of the same 2D bar-joint constraint graph (a C2 × C3 , or doublet.) The first is the
worst possible DR-plan, decomposing into the individual edges of the graph. The second is much
better, taking advantage of the rigid subgraphs (the two triangles.)

optimality is chosen because it measures how efficiently a GCS can be realized. The time is dom-
inated by the largest subsystem to be recombined (the maximum fan-in of the DR-plan). If the
optimal DR-plan has a fan-in of |E|, the number of constraints, the system is called indecompos-
able; the best that can be done is breaking the system down into its constituent constraints, and there
is no improvement over solving the entire system simultaneously.
Finding an optimal DR-plan is NP-hard in general [12]. When dealing with a class for which
optimal plans cannot be readily constructed, an algorithm can instead build DR-plans that satisfy
other desirable properties. Several different meaningful properties have been studied. For example,
see Section 7.1.5 for an algorithm that finds a DR-plan with cluster minimality, i.e. the union of
no subset of children forms the parent. Section 7.1.6 discusses proper vertex-maximality, i.e., each
child is a rigid vertex-maximal proper subgraph of the parent, which happens to be optimal if the
input is independent.
In addition to optimality, a “good” DR-planner should adhere to the following principles [11]:
• Church-Rosser property (or Confluence) – If the DR-planner can follow multiple computa-
tional paths on the same input they all lead to the same outcome: either they all generate a
valid DR-plan for the graph or they don’t.
• Completeness – A DR-plan is found if one exists.
• Low time complexity.
184 Handbook of Geometric Constraint Systems Principles

• Correctness – The output is a valid DR-plan.

There are several natural use-cases of DR-plans, which include:
• Determining complexity of realization. The largest subsystem that must be solved simulta-
neously determines the time complexity.
• Guiding realization. The partial order of the tree establishes the partial-order in which to
solve the system.
• Decomposing the rigidity matrix of the constraint graph, and thereby the self-stress and flex
matrices. Notice also that the DR-planner is finding independent systems. Therefore, a DR-
plan corresponds to a block decomposition of the rigidity matroid and matrix. As mentioned
in Chapter 1, the left and right nullspaces of the rigidity matrix are the space of self-stresses
and flexes, respectively. Therefore, this is also a decomposition of stresses and flexes.
• Completion of under-constrained graphs. See Section 7.3.
• Guided removal of constraints from over-constrained graphs. See Section 7.3.
The first and second items are the main motivation for DR-planning. The first was treated above
and the second is discussed in Algorithm 1. This is a simple and naı̈ve algorithm for using the DR-
plan. More complex methods are discussed in Section 7.2. Next we survey different methods for
DR-planning.

Algorithm 1 Realization using a DR-plan

1: Recursively realize the children of the root. Base case: a child is a single constraint and is
realized in constant time.
2: Recombine the children into a unified realization. This can be done by choosing a single child
as a home coordinate system and then expressing the variables of the other children in terms of
the chosen one. Then the polynomial system is solved.

7.1.2 Triangle-Decomposition
The basic algorithm (first appearing in [4] with proof of correctness in [6]) was originally worked
for 2D point-line systems, but here it will be discussed in the context of 2D bar-joint. Triangle-, or
tree-, decomposition is an older form of decomposition that only works on a specific class of graphs.
In practice, this class arises often.
Algorithm 2 explains the bottom-up algorithm to construct the DR-plan. If Step 4 fails, then
there is no triangle-decomposition. This algorithm will find a triangle-decomposition if one exists.
Conceptually, this can also be thought of in a top-down recursive manner. A triangle-
decomposable system is one that can be separated into 3 subsystems that (a) pairwise intersect
on a single vertex and (b) are each triangle-decomposable.
Realizing a graph with a triangle-decomposition has a simple, efficient algorithm. In fact, the
class of graphs with a triangle-decomposition is exactly the class of graphs that are quadratically
radically solvable (QRS). Each node can be found with a simple ruler-and-compass construction.
The constituent cluster has some distance between the participating vertices (those shared between
other sibling clusters) which can be used in simply solving a simple triangle. A single cluster is
chosen as a frame of reference and the other two are translated to line up with the vertex they share
and then rotated to solve the triangle system.
Geometric Constraint Decomposition: The General Case 185

Algorithm 2 Triangle Decomposition

1: Every edge is registered as a cluster. . Intuitively, a cluster will be a subgraph for which a
realization is known. An edge is the base case.
2: Create a node for each cluster. . These initial clusters will be the leaves of the DR-plan.
3: while there is more than one cluster do
4: Find three clusters, Ci ,C j ,Ck , such that any two share a single vertex.
5: Add a new cluster Cm that is a union of Ci ,C j ,Ck .
6: Add a new node for Cm , and set it as the parent of Ci ,C j ,Ck .
7: Remove Ci ,C j ,Ck from the list of clusters.
8: end while
. Ultimately, the graph will be reduced to a single cluster.

However, there are many rigid graphs that are not triangle-decomposable, including some simple
ones such as C2 × C3 and K3,3 . The idea of shape recognition is to look for specific shapes in
the graph from a finite list of shapes, then recursively decompose the graph as discussed above.
Triangle-decomposition is shape recognition with a triangle as the only shape in the list. I.e. the
recursive definition would be to choose a shape from a list with n edges and to find n subsystems that
(a) intersect on a single vertex in the same configuration as the shape and (b) are each decomposable
by the same list of shapes. However, there is no finite list of shapes that can characterize rigidity in
even two dimensions.

7.1.3 Dulmage-Mendelsohn Decomposition

Inspired by the algorithms for rigidity detection as discussed in Chapter 1, these algorithms use a
form of maximum matching to generate the DR-plan. Such algorithms are referred to as generalized
maximum matching, generalized network flow, or Dulmage-Mendelsohn decompositions.
Consider the equation graph (Y ∪ U, E) made by the system of equations of a GCS, where |Y |
is the number of equations, |U| is the number of unknowns, and for all unknowns contained in an
equation there is an edge between their associated vertices. This graph is naturally bipartite. How-
ever, this does not account for trivial motions. So often, when working with geometric constraint
systems, the trivial motions should be factored out. This is done by pinning the system to the em-
bedding space. This means that some of the variables are removed. In a bar-joint system, a single
edge and the two associated vertices are removed, essentially fixing one vertex at the origin, and the
other on some axis.
The system is well-constrained if and only if the equation graph satisfies the König-Hall relation:
for any subset of equations Y 0 of set Y , then |Γ(Y 0 )| ≥ |Y 0 | where Γ(Y 0 ) is the set of all neighbors
of Y 0 , i.e. the set of unknowns appearing in Y 0 . This is equivalent to the existence of a perfect
matching. A system is irreducible if every subsystem is under-constrained. An under-constrained
system satisfies |Γ(Y 0 )| |Y 0 |
Ref. [1] introduced a new form of decomposition that works for any independent input. If the
input G is well-constrained, then it is bipartite and has a perfect matching, making the algorithm
much simpler. The well-constrained variant is presented in Algorithm 3, their paper gives more
details regarding the complete algorithm.
A directed graph is strongly connected if, for all i and j in the vertex set, there is a path from i
to j and from j to i. A strongly connected component of a graph is a maximal strongly connected
subgraph. Finding all strongly connected components can be done in O(|E|), via several well-known
algorithms (such as Tarjan’s.) Note that the strongly connected components are indecomposable
(irreducible) subgraphs of G.
The resulting graph R is a DAG. A DR-plan can be made from R by Algorithm 4.
186 Handbook of Geometric Constraint Systems Principles

Algorithm 3 Dulmage-Mendelsohn decomposition of well-constrained graphs

1: Find a maximum (perfect) matching M of G = (Y ∪U, E).
2: Build a directed graph G0 by replacing all edges (s,t) ∈ M with the directed edges (s,t) and
(t, s) and directing all other edges from Y to U.
3: Compute the strongly connected components of G0 .
4: Contract each strongly connected component into a single vertex to get R.

Algorithm 4 DR-plan from DM decomposition

1: Run some topological sort to get a total ordering, with the largest element being some sink in
the DAG.
2: Add a node whose children are the equations (constraints) in the largest element.
3: for each preceding element ai−1 in the ordering do
4: Add a node whose children are the equations in ai−1 and the node that was added for ai .
5: end for

7.1.4 Assur Decomposition

Assur graphs (traditionally Assur Groups) were introduced by Leonid Assur in 1914 [2]. Originally
a tool in mechanical engineering, the first paper to use the mathematics of rigidity theory was [15].
The terminology used in this section will be from the rigidity community and is mostly due to [15].
Since their introduction, they have become a mainstay in the design and analysis of mechanisms
(1-dof linkages). Originally designed for 2D pinned linkages, the concepts have been generalized to
all dimensions and symmetric settings [13].
The type of system considered in this section are bar-joint graphs with additional unary pin
constraints and the restriction that no two vertices with a pin can share an edge. A pin is a unary
constraint that fixes both degrees-of-freedom of a joint, giving it an absolute position relative to the
embedding space. Joints without pin constraints are called unpinned, free, or inner vertices. The
rigidity of a pinned graph (or grounded graph ) is different, as the system must not only be rigid but
also have no trivial motions. Furthermore, a generically rigid pinned graph only requires that inner
vertices be placed generically and pinned vertices be distinct. The rationale behind this is motivated
by mechanical engineering, the trivial motions are uninteresting, only the internal motions of the
mechanism itself are of concern. Also, while this section defines everything in terms of graphs, the
theory is easily generalized to multigraphs.
A d-Assur graph is a minimally rigid (in d dimensions) pinned graph whose proper subgraphs
are all flexible. An Assur decomposition is a partial order of the Assur graphs in a minimally rigid
pinned graph, which is unique [15].
Given input graph G = (I ∪ P, E), where I is the set of inner vertices and P is the set of pinned
vertices; therefore, E ⊆ (I × I) ∪ (I × P). See Algorithm 5 for Assur decomposition.

Algorithm 5 Assur Decomposition

1: Find a d-directed orientation of G . see definition below
2: Combine all pinned vertices into a single sink . referred to as grounding
3: Find the strongly connected components . see Section 7.1.3 for details

The first step is finding a d-directed orientation of G, which is an assignment of directions to

the edges such that every inner vertex has out-degree d and every pinned vertex has out-degree zero.
Every minimally rigid d-dimensional pinned graph has a d-directed orientation. If d = 2, this can be
Geometric Constraint Decomposition: The General Case 187

found quickly via algorithms such as a slightly modified pebble game [10]. In higher dimensions,
no combinatorial polynomial time algorithm is known.
While the d-directed orientations of a graph are not unique, they are unique modulo the orien-
tation within a strongly connected component. Any two direction assignments on the same graph
with the same out-degree at every vertex can be obtained from one another simply by the reversal
of directed cycles. Thus, any d-directed orientation is equivalent to another modulo orientation of
cycles. This means that the strongly connected components are the same and the decomposition is
invariant to the choice of d-directed orientation on those graphs that admit them.
The strongly connected components are the d-Assur graphs of G. The directed edges between
the components are a partial order on the Assur graphs to guide decomposition. A DR-plan can be
made in a fashion similar to Algorithm 4. Simply perform a topological sort to get a total ordering,
and at each level of the DR-plan recombine the growing graph with another component (Assur
graph), namely the next one in the ordering.
When d = 2, a pinned graph admits a 2-directed orientation if and only if it is minimally rigid.
Furthermore, the response of inner vertices to using a driver is not intuitive, as discussed above,
unless d = 2. Nevertheless, Assur decomposition can be generalized to d > 2, as in [16]. In higher
dimensions, a d-directed orientation is a necessary condition for d-dimensional rigidity of pinned
graphs.
If, by the removal of a single single edge, all inner vertices of a d-Assur graph go into motion
(i.e., there is some satisfying infinitesimal flex), the graph is called strongly d-Assur. In the case
of 2-Assur graphs, the graph is always strongly 2-Assur; for dimensions greater than two, it is not
always the case [16].
Originally used in mechanical engineering, Assur decomposition is well-suited to study the mo-
tion of under-constrained systems with 1 dof. A non-edge is an intuitive term that refers to a pair of
vertices without an edge between them. In a system with 1 dof, setting the length of a non-edge that
is independent of all edges gives a well-constrained graph from which a realization can be obtained.
By continuously varying the length of this non-edge, the graph continuously changes realizations.
Such a non-edge is called a driver, as it drives the motion of the mechanism. By performing Assur
decomposition on minimally rigid graphs, the partial ordering imposed on the edges will describe
the motion of the entire system given some edge were used as a driver.
Instead of using a rigid graph, one edge can be chosen to be a driver, to which some velocity
is assigned. The position of all vertices can be easily calculated when changing the length of the
driver. A driver only affects the positions of vertices in Assur graph components higher in the partial
ordering. The velocities of the vertices in the component containing the driver can be calculated, and
then used as driver velocities in the calculation of movement higher up the DR-plan tree.
For a better understanding of the configuration space of under-constrained systems (such as
these 1 dof systems), see Chapter 10.

7.1.5 The Frontier Vertex Algorithm

Frontier is a DR-planner that operates via a bottom-up, network-flow based algorithm. Algorithm
6 describes Frontier. The problem of finding the minimum dof-rigid (see Chapter 1) subgraph is
NP-hard [12] even for 2D. The algorithm dense can be used to find a minimal, non-trivial (i.e., more
than a single edge) dof rigid subgraph (see Line 3 in Algorithm 6). See Chapter 1 for an explanation
of dof-rigid subgraphs and [7] for an explanation of algorithm dense.
Some terminology used in the algorithm is defined here. Given some input G and a subgraph
S = (I ∪ F, E), those vertices F which are adjacent to at least one vertex not in I ∪ F are called
frontier vertices, and those vertices I that are adjacent to only frontier and other vertices in I are
called inner vertices.
Frontier works for independent and over-constrained systems. For over-constrained systems,
output is certainly not an optimal DR-plan, in that it does not guarantee optimal fan-in. For well-
188 Handbook of Geometric Constraint Systems Principles

Algorithm 6 Frontier
1: G0 = G.
2: while Gi is not a single vertex do
3: Find Si = (I ∪ F, ESi ), an arbitrary dense subgraph of Gi .
4: Gi+1 = Gi .
5: Remove all edges in ESi from Gi+1 .
6: Remove all inner vertices I from Gi+1 .
7: Add a single vertex (the core) to Gi+1 .
8: Add an edge from every frontier vertex F to the core. The weight of the new edge is the sum
of the weights of the original edges that got merged into the new edge, i.e. all those that went
from this frontier vertex to some P inner vertex. Call
P this set of edges EF .
9: The core is given the weight v∈F w(v) − e∈EF w(e) − D. . i.e. the weight such that this
new subgraph (F ∪ core, EF ) is dense.
10: end while

constrained systems, there is no guarantee of optimality, although whether or not the output is op-
timal is unknown. What Frontier does ensure is that the output DR-plan has cluster minimality.
Cluster minimality ensures that there is no proper subset (of size greater than 1) of children whose
union is dense.

7.1.6 Canonical Decomposition

Thus far, several DR-planning algorithms have been presented that create DR-plans in polynomial
time for independent input graphs with an underlying sparsity matroid (see Chapter 1). However,
the bottom-up nature of those algorithms made it difficult to provide formal guarantees about op-
timality; instead they made guarantees about other desirable properties, discussed previously. Here
we discuss a top-down O(|V |3 ) algorithm that produces an optimal DR-plan for such graphs, written
by the author of this survey [3].
In this section, we discuss only 2D bar-joint systems, although the paper provides extensions to
body-bar, body-hyperpin, and pinned line-incidence systems. The unifying structure of these diverse
systems is: a combinatorial abstract rigidity matroid, and for the algorithmic results an underlying
sparsity matroid generalization.
We define a canonical DR-plan as:

Definition 7.2 Canonical DR-plan A DR-plan that satisfies the two additional properties:
(a) Children are rigid vertex-maximal proper subgraphs of the parents, and
(b) If all pairs of rigid vertex-maximal proper subgraphs intersect trivially then all of them are
children, otherwise exactly two that intersect non-trivially are children.

A trivial subgraph is taken to mean a single vertex. See Figure 7.2 for an example of a canonical
DR-plan. This definition gives gives the canonical DR-plan a surprisingly strong Church-Rosser
property, stated in the following theorem.

Theorem 7.1 A canonical DR-plan exists for a graph G and any canonical DR-plan is an optimal
DR-plan if G is independent.

As a result of the preceding theorem, any canonical DR-plan we create will be optimal (of which
there are potentially exponentially many.) Furthermore, if we construct a DR-plan where all nodes
have fan-in less than or equal to the fan-in of some canonical DR-plan, we have made an optimal
Geometric Constraint Decomposition: The General Case 189

Figure 7.2
Two DR-plans of the same 2D bar-joint constraint graph (three C2 × C3 graphs, or doublets, inter-
secting on a single triangle.) The first is a canonical DR-plan, the first and second levels demon-
strating non-trivial intersections. The second is a pseudosequential DR-plan.

DR-plan. Such an alternative DR-plan (which we show to be optimal) is the pseudosequential DR-
plan:

Definition 7.3 Pseudosequential DR-plan A DR-plan recursively defined as follows:

(a) Base case: When G is an edge, the pseudosequential DR-plan for G is G itself.
(b) In case the pairwise intersections of the rigid vertex-maximal proper subgraphs Ci of G are
all trivial, take the children of G to be the roots of the pseudosequential DR-plans for Ci .
(c) In case there are two rigid vertex-maximal proper subgraphs Ci and C j of G with non-trivial
intersection, take the children of G to be the roots of pseudosequential DR-plans for C j \ Ci
(called an appendage), and Ci (called its partner).

See Figure 7.2 for an example of a pseudosequential DR-plan. Note the difference with canoni-
cal, when the intersection is nontrivial instead of the two proper vertex-maximal rigid subgraphs Ci
and C j both becoming children, the appendage and partner become children. This has the impor-
tant affect of making every node in the DR-plan unique, leading to a much smaller size. Whereas
canonical could be exponential in the size of the graph, pseudosequential will be linear while still
being optimal. Additionally, with its much smaller size, finding a pseudosequential DR-plan has an
efficient algorithm, as expressed by the following theorem.

Theorem 7.2 Computing a pseudosequential DR-plan for an independent graph G = (V, E) has
time complexity O(|V |3 ).

The algorithm involves recursively “growing” the DR-plan tree one branch at a time. We call a
190 Handbook of Geometric Constraint Systems Principles

G f f

L L

Figure 7.3
An illustration of the running of the algorithm to find a pseudosequential DR-plan. The DR-plan
begins as just the input graph. A leaf is chosen (only can choose G) and the branch is computed
down to arbitrary edge f ∈ L. Then another leaf is chosen, and its branch is computed, etc. In this
way, the entire tree is built.

branch(T, a, b) of tree T the subtree that consists of the path from node a to b plus all children of
the nodes on the path.
Given some graph G and edge e ∈ G, there exists a pseudosequential DR-plan PG where the
leaves of branch(PG , G, e) is exactly, ignoring some exceptions here, the rigid vertex-maximal proper
subgraphs of G \ e. The branch PG can be built with Algorithm 7.

Algorithm 7 Branch from rigid vertex-maximal proper subgraphs

1: Compute the rigid vertex-maximal proper subgraphs of G \ e, {Li }.
2: for each L ∈ {Li } do
3: Choose an arbitrary edge f ∈ L and compute the rigid vertex-maximal proper subgraphs of
G \ f , {Mi }.
4: Compare the intersection of L with each M to get its position relative to the other leaves.
5: Compute nodes on the path from G to e.
6: end for

Assuming a preprocessing step of finding all of the rigid vertex-maximal proper subgraphs of G \ e,
for all e, then the above algorithm runs in time O(|V |2 ). The details of Step 4 are omitted, however,
each loop L is being used as a pivot; after running this on each leaf the position of each leaf relative
to one another is determined. Step 5 is simple because the union of the nodes is the parent, and we
have all but one node and we know how they intersect.
See Algorithm 8 for how to build the entire DR-plan PG .
For an illustration of this algorithm, see Figure 7.3. Computing rigid vertex-maximal proper sub-
graphs is time O(|V |2 ), via the pebble game algorithm [10]; thus, preprocessing takes time O(|V |3 ).
Geometric Constraint Decomposition: The General Case 191

Algorithm 8 Pseudosequential DR-plan from branches

1: Preprocessing: Compute the rigid vertex-maximal proper subgraphs of G \ e, for all e.
2: Start with G as the single node in the DR-plan.
3: Recursively compute a branch for each leaf in the DR-plan (that is not a single edge.)

Observe that the size of a pseudosequential DR-plan is O(|V |); thus, the complexity of the last two
steps will be O(|V |3 ) as well.
We conjecture that the, for independent input, Frontier (see Section 7.1.5) produces a pseudose-
quential DR-plan and is therefore optimal as well. However, the bottom-up method of computation
makes analysis much more difficult. Both the complexity and the optimality in the case of well-
constrained input is difficult to understand for Frontier.

7.2 Efficient Realization

Given a geometric constraint system, a realization can naı̈vely be found by directly finding the zeros
of the corresponding polynomial system. If the independently solvable subsystems are found, then
those can be solved separately and then recombined. This process applied recursively can be done
with DR-planning. At this point, one may solve this system bottom-up; recombining constraints in
the leaves into the parent node, working up to the root. The DR-plan tree is a partial order in which
the solutions can be calculated.
While some subclasses of systems have efficient realization schemes, such as triangle-
decomposable systems, this is not the case in general. Upon decomposition, there can be arbitrarily
large systems that must be solved simultaneously. Even in an optimal DR-plan, the fan-in at any
node can be as large as the number of constraints in the corresponding subgraph; such a system is
called indecomposable . Therefore, it may be desirable to further reduce the algebraic complexity
of the subsystems that must be solved.

7.2.1 Numerical Instability of Rigid Body Incidence and Seam Matroid

Consider bar-joint systems, where a trivial intersection is taken to mean a vertex in 2D, an edge in
3D, etc. When recombining rigid bodies whose intersection is non-trivial, and necessarily rigid, it
can be done by simply rotating the reference frame of one system so that the overlapped primitives
align. However, given a trivial intersection, there must be some complex system of rigid bodies.
This system is its own polynomial system that can be solved.
Consider two rigid bodies in 3D that intersect on two vertices which represent six incidence
constraints. The two bodies individually have six dof each, and the count indicates the system is
well constrained . However, one of these incidence constraints is dependent as the two body system
clearly has one degree of freedom. All six dependencies will result in numerical instability when
the constraint system is being solved.
To address this, one incidence constraint can be dropped. As it turns out, there is an underlying
matroid [18], the seam matroid, characterizing the dependencies of these constraints. This matroid
admits an efficient greedy algorithm (the well-formed incidences algorithm) to find a maximal in-
dependent set of constraints.
192 Handbook of Geometric Constraint Systems Principles

Figure 7.4
(Top,Left) shows the seam graph for a decomposition; this is the ground set of the seam matroid.
(Top Right) shows a seam graph that is independent but not maximal. (Bottom, Left) shows a de-
pendent seam graph. (Bottom, Middle) and (Bottom, Right) Show two maximally independent seam
graphs.

7.2.2 Optimal Parameterization in Recombination

Obtaining a realization of the constraint system of an internal node in a DR-plan can be done by
choosing a home coordinate system, that of some child node, and repositioning the rest of the chil-
dren (by using the allowable rigid motions of the system) in the chosen coordinate system such that
all incidences are satisfied.
To minimize the algebraic complexity of this recombination system, Sitharam et. al. in [20] in-
troduce a combinatorial class of incidence tree parametrizations for a collection of rigid bodies.
Minimizing the algebraic complexity over this class reduces this to a purely combinatorial op-
timization problem that is a special case of the set cover problem. They also quantify the exact
improvement of algebraic complexity obtained by optimization.
Consider DR-plans of bar-joint systems where all pairwise intersections are trivial, e.g., a vertex
in 2D, a vertex or edge in 3D, etc. Then the children of a node from an overlap graph, where vertices
are the set of nodes and edges represent incidence between the child subgraphs. A spanning tree of
this overlap graph, with a specified root, is an incidence tree [20]. Using an incidence tree, the
system of polynomials describing the constraint system can be rewritten, leading to reduction in the
complexity of teh resulting parameterized system.
The choice of origin and spanning tree can cause significant improvements in the complexity
of the polynomial system. The total weight of the incidence tree corresponds to the total number
of parameters in the new polynomial system. The depth of an incidence tree corresponds to the
degree of the polynomial system. Ref. [20] shows how to simultaneously optimize over both of
these properties. The overlap graph is used to choose and order the elimination of incidences, i.e.,
to select an incidence tree parametrization that minimizes the number of variables and the algebraic
degree of the parametrized core system (minimizing algebraic complexity)
We first define some background terms and describe the algorithm using Figure 7.5
Geometric Constraint Decomposition: The General Case 193

Figure 7.5
(Top, Left) Standard collections of rigid bodies and implied internal distances. (top, right) The
corresponding weighted overlap graph (edges of weight 6 representing empty overlaps are omitted).
(Middle) Five spanning trees of the overlap graphs of covering sets S(c). The first and the fourth
trees show implied non-tree distance constraints (shown in dotted lines) - they rigidify the remaining
bodies in c \ S(c). The label i in a circle stands for ci . (Bottom) Two of the eight possible solutions
for the input collection of rigid bodies shown in (top, left). For each tetrahedral rigid body, c1 , c2 ,
c3 , c4 , only two triangle facets are shown. The home rigid body c5 is a triangle. The dashed line
represents the parametrized distance constraint. The parametrized incidence constraint is marked
as a black sphere. It forms a triangle with the distance constraint, i.e., together, they constitute the
triangle rigid body c6 (shown in top, left), that is not part of the current covering set (see Middle,
Left). (Bottom, Left) This configuration can be recognized as representing two pentagons in the
z = 0 and z = 1 planes respectively and rotated by 2pi/10 with respect to one another. (Bottom,
Right) This realization is self-penetrating.
194 Handbook of Geometric Constraint Systems Principles

Collection of Rigid Bodies, Covering Set: Let c := c1 , . . . , cn be a set of rigid bodies and X a set
of shared (coordinate free) points in R3 that imply incidences, i.e., points that occur in two or
more of the rigid bodies. The pair (X, c) is a valid collection of rigid bodies (short: collection),
if for i 6= j, ci contains at least one point in X not in c j , and ci ( X . We call c0 ⊂ c a covering
set of X if every point in X lies in at least one rigid body in c0 .
Proper-maximal: A rigid body ci is proper-maximal in X if there is no subset u of X with ci ( u (
X that represents a rigid body. That is, no such u is rigidified by the incidences within u and the
fixed distances within the bodies outside u.
Standard Collection of Rigid Bodies: The collection (X, c) is a standard collection of rigid bodies
if the following hold.

(a) No pair ci 6= c j intersects in more than two points.

(b) (c is complete in X ) All proper-maximal rigid bodies are in c.
(c) All ci are proper-maximal.
Overlap Graph: An overlap graph G(X, c) of a standard collection of rigid bodies (X, c) is a
weighted undirected graph whose vertices are the rigid bodies c j in c and whose edges rep-
resent incidences between pairs of rigid bodies. If an edge between a pair (ci , c j ) represents k
incidences, the weight w(k) assigned to the edge is the number of remaining dofs of c j after
fixing ci ’s position and orientation and resolving the k incidences between them.

Figure 7.5 shows an example where the Optimized Incidence Tree Parametrization Algorithm is
used on an example standard collection of 6 rigid bodies. These rigid bodies can be viewed as the
internal nodes in a DR-plan and the combined structure (bottom) as the parent of these nodes whose
solution is being recombined from the solutions of its children. The figure in the top right shows its
corresponding overlap graph. The figures in the middle show different spanning trees of the covering
set. The figures at the bottom show the two realizations of the possible eight.

7.2.3 Reconciling Conflicting Combinatorial Preprocessors

The work in [21] reconciles the matroid of Section 7.2.1 with the work in this section to get a single
matroid that can be used for an algorithm to solve both problems.
Given a collection of incident rigid bodies, a seam matroid is used to select a well-formed subset
of incidences (maximal independent set), which ensures a correspondence between each solution of
the original system and a solution to the numerically perturbed, well-formed system [19]. On the
other hand, the overlap graph, defined earlier, is used to select incidence tree parameterization that
minimizes the number of variables and hence algebraic complexity.
In [21], they show that while preprocessing, choosing parameters using seam graphs ignores
algebraic complexity and choosing parameters, so as to reduce complexity does not guarantee
well-formedness. They present a new algorithm, Optimal Well-Formed Incidence Selection Algo-
rithm, that combines the key elements of the Optimal Incidence Tree Algorithm [20] with the
Well-formed Incidences Algorithm [18] to guarantee both optimal algebraic complexity and well-
formedness. This greedy and efficient algorithm is the result of a careful construction and proof of
an independence-preserving map between the cycle matroid associated with the overlap graph and
the seam matroid underlies well-formed systems of incidences.
Geometric Constraint Decomposition: The General Case 195

7.3 Handling Under- and Over-Constrained Systems

Most of the content in Sections 7.1 and 7.2 considered independent systems (i.e., not over-
constrained.) Even more so, the focus was predominantly on well-constrained systems; when deal-
ing with under-constrained systems, the response was typically to operate only on the rigid compo-
nents. However, there are several interesting concerns regarding under- and over-constrained sys-
tems.
For under-constrained systems the question of which constraints to add to make the system
well constrained and what the constraint parameter values should be to guarantee a realization is
discussed Chapter 10.
As discussed in Chapter 1, an over-constraint is a dependency in the rigidity matroid. An n-
over-constrained system is a system with n over-constraints. Detecting the minimal set of edges
containing a 1-over-constrained system corresponds to finding a circuit in the rigidity matroid. As
such, there is a unique minimal 1-over-constrained subsystem in a 1-over-constrained system. For
an in-depth discussion on rigidity circuits, see [14].
Ref. [9] presents an algorithm for discovering this minimal subsystem in a 1-over-constrained
system in the 2D bar-joint case. This algorithm uses the DR-plan for efficient detection, and more-
over provides a method for quickly updating the DR-plan to reflect the dropped over-constraint.
The 1-over-constrained case arises naturally in practice; as a user incrementally adds constraints
to a system single over-constraints can be identified as they occur. However, when considering a
complete constraint graph, the k-over-constrained case can occur.

7.4 User Intervention in DR-Planning

In addition to mathematical properties that are important to satisfy, a useful DR-planner must also
be able to handle interaction with a user. In practice, DR-planners are used in live environments
where the designer is often tweaking and updating the geometric constraint system on which they
are working. Often beginning with some under-constrained sketch of the final product and building
up, they will frequently be adding and removing constraints. Additionally, the final product may be
inconsistently over-constrained, requiring changes before the system can be solved.
Whenever intervention is required, while offering guidance to the user the constraint solver [9]:
• Should not require that the end-user be familiar with the mathematics involved behind the
scenes.
• Should not artificially limit choices or include irrelevant choices.
• Should present intuitive and consistent choices that are unique in some well-defined manner.
One concern that is explicitly addressed in Section 7.1 and 7.2 is efficiency. As in all real-time appli-
cations, the user should wait as little as possible. Using optimal DR-plans, with quick recombination
schemes and methods of solving, reduces the impact of redecomposing and solving the system and
any subsystem when making changes.

7.4.1 Root Selection and Navigation

A rigid geometric constraint system may have more than one solution. In fact, it can be exponential
in the number of geometric primitives. Consider a 2D bar-joint system, whose constraint graph
196 Handbook of Geometric Constraint Systems Principles

is a 2-tree. At the placement of each vertex, there will be two positions that satisfy the distance
constraint, resulting in O(2|V | ) realizations. This, of course, does not only happen with 2-trees.
These non-congruent realizations are often called orientations. Choosing a specific realization may
not only be of interest to the user, but may also be necessary for reasonable computation times.
In practice, often when designing systems with real-world application, self-intersection is a
concern. Therefore, it may be natural to consider imposing that only non-intersection realizations
are returned [5]. However, finding non-intersecting results for even a simple polygon is NP-hard. So
this approach cannot be applied in general.
Another possible approach is to allow the user to offer additional over-constraints. This can
significantly reduce the number of acceptable solutions [5]. While finding an optimal DR-plan is
NP-hard in the presence of over-constraints, they can only improve the maximum fan-in since the
original well-constrained systems DR-plan can be used by default. However, requiring additional
work by the user may be undesirable.
Another option is to use heuristics to try to find the realization that most closely matches the user
input sketch [5]. As the realization is being computed and the solver chooses orientations, make the
choice that preserves the orientation in the sketch; if a point is on one side of an edge in the drawing,
choose the solution that matches this. However, the orientation as it appears in the sketch may not
have a real solution, in which case the algorithm must make an arbitrary choice.
A different approach is to instead involve the user in the recombination process. As the solver
finds roots for the subsystems, the user may choose at this time. This choice then limits the space
of real solutions as the subsystem is incorporated into its ancestors in the DR-plan. However, the
choice the user made may not allow for any real solutions, so the algorithm must be able to identify
the problematic subsystem.

7.4.2 Changing Constraint Parameters

A DR-plan is combinatorial, working with the underlying constraint graph. Therefore, a change in
the constraint parameters of a geometric constraint system has no effect on its optimal DR-plan.
Thus, the user is free to change these parameters without necessitating additional computation to
modify the DR-plan. However, the realization will change, and in real-time applications it may be
undesirable to have to re-solve the entire constraint system when changing a single constraint.
The structure of a DR-plan gives insight in how to eliminate repeated computation. The altered
constraint appears in some leaf (or leaves) of the DR-plan. Only the ancestors of this leaf need be
recalculated. So long as the subsystem in its own coordinate system is stored at each node in the
DR-plan data structure, these results can be reused.
However, there is the underlying assumption that the system is generic, otherwise there is no
guarantee that the nodes are truly rigid or that there is any realization at all. Therefore, this must be
checked upon user input.

7.4.3 Dynamic Maintenance

Dynamic maintenance refers to the adding and removing of entire constraints. Unlike changing the
values of parameters, this necessitates recomputation of the DR-plan. The challenges involved arise
from the discussion in Section 7.3. Additional consideration comes from the assumption that a DR-
plan already exists prior to the alterations, since this is happening in a live environment. Therefore,
reusing portions of the DR-plan may be an option to save computation.
References 197

7.5 Conclusion
This survey discusses DR-planning and overviews several important DR-planning algorithms (Sec-
tion 7.1.) Then, focusing on the primary use of DR-planning, i.e. realizing the graph, several dif-
ferent ways to further improve the time complexity were examined (Section 7.2.) Then, related
questions regarding how to handle under- and over-constraints were addressed (Section 7.3.) Fi-
nally, the software design perspective was considered, regarding real-world concerns of user input
and intervention (Section 7.4.)
The main goal of this survey was to motivate the importance of DR-planning. With its very
general definition, DR-planning can be applied to any geometric constraint system. The types of
primitives and constraints possible to use are varied, allowing for the modeling of a diverse cast of
applied problems. Thus, work in the field has far-reaching impact.
Of particular interest to the author for future research is isolating the relevant combinatorial
structure that allows for efficient DR-planning. What is the most minimal and still conceptually
useful structure that unifies the diverse group of constraint systems for which DR-planning is effi-
cient? Seemingly, the class must include those systems with an underlying abstract rigidity matroid,
and further there must be an efficient algorithm for detecting rigidity. Is this so? Must there be an
underlying sparsity matroid, as [3] requires? The author conjectures that even (3, 6)-sparse graphs
(which do not have and underlying sparsity matroid) may still admit a canonical DR-plan.

References
[1] Samy Ait-Aoudia, Roland Jegou, and Dominique Michelucci. Reduction of constraint sys-
tems. Compugraphics, 25:83–92, December 1993.
[2] Leonid Assur. Issledovanie ploskih sterznevyh mehanizmov s nizsimi parami s tocki zreniya
ih struktury i klassikacii. Izdat. Akad. Nauk SSSR, Edited by I. I. Artobolevski, 1952.
[3] T. Baker, M. Sitharam, M. Wang, and J. Willoughby. Optimal decomposition and recombi-
nation of isostatic geometric constraint systems for designing layered materials. Computer-
Aided Geometric Design, 40:1–25, 2015.
[4] William Bouma, Ioannis Fudos, Christoph Hoffmann, Jiazhen Cai, and Robert Paige. Geo-
metric constraint solver. Computer Aided Design, 27(6):487–501, 1995.
[5] Ioannis Fudos. Editable representations for 2D geometric design. Master’s thesis, Purdue
University, December 1993.
[6] Ioannis Fudos and Christoph M. Hoffmann. Correctness proof of a geometric constraint
solver. International Journal of Computational Geometry & Applications, 06(04):405–420,
1996.
[7] Christoph M. Hoffman, Andrew Lomonosov, and Meera Sitharam. Decomposition plans for
geometric constraint problems, part II: new algorithms. Journal of Symbolic Computation,
31(4):409–427, 2001.
[8] Christoph M. Hoffman, Andrew Lomonosov, and Meera Sitharam. Decomposition plans for
geometric constraint systems, part I: Performance measures for CAD. Journal of Symbolic
Computation, 31(4):367–408, 2001.
198 References

[9] Christoph M. Hoffmann, Meera Sitharam, and Bo Yuan. Making constraint solvers more
usable: overconstraint problem. Computer-Aided Design, 36(4):377–399, 2004.
[10] Donald J. Jacobs and Bruce Hendrickson. An algorithm for two-dimensional rigidity perco-
lation: The pebble game. Journal of Computational Physics, 137(2):346–365, 1997.
[11] Christophe Jermann, Gilles Trombettoni, Bertrand Neveu, and Pascal Mathis. Decomposition
of geometric constraint systems: A survey. International Journal of Computational Geometry
& Applications, 16(05n06):379–414, 2006.
[12] Andrew Lomonosov. Graph and combinatorial analysis for geometric constraint graphs.
PhD thesis, University of Florida, 2004.
[13] Anthony Nixon, Bernd Schulze, Adnan Sljoka, and Walter Whiteley. Symmetry adapted
Assur decompositions. Symmetry, 6(3):516, 2014.
[14] Brigitte Servatius. Planar Rigidity. PhD thesis, Syracuse University, 1987.
[15] Brigitte Servatius, Offer Shai, and Walter Whiteley. Combinatorial characterization of the
Assur graphs from engineering. European Journal of Combinatorics, 31(4):1091–1104, 2010.
Rigidity and Related Topics in Geometry.
[16] Offer Shai, Adnan Sljoka, and Walter Whiteley. Directed graphs, decompositions, and spatial
linkages. Discrete Applied Mathematics, 161(18):3028–3047, 2013.
[17] Meera Sitharam. Combinatorial approaches to geometric constraint solving: Problems,
progress, and directions. In Ravi Janardan, Michiel Smid, and Debasish Dutta, editors, Geo-
metric and Algorithmic Aspects of Computer-Aided Design and Manufacturing, volume 67 of
DIMACS Series in Discrete Mathematics and Theoretical Computer Science, chapter 5, pages
117–164. American Mathematical Society, 2005.
[18] Meera Sitharam. Well-formed systems of point incidences for resolving collections of rigid
bodies. International Journal of Computational Geometry & Applications, 16(05n06):591–
615, 2006.
[19] Meera Sitharam. Well-formed systems of point incidences for resolving collections of rigid
bodies. International Journal of Computational Geometry & Applications, 16(05n06):591–
615, 2006.
[20] Meera Sitharam, Jörg Peters, and Yong Zhou. Optimized parametrization of systems of inci-
dences between rigid bodies. J. Symb. Comput., 45(4):481–498, April 2010.
[21] Meera Sitharam, Yong Zhou, and Jörg Peters. Reconciling conflicting combinatorial prepro-
cessors for geometric constraint systems. International Journal of Computational Geometry
& Applications, 20(06):631–651, 2010.
 Part II

Distance Geometry, Real Algebraic

Geometry, and Configuration Spaces
Chapter 8
Dimensional and Universal Rigidities of
Bar Frameworks
A. Y. Alfakih
University of Windsor, Windsor, Ontario, Canada

CONTENTS
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
8.2 Stress Matrices and Gale Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
8.3 Dimensional Rigidity Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
8.4 Universal Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
8.4.1 Affine Motions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
8.4.2 Universal Rigidity Basic Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
8.4.3 Universal Rigidity for Special Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
8.5 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

8.1 Introduction
Suppose that we are given a configuration p = (p1 , . . . , pn ) of points in r-Euclidean space where a
specified subset of the inter-point distances is to be preserved. A natural question to ask is whether
there exists another configuration q that satisfies these specified distances. To put it differently,
we are interested in the problem of determining whether a given proper subset of the inter-point
distances in a configuration p uniquely determines p up to a rigid motion. This problem can be
conveniently described in terms of bar frameworks.
A bar framework (G, p) (or simply a framework∗ ) in dimension r is a connected simple (no loops
or multiple edges) graph G whose vertices are points p1 , . . . , pn in r-Euclidean space, and whose
edges are line segments corresponding to the specified inter-point distances to be preserved. Two
frameworks (G, p) and (G, q) are equivalent if the corresponding edges of (G, p) and (G, q) have the
same (Euclidean) length; and they are congruent if all corresponding inter-point distances of (G, p)
and (G, q) are the same. A given framework (G, p) is universally rigid if every framework that is
equivalent to (G, p) is in fact congruent to it. Therefore, a proper subset of the inter-point distances
in a configuration p uniquely determines p, up to a rigid motion, if and only if the corresponding
bar framework (G, p) is universally rigid.
Representing a point configuration by a Gram matrix renders the universal rigidity problem
amenable to semidefinite programming methodology [23]. In fact, many universal rigidity results are
∗ We are only interested in bar frameworks in this chapter.

201
202 Handbook of Geometric Constraint Systems Principles
4c c2 4c
HHH 
H 
 HH
c

 H
Hc 1c 2 c c3
1 3
(a) (b)
Figure 8.1
Two 2-dimensional bar frameworks in the plane. The edge {1, 3} in framework (b) is drawn as an
arc to make edges {1, 2} and {2, 3} visible. Framework (a) is not dimensionally rigid and has no
affine motions, while framework (b) is dimensionally rigid and has an affine motion.

direct application of known facts in semidefinite programming. In addition to being a fundamental

problem in distance geometry, the universal rigidity problem of bar frameworks has many important
applications in multidimensional scaling in statistics, molecular conformations in chemistry and in
ad hoc wireless sensor networks.
The notions of dimensional rigidity and affine motions turn out to be useful in the study
of universal rigidity. Framework (G, p) is r-dimensional if the points p1 , . . . , pn affinely span r-
dimensional Euclidean space. An r-dimensional framework (G, p), r ≤ (n − 2), is dimensionally
rigid if there is no s-dimensional framework (G, q), equivalent to (G, p), for any s ≥ r + 1. We say
that (G, p) has no affine motions if there is no framework (G, q), equivalent but not congruent, to
(G, p) such that qi = Api + b for all i = 1, . . . , n for some matrix A and vector b. Figure 8.1 depicts
two 2-dimensional frameworks in the plane. Framework (a) is not dimensionally rigid and has no
affine motion. On the other hand, framework (b) is dimensionally rigid and has an affine motion.
Both frameworks are not universally rigid.
Universal rigidity is characterized in terms of dimensional rigidity and affine motions as follows.
Theorem 8.1 (Alfakih [2]) An r-dimensional bar framework (G, p) on n vertices, n ≤ (n − 2), is
universally rigid if and only if (G, p) is both dimensionally rigid and has no affine motions.
Thus, the universal rigidity problem can be addressed by separately addressing its two easier
subproblems, namely dimensional rigidity and affine motions. For instance, sufficient conditions
for dimensional rigidity and sufficient conditions for the absence of affine motions can be combined
to establish sufficient conditions for universal rigidity.

8.2 Stress Matrices and Gale Matrices

The notion of a stress matrix is a useful tool in the study of dimensional and universal rigidities. An
equilibrium stress (or simply a stress) of bar framework (G, p) is a real-valued function ω on E(G)
such that: X
ωi j (pi − p j ) = 0 for all i = 1, . . . , n. (8.1)
j:{i, j}∈E(G)

Let ω be a stress of (G, p) and let E(G) denote the set of missing edges of G, i.e.,
E(G) = {{i, j} : i 6= j, {i, j} 6∈ E(G)}. (8.2)
Then the n × n matrix Ω where

 −ωi j if {i, j} ∈ E(G)
Ωi j = 0 if {i, j} ∈ E(G) (8.3)
 P if i = j
k:{i,k}∈E(G) ωik
Dimensional and Universal Rigidities of Bar Frameworks 203

is called the stress matrix associated with ω, or simply a stress matrix of (G, p). Let (G, p) be an
r-dimensional bar framework on n vertices in r-Euclidean space. The n × r matrix P whose ith row
is (pi )T , i.e.,  1 T 
(p )
P=
 .. 
(8.4)
. 
(pn )T
is called the configuration matrix of (G, p). Observe that P has full column rank since the points
p1 , . . . , pn affinely span the r-Euclidean space. Let e denote the n-vector of all 1’s, then it imme-
diately follows from (8.3) that e and the columns of P are in the null space of Ω. Consequently, a
symmetric matrix Ω is a stress matrix of (G, p) if and only if

ΩP = 0, Ωe = 0 and Ωi j = 0 for all {i, j} ∈ E(G). (8.5)

Hence, the rank of Ω is ≤ n − r − 1. In addition to being interesting in their own right, the following
two theorems are frequently used in the proofs of subsequent results.

Theorem 8.2 (Alfakih [4]) Let (G, p) be an r-dimensional bar framework on n vertices, r ≤ (n −
2), and let Ω be a non-zero positive semidefinite stress matrix of (G, p). Then Ω is a stress matrix
for any bar framework (G, q) that is equivalent to (G, p).

Theorem 8.3 (Alfakih [4]) Let (G, p) be an r-dimensional bar framework on n vertices, r ≤ (n −
2). Then (G, p) admits a non-zero positive semidefinite stress matrix Ω if and only if there exists no
(n − 1)-dimensional bar framework (G, q) that is equivalent to (G, p).

Note that if an r-dimensional framework (G, p) on n vertices, r ≤ (n − 2), does not have a
non-zero positive semidefinite stress matrix, then it follows from Theorem 8.3 that (G, p) is not
dimensionally rigid. As a result, Theorem 8.3 can be used to provide a necessary condition for
dimensional rigidity and, consequently, universal rigidity.
 T  [16, 19]. Any n × (n −
Stress matrices are closely related to Gale matrices (or Gale transform)
P
r − 1) matrix Z whose columns form a basis of the null space of is called a Gale matrix of
eT
(G, p). Sometimes it is more convenient to define Gale transform, as is usually done in the polytope
literature [16, 19]. Let (zi )T be the ith row of Z, then zi is called a Gale transform of pi . The next
theorem establishes the relationship between Gale matrices and stress matrices.
Theorem 8.4 (Alfakih [3]) Let (G, p) be an r-dimensional bar framework on n vertices and let Z
be a Gale matrix of (G, p). Then Ω is a stress matrix of (G, p) if and only if

Ω = ZΨZ T and Ωi j = 0 for all {i, j} ∈ E(G),

where Ψ is a symmetric matrix of order n − r − 1.

Theorem 8.4 shows that stress matrix Ω can be decomposed into a geometric part (Z), and a combi-
natorial one (Ψ).
Gale matrices have useful properties when the points p1 , . . . , pn are in general position. We say
that points p1 , . . . , pn in r-Euclidean space are in general position if every r + 1 of these points are
affinely independent. For instance, points in the plane are in general position if no three of them are
collinear.

Lemma 8.1 [6]

Let (G, p) be an r-dimensional bar framework and let z1 , . . . , zn be Gale transforms of p1 , . . . , pn .
Let J be a subset of {1, . . . , n} of cardinality r + 1, and let the set of points {pi : i ∈ J} be affinely
204 Handbook of Geometric Constraint Systems Principles

b X
3 X
XXX
XX
4 XXX
2 b b b 5
X
X

 

1 b
 
Figure 8.2
A dimensionally rigid 2-dimensional bar framework on 5 vertices which does not admit a positive
semidefinite stress matrix of rank 2.

¯ is linearly independent where J¯ = {1, . . . , n}\J, where \ denotes

independent. Then the set {zi : i ∈ J}
the set theoretic difference.

Hence, if Z is a Gale matrix of a framework in general position, then every submatrix of Z of

order n − r − 1 is nonsingular.

Example 35 Consider the 2-dimensional bar framework (G, p) on 5 vertices depicted in Figure 8.2.
The set of missing edges is given by E(G) = { {1, 4}, {3, 4}, {2, 5} }. Moreover, the configuration
matrix and a Gale matrix of (G, p) are given by:
   
−1 −1 1 0
 −1 0   −2 3 
   
P =  −1
 1  and Z =  1
  0 .
 0 0   0 −4 
3 0 0 1
   
−2 0
Notice that the points p1 , p3 , p5 are affinely independent and the set {z2 = , z4 = }
3 −4
is linearly independent. To find a stress matrix Ω, we have to find a 2 × 2 symmetric matrix Ψ such
that:
(z1 )T Ψz4 = 0, (z3 )T Ψz4 = 0, and (z2 )T Ψz5 = 0.
 
1
   −2 
1 0 T
  
Hence, Ψ = . Accordingly, Ω = ZΨZ =   1  1 −2 1 0 0 . This should

0 0  0 
0
come as now surprise, since the only edges of (G, p) with non-zero stresses are {1, 2}, {2, 3} and
{1, 3}.

8.3 Dimensional Rigidity Results

This section presents sufficient and necessary conditions for dimensional rigidity. These conditions
are expressed in terms of stress matrices or, equivalently, Gale transform.

Theorem 8.5 (Alfakih [2]) Let (G, p) be an r-dimensional bar framework on n vertices, r ≤
(n − 2). If (G, p) admits a positive semidefinite stress matrix Ω with rank n − r − 1, then (G, p)
is dimensionally rigid.
Dimensional and Universal Rigidities of Bar Frameworks 205

In fact, Theorem 8.5 in [2] is stated in terms of Gale transform as given next. Obviously, these two
statements are equivalent.

Theorem 8.6 Let (G, p) be an r-dimensional bar framework on n vertices, r ≤ (n − 2). Let zi be a
Gale transform of pi for i = 1, . . . , n. If there exists a positive definite symmetric matrix Ψ such that:

(zi )T Ψz j = 0 for all {i, j} ∈ E(G),

then (G, p) is dimensionally rigid.

Unfortunately, the sufficient condition in Theorem 8.5 is not necessary. Consider the bar frame-
work depicted in Figure 8.2. Even though this framework is dimensionally rigid (in fact, it is also
universally rigid), as shown in Example 35, this framework admits a positive semidefinite stress
matrix of rank 1. But in this case, n − r − 1 = 2.
A necessary and sufficient condition for dimensional rigidity was given by Connelly and Gortler
[15] based on the Borwein–Wolkowicz facial reduction algorithm [10, 11]. In particular, they
showed that an r-dimensional bar framework (G, p) is dimensionally rigid if and only if (G, p) ad-
mits a sequence of quasi-stress matrices such that: (i) the ranks of these matrices add up to n − r − 1
and (ii) these matrices are positive semidefinite on a chain of nested subspaces.
A symmetric matrix Ω is called a quasi-stress matrix of a (G, p) if and only if

PT ΩP = 0, Ωe = 0 and Ωi j = 0 for all {i, j} ∈ E(G), (8.6)

where P is the configuration matrix of (G, p). Observe that every stress matrix is a quasi-stress
matrix. On the other hand, if a quasi-stress matrix is positive semidefinite, then it is a stress matrix.
The following theorem is a refined version of Connelly-Gortler result. The positive semidefi-
niteness of a symmetric matrix A is denoted by A  0, and the column space and the null space of A
are denoted by R(A), and N (A) respectively.
Theorem 8.7 ([1]) Let (G, p) be an r-dimensional bar framework on n vertices in r-Euclidean
space, r ≤ (n − 2). Then (G, p) is dimensionally rigid if and only if there exist nonzero quasi-stress
matrices Ω0 , Ω1 , . . . , Ωk , for some k ≤ n − r − 2, such that:

(a) Ω0  0, U1T Ω1 U1  0, . . . , UkT Ωk Uk  0,

(b) rank Ω0 + rank (U1T Ω1 U1 ) + · · · + rank (UkT Ωk Uk ) = n − r − 1,
(c) PT Ω1 ρ1 = 0, . . . , PT Ωk ρk = 0,

 ρ01 , U
where 1 , . . . , Uk and ξ1 , . . . , ξk are full column rank matrices defined as follows. R(ρ1 ) =
Ω
N ( PT ). R(ξi ) = N (ρiT Ωi ρi ), Ui = [P ρi ] for i = 1, . . . , k and ρi+1 = ρi ξi for all i = 1, . . . , k −1.
eT

Example 36 Consider the bar framework (G, p) depicted in Figure 8.2. Recall from Example 35
that a stress matrix Ω0 of (G, p) is given by:
 
1
 −2 
0
  
Ω =  1  1 −2 1 0 0 .

 0 
0
206 Handbook of Geometric Constraint Systems Principles
 
−1 −1 1
 −1 0 1 
T
 
Thus, ρ1 = [1 1 1 − 4 1] and hence U1 = [P ρ1 ] =  −1
 1 1 
.
 0 0 −4 
3 0 1
Let
 
3 −6 0 0 3  
 −6 30 0 −24 0  0 0 0
Ω1 =  T 1
 
 0 0 −3 0 3   . Hence, U1 Ω U1 = 0 0
 0 .
 0 −24 0 32 −8  0 0 800
3 0 3 −8 2

Therefore, PT Ω1 P = 0, Ω1 e = 0 and hence Ω1 is a quasi-stress matrix. Furthermore, PT Ω1 ρ1 = 0,

U1T Ω1 U1  0 and rank Ω0 + rank (U1T Ω1 U1 ) = 2. Therefore, framework (G, p) is dimensionally
rigid.

8.4 Universal Rigidity

This section begins with a brief summary of affine motion results. These results are then combined
with dimensional rigidity results in the previous section to obtain sufficient conditions for universal
rigidity. The section ends with several universal rigidity results for some special classes of graphs,
namely r-lateral graphs, chordal graphs and complete bipartite graphs.

8.4.1 Affine Motions

Affine motions of a bar framework (G, p) can be characterized in terms of p1 , . . . , pn and the edges
of G.

Lemma 8.2 [14]

Let (G, p) be an r-dimensional bar framework on n vertices in r-Euclidean space, r ≤ (n − 2). Then
(G, p) has an affine motion if and only if there exists a nonzero symmetric r × r matrix Φ such that:

(pi − p j )T Φ(pi − p j ) = 0 for all {i, j} ∈ E(G),

i.e., (G, p) has an affine motion if and only if its edges directions lie on a conic at infinity.

Equivalently, affine motions can be characterized in terms of Gale matrix and the missing edges
of G. Let E i j be the n × n symmetric matric with 1’s in the (i, j)th and ( j, i)th entries and 0’s
elsewhere.

Lemma 8.3 [3]

Let (G, p) be an r-dimensional bar framework on n vertices in r-Euclidean space, r ≤ (n − 2). Let
Z be a Gale matrix of (G, p). Then (G, p) has an affine motion if and only if there exists a nonzero
y = (yi j ) such that:
X
yi j E i j Z = eξ T ,
{i, j}∈E(G)

for some ξ in (n − r − 1)-Euclidean space.

Dimensional and Universal Rigidities of Bar Frameworks 207
1 3 1 2
Example 37 Consider  framework (b)  depicted  case p − p = 2(p − p ) =
 in Figure8.1. In this
−2 0 0 1
2(p2 − p3 ) = , while p2 − p4 = . Let Φ = . Then (pi − p j )T Φ(pi − p j ) = 0
0 −1 1 0
for all {i, j} ∈ E(G). Hence, it follows from Lemma 8.2 that this framework has an affine motion.
On the other hand, the set of missing edges is given by E(G) = { {1, 4}, {3, 4} } and a Gale
matrix is Z = [1 − 2 1 0]T . Thus
    
0 0 0 y14 1 0
  −2  
 0 0 0 0  0
(y14 E 14 + y34 E 34 )Z = 
   
 0 0 0 y34   1  = 
 = ξ1 e
0 
y14 0 y34 0 0 y14 + y34
has the solution ξ1 = 0 and y14 = −y34 = 1. Hence, again it follows from Lemma 8.3 that this
framework has an affine motion.
We say that (G, p) is generic if the coordinates of p1 , . . . , pn are algebraically independent over
the integers, i.e., if these coordinates do not satisfy any nonzero polynomial with integer coefficients.
The following lemmas present sufficient conditions for the absence of affine motions.

Lemma 8.4 [14]

Let (G, p) be an r-dimensional bar framework on n vertices in r-Euclidean space, r ≤ (n − 2).
Assume that: (i) (G, p) is generic and (ii) each vertex of G has degree ≥ r. Then (G, p) has no affine
motion.

An alternative proof of Lemma 8.4 is given in [21]. Next, the generic assumption is replaced
with the assumption of points in general position.

Lemma 8.5 [8]

Let (G, p) be an r-dimensional bar framework on n vertices in r-Euclidean space, r ≤ (n − 2).
Assume that: (i) (G, p) admits a stress matrix Ω with rank n − r − 1 and (ii) (G, p) is in general
position. Then (G, p) has no affine motion.

The assumptions in Lemma 8.5 are weakened in the following lemma.

Lemma 8.6 [6]

Let (G, p) be an r-dimensional bar framework on n vertices in r-Euclidean space, r ≤ (n − 2).
Assume that: (i) (G, p) admits a stress matrix Ω with rank n − r − 1 and (ii) for each vertex i, the set
{pi } ∪ {p j : {i, j} ∈ E(G)} affinely span r-Euclidean space. Then (G, p) has no affine motion.

The following lemma is an immediate consequence of Lemma 8.6.

Lemma 8.7 [6]

Let (G, p) be an r-dimensional bar framework on n vertices in r-Euclidean space, r ≤ (n − 2).
Assume that: (i) (G, p) admits a stress matrix Ω with rank n − r − 1 and (ii) for each vertex i, the
points {pi } ∪ {p j : {i, j} ∈ E(G)} are in general position. Then (G, p) has no affine motion.

8.4.2 Universal Rigidity Basic Results

The following theorem is an immediate consequence of Theorems 8.1 and 8.5.
Theorem 8.8 (Connelly [12, 13], Alfakih [2]) Let (G, p) be an r-dimensional bar framework on n
vertices, r ≤ (n − 2). If the following two conditions hold:
208 Handbook of Geometric Constraint Systems Principles

(a) (G, p) admits a positive semidefinite stress matrix Ω with rank n − r − 1,

(b) (G, p) has no affine motion,
then (G, p) is universally rigid.

Universal rigidity of bar frameworks has a complete characterization under the generic assump-
tion.

Theorem 8.9 (Alfakih [3], Connelly [14], Gortler and Thurston [18]) Let (G, p) be an r-dimensional
generic bar framework on n vertices in r-Euclidean space, r ≤ (n − 2). Then (G, p) is universally
rigid if and only if it admits a positive semidefinite stress matrix Ω with rank n − r − 1,

The sufficiency part of Theorem 8.9 was proved independently in [12] and [3] while the necessity
part was conjectured in [3] and proved in [18].
If the generic assumption is replaced by the weaker assumption of points in general position,
then only sufficient conditions are known. These conditions, which are not necessary, follow from
Theorem 8.8 and Lemmas 8.5, 8.6, and 8.7.

Theorem 8.10 (Alfakih and Ye [8]) Let (G, p) be an r-dimensional bar framework on n vertices
in r-Euclidean space, r ≤ (n − 2). If the following two conditions hold:
(a) (G, p) admits a positive semidefinite stress matrix Ω with rank n − r − 1,
(b) (G, p) is in general position.
Then (G, p) is universally rigid.

Theorem 8.11 (Alfakih and Nguyen [6]) Let (G, p) be an r-dimensional bar framework on n ver-
tices in r-Euclidean space, r ≤ (n − 2). If the following two conditions hold:
(a) (G, p) admits a positive semidefinite stress matrix Ω with rank n − r − 1,
(b) For each vertex i, the set {pi } ∪ {p j : {i, j} ∈ E(G)} affinely span r-Euclidean space.
Then (G, p) is universally rigid.

Theorem 8.12 (Alfakih and Nguyen [6]) Let (G, p) be an r-dimensional bar framework on n ver-
tices in r-Euclidean space, r ≤ (n − 2). If the following two conditions hold:
(a) (G, p) admits a positive semidefinite stress matrix Ω with rank n − r − 1,
(b) For each vertex i, the points {pi } ∪ {p j : {i, j} ∈ E(G)} are in general position.
Then (G, p) is universally rigid.

8.4.3 Universal Rigidity for Special Graphs

We focus in this section on universal rigidity results for r-lateration graphs, chordal graphs and
complete bipartite graphs. A graph G on n vertices is called an r-lateration graph if there exists a
permutation π of the vertices of G such that: (i) the first r vertices, π(1), . . . , π(r) induce a clique
in G and (ii) each remaining vertex π( j), for j = r + 1, . . . , n, is adjacent to exactly r vertices in
{π(1), . . . , π( j − 1)}. It is known [24] that an r-dimensional bar framework (G, p) in general posi-
tion in r-Euclidean space, where G is an (r + 1)-lateration graph is universally rigid. It turns out that
more can be said about such frameworks.
Dimensional and Universal Rigidities of Bar Frameworks 209

Theorem 8.13 (Alfakih et al [7]) Let (G, p) be an r-dimensional bar framework on n vertices in
general position in r-Euclidean space, r ≤ (n − 2), where G is an (r + 1)-lateration graph. Then
(G, p) admits a positive semidefinite stress matrix Ω of rank n − r − 1.
The proof of Theorem 8.13 is constructive. The desired stress matrix is obtained via an algorithm
which starts with the matrix ZZ T , where Z is a Gale matrix, and iteratively zeros out the entries
corresponding to the missing edges of G.
A graph G is chordal if it has no induced cycle of length ≥ 4. Chordal graphs have many well-
known characterizations [9, 17]. An incomplete graph G is k-vertex connected if G has at least k + 2
vertices and the deletion of any k − 1 vertices leaves G connected.
Theorem 8.14 (Alfakih [5]) Let (G, p) be an r-dimensional bar framework on n vertices in gen-
eral position in r-Euclidean space, r ≤ (n − 2), where G is a chordal graph. Then the following
statements are equivalent:
(a) (G, p) is universally rigid,
(b) G is (r + 1)-vertex connected,
(c) (G, p) admits a positive semidefinite stress matrix Ω of rank n − r − 1.
The proof of Theorem 8.14 is based on the characterization of chordal graphs in terms of perfect
elimination order. Let G1 = (V1 , E1 ) and G2 = (V2 , E2 ) be two graphs such that V1 ∩ V2 induces a
clique in both G1 and G2 . The clique sum of G1 and G2 is the graph G = (V1 ∪V2 , E1 ∪ E2 ). The next
theorem easily follows from the clique-sum characterization of chordal graphs, namely a graph is
chordal if and only if it is the clique sum of cliques.
Theorem 8.15 (Ratmanski [22]) Let (G1 , p1 ) and (G2 , p2 ) be, respectively, two r1 and r2 -
dimensional universally rigid bar frameworks in general position in r-Euclidean space. Further,
let (G, p) be the bar framework such that V (G) = V (G1 ) ∪ V (G2 ) and E(G) = E(G1 ) ∪ E(G2 ).
Then (G, p) is universally rigid if and only if V (G1 ) ∩V (G2 ) has at least r + 1 vertices. Here, V (G)
denotes the vertex set of a graph.
The complete bipartite graph Km,n is the graph whose nodes can be partitioned into two subsets
V1 and V2 with m and n, nodes, respectively such that: (i) every node in V1 is adjacent to every
node in V2 and (ii) no two nodes in V1 or in V2 are adjacent. Thus Km,n has exactly mn edges. The
following theorem is a characterization of 1-dimensional bar frameworks on the line.
Theorem 8.16 (Jordán and Nguyen [20]) Let (Km,n , p ∪ q) be a 1-dimensional complete bipartite
bar framework on at least three vertices. If
(a) p1 < · · · < pm < q1 < · · · < qn , or
(b) p1 < · · · < pk < q1 < · · · < qn < pk+1 < · · · < pm .
Then (Km,n , p ∪ q) is not universally rigid. Otherwise, if Conditions 1 and 2 do not hold, then
(Km,n , p ∪ q) is universally rigid.

8.5 Glossary
Bar Framework: A bar framework (G, p) in dimension r is a connected simple graph G whose
vertices are points p1 , . . . , pn in r-dimensional Euclidean space and whose edges are line seg-
ment between pairs of these points.
210 Handbook of Geometric Constraint Systems Principles

Equivalent Frameworks: Bar frameworks (G, p) and (G, q) are equivalent if ||pi − p j || = ||qi −
q j || for all edges {i, j} of graph G.
Congruent Frameworks: Bar frameworks (G, p) and (G, q) are congruent if ||pi − p j || = ||qi −
q j || for all i, j = 1, . . . , n.
Universal Rigidity: Bar framework (G, p) is universally rigid if every framework (G, q) that is
equivalent to (G, p) is in fact congruent to it.

Dimensional Rigidity: An r-dimensional framework (G, p) is dimensionally rigid if there is no

s-dimensional framework (G, q), equivalent to (G, p), for any s ≥ r + 1.
Affine Motion: Bar framework (G, p) has no affine motions if there is no framework (G, q), equiv-
alent but not congruent, to (G, p) such that qi = Api + b for all i = 1, . . . , n for some matrix A
and vector b.
Configuration matrix: The configuration matrix P of an r-dimensional bar framework in dimen-
sion r is the n × r matrix whose its ith row is (pi )T .
Stress Matrix: A symmetric matrix Ω is a stress matrix of a bar framework (G, p) if and only if
ΩP = 0, Ωe = 0 and Ωi j = 0 for each missing edge {i, j} of G, where P is the configuration
matrix of (G, p) and e is the vector of all 1’s.
Quasi-Stress Matrix: A symmetric matrix Ω is a quasi-stress matrix of a bar framework (G, p) if
and only if PT ΩP = 0, Ωe = 0 and Ωi j = 0 for each missing edge {i, j} of G.
Gale matrix: A Gale matrix of an r-dimensional
 T  bar framework is any matrix Z whose columns
P
form a basis of the null space of .
eT

Points in General Position: points p1 , . . . , pn in r-Euclidean space are in general position if every
r + 1 of these points are affinely independent.

Generic Points: points p1 , . . . , pn are generic if their coordinates are algebraically independent
over the integers.
r-lateration Graph: A graph G is an r-lateration graph if there exists a permutation π of the ver-
tices of G such that: (i) the first r vertices, π(1), . . . , π(r) induce a clique in G and (ii) each
remaining vertex π( j) is adjacent to exactly r vertices in {π(1), . . . , π( j − 1)}.
Chordal Graph: A graph G is chordal if each cycle of G of length ≥ 4 has a chord.
k-vertex Connected Graphs: An graph G is k-vertex connected iff G is the complete graph on
k + 1 vertices or G has at least k + 2 vertices and the deletion of any k − 1 vertices leaves G
connected

Complete Bipartite Graph: A complete bipartite graph is a graph whose nodes are partitioned
into two subsets V1 and V2 such that: (i) every node in V1 is adjacent to every node in V2 and (ii)
no two nodes in V1 or in V2 are adjacent.

Acknowledgment: Research partially supported by the Natural Sciences and Engineering Research
Council of Canada.
References 211

References
[1] A. Y. Alfakih. On Farkas lemma and dimensional rigidity of bar frameworks.
arXiv/1405.2301.
[2] A. Y. Alfakih. On dimensional rigidity of bar-and-joint frameworks. Discrete Appl. Math.,
155:1244–1253, 2007.
[3] A. Y. Alfakih. On the universal rigidity of generic bar frameworks. Contrib. Disc. Math.,
5:7–17, 2010.
[4] A. Y. Alfakih. On bar frameworks, stress matrices and semidefinite programming. Math.
Program. Ser. B, 129:113–128, 2011.
[5] A. Y. Alfakih. On stress matrices of chordal bar frameworks in general positions, 2012.
arXiv/1205.3990.
[6] A. Y. Alfakih and V.-H. Nyugen. On affine motions and universal rigidity of tensegrity frame-
works. Linear Algebra Appl., 439:3134–3147, 2013.
[7] A. Y. Alfakih, N. Taheri, and Y. Ye. On stress matrices of (d + 1)-lateration frameworks in
general position. Math. Program., 137:1–17, 2013.
[8] A. Y. Alfakih and Y. Ye. On affine motions and bar frameworks in general positions. Linear
Algebra Appl., 438:31–36, 2013.
[9] J. R. S. Blair and B. Peyton. An introduction to chordal graphs and clique trees. In J. A.
George, J. R. Gilbert, and J. W-H. Liu, editors, Graph Theory and Sparse Matrix Computation,
IMA Vol. Math. Appl.,56, pages 1–29. Springer, New York, 1993.
[10] J. M. Borwein and H. Wolkowicz. Facial reduction for a non-convex programming problem.
J. Austral. Math. Soc., Ser. A, 30:369–380, 1981.
[11] J. M. Borwein and H. Wolkowicz. Regularizing the abstract convex program. J. Math. Anal.
Appl., 83:495–530, 1981.
[12] R. Connelly. Rigidity and energy. Invent. Math, 66:11–33, 1982.
[13] R. Connelly. Tensegrity structures: Why are they stable? In M. F. Thorpe and P. M. Duxbury,
editors, Rigidity theory and applications, pages 47–54. Kluwer Academic/Plenum Publishers,
1999.
[14] R. Connelly. Generic global rigidity. Discrete Comput. Geom., 33:549–563, 2005.
[15] R. Connelly and S. J. Gortler. Iterative universal rigidity. Technical report,
arXiv:1401.7029v1, 2014.
[16] D. Gale. Neighboring vertices on a convex polyhedron. In Linear inequalities and related
system, pages 255–263. Princeton University Press, 1956.
[17] M. C. Golumbic. Algorithmic graph theory and perfect graphs. Ann. Disc. Math 57, Elsevier,
2004.
[18] S. J. Gortler and D. P. Thurston. Characterizing the universal rigidity of generic frameworks.
Discrete Comput. Geom., 51:1017–1036, 2014.
[19] B. Grünbaum. Convex polytopes. John Wiley & Sons, 1967.
212 References

[20] T. Jordán and V-H. Nguyen. On universally rigid frameworks on the line. Technical report,
Egerváry Research Group, 2012.
[21] M. Laurent and A. Varvitsiotis. Positive semidefinite matrix completion, universal rigidity
and the strong Arnold property. Linear Algebra Appl., 452:292–317, 2014.
[22] K. Ratmanski. Universally rigid framework attachments. Technical report, arXiv: 1011-
4094v2, 2010.
[23] H. Wolkowicz, R. Saigal, and L. Vandenberghe, editors. Handbook of Semidefinite Program-
ming. Theory, Algorithms and Applications. Kluwer Academic Publishers, Boston MA, 2000.
[24] Z. Zhu, A. M-C So, and Y. Ye. Universal rigidity: Towards accurate and efficient localization
of wireless networks, 2010. Proc. IEEE INFOCOM.
Chapter 9
Computations of Metric/Cut Polyhedra
and Their Relatives
Mathieu Dutour Sikirić
Bos̆ković Institute, Zagreb

Michel-Marie Deza
École Normale Supérieure, Paris

Elena I. Deza
Moscow State Pedagogical University

CONTENTS
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
9.2 Metric and Cut Cones and Polytopes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
9.3 Hypermetric Cone and Hypermetric Polytope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
9.4 Cut and Metric Polytopes of Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
9.5 Quasi-Semimetric Polyhedra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
9.6 Partial Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
9.7 Supermetric and Hemimetric Cones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
9.8 Software Computations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228

9.1 Introduction
Here we consider cones and polytopes, the points of which are various specifications, analogs and
generalisations of finite semimetrics, i.e., for some finite set X (usually, X = {1, . . . , n}), the sym-
metric functions d = ((di j )) : X 2 → R≥0 with dii = 0 and triangle inequalities dik + dk j − di j ≥ 0
for i, j, k ∈ X. We also will call a finite semimetric n-points semimetric, or distance matrix.
A semimetric is called metric if di j > 0 for i 6= j, and (X, d) is called metric space. Metrics
are ubiquitous in mathematics; the polyhedra considered here are important in analysis, measure
theory, geometry of numbers, combinatorics, optimization, etc. Given two metric spaces (X, d) and
(X 0 , d 0 ), we say that (X, d) is isometrically embeddable into (or is a metric subspace of) (X 0 , d 0 ) if
there exists a mapping φ : X → X 0 such that d(x, y) = d(φ (x), φ (y)) for all x, y ∈ X.
Given 1 ≤ p ≤ ∞ and integer m ≥ 1, the l m m
p -space is the metric space (R , d p (x, y) = kx − yk p ),
where the l p -norm is

213
214 Handbook of Geometric Constraint Systems Principles

  1p
X
kxk p =  xip  for finite p and , kxk∞ = max xi .
1≤i≤m
1≤i≤m

A metric d is called l p -metric if (X, d) is a metric subspace of the l m p -space, for some X ⊂ R . A
m

metric d on X is (cf. [13]) l p -metric if and only if its restriction on any finite subset of X is l p -metric.
The unit ball Bmp = {x ∈ Rm : kxk p ≤ 1} is m-hyperoctahedron, m-hypercube [−1, 1]m , m-sphere,
m-spheroid, if p = 1, ∞, 2, otherwise. For any p > 1, it holds Bm m m 2
1 ⊆ B p ⊆ B∞ , but l p -spaces are
isometric.
The most prominent is l2 -metric, called also Euclidean (or Pythagorean, as-the-crow-flies, bee-
line) distance. The squared n-points l2 -semimetrics form the convex cone EDMn of Euclidean Dis-
tance Matrices ((d22 ))n , introduced in Chapter 1 and discussed further in Chapters 8 and 10.
Every l2 -metric is a l p -metric for any p ≥ 1, while the n-points l∞ -semimetrics form a convex
cone, which coincides with the cone (called metric cone and denoted METn ) of all n-points semi-
n−1 -space by n points (d , . . . , d
metrics. Indeed, any n-points semimetric d is realized in the l∞ 1i n−1 i )
for i = 1, . . . , n.
The 2nd by prominence is l1 -metric, also known as taxicab, rectilinear, city block or Manhattan
distance. Every l p -metric with 1 < p ≤ 2 is ([51]) a l1 -metric. The n-points l1 -semimetrics form
a convex cone, called cut cone and denoted CUTn . In fact, it is [1, 4] generated by 2n−1 − 1 cut
semimetrics δ (S) = δ (S) = ((di j )), defined, for any subset S ⊂ {1, . . . , n}, by di j = 1 if |{i, j} ∩ S| =
1 and di j = 0, otherwise.
Polyhedral (unlike to EDMn ) cones CUTn and (defined by 3 n3 triangle inequalities) METn

are the central prototypes for polyhedra considered here. We will also consider corresponding poly-
topes, say, CUT Pn and MET Pn , since they have much more symmetries and so are easier to describe.
The cut cone and polytope also admit [3, 5] a characterization in terms of measure spaces:
d ∈ CUTn (or d ∈ CUT Pn ) if and only if there exist a measure (respectively, a probability) space
(Ω, A, µ) and n events A1 , . . . , An ∈ A such that d(i, j) = µ(Ai ∆A j ) for all 1 ≤ i, j ≤ n.
Determining membership in CUTn is [6] NP-complete problem. Also, the connection of dual
CUTn and METn to multi-commodity flows was given in [6].
Under a linear mapping d → c(d) with cii = d1n for 1 ≤ i < n and ci j = 12 (din + d jn − di j ) for
1 ≤ i < j < n, we get the correlation cone c(CUTn ) and boolean quadric polytope c(CUT Pn ), which
are important in mathematical physics, quantum logic and optimization, say, in the Boole problem
and unconstrained quadratic {0, 1}-programming problem.
Connection of CUTn P to geometry of numbers is given by the hypermetrics, i.e., distance ma-
trices d = ((di j ))n with 1≤i, j≤n bi b j di j ≤ 0 for any (b1 , . . . , bn ) ∈ Zn . Clearly, permutations of
b = (1, 1, −1, 0, . . . , 0) give that d is an n-points semimetric. Any l1 -metric is [54] a hypermetric.
The hypermetrics are [2] exactly squared l2 -metrics of the constellations in Euclidean lattices, i.e.,
vertex sets of Delaunay polytopes.
So, l1 -metrics are l22 -metrics, while generalizing l2 -metrics. The study of l1 -metrics meets ongo-
ing (esp., in Riemannian geometry, real analysis, approximation theory, statistics) process: to extend
many prominent l2 -defined notions and results on more general l1 -setting.
The main connection of CUTn to combinatorics was given in [3]: the rational-valued l1 -metrics
d are exactly those, for which λ d isometrically embeds into the path-metric of an N-cube K2N for
some integers N, λ ≥ 1.
In this chapter, we collect known computations of cut and metric polyhedra and their relatives
(hypermetric, quasi-metric, and m-hemimetric polyhedra) and generalizations of such polyhedra on
general (instead of Kn ) connected finite graphs. Such polyhedra over graphs are related to maximum-
cut problem, which asks to find a cut of maximum weight in a graph. It is a prominent problem in
combinatorial optimization with many applications.
Computations of Metric/Cut Polyhedra and Their Relatives 215

In Chapter 8 a related viewpoint is considered for metric rigidities. There the distances are
considered as faces of the cone of positive semidefinite forms. Our viewpoint is different since
in the case of hypermetrics we consider positive definite matrices, not just positive semidefinite
matrices. Also the algebraic structure is different since the lattice structures makes the relevant
cones polyhedral.
Quasi-metrics, failing the condition di j = d ji for metrics, are encountered, for example, as quasi-
distances on one-way routes, mountain walks or rivers with quick flow. Quasi-metrics and oriented
cuts are used in semantics of computations and in computational geometry.
Partial metrics, where a weight dii ≥ 0 is attributed to each point, are another generalization of
metrics, which are equivalent to special quasi-metrics ((di j − dii )). Hemimetrics/supermetrics are
a generalization of metrics where instead of axiomatizing length, we axiomatize area and volume.
For each of those generalizations, there is an analog of triangle inequality and corresponding cones.
The polytopal viewpoint, which is pursued within this chapter, can be useful for rigidity ques-
tions. Whenever a metric (or hypermetric, or any generalization considered here) in a cone C is
incident to an inequality defining C, the metric has less degree of freedom left, i.e., in other words,
it is more rigid. This rigidity viewpoint is used in geometry of numbers for Delaunay polytopes: the
extreme rays of the hypermetric cone correspond to rigid Delaunay polytopes. This viewpoint could
be useful for other metric cones.

9.2 Metric and Cut Cones and Polytopes

Metric cone METn : the set of all semimetrics on n points, i.e., the symmetric functions d :
{1, . . . , n}2 → R satisfying all dii = 0 and all triangle inequalities

dik ≤ di j + d jk .

Metric polytope MET Pn : the set of all d ∈ METn satisfying, in addition, all perimeter inequalities

dik + di j + d jk ≤ 2.

Cut semimetric δS for a set S ⊆ {1, . . . , n}: a vector defined as


1 if |S ∩ {i, j}| = 1,
δS (i, j) =
0 otherwise.

Clearly, δS = δS , and δS is the adjacency matrix of a cut subgraph of Kn .

Cut cone CUTn : the positive span of all 2n−1 − 1 non-zero cut semimetrics.
Cut polytope CUT Pn : the convex hull of all 2n−1 cut semimetrics.
m-multicut δS1 ,...,Sm for a partition ∪m i=1 Si = {1, . . . , n}: a vector defined as

1 if no two ji belong to the same Sa ,
δS1 ,...,Sm ( j1 , . . . , jm ) =
0 otherwise.

The 2-multicuts are exactly cuts; all other

P multicuts on n points are interior points of CUTn ,
because of evident formula 2δS1 ,...,Sm = mi=1 δSi ,Si .

Ridge graph of the cone C or polytope P: the skeleton of the dual cone C∗ or dual polytope P∗ ,
respectively.
216 Handbook of Geometric Constraint Systems Principles

Table 9.1
The number of extreme rays and facets of cones C = CUTn , HY Pn , METn and the number of vertices
and facets of polytopes P = CUT Pn , HY PPn , MET Pn
C/P n=3 n=4 n=5 n=6 n=7 n=8
CUTn , e 3(1) 7(2) 15(2) 31(3) 63(3) 127(4)
CUTn , f 3(1) 12(1) 40(2) 210(4) 38, 780(36) 49, 604, 520(2, 169)
HY Pn , e 3(1) 7(2) 15(2) 31(3) 37, 170(29) 242, 695, 427(9, 003)
HY Pn , f 3(1) 12(1) 40(2) 210(4) 3, 773(14) 298, 592(86)
METn , e 3(1) 7(2) 25(3) 296(7) 55, 226(46) 119, 269, 588(3, 918)
METn , f 3(1) 12(1) 30(1) 60(1) 105(1) 168(1)
CUT Pn , v 4(1) 8(1) 16(1) 32(1) 64(1) 128(1)
CUT Pn , f 4(1) 16(1) 56(2) 368(3) 116, 764(11) 217, 093, 472(147)
HY PPn , v 4(1) 8(1) 16(1) 32(1) 13, 152(6) 1, 388, 383, 872(581)
HY PPn , f 4(1) 16(1) 56(2) 68(3) 10, 396(7) 1, 374, 560(22)
MET Pn , v 4(1) 8(1) 32(2) 554(3) 275, 840(13) 1, 550, 825, 600(533)
MET Pn , f 4(1) 16(1) 40(1) 80(1) 140(1) 224(1)

The full symmetry group of the cones CUTn and METn is the symmetric group Sym(n) for n 6= 4
and Sym(4)×Sym(3) for n = 4. The polytopes CUT Pn , MET Pn , are invariant, besides permutations,
under the following switching operation:

 {1, . . . , n}2 → R
US (d) = 1 − di j if |S ∩ {i, j}| = 1,
(i, j) 7→
di j otherwise.


Together, those define a group of order 2n−1 × n!. For n 6= 4, this is the full symmetry group. For
n = 4, the full group is Aut(K4,4 ) and so its order is 23 × 144.
It holds CUTn ⊆ METn and CUT Pn ⊆ MET Pn with equality only for n = 3, 4.
Table 9.1 gives known results of the number f of facets and extreme rays or vertices (e or v)
for small cut, metric and (considered in next section) hypermetric cones/polytopes. The numbers of
orbits, given there in parentheses, is under Sym(n) for cones and under Sym(n) and 2n−1 switchings
for polytopes.
The enumeration of orbits of facets of CUTn for n ≤ 7 was done in [8, 43, 54] for n = 5, 6,
and 7, respectively. For CUT8 and CUT P8 , sets of facets were found in [15]; completeness of these
sets was shown in [20]. The enumeration of orbits of extreme rays of METn for n ≤ 8 was done in
[17, 18, 44].

9.3 Hypermetric Cone and Hypermetric Polytope

Hypermetric cone HY Pn : the set of all hypermetrics on n points, i.e., the symmetric functions
d : {1, . . . , n}2 → R satisfying all dii = 0 and the inequalities
X X
H(b, d) = bi b j di j ≤ 0 for all b ∈ Zn , bi = 1.
1≤i< j≤n i

One obtains METn using only b of the form (1, 1, −1, 0n−3 ).
Computations of Metric/Cut Polyhedra and Their Relatives 217

Table 9.2
All types of Delaunay simplices S in R7 , their volume, the order of automorphism group and the
number of facets of the cone BarS
Si Vol(Si ) | Aut(Si )| No. of facets (orbits)
S1 1 40,320 298,592(86)
S2 2 40,320 5,768(9)
S3 2 1,440 6,590(62)
S4 3 540 966(9)
S5 3 1,152 728(9)
S6 3 240 640(39)
S7 4 1,440 28(3)
S8 4 240 153(11)
S9 4 144 131(10)
S10 5 72 28(6)
S11 5 48 28(8)

(2k + 1)-gonal inequality: an inequality H(b, d) ≤ 0 with n = 2k +P 1 and {0, ±1}-valued b. The
case of general b can be seen as such inequality on a multiset of ni=1 |bi | points with different
points occuring |b1 |, . . . , |bn | times, respectively.
Hypermetric polytope HY PPn : the polytope defined by all hypermetric inequalities
X
bi b j d(i, j) ≤ s(s + 1),
1≤i< j≤n
Pn
where b = (b1 , . . . , bn ) ∈ Zn with i=1 bi = 2s + 1 and s ∈ Z.
One obtains MET Pn using only b of the form (1, 1, −1, 0n−3 ) and (1, 1, 1, 0n−3 ).

There is an infinity of inequalities, defining the hypermetric cone HY Pn ; so, it is not obvious
that this cone is polyhedral. The proof of this [30, 31, 32] was achieved through the connection with
geometry of numbers that we now explain.
Given a quadratic form q, one can define the induced Delaunay tessellation with point set Zn
([41, 40, 55, 52]). It is well known that there are only a finite number of such tessellations, up to the
action of the group GLn (Z).
For a generic quadratic form, the tessellation is formed by simplices only; but, importantly, when
it is not formed of simplices, this induces linear conditions on the coefficients of the quadratic form.
There are a finite number of simplices, up to GLn (Z) action, and they have been classified in [38]
for n = 7, extending previous classification for n ≤ 6 [10, 49, 50]. Table 9.2 gives key information
on all 11 types of simplices in R7 .
Given a simplex S, denote by BarS the set of quadratic forms, for which S is contained in a
Delaunay polytope of the Delaunay tessellation. It is a polyhedral cone, called a Baranovskii cone
in [52].
For a quadratic form inside BarS , the Delaunay tessellation contains S as a simplex. For a
quadratic form on a facet of BarS , the simplex S is a part of a repartitioning polytope, i.e., a Delaunay
polytope with only n + 2 vertices.
If the simplex S has volume 1, then it is equivalent to the simplex formed by the vertices v1 = 0,
v2 = e1 , . . . , vn+1 = en . The quadratic form q is described uniquely by the distance function d(i, j) =
q(vi − v j ) on the vertices vi . For a given positive definite quadratic form q, denote by c(q) the center
of the sphere, circumscribing S, and by r(q) the radius of this sphere. Since S is of volume 1, a given
218 Handbook of Geometric Constraint Systems Principles

point v ∈ Zn can be uniquely expressed in barycentric coordinates as v = i bi vi with 1 = i bi and

P P
bi ∈ Z. In [32], the following formula is proved:

H(b, d) = kv − c(q)k2 − (r(q))2 .

So, the distance function d corresponds to a quadratic form q, for which S is a part of the Delaunay
tessellation, if and only if it belongs to HY Pn+1 .
With this connection, it is possible to prove the polyhedrality of the cone HY Pn . In general,
we have CUTn ⊆ HY Pn with equality only for 3 ≤ n ≤ 6 [32]. The first proper HY Pn , HY P7 , was
described in [24]; HY P8 was computed in [27].
The hypermetric polytope HY PPn is defined [27], by analogy with the metric and cut polytopes,
as a polytope invariant under the switching operations US . It turns out, that it is polyhedral with the
link being established with the centrally symmetric Delaunay polytopes of dimension n, see [27]
for details. One can check easily whether a distance matrix d belongs to HY Pn or HY PPn ; this has
been used in [27] to determine the facets and vertices of HY PP7 and HY PP8 . See in Table 9.3 all 22
orbits of facets of HY PP8 ; the simplicial ones there are marked by ∗ .
The skeletons of HY PPn , HY Pn contain a clique consisting of all cuts and all nonzero cuts,
respectively. We expect that any vertex is adjacent to a cut vertex (it holds for n ≤ 8); if true, it will
imply that each of the above skeletons has diameter 3. The ridge graphs of MET Pn , METn with n ≥ 4
have diameter 2 [16]. We expect that any facet of HY PPn , HY Pn is adjacent to a triangle/perimeter
facet (it holds for n ≤ 7); if true, it will imply that the ridge graphs of HY PPn , HY Pn have diameter
4.
The hypermetric polytope HY PP8 has 581 orbits of vertices, which are in details:
• 1 orbit of 27 cuts, corresponding to the Delaunay polytope [0, 1];
• 24 orbits, corresponding to the Delaunay polytopes 221 and 321 ;
• 556 orbits, corresponding the Delaunay polytope ER7 .

9.4 Cut and Metric Polytopes of Graphs

Cut polytope of the graph CUT P(G): the convex hull of all cut vectors of a given graph G =
(V, E), or, equivalently, the projections of CUT P|V | on the subspace R|E| indexed by the edge-
set E. Cut cone CUT (G) is such projection of CUT|V | .

Metric polytope MET P(G) of the graph G: the projection of MET P|V | on R|E| indexed by the
edges of G. Metric cone MET (G) is such projection of MET|V | .
Minor of graph G: a graph, which can be obtained from G by deleting edges and vertices and
contracting edges.

Clearly, it holds CUT (G) ⊆ MET (G) and CUT (G) ⊆ MET (G).

Theorem 9.1 CUT (G) = MET (G) or, equivalently, CUT P(G) = MET P(G) if and only if G does
not have any K5 -minor.

It was proved in [53] for cones and in [9] for polytopes.

Remark 9.1 By Wagner’s theorem [56], a finite graph is planar if and only if it has no minors K5
and K3,3 .
Computations of Metric/Cut Polyhedra and Their Relatives 219

Table 9.3
Orbits of facets of the hypermetric polytope HY PP8
|Fi |
Fi Representative 32 No. of classes Inc.([0, 1], {221 , 321 }, ER7 )
F1 (0, 0, 0, 0, 0, 1, 1, 1) 7 2 (96, 1598784, 80836608)
F2 (0, 0, 0, 1, 1, 1, 1, 1) 28 3 (80, 383040, 14300640)
F3 (0, 1, 1, 1, 1, 1, 1, 1) 16 4 (70, 131712, 3975552)
F4 (0, 0, 1, 1, 1, 1, 1, 2) 168 6 (60, 32160, 590960)
F5 (0, 1, 1, 1, 1, 1, 2, 2) 336 9 (52, 9600, 122160)
F6 (1, 1, 1, 1, 1, 1, 1, 2) 32 8 (56, 19656, 370272)
F7 (0, 1, 1, 1, 1, 1, 1, 3) 112 7 (42, 840, 1120)
F8 (1, 1, 1, 1, 1, 2, 2, 2) 224 12 (46, 3528, 39906)
F9 (0, 1, 1, 1, 1, 2, 2, 3) 1, 680 15 (40, 656, 2686)
F10 (1, 1, 1, 1, 1, 1, 2, 3) 224 14 (42, 1323, 6489)
F11 (1, 1, 1, 1, 1, 1, 1, 4) 32 8 (28, 0, 0)∗
F12 (1, 1, 1, 1, 1, 2, 3, 3) 672 18 (36, 252, 464)
F13 (1, 1, 1, 1, 2, 2, 2, 3) 1, 120 20 (38, 585, 3210)
F14 (1, 1, 1, 1, 1, 2, 2, 4) 672 18 (32, 66, 36)
F15 (1, 1, 1, 2, 2, 2, 3, 3) 2, 240 24 (33, 120, 302)
F16 (1, 1, 1, 1, 2, 2, 3, 4) 3, 360 30 (31, 62, 82)
F17 (1, 1, 1, 1, 2, 2, 2, 5) 1, 120 20 (25, 3, 0)∗
F18 (1, 1, 1, 1, 1, 3, 3, 4) 672 18 (27, 0, 1)∗
F19 (1, 1, 1, 2, 2, 3, 3, 4) 6, 720 36 (28, 22, 22)
F20 (1, 1, 1, 1, 2, 3, 3, 5) 3, 360 30 (25, 2, 1)∗
F21 (1, 1, 2, 2, 2, 3, 3, 5) 6, 720 36 (24, 3, 1)∗
F22 (1, 1, 1, 2, 2, 3, 4, 5) 13, 440 48 (24, 3, 1)∗

For a connected graph G = (V, E), the polytope CUT P(G) has 2|V |−1 vertices and dimension
|E|. The switching US still act on CUT P(G) and the automorphisms of G as well. Together, this
defines a group ARes(G) of automorphisms of CUT P(G), and |ARes(G)| = 2|V |−1 | Aut(G)|. The
full group Aut(CUT P(G)) might be larger than that. In [20] the dual description of CUT P(G) was
computed for several graphs; see data on some of them in Tables 9.4 and 9.5. In the 3rd column
there, A(G) denotes 21−|V | | Aut(CUT P(G))|; it is | Aut(G)| for all, but K4 , graphs there.
For a cycle C in G and an odd-sized set F of edges in C, the cycle inequality is
X X
fC,F = xe − xe ≤ |F| − 1.
e∈C−F e∈F

In fact, all facets of MET P(G) are defined [9] by all chordless cycle inequalities fC,F and all in-
equalities 0 ≤ xe ≤ 1 for edges e not belonging to a triangle in G.

Besides above inequalities of MET P(G), some valid inequalities of hypermetric type show up
on CUT P(G). Below we give a way [27] to get such inequalities.

Proposition 9.1
Let us take a valid inequality on CUT Pn of the form
X
f (x) = ai j xi j ≤ A.
1≤i< j≤n

Given a graph G with the vertex-set v1 , . . . , vn , suppose that any two vertices vi , v j are joined by a
such path Pi j , that:
220 Handbook of Geometric Constraint Systems Principles

Table 9.4
The number of facets of CUT P(G) = MET P(G) for some K5 -minor-free graphs G (2 non-planar
graphs and the skeletons of Platonic and some semiregular polyhedra)
G = (V, E) |V |, |E| A(G) No. of facets (orbits) Cycles
Wagner graph M8 8, 12 16 184(4) 2, 2, 4, 5
K3,3 = M6 6, 9 72 90(2) 2, 4
Dodecahedron 20, 30 120 167, 164(8) 2, 5, 9, 10, 10, 11, 12, 12
Icosahedron 12, 30 120 1, 552(4) 3, 5, 6, 6
Cube K22 8, 12 48 200(3) 2, 4, 6
Octahedron K2,2,2 6, 12 48 56(2) 3, 4
Tetrahedron K4 4, 6 144 12(1) 3
Prism7 14, 21 28 7, 394(6) 2, 2, 4, 7, 9, 9
APrism6 12, 24 24 2, 032(5) 3, 6, 7, 7, 8
Cuboctahedron 12, 24 48 1, 360(5) 3, 4, 6, 6, 8
Tr. Tetrahedron 12, 18 24 540(4) 2, 3, 6, 8

• the edge-sets of all paths Pi j are disjoint;

• if ai j > 0, then Pi j is reduced to an edge.
Then, the following inequality is valid on CUT P(G):
 
X X
H f ,G (x) = ai j  xe  ≤ A.
1≤i< j≤n e∈Pi j

Remark 9.2 When applied to the triangle inequalities of CUT P(Kn ) and taking switchings,
the above proposition gives us MET P(G). So, it is temping to define the hypermetric polytope
HY PP(G) for general graph G by the switchings of the extensions of all hypermetric inequalities
obtained from above proposition. What is not clear is when CUT P(G) = HY PP(G) and whether
there is a nice characterization of such hypermetrics. The above discussion is applied also to cones.

Any Kn -subgraph of G will satisfy the hypothesis and the facets of CUT Pn will give facets of
CUT P(G). Proposition 9.1 gives valid inequality induced by a class of graphs homeomorphic Kn .
(A graph H is homeomorphic to a subgraph of G if H can be mapped to G so that the edges of H
are mapped to disjoint paths in G.) A graph homeomorphic to Kn is a special case of a Kn -minor.

Remark 9.3 The proof of Theorem 9.1 [53] does not give hypermetric inequalities, or their gen-
eralizations, in a straightforward way; it appears nonconstructive.

Remark 9.4 In quantum information theory, general Bell inequalities, involving joint probabili-
ties of two probabilistic events, are exactly valid inequalities of the correlation polytope CORP(G)
(called also Boolean quadric polytope) of a graph, say, G. In particular, CORP(Kn,m ) is seen there as
the set of possible results of a series of Bell experiments with a separable (nonentangled) quantum
state shared by two distant parties, where one party has n choices of possible two-valued measure-
ments and the other party has m choices. A valid inequality of CORP(Kn,m ) is called a Bell inequality
and if facet inducing, a tight Bell inequality. This polytope is linearly isomorphic (via the covari-
Computations of Metric/Cut Polyhedra and Their Relatives 221

Table 9.5
The number of facets of CUT P(G) for some graphs G with K5 -minor
G = (V, E) |V |, |E| A(G) Number of facets (orbits) Cycles
Heawood graph 14, 21 336 5, 361, 194(9) 2, 6, 8
Petersen graph 10, 15 120 3, 614(4) 2, 5, 6
Pyr(APrism4 ) 9, 24 16 389, 104(17) 3, 3, 3, 4, 5
Möbius ladder M14 14, 21 28 369, 506(9) 2, 2, 4, 8, 10
Tr.Octahedron on P2 12, 18 48 62, 140(7) 2, 2, 4, 6, 6
K5,5 10, 25 2
2(5!) 16, 482, 678, 610(1, 282) 2, 4
K4,7 11, 28 4!7! 271, 596, 584(15) 2, 4
K4,6 10, 24 4!6! 23, 179, 008(12) 2, 4
K4,5 9, 20 4!5! 983, 560(8) 2, 4
K3,3,3 9, 27 (3!)4 624, 406, 788(2, 015) 3, 4
K1,4,4 9, 24 2(4!)2 36, 391, 264(175) 3, 4
K1,3,5 9, 23 3!5! 71, 340(7) 3, 4
K1,3,4 8, 19 3!4! 12, 480(6) 3, 4
K1,1,3,3 8, 21 4(3!)2 432, 552(50)
3, 3, 4
8 + 20m + 8 m2
K1,1,2,m , m > 2 m+4, 4m+5 4m! (16m − 5(7) 3, 3, 3, 4
Km+4 − Km , m>1 m+4, 4m+6 4!m! 8(8m2 − 3m + 2)(4) 3, 3
K8 − K3 8, 25 360 2, 685, 152(82) 3, 3
K7 − K2 7, 20 240 31, 400(17) 3, 3

ance map) to CUT P(K1,n,m ) [32, Section 5.2]. Similarly, CUT P(K1,n,m,l ) represents three-party Bell
inequalities.
The symmetry group of CORP(Kn,n ) and CUT P(K1,n,m ) has order 21+n+m n!m!. Table 9.5
gives the number of facets of CORP(Kn,m ) with (n, m) = (4, 4), (3, 5), (3, 4). The cases (n, m) =
(2, 2), (3, 3) were settled in [42] and [48], respectively.
In contrast to the Bell’s inequalities, which probe entanglement between spatially-separated
systems, the Leggett-Garg inequalities test the correlations of a single system measured at different
times. The polytope, defined by those inequalities for n observables, is equivalent ([7]) to the cut
polytope CUT Pn .

Remark 9.5 Diversity cone DIVn is the set of all diversities on n points, i.e., [14] the functions
f : {A : A ⊆ {1, . . . , n}} → R satisfying all f (A) ≥ 0 with equality if |A| ≤ 1 and all

f (A ∪ B) + f (B ∪C) ≥ f (A ∪C) if B 6= ∅.
The induced diversity metric di j is f ({i, j}).
Cut diversity cone CDIVn is the positive span of all cut diversities δ (A), where A ⊆ {1, . . . , n},
which are defined, for any S ⊆ {1, . . . , n}, by

1 if A ∩ S 6= ∅ and A \ S 6= ∅,
δS (A) =
0 otherwise.

CDIVn is [14] the set of all diversities from DIVn , which are isometrically embeddable into an l1 -
222 Handbook of Geometric Constraint Systems Principles

diversity, i.e., one defined on Rm with m ≤ b nn c by

2

m
X
fm1 (A) = max {|ai − bi |}.
a,b∈A
i=1

These two cones are natural “hypergraph” extensions of MET (G) and CUT (G).

9.5 Quasi-Semimetric Polyhedra

Quasi-semimetric cone QMETn : the set of all quasi-semimetrics on n points, i.e., the functions
q : {1, . . . , n}2 → R satisfying all qii = 0, all qi j ≥ 0 and all oriented triangle inequalities

qik ≤ qi j + q jk .

Quasi-semimetric polytope QMET Pn : the set of all q ∈ QMETn , satisfying all

qki + qi j + q jk ≤ 2.

Oriented m-multicut δS01 ,...,Sq for an ordered partition ∪m i=1 Si = {1, . . . , n}: a vector (actually, a
quasi-semimetric) defined as

1 if i ∈ Sa , j ∈ Sb and a < b,
δS01 ,...,Sm (i, j) =
0 otherwise.

The number of all oriented multicuts on n points is the Fubini number (called also ordered Bell
number) p0 (n) of all ordered partitions of n.
Oriented multicut cone OMCUTn : the positive span of all p0 (n) − 1 nonzero oriented multicuts
on n points. By δS1 ,...,Sm = δS01 ,...,Sm + δS0m ,...,S1 , we have

{q + qT : q ∈ OMCUTn } = CUTn .

Oriented cut cone OCUTn : the positive span of all 2n − 2 nonzero oriented cuts, i.e., 2-multicuts
0 , on n points.
δS,S

Weightable quasi-semimetric cone W QMETn : the set of all weightable quasi-semimetrics on n

points, i.e., quasi-semimetrics q ∈ QMETn , for which exists a (weight) function w = (wi ) :
{1, . . . , n} → R≥0 satisfying all
qi j + wi = q ji + w j .

Weightable quasi-semimetric polytope W QMET Pn : the set W QMETn ∩ QMET Pn .

QMETn and OMCUT n are full-dimensional cones in Rn(n−1) , while W QMETn , OCUTn are of di-
n+1
n+1

mension 2 − 1 and PMETn is of dimension 2 .
Besides strict inclusions W QMETn ⊂ QMETn and OCUTn ⊂ OMCUTn , we have, with equalities
only for n = 3, the following inclusions

OMCUTn ⊆ QMETn and OCUTn ⊆ W QMETn .

Computations of Metric/Cut Polyhedra and Their Relatives 223

Remark 9.6 Quasi-semimetrics were studied in [21, 22, 26, 33]. A quasi-semimetric q is
weightable if and only if it has relaxed symmetry, i.e., for distinct i, j, k it holds

qi j + q jk + qki = qik + qk j + q ji .

Also, W QMETn consists of all ((di j + dio − d jo )), where ((di j )) is as a semimetric on {0, 1, . . . , n},
but none of the inequalities dio + d jo − di j ≥ 0 is required.

Example 38 Consider random walks on a connected graph G = (V, E), where at each step walk
moves with uniform probability from current vertex to a neighboring one. The hitting time quasi-
metric H(u, v) is defined as the expected number of steps (edges) for a walk on G beginning at vertex
u to reach v for the first time; put H(u, u) = 0. This quasi-metric is weightable. The commuting time
metric is
C(u, v) = H(u, v) + H(v, u) = 2|E|Q(u, v),
where Q(u, v) is called the effective resistance metric.

It is shown in [29] that OCUTn and W QMETn on {1, . . . , n} are projections of CUTn+1 and
METn+1 (defined on {0, 1, . . . , n}) on the subspace orthogonal to δ{0} . Moreover, the quasi-
hypermetric cone W QHY Pn is defined as such projection of HY Pn+1 , and it holds OCUTn ⊆
W QHY Pn ⊆ W QMETn with equalities only for n = 3 and OCUTn = W QHY Pn only for 3 ≤ n ≤ 5.
See data on W QHY P5 in Table 9.6.

Remark 9.7 OCUTn is the set of all n-vertex l1 -quasi-semimetrics, i.e., quasi-metrically embed-
dable ones into a quasi-metric space (Rm , kx − yk) with oriented l1 -norm
m
X
kz = (z1 , . . . , zm )k = max(zi , 0).
i=1

On a measure space, such kx − yk is µ(B \ A), where the sets B, A correspond to x, y.

The cones QMETn and OMCUTn have Sym(n) as a symmetry group. Another symmetry, called
reversal, exists also: associate to each ray q the ray qT defined by qTij = q ji , i.e., in matrix terms, the
reversal corresponds to the transposition of matrices. This yields that Z2 × Sym(n) is a symmetry
group of the cones QMETn and OMCUTn . We expect that this is their full symmetry group. It is so
([29]) for OCUTn .
For QMET Pn , one can define the following analog of the switching operation of MET Pn : given
S ⊂ {1, . . . , n}, call oriented switching the operation

 {1, . . . , n}2 → R
US (q) = 1 − q ji if |S ∩ {i, j}| = 1,
(i, j) 7→
qi j otherwise.


This, together with reversal and Sym(n), gives a group of order 2n n!, expected to be the full sym-
metry group of QMET Pn and W QMET Pn (checked for n ≤ 9).

Oriented cut polytope OCUT Pn : the convex hull of all the oriented cuts and their images under
oriented switchings.

Oriented multicut polytope OMCUT Pn : the convex hull of all the oriented multicuts and their
images under oriented switchings.
224 Handbook of Geometric Constraint Systems Principles

Table 9.6
The number of extreme rays and facets in some quasi-metric cones for 3 ≤ n ≤ 6
No. of ext. rays No. of facets
Cone Dimension (orbits) (orbits) Diameters
OCUT3 =W QMET3 5 6(2) 9(2) 1; 2
OCUT4 =QHY P4 9 14(3) 30(3) 1; 2
OCUT5 14 30(4) 130(6) 1; 3
OCUT6 20 62(5) 16,460(61) 1; 3
QHY P5 14 70(6) 90(4) 2; 2
W QMET4 9 20(4) 24(2) 2; 2
W QMET5 14 190(11) 50(2) 2; 2
W QMET6 20 18,502(77) 90(2) 3; 2
OMCUT3 =QMET3 6 12(2) 12(2) 2; 2
OMCUT4 12 74(5) 72(4) 2; 2
OMCUT5 20 540(9) 35,320(194) 2; 3
QMET4 12 164(10) 36(2) 3; 2
QMET5 20 43,590(229) 80(2) 3; 2

In [28] the oriented quasi-semimetric cone and polytope on an undirected graph G is defined and
studied. A description by inequalities is given.
We expect, that OCUT Pn have exactly 22n−2 vertices, which is much higher than 2n − 2, the
number of extreme rays of OCUTn . Both polytopes have the same symmetry group as QMET Pn . In
Tables 9.6 and 9.7, are given computation for small quasi-metric cones and polytopes, respectively;
the orbits are under Sym(n) for cones and under Sym(n) × Z2 for polytopes.
It is conjectured that the diameters of OCUTn and OMCUTn are 1 and 2, respectively. Further-
more, if f ≥ 0 defines a facet of OMCUTn , then the zero-extension of f to OMCUTn+1 is still a
facet, just as in the CUTn case.
The extreme rays of QMETn have been studied in [19, 26]. There it was proved that they are
not symmetric and have at least n − 1 zeros, implying that no one is the directed path distance of an
oriented graph. The oriented multicuts define extreme rays of QMETn . Also, the vertex-splitting of
an extreme ray is still an extreme ray.
A list of non-adjacencies between facets of QMETn is given in [26] and it is conjectured there
that in all other cases the facets are adjacent, implying that the diameter of QMETn∗ is 2. The
diameter of OMETn is 3 for n = 4, 5.

9.6 Partial Metrics

Partial semimetric cone PMETn : the set of all partial semimetrics on n points, i.e., the symmetric
functions p : {1, . . . , n}2 → R, satisfying all 0 ≤ pii ≤ pi j and all sharp triangle inequalities

pik ≤ pi j + p jk − p j j .

Partial semimetric convex body PMET Pn : the set of all d ∈ PMETn satisfying, in addition, all
pi j ≤ 1 + pii and the perimeter inequalities

pi j + p jk + pki ≤ 2 + pii + p j j + pkk .

Computations of Metric/Cut Polyhedra and Their Relatives 225

Table 9.7
The number of vertices and facets in some quasi-metric polytopes for 3 ≤ n ≤ 6
Polytope Dimension No. of vertices (orbits) No. of facets (orbits)
OCUT P3 =W QMET P3 5 16(2) 16(2)
OCUT P4 =W QMET P4 9 64(3) 40(2)
OCUT P5 14 256(3) 1,056(5)
OCUT P6 20 1,024(4) 1,625,068(97)
W QMET P5 14 2,656(8) 80(2)
W QMET P6 20 1,933,760(120) 140(2)
OMCUT P3 =QMET P3 6 22(3) 20(2)
OMCUT P4 12 136(5) 1,160(9)
QMET P4 12 544(8) 56(2)
QMET P5 20 1,155,136(392) 120(2)

Remark 9.8 Partial semimetrics were introduced in [47]; they are used for treatment of partially
defined/computed objects in semantics of computation.
Clearly, ((pi j )) is a partial semimetric if and only if ((pi j − pii )) is a weighable quasi-
semimetric.

The symmetry group of PMETn should be Sym(n) (checked for n ≤ 9) but PMET Pn has addi-
tional symmetries: for S ⊂ {1, . . . , n} we define the switching US :

 {1, . . . , n}2 → R 
US (p) = 1 + pii + p j j − p ji if |S ∩ {i, j}| = 1,
(i, j) 7→
pi j otherwise.


We expect that, together with Sym(n), this defines the full symmetry group of PMET Pn (checked
for n ≤ 9).

Remark 9.9 In PMET Pn , the entries pi j are controlled by the pii , but the pii entriesP are not
bounded; so, PMET Pn is not a polytope. One may get a polytope by adding the inequalities i pii ≤
1; then the vertices of the obtained polytope are the vertices of MET Pn together with the extreme
rays of PMET Pn .

In the Table 9.8, the orbits of PMETn are under Sym(n), while for PMET Pn , the orbits are under
the group (of order n!2n−1 ) generated by switchings and permutations. But both have 3 orbits of
facets and, maybe, ridge graphs of diameter 2.

9.7 Supermetric and Hemimetric Cones

(m, s)-supermetric cone SMETnm,s : the set of all (m, s)-supermetrics on n points, i.e., for a number
s > 0 and an integer m ≥ 1, the functions h : {1, . . . , n}m+1 → R≥0 , such that
• hi1 ,...,im+1 = 0 if ia = ib for a 6= b;
• hi1 ,...,im+1 = hiσ (1) ,...,iσ (m+1) for all σ ∈ Sym(m + 1) (total symmetry);
226 Handbook of Geometric Constraint Systems Principles

Table 9.8
The number of vertices and facets in partial metric cone PMETn and body PMET Pn for 3 ≤ n ≤ 6
Cone/body Dimension No. of ext. rays (orbits) No. of facets (orbits) Diameters
PMET3 6 13(5) 12(3) 3; 2
PMET4 10 62(11) 28(3) 3; 2
PMET5 15 1,696(44) 55(3) 3; 2
PMET6 21 337,092(734) 96(3) 3; 2
PMET P3 6 17(4) 19(3) 2; 2
PMET P4 12 97(6) 44(3) 2; 2
PMET P5 20 7953(24) 85(3) 3; 2
PMET P6 12 5090337(427) 146(3) ?; 2

• For all distinct i1 , . . . , im+2 , it holds the (s; m)-simplex inequality


s × hi2 ,...,im+2 ≤ hi1 ,i3 ,...,im+2 + · · · + hi1 ,...,im+1 .

Binary (m, s)-supermetric cone SCUTnm,s : the positive span of all {0, 1}-valued extreme rays of
SMETnm,s .
m-hemimetric cone HMETnm : the set of all m-hemimetrics on n points, i.e., (m, 1)-supermetrics.
Multicut hemimetric cone HCUTnm : the positive span of all (m + 1)-multicuts.

In [28] a different definition of hemimetric cone is defined. Instead of (1; m)-simplex inequality a
different inequality using manifold is defined and it allows to define a version of hemimetric cone
on a simplicial complex.
Above notions has been studied in [19, 23, 34, 35].
The m-multicuts determine extreme rays of the cone HMETnm . An adjacency relation between
m-multicuts is conjectured in [19, 23] and the dual-description of many such cones are computed
there; see Table 9.6. The orbits in it are under Sym(n).
Given an m-hemimetric h on X, denote by s(X, h) the maximal s, such that it is a (m, s)-
supermetric. Easy to see that s(X, h) ≤ m + 1, if d is not identically zero.

Example 39 For a set X ⊂ Rd and x1 , . . . , xm+1 ∈ X, denote by µm (x1 , . . . , xm+1 ) the m-dimensional
volume of the convex hull of {x1 , . . . , xm+1 }. Clearly, µm is an m-hemimetric on X. Let sm (X) :=
s(X, µ) and let αn , βn , γn denote the vertex-set of regular n-simplex, n-hyperoctahedron and n-cube,
respectively. In [25], it was shown:
• sm (αn ) = m + 1 for all n ≥ 2, m ≥ 1.
√
• s2 (βn ) = 1 + 3 for all n and sm (βn ) = 3 for all m ≥ 3 and n.
√
1+2
√ n−2
• s2 (γn ) = 2n−3
for n ≤ 8, while for any fixed m, limn→∞ sm (γn ) = 1.

Remark 9.10 We can define the m-hemimetric polytope by adding to the definition of HMETnm
the following m-perimeter inequality:

h(x2 , . . . , xm+2 ) + h(x1 , x3 , . . . , xm+2 ) + · · · + h(x1 , . . . , xm+1 ) ≤ 2.

However, there is no equivalent of the switching in that case and so, the studies have been limited
to cones so far.
Computations of Metric/Cut Polyhedra and Their Relatives 227

Table 9.9
The number of extreme rays and facets of small supermetric cones

Cone Dim. No. of ext. rays (orbits) No. of facets (orbits) Diameters
m,s m+2

SMETm+2 , m+2 s+1 (1) 2m+4(2) min(s+1,m-s+1); 2
1 ≤ s ≤ m−1 but 2; 3 if m=2,s=1
m,m
SMETm+2 m+2 m+2(1) m+2(1) 1; 1
HCUT52 10 25(2) 120(4) 2; 3
HMET52 10 37(3) 30(2) 2; 2
HCUT63 15 65(2) 4,065(16) 2; 3
HMET63 15 287(5) 45(2) 3; 2
HCUT74 21 140(2) 474,390(153) 2; 3
HMET74 21 3,692(8) 63(2) 3; 2
HMET85 28 55,898(13) 84(2) 3; 2
HMET96 36 864,174(20) 108(2) ?; 2
HCUT62 20 90(3) 2,095,154(3,086) 2; ?
HMET62 20 12,492(41) 80(2) 3; 2
SMET52,2 10 132(6) 20(1) 2; 1
SCUT52,2 10 20(2) 220(6) 1; 3
3,3/2
SMET6 15 331,989(596) 45(2) 6; 2
SMET63,2 15 12,670(40) 45(2) 4; 2
3,5/2
SMET6 15 85,504(201) 45(2) 6; 2
SMET63,3 15 1,138(12) 30(1) 3; 1
SCUT63,3 15 21(2) 150(3) 1; 3
SMET74,2 21 2,561,166(661) 63(2) ?; 2
SMET74,3 21 838,729(274) 63(2) ?; 2
SMET74,4 21 39,406(37) 42(1) 3; 1
SCUT74,4 21 112(2) 148,554(114) 1; 4
SMET85,5 28 775,807(92) 56(1) ?; 1

Remark 9.11 Given a set X, the following cone on it was introduced in [46]: regular multidistance
cone MDISX is the set of all functions D : ∪m>1 X m → R≥0 , such that for all x1 , . . . , xm , y ∈ X it holds
• D(x1 , . . . , xm ) = 0 if x1 = · · · = xm ;

• D(x1 , . . . , xm ) = D(xσ (1) , . . . , xσ (m) ) for all σ ∈ Sym(m)) (total symmetry);

• D(x1 , . . . , xm ) ≤ m
P
i=1 D(xi , y) (multidistance m-star inequality);

• D(x1 , . . . , xm ) ≤ D(x1 , . . . , xm , y) (regularity).

Clearly, the restriction of D on X 2 is a semimetric.
228 References

9.8 Software Computations

The polyhedral cones and polytopes, that show up in the study of finite semimetrics, have a com-
plexity that grows very fast with the number of points. For example, it does not appear hopeful that
one could compute the facets of, say, CUT P20 in any reasonable time. On the other hand, for too
small number of points, no useful information is available. Therefore, when searching for new kind
of objects, it is advantageous to search for what the limit of the computational power offers.
One striking feature of the considered polytopes, such as CUT Pn , is their fairly large symmetry
group. The methodology of choice for this computation is the Adjacency Decomposition Method,
which proceeds when one has a facet to compute the adjacent ones and add it to the list if not
equivalent to a known one. The methodology was first introduced in [15] and applied in [23, 26]
to quasi-semimetric and hemimetric cones. Then it was applied in [39, 40, 41] to several polytopes
coming from geometry of numbers; see [12] for an overview of this kind of techniques. One should
keep in mind that yes, symmetry helps the computational tasks but it is not a panacea: some cases
are simply too hard to treat. Similarly to using a faster programming language (like C++) or having
a parallel program, it helps but does not radically change the nature of the problem treated.
However, for other tasks the picture is not so bleak. One is computing automorphism groups of
polytopes and cones, while knowing only their vertices (or facets). Cones defined by 1000 vertices
do not pose any problems; see [11] for details on the procedure used. Another area which is not so
much affected by combinatorial explosion is computing the adjacency between vertices (or facets).
There are several techniques for doing it, that rely on linear programming and are able to scale up
to dimension, say, 100
The software is programmed in the GAP programming language [45] and is available on [36]. A
faster C++ parallel program has been written [37].

References
[1] Patrice Assouad. Plongements isométriques dans L1 : aspect analytique. In Initiation Seminar
on Analysis: G. Choquet-M. Rogalski-J. Saint-Raymond, 19th Year: 1979/1980, volume 41 of
Publ. Math. Univ. Pierre et Marie Curie, pages Exp. No. 14, 23. Univ. Paris VI, Paris, 1980.
[2] Patrice Assouad. Sur les inégalités valides dans L1 . European J. Combin., 5(2):99–112, 1984.
[3] Patrice Assouad and Michel-Marie Deza. Espaces métriques plongeables dans un hypercube:
aspects combinatoires. Ann. Discrete Math., 8:197–210, 1980. Combinatorics 79 (Proc. Col-
loq., Univ. Montréal, Montreal, Que., 1979), Part I.
[4] Patrice Assouad and Michel-Marie Deza. Metric subspaces of l1 . Pub. Math. Orsay, 82-03,
1982.
[5] David Avis. Some Polyhedral Cones Related to Metric Spaces. Ph.D. thesis, Stanford Uni-
versity, 1977.
[6] David Avis and Michel-Marie Deza. The cut cone, L1 embeddability, complexity, and multi-
commodity flows. Networks, 21(6):595–617, 1991.
[7] David Avis, Patrick Hayden, and Mark M. Wilde. Leggett-Garg inequalities and the geometry
of the cut polytope. Phys. Rev. A (3), 82(3):030102, 4, 2010.
References 229

[8] David Avis and Mutt. All the facets of the six-point Hamming cone. European J. Combin.,
10(4):309–312, 1989.
[9] Francisco Barahona and Ali R. Mahjoub. On the cut polytope. Math. Programming,
36(2):157–173, 1986.
[10] Evgenii P. Baranovskiı̆. The conditions for a simplex of 6-dimensional lattice to be l-simplex
(in russian). Ivan. Univ., 2(3):18–24, 1999.
[11] David Bremner, Mathieu Dutour Sikirić, Dmitrii V. Pasechnik, Thomas Rehn, and Achill
Schürmann. Computing symmetry groups of polyhedra. LMS J. Comput. Math., 17(1):565–
581, 2014.
[12] David Bremner, Mathieu Dutour Sikirić, and Achill Schürmann. Polyhedral representation
conversion up to symmetries. In Polyhedral computation, volume 48 of CRM Proc. Lecture
Notes, pages 45–71. Amer. Math. Soc., Providence, RI, 2009.
[13] Jean Bretagnolle, Didier Dacunha-Castelle, and Jean-Louis Krivine. Lois stables et espaces
L p . Ann. Inst. H. Poincaré Sect. B (N.S.), 2:231–259, 1965/1966.
[14] David Bryant and Paul F. Tupper. Diversities and the geometry of hypergraphs. Discrete
Math. Theor. Comput. Sci., 16(2):1–20, 2014.
[15] Thomas Christof and Gerhard Reinelt. Decomposition and parallelization techniques for enu-
merating the facets of combinatorial polytopes. Internat. J. Comput. Geom. Appl., 11(4):423–
437, 2001.
[16] Antoine Deza and Michel-Marie Deza. The ridge graph of the metric polytope and some
relatives. In Polytopes: abstract, convex and computational (Scarborough, ON, 1993), volume
440 of NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., pages 359–372. Kluwer Acad. Publ.,
Dordrecht, 1994.
[17] Antoine Deza, Michel-Marie Deza, and Komei Fukuda. On skeletons, diameters and volumes
of metric polyhedra. In Combinatorics and computer science (Brest, 1995), volume 1120 of
Lecture Notes in Comput. Sci., pages 112–128. Springer, Berlin, 1996.
[18] Antoine Deza, Komei Fukuda, Tomohiko Mizutani, and Cong Vo. On the face lattice of the
metric polytope. In Discrete and computational geometry, volume 2866 of Lecture Notes in
Comput. Sci., pages 118–128. Springer, Berlin, 2003.
[19] Elena Deza, Michel Deza, and Mathieu Dutour Sikirić. Generalizations of finite metrics and
cuts. World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2016.
[20] Michel Deza and Mathieu Dutour Sikirić. Enumeration of the facets of cut polytopes over
some highly symmetric graphs. Int. Trans. Oper. Res., 23(5):853–860, 2016.
[21] Michel-Marie Deza and Elena I. Deza. Cones of partial metrics. Contrib. Discrete Math.,
6(1):26–47, 2011.
[22] Michel-Marie Deza, Elena I. Deza, and Janoš Vidali. Cones of weighted and partial metrics.
In Proceedings of the International Conference on Algebra 2010, pages 177–197. World Sci.
Publ., Hackensack, NJ, 2012.
[23] Michel-Marie Deza and Mathieu Dutour. Cones of metrics, hemi-metrics and super-metrics.
Ann. Eur. Acad. Sci., 1:141–162, 2003.
[24] Michel-Marie Deza and Mathieu Dutour. The hypermetric cone on seven vertices. Experi-
ment. Math., 12(4):433–440, 2003.
230 References

[25] Michel-Marie Deza, Mathieu Dutour, and Hiroshi Maehara. On volume-measure as hemi-
metrics. Ryukyu Math. J., 17:1–9, 2004.
[26] Michel-Marie Deza, Mathieu Dutour, and Elena I. Panteleeva. Small cones of oriented semi-
metrics. In Forum for Interdisciplinary Mathematics Proceedings on Statistics, Combinatorics
& Related Areas (Bombay, 2000), volume 22, pages 199–225, 2002.
[27] Michel-Marie Deza and Mathieu Dutour Sikirić. The hypermetric cone and polytope on eight
vertices and some generalizations. preprint at arxiv:1503.04554, March 2013.
[28] Michel-Marie Deza and Mathieu Dutour Sikirić. Generalized cut and metric polytopes of
graphs and simplicial complexes. preprint at arxiv:1706.02516, March 2017.
[29] Michel-Marie Deza, Vyacheslav P. Grishukhin, and Elena I. Deza. Cones of weighted quasi-
metrics, weighted quasi-hypermetrics and of oriented cuts. In Mathematics of Distances and
Applications, ITHEA, Sofia, pages 31–53. 2012.
[30] Michel-Marie Deza, Vyacheslav P. Grishukhin, and Monique Laurent. Extreme hypermetrics
and L-polytopes. In Sets, graphs and numbers (Budapest, 1991), volume 60 of Colloq. Math.
Soc. János Bolyai, pages 157–209. North-Holland, Amsterdam, 1992.
[31] Michel-Marie Deza, Vyacheslav P. Grishukhin, and Monique Laurent. Hypermetrics in geo-
metry of numbers. In Combinatorial optimization (New Brunswick, NJ, 1992–1993), vol-
ume 20 of DIMACS Ser. Discrete Math. Theoret. Comput. Sci., pages 1–109. Amer. Math.
Soc., Providence, RI, 1995.
[32] Michel-Marie Deza and Monique Laurent. Geometry of cuts and metrics, volume 15 of Algo-
rithms and Combinatorics. Springer, Heidelberg, 2010. First softcover printing of the 1997
original [MR1460488].
[33] Michel-Marie Deza and Elena I. Panteleeva. Quasi-semi-metrics, oriented multi-cuts and
related polyhedra. European J. Combin., 21(6):777–795, 2000. Discrete metric spaces (Mar-
seille, 1998).
[34] Michel-Marie Deza and Ivo G. Rosenberg. n-semimetrics. European J. Combin., 21(6):797–
806, 2000. Discrete metric spaces (Marseille, 1998).
[35] Michel-Marie Deza and Ivo G. Rosenberg. Small cones of m-hemimetrics. Discrete Math.,
291(1-3):81–97, 2005.
[36] Mathieu Dutour Sikirić. Polyhedral. http://mathieudutour.altervista.org/
Polyhedral/.
[37] Mathieu Dutour Sikirić. Polyhedral cpp. https://github.com/MathieuDutSik/
DualDescriptionADM.
[38] Mathieu Dutour Sikirić. The seven dimensional perfect Delaunay polytopes and Delaunay
simplices. Canad. J. Math., 69(5):1143–1168, 2017.
[39] Mathieu Dutour Sikirić, Achill Schürmann, and Frank Vallentin. Classification of eight-
dimensional perfect forms. Electron. Res. Announc. Amer. Math. Soc., 13:21–32 (electronic),
2007.
[40] Mathieu Dutour Sikirić, Achill Schürmann, and Frank Vallentin. Complexity and algorithms
for computing Voronoi cells of lattices. Math. Comp., 78(267):1713–1731, 2009.
[41] Mathieu Dutour Sikirić, Achill Schürmann, and Frank Vallentin. The contact polytope of the
Leech lattice. Discrete Comput. Geom., 44(4):904–911, 2010.
References 231

[42] Arthur Fine. Hidden variables, joint probability, and the Bell inequalities. Phys. Rev. Lett.,
48(5):291–295, 1982.
[43] Vyacheslav P. Grishukhin. All facets of the cut cone Cn for n = 7 are known. European J.
Combin., 11(2):115–117, 1990.
[44] Vyacheslav P. Grishukhin. Computing extreme rays of the metric cone for seven points.
European J. Combin., 13(3):153–165, 1992.
[45] The GAP group. GAP — Groups, Algorithms, and Permutations, Version 4.4.6.
[46] Martin Javier and Major Gaspar. Regular multidistances. In XV Congresso Espagnol sombre
Technologias y Logica Fuzzy, Huelva, pages 297–301. 2010.
[47] Steve G. Matthews. Partial metric topology. In Papers on general topology and applications
(Flushing, NY, 1992), volume 728 of Ann. New York Acad. Sci., pages 183–197. New York
Acad. Sci., New York, 1994.
[48] Itamar Pitowsky and Karl Svozil. Optimal tests of quantum nonlocality. Phys. Rev. A (3),
64(1):014102, 4, 2001.
[49] Sergey S. Ryshkov and Evgenii P. Baranovskiı̆. Repartitioning complexes in n-dimensional
lattices (with full description for n ≤ 6). In Voronoi impact on modern science, pages 115–124.
Institute of Mathematics, Kyiv, 1998.
[50] Sergey S. Ryškov and Evgenii P. Baranovskiı̆. C-types of n-dimensional lattices and 5-
dimensional primitive parallelohedra (with application to the theory of coverings). Proc.
Steklov Inst. Math., (4):140, 1978. Cover to cover translation of Trudy Mat. Inst. Steklov 137
(1976), Translated by R. M. Erdahl.
[51] Isaac J. Schoenberg. Metric spaces and positive definite functions. Trans. Amer. Math. Soc.,
44:522–536, 1938.
[52] Achill Schürmann. Computational geometry of positive definite quadratic forms, volume 48 of
University Lecture Series. American Mathematical Society, Providence, RI, 2009. Polyhedral
reduction theories, algorithms, and applications.
[53] Paul D. Seymour. Matroids and multicommodity flows. European J. Combin., 2(3):257–290,
1981.
[54] Mikhail E. Tylkin (=M. Deza). On Hamming geometry of unitary cubes. Soviet Physics.
Dokl., 5:940–943, 1960.
[55] George Voronoi. Nouvelles applications des paramètres continus à la théorie des formes
quadratiques. Deuxième Mémoire. Recherches sur les parallélloèdres primitifs. J. Reine
Angew. Math, 134(1):198–287, 1908.
[56] Klaus Wagner. über eine Eigenschaft der ebene Komplexe. Math. Annal., 114:570–590, 1937.
Chapter 10
Cayley Configuration Spaces
Meera Sitharam
University of Florida

Menghan Wang
University of Florida

Joel Willoughby
University of Florida

Rahul Prabhu
University of Florida

CONTENTS
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
10.1.1 Euclidean Distance Cone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
10.2 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
10.3 Related Chapters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
10.4 Characterizing 2D Graphs with Convex Cayley Configuration Spaces . . . . . . . . . . . . . . . 236
10.5 Extension to other Norms, Higher Dimensions, and Flattenability . . . . . . . . . . . . . . . . . . . 238
10.5.1 Computing Bounds of Convex Cayley Configuration Spaces in 3D for
Partial 3-Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
10.5.2 Some Background on the Distance Cone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
10.5.3 Genericity and Independence in the Context of Flattenability . . . . . . . . . . . . . . . 242
10.6 Efficient Realization through Optimal Cayley Modification . . . . . . . . . . . . . . . . . . . . . . . . . 245
10.7 Cayley Configuration Spaces of 1-Dof Tree-Decomposable Linkages in 2D . . . . . . . . . 246
10.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250

10.1 Introduction
A Euclidean Distance Constraint System (EDCS) or linkage (G, δ ) is a graph G = (V, E) together
with a distance vector δ that assigns a distance δe or a distance interval [δel , δer ] to each edge e ∈ E.
A d-dimensional realization of the EDCS is the assignment p of points in Rd to the vertices in V
such that the distance equality (resp. inequality) constraints are satisfied: δ(u,v) = ||p(u) − p(v)||,
l
respectively, δ(u,v) r
≤ ||p(u) − p(v)|| ≤ δ(u,v) .
Describing, exploring, and sampling the realization space of a linkage is a difficult problem that
arises in many classical areas of mathematics and computer science and has a wide variety of ap-

233
234 Handbook of Geometric Constraint Systems Principles

plications in computer aided design for mechanical engineering, robotics, and molecular modeling.
Especially for under-constrained (or independent and not rigid) linkage whose realizations have
one or more degrees of freedom of motion, progress on this problem has been very limited. Exist-
ing methods for sampling linkage realization spaces often use Cartesian representations, factoring
out the Euclidean group by arbitrarily “pinning” or “grounding” some of the points’ coordinate
values. Even when the methods use “internal” representation parameters such as Cayley parame-
ters (nonedges) or angles between unconstrained objects, the choice of these parameters is ad hoc.
We characterize the graphs of linkages for which parameters can be judiciously chosen so that the
parametrized configuration space is convex and therefore easier to explore.
We define a representation of the realization space of a linkage in Rd as (i) a choice of parameter
set, specifically a choice of a set F of nonedges of G (F ⊆ V ×V \ E) called Cayley parameters, and
(ii) a set ΦdF (G, δ ), called the Cayley configuration space (a term first coined in [31]), consisting
of realizable distance vector δF for F. I.e., the augmented linkage (G ∪ F, δE , δF ) has at least one
d-dimensional realization. In the presence of inequalities, the Cayley configuration space is denoted
ΦdF (G, [δ l , δ r ]) and the augmented linkage is: (G ∪ F, [δEl , δEr ], δF ). Here G ∪ F refers to a graph
(V, E ∪F). In other words, our representations are in Cayley parameters or nonedge distances: the set
ΦdF (G, δ ) is the projection onto the Cayley parameters in F, of the Cayley-Menger semi-algebraic
set with fixed (G, δ ) [10, 12]. Provided G ∪ F is minimally rigid, for every Cayley configuration
in ΦdF (G, δ ), generically there exists at least one and at most finitely many Cartesian realizations,
namely, the realizations of (G ∪ F, δ , δF ).
We survey results on the study of exact, efficient representations of realization spaces of link-
ages. These results completely characterize linkages that have connected, convex Cayley configura-
tion spaces with efficiently computable bounds based on precise and formal measures of efficiency.
These results do not rely on the genericity of the linkages. Results presented here employ (a) con-
vexity results and positive semi-definiteness of Euclidean distance matrices (we refer the reader to
Chapter 1 for a discussion on Euclidean distance matrices), with (b) forbidden minor characteriza-
tions and algorithms related to d-flattenability of graphs.

10.1.1 Euclidean Distance Cone

The structure of the cone Φn,l p , its strata and faces are well-studied. For l2 this is called the Euclidean
distance matrix (EDM) cone (we refer the reader to Chapter 1 for a discussion on EDM) [14, 16, 37],
which is a simple, relationship to the cone of positive semidefinite matrices, a fact first observed by
Schoenberg [30]. Consequently, understanding its structure is important in semidefinite program-
ming relaxations and the so-called sums of squares method with numerous applications [6, 20, 28].
Connections between combinatorial rigidity and the structure of the EDM have been investigated
extensively by Alfakih [1, 2]. The reader is additionally referred to [15] for a comprehensive survey
of key results about the EDM cone, including observations about the face structure and dimensional
strata of the EDM cone.

Organization: The rest of the chapter is organized as follows. Section 10.2 introduces the vari-
ous background definitions and terms. Section 10.3 lists the related chapters. Section 10.4 gives
the characterization of graphs in 2D that have convex Cayley configuration spaces. Section 10.5
discusses the connection between inherent Cayley configuration space and flattenability, which ex-
tends beyond the Euclidean norm. The section further discusses genericity and independence in the
context of flattenability. Section 10.6 discusses how a linkage can be modified for efficient realiza-
tion. Section 10.7 discusses the Cayley configuration spaces of a special class of linkages, called
1-Dof tree-decomposable linkages, in 2D. Section 10.8 gives the conclusion.
Cayley Configuration Spaces 235

10.2 Glossary
Cayley Configuration Space: The d-dimensional Cayley configuration space of a linkage (G, δ ),
over some set of nonedges, F, under norm l p is denoted ΦdF,l p (G, δ ) and is defined as the set of
vectors δF of lengths attained by the nonedges F over all the realizations of the linkage (G, δ )
in Rd . This same space is also sometimes referred to as the Cayley configuration space of a
realization or framework (G, p) whose edge lengths are δ .
Graphs with convex d-dimensional Cayley configuration space: Graph G has a convex d-
dimensional Cayley configuration space under norm l p if there exists a set of nonedges F such
that for all edge length vectors δ , ΦdF,l p (G, δ ), is convex.

Graphs with inherent convex d-dimensional Cayley configuration space: Graph G has an in-
herent convex d-dimensional Cayley configuration space under norm l p if for every partition of
the edges of G into H and F, and every edge length vector δH , ΦdF,l p (H, δ ), is convex.

Graph Minor: A graph G has a graph K as minor if G can be reduced to K via vertex/edge deletions
and edge contractions (coalescing or identifying the two vertices of an edge).
Minor Closed: If a property of G remains consistent under the operation of taking minors, that
property is minor-closed. A useful result due to [29] is that if a property is minor-closed, then
there is a finite set of forbidden minors (see next definition)
Finite Forbidden Minor Characterization: A class C of graphs has a finite forbidden minor char-
acterization, if there exists a fixed, finite set M of minors, such that G ∈ C if and only if G doesn’t
contain any K ∈ M as a minor.
k-sum: A k-sum is a way of combining two graphs by gluing them together at a Kk or a complete
graph on k vertices.
(Complete) k-Trees: A graph is a (complete) k-tree if it is formed by k-sums of Kk+1 s. A partial
k-tree is a proper subgraph of a complete k-tree. These are often useful in forbidden minor
characterization.

10.3 Related Chapters

Chapter 8 Dimensional and Universal Rigidity of Bar Frameworks: We use notions of rigidity,
genericity, and independence defined in terms of the so-called rigidity matrix.
Chapter 24 Rigidity with Polyhedral Norms: Many of the results in this chapter extend readily
to different norms with appropriate definitions of rigidity and realization developed by Kitson.
Chapter 9 Computations of Metric/Cut Polyhedra and Their Relatives: Many of the ideas re-
lated especially to d-flattenability rely heavily on the structure of the metric cone and its sec-
tions and stratum.
236 Handbook of Geometric Constraint Systems Principles
y
4 p12 p22
d
d
5 3
p01

p21
d
d
d x
p00
1 2 p10
(a) (b)

Figure 10.1
An example of a simple Cayley configuration space on 2 nonedges. On the left is a pentagon graph
G. We choose to look at the Cayley configuration space Φ2F (G, δ ) where F = {(v1 , v4 ), (v2 , v4 )} and
δ = (1, 1, 1, 1, 1). The result is the shaded region to the right where the x-axis are the edge lengths
for (v1 , v4 ) and the y-axis edge lengths for (v2 , v4 ).

10.4 Characterizing 2D Graphs with Convex Cayley Configura-

tion Spaces
For a graph, the choice of Cayley parameter nonedges is important, as this determines whether the
resulting Cayley configuration space is convex or not. Figure 10.1 shows a choice of Cayley param-
eters for a graph which leads to a convex Cayley configuration space while Figure 10.2 shows an
example where the choice of a different Cayley parameter for the same graph leads to a disconnected
Cayley configuration space.
In this section, we present results from [31] that characterize 2D graphs that have convex Cayley
configuration spaces. Specifically, we are interested in characterizing graphs G for which there
is a set F of nonedges such that for all distance assignments δE , the Cayley configuration space
Φ2f , (G, δ ) is convex. We first define some background terms.

Minimal 2-sum component: Recursively decompose a graph using 2-sums. The resulting inde-
composible graphs in the decomposition are minimal 2-sum components.

Theorem 10.1 characterizes 2D graphs in which, for a single nonedge chosen as a Cayley parameter,
the Cayley configuration space is a single interval.

Theorem 10.1 [31] Given a graph G = (V, E) and a nonedge f , the Cayley configuration space
Φ2f , (G, δ ) is a single interval for all δ if and only if all the minimal 2-sum components of G ∪ f that
contain f are partial 2-trees.

In 3-dimensions (and higher), however, the above theorem does not hold, as a simple counterex-
ample shows (see Figure 10.3). In this example the specified nonedge must be contracted to obtain
a forbidden minor from G ∪ f . Thus, to fully characterize the property, the following conjecture was
raised for graphs G and nonedges f such that G ∪ f is not d-flattenable.

Conjecture 1 An nonedge f of a graph G has a single interval of attainable values for all edge
Cayley Configuration Spaces 237

Figure 10.2
When parameters for the EDCS in (left) are chosen, to be the dashed edge, we get a disconnected
Cayley configuration space shown in (right). The realization p(v1 ) can lie in either of the two solid
arc segments of the circle, yielding a disconnected Cayley configuration space.

length vectors of G \ f in d-dimensions if and only if for any minimal 2-sum components of G
containing f that are not d- flattenable, f must always (i) be removed or (ii) duplicated or (iii)
contracted in order to obtain a forbidden minor of d-flattenability from G.

Building on Theorem 10.1, Theorem 10.2 gives an exact characterization of the class of graphs
whose corresponding linkages admit a 2D convex Cayley configuration space for a set of nonedges
F.

Theorem 10.2 For a graph G = (V, E), the following four statements are equivalent:
1. There exists a nonempty set of nonedges F such that for all δ , Φ2F, (G, δE ) is connected;

2. There exists a nonempty set of nonedges F such that for all δ , Φ2F, (G, δE ) is convex;

3. There exists a nonempty set of nonedges F such that for all δ , Φ2F, (G, δE ) is a polytope linear
faces.
4. G has a 2-sum component that is a flexible partial 2-tree.

For the class of graphs described in Theorem 10.2, Theorem 10.3 gives an exact combinatorial
characterization of the choices of Cayley parameters that ensure a convex Cayley configuration
space.

Theorem 10.3 Given a graph G = (V, E) and nonempty set of nonedges F, the 2D Cayley config-
uration space Φ2F, (G, δ ) is a linear polytope, connected or convex for all δ if and only if all the
minimal 2-sum components of G ∪ F containing any subset of F are partial 2-trees.

All of the above theorems hold with edge length intervals instead of edge lengths.
238 Handbook of Geometric Constraint Systems Principles

5 6

3 4

1 f 2

Figure 10.3
G’s 3D Cayley configuration space on nonedge f is one interval, although G ∪ { f } has a K5 minor.

Theorem 10.4 Theorems 10.1–10.3 hold with Φ2F (G, (δ l , δ r )) replacing Φ2F (G, δ ).

Convex Cayley configuration spaces can be extended to other norms. Many of these key results
follow in the next sections. In particular, Section 10.5 gives the extension to general l p norms and
its connection to flattenability.

10.5 Extension to other Norms, Higher Dimensions, and Flatten-

ability
In this section we show the connection between flattenability and Cayley configuration spaces [36].
We start by defining some background terms.

d-Flattenability: A graph G is d-flattenable if for every realization r of G under the l p norm,

the linkage (G, δ ) where δvw := ||r(v) − r(w)|| p also has a realization in the d-dimensional l p -
normed space. This definition does not imply that there is a continuous path of realizations start-
ing from a realization of (G, δ ) in some higher dimension to the realization in d-dimensions,
nor does it refer to realizations of intrinsic dimension d in some higher dimensional space.

The concept of d-flattenability was first introduced in a series of papers [8, 9] for the Euclidean or
l2 norm. However they called it “d-realizability,” which can be confused with the realizability of a
given linkage in d-dimensions. This is one of reasons we introduced the term: flattenability.
The term flattening has also been used by Matousek [21] in the context of non-isometric em-
beddings (with low distortion via Johnson-Lindenstrauss lemma in l2 [24], impossibility of low
distortion in l1 [11], etc). We admit arbitrary distortions of nonedge lengths, but force edge lengths
to remain undistorted.
Immediately by definition, d-flattenability is a minor-closed property under any norm. A full
characterization for 3-flattenable graphs was given for the Euclidean or l2 norm by [8, 9]. A sum-
mary of the known forbidden minor results for the l2 and l1 norms is presented in Table 10.1. The
proofs for the l2 3-dimensional cases mostly involved exhaustion of all other possibilities. Later in
this section, tools are given (see Theorem 10.5) which lead to more direct proofs of such properties.
A close connection between Cayley configuration spaces and flattenability was shown in [36]:
Cayley Configuration Spaces 239

Table 10.1
Summary of known forbidden minor characterizations for l1 and l2 norms.
Norm Dimension Forbidden Minor(s) Characterization References
l2 1 K3 Forests [9]
2 K4 Partial 2-trees [9]
3 K5 , Octahedron Partial 3-trees and others [9]
n>3 Kn+1 + others Partial n-trees and others [5]
l1 1 K3 Forests [36]
2 W4 + 2K4 − 1 Partial 2-trees and others [17]

I1

I2

Figure 10.4
On the left is an illustration of an equidistant K4 with the possible intervals of Case 1. On the right
is the same for Case 2

Theorem 10.5 For l p norms, G is d-flattenable if and only if G has a d-dimensional, inherent
convex Cayley configuration space. As a direct corollary, it follows that both properties are minor-
closed for general l p norms.

This “only if” direction of this statement was shown in [31] for the l2 norm. The argument only
required the fact that the cone of squared distance vectors is convex. Hence, we can use the same
proof if we can show Φn,l p is convex. The proof of the “if” direction requires that the cone is the
convex hull of l pp distance vectors in any dimension d.
The following result provides a nice link between d-flattenability and convex Cayley configura-
tion spaces.

Corollary 10.6 Having a d-dimensional convex Cayley configuration space on all subgraphs is a
minor-closed property.

Another immediate result is that d-flattenability and convex Cayley configuration spaces have
the same forbidden minor characterizations for given d under the same l p -norm. This gives us a nice
tool when trying to find forbidden minors for other l p norms:

Observation 1 If for some assignment of distances δ to edges of G leads to a non-convex

ΦdF,l p (G, δ ), then G is not d-flattenable.

Example Usage: This is an example to show how Theorem 10.5 can be used. We will use it to show
a specific graph is not 2-flattenable under the l1 (or Manhattan) norm. This example comes from
[36].

Theorem 10.7 The so called “banana” graph or K5 minus one edge is not 2-flattenable under the
l1 norm.
240 Handbook of Geometric Constraint Systems Principles

Proof 10.1 We will invoke Observation 1 to show this. Consider a distance vector for the banana
with unit distances for all except one edge, f . This has a realization in 3-dimensions as K5 is 3-
flattenable for the l1 norm (see [5]). Then, we have an equidistant K4 as a subgraph. The only
realization for such a K4 in 2-dimensions is to have all 4 points arranged as the vertices of the unit
ball centered at the origin. The two remaining unit edges then connect a new vertex to two of these
points. Here we have 2 cases: the two vertices border the same quadrant or they lie across one of the
axes from each other.
Case 1: Without loss of generality, we assume the two vertices are the upper right of the K4 . In
Figure 10.4 (left), it can be seen that the new vertex can lie anywhere in I1 or I2 . If it lies in I1 , the
remaining edge of the banana can take lengths in the range [0, 1]. If it lies in I2 , the only length it
can be is two.
Case 2: Without loss of generality, assume the 2 vertices are the top-most and bottom-most.
Again from Figure 10.4 (right), the new vertex only has 2 positions it can be in, each leading to a
length of 1 for the remaining edge.
Hence, Φ2F,l1 (G \ F, δ G\F ) = [0, 1] ∪ {2}, where G is the banana and F = { f }. This is not convex
and thus by Theorem 10.5, the banana is not 2-flattenable.

10.5.1 Computing Bounds of Convex Cayley Configuration Spaces in 3D for Partial

3-Trees
For the case of l2 , partial 3-trees are 3-flattenable and by Theorem 10.5 have inherent convex Cayley
configuration spaces. Cayley configurations are bound by tetrahedral inequalities. However, it is
possible to achieve tighter bounds, as shown in [13], where they give algorithms to find efficient
bounds on Cayley parameters for partial 2-trees in 3D (a subclass of partial 3-trees), and show that
bounds can be computed in quadratic time in the size of the graph. In cases where there are multiple
Cayley parameters, they additionally show that the order in which Cayley parameters are fixed have
an effect on the efficiency of the range computation.
To prove Theorem 10.5 and gain some insight into the connection, it is necessary to have some
information on the so-called distance cone.

10.5.2 Some Background on the Distance Cone

The l pp
Distance Cone: Many of the proofs given for the results in this section rely heavily on the
structure of the cone Φn,l p consisting of vectors δq of pairwise l pp -distances of n-point config-
urations q. (A proof that this set is a cone can be found in [5], which also applies to infinite
dimensional settings).

Strata of the Cone : The d-dimensional stratum of this cone consists of pairwise distance vectors
of d-dimensional point configurations and is denoted Φdn,l p .

Projections of the Cone: The projection or shadow of this cone (resp. stratum) on a subset of co-
ordinates i.e., pairs corresponding to the edges of a graph G is denoted ΦG,l p (resp. ΦdG,l p ).
This projection is the set of realizable edge-length vectors δ of linkages (G, δ ) in l pp (resp. in
d-dimensions) (see Figure 10.5).
Flattening Dimension: The l p -flattening dimension of a graph G (resp. class C of graphs) is the
minimum dimension d for which G (resp. all graphs in C) is flattenable in l p .

Notice that Φn,l p is the same as ΦKn ,l p , where Kn is the complete graph on n vertices. Let n p be the
n
flattening dimension of K . It is not hard to show [15] that in fact n ≤ R(2) . For the Euclidean or
n p
Cayley Configuration Spaces 241

Figure 10.5
On the left we have the cone of realizable distance vectors under l p . It is shown here as a polytope,
but in general that is not the
case; these are not rigorous figures – their purpose is intuitive visual-
ization. The cone lives in n2 -dimensional space where each dimension is a pairwise distance among
n points. In the middle is a projection of the cone onto the edges of some graph. This yields a lower
dimensional object (unless G is complete). On the right, a d-dimensional stratum is highlighted and
lines show the projection onto coordinates representing edges of a graph. In general this stratum is
not just a single face. Note that this projection is equal to the projection as the whole cone (middle)
iff G is d-flattenable.

p dimension of any graph G = (V, E)

l2 case, a further result of Barvinok [7] shows that the flattening
(although he did not use this terminology), is at most O( |E|). Notice additionally that the d-
dimensional Cayley configuration space on a set of non-edges F of G ΦdF,l p (G, δ ) is the coordinate
shadow of the (G, δ )-fiber of ΦdG∪F,l p , i.e., all linkages (G ∪ F, δG∪F ) that have δ assigned to the
edges of G, on the coordinate set F (see Figure 10.6).
Now we outline the proof of Theorem 10.5. The full proof can be found in [36]. We will be
using the convexity properties of the cone and its strata described above to achieve the result.

Proposition 10.1
Φn,l p for general l p is contained in the convex hull of the l pp distance vectors of the 1-dimensional
n-point configurations in R.

To see this, take some realization of a complete linkage (Kn , δ ), which from [5], we know to be
finite dimensional. Then we can simply treat each dimension of the realization as a 1-dimensional
point configuration in R. We simply build the distance vector δ as a convex sum of vectors only
defined by one dimension of the realization. It is well known that Φn,l p is convex. This along with
10.1 leads to the following:

Proposition 10.2
Φn,l p , 1 ≤ p ≤ ∞ is the convex hull of the l pp distance vectors of the 1-dimensional, n-point config-
uration vectors in R.

Now we can move on to proving Theorem 10.5. Suppose G is d-flattenable under some l p -norm.
Because G is d-flattenable, ΦdG,l p is convex. Given a subgraph G0 of G, if we break G into H and
242 Handbook of Geometric Constraint Systems Principles

Figure 10.6
This is an example of ΦdG∪F,l p that is not convex. The linkage (G, δ ) and its fiber in ΦdG∪F,l p are
shown on the left. Note that the fiber is not convex. In the middle, this fiber is then projected onto
the remaining edges of G ∪ F to form ΦdF,l p (G, δ ). Note that it is not convex either. On the right,
Φdn,l p is projected onto the edges of some d-flattenable G (note that this is the same as projection of
Φn,l p ). The inherent Cayley configuration space corresponding to some subgraph G \ H of G is then
shown projected onto the edges of G \ H. This projection is convex.

G0 and fix the values of E corresponding to a linkage (H, δH ), we are taking a section of ΦdH∪G0 ,l p ,
which is again convex. Then any inherent Cayley configuration of G is convex.
For the other direction, we know the Cayley configuration space on the empty set is convex.
This is just ΦdG,l p and because it is convex, it is its own convex hull. From Proposition 10.2, we see
that ΦG,l p is exactly the convex hull of its d-dimensional stratum, ΦdG,l p . Thus, G is d-flattenable.

10.5.3 Genericity and Independence in the Context of Flattenability

Generic d-Flattenability: Given an l p framework (G, r), with n vertices, in arbitrary dimension,
consider its pairwise length vector, δr , in the cone Φn,l p . A framework (G, r) of n vertices is
generic with respect to d-flattenability if the following hold:
(i) There is an open neighborhood Ω of δr in the (interior of the) cone Φn,l p , (recalling that n p
is the flattening dimension of the complete graph Kn , Ω corresponds to an open neighborhood
of n p -dimensional point-configurations of r); and
(ii) (G, r) is d-flattenable if and only if all the frameworks in Ω are.
Generic Convex Cayley Configuration: Let δr be as in the above definition. A framework (G, r)
of n vertices in d-dimensions is generic with respect to the property of convexity of ΦdF,l p (G, δ )
if
(i) There is an open neighborhood Ω of δr in the stratum Φdn,l p (this corresponds to an open
neighborhood of d-dimensional point-configurations of r); and
(ii) (G, r) has convex Cayley configuration space over F if and only if all the frameworks in Ω
do.

Generic Property: A property of frameworks is said to be generic if the existence of a generic

framework with the property implies that the property holds for all generic frameworks.

We refer to Chapter 1 for the definitions of rigidity and independence. For frameworks in polyhedral
Cayley Configuration Spaces 243

norms (including the l p norms), Kitson [25] has defined properties such as well-positioned, regu-
lar, analogous to the above, which have been used to show (infinitesimal) rigidity to be a generic
property of frameworks. Intuitively, a well-positioned d-dimensional framework under the l p norm
is one in whose d-dimensional neighborhood in l p -normed space the pairwise distances between
points can be expressed in polynomial form. We refer the reader to Kitson’s paper for a precise
definition [25].
In this section we show relationships between d-flattenability (and thus existence of convex
inherent Cayley configuration spaces) and combinatorial rigidity concepts via the cone Φn,l p . Note
that the definition of d-flattenability of a graph G in l p requires every l p framework of the graph G
– in an arbitrary dimension – to be d-flattenable. The next theorem weakens this requirement.

Theorem 10.8 Every generic framework of G is d-flattenable if and only if G is d-flattenable.

The “if” direction follows immediately from the definition of d-flattenability. For the “only
if” direction, notice that a nongeneric, (bounded) framework (G, r) is a limit of a sequence Q of
generic, bounded frameworks {(G, ri )}i , with a corresponding sequence of pairwise distance vectors
in ΦG,l p , i.e., a sequence Q0 of bounded linkages of G. Because each (G, ri ) is d-flattenable, each
linkage in Q0 must be realizable as some generic bounded framework (G, ri0 ) in d-dimensions The
projection of the limit framework (G, r) of the sequence Q is the limit linkage of the projected
sequence Q0 of linkages with bounded edge lengths.
Although d-flattenability is equivalent to the presence of an inherent convex Cayley configu-
ration space for G, we now move beyond inherent convex Cayley configuration spaces to Cayley
configuration spaces over specified nonedges F. These could be convex even if G itself is not d-
flattenable (simple examples can be found for d = 2, 3 for l2 in [31]). A complete characterization
of such G, F is shown in [31], in the case of l2 norm for d = 2, conjectured for d = 3, and completely
open for d > 3.
An analogous theorem to Theorem 10.8 can be proven for the property of a d-dimensional
framework (G, r) having a convex Cayley configuration space over specified non-edge set F.

Theorem 10.9 Every generic d-dimensional framework (G, r) has a convex Cayley configuration
space over F if and only if for all δ , the linkage (G, δ ) has a d-dimensional, convex Cayley config-
uration space over F.

The proof of this theorem follows very closely the proof of Theorem 10.8.
Neither of the properties discussed above is a generic property of frameworks even for l2 .

Theorem 10.10 d-flattenability and convexity of Cayley configuration spaces over specific non-
edges F are not generic properties of frameworks (G, r).

See Figure 10.7 for the d-flattenability argument. For convexity of Cayley configuration spaces:
there are minimal, so-called Henneberg-I graphs [33] G, constructed on a base or initial edge f with
the following property: for some 2-dimensional frameworks (and neighborhoods) (G, r) with edge
length vector δ , the 1-dimensional Cayley configuration space Φ2f (G \ f , δG\ f ) (i.e, the attainable
lengths for f ) is a single interval, while for other such frameworks (and neighborhoods) it is two
intervals. Please see Appendix in [33].
Next, we consider the implication of the existence of a generic d-flattenable framework. Specif-
ically, we prove two theorems connecting the d-flattenability with independence in the rigidity ma-
troid:
The “if” direction of this next theorem is a restatement of Proposition 2 in Asimow and Roth [3].

Theorem 10.11 For general l p norms, there exists a generic d-flattenable framework of G if and
only if G is independent in the d-dimensional generic rigidity matroid.
244 Handbook of Geometric Constraint Systems Principles

Figure 10.7
Two linkages of the same graph. In the first figure (left), we have edge lengths for (a, e) and (d, e)
that do not allow G to be flattened. The second graph is realized in 3-dimensions, but by “unfolding
it” as shown, we can flatten it into 2-dimensions. It is easy to see that there is a full dimensional
neighborhood of each linkage such that flattenability is maintained

Note that existence of a generic d-flattenable framework (G, r) is equivalent to the statement that
the pairwise distance vector δr has an open neighborhood Ωr in the relative interior of the stratum
Φdn,l p . The fact that Ωr is in the relative interior means it has full dimension.
Now observe that the generic rigidity matrix of G is the Jacobian of the distance map from the
d-dimensional point-configuration s to the edge-length vector δsG at the point s. For l2 and integral
p > 1, this map is clearly specified by polynomials. Because Ωr has dimension equal to the number
of edges in G, these polynomials are algebraically independent. Hence, their Jacobian has rank equal
to the number of edges in G, meaning the rows of the generic rigidity matrix – that correspond to
the edges of G – are independent.
Corollary 10.12 For general l p norms, a graph G is d-flattenable only if G is independent in the
d-dimensional rigidity matroid.
The following theorem and corollary utilize the dimension of the projection of the d-dimensional
stratum on the edges of G from the above proof. Note that in the above proof, if G is an n-vertex
graph, the neighborhood Ωr has dimension equal to the flattening dimension of Kn ; Ωs has dimen-
sion equal to that of the stratum Φdn,l p , and ΩG has dimension equal to the number of edges of G
(see Figure 10.9).
Theorem 10.13 For general l p norms, a graph G is
(a) Independent in the generic d-dimensional rigidity matroid (i.e., the rigidity matrix of a well-
positioned and regular framework has independent rows), if and only if coordinate projection
of the stratum Φdn,l p onto G has dimension equal to the number of edges of G;

(b) Maximal independent (minimally rigid) if and only if projection of the stratum Φdn,l p onto G
is maximal (i.e., projection preserves dimension) and is equal to the number of edges of G;
(c) Rigid in d-dimensions if and only if projection of the stratum Φdn,l p onto G preserves its di-
mension;
(d) Not independent and not rigid in the generic d-dimensional rigidity matroid if and only if
the projection of Φdn,l p onto G is strictly smaller than the minimum of: the dimension of the
stratum and the number of edges in G.
Corollary 10.14 For l p norms, the rank of a graph G in the d-dimensional rigidity matroid is equal
to the dimension of the projection ΦdG,l p on G of the d-dimensional stratum Φdn,l p .
Cayley Configuration Spaces 245

Figure 10.8
On the left we have two neighborhoods Ωr and Ωr0 of two distance vectors δr and δr0 in the cone.
We then project Ωr and Ωr0 onto the edges of G to obtain Ωl and Ωl 0 , which are essentially the
neighborhoods of (G, δrG ) and (G, δrG0 ). On the right, we then take the fiber of Ωl and Ωl 0 on Φdn,l p .
The fiber of Ωl is completely contained in the stratum while that of Ωl 0 misses (does not intersect)
the stratum.

10.6 Efficient Realization through Optimal Cayley Modification

Realizing a geometric constraint systems (GCS) requires finding real solutions to a large multi-
variate polynomial system (of equalities and inequalities representing the constraints); this requires
double exponential time in the number of variables, even if the type or orientation of the solution
is specified. Thus, geometric constraint systems are typically recursively decomposed into locally
rigid sub-systems and their solutions recombined to get the solution to the entire system. After
decomposition, the complexity of realizing the system typically gets dominated by the largest inde-
composible subsystem.
The Optimal Cayley Modification (OCM) [4] problem asks if the system to be realized can be
modified, by dropping some constraints and adding certain others, so as to make the system easily
realizable. This typically requires the modified system to have a convex Cayley configuration space.
To find realizations of the original GCS, we find generate the Cayley configuration space of the
modified GCS and search through it for realizations where the dropped edges have attained their
original length. An important thing to note here is that in general, the modified graphs that yield
“nice” Cayley spaces are also easy to realize. More formally:

Definition 10.1 Optimal Cayley Modification (OCM) Problem Given a geometric constraint
system G = (V, E) and two constants k and s, find some set of at most k constraints Edrop and some
set of constraints Eadd such that H = (V, (E \ Edrop ) ∪ Eadd ) is well-constrained and there exists
some DR-plan with maximum fan-in of s.

Let G0 = (V, E \ Edrop ), with parameters δ 0 , and F = Eadd . If ΦF (G0 , δ 0 ) is restricted to convex
Cayley configuration spaces, then the space can be searched efficiently. If H is further restricted to
be a class of efficiently realizable graphs, such as triangle decomposable graphs, then the Euclidean
coordinates of each point in the configuration space can quickly be determined. In this way, the
246 Handbook of Geometric Constraint Systems Principles

Figure 10.9
These are visualizations of when frameworks are isostatic and independent. In all of these cases
dim(Ωr ) ≥ max{dim(Ωs , dim(Ωl ))}. We only show 2 and 3 dimensions here, but in general the
dimensions will be much higher. See Figure 10.8 for explanation of what each is. In the following,
when we use equality or inequality, we are referring to dimension. On the left, Ωs = Ωl < Φdn,l p
meaning δr is independent but not isostatic. Middle left: Ωs = Ωl = Φdn,l p , so δr is maximal inde-
pendent or isostatic. Middle right: Ωs = Φdn,l p < Ωl meaning δr is rigid but not independent. Right:
Ωs < Ωl and Ωs < Φdn,l p meaning δr is neither independent nor rigid.

e1 e1 f1 f3

e2 e2 f2 e2

Figure 10.10
The first graph is a K3,3 . The second removes edges e1 and e2 , making a partial 2-tree. The third
adds in edges f1 and f2 , making a 2-tree. The fourth removes just edge e2 . The fifth adds edge f3 ,
creating a low Cayley-complexity graph.

space can be efficiently searched for realizations that additionally satisfy the dropped constraints.
This solution space must contain all solutions to the original constraint system G.
Such convex Cayley configuration spaces exist for partial 2-trees parametrized by edges that
make a 2-tree; because 2-trees are triangle-decomposable, these have fast realization schemes as
well. Alternatively, and perhaps more difficult to find, graphs whose added edges make low Cayley
complexity graphs (see Section 10.7). Consider Figure 10.10 for an example of each of these OCM
solutions.

10.7 Cayley Configuration Spaces of 1-Dof Tree-Decomposable

Linkages in 2D
Degree-of-Freedom (Dof) of a Linkage: The degree-of-freedom of a linkage is the minimum
number of bars that need to be added to the linkage to make it rigid.
Cayley Configuration Spaces 247

Tree-Decomposable Graphs: A graph G is tree-decomposable if it is a single edge, or it can be

divided into three tree-decomposable subgraphs G1 , G2 and G3 , such that G = G1 ∪ G2 ∪ G3 ,
G1 ∩ G2 = ({v3 }, ∅), G2 ∩ G3 = ({v2 }, ∅) and G1 ∩ G3 = ({v1 }, ∅), with v1 , v2 and v3 being
distinct vertices
1-Dof Tree-Decomposable Graphs: A graph G is 1-dof tree-decomposable if there exists a
nonedge f = (u, v) of G such that G ∪ f is tree-decomposable.

Base Nonedge of a 1-Dof Tree-Decomposable Graph: A nonedge f = (u, v) of a 1-dof tree-

decomposable graph G is a base nonedge of G if G ∪ f is tree-decomposable.
1-Dof Tree-Decomposable Linkages: A linkage (G, δ ) is 1-dof tree-decomposable if G is 1-dof
tree-decomposable.

Quadratically-Radically Solvable (QRS) Values: A real number is a QRS value if it is the solu-
tion a triangularized quadratic system with coefficients in Q (i.e., it belongs to an extension
field over Q obtained by nested square-roots).
Quadratically-Radically Solvable (QRS) graph: A graph G is QRS if for any linkage (G, δ )
where l is in Q, the coordinate values of a realization of (G, δ ) are QRS values.

Generic 1-Dof Tree-Decomposable Linkages: A 1-dof tree-decomposable linkage (G, δ ) is

generic if no bar length is zero, all bars have distinct lengths and at most one pair of adjacent
bars can be collinear in any realization.
Cluster: A cluster T of a 1-dof tree-decomposable graph G is a maximal tree-decomposable sub-
graph of G.

Graph Construction of a 1-Dof Tree-Decomposable Graph: Any 1-dof tree-

decomposable graph G can be constructed iteratively as follows, starting from a given base
nonedge f . At the kth construction step, two new clusters T1 and T2 sharing a single vertex vk
are appended to the previously constructed graph G f (k −1), such that T1 and T2 each has exactly
one shared vertex, uk and wk , respectively, with G f (k − 1), where uk 6= wk . Such a construction
step is denoted as vk / (uk , wk ).
Realization Type: A construction step v / (u, w) of a 1-dof tree-decomposable graph G can be
associated with a local orientation for the corresponding realization step of a linkage
 (G, δ ),

 pw − pu 
which takes a value in {+1, −1, 0} and represents the sign of the determinant ∆ =  .
pv − pu 
A realization type is a set of local orientations for all construction steps of G. An oriented
Cayley configuration space of a linkage (G, δ ) is the Cayley configuration space of (G, δ )
restricted to configurations respecting a given realization type.
Base Pair of Vertices of a Construction Step: The base pair of vertices of a construction step vk /
(uk , wk ) is the pair (uk , wk ).
Extreme Graphs and Extreme Edges: The kth extreme graph for a base nonedge f of a 1-dof
tree-decomposable graph G, denoted Ĝ f (k), is the graph obtained by adding a new edge ek =
(uk , wk ) in G f (k − 1), where (uk , wk ) is the base pair of vertices of the kth construction step of
G from f . The edge (uk , wk ) is called the extreme edge of the extreme graph.
Extreme Linkages: For a 1-dof tree-decomposable linkage (G, δ ), the kth extreme linkages are
(Ĝ f (k), l min ) and (Ĝ f (k), l max ), where min and max represent the two possible extreme exten-
sions of l for the extreme edge ek = (u, w), obtained from triangle inequalities.
248 Handbook of Geometric Constraint Systems Principles

Low Cayley Complexity: A 1-dof tree-decomposable graph G is said to have low Cayley (end-
point) on a base nonedge f if all extreme graphs of G for f are tree-decomposable.
Four-Cycle: A four-cycle of clusters consists of four clusters T1 , T2 , T3 , T4 such that each consecu-
tive pair has exactly one shared vertex. In other words, we have T1 ∩ T2 = {v1 }, T2 ∩ T3 = {v2 },
T3 ∩ T4 = {v3 }, T4 ∩ T1 = {v4 }, where v1 , v2 , v3 , v4 are distinct vertices.
Last Level Vertex: Given a 1-dof tree-decomposable graph G, a vertex v is a last level vertex if (a)
there are exactly two clusters T1 and T2 sharing v, and (b) each of T1 and T2 has exactly one
shared vertex with the rest of the graph.
1-Path Graph: A 1-dof tree-decomposable graph G has a 1-path construction from base nonedge
f = (v0 , v00 ) if there is only one last level vertex, other than possibly v0 and v00 . A 1-dof tree-
decomposable graph is 1-path if there exists a base nonedge permitting a 1-path construction.
Minimal Complete Cayley Vector: Given a general 1-dof tree-decomposable linkage (G, δ ) with
low Cayley complexity on a base nonedge f = (v0 , v00 ), a minimal complete Cayley vector F is
a list of O(|V (G)|) nonedges of G, whose addition to G as edges makes G globally rigid (see
the paper [34] for the detailed definition).

Minimal Complete Cayley Distance Vector: The minimal complete Cayley distance vector of a
configuration of a 1-dof tree-decomposable linkage is the vector of distances for the nonedges
in the complete Cayley vector.

The 1-degree-of-freedom (1-dof) tree-decomposable linkages [18, 23, 26] in R2 , which generically
has one degree-of-freedom, are the simplest natural generalization of partial 2-tree linkages that
have nonconvex Cayley configuration spaces, obtained by measuring the attainable distances of a
base nonedge. Since the underlying graphs are obtained by dropping an edge from the minimally
rigid and QRS tree-decomposable graphs, there exists a simple linear time ruler and compass re-
alization to convert from a Cayley configuration to a corresponding Cartesian configuration, given
a specified realization type. For planar graphs, QRS and tree-decomposability have been shown
[22, 27] to be equivalent, and the equivalence has been strongly conjectured for all graphs.
The following theorem, formalized in the paper [19], describes the structure of Cayley config-
uration spaces of generic 1-dof tree-decomposable linkages by associating a linkage configuration
with each endpoint of an (oriented) Cayley configuration space. The proof of theorem follows from
elementary algebraic geometry and can be considered folklore.

Theorem 10.15 (Structure of Cayley configuration space [19]) For a generic 1-dof tree-decomposable
linkage (G, δ ) with base nonedge f = (v0 , v00 ), the following hold:

(a) The (oriented) Cayley configuration space over f is a set of disjoint closed real intervals or
empty.
(b) Any interval endpoint in the (oriented) Cayley configuration space corresponds to the length
of f in a configuration of an extreme linkage.
(c) For any vertex v, pv is a continuous function of l f on each closed interval of the oriented
Cayley configuration space. Consequently, for any nonedge (u, w), l(u, w) is a continuous
function of l f on each closed interval of the oriented Cayley configuration space.

Theorem 10.15 gives a straightforward algorithm [19, 34] which computes the (oriented) Cayley
configuration space for a generic 1-dof tree-decomposable linkage (G, δ ) by realizing all the ex-
treme linkages. However, the algorithm could take time exponential in both the size of the linkage
and the final Cayley configuration space.
Cayley Configuration Spaces 249

The following theorem shows that low Cayley (endpoint) complexity is a robust measure of
the complexity of Cayley configuration spaces of 1-dof tree-decomposable graphs, as it does not
depend on the choice of base nonedge. That is, if we can characterize low Cayley complexity for
a 1-dof tree-decomposable G for a convenient base nonedge f , it translates to a characterization of
low Cayley complexity of G (for all possible base nonedges).

Theorem 10.16 (equivalence of base nonedges [19, 35]) A 1-dof tree-decomposable graph G ei-
ther has low Cayley complexity on all possible base nonedges, or on none of them.

The proof of Theorem 10.16 is via a minimal counterexample.

The following theorem characterizes 1-dof tree-decomposable graphs with low Cayley com-
plexity by looking at the “base four-cycle” of each construction step.

Theorem 10.17 (Four-cycle Theorem [35]) Given a 1-dof tree-decomposable graph G on a base
nonedge f with six or more clusters, G has low Cayley complexity, if and only if every construction
step vk / (uk , wk ) from f has its base pair of vertices taken from an adjacent pair of clusters in a four-
cycle of clusters. Note that a graph with less that six clusters trivially has low Cayley complexity.

One direction of Theorem 10.17 directly follows from the structure of extreme graphs, and the other
direction is proved by induction on the number of construction steps.
Theorem 10.17 yields an algorithm to verify whether a given 1-dof tree-decomposable graph G
has low Cayley complexity on base nonedge f , with O(|V |2 ) time complexity [35]. This is more
efficient than the algorithm that follows the definition of low Cayley complexity, i.e., checking if all
extreme graphs are tree-decomposable, which takes O(|V |3 ) time (checking O(|V |) extreme graphs,
each taking O(|V |2 ) time using existing algorithm [18]).
To go beyond the algorithmic characterization given by Theorem 10.17, it is desirable to obtain a
finite forbidden minor characterization of low Cayley complexity. Unfortunately, such a characteri-
zation is shown to be impossible for general 1-dof tree-decomposable graphs [19, 35]. Nevertheless,
for a natural subclass of 1-dof tree-decomposable graphs that are both 1-path and triangle-free, low
Cayley (algebraic) complexity is shown to be equivalent to planarity.

Theorem 10.18 (Low Cayley complexity and planarity [35]) A 1-path, triangle-free, 1-dof tree-
decomposable graph G has low Cayley complexity if and only if G is planar.

The proof of Theorem 10.18 is by minimal counterexample and observations on the recursive struc-
ture of 1-path, triangle-free graphs.
The following theorem shows that for generic 1-dof tree-decomposable linkages with low Cay-
ley complexity, a path of continuous motion between two given configurations can be efficiently
obtained from the complete description of the Cayley configuration space.

Theorem 10.19 (Continuous motion path Theorem [32, 34]) For a generic linkage (G, δ ) where
G has low Cayley complexity on f ,
(a) There exist at most two paths of continuous motion between any two given configurations,
and the time complexity of finding such a path (provided one exists) is linear in the number of
interval endpoints of oriented Cayley configuration spaces that the path contains.
(b) There is an algorithm that generates the entire set of connected components of the configura-
tion space, where generating each connected component takes time linear in the number of interval
endpoints of oriented Cayley configuration spaces that connected component contains.

The proof of Theorem 10.19 is by showing certain local orientations of the configuration, under the
genericity condition, can only be changed via at interval endpoints of oriented Cayley configuration
spaces.
250 References

For a 1-dof tree-decomposable linkage with low Cayley complexity, assume that the clusters
are globally rigid, and clusters sharing only two vertices with the rest of the graph be reduced
into edges, there is a bijective correspondence between the Cartesian configuration space and a
minimal complete Cayley distance vector. This yields a canonical representation and a meaningful
visualization of the configuration space and continuous motion.

Theorem 10.20 (bijectivity of representation [32, 34]) (1) For a generic 1-path, 1-dof tree-
decomposable linkage with low Cayley complexity, there exists a bijective correspondence between
the set of Cartesian configurations and points on a curve in R2 , whose points are the minimum
complete Cayley distance vectors.
(2) For a generic 1-dof tree-decomposable linkage with low Cayley complexity, there exists a bi-
jective correspondence between the set of Cartesian configurations and points on a curve in n-
dimension, whose points are the minimal complete Cayley distance vectors, where n is the number
of last level vertices of the underlying graph.

The complete Cayley distance vector also provides a meaningful way to define a canonical distance
between any two configurations of 1-dof tree-decomposable linkages, as well as distances between
different connected components of the configuration space [32, 34].

10.8 Conclusion
We characterized graphs whose linkages have convex Cayley configuration spaces, connected the
property to flattenability for arbitrary l p norms and dimensions, applied the results towards efficient
realization of linkages, studied genericity and independence in the context of d-flattenability, and
characterized a larger class of graphs but with low complexity Cayley configurations.

References
[1] Abdo Y. Alfakih. Graph rigidity via euclidean distance matrices. Linear Algebra and its
Applications, 310(13):149–165, 2000.
[2] A.Y. Alfakih. On bar frameworks, stress matrices and semidefinite programming. Mathemat-
ical Programming, 129(1):113–128, 2011.
[3] L. Asimow and B. Roth. The rigidity of graphs. Trans. Amer. Math. Soc., 245, 1978.
[4] T. Baker, M. Sitharam, M. Wang, and J. Willoughby. Optimal decomposition and recombi-
nation of isostatic geometric constraint systems for designing layered materials. Computer
Aided Geometric Design, 40:1–25, 2015.
[5] Keith Ball. Isometric embedding in lp-spaces. European Journal of Combinatorics,
11(4):305–311, 1990.
[6] Boaz Barak, Prasad Raghavendra, and David Steurer. Rounding semidefinite programming
hierarchies via global correlation. CoRR, abs/1104.4680, 2011.
[7] A.I. Barvinok. Problems of distance geometry and convex properties of quadratic maps.
Discrete & Computational Geometry, 13(1):189–202, 1995.
References 251

[8] Maria Belk. Realizability of graphs in three dimensions. Discrete Comput. Geom., 37(2):139–
162, February 2007.
[9] Maria Belk and Robert Connelly. Realizability of graphs. Discrete Comput. Geom.,
37(2):125–137, February 2007.
[10] L.M. Blumenthal. Theory and applications of distance geometry. Chelsea Pub. Co., 1970.
[11] Bo Brinkman and Moses Charikar. On the impossibility of dimension reduction in l1. J. ACM,
52(5):766–788, September 2005.
[12] Arthur Cayley. On a theorem in the geometry of position. Cambridge Mathematical Journal,
2:267–271, 1841.
[13] Ugandhar Reddy Chittamuru. Efficient bounds for 3d cayley configuration space of partial
2-trees. Master’s thesis, University of Florida, 2010.
[14] Jon Dattorro. Convex Optimization & Euclidean Distance Geometry. Meboo Publishing USA,
2011.
[15] Michel Marie Deza and Monique Laurent. Geometry of Cuts and Metrics. Algorithms and
Combinatorics. Springer, Dordrecht, 2010.
[16] Dmitriy Drusvyatskiy, Gbor Pataki, and Henry Wolkowicz. Coordinate shadows of semidefi-
nite and euclidean distance matrices. SIAM Journal on Optimization, 25(2):1160–1178, 2015.
[17] Samuel Fiorini, Tony Huynh, Gwenal Joret, and Antonios Varvitsiotis. The excluded minors
for isometric realizability in the plane. SIAM Journal on Discrete Mathematics, 31(1):438–
453, 2017.
[18] Ioannis Fudos and Christoph M. Hoffmann. A graph-constructive approach to solving systems
of geometric constraints. ACM Trans. Graph., 16(2):179–216, April 1997.
[19] Heping Gao and Meera Sitharam. Characterizing 1-dof Henneberg-I graphs with efficient
configuration spaces. In Proceedings of the 2009 ACM symposium on Applied Computing,
SAC ’09, pages 1122–1126, New York, NY, 2009. ACM.
[20] Karin Gatermann and Pablo A. Parrilo. Symmetry groups, semidefinite programs, and sums
of squares. Journal of Pure and Applied Algebra, 192(13):95–128, 2004.
[21] Piotr Indyk and Jiri Matousek. Low-distortion embeddings of finite metric spaces. In in
Handbook of Discrete and Computational Geometry, pages 177–196. CRC Press, 2004.
[22] Bill Jackson and JC Owen. Radically solvable graphs. arXiv preprint arXiv:1207.1580, 2012.
[23] R. Joan-Arinyo, A. Soto-Riera, S. Vila-Marta, and J. Vilaplana-Pastó. Revisiting decomposi-
tion analysis of geometric constraint graphs. Computer-Aided Design, 36(2):123–140, 2004.
Solid Modeling and Applications.
[24] William Johnson and Joram Lindenstrauss. Extensions of Lipschitz mappings into a Hilbert
space. In Conference in modern analysis and probability (New Haven, Conn., 1982), vol-
ume 26 of Contemporary Mathematics, pages 189–206. American Mathematical Society,
1984.
[25] D. Kitson. Finite and infinitesimal rigidity with polyhedral norms, 2014.
[26] J. C. Owen. Algebraic solution for geometry from dimensional constraints. In Proceedings
of the First ACM Symposium on Solid Modeling Foundations and CAD/CAM Applications,
SMA ’91, pages 397–407, New York, NY, USA, 1991. ACM.
252 References

[27] J.C. Owen and S.C. Power. The non-solvability by radicals of generic 3-connected planar
laman graphs. Transactions of the American Mathematical Society, 359(5):2269–2304, 2007.
[28] Pablo A. Parrilo and Bernd Sturmfels. Minimizing polynomial functions. In Proceedings of
the Dimacs Workshop on Algorithmic and Quantitative Aspects of Real Algebraic Geometry
in Mathematics and Computer Science, pages 83–100. American Mathematical Society, 2003.
[29] Neil Robertson and P.D. Seymour. Graph minors. xx. wagner’s conjecture. Journal of Combi-
natorial Theory, Series B, 92(2):325 – 357, 2004. Special Issue Dedicated to Professor W.T.
Tutte.
[30] I. J. Schoenberg. Remarks to maurice frchets article sur la dfinition axiomatique dune classe
despaces distancis vectoriellement applicable sur lespace de hilbert. annals of mathematics
36(3), 1935.
[31] Meera Sitharam and Heping Gao. Characterizing graphs with convex and connected cayley
configuration spaces. Discrete & Computational Geometry, 43(3):594–625, 2010.
[32] Meera Sitharam and Menghan Wang. How the beast really moves: Cayley analysis of mecha-
nism realization spaces using caymos. Computer-Aided Design, 46(0):205 – 210, 2014. 2013
SIAM Conference on Geometric and Physical Modeling.
[33] Meera Sitharam, Menghan Wang, and Heping Gao. Cayley configuration spaces of 1-dof
tree-decomposable linkages, part II: combinatorial characterization of complexity. CoRR,
abs/1112.6009, 2011.
[34] Meera Sitharam, Menghan Wang, and Heping Gao. Cayley configuration spaces of 2d mecha-
nisms Part I: extreme points, continuous motion paths and minimal representations. Submitted
to Discrete and Computational Geometry, 2013.
[35] Meera Sitharam, Menghan Wang, and Heping Gao. Cayley configuration spaces of 2d mech-
anisms Part II: combinatorial characterization. Submitted to Discrete and Computational
Geometry, 2013.
[36] Meera Sitharam and Joel Willoughby. On flattenability of graphs. In Automated Deduction
in Geometry - 10th International Workshop, ADG 2014, Coimbra, Portugal, July 9-11, 2014,
Revised Selected Papers, pages 129–148, 2014.
[37] Pablo Tarazaga. Faces of the cone of euclidean distance matrices: Characterizations, structure
and induced geometry. Linear Algebra and its Applications, 408(0):1–13, 2005.
Chapter 11
Constraint Varieties in Mechanism
Science
Hans-Peter Schröcker, Martin Pfurner, and Josef Schadlbauer
University of Innsbruck

CONTENTS
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
11.1.1 Linkages and Joints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
11.1.2 Base and End-Effector Frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
11.2 Mechanisms and Algebraic Varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
11.2.1 Geometric Constraint Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
11.2.2 Study Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
11.2.3 Dual Quaternions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
11.2.4 Analyzing Mechanisms via Algebraic Varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
11.3 Serial Manipulators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
11.3.1 Direct and Inverse Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
11.3.2 Synthesis of Open and Closed Serial Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
11.3.3 Singularities and Path Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
11.3.4 Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
11.4 Parallel Manipulators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
11.4.1 Direct and Inverse Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
11.4.2 Singularities and Self-Motions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
11.4.3 Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269

11.1 Introduction
In this chapter we look at constraint varieties of linkages from the viewpoint of mechanism science.
There, not only abstract properties of linkages like flexibility (finite or infinitesimal) or configuration
space topology are studied but also quantities that pertain to a concrete realization and to applica-
tions are of relevance. Examples include transmission ratios, joint forces, stiffness, collisions, or
size. Moreover, additional problems like the construction of a linkage to accomplish a certain task
(“linkage synthesis”) or the computation and avoidance of “singular” configurations appear. Another

253
254 Handbook of Geometric Constraint Systems Principles

specialty is the relevance of linkages with several degrees of freedom and the presence of different
joint types. Linkages with as many degrees of freedom as the underlying motion group (three in
case of SE(2) or SO(3) and six in case of SE(3)) are quite common. Redundant manipulators with
even more degrees of freedom are an active research topic.
Mechanism science is a rewarding field for many branches of applied mathematics. Here,
we mainly focus on algebraic and geometric aspects via Study parameters and dual quaternions.
Progress in the field of computational algebraic geometry made this approach popular over recent
years but, of course, traditional techniques from differential geometry, vector calculus or numerical
mathematics remain indispensable as well.

Glossary
Fixed frame/base: Space or coordinate frame attached to one link that is considered to be immo-
bile.
Joint: Abstract representation of a mechanical device, attached to two or more links and constrain-
ing their relative positions. For example revolute or prismatic joint.
Link: Rigid body, abstractly representing part of a linkage.
Linkgraph: Graph representing linkage topology. The vertices are the links, the edges are the
attached joints.
Moving frame/end effector: Space or coordinate attached to the link whose motion is of primary
interest.
Parallel linkage: Linkage whose linkgraph has (several) cycles. Also called “parallel manipulator”
or “parallel mechanism.”
Serial linkage: Linkage whose linkgraph is linear. Also called “serial manipulator” or “serial
mechanism.”

11.1.1 Linkages and Joints

A linkage or mechanism is a set of rigid bodies, called links whose relative position is constrained
by joints which are said to “connect” links. In order to fully specify a linkage, one must prescribe
which links are connected by joints and also the type of the joint and its position with respect to
the link. The abstract topology can be captured in the linkgraph whose vertices correspond to links.
Two vertices are connected by an edge precisely if their links are connected by a joint. Mechanisms
whose link graph have cycles are called parallel while mechanisms with linear link graph are called
serial.
The predominant joint type in engineering applications is the revolute joint but other joint types
are considered as well. The most common schemes classify joints by continuous subgroups of
SE(3). Besides revolute joints, this gives rise to prismatic, helical, cylindrical, spherical and pla-
nar joints (Figure 11.1).
Note that some of these joints can be modelled as a concatenation of others. We may, for exam-
ple, replace a cylindrical joint by a revolute joint and a prismatic joint provided the revolute axis is
parallel to the translation direction. A spherical joint is equivalent to three revolute joints with in-
tersecting axes. Other “compound joints” are possible as well. A common example is the universal
or Cardan joint which is the composition of rotations about two perpendicular axes. In this text, we
are mostly concerned with revolute and prismatic joints. Helical joints are not easily treated within
an algebraic framework because they do not correspond to an algebraic subgroup of SE(3) (but see
[3] for a remarkable exception).
Constraint Varieties in Mechanism Science 255

Figure 11.1
Revolute, prismatic, and cylindrical joint (top row); spherical, planar, and helical joint (bottom row).

11.1.2 Base and End-Effector Frame

When studying linkages from a purely mathematical viewpoint, it is good and common practice
to consider isometric realizations of the linkage as equivalent. From the viewpoint of mechanism
science, this is not always desirable. Industrial specifications of mechanisms or robots require co-
ordinate frames attached to each link. Moreover, two links often play a particular role and deserve
special treatment. The end-effector, is assumed to fulfill a certain task which is often described in
terms of coordinates with respect to a moving frame, that is, a coordinate system attached to the
end-effector link. Moreover, a mechanical linkage is usually attached to an immobile object via
one of its links, the base. Its position is again described in some coordinate frame, often called base
frame, fixed frame, or world coordinate system. The spaces attached to the moving and fixed frames,
respectively, are called the moving space and fixed space.

11.2 Mechanisms and Algebraic Varieties

Glossary
Point model: A usually projective or affine space or a subset thereof, whose elements (points) are
in bijection with the elements of a motion group. For example, projective three-space P3 and the
space of orthogonal 3 × 3 matrices with positive determinant are both point models for SO(3).
Constraint equations: Equations describing a subset of SE(3) in a suitable point model that can
be reached by a linkage. Algebraic constraint equations give rise to constraint varieties.

Dual quaternions: Algebra obtained from quaternions via scalar extension from real to dual num-
bers. Can be viewed as Study parameters if they satisfy the Study condition.
Four-bar linkage: Linkage, usually with revolute joints, whose linkgraph is a cycle of length four.
Homogeneous transformation matrix: A 4 × 4 matrix representing a projective transformation of
P3 . Usually used for representing rigid body displacements in mechanism science.
256 Handbook of Geometric Constraint Systems Principles

Figure 11.2
A planar four-bar linkage.

Study parameters: Homogeneous coordinates in P7 representing points on the Study quadric. May
be identified with certain dual quaternions.
Study quadric: Hyperquadric in P7 providing a point model of SE(3) via Study’s kinematic map.

11.2.1 Geometric Constraint Equations

A mathematical description of a linkage and its possible configurations can be accomplished by
translating the joint constraints into constraint equations. We are particularly interested in constrain-
ing position and orientation of the moving end-effector link. Consider the planar four-bar linkage
shown in Figure 11.2. It will serve as our model example in this section. Four links are cyclically
connected by revolute joints at points A, B, C, and D. The link of A and B is assumed to be fixed
while the link of C and D is the moving end-effector.
The possible positions of the end-effector are completely described by two “point on circle”
constraints: The point D lies on a circle of radius a around A and the point C lies on a circle of
radius b around B. Thus, we may write

D = A + a(cos ϕ, sin ϕ)| , C = B + b(cos ψ, sin ψ)| ,

where ϕ and ψ denote the angles the link A-B forms with A-D and B-C, respectively. The kinematics
of the four-bar linkage follow from the equation kC − Dk2 = h2 which relates the angles ϕ and
ψ. Note that this equation is trigonometric but can easily be converted to an algebraic form. One
method is to introduce new variables C1 := cos ϕ, S1 := sin ϕ, C2 := cos ψ, S2 := sin ψ and equations
S12 + C12 = S22 + C22 = 1. Another possibility is the tangent half-angle substitution u := tan ϕ2 , v :=
tan ψ2 whence cos ϕ = (1 − u2 )/(1 + u2 ), sin ϕ = 2u/(1 + u2 ) and similar for cosine and sine of ψ.
We will present an alternative and more elaborate algebraic formulation later (see Equa-
tion (11.11)). Here it suffices to say that circle and sphere constraints, their spatial counterparts,
are ubiquitous in mechanism science. Several examples will be given in Section 11.4. However,
numerous other constraints like “point in plane” or “line through point” are conceivable as well.

11.2.2 Study Parameters

Now we introduce a mathematical tool to handle geometric constraints imposed by kinematic joints.
The idea is that any geometric constraint restricts the relative positions of the attached links. Thus,
it can be described by a subset of a suitable point model of SE(3). We primarily focus on geometric
constraints that impose algebraic constraints on the motion group. Luckily, these comprise numer-
ous constraints of practical relevance. The point model we propose is obtained via Study’s kinematic
Constraint Varieties in Mechanism Science 257

map, – a bijection between SE(3) and the points of a hyperquadric S ⊂ P7 , the Study quadric, minus
the points of a projective three-space E, the exceptional generator.
Recall that any rigid body displacement maps a point with Cartesian coordinate vector x (with
respect to the moving coordinate system) to a point with Cartesian coordinates x0 in the fixed system
according to x0 = A · x + a. Here the matrix A is orthogonal, of dimension 3 × 3, and has positive
determinant. It is responsible for the orientation component of the displacement and the vector
a ∈ R3 gives its translation. Often, the rotation matrix A and translation vector a are merged into a
four-by-four matrix according to  0    
x A a x
= · .
1 0 1 1
In this way, it may also act on homogeneous coordinate vectors and one speaks of a homogeneous
transformation matrix. Composition of rigid body displacements corresponds to matrix multiplica-
tion.
It can be shown that there exists a vector p = (p0 , p1 , p2 , p3 )| ∈ R4 such that
 2
p + p21 − p22 − p23 −2p0 p3 + 2p1 p2

2p0 p2 + 2p1 p3
1 0
A= 2p0 p3 + 2p1 p2 p20 − p21 + p22 − p23 −2p0 p1 + 2p2 p3  , (11.1)
∆
−2p0 p2 + 2p1 p3 2p0 p1 + 2p2 p3 p20 − p21 − p22 + p23

where ∆ = p20 + p21 + p22 + p23 . Clearly, replacing p by any non-zero real multiple gives the same
matrix. Conversely, any non-zero vector p ∈ R4 yields, via (11.1), an orthogonal matrix with deter-
minant one. Thus, we have established a bijection between SO(3) and the points of real projective
three-space P3 . The homogeneous coordinates (p0 , p1 , p2 , p3 )| are called the rotation’s Euler pa-
rameters.
Now we also include translations. For any translation vector a = (a1 , a2 , a3 )| , there exists a
unique vector q = (q0 , q1 , q2 , q3 )| ∈ R4 such that

p0 q0 + p1 q1 + p2 q2 + p3 q3 = 0
2(p1 q0 − p0 q1 + p3 q2 − p2 q3 ) = a1
(11.2)
2(p2 q0 − p3 q1 − p0 q2 + p1 q3 ) = a2
2(p3 q0 + p2 q1 − p1 q2 − p0 q3 ) = a3 .

Indeed, (11.2) is a linear system for computing q and its determinant −8(p20 + p21 + p22 + p23 )2 is
different from zero. Equations (11.1) and (11.2) together define a bijection

κ : S \ E → SE(3)
(11.3)
(p0 , p1 , p2 , p3 , q0 , q1 , q2 , q3 ) 7→ (A, a)

from the Study quadric S ⊂ P7 : p0 q0 + p1 q1 + p2 q2 + p3 q3 = 0 minus the exceptional generator

E ⊂ S : p0 = p1 = p2 = p3 = 0 to SE(3). It is called Study’s kinematic map [46]. The defining
equation of S (first equation in (11.2)) is called the Study condition. The coefficients p0 , . . . , p3 ,
q0 ,. . . ,q3 are called Study parameters. Note that they depend on the coordinate systems in the fixed
and moving frame.
Study’s construction may seem rather weird at first sight. However, it is well-motivated by two
important properties. The point model for SE(3) requires a minimal number of parameters (eight
homogeneous coordinates) while providing a bilinear composition law (see Section 11.2.3). Study
showed that these two properties determine the point model up to isomorphism [46]. Other common
point models violate the first or the second condition. For example the composition of homogeneous
transformation matrices is bilinear but each matrix requires twelve parameters, even if we do not
count the constant last row. In computer graphics and gaming industry, Euler angles (“roll,” “pitch,”
258 Handbook of Geometric Constraint Systems Principles

and “yaw”) plus translation vectors are popular. They require only six parameters but have a non-
linear composition.
Let us proceed with some useful formulas for computations in Study parameters: A procedure
for obtaining the vector p = (p0 , p1 , p2 , p3 )| from the matrix A = (ai j )i, j=1,...,3 goes back to Study.
The ratio of the coefficients of p is given by any of the four proportions

p0 : p1 : p2 : p3 = 1 + a11 + a22 + a33 : a32 − a23 : a13 − a31 : a21 − a12

= a32 − a23 : 1 + a11 − a22 − a33 : a12 + a21 : a31 + a13
(11.4)
= a13 − a31 : a12 + a21 : 1 − a11 + a22 − a33 : a23 + a32
= a21 − a12 : a31 + a13 : a23 + a32 : 1 − a11 − a22 + a33 .

In general, all four proportions of (11.4) give the same result. It is, however, possible that up to
three proportions yield 0 : 0 : 0 : 0 and become invalid. A related procedure paying more attention
to numeric data is described in [34]. Having computed p, we may use the equations in (11.2) to find
q for given a.
It is useful to know the Study parameters of some fundamental rigid body displacements. The ro-
tation about the unit vector (v1 , v2 , v3 )| (line through the origin) with rotation angle 2ϕ corresponds
to
p0 = cos ϕ, p1 = v1 sin ϕ, p2 = v2 sin ϕ, p3 = v3 sin ϕ (11.5)
and q = (0, 0, 0, 0)| . The translation by the vector (t1 ,t2 ,t2 )| corresponds to p = (1, 0, 0, 0)| and
q = − 12 (0,t1 ,t2 ,t3 )| . Any displacement of SE(3) is a suitable and intuitive composition of rotations
about lines through the origin and translations. Hence, we need the composition formula for two
displacements in Study parameters. This we do by relating Study’s kinematic map to quaternions
and dual quaternions.

11.2.3 Dual Quaternions

In the Euclidean vector space R4 we pick an orthonormal basis (1, i, j, k) and we embed R3 into R4
by identifying it with the subspace spanned by i, j and k via (x1 , x2 , x3 )| ∈ R3 ,→ (0, x1 , x2 , x3 )| ∈ R4 .
Moreover, we embed R into R4 via x ∈ R ,→ (x, 0, 0, 0)| ∈ R4 . The ambiguous use of the symbol “1”
will be justified as soon as we have extend the multiplication of real numbers to R4 . More precisely,
we define an associative and bilinear multiplication that also extends the familiar multiplication of
real numbers and vectors by the rules

i2 = j2 = k2 = ijk = −1. (11.6)

From (11.6) we may derive a multiplication table and multiplication formula:

1 i j k ab = a0 b0 − a1 b1 − a2 b2 − a3 b3
1 1 i j k +(a0 b1 + a1 b0 + a2 b3 − a3 b2 )i
(11.7)
i i −1 k −j +(a0 b2 − a1 b3 + a2 b0 + a3 b1 )j
j j −k −1 i +(a0 b3 + a1 b2 − a2 b1 + a3 b0 )k,
k k j −i −1
where a = a0 + a1 i + a2 j + a3 k and b = b0 + b1 i + b2 j + b3 k. The elements of the thus defined
associative algebra are called quaternions and, in remembrance of their discoverer W. R. Hamilton,
the algebra itself is denoted by H. It contains two important and familiar sub-algebras: For a, b ∈ R,
quaternion multiplication reduces to the multiplication of real numbers. For a = a0 + a1 i, b = b0 +
b1 i, quaternion multiplication reduces to the multiplication of complex numbers. This is also true if
the quaternion unit i is replaced by j or k.
Every quaternion can be written as h = h0 +h1 i+h2 j+h3 k where h0 , h1 , h2 , h3 are real numbers.
Constraint Varieties in Mechanism Science 259

The quaternion h := p h0 − h1 i − h2 j − h3 k is called the conjugate quaternion of h, the norm of h is

defined as khk := N(h) where N(h) := hh. This definition is possible because N(h) = h20 + h21 +
h22 + h23 is a non-negative real number. For any two quaternions a and b we have N(ab) = N(a)N(b),
kabk = kak kbk, and (ab) = b a. If a 6= 0 there exists a unique quaternion, denoted by a−1 , such that
aa−1 = 1. It is given by a−1 = a/N(a) and also satisfies a−1 a = 1.
One of the main applications of quaternions in mechanism science is the modelling of SO(3).
The action of a quaternion p = p0 + p1 i + p2 j + p3 k ∈ H on the vector x = x1 i + x2 j + x3 k ∈ R3 ⊂ H
is defined as
x 7→ y = y1 i + y2 j + y3 k := pxp/N 2 (p). (11.8)
The right-hand side is linear in x1 , x2 , and x3 . Hence, there exists a matrix A such that (y1 , y2 , y3 )| =
A · (x1 , x2 , x3 )| . A short computation confirms that A is the matrix given in (11.1). Combing with
(11.5), this implies:
• The action (11.8) describes a rotation about the vector (p1 , p2 , p3 ) whose rotation angle 2ϕ
satisfies cos ϕ = p0 /kpk.
• The composition of rotations in Study parameters corresponds to multiplication of quater-
nions. Indeed, we have (rp)x(rp) = r(pxp)r for any quaternions p, r ∈ H and any vector
x ∈ R3 .
We thus see that Study parameters for displacements from SO(3) and quaternions are essentially
the same. This provides us with a convenient algebraic interpretation for Study parameters.
The quaternion model exhibits quite some advantages over other orthogonal matrices. Apart
from the already mentioned smaller number of parameters, the composition of rotations in Study
parameters is more efficient than the multiplication of orthogonal matrices. Moreover, quaternions
provide a very intuitive representation of SO(3). Their vector part already gives the oriented direc-
tion of the rotation axis. The cosine of half the rotation angle can be read off after normalization.
Finally “orthogonal” matrices in applications usually have floating point entries and therefore re-
quire periodic re-orthogonalization. Similarly, quaternions require normalization because one usu-
ally wants to avoid the division in (11.8). However, the latter is a simple task while a geometrically
meaningful orthogonalization of near-orthogonal matrices is not obvious.
Now we extend the standard quaternion algebra by allowing dual numbers as coefficients. This
extension of scalars will provide us with the desired point model for SE(3). The resulting algebra
DH of dual quaternions shares many properties with the quaternions H but some essential details
are different. Most notably, non-invertible elements do exist.
The ring of dual numbers is defined as D := R[ε]/hε 2 i. For every d ∈ D there exist unique
a, b ∈ R (primal and dual part) such that d = a + b · ε. The product of the two dual numbers
d1 = a1 + b1 ε and d2 = a2 + b2 ε equals d1 d2 = (a1 + b1 ε)(a2 + b2 ε) = a1 a2 + ε(a1 b2 + b1 a2 ). The
inverse to d = a + bε exists if and only if a 6= 0. In this case it is d −1 = a−1 − εba−2 . The algebra
of dual quaternions is defined as the scalar extension DH := D ⊗R H. For every h ∈ DH there exist
unique p, q ∈ H (primal and dual part) such that h = p + εq. The product of two dual quaternions
h1 = p1 + εq1 and h2 = p2 + εq2 equals

h1 h2 = (p1 + εq1 )(p2 + εq2 ) = p1 p2 + ε(p1 q2 + q1 p2 ).

Note that addition and multiplication of dual quaternions commute with the projection on the primal
part. The inverse to h = p + εq exists if and only if p 6= 0, whence we have h−1 = p−1 − ε p−1 qp−1 .
Thep dual quaternion conjugate to h = p + εq is h := p + εq. The dual quaternion norm is
khk := N(h), where N(h) := hh. This definition of the dual quaternion norm requires extension
of the analytic square root function to dual numbers because of N(h) = hh = (p + εq)(p + εq) =
pp + ε(pq + qp) ∈ D. Fortunately, this is not necessary in kinematics because for dual quaternions
satisfying the Study condition we have N(h) ∈ R.
260 Handbook of Geometric Constraint Systems Principles

The motivation for the scalar extension from quaternions to dual quaternions is the modelling of
SE(3) and not just SO(3). To this end, we identify Euclidean three space R3 (with fixed Cartesian
coordinates) with the affine three-space 1 + hεi, εj, εki via the inclusion map (x1 , x2 , x3 ) ,→ 1 +
ε(x1 i + x2 j + x3 k). The action of the dual quaternion h = p + εq on x = 1 + ε(x1 i + x2 j + x3 k) is
defined as
x 7→ hε xh, (11.9)
where hε := p − εq. With x = 1 + ε(x1 i + x2 j + x3 k), the right-hand side of (11.9) is linear in x1 ,
x2 , x3 and it is elementary to verify that it may be written as A · (x1 , x2 , x3 )| + a, where A is given
by (11.1) and a = (a1 , a2 , a3 )| by (11.2). Hence, the Study parameters for rigid body displacements
may be interpreted as dual quaternion coefficients. In particular, we obtain the composition law for
Study parameters which again is given by dual quaternion multiplication due to

(hk)ε x (hk) = hε (kε x k)h

for all h, k ∈ DH. Moreover, the restriction of (11.9) to H gives the action of quaternions on vectors
via (11.8), as expected.
We may also use the composition of displacements in Study parameters via dual quaternion
multiplication for coordinate changes in the moving and fixed frame, respectively. Suppose y = hε xh
are the coordinates (in the fixed frame) of the image of a point x (in the moving frame) and select new
coordinates y0 , and x0 in the fixed and moving frame, respectively. Then there exist dual quaternions
f , m that satisfy the Study condition and y0 = fε y f , x = mε x0 m. The same displacement in new
coordinates then reads y0 = h0ε x0 h0 where h0 = f hm because of

y0 = fε y f = fε (hε xh) f = fε (hε (mε x0 m)h) f = ( f hm)ε x0 ( f hm).

We see that coordinate changes in the fixed or moving coordinate frame are done by dual quaternion
left or right multiplication, respectively. These are projective transformations of P7

ϕ : P7 → P7 , [h] 7→ [ f h]; µ : P7 → P7 , [h] 7→ [hm] (11.10)

that leave invariant the Study quadric S and the exceptional generator E by construction. A com-
plete system of invariants of the transformation group generated by (11.10) has only recently been
described in [39]. A more detailed introduction to Study parameters and dual quaternions may be
found in [23] or [43].

11.2.4 Analyzing Mechanisms via Algebraic Varieties

In the previous section we have constructed a point model of SE(3) with rich algebraic and geo-
metric structure (the algebra DH of dual quaternions and the Study quadric S ⊂ P7 , respectively).
This model can be used to systematically analyze the kinematics of linkages based on their con-
straint varieties. A constraint variety is an algebraic variety V in the Study quadric S ⊂ P7 such that
all rigid body displacements in V satisfy certain kinematic conditions that are characteristic of the
given linkage. This makes numerous aspects of computational kinematics accessible to the power-
ful framework of computational algebraic geometry. Note that the minimal number of parameters in
Study’s point model is advantageous in this context. Fundamental concepts of algebraic geometry
like “dimension,” “singularity,” or “prime decomposition” have a concrete kinematic meaning and
are important in theoretical and applied mechanism science.
In this section we mostly confine ourselves to a simple example, the planar four-bar linkage
(Figure 11.2). More comprehensive studies of constraint varieties can be found in [44, 45] and
also later in this chapter. We already recognized the importance of the “point on circle” constraint
for the kinematic description of the planar four-bar. Hence, we start by looking at the constraint
variety of a linkage with a single revolute joint. Since we are only considering planar kinematics,
Constraint Varieties in Mechanism Science 261

Figure 11.3
Constraint varieties of four-bar linkages: General four-bar with two assembly modes; rational con-
straint variety; twisted cubic and straight line; two conics (from left to right).

we can restrict to the sub-algebra of DH which is generated by 1, k, εi, and εj (rotations about
the third coordinate axis and translations in direction of first and second coordinate axis). Write
h = p0 + p3 k + ε(q1 i + q2 j) and consider the action (11.9) on a point x = 1 + ε(x1 i + x2 j) in the
moving plane. By appropriately choosing coordinate origins in the moving and the fixed frame,
respectively, we may assume x1 = x2 = 0 and consider the condition that the image of x lies on the
circle with center (0, 0) and radius r. This gives the prototype of a circle constraint variety

V : (p20 + p23 )r2 − 4(q21 + q2 )2 = 0. (11.11)

The general form, without referring to special coordinates, can be obtained in a similar fashion or
by subjecting the variety (11.11) to the projective transformations (11.10).
We see that V is a regular, ruled quadric that contains the conjugate complex points
(p0 , p3 , q1 , q3 )| = (0, 0, 1, ±i)| – a property which is, in fact, invariant with respect to the transfor-
mations (11.10) and hence of kinematic relevance, see for example the direct kinematics problem
of the planar RPR-manipulator in Section 11.4.1.
The constraint variety of a planar four-bar linkage is the intersection of two circle constraint
varieties (Figure 11.3). This allows the immediate derivation of a number of relevant properties of
four-bar linkages. We give a detailed discussion at an intuitive level. To begin with, the constraint
variety of a planar four-bar linkage is an algebraic curve of dimension one and degree four. In
general it is of genus one and may hence have two disconnected real components. These correspond
to what is known as “assembly modes” in kinematics (leftmost image in Figure 11.3, left row in
Figure 11.4).
The constraint variety of a planar four-bar linkage can be rational if it has precisely one singu-
larity (second image from the left in Figure 11.3 and Figure 11.4, middle column). This is both a
singularity in algebraic geometry and also in the kinematic sense. It corresponds to a folded position
of the four-bar linkage where its behaviour is undefined. If one of the two arms is actuated, it is pos-
sible that the other arm follows in one or the other direction. For automatically controlled linkages,
this is a serious problem.
Also reducible constraint varieties are possible. Two general quadrics can intersect in a twisted
cubic and a straight line, two conic sections, a pair of lines and a conic section, or four straight
lines. Here, only the first three cases can occur (two images on the right of Figure 11.3). They
correspond to a deltoid linkage and a parallelogram linkage, respectively. The former has purely
rotational motion mode which corresponds to a straight line in the constraint variety and a rational
component corresponding to a twisted cubic. The latter has a translational mode along a circle and
262 Handbook of Geometric Constraint Systems Principles

E E
E
C
C
C D D C
C
C
D C
C
C
A B
B
A B A

D D

E E
A B E
A DB
D B B
A
A
C
C C

Figure 11.4
Four-bar linkages: Two assembly modes (left column), rational four-bar (middle column), reducible
four bar (right column).

a second mode which is called “anti-parallelogram” mode (Figure 11.4, right column). Finding
irreducible components of an algebraic variety amounts to a prime decomposition of their ideal.
This is nowadays a standard tool of computational kinematics, see for example [42, 47].

11.3 Serial Manipulators

A serial mechanism is characterized by having a linear linkgraph. The number n of joints equals
the dimension of its constraint variety which is also called the mechanism’s degree of freedom.
Traditionally, one assumes n ≤ 6 because this is already enough for producing a full-dimensional
subset of SE(3). But recently redundant serial chains (n > 6) have gained attention of researchers
and industry because one can use the extra degree of freedom for imposing additional requirements
such as better joint or end-effector velocity or collision avoidance.

Glossary
Bennett linkage: The only nontrivial spatial four-bar linkage with one degree of freedom.
Closed nR linkage: Linkage with only revolute joints whose linkgraph is a cycle of length n.
Denavit-Hartenberg Parameters: Geometric parameters providing a specification of a serial link-
age independent of its configuration.

Direct kinematics: Task of finding the end-effector pose for given joint parameters (rotation an-
gles, translation distances).
Goldberg linkage: The only nontrivial spatial five-bar linkage with one degree of freedom. Ob-
tained by merging two Bennett linkages.
Constraint Varieties in Mechanism Science 263

α
a
Σ1
d
Σ0

Figure 11.5
Relative position of two lines in space.

Inverse kinematic: Task of finding the joint parameters (rotation angles, translation distances) for
given end-effector pose.
Linkage synthesis: Interpolation or approximation of motions by means of linkages. Construction
of linkages to fulfill certain tasks.
Redundant manipulator: Linkage whose degrees of freedom exceed the dimension of the under-
lying motion group (six in case of SE(3)).
Singular position: “Problematic” position of a linkage that should be avoided. Depending on link-
age type and description, several different definitions are used.
In a serial manipulator each link connects exactly two joints. If the joints are viewed as lines
(the rotation axis for revolute joints, some line in translation direction for prismatic joints), the link
may be attached to their common normal. They allow a convenient description of the manipulator’s
geometry that is independent of its current configuration. The relative transformation G between
two coordinate frames attached to the joints (see Figure 11.5) is then the composition of a rotation
with angle α about the first coordinate axis and a translation by the vector (a, 0, d)| . Here, a is the
length of the common normal, d is the offset between the origin of the frame attached to the first
line and the foot point of the common normals of the lines, and α denotes the twist angle of the two
lines. The transformation G is commonly written in terms of a homogeneous transformation matrix
but may as well be given in terms of dual quaternions.
The Denavit Hartenberg (DH) parameters of a linkage consist of all common normal lengths,
offsets, and twist angles of consecutive frames in a serial chain. They provide a geometrically in-
variant way for specifying a serial manipulator, regardless of its configuration and can also be used
for a convenient description of the motion of the end-effector of a serial chain with revolute and
prismatic joints or combinations of these.
After attaching a coordinate frame to each of the n joints of the serial manipulator (Figure 11.6),
it is possible to describe the motion of the end-effector of the serial manipulator with respect to the
base frame as the product
E = M1 G1 M2 G2 · · · Mn−1 Gn−1 Mn Gn . (11.12)
Here, Gi is the transformation between consecutive links Σi and Σi+1 as illustrated by Figure 11.5,
and Mi is a rotation about the first coordinate axis in the local frame (the manipulator’s axis) whose
rotation angle ui is a motion parameter. In the matrix model, we have for example
   
1 0 0 0 1 0 0 0
ai 1 0 0  0 cos ui − sin ui 0
Gi =  0 0 cos αi − sin αi  , Mi = 0 sin ui cos ui 0 .
   (11.13)
di 0 sin αi cos αi 0 0 0 1
An extension to include prismatic joints is quite straightforward but not unique.
264 Handbook of Geometric Constraint Systems Principles

Σ2
Σn−1 end-effector
frame

Σ1

Σn
Σ0 =base frame

Figure 11.6
Coordinate frames attached to a general nR-mechanism.

11.3.1 Direct and Inverse Kinematics

In the direct kinematics problem the joint parameters u1 , . . . , un are known and the pose (position and
orientation) of the end-effector is to be determined. Conversely in the inverse kinematics problem,
the end-effector pose is known while one is looking for joint parameters that bring the mechanism’s
end-effector in exactly that given pose. In both cases the Denavit-Hartenberg parameters (normal
distances ai , offsets di , twist angles αi ) are given. In calibration problems, these parameters are to
be determined from information on joint parameters and end-effector pose. In the case of serial
manipulators the direct kinematics is easily solved via (11.12) and has a unique solution.
In contrast, the inverse kinematics problem of serial manipulators is often a difficult and highly
nonlinear problem. The end-effector position is known and the joint parameters are to be deter-
mined. If the manipulator has n ≤ 5 degrees of freedom, this problem only has a solution if the
end-effector position lies in the constraint variety of the corresponding nR chain. In the case of
n = 6 only “practical” constraints like link lengths, joint limits, or self-collisions come into play.
The inverse kinematics of the 6R manipulators once gained much attention and was known as the
“Mount Everest” problem of kinematics. Nowadays it is considered as solved (see [36]). One pos-
sible solution algorithm makes use of constraint varieties [22]. The fundamental idea is to divide
the mechanism into two chains of three revolute joints. These two chains generate (parametric)
constraint varieties of open 3R chains that can be efficiently computed and intersected. This inter-
section yields exactly the configurations of the sub-chains that allow a re-assembling at the middle
link, thus providing a configuration of the complete serial chain. There exist up to 16 solutions all
of which can be real. Recently, hundreds of manipulators with 16 real inverse kinematics solutions
were found [6]. Some of them have all real solutions on a whole path in their workspace. The pos-
sibility to select among 16 possible paths with different properties concerning collisions, required
joint limits, or energy consumption is beneficial for the path planning.

11.3.2 Synthesis of Open and Closed Serial Chains

In this part we discuss the construction of serial chains. In mechanism science, this is called syn-
thesis. We may also view it as an interpolation problem on the Study quadric by certain constraint
varieties. Problems of this type have a long tradition in kinematics [33]. In the case of serial chains,
synthesis means finding the Denavit-Hartenberg parameters of a manipulator for a prescribed task,
Constraint Varieties in Mechanism Science 265

Figure 11.7
A Bennett mechanism (left), two Bennett mechanisms sharing one link and two joints (middle), and
the corresponding Goldberg 5R mechanism (right).

for example the capability to reach a number of prescribed end-effector poses. One may be inter-
ested in approximate or exact solutions. In the second case, the prescribed task yield an algebraic
system of equations which has to be solved. In case of approximate synthesis the task may lead to
an over-constrained system of equations. Its approximate solution requires a concept for the “dis-
tance” between rigid body displacements. This is quite problematic as incompatible units (angles
and distances) are involved. In this text we focus on exact and algebraic synthesis methods.
There exist several methods for the exact synthesis of 2R chains [1, 2, 8, 33]. For serial chains
with three revolute joints, only numerical solutions are known. Homotopy continuation is used in
[27, 28, 29] while [30] proposed interval analysis. The most recent solution [15] uses numerical
algebraic geometry via the software package Bertini [4]. There it was shown that the exact synthesis
of 3R chains with five prescribed end-effector poses can have at most 456 solutions over the complex
numbers. The number of possible real solutions is still unknown. Virtually nothing seems to be
known about the synthesis of arbitrary serial chains with four or five joints.
A task which is not yet been extensively explored is the synthesis of closed serial chains. One
searches for mechanisms that are movable even if the end-effector is restricted to one fixed pose
and in addition fulfill a certain prescribed task. Closed serial chains with seven or more joints are
movable, whereas such chains with three or less joints are rigid. Therefore, the cases of four, five,
or six joints are of particular interest. Their mobility requires a special geometric design. Flexible
examples with only four revolute joints are planar and spherical four-bar linkages (all axes are
parallel or concurrent, respectively), the Bennett linkage [5] and degenerate cases with coinciding
axes. The Bennett linkage (see Figure 11.7 left) is a mobile spatial four-bar linkage. The geometric
condition for flexibility is that the common normals between any two consecutive axes form a closed
spatial parallelogram, that is, consecutive sides intersect and the distances between opposite point
pairs are equal. An algebraic synthesis algorithm for four-bar linkages with revolute and prismatic
joints using constraint varieties was recently presented in [37].
Essentially the only nontrivial flexible example for n = 5 is known as Goldberg’s linkage. We
may think of it as the concatenation of two Bennett linkages with one common link and two common
joints such that the one Bennett linkage “drives” the other. Removing one common joint and merging
certain links one obtains a 5R linkage with one degree of freedom. This is shown in Figure 11.7.
The necessity of this construction was first proved in [25]. A simpler proof using bond theory is in
[17].
The cases n = 4 and n = 5 are more or less fully understood. This is not true for n = 6. In fact,
a full classification of all mobile closed 6R linkages is one of the big open problems in the field.
A precise problem formulation can be found in [32], the current state of research is summarized
in [31]. Recent progress was made using bond theory and bond diagrams [17, 32]. Only little is
known about the synthesis of closed-loop linkages with six joints. The synthesis of certain closed
six-bar linkages to four given poses is the topic of [18]. It is based on a factorization theory of
266 Handbook of Geometric Constraint Systems Principles

motion polynomials (parametric equations of rational curves on the Study quadric) [16] where the
factorization into linear factors corresponds to a decomposition of the parametric curve/motion into
rotations. In general, there exist several factorizations, each of them giving rise to a serial nR chain.
Combining two such chains it is possible to construct a closed chain with 2n revolute joints. The
synthesis method of [18] can be specialized to the synthesis of closed five-bar linkages [40]. Still,
there is a lot of room for future research. One of the reasons for this is the missing classification of
movable closed 6R linkages.

11.3.3 Singularities and Path Planning

In industrial applications, it is necessary to carefully plan the desired path of a mechanical device.
It is not only necessary to ensure a desired trajectory but also to limit velocities and accelerations
[14]. In addition, it is necessary to avoid configurations of the mechanism where it is in a singular
position. In case of serial manipulators such configurations can be defined in different ways. One
possibility is coincidence of two of the inverse kinematics solutions. Alternatively, the Plücker co-
ordinates of the joint axes satisfy a linear relation [24] or [38, Section 3.1]. Singularities are very
important in practice for several reasons. The manipulator looses degrees of freedom in such config-
urations. Moreover, control of the mechanism in the vicinity of singularities is very difficult because
numerical calculations become unstable. Therefore, singularity avoiding algorithms for path plan-
ning are an important topic.

11.3.4 Open Problems

• The synthesis of serial manipulators with three, four, or five joints to a prescribed number
of task poses. Only numerical solutions for three revolute joints are known.
• The synthesis of closed serial chains which is related to yet another open issue, the complete
classification of this linkage type.
• It is believed that algebraic methods could also be beneficial for singularity avoiding path
planning.
• The factorization theory for motion polynomials still is not fully understood. Some non-
generic cases are topic of current research and details may be rather tricky (see for example
[13]). A factorization theory for multivariate motion polynomials would be highly interest-
ing for applications in mechanism science but is probably rather difficult.

11.4 Parallel Manipulators

A parallel manipulator consists of a fixed base and a moving end-effector (platform), connected by
multiple chains of links and joints called limbs. In other words, its linkgraph contains multiple loops.
In many technical applications the limbs have the same structure. One of the most important parallel
manipulators is the Stewart-Gough platform (Figure 11.9) used for example in flight simulators. The
Stewart-Gough platform has six degrees of freedom and is hence called a “full degree of freedom
parallel manipulator”. A second example is the Delta-Robot, which is used for high-speed pick and
place tasks [35]. The Delta-Robot has three degrees of freedom and is hence called a “lower degree
of freedom parallel manipulator.” Also planar parallel manipulators are interesting, as the following
sections will show.
Constraint Varieties in Mechanism Science 267

A3 A3
B3
B3 SSS

B1 B1
B2 B2

A1 A2 A1 A2

Figure 11.8
Planar 3-RPR parallel manipulator, singular configuration, and singularity surface.

Glossary
3-RPR manipulator: Parallel manipulator with three limbs, each consisting of a revolute, a pris-
matic, and another revolute joint.
Stewart-Gough platform: Parallel manipulator with six limbs, each consisting of two passive
(nonactuated) spherical joints connected by an active (actuated) prismatic joint.
Self-motion: Motion of a generically rigid parallel manipulator. Only possible in special singular
configurations.

11.4.1 Direct and Inverse Kinematics

In contrast to serial manipulators, solving the direct kinematics problem of a parallel manipulator
takes more effort than the inverse kinematic problem. In case of the direct kinematic problem, the
input parameters, in this case some of the joint parameters, are given, and the position of the platform
has to be determined. This will be discussed for two examples in more detail.
For the planar case, consider the 3-RPR parallel manipulator. This manipulator consists of a
fixed base, given by the triangle A1 A2 A3 as shown in Figure 11.8, left, and a moving platform given
by the triangle B1 B2 B3 . Corresponding vertices are connected by RPR limbs, this means revolute
joints are located in Ai and Bi . To move the platform, the distance between Ai and Bi is changed
by an actuated prismatic joint and the variable distance is denoted by ri . The direct kinematics
problem is: Determine the position of the platform for given r1 , r2 and r3 . The vertex B1 has to
be located on a circle with center A1 and radius r1 . Similar constraints hold for B2 and B3 . So the
reformulated geometric task is: Place the triangle B1 B2 B3 on three circles centered at A1 , A2 and A3
with respective radii r1 , r2 and r3 . To solve this task using Study’s kinematic mapping, we have to
intersect three circle constraint varieties with prototype equation (11.11). The intersection of three
quadrics in projective three-space P3 gives, in general, eight points. But in Section 11.2 we saw that
the invalid solutions (0, 0, 1, ±i)| are always among them. Hence, the direct kinematics problem has
only six solutions.
A more general mathematical approach uses computational algebraic geometry for intersecting
the constraint varieties of each limb. Often the number of possible solutions is of interest. In this case
it is useful to eliminate all unknowns until a univariate polynomial is achieved. This maybe done
by computing step wise resultants or a Gröbner basis [10]. In case of the 3-RPR manipulator this
univariate polynomial has, indeed, degree six. In general the degree of the univariate polynomial is
268 Handbook of Geometric Constraint Systems Principles
B6

B1
B4
B5
r6
r1
B3
B2
A6 r5 r4
A1
r2 r3

A3 A5
A4
A2

Figure 11.9
Stewart-Gough platform

only an upper bound, because not all solutions have to be real, but for the 3-RPR parallel manipulator
there are examples with six real solutions as shown in [20].
A spatial example of a parallel manipulator is the Stewart-Gough platform shown in Figure 11.9.
This parallel manipulator consists of six anchor points Ai , i = 1, . . . , 6 in the fixed base and six points
Bi , i = 1, . . . , 6 in the moving platform. Any two points Ai and Bi are connected by an SPS-limb, i.e.
a spherical joint in Ai followed by a prismatic joint and again a spherical joint in Bi . The actuated
joints are the prismatic joints and the variable lengths are denoted by ri . In this case there are six
sphere constraint equations kAi − Bi k2 − ri2 = 0, i = 1, . . . , 6. The constraint equations in this case
are again quadratic, but this time with eight homogeneous unknowns p0 , . . . , p3 and q0 , . . . , q3 . A
univariate polynomial for this problem was computed for the first time in [19]. It is of degree 40, so
for given lengths ri there exist up to 40 positions of the platform. All of them can be real [12].
More recently, lower degree of freedom parallel manipulators draw attention in mechanism sci-
ence. The reason is that for some tasks not the full six degrees of freedoms are required. One
example is the already mentioned Delta-Robot which has three degrees of freedom and is specially
designed for fast “pick and place” tasks.

11.4.2 Singularities and Self-Motions

Some advantages of parallel manipulators over serial manipulators are the high stiffness and the
high accuracy. Because of this, a parallel manipulator is capable of taking high loads. But in some
positions of the platform this stiffness is lost, causing some unwanted behaviour during the motion
of the platform. In rigidity theory, one speaks of infinitesimally flexible configurations. In mecha-
nism science, these positions are called “singular.” Depending on the structure of the manipulator,
computing singular positions can be a challenging task. In general, if the constraint equations of the
parallel manipulator are known, they can be computed as the variety to the system’s Jacobian. This
gives rise to a singularity surface in kinematic image space (Figure 11.8, right-hand side).
Sometimes the singularity conditions have a geometrical meaning, like in case of the planar
3-RPR manipulator. Here, singular positions are characterised by concurrency of the three legs
(Figure 11.8, middle). The spatial Stewart-Gough platform is in a singular position, iff the six lines
joining corresponding vertices Ai , Bi belong to a linear line complex [38, Section 3.1]. Some special
singular configurations can result in self-motions of the parallel manipulator. These motions appear,
when the input parameters are fixed, but the moving platform is able to perform a motion with at
References 269

least one degree of freedom. For the planar 3-RPR manipulator these are described in [7], for the
Stewart-Gough platform see [26].

11.4.3 Open Problems

• For lower degree of freedom parallel manipulators the path planing task is still an open
problem. This is due to the fact that the workspace is a subset of SE(3) and the description
via constraint equations is an implicit representation. In this case it has to be ensured that the
path, a curve in the intersection of the constraint varieties, lies completely in the workspace.
Avoiding singular positions is also necessary. Therefore, additional side conditions have to
be included. In some special cases, like the 3-RPS considered in [41], a parameterization of
the complete workspace is possible, which makes the task of path planing much easier and
relates mechanism science to parameterization issues of algebraic varieties.
• Recently a new class of parallel manipulators, cable driven manipulators, became very pop-
ular. For cable driven manipulators, the rigid limbs of a parallel manipulator are replaced
by cables that control the moving platform. A physical constraint of cables is, that they are
only able to contribute to the motion if they are under tension. This fact leads to some extra
constraints in each cable, giving rise to semi-algebraic equations. The planar case of the
so-called “hanging box” problem was investigated in [9]. where a bound of six solutions for
the direct kinematics problem is given.
• An interesting problem of parallel manipulators are “non-singular assembly mode changes”.
This means the continuous changing between two direct kinematics solutions without pass-
ing through a singular configuration [21]. In this way, it is possible to permute direct kine-
matics solutions. The thus defined permutation group should capture a lot of information on
reconfigurability of the robot. Apparently, this has not yet been considered in mechanism
but similar ideas appear in other classes of geometric problems [11].

References
[1] Kassim Abdul-Sater, Franz Irlinger, and Tim C. Lueth. Two-configuration synthesis of
origami-guided planar, spherical and spatial revolute-revolute chains. ASME Journal of Mech-
anisms and Robotics, 5(3):031005, 2013.
[2] Kassim Abdul-Sater, Manuel M. Winkler, Franz Irlinger, and Tim C. Lueth. Three-position
synthesis of origami-evolved, spherically constrained spatial revolute-revolute chains. ASME
Journal of Mechanisms and Robotics, 8(1):011012, 2015.
[3] Hamid Ahmadinezhad, Zijia Li, and Josef Schicho. An algebraic study of linkages with
helical joints. J. Pure Appl. Algebra, 219(6):2245–2259, 2015.
[4] Daniel J. Bates, Jonathan D. Hauenstein, Andrew J. Sommese, and Charles W. Wampler.
Bertini: Software for numerical algebraic geometry. Available at bertini.nd.edu.
[5] Geoffrey T. Bennett. A new mechanism. Engineering, 76:777–778, 1903.
[6] Mathias Brandstötter. Adaptable Serial Manipulators in Modular Design. PhD thesis, UMIT
- University for Health Sciences, Medical Informaticvs and Technology, Hall, Austria, 2016.
270 References

[7] Sébastien Briot, Ilian Bonev, Damien Chablat, Philippe Wenger, and Vigen Arakelian. Self-
motions of general 3-RPR parallel robots. International Journal of Robotics Research,
27(7):855–866, 2008.
[8] Katrin Brunnthaler, Hans-Peter Schröcker, and Manfred L. Husty. A new method for the
synthesis of Bennett mechanisms. In Proceedings of CK 2005, International Workshop on
Computational Kinematics, Cassino, 2005.
[9] Marco Carricato and Jean-Pierre Merlet. Stability analysis of underconstrained cable-driven
parallel robots. IEEE Transactions on Robotics, 29(1):288–296, 2013.
[10] David Cox, John Little, and Donal O’Shea. Ideals, Varieties, and Algorithms. Springer, 2007.
[11] Abraham Martin del Campo Sanchez. Galois groups of Schubert problems. PhD thesis, Texas
A&M University, 2012.
[12] Peter Dietmaier. The stewart-gough platform of general geometry can have 40 real postures.
In Jadran Lenarčič and Manfred L. Husty, editors, The 6th conference on Advances in Robot
Kinematics 1998, Stobl, pages 7–16, 1998.
[13] Matteo Gallet, Christoph Koutschan, Zijia Li, Georg Regensburger, Josef Schicho, and Nelly
Villamizar. Planar linkages following a prescribed motion. Mathematics of Computation,
86(303):473–506, 2016.
[14] Francisco Geu Flores and Andrés Kecskeméthy. Time-optimal path planning along speci-
fied trajectories. In Hubert Gattringer and Johannes Gerstmayr, editors, Multibody System
Dynamics, Robotics and Control, pages 1–16. Springer, 2013.
[15] Jonathan D. Hauenstein, Charles W. Wampler, and Martin Pfurner. Synthesis of three-revolute
spatial chains for body guidance. Mechanism and Machine Theory, 2016.
[16] Gábor Hegedüs, Josef Schicho, and Hans-Peter Schröcker. Factorization of rational curves in
the Study quadric. Mechanism and Machine Theory, 69:142152, 2013.
[17] Gábor Hegedüs, Josef Schicho, and Hans-Peter Schröcker. The theory of bonds: A new
method for the analysis of linkages. Mechanism and Machine Theory, 70:407424, 2013.
[18] Gábor Hegedüs, Josef Schicho, and Hans-Peter Schröcker. Four-pose synthesis of angle-
symmetric 6R linkages. ASME Journal of Mechanisms and Robotics, 7(4):041006, 2015.
[19] Manfred L. Husty. An Algorithm for Solving the Direct Kinematics of General Stewart-
Gough Platforms. Mechanism and Machine Theory, 31(4):365–380, 1996.
[20] Manfred L. Husty. Non-singular assembly mode change in 3-RPR-parallel manipulators.
In Andrés Kecskeméthy and Andreas Müller, editors, Computational Kinematics. Springer,
2009.
[21] Manfred L. Husty and Clément Gosselin. On the singularity surface of planar 3-RPR parallel
mechanisms. Mechanics Based Design of Structures and Machines, 36(4):411–425, 2008.
[22] Manfred L. Husty, Martin Pfurner, and Hans-Peter Schröcker. Algebraic methods in mecha-
nism analysis and synthesis. Robotica, 25(6):661–675, 2007.
[23] Manfred L. Husty and Hans-Peter Schröcker. Kinematics and algebraic geometry. In
J. Michael McCarthy, editor, 21st Century Kinematics. The 2012 NSF Workshop, pages 85–
123. Springer, London, 2012.
[24] Adolf Karger. Singularity analysis of serial robot-manipulators. Journal of Mechanical De-
sign, 118(4):520–525, 1996.
References 271

[25] Adolf Karger. Classification of 5R closed kinematic chains with self mobility. Mechanism
and Machine Theory, 33(1-2):213222, 1998.
[26] Adolf Karger and Manfred L. Husty. Classification of all self-motions of the original Stewart-
Gough platform. Computer Aided Design, 30(3):205–215, 1998.
[27] Eric Lee and Constantinos Mavroidis. Solving the geometric design problem of spatial 3R
robot manipulators using polynomial homotopy continuation. Journal of Mechanical Design,
124:652–661, 2002.
[28] Eric Lee and Constantinos Mavroidis. An algebraic elimination based algorithm for solving
the geometric design problem of spatial 3R manipulators. In Proc. ASME Design Eng. Tech.
Conf., Mech. & Robotics Conf. ASME, 2004.
[29] Eric Lee and Constantinos Mavroidis. Geometric design of 3R robot manipulators for reach-
ing four end-effector spatial poses. The International Journal of Robotics Research, 23(3):247
– 254, 2004.
[30] Eric Lee, Constantinos Mavroidis, and Jean-Pierre Merlet. Five precision point synthesis of
spatial RRR manipulators using interval analysis. Journal of Mechanical Design, 126:842–
849, 2004.
[31] Zijia Li. Closed linkages with six revolute joints. Phd thesis, Johannes Kepler University,
Linz, 2015.
[32] Zijia Li, Josef Schicho, and Hans-Peter Schröcker. A survey on the theory of bonds. IMA J.
Math. Control Inform., 2016.
[33] J. Michael McCarthy. Geometric Design of Linkages, volume 11 of Interdisciplinary Applied
Mathematics. Springer, 2000.
[34] Johan Ernest Mebius. Derivation of the Euler-Rodrigues formula for three-dimensional rota-
tions from the general formula for four-dimensional rotations.
[35] Jean-Pierre Merlet. Parallel Robots. Kluwer Academic Publishers, 2000.
[36] Martin Pfurner. Analysis of spatial serial manipulators using kinematic mapping. PhD thesis,
University of Innsbruck, Innsbruck, Austria, 2008.
[37] Martin Pfurner, Thomas Stigger, and Manfred L. Husty. Overconstrained single loop four link
mechanisms with revolute and prismatic joints. In Philippe Wenger and Paulo Flores, editors,
New Trends in Mechanism and Machine Science, Theory and Industrial Applications, pages
71–79, Nantes, France, 2016.
[38] Helmut Pottmann and Johannes Wallner. Computational Line Geometry. Springer, 2001.
[39] Tudor-Dan Rad, Daniel F. Scharler, and Hans-Peter Schröcker. The kinematic image of RR,
PR, and RP dyads. Submitted for publication, 2016.
[40] Tudor-Dan Rad and Hans-Peter Schröcker. The kinematic image of 2R dyads and exact syn-
thesis of 5R linkages. In Proceedings of the IMA Conference on Mathematics of Robotics,
2015.
[41] Josef Schadlbauer, Latifah Nurahmi, Manfred L. Husty, Stéphane Caro, and Philippe Wenger.
Operation modes of lower-mobility parallel manipulators. In Second Conference on Interdis-
ciplinary Applications in Kinematics, pages 3–10, 2013.
[42] Josef Schadlbauer, Dominic R. Walter, and Mandfred L. Husty. The 3-RPS parallel manipu-
lator from an algebraic viewpoint. Mechanism and Machine Theory, 75:161176, 2014.
272 References

[43] Jon M. Selig. Geometric Fundamentals of Robotics. Monographs in Computer Science.

Springer, 2 edition, 2005.
[44] Jon M. Selig. On the geometry of point-plane constraints on rigid-body displacements. Acta
Appl. Math., 116(2):133155, 2011.
[45] Jon M. Selig. Some rigid-body constraint varieties generated by linkages. In Jadran
Lenarčič and Manfred L. Husty, editors, Latest Advances in Robot Kinematics, pages 293–
300. Springer, 2012.
[46] Eduard Study. Von den Bewegungen und Umlegungen. Math. Ann., 39:441–566, 1891.
[47] Dominic R. Walter, Manfred L. Husty, and Martin Pfurner. A complete kinematic analysis of
the SNU 3-UPU parallel manipulator. In Daniel J. Bates, Gian M. Besana, Sandra Di Rocco,
and Charles W. Wampler, editors, Contemporary Mathematics, volume 496, pages 331–346.
American Mathematical Society, 2009.
Chapter 12
Real Algebraic Geometry for Geometric
Constraints
Frank Sottile
Texas A&M University, College Station, Texas

CONTENTS
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
12.2 Ideals and Varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
12.3 ... and Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
12.4 Structure of Algebraic Varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
12.4.1 Zariski Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
12.4.2 Smooth and Singular Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
12.4.3 Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
12.5 Real Algebraic Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
12.5.1 Algebraic Relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
12.5.2 Semi-Algebraic Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
12.5.3 Certificates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284

Real algebraic geometry adapts the methods and ideas from (complex) algebraic geometry to study
the real solutions to systems of polynomial equations and polynomial inequalities. As it is the real
solutions to such systems modeling geometric constraints that are physically meaningful, real alge-
braic geometry is a core mathematical input for geometric constraint systems.

12.1 Introduction
Algebraic geometry is fundamentally the study of sets, called varieties, which arise as the common
zeroes of a collection of polynomials. These include familiar objects in analytic geometry, such
as conics, plane curves, and quadratic surfaces. Combining intuitive geometric ideas with precise
algebraic methods, algebraic geometry is equipped with many powerful tools and ideas. These may
be brought to bear on problems from geometric constraint systems because many natural constraints,
particularly prescribed incidences, may be formulated in terms of polynomial equations.
Consider a four-bar mechanism; a quadrilateral in the plane with prescribed side lengths a, b,
c, and d, which may rotate freely at its vertices and where one edge is fixed as in Figure 12.2. The
points x and y are fixed at a distance a apart, the point p is constrained to lie on the circle centered

273
274 Handbook of Geometric Constraint Systems Principles

Figure 12.1
A cubic plane curve and a quadratic surface (a hyperbolic paraboloid).
p
q
d
b c

x a y
fixed side

Figure 12.2
A four-bar mechanism.

at x with radius b, the point q lies on the circle centered at y with radius c, and we additionally
require that p and q are a distance d apart. Squaring the distance constraints gives a system of three
quadratic equations whose solutions are all positions of this four-bar mechanism.
Algebraic geometry works best over the complex numbers, because the geometry of a complex
variety is controlled by its defining equations. (For instance, the Fundamental Theorem of Algebra
states that a univariate polynomial of degree n always has n complex roots, counted with multiplic-
ity.) Geometric constraint systems are manifestly real (as in real-number) objects. For this reason,
the subfield of real algebraic geometry, which is concerned with the real solutions to systems of
equations, is most relevant for geometric constraint systems. Working over the real numbers may
give quite different answers than working over the complex numbers.
This chapter will develop some parts of real algebraic geometry that are useful for geometric
constraint systems. Its main point of view is that one should first understand the geometry of corre-
sponding complex variety, which we call the algebraic relaxation of the original problem. Once this
is understood, we then ask the harder question about the subset of real solutions.
2 y2 = 1 and x2 + y2 = −1 define isomorphic curves in the complex plane—
√ x +√
For example,
send (x, y) 7→ ( −1x, −1y)—which are quite different in the real plane. Indeed, x2 + y2 = 1 is the
unit circle in R2 and x2 + y2 = −1 is the empty set. Replacing ±1 by 0 gives the pair of complex
conjugate lines √ √
x2 + y2 = (x + −1y)(x − −1y) = 0 ,
whose only real point is the origin (0, 0). The reason for this radically different behavior amongst
these three quadratic plane curves is that only the circle has a smooth real point—by Theorem 12.2,
when a real algebraic variety has a smooth real point, the salient features of the underlying complex
variety are captured by its real points.
Real Algebraic Geometry for Geometric Constraints 275

12.2 Ideals and Varieties

C[x1 , . . . , xd ]: The ring of all polynomials with complex coefficients in the variables x1 , . . . , xd .
Variety: A subset V(S) ⊂ Cd where all polynomials in a set S ⊂ C[x1 , . . . , xd ] vanish.
Ideal: Set of polynomials in C[x1 , . . . , xd ] closed under addition and multiplication by other poly-
nomials.
I(X): For a subset X of Cd , this is the set of all polynomials that vanish identically on X.
Hilbert’s Nullstellensatz: A fundamental theorem in algebraic geometry; the variety of a set S of
polynomials is empty if and only if the set S generates the trivial idea h1i = C[x1 , . . . , xd ].

The best accessible introduction to algebraic geometry is the classic book of Cox, Little, and
O’Shea [7]. Many thousands find this an indispensable reference. We assume a passing knowledge
of some aspects of the algebra of polynomials, or at least an open mind. We work over the complex
numbers, C, for now. A collection S ⊂ C[x1 , . . . , xd ] of polynomials in d variables defines a variety,

V(S) := {x ∈ Cd | f (x) = 0 for all f ∈ S} .

We may add to S any of its polynomial consequences, g1 f1 + · · · + gs fs where gi ∈ C[x1 , . . . , xd ]

and fi ∈ S, without changing V(S). This set of polynomial consequences is the ideal generated by
S, and so it is no loss to assume that S is an ideal. Hilbert’s Basis Theorem states that any ideal in
C[x1 , . . . , xd ] is finitely generated, so it is also no loss to assume that S is finite. We pass between
these extremes when necessary.
Dually, given a variety X ⊂ Cd (or any subset), let I(X) be the set of polynomials which
vanish on X. Any polynomial consequence of polynomials that vanish on X also vanishes on X.
Thus I(X) is an ideal in the polynomial ring C[x1 , . . . , xd ]. Let C[X] be the set of functions on X
that are restrictions of polynomials in C[x1 , . . . , xd ]. Restriction is a surjective ring homomorphism
C[x1 , . . . , xd ]  C[X] whose kernel is the ideal I(X) of X, so that C[X] = C[x1 , . . . , xd ]/I(X). Call
C[X] the coordinate ring of X.
To see this connection between algebra and geometry, consider the two plane curves V(y − x2 )
and V(y2 − x3 ) of Figure 12.3. One is the familiar parabola, which is smooth, and the other the

Figure 12.3
Plane curves V(y − x2 ) and V(y2 − x3 ).

semicubical parabola or cuspidal cubic, which is singular (see Section 12.4.2) at the origin. Their
coordinate rings are C[x, y]/hy − x2 i and C[x, y]/hy2 − x3 i, respectively. The first is isomorphic to
C[t], which is the coordinate ring of the line C, while the second is not—it is isomorphic to C[t 2 ,t 3 ].
276 Handbook of Geometric Constraint Systems Principles

The isomorphisms come from the parameterizations t 7→ (t,t 2 ) and t 7→ (t 2 ,t 3 ). This illustrates
another way to obtain a variety—as the image of a polynomial map.
Thus begins the connection between geometric objects (varieties) and algebraic objects (ide-
als). Although they are different objects, varieties and ideals carry the same information. This is
expressed succinctly and abstractly by stating that there is an equivalence of categories, which is a
consequence of Hilbert’s Nullstellensatz, whose finer points we sidestep. For the user, this equiv-
alence means that we may apply ideas and tools either from algebra or from geometry to better
understand the sets of solutions to polynomial equations.

12.3 ... and Algorithms

Because the objects of algebraic geometry have finiteness properties (finite-dimensional, finitely
generated), they may be faithfully represented and manipulated on a computer. There are two main
paradigms: symbolic methods based on Gröbner bases and numerical methods based on homotopy
continuation. The first operates on the algebraic side of the subject and the second on its geometric
side.

Gröbner basis: An algorithmically optimal generating set of an ideal.

Buchberger Algorithm: An algorithm to compute a Gröbner basis for an ideal from a generating
set.
Numerical algebraic geometry: Computational paradigm that applies tools from numerical anal-
ysis to study algebraic varieties.

A consequence of the Nullstellensatz is that we may recover any information about a variety
X from its ideal I(X). By Hilbert’s Basis Theorem, I(X) is finitely generated, so we may repre-
sent it on a computer by a list of polynomials. We emphasize computer because expressions for
multivariate polynomials may be too large for direct human manipulation or comprehension. Many
algorithms to study a variety X through its ideal begin with a preprocessing: a given list of gener-
ators ( f1 , . . . , fm ) for I(X) is replaced by another list (g1 , . . . , gs ) of generators, called a Gröbner
basis for I(X), with optimal algorithmic properties. Buchberger’s Algorithm is a common method
to compute a Gröbner basis.
Many algorithms to extract information from a Gröbner basis have reasonably low complexity.
These include algorithms that use a Gröbner basis to decide if a given polynomial vanishes on
a variety X or to determine the dimension or degree (see Section 12.4.1 and Section 12.4.2) of
X. Consequently, a Gröbner basis for I(X) transparently encodes much information about X. We
expect, and it is true, that computing a Gröbner basis may have high complexity (double exponential
in d in the worst case), and some computations do not terminate in a reasonable amount of time.
Nevertheless, symbolic methods based on Gröbner bases easily compute examples of moderate size,
as the worst cases appear to be rare.
Several well-maintained computer algebra packages have optimized algorithms to compute
Gröbner bases, extensive libraries of implemented algorithms using Gröbner bases, and excellent
documentation. Two in particular—Macaulay2 [11, 12] and Singular [8, 9]—are freely available
with dedicated communities of users and developers. Commercial software, such as Magma, Maple,
and Mathematica, also compute Gröbner bases and implement some algorithms based on Gröbner
bases. Many find SageMath [10], an open-source software connecting different software systems
together, also to be useful.
Real Algebraic Geometry for Geometric Constraints 277

The other computational paradigm—numerical algebraic geometry—uses methods from numer-

ical analysis to manipulate varieties on a computer [18]. Numerical homotopy continuation is used
to solve systems of polynomial equations, and Newton’s method may be used to refine the solutions.
These methods were originally developed as a tool for mechanism design in kinematics [15], which
is closely related to geometric constraint systems.
In numerical algebraic geometry, a variety X of dimension n (see Section 12.4.2) in Cd is rep-
resented on a computer by a witness set, which is a triple (W, S, L), where L is a general affine
plane in Cd of dimension d−n, S is a list of polynomials defining X, and W consists of numerical
approximations to the points of X ∩ L (the number of which is the degree of X, see Section 12.4.2).
Following the points of W as L varies using homotopy continuation samples points of X, and may
be used to test for membership in X. Other algorithms, including computing intersections and the
image of a variety under a polynomial map, are based on witness sets.
Two stand-alone packages—PHCPack [22] and Bertini [2, 3]—implement the core algorithms
of numerical algebraic geometry, as does the Macaulay2 package NAG4M2 [14]. Both PHCPack
and Bertini may be accessed from Macaulay2, Singular, or SageMath.
Each computational paradigm, symbolic and numerical, has its advantages. Symbolic compu-
tations are exact and there are many implemented algorithms. The inexact numerical computations
give refinable approximations, yielding a family of well-behaved relaxations to exact computation.
Also, numerical algorithms are easily parallelized and in some cases the results may be certified to
be correct [13, 17].

12.4 Structure of Algebraic Varieties

Varieties and their images under polynomial maps have well-understood properties that may be
exploited to understand objects modeled by varieties. We discuss some of these fundamental and
structural properties.

Zariski topology: Topology on Cn and on varieties whose closed sets are varieties.
Subvariety: A variety that is a subset of another.
Irreducible variety: A variety that is not the union of two proper subvarieties.
Generic: A property that holds on a dense Zariski open set.
General: A point of a variety where a generic property holds.
Smooth: A point of a variety where it is a manifold.
Singular: A point of a variety where it is not a manifold.
Dimension of X: The dimension of the smooth (manifold) points of a variety X.

Degree of X: The number of points in the intersection of a variety X with a general affine subspace
of dimension n − dim X.
Locally closed: A set that is open in its closure.
Constructible: A set that is a finite union of locally closed sets.
278 Handbook of Geometric Constraint Systems Principles

12.4.1 Zariski Topology

Algebraic varieties in Cd are closed subsets in the usual (classical) topology because polyno-
mial functions are continuous. Varieties possess a second, much coarser topology—the Zariski
topology—whose value is that it provides the most natural language for expressing many properties
of varieties. The Zariski topology is determined by its closed sets, which are simply the algebraic
varieties, and therefore its open sets are complements of varieties.
Closure in the Zariski topology is easily expressed: The Zariski closure U of a set U ⊂ Cd
is V(I(U)), the set of points in Cd where every polynomial that vanishes identically on U also
vanishes. A non-empty Zariski open subset U of Cd is dense in Cd in the classical topology, and a
classical open subset of Cd (e.g. a ball) is dense in the Zariski topology. The Zariski topology of a
variety X in Cd is induced from that of Cd .
We use the Zariski topology to express the analog of unique factorization of integers for vari-
eties. A variety X is irreducible if cannot be written as a union of proper subvarieties. That is, if
X = Y ∪ Z with Y, Z subvarieties of X, then either X = Y or X = Z. A variety X has an irredundant
decomposition into irreducible subvarieties, X = X1 ∪ · · · ∪ Xm , which is unique in that each Xi is
an irreducible subvariety of X and if i 6= j, then Xi 6⊂ X j . We call the subvarieties X1 , . . . , Xm the
(irreducible) components of X.
This decomposition for a hypersurface is equivalent to the factorization of its defining polyno-
mial into irreducible polynomials. For the curve on the left in Figure 12.4 we have,

x3 + x2 y − xy − y2 = (x2 − y)(x + y) ,

showing that its components are the parabola y = x2 and the line y = −x. Both V(x3 +x2 y−xy−y2 )

C2 C3

Figure 12.4
V(x3 +x2 y−xy−y2 ) and V(z−xy, xz−y2 −x2 +y).

and V(z−xy, xz−y2 −x2 +y) are curves with two components, as we see in Figure 12.4.
Zariski open sets are quite large. Any nonempty Zariski open subset U of an irreducible variety
X is Zariski dense in X. Indeed, X = U ∪ (X rU), the union of two closed subsets. Since X 6= X rU,
we have U = X. In fact, U is dense in the classical topology, and any subset of X that is dense in the
classical topology is Zariski dense in X.
A property of an irreducible variety X is generic if the set of points where that property holds
contains a Zariski open subset of X. Generic properties of X hold almost everywhere on X in a very
strong sense, as the points of X where they do not hold lie in a proper subvariety of X. A point of a
variety where a generic property holds is general (with respect to that property).
Real Algebraic Geometry for Geometric Constraints 279

12.4.2 Smooth and Singular Points

Algebraic varieties are not necessarily manifolds, as may be seen in Figures 12.3 and 12.4. However,
the set of points where a variety fails to be a manifold is a proper subvariety. To see this, suppose that
X ⊂ Cd is a variety whose ideal I(X) has generators f1 , . . . , fs . At each point x of X, the Jacobian
j=1,...,d
matrix J = (∂ fi /∂ x j )i=1,...,s has rank between 0 and d. The set Xi of points of X where the rank of
J is at most i is a subvariety which is defined by the vanishing of all (i+1) × (i+1) minors of J. If i
is the smallest index such that Xi = X, so that Xi−1 ( X, then at every point of Xsm := X r Xi−1 the
Jacobian has rank i. Differential geometry informs us that Xsm is a complex manifold of dimension
d−i.
When X is irreducible, Xsm is the set of smooth points of X and Xsing := X r Xsm is the singular
locus of X. A point being smooth is a generic property of X. The dimension dim X of an irreducible
variety X is the dimension of Xsm . When X is reducible, its dimension is the maximum dimension
of an irreducible component. The singular locus of a variety X always has smaller dimension than
X.
For algebraic varieties, dimension has the following properties. If X and Y are subvarieties of
Cd of dimensions m and n, respectively, then either X ∩Y is empty or every irreducible component
of X ∩Y has at least the expected dimension m+n−d. For a general translate Y 0 of Y , dim(X ∩Y 0 ) =
m+n−d. More precisely, there is a Zariski open subset U of the group Cd o GL(d, C) of affine
transformations of Cd such that if g ∈ U then X ∩ gY has dimension m+n−d and is as smooth as
possible in that its singular locus is a subset of the union of Xsing ∩ gY with X ∩ gYsing .
Similarly, Bertini’s Theorem states that there is a Zariski open subset U of the set of polynomials
of a fixed degree such that for f ∈ U, X ∩ V( f ) has dimension dim X−1 and is as smooth as possible.
A consequence of all this is that if L is a general affine linear subspace of dimension d− dim X, then
X ∩ L is a finite set of points contained in Xsm . The number of points is the maximal number of
isolated points in any intersection of X with an affine plane of this dimension and is called the
degree of X. These facts underlie the notion of witness set in numerical algebraic geometry from
Section 12.3.

12.4.3 Maps
We often have a map ϕ : Cd → Cn given by polynomials, and we want to understand the image of
a variety X ⊂ Cd under this map. Algebraic geometry provides a structure theory for the images of
polynomial maps. We begin with an example. Consider the hyperbolic paraboloid V(y − xz) in C3
and its projection to the xy-plane, which is a polynomial map. This image is the union of all lines
through the origin, except for the y-axis, V(x). Figure 12.5 shows both the hyperbolic paraboloid
and a schematic of its image in the xy-plane. This image is (C2 r V(x)) ∪ {(0, 0)}, the union of a
Zariski open subset of C2 and the variety {(0, 0)} = V(x, y).
A set is locally closed if it is open in its closure. In the Zariski topology, locally closed sets are
Zariski open subsets of some variety. A set is constructible if it is a finite union of locally closed
sets. What we saw with the hyperbolic paraboloid is the general case.

Theorem 12.1 The image of a constructible set under a polynomial map is constructible.

Suppose that X ⊂ Cd and ϕ : Cd → Cn is a polynomial map. Then the closure ϕ(X) of the
image of X under ϕ is a variety. When X is irreducible, then so is ϕ(X). (The inverse image of a
decomposition ϕ(X) = Y ∪ Z under ϕ is a decomposition of X.) Theorem 12.1 then implies that
ϕ(X) contains a nonempty Zariski open and therefore a Zariski dense subset of ϕ(X). Applying this
to each irreducible component of a general variety X ⊂ Cd implies that each irreducible component
of ϕ(X) has a dense open subset contained in the image ϕ(X).
280 Handbook of Geometric Constraint Systems Principles
z

y
x

Figure 12.5
The hyperbolic paraboloid and its image in the plane.

12.5 Real Algebraic Geometry

Real algebraic geometry predates its complex cousin, having its roots in Cartesian analytic geometry
in R2 . Its importance for applications is evident, and applications have driven some of its theoretical
development. A comprehensive treatment of the subject is given in the classic treatise of Bochnak,
Coste, and Roy [4]. Real algebraic geometry has long enjoyed links to computer science through
fundamental questions of complexity. There are also many specialized algorithms for treating real
algebraic sets. The equally classic book by Basu, Pollack, and Roy [1] covers this landscape of
complexity and algorithms.

Real algebraic geometry: Study of real solutions to systems of polynomial equations.

Real algebraic variety: A variety defined by real polynomials; its subset of real points.
X(R): The real points of a variety X defined by real polynomials.
Semi-algebraic set: A set defined by a system of polynomial equations and inequalities.
Tarski-Seidenberg Theorem: Quantifier elimination for semi-algebraic sets; the image of a semi-
algebraic set under a polynomial map is a semi-algebraic set.
Cylindrical algebraic decomposition: Algorithm to decompose a semi-algebraic set into a cell
complex of semi-algebraic cells adapted to quantifier elimination.
Positivestellensatz: Real-algebraic version of Nullstellensatz.

12.5.1 Algebraic Relaxation

A complex variety X ⊂ Cd defined by real polynomials has a subset X(R) := X ∩ Rd of real points.
Both X and (more commonly) X(R) are referred to as real algebraic varieties. In the Introduction,
we claimed that it is fruitful to study a real algebraic variety X(R) by first understanding the com-
plex variety X, and then asking about X(R). We consider studying the complex variety X to be
an algebraic relaxation of the problem of studying the real variety. The fundamental reason this
approach is often successful is the following result.

Theorem 12.2 Let X ⊂ Cd be an irreducible variety defined by real polynomials. If X has a smooth
real point, then X(R) is Zariski dense in X.
Real Algebraic Geometry for Geometric Constraints 281

To paraphrase, suppose that X ⊂ Cd is an irreducible variety defined by real polynomials. If X

has a smooth real point, then all algebraic and geometric information about X is already contained
in X(R), and vice-versa.
The reader may have noted that we used pictures of the real algebraic variety X(R) to illustrate
properties of the complex variety X in most of our figures. Theorem 12.2 justifies this sleight of
hand.
A proof of Theorem 12.2 begins by noting that when X has a smooth real point, then the set of
smooth real points Xsm (R) forms a real manifold of dimension dim X. Consequently, the derivatives
at a point of Xsm (R) of a polynomial f restricted to X are determined by the restriction of f to
Xsm (R), which implies that if a polynomial vanishes on Xsm (R), then it vanishes on X.
The two cones V(x2 +y2 −z2 ) and V(x2 +y2 +z2 ) serve to illustrate the hypotheses √ of Theo-
rem 12.2. In C3 , these cones are isomorphic to each other under the substitution z 7→ −1z. In
R3 , the first is the familiar double cone, with real smooth points the complement of the origin, while
the other is the single isolated (and hence singular) point {(0, 0, 0)}. We display the double cone on
the left in Figure 12.6. On the right is the Whitney umbrella. This is the Zariski closure of the image

Figure 12.6
Double cone and Whitney umbrella in R3 .

of R2 under the map (u, v) 7→ (uv, v, u2 ), and is defined by the polynomial x2 − y2 z. The image of
R2 is the canopy
√ of the umbrella. Its handle is the image of the imaginary part of the u-axis of C ,
2

the points (R −1, 0). The Whitney umbrella is singular along the z-axis, which is evident as the
canopy has self-intersection along the positive z-axis. This singularity along the negative z-axis is
implied by its having local dimension 1: were it smooth, it would have local dimension 2.
Theorem 12.2 also leads to the following cautionary example. The cubic y2 −x3 +x is irreducible
and its set of complex zeroes is a torus (with one point removed). Its set of real zeroes has two path-
connected components. Each is Zariski-dense in the complex cubic. Thus the property x ≤ 0 which
holds on the oval is not a generic property, even though it holds on a Zariski dense subset, which is
neither Zariski open or closed.

12.5.2 Semi-Algebraic Sets

The image of R2 in the Whitney umbrella is only its canopy, and not the handle. More interestingly,
the image under projection to the xy-plane of the sphere V(x2 + y2 + z2 − 1) of radius 1 and cen-
ter (0, 0, 0) is the unit disc {(x, y) ∈ R2 | 1−x2 −y2 ≥ 0}. Similarly, by the quadratic formula, the
282 Handbook of Geometric Constraint Systems Principles
y

Figure 12.7
Reprise: cubic plane curve.

polynomial x2 +bx+c in x has a real root if and only if b2 −4c ≥ 0. Thus, if we project the surface
V(x2 +bx+c) to the bc-plane, its image is {(b, c) ∈ R2 | b2 −4c ≥ 0}. We illustrate these examples in
Figure 12.8. They show that the image of an irreducible real variety under a polynomial map need

Figure 12.8
Projection of the sphere and the quadratic formula.

not be dense in the image variety, even though it will be dense in the Zariski topology. We describe
the image of a real variety by enlarging our notion of a real algebraic set.
A subset V of Rd is a semi-algebraic set if it is the union of sets defined by systems of polynomial
equations and polynomial inequalities. Technically, a set V is semi-algebraic if it is given by a
formula in disjunctive normal form, whose elementary formulas are of the form f (x) = 0 or f (x) >
0, where f is a polynomial with real coefficients. This is equivalent to V being given by a formula
that involves only the logical operations ‘and’ and ‘or’ and elementary formulas f (x) = 0 and f (x) >
0. Tarski showed that the image of a real variety under a polynomial map is a semi-algebraic set [20,
21], and Seidenberg gave a more algebraic proof [16].

Theorem 12.3 (Tarski-Seidenberg) The image of a semi-algebraic set under a polynomial map is
a semi-algebraic set.

The astute reader will note that our definition of a semi-algebraic set was in terms of propo-
sitional logic, and should not be surprised that Tarski was a great logician. The Tarski-Seidenberg
Theorem is known in logic as quantifier elimination: its main step is a coordinate projection, which
is equivalent to eliminating existential quantifiers.
Real Algebraic Geometry for Geometric Constraints 283

Example 40 We give a simple application from rigidity theory. Let G be a graph with n vertices
V and m edges. An embedding of G into Rd is simply a map ρ : V → Rd , and thus the space of
embeddings is identified with Rnd . The squared length of each edge of G in an embedding ρ defines
a map f : Rnd → Rm with image some set M. By the Tarski-Seidenberg Theorem, M is a semi-
algebraic set and so it contains an open subset of the real points of its Zariski closure, M. By Sard’s
Theorem, M contains a smooth point of its Zariski closure, and thus M has an open (and dense in
the classical topology) set of smooth points. These are images of embeddings where the Jacobian of
f (which is the rigidity matrix) has maximal rank (among all embeddings). 

Remark 12.1 Semi-algebraic sets are needed to describe more general frameworks involving
cables and struts. In an embedding, the length of an edge corresponding to a cable is bounded above
by the length of that cable, and the length of an edge corresponding to a strut is bounded below by
the length of that strut. In either case, inequalities are necessary to describe possible configurations.

The Tarski-Seidenberg Theorem is a structure theorem for images of real algebraic varieties
under polynomial maps. Much later, this existential result was refined by Collins, who gave an ef-
fective version of quantifier elimination for semi-algebraic sets, called cylindrical algebraic decom-
position [6]. This uses successive coordinate projections to build a description of a semi-algebraic
set as a cell complex whose cells are semi-algebraic sets. While implemented in software [5], it
suffers more than many algorithms in this subject from the curse of complexity and is most effec-
tive in low (d . 3) dimensions. There are however several software implementations of cylindrical
algebraic decomposition. In the worst case, the complexity of a cylindrical algebraic decomposition
is doubly exponential in d, and this is achieved for general real varieties. A focus of [1] and subse-
quent work is on stable algorithms with better performance to compute different representations of
a semi-algebraic set.

12.5.3 Certificates
We close with the Positivestellensatz of Stengele [19], which states that a semi-algebraic set is
empty if and only if there is a certificate of its emptiness having a particular form. A polynomial σ
is a sum of squares if it may be written as a sum of squares of polynomials with real coefficients.
Such a polynomial takes only nonnegative values on Rd . We may use semidefinite programming to
determine if a polynomial is a sum of squares.

Theorem 12.4 (Positivestellensatz) Suppose that f1 , . . . , fr , g1 , . . . , gs , and h are real polynomials.

Then the semi-algebraic set

{x ∈ Rd | fi (x) = 0, i = 1, . . . , r and g j (x) ≥ 0, k = 1, . . . , s and h(x) 6= 0} (12.1)

is empty if and only if there exist polynomials k1 , . . . , kr , sums of squares σ0 , . . . , σs , and a positive
integer n such that

0 = f1 k1 + · · · + fr kr + σ0 + g1 σ1 + · · · + gs σs + h2n . (12.2)

Remark 12.2 To see that (12.2) is a sufficient condition for emptiness, suppose that x lies in
the set (12.1), and then evaluate the expression (12.2) at x. The terms involving fi vanish, those
involving g j are nonnegative, and h(x)2n > 0, which is a contradiction. If h does not appear in a
description (12.1), then we take h = 1 in (12.2).
284 References

References
[1] Saugata Basu, Richard Pollack, and Marie-Françoise Roy. Algorithms in real algebraic geo-
metry, volume 10 of Algorithms and Computation in Mathematics. Springer-Verlag, Berlin,
second edition, 2006.
[2] Daniel J. Bates, Jonathan D. Hauenstein, Andrew J. Sommese, and Charles W. Wampler.
Bertini: Software for numerical algebraic geometry. www.nd.edu/~sommese/bertini.
[3] Daniel J. Bates, Jonathan D. Hauenstein, Andrew J. Sommese, and Charles W. Wampler.
Numerically solving polynomial systems with Bertini, volume 25 of Software, Environments,
and Tools. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2013.
[4] Jacek Bochnak, Michel Coste, and Marie-Françoise Roy. Real algebraic geometry, volume 36
of Ergebnisse der Mathematik und ihrer Grenzgebiete (3). Springer-Verlag, Berlin, 1998.
[5] Christopher W Brown. Qepcad b: a program for computing with semi-algebraic sets using
cads. ACM SIGSAM Bulletin, 37(4):97–108, 2003.
[6] George E. Collins. Quantifier elimination for real closed fields by cylindrical algebraic de-
composition. In Automata theory and formal languages (Second GI Conf., Kaiserslautern,
1975), pages 134–183. Lecture Notes in Comput. Sci., Vol. 33. Springer, Berlin, 1975.
[7] David Cox, John Little, and Donal O’Shea. Ideals, varieties, and algorithms. Undergraduate
Texts in Mathematics. Springer, New York, third edition, 2007.
[8] Wolfram Decker, Gert-Martin Greuel, Gerhard Pfister, and Hans Schönemann.
S INGULAR 4-0-2 — A computer algebra system for polynomial computations.
www.singular.uni-kl.de, 2015.
[9] Wolfram Decker and Christoph Lossen. Computing in algebraic geometry, volume 16 of
Algorithms and Computation in Mathematics. Springer-Verlag, Berlin, 2006.
[10] The Sage Developers. Sage Mathematics Software, 2015. www.sagemath.org.
[11] David Eisenbud, Daniel R. Grayson, Michael Stillman, and Bernd Sturmfels, editors. Com-
putations in algebraic geometry with Macaulay 2, volume 8 of Algorithms and Computation
in Mathematics. Springer-Verlag, Berlin, 2002.
[12] Daniel R. Grayson and Michael E. Stillman. Macaulay2, a software system for research in
algebraic geometry. www.math.uiuc.edu/Macaulay2/.
[13] Jonathan D. Hauenstein and Frank Sottile. Algorithm 921: alphaCertified: certifying solutions
to polynomial systems. ACM Trans. Math. Software, 38(4):Art. ID 28, 20, 2012.
[14] Robert Krone and Anton Leykin. NAG4M2: Numerical algebraic geometry for Macaulay 2.
people.math.gatech.edu/~aleykin3/NAG4M2.
[15] Alexander Morgan. Solving polynomial systems using continuation for engineering and sci-
entific problems. Prentice Hall Inc., Englewood Cliffs, NJ, 1987.
[16] Abraham Seidenberg. A new decision method for elementary algebra. Ann. of Math. (2),
60:365–374, 1954.
[17] Stephen Smale. Newton’s method estimates from data at one point. In The merging of dis-
ciplines: new directions in pure, applied, and computational mathematics (Laramie, Wyo.,
1985), pages 185–196. Springer, New York, 1986.
References 285

[18] Andrew J. Sommese and Charles W. Wampler, II. The numerical solution of systems of poly-
nomials. World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2005.
[19] Gilbert Stengle. A nullstellensatz and a positivstellensatz in semialgebraic geometry. Math.
Ann., 207:87–97, 1974.
[20] Alfred Tarski. A Decision Method for Elementary Algebra and Geometry. RAND Corpora-
tion, Santa Monica, CA., 1948.
[21] Alfred Tarski. A decision method for elementary algebra and geometry. In Quantifier elimina-
tion and cylindrical algebraic decomposition (Linz, 1993), Texts Monogr. Symbol. Comput.,
pages 24–84. Springer, Vienna, 1998.
[22] Jan Verschelde. Algorithm 795: PHCpack: general-purpose solver for polynomial systems by
homotopy. ACM Trans. Math. Software, 25(2):251–276, 1999.
 Part III

Geometric Rigidty
Chapter 13
Polyhedra in 3-Space
Brigitte Servatius
WPI

CONTENTS
13.1 Euler’s Conjecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
13.2 Cauchy’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
13.3 Co-Dimension 2 Results – Bricard Octahedra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
13.4 Polyhedral Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297

13.1 Euler’s Conjecture

We give a brief history of rigidity results on polyhedra and point to some open problems. In 1766,
Euler conjectured: “A closed spatial figure allows no changes, as long as it is not ripped apart,”
see [12]. He ends the paragraph describing the problem with the sentence:
Interim patet hemispherii figuram certe esse mutabilem; cujusmodi autem mutationes
recipere possit, problema videtur difficillimum
indicating that Euler’s approach to the problem of finding motion in the sphere, considering how to
match the motions independently obtainable on the upper and lower hemispheres, anticipated the
approaches of Cauchy, Bricard, Connelly, Steffen, and others.

13.2 Cauchy’s Theorem

A first step toward proving the Euler Conjecture and the first major result published in rigidity
theory was a theorem of Cauchy in the early 19’th century:
Theorem 13.1 (A. L. Cauchy, 1813 [5]) If there is an isometry between the surfaces of two strictly
convex polyhedra which is an isometry on each of the faces, then the two polyhedra are congruent.
By a strictly convex polyhedron, we mean a polyhedron which has the property that, for each ver-
tex, there exists a plane which intersects the polyhedron at that vertex and no other point of the
polyhedron. Cauchy’s proof involved a mixture of topological arguments and arguments from el-
ementary geometry. Unfortunately, there were several minor errors in the proof; the first complete

289
290 Handbook of Geometric Constraint Systems Principles

proof was published by Steinitz and Rademacher in 1934 [21]. A simpler complete proof of a slight
generalization of Cauchy’s Theorem was given by Alexandrov in 1950 [1].
It is easy to see that the convexity condition cannot be dropped from the hypothesis of Cauchy’s
Theorem. Consider the pair of polyhedra constructed as follows. Take a cube and a pyramid with
base congruent to a face of the cube and with height less than the height of the cube. Remove the
base from the pyramid and remove one face from the cube. Then join these two surfaces along their
boundaries, once with the apex of the pyramid pointing out and once with it pointing in. There is an
obvious homeomorphism between these two surfaces which restricts to an isometry on each face;
but these two polyhedra are not congruent. See Figure 13.1.

Figure 13.1
Two isomorphic but non-congruent polyhedra with corresponding faces congruent.

The boundary of the 3-dimensional polyhedron of Cauchy’s theorem is the set of 0, 1, and 2-
dimensional faces, the 2-skeleton. The following corollary is an immediate consequence of Cauchy’s
Theorem.

Corollary 13.2 The 2-skeleton of a strictly convex polyhedron in 3-space is rigid.

13.3 Co-Dimension 2 Results – Bricard Octahedra

With the Euler Conjecture in mind, it is natural to ask if “strictly convex” may be dropped from the
hypothesis of the corollary.
In 1897 Bricard published a complete study of the rigidity properties of the octahedron. The
main result of that study can be stated as follows.
Theorem 13.3 (R. Bricard, 1897, [4]) The 2-skeleton of any polyhedral embedding of the octahe-
dron in 3-space is rigid. However, the 1-skeleton of the octahedron has a non-rigid embedding in
3-space.
At first glance the two parts of this theorem seem to be in contradiction to one another: since
the faces of the octahedron are all triangular, deleting them could not change a rigid plate and
hinge framework into a non-rigid rod and joint framework, see Chapter 13. It would then seem to
follow from the first part of the theorem that any embedding of the 1-skeleton of the octahedron
would also be rigid, however, not every embedding of the 1-skeleton extends to an embedding of
Polyhedra in 3-Space 291

the 2-skeleton. Non-rigid embeddings of the 1-skeleton of the octahedron are easy to describe. Note
first that any quadrilateral framework in R3 whose opposite sides are parallel must have a 180◦
axis of symmetry. Consider a non-regular pyramid with, say, a square base and assume that the
perpendicular projection of the apex onto the base does not lie any line of symmetry of the square;
see Figure 13.2a, b. This last condition is just to ensure that the added bars do not intersect any of the

a) b) c) d)

Figure 13.2
How to embed an octahedral framework in R3 such that it has a motion.

original bars in their interior. Now erect a second pyramid on the same base which is isometric to the
first under the 180◦ rotation about the line perpendicular to the base at its center; see Figure 13.2c,
d. Note that both apices are on the same side of their common base. The reason that this octahedral
1-skeleton is non-rigid is that the 180◦ rotation of the base framework will always interchange the
pyramids, and the motion of the pyramid extends to the larger framework. In other words both
pyramids produce the same family of distortions of the rectangular base as they flex.
To understand the relationship between the two parts of Bricard’s Theorem, consider the 1-
skeleton of the octahedron embedded as just described. Since the faces of the octahedron are trian-
gles, adding these faces to this 1-skeleton can be done without destroying the planarity of those faces
and without limiting any motions of the structure. We thus have a copy of the 2-skeleton of the oc-
tahedron in 3-space which is not rigid. However, it is clearly not embedded: it has self-intersections
which always involve interior points of some triangles.
The next new result on the Euler Conjecture did not occur until many years later when Alexan-
drov gave his generalization of Cauchy’s Theorem.
Theorem 13.4 (A. D. Alexandrov, 1950 [1]) If vertices are inserted in the edges of a strictly con-
vex polyhedron and the faces are triangulated, then the 1-skeleton of the resulting polyhedron is
infinitesimally rigid.
This result has the following corollary:
Corollary 13.5 If a convex polyhedron in 3-space has the property that the collection of faces
containing a given vertex do not all lie in the same plane, then the 2-skeleton of that polyhedron is
infinitesimally rigid. See Figure 13.3.
In 1978, Asimow and Roth showed that the triangulation condition in the hypothesis of Alexan-
drov’s Theorem is necessary.

Theorem 13.6 (L. Asimow and B. Roth, 1978 [3]) The 1-skeleton of a strictly convex polyhedron
embedded in 3-space which has at least one non-triangular face is not rigid.

Asimow and Roth’s proof of this theorem has two parts. The first combinatorial part uses Euler’s
formula for the sphere to show that the number of edges in such a 1-skeleton is less than 3v − 6, the
number of edges in a triangulation of a sphere with with v vertices. Then, by dimension arguments
292 Handbook of Geometric Constraint Systems Principles

Figure 13.3
This convex but not strictly convex polyhedron has an infinitesimally rigid 2-skeleton by the Corol-
lary to Alexandrov’s Theorem. Since all its faces are triangles, the 1-skeleton is also infinitesimally
rigid.

they conclude that there are not enough constraints to make this framework rigid unless it represents
a singular point on the algebraic variety C derived from the edge length constraints, see Chapter 18.
In the second, geometric part of their proof, Asimow and Roth show that a strictly convex framework
cannot correspond to a singular point.
To see that strict convexity is essential to their result, consider a tetrahedron with equilateral
faces. Insert, centered in one face, a smaller equilateral triangle with sides parallel to the sides of
the face. Now, join by edges the corresponding vertices; see Figure 13.4.

s
ppp@
 ps
  pp @ @
@

s
 p
p @H s @
pp p
 
  H@
sa

 
pp @s
HH
aa pp 

pp
aa 

psp 
aa 
aap

Figure 13.4
This is a convex non-triangulated polyhedron which is rigid in R3 , but not infinitesimally rigid as a
panel and hinge structure. The polyhedron has three non-triangular faces, yet the 1-skeleton is easily
seen to be rigid (second order) with the small triangle held in place by tension.

13.4 Polyhedral Surfaces

Stoker [22] extends Cauchy’s rigidity theorem to some more general 3-dimensional polyhedra, in-
cluding special types of non-convex ones and polyhedra with boundary. Cauchy’s original proof is
refined, together with some lemmas of Alexandrov to obtain the results. Stoker views the underlying
problem as that of determining when two polyhedra are isogonal, that is, when corresponding faces
are parallel. Additional conditions are then imposed to ensure congruence.
Conjecture 2 (J. J. Stoker) Two combinatorially equivalent polyhedra with equal corresponding
dihedral angles are isogonal.
Polyhedra in 3-Space 293

a a'
17
a a'
12
10
c' c'''
5
c' c c'' 11 c'''

c c''

b' b'
b b

Figure 13.5
A net for Steffen’s construction of a flexible sphere with only 9 vertices, 14 triangles in 4 congruence
classes, and 21 edges. The valley folds are dashed.

Karcher [18] gives an affirmative answer to Stoker’s problem in some special cases, in particular for
simple 3-polytopes.
The Euler Conjecture was finally settled in 1977, but just prior to that a result was published
which greatly extended the set of configurations for which the Euler Conjecture was known to be
valid.
Theorem 13.7 (H. Gluck, 1975 [15]) Every closed simply connected polyhedral surface embed-
ded in 3-space is generically rigid.
Gluck’s Theorem tells us that the Euler Conjecture is almost always true for closed simply
connected polyhedral surfaces. Just two years later, a counterexample to the Euler Conjecture was
found by R. Connelly [7]. The counterexample is based on Bricard’s flexible octahedron and, of
course, it is non-convex and non-generic. Topologically Connelly’s surface is a sphere, thus showing
that Gluck’s Theorem is best possible. Connelly’s flexible sphere, described in [8], has 11 vertices,
18 triangular faces and 27 edges. Klaus Steffen gave a smaller example of a “Connelly sphere” with
only 9 vertices, and it is very often built according to Steffen’s original instructions communicated
to Connelly in a handwritten note in 1977. A net for Steffen’s example is given in Figure 13.5 which
may be somewhat challenging to assemble.
Connelly showed that the volume enclosed by his flexible spheres remains constant under the
motion. The conjecture, by Connelly and Sullivan, that every orientable closed polyhedral surface
flexes with constant volume [9] is known as the bellows conjecture, since it implies that a mathe-
matical bellows is impossible to construct.
The motion of Steffen’s polyhedron is not so easy to observe, he suggests to cut a hole in one
of the triangles to watch the small motion that stops when self intersection occurs. Euler sensed
the difficulty of finding a truly flexible sphere and Connelly’s spheres are ingenious examples. One
may, however, easily create families of infinitesimally flexible polyhedra. For example, starting
with a regular tetrahedron of Figure 13.6, which in the illustration is placed in a unit cube for visual
reference, then subdivide the top and bottom edges and displace by one unit as shown. The middle
framework has an infinitesimal motion which may be extended to a new point, placed symmetrically
and connected to all the vertices of one of the four quadrilaterals. If the point is placed at (p, p,t),
and has infinitesimal motion (p0 , p0 ,t 0 ), then the parameters must satisfy

t 0 = 6 − t, ss0 = (t − 2)2 , (s − 1)(s0 − 1) = (t + 1)(t − 3)

294 Handbook of Geometric Constraint Systems Principles

Figure 13.6
(a) Tetrahedron with vertices (1, 1, −1), (1, −1, 1), (−1, 1, −1), (1, 1, −1); (b) adding (0, 0, 2) and
(0, 0, −2). A symmetric infinitesimal motion with the top and bottom velocities both√(0, 0, 2) and
the middle having zero vertical component; (c) a horn placed at (p, p, p) with p = ( 1+2 5 )2 .

If one places a point symmetrically for each of the other four quadrilaterals, then this yields a trian-
gulated sphere, see Figure 13.7, which has dihedral symmetry and is infinitesimally flexible. In [24]

Figure 13.7
A four-horn. If one pinches the bottom horns infinitesimally together, the top two are infinitesimally
pulled apart.

Wunderlich and Schwabe construct a family of polyhedra, called four-horns, which are combina-
torially isomorphic to the one in Figure 13.7. Four-horns have congruent faces and are “almost
movable” in the sense that each four-horn has three non-congruent embeddings, see Figure 13.8,
and the “snapping” of the model from one form to the other requires just a small deformation of the
edges, for some examples less than 1%. Two of the embeddings are perfectly flat, while the third
embedding encloses positive volume. Cycling through the fake motion involves connecting four true
conformations, since the non-flat conformation will be transformed into its mirror image, provided
we allow the panels to pass through one another in the flat position. Like so many examples consid-
ered in this section, four-horns are constructed just as Euler suggested in the Latin sentence of the
introduction, by matching a real motion of the top two horns with those of the bottom. A physical
model appears more flexible than Steffen’s polyhedron and feels like a genuine counterexample to
the bellows conjecture.
Polyhedra in 3-Space 295

Figure 13.8
The three forms of the Wunderlich-Schwabe four-horn.

A similar phenomenon of apparent model flexibility is observed in the Jessen Icosahedron [17],
see Figure 13.9. The Jessen icosahedron starts with an irregular icosahedron drawn in the unit cube.

Figure 13.9
The Jessen icosahedron.

It may be taken to have vertices (1, c, 0) and all 12 permutations thereof by the rotational group
of one of the tetrahedra embedded in the unit cube, just as for the regular icosahedron, except that
instead of taking c to be the golden ration, Jessen takes c = .5. The resulting irregular icosahedron is
vertex but not face transitive, and has two congruence classes of triangles, equilateral at the corners
of the cube, and six pairs of isosceles triangles, one at each cube face. To form the Jessen icosa-
hedron, the isosceles pairs are removed and replaced with the pair of isosceles triangles formed by
adding the other diagonal to each quadrilateral boundary. These six “beaks” have a right angled
valley fold, and indeed all angles of the non-convex icosahedron are right angles, and there is an
infinitesimal motion with velocities along the deleted edges, simultaneously closing all the beaks.
It is interesting to note that this motion, infinitesimally closing the breaks and screwing-in the equi-
lateral triangles, half clockwise and half counter-clockwise, obviously decreases infinitesimally the
enclosed volume, so the Jessen icosahedron, unlike the four-horn, is an infinitesimal bellows. The
Jessen icosahedron takes a special place in the family of Douday Shaddocks [11], which Douady
calls shaddocks with six beaks. Combinatorially, members of the shaddock family are isomorphic
to icosahedra. Just as the Wunderlich-Schwabe model is rigid and only infinitesimally flexible, all
members of the Shaddock family are rigid, and only the Jessen Icosahedron is infinitesimally flexi-
ble. In [16] it is shown that a mere .01% change in edge length in the Jessen Icosahedron produces a
change in the dihedral angles of almost 10%, which is a good explanation of the observed flexibility:
296 Handbook of Geometric Constraint Systems Principles

The change in the faces goes unnoticed, while the change in the angles is apparent in manipulating
the model. As a second explanation, the beaks may be replaced by broken beaks, thereby changing
the combinatorial properties through the introduction of new crease-lines, but these new crease-
lines are arbitrarily close to the original edges. The new object is truly flexible, and again optically
indistinguishable from the original.
By putting three 2-horns together Wunderlich and Schwabe obtain a 6-horn, which they show
to be second order flexible. However, 2n-horns for n > 3 are shown to be rigid.
The bellows conjecture was proven by Sabitov [19] for spherical polyhedra in 1995 and for the
general case of triangulated oriented surfaces by Connelly, Sabitov, and Walz [6] in 1997. A short
proof of the Bellows Theorem is given in [20], together with a good overview of related research
following the original proof and many open problems.
It is well known that two embedded polyhedra with the same volume and the same Dehn in-
variants are scissors congruent [10, 23]. The strong bellows conjecture [2] states that if there is a
continuous flex from one, possibly singular, polyhedron P to another, possibly singular, polyhedron
Q, then P and Q have the same Dehn invariants.
Gaı̆fullin, in [14], proves the bellows conjecture for all odd dimensional Lobachevsky spaces.
In [13] he finds, for all n ≥ 2, embedded flexible cross polytopes in the unit sphere Sn with non-
constant volume and formulates the modified bellows conjecture stating that some vertices of a
flexible polyhedron in Sn can be replaced with their antipodes so that the generalized volume of the
resulting flexible polyhedron will remain constant during the flex.

Glossary
n-skeleton: Then n skeleton of a cell complex n is the subcomplex consisting of all cells of dimen-
sion n or less.
Bricard octahedra: One of the collection of bar and joint frameworks whose graph is that of the
octahedron, and which flexes in R3 .
Connelly sphere: A triangulated sphere, embedded in R3 which is flexible. It is necessarily non-
convex.
Dehn invariant: An algebraic measurement assigned to a three-dimensional polyhedron in R3
which is defined using the dihedral angles, see [10]. It is invariant under the operation of dis-
secting the polyhedron into a finite number of pieces and reassembling them to form a new
polyhedron, which is therefore scissors congruent to the first.
isogonal: Relation between two combinatorially isomorphic polyherda whose corresponding faces
are parallel.
Jessen icosahedron: A construction of an infinitesimally flexible sphere whose 1-skeleton is the
graph of the icosahedron.
polyhedral surface: A 2-dimensional cell complex homeomorphic to a sphere and embedded in
R3 such that the faces are plane polygons.
4-horn: One of a collection of infinitesimally flexible polyhedral surfaces due to Wunderlich.
scissors congruent: Two polyhedra scissors-congruent if the first can be cut into finitely many
polyhedral pieces that can be reassembled to yield the second. The Dehn invariant is invariant
under scissors congruence.
References 297

strictly convex polyhedron: A polyhedron each of whose vertices is the sole point of the polyhe-
dron which intersects some plane.

References
[1] A. D. Alexandrov. Convex polyhedra. Springer Monographs in Mathematics. Springer-Verlag,
Berlin, 2005. Translated from the 1950 Russian edition by N. S. Dairbekov, S. S. Kutateladze
and A. B. Sossinsky, With comments and bibliography by V. A. Zalgaller and appendices by
L. A. Shor and Yu. A. Volkov.
[2] Victor Alexandrov and Robert Connelly. Flexible suspensions with a hexagonal equator.
Illinois J. Math., 55(1):127–155 (2012), 2011.
[3] L. Asimow and B. Roth. The rigidity of graphs. Trans. Amer. Math. Soc., 245:279–289, 1978.
[4] Raoul Bricard. Mémoire sur la théorie de l’octahèdre articulé. J. math. pures et appliquées,
3:113–150, 1897.
[5] A.L. Cauchy. Recherches sur les (polygones et les) polyèdres: .... Mémoire I. 1813.
[6] R. Connelly, I. Sabitov, and A. Walz. The bellows conjecture. Beiträge Algebra Geom.,
38(1):1–10, 1997.
[7] Robert Connelly. A counterexample to the rigidity conjecture for polyhedra. Inst. Hautes
Études Sci. Publ. Math., (47):333–338, 1977.
[8] Robert Connelly. A flexible sphere. Math. Intelligencer, 1(3):130–131, 1978/79.
[9] Robert Connelly. Conjectures and open questions in rigidity. In Proceedings of the Inter-
national Congress of Mathematicians (Helsinki, 1978), pages 407–414. Acad. Sci. Fennica,
Helsinki, 1980.
[10] M. Dehn. Ueber den Rauminhalt. Math. Ann., 55(3):465–478, 1901.
[11] A. Douady. Le shaddock à six becs. Bulletin A.P.M.E.P., 281:699–701, 1971.
[12] L. Euler and Academiae Scientiarum Petropolitanae. Leonhardi Euleri Opera postuma math-
ematica et physica: anno MDCCCXLIV detecta. Number v. 2 in Leonhardi Euleri Opera
postuma mathematica et physica: anno MDCCCXLIV detecta. Eggers, 1862.
[13] Alexander A. Gaifullin. Embedded flexible spherical cross-polytopes with nonconstant vol-
umes. Proc. Steklov Inst. Math., 288(1):56–80, 2015.
[14] A. A. Gaı̆ fullin. The analytic continuation of volume and the bellows conjecture in
Lobachevskiı̆ spaces. Mat. Sb., 206(11):61–112, 2015.
[15] Herman Gluck. Almost all simply connected closed surfaces are rigid. Lecture Notes in Math.,
438:225–239, 1975.
[16] V. Gorkavyy and D. Kalinin. On model flexibility of the Jessen orthogonal icosahedron. Beitr.
Algebra Geom., 57(3):607–622, 2016.
[17] Bø rge Jessen. Orthogonal icosahedra. Nordisk Mat. Tidskr, 15:90–96, 1967.
[18] Hermann Karcher. Remarks on polyhedra with given dihedral angles. Comm. Pure Appl.
Math., 21:169–174, 1968.
298 References

[19] I. Kh. Sabitov. On the problem of the invariance of the volume of a deformable polyhedron.
Uspekhi Mat. Nauk, 50(2(302)):223–224, 1995.
[20] I. Kh. Sabitov. Algebraic methods for the solution of polyhedra. Uspekhi Mat. Nauk,
66(3(399)):3–66, 2011.
[21] Ernst Steinitz and Hans Rademacher. Vorlesungen über die Theorie der Polyeder unter Ein-
schluss der Elemente der Topologie. Springer-Verlag, Berlin-New York, 1976. Reprint der
1934 Auflage, Grundlehren der Mathematischen Wissenschaften, No. 41.
[22] J. J. Stoker. Geometrical problems concerning polyhedra in the large. Comm. Pure Appl.
Math., 21:119–168, 1968.
[23] J.-P. Sydler. Conditions nécessaires et suffisantes pour l’équivalence des polyèdres de l’espace
euclidien à trois dimensions. Comment. Math. Helv., 40:43–80, 1965.
[24] W. Wunderlich and C. Schwabe. Eine Familie von geschlossenen gleichflächigen Polyedern,
die fast beweglich sind. Elem. Math., 41(4):88–98, 1986.
Chapter 14
Tensegrity
Robert Connelly
Department of Mathematics, Cornell University, Ithaca, NY

Anthony Nixon
Department of Mathematics and Statistics, Lancaster University, U.K.

CONTENTS
14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
14.2 Tensegrity Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
14.2.1 Combinatorics of Tensegrities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
14.2.2 Geometric Interpretations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
14.2.3 Packings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
14.3 Types of Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
14.3.1 Global Rigidity and Stress Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
14.3.2 Universal and Dimensional Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
14.3.3 Operations on Tensegrities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
14.4 Examples and Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
14.4.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
14.4.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311

14.1 Introduction
Tensegrities were invented and implemented by Kenneth Snelson in 1947 and other artists who
were intrigued by the almost magical way the sticks could be suspended in midair. They were
called tensegrities by R. Buckminster Fuller [31, 32] because of their “tensional integrity” and he
popularised their use.
As well as their mathematical interest and artistic beauty, tensegrities arise in a number of ap-
plication areas. In engineering (see, among others, [7, 36]), tensegrity structures provide efficient
solutions for applications in deployable structures [51, 59], mechanisms [8, 43], multi agent systems
[45], interesting examples of form finding [42, 66], algorithms for synthesis and analysis [27, 52]
and smart sensors [53] as well as being intriguingly light relative to their stability. In biology, tenseg-
rity structures are employed as models underlying the behavior of entities such as the cytoskeleton
[38, 39]. Let us also mention Skelton’s Type 1 (when no pair of bars is adjacent) and Type 2 tenseg-
rities (where bar adjacencies are permitted) [65]. Skelton’s interest in tensegrities arose from engi-
neering and control theory. We focus on the mathematics of tensegrities in this chapter. A popular
attempt at a logical explanation of Fuller’s idea’s for tensegrities was given by Edmondson [29].

299
300 Handbook of Geometric Constraint Systems Principles

Mathematically, a fundamental problem in geometry is to determine when selected distance

constraints, on a finite number of points, fix these points up to congruence, at least for small pertur-
bations. In the language of rigidity theory, we consider a realization of a graph as a finite set of points
where certain pairs (corresponding to the edges) represent one of three types of constraint: cables
are constrained not to get further apart; struts not to get closer together; and bars are constrained to
stay the same distance apart.
We are then interested first, in whether there is a continuous motion of the points satisfying these
constraints which is not simply the restriction of a rigid motion of the ambient Euclidean space. This
is the rigidity question. Stronger is the global rigidity question which asks if the realization is unique
up to rigid motions. In this chapter we will review the theory of tensegrity frameworks and point out
connections with other chapters.

14.2 Tensegrity Frameworks

Tensegrity: A tensegrity framework (G, p) in Rd is a graph G = (V, E) where E has been parti-
tioned into 3 sets E− , E0 and E+ , together with a map p : V → Rd realising the vertices as
points in Rd . The members in E− are known as cables, the members in E0 are bars, and the
members in E+ are struts. If E = E0 then we have a standard bar-joint framework. Cables
cannot increase in length, bars have fixed length and struts cannot decrease in length.

Dominates: A tensegrity framework (G, p) dominates the tensegrity framework (G, q), written
(G, p) ≥ (G, q), if

kpi − p j k ≥ kqi − q j k when i j ∈ E− ,

kpi − p j k = kqi − q j k when i j ∈ E0 ,
kpi − p j k ≤ kqi − q j k when i j ∈ E+ .

Locally rigid: A tensegrity framework (G, p) is locally rigid (or continuously rigid, or often just
rigid) if there is an ε > 0 such that whenever (G, p) dominates (G, q) and kp − qk < ε then p is
congruent to q.
Infinitesimal flex: An infinitesimal flex of a tensegrity (G, p) is an assignment p0 : V → Rd such
that for each edge i j ∈ E we have

(pi − p j ) · (p0i − p0j ) ≤ 0 for cables i j ∈ E− ,

(pi − p j ) · (p0i − p0j ) = 0 for bars i j ∈ E0 ,
(pi − p j ) · (p0i − p0j ) ≥ 0 for struts i j ∈ E+ .

Infinitesimally rigid: A tensegrity framework (G, p) is infinitesimally rigid if every infinitesimal

flex is an infinitesimal isometry of Rd .

Stress: A stress ω = (. . . , ωi j , . . . ) ∈ R|E| on a tensegrity framework (G, p) is an assignment of

scalars to the edges.
Tensegrity 301

Equilibrium stress: A stress is an equilibrium stress if

X
ωi j (p j − pi ) = 0 for all i ∈ V,
j

where the sum is taken over all j ∈ V − i, and if i j ∈

/ E then ωi j = 0.
Proper stress: An equilibrium stress is a proper stress if ωi j ≥ 0 for all i j ∈ E− and ωi j ≤ 0 for all
i j ∈ E+ (there is no condition for the stresses on the bars).

Strict proper stress: An equilibrium stress is a strict proper stress if ωi j > 0 for all i j ∈ E− and
ωi j < 0 for all i j ∈ E+ (there is no condition for the stresses on the bars).
Underlying bar-joint framework: For a tensegrity framework (G, p), let (Ḡ, p) denote the under-
lying bar-joint framework that arises by replacing all members of (G, p) by bars.

Equilibrium load: An equilibrium load on a tensegrity framework (G, p) is an assignment F : V →

Rd such that for each trivial infinitesimal flex p0 ∈ Rd|V | we have F · p0 = 0.
an equilibrium load F by a tensegrity framework (G, p) is an equilib-
Resolution: A resolution ofP
rium stress ω such that j ωi j (p j − pi ) + Fi = 0.

Statically rigid: A tensegrity framework (G, p) is statically rigid if every equilibrium load is re-
solvable.
Prestress stable: A tensegrity framework (G, p) in Rd is prestress stable if there is a proper
equilibrium stress ω such that for every non-trivial infinitesimal flex p0 of (G, p) we have
0 − p0 )2 > 0.
P
ω (p
i< j i j i j

Second-order flex: A second-order flex (p0 , p00 ) for a tensegrity framework (G, p) is a solution to
the following constraints:
(a) for a bar i j: (pi − p j ) · (p0i − p0j ) = 0 and kp0i − p0j k2 + (pi − p j ) · (p00i − p00j ) = 0;
(b) for a cable i j: (pi − p j )(p0i − p0j ) = 0 and ||p0i − p j ||2 + (pi − p j )(p00i − p0 j ) ≤ 0 or (pi −
p j )(p0i − p0j ) < 0; and
(c) for a strut i j: (pi − p j )(p0i − p0j ) = 0 and ||p0i − p j ||2 + (pi − p j )(p00i − p0 j ) ≥ 0 or (pi −
p j )(p0i − p0j ) > 0.
Second-order rigid: A tensegrity framework (G, p) is second-order rigid if all second-order in-
finitesimal flexes (p0 , p00 ) have p0 as a trivial infinitesimal flex.

Tensegrity frameworks arise from bar-joint frameworks by replacing some of the fixed distance
constraints by inequalities. When the distance is forced not to increase we call the edges cables and
when the distance is forced not to decrease we call the edges struts.
The basic definitions of rigidity and infinitesimal rigidity can be easily extended to the setting
of tensegrities. In Figure 14.1 we give an example of a rigid tensegrity framework and a different
realization of the same tensegrity graph which is not rigid. This example makes it immediately clear
that the rigidity or infinitesimal rigidity of the underlying bar-joint framework is not sufficient for
a tensegrity framework to be rigid or infinitesimally rigid (see also Figure 14.3). We also note the
following partial equivalence of rigidity and infinitesimal rigidity.

Theorem 14.1 (Connelly and Whiteley, 1996 [26]) Let (G, p) be a tensegrity framework where:
the vertices of G are realised as a strictly convex polygon; the bars form a Hamilton cycle on the
boundary of this polygon; and there are no struts. Then (G, p) is rigid if and only if it is infinitesi-
mally rigid.
302 Handbook of Geometric Constraint Systems Principles

p1 1 p3 p1 p3

p4
1 1
−1 −1

p2 1 p4 p2

Figure 14.1
A simple tensegrity framework in R2 with a proper equilibrium stress indicated. Cables are repre-
sented by dashed lines and struts by doubled lines. This tensegrity is rigid. A second realization of
the same tensegrity graph, again in R2 , which is flexible.

p1 p1

p3 p5 p3 p5

p2 p4 p2 p4

Figure 14.2
Two infinitesimally rigid tensegrities (in the plane). The tensegrity on the right comes from the one
on the left by interchanging the cables and struts.

p1 p1

p3 p3
p2 p4 p2 p4

Figure 14.3
The tensegrity on the left is rigid but infinitesimally flexible (in R3 ). On the right the tensegrity with
cables and struts interchanged is flexible and infinitesimally flexible (again in R3 ).
Tensegrity 303
p1 p1

p6
p5 p4 p3
p2 p3 p2 p5 p4

Figure 14.4
Examples of pre-stress stable tensegrity frameworks (in the plane) which are not infinitesimally
rigid.

Interchanging cables and struts preserves infinitesimal rigidity, as we illustrate in Figure 14.2.
Note however that rigidity is not preserved by this process, see Figure 14.3.
Equilibrium stresses (see Chapter 16), with their physical interpretation as forces on the edges,
can also easily be extended to tensegrities (typically with positive stresses on cables and negative
stresses on struts), in fact they were originially introduced to the rigidity literature in the context
of tensegrities [12]. The extra condition is simply that for an equilibrium stress to be proper, the
stress on each cable must be positive and the stress on each strut must be negative. The fundamental
importance of equilibrium stresses to the rigidity of tensegrities is indicated in the following basic
result.
Theorem 14.2 (Connelly, 1982 [12]) Let (G, p) be a tensegrity framework with at least one cable
or strut. Suppose (G, p) is rigid, then there is a non-zero proper equilibrium stress.
Connelly and Whiteley [26] proved that infinitesimal rigidity implies prestress stability which in
turn implies second-order rigidity which then implies rigidity for any tensegrity framework. Exam-
ples show that none of these implications are reversible. Figure 14.4 shows examples of pre-stress
stable tensegrity frameworks which are not infinitesimally rigid. In Figure 14.4a there is a rotational
motion on p4 , p5 , p6 and in Figure 14.4b there is the obvious motion on p5 . In both cases there is a
strict proper equilibrium stress such that these non-trivial motion components only occur on vertices
adjacent to members with positive stresses. The pre-stress stability follows from this.
It can be convenient to translate infinitesimal rigidity of a tensegrity framework into a statement
about the underlying bar-joint framework as in the following theorem of Roth and Whiteley.
Theorem 14.3 (Roth and Whiteley, 1981 [49]) Let (G, p) be a tensegrity framework with at least
one cable or strut. Then (G, p) is infinitesimally rigid if and only if the underlying bar framework
(Ḡ, p) is infinitesimally rigid and there is a strict proper stress on (G, p).
A simple consequence of this theorem is that an infinitesimally rigid tensegrity framework (G, p)
is either a bar-joint framework or has at least one cable or strut and satisfies |E| > d|V | − d+1


2 .
(Contrast with the Maxwell counts, see Chapter 18.) A similar fact was observed by Roth and
Whiteley [49]: if (G, p) is an infinitesimally rigid tensegrity framework in Rd (with at least one cable
or strut) and (G0 , p) is obtained from (G, p) by deleting a single cable or strut and then replacing all
remaining cables and struts with bars, then (G0 , p) is infinitesimally rigid in Rd .
We can use equilibrium stresses to prove rigidity in the absence of infinitesimal rigidity. See also
[21] for more information.
The following is an equivalent dual statement of second-order rigidity. The idea is that the
framework has a demon who anticipates, for each possible non-trivial infinitesimal flex, a proper
blocking equilibrium stress that prevents the extension to a second order flex. This is opposed to
prestress stability, where one stress works for all non-trivial infinitesimal flexes.
304 Handbook of Geometric Constraint Systems Principles

Theorem 14.4 (Connelly and Whiteley, 1996 [26]) Let (G, p) be a tensegrity framework in Rd .
Suppose that for all nontrivial infinitesimal motions p0 of (G, p), there exists a proper equilibrium
stress ω such that X
ωi j (p0i − p0j ) · (p0i − p0j ) > 0.
ij

Then (G, p) is second order rigid in Rd .

14.2.1 Combinatorics of Tensegrities

It is well known that (infinitesimal) rigidity of tensegrity frameworks is not a generic property,
unlike bar-joint frameworks (see Chapter 18). In particular, as Figure 14.1 illustrates, there are
graphs and assignments of cables, struts and bars to the edges such that there are open sets of
realizations as a tensegrity framework which are rigid but also open sets of realizations such that
the tensegrity framework is not rigid [60]. It is an open problem to characterize those graphs which
have infinitesimally rigid realizations as tensegrity frameworks. See also [47].
Define a graph G to be strongly rigid if every generic realization of G in Rd as a tensegrity
framework is infinitesimally rigid. Strongly rigid examples include any tensegrity which contains a
spanning generic infinitesimally rigid bar-joint framework. See Figure 14.1 for an example which
is not strongly rigid. It was shown in [40] that testing whether a graph is strongly rigid is NP-hard
even in R1 . Recski and Shai [48] characterised infinitesimal rigidity for 1-dimensional tensegrities
and showed for such frameworks, infinitesimal rigidity can be decided in polynomial time.
In [41] it was shown that the edges of a graph can be assigned as cables and struts such that there
is some realization in Rd as a tensegrity framework which is infinitesimally rigid if and only if the
graph is redundantly rigid in Rd as a generic bar-joint framework (see Chapter 16 for information
about redundantly rigid graphs).

14.2.2 Geometric Interpretations

Dual to the infinitesimal theory for tensegrities is the theory of statics for tensegrity frameworks
which was developed by Whiteley [63] including results for placing tensegrities within faces of
convex polyhedra without breaking static rigidity.
Whiteley considered one story buildings with cables as “roof” edges [61] and also explored
links between tensegrities and scene analysis [62].
The projective invariance of infinitesimal rigidity for bar-joint frameworks extends to tensegrity
frameworks in the following way. Suppose a projective transformation sends a cable to infinity, then
we must replace it by a (finite) strut, and similarly a strut sent to infinity by the transformation must
be replaced by a (finite) cable [63].
While taking a projective viewpoint, we mention very recent results of Eftekhari [30] devel-
oping a tensegrity theory for spherical frameworks and point-hyperplane frameworks (the higher
dimensional analogue of point-line frameworks); see Chapters 22 and 17, Section 17.2.4.
Projective interpretations also lead to the problem of polarity for weaving lines [64]. Here
tensegrity appears unexpectedly. The lines correspond to points, the intersection of the lines cor-
respond to the edge constraints with a choice needed for the designation of whether the constraints
are cables or struts.

14.2.3 Packings
Packings of spherical disks in a polyhedral container can be regarded as tensegrities with all struts
connecting centers of touching disks to each other and the boundary of the container. With this
Tensegrity 305

viewpoint tensegrities have been used to prove results about packings, see, for example, [15, 19].
The local maximal density of the configuration of disks is determined by the rigidity and infinitesi-
mal rigidity of the underlying tensegrity framework. Examples of this are [4, 5, 33] and [20, 25] for
the periodic case (see Chapter 25 for various results on periodic rigidity for bar-joint frameworks).
A jammed packing thought of as a tensegrity can be detected, even for large sizes, using linear
programming [28] to verify rigidity.
The Kneser-Poulsen conjecture states that a re-arrangement of a configuration of spherical balls
in Rd in which the distance between every pair of centres (of the balls) does not decrease has the
property that the volume of the union of the balls does not decrease. This was proved in the case
when d = 2 by Bezdek and Connelly [6]. (See also [35].) One difficulty in the proof is that there
are configurations of centers that have another expansive arrangement in Rd but moving between
these arrangements requires using a path in R2d . In [6] it was shown that you can “leapfrog” one
arrangement to the other expansively using a path in only 2 extra dimensions, thus proving the con-
jecture in dimension 2. In [3], examples were analysed using second order rigidity and in particular
the “demon characterization.” This is a condition for a first order infinitesimal motion to extend to a
second order motion. In particular, suppose we have a first order motion p0 but the tensegrity frame-
work is not second order rigid. If there exists an equilibrium stress ω such that (p0 )T (Ω ⊗ Id )p0 6= 0
(where Ω is the stress matrix defined in the following section and the operator ⊗ denotes the tensor
product) then a brief calculation gives a contradiction. However it may be that there are different
equilibrium stresses for different possible motions.

14.3 Types of Rigidity

Energy function: The energy function associated to a stress ω for a tensegrity framework (G, p) is
defined as X
Eω (p) = ωi j kpi − p j k2 ,
1≤i< j≤n

where n = |V |.

Pmatrix Ω is the n × n symmetric matrix with off-diagonal entries −ωi j

Stress matrix: The stress
and diagonal entries j ωi j .

Globally rigid: A tensegrity framework (G, p) in Rd is globally rigid if every tensegrity framework
(G, q) in Rd , with the same edge labeling, dominated by (G, p) is congruent to (G, p).
Universally rigid: A tensegrity framework (G, p) in Rd is universally rigid if every tensegrity
framework (G, q) in RD for any D ≥ d, with the same edge labeling, dominated by (G, p) is
congruent to (G, p).
Super stable: A tensegrity framework (G, p) is super stable if it has a proper equilibrium stress ω
such that Ω is positive semi-definite, rank Ω = n−d −1 and the underlying bar-joint framework
(Ḡ, p) is rigid.
Dimensionally rigid: A tensegrity (G, p) in Rd is dimensionally rigid if any other framework
(G, q) in RD , for any D satisfying the edge constraints of (G, p) has an affine span of dimension
at most d.
Conic at infinity: A finite set of (non-zero) vectors in Rd lie on a conic at infinity if when regarded
as points in projective d − 1-dimensional space, they lie on a conic.
306 Handbook of Geometric Constraint Systems Principles

14.3.1 Global Rigidity and Stress Matrices

We now turn to global rigidity for tensegrities. Computationally, as with bar-joint frameworks, test-
ing global rigidity for tensegrities is NP-hard [50], essentially equivalent to the subset-sum problem.
Note that when (G, p) dominates (G, q) and ω is a proper equilibrium stress for (G, p) then
Eω (p) ≥ Eω (q) and when ω is strict and Eω (p) = Eω (q) then the kpi − p j k = kqi − q j k for all
i j ∈ E. P
Consider now the energy function Eω introduced in [12]. At a critical point we have j ωi j (p j −
pi ) = 0. This leads us to equilibrium stresses and the stress matrix. Specifically, by regarding Eω as
a quadratic form and considering the associated (symmetric) matrix we get the “big” stress matrix
Ω ⊗ Id . The fact that the stress matrix Ω is n × n thus arises from regarding the realization p (in the
tensegrity (G, p) in Rd ) as a column vector. Notice further that with
 
p1 p2 . . . pn
P=
1 1 ... 1

the critical point condition can be interpreted as the condition PΩ = 0. See also Chapter 16.
For example, recall the tensegrity in Figure 14.1. With the equilibrium stress as indicated we
have the stress matrix  
1 −1 −1 1
−1 1 1 −1
 ,
−1 1 1 −1
1 −1 −1 1
which has rank 1 = 4 − 3 = n − (d + 1).
The following theorem extends a result of Connelly showing that equivalent generic bar-joint
frameworks have the same equilibrium stresses, see Chapter 16, but it does use the additional as-
sumption of positive semidefiniteness.

Theorem 14.5 (Alfakih and Nguyen, 2013 [1]) Let (G, p) be a given tensegrity framework and let
Ω be a proper positive semidefinite stress matrix of (G, p). Then Ω is a proper stress matrix for all
tensegrity frameworks (G, q) dominated by (G, p).

Hendrickson’s necessary conditions for global rigidity [37] (see Chapter 16) also apply to
tensegrities. That is, if (G, p) is a generic globally rigid tensegrity framework in Rd then G is (d +1)-
connected and (G, p) is redundantly rigid in Rd . We also have the following sufficient condition.

Theorem 14.6 (Connelly, 2005 [14]) Let (G, p) be a generic tensegrity framework with a proper
equilibrium stress ω and stress matrix Ω of rank n − d − 1. Then (G, p) is globally rigid in Rd .

14.3.2 Universal and Dimensional Rigidity

Next we discuss universal rigidity. We have the following fundamental result showing how to use
stress matrices to confirm universal rigidity. This follows from the results of [12], see also [21].

Theorem 14.7 (Connelly, 2013 [17]) Let (G, p) be a tensegrity framework whose affine span of p
is all of Rd , with a proper equilibrium stress ω and stress matrix Ω. Suppose further that
(a) Ω is positive semi-definite;
(b) the rank of Ω is n − d − 1; and
(c) The member directions of (G, p) with a non-zero stress, and bars, do not lie on a conic at
infinity.
Tensegrity 307
p1

p3 p4

p2

Figure 14.5
A universally rigid tensegrity in the plane.

Then (G, p) is universally rigid.

Note that the third condition can be replaced by there being no non-trivial affine image of (G, p).
Consider the tensegrity in Figure 14.5. This tensegrity framework has a maximum rank positive
semi-definite stress with non-zero stresses only on the collinear triangle. Therefore the stress direc-
tions of the members with a non-zero stress do lie on a conic at infinity. However this tensegrity is
universally rigid.
Alfakih and Nguyen [1] used dimensional rigidity to help in understanding universal rigidity of
tensegrities. In particular they proved that the assumptions of the above theorem without the conic
condition give sufficient conditions for dimensional rigidity.

Theorem 14.8 (Alfakih and Nguyen, 2013 [1]) Let (G, p) be a tensegrity framework whose affine
span of p is all of Rd , with a proper equilibrium stress ω and stress matrix Ω. Suppose further that
(a) Ω is positive semi-definite and
(b) the rank of Ω is n − d − 1.
Then (G, p) is dimensionally rigid.

Theorem 14.9 (Connelly, 2009 [16]) Suppose a tensegrity framework (G, p) in Rd has a proper
equilibrium stress such that the underlying bar-joint framework (Ḡ, p) is super stable and infinites-
imally rigid. Then (G, q) is globally rigid for any generic q.

We also have the following counterpoint to the characterisation of global rigidity for bar-joint
frameworks in the plane, see Chapter 21.

Theorem 14.10 (Connelly, 2009 [16]) If G is 3-connected and generically redundantly rigid as a
bar-joint framework in R2 , then there is a realization (G, p) in R2 as a tensegrity framework which
is super stable.

In the next theorem we will need some definitions. An iterated affine set C = A0 ⊃ A1 ⊃ A2 ⊃
. . . Ak is a sequence of affine sets. For each Ai we take a basis matrix Bi and define a restricted stress
matrix Ω∗i = Bi−1 Ωi BTi−1 . In particular each Ω∗i is positive semi-definite.

Theorem 14.11 (Connelly and Gortler, 2015 [22]) Suppose C = A0 ⊃ A1 ⊃ A2 ⊃ . . . Ak is an

iterated affine set for a tensegrity (G, p) with n vertices in Rd , with an associated iterated proper
positive semidefinite stress described by positive semidefinite restricted stress matrices Ω∗i . Let ri
308 Handbook of Geometric Constraint Systems Principles
p1 p2

p5 p6

p3 p4

Figure 14.6
An example of a super stable tensegrity.

p2 p2

p4 p4
p5
p3 p1 p3 p1

Figure 14.7
A 1-extension on the edge wx adding new vertex u.

be the rank of Ω∗i . If ki=1 ri = n − d − 1 and the member directions with non-zero stress directions
P
and bars do not lie on a conic at infinity, then (G, p) is universally rigid.
Conversely if (G, p) is universally rigid in Rd , then there is an iterated affine set with an as-
sociated iterated positive semidefinite stress determined by proper stresses, the dimension of Ak is
(d + 1) d+1


2 , and the members with non-zero stress directions and bars do not lie on a conic at
infinity.

The theorem gives a certificate for universal rigidity.

Figure 14.6 gives an example of a super stable tensegrity. A second example comes from choos-
ing the vertices of a cube with cables for each edge of the cube and 4 struts connecting antipodal
vertices.

14.3.3 Operations on Tensegrities

We finish this section by commenting on methods for building larger tensegrities with certain rigid-
ity properties from smaller ones. First let us discuss the 1-extension operation (see Figure 14.7 and
Chapter 19, Section 19.2.1).
A simple three-node tensegrity is when pi , p0 and p j are ordered on a line so that i j is a strut
and both i0 and 0 j are cables. This tensegrity framework has an equilibrium stress where ωi j <
0, ωi0 > 0 and ω j0 > 0. The stress matrix for this tensegrity is easily seen to be positive semi-
definite with one positive eigenvalue and two zero eigenvalues. So in the 1-extension operation,
before the perturbation, one can choose the insertion of p0 and the two new edges as the addition of
a small stress matrix to the original stress matrix in such a way that the cable stress say on the edge
i j cancels with the strut stress on the three node tensegrity. The effect is that the new eigenvalue will
be positive, while the others will only be perturbed slightly. The same argument works when i j is a
strut and p0 is on the line through pi and p j , but not in the interval between them.
Tensegrity 309
p1 p1

p0

p3 p2 p3 p2

Figure 14.8
A delta-Y operation adding a new point p0 .

Second we comment on vertex splitting (again see Chapter 19, Section 19.2.1). Recall that a
vertex splitting operation on a vertex v, deletes v and the edges 2, 3 . . . ,t incident to v and adds two
vertices 0, 1 along with edges 20, 21, 30, 31, 01, i0 for 3 < i ≤ k and i1 for k < i ≤ t. The vertex split
is non-trivial for any t, k with k > 2 and t > k.

Theorem 14.12 (Connelly, 2009 [16]) Let (G, p) be a tensegrity framework which is super stable
in Rd . Let G0 be formed from G by a non-trivial vertex splitting operation which splits v into two
vertices v0 , v1 and let q be the realization of G in which q(x) = p(x) for all x ∈ V − v and q(v0 ) =
q(v1 ) = p(v). Suppose that (G0 − v0 v1 , q) is infinitesimally rigid in Rd . Then (G0 , p̂) is super stable
in Rd for any generic p̂.

The theorem also holds if we replace super stable by globally rigid since the proof keeps track
of the rank of the stress matrix independently of the stress matrix being positive semi-definite.
It is also worth pointing out that the technique of coning (see Chapter 17, Section 17.2.2) extends
to tensegrities by simply choosing all edges to the cone vertex as bars.
Next, let us discuss methods for combining globally rigid tensegrity frameworks. For rigid-
ity we note that the 2-sum operation (see Chapter 19) preserves generic infinitesimal rigidity for
tensegrities, see [41]. To move to global rigidity, it is easy to see that adding two positive semi-
definite matrices results in a positive semi-definite matrix. Moreover it follows from the definition
that adding two generic globally rigid tensegrity frameworks in Rd with at least d + 1 vertices in
common preserves generic global rigidity. In [9] it is shown that in this process, for bar-joint frame-
works, a single common edge can also be removed. However, extending this to tensegrities is an
open problem.
Finally, recall that the delta-Y operation on a graph G removes a triangle of edges 12, 23, 31
and adds a new vertex 0 and 3 new edges 01, 02, 03, see Figure 14.8. Applying this operation to
tensegrities is discussed in [17].

14.4 Examples and Applications

Tensegrity polygon: A tensegrity polygon is a planar tensegrity framework in which the cables
form a convex polygon containing all the vertices and the set of bars is empty.
Abstract tensegrity polygon: An abstract tensegrity polygon is a graph G = (V, E) where E is
partitioned into two sets E− and E+ and the edges in E− form a Hamilton cycle of G.
Convex realization: A convex realization of an abstract tensegrity polygon is a planar realization
in which the Hamilton cycle of cables forms a convex polygon.
310 Handbook of Geometric Constraint Systems Principles
p1 p2 p1 p2

p5 p6 p5 p6

p3 p4 p3 p4

Figure 14.9
A Cauchy polygon and a Grunbaum polygon.

Strong: An abstract tensegrity polygon G is strong if every convex realization (G, p) has a non-
trivial proper stress.
Robust: An abstract tensegrity polygon G is robust if every convex realization (G, p) has a strict
proper stress.
Stable: An abstract tensegrity polygon G is stable if every convex realization (G, p) is infinitesi-
mally rigid.
Spider web: A spider web is a tensegrity framework were some subset of the vertices are pinned
and all edges are cables.

14.4.1 Examples
We now describe examples of super stable and thus universally rigid tensegrities.
Suppose we have an abstract tensegrity polygon in which the Hamilton cycle vertices are la-
belled 1, 2, . . . , n. A Cauchy polygon is an abstract tensegrity polygon in which the struts are of the
form i(i + 2) for 1 ≤ i ≤ n − 2. A Grunbaum polygon is an abstract tensegrity polygon in which the
struts are 13 and 2i for 4 ≤ i ≤ n. Figure 14.9 gives examples of Cauchy and Grunbaum polygons,
respectively. Note also that Figure 14.1 is the smallest Cauchy and Grunbaum polygon.
Roth and Whiteley [49] proved that Generalized Grunbaum polygons are robust and stable.
Geleji and Jordán [34] proved that, for an abstract tensegrity polygon G, the properties strong,
robust and stable are equivalent.
Theorem 14.13 (Connelly, 1982 [12]) Suppose a tensegrity framework (G, p) in R2 consists of a
convex polygon, with cables on the boundary and struts inside. If there is a nonzero stress then the
stress matrix has rank n − 3 and is positive semidefinite (and hence universally rigid).
The special class of tensegrity frameworks where every edge is a cable, known as spiderwebs,
admits a more fully understood theory. In particular Connelly [12] proved the following theorem.
Theorem 14.14 Let (G, p) be a pinned spider web framework in Rd , where G is connected, with a
strict equilibrium stress. Then (G, p) is globally rigid in Rd .
Observe that this theorem can be converted to a statement about the global rigidity of unpinned
tensegrities consisting of a spider web plus a complete graph of struts on the previously pinned
vertices. See, for example, Figure 14.10.
A number of further classes of super stable tensegrities are discussed in [16] such as tensegrities
with a triangle of struts and then only cables inside with emphasis on the potential to use inductive
constructions to show that all graphs in the class that are also 3-connected and redundantly rigid
have super stable realizations.
References 311

p2
p2

p4 p5 p4 p5
p6 p6
p3 p1 p3 p1

Figure 14.10
A spiderweb with a complete graph of struts around the outside and the pinned version of this
spiderweb.

14.4.2 Applications
The chapter [13] discusses the use of global rigidity for tensegrities in one proof of Cauchy’s arm
lemma, which is of course used in the proof of Cauchy’s famous theorem on convex polyhedra.
See also Chapter 13. In particular the idea is to use a Cauchy polygon with struts “representing”
angles and prove the universal rigidity of the Cauchy polygon using induction and Theorem 14.7.
The problem is then translated back to Cauchy’s arm problem by replacing the cables by bars and
noticing that the increases in length of (pairs of incident) struts correspond to increases in the size
of the corresponding angles. Further applications of tensegrities to distance geometry problems can
be found in [2, 24, 54].
A wealth of further information on tensegrities can be found in the literature. For complete
bipartite examples, see [23]; for highly symmetric tensegrities see the catalogue [11] (also [10]); for
dihedrally symmetric examples including Snelson’s example, see [67].
Pak and Vilenchik used Theorem 14.7 to show particular instances of the graph realization
problem are tractable and deduced from their work a method of constructing uniquely k-colorable
graphs [46]. Laurent and Varvitsiotis [44] studied tensegrity frameworks in the context of universal
completability and semidefinite programming. Similarly Tanigawa studied spherical tensegrities
[58] in the context of semidefinite programming and matrix completion.
Szabadka [57] conjectured that there exists an integer k such that if (G, p) is a generic tensegrity
framework with at least k cables and at least k struts, G is 3-connected and the underlying bar-joint
framework (Ḡ, p) is redundantly rigid, then (G, p) is rigid in R2 . Szabadka also proved several
results about tensegrities relating to 1-extensions, and to complete graphs and wheels.
We finish by mentioning that the Carpenter’s rule problem [18, 55] is an example of a funda-
mental problem in discrete geometry which can be interpreted as a tensegrity problem. Here the
opening of a polygonal chain can be seen to be an ‘expansive motion’ where distances only increase
so can be replaced by struts. Indeed any context with expansive motions fits this philosophy (e.g.,
[56]).

References
[1] A. Y. Alfakih and Viet-Hang Nguyen. On affine motions and universal rigidity of tensegrity
frameworks. Linear Algebra Appl., 439(10):3134–3147, 2013.
312 References

[2] Maria Belk. Realizability of graphs in three dimensions. Discrete Comput. Geom., 37(2):139–
162, 2007.
[3] Maria Belk and Robert Connelly. Making contractions continuous: a
problem related to the kneser-poulsen conjecture. Technical report,
http://inside.bard.edu/academic/programs/math/mbelk/Contractions.pdf, 2007.
[4] A. Bezdek, K. Bezdek, and R. Connelly. Finite and uniform stability of sphere coverings.
Discrete Comput. Geom., 13(3-4):313–319, 1995.
[5] A. Bezdek, K. Bezdek, and R. Connelly. Finite and uniform stability of sphere packings.
Discrete Comput. Geom., 20(1):111–130, 1998.
[6] Károly Bezdek and Robert Connelly. Pushing disks apart—the Kneser-Poulsen conjecture in
the plane. J. Reine Angew. Math., 553:221–236, 2002.
[7] C. R. Calladine. Buckminster fuller’s “trensegrity” structures and clerk maxwell’s rules for the
construction of stiff frameworks. International Journal of Solids and Structures, 14:161–172,
1978.
[8] C. R. Calladine and S. Pellegrino. First-order infinitesimal mechanisms. Internat. J. Solids
Structures, 27(4):505–515, 1991.
[9] R. Connelly. Combining globally rigid frameworks. Tr. Mat. Inst. Steklova, 275(Klassich-
eskaya i Sovremennaya Matematika v Pole Deyatelnosti Borisa Nikolaevicha Delone):202–
209, 2011.
[10] R. Connelly and M. Terrell. Tenségrités symétriques globalement rigides. Structural Topol-
ogy, (21):59–78, 1995. Dual French-English text.
[11] Robert Connelly. Highly symmetric tensegrity structures. http://www.math.cornell.edu/ tens/.
[12] Robert Connelly. Rigidity and energy. Invent. Math., 66(1):11–33, 1982.
[13] Robert Connelly. Rigidity. In Handbook of convex geometry, Vol. A, B, pages 223–271.
North-Holland, Amsterdam, 1993.
[14] Robert Connelly. Generic global rigidity. Discrete Comput. Geom., 33(4):549–563, 2005.
[15] Robert Connelly. Rigidity of packings. European J. Combin., 29(8):1862–1871, 2008.
[16] Robert Connelly. Questions, conjectures and remarks on globally rigid tensegrities.
http://www.math.cornell.edu/ connelly/09-Thoughts.pdf, 2009.
[17] Robert Connelly. Tensegrities and global rigidity. In Shaping space, pages 267–278. Springer,
New York, 2013.
[18] Robert Connelly, Erik D. Demaine, and Günter Rote. Straightening polygonal arcs and con-
vexifying polygonal cycles. Discrete Comput. Geom., 30(2):205–239, 2003. U.S.-Hungarian
Workshops on Discrete Geometry and Convexity (Budapest, 1999/Auburn, AL, 2000).
[19] Robert Connelly and William Dickinson. Periodic planar disc packings. Philos. Trans. R.
Soc. Lond. Ser. A Math. Phys. Eng. Sci., 372(2008):20120039, 17, 2014.
[20] Robert Connelly, Matthew Funkhouser, Vivian Kuperberg, and Evan Solomonides. Packings
of equal disks in a square torus. arXiv:1512.08762.
[21] Robert Connelly and Steven Gortler. Prestress stability of triangulated convex polytopes and
universal second order rigidity. arXiv:1510.04185.
[22] Robert Connelly and Steven Gortler. Iterative universal rigidity. Discrete Comput. Geom.,
53(4):847–877, 2015.
References 313

[23] Robert Connelly and Steven Gortler. Universal rigidity of complete bipartite graphs. arXiv:
1502.02278, 2016.
[24] Robert Connelly and Jean-Marc Schlenker. On the infinitesimal rigidity of weakly convex
polyhedra. European J. Combin., 31(4):1080–1090, 2010.
[25] Robert Connelly, Jeffrey D. Shen, and Alexander D. Smith. Ball packings with periodic
constraints. Discrete Comput. Geom., 52(4):754–779, 2014.
[26] Robert Connelly and Walter Whiteley. Second-order rigidity and prestress stability for tenseg-
rity frameworks. SIAM J. Discrete Math., 9(3):453–491, 1996.
[27] Miguel de Guzmán and David Orden. From graphs to tensegrity structures: geometric and
symbolic approaches. Publ. Mat., 50(2):279–299, 2006.
[28] Aleksandar Donev, Salvatore Torquato, Frank H. Stillinger, and Robert Connelly. A lin-
ear programming algorithm to test for jamming in hard-sphere packings. J. Comput. Phys.,
197(1):139–166, 2004.
[29] Amy C. Edmondson. A Fuller explanation. Design Science Collection. Birkhäuser Boston,
Inc., Boston, MA, 1987. The synergetic geometry of R. Buckminster Fuller, A Pro Scientia
Viva Title.
[30] Yaser Eftekhari. Geometry of point-hyperplane and spherical frameworks. PhD thesis, York
University, 2017.
[31] Richard Buckminster Fuller. Synergetics 2: Further Explorations in the Geometry of Thinking,
Volume 2. Macmillan, 1983.
[32] Richard Buckminster Fuller and E. J. Applewhite. Synergetics: Explorations in the geometry
of thinking. Macmillan, 1982.
[33] Zsolt Gáspár, Tibor Tarnai, and Krisztián Hincz. Partial covering of a circle by equal circles.
Part II: the case of 5 circles. J. Comput. Geom., 5(1):126–149, 2014.
[34] János Geleji and Tibor Jordán. Robust tensegrity polygons. Discrete Comput. Geom.,
50(3):537–551, 2013.
[35] Igors Gorbovickis. Strict Kneser-Poulsen conjecture for large radii. Geom. Dedicata, 162:95–
107, 2013.
[36] S. D. Guest. The stiffness of tensegrity structures. IMA J. Appl. Math., 76(1):57–66, 2011.
[37] Bruce Hendrickson. Conditions for unique graph realizations. SIAM J. Comput., 21(1):65–84,
1992.
[38] D.E. Ingber. Cellular tensegrity: defining new rules of biological design that govern the cy-
toskeleton. J. Cell Sci., 104:613–627, 1993.
[39] Donald E. Ingber, Ning Wang, and Dimitrije Stamenović. Tensegrity, cellular biophysics, and
the mechanics of living systems. Rep. Progr. Phys., 77(4):046603, 21, 2014.
[40] Bill Jackson, Tibor Jordán, and Csaba Király. Strongly rigid tensegrity graphs on the line.
Discrete Appl. Math., 161(7-8):1147–1149, 2013.
[41] Tibor Jordán, András Recski, and Zoltán Szabadka. Rigid tensegrity labelings of graphs.
European J. Combin., 30(8):1887–1895, 2009.
[42] Yoshihiro Kanno. Exploring new tensegrity structures via mixed integer programming. Struct.
Multidiscip. Optim., 48(1):95–114, 2013.
314 References

[43] E. N. Kuznetsov. On immobile kinematic chains and a fallacious matrix analysis. Trans.
ASME J. Appl. Mech., 56(1):222–224, 1989.
[44] M. Laurent and A. Varvitsiotis. Positive semidefinite matrix completion, universal rigidity
and the strong Arnold property. Linear Algebra Appl., 452:292–317, 2014.
[45] Benjamin Nabet and Naomi Ehrich Leonard. Shape control of a multi-agent system using
tensegrity structures. In Lagrangian and Hamiltonian methods for nonlinear control 2006,
volume 366 of Lect. Notes Control Inf. Sci., pages 329–339. Springer, Berlin, 2007.
[46] Igor Pak and Dan Vilenchik. Constructing uniquely realizable graphs. Discrete Comput.
Geom., 50(4):1051–1071, 2013.
[47] András Recski. Combinatorial conditions for the rigidity of tensegrity frameworks. In Hori-
zons of combinatorics, volume 17 of Bolyai Soc. Math. Stud., pages 163–177. Springer, Berlin,
2008.
[48] András Recski and Offer Shai. Tensegrity frameworks in one-dimensional space. European
J. Combin., 31(4):1072–1079, 2010.
[49] B. Roth and W. Whiteley. Tensegrity frameworks. Trans. Amer. Math. Soc., 265(2):419–446,
1981.
[50] J. B. Saxe. Embeddability of weighted graphs in k-space is strongly np-hard. In Proceedings
of the 17th Allerton Conference in Communications, Control and Computing, pages 480–489,
1979.
[51] M. Schenk, S. D. Guest, and J. L. Herder. Zero stiffness tensegrity structures. Internat. J.
Solids Structures, 44(20):6569–6583, 2007.
[52] Meera Sitharam and Mavis Agbandje-Mckenna. Modeling virus self-assembly pathways:
avoiding dynamics using geometric constraint decomposition. J. Comput. Biol., 13(6):1232–
1265, 2006.
[53] R.T. Skelton and C. Sultan. Controllable tensegrity: a new class of smart structures. Proc.
SPIE, 3039:166–177, 1997.
[54] Anthony Man-Cho So and Yinyu Ye. A semidefinite programming approach to tensegrity
theory and realizability of graphs. In Proceedings of the Seventeenth Annual ACM-SIAM
Symposium on Discrete Algorithms, pages 766–775. ACM, New York, 2006.
[55] Ileana Streinu. Pseudo-triangulations, rigidity and motion planning. Discrete Comput. Geom.,
34(4):587–635, 2005.
[56] Ileana Streinu and Walter Whiteley. Single-vertex origami and spherical expansive motions.
In Discrete and computational geometry, volume 3742 of Lecture Notes in Comput. Sci.,
pages 161–173. Springer, Berlin, 2005.
[57] Zoltan Szabadka. Globally rigid frameworks and rigid tensegrity graphs in the plane. PhD
thesis, Eotvos Lorand University, 2010.
[58] Shin-Ichi Tanigawa. The signed positive semidefinite matrix completion problem for odd-k4
minor free signed graphs. arXiv: 1603.08370, 2016.
[59] G. Tibert. Deployable tensegrity structures for space applications. PhD thesis, Royal institute
of technology, Stockholm, 2002.
[60] Neil L. White and Walter Whiteley. The algebraic geometry of stresses in frameworks. SIAM
J. Algebraic Discrete Methods, 4(4):481–511, 1983.
References 315

[61] Walter Whiteley. Cones, infinity and 1-story buildings. Structural Topology, (8):53–70, 1983.
With a French translation.
[62] Walter Whiteley. A correspondence between scene analysis and motions of frameworks. Dis-
crete Appl. Math., 9(3):269–295, 1984.
[63] Walter Whiteley. Infinitesimally rigid polyhedra. I. Statics of frameworks. Trans. Amer. Math.
Soc., 285(2):431–465, 1984.
[64] Walter Whiteley. Rigidity and polarity. II. Weaving lines and tensegrity frameworks. Geom.
Dedicata, 30(3):255–279, 1989.
[65] Darrell Williamson, Robert E. Skelton, and Jeongheon Han. Equilibrium conditions of a
tensegrity structure. Internat. J. Solids Structures, 40(23):6347–6367, 2003.
[66] J. Y. Zhang and M. Ohsaki. Self-equilibrium and stability of regular truncated tetrahedral
tensegrity structures. J. Mech. Phys. Solids, 60(10):1757–1770, 2012.
[67] J.Y. Zhang, S.D. Guest, R. Connelly, and M. Ohsaki. Dihedral ‘star’ tensegrity structures.
International Journal of Solids and Structures, 47(1):1 – 9, 2010.
Chapter 15
Geometric Conditions of Rigidity in
Nongeneric Settings
Oleg Karpenkov
University of Liverpool

CONTENTS
15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
15.2 Configuration Space of Tensegrities and its Stratification . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
15.2.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
15.2.2 Definition of a Tensegrity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
15.2.3 Stratification of the Space of Tensegrities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
15.2.4 Tensegrities on 4 Points in the Plane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
15.3 Extended Cayley Algebra and the Corresponding Geometric Relations . . . . . . . . . . . . . . 322
15.3.1 Extended Cayley Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
15.3.2 Geometric Relations on Configuration Spaces of Points and Lines . . . . . . . . . . 324
15.4 Geometric Conditions of Infinitesimal Flexibility in Terms of Extended Cayley
Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
15.4.1 Examples in the Plane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
15.4.2 Frameworks in General Position . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
15.4.3 Non-parallelizable Tensegrities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
15.4.4 Geometric Conditions for Existence Non-parallelizable Tensegrities . . . . . . . . 327
15.4.5 Conjecture on Strong Geometric Conditions for Tensegrities . . . . . . . . . . . . . . . 327
15.5 Surgeries on Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
15.6 Algorithm to Write Geometric Conditions of Realizability of Generic Tensegrities . . 330
15.6.1 Framed Cycles in General Gosition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
15.6.1.1 Basic Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
15.6.1.2 Geometric Conditions for Framed Cycles . . . . . . . . . . . . . . . . . . . . 331
15.6.1.3 Geometric Conditions for Trivalent Graphs . . . . . . . . . . . . . . . . . . 332
15.6.2 Resolution Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
15.6.2.1 Definition of Resolution Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
15.6.2.2 Resolution of a Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
15.6.2.3 HΦ-Surgeries on Completely Generic Resolution Schemes . . . 333
15.6.3 Construction of Framing for Pairs of Leaves in Completely Generic
Resolution Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
15.6.4 Framed Cycles Associated to Generic Resolutions of a Graph . . . . . . . . . . . . . 335
15.6.5 Natural Correspondences Between ΞG (P) and the Set of all Resolutions for
G(P) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
15.6.6 Techniques to Construct Geometric Conditions Defining Tensegrities . . . . . . 336
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337

317
318 Handbook of Geometric Constraint Systems Principles

15.1 Introduction
In this chapter we discuss geometric approach to infinitesimal rigidity in non-generic settings. Cur-
rently this approach is developed for the case of the plane and there is almost nothing known for
three- and higher-dimensional spaces. Recall that a framework (i.e., a realization of a graph in the
plane, or in the space) is infinitesimally rigid if every infinitesimal isometric deformation of the
framework is an infinitesimal isometric deformation for the corresponding complete graph on the
vertices. One of the fundamental characterizations of infinitesimal rigidity is via the linear indepen-
dence of the rows of the rigidity matrix. Namely the configuration is infinitesimally rigid if and only
if the corresponding rigidity matrix is of full rank.
Since the rank of the rigidity matrix is defined via algebraic equations (by the determinants of
a given rank submatrices of the rigidity matrix), a rigidity matrix of a generic graph realization in
the plane is always either full rank or not full rank. In this chapter we study the case of graphs when
generic realizations of a given graph have a full rank rigidity matrix. In this situation infinitesimal
flexibility is achieved only for specific frameworks, forming a positive codimension orbifold. As we
see, such graphs are infinitesimally flexible only in nongeneric settings. A geometric description of
such setting (conditions) is the main goal of this chapter.
First algebraic characterization of nongeneric settings were developed in 1983 by N. L. White
and W. Whiteley in [31]. The authors proposed the techniques of tie-downs, where one supplements
the rigidity matrix by additional lines up to a square matrix. Then the determinant of the obtained
matrix gives an equation for infinitesimal flexibility configurations. Finally, one should rewrite it
in terms of bracket ring and factorize (this is important for removing extra lines used to write the
determinant). Several of the obtained factors would be independent on the choice of the supplemen-
tary lines. The locus of these factors is precisely the set of all infinitesimally flexible configurations.
Here we should notice that the factorization of such bracket expressions is a hard open problem
even in the two-dimensional case.
In this chapter we consider an alternative geometric approach where the conditions of infinites-
imal flexibility are written in terms of meet and join operations and relations of extended Cayley
algebra (see in [16]). It is based on the study of tensegrities, where the characterization of nongeneral
settings is given by simple geometric conditions on the vertices of frameworks.
To familiarize the reader with the problem of geometric conditions of infinitesimal rigid-
ity/flexibility in nongeneric settings we discuss one simple example. Recall that the bracket for
three points in the plane p1 , p2 , p3 is the following expression

[p1 , p2 , p3 ] = det(p2 − p1 , p3 − p1 ),

where det(v, w) is the determinant of a matrix whose columns are v and w respectively. Here the
bracket is considered as a degree 2 polynomial on 6 coordinates of the points involved.

Example 41 Consider a tensegrities on 6 distinct points with the graph G as on Figure 15.1 on the
left. What are all 6-point configurations P = (p1 , . . . , p6 ) in the plane whose frameworks G(P) are
not infinitesimally rigid? In geometric terms we have the following conditions (later we describe
such geometric conditions in terms of extended Cayley algebra):

• The lines p1 p2 , p3 p4 , p5 p6 have a common point or parallel to each other (see Figure 15.1
at the middle);
• the points p1 , p4 , p5 are in a line;
• the points p2 , p3 , p6 are in a line.
Geometric Conditions of Rigidity in Nongeneric Settings 319

p4 p4

p3 p3
4 3
p5 p6 p6 O
5 6 p5
p2 p2
1 2
G p1 p1

Figure 15.1
The graph G (on the left); one of its realization with nonzero equilibrium stresses for G (in the
middle); a corresponding infinitesimally isometric flexion (on the right).

Each of these conditions define a subset (strata) on the configuration space of all 6-point con-
figurations. The corresponding bracket ring expression for this graph is:


[p4 , p1 , p2 ][p3 , p6 , p5 ] − [p3 , p1 , p2 ][p4 , p6 , p5 ] [p1 , p4 , p5 ][p2 , p3 , p6 ] = 0.

Here we have three factors: the first, the second, and the third factors correspond to the first, the
second, and the third items, respectively.

Remark 15.1 Notice that the direction of infinitesimal flexion for the above example is very easy
to see. Fix one of the triangles (say p2 p6 p3 ) and infinitesimally isometrically rotate another triangle
p1 p4 p5 about the intersection O (see Figure 15.1 on the right).
In fact if one of the triangles is obtained by a translation from another we have a finite flexion.

Further organization of this chapter. In Section 15.2 we recall the definition of tensegrity, de-
fine the configuration space of tensegrities and discuss its stratification; as an example we study
here a simple case of tensegrities on 4 points in the plane. Further in Section 15.3 we study the
extended Cayley algebra and introduce the corresponding geometric relations on the elements of
the configuration spaces of lines and points in the plane. After that we study geometric conditions
of infinitesimal rigidity in terms of extended Cayley algebra in Section 15.4. We continue in Sec-
tion 15.5 with an introduction of the techniques of local surgeries on graphs that preserve geometric
conditions. Finally in Section 15.6 we conclude with the algorithmic questions in the subject.

15.2 Configuration Space of Tensegrities and its Stratification

Let us start with some definitions and background. Within this chapter we will employ the tech-
niques of tensegrities that is very useful in the study of infinitesimal rigidity. Nonzero tensegrities
and infinitesimal flexions are closely related. In fact, the following statement holds:
A realization of the graph is infinitesimally flexible if and only if it admits a non-zero tensegrity.
One of the benefits of tensegrities is an additional structure of a force-load for the edges of the
realizations. The force-loads are extremely helpful for understanding geometric conditions describ-
ing infinitesimally flexible configurations in nongeneric settings.
320 Handbook of Geometric Constraint Systems Principles

15.2.1 Background
The story of tensegrities goes back to J.C. Maxwell who studied equilibrium states for frames under
the action of static forces, see [20]. In the second half of the twentieth century tensegrities reap-
peared in the art, K. Snelson constructed several surprising cable-bar sculptures that are actually
tensegrities [26]. Recently TensegriTree was erected (see [7]) to celebrate the 50th anniversary of
the University of Kent (see [21] for an overview and history of tensegrity constructions). The term
“tensegrity” (a combination of words “tension” and “integrity”) was introduced by R. Buckminster
after erection of first such constructions. There are various applications of tensegrities in different
branches of science. In particular they were used in the study of viruses [1, 24], cells [12, 13], and
deployable mechanisms [25, 28].
In mathematics, tensegrities were studied in different contexts: rigidity and flexibility of tenseg-
rities [2, 3, 22, 34]; minimal rigidity [29]; mechanical properties [14]. Classical tensegrities were
extended to spherical and projective geometry [23]; to other non-Euclidean geometries and special
normed spaces [15, 19]; to surfaces in R3 [15]. See also a general introductory paper on tensegrities
in mathematics by R. Connelly [4].
In [31, 32] N.L. White and W. Whiteley studied the existence of non-zero tensegrities for a
given n-tuple of points. They have introduced algebraic conditions for the existence of nontrivial
tensegrities (see also [33]). Later in the papers [8, 9, 10] M. de Guzmán and D. Orden introduce atom
decomposition techniques and describe several conditions for some further examples of graphs.
Finally in [16] the existence of non-zero tensegrities was described in terms of extended Cayley
algebra.

15.2.2 Definition of a Tensegrity

Let us bring together main notions and definitions in theory of tensegrities.

Definition 15.1 Fix a positive integer d. Let G = (V, E) be an arbitrary graph without loops and
multiple edges. Let it have n vertices.
• A framework G(P) in Rd is a map of the graph G with vertices v1 , . . . , vn on a finite point
configuration P = (p1 , . . . , pn ) in Rd with straight edges, such that G(P)(vi ) = pi for i =
1, . . . , n.
• A stress w on a framework is an assignment of real scalars wi, j (called tensions) to its edges
pi p j . We also put wi, j = 0 if there is no edge between the corresponding vertices. Observe
that wi, j = w j,i , since they refer to the same edge.
• A stress w is called a self-stress if, in addition, the following equilibrium condition is ful-
filled at every vertex pi : X
wi, j pi p j = 0.
{ j| j6=i}

By pi p j we denote the vector from the point pi to the point p j .

• A pair (G(P), w) is called a tensegrity if w is a self-stress for the framework G(P).
• We say that F is a force-load for a tensegrity (G(P), w) if for every edge pi p j there exists a
force Fi, j acting along this edge such that

Fi, j = wi, j pi p j .
Geometric Conditions of Rigidity in Nongeneric Settings 321

15.2.3 Stratification of the Space of Tensegrities

Let us consider the configuration space of all ordered n-tuples of points in Rd , i.e., (Rd )n . Let also
G be some graph on n vertices. Denote by TensG (P) the space of all tensegrities on G(P). Note that
TensG (P) is a linear space whose dimension coincide with codimension of the rigidity matrix.

Definition 15.2
• A maximal connected component of (Rd )n whose all points have the same dimension
TensG (∗) is called a stratum for the graph G.
• A disjoint decomposition of (Rd )n into strata with respect to G is called the stratification of
(Rd )n related to G
• The universal tensegrity stratification of (Rd )n is the intersection of all the stratifications for
all possible graphs G on n vertices. The strata of universal tensegrity stratification are called
the universal tensegrity strata.

It is known that all the strata are semialgebraic sets [5].

15.2.4 Tensegrities on 4 Points in the Plane

In this subsection we consider the universal stratification of the space of all tensegrities on 4 points
in the plane. For a generic 4-tuple of points P we have: no three points in a line, no two points
coincide. A generic P admits the unique up to a scalar multiplication tensegrity for a complete graph
K4 and no tensegrities for all other graphs on 4 points. There are exactly 14 connected components
of generic points.
The universal strata of codimension 1 correspond to the configurations P where three out of four
vertices of the graph lie in a line. For such configurations there are some K3 ⊂ K4 graphs that admits
a nonzero tensegrity. The number of such strata is 24.
Combinatorial adjacency structure of the full dimension and codimension 1 universal strata is
shown on Figure 15.2. Each oval corresponds to the union of 6 universal strata of codimension 1
having the same triples of points in a line. The ovals divide the plane into 14 connected component
representing universal strata of full dimension. Finally the large dots represent the universal strata
of higher codimension (which is rather non-trivial to show on the picture).
All the universal strata of codimension greater than 1 are the intersections of the closures of the
codimension 1 universal strata. The table of all these strata follows.

Stratum description codim quantity

Four points in a line 2 12
Two points coincide 2 12
Four points in a line, two of which coincide 3 18
Three points coincide 4 4
Two pairs of points coincide 4 3
All points coincide 6 1

For a more detailed description of the configuration spaces we refer to papers [5]. The case
of planar tensegrities on 5 points is exhaustively studied in [17]. It has the following amount of
universal strata. The universal stratum of codimension 8 is the stratum corresponding to the case
when all the points coincide, it is of dimension 2.
322 Handbook of Geometric Constraint Systems Principles

Figure 15.2
Adgacency of codimension 1 and full dimension strara of tensegrities on 4 points in the plane.

Codimension of a stratum 0 1 2 3 4 5 6 7 8
Number of strata 264 600 810 300 170 75 15 0 1

15.3 Extended Cayley Algebra and the Corresponding Geometric

Relations
If the number of points is greater than 5 then a more interesting universal strata appear. They are
defined by geometric conditions of the extended Cayley algebra which we study in this section.
The elements of the extended Cayley algebra are points, lines and a special element which we
denote as “true.” The element true is similar to element 1 in the algebra of logic. It is introduced in
order not to consider multiple cases when certain non-generality of the configuration occur. Usually
true means that either certain lines or points coincide. Whenever the element true is involved in an
expression on Cayley algebra, this expression is true itself.

15.3.1 Extended Cayley Algebra

For simplicity we work in the projectivization of the Euclidean plane, i.e., in real projective plane
RP2 . First of all we define four elementary operations on points and lines in RP2 . These operations
have much in common with join and meet operations of Cayley algebra (i.e., ∨ and ∧). For more
information on Cayley algebras we refer to [6], [30], and [11]. Note that it is much more convenient
to work in the projective settings since every two lines here intersect.
Geometric Conditions of Rigidity in Nongeneric Settings 323

Operation I (2-point operation). Denote the first operation by (∗, ∗). This operation is a binary
operation defined on the set of all points in the projective plane and the additional element true, as
follows.
(∗, ∗) p1 p2 (6= p1 ) true
p1 true p1 ∨ p2 true
p2 (6= p1 ) p1 ∨ p2 true true
true true true true

Here p1 and p2 are arbitrary distinct points and p1 ∨ p2 denotes the line through these points.
In case if there is no confusion we write p1 p2 instead of (p1 , p2 ).
Operation II (2-line operation). Similarly we define the binary operation ∗ ∩ ∗ on the the set of all
lines in the projective plane and the additional element true.

∗∩∗ `1 `2 (6= `1 ) true

`1 true `1 ∧ `2 true
`2 (6= `1 ) `1 ∧ `2 true true
true true true true

Here `1 and `2 are arbitrary distinct lines and `1 ∨ `2 denotes the intersection point of these lines.
Choice operations. Further we need two choice operations:

Operation III Operation IV

Point choice Line choice

Pick a point on a given line avoiding a Pick a line through a given point avoid-
given discrete subset of points ing a given discrete subset of lines

Let us study the following curious example.

Remark 15.2 Consider the following 4-tiple of points of projective plane:

p1 = [0 : 0 : 1], p2 = [1 : 0 : 1], p3 = [0 : 1 : 1], p4 = [1 : 1 : 1]

and let p = [a : b : c]. What can we obtain starting with the points p1 , p2 , p3 , and p4 applying
consequently Operations I and II? The answer to this question is very surprising. One can construct
precisely all points of p ∈ QP2 (i.e., p = (a, b, c) is proportional to a triple of rational numbers).

Relations. Finally we define several basic relations for points and lines in RP2 which will be used
to generate equations in extended Cayley algebra:

Relation Notation It is fulfilled in the following cases

1) One of the entries is true;
3-point (p1 , p2 , p3 ) = true 2) At least two points coincide;
3) p1 , p2 , p3 are in a line.
1) One of the entries is true;
point-line p ∈ ` = true 2) p is contained in `.
1) One of the entries is true;
3-line `1 ∩ `2 ∩ `3 = true 2) at least two lines coincide;
3) `1 , `2 , `3 are concurrent.
324 Handbook of Geometric Constraint Systems Principles

We conclude this subsection with an important remark.

Remark 15.3 In fact every extended Cayley algebra expression is well defined for points and
lines of the Euclidean plane (which is considered as an affine plane in the projective plane). For
that reason we will use extended Cayley algebra in Euclidean settings as well. While operating with
points and lines of the Euclidean plane, it is useful to keep in mind that the resulting object can be
the line at infinity or a point at the line at infinity.

15.3.2 Geometric Relations on Configuration Spaces of Points and Lines

Let us consider geometric relations for special configuration spaces of lines. For a fix n-point con-
figuration P consider the configuration space of (non-fixed) lines passing through prescribed points,
namely  
(`1 , . . . , `m )`i passes through p j(i) , i = 1, . . . , m}
We denote it by ΞP (R),
 where R is the list of inclusion conditions defining the configuration space
(i.e., R = p j(i) ∈ `i i = 1, . . . , m ).

Remark 15.4 One may think of lines `1 , . . . , `m to be variables of equations while points
p1 , . . . , pn to be parameters on which equations depend.

Let us give the following general definitions.

Definition 15.3 Consider ΞP (R) as above.

• A geometric condition on ΞP (R) is a composition of several geometric operations and one
geometric relation on points (p1 , . . . , pn ) and lines (`1 , . . . , `m ) of ΞP (R).
• We say that a system of geometric conditions on ΞP (R) is fulfilled at P if there exists a choice
of m lines satisfying all the conditions R such that every geometric condition is “true” for
this choice of lines.
• Two systems of geometric conditions on n-tuples of points and m-tuples of lines satisfying
conditions R are equivalent if for every configuration P these systems are simultaneously
either fulfilled or not fulfilled for ΞP (R) at P.

Let us consider the following simple example.

Example 42 For the configuration space
Ξ(p1 ,...,p6 ) (p5 ∈ `1 )
we consider the following system of geometric conditions:

p1 p4 ∩ p2 p3 ∩ ` = true
.
p6 ∈ ` = true
Then a generic configuration of 6 points does not fulfill this system, while the following configuration
does.
p2 p4

p5 p6 ‘

p1 p3
Geometric Conditions of Rigidity in Nongeneric Settings 325

15.4 Geometric Conditions of Infinitesimal Flexibility in Terms of

Extended Cayley Algebra
Recall that the geometric conditions of infinitesimal flexibility coincide with geometric conditions
on realizability of non-zero tensegrities for the corresponding frameworks.
In Subsection 15.4.1 we study examples of such conditions for graphs of 6 and 7 vertices.
Further in Subsections 15.4.2 and 15.4.3 we discuss the generality assumptions for frameworks and
tensegrities respectively. In Subsection 15.4.4 we formulate the theorem which states that conditions
on realizability of a generic (i.e., non-parallelizable) tensegrity for a given graph or framework are
expressed in terms of extended Cayley algebra. This theorem is one of the central statements of this
chapter. Finally in Subsection 15.4.5 we discuss a potential way to improve this theorem.

15.4.1 Examples in the Plane

In all the examples of this subsection the list of non-fixed lines is empty. So we have simply ΞP ().
There are two interesting examples on 6 vertices, they are listed below.

Graph (6 vert.) Sufficient geometric conditions

p4 p3
p5 p6
p1 p2 ∩ p3 p4 ∩ p5 p6 = true
p1 p2
p4 p3

p1 p2 ∩ p4 p5 , p2 p3 ∩ p5 p6 , p3 p4 ∩ p6 p1 = true
p5 p6
(Equivalently: the six points p1 , p2 , p3 , p4 , p5 , and
p1 p2 v6 are on a conic)

In the right column of the above table we write a sufficient condition for a 6-tuple of points to
admit a non-zero tensegrity (or, equivalently, to have a non-trivial infinitesimal isometric flexion)
for the corresponding graph.
We have the following collection of graphs on 7 vertices together with geometric conditions for
them.
Graph (7 vert.) Sufficient geometric conditions
p5 p3 p4

p6 p7
(p1 , p2 , p3 ) = true
p1 p2
p4 p3
p5 p6
p7 p1 p2 ∩ p3 p4 ∩ p5 p6 = true
p1 p2
p4 p3
p5 p7
p6 p1 p2 ∩ p3 p4 ∩ (p5 , p2 p6 ∩ p3 p7 ) = true
p1 p2
p4 p3
p5 p7
p6 p1 p2 ∩ p4 p5 , p2
p3 ∩ (p5 , p1 p6 ∩ p3 p7 ), p3 p4 ∩
p1 p2 (p1 , p1 p6 ∩ p3 p7 ) = true
326 Handbook of Geometric Constraint Systems Principles

15.4.2 Frameworks in General Position

Recall that a cycle is a graph homeomorphic to the circle. A simple cycle in a graph G is a subgraph
of G homeomorphic to the circle.

Definition 15.4 Let us consider the following notions of general position.

• An n-tuple of lines in the Euclidean plane are said to be in general position if the following
two conditions hold:
— no three lines meet in a point;
— no three lines are parallel.
• Let C(P) be a realization of a cycle C in the plane, where P = (p1 , . . . , pn ). We say that C(P)
is in general position if the lines passing through the edges of C(P) are in general position.
• A framework G(P) in the plane is said to be in general position if every simple cycle of at
most n − 1 vertex is in general position (recall that G has n vertices).

In particular if a cycle is in general position, then all its edges are of nonzero length.

15.4.3 Non-parallelizable Tensegrities

When study conditions of realisability we will assume some natural general position assumptions
for the force-loads of tensegrities. These assumptions are collected in the following definition.

Definition 15.5 Let G be a graph on n vertices, (G(P), w) be a tensegrity and F be an equi-

librium force load for the tensegrity (G(P), w). We say that the equilibrium force-load F is non-
parallelizable at vertex p, if the following two conditions are fulfilled.
Suppose that the forces of F at all edges adjacent to p are F1 , . . . , Fs .
• Let εi ∈ {0, 1} for i = 1, . . . , s. Then
s
X
εi Fi = 0 if and only if ε1 = . . . = εs .
i=1

• All the following forces

s
X
F1 + εi Fi , where (ε2 , . . . εs ) ∈ {0, 1}s \ {(1, . . . , 1)}
i=2

are nonproportional to each other. (Recall that here we consider precisely 2s−1 − 1 forces.)
We sat that a tensegrity (G(P), F) is non-parallelizable if it is non-parallelizable at every its vertex.

Remark 15.5 If the first condition of Definition 15.5 is not fulfilled at p, then one of the following
surgeries on the graph can be done:
— either some edge can be deleted
— or p can be splitted in two vertices (each edge is adjacent either to one copy of p or to
another).
The resulting framework admits tensegrity with the same non-zero forces along the corresponding
edges. So in some sense this tensegrity is realizable for a framework of a “simpler” graph (i.e., such
graph has a smaller first Betti number).
Geometric Conditions of Rigidity in Nongeneric Settings 327

15.4.4 Geometric Conditions for Existence Non-parallelizable Tensegrities

In this subsection we briefly describe the objects that are involved in the geometric conditions for
non-parallelizable tensegrities. The actual algorithm to write them would be given in the last section
of this chapter.
Consider an arbitrary graph G on n vertices, and let G(P) be a framework with distinct points
P = (p1 , . . . , pn ). Consider the following arrangement of lines: At each point pi ∈ P we choose
deg pi − 3 ordered lines passing through pi . Denote by ΞG (P) the configuration space of all such
arrangements.

Remark 15.6 We have

n
X

dim(ΞG (P)) = deg(pi ) − 3 ,
i=1

In particular, if all vertices are of degree 3, then the configuration space ΞG (P) is empty.

We use the configuration space ΞG (P) for the detection non-parallelizable tensegrities. Let us
formulate one of the central theorems of this chapter.

Theorem 15.1 A framework G(P) in general position admits a non-parallelizable tensegrity if and
only if this framework satisfies a certain system of geometric conditions on points of P and lines of
ΞG (P) for all simple cycles of G.

Remark 15.7 The system of geometric conditions is explicitly described by the algorithm of
Subsection 15.6.6, see also Theorem 15.3.

The proof of this theorem is rather technical, we refer an interested reader to [16].

15.4.5 Conjecture on Strong Geometric Conditions for Tensegrities

As we have seen in the examples of the tables above, all the geometric conditions are written entirely
in terms of the vertices of the framework P, none of the lines of ΞG (P) for them are involved. One
might expect that this situation is general for planar tensegrities. Let us briefly discuss this here.

Definition 15.6 Let G be a graph and let G(P) be one of the frameworks for G. We say that
a geometric condition on points P is a strong geometric condition for G(P) if it does not involve
choice operations (i.e., operations III and IV).

In many cases geometric conditions on points and lines passing through them are equivalent to
certain strong geometric conditions on points only. However this is not always the case, we illustrate
this with the following example.

Example 43 In this example we deal with the configuration space

Ξ(p1 ,p2 ,p3 ,p4 ,p5 ,p6 ) p1 ∈ `1 , p2 ∈ `2 , p3 ∈ `3

and the following system of geometric conditions for it:


 `1 ∩ `2 ∩ p4 p5 = true
`2 ∩ `3 ∩ p5 p6 = true
`3 ∩ `1 ∩ p6 p4 = true


Below is the example of a 9-point and 3-line configuration satisfying the above system of condition.
328 Handbook of Geometric Constraint Systems Principles

p5

‘2
p2
p6
p1 p3
‘3
‘1
p4

From the one hand, this system is not equivalent to any system of strong geometric conditions on P.
From the other hand this system is not related to any graph G.
In the view of the last example it would be interesting to check if the following statement holds.
Conjecture ([16]). For every graph G there exists a system of strong geometric conditions such that
a framework G(P) in general position admits a non-parallelizable tensegrity if and only if P satisfies
this system of strong geometric conditions.
In other words, this conjecture implies that all the non-parallelizable tensegrities are described
in terms of Cayley algebra operations on the vertices of frameworks.
We would like to conclude this section with the following interesting question.
Problem. Develop a similar geometric approach to the study of infinitesimal flexibility in three- and
higher-dimensional cases.

15.5 Surgeries on Graphs

Surgeries on graphs provide a techniques to obtain geometric conditions on graphs while knowing
geometric conditions on some another graphs. Here we change the structure of the graph locally
leaving most of the vertices and edges of the original graph unchanged. The more surgeries one
knows, the smaller the set of initial graphs one needs to study in order to develop a list of geometric
conditions for all graphs. This techniques is rather successful in the planar case and rather unstudied
in higher dimensional cases. Below we describe most of the currently known graph surgeries.
In the diagrams we show a part of a graph. Black dots indicate vertices that might have some
edges that we do not see on the corresponding diagram (i.e., the edges connecting such vertices with
the vertices that are not in this part of the graph). The edges of white dots are precisely the edges
shown in the diagram. Finally we would like to mention that surgeries nay be applied only under
certain genericity conditions. Such conditions are indicated in the last column.
Basic surgeries. We start with the simplest possible type of surgeries. These surgeries remove the
points of degree 1 and 2. Here is the complete list of them.
Source Target Conditions
p1 p2 p1 p1 6= p2 .
p1 p2 p3 p1 p3 The points p1 , p2 and p3
are not in a line.
p1 p2 p3 p1 p3 The points p1 , p2 and p3
are in a line and distinct
to each other.
Geometric Conditions of Rigidity in Nongeneric Settings 329

Dimension 1 subgraph surgeries. The second class of surgeries is rather wide. First of all we give
the following definition.

Definition 15.7 We say that a triple (G, e1 , e2 ) where G is a graph and e1 and e2 are its edges
fulfills the condition hG, e1 , e2 i at a framework (G, P) if the following two conditions hold.
— The graph G has a unique (up to a scalar) non-zero tensegrity.
— The stresses at edges e1 and e2 for non-zero tensegrities on (G, P) are non-zero.

Here we have the following general statement.

Proposition 15.1
If a (G, e1 , e2 ) fulfills the condition hG, e1 , e2 i then the tensegrities for the graph G ∪ e1 admits
non-zero tensegrities at P if and only if the graph G ∪ e2 admits non-zero tensegrities at P.

The last proposition give rise to the following list of surgeries.

Source Target Conditions

p1 p4 p1 p4
All the triples of points are not in a line.

p2 p3 p2 p3
p1 p4 p1 p4 The triples of points are not in a line:
p5 (p4 , p1 , p5 ) and (pi , pi+1 , p5 ) where i =
p5 1, 2, 3, and the points (p1 , p2 , p3 , p4 ) are
p2 p3 p2 p3 not in one line.
... ... ...
H \ e1 H \ e2 There exists G ⊂ H with e1 , e2 ∈ G satis-
fying Consdition hG, e1 , e2 i.

One should be careful while using these surgeries, since the condition hG, e1 , e2 i might already
contain nontrivial configuration with non-zero tensegrities. So while removing such cases one might
remove possible realizations.
Degree 3 vertex surgeries. The next class of surgeries plays an important role in the planar case.

Source Target Relations Conditions

p1 p1 the triples
of points are
q2 q3 r r=p2 q2 ∩ p3 q3 not in a line:
p2 p3 p2 p3 (p1 , q2 , q3 ),
(p2 , p3 , qi ),
and (p1 , pi , qi ),
where i = 2, 3.
p1 p4
p1 p4 the triples
q1 r1 of points are
r1 =p1 q1 ∩ p4 q2
q2 r2 r2 =p2 q1 ∩ p3 q2 not in a line:
p2 p3 (p1 , p2 , q1 ),
p2 p3 (p3 , p4 , q2 ),
and (pi , q1 , q2 )
for i=1, . . . , 4.
330 Handbook of Geometric Constraint Systems Principles

The last surgery is called an HΦ-surgery. It is essentially used in the study of planar tensegrities,
see in [16].
Remark on strong geometric conditions for graphs on small number of vertices. Using these surg-
eries one might find strong geometric conditions of infinitesimal isometric flexibility to all graphs
having 9 or less vertices (see in [16]). The complete list of codimension 1 graphs for 8 or less ver-
tices can be found in [5]. The first example of a graph which we unable to reduce to 9-point graphs
using the above surgeries is as follows:

This graph is a good candidate as a counterexample to the conjecture of Subsection 15.4.5.

15.6 Algorithm to Write Geometric Conditions of Realizability of

Generic Tensegrities
In this section we show how to construct geometric conditions for cycles mentioned in Theo-
rem 15.1.
We start this section with the study of framed cycles, here for each framed cycle in general
position we introduce a certain geometric condition related to it. This already allows us to define
geometric conditions for trivalent graphs (see Subsection 15.6.1). For the graphs having vertices of
degree greater than 3 we introduce resolution schemes for such vertices (see Subsection 15.6.2). Fur-
ther in Subsections 15.6.3 and 15.6.4 using resolution schemes we construct the framings for simple
cycles of a general graph G, which provides us with desired geometric conditions for G. Finally we
summarize the construction techniques of geometric conditions detecting non-zero tensegrities in
Subsection 15.6.5 (see also Theorem 15.3).

15.6.1 Framed Cycles in General Gosition

15.6.1.1 Basic Definitions
Let us start with the following general definition.

Definition 15.8 Let P = (p1 , . . . , pk ).

• A realization of the cycle C(P) in the Euclidean plane has a framing if every vertex pi is
equipped with a line `i passing through it.
• The realization C(P) together with its framing is called the framed cycle. Denote it by


C(P, L) = (p1 , . . . , pk ), (`1 , . . . , `k ) .

We use the following notion of genericity for framed cycles.

Geometric Conditions of Rigidity in Nongeneric Settings 331

p4
p3 ‘3

p02
p2 p1 ‘2
‘02

Figure 15.3
A projection operation ω2 . On the left we have: p02 = p1 p2 ∩ p3 p4 ; on the right: `02 = (p02 , `2 ∩ `3 ).

Definition 15.9 A framed cycle C(P, L) is in general position if

— the cycle C(P) is in general position;
— for every admissible i we have: the line `i does not contain the points pi−1 and pi+1 .

Let C(P, L) be a cycle with k ≥ 4 vertices. for an arbitrary i ∈ {1, 2, . . . , k} we set

p0i = pi−1 pi ∩ pi+1 pi+2 ;
`0i = p0i (`i ∩ `i+1 ).
(Here we set p0 = pk , pk+1 = p1 , and pk+2 = p2 .)
The following surgery on the framed cycle C(P, L) is called a projection operation of a cycle:
ωi (C(P, L)) = C(P0 , L0 ),
where
P0 = (p1 , p2 , . . . , pi−1 , p0i , pi+2 , . . . , pk ), and
(15.1)
L0 = (`1 , `2 , . . . , `i−1 , `0i , `i+2 , . . . , `k ).
In other words the projection operation ωi (C(P, L)) removes the edge pi pi+1 by prolonging the
edges pi−1 pi and pi+1 pi+2 towards the intersection point of their lines (i.e., p0i ) and introduce a
framing `0i to p0i . See an example of ω2 (C(P, L)) on Figure 15.3.

Remark 15.8 It is clear that the projection operation is entirely expressed by the elementary
operations (Operations II and I).

15.6.1.2 Geometric Conditions for Framed Cycles

Consider a framed cycle

C(P, L) = C (p1 , . . . , pk ), (`1 , . . . , `k )
in general position. Then the composition of any admissible k − 3 projection operations applied to
C(P, L) will result in a framed triangular cycle in general position. Denote the resulting cycle by
C(P̂, L̂) = C(( p̂1 , p̂2 , p̂3 ), (`ˆ1 , `ˆ2 , `ˆ3 )).

Definition 15.10 Let C(P, L) and C(P̂, L̂) be as above. Then the condition
`ˆ1 ∩ `ˆ2 ∩ `ˆ3 = true (15.2)
is the geometric condition defined by C.

Note the following.

332 Handbook of Geometric Constraint Systems Principles

• The geometric condition for C(P, L) does not depend on the choice of projection operations.
The resulting geometric relations are equivalent.
• The geometric condition for C(P, L) is a combination of 9k − 9 Operations I; 6k − 6 Op-
erations II; and one 3-line relation on points P and lines L. (Hence it is a true geometric
condition of extended Cayley algebra).

For simplicity one might always fix the following composition of projection operators:

ω1 ◦ . . . ◦ ω1 (C(P, L)).
| {z }
k − 3 times
The expression in terms of Operations I, II and one 3-line relation are written directly from expres-
sions (15.1).

Example 44 For the cycles on 3, 4, and 5 vertices we have the following geometric relations:

k=3: `1 ∩ `2 ∩ `3 = true;
k=4: (`1 ∩ `4 , `2 ∩ `3 , p1 p2 ∩ p3 p4 ) = true;
k=5: (`2 ∩ `3 , p1 p2 ∩ p3 p4 ) ∩ `1 ∩ (`4 ∩ `5 , p1 p3 ∩ p3 p4 ) = true.

15.6.1.3 Geometric Conditions for Trivalent Graphs

In the case of trivalent graphs we are in position write down all the geometric conditions for cycles
mentioned in Theorem 15.1.
Let G be a trivalent graph and G(P) be a framework in general position. Consider a cycle C in
G and and set the natural framing for C(P(C)) (here P(C) ⊂ P) as follows. Let the vertex pi ∈ C by
adjacent to the edges pi pi,1 , pi pi,2 , and pi pi,3 of the framework G(P). Without loss of generality we
assume that the edges pi pi,1 , pi pi,2 are edges of C(P) at vertex pi . Then we set `i (C) = pi , pi,3 . Set

L(C) = (`1 (C), . . . , `k (C)).

The geometric condition for the cycle C ∈ G in Theorem 15.1 is precisely the geometric condi-
tion (15.2) of Definition 15.10 for C(P(C), L(C)). Now Theorem 15.1 for 3-valent graphs can be
reformulated as follows.

Theorem 15.2 Let G be a trivalent graph. A framework G(P) in general position admits a non-
parallelizable tensegrity if and only if every simple (framed) cycle C(P(C), L(C)) of G satisfies the
geometric condition (15.2) for C(P(C), L(C)).

15.6.2 Resolution Schemes

Suppose now G has vertices of degree greater than 3, so we cannot write geometric conditions for
cycles using Theorem 15.2. The main idea here is to consider resolutions at vertices of degree k > 3
replacing them by unrooted full binary trees with k leaves. Then one can define a similar geometric
conditions for the resulting graph. We show the main steps to do this below.

15.6.2.1 Definition of Resolution Schemes

Let us study how to replace a vertex of the framework by a unrooted full binary tree. Recall that an
unrooted full binary tree is a tree without the root where the degree of every vertex of T is either 1
or 3. Recall also that an edge of a tree is a leaf if one of its vertices is of degree 1. All other edges
are interior edges of a tree.
Geometric Conditions of Rigidity in Nongeneric Settings 333

Denote by Gr(1, RP2 ) the Grassmannian of 2-dimensional planes in R3 (i.e., Gr(1, RP2 ) is the
set of all lines in the projective plane). Here as usual when we work in affine settings we consider
an affine Grassmannian as a subset of the projective Grassmannian.

Definition 15.11 Consider an unrooted full binary tree T and let

L : E(T ) → Gr(1, RP2 ).

We say that a pair (T, L) is a resolution scheme at point p in the Euclidean plane if for every edge
e ∈ T it holds p ∈ L(e). Denote it by (T, L) p .

15.6.2.2 Resolution of a Framework

In what follows we restrict ourselves to graphs whose vertices are all of degree 3 or greater.

Definition 15.12 Let G be a graph on n vertices and let G(P) be its framework on P = (p1 , . . . , pn ).
We say that the collection
(G(P), ((T1 , L1 ) p1 , . . . , (Tn , Ln ) pn ))
is a resolution of G(P) if for every i we have:
• the resolution scheme (Ti , Li ) pi has exactly deg pi leaves.
• the edges of G adjacent to pi are enumerated by the leaves (Ti , Li ) pi (i.e., the one-to-one
correspondence between the adjacent edges and the leaves is fixed).

• let v be a leaf at of Ti corresponding to an edge pi p j then

Li (v) = (pi , p j ).

We denote it by G(P)TL .

15.6.2.3 HΦ-Surgeries on Completely Generic Resolution Schemes

In this section we describe a surgery for a certain class of generic resolution schemes. This surgery
is similar to HΦ-surgery (the second degree 3 vertex surgery shown on page 329) for graphs.
Let (T, L) p be a resolution scheme. Consider the tree T 0 obtained from T by the following flip
operation:
v4 v6
v4 v6
v2 v20
v1 v10
v3 v5
v3 v5
Here white vertices are interior while black vertices can be both interior or leaves.
Let us construct the line ` which we further associate to v01 v02 . This is done geometrically using
the following sequence of Operations I–IV.

• Operation IV: Pick a point p∞ 6= p ∈ L(v1 v2 );

• Operation III: Pick a line `∞ 6= L(v1 v2 ) through p∞ ;
334 Handbook of Geometric Constraint Systems Principles

L 0(v10v30)
L(v2v6)
L(v1v4)

p L(v1v2)

‘00 ‘0
p00 p0
^
‘

L(v2v5)
p000 L(v1v3)

Figure 15.4
Geometric construction of ` = L0 (v01 v02 ).

• Operation III: Pick a point p0 ∈

/ {p, p∞ };
• Operations I and II: Define `ˆ = p(L(v1 v2 ) ∩ `∞ );
• Operations I: p00 = `ˆ ∩ L(v2 v5 );

• Operations I and II: `0 = p0 (L(v1 v4 ) ∩ `∞ );

• Operations I and II: `00 = p00 (L(v2 v6 ) ∩ `∞ );
• Operations I: p000 = `0 ∩ `00 ;
• Operations II: ` = pp000 .
See Figure 15.4.

Remark 15.9 In this construction we recommend to consider the line `∞ to the the line at infinity
and p∞ be one of its points. As we have already mentioned in Remark 15.3 the lines and points
involved in our construction may be at infinity.

Definition 15.13 An HΦ-surgery on (T, L) p at the interior edge v1 v2 is the operation that replaces
(T, L) with (T 0 , L0 ) where L0 is defined as follows:

L0 (v01 v02 ) = `; L0 (v01 v3 ) = L(v1 , v3 );

L0 (v01 v5 ) = L(v2 , v5 ); L0 (v02 v4 ) = L(v1 , v4 );
L0 (v02 v6 ) = L(v2 , v6 ); L0 (e) = L(e) for any other edge e.

In fact, the resulting resolution scheme is not always well-defined due to some non-genericity
phenomena (e.g., when the lines v1 v3 and v2 v5 coincide). So the following definition is actual here.

Definition 15.14 We say that a resolution scheme (T, L) is completely generic if every composi-
tion of HΦ-surgeries is well-defined.
Geometric Conditions of Rigidity in Nongeneric Settings 335

For a geometric description of completely generic resolution trees we refer to [16].

Remark 15.10 Let (T, L) p be a completely generic resolution scheme (where T has n leaves).
Then applying all possible different compositions of HΦ-surgeries one gets precisely (2n − 1)!!
distinct resolutions scheme, which is the number of the all unrooted binary full trees with n marked
leaves (e.g., see ex. 5.2.6 in [27]). In fact, these schemes are in a natural one-to-one correspondence
with the set of all unrooted binary full trees with n marked leaves ((T, L) → T ).

Definition 15.15 We say that a resolution of a graph is generic if every its resolution scheme is
completely generic.

15.6.3 Construction of Framing for Pairs of Leaves in Completely Generic Resolu-

tion Schemes
Assume that we are given by a completely generic resolution scheme (T, L) p . Let us define a special
line for every pair of leaves u, v.

Definition 15.16 Let (T, L) be a completely generic resolution scheme. Consider a composition
φ of HΦ-surgeries and a resolution scheme (T 0 , L0 ) p such that the following two conditions hold
• φ (T, L) p = (T 0 , L0 ) p ;

• the leaves u and v are adjacent edges in the tree T 0 (note that any HΦ-surgery does not affect
leaves, so u and v are leaves for both trees T and T 0 ).
Assume that w is the third edge of T 0 adjacent to the common point of the leaves u and v. Set

` p (u, v) = L0 (w).

The following two statements clarify the correctness of Definition 15.16.

Proposition 15.2
For a completely generic resolution scheme (T, L) we have:

• there exists a pair φ , (T 0 , L0 ) p satisfying both conditions of Definition 15.16.

• the line ` p (v, w) does not depend on the choice of φ and (T 0 , L0 ) p .

We will skip the
proof of the second statement and provide the algorithm to construct a certain
pair φ , (T 0 , L0 ) p below.

Remark 15.11 (Construction of ` p (v, w).) Let v1 . . . vs be a simple path connecting the leaf u =
v1 v2 where L(v1 v2 ) = pi p j and the leaf v = vs−1 vs with framing to L(vs−1 vs ) = pi pk . Then we
consequently apply s−3 HΦ-surgeries along the edges v2 v3 , v3 v4 , . . . , vs−2 vs−1 . As a result we have
a resolution scheme (Ti0 , L0 ) pi whose leaves L0−1 (ei j ) and L0−1 (eik ) share a common vertex.

15.6.4 Framed Cycles Associated to Generic Resolutions of a Graph

First, we define the framing for two adjacent edges of frameworks.
336 Handbook of Geometric Constraint Systems Principles

Definition 15.17 Consider a generic resolution G(P)TL of a framework G(P). Let pi p j and pi pk
be two edges in G(P) with a common vertex pi and let v and w be the associated leaves in the
resolution scheme (Ti , Li ) pi . Then the line ` pi (v, w) introduced in Definition 15.16 is the associated
framing for the pair of edges (ei j , eik ) at pi . We denote it by ` j,i,k .

The above definition leads to the natural notion of framed cycles associated to generic resolu-
tions of a graph.

Definition 15.18 Consider a generic resolution G(P)TL of a framework G(P). Let C = p1 . . . ps be

a cycle in G(P). Denote by C(G, T , P, L) the framed cycle with vertices p1 . . . ps such that

`i−1,i,i+1 is a framing at qi (for i = 1, . . . , s).

We say that this cycle is a framed cycle associated to G(P)TL .

Definition 15.19 Any line ` j,i,k in the framing associated to a G(P)TL is explicitly expressed in
terms of Operations I–IV on the lines of L(E(GT )) (by Definition 15.13 and Remark 15.11). Let
us fix one of the possible composition of Operations I–IV defining the framing ` j,i,k and call it the
sequence of geometric operations defining ` j,i,k .

15.6.5 Natural Correspondences Between ΞG (P) and the Set of all Resolutions for
G(P)
Given a framework G(P) and the corresponding configuration space ΞG (P). Let us fix the following
data and notation:

• Fix a resolution tree Ti at each vertex pi and denote T = (T1 , . . . , Tn ).

• Denote by G(P)T∗ the configuration space of all resolutions for the framework G(P) whose
resolution schemes at pi has a tree Ti for i = 1 . . . , n.
• Enumerate all interior edges for every tree Ti .
• Enumerate all the lines of ΞG (P) passing through every point pi .

Once the above is done, we have a natural isomorphism between G(P)T∗ and ΞG (P). Here the
line ` j at point pi of the configuration in ΞG (P) corresponds to the line Li (v j ) of the j-th interior
edge of the resolution scheme (Ti , Li ) pi .

15.6.6 Techniques to Construct Geometric Conditions Defining Tensegrities

Finally let us show step by step how to write down the system of geometric conditions for the
existence of non-parallelizable tensegrities.
Input Data. We start with a framework G(P) in general position.
Step 1. Fix T = (T1 , . . . , Tn ) in resolution schemes at all vertices and associate the configuration
space ΞG (P) with G(P)T∗ (see Subsection 15.6.5).
Step 2. Pick all simple cycles C1 , . . . ,CN in G that does not pass through all the points of G.
Step 3. Write all lines ` j,i,k in terms of compositions of Operations I–IV on the points of P and the
lines corresponding to the interior edges of G(P)T∗ . (See Definition 15.19.)
Step 4. Define framed cycles Ci (G, T , P, L) related to Ci for i = 1, . . . , N. Here we use the lines
obtained in Step 3 for framings.
References 337

Step 5. Write down geometric conditions for Ci (G, T , P, L) for i = 1, . . . , N in terms of lines `i, j,k .
(See the construction of Section 15.6.1.2.)
Step 6. Combining together Step 3 and Step 5 we write down geometric conditions for framed cycles
Ci (G, T , P, L) for i = 1, . . . , N in terms of P and the lines of ΞG (P) (which is isomorphic to G(P)T∗ ,
see Section 15.6.5).
Output data. As an output we get the system of geometric conditions on the space ΞG . By Theo-
rem 15.1 this system is fulfilled if and only if there exists a non-parallelizable tensegrity at ΞG .

Theorem 15.3 The above algorithm produces geometric conditions for Theorem 15.1.

Remark 15.12 In fact, at Step 2 it is sufficient to pick only the simple cycles generating H1 (G).
In practice it is sufficient to choose even less cycles to get the corresponding geometric existence
condition of a non-parallelizable tensegrity.

For further details and justification of the above algorithm we refer to [16].

References
[1] D.L.D. Caspar and A. Klug, Physical principles in the construction of regular viruses, in Pro-
ceedings of Cold Spring Harbor Symposium on Quantitative Biology, vol. 27 (1962), pp. 1–24.
[2] R. Connelly and W. Whiteley, Second-order rigidity and prestress stability for tensegrity
frameworks, SIAM Journal of Discrete Mathematics, vol. 9, no. 3 (1996), pp. 453–491.
[3] R. Connelly, Tensegrities and global rigidity, Shaping space, Springer, New York (2013),
pp. 267-278.
[4] R. Connelly, What is . . . a tensegrity?, Notices Amer. Math. Soc., vol. 60, no. 1 (2013), pp. 78-
80.
[5] F. Doray, O. Karpenkov, J. Schepers, Geometry of configuration spaces of tensegrities, Disc.
Comp. Geom., vol. 43, no. 2 (2010), pp. 436–466.
[6] P. Doubilet, G.-C. Rota, J. Stein, On the foundations of combinatorial theory. IX. Combinato-
rial methods in invariant theory, Studies in Appl. Math., vol. 53 (1974), pp. 185-216.
[7] D. Gray, http://expedition.uk.com/projects/tensegritree-university-of-kent/
[8] M. de Guzmán, Finding Tensegrity Forms, preprint (2004).
[9] M. de Guzmán, D. Orden, Finding tensegrity structures: Geometric and symbolic aproaches,
in Proceedings of EACA-2004, pp. 167–172 (2004).
[10] M. de Guzmán, D. Orden, From graphs to tensegrity structures: Geometric and symbolic
approaches, Publ. Mat. 50 (2006), pp. 279–299.
[11] H. Li, Invariant algebras and geometric reasoning. With a foreword by David Hestenes.
World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2008, xiv+518 pp.
[12] D. E. Ingber, Cellular tensegrity: defining new rules of biological design that govern the cy-
toskeleton, Journal of Cell Science, vol. 104 (1993), pp. 613–627.
[13] D.E. Ingber, N. Wang, D. Stamenović, Tensegrity, cellular biophysics, and the mechanics of
living systems, Rep. Progr. Phys., vol. 77, no. 4 (2014), 046603, 21 p.
338 References

[14] B. Jackson, T. Jordán, B. Servatius, H. Servatius, Henneberg moves on mechanisms, Beitr.

Algebra Geom., vol. 56, no. 2 (2015), pp. 587-591.
[15] B. Jackson, A. Nixon, Stress matrices and global rigidity of frameworks on surfaces, Discrete
Comput. Geom. vol. 54, no. 3 (2015), pp. 586-609.
[16] O. Karpenkov, The combinatorial geometry of stresses in frameworks, preprint (2015),
arXiv:1512.02563 [math.MG].
[17] O. Karpenkov, J. Schepers, B. Servatius, On stratifications for planar tensegrities with a small
number of vertices, ARS Mathematica Contemporanea, vol. 6, no. 2 (2013), pp. 305–322.
[18] D. Kitson, S.C. Power, Infinitesimal rigidity for non-Euclidean bar-joint frameworks, Bull.
Lond. Math. Soc., vol. 46, no. 4, (2014), pp. 685-697.
[19] D. Kitson, B. Schulze Maxwell-Laman counts for bar-joint frameworks in normed spaces,
Linear Algebra and its Applications vol. 481, (2015), pp. 313–329.
[20] J. C. Maxwell, On reciprocal figures and diagrams of forces, Philos. Mag. vol. 4, no. 27
(1864), pp. 250–261.
[21] R. Motro, Tensegrity: Structural systems for the future, Kogan Page Science, London, 2003.
[22] B. Roth, W. Whiteley, Tensegrity frameworks, Trans. Amer. Math. Soc., vol. 265, no. 2 (1981),
pp. 419–446.
[23] F.V. Saliola, W. Whiteley, Some notes on the equivalence of first-order rigidity in various
geometries, arXiv:0709.3354 [math.MG], 15 p.
[24] C. Simona-Mariana, B. Gabriela-Catalina, Tensegrity applied to modelling the motion of
viruses, Acta Mech. Sin. vol. 27, no. 1 (2011), pp. 125-129.
[25] R.E. Skelton, Deployable tendon-controlled structure, United States Patent 5642590, July 1,
1997.
[26] K. Snelson, http://www.kennethsnelson.net.
[27] R.P. Stanley, Enumerative combinatorics. Vol. 2. With a foreword by Gian-Carlo Rota and
appendix 1 by Sergey Fomin. Cambridge Studies in Advanced Mathematics, 62. Cambridge
University Press, Cambridge, 1999, xii+581 pp.
[28] A. G. Tibert. Deployable tensegrity structures for space applications, Ph.D. Thesis, Royal
Institute of Technology, Stokholm 2002.
[29] M. Wang, M. Sitharam, Combinatorial Rigidity and Independence of Generalized Pinned
Subspace-Incidence Constraint Systems, in Automated Deduction in Geometry–10th Inter-
national Workshop, Coimbra, Portugal (2014), pp. 166–180, arXiv:1503.01837 [cs.CG].
[30] N.L. White, T. McMillan, Cayley factorization. Symbolic and algebraic computation, Rome,
(1988), pp. 521-533, Lecture Notes in Comput. Sci., 358, Springer, Berlin, 1989.
[31] N. L. White, W. Whiteley, The algebraic geometry of stresses in frameworks, SIAM J. Alg.
Disc. Meth., vol. 4, n. 4 (1983), pp. 481–511.
[32] N. L. White, W. Whiteley, The algebraic geometry of motions of bar-and-body frameworks,
SIAM J. Algebraic Discrete Methods, vol. 8, no. 1 (1987), pp. 1-32.
[33] N. White, The bracket ring of a combinatorial geometry, I, Trans. Amer. Math. Soc., vol. 202
(1975), pp. 79–95.
References 339

[34] W. Whiteley, Rigidity and scene analysis, in J.E. Goodman and J. O’Rourke, editors, Hand-
book of Discrete and Computational Geometry, chapt. 49, pp. 893–916, CRC Press, New
York, 1997.
Chapter 16
Generic Global Rigidity in General
Dimension
Steven J. Gortler
[email protected], SEAS, Harvard University

CONTENTS
16.1 Basic Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
16.2 Connelly’s Sufficiency Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
16.3 Hendrickson’s Necessary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
16.3.1 Nonsufficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
16.4 Necessity of Connelly’s Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
16.5 Randomized Algorithm for Testing Generic Global Rigidity . . . . . . . . . . . . . . . . . . . . . . . . 347
16.6 Surgery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
16.7 Other Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349

16.1 Basic Setup

In this chapter we discuss the concept of global rigidity for bar and joint frameworks in an arbitrary
fixed dimension, d. This is stronger than the plain “rigidity” property (also called local rigidity),
in that it requires that, up to congruence, there is a unique framework satisfying the given length
constraints. In particular, even if there are a only finite number of equivalent and noncongruent
frameworks, if that number is greater than one, then such frameworks are not considered to be
globally rigid.
In this chapter, we will always assume that d is a fixed dimension, G is a graph with n vertices
and m edges, and n ≥ d + 2. A (bar and joint) framework in Rd , denoted as (G, p), is a graph G
together with a configuration p = (p1 , . . . , pn ) of points in Rd .

Globally rigid: A bar and joint framework (G, p) in Rd is globally rigid if any equivalent frame-
work (G, q) in Rd is, in fact, congruent to (G, p). Otherwise, we say that the framework is
globally flexible.
Generic: We say that a set of configuration p (or a framework (G, p)) is generic if its coordinates
do not satisfy any nontrivial algebraic relations with rational coefficients.
Generically globally rigid: We say that a graph G is generically globally rigid in Rd , if all generic
frameworks are globally rigid.

341
342 Handbook of Geometric Constraint Systems Principles

Generically globally flexible: We say that a graph G is generically globally flexible in Rd , if all
generic frameworks are globally flexible. We can also call this property: generically not globally
rigid.
Generic property: A property P is generic if for every graph, either all generic frameworks satisfy
P or none do.

Global rigidity naturally arises in the context of attempting to find a d-dimensional framework
from an input set of inter-point length constraints. Only when the underlying framework is globally
rigid is such an inverse problem well posed. Note that even when the framework is globally rigid, it
may not be computationally easy to solve such an inverse problem.
The bad news is that it is NP-Hard to determine if a framework (G, p) (say with integer valued
coordinates) is globally rigid [11]. For some initial insight into why this may be so, consider the case
where the graph is an n-cycle and we have a framework in R1 . Any one-dimensional framework of
this graph gives rise to a set of signs on the edge lengths that solves a SUBSET-SUM problem
(since the cycle must close up). The ability to efficiently determine global rigidity would give one
the ability to efficiently determine if there is a second solution to a SUBSET-SUM problem that has
one known satisfying set of signs. This is an NP-HARD problem.
This pessimism leads us to look instead on understanding the so-called “generic” case. Be-
fore defining this carefully, we first describe what we are after. Let us consider the case of a one-
dimensional framework (G, p) of some graph G. Suppose that this graph is not 2-connected, then
we can find some disconnecting vertex, i, and “flip” the embedding of one of the components across
the point pi while maintaining all of the edge lengths. This will give us some equivalent framework
(G, q). This new framework will not be congruent to (G, p) unless all but one of the components are
collapsed into a single point. The special inputs, p, that have such collapses can easily be described
using a set of algebraic relations on the coordinates of p.
On the other hand consider the case where G is an n-cycle (modeling a SUBSET-SUM problem
with real edge lengths). Unless the points (or lengths) are chosen very carefully, we would not
expect to find a second way of solving the SUBSET-SUM problem. Indeed if such a solution existed,
there must be some specific non-trivial algebraic relation on the p that allows for this unexpected
second solution. If we were sufficiently dedicated, We could spend some effort to write down this
algebraic relation, but we are not so inclined; we are content to know that inputs allowing such
second solutions are very special and hard to come by. They certainly have measure 0, but as they
arise from algebraic solutions, there are even more special than that.
With a bit more effort, we can maybe convince ourselves that if G is 2-connected, then the exis-
tence of all of the overlapping cycles will ensure that (G, p) is globally rigid, unless we choose our p
very carefully (so that they satisfy some specify-able, but perhaps unspecified, algebraic equations
with rational coefficients).
Doing this, we conclude that if G is 2-connected then (G, p) in R1 will be globally rigid, unless
p is “special”. Conversely, if G is not 2-connected, then (G, p) in R1 will be globally flexible, unless
p is “special”.
There are a bunch of ways of making this notion formal, and the one typically used in the rigidity
literature is through the notion of a generic framework.
Given this notion, we can state

Theorem 16.1 A graph G is generically globally rigid in R1 if it is 2-connected. A graph G is

generically globally flexible in R1 if it is not 2-connected.

Note that our formal definition is a bit of nominal overkill, as it has ruled out saying anything
about any p that satisfy any algebraic equations defined over the rationals. While in fact, as claimed
above we really know about the behavior of (G, p) outside of one specific algebraically definable
set. But it is immediately clear that these singular sets are at the very least semi-algebraic, defined
Generic Global Rigidity in General Dimension 343

using real algebraic equalities and inequalities with rational coefficients. Thus if all generic points
lie outside such a singular set, then the singular set must be of lower dimension, and indeed lie in
some algebraic subset (defined using equalities only). Thus we could also state the theorem in the
form

Theorem 16.2 If G is 2-connected, then (G, p) is globally rigid in R1 over a Zariski-open set of
configurations. If G is not 2-connected, then (G, p) in R1 is globally flexible over a Zariski-open set
of p. Moreover, these Zariski open sets can be expressed over the rationals.

And so when you read theorems in the form of Theorem 16.1 in this chapter, you can always
think of them in the form of Theorem 16.2.
We now notice something important in this theorem. Since a graph G is either 2-connected, or it
is not-2-connected, we see that in R1 , G is either generically globally rigid, or it is generically glob-
ally flexible. In other words, if G is not generically globally rigid in R1 then it must be generically
globally flexible in R1 .
Apriori, we might have imagined that there could be some graphs G that are neither generically
globally rigid nor generically globally flexible in R1 . Perhaps, half the frameworks are globally rigid
while the other half are globally flexible. Theorem 16.2 states that this does not happen.
We formalize this with the concept of a generic property, and we state the corollary

Corollary 16.3 Global rigidity in R1 is a generic property.

The goal of this chapter is to explore the property of generic global rigidity in any arbitrary but
fixed dimension, denoted as d. In fact, as we will describe below, global rigidity in Rd is a generic
property as well.

16.2 Connelly’s Sufficiency Theorem

In this section we will describe Connelly’s sufficiency condition for generic global rigidity in Rd [3].

Stress vector: Given a graph G, a stress vector ω = (. . . , ωi j , . . . ), is an assignment of a real scalar

ωi j = ω ji to each edge, {i, j} in G. (We have ωi j = 0, when {i, j} is not an edge of G).
Equilibrium stress vector: We say that ω is an equilibrium stress vector for (G, p) if the vector
equation X
ωi j (pi − p j ) = 0 (16.1)
j

holds for all vertices i of G. The equilibrium stress vectors of (G, p) form the co-kernel of its
rigidity matrix R(p).
Stress matrix: We associate an n-by-n stress matrix Ω to a stress vector ω, by saying that i, j entry
of Ω is −ωi j , for i 6= j, and the diagonal entries of Ω are such that the row and column sums of
Ω are zero. The stress matrices of G are simply the symmetric matrices with zeros associated
to nonedge pairs, and such that the vector of all-ones is in its kernel.
Equilibrium stress matrix: If ω is an equilibrium stress vector for (G, p) then we say that the
associated Ω is an equilibrium stress matrix for (G, p).
For each of the d spacial dimensions, if we define a vector v in Rn by collecting the the associ-
ated coordinate over all of the points in p, we have Ωv = 0. Thus if the dimension of the affine
span of the vertices p is d, then the rank of Ω is at most n − d − 1, but it could be less.
344 Handbook of Geometric Constraint Systems Principles

Figure 16.1
On the left we have the configuration space, Rnd . The map f measures the squared lengths along
each of the m edges and maps to a point in Rm . The image of f over all configurations is the mea-
surement set Md . A generic configuration maps to a smooth point of Md where it has a well defined
tangent and normal space. At smooth points f (p), normal vectors are equilibrium stress vectors for
(G, p). If (G, p) and (G, q) are equivalent, they map to the same point on the measurement set, and
share the equilibrium stress matrices. If this matrix has rank n − d − 1, then such p and q must be
related by an affine transform.

Conic at infinity: We say that the edge directions of (G, p) are on a conic at infinity in Rd if
there exists some nonzero d-by-d symmetric matrix Q, such that for all edges i j, we have
(pi − p j )t Q(pi − p j ) = 0.
With these definitions, we can now state Connelly’s sufficiency condition [3].
Theorem 16.4 Suppose that there is a generic configuration p in Rd such that (G, p) has a equilib-
rium stress matrix of rank n − d − 1. Then G is generically globally rigid in Rd .

Remark 16.1 If one generic (G, p) in Rd has such an equilibrium stress matrix, then so too must
any other generic framework (G, q) in Rd .

In fact, due to the properties of matrix ranks the theorem can be slightly strengthened [5, 7] to
say
Theorem 16.5 Suppose that there is an infinitesimally rigid configuration p in Rd such that (G, p)
has a nonzero equilibrium stress matrix of rank n − d − 1. Then G is generically globally rigid in
Rd .

Remark 16.2 Caveats [5]:

Even if a infinitesimally rigid framework (G, p) has an equilibrium stress matrix of rank n − d −
1, this does not imply that (G, p) itself is globally rigid in Rd . It just implies that G is generically
globally rigid in Rd .
Even when G is generically globally rigid, there can still be (nongeneric) infinitesimally rigid
frameworks that are globally flexible.
Even if a generic framework (G, p) is globally rigid, there may be another framework (G, q) that
is globally flexible, where q is obtained from p by applying an affine transform in Rd .

It is worthwhile outlining the central ideas used in the proof of this theorem. (See Figure 16.1)
For a fixed graph G and dimension d, let f be the map that takes a configuration of n points in
Rd , and measures the squared Euclidean lengths on the m edges, thought of as a point in Rm .
Generic Global Rigidity in General Dimension 345

We will refer to the image of f , acting on all possible configurations in Rd as the measure-
ment set Md . The measurement set is a semi-algebraic set and so most of its neighborhoods look
like smooth manifolds in Rm . The nonsmooth points of Md must satisfy some additional algebraic
equations in Rm that do not vanish identically on Md , and these pull back through f to some non-
trivial algebraic equations on configuration space. Thus the image of f , of a small ball around some
generic p, must be a smooth manifold, with a well defined tangent and normal space in Rm .
Recall that any equilibrium stress vector ω of (G, p) is in the cokernel of the rigidity matrix of
(G, p) which represents the Jacobian of f . Thus ω is orthogonal in Rm to image of the differential
of f at p. When p is generic, the image of the differential spans the tangent space of Md at f (p),
placing ω within its normal space.
With this picture in mind, if (the generic) (G, p) and (G, q) are equivalent, then f (p) = f (q);
both configurations map to the same smooth point in Md . Thus ω (as a normal vector) must be
orthogonal to the columns of the rigidity matrix of (G, q). Thus ω must also be an equilibrium
stress for (G, q).
Suppose that the the rank of Ω is n − d − 1, then this implies that the configuration q must arise
as a d-dimensional affine transform of the configuration p.
At this point we are almost done, all we need to rule out is the possibility that there is an
equivalent (G, q) which arises as an affine (but non Euclidean) transform of (G, p). This possibility
is formalized using the notion of a conic at infinity.
The following simple proposition is proved in [3].

Proposition 16.1
Every affine transform of p preserving edge lengths of (G, p) is a congruence, if and only if the edge
directions of (G, p) are not on a conic at infinity.

And the last piece of the puzzle is then the following proposition from [3].

Proposition 16.2
Suppose that (G, p) is a generic framework in Rd and each vertex has degree at least d, then the
edge directions of (G, p) are not on a conic at infinity.

Putting these pieces together, one can complete Connelly’s proof.

Here is a useful variant of Proposition 16.2 that does not require genericity [1]:

Proposition 16.3
Suppose that (G, p) is a framework in Rd and for each vertex v, the affine span of v and its neighbors
in (G, p) span all of Rd . Suppose also that (G, p) has an equilibrium stress matrix of rank n − d − 1.
Then the edge directions of (G, p) are not on a conic at infinity.

16.3 Hendrickson’s Necessary Conditions

In this section we describe Hendrickson’s necessary conditions for generic global rigidity [9]. As
described in Chapter 21, these condition are also sufficient for generic global rigidity in R2 . In three
dimensions and higher, though, these conditions are not sufficient.

Redundantly rigid: A graph G is redundantly rigid in Rd if it remains rigid even after removing
any single edge.
346 Handbook of Geometric Constraint Systems Principles

Figure 16.2
The closed curve (topological circle) represents a set of configurations that are equivalent once edge
ei j is removed from G. The vertical projection, π, represents the measurement of the squared length
between vertex i and vertex j. A generic point in the range of this measurement, π(p), must have an
even number of points in the same fiber, proving the existence of the equivalent (G, q).

Theorem 16.6 Let (G, p) be a generic framework in Rd . If (G, p) is globally rigid then both of the
following must hold
(a) G is vertex (d + 1)-connected.
(b) G is redundantly rigid in Rd .

This means that if G does not satisfy one of these two conditions, then G is generically globally
flexible in Rd .
Again, it is worthwhile outlining the central ideas used in the proof of this theorem. (See Fig-
ure 16.2.)
The necessity of the first condition is something we already saw above when discussing the
necessity of 2-connectivity for generic global rigidity in one dimension. This generalizes easily
to d dimensions. If the graph is not d + 1 connected, then we can find d vertices whose removal
would disconnect the graph into multiple components. Generically, we can simply reflect one of
these components across the hyperplane spanned by the cut vertices to obtain an equivalent but
noncongruent second framework.
The second condition is more interesting. Suppose that G is not redundantly rigid. We remove
one edge ei j that makes the resulting graph G0 flexible. Henrickson then argues (using genericity,
Sard’s theoerem and the implicit function theorem) that the set of equivalent frameworks (modulo
congruence) to (G0 , p) must form a smooth compact manifold, and thus must contain a circle. One
then looks at the squared distance between the pair pi and p j (corresponding to the removed edge)
as we travel along this circle in configuration space. This gives a map from a circle to the real line.
Such a map must have an mod-2-degree of zero. Thus any generic point in its image must have at
least one other point its fiber. Any such other point will be an equivalent but incongruent framework
with respect to the original graph G.

16.3.1 Nonsufficiency
In three and higher dimensions, there are examples of graphs that are d + 1 connected and redun-
dantly rigid while not being generically globally rigid. Connelly found [2] a family of such examples
that were complete bipartite graphs in various dimensions The smallest such instance is K5,5 in R3 .
More kinds of examples have been found by Frank and Jiang [6] and more recently by Jordan et
al. [10].
Generic Global Rigidity in General Dimension 347

16.4 Necessity of Connelly’s Condition

Connelly conjectured that the conditions of Theorem 16.4 were in fact necessary and this was proven
by Gortler, Healy, and Thruston [7].

Theorem 16.7 Suppose that a generic configuration p in Rd is globally rigid in Rd . Then (G, p)
must have a equilibrium stress matrix of rank n − d − 1.

Moreover, in light of Remark 16.1, one generic framework has an equilibrium stress matrix of
rank n − d − 1 iff all generic frameworks do. So the theorem tells us that if one generic framework
does not have an equilibrium stress matrix of rank n − d − 1 then G is generically globally flexible
in Rd .
Together with Theorem 16.4, this gives us
Corollary 16.8 Global rigidity in Rd is a generic property.
Additionally, using the corollary one can prove [5, 7]:
Theorem 16.9 Suppose that an infinitesimally rigid configuration p in Rd is globally rigid in Rd .
Then G must be generically globally rigid. And conversely if G is generically globally flexible, then
all infinitesimally rigid frameworks must be globally flexible.

Remark 16.3 Caveat [5]: G can be generically globally rigid, and (G, p) can be infinitesimally
rigid, while (G, p) still does not have an equilibrium stress matrix of rank n − d − 1.

The proof of theorem 16.7 is inspired by the mod-2 degree argument used in the proof of Hen-
drickson’s theorem, though the details get a bit more complicated. In this case they look at the
degree of a certain map with a certain range and certain domain: The domain is the be the set of
frameworks (G, q) that are in equilibrium under some chosen Ω, an equilibrium stress of (G, p).
The range is Rm where m is the number of edges. The map simply measures the squared length of
each edge in a framework.
Under the assumption of genericity and that (G, p) does not have an equilibrium stress matrix
of rank n − d − 1, they show that this map has a well defined degree-mod-2. The proof of this step
uses some non-trivial properties of ruled algebraic varieties. We note, for further reference that it
also uses the fact that if the graph is connected and one-vertex has been pinned down (to mod out
translations), then the squared length map is proper.

16.5 Randomized Algorithm for Testing Generic Global Rigidity

The above theory leads naturally to a randomized efficient algorithm for checking if a graph G is
generically globally rigid in Rd .
The basic idea is that if G is generically globally rigid, then a randomly selected equilibrium
stress matrix of a randomly selected configuration p will (with high probability) have rank n −
d − 1. (Keep in mind that a randomly select matrix from a linear space of matrices will, with high
probability, exhibit the largest rank of any matrix in the space).
The algorithm is outlined as follows:
(a) Pick a random configuration p in Rd .
348 Handbook of Geometric Constraint Systems Principles

(b) If (G, p) is infinitesimally flexible, then output “generically globally flexible” and exit.
(c) Calculate the space of equilibrium stresses of (G, p).
(d) Pick a random equilibrium stress from this space.
(e) Calculate the rank of this equilibrium stress.
(f) If this rank is less than n − d − 1 then output “generically globally flexible” and exit.
(g) Output “generically globally rigid” and exit.

in step (2) is done by computing the rigidity matrix of (G, p) and comparing its rank to
The test
nd − d+1 2 as in Chapter 18.
Step (3) is done by computing the co-kernel of the rigidity matrix of (G, p). Any vector in the
co-kernel is an equilibrium stress vector.
If the algorithm has reached step (7), then it must have found an equilibrium stress matrix of
rank n − d − 1, for an infinitesimally rigid (G, p). In this case, G is certifiably generically globally
rigid using Theorem 16.5. Thus the algorithm is always correct when it outputs “generically globally
rigid.” On the other hand, when the algorithm outputs “generically globally flexible,” there is (only)
a small chance that it picked an exceptional framework (G, p), or an exceptional equilibrium stress
matrix Ω.
In practice, one might perform the above steps numerically using floating point numbers. To
obtain theoretical guarantees [7], one can perform the above steps using integers modulo some
random prime larger than 4mn.

16.6 Surgery
There are various combinatorial operations on graphs that preserve generic global rigidity. We sum-
marize some of these results here.
Connelly and Whiteley [5] discuss generic global rigidity under the coning operation. Starting
with a graph G, one new vertex v is added and edges are added between v and all of the other
vertices.

Theorem 16.10 A graph G is generically globally rigid in Rd iff its cone graph is generically
globally rigid in Rd+1 .

Connelly [4] showed the following

Theorem 16.11 Suppose that G1 and G2 are generically globally rigid in Rd , each with at least
d + 2 vertices. Suppose that d + 1 vertices of G1 are identified with d + 1 vertices of G2 . Next, Let G
be the union of G1 and G2 . Next, up to one edge can be removed from G, if it is common to both G1
and G2 . Then G is generically globally rigid in Rd .

This theorem has been generalized by Tanigawa [13]. That paper also includes the following
interesting theorem:

Theorem 16.12 If G is vertex redundantly rigid in Rd , then G is generically globally rigid in Rd ,

Szabadka [12] shows the following

References 349

Theorem 16.13 A graph obtained from a generically globally rigid graph in Rd by a 1-extension
is generically globally rigid in Rd .
A 1-extension in Rd removes an existing edge e and adds a new vertex with d + 1 new incident
edges so that the new vertex is incident with both the endpoints of e.

16.7 Other Spaces

Connelly and Whiteley [5] studied the relationship between generic global rigidity in Rd and on the
sphere Sd .

Theorem 16.14 A graph G is generically globally rigid in Rd iff G is generically globally rigid in
Sd . Additionally, generic global rigidity in Sd is a generic property.

Similarly, Gortler and Thurston [8] studied the relationship between generic global rigidity in
various spaces.

Theorem 16.15 A graph G is generically globally rigid in Rd iff G is generically globally rigid in
Cd . Additionally, generic global rigidity in Cd is a generic property.

In the above notion, the complex edge squared length is measured as the complex (pi − p j )2 ,
with no conjugation. (If conjugation is used, then this problem simply reduces to that of global
rigidity in R2d .)
They also looked at generic global rigidity in pseudo-Euclidean spaces such as Minkowski
space.

Theorem 16.16 A graph G is generically globally rigid in Rd iff G is generically globally rigid in
any and all d-dimensional pseudo-Euclidean spaces and hyperbolic space.

Remark 16.4 It is not known if generic global rigidity in a pseudo-Euclidean space is a generic
property. It is possible that there are graphs that are not generically globally rigid in Rd (equiv. Cd )
but are not generically globally flexible in a pseudo-Euclidean space (rather it has generic points
exhibiting both behaviors).

References
[1] A.Y. Alfakih and Viet-Hang Nguyen. On affine motions and universal rigidity of tensegrity
frameworks. Linear Algebra and its Applications, 439(10):3134–3147, 2013.
[2] Robert Connelly. On generic global rigidity. In Applied geometry and discrete mathematics,
volume 4 of DIMACS Ser. Discrete Math. Theoret. Comput. Sci., pages 147–155. Amer. Math.
Soc., Providence, RI, 1991.
[3] Robert Connelly. Generic global rigidity. Discrete Comput. Geom, 33(4):549–563, 2005.
[4] Robert Connelly. Combining globally rigid frameworks. Proceedings of the Steklov Institute
of Mathematics, 275(1):191–198, 2011.
350 References

[5] Robert Connelly and WJ Whiteley. Global rigidity: the effect of coning. Discrete & Compu-
tational Geometry, 43(4):717–735, 2010.
[6] Samuel Frank and Jiayang Jiang. New classes of counterexamples to hendricksons global
rigidity conjecture. Discrete & Computational Geometry, 45(3):574–591, 2011.
[7] Steven J. Gortler, Alexander D. Healy, and Dylan P. Thurston. Characterizing generic global
rigidity. American Journal of Mathematics, 132(4):897–939, 2010.
[8] Steven J. Gortler and Dylan P. Thurston. Generic global rigidity in complex and pseudo-
euclidean spaces. In Rigidity and symmetry, pages 131–154. Springer, 2014.
[9] Bruce Hendrickson. Conditions for unique graph realizations. SIAM J. Comput., 21(1):65–84,
February 1992.
[10] Tibor Jordán, Csaba Király, and Shin-Ichi Tanigawa. Generic global rigidity of body-hinge
frameworks. Technical report, EGRES Technical Reports, TR2014-06, 2014.
[11] J. B. Saxe. Embeddability of weighted graphs in k-space is strongly NP-hard. In Proc. 17th
Allerton Conf. in Communications, Control, and Computing, pages 480–489, 1979.
[12] Zoltán Szabadka. Globally rigid frameworks and rigid tensegrity graphs in the plane. PhD
thesis, Department of Operations Research, Eötvös Loránd University, 2010.
[13] Shin-Ichi Tanigawa. Sufficient conditions for the global rigidity of graphs. Journal of Com-
binatorial Theory, Series B, 113:123–140, 2015.
Chapter 17
Change of Metrics in Rigidity Theory
Anthony Nixon
Department of Mathematics and Statistics, Lancaster University, U.K.

Walter Whiteley
Department of Mathematics, York University, Canada

CONTENTS
17.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
17.2 Projective Transfer of Infinitesimal Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
17.2.1 Coning and Spherical Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
17.2.2 Rigidity Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
17.2.2.1 Spherical to Affine Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
17.2.3 Equilibrium Stresses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
17.2.4 Point-Hyperplane Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
17.2.5 Tensegrity Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
17.3 Projective Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
17.4 Pseudo-Euclidean Geometries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
17.4.1 Hyperbolic and Minkowski Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
17.5 Transfer of Symmetric Infinitesimal Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
17.5.1 Symmetric Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
17.6 Global Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
17.6.1 Universal Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
17.6.2 Projective Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
17.6.3 Pseudo-Euclidean Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
17.7 Summary and Related Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372

17.1 Introduction
In this chapter we abstract the basic notions of rigidity and flexibility to apply in various geometries
which live in a shared projective space. We show the equivalence of infinitesimal rigidity across the
Cayley-Klein geometries using projective transformations.
Given rigidity properties of a framework (G, p) in Rd , we can often lift these properties to
a coned framework in Rd+1 , and then re-project to a different hyperplane to obtain another d-
dimensional realization (G, q) as a slice of the cone, preserving key rigidity properties. This is a
concrete form of a projective transformation from (G, p) to (G, q). Alternatively, there may be a di-
rect analysis of the impact of a projective transformation on the various properties. These projective

351
352 Handbook of Geometric Constraint Systems Principles

transformations preserve the infinitesimal rigidity of (G, p), and have a clearly described impact
on the coefficients of any equilibrium-stress of (G, p). These processes also provide the tools to
confirm the transfer of these basic properties among Euclidean, Spherical, Pseudo-Euclidean, and
Hyperbolic metrics –all of which live in a common projective space. The geometric transfer, for
a specific projective configuration p̃ in any of the metrics, then gives corresponding combinatorial
transfers of generic properties.
We also consider global (and universal) rigidity in these spaces by analysing the stress matrix.
In this setting a number of key transfer results are known generically, or combinatorially, but the
transfers are less geometrically robust.
In Section 17.2 we present the equivalence of infinitesimal rigidity for Euclidean and Spherical
frameworks by passing through cone frameworks and then in Section 17.3 we consider projective
frameworks directly. Section 17.4 considers extensions to Minkowski and Hyperbolic geometries
and analogues for symmetric frameworks are discussed in Section 17.5. In Section 17.6 we consider
the more detailed transfer of global and universal rigidity before concluding by mentioning some
related work and summarising the results of the chapter in Section 17.7.

17.2 Projective Transfer of Infinitesimal Rigidity

Infinitesimal flex: Let G = (V, E). We say that u : V → Rd is an infinitesimal flex of (G, p) if u
satisfies
hpi − p j , ui − u j i = 0 ({i, j} ∈ E).
Infinitesimally rigid: A framework (G, p) in Rd is infinitesimally rigid if the dimension of the
d+1

space of infinitesimal flexes of (G, p) is equal to 2 (assuming that the points p(V ) affinely
span Rd ).
Rigidity matrix: The rigidity matrix Rd (G, p) of a framework (G, p) in Rd is the |E| × d|V | matrix
with rows indexed by edges and columns by d-tuples of vertices with the following form:
i j
.. ..
 
 . . 
Rd (G, p) = {i, j} 
 0 ... 0 pi − p j 0 ... 0 p j − pi 0 ... 0
.
.. ..
. .

Cone graph: Let G = (V, E) be a graph. The cone of G is the graph G ∗ o which has vertex set V ∪ o
and edge set E ∪ {{x, o} : x ∈ V }.
Cone framework: Given a framework (G, p) in Rd , the cone framework (G ∗ o, p̄∗ ) is defined by
the realization p̄∗ : V ∪ o → Rd+1 such that p̄i = (pi , 1) ∈ Rd+1 and p̄∗ (o) = (0, . . . , 0).
Spherical framework: Define the unit sphere Sd in Rd+1 by Sd = {(x1 , x2 , . . . , xd+1 ) ∈ Rd+1 :
x12 + x22 + · · · + xd+1
2 = 1}. Let G = (V, E) be a graph, a framework (G, p) on Sd is an ordered
pair consisting of a graph G and a realization p such that pi ∈ Sd for all i ∈ V .
Spherical infinitesimal flex: A map u : V → Rd+1 is said to be an infinitesimal flex of a Spherical
framework (G, p) if it satisfies
hpi , u j i + hp j , ui i = 0 ({i, j} ∈ E) and
hpi , ui i = 0 (i ∈ V ).
Change of Metrics in Rigidity Theory 353

Projective transformation: A d-dimensional framework (G, q) is a projective transformation of a

d-dimensional framework (G, p) if there is an invertible (d + 1) × (d + 1) matrix T for which
T (pi , 1) = λi (qi , 1) for all 1 ≤ i ≤ n (where λi is a scalar and (pi , 1) are the affine coordinates
of pi ).
Equilibrium stress: An equilibrium stress ω of a framework (G, p) in Rd is a vector ω in the
cokernel of Rd (G, p). Equivalently, ω is an equilibrium stress if
n
X
ωi j (pi − p j ) = 0 for all 1 ≤ i ≤ n, (17.1)
j=1

where ωi j is taken to be equal to ωe if e = {i, j} ∈ E and to be equal to 0 if {i, j} 6∈ E.

Sliding: Given a cone framework (G ∗ o, p̄∗ ) in Rd+1 , the process of sliding moves the points
p̄i , i ∈ V , along the unique lines connecting them to the cone vertex (while avoiding the cone
vertex itself). Algebraically, p̄i goes to αi p̄ for α 6= 0.

17.2.1 Coning and Spherical Frameworks

We embed Rd into Rd+1 , in the hyperplane xd+1 = 1 (where xd+1 is the last coordinate of vectors
in Rd+1 ). In this manner we consider a framework (G, p) in Rd as a framework (G, p) on the
hyperplane xd+1 = 1 by setting p̄i = (pi , 1) for 1 ≤ i ≤ |V |.
With this embedding we then cone to (G ∗ o, p̄∗ ) by adding a new vertex o and setting p∗ (o) =
(0, . . . , 0). Now slide the points pi along the line through the origin o and pi such that the lengths
of the edges oi are all equal. See Figure 17.1 for an example. We get a new framework which is
infinitesimally rigid if and only if the original was. If we assume the cone edges are all length 1 then
we have a framework in Rd+1 which we can view as a framework on the sphere.

Figure 17.1
An infinitesimally rigid framework in R2 realized in a hyperplane in R3 then coned to an infinitesi-
mally rigid framework in R3 . Each point x in the cone (minus its vertex), determines a unique line.
Sliding x along this line (minus the cone vertex) preserves infinitesimal rigidity.

This motivates the following presentation. Following [32] we focus on infinitesimal flexes. The
transfer of infinitesimal rigidity can also be seen at the level of matrix operations on the rigidity ma-
trices which we will subsequently define. Define the upper-hemisphere Sd>0 = {x ∈ Sd : hx, ei > 0}
where e = (0, . . . , 0, 1) ∈ Rd+1 . Suppose we have a framework (G, p) on Sd>0 . If p(t) is a con-
tinuous flex of (G, p) on Sd>0 , then for all t in some interval and for all {i, j} ∈ E, we have
354 Handbook of Geometric Constraint Systems Principles

dSd (pi (t) · p j (t)) = ci j , where dSd denotes Spherical distance, ci j is constant for all {i, j} ∈ E,
>0 >0
and for all t and k ∈ V , pk (t) · pk (t) = 1. Equivalently, for all t, {i, j} ∈ E and k ∈ V , we have the
system

pi (t) · p j (t) = cos ci j (17.2)

pk (t) · pk (t) = 1. (17.3)

If the flex p(t) is differentiable at t = 0, then p(t) must satisfy pi · p j (0) + pi (0) · p j = 0 and pk ·
pk (0) = 0.
It follows that an infinitesimal flex of the framework (G, p) on Sd>0 is a map u : V → Rd+1
satisfying, for each {i, j} ∈ E and for each k ∈ V ,

pi · u j + p j · ui = 0 and pk · uk = 0. (17.4)

A trivial infinitesimal flex of Sd>0 is an infinitesimal flex which is also an infinitesimal isometry of
Rd+1 . The framework (G, p) is infinitesimally rigid on Sd>0 if all infinitesimal flexes of (G, p) are
restrictions of trivial infinitesimal flexes.
If (G, p) is a bar-and-joint framework on Sd>0 , then the cone framework (G ∗ o, p̄∗ ) is infinites-
imally rigid in Rd+1 if and only if (G, p) is infinitesimally rigid on Sd+1 >0 . That is, frameworks on
d
S>0 can be modeled by the cone on the same framework in R . d+1

We now present two maps, a map carrying a framework (G, p) on Sd>0 into a framework (G, q)
in Rd , and a map carrying the infinitesimal flexes of (G, p) into infinitesimal flexes of (G, q). The
latter map carries trivial infinitesimal flexes of Sd>0 to trivial infinitesimal flexes of Rd , yielding the
equivalence of infinitesimal rigidity in these spaces.
If (G, p) is a framework on Sd>0 , then G(φ ◦ p) is a framework in Rd , where φ : Sd>0 → Rd is
given by
x
φ (x) = .
e·x
The inverse of φ is given by
x
φ −1 (x) = .
x·x

pi ui

φ (pi ) -hui , pi ie

ψ pi (ui )

Figure 17.2
Transfer of infinitesimal flexes between Rd and Sd>0 .

If u is an infinitesimal flex of the framework (G, p) on Sd>0 , let ψ denote the map

1
ψ : ui → (ui − (ui · e)e).
e · pi
Change of Metrics in Rigidity Theory 355

If (G, q) is a framework in Rd with infinitesimal flex u, then ψ −1 is given by

1
ψ −1 : ui → √ (ui (ui · qi )e).
qi · qi

These maps are illustrated in Figure 17.2. Observe that ψ and ψ −1 map into the appropriate tangent
spaces: φ −1 (qi ) · ψ −1 (ui ) = 0 and ψ(ui ) · e = 0. By direct computation we have the following
theorem.

Theorem 17.1 A vector u ∈ Rd is an infinitesimal flex of the framework (G, p) on Sd>0 if and only if
ψ ◦ u is an infinitesimal flex of the framework (G, φ ◦ p) in Rd . Moreover, u is a trivial infinitesimal
flex if and only if ψ ◦ u ◦ φ −1 is a trivial infinitesimal flex.

Corollary 17.2 A framework (G, p) is infinitesimally rigid on Sd>0 if and only if (G, φ ◦ p) is in-
finitesimally rigid in Rd .

17.2.2 Rigidity Matrices

One can take multiple viewpoints on the Spherical rigidity matrix. Consider the following matrix:

i j
 
{i, j} 0 ... 0 pi − p j 0 ... 0 p j − pi 0 ... 0
 .. .. 

 . . 

 0 ... 0 0 ... 0 0 ... 0
i  pi 0 .
 .. .. 
 . . 
j 0 ... 0 0 0 ... 0 pj 0 ... 0
Firstly, as above this is the rigidity matrix as a cone framework in Rd+1 with the cone vertex at
the origin and the columns for the cone vertex deleted. Secondly we can view this as the rigidity
matrix for a framework on the sphere with vertex rows encoding the fact that infinitesimal flexes
are tangent to the surface of the sphere (here we can interpret the edges as either Euclidean [25] or
geodesic [43]). Thirdly, and we will take this viewpoint in what follows, the following row equiva-
lent form of the rigidity matrix which is more natural for projective reasoning:

i j
 
{i, j} 0 ... 0 pj 0 ... 0 pi 0 ... 0
 .. .. 

 . . 

RSd (G, p) = i  0 ... 0
 pi 0 ... 0 0 0 ... 0.
 .. .. 
 . . 
j 0 ... 0 0 0 ... 0 pj 0 ... 0
A more subtle task is to extend this transfer of infinitesimal flexes to a transfer of finite flexes.
This is more challenging because flexibility is not a projective invariant, unlike infinitesimal flexi-
bility, see [30]. We illustrate this with an example in Figure 17.3.
However, rigidity does transfer for regular configurations, which form an open dense subset of
all configurations. A configuration p in Sd>0 is regular for G if RSd (G, p) has the maximum rank
over all configurations x in Sd>0 . Similarly q in Rd is regular for G if Rd (G, p) has the maximum
rank over all x in Rd . Note that p is regular on Sd>0 if and only if (G, φ ◦ p) is regular for G in Rd .
This gives a strong transfer of rigidity and infinitesimal rigidity at regular configurations.
356 Handbook of Geometric Constraint Systems Principles

(a) (b)

Figure 17.3
Two realizations of the complete bipartite graph K3,3 . Both realizations are infinitesimally flexible
but only the realization on the left (with the lines perpendicular) has a continuous flex.

Theorem 17.3 If p is a regular point for (G, p) in Sd>0 the the following are equivalent:

(a) (G, p) is rigid on Sd>0 ;

(b) (G, p) is infinitesimally rigid on Sd>0 ;

(c) (G, φ ◦ p) is rigid in Rd ;

(d) (G, φ ◦ p)) is infinitesimally rigid in Rd .

Since flexibility is the negation of rigidity; we conclude that, at a regular point, (G, p) is flexible
on Sd>0 if and only if (G, φ ◦ p) is flexible in Rd . See also [34, Section 4.2] for an example at a
non-regular point where flexibility is not transferred under coning.
We will consider affine space Ad as the hyperplane xd+1 = 1 in Rd+1 . Recall φ −1 : Ad → Sd>0
is defined by φ −1 (x) = x·x
x
(central projection).
For notational simplicity we will define the affine rigidity matrix only for A2 . The rigidity matrix
for (G, p) in A2 has the form:

p̃i p̃ j
 
xi xj yi yj xj xi yj yi
edge {i, j}
 ... zi − zj zi − zj 0 ... zj − zi zj − zi 0 ...  
 .. .. .. .. .. .. 
. . . . . .
 
 
xi yi
RA2 (G, p̃) = vertex i  . . . 1 ··· 0 0 0 · · · .
 
 zi zi 
 .. .. .. .. .. .. 
 . . . . . . 
 xj yj 
vertex j  . . . 0 0 0 ··· zj zj 1 ··· 

Theorem 17.4 (Saliola and Whiteley, 2007 [32]) A bar-joint framework (G, p) is infinitesimally
rigid in Ad if and only if (G, φ ◦ p) is infinitesimally rigid on Sd>0 .
Change of Metrics in Rigidity Theory 357

Figure 17.4
(a) A Spherical bar-joint framework (G, p) on S2>0 and (b) the corresponding bar-joint framework
in R2 .

As an example we present Figure 17.4. This gives an infinitesimally rigid framework on S2>0
and the corresponding framework in R2 .
If we focus on the framework (G, p) on the upper hemisphere S2>0 and scale each pi = (xi , yi , zi )
with zi 6= 0 by z1i to obtain p̃i = ( xzii , yzii , 1) to create a framework (G, p̃) which is row equivalent to
RS2 (G, p) (see, also, [34]). In particular they have isomorphic kernels. The corresponding matrix
>0
transformations result in the affine rigidity matrix of the framework.
Given a framework (G, q) on Sd and I ⊆ V , the inversion ι (with respect to I) is an operator
acting on q such that (ι ◦ q)i = −qi if i ∈ I and (ι ◦ q)i = qi otherwise. Note that u is an infinitesimal
flex of (G, q) if and only if ι ◦ u, defined by (ι ◦ u)i = −ui (i ∈ I) and (ι ◦ u)i = ui (i ∈ V \ I), is an
infinitesimal flex of (G, ι ◦ q), which again means that ι preserves infinitesimal rigidity. See Figure
17.5. Inversion also takes a regular point q to a regular point (ι ◦ q). More generally, inversion takes
a flexible framework (G, q) to a flexible framework (G, ι ◦ q).
We use inversion to flip points in Sd<0 to Sd>0 so that a framework (G, q) on Sd (with no points
on the equator) is transferred to a framework (G, ι ◦ q) on Sd>0 . It follows that a framework (G, q)
on Sd can be transformed to a framework (G, ι ◦ γ ◦ q) on Sd>0 by first applying a rotation γ which
moves all points off the equator, and then applying ι to flip points to Sd>0 . In this way all of the
results about Sd>0 extend to Sd .

17.2.2.1 Spherical to Affine Transfer

We can move from the Spherical rigidity matrix to the affine rigidity matrix by a sequence of row and
column operations. These do not affect the dimension of the kernel. They are just multiplying the
matrix on the left by an invertible matrix TA,S . The only column operation required is to multiply the
three columns corresponding to the vertex pi by zi (and similar situation for each vertex p j ) which
shows that velocities (elements of the kernel) will transfer by ui → ( z1i )ui = ũi from the Spherical
matrix RS (G, p) to the affine matrix RA (G, p̃) and conversely. The kernels are isomorphic, and a
framework (G, p) is infinitesimally rigid on the sphere if and only if the corresponding framework
G, p̃) is infinitesimally rigid.
Suppose, for simplicity, that d = 2. We can go a step further and isolate the framework on the
plane z = 1 as a plane framework, (G, p̂). When we compare the matrix RA2 (G, p) with the usual
358 Handbook of Geometric Constraint Systems Principles

(a) (b)

Figure 17.5
A framework on S2 with points in the lower and upper-hemispheres and on the equator (a). The
dotted lines illustrate inversion applied to part of the Spherical framework to move to S2≥0 , and the
result of inversion is shown in (b).

rigidity matrix R2 (G, p̂) =

p̂i p̂ j
 
 .. .. .. .. 
 . . . . 
edge {i, j}  . . . x̂i − x̂ j ŷi − ŷ j ... x̂ j − x̂i ŷ j − ŷi . . . ,
 
.. .. .. ..
 
 
 . . . . 

where p̂i = (x̂i , ŷi ) = ( xzii , yzii ) for i ∈ V , we notice that they record equivalent information. That is,
there exists a one-one correspondence between the solution spaces of the Spherical matrix RS2 (G, p)
and the plane matrix R2 (G, p̂). Specifically, the kernels of the affine matrix RA (G, p̃) and the plane
matrix R2 (G, p̂) correspond in the following way. Define the maps: û : V → R2|V | by ûi = (ûi,x , ûi,y );
ũ : V → R3|V | by ũi = (ûi,x , ûi,y , −hûi , p̂i i); and u : V → R3|V | by ui = αi ũi (where αi = k p̃i k is a
scalar). Then û is an infinitesimal flex of (G, p̂) if and only if ũ is an infinitesimal flex of (G, p̃)
which holds if and only if u is an infinitesimal flex of (G, p).
A summary of the transfer of infinitesimal rigidity between the different geometries is given in
Figure 17.6.

17.2.3 Equilibrium Stresses

Having examined the theory of first-order flexes, we pause to present the flexes as the solutions to a
matrix equation Rd (G, p)x = 0. In this setting, we have the equivalent theory of static rigidity work-
ing with the row space and row dependences (the equilibrium stresses) of these matrices, instead of
the column dependencies (the infinitesimal flexes). The correspondence is immediate, but it takes a
particularly nice form for the “projective” models in Euclidean space of the standard metrics [32].
In this setting, the rigidity correspondence is a simple matrix multiplication:

RSd (G, p)[TXY ] = Rd (G, p)

>0
Change of Metrics in Rigidity Theory 359
Euclidean plane
hyperbolic plane
R2

H2
Projection of sphere

S2
D2

de Sitter plane

Figure 17.6
Visualization of the transfer between metrics.

for the same underlying configuration p, where TXY is a block diagonal matrix with a block entry
for each vertex, based on how the sense of perpendicular is twisted at that location from one metric
to the other. As a consequence of this simple correspondence of matrices, we see that row depen-
dencies (the static equilibrium stresses) are completely unchanged by the switch in metric. As a
biproduct of this static correspondence, there is a correspondence for the infinitesimal rigidity of the
structures with inequalities, the tensegrity frameworks, which are well understood as a combination
of first-order theory and equilibrium stresses of the appropriate signs for the edges with pre-assigned
inequality constraints (see Chapter 14).

Lemma 17.1 Saliola and Whiteley, 2007 [32]

There is an equilibrium stress ω for a framework (G, p) in Rd if and only if there is an equilibrium
stress ω̄ for the corresponding framework (G, p̄) on Sd>0 .

We note that the static theory of rigidity was studied in detail by Crapo and Whiteley [12], in
a projective setting. The projective invariance of statics was known back in the 1860’s, at the time
that projective geometry reached the U.K. [29].

17.2.4 Point-Hyperplane Frameworks

We now extend the transfer of infinitesimal rigidity to include points on the equator of the sphere
and hence points at infinity in the corresponding affine plane. Here a surprising link to point-line
frameworks (see Chapter 22) was developed in [14].
Let us denote the underlying graph of a point-hyperplane framework in Rd by G = (VP ∪
VL , EPP ∪ EPL ∪ ELL ), where VP and VL represent the set of points and the set of hyperplanes, re-
spectively. The edge set is partitioned into EPP , EPL , ELL according to the bipartition {VP ,VL } of
the vertex set. Alongside the usual point configuration we have the line configuration ` = (a, r) :
VL → Sd−1 × R. Then each i ∈ VP is associated with pi ∈ Rd while each ` j ∈ VL is associated with
a hyperplane {x ∈ Rd : ha j , xi + r j = 0} for some a j ∈ Sd−1 and r j ∈ R.
Point-hyperplane rigidity is then the natural generalisation of point-line rigidity to higher di-
mensions with constraints representing the distances between some pairs of points, the distances
360 Handbook of Geometric Constraint Systems Principles

between some pairs of points and hyperplanes and angles between some pairs of hyperplanes. More
formally the constraints, after differentiating, are determined by the following system of equations:

hpi − p j , ui − u j i = 0 ({i, j} ∈ EPP ) (17.5)

hpi , ȧ j i + hui , a j i − ṙ j = 0 ({i, j} ∈ EPL ) (17.6)
hai , ȧ j i + hȧi , a j i = 0 ({i, j} ∈ ELL ) (17.7)
hai , ȧi i = 0 (i ∈ VL ). (17.8)

˙ with `˙ = (ȧ, ṙ), is said to be an infinitesimal flex of (G, p, `) if it satisfies the system
A map (u, `),
(17.5)–(17.8), and (G,
p, `) is infinitesimally rigid if the dimension of the space of its infinitesimal
flexes is equal to d+1
2 , assuming the points p(VP ) and hyperplanes `(VL ) affinely span R .
d

Translate the point-hyperplane framework (G, p, `) to a point-hyperplane framework (G, p̂, `) in

affine space Ad by taking p̂i = (pi , 1) for all i ∈ Vp . It becomes clear that the linear constraints for an
infinitesimal flex of a point-hyperplane framework in Ad are almost identical to those for a bar-joint
framework in Sd . (See [14] for more details.)

Theorem 17.5 (Eftekhari et al., 2017 [14]) Let (G, p̂, `) be a point-hyperplane framework in Rd
with G = (VP ∪ VL , E). Let (G, p) be the corresponding framework on Sd with the points p(v) for
v ∈ VL lying on the equator ∗ and let (G, q) be the corresponding framework in Rd with the points
p(v) for v ∈ VL lying on a hyperplane. Then the following are equivalent:
(a) the point-hyperplane framework (G, p̂, `) is infinitesimally rigid in Rd ;
(b) the bar-joint framework (G, p) on Sd is infinitesimally rigid; and
(c) the bar-joint framework (G, q) in Rd with the points in VL lying on a hyperplane is infinitesi-
mally rigid.

To illustrate how this transfer can be useful, we note that it leads directly to a combinatorial
characterisation of bar-joint frameworks in the plane in which a subset of points are collinear. This
follows from the equivalence above because of a combinatorial characterisation of rigidity for point-
line frameworks due to Jackson and Owen [21]. This result extends a theorem of Jackson and Jordán
[20] dealing with the case of 3 collinear points.
In [14] a number of further combinatorial results are given for generic point-hyperplane frame-
works with various constraints on the hyperplanes (fixed, fixed intercept and fixed normal).

17.2.5 Tensegrity Frameworks

As well as bar-joint frameworks one may consider the generalisation to tensegrity frameworks where
some bars are allowed to increase but not decrease in length and some are allowed to decrease but
not increase (see Chapter 14). We point out two theorems that show that infinitesimal rigidity for
tensegrities can be understood by coning to the sphere.

Theorem 17.6 (Schulze and Whiteley, 2012 [34]) Given a tensegrity framework (G, p) in Rd and
a corresponding cone framework (G ∗ o, p̄∗ ) in Rd+1 (where all edges in G remain as bars, cables
and struts respectively and the cone edges are all bars). Then (G, p) is an infinitesimally rigid
tensegrity framework in Rd if and only if (G ∗ o, p̄∗ ) is an infinitesimally rigid tensegrity framework
in Rd+1 .
∗ A calculation translates the system (17.5)–(17.8) into an equivalent linear system where r does not appear [14]. In other
j
words the last coordinate of ` j does not affect the infinitesimal rigidity so we may assume that ` : VL → Sd−1 × {0}.
Change of Metrics in Rigidity Theory 361

Theorem 17.7 (Schulze and Whiteley, 2012 [34]) Given a tensegrity framework (G, p) in Rd and
a corresponding framework (G ∗ o, q) on Sd , then (G, p) is an infinitesimally rigid tensegrity frame-
work in Rd if and only if (G ∗ o, q) is an infinitesimally rigid tensegrity framework on Sd .

We note that Figure 17.7 (b) and (e) give a sample of how the cables and struts of a tensegrity
framework are transformed in a projective transformation.

17.3 Projective Frameworks

When we look again at the matrix RSd (G, p) for a Spherical framework (G, p), we can view it
through a projective lens. We are now working in Pd and considering the rigidity matrix for a
configuration p̌i ∈ Pd .

i j
 
{i, j} 0 ... 0 p̌ j 0 ... 0 p̌i 0 ... 0
 .. .. 

 . . 

RPd (G, p̌) = i  0 ... 0
 p̌i 0 ... 0 0 0 ... 0
 .. .. 
 . . 
j 0 ... 0 0 0 ... 0 p̌ j 0 ... 0

In this projective space, we can apply scaling to individual vertices, replacing p̌i with αi p̌i where
αi 6= 0, and preserve the rank of the rigidity matrix:
- multiply the columns for i by α1i ;
- the rows for edges {i, j} by αi ; and
- the row for p̌i by αi2 .
The result is the desired matrix for the scaled framework:
i j
 
{i, j} 0 . . . 0 p̌ j 0 . . . 0 αi p̌i 0 ... 0
 .. .. 

 . . 

RPd (G, p̌) = i  0 . . . 0 αi p̌i
 0 ... 0 0 0 ... 0.
 .. .. 
 . . 
j 0 ... 0 0 0 ... 0 p̌ j 0 ... 0

Notice that this matrix operation includes inversion in the sphere for p̌i , with αi = −1. However,
our focus is currently on the shared projective framework.
We can trace the impact of an arbitrary projective transformation with an invertible matrix
Td+1×d+1 by multiplying the entire rigidity matrix on the right by the block diagonal |V |(d + 1) ×
|V |(d + 1) matrix T with block entries Td+1×d+1 . This multiplication by an invertible matrix pre-
serves the rank of the rigidity matrix. Combined with the scaling above, we have a matrix form for
an arbitrary projective transformation.
In these combined operations, an equilibrium stress ωi j is transformed to αi1α j ωi j and ωi is
1
transformed to ω.
αi2 i
If we apply a Cayley algebra join ∨pi on the
P right to the columns associated to p̌i , the equilibrium
equation for an equilibrium stress becomes { j|{i, j}∈E} p̌ j ∨ p̌i = 0, since p̌i ∨ p̌i = 0. This is the
simple form of an equilibrium stress in terms of the 2-extensors for the forces at p̌i [12]. (See
362 Handbook of Geometric Constraint Systems Principles

Chapters 3 and 4 for Cayley algebra background.) Figure 17.7 (b) and (d) illustrates how the signs
for an equilibrium stress change under a projective transformation in the plane. The idea is that
when the projective transformation applies a negative weight to one end, but not the other, of an
edge, the sign of the equilibrium stress (compression or tension) is changed.

u2 M1 M6
1 u6
1 M2 u1
2 6
2 6

line to infinity

u3 u5
M3

M5
3 5 3 5 u4
4 4

(a) (b) (c)

u5 u3
M5 M3

u4 M4
5 3 5 3
4 4

line from infinity u2 M1

M2 u6
M6
1 1
u1
2 6
2 6

(d) (e) (f)

Figure 17.7
The framework with graph K3,3 on a circle (a) has an equilibrium stress with tension on dotted bars
and compression on solid bars (b). In (c) we see a non-trivial infinitesimal flex and the correspond-
ing (perpendicular) projective momenta. After a projective transformation, the framework lies on
a hyperbola (d) with the transformed equilibrium stress pattern (e) and the transformed projective
momenta and corresponding perpendicular infinitesimal flexes (f).

The projective interpretation of the infinitesimal flexes is less well known but it provides signif-
icant geometric insight into how the infinitesimal flexes transform.
We begin with the projective re-presentation of infinitesimal flexes in Rd as solutions of the
matrix equation RPd (G, p̌)M = 0. It is implicit in the coning theorems, that there is an isomorphism
of the solutions to RRd (G, p)v = 0 and RPd (G, p̌)M = 0 which fix the origin, We now give an explicit
correspondence, and interpret the entries Mi , as hyperplane coordinates in Pd for each vertex. The
equation
h p̌i , Mi i = p̌i · Mi = 0
says that Mi represents a hyperplane through the projective point p̌i .
Change of Metrics in Rigidity Theory 363

Proposition 17.1
For a framework (G, p) in Rd and the corresponding projective (affine) framework (G, p) with
weight (last coordinate) 1 in Pd , (. . . , ui , . . .) is an infinitesimal flex of (G, p) if and only if
(. . . , Mi , . . .) = (. . . , (ui , −ui · pi ), . . .) is a projective motion.

Proof. For every edge {i, j} we find the correspondence:

(pi − p j ) · (ui ) + (p j − pi ) · v j = (pi − p j ) · (ui ) + (p j − pi ) · v j = 0 (17.9)

⇐⇒ [(p j ui − pi ui ) + (pi v j − p j v j )] = h p̌i , Mi i + h p̌i , M j i = 0. (17.10)

Combined with the observation above, extended we confirm,

h p̌i , Mi i = (pi , 1) · (ui , −pi ui ) = pi ui − pi ui = 0

This gives the full correspondence for configurations in Rd . 

We call the hyperplanes Mi the momentum of the point. See Figure 17.7 (c), (f) for the pairings
of ui and Mi . If we change p̌i to α p̌i then tracking the change through the matrix operations above,
transforms Mi to α1i Mi . Above, a projective transformation is applied by right multiplication of the
rigidity matrix by T . Left multiplication by T −1 preserves the solution space of M.
In Figure 17.7, the original infinitesimal flexes for the circle are perpendicular to the circle,
inward for 1, 3, 5 and outward for 2, 4, 6, and the momenta are therefore tangent to the circle, but with
different “signs.” After transformation, the momenta remain tangent to the conic and the rescaled
infinitesimal flexes are perpendicular to the new conic. This is a sample of how this representation
provides additional geometric insight.
When transforming frameworks and infinitesimal flexes to other metrics, we can continue to fol-
low the momenta, and then use the inverse transformation to corresponding normals to the momenta
–using the orthogonality of the new metric.
A valuable exploration of the underlying projective geometry of bar and joint frameworks is the
study of pure conditions which present projective polynomials to capture which realizations of a
generically isostatic framework are singular (lowering the rank of the rigidity matrix) [37]. These
polynomials apply across all of the metrics being explored here (see Chapter 4).

17.4 Pseudo-Euclidean Geometries

Pseudo-Euclidean space: Consider Rd as equipped with a pseudo Euclidean metric. That is we
define the length of a vector w ∈ Rd as
s
X d
X
kwk2 = − w2i + w2i .
i=1 i=s+1

This defines a Pseudo-Euclidean space Fd . Note also that the case when s = 0 is the usual
Euclidean case we have already discussed. These are sometimes called Minkowski Spaces, and
d , where t = d − s, see [34].
written as Mt,s

Minkowski space: Minkowski space Md is the Pseudo-Euclidean space with s = 1.

364 Handbook of Geometric Constraint Systems Principles

Spheres in Pseudo Euclidean space: Similarly, for x, y ∈ Rd+1 , let hx, yik denote the function
hx, yik = x1 y1 + · · · + xd−k+1 yd−k+1 − xd−k+2 yd−k+2 − · · · − xd+1 yd+1 ,
d denote the set,
and let Xc,k
d
Xc,k = {x ∈ Rd+1 : hx, xik = c},
d
for some constant c 6= 0 and k ∈ N. The space X−1,1 is the d-dimensional Hyperbolic space,
d d
which we denote by H . The space X1,1 is the d-dimensional de Sitter space which we denote
by Dd . (Note that Spherical space Sd is the case when k = 0, c = 1.) Some literature asks that
xd+1 > 0, corresponding to the upper hemisphere [32], but we choose to be more general, using
inversion to complete the picture [34].

17.4.1 Hyperbolic and Minkowski Spaces

Now we consider the transfer of rigidity results for Hyperbolic and Pseudo-Euclidean (in particular
Minkowski) spaces. More details on non-Euclidean, or Cayley-Klein, geometries may be found in
[36].
Define φ 0 : Fd → Sd>0 by φ 0 (x) = kxk
x
.

Theorem 17.8 (Saliola and Whiteley, 2007 [32]) A bar-joint framework (G, p) is infinitesimally
rigid in Fd if and only if (G, φ 0 ◦ p) is infinitesimally rigid on Sd>0 .
Recall Figure 17.6. This shows the map from a Spherical framework to Hyperbolic and cor-
responding infinitesimal flexes. Similarly to the Euclidean case, we can cone in Pseudo-Euclidean
space to pass to any of spheres (in particular to Hyperbolic and de Sitter spheres).
Theorem 17.9 (Saliola and Whiteley, 2007 [32]) A bar-joint framework (G, p) is infinitesimally
rigid in Hd if and only if the corresponding Spherical framework is infinitesimally rigid on Sd>0 .
We may also consider “points at infinity” in the Hyperbolic geometry: points on the absolute,
a d-sphere in the Klein model in Rd+1 . Since we are coning from the sphere, in principle through
the affine plane xn+1 = 1 which has a Euclidean metric even in Minkowski space, the “absolute” is
not special in Minkowski space, though the metric along rays of the cone has all distances = 0. This
should not cause major changes in the rigidity of the framework.
It is, however, an open problem to extend the results of [14] to frameworks in the de Sitter space
Dd with points on the “equator” xn+1 = 0. We anticipate that the methods and results of [14] will
extend with appropriate adjustments.
We can also consider tensegrity frameworks in Hyperbolic space.
Theorem 17.10 (Schulze and Whiteley, 2012 [34]) Given a tensegrity framework (G, p) in Rd
and a corresponding framework (G ∗ o, q) in Hd , then (G, p) is an infinitesimally rigid tensegrity
framework in Rd if and only if (G ∗ o, q) is an infinitesimally rigid tensegrity framework in Hd .

17.5 Transfer of Symmetric Infinitesimal Rigidity

S-symmetric graph: An action of a group S on G is a group homomorphism θ : S → Aut(G),
where Aut(G) denotes the automorphism group of the graph G. If S acts on G by θ , then we
say that the graph G is S-symmetric (with respect to θ ).
Change of Metrics in Rigidity Theory 365

S-gain graph: For an S-symmetric graph G = (V, E) with a free action θ , the S-gain graph is an
oriented labelled quotient graph G/S. The quotient is the multi-graph (possibly with loops)
which has the set V /S = {Si : i ∈ S} of vertex orbits as its vertex set and the set E/S = {Se :
e ∈ E} of edge orbits as its edge set. An edge orbit connecting Si and S j can be written as
{(θ (γ)(i), θ (γ) ◦ θ (α)( j)) : γ ∈ S} for a unique α. For each such edge orient it from Si to S j
and assign the label α.
S-symmetric framework: Let S be an abstract group and G = (V, E) be an S-symmetric graph with
respect to an action θ : S → Aut(G). Suppose also that S acts on Rd via the homomorphism
τ : S → O(Rd ), where O(Rd ) denotes the orthogonal group. Then, we say that a framework
(G, p) is S-symmetric (with respect to θ and τ) if:
τ(x)(p(v)) = p(θ (x)v) for all x ∈ S and all v ∈ V.

Forced symmetric infinitesimal flex: Let S be a symmetry group. An infinitesimal flex u : V → Rd

on an S-symmetric graph (G, p) is a forced symmetric infinitesimal flex if x(ui ) = ux(i) for all
i ∈ V and x ∈ S.
Forced symmetrically rigid: An S-symmetric framework is forced symmetrically infinitesimally
rigid if every forced symmetric infinitesimal flex is trivial.
S-regular: A symmetric framework (G, p) in Rd (resp. on Sd ) is S-regular if the rank of the rigidity
matrix Rd (G, p) (resp. RSd (G, p)) is maximal over all S-symmetric configurations q ∈ Rd|V |
(resp. q ∈ Sd|V | ).

17.5.1 Symmetric Frameworks

Let S be a symmetry group, let x ∈ S and let Fx denote the linear subspace of Rd which consists of
all points a ∈ Rd with x(a) = a. If (G, p) is S-symmetric then pi lies in the subspace
\
U(pi ) = Fx .
x∈S:x(pi )=pi

For given bases Bi and B j of U(pi ) and U(p j ), let Mi and M j be the matrices whose columns are
the coordinate vectors of Bi and B j relative to the canonical basis of Rd .
Let (G, p) be an S-symmetric framework in Rd which has no joint that is “fixed” by a non-trivial
symmetry operation in S. Further, let (G0 , ψ) be the quotient S-gain graph of (G, p). For each edge
e ∈ E(G0 ), the orbit rigidity matrix O(G, p, S) of (G0 , p) is the |E(G0 )| × d|V (G0 )| matrix of the
form

i j
 .. 
.
p j − x−1 (pi ) M j



{i, x( j)}  pi − x(p j ) Mi

 
 0 ... 0 0 ... 0 0 ... 0
.
{i, x(i)}  0 ... 0 2pi − x(pi ) − x−1 (pi ) Mi 0 ... 0 0 0 ... 0

..
.
We refer the reader to Chapter 25 and [34] for more details. Similarly, for an S-symmetric Spher-
ical framework (G, p) with p : V → Sd , we can also write down a standard Spherical orbit matrix.
This is the (|E(G0 )| + |V (G0 )|) × (d + 1)|V (G0 )| matrix OS (G, p):
366 Handbook of Geometric Constraint Systems Principles

i j
 .. 
.
p j − x−1 (pi ) M j



{i, x( j)} 

0 ... 0 pi − x(p j ) Mi
0 ... 0 0 ... 0
2pi − x(pi ) − x−1 (pi ) Mi

{i, x(i)} 0 ... 0 0 ... 0 0 0 ... 0

i 0
 ... 0 pi 0 ... 0 0 0 ... 0
.
 
 
j 0
 ... 0 0 0 ... 0 pj 0 ... 0

..
.
Our next theorem can be proved by symmetrising the coning transformation from the Euclidean to
the Spherical setting we have already presented, see [34] for details.

Theorem 17.11 [34] Let q be a configuration of points on Sd>0 such that the projection π(q) from
the centre of Sd onto Rd (via the affine plane) satisfies π(q) = p. Let S∗ be a symmetry group of Sd>0
and let S be the corresponding symmetry group (under π) of Rd . Then:
• the space of S-symmetric infinitesimal flexes of (G, p) in Rd is isomorphic to the space of
S∗ -symmetric infinitesimal flexes of (G, q) on Sd>0 ;

• the space of S-symmetric equilibrium stresses of (G, p) in Rd is isomorphic to the space of

S∗ -symmetric equilibrium stresses of (G, q) on Sd>0 .

We also note that the transfer takes an S-regular configuration p in Rd to an S∗ -regular config-
uration q on Sd . With this, symmetric flexes also transfer from Rd to Sd through coning. See [34,
Section 4.1] for details and also cautionary examples.
We mention that the transfer described in Theorem 17.11 was geometric and hence, noting that
the transfer operations preserve the symmetry group (as long as the group exists in both Spherical
and Euclidean space), gives us the transfer for incidentally symmetric frameworks. One can also
consider the transfer at a more detailed level. In particular Schulze and Tanigawa developed a block
decomposition of the rigidity matrix (see Chapter 25 and [33]). With this, we expect that the rank
preserving matrix operations can be applied to each block of the decomposition of the rigidity matrix
for incidental symmetry.
In Figure 17.8 we give an example, adapted from Figure 17.7, to illustrate how the analysis
works, even with fixed vertices and fixed edges.
In this subsection we explain how the transfer of infinitesimal rigidity extends to symmetric
frameworks. When this is done, then the general transfer of infinitesimal rigidity can be thought of
as the special case when the symmetry group is the identity.
In fact this symmetric transfer extends to Pseudo-Euclidean spaces and to Hyperbolic and de
Sitter spaces [34] whenever the symmetry group exists. The basic examples of symmetry groups
that exist throughout the range of Pseudo-Euclidean geometries are the reflection and half turn
rotation groups. It is an open problem, currently being investigated in [4], to confirm that this transfer
extends to include points on the equator of Sd (points at infinity). We also mention that the results
in Theorem’s 17.6 and 17.7 were extended to symmetric tensegrities in [34].

17.6 Global Rigidity

Globally rigid framework: A bar-joint framework (G, p) in Rd is globally rigid if every equiva-
lent framework (G, q) arises from (G, p) by a congruence of Rd .
Change of Metrics in Rigidity Theory 367

1 b 1
a a
3
4 2
3 2
a
Mirror1 2
Half-turn Center
3 1

a 1 1 a b
Mirror2 b 4
(a) (b) (c) (d)

Figure 17.8
A framework, with underlying graph K3,3 , with mirrors and a half-turn (a) has two vertex orbits
a, b and four edges orbits (b). The corresponding gain graph (c) has the count of 3 columns (2 for
a, 1 for b fixed on the mirror) and 4 rows (edge orbits). Since 4 > 3 there is a fully symmetric
equilibrium stress (d). The corresponding infinitesimal flex in Figure 17.7 is not symmetric, and
the framework is rigid and globally rigid [11]. Note this analysis also applies without change to the
projected version in Figure 17.7.

Globally rigid graph: A graph is globally rigid in Rd if all generic realizations (G, p) are globally
rigid in Rd .
Quasi-generic Spherical framework: A framework on the sphere is quasi-generic if it is the cen-
tral projection of a generic framework in Rd+1 .
Universally rigid: A framework (G, p) in Rd is universally rigid if every equivalent framework
(G, q) in RD for any D ≥ d arises from (G, p) by a congruence.
Dimensionally rigid: A framework (G, p) in Rd is dimensionally rigid if there are no equivalent
frameworks with a higher dimensional affine span.

Pmatrix Ω is the n × n symmetric matrix with off-diagonal entries −ωi j

Stress matrix: The stress
and diagonal entries j ωi j .

Conic at infinity: A finite set of (non-zero) vectors in Rd lie on a conic at infinity if when regarded
as points in projective d − 1-dimensional space, they lie on a conic.

Super stable: A framework (G, p) in Rd with a full dimensional affine span is called super stable if
it has an equilibrium stress matrix Ω which is PSD with maximum rank and its edge directions
do not lie on a conic at infinity.
Slicing: Given a cone framework (G, p) in Rd+1 , the process of slicing means to slide all points
of p to the affine plane and then consider the resulting subframework (without the cone) as a
framework in Rd .

Let us now consider the effect of changing the metric on the global rigidity of the corresponding
frameworks (see also [23]). We have already seen that equilibrium stresses transfer, and it is clear
that these necessary conditions also transfer. Hendrickson gave necessary conditions for the global
rigidity of a framework G(p) in dimension d: (i) the graph is (d + 1)-connected and (ii) G is redun-
dantly rigid in Rd . In dimensions d = 1 and d = 2 these are also sufficient in Rd . The conditions are
well-known not to be sufficient for d > 3.
It follows from the main result of [16] that global rigidity in Euclidean spaces is a generic
property. That is, if some generic realization (G, p) of G is globally rigid in Rd then all generic
368 Handbook of Geometric Constraint Systems Principles

(a) (b)

Figure 17.9
(a) A globally rigid realization of the 4-cycle on the line is coned to a framework in R2 that is not
globally rigid. The angles α = 60◦ , β = 120◦ , and the other two angles at the cone vertex are 90◦
induce a second realization. (b) An equivalent but non-congruent realization.

realizations (G, q) in Rd are also globally rigid. A fundamental result of Connelly and Whiteley
transfers this to Spherical frameworks, which also appears in the work of Pogorelov [27, 28].
Theorem 17.12 (Connelly and Whiteley, 2010 [6]) A graph G is generically globally rigid in Rd
if and only if G is quasi-generically globally rigid on the sphere Sd .
To understand this theorem lets start by considering coning. It is not hard to see that this process
takes a redundantly rigid framework in Rd to a redundantly rigid framework in Rd+1 and takes a
d-connected graph to a d + 1 connected graph. Thus Hendrickson’s necessary conditions transfer,
and the sufficiency for d = 1 and d = 2 transfers to the sphere, and also to H2 and D2 (private
communication from Stephen Gortler). One of the key ideas to the stronger transfer of global rigidity
is the averaging technique which is implicit in the work of Pogorelov [27, 28], and appears in [35] .
Theorem 17.13 (Connelly and Whiteley, 2010 [6]) Given a globally rigid and infinitesimally
rigid framework (G, p) in Rd , there exists an open neighborhood N p such that for any q ∈ N p ,
the framework (G, q) is globally rigid and infinitesimally rigid in Rd .
The other key idea is the stress matrix characterisation of generic global rigidity given in [7, 16]
(see Chapter 16). In particular the fact that projective transformations preserve infinitesimal rigidity
is extended in [6] to show that projective transformations also preserve the property of having a
maximum rank stress matrix. For coning, in particular, we have the following result.
Theorem 17.14 (Connelly and Whiteley, 2010 [6]) Let (G, p) and (G ∗ o, p̄∗ ) be frameworks in
Rd of full affine span. Then (G, p) has a stress matrix Ω with rank |V |−d −1 if and only if (G∗o, p̄∗ )
has a stress matrix Ω∗ with rank |V | − d − 1.
By combining this result with [7, 16] we have the following.
Corollary 17.15 A framework (G, p) in Rd is globally rigid if and only if the cone framework
(G ∗ o, p̄∗ ) in Rd+1 is globally rigid.
The example in Figure 17.9 is based on a cycle (generically globally rigid on the line) whose
cone is (generically) globally rigid in the plane. However the non-generic realization of the cycle
results in an equivalent but non-congruent realization.

17.6.1 Universal Rigidity

While global rigidity in Euclidean space is a generic property, universal rigidity is not. Even on the
real line a cycle Ck , with k ≥ 4, may be universally rigid or flexible in R2 depending on the ordering
of the vertices [22].
Change of Metrics in Rigidity Theory 369

In the next three results we present coning for dimensional rigidity, super stability and universal
rigidity [10].

Theorem 17.16 The framework (G, p) is dimensionally rigid in Rd if and only if the cone frame-
work (G ∗ o, p̄∗ ) is dimensionally rigid in Rd+1 .

Theorem 17.17 The framework (G, p) is super stable in Rd if and only if the cone framework
(G ∗ o, p̄∗ ) is super stable in Rd+1 .

Theorem 17.18 If the framework (G, p) is super stable in Rd then the cone framework (G ∗ o, p̄∗ )
is universally rigid in Rd+1 . If the cone framework (G ∗ o, p̄∗ ) is universally rigid in Rd+1 then the
framework (G, p) is dimensionally rigid in Rd .

The only failure for projecting a universally rigid cone framework to a universally rigid frame-
work can come from the appearance of affine flexes due to the projection having member directions
on a conic at infinity. If, for example, the framework (G, p) is in general position, this cannot happen,
as was observed by Alfakih and Ye [2].

Lemma 17.2 [8]

Let (G ∗ o, p̄∗ ) be a cone framework in Rd+1 with an equilibrium stress matrix Ω. Then any frame-
work obtained from (G ∗ o, p̄∗ ) by sliding has an equilibrium stress matrix with the same rank and
signature as Ω.

In fact sliding, as we illustrated in Figure 17.1, preserves global rigidity. One of the key aspects
here is how the equilibrium stress changes under sliding the framework. When the framework is
slid to a hyperplane then a new equilibrium stress emerges from the flatness, but the cone edges
simultaneously become unstressed. In [8] it is also shown that super stability is invariant under
coning and slicing. A consequence being that projective transformations that do not send points to
infinity preserve super stability.
It is also worth pointing out explicitly that inversion is a special case of sliding so the above
discussion implies that inversion preserves global rigidity, superstability and universal rigidity.

17.6.2 Projective Transformations

A general projection in Rd can be formed by coning to Rd+1 , rotating, and re-projecting, perhaps
several times. As a result, the coning results guarantee the projective invariance of key properties.

Theorem 17.19 A framework (G, p) is super stable (resp. dimensionally rigid) in Rd if and only if
every invertible projective transformation which keeps vertices finite is super stable (resp. dimen-
sionally rigid).

This projective invariance is almost always true for universal rigidity. The key is whether the
directions of the members meet the set X which is going to infinity, in a conic.

Theorem 17.20 ([9]) If a framework (G, p) is universally rigid in Rd , and X is a hyperplane avoid-
ing all vertices, such that the members do not meet it in a conic, then any invertible projective
transformation T which takes X to infinity makes (G, T (p)) universally rigid.

An illustrative example is given in [9, Figures 8 and 9]. Note that this theorem also extends to
tensegrity frameworks, [9, Theorem 13.1].
Alfakih and Nguyen [1] showed that having a maximum rank PSD stress matrix is a sufficient
condition for dimensional rigidity. When a nested sequence of affine spaces, with a correspond-
ing sequence of PSD stress matrices, is used to iteratively demonstrate the dimensional rigidity of a
370 Handbook of Geometric Constraint Systems Principles

given framework [9], the projected framework is also dimensionally rigid, since the projected frame-
work has the projected affine sequence to demonstrate its dimensional rigidity. It is well known that
there are special positions for which global rigidity is not projectively invariant. In fact it is not even
an affine invariant, see [6, Figure 4 and Example 8.3]. However, the rank of any stress matrix Ω on
the framework is invariant under projective transformations [6].

17.6.3 Pseudo-Euclidean Metrics

Global rigidity has a natural extension to all the Pseudo-Euclidean metrics and their spheres. There
are two distinct ways that global rigidity of G can be “generic” within a metric: (i) there are some
generic configurations at which (G, p) is globally rigid; and (ii) if (G, p) is globally rigid at one
generic configuration, then (G, q) is globally rigid for all generic q. We use the language generically
globally rigid for the second sense. In arbitrary Pseudo-Euclidean spaces we have the following
result of Gortler and Thurston [15].

Theorem 17.21 A graph G is generically globally rigid in Rd if and only if it is globally rigid for
some generic configurations in Fd .

However this does not complete the story. It is not clear if global rigidity is a generic property
in Pseudo-Euclidean spaces. It is possible that there are graphs which are not generically globally
rigid but nevertheless do have some generic realizations which are globally rigid. A partial result
given in [15] resolves this positively when G has a (generically) globally rigid subgraph on at least
d + 1 vertices.
Here Hyperbolic space was modelled as the “’-1” sphere in Minkowski space. More formally
Minkowski space Md+1 is the (d + 1)-dimensional Pseudo-Euclidean space with one negative coor-
dinate in its signature. We then model Hyperbolic space as the vectors x ∈ Md+1 such that |x|2 = −1
in the Minkowski metric.
It is an open problem to understand the transfer of universal rigidity for Pseudo-Euclidean
spaces. Generic universal rigidity was characterised in Euclidean spaces in [17] as graphs which
admit maximum rank PSD stress matrices. We have already seen how maximum rank equilibrium
stresses transfer.

17.7 Summary and Related Topics

We conclude the chapter with Table 17.1 summarizing the key results we have presented and then
by briefly mentioning a number of related pieces of work. The table reveals a number of problems
that have not been studied in detail.
As we assembled the table we recognized some gaps that have not been carefully explored for
a transfer of metric. One example is the concepts of second-order rigidity and pre-stress stability
[11]. Second-order rigidity is not projectively invariant, and will only be of interest in the presence
of some equilibrium stress. There is some evidence these results will extend geometrically to the
Spherical metric through coning, but we are not aware of work transferring these concepts carefully
to the Spherical metric or to other metrics. Even dimensional rigidity, and universal rigidity, have
not been carefully defined for other metrics. Though they appear amendable to generalizations and
transfers, with appropriate definitions.
We have discussed that infinitesimal rigidity, as a projective invariant, is equivalent in the
Cayley-Klein metrics (Euclidean, Spherical, Minkowski, Hyperbolic spaces, etc.). Let us make two
comments on the limitations of this approach. First, one may wonder if the alternative viewpoint
Change of Metrics in Rigidity Theory 371

Table 17.1
We compare the properties of Euclidean frameworks with frameworks in other geometries. “Yes”
refers to the equivalence of the property in the given geometry with the Euclidean case and “no”
refers to the non-equivalence.
Type of Nature Cone Spherical Pseudo-Euclidean Hyperbolic
rigidity of transfer Sd Fd Hd
Infinitesimal geometric Yes, [32] Yes, [32] Yes, [32] Yes, [32]
Rigidity generic Yes Yes Yes Yes
S-symmetric geometric Yes, [34] Yes, [34] Yes, [34] Yes, [34]
Super stability geometric Yes, [10] ? ? ?
Dimensional geometric Yes, [10] ? ? ?
Global generic Yes, [6] ? ? Yes d = 2
Universal geometric No, [23] No, [23] ? ?

these other geometries provide can help with conjectures about Euclidean frameworks. We know of
no results in this direction. Second, one may consider more abstract metric spaces. In general the
transfer processes we have described break down since such spaces have different numbers of “triv-
ial” infinitesimal flexes and hence different combinatorics required for rigidity. Rigidity in various
other metric space contexts have been considered [24, 25, 26], see also Chapter 24.
We now mention some relevant work in progress [4]. A half turn symmetric framework in Rd
is projectively equivalent to a mirror reflection, in all dimensions, which guarantees all infinitesi-
mal and static properties are preserved. This same projective transformation applies in all Pseudo
Euclidean spaces and will preserve all infinitesimal rigidity properties. This projective transforma-
tion also shows that forced-symmetric rigidity is preserved under the transformation. The reader
may like to compare and contrast this with combinatorial results for symmetric frameworks given
in Chapter 25. In [4] more general pairings of symmetry groups with equivalent rigidity properties
are explored.
Saliola and Whiteley [31] analysed arrangements of lines and circles in the plane, and more
generally arrangements of hyperplanes and hyperspheres in any dimension d. This resulted in a
direct, geometric correspondence to points and distances in Dd+1 , which preserves infinitesimal
rigidity and flexibility. This transformation relies on stereographic projection and lifting between
the Euclidean space and the sphere, which takes spheres and hyperplanes into spheres on Sd+1 and
preserves angles. This correspondence is interesting, as it embeds the entire infinitesimal theory
of de Sitter space into a set of geometric constraints in a lower dimensional Euclidean space, as
well as giving an example where we have a clear analysis for angles. See [31] for further results
and comments, including the connections of Cauchy’s Theorem on the infinitesimal rigidity for
convex polyhedra in all of these spaces, along with Andreev’s Theory [3] for angles in polyhedra in
Hyperbolic space.
Since the basic infinitesimal rigidity is projectively invariant, this invites exploration of the im-
pact of polarity. The situation is very different for Spherical, Hyperbolic, and de Sitter spaces, where
points and distances polarize to hyperplanes and angles, and the setting of Euclidean space. In R3
polarity connects bar-joint frameworks to a form of infinitesimal rigidity for interesting structures
call sheet works [41]. It is now natural to ask whether this same correspondence appears in M3 ?
There is a related special correspondence of first-order rigidity in R2 , with a lifting matroid of lines
in R2 [44]. Again, we now propose extending this to M2 and liftings into M3 .
Izmestiev [18] provides an alternative viewpoint, with new proofs, on the projective theory of
rigidity and in [19] constructed examples of infinitesimally flexible Hyperbolic cone manifolds.
Much of the work on body-hinge frameworks is presented in projective form (e.g. [12]). In
addition the work on the algebraic geometry of infinitesimal body-bar frameworks [38] is presented
372 References

in a completely projective form, that will transfer across metrics. In addition, the inductive proof
of rigidity for body-bar frameworks in Rd carries over directly to the Spherical metric, and these
appear in a chapter on CAD frameworks in [13]. We note that for body-bar frameworks there is an
efficient combinatorial algorithm for generic global rigidity in all dimensions [5], something that
does not exist for bar and joint frameworks.
Further relevant work on cones and buildings was considered in [40], where coning is presented
with projective algebra, and the cone-point at infinity. An extensive early presentation of both the
statics and infinitesimal theory, with many worked examples, is available on the web at [39].
Turning the infinitesimal velocities 90 degrees in R2 has a second interpretation as the vectors of
a parallel drawing in R2 [35]. This theory of parallel drawings has a number of analogous properties
to first-order rigidity: (i) it generalizes to all dimensions with a rigidity-like matrix; (ii) the rank
of the matrix is projectively invariant; and (iii) it transfers to all metrics which have the full space
of translations: the Pseudo-Euclidean metrics, including Minkowski space. In addition, this theory
has a dual theory of projections and liftings, and stronger combinatorial algorithms [42]. Since
these theories are already fully linear, infinitesimal rigidity and global rigidity coincide for parallel
drawing and scene analysis.

References
[1] A. Y. Alfakih and Viet-Hang Nguyen. On affine motions and universal rigidity of tensegrity
frameworks. Linear Algebra Appl., 439(10):3134–3147, 2013.
[2] A. Y. Alfakih and Yinyu Ye. On affine motions and bar frameworks in general position. Linear
Algebra Appl., 438(1):31–36, 2013.
[3] E. M. Andreev. Convex polyhedra in Lobačevskiı̆ spaces. Mat. Sb. (N.S.), 81 (123):445–478,
1970.
[4] Katharine Clinch, Anthony Nixon, Bernd Schulze, and Walter Whiteley. Pairing symmetries
for projective and spherical frameworks. in preparation, 2018.
[5] R. Connelly, T. Jordán, and W. Whiteley. Generic global rigidity of body-bar frameworks. J.
Combin. Theory Ser. B, 103(6):689–705, 2013.
[6] R. Connelly and W. J. Whiteley. Global rigidity: the effect of coning. Discrete Comput.
Geom., 43(4):717–735, 2010.
[7] Robert Connelly. Generic global rigidity. Discrete Comput. Geom., 33(4):549–563, 2005.
[8] Robert Connelly, Steven Gortler, and Louis Theran. Affine rigidity and conics at infinity.
arXiv: 1605.07911, 2016.
[9] Robert Connelly and Steven J. Gortler. Iterative universal rigidity. Discrete Comput. Geom.,
53(4):847–877, 2015.
[10] Robert Connelly and Steven J. Gortler. Universal rigidity of complete bipartite graphs. Dis-
crete Comput. Geom., 57(2):281–304, 2017.
[11] Robert Connelly and Walter Whiteley. Second-order rigidity and pre-stress stability for
tensegrity frameworks. SIAM J. on Discrete Methods, 9:453–492, 1986.
[12] Henry Crapo and Walter Whiteley. Statics of frameworks and motions of panel structures,
a projective geometric introduction. Structural Topology, 6:43–82, 1982. With a French
translation.
References 373

[13] Yaser Eftekhari. Geometry of point-hyperplane and spherical frameworks. PhD thesis, York
University, 2017.
[14] Yaser Eftekhari, Bill Jackson, Anthony Nixon, Bernd Schulze, Shin-Ichi Tanigawa, and Walter
Whiteley. Point-hyperplane frameworks, slider joints, and rigidity preserving transformations.
arXiv: 1703.06844, 2017.
[15] Steven Gortler and Dylan Thurston. Generic global rigidity in complex and pseudo-euclidean
spaces. In Rigidity and Symmetry. Fields Institute Communications, 2014.
[16] Steven J. Gortler, Alexander D. Healy, and Dylan P. Thurston. Characterizing generic global
rigidity. Amer. J. Math., 132(4):897–939, 2010.
[17] Steven J. Gortler and Dylan P. Thurston. Characterizing the universal rigidity of generic
frameworks. Discrete Comput. Geom., 51(4):1017–1036, 2014.
[18] Ivan Izmestiev. Projective background of the infinitesimal rigidity of frameworks. Geom.
Dedicata, 140:183–203, 2009.
[19] Ivan Izmestiev. Examples of infinitesimally flexible 3-dimensional hyperbolic cone-
manifolds. J. Math. Soc. Japan, 63(2):581–598, 2011.
[20] Bill Jackson and Tibor Jordán. Rigid two-dimensional frameworks with three collinear points.
Graphs Combin., 21(4):427–444, 2005.
[21] Bill Jackson and J. C. Owen. A characterisation of the generic rigidity of 2-dimensional
point-line frameworks. J. Combin. Theory Ser. B, 119:96–121, 2016.
[22] Tibor Jordán and Viet-Hang Nguyen. On universally rigid frameworks on the line. Contrib.
Discrete Math., 10(2):10–21, 2015.
[23] Tibor Jordan and Walter Whiteley. Global rigidity. In Csaba D Toth, Joseph O’Rourke, and
Jacob E Goodman, editors, Handbook of discrete and computational geometry 3rd Edition.
Chapman and Hall/CRC, 2018.
[24] D. Kitson and S. C. Power. Infinitesimal rigidity for non-Euclidean bar-joint frameworks.
Bull. Lond. Math. Soc., 46(4):685–697, 2014.
[25] A. Nixon, J. C. Owen, and S. C. Power. Rigidity of frameworks supported on surfaces. SIAM
J. Discrete Math., 26(4):1733–1757, 2012.
[26] Anthony Nixon and Stephen Power. Double-distance frameworks and mixed sparsity graphs.
arXiv: 1709.06349, 2017.
[27] A. V. Pogorelov. Topics in the theory of surfaces in elliptic space. Translated by Royer and
Roger, Inc. Edited and with a preface by Richard Sacksteder. Russian Tracts on Advanced
Mathematics and Physics, Vol. I. Gordon and Breach, New York, 1961.
[28] A. V. Pogorelov. Extrinsic geometry of convex surfaces. American Mathematical Society,
Providence, R.I., 1973. Translated from the Russian by Israel Program for Scientific Transla-
tions, Translations of Mathematical Monographs, Vol. 35.
[29] W. J. M. Rankine. On the application of barycentric perspective to the transformation of
structures. Phil. Mag., 26:387–388, 1863.
[30] B. Roth and W. Whiteley. Tensegrity frameworks. Trans. Amer. Math. Soc., 265(2):419–446,
1981.
[31] Franco Saliola and Walter Whiteley. Constraining plane configurations in CAD: circles, lines,
and angles in the plane. SIAM J. Discrete Math., 18(2):246–271, 2004.
374 References

[32] Franco Saliola and Walter Whiteley. Some notes on the equivalence of first-order rigidity in
various geometries. arXiv: 0709.3354, 2007.
[33] Bernd Schulze and Shin-ichi Tanigawa. Infinitesimal rigidity of symmetric bar-joint frame-
works. SIAM J. Discrete Math., 29(3):1259–1286, 2015.
[34] Bernd Schulze and Walter Whiteley. Coning, symmetry and spherical frameworks. Discrete
Comput. Geom., 48(3):622–657, 2012.
[35] Bernd Schulze and Walter Whiteley. Rigidity and scene analysis. In Csaba D Toth, Joseph
O’Rourke, and Jacob E Goodman, editors, Handbook of discrete and computational geometry
3rd Edition. Chapman & Hall/ CRC, 2018.
[36] Horst Struve and Rolf Struve. Non-Euclidean geometries: the Cayley-Klein approach. J.
Geom., 98(1-2):151–170, 2010.
[37] Neil White and Walter Whiteley. Algebraic geometry of stresses in frameworks. SIAM J. Alg.
Disc. Math., 4:53–70, 1983.
[38] Neil White and Walter Whiteley. The algebraic geometry of motions of bar and body frame-
works. SIAM J. Alg. Disc. Math., 8:1–32, 1987.
[39] Walter Whiteley. Introduction to structual topology.
http://wiki.math.yorku.ca/index.php/Resources in Rigidity Theory, 1978.
[40] Walter Whiteley. Cones, infinity and 1-story buildings. Structural Topology, 8:53–70, 1983.
With a French translation.
[41] Walter Whiteley. Rigidity and polarity. I. Statics of sheet structures. Geom. Dedicata,
22(3):329–362, 1987.
[42] Walter Whiteley. A matroid on hypergraphs, with applications in scene analysis and geometry
75-95. Disc. and Comp. Geometry, 4:75–95, 1988.
[43] Walter Whiteley. The union of matroids and the rigidity of frameworks. SIAM J. Discrete
Math., 1(2):237–255, 1988.
[44] Walter Whiteley. Rigidity and polarity. II. Weaving lines and tensegrity frameworks. Geom.
Dedicata, 30(3):255–279, 1989.
 Part IV

Combinatorial Rigidity
Chapter 18
Planar Rigidity
Brigitte Servatius
Mathematics Department, Worcester Polytechnic Institute, Worcester, MA

Herman Servatius
Mathematics Department, Worcester Polytechnic Institute, Worcester, MA

CONTENTS
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
18.1 Rigidity of Bar and Joint Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
18.1.1 Rigidity Matrix and Augmented Rigidity Matrices . . . . . . . . . . . . . . . . . . . . . . . . . 380
18.1.2 Rigidity Matrix as a Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
18.1.3 The Infinitesimal Rigidity Matroid of a Framework . . . . . . . . . . . . . . . . . . . . . . . . 384
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
18.2 Abstract Rigidity Matroids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
18.2.1 Characterizations of A2 and (A2 )⊥ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
18.2.2 The 2-Dimensional Generic Rigidity Matroid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
18.2.3 Cycles in G2 (n) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
18.2.4 Rigid Components of G2 (G) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 391
18.2.5 Representability of G2 (n) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
18.3 Rigidity and Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
18.3.1 Birigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
18.3.2 Tree Decomposition Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397
18.3.2.1 Computation of Independence in G2 (n) . . . . . . . . . . . . . . . . . . . . . . 398
18.3.3 Pinned Frameworks and Assur Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400
18.3.3.1 Isostatic Pinned Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400
18.3.4 Body and Pin Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
18.3.5 Rigidity of Random Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 408
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 408

Introduction
The consideration of geometric constraint systems as a general subject is a relatively recent devel-
opment, although many specific embodiments do have a long history. The oldest examples can be
expressed in the form of a bar and joint framework, considered by Watt, Maxwell, and Cremona in

377
378 Handbook of Geometric Constraint Systems Principles

the nineteenth century [10, 37]. Frameworks which have been most successfully studied are those
which lie in the plane, such as the mechanisms studied by Kempe [19].
In fact, many of the ruler and compass constructions of classical geometry, which go back to
antiquity, can be viewed in the context of bar and joint frameworks in the plane.

18.1 Rigidity of Bar and Joint Frameworks

We want to consider the rigidity of a framework F = ((V, E), p) in d-space with nodes corresponding
to the vertices V , length constraints corresponding to the edges E, and a placement of the vertices
into Euclidean space by p : V → R|V |d , particularly the case d = 2, the plane. For this considera-
tion, it is important to distinguish the different and related forms of rigidity. The most natural and
straightforward notion is local rigidity, or simply, rigidity. A framework is rigid if every placement
q : V → Rd sufficiently close to p and preserving the distances between the placements of adjacent
vertices, is related to p by a congruence of Rd .
Regarding p as a point in real |V |d-space, the rigidity function corresponding to the graph G =
(V, E) is the map ρ : R|V |d → R|E| where the coordinate of ρ(p) corresponding to e = (v, w) ∈ E is
kp(v) − p(w)k2 . We can now describe the set of all frameworks on G with edge lengths equivalent
to F as
C(F) = {(G, q) | ρ(q) = ρ(p)} (18.1)
Placements in C(F) correspond to the solutions of the constraint system consisting of |E| quadratic
equations
(p(v) − p(w))2 = (q(v) − q(w))2 (v, w) ∈ E, (18.2)
so C(F) is an algebraic set and it is called the configuration space of F. Clearly p ∈ C(F) and we
may describe a physical movement of the framework in Rd , that is, a movement of the vertices
which preserves the lengths of the edges, by a continuous path in C starting at p. Moreover, since
C(F) is an algebraic set, it follows that if the framework F is not rigid, then there is in fact a motion
of F, see [38]. For example, an equilateral triangle framework is rigid in the plane, while a square
framework is non-rigid, deformable through a family of rhombi.
In the 1970s, Bolker and Crapo [3, 4] developed a criterion for how to rigidify a grid of squares
in the plane using diagonal braces along a selection of the square diagonals. Given a braced grid,
they defined a bipartite graph whose vertex set is the set of rows and columns of the grid, and let row
i be adjacent to column j if there is a brace in the square (i, j). See Figure 18.1 for some examples
of braced grids and their brace graphs.

Theorem 18.1 (Bolker) A braced grid is rigid if and only if the bipartite graph of braces is con-
nected.

In [3] the circuits in the structure matroid of face diagonal braces for a grid of cubes in space are
characterized. In [49] necessary conditions for the bases are given and it is pointed out that these
conditions may be checked in polynomial time using Edmonds’ matroid partition algorithm [12].
A strictly stronger notion than rigidity is global rigidity, which requires that all edge distance
preserving placements of (V, E) be related to p by a congruence of Rd . For a simple example,
a square framework with a diagonal brace is rigid, but not globally rigid, since it has an alter-
nate placement folded into an isosceles right triangle with two vertices placed together. The square
framework with two diagonal braces, whose underlying graph is the complete graph on 4 vertices,
is globally rigid. Every placement of a complete graph is globally rigid, so every rigid framework
can be made globally rigid by adding a sufficient number of braces.
Planar Rigidity 379
c1 c2 c3
r1r cr1 r2r
r1
c1 c2 c3 c4 c5 @@
r ` rc1 r2
r1 r1 ` r r r
@
@ r3 @
r ` ```
`rc2
`
r2 @ @ r2 ` @ c3 r3 c2
r ` ```
`rc3
`
r3 @ @ r3 ` @

r4 r``` ```rc4
`` @
r4 @ @  A
 AA A
r5
@ @
r5 r `rc5
``` AA AA

@
@ @@ 

Figure 18.1
A rigid braced 4 × 4 grid with its 2-connected brace graph, so deleting any brace still leaves a
rigid grid. A braced 3 × 3 grid with disconnected brace graph and a deformation derived from its
components.

A more theoretically tractable notion is infinitesimal rigidity, which forbids not actual motions of
the vertices, but infinitesimal ones. Given a framework F, hence a fixed placement p, we have seen
that the constraint system for rigidity and global rigidity is the system of |E| quadratic equations
(18.2) whose derivative, evaluated at p, gives

(p(v) − p(w)) · (p0 (v) − p0 (w)) = 0 (v, w) ∈ E, (18.3)

a linear system of constraints on the initial velocities {p0 (v)} of any motion of F, geometrically
requiring that the velocities of an adjacent pair of vertices must have equal projections onto the
edge between them, see Figure 18.2a.

(a) (b) (c)

Figure 18.2
(a) The infinitesimal constraint on each edge. (b) and (c) Two globally rigid frameworks which are
not infinitesimally rigid.

It is common practice when moving to infinitesimal considerations to assume that the placement
is an embedding, a one-to-one map on the vertex set, so that none of the equations in (18.3) is
trivial; and that the framework contains a set of at least d vertices
in general position, so that the
system (18.3) always has a solution space of rank at least d+1 2 , corresponding to the linear space
I0 of infinitesimal isometries of d-space, the trivial solutions. A convenient basis for I0 consists
of the d vectors (ei
, ei , . . . , ei ), where ei is the i’th standard basis element of Rd (written as a row
vector), and the d2 vectors (Mi, j p(v1 ), Mi, j p(v2 ), . . . , Mi, j p(v|V | )), where Mi, j , 1 ≤ i < j ≤ d, is
a d × d skew symmetric matrix which has exactly two non-zero entries, a 1 in the (i, j) position,
and a −1 in position ( j, i), i = 1, . . . d. If the vectors e1 . . . ed are
the principle axes of the point set
p(v1 ) . . . p(v|V | ) and their centroid is at the origin then these d+1 2 vectors form an orthogonal basis
for I0 . A framework for which the elements of I0 are the only solutions is said to be infinitesimally
rigid.
380 Handbook of Geometric Constraint Systems Principles

It is not true that every framework which has a non-trivial motion, that is, a motion changing
the distance between at least one pair of non-adjacent vertices, has a motion whose initial velocities
are also non-trivial, see [7]. Nevertheless, since the solutions to (18.2) form an algebraic set, any
motion may be assumed to be analytic, and by considering higher derivatives the following key
result follows, see [17].
Theorem 18.2 If a framework in d-space is infinitesimally rigid, then it is rigid.
The converse is not true, see Figures 18.2b and c in which the existence of the non-trivial in-
finitesimal motion of a rigid framework correlates to the placement being in a special position. For
b, the placement is not in general position, and for c there is a separating set of three collinear edges.
In [2] Bolker and Roth prove (among many other things) that a placement of K3,3 in R2 is
infinitesimally rigid unless its six vertices lie on a conic. Actual motions of planar bipartite frame-
works were studied in [36, 45].
In [15] it is shown that there is an open dense set of placements, the generic embeddings, for
which neither rigidity nor infinitesimal rigidity depend on the placement, and, in addition, the con-
verse of Theorem 18.2 is true.
Theorem 18.3 If F = (G, p) is a rigid framework in d-space with a generic placement p, then F is
infinitesimally rigid.
For simplicity, one can take the generic embeddings to be those for which the |V |d coordinates
are algebraically independent from one another, or, practically, a “random” embedding. If a frame-
work has a rigid generic embedding, we say its graph is generically rigid. Table 18.1 illustrates the
connections between these concepts. It is often stated that the framework F = ((V, E), p) is gener-
ically rigid if for any generic embedding q of V , the framework ((V, E), q) is rigid, however much
confusion can be avoided if one reserves the concept of generic rigidity to graphs, so that the first
three columns in Table 18.1 are referring to the frameworks F = ((V, E), p) as drawn, while for the
fourth column, only the graph (V, E) is relevant.

18.1.1 Rigidity Matrix and Augmented Rigidity Matrices

Given an ordering on the vertices and the edges, the linear constraint system of the equations in
(18.3) for infinitesimal rigidity is encoded in a |V |d × |E| matrix, R(F), called the rigidity matrix.
The infinitesimal motions of the framework, also called flexes, are the elements of the kernel of this
matrix.
The placement p(u) = (−2, 0), p(v) = (1, 3), p(w) = (1, −3) corresponds to the rigidity matrix
R(F),

 u v w 
R(({u, v, w}, {(u, v), (u, w)}), p) = (u, v) 3 3 −3 −3 0 0 ,
(u, w) 3 −3 0 0 −3 3

a basis for whose flexes is illustrated in Figure 18.4, where the first three flexes are a basis for I0 ,
and the fourth flex spans the orthogonal complement of I0 . So the fourth flex is non-trivial and
corresponds to a displacement which does not change the centroid of p(V ) and has zero moment.
If the centroid of p(V ) is at the origin, the space of such representatives for the non-trivial flexes

can be found by computing the kernel of R(F) augmented by rows corresponding to the d+1 2
normalization equations
|V | |V |
X X
0= ei · p0 (vk ), 1 ≤ i ≤ d, 0= Mi j p(vk ) · p0 (vk ), 1 ≤ i < j ≤ d. (18.4)
k=1 k=1
Planar Rigidity 381

Table 18.1
A comparison of types of rigidity. All frameworks are in the plane.
rigid globally infinitesimally generically
rigid rigid rigid

No No No No

No No No Yes

Yes No No No

Yes No No Yes

Yes No Yes Yes

Yes Yes No No

Yes Yes No Yes

Yes Yes Yes Yes

For the example above, the augmented rigidity matrix is

 u v w 
(u, v) 3 3 −3 −3 0 0
(u, w) 
 3 −3 0 0 −3 3 
 .
(e1 , e1 , e1 ) 
 1 0 1 0 1 0 

(e2 , e2 , e2 )  0 1 0 1 0 1 
M1,2 (u, v, w) 0 2 3 −1 −3 −1

Its kernel has dimension 1 containing the vector (−2, 0, 1, −3, 1, 3) corresponding to a non-trivial
infinitesimal motion leaving the centroid at the origin.
An augmentation of the constraint system eliminating the trivial motions is often called pinning
the framework.
To apply the above method of pinning to a given framework, first compute the centroid and
translate the framework so that the centroid is at the origin.
A more direct method of pinning is to choose a set {v1 , . . . , vd } of vertices placed in general
position and to constrain any flex to require that p0 (v1 ) = 0 and also for 2 ≤ i ≤ d that p0 (vi ) be
382 Handbook of Geometric Constraint Systems Principles

v v v v

u u u u

a) w b) w c) w d) w

Figure 18.3
A basis for the flexes of a two-bar framework.

contained in the affine span of {p(v1 ), . . . p(vi−1 )}. If we assume, without loss of generality, that
p(vi ) ∈ he1 , . . . , ei−1 i, for i = 1, . . . , d, these restricted flexes can be found by augmenting R(F) by
rows corresponding to the system

e j · p0 (vi ) = 0, 1 ≤ i ≤ j ≤ d, (18.5)

This constrains any flex to satisfy p0 (v1 ) = 0 and for 2 ≤ i ≤ d to require p0 (vi ) to be contained in the
affine span of {p(v1 ), . . . p(vi−1 )}. See Figure 18.4 in which the second example illustrates pinning

v v

u u

(a) w (b) w

Figure 18.4
Normalized flexes for frameworks pinned with the equations in (18.5) using (a) v1 = w and v2 = v
and (b) v1 = w and v2 = u.

the vertices of a complete subgraph Kd of G, in the plane just an edge. This forces p0 (vi ) = 0 for
1 ≤ i ≤ d, and in this case values of the normalized flexes on the remaining vertices can be found
using the reduced rigidity
matrix, obtained from R(F) by deleting the d 2 columns corresponding
to {v1 , . . . , vd }, and the d2 rows corresponding to the edges between them. The reduced rigidity
matrix in the above example for pinning u and w is just [3, 3].

18.1.2 Rigidity Matrix as a Transformation

It is very useful when considering the abstract structures which follow to regard the rigidity matrix
of a framework, R(F), as a linear transformation from Rd|V | to R|E| . The domain is the set Rd|V |
of vertex displacements, ∆p ∈ Rd|V | , written as 1 × |V |d column vector. The range is the space of
edge elongations, ∆e ∈ R|E| . A 1 × |E| column vector ∆e of edge elongations is in the image of
R(F) if it is the linear approximation to the result of a vertex displacement, ∆e = R(F)p, in which
case ∆e is called a strain. Note that here strains have units of square length, which differs from
common engineering practice. Of course, a displacement ∆p corresponds to an infinitesimal motion
if its strain is the zero vector, R(F)∆p = 0.
Planar Rigidity 383

The dual transformation carries just as much information about the framework, although the
physical interpretation is different. Since each edge constrains the motion of its incident vertices, it
is natural to regard the role of that edge as to exert force on its endpoints, equal in magnitude but
opposite in direction, depending on whether the edge is under tension or compression. The state
of these edge forces in the framework is given by a row vector ω ∈ R|E| , whose coordinates have
units of force per unit length, called a stress vector. One easily computes that the d coordinates
of f = ωR(F) corresponding to vertex v form the resultant force at vertex v from the forces at its
incident edges. We think of a resultant f as being available to balance, or resist, an external force
−f on the vertices of the framework. In short, the domain of the dual transformation is the space of
edge stresses, the codomain is the space of vertex forces, and the image is the space of resolvable
forces.
Of course the framework will be unable to resist forces which tend to translate or rotate the
framework as a whole, so we are only concerned with vertex forces which are equilibriated, that
is, orthogonal to I0 . The framework F is rigid in this context, we say statically rigid, if any applied
equilibriated force can be balanced by vertex forces
arising from a stress assignment to the edges.
In other words, the rank of R(F) must be |V |d − d+1 2 , which is the same rank that R(F) must have
in order for every flex to be trivial, so these dual notions of rigidity, static and infinitesimal rigidity,
are the same. The kernel of the dual transformation consists of those stresses which resolve to the
zero vector on all vertices. These stresses are called variously, equilibrium stresses, prestresses,
resolvable stresses, and, unfortunately, stresses. (This ambiguity is usually resolved by the context
–if it is stated that a framework has a stress, or the rank of the space of stresses is computed,
equilibrium stresses are meant). Table 18.2 summarizes the dual terminology, kinematic and static,
surrounding the rigidity matrix.

Table 18.2
The rigidity nomenclature for the linear algebra of the rigidity matrix.
linear algebra framework
domain R|V |d placements
codomain R|E| stresses
range edge elongations
corange vertex forces
row infinitesimal constraint
resultant force from a single stressed edge
column edge strain
columns for v ∈ V vector star at p(v)
column space strains
row space resolvable forces
kernel flexes
d+1

kernel dimension 2 + degree of freedom
cokernel resolvable stresses
constraint dependence
rank strain dimension
dimension of resolvable forces

The connection between these dual points of view is in the consideration of the work done by a
displacement of the vertices under the vertex forces,
W = f · ∆p = (ωR(F)) ∆p = ω (R(F)∆p) = ω · ∆e (18.6)
384 Handbook of Geometric Constraint Systems Principles

which will be examined more thoroughly in a later Chapter 14. Here we only want to note the
roles played by the various components of the linear transformation R(F) in terms of infinitesimal
rigidity.

Theorem 18.4 A framework F is infinitesimally rigid if and only if any one of the following condi-
tions is satisfied:
(a) F has no non-trivial flexes.
d+1

(b) The rank of R(F) is d|V | − 2 .

(c) Every set of equilibriated forces on the vertices is equal to the resultant forces from some
stress on the edges.
(d) The number of independent resolvable stresses is |E| + d+1


2 − d|V |.

(e) For each pair of vertices v and w, there is a stress ω whose resultant force f = ωR(F) is zero
except at vertices v and w, where the forces are equal and opposite, directed along the line
between p(v) and p(w).

18.1.3 The Infinitesimal Rigidity Matroid of a Framework

Given a framework F = ((V, E), p), the rows of R(F) correspond to the infinitesimal constraints
on the vertices. Independence of these constraints is a key feature of the system. If one row is
linearly dependent on the remaining rows, that row may be deleted without changing the flexes of the
framework, and the corresponding bar is implied by the other bars, or redundant, and the framework
overbraced. The proper combinatorial structure for studying independence is the matriod. The rows
of R(F) comprise a matroid, the infinitesimal rigidity matroid of F, F(F), whose independent sets
are those subsets of rows which are linearly independent. The following lists the important structures
in this matroid, contrasted with language of linear algebra, and that of frameworks.
Independence: A set E 0 is independent in F(F). — The rows in E 0 are linearly independent. —
The subframework ((V (E 0 ), E 0 ), p) is nowhere overbraced.
Dependence: A set E 0 is dependent in F(F). — The rows in E 0 are linearly dependent. — The
subframework ((V (E 0 ), E 0 ), p) is somewhere overbraced.
Cycle: A set E 0 is a cycle in F(F), a minimally dependent set. — The rows in E 0 are mini-
mally linearly dependent. — The subframework ((V (E 0 ), E 0 ), p) is overbraced, but all of its proper
subframeworks are not overbraced.
Basis: A set E 0 is a basis in F(F), a maximal independent set. — The rows in E 0 are linearly
independent and span the rowspace. — The subframework ((V, E 0 ), p) is a minimal subframework
with the same flexes as F.
Rank: The rank of E 0 , rank(E 0 ) in F(F), the cardinality of the largest independent subset of E 0 .
— The rank of the row vectors in E 0 . — The dimension of the space of resultant forces from stresses
on E 0 .
Closure: The closure of E 0 in F(F) is the set of all e ∈ E with rank(E 0 + e) = rank(E 0 ). — The
set of all rows in E which lie in the span of E 0 . — The set of bars in E which are implied by the bars
of E 0 .
Planar Rigidity 385

Glossary
Framework, linkage: A pair F = (G, p) where G = (V, E) is a graph and p : V → Rd is a placement
of the vertices into d-dimensional Euclidean space. It is also common to notate the framework
((V, E), p) as (V, E, p).
Joint, node: A point of the framework, represented by a vertex of the graph, where bars may meet,
and whose motion is constrained by the bars.
Bar, link, brace: A connection between two joints, represented by an edge of the graph, whose
role is to fix the distance between that pair of nodes.
Configuration Space: The configuration space of the framework F is the space C(F) of all place-
ments q : V → Rd such that kp(v) − p(w)k = kq(v) − q(w)k for every edge (v, w) ∈ E. The
configuration space C(F) inherits a natural metric and differentiable structure from R|V |d .
Pinned framework: A framework normalized to exclude trivial motions and flexes.
Rigid framework, locally rigid framework: A framework (G, p) such that p has a neighborhood
in C(p) containing only embeddings q such that q = pI, for some isometry I of d-space.
Rigidity function, length function: The rigidity function of the graph G is ρ : R|V |d → R|E| which
takes each placement p to the vector ρ(p) of square bar lengths; the coordinate of R|E| corre-
sponding to the edge (i, j) ∈ E is (p(i) − p( j))2 .
Rigidity matrix: The rigidity matrix R(F) of a framework is an |E| × d|V | matrix, the Jacobian
matrix of the rigidity function of the graph, evaluated at the embedding p, divided by 2.
Motion of the framework F, deformation: A continuous one parameter family Ft of frameworks
with F0 = F, not all of which are congruent to F.
Infinitesimal motion, first order motion, flex: Solution p0 (v) to the linear system

(p(v) − p(w)) · (p0 (v) − p0 (w)) = 0 (v, w) ∈ E, (18.3)

encoded by the rigidity matrix.

Infinitesimally rigid framework, first order rigid framework: A framework F = ((V, E), p)
whose rigidity matrix R(F) has rank d|V | − d+1


2 .
General position embedding: Embedding p such that no n vertices, n ≥ 2, are placed on any n − 2
dimensional affine subspace.
Generic embedding: Embedding p : V → Rd is generic if no sub-determinant of the rigidity matrix
is zero unless it is identically equal to zero.
Generically rigid graph: Graph with an (infinitesimally) rigid generic embedding.
(Internal) degree of freedom: Number of independent flexes of the framework. Rank of the kernel
of the rigidity matrix, df(F),
 
d +1
df(E) = r(E) − m|V (E)| + .
2
386 Handbook of Geometric Constraint Systems Principles

18.2 Abstract Rigidity Matroids

Let K = K(V ) denote the edge set of the complete graph on the set V , |V | = n. A matroid Ad on
K with closure operator h·i is called a d-dimensional abstract rigidity matroid for V if, besides the
usual closure axioms for a matroid:
C1 T ⊆ hT i;
C2 If R ⊆ T , then hRi ⊆ hT i;
C3 hhT ii = hT i;
C4 If s,t ∈ (E − hT i), then s ∈ hT ∪ {t}i if and only if t ∈ hT ∪ {s}i;
it satisfies the additional conditions C5 and C6 below:
C5 If E, F ⊆ K and |V (E) ∩V (F)| < d, then hE ∪ Fi ⊆ (K(V (E)) ∪ K(V (F))).
C6 If hEi = K(V (E)), hFi = K(V (F)) and |V (E) ∩V (F)| ≥ d, then hE ∪ Fi = K(V (E ∪ F)).
Note that Axiom C5 implies hEi ⊆ K(V (E)) so, motivated by the infinitesimal case, we say E ⊆ K
is rigid if hEi = K(V (E)). With this definition C5 expresses the idea that any separating set of
codimension less than d allows a hinge motion, and C6 reads: if two rigid sets have d or more
vertices in common then their union is rigid.
All infinitesimal rigidity matroids satisfy C6. If p embeds the vertices of a framework in general
position, i.e., no m vertices, m ≤ d + 1, lie on any m − 2 dimensional affine subspace, then F(p)
satisfies C6 also and is an abstract rigidity matroid. There exist, however, abstract rigidity matroids
which are not infinitesimal, [17], and abstract rigidity may also arise in other contexts, see [26]
or [64].
Let Ad be a d-dimensional abstract rigidity matroid. If E is any edge set such that |V (E)| ≤ d +1
then, by C5, E is independent. In particular, Kd+1 is independent. If v 6∈ V (E) and F is a set of d
edges joining v to vertices in V (E), then E ∪ F is called a 0-extension of E. Again by C5, 0-
extensions of independent sets are independent. In particular, Kd+2 − e is a 0-extension of Kd+1 , and
hence independent. Moreover, Kd+2 − e is the union of two Kd+1 ’s along a Kd , thus hKd+2 − ei =
Kd+2 by C6, and we have that Kd+2 is a cycle in Ad . The same kind of argument shows that, in fact,
if |V (E)| ≥ d + 1, then  
d +1
r(E) ≤ m|V (E)| −
2
with equality when hEi is complete. This inequality is fundamental for inductive constructions of
independent sets, which are treated in Chapter 19. If hEi is not complete, then the discrepancy
measures the (internal) degree of freedom, df(E), of the edge set E,
 
d +1
df(E) = r(E) − m|V (E)| +
2
.
Abstract rigidity matroids were introduced by Graver in [16], and further investigated in [17,
44, 41].
In dimension 1 the generic rigidity matroids are the only examples of abstract rigidity matroids
and they are known as connectivity matroids, or graphic matroids.
There are 2-dimensional infinitesimal rigidity matroids which are not generic, for example see
Bolker and Roth [2], where it is shown that if V = {1, . . . , 6} is embedded on a conic in R2 , then
each copy of K3,3 in this 2-dimensional infinitesimal rigidity matroid is dependent.
Planar Rigidity 387

There also exist 2-dimensional abstract rigidity matroids which are not infinitesimal rigidity ma-
troids. Take 6 generically embedded vertices, but remove one particular K3,3 subgraph from the set
of bases. It is easy to verify, that this yields another abstract rigidity matroid, which, according to
the Bolker-Roth result is not infinitesimal. A second example is based on the observation that, in
the 2-dimensional infinitesimal rigidity matroid for V = {1, . . . , 6}, not all prisms can be dependent.
For example in Figure 18.2 there are two prisms on the same symmetrically embedded vertex set.
Figure 18.2a has an infinitesimal motion, hence its edge set is dependent in the corresponding in-
finitesimal rigidity matroid, while the prism in Figure 18.2b is isostatic. One can show that starting
with the bases for the 2-dimensional generic rigidity matroid of V and deleting all prisms results in
the bases for a 2-dimensional abstract rigidity matroid for V which is clearly not infinitesimal.

v v v
H v
@ HH
@ H
@v v v HHv
@ HH
H
@ HH
a) v @v b) v Hv

Figure 18.5
Two prisms on the same vertex set.

We know that all 2-dimensional abstract rigidity matroids on n vertices have rank 2n − 3 and that
those edge sets obtained from a single edge by a sequence of 0-extensions must therefore be bases.
Any 2-dimensional abstract rigidity matroid on n vertices, and hence any minimal 2-dimensional
abstract rigidity matroid on n vertices, must include these edge sets as bases. In general, however,
additional bases are necessary, as illustrated in Figure 18.2. The sets B1 and B2 are bases in any

' $ ' $
e f
s s s s s s
s @s @s s @s @s s @s @s
@ @ @ @ @ @

@s @s @s @s @s @s
@ @ @ @ @ @
B1 B2 B3
Figure 18.6
B3 is obtained by basis exchange.

abstract rigidity matroid on seven vertices since they may be obtained by five 0-extensions on a
single edge. The basis axioms assert that B1 − e can be augmented to a basis with an edge from
B2 , which, since tetrahedra are dependent, must be B3 = B1 − e + f . Thus B3 is also a basis in any
2-dimensional abstract rigidity matroid, but B3 is not the result of a sequence of 0-extensions of an
edge since it has no vertex of valence 2.

18.2.1 Characterizations of A2 and (A2 )⊥

The dual matroid, although it contains the same mathematical information, has structures with in-
teresting and important interpretations. The abstract rigidity matroid generalizes the infinitesimal
rigidity matroid, which encodes the independence structure of the rigidity transformation. The dual
388 Handbook of Geometric Constraint Systems Principles

matroid does not directly generalize the dual transformation, but the duality in the sense of homol-
ogy versus cohomology.
For every matroid on groundset E, complements of bases are the bases of the dual matroid.
Matroid cocycles, minimally dependent sets in the dual matroid, are the minimal sets which intersect
every basis of the matroid. Using the rigidity matrix it is easy to verify that in the 2-dimensional
infinitesimal rigidity matroid on the edge set of a complete graph on at least 4 vertices, the set of
edges, S(v), incident to a vertex v ∈ V is dependent in the dual matroid and deleting any element
e ∈ S(v) yields a cocycle. We call these cocycles the vertex cocycles of v; and, for each edge e ∈ S(v),
we denote the vertex cocycle S(v) − e by Se (v)
Se (v) is actually a cocycle in any 2-dimensional abstract rigidity matroid and, in fact, these
cocycles are of use in characterizing 2-dimensional abstract rigidity matroids.

Theorem 18.5 Let (V, K) be the complete graph on n vertices. If M is a 2-dimensional abstract
rigidity matroid on K, then each of the following three conditions hold:
a. For each v ∈ V and each e ∈ S(v), Se (v) is a cocycle of M.
b. No cycle of M contains a vertex of valence less than three.
c. Each 2-valent 0-extension of an independent set of M is also an independent set of M.
Conversely, if any one of the conditions a, b or c hold then M is a 2-dimensional abstract rigidity
matroid on K.

Theorem 18.5, proved in [15], can be used to construct abstract rigidity matroids that are not
infinitesimal. It was extended to higher dimensions by Nguyen in [41].
Observe that the existence of the vertex cocycles in Theorem 18.5 forces the rank of the matroid
to be at least 2n − 3. Also all K4 ’s must be cycles, bounding the rank from above.
Both of these conditions are necessary. The uniform matroid on K of rank 5 has each K4 as a
cycle but, when |V | > 4, it is not a 2-dimensional abstract rigidity matroid; and the uniform matroid
on K of rank |V2| − (|V | − 3) has each vertex star minus an edge as a cocycle but, when |V | > 3, it

is not a 2-dimensional abstract rigidity matroid.

We now turn our attention to an arbitrary rigidity matroid A2 for V = {1, . . . , n}. The complete
set of cocycles can be described. Recall that we say that a spanning edge set E ⊆ K(V ) has degree
of freedom 1 when r(E) = 2n − 4 = r(K(V )) − 1.

Theorem 18.6 Let V = {1, . . . , n} be given. The cocycles of A2 are the complements of closed
spanning subsets with degree of freedom 1 in A2 .

It is easily verified that all cocycles of K5 are isomorphic to one of the following four graphs.
The first graph is a vertex cocycle.

t t t
tHH Ht t t t
HHHt u H Ht
H
 
 AA
AA

t
 At t

t t At
t

Figure 18.7
Cocycles of G2 (K5 ).

The reason for studying cocycles, and in particular vertex cocycles, is to get information about
the graph from the matroid defined on the edge set. In dimension 1 we have the following nice result
Planar Rigidity 389

by Whitney [66] that the cycle matroid of a 3-connected graph uniquely determines the graph.
In this case the vertex cocycles span the cocycle space, which allows us to draw a graph from
a set of cocycles with the property that every edge is contained in exactly two of them. Jordán
and Kaszanitzky extended this result to rigidity matroids in [25]. They show that if G is 7-vertex-
connected then it is uniquely determined by its two-dimensional rigidity matroid, and if a two-
dimensional rigidity matroid is (2k − 3)-connected then its underlying graph is k-vertex-connected.

18.2.2 The 2-Dimensional Generic Rigidity Matroid

If v 6∈ V (E), e ∈ E and F is a set of d + 1 edges joining v to vertices in V (E), including both
endpoints of e, then E ∪ F − e is called a 1–extension of E.

3 3 3 5
4 4

2 2 2 2
0 1 0 1 0 1 0 1 0 1

2 2 2 2
3 3 3
4 4

0 1 0 1 0 1 0 1 0 5 1

Figure 18.8
Constructing frameworks in the plane on the triangular prism and K3,3 from a single edge via a
sequence of 0-extensions and 1-extensions.

In dimension two, we can re-state the rank bound (18.2) using the 1-extension. If E ⊆ K is
independent, then |F| ≤ 2|V (F)|− 3, for each nonempty subset F of E. We say that independent sets
satisfy Laman’s condition. An abstract rigidity matroid is said to have the 1-extendability property
if, given S
an independent set E on vertex set V , every 1–extension of E yields an independent edge
set on V v.

Theorem 18.7 (Laman’s Theorem) Let A2 be an abstract rigidity matroid on the complete graph
on n vertices. The following are equivalent:
(a) A2 satisfies Laman’s condition,
(b) A2 has the 1-extendability property,
(c) A2 = G2 (n).

This characterization of G2 (n) yields the following.

Theorem 18.8 G2 (n) is the unique maximal 2-dimensional abstract rigidity matroid on n vertices.
The Laman characterization of the isostatic sets in G2 (n) is a powerful tool in the study of
isostatic sets and, in view of the maximality of G2 (n), tells us much about the structure of isostatic
sets in an arbitrary 2-dimensional abstract rigidity matroid.
Given a graph G, with edge set E, one can show directly that the collection of subsets of E which
390 Handbook of Geometric Constraint Systems Principles

satisfy Laman’s condition in dimension 2 are the independent sets of a matroid on E. Laman’s fa-
mous 1970 paper [33] used linear algebra techniques to prove Theorem 18.7. A paper by Pollaczek-
Geiringer [46] in 1927 gave an equivalent theorem to Laman’s theorem, where an induction proof
using 0–extensions and 1–extensions is provided. Moreover, Pollaczek-Geiringer extends Theo-
rem 18.7 to mechanisms: A framework on n vertices and e edges has 2n − 3 − e generic degrees
of freedom if no subframework is overbraced. She also formulates an algebraic interpretation of
Theorem 18.7 in the spirit of Frobenius [14] and points out interesting references to special position
frameworks, see [31].
It is natural to ask if the edge sets that satisfy Laman’s inequalities for dimension m > 2 also
form the collection of independent sets of some matroid, since such a matroid would be a logical
candidate for the generic rigidity matroid in dimension m. Unfortunately we have the following
negative result.

Theorem 18.9 The collection of edge sets that satisfy Laman’s condition for dimension m cannot
be the collection of independent sets of any matroid on K if m ≥ 3.

In the case d = 3, one may observe that the proposed rank bound, r(E) ≤ 3|V (E)| − 6 fails for
single edges. On the other hand, the bound r(E) ≤ 3|V (E)| − 5 works for single edges.
Whiteley, in [65], presents an important survey on an array of matroids drawn from three sources
in discrete applied geometry: (i) static (or first-order) rigidity of frameworks and higher skeletal
rigidity; (ii) parallel drawings (or, equivalently, polyhedral pictures); and (iii) Crr−1 -cofactors ab-
stracted from multivariate splines in all dimensions. In dimension 2 all of these three matroids are
the same, in fact all equal G2 (n).

18.2.3 Cycles in G2 (n)

Laman’s characterization of G2 (n) allows us to combinatorially identify the independent sets, from
which it is easy to deduce a characterization for the cycles of G2 (n).

Theorem 18.10 An edge set C is a cycle in G2 (n) if and only if |E(C)| = 2|V (C)| − 2 and |F| ≤
2|V (F)| − 3 for every proper subset F of E(C).

In Figure 18.9, all cycles in G2 (n) having fewer than 7 vertices are described.

t t t t t t t t t t

J @t ttHHt @B t
@ t Z
JZ
J 

t t BBtZt
Z

tJ @ A
A
B J



t Jt t @t t At
Bt t
Jt
Z
t Jt ZBt
 Z J Z B 
HH

A

Figure 18.9
The generic rigidity cycles on up to 6 vertices.

Note that A2 admits no nongeneric cycles on fewer than six vertices. The only possible non-
generic cycles on six vertices are K3,3 and the triangular prism.
Also note that the wheel Wn , the graph obtained from an n-gon by attaching a single new vertex
which is adjacent to all the vertices of the n-gon, is a cycle in any abstract rigidity matroid in which
it is contained. In Figure 18.9 the first three graphs are wheels.
It follows immediately from Theorem 18.10 that generic cycles are rigid, in fact over braced in
a homogeneous way, specifically, the removal of any edge leaves a rigid graph whose edge set is
independent in G2 (n). In fact, the rigidity of the cycles characterizes G2 (n).
Planar Rigidity 391

Theorem 18.11 Let A2 be a 2-dimensional abstract rigidity matroid on n vertices. Then A2 =

G2 (n) if and only if all cycles in A2 are rigid.

The following lemma lists some simple but useful properties of cycles in G2 (n).

Lemma 18.1
Let C be a cycle in G2 (n). Then:
a. (V (C),C) is 2-connected.
b. (V (C),C) is 3-edge-connected.

c. If the removal of 3 edges disconnects (V (C),C), then the 3 edges have a common endpoint v,
and v is of valence 3 in C.

It is easy to verify that a 1-extension of a cycle is a cycle. The question, posed by R. Connelly,
whether or not all 3-connected cycles in G2 (n) can be obtained from the tetrahedron by a sequence
of 1-extensions was answered affirmatively by Berg and Jordán in [1].

Theorem 18.12 (Berg-Jordán) All 3-connected cycles in G2 (n) can be obtained from the tetrahe-
dron by a sequence of 1-extensions.

The Berg-Jordán result is a very useful tool in investigating the structure of generic cycles. For
independent sets, all 1-extensions yield independent sets, so one might suspect that all 1-extensions
of generic cycles are generic cycles, but this turns out to be false: Consider C, the fifth cycle in
Figure 18.9. Let v be the upper right-hand vertex of C. There is no 1-extension giving C with v as
the new vertex being attached, since C − v has 3 vertices of degree 2. Connelly in [6] has intriguing
examples with many adjacent vertices of degree 3 which he calls spider webs.
Let C1 and C2 be cycles in Gd (n) such that E(C1 ) ∩ E(C2 ) = {e} and |V (E1 ) ∩V (E2 )| = 2, i.e.,
C1 and C2 have exactly one edge and no vertices other than the endpoints of this edge in common.
Then the symmetric difference of C1 and C2 is an d-cycle for all d. This means that for d > 2 there
must be nonrigid cycles, in fact cycles with arbitrarily large degree of freedom.
If a cycle in G2 (n) is not 3-connected, then it is the symmetric difference of two cycles in G2 (n)
which have exactly one edge and its endpoints in common. This operation is called the 2-sum. So
all cycles in G2 (n) are obtained by the operations of 1-extension and 2-sum from a set of disjoint
tetrahedra. However, G2 (n) is not closed under 2-sum decomposition [52].
In [20] Jackson and Jordán proved that in the plane a graph G is globally rigid if G is vertex
3-connected and G2 (G) is connected (i.e., for any two edges of G there is a cycle of G2 (G) contain-
ing both of them), so the obvious two necessary conditions for a unique embedding of G are also
sufficient in the plane.
While the rigidity of a graph and the connectivity of a graph are closely related, planarity
and rigidity seem quite unrelated. However, we have the following curious property in the plane,
see [51].
Theorem 18.13 Let C be a cycle in G2 (n) such that (V (C),C) is planar. Then the edge set of the
geometric dual of (V (C),C) is also a generic cycle.

Wheels, Wn are self-dual planar generic cycles.

18.2.4 Rigid Components of G2 (G)

We regard a graph G as a subgraph of Kn for some n > |V (G)| and denote the restriction of G2 (n) to
E(G) by G2 (G). Since G2 (n) is a submatroid of G2 (m) for m > n, this definition is independent of n,
392 Handbook of Geometric Constraint Systems Principles

and we may define a set F in G2 (G) to be rigid if r(F) = 2|V (F)| − 3, where r is the rank function
of G2 (n) restricted to E(G).
A maximal rigid subgraph of G is called an r-component. It is an immediate consequence of
Axiom C6 that two r-components have at most one vertex in common, hence the r-components may
be regarded as a partition of the edges of G. A non-rigid graph has at least two r-components.
Since the cycles of G2 (G) consist of those cycles of G2 (n) which are completely contained in
E(G), and since cycles are rigid by Theorem 18.11, it follows that every cycle of G2 (n) is completely
contained in some rigid component. This gives the following theorem.

Theorem 18.14 Let G be a graph and let E1 , . . . , Ek be the rigid components of G. Then G2 (G) =
G2 (G1 ) ⊕ . . . ⊕ G2 (Gk ), where Gi = (V (Ei ), Ei ).

If G0 = (V 0 , E 0 ) is a subgraph of G, and G0i = (Vi0 , Ei0 ) are the rigid components of G0 , then the
closure of E 0 in G2 (G) is E 0 together with all edges in E both of whose endpoints lie in some G0i ,
and the closure of E 0 in G2 (n) is the union of the cliques K(V (E10 )). This leads to the following
characterization of G2 (n).

Theorem 18.15 Let A2 be a 2-dimensional abstract rigidity matroid on n vertices. Then A2 =

G2 (n) if and only if for any closed set E with cliques E1 , . . . , Ek , and E = ∪ki=1 Ei we have r(E) =
r(E1 ) + · · · + r(Ek ).

It follows from this result that, if A2 is a 2-dimensional abstract rigidity matroid which is not
generic, then it must contain a closed set violating the the equality in this theorem. Look, for exam-
ple, at Figure 18.2a, where any additional edge would kill the existing infinitesimal motion, so the
edge set is closed, the rigid components are the two triangles and the three parallel edges, but the
rank is only 8.
Consider V = {1, . . . , 6} embedded on a conic. Then, as we have noted, each copy of K3,3 in
K(V ) is a cycle. K3,3 is closed and its r-components are simply its edges, so the equality in the
theorem is violated for K3,3 .
In many ways, the r-components are analogous to the connected components of a graph. How-
ever, the two concepts do have some important differences: The removal of any edge from a graph
increases the number of connected components by at most one. By contrast, the removal of an edge
from a rigid graph can result in a graph with many r-components. For example, if G is a quadrilateral
with 1 diagonal, then the removal of the diagonal results in a graph with four r-components, all of
which are single edges. Removing any other edge of G leaves two r-components, a triangle and a
single edge. It is not difficult to show, however, that deleting an edge from a minimally rigid graph
in the plane always yields an even number of r-components.
The number of r-components does not give us any information on the rank of a given set. Of
course, we could compute the rank by summing the ranks of the rigid components. Sometimes it is
of interest to know how many additional edges are needed to achieve rigidity of (V (E), E). Recall
that the degree of freedom of an edge set E, df(E), in G2 (n) is defined by

df(E) + r(E) = 2|V (E)| − 3.

A graph G = (V, E) with degree of freedom 1 and 2k r-components, each of which has ni vertices,
has
2k
X
|V | = ni − 3k + 2.
i=1

Using the direct sum decomposition of G2 (G) induced by the r-components we can give a com-
binatorial description of G2 (G) via the rank function.
Planar Rigidity 393

Theorem 18.16 Let G = (V, E) be a graph and let r denote the rank function of G2 (G). Then
k
X
r(E) = min (2|V (Ei )| − 3),
i=1

where the minimum is taken over all collections {Ei } such that E = ∪Ei .

The minimum in (18.16) is achieved if {Ei } coincides with the collection of rigid components
of G, however other collections may also give the minimum, as seen in Figure 18.10, in which the

tP
t PP Pt
t t t t
 

t tPPP
t t t @ Pt t
PPP
 P  B 
Pt t  tB
P 
B PP @  B  B
 B  tB  B t @Bt
t Bt t @B t t Bt t t
BB t 
@ BB  t t B 
B   tPP B 
 B t 
@ t B 
Bt t t B  t tP BBt t
P
  @ P B  

Pt BBt @ tP t
P
PP   P
P
tP t
PPPt
4(2 · 5 − 3) 4(2 · 4 − 3) + 8(2 · 2 − 3)
Figure 18.10
P
Computing (2|V (Ei )| − 3) for different edge partitions.

count at the left is via the rigid components. From Theorem 18.16 there follows the following result
first due to Lovász and Yemini, see [35].

Corollary 18.17
k
X
df(E) = 2|V (E)| − 3 − min (2|V (Ei )| − 3),
i=1

where the minimum extends over all systems {E1 , . . . , Ek } of subsets of E such that E1 ∪ . . . ∪ Ek = E.

18.2.5 Representability of G2 (n)

The graphic matroid, in our notation G1 (n), has the nice property of being binary, in fact, regular,
see [43, 63]. We know that G2 (n) is representable over the real numbers, since a representation is
obtained from a generic embedding p of the vertices V of a complete graph K, by the rigidity matrix
R(p) = R(K, p).

Theorem 18.18 No 2-dimensional abstract rigidity matroid on more than 4 vertices is binary.

Proof 18.1 Recall that a matroid is binary if and only if the symmetric difference of two cycles
is the disjoint union of cycles. If n > 4, we can find in Kn two K4 ’s having 3 vertices in common so
that their disjoint union is independent. See Figure 18.11.
394 Handbook of Geometric Constraint Systems Principles
1 sH 1 sH
L HH4s 4s L HH4s
 L "
"
 L
 L" "  L
s"" Ls
+ s"" s
= s Ls
2 3 2 bb 3J 2 bb3 J
Js 5 bs 5
bb bJ

Figure 18.11
Two cycles whose symmetric difference is independent.

In any matroid M on E representable over a finite field F, the number of cycles is bounded by
|F|k −1
|F|−1 ,where k = |E| − r(E). So if k is small, there are relatively few cycles.
The next theorem uses this fact to show that no finite field is large enough to represent all generic
rigidity matroids in dimension 2.

Theorem 18.19 There is no finite field F such that G2 (n) is representable over F for all n.

Proof 18.2 The graph (V, E) in Figure 18.12 is rigid and has k = |E| − r(E) = 2. The removal

s s
Q
B Qs s s s s s s s s  B
B Q B

B AA
s A

s A

s A

s A

s A
s A
s A
s As B

A
A
A
A
A
A
A
A
B  Q B
Bs QQB s

Figure 18.12
A rigid graph with k = |E| − r(E) = 2.

of any edge with two endpoints of valence 4, as well as removal of a vertex of valence 3, yields a
cycle. Clearly, we can make the string of triangles as long as we wish, so if a graph G of this form
has 2n vertices, G2 (n) contains 2(n + 1) cycles. Now assume that G2 (n) is representable over F. We
have:
|F|2 − 1
|F| + 1 = ≥ 2(n + 1).
|F| − 1

However, G2 (n) is representable over the rational numbers. Representability of abstract rigidity
matroids over fields of prime characteristic has not yet been studied.

Glossary
0-extension, Henneberg-1 move: The 0-extension of a framework ((V, E), p) in dimension d is a
new framework consisting of ((V, E), p) augmented by a new vertex v together d edges joining
v to vertices in V .
1-extension, Henneberg-2 move, edge split: The 1-extension of a framework ((V, E), p) at an
edge e ∈ E in dimension d is a new framework consisting of ((V, E − e), p) augmented by a
Planar Rigidity 395

new vertex v together with d + 1 edges joining v to vertices in V , including the endpoints of the
deleted (split) edge e.
1-Extendability property: An abstract rigidity matroid has the 1-extendability property if, given
anSindependent set E on vertex set V , every 1–extension of E yields an independent edge set on
V v.
Abstract rigidity matroid: A matroid Ad on the edges of a complete graph K with closure oper-
ator h·i is called a d-dimensional abstract rigidity matroid for V if it satisfies the usual four
closure axioms for a matroid as well as
C5 If E, F ⊆ K and |V (E) ∩V (F)| < d, then hE ∪ Fi ⊆ (K(V (E)) ∪ K(V (F))).
C6 If hEi = K(V (E)), hFi = K(V (F)) and |V (E) ∩V (F)| ≥ d, then hE ∪ Fi = K(V (E ∪ F)).

Cocycles: Minimally dependent sets in the M∗ , i.e., minimal sets which intersect every basis of
M.
Generic rigidity matroid: the matroid Gd (n), the infinitesimal rigidity matroid on the edges of the
complete graph on n vertices placed generically in Rd .

Infinitesimal rigidity matroid: Matroid on the edges of a framework in which the independent sets
are those edge sets whose corresponding rows in the rigidity matrix are linearly independent.
Isostatic framework (infinitesimally, generically): (Infinitesimally, generically) rigid framework
such that the removal of any bar destroys the rigidity.

Laman’s condition: An edge set satisfies Laman’s condition if |F| ≤ 2|V (F)| − 3 for all all non-
empty subsets F ⊆ E. Independent sets of G2 (n) satisfy Laman’s condition.
Rigid component, r-component: A maximal rigid subgraph of G. The r components of an abstract
rigidity matroid partition E.
Vertex cocycles: Cocycles of an abstract rigidity matroid on (V, E) consisting only of edges inci-
dent at one vertex. In A2 , the vertex cocycles are sets of the form Se (v) consisting of all edges
incident to v except e.

18.3 Rigidity and Connectivity

The matroid G1 (G) is the connectivity matroid of the graph G, and is closely related to the con-
nectivity of the graph. As such, generic rigidity in higher dimensions has often been regarded as
a generalization of connectivity. Both concepts require a certain number of edges to be well dis-
tributed over the graph. The purpose of this section is to collect results on the relation between the
generic rigidity of graphs, the connectivity of graphs, and the connectivity of the relevant matroids.
The connectivity of a graph may be measured in many ways. In general, a graph is (vertex) k-
connected if it has at least k + 1 vertices and the removal of any k − 1-vertices results in a connected
graph. A graph is edge k-connected if it has at least k edges and the removal of any k −1 edges results
in a connected graph. Except for trivial examples, if k = 1 the edge and vertex versions coincide and
connectivity is characterized by a counting condition completely analogous to Laman’s Theorem:
396 Handbook of Geometric Constraint Systems Principles

Theorem 18.20 A graph G = (V, E) is connected if and only if there is a subset F ⊆ E such that
(a) |F| = |V | − 1, and
(b) |F 0 | ≤ |V (F 0 )| − 1 for all subsets F 0 ⊆ F.

The relation between vertex and edge k-connectivity is well studied, see for example the fundamen-
tal work of Tutte [61], or the modern text by Diestel [11].
The higher the vertex connectivity of the graph, the more likely it is that the graph is generically
rigid in dimension d. However, even in dimension 2, this heuristic is often violated: there are gener-
ically rigid graphs which are 2-connected but not 3-connected, for instance the rightmost graph in
Figure 18.9, yet in Figure 18.13a we have a graph which is 5-connected but generically flexible.

(a) (b)

Figure 18.13
(a) A 5-regular graph whose vertices are incident vertex-edge pairs (v, e) in K5,5 , with two pairs
being adjacent if they have an element in common.(b) An edge-birigid graph which is the union of
three cycles.

Each of the copies of K5 is over-braced by 3 edges, and removing these 30 redundant edges leaves
the remainder under braced. The next result shows that the region of uncertainty for generic rigidity
in the plane is from 2-connected to 5-connected.
Theorem 18.21 (Lovász and Yemini [35]) Every vertex 6-connected graph is generically rigid in
R2 .
There is no similar result for edge connectivity alone, since taking the union of two copies of Kn+1
at a vertex yields a graph which is n-edge-connected but generically flexible in R2 .

18.3.1 Birigidity
A graph is vertex k-rigid if it has at least k +1 vertices and the removal of any k − 1 vertices, together
with their incident edges, leaves a rigid graph A graph is edge k-rigid if the removal of any k − 1
edges leaves a rigid graph. For k = 2 the term birigidity is used. Every edge in an edge birigid graph,
that is a 2-edge rigid graph, is redundant, and the term redundantly rigid is also used for G. With
Planar Rigidity 397

these definitions, the generic (edge) k-rigidity of a graph in R1 is exactly its k-connectivity. A cycle
in G2 (n) is edge-birigid, but Figure 18.13 illustrates the general case.
Parallel to vertex- and edge connectivity in graphs, Tutte [62] also introduced the concept of
connectivity in matroids. A matroid M on S is connected if r(F) + r(E − F) > r(E) holds for every
non-empty proper subset F of E. If there is a subset F such that r(F) + r(E − F) ≤ r(E), then
necessarily r(F) + r(E − F) = r(E) and F and E − F are said to separate the matroid, and we write
M = F ⊕ (E − F). The connectivity matroid, G1 (G), of a graph G is connected if and only if G is
2-connected.
The situation in the plane is more complex. Suppose that G = (V, E) is vertex birigid in the
plane. Let v ∈ V and let v be incident to edges e1 , . . . , em . If G is not simply a triangle, we must have
3 ≥ m, otherwise deleting one of the neighbors of v would leave a non-trivial motion of v. Since
G is birigid, E 0 = E − {e1 , . . . , em } is rigid, as well as E 0 ∪ {ei , e j} for any pair 1 ≤ i < j ≤ m, and
E 0 ∪ {ei , e j , ek } must be dependent, with a cycle of G2 (G) containing both ei and e j . So every pair
of edges incident to v belong to a cycle of G2 (G), and hence to the same direct summand of G2 (G).
Since the graph G is connected, G2 (G) can have only one direct summand, so G2 (G) is a connected
matroid.
Suppose, on the other hand, that G2 (G) is connected. Then by Theorem 18.14 it has only one
rigid component, so G is rigid, and hEi = 2|V | − 3. If any edge e of G = (V, E) were not redundant
then hE − ei = 2|V | − 4, and E − e and {e} would be a separating partition of the matroid G2 (G),
hE − ei + hei = hEi, so G is edge birigid.

Theorem 18.22 (Graver, Servatius, and Servatius, 1993 [15]) If G is vertex birigid in the plane,
then G2 (G) is connected.
If G2 (G) is connected then G is edge-birigid.

The argument above implies this birigidity result is also true in any abstract rigidity matroid,
but neither implication can be reversed. The graph of Figure 18.13b is edge birigid but its generic
rigidity matroid is the direct sum of three connected components. The generic rigidity matroid on
the wheel Wn , n ≥ 4 is connected, since G2 (Wn ) is just a cycle, but removing the center vertex yields
a non-rigid graph.
In reviewing the proof of the Lovász Yemini Theorem, we see that, if a graph is 6-connected,
then it is not only rigid but over-braced. In [22] 6-connectivity is replaced by 6-mixed connectivity,
where a graph G is 6-mixed connected if G −U − D is connected for all vertex subsets U and edge
subsets D satisfying 2|U| + |D| ≤ 5.
The connectivity of the rigidity matroid together with the fact that cycles are rigid was used by
Jackson and Jordàn in [20] to characterize global rigidity in the plane.

Theorem 18.23 (Jackson and Jordán, 2005 [20]) A graph G is generically globally rigid in the
plane if and only if G is 3-connected and edge birigid.

18.3.2 Tree Decomposition Theorems

Tutte, resp. Nash-Williams, see [39, 40, 60] proved (1961–1964) that a graph G = (V, E) is de-
composable into k spanning forests if and only if |F| ≤ k(|V (F)| − 1) for all nonempty subsets F
of E. For k = 2 this looks remarkably similar to Laman’s characterization of independent sets in
G2 (n). The graphs described by Tutte/Nash-Williams have the same number of edges, relative to the
number of vertices, as a cycle in G2 (n).
Two spanning trees of G which are edge disjoint and have G as their union are called a 2-tree
decomposition of G.
A 2-tree decomposition of G is said be a proper 2-tree decomposition if no pair of proper sub-
trees, excepting single vertices, have the same span.
398 Handbook of Geometric Constraint Systems Principles

Choosing two paths of length n randomly on the same vertex set of size n will yield a graph
with 4 vertices of degree three (the respective endpoints of the two paths, and all other vertices of
degree 4. By construction this graph has a proper two tree (= path) decomposition, so it will be a
cycle in G2 (n).
However, it is not true that every cycle with vertices of valence 3 and 4 only, has a two tree
decomposition into two paths. Kijima and Tanigawa, [29] construct infinitely many planar coun-
terexamples. See also [30].
If an edge set E is isostatic, x, y ∈ V (E), (x, y) 6∈ E, then E + (x, y) contains exactly one cycle
of G2 (n) and admits a 2-tree decomposition. To test isostaticity by using tree decompositions it is
more economical to use the following result of Recski [47, 48, 50].

Theorem 18.24 Let G = (V, E) be a graph. Then E is isostatic in G2 (n) if and only if, doubling any
edge e ∈ E, results in a multigraph which admits a 2-tree decomposition.

Theorem 18.24 implies that an isostatic set is the union of three trees. One of the trees is span-
ning, the other two arise from removing the extra edge from the second tree in the decomposition,
so one of the small trees could actually be an isolated vertex. Crapo [9] noted that many other 3-tree
decompositions exist and went on to prove the 3-tree decomposition theorem below. In Figure 18.14
we illustrate his result by listing three 3-tree decompositions for K3,3 .

r r r r r r
r
Br Br r A r r
AA
B B A


B
A   B
A
Br
Ar r r B
r Ar
r r r r r r
r
Br r r r Br
AA
B A

 B
A

B
A 
Br
Ar Br r r
Ar
B  A
r r r r r r
r
Br r Br r r
AA
B  B A


B
A 
A
B
r Ar Br r
r Ar
B  A

Figure 18.14
Three 3-tree decompositions of K3,3 .

Theorem 18.25 A graph G = (V, E) is isostatic if and only if G is the edge disjoint union of three
trees such that each vertex of G is contained in exactly two of the trees, and no two subtrees have
the same span.

18.3.2.1 Computation of Independence in G2 (n)

Given a graph G = (V, E), to decide whether E is independent in G2 (G) we can do better than naı̈vely
check the inequality of Laman’s Theorem for all subsets of E. The overbrace function φ (F) =
2|V (F)| − 3 − |F| is submodular on the subsets F of E, φ (F1 ∪ F2 ) + φ (F1 ∩ F2 ) ≤ φ (F1 ) + φ (F2 ) so
we may use any one of the polynomial time algorithms which can find the minimum of a submodular
set function, see [18]. If this minimum is non-negative, then E is independent. The following is a
modification of an algorithm by Edmonds, [12], and makes use directly of the tree decomposition
of Theorem 18.25.
Planar Rigidity 399

INPUT A graph G = (V, E)

OUTPUT A 2-forest decomposition of G when G is independent and the message “DEPENDENT”
when G is dependent.

Step 1 Set F0 = F1 = ∅.
Step 2 If F0 ∪ F1 = E, GOTO Step 7; otherwise, let e be the edge of least index in E − (F0 ∪ F1 ).
Step 3 If e 6∈ hF0 i, replace F0 by F0 + e and GOTO Step 2.

Step 4 If e 6∈ hF1 i, replace F1 by F1 + e and GOTO Step 2.

Step 5 Construct the sequences of nested sets:

F0 ⊇ F2 = F0 ∩ hF1 i ⊇ F4 = F0 ∩ hF3 i ⊇ · · · ;

F1 ⊇ F3 = F1 ∩ hF2 i ⊇ F5 = F1 ∩ hF4 i ⊇ · · · ;
until the sequences become stationary. If the stationary sets are not empty, STOP and output
“DEPENDENT”.

Step 6 Let j be the first index so that e 6∈ hFj i. Let C be the unique cycle contained in Fj (mod 2) + e
and let j0 be the first index so that C − e 6⊆ hFj0 i. Next let e0 be the edge in C − e with smallest
index such that e0 6∈ hFj0 i. Replace Fj (mod 2) by Fj (mod 2) + e − e0 , replace e by e0 and GOTO
Step 3.
Step 7 Construct the sequences of nested sets:

F0 ⊇ F2 = F0 ∩ hF1 i ⊇ F4 = F0 ∩ hF3 i ⊇ · · · ;

F1 ⊇ F3 = F1 ∩ hF2 i ⊇ F5 = F1 ∩ hF4 i ⊇ · · · ;
until the sequences become stationary. If the stationary sets are not empty, STOP and output
“DEPENDENT”, otherwise STOP and output F0 and F1 .

We apply algorithm to the dependent graph in Figure 18.15.

t 7 t
t 8
```
S
A `
5 ` 

A
AS 4 1 9 

ASSt3tHH
2 t



A B 
A 

6 B 11 10
12
AB 
AB 

AB 

ABt

Figure 18.15
A 3-connected planar cycle in G2 (n).
400 Handbook of Geometric Constraint Systems Principles

Example 45 Starting with F0 = F1 = ∅, the algorithm iterates Step 2 eleven times before it can
no longer simply add an edge to either F0 or F1 . At this point, we have F0 = {1, 2, 3, 4, 6, 7} and
F1 = {5, 8, 9, 10, 11}. (See the table at the end of this example.) Edge 12 is in the closures of both F0
and F1 ; so, the algorithm moves on to Step 5 (for the first time). The closures of the sets it generates
are:

hF0 i = E,
hF1 i = {4, 5, 6, 7, 8, 9, 10, 11, 12},
hF2 i = hF0 ∩ hF1 ii = {4, 6, 7, 11, 12},
hF3 i = hF1 ∩ hF2 ii = {11},
hF4 i = hF0 ∩ hF3 ii = ∅.

In step 6, we note j = 3; we add edge 12 to F1 and find the cycle C = {9, 10, 12}; we note j0 = 2: we
then replace edge 9 by edge 12 in F1 and return to Step 3 with e equal to edge 9. Since edge 9 is in
the closures of both F0 and the new F1 , we again proceed to Step 5. This time we get the sequence:

hF0 i = E,
hF1 i = {4, 5, 6, 7, 8, 9, 10, 11, 12},
hF2 i = hF0 ∩ hF1 ii = {4, 6, 7, 11, 12},
hF3 i = hF1 ∩ hF2 ii = {11, 12},
hF4 i = hF0 ∩ hF3 ii = ∅.

This time j equals the “old” j0 or 2. Edge 9 is added to F0 creating the cycle C = {2, 3, 4, 7, 9}.
Then j0 = 1 and edge 9 replaces edge 2 in F0 . We move to Step 3 with e equal to edge 2 and then to
Step 4 where edge 2 is added to F1 . The algorithm then returns to Step 2 and moves directly to Step
7. We have a 2-forest decomposition of E. But, since hF0 i = hF1 i = E, the algorithm terminates with
the message “DEPENDENT”.
In Table 18.3 we list the sequence of sets F0 and F1 which occur.

The algorithm can be altered to check for cycles by noting that in a graph G = (V, E) with
|E| = 2|V | − 2 edges any proper subcycle must be contained in one of the subgraphs of G obtained
by deleting a vertex.
Jacobs and Hendrickson, see [24], use yet another version, similar to Recski’s Theorem, of
Laman’s Theorem, namely the fact that an edge set E is independent in G2 (G(E)) if quadrupling
any one of its edges produces no induced subgraph of average degree larger than 2. They show that
the existence of a pebble covering is equivalent to the independence condition in the quadrupled
edge formulation, where a pebble covering is the result of the Pebble game: Each vertex is given
two pebbles. A vertex can use its pebbles to cover any two edges which are incident to that vertex.
An assignment covering all edges is a pebble covering.
Pebble games were further investigated in [34, 56] and generalized to hypergraphs in [55] and
Cad systems [13].

18.3.3 Pinned Frameworks and Assur Decomposition

18.3.3.1 Isostatic Pinned Framework
A mechanism is a framework with one internal degree of freedom. Given a mechanism, we are
interested in its internal motions, not the trivial ones, so following the mechanical engineers we
pin the framework by prescribing, for example, the coordinates of the endpoints of an edge, or in
general by fixing the position of the vertices of some rigid subgraph, see Figure 18.16. Vertices
with fixed positions are called pinned, the others inner. (Inner vertices are sometimes called free or
Planar Rigidity 401

Table 18.3
Steps of the modified Edmonds Algorithm
iteration F0 F1
of Step 2
0 ∅ ∅
1 {1} ∅
2 {1, 2} ∅
3 {1, 2, 3} ∅
4 {1, 2, 3, 4} ∅
5 {1, 2, 3, 4} {5}
6 {1, 2, 3, 4, 6} {5}
7 {1, 2, 3, 4, 6, 7} {5}
8 {1, 2, 3, 4, 6, 7} {5, 8}
9 {1, 2, 3, 4, 6, 7} {5, 8, 9}
10 {1, 2, 3, 4, 6, 7} {5, 8, 9, 10}
11 {1, 2, 3, 4, 6, 7} {5, 8, 9, 10, 11}
12 {1, 2, 3, 4, 6, 7} {5, 8, 10, 11, 12}
120 {1, 3, 4, 6, 7, 9} {5, 8, 10, 11, 12}
1200 {1, 3, 4, 6, 7, 9} {2, 5, 8, 10, 11, 12}

unpinned in the literature.) Edges among pinned vertices are irrelevant to the analysis of a pinned
framework. We will denote a pinned graph by G(I, P; E), where I is the set of inner vertices, P is
the set of pinned vertices, and E is the set of edges, where each edge has at least one endpoint in I.
A pinned graph G(I, P; E) is said to satisfy the pinned framework conditions if |E| = 2|I| and
for all subgraphs G0 (I 0 , P0 ; E 0 ) the following conditions hold:
(a) |E 0 | ≤ 2|I 0 | if |P0 | ≥ 2,
(b) |E 0 | ≤ 2|I 0 | − 1 if |P0 | = 1 , and
(c) |E 0 | ≤ 2|I 0 | − 3 if P0 = ∅.
We call a pinned graph G(I, P; E) pinned isostatic if E = 2|I| and G ∪ KP is rigid as an unpinned
graph, where KP is a complete graph on a vertex set containing all pins (but no inner vertices).
In other words, we “replace” the pinned vertex set by a complete graph containing the pins and
call G(I, P; E) isostatic, if choosing any basis in that replacement produces an (unpinned) isostatic
graph.
A pinned graph G(I, P; E) realized in the plane, with P for the pins, and p for all the vertices,
is a pinned framework. A pinned framework is rigid if the matrix R(G ∪ KP ) has rank 2|I|, with
the columns corresponding to the vertex set of K p removed, independent if the rows of R(G ∪ KP )
corresponding to E are independent, and isostatic, if it is rigid and independent. The vertices I of a
pinned framework are in generic position if any submatrix of the rigidity matrix is zero only if it is
identically equal to zero with the coordinates of the inner vertices as variables. The coordinates of
the pins are prescribed constants.
Figure 18.16 shows an example of a pinned isostatic G and a corresponding basis of R(G ∪ KP ).

It is common in engineering to choose pins in advance and their placement P might not be
generic, in fact not even in general position, as it is sometimes necessary to have all pins on a line.
The following result shows that this is not a problem.
402 Handbook of Geometric Constraint Systems Principles

G G

a) b)

Figure 18.16
Framework (a) is pinned isostatic because framework (b) is isostatic.

Theorem 18.26 Given a pinned graph G(I, P; E), the following are equivalent:
(i) There exists an isostatic realization of G.
(ii) The Pinned Framework Conditions are satisfied.
(iii) For all placements P of P with at least two distinct locations and all generic positions of I
the resulting pinned framework is isostatic.

A pinned graph G(I, P; E) satisfying the Pinned Framework Conditions must have at least two
pins and in every isostatic realization of G there must be at least two distinct pin locations. Placing
all pins in the same location never yields an isostatic framework, but we can make an important
observation about the degree of freedom of such a “pin collapsed” framework.

Theorem 18.27 Let G(I, P; E) be a pinned graph satisfying the Pinned Framework S Conditions.
Identifying the pinned vertices to one vertex p∗ yields a graph G∗ (V, E), V = I {p∗ } and the
degree of freedom of G∗ is one less than the number of circuits contained in G( (G)∗ ).

Theorem 18.28 Assume G = (I, P; E) is a pinned isostatic graph. Then the following are equiva-
lent:
(i) G = (I, P; E) is minimal as a pinned isostatic graph: that is for all proper subsets of vertices
I 0 ∪ P0 , I 0 ∪ P0 induces a pinned subgraph G0 = (I 0 ∪ P0 , E 0 ) with |E 0 | ≤ 2|I 0 | − 1.
(ii) If the set P is contracted to a single vertex p∗ , inducing the unpinned graph G∗ with edge
set E, then G∗ is a cycle.
(iii) Either the graph has a single inner vertex of degree 2 or each time we delete a vertex, the
resulting pinned graph has a motion of all inner vertices (in generic position).
(iv) Deletion of any edge from G results in a pinned graph that has a motion of all inner vertices
(in generic position).

Condition (i) is a refinement of the Grübler count [42], in a form which is now necessary and
sufficient.
Condition (ii) translates the minimality condition to minimal dependence in G2 (n) and thus
serves as a purely combinatorial description of Assur graphs and may be checked for example by
the modified Edmonds algorithm described in Section 18.3.2.1.
Conditions (iii) and (iv) are similar in nature. Condition (iii) provides the engineer with a quicker
check for the Assur property for smaller graphs than (iv), since there are fewer vertices than edges
to delete. However, condition (iv) tells the engineer that a driver inserted for an arbitrary edge will
(generically) move all inner vertices.
Some examples of Assur graphs are drawn in Figure 18.17 and their corresponding generic
cycles in Figure 18.18.
A general isostatic framework can be decomposed into a partially ordered set of Assur graphs.
This partial order can be represented in an Assur scheme as in Figure 18.22.
Planar Rigidity 403

Figure 18.17
Assur graphs.

Figure 18.18
Corresponding cycles for Assur graphs.

Figure 18.19 shows isostatic pinned frameworks and Figure 18.21 indicates their decomposition.

Figure 18.19
Decomposable (not Assur) graphs.

Under this operation, the Assur graphs will be the minimal, indecomposable graphs. Every
pinned isostatic graph G is a unique composition of Assur graphs, called Assur components of
G. Decomposition by identifying the pins, then deleting the resulting generic cycle and pinning the
vertices of attachment naturally induces a partial order on the Assur components of an isostatic
graph: component A ≤ B if B occurs at a higher level, and B has at least one vertex of A as a pinned
vertex. The algorithm for decomposing the graph guarantees that A ≤ B means that B occurs at a
later stage than A.
This partial order, with the identifications needed for linkage composition, can be used to re-
assemble the graph from its Assur components.
Deleting any edge in an isostatic framework produces a mechanism. Its decomposition into
Assur components permits the analysis of this mechanism in layers. In fact, we can delete an edge
in each Assur component to obtain a pinned framework with several degrees of freedom whose
complex behavior can be simply described by analyzing the individual Assur components. The
404 Handbook of Geometric Constraint Systems Principles

Figure 18.20
The first step of a decomposition for isostatic frameworks in 18.19 –with identified subcircuit(s).

Figure 18.21
Recomposing the pinned isostatic graphs in Figure 18.19 from their Assur components.

engineer thinks of edge deletion as replacing an edge by a driver. This process of adding drivers is
studied in [53].
The dyad is the only Assur graph on three vertices. There is no Assur graph on four vertices. An
Assur graph, whose corresponding generic cycle is K4 is called a basic Assur graph.
To generate all Assur graphs (on five or more vertices) we use Theorem 18.28(ii) together with
Theorem 18.12 to generate all rigidity circuits. To get from a rigidity circuit C to an Assur graph,
we choose a vertex p∗ of C and split it into two or more pins. The choice of p∗ , the splitting of p∗
into a set P of pins (2 ≤ |P| ≤ val(p∗ )), and choosing for each edge incident to p∗ an endpoint from
P allows us to construct several Assur graphs from one generic cycle, see Figure 18.23. We say that
G(I, P; E) and G0 (I, P0 , E) are related by pin rearrangement if G∗ = G0∗ (see Figure 18.23).

I II I

II
IV
IV III
III
V

V
a) b) c)

Figure 18.22
An isostatic pinned framework (a) has a unique decomposition into Assur graphs (b) which is rep-
resented by a partial order or Assur scheme (c).
Planar Rigidity 405

Figure 18.23
Pin rearrangement (maintaining at least two pins).

Theorem 18.29 All Assur graphs on 5 or more vertices can be obtained from basic Assur graphs
by a sequence of edge-splits, pin-rearrangements and 2-sums of smaller Assur graphs.

There are additional operations under which the class of Assur graphs is closed, which are of
interest to the mechanical engineer, for example vertex-split (Figure 18.24).

Figure 18.24
Vertex split taking an Assur graph to an Assur graph.

The inductive constructions for Assur graphs can be used to provide a visual certificate sequence
for an Assur graph. If we are given a sequence of edge-splits and 2-sums starting from a dyad and
ending with G, see Figure 18.25, it is trivial to verify that G is an Assur graph. It is well known,
see [59], that there are exponential algorithms to produce such a certificate. However, all algorithms
mentioned in section 18.3.2.1 can be adapted to verify the Assur property.

Figure 18.25
Certificate sequence for the final Assur graph.
406 Handbook of Geometric Constraint Systems Principles

18.3.4 Body and Pin Structures

Given a simple graph G = (V, E), we may regard G as a body-and-pin graph of a structure in the
plane: The vertices are bodies, edges denote pins. Each pin connects just two bodies. As realizations
consider amorphous bodies. An embedding specifies the location of the pins. One possibility for
bodies in the plane are (long enough) rods, i.e., line segments. In this case all the pins on a body are
collinear and we speak of a pin-collinear body-and-pin-structure. Note that if G = (V, E) is simple,
then a pin collinear structure exists. And can be constructed from any generic embedding of the
structure graph G = (V, E) in R2 . by forming the polar of that embedding, see Figure 18.26.
Jackson and Jordán show in [21] that the body-and-pin and rod-and-pin 2-polymatroids of a
graph are identical.

Theorem 18.30 Let G(V, E) be a multigraph. Then the following statements are equivalent:
(a) G has a realization as an infinitesimally rigid body-and-hinge framework in R2 .
(b) G has a realization as an infinitesimally rigid body-and-hinge framework (G, q) in R2 with
each of the sets of points {q(e) : e ∈ EG (v)}, v ∈ V , collinear.
(c) 2G contains 3 edge disjoint spanning trees.

g
f
e G
A
h
a d
c D
b C H

B
F

Figure 18.26
A graph and its polar with respect to a circle.

Figure 18.26 shows a simple graph. We want to interpret here the vertices as bodies and the edges
prescribing incidences between bodies. If the bodies are to be realized as line segments, the polar of
the graph G with respect to some conic yields the desired realization. It is rigid, because the graph
contains three spanning trees on the edge sets {a, b, d, f , h}, {a, c, e, f , g}, {b, c, d, g, h}, collectively
using any edge of G at most twice, which means that 2G contains 3 edge disjoint spanning trees.
Theorem 18.30 is a special case of the Tay-Whiteley theorem, [58], where the molecular conjec-
ture was formulated: A multigraph has a generically rigid realization as a hinged structure in n-space
if and only it has a rigid realization as a hinged structures in n-space with all hinges of body vi in
a hyperplane Hi of the space. The molecular conjecture was proved by Katoh and Tanigawa [28] in
all dimensions. Tay, [57], extended the Tay-Whiteley Theorem to the case where each hinge can be
shared by more than two bodies. It would be interesting to know if the Katoh-Tanigawa Theorem
can also be extended in the same way.
Planar Rigidity 407

18.3.5 Rigidity of Random Graphs

Let Gn,d denote the probability space of all d-regular graphs on n vertices with the uniform proba-
bility distribution. A sequence of graph properties An holds asymptotically almost surely, or a.a.s.
for short, in Gn,d if limn→∞ PrGn,d (An ) = 1. Graphs in Gn,d are known to be a.a.s. highly connected.
It was shown by Bollobás [5] and Wormald [68] that if G ∈ Gn,d for any fixed d ≥ 3, then G is a.a.s.
d-connected. This result was extended to all 3 ≤ d ≤ n − 4 by Cooper et al. [8] and Krivelevich
et al. [32]. Stronger results hold if we discount “trivial” cutsets. In [67], Wormald shows that if if
G ∈ Gn,d for any fixed d ≥ 3, then G is a.a.s. cyclically (3d − 6)-edge-connected. Together with the
fact, shown in [23] that if G = (V, E) is a cyclically 5-edge-connected 4-regular graph, then G is
globally rigid, this immediately gives:

Theorem 18.31 ( [23]) If G ∈ Gn,d and d ≥ 4 then G is a.a.s. globally rigid.

For a random d-regular graph G on n vertices the expected number, EX j , of subgraphs H of

minimum degree three of order j and size 2 j − k in an m-edged subgraph of G was calculated
in [54].
Considering log(EX j )/n as a function of m/n and j/n the graphs in Figure 18.27 for d = 5 and
d = 6 were obtained. To see for which m/n and j/n the expected number EX j is roughly one, or
log(EX j )/n = 0, we have inserted the plane z = 0 in the graphs.

m j m j
n n n n
Figure 18.27
log(EX j ) as function of m/n and j/n; d = 5 (left), d = 6 (right).

From Figure 18.27 one might predict that as edges are percolated randomly into a graph to obtain
a d-regular graph, that the rigidity phase transition will be first order, since small rigid subgraphs
are unlikely.
Let G(n, p) denote the probability space of all graphs on n vertices in which each edge is chosen
independently with probability p.

Theorem 18.32 (Jackson, Servatius, and Servatius, 2007 [23]) Let G ∈ G(n, p), where p =
(log n + k log log n + w(n))/n, and limn→∞ w(n) = ∞.
(a) If k = 2 then G is a.a.s. rigid.
(b) If k = 3 then G is a.a.s. globally rigid.

Rigidity percolation in this Erdős-Rényi model was studied in [27], where it is shown that there
exists a sharp threshold for a giant rigid component to emerge.
408 References

Glossary
2-tree decomposition: Partition of the edge set of G into two spanning trees.
Body pin graph: A graph in which the vertices represent rigid bodies and an edge indicates that
the two bodies are pinned together at a point.
Grübler count: An engineering rule of thumb to measure the degree of freedom (ignoring con-
straint independence) of a system. The Grübler count of a body pin framework in the plane
adds 3 for each body and subtracts 2 for each pin, and subtracts 3 for the isometries of the
plane.
Mechanism: a framework with one internal degree of freedom, (k’th order mechanism, with k
degrees of freedom, is also sometimes used.)
Pinned Framework: A framework with at least two vertices of fixed position.
proper 2-tree decomposition: A 2-tree decomposition of G such that no pair of proper subtrees,
excepting single vertices, have the same span.
edge k-rigid: A rigid graph such that the removal of any k − 1 edges yields a framework with the
same type of rigidity.
edge birigid, edge 2-rigid, redundantly rigid: A k-rigid framework (graph) with k = 2.
(vertex) k-rigid: G has at least k + 1 vertices such that the removal of any k − 1 vertices, together
with their incident edges, leaves a rigid graph.
(vertex) birigid, vertex 2-rigid: A vertex 2-rigid framework (graph).
spanning forest of G: A subgraph of a graph G which is cycle free and contains all the vertices of
G. Its connected components are trees and isolated vertices.
submodular function: Function defined on P(E) satisfying φ (F1 ∪ F2 ) + φ (F1 ∩ F2 ) ≤ φ (F1 ) +
φ (F2 ).

References
[1] Alex R. Berg and Tibor Jordán. A proof of Connelly’s conjecture on 3-connected circuits of
the rigidity matroid. J. Combin. Theory Ser. B, 88(1):77–97, 2003.
[2] E. D. Bolker and B. Roth. When is a bipartite graph a rigid framework? Pacific J. Math.,
90(1):27–44, 1980.
[3] Ethan D. Bolker. Bracing rectangular frameworks. II. SIAM J. Appl. Math., 36(3):491–508,
1979.
[4] Ethan D. Bolker and Henry Crapo. Bracing rectangular frameworks. I. SIAM J. Appl. Math.,
36(3):473–490, 1979.
[5] Béla Bollobás. Random graphs, volume 73 of Cambridge Studies in Advanced Mathematics.
Cambridge University Press, Cambridge, second edition, 2001.
References 409

[6] Robert Connelly. Rigidity. In Handbook of convex geometry, Vol. A, B, pages 223–271.
North-Holland, Amsterdam, 1993.
[7] Robert Connelly and Herman Servatius. Higher-order rigidity—what is the proper definition?
Discrete Comput. Geom., 11(2):193–200, 1994.
[8] Colin Cooper, Alan Frieze, and Bruce Reed. Random regular graphs of non-constant degree:
connectivity and Hamiltonicity. Combin. Probab. Comput., 11(3):249–261, 2002.
[9] Henry Crapo. On the generic rigidity of plane frameworks. Research Report RR-1278, IN-
RIA, August 1990. Projet ICSLA.
[10] Luigi Cremona. Le figure reciproche. Civilità delle Macchine, 4(5):55–62, 1956.
[11] Reinhard Diestel. Graph theory, volume 173 of Graduate Texts in Mathematics. Springer,
Heidelberg, fourth edition, 2010.
[12] Jack Edmonds. Minimum partition of a matroid into independent subsets. J. Res. Nat. Bur.
Standards Sect. B, 69B:67–72, 1965.
[13] James Farre, Helena Kleinschmidt, Jessica Sidman, Audrey St. John, Stephanie Stark, Louis
Theran, and Xilin Yu. Algorithms for detecting dependencies and rigid subsystems for CAD.
Comput. Aided Geom. Design, 47:130–149, 2016.
[14] F. G. Frobenius. Über zerlegbare Determinanten. Sitzungsberichte der Berl. Akademie, XVIII,
1917.
[15] Jack Graver, Brigitte Servatius, and Herman Servatius. Combinatorial rigidity, volume 2 of
Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 1993.
[16] Jack E. Graver. Rigidity matroids. SIAM J. Discrete Math., 4(3):355–368, 1991.
[17] Jack E. Graver, Brigitte Servatius, and Herman Servatius. Abstract rigidity in m-space. In
Jerusalem combinatorics ’93, volume 178 of Contemp. Math., pages 145–151. Amer. Math.
Soc., Providence, RI, 1994.
[18] Martin Grötschel, László Lovász, and Alexander Schrijver. Geometric algorithms and combi-
natorial optimization, volume 2 of Algorithms and Combinatorics: Study and Research Texts.
Springer-Verlag, Berlin, 1988.
[19] E. W. Hobson, H. P. Hudson, A. N. Singh, and A. B. Kempe. Squaring the Circle and Other
Monographs. Chelsea Publishing Company, New York, NY, 1953.
[20] Bill Jackson and Tibor Jordán. Connected rigidity matroids and unique realizations of graphs.
J. Combin. Theory Ser. B, 94(1):1–29, 2005.
[21] Bill Jackson and Tibor Jordán. On the rigidity of molecular graphs. Combinatorica,
28(6):645–658, 2008.
[22] Bill Jackson and Tibor Jordán. A sufficient connectivity condition for generic rigidity in the
plane. Discrete Appl. Math., 157(8):1965–1968, 2009.
[23] Bill Jackson, Brigitte Servatius, and Herman Servatius. The 2-dimensional rigidity of certain
families of graphs. J. Graph Theory, 54(2):154–166, 2007.
[24] Donald J. Jacobs and Bruce Hendrickson. An algorithm for two-dimensional rigidity perco-
lation: the pebble game. J. Comput. Phys., 137(2):346–365, 1997.
[25] Tibor Jordán and Viktória E. Kaszanitzky. Highly connected rigidity matroids have unique
underlying graphs. European J. Combin., 34(2):240–247, 2013.
410 References

[26] Gil Kalai. Hyperconnectivity of graphs. Graphs Combin., 1(1):65–79, 1985.

[27] Shiva Prasad Kasiviswanathan, Cristopher Moore, and Louis Theran. The rigidity transition
in random graphs. In Proceedings of the Twenty-Second Annual ACM-SIAM Symposium on
Discrete Algorithms, pages 1237–1252. SIAM, Philadelphia, PA, 2011.
[28] Naoki Katoh and Shin-ichi Tanigawa. A proof of the molecular conjecture. Discrete Comput.
Geom., 45(4):647–700, 2011.
[29] Shuji Kijima and Shin-ichi Tanigawa. Sparsity and connectivity of medial graphs: concerning
two edge-disjoint Hamiltonian paths in planar rigidity circuits. Discrete Math., 312(16):2466–
2472, 2012.
[30] Csaba Király and Ferenc Péterfalvi. Balanced generic circuits without long paths. Discrete
Math., 312(15):2262–2271, 2012.
[31] E. Kötter. Über die Möglichkeit n Punkte in der Ebene oder im Raum durch weniger als
2n − 3 oder 3n − 6 Stäbe von ganz unveränderlicher Länge unverschieblich miteinander zu
verbinden. Festschrift für H. Mueller-Breslau, 1912.
[32] Michael Krivelevich, Benny Sudakov, Van H. Vu, and Nicholas C. Wormald. Random regular
graphs of high degree. Random Structures Algorithms, 18(4):346–363, 2001.
[33] G. Laman. On graphs and rigidity of plane skeletal structures. J. Engrg. Math., 4:331–340,
1970.
[34] Audrey Lee and Ileana Streinu. Pebble game algorithms and sparse graphs. Discrete Math.,
308(8):1425–1437, 2008.
[35] L. Lovász and Y. Yemini. On generic rigidity in the plane. SIAM J. Algebraic Discrete
Methods, 3(1):91–98, 1982.
[36] H. Maehara and N. Tokushige. When does a planar bipartite framework admit a continuous
deformation? Theoret. Comput. Sci., 263(1-2):345–354, 2001. Combinatorics and computer
science (Palaiseau, 1997).
[37] J. Clerk Maxwell. On Reciprocal Diagrams in Space, and their relation to Airy’s Function of
Stress. Proc. London Math. Soc., S1-2(1):58.
[38] John Milnor. Singular points of complex hypersurfaces. Annals of Mathematics Studies, No.
61. Princeton University Press, Princeton, NJ, 1968.
[39] C. St. J. A. Nash-Williams. Edge-disjoint spanning trees of finite graphs. J. London Math.
Soc., 36:445–450, 1961.
[40] C. St. J. A. Nash-Williams. Decomposition of finite graphs into forests. J. London Math. Soc.,
39:12, 1964.
[41] Viet-Hang Nguyen. On abstract rigidity matroids. SIAM J. Discrete Math., 24(2):363–369,
2010.
[42] Robert L. Norton. Design of Machinery: An Introduction To The Synthesis and Analysis of
Mechanisms and Machines. McGraw Hill, New York, 2004.
[43] James G. Oxley. Matroid theory. Oxford Science Publications. The Clarendon Press, Oxford
University Press, New York, 1992.
[44] Sachin Patkar, Brigitte Servatius, and K. V. Subrahmanyam. Abstract and generic rigidity in
the plane. J. Combin. Theory Ser. B, 62(1):107–113, 1994.
References 411

[45] Kevin Peterson. The stress spaces of bipartite frameworks. Pacific J. Math., 197(1):173–182,
2001.
[46] H. Pollaczek-Geiringer. Über die Gliederung ebener Fachwerke. ZAMM - Journal of Ap-
plied Mathematics and Mechanics / Zeitschrift für Angewandte Mathematik und Mechanik,
7(1):58–72, 1927.
[47] András Recski. A network theory approach to the rigidity of skeletal structures. I. Modelling
and interconnection. Discrete Appl. Math., 7(3):313–324, 1984.
[48] András Recski. A network theory approach to the rigidity of skeletal structures. II. Laman’s
theorem and topological formulae. Discrete Appl. Math., 8(1):63–68, 1984.
[49] András Recski. Bracing cubic grids—a necessary condition. In Proceedings of the Oberwol-
fach Meeting “Kombinatorik” (1986), volume 73, pages 199–206, 1989.
[50] András Recski. Matroid theory and its applications in electric network theory and in stat-
ics, volume 6 of Algorithms and Combinatorics. Springer-Verlag, Berlin; Akadémiai Kiadó
(Publishing House of the Hungarian Academy of Sciences), Budapest, 1989.
[51] Brigitte Servatius and Peter R. Christopher. Construction of self-dual graphs. Amer. Math.
Monthly, 99(2):153–158, 1992.
[52] Brigitte Servatius and Herman Servatius. On the 2-sum in rigidity matroids. European J.
Combin., 32(6):931–936, 2011.
[53] Brigitte Servatius, Offer Shai, and Walter Whiteley. Geometric properties of Assur graphs.
European J. Combin., 31(4):1105–1120, 2010.
[54] Brigitte Servatius and Nicholas Wormald. On the size and number of rigid subgraphs of
d-regular graphs. Preprint.
[55] Ileana Streinu and Louis Theran. Sparse hypergraphs and pebble game algorithms. European
J. Combin., 30(8):1944–1964, 2009.
[56] Ileana Streinu and Louis Theran. Sparsity-certifying graph decompositions. Graphs Combin.,
25(2):219–238, 2009.
[57] Tiong-Seng Tay. Linking (n − 2)-dimensional panels in n-space. II. (n − 2, 2)-frameworks and
body and hinge structures. Graphs Combin., 5(3):245–273, 1989.
[58] Tiong-Seng Tay and Walter Whiteley. Recent advances in the generic rigidity of structures.
Structural Topology, (9):31–38, 1984. Dual French-English text.
[59] Tiong-Seng Tay and Walter Whiteley. Generating isostatic frameworks. Structural Topology,
(11):21–69, 1985. Dual French-English text.
[60] W. T. Tutte. On the problem of decomposing a graph into n connected factors. J. London
Math. Soc., 36:221–230, 1961.
[61] W. T. Tutte. Connectivity in graphs. Mathematical Expositions, No. 15. University of Toronto
Press, Toronto, Ont.; Oxford University Press, London, 1966.
[62] W. T. Tutte. Connectivity in matroids. Canad. J. Math., 18:1301–1324, 1966.
[63] D. J. A. Welsh. Matroid theory. Academic Press [Harcourt Brace Jovanovich, Publishers],
London-New York, 1976. L. M. S. Monographs, No. 8.
[64] Walter Whiteley. Matroids and rigid structures. In Matroid applications, volume 40 of Ency-
clopedia Math. Appl., pages 1–53. Cambridge Univ. Press, Cambridge, 1992.
412 References

[65] Walter Whiteley. Some matroids from discrete applied geometry. In Matroid theory (Seattle,
WA, 1995), volume 197 of Contemp. Math., pages 171–311. Amer. Math. Soc., Providence,
RI, 1996.
[66] Hassler Whitney. 2-Isomorphic Graphs. Amer. J. Math., 55(1-4):245–254, 1933.
[67] N. C. Wormald. Models of random regular graphs. In Surveys in combinatorics, 1999 (Can-
terbury), volume 267 of London Math. Soc. Lecture Note Ser., pages 239–298. Cambridge
Univ. Press, Cambridge, 1999.
[68] Nicholas C. Wormald. The asymptotic connectivity of labelled regular graphs. J. Combin.
Theory Ser. B, 31(2):156–167, 1981.
Chapter 19
Inductive Constructions for
Combinatorial Local and Global Rigidity
Anthony Nixon
Department of Mathematics and Statistics, Lancaster University, U.K.

Elissa Ross
MESH Consultants Inc., Fields Institute, Toronto, Canada

CONTENTS
19.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
19.2 Rigidity in Rd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
19.2.1 Inductive Operations on Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
19.2.2 Recursive Characterizations of Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 418
19.2.3 Combinatorial Characterizations of Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
19.3 Body-Bar, Body-Hinge, Molecular, etc. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422
19.3.1 Geometry and Combinatorics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
19.3.2 Characterizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
19.4 Further Rigidity Contexts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
19.4.1 Frameworks with Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
19.4.2 Infinite Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
19.4.3 Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
19.4.4 Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
19.4.5 Applications of Rigidity Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
19.4.6 Direction-Length Frameworks and CAD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
19.4.7 Nearly Generic Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
19.5 Summary Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 430

19.1 Introduction
A bar-joint framework is a geometric realization of a graph. Formally, a framework is a pair (G, p)
where G = (V, E) is a (finite, simple) graph and p : V → Rd . The framework (G, p) is rigid if every
edge-length-preserving continuous deformation of the vertices arises from an isometry of Rd (see
Chapter 18). A stronger condition is global rigidity, where (G, p) is the only realization of G in Rd ,
up to isometries, with the edge lengths prescribed by p (see Chapter 21).
Determining the rigidity, or global rigidity, of a given framework is NP-hard [1, 69]. However,
the situation improves for generic frameworks where one can linearize the problem and charac-

413
414 Handbook of Geometric Constraint Systems Principles

terize generic rigidity via the rank of the rigidity matrix [3]. A framework (G, p) is generic if the
coordinates of p form an algebraically independent set over Q.
A key topic in rigidity theory, perhaps the fundamental topic, is to characterize generic rigidity,
and generic global rigidity, in purely combinatorial terms. That is, we seek to characterize generic
rigidity as a property of the graph alone. Cornerstone theorems in rigidity theory are such character-
izations of rigidity [51] and global rigidity [8, 28] for generic frameworks in R2 . In both cases the
proof was obtained via inductive, or recursive, constructions.
Our focus will be on such constructions both for these fundamental results and for related prob-
lems in various areas of rigidity theory. Since the first of these results in 1970 [51], there has been a
multitude of papers using inductive constructions for a variety of problems. We will briefly describe
a number of these. (In a number of the results we state, the geometric operations do not need the full
strength of generic; general position, where no d + 1 points define a (d − 1)-dimensional hyperplane
often suffices.) One of our aims is to show the versatility of proofs via inductive construction. In
particular, how similar arguments can be adapted to seemingly disparate areas of rigidity theory.
An inductive construction has two distinct aspects to it. The first aspect is to prove a constructive
characterization of the class of graphs. That is, to show that a short list of (typically) local operations
is sufficient to characterize a class of graphs, in the sense that every graph in the class, and no others,
can be built recursively from a (or a few) “base graph(s)” using only these local operations. A key
starting point in such characterizations is often to establish what the minimum degree is for such
a graph and for each possibility provide a way of adding a vertex with such a degree. The second
aspect then considers these graph operations as framework operations. Here the goal is to show that
the operation preserves generic rigidity. This is done by induction, first by showing directly that a
generic realization of the base graph is rigid and then by showing that applying any of the operations
to a rigid framework results in a larger rigid framework. A typical technique for this is to use the
matrix characterization of rigidity and often involves finding a special, non-generic, realization that
is rigid and then using the fact that generic frameworks have maximal rank to conclude the rigidity
of a generic realization.
Alongside the other chapters in this book, we direct the reader to [19, 39, 62, 84, 85] for a wealth
of further information and references about rigidity and to [16, 49] for details on related constructive
characterizations in combinatorial optimization. Since we will not put any focus on 1-dimensional
frameworks we direct the interested reader to [27].
We close the introduction by briefly outlining what follows. In Section 19.2 we describe results
relevant to the rigidity and global rigidity of generic (bar-joint) frameworks in Rd . We separate
these results into three kinds: geometric operations on frameworks; combinatorial results about
graphs; and characterizations of rigidity in combinatorial terms (for this last point we can only
discuss R2 ). In Section 19.3 we repeat this 3-step analysis for special classes of frameworks, such
as body-bar frameworks in Rd where we have a much fuller combinatorial understanding of rigidity.
In Section 19.4 we try to survey all corners of the rigidity literature that have been influenced by
inductive construction techniques. Necessarily this means that, in most cases, we give little more
than references for the interested reader to follow.
Notation: G = (V, E) is a finite simple graph. When loops and multiple edges are permitted we
will explicitly use the term multigraph.

19.2 Rigidity in Rd
In this section we will describe results relevant to inductive constructions used in the study of rigidity
of bar-joint frameworks in Euclidean spaces. We split our analysis into two sections: geometric and
combinatorial.
Inductive Constructions for Combinatorial Local and Global Rigidity 415

0-extension: The operation of (d-dimensional) 0-extension forms a new graph G0 from G by adding
a new vertex vn+1 with d edges incident to distinct vertices of G.
1-extension: The operation of (d-dimensional) 1-extension forms a new graph G0 from G by delet-
ing an edge vi v j and adding a new vertex vn+1 and d + 1 new edges incident to vi , v j and d − 1
further distinct vertices of G.
0-reduction: The operation of (d-dimensional) 0-reduction forms a new graph G0 from G by delet-
ing a degree d vertex v from G.
1-reduction: The operation of (d-dimensional) 1-reduction forms a new graph G0 from G by delet-
ing a degree d + 1 vertex v from G and adding an edge between two vertices formerly adjacent
to v.

X-replacement: The operation of (d-dimensional) X-replacement forms a new graph G0 from G by

deleting two non-adjacent edges vi v j , vk v` and adding a new vertex vn+1 and d + 2 new edges
incident to vi , v j , vk , v` and d − 2 further distinct vertices of G.
vertex splitting: Let v ∈ V have N(v) = {u1 , . . . , um }. A vertex splitting operation (in d-
dimensions) on v removes v and its incident edges and adds vertices v0 , v1 and edges
u1 v0 , u2 v0 , u1 v1 , u2 v1 , . . . , ud−1 v0 , ud−1 v1 , v0 v1 and rearranges the edges ud v, . . . , um v in some
way into edges ui v j for i ∈ {d, . . . , m} and j ∈ {0, 1}.

For each of the above definitions, the dimension corresponds to the dimension we intend to use
these graph operations as operations on frameworks. Since this is usually clear from the context we
typically drop “d-dimensional” for brevity when referring to these operations.

V -replacement: The operation of V -replacement (note only applicable in 3-dimensions) forms a

new graph G0 from G by deleting two adjacent edges vi v j , v j vk and adding a new vertex vn+1
and 5 new edges incident to vi , v j , vk and 2 further distinct vertices of G.
vertex-to-4-cycle: Let v ∈ V have N(v) = {u1 , . . . , um }. A vertex-to-4-cycle operation (note only
applicable in 2-dimensions) on v removes v and its incident edges and adds vertices v0 , v1 and
edges u1 v1 , v1 u2 , u2 v0 , v0 u1 and rearranges the edges u3 v, . . . , um v in some way into edges ui v j
for i ∈ {3, . . . , m} and j ∈ {0, 1, }.

Our remaining definitions concern graphs.

(2, 3)-sparse: G is (2, 3)-sparse if |E 0 | ≤ 2|V 0 | − 3 for all subgraphs (V 0 , E 0 ) of G with |E 0 | > 0.
(2, 3)-tight: G is (2, 3)-tight if |E| = 2|V | − 3 and G is (2, 3)-sparse.
(k, `)-sparse: G is (k, `)-sparse if |E 0 | ≤ k|V 0 | − ` for all subgraphs (V 0 , E 0 ) of G with |V 0 | ≥ k.
(k, `)-tight: G is (k, `)-tight if |E| = k|V | − ` and G is (k, `)-sparse.

(2, 3)-circuit: G is a (2, 3)-circuit if |E| = 2|V | − 2 and every proper subgraph of G is (2, 3)-sparse.
admissible: A degree 3 vertex v in a (2, 3)-circuit is admissible if there is a 1-reduction at v which
results in a (2, 3)-circuit.
feasible: A degree 3 vertex v in a 3(-vertex)-connected (2, 3)-circuit is feasible if there is a 1-
reduction at v which results in a 3-connected (2, 3)-circuit.
redundantly rigid graph: G is redundantly rigid if G − e contains a spanning (2, 3)-tight subgraph
for all e ∈ E.
416 Handbook of Geometric Constraint Systems Principles

vn+1
vi vi vn+1

vj vj

(a) 0-extension (b) 1-extension

Figure 19.1
The fundamental moves in two dimensions.

planar 1-extension: A 1-extension is planar if G and G0 are both planar.

planar vertex splitting: A vertex split is planar if G and G0 are both planar.
Laman plus one graph: G is a Laman plus one graph if there exists e ∈ E such that G − e is
(2, 3)-tight.
2-sum: A 2-sum of two graphs Gi = (Vi , Ei ), for i = 1, 2, over the pair x, y forms the graph G =
(V, E) where V = V1 ∪V2 with V1 ∩V2 = {x, y}, xy ∈ Ei for i = 1, 2 and E = (E1 ∪ E2 ) − xy. A
2-separation is the inverse of a 2-sum.

M-connected: A graph G is M-connected if the rigidity matroid of a generic framework (G, p) is

connected.

19.2.1 Inductive Operations on Frameworks

In this subsection we will consider a variety of local operations on graphs and when they are known
to preserve the rigidity or global rigidity of frameworks. Our focus will be on rigidity (and global
rigidity) of bar-joint frameworks in Rd .
0- and 1-extensions (see Figure 19.1) are so fundamental they have been described in the litera-
ture under a number of different names, such as Henneberg-1 and -2 moves [24], or vertex addition
and edge-splitting [84]. Together they may be called Henneberg moves [24]. Both are well known to
preserve rigidity in all dimensions. The first is trivially proved via considering the rank of a matrix.
The second has been proved in several different ways, see, for example, [51, 79].

Lemma 19.1
Let (G, p) be a rigid framework in Rd . Let G0 be formed from G by a 0-extension. Then (G0 , p0 ) is
generically rigid in Rd .

Lemma 19.2
Let (G, p) be a generically rigid framework in Rd . Let G0 be formed from G by a 1-extension. Then
(G0 , p0 ) is generically rigid in Rd .

Several proofs of Lemma 19.2 can be extended to prove our next result, which is only known in
R2 . Indeed an example from [19] confirms the analogue in dimension at least 4 is false.

Lemma 19.3
Let (G, p) be a generically rigid framework in R2 . Let G0 be formed from G by X-replacement.
Then (G0 , p0 ) is generically rigid in R2 .
Inductive Constructions for Combinatorial Local and Global Rigidity 417

vi vi vn+1 vi vi vn+1
vk vk vk vk
vj v vj v vj
vj
(a) X-replacement (b) V -replacement

Figure 19.2
X-replacement (a) and V -replacement (b) in three dimensions.

v1
v v0 v1 v v0

v v0 v1

(a) Vertex Split 2D (b) Vertex to four-cycle (c) VertexSplit 3D

Figure 19.3
Vertex splitting in two and three dimensions, and the two-dimensional vertex-to-four-cycle move.

The 3-dimensional version of X-replacement (see Figure 19.2(a)) is unclear.

Conjecture 3 ([79]) Let (G, p) be a generically rigid framework in R3 . Let G0 be formed from G
by X-replacement. Then (G0 , p0 ) is generically rigid in R3 .

A number of special cases have now been proved [10, 19, 26], however the general case still
seems to be difficult. A resolution to this conjecture would shed great light on the 3D rigidity
problem. The sister operation, V -replacement (see Figure 19.2(b)), is known not to preserve rigidity
but for purely graph theoretical reasons (for example it may turn a degree 3 vertex at the point of
the V into a degree 2 vertex). Tay and Whiteley [79] proposed a “double V ” conjecture which gets
around the obvious combinatorial defect and leaves a geometric conjecture analogous to Conjecture
3. Positive resolution of both of these conjectures would lead to a combinatorial characterization of
generic rigidity in R3 .
We mention one more fundamental operation which is known to preserve rigidity in all dimen-
sions, Whiteley’s vertex splitting operation, Figure 19.3(a).

Proposition 19.1 Whiteley [83]

Let (G, p) be a generically rigid framework in Rd . Let G0 be formed from G by vertex splitting.
Then (G0 , p0 ) is generically rigid in Rd .

Vertex splitting was also crucial in work of Finbow-Singh, Ross and Whiteley analyzing block
and hole polyhedra [15]. There has been recent work on triangulated surfaces in [11] using inductive
techniques (in particular vertex splitting) to extend this. Also recently global rigidity has been char-
acterized for a number of triangulated surfaces (sphere, torus, projective plane) again using vertex
splitting [44].
A variant of the 2-dimensional version of vertex splitting, which we call a vertex-to-4-cycle
move (see Figure 19.3(b)), is also known to preserve rigidity.
418 Handbook of Geometric Constraint Systems Principles

Proposition 19.2 Lomeli, Moshe and Whiteley [50]

Let (G, p) be a generically rigid framework in R2 . Let G0 be formed from G by a vertex-to-4-cycle
move. Then (G0 , p0 ) is generically rigid in R2 .

This move has also been extended to other rigidity contexts [58].
We now move on to consider global rigidity. It is quickly apparent that 0-extension does not
preserve global rigidity. However 1-extension does.

Theorem 19.1 (Connelly [8]) Let (G, p) be a generically globally rigid framework in Rd . Let G0
be formed from G by 1-extension. Then (G0 , p0 ) is generically globally rigid in Rd .

Here the proof is far more intricate than the proof of Lemma 19.2. Connelly [8] proved this result
as a corollary of his sufficient condition for a generic framework to be globally rigid. He did this by
utilizing equilibrium stresses (also known as self-stresses) and, in particular, a natural stress matrix
whose rank being maximal, along with the framework being rigid, guarantees global rigidity. An
alternative, more direct proof, was later given by Jackson, Jordán, and Szabadka [33] in their work
on globally linked pairs. Both proofs rely essentially on the coordinates of p being algebraically
independent.
It is easy to check that (d +1)-connectivity can fail when we apply X-replacement and, this being
a necessary condition for global rigidity [22], X-replacement does not preserve global rigidity.
Vertex splitting can create a degree d vertex so clearly does not, in general, preserve global rigid-
ity. Ruling this out is known to be sufficient for vertex splitting to preserve generic global rigidity
in R2 . However the proof, found in [43], is as a consequence of combinatorial characterizations of
rigidity and global rigidity (see Subsection 19.2.3). In higher dimensions the problem is open.

Conjecture 4 (Connelly and Whiteley [7]) Let (G, p) be a generically globally rigid framework
in Rd . Let G0 be formed from G by a vertex splitting operation such that G0 has minimum degree
d + 1. Then (G0 , p0 ) is generically globally rigid in Rd .

Insight into why this is difficult is given by Connelly in [9]. In particular he proves a partial
result [9, Theorem 29] which reduces the problem to proving that (G0 , p0 ) is generically redundantly
rigid (when (G, p) is generically globally rigid). There has been very recent progress toward solving
Conjecture 4. Jordán and Tanigawa [44] have used a “non-degenerate stress” idea to prove that every
graph generated from Kd+2 by a sequence of vertex splitting operations (that preserve the minimum
degree requirement) is globally rigid.
Finally let us mention that Connelly proved a method of combining two generically globally
rigid frameworks in dimension d, by identifying d + 1 common vertices, that results in a generically
globally rigid framework [5].

19.2.2 Recursive Characterizations of Graphs

In this subsection we consider our operations graph theoretically. The key aim being to understand
intricate classes of graphs by building any member using these simple operations. The majority of
this subsection focuses on two dimensions, corresponding results on the line are typically easy and
analogues for dimension ≥ 3 are only partially understood. We do discuss higher dimensions: for
bar-joint frameworks briefly at the end of this subsection; and for special classes of frameworks in
the next section.
Our first result shows that 0- and 1-extensions are sufficient to characterize the class of graphs
relevant to rigidity in R2 .

Theorem 19.2 (Henneberg, Laman [24, 51]) G is (2, 3)-tight if and only if G can be generated
recursively from K2 using only the operations of 0- and 1-extension.
Inductive Constructions for Combinatorial Local and Global Rigidity 419

(a) (2, 3)-tight (b) (2, 3)-tight (c) Not (2, 3)-tight

Figure 19.4
Examples (and non-examples) of (2, 3)-tight graphs.

Figure 19.5
Sample construction sequence of the triangular prism: 0-extension followed by three 1-extensions.

Jackson and Jordán [28] mildly strengthened this result by proving that a (2, 3)-tight graph can
be generated from any (2, 3)-tight subgraph using 0- and 1-extensions.
The proof of Theorem 19.2, starts with the simple fact that every (2, 3)-tight graph, with |V | ≥ 3,
has minimum degree in the set {2, 3}. Consider a vertex of minimum degree. If it is degree 2, apply
a 0-reduction and the result is always (2, 3)-tight. If it is degree 3, we need a modicum of care: there
are three possible new edges in a 1-reduction and it may be that only one of those results in a (2, 3)-
tight graph. Nevertheless every degree 3 vertex is reducible; for example if every possible new edge
already exists, there is a copy of K4 and hence the graph is not (2, 3)-tight, (Figure 19.4(c)). Figure
19.5 constructs the triangular prism from K2 .
This result has been extended to prove the stronger fact that we can generate all planar (2, 3)-
tight graphs using these operations: insisting that every intermediate graph is also planar. (To do this
we can no longer apply arbitrary 1-extensions, but 1-extensions that preserve planarity suffice.)

Theorem 19.3 (Haas et al [21]) G is planar and (2, 3)-tight if and only if G can be generated re-
cursively from K2 using only the operations of planar 0- and planar 1-extension.

For planar (2, 3)-tight graphs there is an alternative characterization using vertex splitting in-
stead. Here the reduction relies on the guarantee of a triangular face (in fact two triangular faces are
guaranteed) in a planar (2, 3)-tight graph distinct from K2 .
420 Handbook of Geometric Constraint Systems Principles
v
x
x
u

w
y y
(a) A (2, 3)-circuit with no (b) In this (2, 3)-circuit, v is
admissible degree 3 vertex. admissible but not feasible
whereas u and w are both
feasible.

Figure 19.6
Examples of (2, 3)-circuits.

Theorem 19.4 (Fekete, Jordan and Whiteley [13]) G is planar and (2, 3)-tight if and only if G
can be generated recursively from K2 using only planar vertex splitting.

This result was also implicit in Owen and Power’s [65] study of 3-connected, planar (2, 3)-tight
graphs; which they showed could be generated using either vertex splitting or one other move. In
[2] an analogue is given for planar (2, 2)-tight graphs (using the vertex-to-K4 move, the inverse of
the operation of contracting a copy of K4 , in addition to vertex splitting) and used to prove results
about contacts of circular arcs in the plane.
The 0- and 1-extension operations are even sufficient to characterize the following special class
of (2, 2)-tight graphs.

Theorem 19.5 (Haas et al [21]) G is Laman plus one if and only if G can be generated recursively
from K4 using only the operations of 0- and 1-extension.

For a related class, the (2, 3)-circuits, the minimum degree is 3 so 0-extension is of no use. In-
stead we have the following theorem, which was the first combinatorial step towards understanding
globally rigid graphs in the plane. Note that it is not true that every degree 3 vertex can be reduced.
For example the graph in Figure 19.6(a) has no admissible vertex. The key to the theorem is isolating
when some degree 3 vertex can be reduced.

Theorem 19.6 (Berg and Jordán [4]) Every (2, 3)-circuit, distinct from K4 , contains at least 3 ad-
missible degree 3 vertices.

In the following the 2-sum operation, as is consistent with its usage in matroid theory, glues two
(2, 3)-circuits together along an edge and deletes the common edge. Servatius and Servatius [73]
showed further that the rigidity matroid is not closed under 2-sum decomposition.

Theorem 19.7 (Berg and Jordán [4]) G is a (2, 3)-circuit if and only if G can be generated recur-
sively from disjoint copies of K4 by applying 1-extensions within connected components and taking
2-sums of connected components.

Berg and Jordán also extended their result to show that 1-extension alone is sufficient to generate
all 3-connected (2, 3)-circuits. The key being that they extended Theorem 19.6 to guarantee at least
2 feasible nodes. (Figure 19.6(b) gives an example with an admissible node v which is not feasible;
the example does, though, contain feasible nodes.)

Theorem 19.8 (Berg and Jordán [4]) Every 3-connected (2, 3)-circuit can be generated from K4
by 1-extensions.
Inductive Constructions for Combinatorial Local and Global Rigidity 421

A detailed argument using ear decompositions of the (2, 3)-sparse matroid extended this to the
following crucial characterization.
Theorem 19.9 (Jackson and Jordán [28]) Let G be M-connected and let G0 be formed from G by
a 1-extension. Then G0 is M-connected. Moreover if G is M-connected and 3-connected then G can
be generated from K4 by 1-extensions and edge additions.
This quickly implies the following result.
Theorem 19.10 (Jackson and Jordán [28]) G is 3-connected and redundantly rigid if and only if
G can be generated recursively from K4 using only the operations of 1-extension and edge addition.
To understand the difficulty in extending Theorem 19.8 to Theorem 19.9, it is instructive to
note, first, that there are 3-connected redundantly rigid graphs which do not contain a spanning
(2, 3)-circuit and, second, that it is not immediately clear that every 3-connected and “minimally”
M-connected graph even contains a vertex of degree 3.
It seems to be hard to find a recursive construction for the class of redundantly rigid graphs.
Weaken this a little and say a graph G (with an associated generic framework (G, p)) is redundant
if every edge is in a (2, 3)-circuit. Then an inductive construction of redundant graphs was given in
[42].

19.2.3 Combinatorial Characterizations of Rigidity

For the two main rigidity theorems in R2 let us briefly discuss how the inductive proof proceeds. In
most similar results, a proof by inductive construction has the same style.
Theorem 19.11 (Laman [51]) A generic framework (G, p) in R2 is rigid if and only if it contains
a spanning subgraph which is (2, 3)-tight.
The inductive proof is for the deeper direction (2, 3)-tight implies minimally rigid. We proceed
by induction on n and the base case is the trivial check that K2 is minimally rigid. Suppose every
(2, 3)-tight graph on at most n vertices is minimally rigid. Consider a (2, 3)-tight graph G on n + 1
vertices. By Theorem 19.2 we can reduce G to a smaller (2, 3)-tight graph using a 0-reduction or a
1-reduction. By induction this smaller graph is minimally rigid. Now apply Lemma 19.1 or 19.2 to
deduce that G was minimally rigid.
Theorem 19.12 (Hendrickson [22], Connelly [8], Jackson and Jordán [28]) Let G be a graph
with |V | ≥ 4. A generic framework (G, p) in R2 is globally rigid if and only if G is 3-connected
and (G, p) is redundantly rigid.
Hendrickson [22] proved that every globally rigid generic framework is 3-connected and re-
dundantly rigid. The sufficiency direction is proved by induction, combining Theorems 19.10 and
19.1.
Tanigawa [76] gave a new sufficient condition for global rigidity using “vertex-redundant rigid-
ity” which allowed him to simplify some parts of the proof of Theorem 19.12.
Let (G, p) be a framework and let u, v ∈ V . Then the pair u, v is globally linked if the distance
between u and v is the same in (G, p) as it is in any other generic realization (G, q). In [33] globally
linked pairs were studied. In particular 1-extension preserves global linkedness when the edge being
deleted is redundant. Subsequently they showed [34] that 1-extensions on non-redundant edges
preserve the property of being not globally linked.
For 3- and higher-dimensional frameworks one of the key combinatorial difficulties in proving
a 3D rigidity theorem is that 0-extensions and 1-extensions are not sufficient. However Jackson and
Jordán proved an analogue of Theorem 19.2 for degree bounded graphs, in arbitrary dimension,
using just these two operations. We say that a framework (G, p) is independent if its rigidity matrix
has maximum possible rank over all realizations of G.
422 Handbook of Geometric Constraint Systems Principles

Theorem 19.13 (Jackson and Jordán [29]) Let G be
a connected graph with minimum degree
d + 1 and maximum degree d + 2. Then G is (d, d+1
2 )-sparse if and only if (G, p) is generically
independent.

They also used 0- and 1-extensions to prove a similar result for sparse graphs (see [29, Corollary
4.3].)
In [59] the rigidity of frameworks on concentric d-spheres whose radii are allowed to vary con-
tinuously was considered. Here the combinatorial objects are vertex colored graphs and analogues
of Laman’s theorem for circles with arbitrarily many radii varying independently were proved using
inductive techniques, in particular 0- and 1-extensions on colored graphs. There results were also
extended to 2-spheres with 1 or 3 radii varying independently. By the equivalence of rigidity in Rd
with rigidity on the d-sphere (see Chapter 17, Section 17.2), these results give interpolation theo-
rems between rigidity characterizations in R and R2 and from R2 , a small step, in the direction of
R3 .
Penne [66] showed that triangle free 1-extensions preserve the property of having an irreducible
pure condition (in the sense of White-Whiteley [81]). He conjectured that an isostatic graph (in the
plane) has an irreducible pure condition if and only if the only proper rigid subgraphs are single
edges. He called such graphs minimally isostatic graphs (MIGs). There are countably many MIG’s
that cannot be generated from K3,3 by triangle free 1-extensions. Proving an inductive construction
for this class of graphs is an interesting open problem. A closely related class of graphs was consid-
ered in [34] in the study of globally loose pairs. These are isostatic graphs in which the only proper
rigid subgraphs are complete graphs. Again finding an inductive construction is an open problem.

19.3 Body-Bar, Body-Hinge, Molecular, etc.

While we have a good understanding of rigidity in R2 it is still true that little is known for bar-joint
frameworks in higher dimensions. For example, combinatorially, no analogue of Theorem 19.2 is
known for (3, 6)-tight graphs with, or without, any consideration of excluding known problematic
examples like the double banana.
However, a number of special classes of frameworks have been studied extensively. In the pre-
vious section we saw some special cases (for example, Theorem 19.13). Here we will focus on
frameworks in Rd with special geometry and show that in these cases a lot can be said combinato-
rially about rigidity.

Body-bar framework: A d-dimensional body-bar framework is a structural model consisting of

rigid bodies and stiff bars. The bodies are free to move continuously in Rd subject to the con-
straints imposed by the bars.
Body-hinge framework: A d-dimensional body-hinge framework is a structural model consisting
of rigid bodies and hinges. Each hinge is a (d − 2)-dimensional affine subspace that joins some
pair of bodies. The bodies are free to move continuously in Rd subject to the constraint that the
relative motion of any two bodies joined by a hinge is a rotation about the hinge.
Molecular framework: A molecular framework is a body-hinge framework in which the hinges
are coplanar. That is, all the (d − 2)-dimensional affine subspaces incident to a body are con-
tained in a common (d − 1)-dimensional affine subspace.
edge pinch: Let 0 ≤ j ≤ m ≤ k. An edge pinch K(k, m, j) pinches j edges of G into a new node z,
places m − j loops on z and connects z to other vertices of G by k − m new edges.
Inductive Constructions for Combinatorial Local and Global Rigidity 423
m − j loops

2j pinched z
edges k − m new
edges

j edges

Figure 19.7
Edge pinch K(k, m, j).

19.3.1 Geometry and Combinatorics

The simplest geometrically special class is that of rigid bodies attached by bars. Since such body-
bar frameworks are special classes of bar-joint frameworks, Lemmas 19.1 and 19.2 clearly apply.
Similar proofs show that the multigraph variants also preserve rigidity. Similarly 0-extension does
not preserve global rigidity for body-bar frameworks, while Theorem 19.1 implies that 1-extension
does. However, Connelly, Jordán, and Whiteley [6], in their study of body-bar global rigidity, had
to work significantly harder including use of a grafting operation (to combine graphs) and a triangle
exchange operation (swapping edges within a collinear triangle).
The following theorem extends the 0- and 1-extension construction to (k, k)-tight graphs for
arbitrary k.

Theorem 19.14 (Tay [78]) A loopless multigraph G is (k, k)-tight if and only G can be formed from
the single vertex by a series of loopless edge pinches (K(k, m, m), where 0 ≤ m ≤ k).

As an aside we note that this was then extended to (k, `)-tight multigraphs by Fekete and Szegő
[14]. Let Pn be the multigraph on one vertex with n incident loops.

Theorem 19.15 (Fekete and Szegő [14]) Let 1 ≤ k ≤ `. A multigraph G is (k, `)-tight if and only
if G can be obtained from Pk−` by edge pinches K(k, m, j) where j ≤ m ≤ k − 1 and m − j ≤ k − `.
A multigraph G is (k, 0)-tight if and only if G can be obtained from Pk by edge pinches K(k, m, j)
where j ≤ m ≤ k and m − j ≤ k.

A classical result of Nash-Williams [54] expresses (k, k)-tightness in terms of packing of k

edge-disjoint spanning trees. This leads to analogues of (2, 3)-tight graphs in terms of nearly k-tree
connected graphs and of redundantly rigid graphs in terms of highly k-tree connected graphs. For
these classes of graphs, Frank and Szegő [17] proved constructive characterizations using analogues
of pinching. There are also a number of constructive characterizations of related classes of graphs,
see [16] and the references therein.
In their work on the molecular conjecture, Katoh and Tanigawa
[45] proved an inductive con-
struction of graphs G such that the graph G̃ formed from taking d+1 2 − 1 copies of each edge con-
tains d+1


2 edge-disjoint spanning trees and deleting any edge from G destroys the spanning tree
property. Their characterization, for such graphs G which are 2-edge-connected uses contraction of
proper rigid subgraphs and the inverse of subdivision.
In [48] an inductive construction of minimally rigid body-hinge graphs was given.
424 Handbook of Geometric Constraint Systems Principles

19.3.2 Characterizations
For body-bar frameworks, rigidity can be elegantly characterized via tree packing in arbitrary di-
mension.

Theorem 19.16 (Tay [77]) A multigraph G is generically minimally rigid in Rd as a body-bar

d+1

framework if and only if G is (D, D)-tight, where D = 2 is the dimension of the Euclidean
group Euc(Rd ).

Tay [77] and Whiteley [82] extended this result to body-hinge frameworks, using non-inductive
means. Tay and Whiteley [79] conjectured that it could be extended to “hinge-concurrent” frame-
works and this became known as the molecular conjecture and was used by material scientists for
many years [38, 80]. Jackson and Jordán [31] used 0-extensions, 1-extensions and vertex splitting in
their proof of the conjecture for 2-dimensions. (Since they worked in dimension 2 the hinges were
actually pins.)

Theorem 19.17 (Jackson and Jordán [31]) A multigraph G has an infinitesimally rigid pin-
collinear body-pin realization if and only if 2G (the graph obtained from G by doubling all edges)
contains three edge-disjoint spanning trees.

The following theorem, proved by Katoh and Tanigawa [45], extended this to d-dimensions
and hence turned the molecular conjecture into a theorem. This is one of the most powerful results
proved using an inductive construction.

Theorem 19.18 (Katoh and Tanigawa [45]) A multigraph G can be realised as an infinitesimally
rigid body-and-hinge framework in Rd if and only if G can be realised as an infinitesimally rigid
panel-and-hinge framework in Rd .

Global rigidity is also well understood for body-bar frameworks.

Theorem 19.19 (Connelly, Jordan, and Whiteley [6]) A body-bar framework is generically glob-
ally rigid in Rd if and only if it is generically redundantly rigid in Rd .

An interesting aspect of Connelly, Jordan and Whiteley’s proof is that they were able to go
outside the inductive class, using looped graphs in intermediate steps of the induction process.
Very recently this has been extended to body-hinge frameworks by Jordán, Kiraly and Tanigawa
[41] (using non-inductive techniques). A global rigidity version of the molecular theorem is an open
problem, although a conjecture was presented in [6].

19.4 Further Rigidity Contexts

Gain graph: A Γ-gain graph is a multigraph with directions whose edges are labelled by elements
of a group Γ.
Balanced: A subgraph H of a Γ-gain graph G is balanced if, for every cycle in H, the product
of the edge labels equals the identity (making the appropriate choice of group element or its
inverse depending on the direction of the edge).
(k, `, m)-gain-sparse/tight: A Γ-gain graph G is (k, `, m)-gain-sparse if |F| ≤ k|V (F)| − ` for every
balanced subset F ⊂ E and |F| ≤ k|V (F)| − ml for every unbalanced subset F ⊂ E. Moreover
a Γ-gain graph G is (k, `, m)-gain-tight if |E| = k|V | − m and G is (k, `, m)-gain-sparse.
Inductive Constructions for Combinatorial Local and Global Rigidity 425

mechanism: A framework with exactly one degree of freedom.

direction-length graph: A loopless multigraph G = (V, D ∪ L) whose edge set is partitioned into
direction edges D and length edges L. Neither D nor L contains parallel edges.
1-extension on direction-length graphs: These are 1-extensions, as defined above, that may be
direction-pure, length-pure (involving only direction or length edges, respectively) or mixed.

In this final section we briefly describe a number of areas of current research in rigidity theory
where inductive constructions have been useful.

19.4.1 Frameworks with Symmetry

There has been significant use of inductive constructions for the study of symmetric frameworks.
These are frameworks, typically in R2 , which admit an action of a symmetry group but are other-
wise as generic as possible (see Chapter 25). Symmetry analysis in rigidity splits into two distinct
strands. First, there is incidental symmetry where a framework happens to possess some symmetry
but deformations which break the symmetry are permitted. The key sources for such considerations
are [70, 71, 72]. Known characterizations of incidental rigidity are, currently, limited to very small
groups but make significant use of inductive constructions.
Second, there is forced symmetry, where the framework is symmetric and only deformations
which preserve that symmetry are considered. We mention here one sample theorem of Jordan,
Kaszanitsky, and Tanigawa [40] which uses symmetrized versions of 0-extension, 1-extension and
X-replacement to characterize symmetry-forced rigidity. They first use these operations to give an
inductive construction of (2, 3, 1)-gain-tight graphs and then show that these operations preserve
symmetric rigidity. See Figures 19.8 and 19.9 for samples of gain graphs, symmetric frameworks,
and the 1-extension operations in these cases.
Moreover of interest across several rigidity contexts are (2, `, m)-gain-tight graphs in the range
0 ≤ m ≤ ` < 4. For the range 1 ≤ m ≤ ` < 4, inductive constructions are actually known in each case
(see [64]). However when m = 0 proving analogues is open. Two hints of the additional difficulty
arises from the facts that the minimum degree may be 4 (contrary to the case when m ≥ 1) and that
(2, 3, 0)-gain-sparsity does not, in general, induce a matroid (see [72]).

19.4.2 Infinite Frameworks

Infinite frameworks which are periodic with respect to some lattice have been considered from an
inductive construction perspective. In particular using the viewpoint of considering periodic frame-
works as frameworks on a torus, Ross [68] proved that, with a fixed lattice, periodic rigidity was
equivalent to the graph being (2, 3, 2)-gain-tight. The proof extended Theorem 19.14, in the case
k = 2, to deal with the group elements in subgraphs. This has been extended inductively in [63]
to a partially variable lattice, in [67] to body-bar periodic frameworks in R3 and extended non-
inductively by Malestein and Theran [53] to a fully variable lattice.
Whiteley [86] has also considered inductions for geometrically special types of repetitive struc-
tures (fragments, substrates, etc.) and suggested potential applications of inductive techniques such
as an extended form of vertex splitting.
Theorem 19.2 has been generalized to countably infinite graphs (with bounded degree vertices),
still using 0- and 1-extensions, by Kitson and Power [46]. Using this, a countably infinite version
of Laman’s theorem has been derived. Their combinatorial object is a containment tower of finite
(2, 3)-tight subgraphs.
426 Handbook of Geometric Constraint Systems Principles
s

2
(a) Cs -labeled (b) Cs -symmetric framework
gain graph

1
(1, 0)

2 3
(c) Z2 -labeled gain graph (d) Z2 -symmetric framework

Figure 19.8
A mirror-symmetric framework (b) and its gain graph (a). A plane-periodic framework (d) and its
gain graph (c).

19.4.3 Surfaces
The rigidity (and global rigidity) of frameworks in R3 which are forced to lie on a 2-dimensional
manifold has also been studied in [25, 37, 57, 58]. Characterizations were given for the rigidity
of generic frameworks on concentric spheres and concentric cylinders [57] and this was extended
to surfaces with exactly one isometry in [58]. These theorems were proved inductively requiring
characterizations of (simple) (2, 2)-tight and (2, 1)-tight graphs, see also [56]. (The simplicity re-
quirement preventing these results from being deducable from Theorems 19.14 and 19.15.) The op-
erations used were 0-extension, 1-extension, vertex-to-4-cycle, vertex-to-K4 , and in the (2, 1)-tight
case, joining two graphs in the class by a bridge.
For example, we briefly describe the (2, 2)-tight characterization given in [58]. The minimum
degree is in the set {2, 3}. By using 0-extensions we may assume that any graph that we cannot
generate using these moves has minimum degree 3. The problem with extending Theorem 19.2
is that K4 is (2, 2)-sparse so it can happen that every vertex of degree 3 cannot be reduced to a
(simple) (2, 2)-tight graph using a 1-reduction; all degree 3 vertices may be contained in subgraphs
isomorphic to K4 . In such a case we can contract a copy of K4 to a single vertex unless there is a
Inductive Constructions for Combinatorial Local and Global Rigidity 427
s

2 s
(a) 1-extension on (b) Symmetric 1-extension
symmetric gain graph
in 19.8(a)

(1, 0)
4 1

(0, 1)

2 3
(c) 1-extension on periodic gain (d) Periodic 1-extension
graph in 19.8(c)

Figure 19.9
1-extensions on the graphs of Figure 19.8.

copy of K3 which intersects the K4 in a single edge. However this structure gives a 4-cycle on the
vertex set of the K3 along with the degree 3 vertex of the K4 to which we can always apply an inverse
vertex-to-4-cycle move.
A recursive construction of (simple) (2, 0)-tight graphs was recently obtained in [20]. The case
of rigidity on a specific surface admitting no isometries such as the ellipsoid remains open due to
the difficulty of understanding X-replacement on such a surface.
Considering global rigidity in this context leads to the study of (2, 2)-circuits, for which a char-
acterization was given in [60]. Also an analogue of the stress matrix for frameworks on surfaces was
developed, using which, the 1-extension operation was shown to preserve global rigidity on surfaces
[25]. In [37] these results were extended to completely characterize global rigidity on the cylinder.
As well as building on these earlier results the proof used a generalized form of vertex splitting
and two further recursive construction results for cylindrical frameworks. Finally, symmetric frame-
works on surfaces were considered giving rise to several inductive constructions for various classes
of (k, `, m)-gain-tight graphs (see [64]).
The above-mentioned construction of (2, 2)-tight graphs was also used to prove an analogue
of Laman’s theorem for frameworks in R2 equipped with a non-Euclidean metric (an ` p norm for
428 Handbook of Geometric Constraint Systems Principles

p ∈ (1, ∞) with p 6= 2). See [47] and Chapter 24 for details. Generalizing these contexts to metric
spaces with multiple types of distance constraints with their combinatorics handled by sparsity
constraints on edge-colored subgraphs has also recently been considered [61].

19.4.4 Mechanisms
Servatius, Shai, and Whiteley [74] use 0- and 1-extensions in their study of pinned frameworks
in the plane and Assur graphs. They give an analogue of Theorem 19.2 in the pinned setting and
present a recursive construction of Assur graphs using 1-extensions and 2-sums. Assur graphs arise,
in engineering, in the synthesis and analysis of mechanisms.
Gao and Sitharam [18] studied the class of graphs constructable by 0-extensions in their analysis
of geometric properties (connectivity, convexity, etc.) of Cayley configuration spaces.

19.4.5 Applications of Rigidity Techniques

Jackson, Jordán, and Tanigawa [35] used inductive construction techniques, including 0-extensions,
1-extensions and what they call double 1-extensions, to prove results about matrix completion prob-
lems for low rank semidefinite matrices.
Algorithmic aspects of rigidity have been considered using inductive constructions. In particular
inductive moves give efficient ways of constructing families of rigid [52] and globally rigid [43]
graphs in the plane.
Recently rigidity has found applications in control of robotic formations. Here inductive tech-
niques have been used, for example, to understand rigidity with direction constraints (known to
control theorists as bearing rigidity) [87]. Versions of the 0- and 1-extension operations were also
used in a similar context to analyse persistence [23], which concerns rigidity for directed graphs
where constraints are “one-way.”

19.4.6 Direction-Length Frameworks and CAD

Rigidity and inductions have also been applied to various problems in computer aided design (CAD)
[75]. Building on a result of Jackson and Jordán on the global rigidity of 2-dimensional direction-
length graphs [32], Nguyen proved that d-dimensional 1-extensions preserve global rigidity of
direction-length graphs [55]. These results are significant because it is not known whether global
rigidity is a generic property of direction-length frameworks even in two dimensions. Therefore, it
is helpful to have inductive methods to generate families of globally rigid direction-length graphs,
and this also provides a tool to verify the global rigidity of certain frameworks.
Owen and Power [65] showed that vertex splitting preserves the property of being radically
solvable; a property, intimately linked to rigidity, of importance in CAD. It seems to be a challenging
open problem to determine if the 1-extension operation preserves the property of being radically
solvable.

19.4.7 Nearly Generic Frameworks

Since genericity is a very strong assumption there is substantial interest in weakening this assump-
tion. We have already discussed such contexts such as symmetric frameworks and frameworks on
surfaces. Here we mention combinatorial results for frameworks in the plane with some speci-
fied non-genericity. Jackson and Jordán [30] characterized rigidity for frameworks in the plane
where three designated vertices are collinear but are otherwise generic. To do this they extended
the Henneberg-Laman technique by avoiding these 3 special vertices in their 0- and 1-extension
arguments. Second, Fekete, Jordán, and Kaszanitzky [12] proved a delete-contract characterization
Inductive Constructions for Combinatorial Local and Global Rigidity 429

of rigidity in the plane where two designated vertices are coincident but the framework is otherwise
generic. Again the technique was to use 0- and 1-extensions but the proofs required analyzing a more
complicated count matroid. Further use of these recursive operations extended the result to a triple
of coincident points [20]. These ideas were very recently extended to frameworks on the sphere
and on the cylinder with two coincident points [36] using similar operations to those described in
Subsection 19.4.3.

19.5 Summary Tables

Tables 19.1 and 19.2 summarize the use of inductive moves for local and global rigidity.

Table 19.1
Summary of inductive moves for local rigidity.
Operation Figure Type of framework Rigidity Class of graphs Characterization
0-, 1-extension Fig. 19.1 2-dim bar-joint Lemmas 19.1, (2, 3)-tight, Theorem 19.2 2-dim rigidity
19.2 Laman-plus-one, Theorem 19.5
d-dim bar-joint Lemmas 19.1, Degree bounded
19.2
Fig. 19.9(a), Symmetric Yes [71] (2, 3, 1)-gain-tight, [40] Forced rotational
(b) Forced reflectional
Odd-order dihedral, [40]
Fig. 19.9(c), Periodic Yes, [68] (2, 3, 2)-gain-tight, [68] Fixed lattice
(d) Partial lattice, [63]
Surfaces Yes, [58] (2, 2)-tight, [58] cylinder, [57]
(2, 1)-tight, [56] 1 isometry surfaces, [58]
Symmetry and surfaces Yes, [64] (2, 3, 3)-gain-tight, [64] Spherical inversion, [64]
(2, 2, 2)-gain-tight, [64] Cylinder with
(2, 2, 1)-gain-tight, [64] Various groups, [64]
(2, 1, 1)-gain-tight, [64] Cone, [64]
Infinite bar-joint Yes, [46] Sequential (2, 3)-tight, [46]
X-replacement Fig. 19.2(a) 2-dim bar-joint Lemma 19.3
Fig. 19.2(a) 3-dim bar-joint Conjecture 3
Fig. 19.2(a) d-dim bar-joint No, [19]
V-replacement Fig. 19.2(b) 3-dim bar-joint No
Vertex splitting Fig. 19.3(a), d-dim bar-joint Proposition 19.1
(c)
Vertex-to-4-cycle Fig. 19.3(b) 2-dim bar-joint Yes, [50]
Surfaces Yes, [58] (2, 2)-tight Cylinder, [57]
(2, 1)-tight 1 isometry surfaces
2-sum 2-dim bar-joint Yes
Pinching Fig. 19.7 d-dim body-bar Yes (D, D)-tight, [78] Body-bar rigidity, [77]

Table 19.2
Summary of inductive moves for global rigidity.
Operation Type of framework Global rigidity Class(es) of graphs Characterization
0-extension d-dim bar-joint No
1-extension 2-dim bar-joint Theorem 19.1 (2, 3)-circuits, [4]
3-connected and
redundantly rigid graphs, [28] 2-dim global, [8, 28]
d-dim bar-joint Theorem 19.1
X-replacement 2-dim bar-joint No, see, for example,
[62, Figure 6]
Vertex splitting 2-dim bar-joint Yes, [13]
d-dim bar-joint Conjecture 4
2-sum 2-dim bar-joint No (2, 3)-circuits
Pinching d-dim body-bar Yes, [6] Highly k-tree connected, [17] Body-bar global, [6]
1-extension Surfaces Yes, [25] (2, 2)-circuits, [60]
1-extension Direction-length Yes, [32]
430 References

References
[1] T. Abbott. Generalizations of Kempe’s universality theorem. Master’s thesis, 2008.
[2] M. Alam, D. Eppstein, M. Kaufmann, S. Kobourov, S. Pupyrev, A. Schulz, and T. Ueckerdt.
Contact representations of sparse planar graphs. arXiv:1501.00318, 2015.
[3] L. Asimow and B. Roth. The rigidity of graphs. Trans. Amer. Math. Soc., 245:279–289, 1978.
[4] Alex R. Berg and Tibor Jordán. A proof of Connelly’s conjecture on 3-connected circuits of
the rigidity matroid. J. Combin. Theory Ser. B, 88(1):77–97, 2003.
[5] R. Connelly. Combining globally rigid frameworks. Tr. Mat. Inst. Steklova, 275(Klassich-
eskaya i Sovremennaya Matematika v Pole Deyatelnosti Borisa Nikolaevicha Delone):202–
209, 2011.
[6] R. Connelly, T. Jordán, and W. Whiteley. Generic global rigidity of body-bar frameworks. J.
Combin. Theory Ser. B, 103(6):689–705, 2013.
[7] R. Connelly and W. J. Whiteley. Global rigidity: the effect of coning. Discrete Comput.
Geom., 43(4):717–735, 2010.
[8] Robert Connelly. Generic global rigidity. Discrete Comput. Geom., 33(4):549–563, 2005.
[9] Robert Connelly. Questions, conjectures and remarks on globally rigid tensegrities. Technical
Report, November 2009.
[10] James Cruickshank. On spaces of infinitesimal motions and three dimensional Henneberg
extensions. Discrete Comput. Geom., 51(3):702–721, 2014.
[11] James Cruickshank, Derek Kitson, and Stephen C. Power. The generic rigidity of triangulated
spheres with blocks and holes. J. Combin. Theory Ser. B, 122:550–577, 2017.
[12] Zsolt Fekete, Tibor Jordán, and Viktória E. Kaszanitzky. Rigid two-dimensional frameworks
with two coincident points. Graphs Combin., 31(3):585–599, 2015.
[13] Zsolt Fekete, Tibor Jordán, and Walter Whiteley. An inductive construction for plane Laman
graphs via vertex splitting. In Algorithms—ESA 2004, volume 3221 of Lecture Notes in Com-
put. Sci., pages 299–310. Springer, Berlin, 2004.
[14] Zsolt Fekete and László Szegő. A note on [k, l]-sparse graphs. In Graph theory in Paris,
Trends Math., pages 169–177. Birkhäuser, Basel, 2007.
[15] Wendy Finbow-Singh and Walter Whiteley. Isostatic block and hole frameworks. SIAM J.
Discrete Math., 27(2):991–1020, 2013.
[16] András Frank. Connections in combinatorial optimization, volume 38 of Oxford Lecture
Series in Mathematics and Its Applications. Oxford University Press, Oxford, 2011.
[17] András Frank and László Szegő. Constructive characterizations for packing and covering with
trees. Discrete Appl. Math., 131(2):347–371, 2003.
[18] Heping Gao and Meera Sitharam. Characterizing 1-dof henneberg-i graphs with efficient
configuration spaces. In Proceedings of the 2009 ACM symposium on Applied Computing,
pages 1122–1126. ACM, 2009.
[19] Jack Graver, Brigitte Servatius, and Herman Servatius. Combinatorial rigidity, volume 2 of
Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 1993.
References 431

[20] Hakan Guler. Rigidity of Frameworks. PhD thesis, Queen Mary, University of London, 2018.
[21] Ruth Haas, David Orden, Günter Rote, Francisco Santos, Brigitte Servatius, Herman Ser-
vatius, Diane Souvaine, Ileana Streinu, and Walter Whiteley. Planar minimally rigid graphs
and pseudo-triangulations. Comput. Geom., 31(1-2):31–61, 2005.
[22] Bruce Hendrickson. Conditions for unique graph realizations. SIAM J. Comput., 21(1):65–84,
1992.
[23] Julien M. Hendrickx, Brian D. O. Anderson, Jean-Charles Delvenne, and Vincent D. Blondel.
Directed graphs for the analysis of rigidity and persistence in autonomous agent systems.
Internat. J. Robust Nonlinear Control, 17(10-11):960–981, 2007.
[24] L. Henneberg. Die Graphische Statik der starren Systeme. (Johnson Reprint), 1911.
[25] B. Jackson and A. Nixon. Stress matrices and generic global rigidity of frameworks on sur-
faces. Discrete Comput. Geom., 54(3):586–609, 2015.
[26] B. Jackson and J. Owen. Notes on henneberg moves. 2013.
[27] Bill Jackson. Notes on the rigidity of graphs. Levico Conference Notes, 2007.
http://www.science.unitn.it/cirm/JacksonLectures.pdf.
[28] Bill Jackson and Tibor Jordán. Connected rigidity matroids and unique realizations of graphs.
J. Combin. Theory Ser. B, 94(1):1–29, 2005.
[29] Bill Jackson and Tibor Jordán. The d-dimensional rigidity matroid of sparse graphs. J.
Combin. Theory Ser. B, 95(1):118–133, 2005.
[30] Bill Jackson and Tibor Jordán. Rigid two-dimensional frameworks with three collinear points.
Graphs Combin., 21(4):427–444, 2005.
[31] Bill Jackson and Tibor Jordán. Pin-collinear body-and-pin frameworks and the molecular
conjecture. Discrete Comput. Geom., 40(2):258–278, 2008.
[32] Bill Jackson and Tibor Jordán. Operations preserving global rigidity of generic direction-
length frameworks. International Journal of Computational Geometry & Applications,
20(06):685–706, 2010.
[33] Bill Jackson, Tibor Jordán, and Zoltán Szabadka. Globally linked pairs of vertices in equiva-
lent realizations of graphs. Discrete Comput. Geom., 35(3):493–512, 2006.
[34] Bill Jackson, Tibor Jordán, and Zoltán Szabadka. Globally linked pairs of vertices in rigid
frameworks. In Robert Connelly, Walter Whiteley, and Asia Weiss, editors, Rigidity and
Symmetry. Fields Institute, 2014.
[35] Bill Jackson, Tibor Jordán, and Shin-ichi Tanigawa. Combinatorial conditions for the unique
completability of low-rank matrices. SIAM J. Discrete Math., 28(4):1797–1819, 2014.
[36] Bill Jackson, Viktoria Kaszanitsky, and Anthony Nixon. Rigid cylindrical frameworks with
two coincident points. arXiv: 1607.02039, 2016.
[37] Bill Jackson and Anthony Nixon. Global rigidity of generic frameworks on the cylinder.
arXiv: 1610.07755, 2017.
[38] Donald J Jacobs, Leslie A Kuhn, and Michael F Thorpe. Flexible and rigid regions in proteins.
In Rigidity theory and applications, pages 357–384. Springer, 2002.
[39] Tibor Jordán. Combinatorial rigidity: graphs and matroids in the theory of rigid
frameworks. Technical Report TR-2014-12, Egerváry Research Group, Budapest, 2014.
www.cs.elte.hu/egres.
432 References

[40] Tibor Jordán, Viktória E. Kaszanitzky, and Shin-ichi Tanigawa. Gain-sparsity and symmetry-
forced rigidity in the plane. Discrete Comput. Geom., 55(2):314–372, 2016.
[41] Tibor Jordán, Csaba Király, and Shin-ichi Tanigawa. Generic global rigidity of body-hinge
frameworks. J. Combin. Theory Ser. B, 117:59–76, 2016.
[42] Tibor Jordán, András Recski, and Zoltán Szabadka. Rigid tensegrity labelings of graphs.
European J. Combin., 30(8):1887–1895, 2009.
[43] Tibor Jordán and Zoltán Szabadka. Operations preserving the global rigidity of graphs and
frameworks in the plane. Comput. Geom., 42(6-7):511–521, 2009.
[44] Tibor Jordan and Shin-Ichi Tanigawa. Global rigidity of triangulations with braces. EGRES
TR-2017-06, 2017.
[45] Naoki Katoh and S.-I. Tanigawa. A proof of the molecular conjecture. Discrete & Computa-
tional Geometry, 45(4):647–700, 2011.
[46] D. Kitson and S. Power. The rigidity of infinite graphs. arXiv:1310.1860, 2013.
[47] D. Kitson and S. C. Power. Infinitesimal rigidity for non-Euclidean bar-joint frameworks.
Bull. Lond. Math. Soc., 46(4):685–697, 2014.
[48] Yuki Kobayashi, Yuya Higashikawa, Naoki Katoh, and Naoyuki Kamiyama. An inductive
construction of minimally rigid body–hinge simple graphs. Theoretical Computer Science,
556(0):2 – 12, 2014. Combinatorial Optimization and Applications.
[49] Erika R. Kovács and László A. Végh. Constructive characterization theorems in combinato-
rial optimization. In Combinatorial optimization and discrete algorithms, RIMS Kôkyûroku
Bessatsu, B23, pages 147–169. Res. Inst. Math. Sci. (RIMS), Kyoto, 2010.
[50] L. Moshe L. Lomeli and W. Whiteley. Bases and circuits for 2-rigidity: constructions via tree
partitions.
[51] G. Laman. On graphs and rigidity of plane skeletal structures. J. Engrg. Math., 4:331 – 340,
1970.
[52] Audrey Lee and Ileana Streinu. Pebble game algorithms and sparse graphs. Discrete Math.,
308(8):1425–1437, 2008.
[53] Justin Malestein and Louis Theran. Generic combinatorial rigidity of periodic frameworks.
Adv. Math., 233:291–331, 2013.
[54] C. St. J. A. Nash-Williams. Decomposition of finite graphs into forests. J. London Math. Soc.,
39:12, 1964.
[55] Viet-Hang Nguyen. 1-extensions and global rigidity of generic direction-length frameworks.
International Journal of Computational Geometry & Applications, 22(06):577–591, 2012.
[56] A. Nixon and J. Owen. An inductive construction of (2,1)-tight graphs. Contributions to
Discrete Math, 9(2):1–16, 2014.
[57] A. Nixon, J. C. Owen, and S. C. Power. Rigidity of frameworks supported on surfaces. SIAM
J. Discrete Math., 26(4):1733–1757, 2012.
[58] A. Nixon, J. C. Owen, and S. C. Power. A characterization of generically rigid frameworks
on surfaces of revolution. SIAM J. Discrete Math., 28(4):2008–2028, 2014.
[59] A. Nixon, B. Schulze, S.I. Tanigawa, and W. Whiteley. Rigidity of frameworks on expanding
spheres. arXiv:1501.01391, 2015.
References 433

[60] Anthony Nixon. A constructive characterisation of circuits in the simple (2, 2)-sparsity ma-
troid. European J. Combin., 42:92–106, 2014.
[61] Anthony Nixon and Stephen Power. Double-distance frameworks and mixed sparsity graphs.
arXiv: 1709.06349, 2017.
[62] Anthony Nixon and Elissa Ross. One brick at a time: a survey of inductive constructions in
rigidity theory. In Robert Connelly, Walter Whiteley, and Asia Weiss, editors, Rigidity and
Symmetry. Fields Institute, 2014.
[63] Anthony Nixon and Elissa Ross. Periodic rigidity on a variable torus using inductive con-
structions. Electron. J. Combin., 22(1), 2015.
[64] Anthony Nixon and Bernd Schulze. Symmetry-forced rigidity of frameworks on surfaces.
Geom. Dedicata, 182:163–201, 2016.
[65] J. C. Owen and S. C. Power. The non-solvability by radicals of generic 3-connected planar
Laman graphs. Trans. Amer. Math. Soc., 359(5):2269–2303, 2007.
[66] Rudi Penne. Isostatic bar and joint frameworks in the plane with irreducible pure conditions.
Discrete Appl. Math., 55(1):37–57, 1994.
[67] Elissa Ross. The rigidity of periodic body-bar frameworks on the three-dimensional fixed
torus. Philos. Trans. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci., 372(2008):20120112, 23,
2014.
[68] Elissa Ross. Inductive constructions for frameworks on a two-dimensional fixed torus. Dis-
crete & Computational Geometry, 54(1):78–109, 2015.
[69] J. Saxe. Embeddability of weighted graphs in k-space is strongly np-hard. In Proceedings
of the 17th Allerton conference in communications, control and computing, pages 480–489,
1979.
[70] B. Schulze. Combinatorial and geometric rigidity with symmetry constraints. PhD thesis,
York University, 2009.
[71] Bernd Schulze. Symmetric versions of Laman’s theorem. Discrete Comput. Geom.,
44(4):946–972, 2010.
[72] Bernd Schulze and Shin-ichi Tanigawa. Infinitesimal rigidity of symmetric bar-joint frame-
works. SIAM J. Discrete Math., 29(3):1259–1286, 2015.
[73] Brigitte Servatius and Herman Servatius. On the 2-sum in rigidity matroids. European J.
Combin., 32(6):931–936, 2011.
[74] Brigitte Servatius, Offer Shai, and Walter Whiteley. Combinatorial characterization of the
Assur graphs from engineering. European J. Combin., 31(4):1091–1104, 2010.
[75] Brigitte Servatius and Walter Whiteley. Constraining plane configurations in computer-aided
design: combinatorics of directions and lengths. SIAM J. Discrete Math., 12(1):136–153
(electronic), 1999.
[76] Shin-ichi Tanigawa. Sufficient conditions for the global rigidity of graphs. J. Combin. Theory
Ser. B, 113:123–140, 2015.
[77] Tiong-Seng Tay. Rigidity of multigraphs I: linking rigid bodies in n-space. J. Combinatorial
Theory B, 26:95 – 112, 1984.
[78] Tiong-Seng Tay. Henneberg’s method for bar and body frameworks. Structural Topology,
17:53–58, 1991.
434 References

[79] Tiong-Seng Tay and Walter Whiteley. Generating isostatic frameworks. Structural Topology,
(11):21–69, 1985. Dual French-English text.
[80] Michael Thorpe. http://flexweb.asu.edu/.
[81] Neil L. White and Walter Whiteley. The algebraic geometry of stresses in frameworks. SIAM
J. Algebraic Discrete Methods, 4(4):481–511, 1983.
[82] Walter Whiteley. The union of matroids and the rigidity of frameworks. SIAM J. Discrete
Math., 1(2):237–255, 1988.
[83] Walter Whiteley. La division de sommet dans les charpentes isostatiques. Structural Topology,
(16):23–30, 1990. Dual French-English text.
[84] Walter Whiteley. Some matroids from discrete applied geometry. In Matroid theory (Seattle,
WA, 1995), volume 197 of Contemp. Math., pages 171–311. Amer. Math. Soc., Providence,
RI, 1996.
[85] Walter Whiteley. Rigidity and scene analysis. In Handbook of discrete and computational
geometry, CRC Press Ser. Discrete Math. Appl., pages 893–916. CRC, Boca Raton, FL, 1997.
[86] Walter Whiteley. Fragmentary and incidental behaviour of columns, slabs and crystals. Philo-
sophical Transactions of the Royal Society of London A: Mathematical, Physical and Engi-
neering Sciences, 372(2008), 12 2013.
[87] Shiyu Zhao and Daniel Zelazo. Bearing rigidity and almost global bearing-only formation
stabilization. IEEE Trans. Automat. Control, 61(5):1255–1268, 2016.
Chapter 20
Rigidity of Body-Bar-Hinge Frameworks
Csaba Király
Department of Operations Research, ELTE Eötvös Loránd University, and MTA-ELTE Egerváry
Research Group on Combinatorial Optimization, Pázmány Péter sétány 1/C, Budapest, Hungary

Shin-ichi Tanigawa
Department of Mathematical Informatics, The University of Tokyo, Hongo, Bunkyo-ku, Tokyo, Japan

CONTENTS
20.1 Rigidity of Body-Bar-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
20.1.1 Body-Bar Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
20.1.2 Body-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
20.1.3 Body-Bar-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
20.2 Generic Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
20.2.1 Body-Bar Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
20.2.2 Body-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
20.3 Other Related Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
20.3.1 Plate-Bar Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
20.3.2 Identified Body-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
20.3.3 Panel-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
20.3.4 Molecular Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
20.3.5 Body-Pin Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
20.3.6 Body-Bar Frameworks with Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
20.3.7 Other Variants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
20.4 Generic Global Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
20.4.1 Body-Bar Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
20.4.2 Body-Hinge Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
20.4.3 Counterexamples to Hendrickson’s Conjecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
20.5 Graph Theoretical Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
20.5.1 Tree Packing and Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
20.5.2 Brick Partitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
20.5.3 Constructive Characterizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
20.5.4 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457

This chapter discusses body-bar-hinge frameworks which are abstract models of structures consist-
ing of rigid bodies connected by bars and/or hinges. A body-bar-hinge framework can be regarded
as a special case of bar-joint frameworks, however, its underlying combinatorics is well understood
even in higher dimensional spaces. In particular its generic rigidity has an exact connection to the

435
436 Handbook of Geometric Constraint Systems Principles

tree packing problem, one of the fundamental problems in graph theory, and this link enables us to
analyze the higher dimensional rigidity by graph theoretical approach.

20.1 Rigidity of Body-Bar-Hinge Frameworks

20.1.1 Body-Bar Frameworks
A body-bar framework models a structure consisting of full-dimensional rigid bodies connected by
stiff bars. The bodies are free to move subject to the constraint that the distance between the points
linked by bars must be constant. We shall summarize the basics of body-bar frameworks.

Body-bar framework: A d-dimensional body-bar framework is a pair (G, b ) of a multi-graph G =

(V, EB ) and b : ÊB 3 (e, v) 7→ be,v ∈ Rd , where ÊB denotes the set of pairs (e, v) of an edge
e in EB and its endvertex v. Here G represents the underlying graph, which is obtained by
regarding each body as a vertex and each bar as an edge. See Figure 20.1 and Figure 20.2(a).
b represents the bar-configuration, where the bar corresponding to e = uv is the line segment
[be,u , be,v ] between be,u and be,v .
Body-bar realization: A body-bar framework (G, b ) with the underlying graph G is called a body-
bar realization of G.
Replacement: A replacement of a d-dimensional body-bar framework (G, b ) is a map r : V 3 v 7→
rv ∈ Euc(d) that satisfies

krv (be,v ) − ru (be,u )k = kbe,v − be,u k (e = uv ∈ EB ),

where Euc(d) denotes the set of isometries of the d-dimensional Euclidean space Rd and k · k
denotes the Euclidean norm. A replacement is nontrivial if ru 6= rv for some pairs of vertices
u, v.
Global rigidity: A d-dimensional body-bar framework (G, b ) is globally rigid if it has no nontrivial
replacement.
Finite motion: A finite motion of a body-bar framework (G, b ) is a continuous family r t (t ∈ [0, 1])
of replacements of (G, b ) parameterized by t such that rv0 is the identity for every v ∈ V . A finite
motion is nontrivial if rvt 6= rut for some u, v ∈ V and t ∈ (0, 1].
Rigidity: A d-dimensional body-bar framework (G, b) is rigid if (G, b) has no nontrivial finite
motion.
Infinitesimal isometry: An infinitesimal isometry i of Rd is i : Rd 3 p 7→ Sp + t ∈ Rd for some
skew-symmetric matrix S of size d and t ∈ Rd .
Infinitesimal motion: An infinitesimal motion m of a d-dimensional body-bar framework (G, b ) is
a map m : V 3 v 7→ mv ∈ Inf(d) that satisfies

hmv (be,v ) − mu (be,u ), be,v − be,u i = 0 (e = uv ∈ EB ), (20.1)

where Inf(d) denotes the set of infinitesimal isometries of Rd . An infinitesimal motion m is

trivial if mu = mv for every u, v ∈ V .
Infinitesimal rigidity: A body-bar framework (G, b ) is infinitesimally rigid if every infinitesimal
motion of (G, b ) is trivial.
Rigidity of Body-Bar-Hinge Frameworks 437

Degree of freedom: The degree of freedom of a body-bar framework is the dimension of the space
of infinitesimal motions modulo trivial motions.
Generic framework: A body-bar framework (G, b ) is generic if the set of coordinates in b is al-
gebraically independent over Q.
(k, `)-sparsity: A multigraph G is (k, `)-sparse if |F| ≤ k|V (F)| − ` for any nonempty F ⊆ E(G),
and a (k, `)-sparse graph G is (k, `)-tight if |E(G)| = k|V (G)| − `.

(a) (b)
a b a b

d c d c

(c) (d)
Figure 20.1
Body-bar frameworks in R2 and the underlying graphs.

Equivalent bar-joint frameworks. A body-bar framework can be defined in an alternative way

by regarding each body as a globally rigid bar-joint subframework. The underlying graph of the
resulting bar-joint framework is called a body-bar graph. Specifically, for a graph G, the (d-
dimensional) body-bar graph induced by G, denoted by GB , is defined as follows:
• GB consists of (d + 1)|V (G)| + 2|E(G)| vertices; for each v ∈ V (G) we have d + 1 vertices
xv,1 , . . . , xv,d+1 and for each e = uv ∈ E(G) we have two vertices xe,u and xe,v ;
• the core C(v) of v ∈ V (G) is the complete graph on {xv,1 , . . . , xv,d+1 };
• the body B(v) of v ∈ V (G) is the complete graph on V (C(v)) ∪ {xe,v : e ∈ δG (v)}, where
δG (v) denotes the set of edges in G incident to v;

• E(GB ) := v∈V (G) E(B(v)) ∪ {{xe,u , xe,v } : e = uv ∈ E(G)}.

S

See Figure 20.2 for an example.

Given a body-bar framework (G, b ), the above construction of GB can be extended at the level
of frameworks to obtain an equivalent bar-joint framework (GB , p ). Here the joint configuration
p : V (GB ) → Rd is taken such that p (xe,v ) = be,v for every (e, v) ∈ ÊB and p (C(v)) is an affinely
independent set of points whose coordinates are algebraically independent from the coordinates
438 Handbook of Geometric Constraint Systems Principles

B(v) C(v)

C(u) B(u)

(a) (b)
Figure 20.2
A body-bar framework in R3 and the body-bar graph induced by the underlying graph.

in b . Then (G, b ) is infinitesimally rigid/rigid/globally rigid if and only if (GB , p ) is infinitesi-

mally rigid/rigid/globally rigid. In this sense, body-bar frameworks form a special class of bar-
joint frameworks, and several basic facts about bar-joint frameworks can be translated in terms of
body-bar frameworks. An example is the following connection between rigidity and infinitesimal
rigidity.

Proposition 20.1
If a body-bar framework (G, b) is infinitesimally rigid, then it is rigid.

Rigidity matrices. The study of body-bar frameworks was initiated by Tay [39], where he pro-
posed a rigidity matrix written in terms of the Plücker coordinates of bars.
Let 2 Rd+1 be a d+1
W

2 -dimensional vector space each of whose coordinate is indexed by a pair
(i, j) with 1 ≤ i < j ≤ d +1, and for p, q ∈ Rd+1 , let p∨q be a vector in 2 Rd+1 whose (i, j)-th entry
W
is the
determinant of the 2 × 2-matrix consisting of the i-th and the j-th rows of (d + 1) × 2-matrix
pq .
For a vector x ∈ Rd , let x̂ = (x> , 1)> ∈ Rd+1 be a vector in Rd+1 obtained from x ∈ Rd by
appending one at the last coordinate. Let [x, y] be the (oriented) line segment from x to y in Rd .
Then the Plücker coordinate vector of [x, y] is defined by x̂ ∨ ŷ. For example, if x = (xi )1≤i≤3 and
y = (yi )1≤i≤3 , then

>
x̂ ∨ ŷ = x1 y2 − x2 y1 x2 y3 − x3 y2 x3 y1 − x1 y3 x1 − y1 x2 − y2 x3 − y3 .
Let m ∈ Inf(d) be an infinitesimal isometry that can be represented as m(x) = Sx + t for some
skew-symmetric matrix S and t ∈ d d
R forevery x ∈ R . We can alternatively represents
 m by using
S t m(x)
the skew-symmetric matrix Ŝ := of size (d +1)×(d +1). Then = Ŝx̂ for every
−t 0 −hm,ti
2 d+1
x ∈ Rd . By regarding Ŝ as a d+1

W
2 -dimensional vector s ∈ R whose (i, j)-the entry is equal
to the (i, j)-th entry of Ŝ, one can canonically represent each infinitesimal isometry m ∈ Inf(d) as a
vector s in 2 Rd+1 . Then for any x, y, z ∈ Rd we have
W

hx − y, m(z)i = hx̂ − ŷ, Ŝẑi = (x̂ − ŷ)> Ŝẑ = hs, (x̂ − ŷ) ∨ ẑi. (20.2)
Consider a body-bar framework (G, b ) and an infinitesimal motion m of (G, b ). If we represent
each mv (v ∈ V (G)) by sv ∈ 2 Rd+1 , then it follows from (20.2) and anticommutativity of ∨ that
W

hmv (be,v ) − mu (be,u ), be,v − be,u i = hsv − su , b̂e,v ∨ b̂e,u i.

In view of this relation, one can alternatively define an infinitesimal motion of (G, b ) as s : V (G) →
W2 d+1
R satisfying
hsv − su , b̂e,v ∨ b̂e,u i = 0 (e = uv ∈ EB ). (20.3)
Rigidity of Body-Bar-Hinge Frameworks 439

Then an infinitesimal motion s is trivial if and only if su = sv for every u, v ∈ V .

Note that (20.3) is a system of linear equations in variable s , where the solution space is the
space
of infinitesimal motions of (G, b ). The rigidity matrix RB (G, b ) is a matrix of size |E(G)| ×
d+1
2 |V (G)| representing the system (20.3), i.e.,

u v
 
e = uv  0 . . . 0 b̂e,u ∨ b̂e,v 0 . . . 0 b̂e,v ∨ b̂e,u 0 . . . 0 .

d+1

Since the set of trivial infinitesimal motions forms a 2 -dimensional linear space, we have the
following.

Proposition 20.2
A body-bar framework (G, b ) in Rd is infinitesimally rigid if and only if rank RB (G, b ) =
d+1
d+1

2 |V (G)| − 2 .

Maxwell-type condition. By Proposition 20.2,

   
d +1 d +1
|E(G)| ≥ |V (G)| −
2 2
if a body-bar framework (G, b) is infinitesimally rigid. This condition can be strengthened as in the
form of Maxwell’s condition for bar-joint frameworks.

Proposition 20.3
d+1

If a body-bar framework (G, b ) is infinitesimally rigid, then G contains 2 edge-disjoint span-
ning trees, or equivalently a spanning ( d+1

d+1

2 , 2 )-tight subgraph.

The equivalence between the two conditions in Proposition 20.3 is due to Nash-Williams’ theo-
rem [32]. (See Section 20.5 for more details.)
Proposition 20.3 is a sufficient condition for the existence of nontrivial infinitesimal motions,
and thus it can be used for proving that a body-bar framework is not infinitesimally rigid. However,
the form of Proposition 20.3 (which is the most popular form in the literature) is not convenient
for this purpose as it is not clear how to show that G does not contain d+1


2 edge-disjoint spanning
trees. The Tutte-Nash-Williams theorem explained in Theorem 20.18 converts Proposition 20.3 to
the following useful form.

Proposition 20.4
A body-bar framework (G, b ) is not infinitesimally rigid if there is a partition P of V (G) satisfying
eG (P) < d+1


2 (|P| − 1), where eG (P) denotes the number of edges connecting different sets in P.

For example, in graph G in Figure 20.1(d), take partition P = {{a}, {b, d}, {c}} of V (G). Then
eG (P) = 5 < 6 = 3(|P| − 1). Hence this partition certifies that the body-bar framework in Fig-
ure 20.1(b) is not infinitesimally rigid.
In fact the connection between infinitesimal rigidity and tree-packings in Proposition 20.3 is ex-
plicit in the combinatorial zero/nonzero pattern of the rigidity matrix RB (G, b ). For d = 1, RB (G, b )
is exactly the incidence matrix of (an orientation) G, and in general RB (G, b ) can be considered as
d+1

the ”union” of 2 copies of the incidence matrix, where the meaning of ”union” can be formal-
ized using matroid union in matroid theory, see [36] for more details. Such a combinatorial pattern
of the rigidity matrix can be exploited even for analyzing singular cases [46]. The details can be
found in Chapter 4.
440 Handbook of Geometric Constraint Systems Principles

20.1.2 Body-Hinge Frameworks

A d-dimensional body-hinge framework models a structure consisting of rigid bodies connected
by hinges. Each hinge is a (d − 2)-dimensional simplex that connects a pair of bodies. The bodies
are free to move in Rd subject to the constraint that the relative motion of any two bodies joined
by a hinge is a rotation about the hinge. We shall summarize basic terminology of body-hinge
frameworks.

Body-hinge framework: A d-dimensional body-hinge framework is a pair (G, h ) of a multigraph

G = (V, EH ) and h : EH → ∆dd−2 , where ∆dd−2 denotes the set of (d − 2)-dimensional simplices
in Rd . Here G represents the underlying graph, where each vertex corresponds to a body and
each edge corresponds to a hinge. h represents the hinge-configuration. See Figure 20.3 and
Figure 20.4(a).
Body-hinge realization: A body-hinge framework (G, h ) with the underlying graph G is called a
body-hinge realization of G.
Replacement: A replacement of a d-dimensional body-hinge framework (G, h ) is a map r : V 3
v 7→ rv ∈ Euc(d) that satisfies

ru (x) = rv (x) (e = uv ∈ EH , x ∈ he ).

A replacement r is nontrivial if ru 6= rv for some pairs of vertices u, v.

Global rigidity: A d-dimensional body-hinge framework (G, h ) is globally rigid if it has no non-
trivial replacements.
Finite motion: A finite motion is a continuous family r t (t ∈ [0, 1]) of replacements of (G, h ) such
that rv0 is the identity for every v ∈ V . A finite motion r t is nontrivial if rvt 6= rut for some t ∈ (0, 1]
and some u, v ∈ V .
Rigidity: A d-dimensional body-hinge framework (G, h ) is rigid if (G, h ) has no nontrivial finite
motion.
Infinitesimal motion: An infinitesimal motion m of a d-dimensional body-hinge framework (G, h)
is a map m : V 3 v 7→ mv ∈ Inf(d) that satisfies

mu (x) = mv (x) (e = uv ∈ EH , x ∈ he ). (20.4)

An infinitesimal motion m is trivial if mu = mv for every u, v ∈ V .

Infinitesimal rigidity: (G, h ) is infinitesimally rigid if every infinitesimal motion of (G, h ) is triv-
ial.
Generic framework: A body-hinge framework (G, h ) is generic if the set of coordinates of the
vertices of the simplices in h is algebraically independent over Q.

Equivalent bar-joint frameworks. As in the case of body-bar frameworks, the rigidity property
of body-hinge frameworks can be captured by looking at equivalent bar-joint frameworks, obtained
by replacing each body by a bar-joint realization of a large enough complete graph. For a graph
G = (V, EH ), the (d-dimensional) body-hinge graph induced by G, denoted by GH , is defined as
follows:
• GH consists of (d + 1)|V (G)| + (d − 1)|E(G)| vertices; for each v ∈ V (G) we have d + 1
vertices xv,1 , . . . , xv,d+1 and for each e ∈ E(G) we have d − 2 vertices xe,1 , . . . , xe,d−1 ;
Rigidity of Body-Bar-Hinge Frameworks 441

a b

d c

(b)
(a)
Figure 20.3
A body-hinge framework in R2 and the underlying graph.

• the hinge H(e) of e is the complete graph on {xe,1 , . . . , xe,d−1 } for each e ∈ E;
• the core C(v) of v is the complete graph on {xv,1 , . . . , xv,d+1 } for each v ∈ V ;
S
• the body B(v) of v is the complete graph on V (C(v)) ∪ e∈δG (v) V (H(e));

• E(GH ) := v∈V E(B(v)).

S

See Figure 20.4.

B(v)
C(v)
H(e)

C(u)
B(u)
(a)
(b)
Figure 20.4
A body-hinge framework in R3 and the body-hinge graph induced by the underlying graph.

Given a body-hinge framework (G, h ), the above construction of GH can be extended at the level
of frameworks to obtain an equivalent bar-joint framework (GH , p). Here p : V (GH ) → Rd is taken
such that he = conv{pp(xe,i ) : 1 ≤ i ≤ d − 1} for every e ∈ EH and p(C(v)) is an affinely independent
set of points independent from the coordinates of the vertices in h. Then (G, h) is infinitesimally
rigid/rigid/globally rigid if and only if (GH , p) is infinitesimally rigid/rigid/globally rigid.

Maxwell-type condition. Suppose that there are two bodies indexed by u and v and connected by
a hinge huv in Rd . The hinge restricts the possible infinitesimal motions mu , mv ∈ Inf(d) assigned to
those bodies by
mu (x) = mv (x) (x ∈ huv ). (20.5)
This is equivalent to
hy − x, mu (x) − mv (x)i = 0 (x ∈ huv , y ∈ Rd ). (20.6)
W2 d+1 d+1

Letting su , sv ∈ R to be the 2 -dimensional vectors representing mu and mv as in the last
section, (20.6) can be written as
hsu − sv , x̂ ∨ ŷi = 0 (x ∈ huv , y ∈ Rd ). (20.7)
442 Handbook of Geometric Constraint Systems Principles

The set of (oriented) segments in Rd is said to be independent if the corresponding Plücker

coordinate vectors form a linearly independent set. For any independent set of ( d+1


2 − 1) segments
[ai , bi ] intersecting huv , we have span{x̂ ∨ ŷ : x ∈ huv , y ∈ Rd } = span{âi ∨ b̂i : 1 ≤ i ≤ d+1


2 − 1}.
Thus (20.7) is further simplified to a system of ( d+1


2 − 1) equations,
 
d +1
hsu − sv , âi ∨ b̂i i = 0 (1 ≤ i ≤ − 1). (20.8)
2

Each equation in (20.7) is nothing but the linear constraint imposed by a bar between
ai and bi .
Therefore the hinge constraint is equivalently regarded as length constraints of ( d+12 − 1) bars
intersecting the hinge.
This replacement of a hinge with bars enables us to regard body-hinge frameworks as a special
case of body-bar frameworks, and a rigidity matrix of a body-hinge framework (G, h ) can be defined
as that of a resulting body-bar framework.
For a graph G and a positive integer k, let kG be the graph obtained by replacingeach edge
with k parallel ones. By the above technique of replacing each hinge with d+1


2 − 1 bars, the
following Maxwell-type condition for body-hinge frameworks follows from Proposition 20.3.

Proposition 20.5
d+1
d+1

If a body-hinge framework (G, h ) is infinitesimally rigid, then ( 2 − 1)G contains 2 edge-
disjoint spanning trees.

20.1.3 Body-Bar-Hinge Frameworks

The combined notion of body-bar frameworks and body-hinge frameworks are useful in some ap-
plications.

Body-bar-hinge framework: A body-bar-hinge framework is a triple (G, b, h) of a 2-edge-colored

graph G = (V, EB ∪ EH ), b : ÊB → Rd , and h : EH → ∆dd−2 . In the underlying graph G = (V, EB ∪
EH ), each vertex corresponds to a body, each edge in EB corresponds to a bar, and each edge in
EH is corresponding to a hinge.
The rigidity/global rigidity/infinitesimal rigidity of a body-bar-hinge framework is defined in an
obvious manner by unifying the definitions of body-bar frameworks and body-hinge frameworks.
The constructions of equivalent bar-joint frameworks are also extendable, and hence the rigidity
question for body-bar-hinge frameworks can be answered by looking at corresponding equivalent
bar-joint frameworks.
For the analysis of infinitesimal rigidity, one may convert each hinge with d+1


2 − 1 independent
bars intersecting the hinge, and the analysis can be carried out in an equivalent body-bar framework.
This in particular implies the following.

Proposition 20.6
If a
body-bar-hinge framework (G = (V, EB ∪ EH ), b , h ) is infinitesimally rigid, then H contains
d+1
2 edge-disjoint spanning trees, where H is the graph obtained from G by replacing each edge
in EH with d+1


2 − 1 parallel edges.
Rigidity of Body-Bar-Hinge Frameworks 443

20.2 Generic Rigidity

In this section we give a characterization of the rigidity of generic body-bar-hinge frameworks.

20.2.1 Body-Bar Frameworks

A remarkable feature of body-bar frameworks is that, in the generic case, the converse direction
of Proposition 20.3 is true in any dimension. This is in contrast to bar-joint frameworks, whose
combinatorial characterization is a central open question in rigidity theory.
Theorem 20.1 (Tay [39]) Let G be a multi-graph.
Then a generic d-dimensional body-bar realiza-
tion of G is rigid if and only if G contains d+1
2 edge-disjoint spanning trees.
For example, for the body-bar framework in Figure 20.1(a), three edge-disjoint spanning trees
in the underlying graph in Figure 20.1(c) can be used for a certificate of the infinitesimal rigidity.
Depending on the context, different combinatorial conditions are used, see Theorem 20.20.
The necessity of Theorem 20.1 follows from Proposition 20.3. Tay’s original proof for the suf-
ficiency is based on a constructive characterization of (k, k)-tight graphs, see Theorem 20.24. Sev-
eral different proofs are known in the literature, and one of the simplest proofs is the following.
Consider proving a slightly generalized statement by induction on the lexicographical ordering of
(V (G),
E(G)): a generic d-dimensional body-bar realization of G has k degrees of freedom if G is
( d+12 , d+1

2 )-sparse with |E(G)| = d+1

2 |V (G)| − d+1

2 − k. We assume E(G) 6= ∅ since other-
wise the claim is trivial. Take any edge e = uv and consider a generic realization (G − e, b ) of G − e,
which has k + 1 degrees of freedom by induction. If a generic realization of G has more than k de-
grees of freedom, then every infinitesimal motion m of (G − e, b ) satisfies mu = mv . Now, consider
the graph G/e obtained by contracting e. Then, as mu = mv holds for every infinitesimal motion m of
(G − e, b ), the degree of freedom of a generic realization of G/e should be
at least
that of (G − e, b ),
which is k + 1. However, one can show that G/e contains a spanning ( d+1 d+1
2 , 2 )-sparse subgraph
H with |E(H)| ≥ d+1 d+1



2 |V (H)| − 2 − k, which is a contradiction since a generic realization of H
has degree of freedom at most k by induction.
Another simple proof was due to Whiteley [47] using tree-decompositions, and the idea
is used
in various different places [21, 22, 26, 35]. Suppose that G can be decomposed into d+1 2 edge-
disjoint spanning connected subgraphs Ti, j (1 ≤ i < j ≤ d + 1). Take an affinely independent set
x1 , . . . , xd+1 of points in Rd and set b to be be,u = xi and be,v = x j if e = uv ∈ Ti, j . Then due to the
specialty of b it can be easily shown that the rigidity matrix RB (G, b ) is equivalent to the direct
sum of the adjacency matrices of Ti, j over all pairs i, j. Thus the rigidity matrix has rank equal to
d+1

2 (|V (G)| − 1), and (G, b ) is infinitesimally rigid.

20.2.2 Body-Hinge Frameworks

Tay [40, 41] and Whiteley [47] independently observed that a similar combinatorial condition as in
Theorem 20.1 can be used to characterize the generic rigidity of body-hinge frameworks.

Theorem 20.2 (Tay [40, 41], Whiteley [47]) Let G be a multi-graph. Then a generic d-
dimensional body-hinge realization of G is rigid if and only if ( d+1 d+1



2 − 1)G contains 2 edge-
disjoint spanning trees.

Whiteley’s original proof [47] is simpler although Tay [40, 41] proved a more general statement
(see Subsection 20.3.2). In fact, a proof of Theorem 20.1 given in the last section is
applicable for
Theorem 20.2. To see this, suppose that ( d+1 can be decomposed into d+1


2 − 1)G 2 edge-disjoint
444 Handbook of Geometric Constraint Systems Principles

spanning connected subgraphs Ti, j (1 ≤ i < j ≤ d + 1). As in the last section, we construct a body-
bar realization (( d+1 d+1



2 − 1)G, b ) of ( 2 − 1)G by taking an affinely independent set x1 , . . . , xd+1
of points in Rd and setting b to be be,u = xi and be,v = x j if e ∈ Ti, j . Then due to the specialty of b ,
one can check that (( d+1 d+1



2 − 1)G, b ) is infinitesimally rigid. Moreover the set of the 2 − 1 bars
associated with an edge f ∈ E(G) intersects a (d − 2)-dimensional simplex, denoted by h f . Thus by
setting h to be h ( f ) = h f for each f ∈ E(G) we get an infinitesimally rigid body-hinge realization
(G, h ), as (G, h ) is infinitesimally rigid if and only if (( d+1


2 − 1)G, b ) is infinitesimally rigid.
As noted by Jackson and Jordán [21], it is possible to extend the characterization to body-bar-
hinge frameworks by applying Whiteley’s proof. Moreover the rank of the rigidity matrix can be
combinatorially described.

Theorem 20.3 (Jackson and Jordán [21]) Let G = (V, EB ∪ EH ) be a 2-edge-colored multigraph,
and let H be a graph obtained from G by replacing each edge in EH by d+1


2 − 1 parallel copies.
d
Then the rank of the rigidity matrix of a generic body-bar-hinge realization of G in R is equal to
  
d +1
min (|V (G)| − |Q|) + eH (Q) : Q is a partition of V (H) .
2

20.3 Other Related Models

In this section we summarize known results on variants of body-bar-hinge frameworks.
k-plate: A k-plate is a compact connected set in Rd whose affine span is k-dimensional.
k-plate-bar framework: A k-plate-bar framework is a structure consisting of k-plates connected
by bars. Each bar can connect two distinct plates. The underlying graph is defined by associating
each k-plate with a vertex and each bar with an edge. See Figure 20.5 for an example.
k-plate-bar realization: A k-plate-bar framework with the underlying graph G is called a k-plate-
bar realization of G.

Identified body-hinge framework: An identified body-hinge framework is a generalized model of

a body-hinge framework, where each hinge may connect more than two bodies.
Panel-hinge framework: A panel-hinge framework is a structure consisting of rigid panels con-
nected by hinges, where a panel means a (d − 1)-plate in Rd .

Square of a graph: The square G2 of a graph G is the graph on V (G) which contains an edge uv
if and only if the distance between u and v in G is at most two. See Figure 20.6.
Molecular framework: A molecular framework is a 3-dimensional body-hinge framework in
which all the hinges incident to each body pass through a point.
Body-pin framework: A body-pin framework is a structure consisting of bodies connected at
points.
Rigidity of Body-Bar-Hinge Frameworks 445

(b)
(a)
Figure 20.5
A 1-plate-bar framework in R3 and the underlying graph.

20.3.1 Plate-Bar Frameworks

The rigidity/infinitesimal rigidity/global rigidity of a k-plate-bar framework is defined in the identi-
cal manner as those of body-bar framework. The only difference is that for a k-plate-bar framework
the space of replacements of each plate is defined by factoring the isometries fixing the plate.
Although plate-bar frameworks appear in real applications, understanding the rigidity property is
also theoretically important. When k = d, those are body-bar frameworks, whose generic rigidity is
characterized combinatorially in Theorem 20.1. When k = 0, those are bar-joint frameworks, whose
generic rigidity has not yet been characterized when d ≥ 3. The concept of k-plate-bar frameworks
interpolates between those two well-studied models.
It follows from the definition that the same combinatorial condition can be used to characterize
the generic rigidity for k = d −1. The following result by Tay [40, 41] solves the case when k = d −2.

Theorem 20.4 (Tay [40, 41]) Let G be a multigraph. Then
G has an
infinitesimally rigid (d − 2)-
plate-bar realization in Rd if and only if G contains a ( d+1 d+1
2 − 1, 2 )-tight spanning subgraph.

Tanigawa [36] gave an alternative inductive proof starting from Theorem 20.1. A constructive
characterization of (k, k + 1)-tight graphs is known by Frank and Szegő [10].
It seems that going beyond Theorem 20.4 is less explored. Understanding (d − 3)-plate-bar
frameworks is certainly difficult for d = 3, but for higher dimensions the Maxwell-type condition
may be necessary and sufficient.

20.3.2 Identified Body-Hinge Frameworks

The body-hinge model given in Section 20.1.2 can capture identified body-hinge frameworks by
locating several hinges at the same position. However the resulting body-hinge framework is no
longer generic and Theorem 20.2 cannot be applied to identified body-hinge frameworks.
In order to capture its combinatorics correctly, the underlying graph of an identified body-hinge
framework is defined by a hypergraph by associating each body with a vertex and each hinge with
a hyperedge, or equivalently a bipartite graph G = (VB ∪ VH , E) whose one side VB corresponds to
the set of bodies and the other side VH corresponds to the set of hinges. There is an edge between
u ∈ VB and v ∈ VH if the hinge of v is incident to the body of u.
Since each hinge is a (d − 2)-plate, each identified body-hinge framework can be considered as
a framework consisting of bodies and (d − 2)-plates linked by bars. Hence the following extension
of Theorem 20.2 follows as a corollary of Theorem 20.4.

Theorem 20.5 (Tay [40, 41]) Let G = (VB ∪VH , E) be a bipartite
graph. Then a generic identified
body-hinge realization of G in Rd is rigid if and only if ( d+1
2 − 1)G contains a ( d+1
d+1

2 , 2 −
446 Handbook of Geometric Constraint Systems Principles

1, 2 )-tight spanning subgraph H, i.e., |E(H)| = d+1

d+1

|VB | + ( d+1 − 1)|VH | − d+1




2 2 2 and
d+1
d+1
d+1

|E(I)| ≤ 2 |VB ∩V (I)| + ( 2 − 1)|VH ∩V (I)| − 2 for any nonempty subgraph I of H.

20.3.3 Panel-Hinge Frameworks

Panel-hinge frameworks are another important variants of body-hinge frameworks that frequently
appear in structural engineering. Theoretically they also have a rich historical background dating
back to Cauchy. Cauchy’s theorem says that the 2-skeleton of a strictly convex polytope is rigid in
R3 if it is regarded as a panel-hinge framework by identifying each face as a panel and each edge as
a hinge. The rigidity of polytopes is explained in detail in Chapter 13.
A panel-hinge framework can be regarded as a body-hinge framework with hinge-coplanarity
condition, i.e., the set of hinges incident to a body lies on a (d − 1)-dimensional subspace. Hence
the Maxwell-type necessary condition for infinitesimal rigidity (Proposition 20.5) is still valid for
any panel-hinge framework. On the other hand the combinatorial characterization for generic body-
hinge frameworks (Theorem 20.2) cannot be applied because of the hinge-coplanarity condition.
Tay and Whiteley [42] conjectured that the same combinatorial condition is still valid for panel-
hinge frameworks. This conjecture was confirmed by Jackson and Jordán [20] for d = 2, and then
by Katoh and Tanigawa [25] for general d.

Theorem 20.6 (Katoh and Tanigawa [25]) Let G be a graph.  Then a generic d-dimensional
panel-hinge realization of G is rigid if and only if d+1 d+1



2 − 1 G contains 2 edge-disjoint span-
ning trees.

Extending Theorem 20.6 to identified panel-hinge frameworks is an important unsolved prob-

lem.

20.3.4 Molecular Frameworks

A molecular framework is a special class of body-hinge frameworks which can be used to under-
stand the rigidity of ideal molecules. Consider a simplest model of ideal molecules in R3 , where
atoms are connected only by covalent bonds. Each bond constrains the distance between two atoms
while two consecutive bonds fixes the angle between the bonds. Hence the rigidity property of an
ideal molecule with the underlying graph G is analyzed by looking at its square G2 . See Figure 20.6.

Figure 20.6
G and G2 , where G consists of the bold edges and G2 consists of the bold edges and the dotted
edges.

Observe that for each vertex v ∈ V (G) the neighbors of v and v itself induce a complete sub-
graph in G2 . By regarding such a complete subgraph as a body associated with v, we see that two
bodies share a hinge if and only if the associated two vertices are adjacent in G. Thus, with this
Rigidity of Body-Bar-Hinge Frameworks 447

identification between complete subgraphs and bodies, a bar-joint framework (G2 , p ) of G2 leads
to a body-hinge framework (G, h ) with the underlying graph G, where h (i j) = [pp(i), p ( j)] for each
edge i j ∈ E(G).
Formally we have the following:

Proposition 20.7
Let (G2 , p ) be a bar-joint framework in general position for a graph G with minimum degree at
least two, and let (G, h ) be the 3-dimensional body-hinge framework with h (i j) = [pp(i), p ( j)] for
each edge i j ∈ E(G). Then (G2 , p ) is infinitesimally rigid in R3 if and only if (G, h ) is infinitesimally
rigid in R3 .

Note that the body-hinge framework (G, h ) in Proposition 20.7 is hinge-concurrent.

Tay and Whiteley [42] observed that a molecular framework in R3 can be converted to a panel-
hinge framework through polarity. Recall that in a three-dimensional projective space and its polar
space a point is associated with a hyperplane while a line is associated with a line keeping the
incidence structure. Hence for each hinge-concurrent body-hinge framework in R3 one can associate
a panel-hinge framework in the polar space. A classical observation by Crapo and Whiteley [6] is
that the infinitesimal rigidity is preserved under projective polarity, which implies the following.
Theorem 20.7 (Crapo and Whiteley [6]) Let G be a graph. Then a molecular framework (G, h ) is
infinitesimally rigid in R3 if and only if the polar panel-hinge framework (G, h ∗ ) is infinitesimally
rigid in R3 .
Combining Theorem 20.6, Proposition 20.7, and Theorem 20.7, we have the following, which was
known as the molecular rigidity conjecture before [25]:
Theorem 20.8 (Katoh and Tanigawa[25]) For a graph G, the following are equivalent:
(a) A generic molecular framework with the underlying graph G is rigid in R3 ;
(b) A generic panel-hinge realization of G is rigid in R3 ;
(c) A generic bar-joint realization of G2 is rigid in R3 ;
(d) 5G contains six edge-disjoint spanning trees.
Jackson and Jordán [18, 19] gave a number of important consequences of the molecular conjec-
ture.
Theorem 20.9 (Jackson and Jordán [19]) Let G = (V, E) be a graph with minimum degree at least
two. Then the rank of the generic 3-dimensional rigidity matroid of G2 is equal to min{3|V |−6|P|−
3 + 5eG (P) : P is a partition of V (G)}.
See Chapter 1 for the definition of the generic rigidity matroid.
Theorem 20.10 (Jackson and Jordán [19]) Let G = (V, E) be a graph. Then G2 is independent
in the 3-dimensional rigidity matroid if and only if G2 satisfies Maxwell’s condition, i.e., for any
F ⊆ E(G2 ) with |V (F)| ≥ 3, |F| ≤ 3|V (F)| − 6.
Recall that a rigid component of a graph H = (V, E) is a rigid subgraph of H which is maximal
with respect to inclusion. In several applications, the task is to identify rigid components rather than
checking just rigidity, and hence characterizing rigid components is important.
Theorem 20.11 (Jackson and Jordán [18]) Let G = (V, E) be a graph of minimum degree at least
two. Then for each rigid component C of G2 there is a 6/5-brick B = G[X] of G with C = G[X ∪
NG (X)]2 . Conversely for each 6/5-brick B = G[X] of G the subgraph G[X ∪ NG (X)]2 is a rigid
component of G2 .
448 Handbook of Geometric Constraint Systems Principles

For the definition of 6/5-brick, see Section 20.5.

As explained in Section 20.5.4 there is an efficient algorithm for computing the collection of
6/5-bricks of G, and Theorem 20.11 leads to an efficient algorithm for computing the rigid compo-
nent decomposition of G2 .
It was also shown in [18] that Theorem 20.8 implies the rank formula of G2 in terms of the rigid
component decomposition of G2 .
Understanding the mechanisms of proteins is a challenging question in molecular biology, and
the theory of statics forms a basics for attacking the problem. Combinatorial approaches based on
Theorem 20.8 are implemented in several places, where the popular pebble game algorithm or its
variant is commonly used for checking the condition of Theorem 20.8. See, e.g., Section 20.5.4 and
[44, 49] for more details. A geometric analysis of various special cases can be found in [48, 49].

20.3.5 Body-Pin Frameworks

Body-pin frameworks also form a natural class of structures in real applications. If a pin may connect
more than two bodies, understanding rigidity properties of body-pin frameworks turns out to be
challenging as any rigidity question of bar-joint frameworks can be formulated in this body-pin
model. Thus in this subsection we shall focus on the model where each pin connects exactly two
bodies. Under this assumption, the underlying graph of a body-pin framework becomes a graph,
where each vertex represents a body and each edge represents a pin connecting the two bodies
associated with the endvertices. A body-pin framework is then denoted by a pair (G, p ) of the
underlying graph G = (V, EP ) and the pin-configuration p : EP → Rd . A replacement of a body-pin
framework (G, p ) is defined as m : V 3 v 7→ mv ∈ Euc(d) that satisfies

mu (pe ) = mv (pe ) (e = uv ∈ EP ).

(G, p ) is said to be rigid if it has no nontrivial finite motion, i.e., a continuous family r t (t ∈ [0, 1])
of replacements of (G, p ) such that m0v is identity and rvt 6= rut for some t ∈ (0, 1].
Dress gave a conjecture for characterizing the rigidity of generic bar-joint frameworks. Although
Dress’ conjecture was disproved by Jackson and Jordán [17], one may wonder if it still holds for
some sub-classes of graphs. Jackson, Jordán and Tanigawa conjecture that Dress’ conjecture is true
for generic body-pin frameworks. The conjecture has the following attractive form for body-pin
frameworks.
Conjecture 5 Let G be a graph. Then a generic three-dimensional body-pin realization of G is
rigid if and only if X
hG (X, X 0 ) ≥ 6(|P| − 1)
{X,X 0 }∈(P
2)

P

for every partition P of V , where 2 denotes the set of pairs of subsets in P

if dG (X, X 0 ) ≥ 3


6
if dG (X, X 0 ) = 2

5
hG (X, X 0 ) =


3 if dG (X, X 0 ) = 1
if dG (X, X 0 ) = 0

0

and dG (X, X) denotes the number of edges in G connecting X and X.

Note that, if hG were defined to be hG (X, X 0 ) = 6 for dG (X, X 0 ) = 2, then the combinatorial condition
is equivalent to the Tutte-Nash-Williams condition in Theorem 20.18 for 3G to contain six edge-
disjoint spanning trees.
Rigidity of Body-Bar-Hinge Frameworks 449
a

c b

v0

(a) (b)
Figure 20.7
A body-bar framework with bar-boundary and the underlying graph.

20.3.6 Body-Bar Frameworks with Boundaries

In engineering applications a structure has some links to the external environment such as the ground
or walls. In rigidity theory, such a structure is commonly understood by using a pinned framework,
where pinned points are fixed in the ambient space. Its rigidity property can be captured by the
conventional rigidity of frameworks (without pinned points) by linking every pair of fixed points by
a bar. However, in the body-bar model, the resulting body-bar frameworks are no longer generic,
and Theorem 20.1 cannot be applied. Katoh and Tanigawa [26] gave an extension of Theorem 20.1
to pinned body-bar frameworks.
The underlying combinatorics of a pinned body-bar framework is captured by a graph G =
(V, EB ) along with a multisubset P of V . Then a pinned body-bar framework is a tuple (G, P, b , p )
with b : ÊB → Rd and p : P → Rd , where b and p encode the locations of bars and pins, respectively.
An infinitesimal motion of (G, P, b , p ) is an infinitesimal motion of the ordinary body-bar frame-
work (G, b ) with an extra pinning constraint:

mv (pp(v)) = 0 (v ∈ P).

We say that (G, P, b , p ) is infinitesimally rigid if (G, P, b , p ) has no nonzero infinitesimal motion.
By extending Tutte-Nash-Williams’ tree-packing theorem, Katoh and Tanigawa [26] proved the
following.
Theorem 20.12 (Katoh and Tanigawa [26]) Let G = (V, EB ) be a graph, P be a multiset P of ver-
tices in V , and p : P → Rd . Then there exists a bar-configuration b : ÊB :→ Rd such that the pinned
body-bar framework (G, P, b, p) is infinitesimally rigid if and only if
  X +1
X dX
d +1
eG (P) ≥ |P| − (d − i + 1)
2
X∈P i=1

for every partition P of V , where dX denotes the dimension of the affine span of p(X ∩ P), which is
−1 if X ∩ P = ∅.

It should be noted that p is given in advance in Theorem 20.12 and it may not be generic.
Theorem 20.12 is a corollary of the following result on body-bar frameworks with bar-boundary. A
body-bar framework with bar-boundary is a body-bar framework, some of whose bodies are linked
to the ground by bars. By regarding the ground as a body fixed in the ambient space, the underlying
combinatorics can be captured by the augmented graph G ∪ {v0 } by adding vertex v0 representing
the fixed body. See Figure 20.7.

Theorem 20.13 (Katoh and Tanigawa [26]) Let G ∪ {v0 } be a graph with a designated vertex v0 ,
E0 be the set of edges in G ∪ {v0 } incident to v0 , and b 0 : Ê0 → Rd . Then there exists an extension
450 Handbook of Geometric Constraint Systems Principles

b : Ê(G ∪ {v0 }) → Rd of b 0 such that the body-bar framework (G ∪ {v0 }, b ) is infinitesimally rigid
if and only if  
d +1 X
eG (P) ≥ |P| − dim span{b0e,u ∨ b0e,v0 : e ∈ E0 (X)}
2
X∈P

for every partition P of V (G), where E0 (X) denotes the set of edges in E0 incident to X.

It should be noted that b 0 is given and may not be generic in Theorem 20.13. Theorem 20.13
also has a nice application to constrained body-bar frameworks, where each body is allowed to move
only specific directions. For example, if a body is a horizontally flat shaped disc, then it is natural to
restrict its motion to those keeping in a horizontal position. Note that body-bar frameworks with bar-
boundary capture such constrained body-bar frameworks. Even if allowable directions of motions
of bodies are non-generic, Theorem 20.13 gives a combinatorial characterization of infinitesimal
rigidity.

20.3.7 Other Variants

There are several further extensions of Theorem 20.1 and Theorem 20.2. Jackson and Nguyen [22]
considered the length-direction rigidity of body-bar frameworks, where some of the edges represent
direction-constraints between points on bodies. They gave an extension of Theorem 20.1 in terms
of sparsity condition of the underlying graphs. Haller et al. [14] discussed further different types of
geometric constraints among bodies. See Chapter 23 for this general model.
Theorem 20.1 and Theorem 20.2 assume generic placements of bars and hinges, but such
generic assumptions sometimes limit applications. In [13] Guest, Schulze, and Whiteley discussed
Maxwell-type condition for body-bar frameworks under point group symmetry. In [35, 37] exten-
sions of Theorem 20.1 and Theorem 20.2 were given for body-bar frameworks with point group
symmetry. In [2, 34, 37] extensions of Theorem 20.1 to periodic infinite body-bar frameworks were
discussed. See Chapter 25 for more details.

20.4 Generic Global Rigidity

One of the most challenging questions in rigidity theory is to give a combinatorial characterization
of the global rigidity of generic bar-joint frameworks. This was solved for d = 2 by Jackson and
Jordán [16] and it remains open for d ≥ 3 (see Chapter 21). Again, in contrast to bar-joint frame-
works, a succinct characterization is known for body-bar-hinge frameworks in general dimension.

20.4.1 Body-Bar Frameworks

As mentioned in Section 20.1, the global rigidity question for body-bar-hinge frameworks can be
reduced to that of equivalent bar-joint frameworks. This conversion allows us to use several related
results from the theory of global rigidity of bar-joint frameworks. The most important result in our
context is Hendrickson’s theorem [15], which states that if a generic bar-joint framework (G, p ) on
at least d +2 vertices is globally rigid in Rd , then G is (d +1)-connected and the framework is redun-
dantly rigid, that is, (G − e, p ) is rigid for every edge of G. If we apply this result for the equivalent
bar-joint frameworks to body-bar frameworks we immediately get the following proposition.

Proposition 20.8
d+1

If a generic body-bar framework (G, b ) is globally rigid in Rd , then G − e contains 2 edge-
disjoint spanning trees for every edge e of G.
Rigidity of Body-Bar-Hinge Frameworks 451

The global rigidity of generic body-bar frameworks was characterized by Connelly, Jordán and
Whiteley [4] by showing the sufficiency of the condition in Proposition 20.8.

Theorem 20.14 (Connelly, Jordán, and Whiteley [4]) Let G be a multi-graph. Then a generic d-
dimensional body-bar realization of G is globally rigid if and only if G − e contains d+1


2 edge-
disjoint spanning trees for every edge e of G.

The original proof of [4] used the constructive characterization of Frank and Szegő [10] (see
Theorem 20.25). A different, simplified proof was given by Tanigawa [38]. It was shown in [38] that
a generic bar-joint framework (G, p ) is globally rigid in Rd if (G, p ) is vertex-redundantly rigid, i.e.,
(G − v, p ) is rigid for every v ∈ V (G). It turns out that if G satisfies the combinatorial condition in
Theorem 20.14 then a generic realization of the body-bar graph GB is vertex-redundantly rigid.

20.4.2 Body-Hinge Frameworks

 
Tanigawa [38] observed that if d+1 G − 2e contains d+1



2 − 1 2 edge-disjoint spanning trees for
every edge e of G, then a generic d-dimensional body-hinge realization of G is vertex-redundantly
rigid and hence globally rigid. Especially, if a body-hinge framework is hinge-redundantly rigid
(that is, it remains rigid after a deletion of an arbitrary hinge), then it is globally rigid. With some
further ideas Jordán, Király and Tanigawa [23] characterized the global rigidity of generic body-
hinge frameworks.

Theorem 20.15 (Jordán, Király, and Tanigawa [23, 24]) Let G be a multigraph and d
≥ 3. Then
a generic d-dimensional body-hinge realization of G is globally rigid if and only if ( d+1
2 − 1)G − e
d+1
d+1

contains 2 edge-disjoint spanning trees for every edge e of ( 2 − 1)G.

Unlike body-bar frameworks, the necessity of Theorem 20.15 is not a direct consequence of
Hendrickson’s theorem (yet the proof in [23] relies on Hendrickson’s theorem). The key idea is
that if d ≥ 3 and the degree is at least 2 in G, then the rigidity/global rigidity of a generic body-
hinge framework (G, h ) is equivalent to the rigidity/global rigidity of a bar-joint framework that
arises from the equivalent bar-joint framework (GH , p ) by removing each core C(v) for every v ∈
V (G). This latter bar-joint framework (SGH , p 0 ) is called a skeleton. It was shown that (SH , p 0 ) is not
G
redundantly rigid if G does not satisfy the combinatorial condition of Theorem 20.15 if d = 3. For
d ≥ 4, a slightly modified bar-joint framework is used for applying the same argument.
For d = 2, a much simpler condition characterizes global rigidity. (See also Corollary 20.19 for
a connection to redundant rigidity.)
Theorem 20.16 (Jordán, Király, and Tanigawa [23]) Let G be a multigraph and d ≥ 3. Then a
generic 2-dimensional body-hinge realization of G is globally rigid if and only if G is 3-edge-
connected.

The proof technique in Theorem 20.15 can be extended to body-bar-hinge frameworks.

Theorem 20.17 (Jordán, Király, and Tanigawa [23]) Let G = (V, EB ∪ EH ) be a 2-edge-colored
multigraph, d ≥ 3, and let H be a graph obtained from G by replacing each edge in EH by d+1


2 −1
parallel copies. Then a generic
d-dimensional body-bar-hinge realization of G is globally rigid if
and only if H − e contains d+1
2 edge-disjoint spanning trees for every edge e of H.
Hinge-redundant rigidity/global rigidity of body-hinge frameworks were discussed by
Kobayashi et al. [29].
In Section 20.3 we have seen various extensions of Tay’s theorem (Theorem 20.1) for body-bar
frameworks and Tay-Whiteley’s theorem (Theorem 20.2) for body-hinge frameworks. To the best
452 Handbook of Geometric Constraint Systems Principles

(a) (b)

Figure 20.8
(a) C6H is not globally rigid because (b) its skeleton is minimally rigid.

of our knowledge the corresponding global rigidity questions have not been addressed. The most
important and interesting question would be the characterization of the global rigidity of molecular
frameworks or generic bar-joint realizations of the square of graphs in R3 . It was shown in [23] that,
contrary to the rigidity case, the combinatorial condition in Theorem 20.15 is no longer sufficient.
Figure 20.6 is an example that satisfies the condition in Theorem 20.15 but its square is not globally
rigid as it is not 4-connected.

20.4.3 Counterexamples to Hendrickson’s Conjecture

Hendrickson [15] conjectured that his necessary conditions are also sufficient for the global rigidity
of bar-joint frameworks. Although the conjecture is true for d ≤ 2 [16], Connelly gave counterex-
amples for d ≥ 3 [3]. Connelly and Whiteley [5] then conjectured that there exist no counterexample
in Rd containing Kd+1 as a subgraph. They also conjectured that the number of counterexamples is
finite in Rd , for all d ≥ 3. Later Frank and Jiang [11] disproved the latter conjecture for d ≥ 5, and
then Jordán, Király and Tanigawa [23] disproved both conjectures for all d ≥ 3.
The counterexamples by Jordán, Király, and Tanigawa [23] were built
based on Theorem 20.15.
By Theorem 20.2, if (G, b ) is rigid, then ( d+1 contains d+1


2 − 1)G 2 edge-disjoint spanning
trees, which in turn implies that a generic bar-joint realization (GH , p ) is (d + 1)-connected and
redundantly rigid. In other words, GH
satisfies Hendrickson’s condition. On the other hand, if
( d+1 − e does not contain d+1 edge-disjoint spanning trees for some edge e, then (GH , p )


2 − 1)G 2
is not globally rigid by Theorem 20.15. Since there are infinitely many graphs G satisfying the above
two combinatorial properties, we get an infinite family of counterexamples.
The simplest example is the case when G is the cycle C6 of size six: a generic realization of C6H
is 4-connected and redundantly rigid but it is not globally rigid by Theorem 20.15. See Figure 20.8.

20.5 Graph Theoretical Aspects

In this section we summarize some graph theoretical results concerning body-bar-hinge frameworks.

k-edge-connectivity: An undirected graph G = (V, E) is k-edge-connected if there are k edge-

disjoint paths between every pair of vertices in V .
Rigidity of Body-Bar-Hinge Frameworks 453

k-connectivity: G = (V, E) is k-connected if there are k internally vertex-disjoint paths between

every pair of vertices in V .
Tree-connectivity: G = (V, E) is (k, `)-tree-connected if G − F contains k edge-disjoint spanning
trees for all F ⊆ E with |F| ≤ `. When ` = 0 or 1 a (k, `)-tree-connected graph is also k-tree-
connected or highly k-tree-connected, respectively.
Rooted (k, `)-arc-connectivity: For positive integers k and `, a digraph D = (V, A) with a root
vertex r0 ∈ V is r0 -rooted (k, `)-arc-connected if there are k arc-disjoint (one-way) paths from
r0 to every other vertex and there are ` arc-disjoint paths from every vertex to r0 . A digraph D
is rooted (k, `)-arc-connected if it is r0 -rooted (k, `)-arc-connected for some r0 ∈ V .
Brick: Let q = b/c with positive integers b and c. A q-brick of a graph G is a maximal subgraph H
such that cH is b-tree-connected.
Superbick: Let q = b/c with positive integers b and c. A q-superbrick of a graph G is a maximal
subgraph H such that cH is highly b-tree-connected.
Pinching: Let G be an undirected graph. By pinching a set F of edges in G, we mean an operation
that inserts a new vertex and subdivides the edges in F by the new vertex.

(k, `)-sparsity matroid: A matroid on the edge set of a graph G is a (k, `)-sparsity matroid (or
(k, `)-count matroid) if its independent set family is the family of the edge sets of the (k, `)-
sparse subgraphs in G.

20.5.1 Tree Packing and Connectivity

Connectivity is one of the most fundamental concepts in graph theory and various variants have
been investigated in the literature. The most relevant one in our context is tree-connectivity. By
Menger’s theorem, k-edge-connectivity (resp., k-connectivity) is equivalent to the property that the
graph remains connected after the removal of at most k − 1 edges (resp., k − 1 vertices). The fol-
lowing theorem by Tutte [45] and Nash-Williams [31] gives a Menger-type characterization for
(k, `)-tree-connected graphs.
Theorem 20.18 (Tutte [45] and Nash-Williams [31]) A graph G = (V, E) is (k, `)-tree-connected
if and only if
eG (P) ≥ k(|P| − 1) + ` (20.9)
holds for every partition P of V .
Theorem 20.18 gives a simple certificate for a graph not being (k, `)-tree-connected. Such a certifi-
cate is useful even in the rigidity context as demonstrated in Proposition 20.4.
 In
therigidity context Theorem 20.18 implies a robust version of Theorem 20.1. Namely G is
d+1
2 , ` -tree-connected if and only if a generic body-bar realization of G is `-edge-redundantly
rigid in Rd , i.e., it is rigid after the removal of any ` edges.
In several papers a graph G = (V, E) satisfying the latter condition of Theorem 20.18 is called
(k, `)-partition-connected. A (k, 0)-partition-connected graph is called k-partition-connected. Ob-
serve that (k, `)-partition-connectivity coincides with (k + `)-edge-connectivity when ` ≥ k, which
means that a generic body-bar realization of G is `-edge-redundantly rigid if and only if G is (k + `)-
edge-connected. Similarly, by a simple edge-counting argument, Theorem 20.18 implies the follow-
ing
Corollary 20.19 G is k-edge-connected if and only if 2G is (k, k)-tree-connected.
454 Handbook of Geometric Constraint Systems Principles

Graph orientations are important subjects even in rigidity theory as they appear in the pebble
game algorithm for checking (k, `)-sparsity. As pointed in [1], the pebble game can be seen as a
variant of Hakimi’s classical orientation theorem. A far reaching generalization of Hakimi’s orien-
tation theorem was shown by Frank [8], which leads to a characterization of (k, `)-tree-connectivity
in terms of graph orientations.
By Menger’s theorem, the r0 -rooted (k, `)-arc-connectivity of a digraph D = (V, E) is equivalent
to the following: δD (X) ≥ k and ρD (X) ≥ ` for every set X ( V with r0 ∈ X where δD (X) and
ρD (X) denote the numbers of out-going arcs from X and in-coming arcs to X in D, respectively.
From a general orientation result of Frank [8], it was shown that (k, `)-arc-connected orientability
is equivalent to (k, `)-partition-connectivity.
We summarize the characterizations we have seen.
Theorem 20.20 For positive integers k, `, d with k ≥ ` and a graph G = (V, E), the following are
equivalent:
• G is (k, `)-tree-connected;
• G is (k, `)-partition-connected (Tutte [45] and Nash-Williams [31]);
• G contains a (k, k)-tight spanning graph after removing any ` edges (Nash-Williams [32]);

• G has a rooted (k, `)-arc-connected orientation (Frank [8]);

• G has an r-rooted (k, `)-arc-connected orientation for any r ∈ V (Frank [8]).

Moreover, if k = d+1


2 , then the above conditions are equivalent to the following:

• A generic d-dimensional body-bar realization of G is `-edge-redundantly rigid (Tay [39]).

If k = d+1


2 and ` ≥ 1, then the above conditions are equivalent to the following:
• A generic d-dimensional body-bar realization of G is (` − 1)-edge-redundantly globally
rigid (Connelly, Jordán, and Whiteley [4]).
Frank’s characterization in terms of orientability leads to the most compact certificate for (k, `)-
tree-connectivity, and thus for `-edge-redundant rigidity.
Nash-Williams’ result [32] in Theorem 20.20 implies that G is (k, k)-tree-connected if and only
if the rank of the (k, k)-sparsity matroid is equal to k|V | − k. Hence one can investigate the properties
of tree-connectivity through the (k, k)-sparsity matroid.
It is important to understand which rigid frameworks are globally rigid. In view of Theorem 20.1
and Theorem 20.14, the following corollary gives a matroidal view for such a question.
Corollary 20.21 Suppose that a graph G = (V, E) is k-tree-connected. Then G is highly k-tree-
connected if and only if every e ∈ E is covered by a circuit of the (k, k)-sparsity matroid of G, that
is, the (k, k)-sparsity matroid of G is connected.
A stronger version of Corollary 20.21 was observed by Király [28].

Theorem 20.22 (Király [28]) Let G = (V, E) be a k-tree connected graph and let T = (V, F) be a
spanning (k, k)-tight subgraph of G. Then G − e is k-tree-connected if and only if e ∈ E − F or e is
induced by a minimal (k, k)-tight subgraph of T covering both u and v for some edge uv ∈ E − F.
Rigidity of Body-Bar-Hinge Frameworks 455

20.5.2 Brick Partitions

In order to find the maximal rigid or globally rigid components of a body-bar-hinge framework, one
would like to determine maximal k-tree-connected and highly k-tree-connected subgraphs. This can
be done using a result of Jackson and Jordán [18]. It follows from Theorem 20.18 that the definitions
of q-bricks and q-superbricks for a rational number q are independent from the representation of q
as b/c. As the union of two (k, `)-tree-connected graphs with at least one common vertex is also
(k, `)-tree-connected, the vertex sets of q-bricks (resp., q-superbricks) of a graph G form a partition
of V (G), called the q-brick-partition (resp., q-superbrick-partition) of G.
For a partition P of V and a rational number q, defG,q (P) = q(|P| − 1) − eG (P) denotes the
deficiency of P in G (with respect to q). Furthermore, let defq (G) = max{defG,q (P) : P is a partition
of V }. A partition P is q-tight if defG,q (P) = defq (G). Jackson and Jordán [18] characterized the
brick- and superbrick-partitions of a graph as follows.

Theorem 20.23 (Jackson and Jordán [18]) Let G = (V, E) be a graph and let q be a rational num-
ber. Let P and Q be q-tight partitions of V such that |P| is as small as possible and |Q| is as large
as possible. Then P is the q-brick-partition of G and Q is the q-superbrick-partition of G.

20.5.3 Constructive Characterizations

Tay [39] used the following constructive characterization of k-tree-connected graphs to prove The-
orem 20.1.

Theorem 20.24 (Tay [39]) An undirected graph G = (V, E) is k-tree-connected if and only if G can
be built up from a single vertex graph by the following operations:
(i) add a new edge,
(ii) pinch i (0 ≤ i ≤ k −1) existing edges with a new vertex v, and add k −i new edges connecting
v with existing vertices.

A constructive characterization of highly k-tree-connected graphs was given by Frank and

Szegő [10], which was used by Connelly, Jordán, and Whiteley [4] to prove Theorem 20.14.

Theorem 20.25 (Frank and Szegő [10]) An undirected graph G = (V, E) is highly k-tree-
connected if and only if G can be built up from a single vertex graph by the following operations:
(i) add a new edge (that can be a loop),
(ii) pinch i (1 ≤ i ≤ k −1) existing edges with a new vertex v, and add k −i new edges connecting
v with existing vertices.

In some applications in rigidity, graphs are required to be simple. A constructive characterization

of simple k-tree-connected graphs is known for k = 2 by Nixon, Owen and Power [33], but the
general case is still open.

20.5.4 Algorithms
We collect here some algorithmic results that are useful in the investigation of body-bar-hinge
frameworks.
456 Handbook of Geometric Constraint Systems Principles

Algorithms for testing k-tree-connectivity. There are several algorithms for deciding whether
a graph is k-tree-connected. The first group of these algorithms is based on the fact that k-tree-
connectivity can be decided by computing the rank of the (k, k)-sparsity matroid. This matroid is the
union of k copies of the graphic matroid, which immediately provides an algorithm to calculate the
rank function (see Gabow and Westerman [12]). There is also an algorithm to calculate the rank in
any (k, `)-sparsity matroid based on degree-constrained orientations, widely called the pebble game
algorithm [30]. This algorithm is a bit less efficient but is much easier to implement and extend to
further applications. For example, it can be used to output the maximal (k, k)-tight subgraphs of a
graph, that is, the k-brick partition. We refer to [30, 43] for more details.
Another way for deciding k-tree-connectivity is due to Frank [7], and is based on rooted (k, 0)-
arc-connected orientations (see Theorem 20.20). A further efficient implementation of Frank’s al-
gorithm was given by Király [27].

Algorithms for testing high k-tree-connectivity. By exploiting the (k, k)-sparsity matroid, one
can also design an efficient algorithm for testing high k-tree-connectivity. Again, the pebble game
algorithm can be modified to determine the connected components of the (k, k)-sparsity matroid of
G (see [30, 43]), which gives the k-superbrick partition of the graph. Hence this also provides an
algorithm to compute the q-superbrick partition for every rational q.
The original proof of Frank’s theorem in [8] is algorithmic, and hence it provides an algorithm
for testing (k, `)-tree-connectivity by using orientations. This algorithm is more involved than that
for k-tree-connectivity. For ` = 1, Király [27] gave a simplified algorithm. The main observation in
[27] is that if we take an arbitrary rooted k-arc-connected orientation and take the set of vertices
from which the root is reachable in a one-way path, then these vertices induce a rooted (k, 1)-arc-
connected sub-digraph. Hence the algorithm is the following. Take an r0 v edge and orient it towards
r0 and take a rooted k-arc-connected orientation of G − r0 v. Take the set R (containing r0 and v)
of vertices from which r0 is reachable by this orientation, shrink it to a new root and restart the
procedure with this graph.

Augmenting to a (highly) k-tree-connected graph. Augmenting a given framework to a rigid

or globally rigid framework with minimum cost is a natural engineering problem, and due to the
combinatorial characterizations we may focus on augmenting a graph to a (highly) k-tree-connected
graph when the framework is generic. By the matroid structure behind, augmenting a graph to a
k-tree-connected graph can be solved easily. Moreover, we can also handle the weighted version of
this problem. However, the augmentation to high tree-connectivity is only solved for the unweighted
case.
Frank and T. Király [9] gave a polynomial algorithm to augment a graph to a (k, `)-tree-
connected graph with minimum number of edges. A rather simple algorithm can be given for the
important subcase when the input is k-tree-connected and ` = 1 by an algorithm of Király [28].
This algorithm uses several results and algorithms mentioned in this section. It first takes the k-
superbrick partition and observes that shrinking the members of this partition to single vertices
results in a (k, k)-tight graph. To augment this latter graph, we need to cover its edge-set by minimal
(k, k)-tight subgraphs by Theorem 20.22. This is done with a two phase greedy algorithm that uses
the pebble game algorithm as a subroutine.

How to decide whether a bar-joint realization of graph is a body-bar-hinge framework. In

order to use the above algorithms to decide the rigidity or global rigidity of some specific generic
bar-joint frameworks, one needs to discover its body-bar-hinge structure. In [23] a simple algorithm
for testing whether a graph is a body-hinge graph is given. This algorithm can be modified for testing
whether a graph has a body-bar-hinge or some other related structure (e.g. a skeleton).
References 457

References
[1] A.R. Berg and T. Jordán. Algorithms for graph rigidity and scene analysis. In U. Zwick
G. Di Battista, editor, Proc. 11th Annual European Symposium on Algorithms (ESA), Springer
Lecture Notes in Computer Science 2832, pages 78–89, 2003.
[2] C. Borcea, I. Streinu, and S. Tanigawa. Periodic body-and-bar frameworks. SIAM J. Discrete
Math., 29(93–112), 2015.
[3] R. Connelly. On generic global rigidity. In P. Gritzmann and B. Sturmfels, editors, Applied
geometry and discrete mathematics, volume 4 of DIMACS Ser. Discrete Math. Theoret. Com-
put. Sci, pages 147–155. AMS, 1991.
[4] R. Connelly, T. Jordán, and W. Whiteley. Generic global rigidity of body-bar frameworks.
Journal of Combinatorial Theory Series B, 103:689–705, 2013.
[5] R. Connelly and W. Whiteley. Global rigidity: The effect of coning. Discrete & Computa-
tional Geometry, 43(4):717–735, 2010.
[6] H. Crapo and W. Whiteley. Statics of frameworks and motions of panel structures, a projective
geometric introduction. Structural Topology, 6(43–82), 1982.
[7] A. Frank. On disjoint trees and arborescences. In Algebraic Methods in Graph Theory, 25,
pages 59–169. Colloquia Mathematica Soc. J. Bolyai, Norh-Holland, 1978.
[8] A. Frank. On the orientation of graphs. J. Comb. Theory, Ser. B, 28(3):251–261, 1980.
[9] A. Frank and T. Király. Combined connectivity augmentation and orientation problems. Dis-
crete Appl. Math., 131(2):401–419, 2003.
[10] A. Frank and L. Szegő. Constructive characterizations for packing and covering with trees.
Discrete Applied Mathematics, 131(2):347–371, 2003.
[11] S. Frank and J. Jiang. New classes of counterexamples to Hendrickson’s global rigidity con-
jecture. Discrete & Computational Geometry, 45:574–591, 2011.
[12] H.N. Gabow and H.H. Westermann. Forests, frames, and games: Algorithms for matroid sums
and applications. Algorithmica, 7(5&6):465–497, 1992.
[13] S.D. Guest, B. Schulze, and W. Whiteley. When is a symmetric body-bar structure isostatic?
Internat. J. Solids Structures, 47:2745–2754, 2010.
[14] K. Haller, A. Lee-St. John, M. Sitharam, I. Streinu, and N. White. Body-and-cad geometric
constraint systems. Comput. Geom. Theory Appl., (45):385–405, 2012.
[15] B. Hendrickson. Conditions for unique graph realizations. SIAM J. Comput., 21(1):65–84,
1992.
[16] B. Jackson and T. Jordán. Connected rigidity matroids and unique realizations of graphs. J.
Comb. Theory, Ser. B, 94:1–29, 2005.
[17] B. Jackson and T. Jordán. The Dress conjectures on rank in the 3-dimensional rigidity matroid.
Advances in Applied Mathematics, 35(4):355–367, 2005.
[18] B. Jackson and T. Jordán. Brick partitions of graphs. Discrete Mathematics, 310(2):270–275,
2008.
[19] B. Jackson and T. Jordán. On the rigidity of molecular graphs. Combinatorica, 28(6):645–
658, 2008.
458 References

[20] B. Jackson and T. Jordán. Pin-collinear body-and-pin frameworks and the molecular conjec-
ture. Discrete and Computational Geometry, 40(2):258–278, 2008.
[21] B. Jackson and T. Jordán. The generic rank of body–bar-and-hinge frameworks. European
Journal of Combinatorics, 31(2):574–588, 2009.
[22] B. Jackson and V.H. Nguyen. Graded sparse graphs and body-length-direction frameworks.
European Journal of Combinatorics, 46:51–67, 2015.
[23] T. Jordán, Cs. Király, and S. Tanigawa. Generic global rigidity of body-hinge frameworks.
Journal of Combinatorial Theory, Series B, 117:59–76, 2016.
[24] T. Jordán, Cs. Király, and S. Tanigawa. Generic global rigidity of body-hinge frameworks:
a correction. Technical Report TR-2017-11, Egerváry Research Group, Budapest, 2017.
www.cs.elte.hu/egres.
[25] N. Katoh and S. Tanigawa. A proof of the Molecular conjecture. Discrete & Computational
Geometry, 45:647–700, 2011.
[26] N. Katoh and S. Tanigawa. Rooted-tree decompositions with matroid constraints and the
infinitesimal rigidity of frameworks with boundaries. SIAM Journal on Discrete Mathematics,
27:155–185, 2013.
[27] Cs. Király. Algorithms for finding a rooted (k, 1)-edge-connected orientation. Discrete Appl.
Math., 166:263–268, March 2014.
[28] Cs. Király. Rigid graphs and an augmentation problem. Technical Report TR-2015-03,
Egerváry Research Group, Budapest, 2015. www.cs.elte.hu/egres.
[29] Y. Kobayashi, Y. Higashikawa, N. Katoh, and A. Sljoka. Characterizing redundant rigidity and
redundant global rigidity of body-hinge graphs. Information Processing Letters, 116(2):175
– 178, 2016.
[30] A. Lee and I. Streinu. Pebble game algorithms and sparse graphs. Discrete Mathematics,
308(8):1425–37, 2008.
[31] C.St.J.A. Nash-Williams. Edge-disjoint spanning trees of finite graphs. J. London Math. Soc.,
36:445–450, 1961.
[32] C.St.J.A. Nash-Williams. Decomposition of finite graphs into forests. Journal of the London
Mathematical Society, 1(1):12, 1964.
[33] A. Nixon, J.C. Owen, and S.C. Power. Rigidity of frameworks supported on surfaces. SIAM
J. Discrete Math., 26(4):1733–1757, 2012.
[34] E. Ross. The rigidity of periodic body-bar frameworks on the three-dimensional fixed torus.
Phil. Trans. Royal Soc. A, 372:2008, 2014.
[35] B. Schulze and S. Tanigawa. Linking rigid bodies symmetrically. European Journal of Com-
binatorics, 42:145–166, 2014.
[36] S. Tanigawa. Generic rigidity matroids with Dilworth truncations. SIAM Journal on Discrete
Mathematics, 26:1412–1439, 2012.
[37] S. Tanigawa. Matroids of gain graphs in applied discrete geometry. Tran. Amer. Math. Soc.,
367:8597–8641, 2015.
[38] S. Tanigawa. Sufficient conditions for globally rigidity of graphs. Journal of Combinatorial
Theory Series B, 113:123–140, 2015.
References 459

[39] T.S. Tay. Rigidity of multi-graphs. I: Linking rigid bodies in n-space. Journal of Combinato-
rial Theory. Series B, 36(1):95–112, 1984.
[40] T.S. Tay. Linking (n − 2)-dimensional panels in n-space II:(n − 2, 2)-frameworks and body
and hinge structures. Graphs and Combinatorics, 5(1):245–273, 1989.
[41] T.S. Tay. Linking (n − 2)-dimensional panels in n-space I:(k − 1, k)-graphs and (k − 1, k)-
frames. Graphs and Combinatorics, 7(3):289–304, 1991.
[42] T.S. Tay and W. Whiteley. Recent advances in the generic rigidity of structures. Structural
Topology, 9:31–38, 1984.
[43] L. Theran. Sparsity matroids for combinatorial rigidity. in this handbook.
[44] M.F. Thorpe, M. Chubynsky, B. Hespenheide, S. Menor, D.J. Jacobs, L.A. Kuhn, M.I. Zavod-
szky, M. Lei, A.J. Rader, and W. Whiteley. Flexibility in biomolecules. In Current topics in
physics, pages 97–112 (Chapter 6). Imperial College Press, 2005.
[45] W.T. Tutte. On the problem of decomposing a graph into n connected factors. Journal of the
London Mathematical Society, 36:221–230, 1961.
[46] N. White and W. Whiteley. The algebraic geometry of motions of bar-and-body frameworks.
SIAM J. Alg. Disc. Math., 8:1–32, 1987.
[47] W. Whiteley. The union of matroids and the rigidity of frameworks. SIAM Journal on Discrete
Mathematics, 1(2):237–255, 1988.
[48] W. Whiteley. Rigidity of molecular structures: geometric and generic analysis. In M.F. Thorpe
and P.M. Duxbury, editors, Rigidity theory and applications, pages 21–46. Kluwer, 1999.
[49] W. Whiteley. Counting out to the flexibility of molecules. Physical Biology, 2:S116–S126,
2005.
Chapter 21
Global Rigidity of Two-Dimensional
Frameworks
Bill Jackson
School of Mathematical Sciences, Queen Mary University of London, Mile End Road, London,
England

Tibor Jordán
Department of Operations Research, Eötvös University, Pázmány Péter sétány 1/C, Budapest, Hun-
gary, and MTA-ELTE Egerváry Research Group on Combinatorial Optimization, Budapest, Hun-
gary

Shin-Ichi Tanigawa
Department of Mathematical Informatics, The University of Tokyo, Hongo, Bunkyo-ku, Tokyo, Japan

CONTENTS
21.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
21.2 Conditions for Global Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
21.2.1 Stress Matrix Characterization in Rd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
21.2.2 Hendrickson’s Necessary Conditions for Global Rigidity . . . . . . . . . . . . . . . . . . . 464
21.3 Graph Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
21.4 Characterization of Global Rigidity in R1 and R2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
21.5 The Rigidity Matroid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
21.5.1 Rd -Independent Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
21.5.2 Rd -Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
21.5.3 Rd -Connected Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
21.6 Special Families of Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
21.6.1 Highly Connected Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
21.6.2 Vertex-Redundantly Rigid Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
21.6.3 Vertex Transitive Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
21.6.4 Graphs of Large Minimum Degree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
21.6.5 Random Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
21.6.6 Unit Disk Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
21.6.7 Squares of Gaphs, Line Graphs, and Zeolites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
21.7 Related Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474
21.7.1 Globally Linked Pairs of Vertices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474
21.7.2 Globally Rigid Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
21.7.3 Globally Loose Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
21.7.4 Uniquely Localizable Vertices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 477
21.7.5 The Number of Non-Equivalent Realizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 477
21.7.6 Stability Lemma and Neighborhood Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 478

461
462 Handbook of Geometric Constraint Systems Principles

21.8 Direction Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479

21.8.1 Parallel Drawings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 480
21.8.2 Direction-Length Global Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481
21.9 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482
21.10 Optimization Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482
Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483

21.1 Introduction
Framework: A d-dimensional framework is a pair (G, p), where G = (V, E) is a graph and p is a
map from V to Rd . We also call (G, p) a realization of G in Rd .
Generic framework: A framework (G, p) is generic if the set of coordinates of the points p(v),
v ∈ V (G), is algebraically independent over the rationals.
Equivalent frameworks: Two d-dimensional frameworks (G, p) and (G, q) are equivalent if
||p(u) − p(v)|| = ||q(u) − q(v)|| holds for all pairs u, v with uv ∈ E, where ||.|| denotes the
Euclidean norm in Rd .
Congruent frameworks: Two d-dimensional frameworks (G, p) and (G, q) are congruent if
||p(u) − p(v)|| = ||q(u) − q(v)|| holds for all pairs u, v with u, v ∈ V . This is the same as saying
that (G, q) can be obtained from (G, p) by an isometry of Rd .
Globally rigid: A d-dimensional framework (G, p) is globally rigid if every d-dimensional frame-
work which is equivalent to (G, p) is congruent to (G, p). The graph G is globally rigid in Rd
if every generic realization of G in Rd is globally rigid. See Figure 21.1.
Rigid: A d-dimensional framework (G, p) is rigid if there exists an ε > 0 such that, if a d-
dimensional framework (G, q) is equivalent to (G, p) and ||p(u) − q(u)|| < ε for all v ∈ V ,
then (G, q) is congruent to (G, p). The graph G is rigid in Rd if every generic realization of G
in Rd is rigid.

It is a hard problem to decide if a given framework is rigid or globally rigid. Indeed Saxe [46]
showed that it is NP-hard to decide if even a 1-dimensional framework is globally rigid and Abbot
[1] showed that the rigidity problem is NP-hard for 2-dimensional frameworks.
The analysis and characterization of rigid and globally rigid frameworks become more tractable,
however, if we consider generic frameworks. The main reason for this is that rigidity and global
rigidity of frameworks in Rd are both generic properties in the sense that they depend only on the

(a) (b) (c)

Figure 21.1
Frameworks in the plane: (a) globally rigid, (b) rigid but not globally rigid, (c) not rigid.
Global Rigidity of Two-Dimensional Frameworks 463

graph G and not the particular realization p, when (G, p) is generic. This follows from results of
Asimow and Roth [2] and Gortler, Healy and Thurston [23], respectively. The problems of finding
a polynomially verifiable characterization for graphs which are rigid or globally rigid in Rd have
been solved for d = 1, 2, but are major open problems for d ≥ 3.
Global rigidity has several real life applications. One such application occurs in the network
localization problem, in which the locations of some nodes (called anchors) of a network as well
as the distances between some pairs of nodes are known, and the goal is to determine the location
of all nodes. This is one of the fundamental algorithmic problems in the theory of wireless sensor
networks and has been the focus of a number of research articles and survey papers, see for example
[3, 19, 48]. When constructing a sensor network, we would like the solution to the localization
problem to be unique, in this case we say that the network is uniquely localizable. It is not difficult
to see that the unique localizability of a generic network is equivalent to the global rigidity of its
underlying distance graph (in which two nodes are adjacent if and only if the distance between them
is known). We will return to this problem in Sections 21.7 and 21.10.
We will mostly be concerned with global rigidity in R2 but will state results for Rd when they
hold in all dimensions. Some of the material in this chapter is discussed in more detail in the survey
by Jordán [37].

21.2 Conditions for Global Rigidity

Equilibrium stress: An equilibrium stress for a framework (G, p) in Rd is an assignment ω : E →
R such that, for each vertex vi ∈ V :
X
ωi, j (p(vi ) − p(v j )) = 0. (21.1)
j:vi v j ∈E

Stress matrix: The stress matrix Ω associated to an equilibrium stress ω is the |V | × |V | symmetric
matrix in which the entries are defined so that Ω[i, j] = −ωi, j for all edges vi v j ∈ E, Ω[i, j] = 0
for all non-adjacent vertex pairs vi , v j ∈ V , and Ω[i, i] is chosen so that each row and column
sum is equal to zero.
Redundantly rigid graph: A graph G is redundantly rigid in Rd if G has at least two vertices and
G − e is rigid in Rd for all e ∈ E(G).
k-connected graph: A graph G is k-connected for some positive integer k if |V (G)| ≥ k + 1 and
G − X is connected for all X ⊂ V (G) with |X| ≤ k − 1.

21.2.1 Stress Matrix Characterization in Rd

Connelly [13] built a theoretical foundation for the analysis of global rigidity based on stress ma-
trices. He observed that the rank of any stress matrix for (G, p) is at most |V | − d − 1 when p(V )
affinely spans Rd . Subsequently, Connelly [15] (sufficiency) and Gortler, Healy and Thurston [23]
(necessity) showed that the property of having a stress matrix with this maximum possible rank
characterizes global rigidity for generic frameworks.

Theorem 21.1 Let (G, p) be a generic framework in Rd on at least d + 2 vertices. Then (G, p)
is globally rigid in Rd if and only if (G, p) has an equilibrium stress ω for which the rank of the
associated stress matrix Ω is |V | − d − 1.
464 Handbook of Geometric Constraint Systems Principles

We will say that (G, p) has a maximum rank stress matrix when its rank is |V | − d − 1.
We can use Theorem 21.1 to deduce that global rigidity in Rd is a generic property. It also gives
rise to a randomized algorithm for deciding if a given graph is globally rigid in Rd , but it does not
give a deterministic polynomial algorithm or a polynomially verifiable characterization for global
rigidity. See Chapter 16 for more details.

21.2.2 Hendrickson’s Necessary Conditions for Global Rigidity

Hendrickson [24] showed that redundant rigidity and (d + 1)-connectivity are both necessary con-
ditions for a graph to be globally rigid in Rd .
Theorem 21.2 Let G be a globally rigid graph in Rd . Then either G is a complete graph on at most
d + 1 vertices, or G is (d + 1)-connected and redundantly rigid in Rd .
These necessary conditions are also sufficient to imply the global rigidity of G in Rd when
d = 1, 2, as we shall see below. This is not the case, however, when d ≥ 3. We say that a graph
G is a Hendrickson graph in Rd if it satisfies the necessary conditions for global rigidity given in
Theorem 21.2, but it is not globally rigid in Rd . For d = 3, Connelly [14] showed that the complete
bipartite graph K5,5 is a Hendrickson graph. He also constructed similar examples (specific complete
bipartite graphs on d+2


2 vertices) for all d ≥ 3. Frank and Jiang [21] found two more (bipartite)
Hendrickson graphs in R4 as well as infinite families in Rd for d ≥ 5. Jordán, Király, and Tanigawa
[38] constructed infinite families of Hendrickson graphs for all d ≥ 3. Further examples can be
obtained by using the observation that the cone of a d-dimensional Hendrickson graph is a (d + 1)-
dimensional Hendrickson graph.

21.3 Graph Operations

1-extension operation: Given a graph H with a designated edge e = v1 v2 , and d − 1 additional
vertices v3 , . . . , vd+1 , the d-dimensional 1-extension operation on e (sometimes called an edge
d-split) adds a new vertex v0 , removes e, and inserts d + 1 new edges v0 v1 , v0 v2 , ..., v0 vd+1 . See
Figure 21.2.
Cone operation: Given a graph H, the cone of H is the graph obtained by adding a new vertex v
and joining v to every vertex of H.
Vertex splitting operation: Given a graph G = (V, E), a vertex v ∈ V , and a tripartition F0 , F1 , F2
of the edges incident to v with |F0 | = d − 1, the d-dimensional vertex splitting operation at
v replaces the vertex v by two new vertices v1 and v2 and an edge v1 v2 , replaces each edge
uv ∈ F0 by two new edges uv1 , uv2 , and replaces each edge wv ∈ Fi by an edge wvi , i = 1, 2, see
Figure 21.3. The vertex splitting operation is said to be non-trivial if F1 , F2 are both non-empty,
or equivalently, if each of the split vertices v1 , v2 has degree at least d + 1.
Inductive constructions for graphs which use global rigidity preserving operations are frequently
used, both to prove that certain families of graphs are globally rigid and to analyze the global rigidity
of a particular graph.
It is easy to see that adding an edge or a vertex of degree d + 1 to a graph will preserve global
rigidity in Rd . By applying the degree d + 1 vertex addition operation recursively to Kd+1 we obtain
the family of (d + 1)-lateration graphs, which are examples of sparse globally rigid graphs in Rd .
The 1-extension operation is a finer and more useful graph operation that preserves global rigid-
ity.
Global Rigidity of Two-Dimensional Frameworks 465
v0

v1 v1
v2 v2

v3 v3

Figure 21.2
A 2-dimensional 1-extension.

u u
v v1
v2

Figure 21.3
2-dimensional vertex splitting.

Theorem 21.3 [15] Suppose that G can be obtained from Kd+2 by a sequence of 1-extensions and
edge additions. Then G is globally rigid in Rd .

Figure 21.4 shows that every wheel can be constructed from K4 by a sequence of 2-dimensional
1-extensions and hence is globally rigid in R2 .
Connelly proved Theorem 21.3 by showing that Kd+2 has an equilibrium stress with a maximum
rank stress matrix and that this property is preserved by the 1-extension operation. He then used
(the sufficiency part of) Theorem 21.1 to deduce that G is globally rigid. Since the necessity part of
Theorem 21.1 is now known to be true, we can deduce the following stronger result using Connelly’s
result on 1-extensions.

Theorem 21.4 Suppose G is obtained from a graph H by the d-dimensional 1-extension operation.
If H is globally rigid in Rd and |V (H)| ≥ d + 2 then G is globally rigid in Rd .

A direct proof of Theorem 21.4 for the case when d = 2 is given in [31]. This proof is extended
to all d in [49].
Connelly and Whiteley used Theorem 21.1 to show that the cone operation transfers global
rigidity from Rd to Rd+1 .

Theorem 21.5 [16] Suppose that G is the cone of H. Then G is globally rigid in Rd+1 if and only
if H is globally rigid in Rd .

We have seen that wheeels are globally rigid in R2 . Since a wheel on n + 1 vertices is the cone
of a cycle on n vertices, Theorem 21.5 implies that cycles are globally rigid in R1 . (This observation
could also be deduced directly from Theorem 21.4.)
Cheung and Whiteley [10] conjecture that the non-trivial d-dimensional vertex splitting op-
eration preserves generic global rigidity in Rd for all d ≥ 1. Jordán and Szabadka [39] used the
characterization of generic global rigidity in R2 , see Theorem 21.10 below, to verify this conjecture
when d = 2.

Theorem 21.6 [39] Suppose G is obtained from H by a non-trivial 2-dimensional vertex splitting
operation. If H is globally rigid in R2 then G is also globally rigid in R2 .
466 Handbook of Geometric Constraint Systems Principles

Figure 21.4
A sequence of 2-dimensional 1-extensions.

Two other graph operations which are sometimes used are the X-replacement and diamond split
operations, see Figure 21.5. In general an X-replacement may not preserve 3-connectivity. It was
pointed out in [39] that a diamond split may not preserve redundant rigidity in R2 . Hence neither
operation is guaranteed to preserve global rigidity in R2 .

(a) (b)
Figure 21.5
(a) X-replacement, and (b) diamond split in R2 .

21.4 Characterization of Global Rigidity in R1 and R2

Generic global rigidity has been completely characterized in R1 and R2 . The 1-dimensional result
is folklore.

Theorem 21.7 A graph is globally rigid in R1 if and only if it is a complete graph on at most two
vertices or is 2-connected.

Note that Hendrickson’s condition of redundant rigidity is implied by 2-connectivity since

connected graphs are rigid in R1 . One way to prove Theorem 21.7 is to use the well-known
ear-decompositions of 2-connected graphs to obtain the following inductive construction for 2-
connected graphs and then apply Theorem 21.3.

Theorem 21.8 A graph is 2-connected if and only if it can be obtained from K3 by a sequence of
(1-dimensional) 1-extensions and edge additions.

A similar proof technique works for global rigidity in R2 . A key step is the following inductive
construction for 3-connected redundantly rigid graphs.

Theorem 21.9 [26] Every 3-connected graph which is redundantly rigid in R2 can be obtained
from K4 by a sequence of (2-dimensional) 1-extensions and edge additions.
Global Rigidity of Two-Dimensional Frameworks 467

The 2-dimensional rigidity matroid plays a pivotal role in extending the proof technique of
Theorem 21.8 to obtain Theorem 21.9. More details will be given in Section 21.5 below.
Theorems 21.2, 21.3 and 21.9 combine to give the following characterization of globally rigid
graphs in R2 .

Theorem 21.10 [26] A graph G is globally rigid in R2 if and only if G is a complete graph on at
most three vertices or G is 3-connected and redundantly rigid.

Theorem 21.9 gives rise to an efficient deterministic algorithm for testing whether a graph is
globally rigid in R2 . More details will be given in Section 21.9.

21.5 The Rigidity Matroid

Rigidity matrix: The rigidity matrix R(G, p) of a d-dimensional framework (G, p) is a matrix of
size |E| × d|V |, where, for each edge e = vi v j ∈ E, in the row corresponding to e, the entries in
the d columns corresponding to vertices vi and v j contain the d coordinates of (p(vi ) − p(v j ))
and (p(v j ) − p(vi )), respectively, and the remaining entries are zeros.
Rigidity matroid: The rigidity matrix of (G, p) defines the rigidity matroid of (G, p) on the ground
set E, where a set of edges F ⊆ E is independent if and only if the rows of the rigidity matrix
indexed by F are linearly independent. (For more details on matroids and related combinatorial
results the reader is referred to [20, 45].) Since the entries of the rigidity matrix are polynomial
functions of the coordinates of the points p(vi ) with integer coefficients, any two generic d-
dimensional frameworks (G, p) and (G, q) have the same rigidity matroid. We call this the
d-dimensional rigidity matroid of the graph G. We denote the rigidity matroid of G by Rd (G)
and its rank by rd (G).

Rd -independent graph: A graph G = (V, E) is Rd -independent if E is independent in Rd (G).

Rd -circuit: A subgraph H = (W,C) of a graph G is an Rd -circuit of G if C is a circuit (i.e., a
minimal dependent set) in Rd (G). In particular, G is an Rd -circuit if E is a circuit in Rd (G).
2-sum operation: Given two graphs H1 = (V1 , E1 ) and H2 = (V2 , E2 ) with V1 ∩ V2 = ∅ and two
designated edges u1 v1 ∈ E1 and u2 v2 ∈ E2 , the 2-sum of H1 and H2 (along the edge pair u1 v1 ,
u2 v2 ) is the graph H1 ⊕2 H2 obtained from H1 − u1 v1 and H2 − u2 v2 by identifying u1 with u2
and v1 with v2 . See Figure 21.6.
Components of a matroid : Given a matroid M = (E, I), we define a relation on E by saying
that e, f ∈ E are related if e = f or if there is a circuit C in M with e, f ∈ C. It is well-known
that this is an equivalence relation. The equivalence classes are called the components of M.
If M has at least two elements and only one component then M is said to be connected. If
M has components E1 , E2 , . . . , Et and Mi is the matroid restriction of M onto Ei then M =
M1 ⊕ M2 . . . ⊕ Mt , where ⊕ denotes the direct sum of matroids, see [45].
Rd -components of a graph: A graph G = (V, E) is Rd -connected if its d-dimensional rigidity ma-
troid Rd (G) is connected. The Rd -components of G are the subgraphs of G induced by the
components of Rd (G), see Figure 21.8. We say that an Rd -component is trivial if it has only
one edge, that is, if it is induced by an edge which belongs to no Rd -circuit. Such an edge is
also called an Rd -bridge.
468 Handbook of Geometric Constraint Systems Principles

u1 u2

H1 H2

v1 v2

Figure 21.6
The 2-sum operation.

Figure 21.7
An R2 -circuit that is not 3-connected. This graph can be constructed by a 2-sum of two copies of
K4 .

21.5.1 Rd -Independent Graphs

It is not difficult to see that a graph G is R1 -independent if and only if it is a forest and hence R1 (G)
is the well known cycle matroid of G. A fundamental theorem of Laman [41] characterizes R2 -
independent graphs. It remains an open problem to find good characterizations for Rd -independence
or, more generally, the rank function of Rd (G) when d ≥ 3.

21.5.2 Rd -Circuits
Wheels, K3,3 plus an edge, and K3,4 are all examples of 3-connected R2 -circuits. An example of an
R2 -circuit that is not 3-connected is given in Figure 21.7. Laman’s characterization of independence
in R2 implies that R2 -circuits are redundantly rigid in R2 . It follows that a graph G is redundantly
rigid in R2 if and only if G is rigid in R2 and each edge of G belongs to an R2 -circuit of G. (This is
not true in higher dimensions, since Rd -circuits may not be rigid in Rd when d ≥ 3.)
Berg and Jordán [5] obtained the following inductive construction for R2 -circuits.

Theorem 21.11 A graph G is an R2 -circuit if and only if G is a connected graph obtained from
disjoint copies of K4 ’s by taking 2-sums and 2-dimensional 1-extensions.

They then used Theorem 21.11 to obtain a simpler construction for 3-connected R2 -circuits.

Theorem 21.12 [5] A graph G is a 3-connected R2 -circuit if and only if G can be constructed from
K4 by a sequence of 2-dimensional 1-extensions.

Theorem 21.12 is an important step in the characterization of generic global rigidity in R2 .

Together with Theorem 21.3, it immediately implies that 3-connected R2 -circuits are globally rigid.

21.5.3 Rd -Connected Graphs

By definition the Rd -components of a graph G are pairwise edge-disjoint subgraphs of G and that
the rigidity matroid Rd (G) can be expressed as the direct sum of the rigidity matroids of its Rd -
components.
Global Rigidity of Two-Dimensional Frameworks 469

Figure 21.8
A graph that is redundantly rigid in R2 but it not R2 -connected. It has three R2 -components given
by the three copies of K4 .

When d = 2, we can also define the rigid components of G to be the maximal subgraphs which
are rigid in R2 , and the redundantly rigid components to be the maximal subgraphs which are
redundantly rigid in R2 , together with the subgraphs which are induced by an R2 -bridge. We refer
to the second type of redundantly rigid component as trivial. Thus the trivial R2 -components and
the trivial redundantly rigid components are the same. Since the non-trivial R2 -components of G
are redundantly rigid, the R2 -components are pairwise edge-disjoint vertex-induced subgraphs of
G and the partition of E(G) given by the R2 -components is a refinement of the partition given
by the redundantly rigid components and hence a further refinement of the partition given by the
rigid components. In addition, the rigidity matroid R2 (G) can be expressed as the direct sum of the
rigidity matroids of either the R2 -components of G, the redundantly rigid components of G, or the
rigid components of G.
We say that a graph G is nearly 3-connected if G can be made 3-connected by adding at most
one new edge.

Theorem 21.13 [26] If G is nearly 3-connected and every edge of G is in some R2 -circuit, then G
is R2 -connected.

As a corollary we obtain:

Theorem 21.14 [26] If G is a 3-connected graph and is redundantly rigid in R2 , then G is R2 -

connected.

Figure 21.8 shows an example of a graph that is redundantly rigid in R2 but is not R2 -connected.

The main result of this subsection is the following inductive construction. Its proof is inductive
and uses Theorem 21.12 as a base case. The inductive step uses an ear-decomposition of the rigidity
matroid of an R2 -connected graph, in a similar way that an ear decomposition of a 2-connected
graph can be used to prove Theorem 21.8.

Theorem 21.15 [26] A graph is 3-connected and R2 -connected if and only if it can be obtained
from K4 by a sequence of (2-dimensional) 1-extensions and edge additions.

Theorem 21.9 follows from Theorems 21.14 and 21.15.

An example of this inductive construction is given in Figure 21.9.
By using the above results one can also obtain an inductive construction for R2 -connected
graphs using 1-extensions, edge-additions and 2-sums [26], cf. Theorem 21.11.
470 Handbook of Geometric Constraint Systems Principles

Figure 21.9
A sequence of 1-extensions and edge additions.

21.6 Special Families of Graphs

Mixed connectivity: In a graph G = (V, E), a pair (U, D) with U ⊆ V and D ⊆ E is a mixed cut if
G − U − D is not connected. The graph G is 6-mixed-connected if |V | ≥ 4 and 2|U| + |D| ≥ 6
for all mixed cuts (U, D) in G.
Cyclic edge-connectivity: A graph G = (V, E) is cyclically k-edge-connected if, for all X ⊆ V such
that G[X] and G[V − X] both contain cycles, we have at least k edges from X to V − X.
Vertex redundant rigidity: A graph G is vertex-redundantly rigid in Rd if G − v is rigid in Rd for
all v ∈ V (G).
k-rigid graph: A graph G is k-rigid in Rd if G −U is rigid in Rd for all U ⊆ V (G) with |U| ≤ k − 1.
Unit disk framework: A unit disk framework (with sensing radius R) is a framework (G, p) in R2
for which uv ∈ E(G) if and only if ||p(u) − p(v)|| ≤ R.
Unit disk graph: A graph G is a unit disk graph if it can be realized as a unit disk framework in
R2 .
Square of a graph: The square G2 of a graph G is obtained from G by adding a new edge uv for
each pair u, v ∈ V (G) of distance two in G.

21.6.1 Highly Connected Graphs

The proof technique used by Lovász and Yemini [44] to show that 6-connected graphs are rigid in
R2 yields the stronger result that these graphs are in fact redundantly rigid. We can combine this
result with Theorem 21.10 to obtain:

Theorem 21.16 [26] Every 6-connected graph is globally rigid in R2 .

An infinite family of 5-connected non-rigid graphs given in [44] shows that the hypothesis on
vertex connectivity in Theorem 21.16 cannot be reduced from six to five. On the other hand, Jackson,
Servatius, and Servatius showed in [34] that the connectivity hypothesis can be replaced by a slightly
weaker hypothesis of “essential-6-vertex-connectivity” which allows vertex cuts of size four or five
as long as they only separate one or at most three vertices, respectively, from the rest of the graph.
It was shown in [28] that the connectivity hypothesis can be weakened in a more substantial way
and still guarantee the rigidity and global rigidity of a graph.

Theorem 21.17 [28] If G is a 6-mixed-connected graph, then G − e is globally rigid in R2 for all
e ∈ E.
Global Rigidity of Two-Dimensional Frameworks 471

Figure 21.10
A graph that is mixed-6-connected but is not 4-connected.

Note that G is 6-mixed-connected if and only if G is 6-edge-connected, G − v is 4-edge-

connected for all v ∈ V , and G − {u, v} is 2-edge-connected for all u, v ∈ V . It follows that 6-vertex-
connected graphs are 6-mixed-connected and 6-mixed-connected graphs are 3-vertex-connected.
See Figure 21.10 for an example.
The final result of this subsection observes that for 4-regular graphs an even weaker connectivity
condition implies global rigidity.

Theorem 21.18 [34] Every cyclically 5-edge-connected 4-regular graph is globally rigid in R2 .

Examples of 4-regular 4-connected graphs and 5-regular 5-connected graphs which are not glob-
ally rigid are given in Theorem 21.20 (c),(d) below. See also Figure 21.11.

21.6.2 Vertex-Redundantly Rigid Graphs

Tanigawa [50] showed that vertex-redundant rigidity is a sufficient condition for global rigidity in
all dimensions.

Theorem 21.19 Every graph which is vertex-redundantly rigid in Rd is globally rigid in Rd .

Kaszanitzky and Király [40] solved a number of extremal problems related to k-rigidity. One can
also consider higher degrees of redundant rigidity with respect to edge removal as well as similar
notions for global rigidity. It is an open problem to decide whether any of these graph properties can
be tested in polynomial time for d ≥ 2.

21.6.3 Vertex Transitive Graphs

Vertex transitive graphs which are rigid or globally rigid were characterized by Jackson, Servatius,
and Servatius [34].

Theorem 21.20 Let G be a connected k-regular vertex transitive graph on n vertices. Then G is not
globally rigid in R2 if and only if one of the following holds:
(a) k = 2 and n ≥ 4.
(b) k = 3 and n ≥ 6.
(c) k = 4 and G has a 3-factor F consisting of s disjoint copies of K4 where s ≥ 3.
(d) k = 5 and G has a 4-factor F consisting of s disjoint copies of K5 where s ≥ 6.

As a corollary they determine all vertex transitive graphs which are rigid but not globally rigid
in R2 . See Figure 21.11 for an example of such a graph.
472 Handbook of Geometric Constraint Systems Principles

Figure 21.11
A 4-regular 4-connected vertex transitive graph that is rigid but not globally rigid in R2 .

21.6.4 Graphs of Large Minimum Degree

Jackson and Jordán obtained a sufficient condition for global rigidity in terms of minimum degree.
n+1
Theorem 21.21 [27] Let G be a graph on n ≥ 4 vertices with minimum degree at at least 2 . Then
G is globally rigid in R2 .

The graph consisting of two complete graphs of equal size with two vertices in common shows
that the bound on the minimum degree in Theorem 21.21 is best possible.

21.6.5 Random Graphs

We consider three different models of random graphs. Throughout this subsection, we assume that
all logarithms are natural.
Our first model is the Erdős-Rényi model of random graphs. Let G(n, p) denote the probability
space of all graphs on n vertices in which each pair of vertices is joined by an edge with independent
probability p, see [6]. A sequence of graph properties An holds asymptotically almost surely, or a.a.s.
for short, in G(n, p) if limn→∞ PrG(n,p) (An ) = 1.

Theorem 21.22 [34] Let G ∈ G(n, p), where p = (log n+k log log n+w(n))/n, and limn→∞ w(n) =
∞.
(a) If k = 2 then G is a.a.s. rigid in R2 .
(b) If k = 3 then G is a.a.s. globally rigid in R2 .

The bounds on p given in Theorem 21.22 are best possible since if G ∈ G(n, p) and p = (log n +
k log log n + c)/n for any constant c, then G a.a.s. does not have minimum degree at least k, see [6].
Our second model is of random regular graphs. Let Gn,d denote the probability space of all d-
regular graphs on n vertices chosen with the uniform probability distribution. (We refer the reader
to [6] for a mathematical procedure for generating the graphs in Gn,d .) Since globally rigid graphs
on at least four vertices are redundantly rigid, the only globally rigid graphs in G(n, d) for d ≤ 3 are
K2 , K3 , and K4 . The situation changes drastically for d ≥ 4.

Theorem 21.23 [34] If G ∈ Gn,d and d ≥ 4 then G is a.a.s. globally rigid in R2 .

Our third model is of geometric random graphs. Let Geom(n, r) denote the probability space of
all graphs on n vertices in which the vertices are distributed uniformly at random in the unit square
and all pairs of vertices of distance at most r are joined by an edge. Suppose G ∈ Geom(n, r). Li
et al. [43] have shown that if nπr2 = log n + (2k − 3) log log n + w(n) for k ≥ 2 a fixed integer and
limn→∞ w(n) = ∞, then G is a.a.s. k-connected. As noted by Eren et al. [19], this result can be
Global Rigidity of Two-Dimensional Frameworks 473

combined with Theorem 21.16 to deduce that if nπr2 = log n + 9 log log n + w(n) then G is a.a.s.
globally rigid. We do not know if this result is best possible. However, it is also shown in [43] that
if nπr2 = log n + (k − 1) log log n + c for any constant c, then G is a.a.s. not k-connected. Thus, if
nπr2 = log n + 2 log log n + c for any constant c, then G is a.a.s. not 3-connected, and hence is a.a.s.
not globally rigid.

21.6.6 Unit Disk Graphs

The family of unit disk graphs is a natural model for sensor networks in which the distance between
two nodes is known if and only if this distance is at most the sensing radius R of the nodes, (c.f.
Section 21.7.4). It is NP-hard to test whether a graph is a unit disk graph [8], and it is also NP-
hard to test whether a unit disk framework is globally rigid [3]. However, it may be possible to use
the unit disk property of a unit disk framework (G, p) and bounds on the sensing radius to deduce
necessary or sufficient conditions which imply that (G, p) is globally rigid in the sense that it is
a unique realization of G as a unit disk framework, with the given edge lengths. This is a largely
unexplored area of research.

21.6.7 Squares of Gaphs, Line Graphs, and Zeolites

The square G2 of a graph G is used in several applications. For example, if we double the sensing
radius of a unit disk framework (G, p), the underlying graph of the new unit disk framework will
contain G2 . It follows from the next result that this doubling operation makes a generic unit disk
framework, whose underlying graph is 2-edge-connected, globally rigid in R2 .

Theorem 21.24 [10] Let G be a connected graph. Then G2 is globally rigid in R2 if and only if one
endvertex of every bridge of G has degree one.

A d-dimensional body-pin framework is a different structural model consisting of full dimen-

sional rigid bodies and pins. Each pin is a joint that connects two bodies. The bodies are free to
move continuously in Rd subject to the constraint that the relative motion of any two bodies joined
by a pin is a rotation about the pin. The framework is rigid if every such motion preserves the dis-
tances between all pairs of points belonging to different rigid bodies, i.e., the motion extends to an
isometry of Rd . The underlying graph of the framework has the bodies as its vertices and the pins
as its edges. We can obtain an equivalent bar-joint framework to a given body-pin framework by
taking a bar-joint realization of the line graph of the underlying graph of the body-pin framework
(this operation replaces each body by a bar-joint realization of a large complete graph in such a way
that two bodies joined by a pin share a joint.
The special case when d = 2 and the underlying graph is 3-regular gives rise to the (2-
dimensional) combinatorial zeolites. The investigation of these structures is motivated by the in-
terest in the flexibility properties of real zeolites, which are molecules formed by corner-sharing
tetrahedra. Their rigidity can be understood by considering their equivalent bar-joint realizations.

Theorem 21.25 [35] Let G = (V, E) be a 3-regular graph. Then the line graph of G is globally
rigid in R2 if and only if G is 3-edge-connected.

It was further shown in [38] that the underlying graphs of globally rigid generic body-pin frame-
works in R2 can be characterized by the same connectivity condition. A generalization to higher
dimensions can also be found in [38], see Chapter 20.
474 Handbook of Geometric Constraint Systems Principles

21.7 Related Properties

In this section we consider properties of generic frameworks which are weaker than global rigidity.
We will assume that all frameworks are 2-dimensional unless explicitly stated otherwise.
Globally linked pair of vertices: A pair of vertices {u, v} in a framework (G, p) is globally linked
in (G, p) if, in all equivalent frameworks (G, q), we have ||p(u) − p(v)|| = ||q(u) − q(v)||. The
pair {u, v} is globally linked in the graph G if it is globally linked in all generic frameworks
(G, p).
Globally rigid cluster: A globally rigid cluster of G is a maximal subset of V in which all pairs of
vertices are globally linked in G.
Globally loose pair of vertices: A pair of vertices {u, v} in a graph G is globally loose if {u, v} is
not globally linked in all generic realizations of G.
Uniquely localizable vertex: Let (G, p) be a generic framework with a designated set of vertices
P. A vertex v ∈ V (G) is uniquely localizable in (G, p) with respect to P if whenever (G, q) is
equivalent to (G, p) and p(b) = q(b) for all vertices b ∈ P, then we also have p(v) = q(v). We
say that v is uniquely localizable in the graph G with respect to P, if v is uniquely localizable
with respect to P in all generic frameworks (G, p).

21.7.1 Globally Linked Pairs of Vertices

It follows from the above definition that a graph G is globally rigid if and only if all pairs of vertices
of G are globally linked. Unlike global rigidity, however, ‘global linkedness’ is not a generic prop-
erty. Figure 21.12 shows an example of a pair of vertices in a rigid graph G which is globally linked
in one generic realization, but not in another. (Note, however that global linkedness is a generic
property for 1-dimensional frameworks since, in this case, {u, v} is globally linked in G if and only
if G has two openly disjoint uv-paths.)

u u
u

v v v

(a) (b) (c)

Figure 21.12
The pair {u, v} is globally linked in the framework (a), but not in (b) since the distance between u
and v is different in the equivalent framework (c).

No complete characterization of globally linked pairs in graphs is known when d ≥ 2. We

will describe several results and conjectures in the 2-dimensional case. The first shows that the
1-extension operation preserves the property that a pair of vertices is globally linked as long as the
deleted edge is redundant.
Theorem 21.26 [31] Let G, H be graphs such that G is obtained from H by a 1-extension operation
on the edge xy and vertex w. Suppose that H − xy is rigid and that {u, v} is globally linked in H.
Then {u, v} is globally linked in G.
Global Rigidity of Two-Dimensional Frameworks 475

Let G = (V, E) be a graph and x, y ∈ V . We use κG (x, y) to denote the maximum number of
pairwise openly disjoint xy-paths in G. Note that if xy ∈/ E then, by Menger’s theorem, κG (x, y) is
equal to the size of a smallest set S ⊆ V − {x, y} for which there is no xy-path in G − S. It is easy to
see that:
Lemma 21.7.1 [31] Let (G, p) be a generic framework, x, y ∈ V (G), xy ∈
/ E(G), and suppose that
κG (x, y) ≤ 2. Then {x, y} is not globally linked in (G, p).
This observation, together with Theorem 21.26, can be used to characterize globally linked pairs
in R2 -connected graphs.
Theorem 21.27 [31] Let G = (V, E) be an R2 -connected graph and x, y ∈ V . Then {x, y} is globally
linked in G if and only if κG (x, y) ≥ 3.
r r

r ur rv r

r r r r
x w

r r

Figure 21.13
An R2 -circuit G. The pairs (u, v), (v, w), (w, x), (x, u) are globally linked in G since each pair of
vertices is connected by three pairwise openly disjoint paths.

Theorem 21.27 is illustrated in Figure 21.13. It has the following immediate corollary.
Corollary 21.28 [31] Let G = (V, E) be a graph and x, y ∈ V . If either xy ∈ E, or there is an R2 -
component H of G with {x, y} ⊆ V (H) and κH (x, y) ≥ 3, then {x, y} is globally linked in G.
It is conjectured in [31] that the sufficient condition for global linkedness given in Corollary
21.28 is also necessary:
Conjecture 6 The pair {x, y} is globally linked in a graph G = (V, E) if and only if either xy ∈ E
or there is an R2 -component H of G with {x, y} ⊆ V (H) and κH (x, y) ≥ 3.
The following two closely related conjectures are also given in [31]. It is shown that together,
these conjectures are equivalent to Conjecture 6.
Conjecture 7 Suppose that {x, y} is a globally linked pair in a graph G. Then there is a redundantly
rigid component R of G with {x, y} ⊆ V (R).
Conjecture 8 Let G be a graph. Suppose that there is a redundantly rigid component R of G with
{x, y} ⊆ V (R) and {x, y} is globally linked in G. Then {x, y} is globally linked in R.
The fact that Conjecture 6 implies both Conjectures 7 and 8 follows from Theorem 21.13.
It is straightforward to show that the 0-extension operation (which adds a vertex of degree 2 to a
graph) preserves the property that a pair of vertices is not globally linked, see [31]. Our next result
(which is a counterpart to Theorem 21.26) shows that a similar result holds for 1-extension.
476 Handbook of Geometric Constraint Systems Principles

Theorem 21.29 [32] Let H = (V, E) be a rigid graph and let G be a 1-extension of H on some
edge uw ∈ E. Suppose that H − uw is not rigid and that {x, y} is not globally linked in H for some
x, y ∈ V . Then {x, y} is not globally linked in G.

This result was used in [32] to deduce:

Theorem 21.30 Let G = (V, E) be a rigid graph, u, v ∈ V , and R = (U, F) be a redundantly rigid
component of G. Suppose that G − e is not rigid for all e ∈ E − F. Then {u, v} is globally linked in
G if and only if uv ∈ E or {u, v} is globally linked in R.

The special case of Theorem 21.30 when G has no non-trivial redundantly rigid components
characterizes globally linked pairs in minimally rigid graphs.

Theorem 21.31 [32] Let G = (V, E) be a minimally rigid graph and u, v ∈ V . Then {u, v} is globally
linked in G if and only if uv ∈ E.

21.7.2 Globally Rigid Clusters

We can identify globally rigid clusters in G by using the following procedure. We first construct the
R2 -components H of G. For each component H, we construct the augmented graph H + by adding
an edge between all pairs of non-adjacent vertices u, v such that H − {u, v} is disconnected. We then
construct the cleavage units of H by taking the maximal subgraphs of H + which are 3-connected or
are equal to K3 , see [26, Section 3] for more details. By Corollary 21.28, each pair of vertices in a
3-connected cleavage unit of an R2 -component of G is globally linked in G. This can be illustrated
by the graph G in Figure 21.13. It is R2 -connected and has five cleavage units which are equal to
the five K4 -subgraphs in G ∪ {uv, vw, wx, xu}. The vertex sets of these five subgraphs are the globally
rigid clusters of G.
The truth of Conjecture 6 would imply that the vertex sets of the 3-connected cleavage units of
the R2 -components of G (and the copies of K2 or K3 not included in any such cleavage unit) are
precisely the globally rigid clusters of G. Note that the vertices of a globally rigid cluster of G need
not induce a globally rigid subgraph in G. For example, in Figure 21.13 none of the globally rigid
clusters induce a globally rigid subgraph.

21.7.3 Globally Loose Pairs

It follows from Lemma 21.7.1 and Theorem 21.27 that if G is R2 -connected then each pair of
vertices {u, v} is either globally linked or globally loose in G, and that {u, v} is globally loose if and
only if κG (u, v) = 2. On the other hand, there exist graphs G with a pair of vertices {u, v} that are
neither globally linked nor globally loose in G.
Our next result gives a sufficient condition for a pair {u, v} to be globally loose in a graph G.

Theorem 21.32 [31] Let G = (V, E) be a globally rigid graph and e = uv ∈ E. Suppose that G − e
is not globally rigid. Then {u, v} is globally loose in G − e.

A minimally rigid graph G is said to be special if its only rigid proper subgraphs are complete
graphs on at most three vertices. Examples of special graphs are K3,3 and the prism (i.e., the com-
plement graph of a cycle of length six). It can be seen that special graphs are 3-connected. It follows
that, if G is special and uv ∈
/ E(G), then G + uv is a 3-connected R2 -circuit. Thus G + uv is globally
rigid by Theorem 21.10. Since G − uv is not globally rigid, Theorem 21.32 implies:

Theorem 21.33 [31] Let G be a special minimally rigid graph and suppose that u, v ∈ V . Then
{u, v} is globally loose in G if and only if uv ∈
/ E.
Global Rigidity of Two-Dimensional Frameworks 477

The following stronger result was proved in [32]: if G is minimally rigid and G + uv is an R2 -
circuit for two non-adjacent vertices u, v of G, then {u, v} is globally loose. The special case when
G + uv is a 3-connected R2 -circuit follows from Theorem 21.32.

21.7.4 Uniquely Localizable Vertices

Unique localizability has direct applications in sensor network localization. We can think of P as
a set of pinned vertices (or anchor nodes) in a sensor network. Vertices in P are, by definition,
uniquely localizable. It is easy to observe that, if some vertex v ∈ V − P is uniquely localizable, then
|P| ≥ 3 and there exist three openly disjoint paths from v to P (cf. Lemma 21.7.1). As was the case
for global linkedness, unique localizability is not a generic property.
Let G + K(P) denote the graph obtained from G by adding all edges bb0 for which bb0 ∈ / E and
0
b, b ∈ P. The following lemma from [31] is easy to prove.

Lemma 21.7.2 [31] Let G = (V, E) be a graph, P ⊆ V and v ∈ V − P. Then v is uniquely localizable
in G with respect to P if and only if |P| ≥ 3 and {v, b} is globally linked in G + K(P) for all (or
equivalently, for at least three) vertices b ∈ P.

Lemma 21.7.2 and Theorem 21.27 imply the following characterization of uniquely localizable
vertices when G + K(P) is R2 -connected.

Theorem 21.34 [31] Let G = (V, E) be a graph, P ⊆ V and v ∈ V − P. Suppose that G + K(P)
is R2 -connected. Then v is uniquely localizable in G with respect to P if and only if |P| ≥ 3 and
κ(v, b) ≥ 3 for all b ∈ P.

Similarly, Lemma 21.7.2 and Conjecture 6 would imply the following characterization of
uniquely localizable vertices in an arbitrary graph.

Conjecture 9 Let G = (V, E) be a graph, P ⊆ V and v ∈ V − P. Then v is uniquely localizable

in G with respect to P if and only if |P| ≥ 3 and there is an R2 -component H of G + K(P) with
P + v ⊆ V (H) and κH (v, b) ≥ 3 for all b ∈ P.

Theorems 21.27 and 21.34 imply that the sets of globally linked pairs and uniquely localizable
vertices can be determined for R2 -connected graphs. Conjectures 6 and 9 would extend this to all
graphs.

21.7.5 The Number of Non-Equivalent Realizations

An infinitesimally rigid framework (G, p) is regular valued if all equivalent frameworks are in-
finitesimally rigid. (See Chapter 18 for the definition of infinitesimal rigidity.) It is known that an
infinitesimally rigid, regular valued framework (G, p) has only finitely many equivalent and pair-
wise non-congruent realizations. We denote this number by R(G, p). This parameter is related to
global linkedness by the fact that two vertices u, v are globally linked in an infinitesimally rigid,
regular valued framework (G, p) if and only if R(G, p) = R(G + uv, p). As for global linkedness,
the value taken by R(G, p) is not a generic property: it is not difficult to see that R(G, p) = 2 and
R(G, q) = 4 for the frameworks (G, p) and (G, q) shown in Figure 21.12(a) and (b), respectively.
On the other hand the following result shows that R(G, p) is constant for all generic realizations
of an R2 -connected graph G = (V, E). For each {u, Pv} ⊂ V , let wG (u, v) denote the number of
connected components of G − {u, v} and put b(G) = (wG (u, v) − 1), where the summation is over
all {u, v} ⊂ V .

Theorem 21.35 [31] Let (G, p) be a generic realization of a R2 -connected graph G. Then
R(G, p) = 2b(G) .
478 Handbook of Geometric Constraint Systems Principles

For example, in the graph G of Figure 21.13, b(G) = 4 and hence R(G, p) = 16 for all generic
realizations (G, p).
Theorem 21.35 enables us to obtain a representative of each distinct congruence class of frame-
works which are equivalent to a given generic framework (G, p) when G is R2 -connected by itera-
tively applying the following operation to (G, p). Choose a 2-vertex-cut {u, v} of G and reflect some,
but not all, of the components of G − {u, v} in the line through the points p(u) and p(v). Thus, even
if a sensor network with an R2 -connected grounded graph is not uniquely localizable, we may still
obtain all possible sets of locations from one set of feasible locations in a straightforward manner.
It would be of interest to find tight bounds on R(G, p) when G belongs to a given family of
graphs and p is generic. To this end we let R(G) be the maximum value taken by R(G, p) over all
generic realizations of G. Borcea and Streinu [7] show that R(G) ≤ 12 2n−4 n for all rigid graphs


n−2 ≈ 4
G with n vertices and construct an infinite family of rigd graphs G with R(G) = 12(n−3)/3 ≈ 2.29n .
A slightly better construction due to Emiris and Moroz [17], see also [18], gives an infinite family
of rigid graphs G with R(G) = 28(n−3)/4 ≈ 2.3n . Capco et al [9] have recently obtained a recurrence
formula which gives an upper bound on R(G) when G is minimally rigid (more precisely their
recurrence formula determines the number of frameworks in C2 which are equivalent to a given
generic realization of G).
We may also consider lower bounds on R(G). Jackson and Owen [33] conjecture that R(G) ≥
2n−3 when G is minimally rigid and verify their conjecture in the special case when G is planar.

21.7.6 Stability Lemma and Neighborhood Results

Connelly and Whiteley [16, Theorem 13] showed that if a d-dimensional framework (G, p) is both
infinitesimally rigid and globally rigid, then the same properties hold for all frameworks (G, q)
which are close enough to (G, p):

Theorem 21.36 Given a framework (G, p) which is globally rigid and infinitesimally rigid in Rd ,
there is an open neighborhood U of p in Rd|V (G)| such that for all q ∈ U the framework (G, q) is
globally rigid and infinitesimally rigid.

We will describe analogues of this result for globally linked pairs and the number of non-
equivalent realizations. The first result shows that, for an infinitesimally rigid, regular valued frame-
work framework (G, p), R(G, p) does not increase in some open neighborhood of p.

Theorem 21.37 [32] Suppose that (G, p) is an infinitesimally rigid, regular valued framework.
Then there exists an open neighborhood U of p such that, for all q ∈ U, (G, q) is an infinitesimally
rigid, regular valued framework with R(G, q) ≤ R(G, p).

Note that Theorem 21.37 generalizes (the 2-dimensional version of) Theorem 21.36 since an
infinitesimally rigid, globally rigid framework (G, p) is regular valued and has R(G, p) = 1. The
following example shows that we can have R(G, q) < R(G, p) for a framework (G, p) satisfying the
hypotheses of Theorem 21.37 and q arbitrarily close to p. Consider the realization (G, p) of a wheel
in which the central vertex and two nonconsecutive rim vertices are collinear. Then R(G, p) = 2 but
R(G, q) = 1 for all generic (G, q) since wheels are globally rigid.
In contrast, our next results shows that R(G, p) is constant in some open neighborhood of p if
either G is minimally rigid or p is generic.

Theorem 21.38 [32] Suppose that (G, p) is an infinitesimally rigid, regular valued realization of a
minimally rigid graph G = (V, E). Then there exists an open neighborhood U of p such that, for all
q ∈ U, (G, q) is infinitesimally rigid, regular valued, and has R(G, q) = R(G, p).
Global Rigidity of Two-Dimensional Frameworks 479

Figure 21.14
The framework (G, p) is an infinitesimally rigid, regular valued realization of a minimally rigid
graph. There are exactly four equivalent realizations which keep the triangle pu pv pw fixed, and they
can be obtained by reflecting x in the line through pu pz and/or reflecting y in the line through pv pz .
The distance between x and y is the same in all such realizations so (x, y) is globally linked in (G, p).
On the other hand, {x, y} is not globally linked in any generic realization (G, q).

Corollary 21.39 Suppose that (G, p) is an infinitesimally rigid, regular valued realization of a min-
imally rigid graph G = (V, E). Then there exists an open neighborhood U of p such that, for all
u, v ∈ V and all q ∈ U, {u, v} is not globally linked in (G, q) if {u, v} is not globally linked in (G, p).

The analogous result for the property that {u, v} are globally linked does not hold in general,
see Figure 21.14.
We next show that R(G, p) remains constant in an open neighborhood of p for any rigid graph
G when p is generic.

Theorem 21.40 [32] Suppose that G = (V, E) is rigid and (G, p) is generic. Then there exists an
open neighborhood U of p such that, for all q ∈ U, (G, q) is infinitesimally rigid, regular valued,
and has R(G, p) = R(G, q).

Corollary 21.41 Suppose that G = (V, E) is rigid and (G, p) is generic. Then there exists an open
neighborhood U of p such that, for all u, v ∈ V and all q ∈ U, {u, v} is globally linked in (G, p) if
and only if {u, v} is globally linked in (G, q).

The realization (G, p) of a wheel in which the central vertex and two nonconsecutive rim vertices
are collinear shows that Corollaries 21.39 and 21.41 become false if we remove the respective
hypotheses that G is minimally rigid or p is generic. The problem is that there are pairs of vertices
which are not globally linked in (G, p) but are globally linked in (G, q) for q arbitrarily close to p.
The example in Figure 21.14 shows that we can also have pairs of vertices which are globally linked
in (G, p) but are not globally linked in (G, q) for q arbitrarily close to p, if we remove the hypothesis
that (G, p) is generic.

21.8 Direction Constraints

Parallel drawings: Two d-dimensional frameworks (G, p) and (G, q) are parallel if p(u) − p(v) is
a scalar multiple of q(u) − q(v) for all uv ∈ E.
Direction congruence: Two d-dimensional frameworks (G, p) and (G, q) are direction congruent
if there exists a scalar λ and a vector t such that q(v) = λ p(v)+t for all v ∈ V . This is equivalent
to saying that (G, q) can be obtained from (G, p) by a translation and dilation.
480 Handbook of Geometric Constraint Systems Principles

Tight framework: A d-dimensional framework (G, p) is tight if every d-dimensional framework

which is parallel to (G, p) is direction congruent to (G, p). The graph G is in Rd if every generic
realization of G in Rd is tight.
Mixed graph: A mixed graph is a graph together with a bipartition D ∪ L of its edge set. We refer
to edges in D as direction edges and edges in L as length edges.
Direction-length framework: A d-dimensional direction-length framework (G, p) is a mixed
graph G = (V ; D, L) together with a map p : V → Rd .
Direction-length equivalence: Two d-dimensional direction-length frameworks (G, p) and (G, q)
are equivalent if p(u) − p(v) is a scalar multiple of q(u) − q(v) for all uv ∈ D and ||p(u) −
p(v)|| = ||q(u) − q(v)|| for all uv ∈ L.
Direction-length congruence: Two d-dimensional direction-length frameworks (G, p) and (G, q)
are congruent if there exists a vector t ∈ Rd and λ ∈ {−1, 1} such that q(v) = λ p(v) + t for all
v ∈ V . This is equivalent to saying that (G, q) can be obtained from (G, p) by an inversion and
a translation.
Direction-length global rigidity: A d-dimensional direction-length framework (G, p) is globally
rigid if every d-dimensional direction-length framework which is equivalent to (G, p) is con-
gruent to (G, p).
Direction-length rigidity: A d-dimensional direction-length framework (G, p) is rigid if there ex-
ists an ε > 0 such that every d-dimensional direction-length framework (G, q) which is equiv-
alent to (G, p) and satisfies kp(v) − q(v)k < ε for all v ∈ V , is congruent to (G, p). The mixed
graph G is rigid in Rd if every generic realization of G in Rd is rigid.
Direction-balanced mixed graph: A 2-separation G1 , G2 of a mixed graph G is is a pair of
subgraphs G1 , G2 such that G = G1 ∪ G2 , |V (G1 ) ∩ V (G2 )| = 2 and V (G1 ) − V (G2 ) 6= ∅ =6
V (G2 ) −V (G1 ). It is direction-balanced if both G1 and G2 contain a direction edge. The mixed
graph G is direction balanced if all 2-separations of G are direction balanced.

21.8.1 Parallel Drawings

The fact that direction constraints are linear implies that the local and global versions of tightness are
equivalent properties of frameworks (if two d-dimensional frameworks (G, p) and (G, q) are parallel
then we can continuously move each vertex v of G on the line joining p(v) to q(v) to transform
(G, p) to (G, q) in such a way that the intermediate frameworks remain parallel). Whiteley [51]
characterized generic tightness in all dimensions.

Theorem 21.42 [51] Let (G, p) be a d-dimensional generic framework. Then (G, p) is tight if and
only if X
(d|V (H)| − d − 1) ≥ d|V (G)| − d − 1
H∈H

for all families H of subgraphs which have at least two vertices and cover E(G).

This implies, in particular, that tightness is a generic property of frameworks.

When d = 2, the characterization of generic tightness given in Theorem 21.42 is the same as
the characterization of generic rigidity given by Lovász and Yemini [44]. This implies that generic
tightness and generic rigidity are equivalent when d = 2.
Global Rigidity of Two-Dimensional Frameworks 481

21.8.2 Direction-Length Global Rigidity

Since the problem of characterizing global rigidity for direction-length frameworks is at least as hard
as the corresponding problem when there are only length constraints, we will restrict our attention
to the 2-dimensional case, unless explicitly stated otherwise.
Servatius and Whiteley [47] developed a rigidity theory for direction-length frameworks in R2
analogous to that for length constrained frameworks. One may construct a |D ∪ L| × 2|V | (direction-
length) rigidity matrix for a direction-length framework (G, p) and use its rows to define the
(direction-length) rigidity matroid RDL (G, p) of (G, p). A generic direction-length framework is
rigid if and only if its rigidity matrix, or matroid, has rank 2|V | − 2. It follows that the rigidity of
direction-length frameworks is a generic property. All generic realizations of a mixed graph G have
the same rigidity matroid and this matroid is defined to be the rigidity matroid RDL (G) of G. The
main result of Servatius and Whiteley [47] characterizes the generic rigidity of mixed graphs in R2 .
The problem of characterizing when a generic direction-length framework (G, p) is globally
rigid is still an open problem, even though some recent results get very close to this target. More
precisely, global rigidity has been characterized for generic frameworks when the rigidity matroid
of the underlying mixed graph is connected. Mixed graphs with the property that all their generic
realizations are globally rigid have also been characterized. (It is not yet known whether global
rigidity of direction-length frameworks is a generic property i.e., we do not know whether there
exist mixed graphs which have both a globally rigid generic realization and a non-globally rigid
generic realization in R2 .)
We first give a necessary condition for global rigidity, which is analogous to the “3-
connectedness condition” of Theorem 21.2.

Theorem 21.43 [29] Suppose that (G, p) is a generic globally rigid direction-length framework.
Then G is 2-connected and direction balanced.

Rigidity is also a necessary condition for global rigidity. Redundant rigidity, however, is no
longer necessary. To see this consider a minimally rigid mixed graph G with exactly one length
edge e. The 2-dimensional version of Theorem 21.42 and the above mentioned characterization of
generic direction-length rigidity imply that G − e is tight, and this in turn implies that every generic
realization of G is globally rigid. On the other hand, G − f is not rigid for all edges f of G.
We next give a result on 1-extensions which is analogous to Theorem 21.4. For a mixed graph
G, the operation 1-extension (on edge uw and vertex z) deletes the edge uw and adds a new vertex
v and new edges vu, vw, vz for some vertex z ∈ V (G), with the provisos that at least one of the new
edges has the same type as the deleted edge and, if z = u, then the two edges from z to u are of
different type.

Theorem 21.44 [30] Let H be a mixed graph with at least three vertices and G be obtained from
H by a 1-extension on an edge uw. Suppose that (G, p) is a generic realization of G. If H − uw is
rigid and (H, p|H ) is globally rigid, then (G, p) is globally rigid.

In mixed graphs, a special kind of 0-extension also preserves global rigidity.

Theorem 21.45 [30] Let G and H be mixed graphs with |V (H)| ≥ 2 and (G, p) be a generic real-
ization of G. Suppose that G can be obtained from H by adding a vertex v incident to two direction
edges. Then (G, p) is globally rigid if and only if (H, p|H ) is globally rigid.

Note that the graph G we obtain in Theorem 21.45 will not be redundantly rigid since it will
contain a vertex of degree two.
We can use Theorems 21.44 and 21.45 to show that a special family of generic direction-length
frameworks are globally rigid. We say that a mixed graph G = (V ; D, L) is a mixed RDL -circuit
482 Handbook of Geometric Constraint Systems Principles

6 L and D ∪ L is a circuit in RDL (G). It is easy to show that mixed RDL -circuits are 2-
if D 6= ∅ =
connected. Jackson and Jordán [29] showed that the necessary condition for global rigidity given in
Theorem 21.43 is also sufficient to imply that mixed RDL -circuits are globally rigid.
Theorem 21.46 Let (G, p) be a generic realization of a mixed RDL -circuit. Then (G, p) is globally
rigid if and only if G is direction balanced.
Clinch [11] extended Theorem 21.46 to RDL -connected mixed graphs, i.e., mixed graphs G for
which the matroid RDL (G) is connected.
Theorem 21.47 Let (G, p) be a generic realization of an RDL -connected mixed graph. Then (G, p)
is globally rigid if and only if G is direction balanced.
A recent result of Clinch, Jackson, and Keevash [12] characterizes mixed graphs with the prop-
erty that all their generic realizations in R2 are globally rigid.
Theorem 21.48 Suppose G = (V, D ∪ L) is a mixed graph. Then every generic realization of G
is globally rigid if and only if G is rigid and either |L| = 1 or G has a direction-balanced RDL -
connected subgraph which contains L.
We close this section by noting that we can characterize global rigidity in all dimensions for
mixed graphs in which every pair of adjacent vertices is connected by both a length and a direction
edge.
Theorem 21.49 [29] Let G be a mixed graph in which every pair of adjacent vertices is connected
by both a length and a direction edge, and (G, p) be a generic realization of G in Rd . Then (G, p)
is globally rigid if and only if G is 2-connected.

21.9 Algorithms
The structural results presented in this chapter give rise to efficient combinatorial algorithms for
testing different global rigidity properties of generic frameworks and for solving a number of related
algorithmic problems in the plane. The major problems are as follows. How can we decide whether
a given graph G is rigid, redundantly rigid, or R2 -connected? More generally, how can we identify
the rigid, redundantly rigid, and R2 -connected components of G?
The key ingredient in solving these problems is an efficient subroutine for checking if a set of
edges is independent in the 2-dimensional rigidity matroid of G. For this matroid, which is known
to be a so-called count matroid, we have fast combinatorial algorithms for testing independence.
This subroutine can be implemented in O(|V |2 ) time by using various alternating path algorithms:
methods from matching theory [24], network flows [25], and graph orientations [4, 42] have been
used to do this. By using additional algorithmic techniques, each of the problems mentioned above
can be solved in O(|V |3 ) time.

21.10 Optimization Problems

The globally rigid pinning problem is motivated by a natural question from network localization:
the goal is to find a smallest set P of vertices in a graph G for which G + K(P) is globally rigid
References 483

in R2 (where K(P) denotes a complete graph on the vertex set P). Although the complexity of this
problem is still open, an efficient approximation algorithm is given by Jordán [36].
The R2 -connected pinning problem is a variant of the globally rigid pinning problem in which
the goal is to find a smallest set P ⊆ V such that G + K(P) is R2 -connected. This problem can
be formulated as finding a largest matroid matching in the hypergraphic matroid defined on V by
the R2 -components of G. Hypergraphic matroids are known to be linear, but it is not known how
to find a suitable linear representation. The complexity status of the matroid matching problem
in hypergraphic matroids is still open. Nevertheless, this formulation can be used to design a 32 -
approximation algorithm (which works for the minimum cost version as well). This approximation
algorithm can be used as a subroutine to design a 52 -approximation algorithm for finding a smallest
(or minimum cost) subset P for which G + K(P) is globally rigid in R2 [36].
The above methods can be used to design a constant factor approximation algorithm for the
corresponding augmentation problem in which the goal is to add a smallest set F of new edges to
G so that G + F is globally rigid. It was shown by Garcı́a and Tejel [22] that the related problem of
augmenting a rigid graph to a redundantly rigid graph by a smallest set of new edges is NP-hard.

Acknowledgment
This work was supported by the Hungarian Scientific Research Fund grant no. K109240 and
K115483. The third author was supported by JSPS KAKENHI Grant Number JP15K15942.

References
[1] T.G. Abbot. Generalizations of Kempe’s universality theorem. Master’s thesis, MIT,
http://web.mit.edu/tabbott/www/papers/mthesis.pdf, 2008.
[2] L. Asimow and B. Roth. The rigidity of graphs. Trans. Amer. Math. Soc., 245:279–289, 1978.
[3] J. Aspnes, T. Eren, D. K. Goldenberg, A. S. Morse, and W. Whiteley. A theory of network
localization. IEEE Transactions on Mobile Computing, 5:1663–1678, 2006.
[4] A.R. Berg and T. Jordán. Algorithms for graph rigidity and scene analysis. In U. Zwick
G. Di Battista, editor, Proc. 11th Annual European Symposium on Algorithms (ESA), Springer
Lecture Notes in Computer Science 2832, pages 78–89, 2003.
[5] A.R. Berg and T. Jordán. A proof of Connelly’s conjecture on 3-connected circuits of the
rigidity matroid. J. Combinatorial Theory Ser. B., 88:77–97, 2003.
[6] B. Bollobás. Random graphs. Academic Press, New York, 1985.
[7] C. Borcea and I. Streinu. The number of embeddings of minimally rigid graphs. Discrete
Comput. Geom., 31:287–303, 2004.
[8] H. Breu and D.G. Kirkpatrick. Unit disk graph recognition is NP-hard. Comput. Geom.,
9:3–24, 1998.
[9] J. Capco, M. Gallet, G. Grassegger, C. Koutschan, N. Lubbes, and J. Schicho. The number of
realizations of a Laman graph. Technical report, arXiv 1701.05500v2, 2017.
484 References

[10] M. Cheung and W. Whiteley. Transfer of global rigidity results among dimensions: graph
powers and coning. Technical report, York University, July 2008.
[11] K. Clinch. Global rigidity of 2-dimensional direction-length frameworks with connected
rigidity matroids. Technical report, arXiv:1608.08559, 2016.
[12] K. Clinch, B. Jackson, and P. Keevash. Global rigidity of 2-dimensional direction-length
frameworks. Technical report, arXiv:1607.00508v2, 2018.
[13] R. Connelly. Rigidity and energy. Invent. Math., 66:11–33, 1982.
[14] R. Connelly. On generic global rigidity. In Applied Geometry and Discrete Mathematics,
volume 4 of DIMACS Ser. Discrete Math, Theoret. Comput. Sci., pages 147–155. Amer. Math.
Soc., 1991.
[15] R. Connelly. Generic global rigidity. Discrete Comput. Geom., 33:549–563, 2005.
[16] R. Connelly and W. Whiteley. Global rigidity: the effect of coning. Discrete Comput. Geom.,
43:717–735, 2010.
[17] I.Z. Emiris and G. Moroz. The assembly modes of 11-bar linkages. In Proc. IFToMM World
Cong. Mechanism and Machine Science, Guanajuato, Mexico, 2011.
[18] I.Z. Emiris and I. D. Psarros. Counting euclidean embeddings of rigid graphs. Technical
report, arXiv:1402.1484v1, 2014.
[19] T. Eren, W. Whiteley, A. Morse, P.N. Belhumeur, and B.D.O. Anderson. Sensor and network
topologies of formations with direction, bearing, and angle information between agents. In
Proc. of the 42nd IEEE Conference on Decision and Control, pages 3064–3069, Maui, HI,
USA, December 2003.
[20] A. Frank. Connections in combinatorial optimization. Oxford University Press, 2011.
[21] S. Frank and J. Jiang. New classes of counterexamples to Hendrickson’s global rigidity con-
jecture. Discrete Comput. Geom., 45(3):574–591, 2011.
[22] A. Garcı́a and J. Tejel. Augmenting the rigidity of a graph in R2 . Algorithmica, 59:145–168,
2011.
[23] S. Gortler, A. Healy, and D. Thurston. Characterizing generic global rigidity. American
Journal of Mathematics, 132(4):897–939, 2010.
[24] B. Hendrickson. Conditions for unique graph realizations. SIAM J. Comput., 21:65–84, 1992.
[25] H. Imai. Network-flow algorithms for lower-truncated transversal polymatroids. J. Operations
Res. Soc. Japan, 26:186–210, 1983.
[26] B. Jackson and T. Jordán. Connected rigidity matroids and unique realizations of graphs. J.
Combinatorial Theory Ser B, 94:1–29, 2005.
[27] B. Jackson and T. Jordán. Graph theoretic techniques in the analysis of uniquely localizable
sensor networks. In G. Mao and B. Fidan, editors, Localization Algorithms and Strategies for
Wireless Sensor Networks. IGI Global, 2009.
[28] B. Jackson and T. Jordán. A sufficient connectivity condition for generic rigidity in the plane.
Discrete Appl. Math., 157:1965–1968, 2009.
[29] B. Jackson and T. Jordán. Globally rigid circuits of the direction-length rigidity matroid. J.
Combinatorial Theory Ser B, 100:1–22, 2010.
References 485

[30] B. Jackson and T. Jordán. Operations preserving global rigidity of generic direction-length
frameworks. International Journal on Computational Geometry and Applications, 20:685–
706, 2010.
[31] B. Jackson, T. Jordán, and Z. Szabadka. Globally linked pairs of vertices in equivalent real-
izations of graphs. Discrete Comput. Geom., 35:493–512, 2006.
[32] B. Jackson, T. Jordán, and Z. Szabadka. Globally linked pairs of vertices in rigid frameworks,
in: Rigidity and symmetry. Fields Institute Communications, 70:177–203, 2014.
[33] B. Jackson and J.C. Owen. The number of equivalent realisations of a rigid graph. Technical
report, arXiv:1204.1228, 2012.
[34] B. Jackson, B. Servatius, and H. Servatius. The 2-dimensional rigidity of certain families of
graphs. J. Graph Theory, 54:154–166, 2007.
[35] T. Jordán. Generically globally rigid zeolites in the plane. Information Processing Letters,
110:841–844, 2010.
[36] T. Jordán. Rigid and globally rigid graphs with pinned vertices. In G.O.H. Katona, A. Schri-
jver, and T. Szőnyi, editors, Fete of Combinatorics and Computer Science, Bolyai Society
Mathematical Studies. Springer, 2010.
[37] T. Jordán. Combinatorial rigidity: graphs and matroids in the theory of rigid frameworks.
In Discrete Geometric Analysis, volume 34 of MSJ Memoirs. The Mathematical Society of
Japan, 2016.
[38] T. Jordán, C. Király, and S. Tanigawa. Generic global rigidity of body-hinge frameworks. J.
Combinatorial Theory Ser B, 117:59–76, 2016.
[39] T. Jordán and Z. Szabadka. Operations preserving the global rigidity of graphs and frame-
works in the plane. Computational Geometry, 42:511–521, 2009.
[40] V.E. Kaszanitzky and C. Király. On minimally highly vertex-redundantly rigid graphs. Graphs
and Combinatorics, 32:225–240, 2016.
[41] G. Laman. On graphs and rigidity of plane skeletal structures. J. Engineering Math., 4:331–
340, 1970.
[42] A. Lee and I. Streinu. Pebble game algorithms and sparse graphs. Discrete Math., 308:1425–
1437, 2008.
[43] X-Y. Li, P-J. Wan, Y. Wang, and C-W. Yi. Fault tolerant deployment and topology control
in wireless networks. In Proceedings of the ACM Symposium on Mobile Ad Hoc Networking
and Computing (MobiHoc), pages 117–128, Annapolis, MD, June 2003.
[44] L. Lovász and Y. Yemini. On generic rigidity in the plane. SIAM J. Algebraic Discrete
Methods, 3:91–98, 1982.
[45] J.G. Oxley. Matroid theory. Oxford University Press, 2nd edition, 2011.
[46] J.B. Saxe. Embeddability of weighted graphs in k-space is strongly NP-hard. Technical report,
Computer Science Department, Carnegie-Mellon University, 1979.
[47] B. Servatius and W. Whiteley. Constraining plane configurations in computer-aided design:
combinatorics of directions and lengths. SIAM J. Discrete Math., 12:136–153, 1999.
[48] A. M. So and Y. Ye. Theory of semidefinite programming for sensor network localization.
Math. Program., 109:367–384, 2007.
486 References

[49] Z. Szabadka. Globally rigid frameworks and rigid tensegrity graphs in the plane. PhD thesis,
Institue of Mathematics, Eötvös Loránd University, Hungary, 2010.
[50] S. Tanigawa. Sufficient conditions for the global rigidity of graphs. J. Combinatorial Theory
Ser B, 113:123–140, 2015.
[51] W. Whiteley. Some matroids from discrete applied geometry. In J. Bonin, J. Oxley, and
B. Servatius, editors, Matroid Theory, volume 197 of Contemp. Math. Amer. Math. Soc.,
1996.
Chapter 22
Point-Line Frameworks
Bill Jackson
School of Mathematical Sciences, Queen Mary University of London, London, England.

J.C. Owen
Siemens, Cambridge, England.

CONTENTS
22.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
22.1.1 Motivation from CAD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
22.1.2 Motivation from Automated Deduction in Geometry (ADG) and Theorem
Proving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
22.1.3 Constraint Graphs and Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
22.2 Point-Line Graphs and Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 490
22.2.1 Point-Line Frameworks and the Rigidity Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
22.2.2 The Rigidity Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492
22.2.3 The Rigidity Matroid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 494
22.2.4 Affine Properties of the Point-Line Rigidity Matrix . . . . . . . . . . . . . . . . . . . . . . . . 494
22.2.5 Fixed-Slope Point-Line Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495
22.3 Characterization of the Generic Rigidity Matroid for Point-Line Frameworks in R2 . . 496
22.3.1 A Count Matroid for Point-Line Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496
22.3.2 A Characterization of Independence in MPL (G) when G is Naturally
Bipartite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
22.3.3 A Characterization of Independence in MPL (G) . . . . . . . . . . . . . . . . . . . . . . . . . . . 501
22.3.4 The Rank Function for MPL (G) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 501
22.4 Extensions to Rd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
22.4.1 Point-Line Frameworks in Rd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
22.4.2 Point-Hyperplane Frameworks in Rd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
22.5 Direction-Length Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503

22.1 Introduction
In this chapter we will describe the rigidity properties of frameworks in which some of the points
(representing joints) in a bar-joint framework are replaced by lines. The graph of the framework thus
becomes a vertex-labelled graph in which some of the vertices are labelled as points and some are
labelled as lines. This gives rise to three different types of edges – those connecting a pair of points,
those connecting a point and a line and those connecting two lines. The geometry of the framework
corresponds to a collection of points and a collection of unbounded lines. The three classes of edges

487
488 Handbook of Geometric Constraint Systems Principles

correspond to a constrained distance between a pair of points, a constrained distance between a point
and a line (where the distance is defined as the Euclidean distance from the point to the nearest point
on the line) and a constrained angle between two lines.

22.1.1 Motivation from CAD

Our intial interest in point-line frameworks arises from applications in computer-aided design
(CAD). The first CAD systems typically represented all of the component geometric elements by a
number of points on the element [16, 11]. For example a line can be represented by two points on
the line or a plane by three points and then the distance between a point and a line is a trivalent con-
straint and the angle between two lines a four-valent constraint. However it became clear [3, 12, 19]
that it is simpler and more efficient to represent the geometric elements more directly so that for
example a line has coordinates corresponding to its position and tangent direction and a plane has
coordinates corresponding to its position and normal direction.

22.1.2 Motivation from Automated Deduction in Geometry (ADG) and Theorem

Proving
Many theorems of Euclidean geometry and projective geometry can be described in terms of geo-
metric elements such as points, lines and circles together with distance relations between pairs of
them [1, 5, 15]. For example the coincidence of a point and line implies that their separation distance
is zero and similarly for a line-circle tangency. Many theorems can be interpreted as a statement that
some of these incidence relations can be deduced from the existence of other relations and recog-
nising this by some automated procedure is an active area of research.
As a specific example the well-known theorem of Desargues on the plane can be represented as
a point-line framework in which certain points are incident to certain lines. If the points and lines
are subject to these constraints but are otherwise generic then the dependency of the rigidity matrix
would imply the theorem. The existence of such theorems shows that charactarising the rigidity
properties of such quasi-generic point-line frameworks is a hard (and unsolved) problem.

22.1.3 Constraint Graphs and Frameworks

In CAD and ADG we are interested in many different types of geometric elements. For example
the representation of a simple machined part may entail representing points, lines, planes, circles,
cylinders, tori, and cones [13, 20]. If we also specify relationships between them, such as the fact
that some pair are tangent or have a constrained separation distance between them, then this is
commonly represented by a constraint-graph in which the nodes in the graph are labelled with
the type of geometric element which they represent and the edges are labelled with the type of
relationship which exists between its incident vertices [9].
If we assign coordinates to the geometric elements then the constraint graph becomes a con-
straint framework. It is rigid if every continuous motion of the geometric elements which preserves
the constraints results in a constraint framework which can be obtained from the initial framework
by an isometry of the space we are working in. In this view the bar-joint frameworks which are the
subject of many of the chapters in this book are constraint frameworks in which all of the geometric
elements are points and all of the constraints are distances between pairs of points.
Figure 22.1 shows all complete point-line graphs on three vertices. In these graphs we have
adopted our usual convention that vertices which represent points are drawn as filled circles and la-
belled ui and vertices which represent lines are drawn as unfilled circles and labelled vi . Figure 22.2
shows point-line frameworks which represent a geometric realization of each graph in Figure 22.1.
The dotted lines in these frameworks represent either a distance constraint or an angle constraint
between the two geometries which they connect. The distance and angle constraints in frameworks
Point-Line Frameworks 489

u1 v1

(a) (b)

u2 u3 u2 u3
v1 v1

(c) (d)

v2 u3 v2 v3

Figure 22.1
The four distinct, complete point-line graphs on three vertices. Filled circles are point-vertices la-
belled ui and unfilled circles are line-vertices labelled vi .

u1
u3
u2
(a) (b)

u2 u3 v1

(c) v2 (d) v3
u3 v2

v1 v1

Figure 22.2
Four frameworks corresponding to the four point-line graphs in Figure 22.1.
490 Handbook of Geometric Constraint Systems Principles

v3

u5
u3 u6
(a) (b)
u4 u5
u1
v2 u6 v3
v1 v2 u2 u4
u1 u3
u2

v1

Figure 22.3
A point-line graph (a) and a corresponding point-line framework (b) in R2 .

(a), (b) and (c) can be specified independently. However the three angle constraints in framework
(d) must sum to 180 degrees and so cannot be specified independently.
Figure 22.3 shows a graph and corresponding framework which is derived from the 3-cycle
of line-vertices shown in Figure 22.1(d) by substituting rigid subgraphs in place of the angle con-
straints. Although this graph has only distance constraints (between points and lines or between
pairs of points), we will see that the distances in this graph cannot be specified independently.
As we observed in Section 22.1.1, a line can also be represented by a pair of points and so we
may ask what is to be gained by allowing additional geometric elements. It is not difficult to see that
many of the constraints can also be represented by bars in an appropriate bar-joint framework. For
example a point may be constrained to be a fixed distance from a line which is represented by a pair
of points using a Peaucellier mechanism which is itself a bar-joint framework.
However, the result of this substitution is a very specialized bar-joint framework and if there are
several such substitutions it is difficult to get any intuition for the rigidity properties of the frame-
work. Many of the important results for bar-joint frameworks refer to specific types of frameworks
and these may not include those specialized bar-joint frameworks which can represent more general
constraint frameworks [17]. The properties of such frameworks may only be obtained by studying
the constraint framework directly. For example Laman’s theorem describes the rigidity properties
of a bar-joint framework for which the point coordinates are generic (algebraically independent
over the rationals). A point-line framework for which the line positions and slopes are also generic
would give a corresponding bar-joint framework in which the point coordinates are not generic and
so Laman’s theorem would not apply. We will describe below how we can obtain a theorem which
corresponds to Laman’s theorem by considering the point-line framework directly.

22.2 Point-Line Graphs and Frameworks

Point-line graph: A point-line graph G = (VP ,VL , E) is a graph G = (V, E) without loops together
with an ordered pair (VP ,VL ) of, possibly empty, disjoint sets whose union is V [9].
Point-Line Frameworks 491

u1
u1 u2 u3
v2 v3

(a) (b)
u3 u2

v1 v2 v3 v1

Figure 22.4
(a) The naturally bipartite point-line graph K(3,3) and (b) a framework on this graph.

We refer to vertices in VP and VL as point-vertices and line-vertices, respectively. We label the

vertices as VP = {u1 , . . . , us } and VL = {v1 , . . . , vt }, and the edges as E = {e1 , e2 , . . . , em }. We use
EPP , EPL and ELL to denote the sets of edges incident to two point-vertices, to a point-vertex and
a line-vertex and to two line-vertices respectively. For e ∈ E, we write e = xy to mean that the
end-vertices of e are x and y.
We supplement the above notation when it is not obvious which graph we are referring to by
using V (G), E(G), etc.
We will assume that a point-line graph is simple (without parallel edges) unless it is explicitly
described as a point-line multigraph. Sometimes we will want to generate a multigraph from a
simple graph G by introducing a second copy of an edge which is already in the graph. We call this
doubling an edge. If e = i j ∈ E(G) then we write G + i j for the graph with e doubled and denote
the new edge by e0 .

Naturally bipartite: A point-line graph G = (VP ,VL , E) is naturally bipartite if G is a bipartite

graph with bipartition VP ,VL .

In this case EPP = ELL = ∅ and E = EPL . For example the point-line graph on K(3, 3) shown in
Figure 22.4(a) is naturally bipartite.
Naturally bipartite point-line graphs are especially important. In particular we can derive a cor-
responding naturally bipartite point-line graph from any point-line graph G by replacing any edge
e = ab in EPP or ELL with a copy of the naturally bipartite K(3, 3) shown in Figure 22.4(a) where
either a pair of point-vertices or a pair of line-vertices in K(3, 3) are identified with the vertices a
and b in G. Figure 22.5(b) shows the result of this substitution into the graph in Figure 22.5(a).
We shall see that this substitution preserves infinitesimal rigidity and this will allow us to restrict to
naturally bipartite point-line graphs.

22.2.1 Point-Line Frameworks and the Rigidity Map

A point-line framework is a realization of a point-line graph which is obtained by assigning appro-
priate coordinates to each point- and line-vertex in the graph. We will henceforth fix our attention
to point-line frameworks in R2 , unless explicitly stated otherwise. In this context, a point-vertex ui
gets coordinates (xi , yi ) ∈ R2 . A line may be described in the (x, y)-plane by the equation x = ay + b
where b gives the intersection of the line with the x-axis and a gives its slope. Hence a line-vertex vi
492 Handbook of Geometric Constraint Systems Principles
v4 u1 v1 u4

u1 v1

(a) v5 u3 v3 u5

(b)

u2 v2
v6 u2 v2 u6

Figure 22.5
The point-line graph (a) converts into the naturally bipartite graph (b).

gets coordinates (ai , bi ) ∈ R2 (assuming we choose a coordinate system such that none of the lines
is parallel to the x-axis).
Point-line framework: A point-line framework is a pair (G, p) where G is a point-line graph and
p : V → R2 .
We put p(ui ) = (xi , yi ) for each ui ∈ VP and p(vi ) = (ai , bi ) for each vi ∈ VL . This gives a geometric
realization of (G, p) by taking the point corresponding to ui to have cartesian coordinates (xi , yi ) and
the line corresponding to vi to be x = ai y + bi .

Rigidity map: The rigidity map of a point-line framework (G, p) is the map fG : R2|V | → R|E|
defined as follows.
For each p ∈ R2|V | we take fG (p) = ( f1 (p), f2 (p), . . . , fm (p)) where the components of fG (p)
are indexed by the edges ei ∈ E and
2 2
 (x j − xk ) + (y j − yk ) if ei = u j uk ∈ EPP

1
fi (p) = (x j − y j ak − bk )(1 + a2k )− 2 if ei = u j vk ∈ EPL
tan−1 a j − tan−1 ak if ei = v j vk ∈ ELL and j < k.


These expressions for fi (p) are: the squared distance between the points represented by u j and
uk when ei = u j uk ∈ EPP ; the signed distance between the point represented by u j and the line
represented by vk when ei = u j vk ∈ EPL ; the angle between the lines represented by v j and vk when
ei = v j vk ∈ ELL .

22.2.2 The Rigidity Matrix

Jacobian matrix: The Jacobian matrix J(G, p) of the rigidity map is the |E| × 2|V |-matrix of par-
tial derivatives of fG evaluated at some point p ∈ R2|V | .

The rows of J(G, p) are indexed by E and pairs of columns of J(G, p) are indexed by V as
follows: the two columns indexed by a vertex ui ∈ VP are labelled ui,x and ui,y and the two columns
indexed by a vertex vi ∈ VL are labelled vi,a and vi,b . The entries in J(G, p) are as follows.
• A row indexed by an edge ei = u j uk ∈ EPP has entries 2(x j − xk ), 2(y j − yk ), 2(xk − x j ) and
2(yk − y j ) in the columns indexed by u j,x , u j,y , uk,x and uk,y , respectively.
Point-Line Frameworks 493
1 1
• A row indexed by an edge ei = u j vk ∈ EPL has entries (1 + a2k )− 2 , −ak (1 + a2k )− 2 , (−x j ak −
3 1
y j + ak bk )(1 + a2k )− 2 and −(1 + a2k )− 2 in the columns indexed by u j,x , u j,y , vk,a and vk,b ,
respectively.
1
• A row indexed by an edge ei = v j vk ∈ ELL with j < k has entries (1 + a2j )− 2 and −(1 +
1
a2k )− 2 in the columns indexed by v j,a and vk,a , respectively.
• All other entries are zero.

Rigidity matrix: The rigidity matrix of a point-line framework (G, p) is the |E| × 2|V |-matrix
R(G, p) with the following entries.
• A row indexed by an edge ei = u j uk ∈ EPP has entries x j − xk , y j − yk , xk − x j and yk − y j
in the columns indexed by u j,x , u j,y , uk,x and uk,y respectively.
• A row indexed by an edge ei = u j vk ∈ EPL has entries 1, −ak , −x j ak − y j and −1 in the
columns indexed by u j,x , u j,y , vk,a and vk,b , respectively.
• A row indexed by an edge ei = v j vk ∈ ELL with j < k has entries 1 and −1 in the columns
indexed by v j,a and vk,a , respectively.
• All other entries are zero.

The Jacobian matrix J(G, p) can be constructed from R(G, p) using the following row and col-
umn operations. For each vi ∈ VL , multiply the column of R(G, p) indexed by vi,b by ai bi and subtract
it from the column indexed by vi,a , then divide the resulting column by (1 + a2i ). For each e ∈ EPP ,
multiply the row indexed by e by 2. For each e = u j vk ∈ EPL , divide the row indexed by e by
1
(1 + a2k ) 2 . This construction shows that J(G, p) and R(G, p) are similar matrices and we have

Lemma 22.2.1 Let (G, p) be a point-line framework. Then rank J(G, p) = rank R(G, p).

Rigid: A point-line framework is rigid if every continuous motion of the points and lines which
satisfies the constraints results in a framework which can be obtained by a continuous isometry
of R2 (i.e., a translation or rotation).
Degenerate: A framework is degenerate if there is an isometry which leaves all the points and lines
in the framework unchanged.

Hence a degenerate framework has only parallel lines (and no points) and is unchanged by
a translation along the lines, or has only coincident points (and no lines) and is unchanged by a
rotation around the points. All other frameworks are nondegenerate.
If (G, p) is nondegenerate then the rotations and translations of R2 generate a 3-dimensional
subspace of the null space of J(G, p). This gives the following result from [9].

Lemma 22.2.2 Let (G, p) be a nondegenerate point-line framework. Then

(a) rank R(G, p) ≤ 2|V | − 3.
(b) If rank R(G, p) = 2|V | − 3 then (G, p) is rigid.
(c) If p is a point where rank(R(G, p)) is maximum (i.e. a regular point of fG ) and rank R(G, p) <
2|V | − 3 then (G, p) is not rigid.

Generic: A framework (G, p) is generic if the coordinates in p are algebraically independent over
Q.

It is straightforward to show that rank R(G, p) is maximum for any generic point-line framework
(G, p).
494 Handbook of Geometric Constraint Systems Principles
u3
(a) (b)

u2
u1 u2 u3

v1

Figure 22.6
(a) A framework on the graph with three point-vertices in which the points u1 , u2 , and u3 are
collinear and (b) a framework on the graph with a line-vertex and two point-vertices in which the
points u2 , u3 are on an axis orthogonal to the line v1 .

22.2.3 The Rigidity Matroid

Row matroid: The row matroid of a matrix R is the matroid M(R) on the set E of rows of R,
in which a set F ⊆ E is independent in M(R) if and only if the corresponding rows of R are
linearly independent.
Ridigity matroid: The rigidity matroid MPL (G, p) of a framework (G, p) is the row matroid of its
rigidity matrix R(G, p). When (G, p) is generic, its rigidity matroid depends only on the graph
G. We refer to this matroid as the rigidity matroid of G and denote it by MPL (G).

Although the generic rigidity matroid MPL (G) is not sufficient to fully describe the rigidity
properties of all point-line frameworks (G, p) it does describe the properties of a typical framework
on G, that is MPL (G, p) = MPL (G) for almost all p. In addition the matroid MPL (G) does give
some information on the rigidity properties of all frameworks (G, p). For example, a set of rows in
R(G, p) is independent only if the corresponding set of edges are independent in MPL (G) (although
the converse is not necessarily true). For these reasons a charactarization of the rigidity matroid
MPL (G) is important.

22.2.4 Affine Properties of the Point-Line Rigidity Matrix

It is well known that the rank of the rigidity matrix of a bar-joint framework is invariant under
general affine transformations of the underlying Euclidean space [23]. This is also true for a point-
line framework with VP = ∅ because the rigidity matrix does not depend on any of the point or line
coordinates. However the rank of the rigidity matrix of a point-line framework is not invariant under
general affine transformations of space when VP 6= ∅ and VL 6= ∅ as the following example shows.
Figure 22.6(a) shows a framework on the complete graph with three point-vertices and Figure
22.6(b) shows a framework on the complete graph with a line-vertex and two point-vertices. In
Figure 22.6(a) the three points are colinear and in Figure 22.6(b) the line v1 is orthogonal to the
direction from u2 to u3 .
The framework in 22.6(a) has an infinitesimal flex and the rank of its rigidity matrix is 2. Al-
though the rank of the rigidity matrix of a generic framework on the same graph is 3 these two
frameworks are not related by an affine transformation because any affine transform preserves the
collinearity of points. Similarly the framework in 22.6(b) has an infinitesimal flex and the rank of
its rigidity matrix is 2 whereas the rank of the rigidity matrix of a generic framework on the same
graph is 3. However in this case there is an affine transformation connecting the two frameworks
and hence this transform does not preserve the rank of the rigidity matrix.
Point-Line Frameworks 495

22.2.5 Fixed-Slope Point-Line Frameworks

For some applications we may consider the slopes of the lines in a framework to be fixed to given
values and not subject to variation when the framework is flexed [13, 14, 22]. For example if all
the lines are joined by a connected network of angle constraints then there can be no variation
of the relative angles of the lines and the rigidity matrix can be simplified by deleting the rows
corresponding to the angle constraints and the columns corresponding to the line slopes. In addition
we will see that if the graph is naturally bipartite and the coordinates of the points in a point-line
framework are generic then the rigidity properties of the framework are fully determined by the
properties of the corresponding fixed-slope point-line matroid.

Fixed-slope rigidity matrix: The fixed-slope rigity matrix R f ixed (G, p) of a point-line framework
(G, p) is the (|EPP |+|EPL |)×(2(|VP |+|VL |) matrix in which the rows are indexed by EPP ∪EPL
and columns or pairs of columns are indexed by VL and VP . The pairs of columns indexed by
ui ∈ VP are labelled as ui,x and ui,y and the column indexed by vi ∈ VL is labelled as vi,a . The
entries are defined as follows.
• A row in R f ixed (G, p) indexed by an edge ei = u j uk ∈ EPP has entries x j −xk , y j −yk , xk −
x j , yk − y j in the columns indexed by u j,x , u j,y , uk,x and uk,y , respectively.
• A row in R f ixed (G, p) indexed by an edge ei = u j vk ∈ EPL has entries 1, −ak , −1 in the
columns indexed by u j,x , u j,y and vk,b , respectively.
• All other entries are zero.
Fixed-slope rigid: A point-line framework (G, p) is fixed-slope rigid if every continuous motion
of the points and lines which satisfies the constraints determined by the edges of G and does
not change the slope of the lines results in a framework which is a translation of (G, p).

Corresponding to Lemma 22.2.2 we have following

Lemma 22.2.3 Let (G, p) be a point-line framework with |VL | 6= ∅. Then
(a) rank R f ixed (G, p) ≤ 2|VP | + |VL | − 2.
(b) If rank R f ixed (G, p) = 2|VP | + |VL | − 2 then (G, p) is rigid.
(c) If p is a point where rank R f ixed (G, p) is maximum and rank R f ixed (G, p) < 2|VP | + |VL | − 2 then
(G, p) is not rigid.
We denote the row matroid of R f ixed (G, p) by M f ixed (G, p), and simplify this to M f ixed (G)
when p is generic.
If G is naturally bipartite then EPP = ∅ and R f ixed (G, p) does not depend on the coordinates of
the point-vertices. Then R f ixed (G, p) = R f ixed (G, a) where a = a1 , . . . , a|VL | are the line-slope coor-
dinates. Whiteley [22] observed that in this case R f ixed (G, a) is the same matrix as he obtained from
the condition that an incidence graph on the line corresponds to the projection of a nontrivial scene in
the plane. In this case, when a is generic, he obtained the following charactarization of independence
in the matriod M f ixed (G). A set F ⊆ E is independent if and only if |F 0 | ≤ 2|VP (F 0 )| + |VL (F 0 )| − 2
for all nonempty subsets F 0 of F, where VP (F 0 ) and VL (F 0 ) denote the sets of point- and line-vertices
which are incident with F 0 .
In the general case, when EPP 6= ∅, Owen and Power [14] obtained the following charactarization
of independence in M f ixed (G).

Theorem 22.2.4 Let G = (VP ,VL , E) be a point-line graph and F ⊆ E. Then F is independent
M f ixed (G) if and only if |F 0 | ≤ 2|VP (F 0 )| + |VL (F 0 )| − 2 for all non-empty subsets F 0 of F, with
strict inequality whenever F 0 ⊆ EPP .

This charactarization may also be deduced from the charactarization of independence in

MPL (G) which we describe below.
496 Handbook of Geometric Constraint Systems Principles

22.3 Characterization of the Generic Rigidity Matroid for Point-

Line Frameworks in R2
22.3.1 A Count Matroid for Point-Line Graphs
Set function: A set function is a map f : 2E → Z which assigns an integer to every subset of a set
E.

Non-decreasing: The set function f is nondecreasing if f (S + e) ≥ f (S) for all e ∈ E and all S ⊆ E.
Intersecting submodular: The set function f is intersecting submodular if f (S1 )+ f (S2 ) ≥ f (S1 ∪
S2 ) + f (S1 ∩ S2 ) for all S1 , S2 ⊆ S with S1 ∩ S2 6= ∅.

Let f be a nondecreasing intersecting submodular set function which is nonnegative on the

nonempty subsets of E.
Induced matroid: The matroid induced by f is the matroid M( f ) on E in which a set I ⊆ E is
independent if and only if |J| ≤ f (J) for all ∅ 6= J ⊆ I.
We will use the term count matroid to loosely describe a matroid on the edge set of a graph
which is induced by a function which ‘counts’ the numbers of vertices incident with each set of
edges in the graph.
Given a point-line graph G = (VP ,VL , E) we will use three nondecreasing, submodular set func-
tions , νP , νL and ν, on E. These functions are defined by putting νP (F) = |VP (F)|, νL (F) = |VL (F)|
and ν(F) = |V (F)| for all F ⊆ E. They can be used to define several well known count matroids:

• M(ν − 1) is the graphic matroid of a graph G (which is the same as the 1-dimensional bar-
joint rigidity matroid of G).

• M(2ν − 3) is the 2-dimensional bar-joint rigidity matroid of a graph G. This implies that
MPL (G) = M(2ν − 3) when G is a point-line graph and VL = ∅.

• M(2νP + νL − 2) is the fixed slope rigidity matroid of a point-line graph G whenever G is

naturally bipartite. (This follows from the results of Whiteley or Owen and Power mentioned
at the end of Section 22.2.5.)

Since both points and lines have two degrees of freedom in R2 we might expect that the prop-
erties of MPL (G) are not changed significantly when some of the points in a bar-joint framework
are replaced with lines. The graphs on three vertices shown in Figure 22.1 show that this is not the
case: the 3-cycle of point-vertices is generically rigid; the graphs in which one or two of these point-
vertices are replaced by line-vertices are also generically rigid; but no realization of the 3-cycle of
line-vertices in R2 is rigid since the three angles in the triangle cannot be specified independently.
The last example easily generalizes to include any graph which has VP = ∅. In this case any
cycle of edges corresponds to the lines in a geometric polygon in which all of the angles are given
and are therefore dependent. Hence if VP (G) = ∅, MPL (G) is the graphic matroid M(νL − 1).
We have seen that MPL (G) = M(2ν − 3) when VL (G) = ∅ and MPL (G) = M(ν − 1) when
VP (G) = ∅. The other graphs on three vertices suggest the possibility that the first case may hold even
when VL (G) 6= ∅, provided that we have VP (G) 6= ∅. This suggests that MPL might be determined
by the set function ft : 2E → Z given by ft (F) = ν(F) − 1 if VP (F) = ∅ and ft (F) = 2ν(F) − 3
otherwise. The graph and framework shown in Figure 22.7 show that this is not the case. It is easy
Point-Line Frameworks 497

v3
v2 v3
u1 u2
(a) (b) u1

v1 v2
v1

u2

Figure 22.7
The graph (a) is a point-line graph with |VP | = 2 and |VL | = 3. (b) Shows a framework on this graph.

to check that G has 2|V | − 3 vertices and that E is independent in M(2ν − 3). On the other hand,
the line labelled v3 has only angle constraints and hence its position is undetermined. This shows
that G is not generically rigid and hence that MPL (G) is not determined by the set function ft .
Another way to see that the edge set of the graph of figure 22.7(a) is not independent in
MPL (G) is to consider the graph G + v1 v2 . The edge v1 v2 is contained in two distinct dependent
sets S1 = {v1 v2 , v2 v3 , v3 v1 } and S2 = {v1 v2 , u1 v1 , u1 v2 , u2 v1 , u2 v2 , u1 u2 } of the point-line rigidity
matroid. The matroid circuit axiom now implies that E(G) = (S1 ∪ S2 ) − v1 v2 is dependent in the
point-line rigidity matroid.
We can also use the graph G+v1 v2 to show that the set function ft does not have the submodular-
ity property required to define a count matroid. We have E(G+v1 v2 ) = S1 ∪S2 and S1 ∩S2 = {v1 v2 }.
By considering the subgraphs induced by S1 , S2 , S1 ∪ S2 and S1 ∩ S2 shown in 22.8 we may deduce
that ft (S1 ) = 2, ft (S2 ) = 5, ft (S1 ) ∪ S2 ) = 7 and ft (S1 ∩ S2 ) = 1. This implies that ft is not an
intersecting submodular function.
Notice however that the set function ft does provide the following necessary condition for a set
of edges in a graph G to be independent in MPL (G).
Lemma 22.3.1 Let G = (VP ,VL , E) be a point-line graph and F ⊂ E. If F is independent in MPL (G)
then |F 0 | ≤ ft (F 0 ) for all ∅ 6= F 0 ⊆ F.
We will use the idea of introducing new edges between pairs of line-vertices which are not
already connected by an edge to obtain a stronger necessary condition for independence in MPL (G),
Lemma 22.3.2 below. This condition is given in terms of a set function which is submodular and
hence is a candidate to induce MPL (G) as a count matroid.
Suppose G = (VP ,VL , E) is a point-line graph. If a subgraph Hi ⊆ G is generically rigid then we
may add a spanning tree of edges on VL (Hi ) to G and remove the same number of edges in EPP (Hi )∪
EPL (Hi ) from G in such a way that we do not change the maximum number of independent edges
in MPL (G). We may do this simultaneously on all subgraphs Hi corresponding to a partition of E
to obtain an upper bound on the rank of MPL (G). The fact that some of the subgraphs Hi may not
be generically rigid means that we may incorrectly increase the count of the maximum number of
independent edges and hence in this way we get only an upper bound on the maximum number of
independent edges in MPL (G). Minimizing this bound over all partitions gives an even better upper
bound. We used this idea in [9] to obtain the following necessary condition for independence in
MPL (G).
498 Handbook of Geometric Constraint Systems Principles
v3 u1

(a) (b)
v1 v2

v1 v2
u2
G1 G2

v3

u1
(c) (d)

v1 v2 v1 v2

u2
G1 ∪ G2 G1 ∩ G2

Figure 22.8
Point-line graphs G1 and G2 together with the graphs G = G1 ∪ G2 and G1 ∩ G2 induced respectively
by the union and intersection of their edge sets.

Lemma 22.3.2 Let G = (VP ,VL , E) be point-line graphs. For F ⊆ E let

( s )
X
ρ(F) = min (2νP (Fi )| + νL (Fi ) − 2)
i=1

where the minimum is taken over all partitions {F1 , . . . , Fs } of F, and let
f (F) = ρ(F) + νL (F) − 1.
If F 6= ∅ and F is independent in MPL (G) then |F| ≤ f (F).
It is straightforward to check that, if VP (F) = ∅, then the optimal partition has s = |F| which gives
ρ(F) = 0 and f (F) = ν(F) − 1. Otherwise the partition with s = 1 shows that ρ(F) ≤ 2νP (F) +
νL (F) − 2 and hence f (F) ≤ 2ν(F) − 3. Thus f (F) ≤ ft (F) for all F ⊆ E.
For the graph in Figure 22.7(a) we consider F = E and the partition E = {F1 , F2 , F3 } where
F1 = {v1 v3 }, F2 = {v2 v3 } and F3 = {u1 v1 , u2 v1 , u1 v2 , u2 v2 , u1 u2 }. This partition gives ρ(E) ≤ 4 and
f (E) ≤ 6. Since |E| = 7 this shows that E is not independent in MPL (G).
Similar calculations for the four graphs in Figure 22.8 give f (S1 ) = 2, f (S2 ) = 5, f (S1 ∩ S2 ) = 1
and f (S1 ∪ S2 ) = 6. Hence f is submodular on these edge sets. Indeed, a theorem of Dunstan [7]
implies that the function ρ is submodular on E and hence f is also submodular on E. We can also
show that f is nondecreasing, and nonnegative on the nonempty subsets of E, see [9, Lemma 3.5].
Thus f induces a count matroid M( f ) on E, and Lemma 22.3.2 tells us that independence in M( f )
is a necessary condition for independence in MPL (G).
We will see that MPL (G) = M( f ). The following properties of M( f ) will be useful. They can
be derived using standard matroid arguments, see [9].
Let M1 and M2 be matroids on the same ground set E.
Point-Line Frameworks 499

Matroid union: The matroid union of M1 and M2 is the matroid M1 ∨ M2 on E defined by the
condition that a set I ⊆ E is independent in M1 ∨ M2 if and only if E is the disjoint union of
I1 and I2 where I1 is independent in M1 and I2 is independent in M2 .

Lemma 22.3.3 Let G = (VP ,VL , E) be a point-line graph. Then

(a) a set S ⊆ E is independent in M(ρ + νL − 1) if and only if S + i j is independent in M(ρ + νL )
for all i j ∈ S,
(b) M(ρ + νL ) = M(2νP + νL − 2) ∨ M(νL ).

Example We can use Lemma 22.3.3 to show that the edge set of the naturally bipartite point-line
graph K(3, 3) is independent in M(ρ + νL − 1). By symmetry, it suffices to show that E + i j is
independent in M(2νP +νL −2)∨M(νL ) for any fixed edge i j ∈ E. Let S be any 1-factor of K(3, 3)
which contains i j and T = (E + i j) − S. Then it is straightforward to check that T is independent in
M(2νP + νL − 2) and S is independent in M(νL ). Hence E + i j = S ∪ T is independent in M(2νP +
νL − 2) ∨ M(νL ).
We will use matroid theory and some elementary linear algebra to characterize independence in
MPL (G) in the next section, in the special case when G is naturally bipartite. This will allow us to
use Lemma 22.3.3 to deduce that MPL (G) = M(ρ + νL − 1) for any naturally bipartite point-line
graph G. Then in Section 22.3.3 we will use the substitution of edges in EPP and ELL with copies of
naturally bipartite K(3, 3) graphs to extend this result to all point-line graphs.

22.3.2 A Characterization of Independence in MPL (G) when G is Naturally Bipar-

tite
Let G = (VP ,VL , E) be a naturally bipartite point-line graph. Let (G, p) be a corresponding point-
line framework and let a = (a1 , . . . , a|VL | ) denote the vector of line slope coordinates. We will first
show that, if the point coordinates p(ui ) for ui ∈ VP are algebraically independent over Q(a), then
the linear independence of the rows of R(G, p) is closely related to the linear independence of the
rows of a simpler matrix A(G, a).
Let c = (c1 , . . . , c|E| ) ∈ R|E| be a vector of generic edge weights and let A(G, a, c) be the matrix
obtained by replacing the entries −x j ak − y j in the columns for vk,a in R(G, p) by c jk for each
e = u j vk ∈ E. We will refer to A(G, a, c) as the frame matrix corresponding to R(G, p). Note that
this frame matrix does not depend on the coordinates of the point-vertices but only on the slopes of
the lines (and on the parameters c).

Lemma 22.3.4 Let G = (VP ,VL , E) be a naturally bipartite point-line graph, let (G, p) be a cor-
responding point-line framework, let a = (a1 , . . . , a|VL | ) denote the vector of line slope coordinates
in p. Suppose that the coordinates p(ui ) = (xi , yi ) of the point-vertices ui ∈ VP are algebraically
independent over Q(a). Then the following statements are equivalent.
(a) The rows of A(G + e0 , a, c) are independent for all e0 = ui v j with e = ui v j ∈ E.
(b) The rows of R(G, p) are independent.

Proof. (a) =⇒ (b). Suppose that (a) holds. We can represent each vector in the null space of
A(G, a, c) as (q, h) where q : VP → R2 and h : VL → R2 . Let q(uk ) = (qk,1 , qk,2 ) for all uk ∈ VP
and h(vk ) = (hk,1 , hk,2 ) for all vk ∈ VL . Choose e = ui v j ∈ EPL . Then rank A(G, a, c) = rank A(G +
e0 , a, c) − 1 so we can find a (q, h) ∈ Null A(G, a, c) such that h j,1 6= 0.
Repeating this argument for each e ∈ E and taking a suitable linear combination of the vectors
we obtain, we can construct a (q, h) ∈ Null A(G, a, c) such that h j,1 6= 0 for all v j ∈ VL . The fact that
(q, h) ∈ Null A(G, a, c) gives

qi,1 − a j qi,2 + ce h j,1 − h j,2 = 0 for all e = ui v j ∈ E. (22.1)

500 Handbook of Geometric Constraint Systems Principles

Construct a point-line framework (G, p) by putting p(ui ) = (−qi,2 , qi,1 ) for all ui ∈ VP and p(v j ) =
(a j , 0) for all v j ∈ VL . Equation (22.1) enables us to transform the rigidity matrix R(G, p) to the
A-matrix A(G, a, c) by subtracting h j,2 times column v j,2 from column v j,1 for all v j ∈ VL , and
then dividing column v j,1 by h j,1 . This implies that rank R(G, p) = rank A(G, a, c) = |E|. It follows
that the rows of the rigidity matrix of any realization of G as a point-line framework with generic
coordinates for the point-vertices will be linearly independent. Hence (b) holds.
(b) =⇒ (a). Suppose (b) holds. The specialization c0e = −x j ak − y j for e = jk ∈ EPL shows
that the rows of A(G, a, c) are independent. The vector (q, h) with qk,1 = −yk , qk,2 = xk , hk,1 = −1
and hk,2 = 0, which corresponds to an infinitesimal rotation of p about the origin, is in the nullspace
of R(G, p). This vector is also in the nullspace of A(G, a, c0 ) for the specialization c0 given above.
However (q, h) is not in the nullspace of A(G + e0 , a, c0 ) for any e ∈ E provided we choose c0e0 6= c0e .
Hence rank A(G + e0 , a, c0 ) > rank (A(G, a, c0 )) for all e ∈ E which implies (a). •

We next show that the row matroid of A(G, a, c) can be expressed as the matroid union of two
count matroids, one of which is the fixed slope rigidity matroid of G. We will use the following result
of Brylawski [4, Lemma 7.6.14(1)], see also [9, Lemma 4.1], which gives a linear representation for
the matroid union of two row matroids.

Lemma 22.3.5 Let Mi be an m × ni -matrix for i = 1, 2. Let X the m × m diagonal matrix with
entries xi in row i where xi are algebraically independent over Q(M1 , M2 ) and let (M1 , XM2 ) be the
m × (n1 + n2 )-matrix whose columns are the combination of the columns of M1 and the columns of
XM2 . Then M(M1 , XM2 ) = M(M1 ) ∨ M(M2 ).

Lemma 22.3.6 Let G be a naturally bipartite graph, let (G, p) be a framework with line slopes
a and let A(G, a, c) be the corresponding frame matrix. Then M(A(G, a, c)) = M f ixed (G, a)) ∨
M(νL ).

Proof. Let RL be the |E| × |VL |-matrix with entry 1 in the row for ui v j and column for v j and 0
elsewhere. Then RL is the (0, 1)-incidence matrix for a bipartite graph and it is well-known that
M(RL ) = M(νL ).
Let X be the |E| × |E| diagonal matrix with generic parameters ce for the diagonal entry
corresponding to e ∈ E. After a reordering of rows and columns, the matrix A can be written
as A(G, a, c) = (R f ixed (G, a), XRL (G)). Then M(A(G, a, c)) = M f ixed (G, a) ∨ M(νL ) by Lemma
22.3.5. •

Lemmas 22.3.4 and 22.3.6 together with the result that M f ixed (G) = M(2ν p + νL − 2) stated at
the end of Section 22.2.5 give the following characterization of independence in MPL (G) when G
is naturally bipartite.

Lemma 22.3.7 Let G = (VP ,VL , E) be a naturally bipartite point-line graph, let (G, p) be a frame-
work on G and let a = (a1 , . . . , a|VL | ) be the vector of line-slopes. Suppose that the coordinates
p(ui ) = (xi , yi ) of the point-vertices ui ∈ VP are algebraically independent over Q(a). Let S ⊆ E.
Then
(a) S is independent in MPL (G, p) if and only if S + ui v j is independent in M f ixed (G + ui u j , a)) ∨
M(νL ) for all ui v j ∈ E.
(b) S is independent in MPL (G) if and only if S + ui v j is independent in M(2νP + νL − 2) ∨ M(νL )
for all ui v j ∈ E.
Point-Line Frameworks 501

22.3.3 A Characterization of Independence in MPL (G)

We described in Section 22.2 how any point-line graph G can be converted to a naturally bipartite
point-line graph G+ by replacing all edges in EPP and ELL with copies of the naturally bipartite
point-line graph K(3, 3). We showed in the example following lemma 22.3.3 that the edge set of
K(3, 3) + i j is independent in M(2νP + νL − 2)) ∨ M(νL ) for any edge i j of K(3, 3). Lemma 22.3.7
now implies that the edge set of K(3, 3) is independent in the generic point-line rigidity matroid.
It is straightforward to check that 2|V (G)| − |E(G)|=2|V (G+ )| − |E(G+ )|. A more detailed anal-
ysis given in [9] shows that E(G) is independent in MPL (G) if and only if E(G+ ) is indepen-
dent in MPL (G), and E(G) is independent in M(ρ + νL − 1) if and only if G+ is independent in
M(ρ + νL − 1).
Together with lemmas 22.3.3 and 22.3.7(b), this gives the following charactarization for inde-
pendence in the rigidity matroid of any point-line graph

Theorem 22.1 Let G = (VP ,VL , E) be a point-line graph and F ⊆ E. The following statements are
equivalent:
(a) F is independent in M(ρ + νL − 1);
(b) for all i j ∈ F, F + i j is independent in M(2νP + νL − 2)) ∨ M(νL );
(c) F is independent in MPL (G).

Property (b) in Theorem 22.1 leads to a polynomial-time algorithm to determine a maximal

independent subset of MPL (G), for any point-line graph G, see [9]. An implementation of the
algorithm as a two-color pebble game is given in an arXiv version of this paper [10].

22.3.4 The Rank Function for MPL (G)

We used Theorem 22.1 to obtain the rank function rPL of MPL (G) for any point-line graph G in
[9]. For any F ⊆ E and any partition F = {F1 , . . . , F|F | } of F we define a graph GF as follows. The
vertices vi of GF correspond to the partition sets Fi and vi v j ∈ E(GF ) if and only if VL (Fi )∩VL (Fj ) 6=
∅. We let cF denote the number of connected components of GF .
Theorem 22.2 Let G = (VP ,VL , E) be a point-line graph and F ⊆ E. Then
 
X 
rPL (F) = νL (F) + minF (2νP (Fi ) + νL (Fi ) − 2) − cF ,
 
Fi ∈F

where the minimum is taken over all partitions F of F.

When |VP | = 0, the minimum in the above expression for rPL (F) occurs when the sets Fi are
all single edges. In this case cF = cF is just the number of connected components in the subgraph
induced by F and rPL (F) = ν(F) − cF is the rank function of the graphic matroid of G.
When |VL | = 0, we have cL (F) = |F| for all F and
 
X 
rPL (F) = minF (2ν(Fi ) − 3)
 
Fi ∈F

is the rank function of the 2-dimensional bar-joint rigidity matroid of G.

502 Handbook of Geometric Constraint Systems Principles

22.4 Extensions to Rd
22.4.1 Point-Line Frameworks in Rd
A line in Rd is uniquely determined by two points and each of these two points can be anywhere on
the line. Hence a line is represented by 2(d − 1) coordinates. If we choose these coordinates to be
the d − 1 components of the line direction and the d − 1 coordinates of the intersection of the line
with a d − 1 dimensional linear subspace (a hyperplane) we can obtain a Jacobean matrix with |E|
rows and d|VP | + 2(d − 1)|VL | columns which has a similar form to the matrix J(G, p) given above
for 2-dimensional frameworks. An additional consideration is that we may specify more than one
type of edge between two lines. For example if d = 3 we can specify both the angle between two
lines and the distance between them.
If VP = ∅ then E = ELL . If in addition d = 3 and there are no line-line distance constraints then
J(G, p) corresponds to the rigidity matrix for points on the surface of a sphere [25] and the generic
rigidity matroid is the same as the generic rigidity matroid for points in the plane.
If d = 3 and we apply the restictions that (a) all vertices occur in rigid subgraphs and (b) each
vertex has at most one neighbor which is not in its rigid subgraph, then the Jacobian matrix J(G, p)
corresponds to the rigidity matrix for a body and CAD framework [20].
A characterization of the generic rigidity matroid for point-line frameworks in R3 on all point-
line graphs would give a characterization for bar-joint frameworks as a special case, which is a
significant unsolved problem. It is conceivable however that the study of naturally bipartite point-
line frameworks in R3 may be more tractable.

22.4.2 Point-Hyperplane Frameworks in Rd

The Jacobean matrix descibed in Section 22.4.1 does not reduce to a simpler form like R(G, p) when
d > 2. We do obtain a simpler rigidity matrix, however, if we replace the lines with planes when
d = 3 or more generally with subspaces of codimension 1 (hyperplanes) for any d. A hyperplane
can be represented by d parameters which we may take to be a unit normal vector n and a distance
b. Then the signed distance from a point with coordinates p j to a hyperplane with coordinates
(nk , bk ) is p j .nk − bk and the rigidity matrix has a form similar to R(G, p). Hence point-hyperplane
frameworks are a natural extension of point-line frameworks when d > 2.
Let VP and VH be the sets of point and hyperplane vertices, respectiviely. For any point-
hyperplane graph G, we can define both a generic rigidity matroid and a generic fixed-slope rigidity
matroid. Since both cases include the generic rigidity matroid for bar-joint frameworks as a spe-
cial case, we may conclude that a characterization of indepedence in either case will be difficult
when d > 2. If G is a naturally bipartite graph, however, then the fixed slope rigidity matrix is the
same as the incidence matrix for points and hyperplanes that is studied in scene analysis, see [22].
In particular, it follows from [22, Theorem 4.1] that the generic d-dimensional fixed-slope rigid-
ity matroid for a naturally bipartite point-hyperplane graph G with vertex partitions VP and VH is
M(d(νP − 1) + νH ) for all d ≥ 1, see [6] for more details.
The main result of Eftekhari et al. [6] gives a rigidity preserving transformation between point-
hyperplane frameworks and bar-joint frameworks. They show that, for a graph G = (V, E) and X ⊆
V , G has an infinitesimally rigid realization as a bar-joint framework in Rd in which the vertices in
X lie on a common hyperplane if and only if G has an infinitesimally rigid realization as a point-
hyperplane framework in Rd in which the vertices in X are represented as hyperplanes and the
vertices in V \ X are represented as points. Combining this result with Theorem 22.1, we obtain a
good characterization for when G can be realized as an infinitesimally rigid bar joint framework
References 503

in R2 in which the vertices in X are colinear. The special case when |X| = 3 had previously been
characterized in [8].

22.5 Direction-Length Frameworks

Instead of introducing lines as new objects in constraint graphs, we may adopt the approach de-
scribed in Section 22.1.2 and extend the usual bar-joint frameworks by introducing directional
constraints between pairs of points in the framework. For a pair of points p1 , p2 this introduces
a constraint of the form (p1 − p2 )/|p1 − p2 | = â12 where â12 is a unit vector describing a fixed
direction. If all of the constraints are direction constraints we have a direction framework. These
frameworks occur naturally in the theory of parallel redrawings and Whiteley [24] showed that the
matroid for a generic direction framework in R2 is isomorphic to the matroid for a generic distance
(bar-joint)framework on the same graph. Under this isomorphism the rotational isometry which is
satisfied by the distance framework is replaced by a scaling isometry for the direction framework.
A constraint configuration which allows both distance and direction constraints between pairs
of points is represented by an edge labelled graph for which all of the vertices represent points
and the edges are partitioned into two sets D and L which represent the direction and distance
(length) constraints, respectively. If the coordinates of the vertices are generic the rigidity of the
framework is determined by a rigidity matroid on the direction-length graph. Servatius and Whiteley
[18] characterized independence in this rigidity matroid by the following condition. A set F ⊆
D ∪ L is independent if and only if |F 0 | ≤ 2ν(F 0 ) − 2 for all nonempty subsets F 0 of F, with strict
inequality whenever F 0 ⊆ D or F 0 ⊆ L.
Although direction-length frameworks constrain the direction between pairs of points they do
this with respect to a fixed coordinate direction and cannot be used to constrain only the relative
angle between two pairs of points. In particular, there is no structure in the rigidity matroid for a
direction-length framework which represents the geometric fact that the three angles in a triangle
are dependent and sum to 180 degrees. It follows that a point-line framework cannot be represented
by a specialized direction-length framework.
On the other hand, any direction-length framework can be represented by a fixed-slope point-line
framework by adding a fixed slope line for each pair of points which are constrained by a direction
constraint, constraining this line to be at distance zero from the two points and deleting the direction
constraint. Note, however, that this substitution will not give a generic point-line framework even
if the initial direction-length framework is generic. Instead it gives rise to the type of quasi-generic
point-line framework discussed in Section 22.1.3. The fact that independence can be characterized
for the generic direction-length rigidity matroid indicates that we may also be able to characterize
independence for such quasi-generic point-line frameworks when we add the condition that each
line is incident with at most two points.

References
[1] Proceedings of Workshops on Automated Deduction in Geometry 1996-2016.
[2] L. Asimow and B. Roth, The rigidity of graphs, Trans. Amer. Math. Soc. 245 (1978), 279-289.
[3] W. Bouma, I. Fudos, C. M. Hoffmann, J. Cai, and R. Paige, A geometric constraint solver,
Computer-Aided Design, 27(1995) 487–501.
504 References

[4] T. Brylawski, Constructions, in Theory of Matroids, ed. N. White, CUP, London, 1986, 127–
223.
[5] S. C. Chou, Mechanical geometry theorem proving, D. Reidel Publishing company (1988).
[6] Y. Eftekhari, B. Jackson, N. Nixon, B. Schulze, S.-I. Tanigawa and W. White-
ley, Point-hyperplane frameworks, slider joints, and rigidity preserving transformations,
https://arxiv.org/pdf/1703.06844
[7] A. Frank, Connections in combinatorial optimization, Oxford Lecture Series in Mathematics
and its Applications, 38, Oxford University Press, Oxford 2011.
[8] B. Jackson and J. Jordán, Rigid two-dimensional frameworks with three collinear points,
Graphs and Combinatorics 21 (2005) 427–444.
[9] B. Jackson and J. C. Owen, A characterization of the generic rigidity of 2-dimensional point-
line frameworks, Journal of Combinatorial Theory, Series B 119 (2016) 96–121.
[10] B. Jackson and J. C. Owen, A characterization of the generic rigidity of 2-dimensional point-
line frameworks, http://arxiv.org/abs/1407.4675v1
[11] V. C. Lin, D. C. Gossard and R. A. Light, Variational geometry in computer aided design,
Proceedings of Siggraph (1981) 171–177.
[12] J. C. Owen, Algebraic solution for geometry from dimensional constraints, ACM Symposium
on Foundations in Solid Modeling (1991), 397–407.
[13] J. C. Owen, Constraints on simple geometry in two and three dimensions, J. Comput. Geom.
Appl. 6 (1996) 421.
[14] J. C. Owen and S.C.Power, Independence conditions for point-line-position frameworks,
preprint (2006).
[15] T. Pisanski, B. Servatius, Configurations from a graphical viewpoint, Birkhauser, 2013.
[16] I. E. Sutherland, Sketchpad: A man-machine graphical communication system, Phd Thesis,
M.I.T., Cambridge, MA., 1963.
[17] B. Servatius and H. Servatius, Combinatorial local rigidity of 2 dimensional bar and joint
frameworks, chapter in this book.
[18] B. Servatius and W. Whiteley, Constraining plane configurations in CAD:Combinatorics of
directions and lengths, SIAM J. Disc. Math. 12 (1999) 136–153.
[19] P. Todd, A k-tree generalisation that characterises consistency of dimensioned engineering
drawings, SIAM Journal of Disc. Math. 2 (1989) 255–261.
[20] A. St.John, Generic rigidity of body and cad frameworks, chapter in this book.
[21] W. Whiteley, The union of matroids and the rigidity of frameworks, SIAM J. Disc. Math. 1
(1988) 237–255.
[22] W. Whiteley, A matroid on hypergraphs with applications in scene analysis and geometry,
Disc. Comput. Geom. 4 (1989) 75–95.
[23] W. Whiteley, Matroids and rigid structures, in Matroid Applications ed. Neil White, Encyclo-
pedia of Mathematics and Its Applications 40 (1992) 1–51.
[24] W. Whiteley, Some Matroids from discrete applied geometry, in Matroid Theory AMS Con-
temporary Mathematics 197 (1996) 171–313.
[25] A. Nixon and W. Whiteley, Change of Metrics in Rigidity Theory, chapter in this book.
Chapter 23
Generic Rigidity of Body-and-Cad
Frameworks
Audrey St. John
Department of Computer Science, Mount Holyoke College, South Hadley, MA

CONTENTS
23.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
23.2 Algebraic Body-and-Cad Rigidity Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506
23.2.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506
23.2.2 Getting to Know Body-and-Cad Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
23.2.3 Formalization of the Algebraic Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
23.2.4 Building a 3D Body-and-Cad Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512
23.3 Infinitesimal Body-and-Cad Rigidity Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
23.3.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
23.3.2 The Pattern of the Rigidity Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
23.3.2.1 Primitive Angular and Blind Constraints . . . . . . . . . . . . . . . . . . . . . 515
23.3.3 Generic Rigidty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
23.4 Combinatorial Body-and-Cad Rigidity Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
23.4.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
23.4.2 The Rigidity Matroid and Sparsity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518
23.4.3 Characterizing Generic Body-and-Cad Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518
23.4.4 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519
23.5 Open Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523

23.1 Overview
Popular computer-aided design (CAD) software, such as SolidWorks or OnShape, enables users to
build designs for 2D and 3D systems, typically by describing and constraining geometric elements
such as points and lines. Underlying solvers check when a system is well-constrained (or minimally
rigid), under-constrained (or flexible) or over-constrained (containing unnecessary dependencies
that may or may not be consistent) to provide feedback to the user. These solvers generally rely
on combinatorial and numerical techniques, but are susceptible to performance issues, prompting
research questions related to speed and instability (especially near special positions). In addition,
capturing design intent is a key goal for CAD software and there is room to improve the feedback
given to the user. For example, simply being told a system is over-constrained is not as useful as

505
506 Handbook of Geometric Constraint Systems Principles

being told how to resolve the dependencies. Users are not experts in geometric constraint systems,
and more intuitive feedback can help users capture and refine their design intent.
In this chapter, we describe the body-and-cad rigidity model, defined to capture the majority of
systems specified by this type of CAD software: a body-and-cad framework is composed of rigid
bodies with pairwise coincidence, angular or distance constraints between them [3]. The derivation
of the theory follows the work of [10, 12] and similarly considers frameworks in dimension d;
we
denote the number of degrees of freedom availabe to a d-dimensional rigid body by D = d+1 2 .
We start with the algebraic setting for rigidity in Section 23.2, walking through examples in
2D and 3D. The rigidity matrix and pure condition defined in Section 23.3 allow us to analyze the
first-order infinitesimal behavior. Section 23.4 describes the combinatorial counting property that
characterizes generic infinitesimal rigidity of a subset of body-and-cad frameworks [6]. Finally, we
conclude in Section 23.5 with discussion around open questions and related applications.
A note to the reader: This chapter is intended to be a “quick reference” guide with an emphasis
on building intuition for body-and-cad rigidity; the more technical underpinnings required for the
results are left in the References for conciseness. The theory relies on: linear algebra, combina-
torics, matroid theory (see, e.g., [8]), sparsity counts (see Chapter 19), and pebble game algorithms
[5].

23.2 Algebraic Body-and-Cad Rigidity Theory

23.2.1 Glossary
bicolored (multi)graph: A multigraph G = (V, E = R t B) with vertex set V = [1..n] and edge set
partitioned into two colors: red edges in R and black edges in B.
body-and-cad framework: F = (G, c, L1 , . . . , L|C| ) defined on a cad graph (G, c) and a family of
length functions L1 , . . . , L|C| (C depends on the embedding dimension), where Li specifies the
geometry of constraints associated with edges assigned color ci .
cad constraint: A constraint of type (coincidence, angle, or distance); imposed on geometric ele-
ments identified on a pair of bodies; the set of possible constraints is denoted by C and depends
on the embedding dimension.
cad graph: A pair (G = (V, E), c) of a multigraph along with a coloring function c : E → C (C de-
pends on the embedding dimension); vertices represent bodies, edges represent cad constraints
and c(e) colors edge e by its cad constraint label.
congruent: Two realizations of a framework are congruent if one can be obtained from the other
by multiplying each Ti by some T 0 ∈ SE(d).

flexible: A framework is (locally) flexible with respect to a realization T if there exists another
realization T0 in some neighborhood of T not congruent to T.
geometric element: A linear subspace (e.g., point, line, plane) rigidly affixed to a body.
internal degrees of freedom: The space of motions available to the system, modulo the trivial
degrees of freedom.
primitive angular constraint: A constraint that may block at most one rotational
degree of free-
dom; corresponds to one row in the rigidity matrix of Section 23.3 with 2 d2 nonzero entries
(in the “rotational” columns).
Generic Rigidity of Body-and-Cad Frameworks 507

primitive blind constraint: A constraint that may block at most one (rotational or translational)
degree of freedom; corresponds to one row in the rigidity matrix of Section 23.3 with 2D
nonzero entries.
primitive cad graph: A bicolored multigraph G = (V, E = R t B) associated to a body-and-cad
framework, where V is the set of rigid bodies, R is the set of red edges corresponding to prim-
itive angular constraints and B is the set of black edges corresponding to primitive blind con-
straints.
primitive constraint: A constraint that may block at most one degree of freedom; corresponds to
one row in the rigidity matrix of Section 23.3.
realization: An assignment T ∈ SE(d)n of reference frames for each body such that the constraint
equations determined by L1 , . . . , L|C| are satisfied, where F = (G, c, L1 , . . . , L|C| ) is a framework
with n bodies in dimension d.
rigid: A framework is (locally) rigid with respect to a realization T if all realizations in any neigh-
borhood of T are congruent.
trivial degrees of freedom: The D-dimensional space of trivial Euclidean motions (for the ambient
space Rd ).

23.2.2 Getting to Know Body-and-Cad Frameworks

The building blocks of a body-and-cad framework are fully dimensioned rigid bodies (see also
Chapter 20) with no internal degrees of freedom, such as a chair leg in 3D. Geometric constraints
are placed between pairs of rigid bodies by specifying geometric elements (e.g., points, lines or
planes) affixed to each body. We consider coincidence, angular and distance (abbreviated cad) con-
straints, which can be used to model the majority of constraints used in CAD software and refer to
constraints by the elements involved and the type of constraint; for instance, we could place a plane-
plane coincidence constraint to force the top surface of a chair leg to coincide with the bottom of
the chair’s seat. Note that, in CAD software, users often work in a 3D “assembly” environment,
which can be modeled as a body-and-cad framework with rigid bodies for the “parts” and geometric
constraints for the “mates.”
We now walk through a simple example to develop the foundations and terminology for body-
and-cad rigidity theory. We follow the same process experienced by a CAD user of adding con-
straints individually to help the user gain intuition about the main rigidity questions.

Example 46 Refer to Figure 23.1(a) throughout this example. Consider a book and a clock placed
on a table. We model this system in 2D, so each body has 3 degrees of freedom: horizontal trans-
lation, vertical translation and rotation. The framework as a whole has 6 degrees of freedom. We
are interested in the relative motion of bodies in the framework, so we fix the book’s position at the
origin to eliminate the framework’s 3 trivial degrees of freedom. We refer to the 3 remaining degrees
of freedom as the internal degrees of freedom and say that the framework is flexible.

• We specify constraint I to be a line-line perpendicular constraint. The clock can no longer

rotate relative to the book, and the framework now has two internal degrees of freedom.
Refer to Figure 23.1(b).
• Constraint II, a point-line coincidence constraint, results in the framework having one
internal degree of freedom; refer to Figure 23.1(c).
508 Handbook of Geometric Constraint Systems Principles

• To block this remaining internal degree of freedom, we can specify constraint III to be a
point-line distance constraint so that the resulting framework, depicted in Figure 23.1(d),
is well-constrained or minimally rigid.

4 4

A1
B 12 12
3 9 3
B
A A B1
9 3
2 12 3 2 12
6 3
Handbook B Handbook B
6

9 9 3
1 1
6
6

(0,0) 1 2 3 4 5 (0,0) 1 2 3 4 5

(a) A book and a clock each have 3 degrees of freedom (b) Constraint I requires a line `B1 on the clock to be
in the plane: horizontal translation, vertical translation perpendicular to a line `A1 on the book. The resulting
and rotation. Fixing the book at the origin eliminates framework has 2 internal degrees of freedom: the clock
the framework’s trivial degrees of freedom, leaving 3 can translate vertically and horizontally.
internal degrees of freedom: the clock can move freely.

4 4
12 12 12
A1
A2 B1
B B A1
A2 B1
B
3 9 9 3 3 3 9 3
pB 1 pB 1
A 6 6
A 6
2 2
Handbook Handbook

1 1

(0,0) 1 2 3 4 5 (0,0) 1 2 3 4 5

(c) Adding constraint II specifying the center point pB1 (d) A final constraint III requires the distance between
of the clock to lie on the short edge `A2 of the book the center point pB1 of the clock and the book’s line `A1
removes a degree of freedom. The resulting framework to be 4 and removes the last internal degree of freedom,
has 1 internal degree of freedom: the clock can translate resulting in a minimally rigid framework.
horizontally.

Figure 23.1
A 2D body-and-cad framework has two rigid bodies (a book A and a clock B) with pairwise con-
straints.

We capture the combinatorics of this framework in a cad graph, as depicted in Figure 23.2.
Now consider constraint IV, a line-line coincidence constraint. A framework with only this con-
straint has one internal degree of freedom: the clock can translate horizontally (see Figure 23.3(a)).
This highlights behavior that cad constraints may exhibit; unlike Constraints I, II and III, which
each block a single degree of freedom, Constraint IV blocks two degrees of freedom. To capture this
behavior, we associate a set of primitive constraints with each cad constraint; a primitive cad con-
straint blocks (at most) one degree of freedom and corresponds to a single row in the rigidity matrix
discussed in Section 23.3. We further refine the classification of a primitive constraint as a primi-
tive angular constraint, which blocks a rotational degree of freedom, or a primitive blind constraint,
which may block a rotational or translational degree of freedom. We make this distinction as an-
gular constraints require special consideration; notice that, while the framework defined with only
Constraint IV has 1 remaining degree of freedom, it cannot be eliminated with angular constraints.
The primitive cad graph associated with a cad graph is bicolored, with red edges for primitive angu-
Generic Rigidity of Body-and-Cad Frameworks 509

eI (line-line perpendicular) eI (line-line perpendicular)

A eII (point-line coincidence) B A eII (point-line coincidence) B

eIII (point-line distance) eIII (point-line distance)

(a) The cad graph has a vertex for each body and a (la- (b) The associated primitive cad graph: the line-line
beled) edge for each cad constraint. perpendicular cad constraint is associated to one prim-
itive angular constraint, while the point-line coinci-
dence and distance constraints are each associated to
one primitive blind constraint.

Figure 23.2
The combinatorics of the body-and-cad framework with Constraints I, II and III.

lar constraints and black edges for primitive blind constraints; refer to Figures 23.2(b) and 23.3(b).

4
12 12
B B
3
A2 B1
9 9 3 3 eIV (line-line
A A B
2
6 6
coincidence)
Handbook

eIV (line-line
A B
(0,0) 1 2 3 4 5
coincidence)
(a) Constraint IV requires the line `B1 on the clock to (b) The cad graph (top) and primitive cad graph (bot-
be coincident to the line `A2 on the book. The resulting tom) for a framework with Constraint IV; a line-line co-
framework has 1 internal degree of freedom: the clock incidence cad constraint is associated to one primitive
can translate horizontally. angular constraint and one primitive blind constraint.

Figure 23.3
A line-line constraint highlights the need to further refine cad constraints.

A framework with Constraints I, II, III and IV is over-constrained and consistent, as Constraint
IV is already implied by Constraints I and II; indeed, a framework with constraints III and IV
is minimally rigid. If, however, we specified constraint V to be a line-line coincidence constraint
between `B1 and `A1 in the framework with constraints I, II, and III, we would obtain an over-
constrained and inconsistent framework; there is no position that satisfies all constraints.

23.2.3 Formalization of the Algebraic Setting

We now state the formal definitions for the algebraic rigidity concepts introduced by this example.
Throughout this chapter, we work in dimensions 2 and 3, though the theory can be generalized fur-
ther. A body-and-cad framework in dimension d is composed of a set of rigid bodies with pairwise
cad constraints specified between geometric elemeents (e.g., a point, line or plane).
While we might intuitively picture a rigid body as a solid object, we associate a frame of ref-
erence (local coordinate system) to each body, as is standard in robotics and graphics approaches.
Then we can formally represent every rigid body i by assigning an element Ti ∈ SE(d), where
SE(d) is the special Euclidean group for dimension d containing rigid body motions; we represent
510 Handbook of Geometric Constraint Systems Principles

elements of SE(d) as (d + 1) × (d + 1) matrices. Since Ti gives the location and orientation of the
reference frame in the global coordinate system, we can compute global coordinates for any geomet-
ric element: left-multiplying a point in homogeneous coordinates by a T ∈ SE(d) gives transformed
coordinates in homogeneous coordinates (see, e.g., [9]). For a point p = (p1 , . . . , pd ) ∈ Rd , denote
its homogeneous coordinates by p b = (p1 : p2 : · · · : pd : 1); to convert back, we denote the extracted
Cartesian coordinates by p b = p. Then, if a point on body i has local coordinates pi ∈ Rd , Tg
e bi ∈ Rd
ip
computes its global coordinates.
The specific sets of geometric elements and cad constraints that we consider depend on the
dimension; to clarify definitions, we will specify the general theory to dimension 2 in this section.
Each geometric element is rigidly affixed to a body i and is represented with coordinates local
to body i’s reference frame: pi for a point or (pi , qi ) for the point-direction form of a line, with
pi , qi ∈ R2 . We denote the set of cad constraints by C and label each by the geometric elements
involved. In 2D, a geometric element is either a point or a line, and C = {point-point distance,
point-point coincidence, point-line distance, point-line coincidence, line-line distance, line-line
coincidence, line-line perpendicular, line-line parallel and line-line fixed angle}. We assume
an ordering of the elements of C and abuse notation by identifying C with [1..|C|]. For example,
assuming the ordering of 2D body-and-cad constraints just given, the point-line distance constraint
type is identified with index 3.
A cad graph is defined as a multigraph G = (V, E) along with a coloring function c : E → C,
where vertices represent rigid bodies and edges represent cad constraints. The coloring function
c(e) for e ∈ E “colors” the edge e with its cad constraint type. Denote by Ei = {e ∈ E|c(e) = i} the
edges of type i; e.g., E3 = {eIII } in Example 46.
A body-and-cad framework F = (G, c, L1 , . . . , L|C| ) is a cad graph (G, c) along with a family
of length functions L1 , L|C| , where Li specifies the geometry of constraints associated with edges
assigned color i. For instance, in 2D, the (squared) distance function L3 : E3 → R2 × (R2 × R2 ) × R
expresses a point-line distance constraint √ L3 (e) = (pi , (p j , q j ), δ ) by requiring a point with local
coordinates pi on body i to be a distance δ from a line defined by point p j and direction q j in local
coordinates on body j. In Example 46, if we assume the local reference frame for both bodies is
aligned with the global reference frame (i.e., the local frame’s origin is at global coordinates (0, 0)
and is not rotated), then L3 (eIII ) = ((4, 3), ((0, 0), (0, 1)), 16).
A realization of F in Rd is a specification T = (T1 , . . . , Tn ) ∈ SE(d)n of reference frames such
that the length functions are satisfied. That is, a realization is a solution to the system of equations
expressing the constraints; the (typically quadratic) equations themselves depend on the type of
constraint.
For example, the (squared) length function L3 for point-line distance constraints requires any
realization T to place the bodies in such a way that the point on one body is a specified distance
from the line on the other body. The distance can be expressed by projecting a vector v from the
v·w
point to the line onto a vector w perpendicular to the line: ||w|| .
We provide this 2D constraint equation in full detail for concreteness. Let e ∈ E3 and L(e) =
(pi , (p j , q j ), δ ), where pi gives the coordinates of the point relative to the reference frame of body
i, p j and q j the coordinates of a point on the line and its direction relative to the reference frame
of body j and δ the squared distance to be satisfied. If qi = (a, b), let q⊥ j = (b, −a) be a vector
perpendicular to q j . Thus, computing the global coordinates for a vector T j pbj − Tg g ip
bi from the point
] ⊥ perpendicular to the line, any realization T must satisfy:
to the line and for a vector T j qcj

] ⊥ )2
j pbj − Ti p
bi ) · T j qc
]
((T g
j
=δ (23.1)
] ⊥ ||2
||T j qcj
Generic Rigidity of Body-and-Cad Frameworks 511

Two realizations in Rd are congruent if they are related by a Euclidean motion, e.g., if one can
be obtained from the other by multiplying each Ti by some T 0 ∈ SE(d). If all realizations of a
framework are congruent, the framework is globally rigid. We are interested in the more intuitive
concept of local rigidity, defined by assuming the framework is given along with a realization T. If
all realizations of F in any neighborhood ∗ of T are congruent to T, then F is (locally) rigid (with
respect to T); otherwise, it is flexible.
For example, let F be a body-and-cad framework with the single point-line distance constraint
(Constraint III) from Example 46, i.e., L3 (eIII ) = ((4, 3), ((0, 0), (0, 1)), 16). Then T = (I3 , I3 ), where
I3 is the 3x3 identity  matrix, is arealization of F, depicted in Figure 23.1(d). For some parameter
1 0 0
t ∈ R, let T0 (t) = (I3 , 0 1 t ); T0 (t) also gives realizations of F, corresponding to translating
0 0 1
the position of the clock vertically by t. For instance, T0 (−2) is the realization obtained by vertically
translating the clock so that its center is at position (4, 1). Since, for any neighborhood of T, there
exists a realization T0 (t 0 ) (for some t 0 ∈ R) that is not congruent to T, the framework is flexible.
Note that, for the classical distance-based rigidity models of bar-and-joint or body-and-bar, min-
imal rigidity is defined in terms of removing any bar constraint, which block at most one degree of
freedom. However, as highlighted by Constraint IV in Example 46, cad constraints may block more
than one degree of freedom. Therefore, we define the analogous notion for body-and-cad in terms of
primitive constraints. For a cad graph G = (V, E), let the bicolored multigraph HG = (V, E 0 = R t B)
denote the primitive cad graph associated to G, where each e ∈ E is associated to a set of primitive
angular constraints Re ∈ R and primitive blind constraints Be ∈ B. The exact numbers of primitive
angular and primitive blind constraints associated to a cad constraint depend on the dimension; for
dimensions 2 and 3, refer to Tables 23.1 and 23.2. A rigid body-and-cad framework is minimally
rigid if the removal of any primitive constraint results in a flexible framework.

Table 23.1
Association of 2D body-and-cad (coincidence, angular, distance) constraints with the number of
primitive angular and blind constraints.

Point Line
Angular Blind Angular Blind
Point
Coincidence 0 2 0 1
Distance 0 1 0 1
Line
Coincidence 1 1
Distance 1 1
Parallel 1 0
Perpendicular 1 0
Fixed angular 1 0

∗ The definition of the metric space in which this neighborhood is defined is outside the scope of this chapter.
512 Handbook of Geometric Constraint Systems Principles

Table 23.2
Association of 3D body-and-cad (coincidence, angular, distance) constraints with the number of
primitive angular and blind constraints.

Point Line Plane

Angular Blind Angular Blind Angular Blind
Point
Coincidence 0 3 0 2 0 1
Distance 0 1 0 1 0 1
Line
Coincidence 2 2 1 1
Distance 0 1 1 1
Parallel 2 0 1 0
Perpendicular 1 0 2 0
Fixed angular 1 0 1 0
Plane
Coincidence 2 1
Distance 2 1
Parallel 2 0
Perpendicular 1 0
Fixed angular 1 0

23.2.4 Building a 3D Body-and-Cad Framework

We conclude with the final example for this section: building a dimmer wall plate with slider con-
trols.
Example 47 In 3D, each rigid body has 6 degrees of freedom (3 rotational and 3 translational), and
we can have planes as geometric elements. Figure 23.4 contains a CAD system for a dimmer wall
plate with 3 slider controls. The body-and-cad framework is composed of 4 parts, each modeled as
a rigid body: the plate A and slider controls B,C and D. There are 6 cad constraints, depicted in
Figures 23.5 and 23.6. We consider the process of adding the parts and constraints one at a time,
as a CAD user would do.
• With the plate A added and fixed in place with its lower left corner at the origin and its base
in the first quadrant of the xy-plane, we add the first slider control B. The framework has 6
internal degrees of freedom, as B can rotate and translate freely.
• After adding Constraint I, a plane-plane coincidence between A and B, the framework has
3 internal degrees of freedom: B can translate in the x- and y- directions and rotate about
the z-axis. Refer to Figure 23.5(a).
• Specifiying Constraint II as a line-plane coincidence between an edge on A and a face on B
achieves the intended design of a framework with 1 internal degree of freedom: B can only
translate in the y-direction. Refer to Figure 23.5(b).

• The second slider control C is added, initially free to move with its 6 degrees of freedom.
Constraint III, a plane-plane coincidence between B and C, is intended to be analogous
to Constraint I, resulting in C having 3 degrees of freedom: C can translate in the x- and
y- directions and rotate about the z-axis. However, due to the overhead of rotating the user
Generic Rigidity of Body-and-Cad Frameworks 513

C
B
B

z
z D
A y
A
y
x x
C
(0,0,0) (0,0,0)

(a) Assume that the base A is fixed at the origin. Then (b) The final framework (with six constraints) has 3 de-
the framework has 18 (internal) degrees of freedom, as grees of freedom, as each control can slide in the y-
each of the controls can rotate and translate freely with direction.
6 degrees of freedom.

Figure 23.4
A dimmer wall plate can be designed with 4 parts: a base A and three slider controls B,C, and D.

interface, the constraint is placed between B and C, as the tops of the sliders are easy to
select. The entire framework has 4 internal degrees of freedom. Refer to Figure 23.6(a).
• Constraint IV is analogous to Constraint II, specifying a line-plane coincidence between
an edge on A and a face on C. This achieves the intended design of a framework with 2
internal degrees of freedom: B and C can only translate in the y-direction. Refer to Figure
23.6(a).
• The last slider control D is added, initially free to move with its 6 degrees of freedom. Con-
straint V, a plane-plane coincidence between B and D is specified similarly to Constraint
III, resulting in D having the 3 degrees of freedom for translating in the x- and y- directions
and rotating about the z-axis. The entire framework has 5 internal degrees of freedom. Refer
to Figure 23.6(b).
• The last constraint is again chosen based on what happened to be easiest to select based
on the CAD user interface. Constraint VI places a plane-plane parallel constraint between
a face on C and a face on D. This achieves the final intended design of a framework with
3 internal degrees of freedom: each slider control B,C and D can only translate in the y-
direction. Refer to Figure 23.6(b).
The combinatorics (cad and primitive cad graphs) of this framework are depicted in Figure 23.7.

23.3 Infinitesimal Body-and-Cad Rigidity Theory

23.3.1 Glossary
(generically) dependent: A framework is (generically) dependent if its (generic) rigidity matrix
does not have full rank.
general: A realization of a generically independent framework is general if it does not lie on the
variety given by the pure condition.
514 Handbook of Geometric Constraint Systems Principles

B (0,0,0)

A x
y B

y
A
x

(0,0,0)

(a) The plane-plane coincidence Constraint I specifies (b) Constraint II is a line-plane coincidence constraint
that the bottom plane of A is coincident to the top plane restricting a line on A to lie in a plane on B.
of the base of B. The view of the CAD software must
be rotated to select the bottom plane of A. The resulting
framework has 3 degrees of freedom, as B can translate
in the x- and y-directions and rotation about z.

Figure 23.5
Two constraints result in a framework with 1 degree of freedom: B can slide in the y-direction.

generic rigidity matrix: The matrix M(G, x) on (a + b)|V | columns and |E| rows, where x maps
E to vectors of length a + b with indeterminates as entries; if e ∈ R, x maps to a vector with
indeterminates in the first a entries and 0 in the remaining b, where G = (V, E = B t R) is a
bicolored graph, and a and b are integers.
(generically) independent: A framework is (generically) independent if its (generic) rigidity ma-
trix has full rank.
(generically) infinitesimally flexible: A framework is (generically) infinitesimally flexible if its
(generic) rigidity matrix has rank < Dn − d.
infinitesimally minimally rigid: An infinitesimally rigid framework is infinitesimally minimally
rigid if the removal of any primitive constraint results in an infinitesimally flexible framework.
infinitesimal motion: A vector of length Dn in ker(M(G, r)), where M(G, r) is the rigidity ma-
trix for a framework (G, r), assigning an instantaneous motion for each body infinitesimally
preserving the constraints.
(generically) infinitesimally rigid: A framework is (generically) infinitesimally rigid if its
(generic) rigidity matrix has rank exactly Dn − d.
instantaneous motion: A vector s ∈ RD , representing an element from se(d), the Lie algebra as-
d
ω , v) with ω ∈ d (the rotational component) and v ∈ RD−(2) .


sociated with SE(d); s = (ω 2

pure condition: A polynomial expressing the determinant of a tied-down generic rigidity matrix
for a primitive cad graph with Dn − D edges.

rigidity matrix: The Jacobian of the algebraic system of cad constraint equations; for a framework
primitive cad graph G
= (V, E), the |E|×D|V | matrix M(G, r) has entries encoded by r : E → D,
where the last D − d2 entries of r(e) are 0 if e ∈ R.
Generic Rigidity of Body-and-Cad Frameworks 515

C C
B B

z z

y
A y
A
x x

(0,0,0) (0,0,0)

(a) Constraint III (orange) is a plane-plane coinci- (b) Similar to Constraint III, Constraint V (orange) is
dence constraint between the tops of B and C. Select- a plane-plane coincidence constraint between the tops
ing these planes is easier than following the design in- of B and D. Constraint V I (green) is a plane-plane dis-
tent of a constraint analogous to Constraint I between A tance constraint between a plane on C and a plane on
and C. Constraint IV (green) is a line-plane coincidence D; again, due to the CAD user interface, this constraint
constraint similar to Constraint II restricting a line on A was chosen instead of a constraint similar to Constraints
to lie in a plane on C. The resulting framework has 2 II and IV , which would have better captured the design
degrees of freedom: B and C can each slide in the y- intent. In fact, Constraint V I overconstrains the system
direction. in a consistent way, as it blocks D’s rotation about y,
already blocked by Constraint V .

Figure 23.6
Four additional constraints complete the design.

23.3.2 The Pattern of the Rigidity Matrix

As with other rigidity models, the infinitesimal rigidity theory captures the first-order behavior of
a body-and-cad framework F. The equations describing the algebraic rigidity, of the type shown in
Equation 23.1, are typically quadratic; solutions to the system are realizations T = (T1 , . . . , Tn ) that
assign an element from SE(d) to each of the n bodies in the framework. We refer to the Jacobian of
this system as the rigidity matrix.
For the technical setup of the infinitesimal rigidity theory, assume that T(t) : n → SE(d) is a
differentiable motion of the bodies and, for simplicity of analysis, assume that T(t0 ) = (ID , . . . , ID ),
where D = d+1


2 . Then elements in the kernel of the rigidity matrix can be interpreted as assign-
ing an instantaneous motionrepresented as an element from se(d), the Lie algebra associated with
SE(d), to each body. We coordinatize se(d) with vectors s ∈ RD . An instantaneous motion s con-
d d
tains a rotational component ω ∈ R(2) ; † let s = (ωω , v) with v ∈ RD−(2) . If S = (ss , . . . , s ) ∈ RDn
1 n
is in the kernel of the rigidity matrix of F, we call S an infinitesimal motion of F.
Every framework will always have infinitesimal motions corresponding to the instantaneous
trivial motions that assign the same element s 0 to each body. Thus, the kernel has dimension at least
D. If the only infinitesimal motions of F are the trivial ones, i.e., if the dimension of the kernel
is exactly D, we say the framework is infinitesimally rigid; otherwise, it is infinitesimally flexible.
Analogous to classical bar-and-joint rigidity [1], infinitesimal rigidity implies local rigidity.

23.3.2.1 Primitive Angular and Blind Constraints

A framework is infinitesimally rigid exactly when its rigidity matrix has rank Dn − D, prompting
closer analysis of the pattern of the matrix. Assume that the Dn columns are ordered with D for
each body i, corresponding to the coordinates of s i . The rows correspond to the linearized con-
straints of the framework; each expresses a linearized primitive cad constraint between bodies i and
j. If the constraint is blind, there may be D non-zero entries in the columns for body i, with their
† The component ω is drawn from the Lie algebra so(d) associated with the special orthogonal group SO(d) of rotations.
516 Handbook of Geometric Constraint Systems Principles

eI B
(plane-plane
coincidence) e eIII eV
II
(line-plane (plane-plane (plane-plane
coincidence) coincidence) coincidence)
eIV eVI
A (line-plane C (plane-plane D
coincidence) distance)
(a) The cad graph has 4 vertices and 6 edges.

B
plane-plane
coincidence
line-plane plane-plane plane-plane
coincidence coincidence coincidence

line-plane plane-plane
A coincidence C distance D
(b) The associated primitive cad graph has 4 vertices and 16 edges, one of which is redundant.

Figure 23.7
The combinatorics of the 3D body-and-cad framework for a dimmer wall plate with 3 slider
switches.

negation

in the columns for body j and zeroes elsewhere. If the constraint is angular, there may
be d2 non-zero entries in the columns corresponding to the rotational component ω i with their
negation in the columns corresponding to ω j and zeroes elsewhere (in particular, the columns for
vi and v j contain zeroes). A schematic for the pattern of the rigidity matrix highlighting the distinc-
tion between angular and blind constraints is depicted below. For 3D body-and-cad constraints, the
equations expressing the constraints can be developed directly in the infinitesimal setting using the
Grassmann-Cayley algebra; for full details, refer to [3]. The analogous development can be carried
through for 2D body-and-cad constraints.
The primitive cad graph G = (V, E = R t B) captures the combinatorics of the rigidity matrix,
with edges in R corresponding to red primitive angular constraint rows and edges in B corresponding
to black primitive blind constraint rows. We can represent the rigidity matrix via a function r : E →
RD that labels an edge e between vertices i and j of the primitive cad graph with a vector r(e) ∈ RD ;
the corresponding row contains r in the D columns for i and −r in the columns for j. If e ∈ R, the
last D − d2 entries of r(e) are 0.

For e ∈ R and f ∈ B with endpoints i and j, r(e) = (a1 , . . . , a(d ) , 0, . . . , 0) and r( f ) = (b1 , . . . , bD ),
2
a schematic of the rigidity matrix rows corresponding to e and f follows.

··· ωi vi ··· ωj vj ···

Primitive angular e ∈ R ··· a1 · · · a(d ) 0 ··· −a1 · · · − a(d ) 0 ···
2 2
Primitive blind f ∈ B ··· b1 · · · bD ··· −b1 · · · bD ···

The pair (G, r) is sufficient to represent the infinitesimal behavior of a body-and-cad framework,
and we denote the rigidity matrix by M(G, r).
To summarize terminology, if the rank of the rigidity matrix M(G, r) is:
• exactly Dn − D, the framework is infinitesimally rigid;
Generic Rigidity of Body-and-Cad Frameworks 517

• < Dn − D, the framework is infinitesimally flexible;

• exactly |E|, the framework is independent;
• < |E|, the framework is dependent

23.3.3 Generic Rigidty

We are interested in the “typical” behavior of a framework, behavior that almost all realizations
share. This is captured by the notion of generic infinitesimal rigidity, formally developed in [6]
with a subsequent presentation in [2]. To reduce terminology for this chapter, we do not provide
the full details here, instead summarizing the salient contributions. The main object of study is a
polynomial called the pure condition expressing the determinant of a tied-down generic rigidity
matrix; a framework is generically infinitesimally minimally rigid exactly when the pure condition
is not identically zero.
Given a primitive cad graph G = (V, E), the generic rigidity matrix M(G, x) is obtained by
replacing nonzero coordinates with indeterminates. Specifically, for each edge e ∈ E, x(e) maps
to a vector with D formal indeterminates if e ∈ B and d2 indeterminates if e ∈ R. The analogous

definitions from above follow. If the rank of the generic rigidity matrix M(G, r) is:
• exactly Dn − D, the framework is generically infinitesimally rigid;
• < Dn − D, the framework is generically infinitesimally flexible;
• exactly |E|, the framework is generically independent;
• < |E|, the framework is generically dependent
To remove the subspace of the kernel of a rigidity matrix corresponding to the D-dimensional
space of “trivial motions,” we can tie down a body i by appending D rows whose only non-zero
entries are specified by embedding the D × D identity matrix in the columns corresponding to body
i. Tying down a body ‡ in the generic rigidity matrix of a framework with primitive cad graph
G = (V, E), where |E| = Dn − D gives a square matrix, whose determinant expresses the body-
and-cad pure condition for all frameworks with the same underlying G. The pure condition is a
polynomial in the indeterminates used to generalize the entries of these frameworks’ rigidity matri-
ces; if it is identically zero, every framework with the combinatorics of G is infinitesimally flexible
(and dependent). For frameworks with a nonzero pure condition, we refer to a realization as gen-
eral (exhibiting generic behavior) if it does not lie on the variety given by the pure condition. An
approach for analyzing special (e.g., non-general) realizations is given in Chapter 4.

23.4 Combinatorial Body-and-Cad Rigidity Theory

23.4.1 Glossary
[a, b]-pebble game: An O(mn2 )-time algorithm for determining independence in the [a, b]-sparsity
matroid.
[a, b]-sparse: A bicolored graph G = (V, R t B) is [a, b]-sparse, where a and b are non-negative
integers, if there exists B0 ⊆ B such that (1) R ∪ B0 is (a, a)-sparse and (2) B \ B0 is (b, b)-sparse.
‡ The choice of tie-down does not matter [2].
518 Handbook of Geometric Constraint Systems Principles

[a, b]-tight: An [a, b]-sparse graph on n vertices is tight if has kn − k edges, where k = a + b.
(k, `)-sparse: A graph is (k, `)-sparse, where k and ` are non-negative integers with 0 ≤ ` < 2k, if
every set of n0 vertices spans at most max(0, kn0 − `) edges.
(k, `)-tight: A (k, `)-sparse graph on n vertices is tight if it has exactly kn − ` edges.
(generic) rigidity matroid for G: A set of edges is independent if the submatrix of M(G, r) (or
M(G, x) for the generic setting) given by the corresponding rows has ful rank.

23.4.2 The Rigidity Matroid and Sparsity

As with other rigidity models, the body-and-cad rigidity matroid of a framework is defined on the
edge set of its primitive cad graph G and infinitesimal constraint function r: a set of edges is in-
dependent if the submatrix of (G, r) given by the corresponding rows has full rank. The generic
rigidity matroid for a framework is defined in terms of the combinatorics of the framework’s prim-
itive cad graph G: a set of is independent if the submatrix of M(G, x) given by the corresponding
rows has full rank.
Notation. Throughout this section, there are several definitions that are modified with parame-
ters, e.g., [a, b]-sparsity and [a, b]-pebble game; the main term is what follows the parametrization.
Note also that both square brackets and parentheses appear in the parameters; square brackets (e.g.,
[a, b]-sparsity) are specific to behavior specific to the body-and-cad setting, while parentheses (e.g.,
(k, `)-sparsity) are consistent with notation standard in the field.
The pattern of M(G, x) leads to a sparsity counting condition, as for other rigidity models (re-
fer to Chapters 1, 18 and 20). A graph is (k, `)-sparse if every set of n0 vertices spans at most
max(0, kn0 − `) edges; a sparse graph is tight if it has exactly kn − ` edges. Results of Nash-Williams
[7] and Tutte [11] show that a graph is (k, k)-tight if and only if it is the edge-disjoint union of k
spanning trees. For more on (k, `)-sparse graphs, see Chapter 19.
M(G, x) has full rank if G is (D, D)-sparse. In addition, since the red edge set R corresponds
to rows in M(G, x) with additional zero entries, G0 = (V, R) (with only red edges) must satisfy
d
d

( 2 , 2 )-sparsity. This nested sparsity condition, while necessary, is not sufficient for rigidity, as
observed in [3]. Intuitively, one can consider independence in the rigidity matroid in terms of par-
titioning the rotational and translational degrees of freedom in a framework. This corresponds to a
partitioning of the primitive constraints as well, with angular constraints only assigned to blocking
rotational degrees of freedom. This perspective leads to the more general notion of [a, b]-sparsity;
note the square brackets make a distinction from the (k, `)-sparsity just discussed.
A bicolored graph G = (V, E = R t B) is [a, b]-sparse if there exists a B0 ⊆ B such that (V, R ∪ B0 )
is (a, a)-sparse and (V, B \ B0 ) is (b, b)-sparse; it is tight if, in addition, it has exactly (a + b)n − (a +
b) edges. By the results of Nash-Williams and Tutte [7, 11], his implies that G is tight if and only
if ∃B0 ⊆ B such that (V, R ∪ B0 ) is the edge-disjoint union of a spanning trees and (V, B \ B0 ) is the
edge-disjoint union of b spanning trees. Figure 23.8 depicts a [1, 2]-tight graph; in Figure 23.8(b),
the solid lines include both red edges and form a spanning tree, while the remaining black edges
form two edge-disjoint (dotted and dashed) spanning trees.

23.4.3 Characterizing Generic Body-and-Cad Rigidity

Independence of M(G, x) can be analyzed via the pure condition defined in Section 23.3, leading to
the following characterization for body-and-cad rigidity in terms of [a, b]-sparsity.
Theorem 23.1 ([6]) A body-and-cad framework F with underlying primitive cad graph G =
(V, E = R t B) is independent if and only if:
Generic Rigidity of Body-and-Cad Frameworks 519

1 4

2 3
(a) The graph has two red edges and 8 black edges. (b) The solid edges (including both red edges) form a
spanning tree; the dotted and dashed edges form two
other spanning trees.

Figure 23.8
A 1, 2-tight graph. Adding the bold black edge between vertices 2 and 4 to the red edge set results
in a (1, 1)-tight subgraph and a (2, 2)-tight subgraph.

• G is [1, 2]-sparse and F is embedded in dimension 2

• G is [3, 3]-sparse and F is embedded in dimension 2 with no point-point coincidence con-

straints
The omission of point-point coincidence constraints in 3D is due to the corresponding rows in the
rigidity matrix not following the generic pattern of M(G, x). In particular, placing two point-point
coincidence constraints between a pair of bodies results in a drop in the rank that does not appear
generically; the resulting framework is essentially the “double banana” (see Chapter 1, Figure 1.4).
The primitive cad graph for the minimally rigid 2D body-and-cad framework in Figure 23.2(b)
is [1, 2]-tight. The primitive cad graph for the flexible, but dependent, 3D body-and-cad framework
in Figure 23.7(b) is not [3, 3]-sparse; the graph with only red edges is not (3, 3)-sparse as it is a
circuit in the (3, 3)-sparsity matroid.

23.4.4 Algorithms
The [a, b]-sparsity condition is matroidal, leading to a greedy algorithm called the [a, b]-pebble
game for determining independence [2]. This pebble game uses the approach of Knuth for matroid
union [4], which generally requires “oracle” calls for determining independence in the (a, a)- and
(b, b)-sparsity matroids. By integrating and maintaining the (k, `)-pebble game algorithm for (k, `)-
sparsity [5], the [a, b]-pebble game achieves a faster time complexity of O(mn2 ). Since the [a, b]-
sparsity condition is matroidal, the pebble game algorithm finds maximum-sized independent sets,
allowing it to determine if the input graph contains a spanning tight graph as well as detecting
dependencies and outputting [a, b]-components (vertex-maximal induced [a, b]-tight subgraphs in a
graph that is not spannning a tight graph).
For containment, Algorithm 9 (reproduced from [2]) describes the [a, b]-pebble game algorithm.
As with other pebble games, one can view the pebbles as tracking degrees of freedom: aqua pebbles
for “angular” degrees of freedom and tan pebbles for “translational.” Each edge is covered by a
pebble, representing the degree of freedom “blocked” by that constraint. Observe that the black
edges may be covered by either an aqua or tan pebble, generalizing the notion of blind constraints
blocking either rotational or translational degrees of freedom, while red edges may only be covered
by aqua pebbles (angular constraints can block only rotational degrees of freedom).
520 Handbook of Geometric Constraint Systems Principles

Algorithm 9 The [a, b]-pebble game algorithm.

Input: A bi-colored graph G = (V, E = B t R), with black B and red R edges.
Output: [a, b]-sparsity property tight, sparse, dependent and contains spanning tight, or dependent.
Setup: Initialize an empty directed graph H on vertex set V . On each vertex, place a aqua pebbles and b tan
pebbles.
Allowed moves:
Add red edge i j [Precondition: ≥ a + 1 aqua pebbles on i and j.]
– Add the new edge, cover it with an aqua pebble from i (there is one by the precondition).
– Orient i j out of i.
Add black edge i j [Precondition: ≥ a + 1 aqua pebbles on i and j or ≥ b + 1 tan pebbles on i and j.]
– Add the new edge; cover it with a pebble from i using aqua (if there are a + 1 aqua) or tan (if there are
b + 1 tan).
– Orient i j out of i.
Edge reversal [Precondition: vertex j has a pebble on it and an in-edge i j covered by the same color.]
– Reverse the edge by orienting it as ji out of j, covering with the pebble from j and returning the (same
color) pebble originally covering i j to i.
Aqua exchange edge reversal [Precondition: vertex j has an aqua pebble on it and a black in-edge i j
covered by a tan pebble; i and j do not belong to the same (a, a)-component of aqua pebble covered
edges.]
– Reverse the edge by orienting it as ji out of j, covering with the aqua pebble from j and returning the
tan pebble originally covering i j to i.
Tan exchange edge reversal [Precondition: vertex j has a tan pebble on it and a black in-edge i j covered
by an aqua pebble; i and j do not belong to the same (b, b)-component of tan pebble covered edges.]
– Reverse the edge by orienting it as ji out of j, covering with the tan pebble from j and returning the
aqua pebble originally covering i j to i.
Method:
(a) For each edge e ∈ E
(a) If e is black: attempt to collect b + 1 tan pebbles on its endpoints with Alg. 10.
(b) If Alg. 10 returns true: insert e with an add black edge move.
(c) Else, or if e is red: attempt to collect a + 1 aqua pebbles on its endpoints with Alg. 10.
(d) If Alg. 10 returns true: insert e with an add black/red edge move.
(e) Else: reject it and highlight the edges returned by Alg. 10 as the fundamental circuit of the edge
(if e is black, this is the union of both calls to Alg. 10).
(b) If every edge is added: output tight if there are a + b pebbles left and sparse otherwise.
(c) Else, there were rejected edges: output dependent and contains spanning tight if there are a + b
pebbles left and dependent otherwise.
Generic Rigidity of Body-and-Cad Frameworks 521

Algorithm 10 The subroutine for finding pebbles for the [a, b]-pebble game.
Input: An [a, b]-pebble game configuration (a directed bi-colored graph), an edge e, and a desired additional
pebble color ce (aqua or tan).
Output: true if a + 1 aqua (if ce is aqua) or b + 1 tan (if ce is tan) pebbles can be collected on the endpoints of
e or false otherwise, along with the set of visited edges.
Method:
(a) Initialize set F = ∅.
(b) Initialize queue Q = ∅. Entries of Q will be of the form ( f , c), recording an edge on which to cover with
a pebble of color c.
(c) Set e.predecessor = NIL.
(d) Enqueue (e, ce ) into Q.
(e) While Q is not empty
(a) Dequeue ( f , c).
(b) If f 6= e and f is red, continue to the next iteration of the loop.
(c) Use the basic pebble game rules to try to collect a + 1 (if c is aqua) or b + 1 (if c is tan) pebbles
on the endpoints of f ; let F 0 be the set of edges visited by that search.
(d) If the pebbles were collected
i. Let g = f .
ii. While g.predecessor 6= NIL
A. Let d be the color of the pebble covering g, d be the opposite color, u and v the source
and target of g.
B. Collect a pebble of color d on v using the basic pebble game rules with edge reversal
moves.
C. Perform a d exchange edge reversal move to reverse the edge from v to u, covering it
with the d-colored pebble and releasing a d-colored pebble back onto u.
D. Set g = g.predecessor.
iii. Collect a + 1 (if c is aqua) or b + 1 (if c is tan) pebbles on the endpoints of g(= e).
iv. Output true and F ∪ F 0 .
(e) Otherwise
i. For each edge g ∈ F 0 that is not in F
A. Set g.predecessor = f ; let c be the opposite of color c.
B. Enqueue (g, c) into Q.
ii. Assign F = F ∪ F 0 .
(f) Output false and F.
522 Handbook of Geometric Constraint Systems Principles

A B
B A B

C
C
C

(a) A flexible 2D body-and-cad framework consisting (b) The well-known “triple banana” 3D bar-and-joint
of 3 bodies with the following 4 constraints: dashed framework spanning 12 joints is flexible as each “ba-
lines on A and B must be parallel; solid lines on A and nana” can rotate. This framework has a contextually
C must be parallel; two bars between bodies B and C fix rigid block {A, B, C} that is flexible as an induced
the distance between pairs of points. The contextually framework, since there are no constraints among A, B
rigid block {B,C} is flexible as an induced framework. and C.

Figure 23.9
Contextually rigid blocks highlight behavior that does not appear in 2D bar-and-joint and d-
dimensional body-and-bar rigidity models.

23.5 Open Questions

We conclude this chapter with a few open questions for body-and-cad rigidity that would improve
feedback for users of CAD software.
Contextually rigid systems. A rigid block is defined to be a set of bodies that do not contain any
relative degrees of freedom. Detecting rigid blocks can improve feedback when CAD users add
a dependent (and inconsistent) constraint. Most CAD software attempts to find a minimal set of
dependent (primitive) constraints (i.e., a circuit in the rigidity matroid) to give the user guidance
on how to resolve the inconsistency. Detecting such dependent sets can be achieved if rigid blocks
can be identified. In the well-understood rigidity models of 2D bar-and-joint and d-dimensional
body-and-bar frameworks, rigid blocks have the property that they remain rigid as induced frame-
works. However, a body-and-cad rigid component may be flexible as an induced framework; see
the example in Figure 23.9(a). We refer to such a component as contextually rigid, as its rigidity
depends on other constraints in the framework. This behavior is also exhibited by 3D bar-and-joint
frameworks; the “triple banana” contains a contextually rigid component on the three connecting
joints (see Figure 23.9(b)). The [a, b]-pebble game of Section 23.4.4 can detect rigid blocks that are
rigid as induced frameworks. However, it remains an open question to find an efficient algorithm
that detects contextually body-and-cad rigid blocks.
Dependent systems. Understanding these contextually rigid components may lead towards a com-
plete combinatorial characterization of the circuits of the [a, b]-sparsity matroid. While the structure
of some circuits is described in [2], it remains open to determine if the categorization is complete.
This has an impact on the algorithm for factoring the pure condition, which is the starting point for
developing approaches towards giving more intuitive feedback for overconstrained systems. Cur-
rent CAD software typically highlights a (not necessarily minimal) set of constraints that cannot be
resolved due to an dependency, without giving additional guidance or information to the user. Using
Cayley factorization to find a geometric interpretation of the dependency may lead to more useful
feedback for the user (e.g., “the design is overconstrained due to faces on Part 1 being parallel”);
for a case study, refer to Section 3 of [2].
Capturing design intent with equivalent systems. While CAD software strives to accurately cap-
ture the design intent of the user, the often cumbersome user interface can prevent the designer
References 523

from easily specifying constraints that match their intent. For example, the design intent for the 3D
dimmer wall plate of Example 47 is that each slider control should move with respect to (only) the
base plate. However, the combinatorics depicted in Figure 23.7 make it clear that the constraints
do not effectively capture this intent. Indeed, the tree-like combinatorics shown in 23.10, which
correspond to an equivalent body-and-cad framework, would better capture this design intent. This
highlights an open challenge for CAD software. Given a design, can the software suggest an equiv-
alent one that better captures design intent? For body-and-cad rigidity theory, this could be posed
as an open question to find a way of generating equivalent frameworks. Further, we seek “simple”
frameworks, which may better capture design intent, and may also lead to more efficient systems
for the embedded numerical solvers to process.

B B
plane-plane plane-plane
coincidence coincidence
line-plane line-plane
coincidence coincidence
plane-plane plane-plane
coincidence coincidence
A C A C
line-plane line-plane
coincidence coincidence

plane-plane line-plane
coincidence coincidence
line-plane plane-plane
coincidence coincidence
D D
(a) The cad graph has 4 vertices and 6 edges, with 2 (b) The associated primitive cad graph has 4 vertices
edges constraining each slider switch to the base. and 15 edges with no redundancies.

Figure 23.10
The combinatorics of an equivalent 3D body-and-cad framework that better captures the “tree-like”
design intent of a dimmer wall plate with 3 slider switches.

Acknowledgments. Research partially supported by NSF IIS-1253146. Table 23.2 reproduced from
[3], Figure 23.8 from [6], and Figure 23.9 from [2]. The author would like to thank Jessica Sidman
for her valuable and insightful feedback on this chapter.

References
[1] Leonard Asimow and Ben Roth. The rigidity of graphs II. Journal of Mathematical Analysis
and Applications, 68:171–190, March 1979.
[2] James Farre, Helena Kleinschmidt, Jessica Sidman, Audrey St. John, Stephanie Stark, Louis
Theran, and Xilin Yu. Algorithms for detecting dependencies and rigid subsystems for cad.
Computer Aided Geometric Design, 47:130–149, 10 2016.
524 References

[3] Kirk Haller, Audrey Lee-St.John, Meera Sitharam, Ileana Streinu, and Neil White. Body-
and-cad geometric constraint systems. Computational Geometry: Theory and Applications,
45(8):385–405, 2012.
[4] Donald E. Knuth. Matroid partitioning. Technical report, Stanford, CA, 1973.
[5] Audrey Lee and Ileana Streinu. Pebble game algorithms and sparse graphs. Discrete Math.,
308(8):1425–1437, 2008.
[6] Audrey Lee-St.John and Jessica Sidman. Combinatorics and the rigidity of cad systems.
Computer-Aided Design, 45(2):473–482, 2013.
[7] C. St. J. A. Nash-Williams. Edge-disjoint spanning trees of finite graphs. Journal London
Math. Soc., 36:445–450, 1961. Characterization graphs containing k edge-disjoint spanning
trees with counting properties (including (k, k)-sparsity) and partition results).
[8] J.G. Oxley. Matroid Theory. Oxford graduate texts in mathematics. Oxford University Press,
2011.
[9] J. M. Selig. Geometric Fundamentals of Robotics. Springer Publishing Company, Incorpo-
rated, 2nd edition, 2010.
[10] Tiong-Seng Tay. Rigidity of multi-graphs. I. Linking rigid bodies in n-space. Combinatorial
Theory Series, B(26):95–112, 1984.
[11] W. T. Tutte. On the problem of decomposing a graph into n connected factors. J. London
Math. Soc., 36:221–230, 1961.
[12] Neil White and Walter Whiteley. The algebraic geometry of motions of bar-and-body frame-
works. SIAM J. Algebraic Discrete Methods, 8(1):1–32, 1987.
Chapter 24
Rigidity with Polyhedral Norms
Derek Kitson
Dept. Math. Stats., Lancaster University, Lancaster, U.K.

CONTENTS
24.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
24.1.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
24.1.2 Statement of the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
24.2 Rigidity of Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
24.2.1 Points of Differentiability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
24.2.2 The Rigidity Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528
24.2.3 Framework Colors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529
24.2.4 Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530
24.2.5 Path Chasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530
24.2.6 Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531
24.3 Rigidity of Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535
24.3.1 Sparsity Counts and Tree Decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
24.3.2 Regular Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
24.3.3 Symmetric Isostatic Placements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
24.3.4 Symmetric Tree Decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 539
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 540

24.1 Introduction
The goal of this chapter is to present a streamlined introduction to rigidity theory for bar-joint
frameworks in Rd where the underlying metric is governed by a polyhedral norm (as opposed to the
Euclidean norm). Non-Euclidean rigidity theory, in which an alternative metric or quadratic form
is used to set geometric constraints, appears to be a relatively new topic. It has been considered
by various authors in the contexts of spherical and hyperbolic geometry and in pseudo-Euclidean
spaces (see [1, 3, 11, 12, 13]). However, the techniques used in each of these cases are different
from those required when the constraints are derived from a polyhedral norm. The theory presented
here is developed in [6, 4, 7, 8, 5]. For the purposes of exposition, proofs are omitted and the results
are not presented in their most general form. To give a broader context, note that we are working
within the geometry of finite dimensional normed linear spaces over R, known in the literature
as Minkowski spaces (not to be confused with Minkowski space-time). One benefit of working in
this setting is that we retain much of the interplay between real analysis and linear algebra that

525
526 Handbook of Geometric Constraint Systems Principles

underpins Euclidean rigidity theory. For further reading on Minkowski geometry see [9, 14] and
references therein.
A norm on Rd is polyhedral if its unit ball is a convex polytope. The most familiar examples of
polyhedral norms are the `1 norm and its dual the `∞ norm. To build a picture of how a framework
can “flex” in such a setting consider a two-dimensional connected framework with one designated
control node pinned at the origin. Suppose that all other nodes are constrained to move in a straight
line following one of only four directions: up, down, left and right. Note that these directions of
motion correspond to the four extreme points of the `1 unit ball. If the nodes of the framework
undergo a continuous motion, with the control node remaining fixed and all other nodes following
one of the four allowed directions, then Euclidean distances between adjacent nodes will in general
not be preserved. However, there may exist a continuous motion which preserves `∞ distances and
so it is natural to ask whether there is a rigidity theory for the `∞ norm which could be applied in
this context.
More generally, we could suppose that the nodes of the framework are constrained to move in
some finite number n of pre-determined directions. If each direction is represented by a vector of
Euclidean norm 1 then the absolutely convex hull of these vectors is the unit ball for a unique norm
on R2 . This norm, and its dual, are examples of polyhedral norms and play roles analogous to the
`1 and `∞ norms above. This motivates the development of a rigidity theory for general polyhedral
norms on Rd .

24.1.1 Glossary
In the following, G = (V, E) is a finite simple graph and k · k is a norm on Rd . The automorphism
group of G is denoted Aut(G) and the group of linear isometries of (Rd , k · k) is denoted Isom(Rd , k ·
k).
Polyhedral norm: A norm on Rd with the property that the closed unit ball B = {x ∈ Rd : kxk ≤ 1}
is a convex polytope.
Bar-joint framework: A pair (G, p) where p = (pv )v∈V ∈ (Rd )V and pv 6= pw for each edge vw in
G.
Rigidity map: The map fG : (Rd )V → RE , (xv )v∈V 7→ (kxv − xw k)vw∈E .
Well-positioned framework: A bar-joint framework (G, p) with the property that the rigidity map
fG is differentiable at p.
Continuous flex: For a bar-joint framework (G, p), a continuous flex is a continuous path in
fG−1 ( fG (p)) which passes through p.
Infinitesimal flex: For a well-positioned bar-joint framework (G, p), an infinitesimal flex is a vec-
tor which lies in the kernel of the differential d fG (p).

Rigid motion: A family of continuous paths αx : (−δ , δ ) → Rd , x ∈ Rd , with the property that
αx (t) is differentiable at t = 0, αx (0) = x for each x ∈ Rd and kx − yk = kαx (t) − αy (t)k for all
x, y ∈ Rd and all t ∈ (−δ , δ ).
Infinitesimal rigid motion: A vector field η : Rd → Rd derived from a rigid motion of Rd by the
formula η(x) = αx0 (0).
Rigid framework: A bar-joint framework (G, p) with the property that every continuous flex arises
as the restriction of a rigid motion to {pv : v ∈ V }.
Flexible framework: A bar-joint framework which is not rigid.
Rigidity with Polyhedral Norms 527

Isostatic framework: A rigid bar-joint framework (G, p) with the property that every bar-joint
framework obtained from (G, p) by removing a single edge from G is flexible.
Symmetric framework: A bar-joint framework (G, p) with a group action θ : Γ → Aut(G) and a
faithful group representation τ : Γ → Isom(X, k · k) such that τ(γ)pv = pθ (γ)v for all γ ∈ Γ and
all v ∈ V .

24.1.2 Statement of the Problem

Let X be a finite dimensional real vector space and let k · k be a norm on X.
(a) Given a framework (G, p) in (X, k · k) with symmetry group Γ, determine whether the frame-
work is rigid, isostatic or flexible.
(b) Given a graph G and a group action θ : Γ → Aut(G), determine whether there exists p ∈ X V
and a representation τ : Γ → Isom(X, k · k) such that (G, p) is isostatic and Γ-symmetric with
respect to θ and τ.
While these problems are stated for general normed spaces we will restrict our attention here to
polyhedral norms on Rd . The first problem regarding the rigidity of frameworks is considered in
Section 24.2. Initially it will be assumed the symmetry group Γ is trivial (in which case no prior
knowledge of symmetry is required). We will then consider non-trivial symmetry groups and for
this the reader may wish to refer to Chapter 25. The second problem regarding the existence of rigid
placements for graphs is considered in Section 24.3.

24.2 Rigidity of Frameworks

In this section we present methods for determining whether a given bar-joint framework is rigid
with respect to a polyhedral norm.

24.2.1 Points of Differentiability

A key feature of the Euclidean norm, from the perspective of rigidity theory, is that given any bar-
joint framework (G, p), the corresponding rigidity map fG is always differentiable at p. This allows
us to pass to the differential d fG (p) and to detect infinitesimal flexibility for any framework (G, p).
In contrast, for polyhedral norms there exist valid framework placements p which are points of
non-differentiability for the rigidity map fG and so to begin we make the following observation.

Theorem 24.1 Let (G, p) be a bar-joint framework in Rd and let k · kP be a polyhedral norm on
Rd . The following statements are equivalent.
(i) fG is differentiable at p.
(ii) For each edge vw ∈ E the vector pv − pw lies in the conical hull of exactly one facet of the
polytope P.

A framework (G, p) may be regarded as well-positioned for a particular choice of polyhedral

norm if the corresponding rigidity map fG is differentiable at p. In this case, the elements of the
kernel of the differential d fG (p) represent infinitesimal flexes of (G, p) and it can be shown that
528 Handbook of Geometric Constraint Systems Principles

(continuous) rigidity is equivalent to infinitesimal rigidity. Two convenient methods for detecting
rigidity emerge from this equivalence. The first is based on computing the rank of the differential
d fG (p) and for this we require an analogue of the Euclidean rigidity matrix. The second method is
to consider properties of an induced edge-labeling of the graph. This method is made possible by
the fact that the polyhedral rigidity matrix has only finitely many possible entries, derived from the
finitely many facets of the polytope P.

Example 48 Consider a degree 2 vertex in a framework (G, p) in (R2 , k · k1 ) as illustrated on the

left of Figure 24.1. The unit ball P in (R2 , k · k1 ) is indicated on the right. If the vertex v is pinned
then, under any motion of the framework which preserves the `1 distance between adjacent vertices,
the vertices a and b are constrained to lie on the indicated polytopes. Note that in this example the
rigidity map fG is not differentiable at p as the vector pv − pb lies in the conical hull of two different
facets of P.

1
a F1

v
b −1 1

F2
−1

Figure 24.1
A framework (G, p) in (R2 , k · k1 ) where p is a point of non-differentiability for the rigidity map fG .

Theorem 24.2 Let G be a graph and let k · kP be a polyhedral norm on Rd . If the rigidity map fG
is differentiable at p then,
(i) rank d fG (p) ≤ d|V | − d, and,
(ii) (G, p) is rigid in (Rd , k · kP ) if and only if rank d fG (p) = d|V | − d.

The appearance of the second d in the above rank formula needs some explanation as in the
Euclidean context the corresponding number is d(d+1) 2 . This number indicates the dimension of the
space of trivial infinitesimal motions which, in Euclidean space, can be attributed to d translational
motions and d(d−1)2 rotational motions. On replacing the Euclidean norm with a polyhedral norm,
rotations are no longer isometric and so do not induce trivial infinitesimal motions. We are left, in
this case, with only d translational flexes generating the space of all trivial infinitesimal motions.

24.2.2 The Rigidity Matrix

In order to compute the rank of d fG (p) we require a convenient matrix representation. Such a
representation is provided below in which non-zero entries are determined by the facets of the
polytope P. To describe these entries note that P is an intersection of half-spaces and so has the
form,
\n
P= {x ∈ Rd : x · F̂j ≤ 1}
j=1
Rigidity with Polyhedral Norms 529

for some F̂1 , . . . , F̂n ∈ Rd . Note that the facets of P have the form,

Fj = {x ∈ P : x · F̂j = 1}

and the vectors F̂1 , . . . , F̂n are the extreme points of the polar set P 4 ,

P 4 = {x ∈ Rd : x · y ≤ 1, ∀ y ∈ P}.

(For example, in the case of the `∞ norm on R2 we may take F̂1 = (1, 0), F̂2 = (0, 1) and their
negatives).

Theorem 24.3 Let G = (V, E) be a graph and let k · kP be a polyhedral norm on Rd . If the rigidity
map fG is differentiable at p then d fG (p) admits a matrix representation with rows indexed by E
and columns indexed by V × {1, 2, . . . , d}. The row entries for an edge vw ∈ E are,

v w
 
vw 0 ··· 0 F̂ 0 ··· 0 −F̂ 0 ··· 0 ,

where F is the unique facet of P with a conical hull that contains pv − pw and F̂ ∈ Rd is the
corresponding extreme point in the polar set P 4 .

Note that the above matrix representation of d fG (p) has the same size and the same identically
zero entries as the standard Euclidean rigidity matrix. It is unique up to the ordering of vertices and
edges.

24.2.3 Framework Colors

Let (G, p) be a bar-joint framework in (Rd , k · kP ) and let vw ∈ E be an edge of G. If pv − pw is
contained in the conical hull of a facet F of P then the pair of facets [F] = {F, −F} is referred to as
an induced framework color for the edge vw. If the rigidity map fG is differentiable at p then each
edge vw has exactly one framework color, denoted Phi(vw). A subgraph of G is monochrome with
framework color [F] if each edge in the subgraph has induced framework color [F]. The maximal
monochrome subgraph with framework color [F] is denoted GF .

Example 49 Consider the placement p of the complete graph K3 in (R2 , k·k∞ ) illustrated in Figure
24.2. The unit ball P is indicated on the left with representative facets labelled F1 and F2 . This
framework is well-positioned for the `∞ norm as each of the three vectors pa − pb , pb − pc and
pa − pc lies in the conical hull of exactly one facet of P (see Prop. 24.1). Each edge has exactly one
framework color,
Φ(ab) = [F1 ], Φ(ac) = [F2 ], Φ(bc) = [F2 ].
The monochrome subgraphs GF1 and GF2 are indicated below in black and gray, respectively. The
polyhedral rigidity matrix is,

a,1 a,2 b,1 b,2 c,1 c,2

 
ab 1 0 −1 0 0 0
bc  0 0 0 1 0 −1 
ac 0 1 0 0 0 −1

Note that the rigidity matrix has rank 3 which, by Theorem 24.2, indicates that (K3 , p) is flexible.
An evident flex is obtained by pinning a and b while translating c horizontally.
530 Handbook of Geometric Constraint Systems Principles
1 F2

F1
c(0, 2)
−1 1

−1
a(−1, 0) b(1, 0)

Figure 24.2
A placement for K3 in (R2 , k · k∞ ) with induced framework colors indicated in black and gray.

24.2.4 Connectivity
The following observation asserts that rigid frameworks satisfy a strong form of connectivity which
may be expressed in terms of the framework coloring.

Theorem 24.4 If (G, p) is rigid in (Rd , k · kP ) then G is connected and any subgraph obtained
from G by removing the edges of fewer than d maximal monochrome subgraphs is connected and
spanning in G.

To see why this is the case, consider a two-dimensional framework (G, p) with a maximal
monochrome subgraph GF , the removal of which disconnects the graph into two connected compo-
nents. A non-trivial flex of (G, p) is obtained by pinning one of these components while translating
the other component in a direction orthogonal to F̂. This argument also extends to d-dimensional
frameworks. The following example demonstrates that the strong connectivity condition alone is
not sufficient for rigidity.

Example 50 Consider the polyhedral norm on R2 given by

kxkP = |x · b1 | + |x · b2 | + |x · b3 |

where b1 = (1, 0), b2 = (0, 1) and b3 = (1, 1) and let (K3 , p) be the framework in (R2 , k · kP )
illustrated in Figure 24.3. The maximal monochrome subgraphs corresponding to the facets F1 , F2 ,
and F3 are indicated by black, gray and dashed lines, respectively. The polyhedral rigidity matrix
is,
a,1 a,2 b,1 b,2 c,1 c,2
 
(ab, F1 ) 2 2 −2 −2 0 0
(bc, F3 )  0 0 −2 0 2 0 
(ac, F2 ) 0 2 0 0 0 −2
Note that G satisfies the strong connectivity condition, however (K3 , p) is not rigid. An exam-
ple of a non-trivial flex, indicated by the arrows in Figure 24.3, is the continuous path α(t) =
(αa (t), αb (t), αc (t)) where αa (t) = (t, 0), αb (t) = (2, 2 + t) and αc (t) = (−1, 3).

24.2.5 Path Chasing

A collection of monochrome subgraphs H1 , . . . , Hk , with framework colors [F1 ], . . . , [Fk ], respec-
tively, is independent if F̂1 , . . . , F̂k are linearly independent in Rd .

Theorem 24.5 Let (G, p) be a well-positioned framework in (Rd , k · kP ) and let u = (uv )v∈V be an
infinitesimal flex of (G, p). If there exist d independent monochrome paths from a vertex v to a vertex
w then uv = uw .

Using the above theorem, rigidity can sometimes be detected by simple path chasing arguments
as illustrated in the following example.
Rigidity with Polyhedral Norms 531
1
F2 2
c(−1, 3)
F1
F3
b(2, 2)
−1 1
2 2

−1
2
a(0, 0)

Figure 24.3
A flexible bar-joint framework with three maximal monochrome subgraphs which satisfy the strong
connectivity condition.

Example 51 Consider the polyhedral norm defined for x = (x1 , x2 ) ∈ R2 by,


2|x2 | if |x1 | ≤ |x2 |
kxkP =
|x1 | + |x2 | if |x1 | ≥ |x2 |

and let (G, p) be the framework in (R2 , k · kP ) illustrated in Figure 24.2.5. The maximal
monochrome subgraphs induced by the facets F1 , F2 , and F3 are indicated by black, gray and dashed
lines, respectively, and the corresponding extreme points of the polar set P 4 are,

F̂1 = (1, 1), F̂2 = (0, 2), F̂3 = (−1, 1).

By computing the rank of the polyhedral rigidity matrix we see that (G, p) is rigid. Alternatively,
we may apply the following path chasing argument. There exist two independent monochrome paths
from vertex a to vertex d and so if u is an infinitesimal flex of (G, p) then ua = ud . Similarly, ud = ue ,
ue = u f and u f = uc . It follows that we may pin the vertices a, c, d, e, f and it only remains to note
that in this case b must also be pinned. Thus the space of infinitesimal flexes consists of translational
flexes only and so (G, p) is rigid.

Isostatic frameworks in (Rd , k · k∞ ) can be completely characterized in terms of maximal

monochrome subgraphs.

Theorem 24.6 Let (G, p) be a well-positioned framework in (Rd , k · k∞ ). The following statements
are equivalent.
(i) (G, p) is isostatic.
(ii) Each maximal monochrome subgraph is a spanning tree for G.

The proof of the above theorem combines the strong connectivity condition for rigid frameworks
and the path-chasing method.

24.2.6 Symmetry
For frameworks with a non-trivial symmetry group, rigidity may also be detected by considering
monochrome subgraph decompositions in an associated gain graph (G0 , ψ). This is particularly
useful when applying constructive methods to characterize graphs which admit rigid symmetric
placements. To illustrate the method, consider a framework (G, p) in (R2 , k · k∞ ) with half-turn
rotational symmetry. The induced monochrome subgraphs of G are themselves symmetric and so
induce two edge-disjoint subgraphs of G0 . These are referred to as the maximal monochrome sub-
graphs of G0 .
 532

1
2 F2 e

F3 F1 f d

−1 1
b
a c
− 12

a,1 a,2 b,1 b,2 c,1 c,2 d,1 d,2 e,1 e,2 f ,1 f ,2
 
(ab, F1 ) 1 1 −1 −1 0 0 0 0 0 0 0 0
(ae, F2 ) 
 0 2 0 0 0 0 0 0 0 −2 0 0 

(a f , F2 ) 
 0 2 0 0 0 0 0 0 0 0 0 −2 

(bc, F3 ) 
 0 0 −1 1 1 −1 0 0 0 0 0 0 

(bd, F1 ) 
 0 0 1 1 0 0 −1 −1 0 0 0 0 

(b f , F3 ) 
 0 0 −1 1 0 0 0 0 0 0 1 −1 

(cd, F2 ) 
 0 0 0 0 0 2 0 −2 0 0 0 0 

(ce, F2 ) 
 0 0 0 0 0 2 0 0 0 −2 0 0 

(de, F3 )  0 0 0 0 0 0 −1 1 1 −1 0 0 
(e f , F1 ) 0 0 0 0 0 0 0 0 1 1 −1 −1

Figure 24.4
An isostatic bar-joint framework in (R2 , k · kP ) as discussed in Example 51. The unit ball P is depicted on the left, the framework (G, p) with its three
induced monochrome subgraphs on the right and a rigidity matrix below. The labelling of the rows of the rigidity matrix indicates an edge of G and the
corresponding facet of P representing the framework color for that edge.
Handbook of Geometric Constraint Systems Principles
Rigidity with Polyhedral Norms 533

Theorem 24.7 Let (G, p) be a well-positioned framework in (R2 , k · k∞ ) with half-turn rotational
symmetry. If the rotation acts freely on the vertex set then the following are equivalent.
(i) (G, p) is rigid.
(ii) The maximal monochrome subgraphs of the gain graph both contain connected spanning
unbalanced map graphs.

A key idea in the proof of the above theorem is that unbalanced map graphs in the gain graph
correspond to subgraphs with symmetric components in the covering graph. In particular, a con-
nected unbalanced map graph corresponds to a connected symmetric subgraph. The result follows
from this observation and Theorem 24.6.

Example 52 Let (G, p) be the well-positioned framework in (R2 , k · k∞ ) illustrated in Figure 24.5.
This framework is symmetric under half-turn rotation and the rotation acts freely on the vertex
set. The maximal monochrome subgraphs of G, indicated in black and gray, respectively, are both
spanning trees and so (G, p) is rigid by Theorem 24.6. Alternatively, note that the induced maximal
monochrome subgraphs of the gain graph are both spanning and contain a single unbalanced cycle
and so (G, p) is rigid by Theorem 24.7.

1 F2
−1

F1
−1 −1
−1 1
−1

−1

Figure 24.5
An isostatic framework in (R2 , k · k∞ ) with half-turn rotational symmetry (center) and associated
gain graph (right). The induced monochrome subgraphs are indicated in black and gray.

Consider a framework (G, p) in (R2 , k · k∞ ) with reflectional symmetry in a coordinate axis.

The induced framework colors are preserved by the reflection and so induce a natural framework
coloring on the edges of the associated gain graph G0 . If the reflection acts freely on the vertex set
then rigidity is again characterized by considering the maximal monochrome subgraphs of G0 .

Theorem 24.8 Let (G, p) be a well-positioned framework in (R2 , k·k∞ ) with reflectional symmetry
in a coordinate axis. If the reflection acts freely on the vertex set then the following are equivalent.
(i) (G, p) is rigid.
(ii) The maximal monochrome subgraphs of the gain graph both contain connected spanning
unbalanced map graphs.

The characterization of rigidity rather than isostaticity in the previous two theorems is deliber-
ate. If the maximal monochrome subgraphs of the gain graph are themselves connected spanning
unbalanced map graphs this does not guarantee that the maximal monochrome subgraphs of the
covering graph are spanning trees. The following example illustrates this point. In fact, in the next
section it is observed that an isostatic framework with reflectional symmetry must contain a fixed
vertex and so any reflection framework with a free action on the vertex set will fail to be isostatic.
534 Handbook of Geometric Constraint Systems Principles

Example 53 Let (G, p) be the well-positioned reflection framework in (R2 , k · k∞ ) illustrated in

Figure 24.6. The maximal monochrome subgraphs of G both contain spanning trees and so (G, p)
is rigid by Theorem 24.6. Alternatively, note that the maximal monochrome subgraphs of the gain
graph are both connected, spanning unbalanced map graphs and so (G, p) is rigid by Theorem 24.8.

1 F2
−1

F1

−1 1
−1

−1
−1

Figure 24.6
A rigid reflection framework in (R2 , k · k∞ ) (center) and an associated gain graph (right). The max-
imal monochrome subgraphs of the gain graph are both spanning unbalanced map graphs.

A flex of a symmetric framework is symmetric if it preserves the symmetry of the framework.

A symmetric framework is symmetrically rigid if every symmetric flex of the framework is trivial.
Consider for example a Z2 -symmetric framework (G, p) in (Rd , k · kP ) with a free action on the
set of vertices. If the rigidity map is differentiable at p then the rigidity operator d fG (p) may be
expressed as a direct sum R1 ⊕ R2 where the kernel of R1 consists of the symmetric infinitesimal
flexes of (G, p). In the following V0 denotes the set of vertex orbits for G and E0 denotes the set of
edge orbits.

Theorem 24.9 Let (G, p) be a well-positioned reflection framework in (R2 , k · k∞ ). If the reflection
acts freely on the vertex set then the following are equivalent.
(i) (G, p) is symmetrically rigid.
(ii) rank R1 = 2|V0 | − 1.

The following matrix representation for R1 is known as an orbit matrix. To define it we must first
fix a choice of vertex orbit representatives. The corresponding gain for an edge orbit [e] is denoted
ψ[e] . For each vertex orbit [v], let ṽ denote the chosen representative vertex in G.

Theorem 24.10 Let (G, p) be a well-positioned and Z2 -symmetric framework in (Rd , k · kP ).

Then R1 admits a matrix representation with rows indexed by E0 and columns indexed by
V0 × {1, 2, . . . , d}.
If an edge orbit [e] = [vw] is not a loop then the corresponding row entries are,

[v] [w]
 
[e] 0 ··· 0 F̂ 0 ··· 0 −F̂ 0 ··· 0 ,

where F is the unique facet of P which has a conical hull that contains pṽ − pψ[e] w̃ .
If [e] is a loop at a vertex [v] then the row entries are,

[v]
 
[e] 0 ··· 0 2F̂ 0 ··· 0 ,

where F is the unique facet of P which has a conical hull that contains pṽ − p−ṽ .
Rigidity with Polyhedral Norms 535

A framework is symmetrically isostatic if it is symmetrically rigid and no proper symmetric

spanning subframework is symmetrically rigid. It is possible to determine when a framework in
(R2 , k·k∞ ) with reflectional symmetry in a coordinate axis is symmetrically isostatic by considering
the induced maximal monochrome subgraphs in the gain graph G0 . In this case the characterization
is that one of the maximal monochrome subgraphs of G0 is a spanning unbalanced map graph and
the other is a spanning tree. Which of the two is a map graph and which is a tree will depend on the
coordinate axis in which the reflection acts. This is illustrated in the example below.

Example 54 Consider the well-positioned reflection framework (G, p) illustrated in Figure 24.7.
Label the vertices of the gain graph a, b, c, d moving anti-clockwise and ending with the loop at d.
Then the polyhedral orbit matrix is,

a,1 a,2 b,1 b,2 c,1 c,2 d,1 d,2

 
(ac, F1 ) 1 0 0 0 −1 0 0 0
(bc, F1 ) 
 0 0 1 0 −1 0 0 0 

(ad, F1 ) 
 1 0 0 0 0 0 −1 0 

(dd, F1 ) 
 0 0 0 0 0 0 2 0 

(ab, F2 ) 
 0 1 0 −1 0 0 0 0 

(bc, F2 )  0 0 0 1 0 −1 0 0 
(cd, F2 ) 0 0 0 0 0 1 0 −1

Note that the rank of the orbit matrix is 7 and so, by Theorem 24.9, the framework is symmetrically
rigid. In fact, (G, p) is symmetrically isostatic as the removal of any orbit of edges results in a sym-
metric framework with a non-trivial symmetric flex. Note that the maximal monochrome subgraphs
in the gain graph consist of a spanning unbalanced map graph and a spanning tree. Also note that
the strong connectivity condition is not satisfied by the induced framework coloring on G and so
(G, p) is not rigid.

1 F2
−1

F1

−1 1

−1
−1

Figure 24.7
A symmetrically isostatic reflection framework in (R2 , k·k∞ ) (center) with an associated gain graph
(right).

24.3 Rigidity of Graphs

In this section we survey properties of graphs which are necessary, and in some cases also sufficient,
for the existence of a rigid placement, with or without symmetry.
536 Handbook of Geometric Constraint Systems Principles

24.3.1 Sparsity Counts and Tree Decompositions

The rank formula for rigid frameworks implies the following counting conditions must hold for any
finite simple graph which admits a well-positioned rigid or isostatic placement. These conditions
are analogous to Maxwell’s counting conditions for Euclidean frameworks.

Theorem 24.11 Let G be a finite simple graph.

(i) If there exists a rigid well-positioned placement of G in (Rd , k · kP ) then, |E| ≥ d |V | − d.
(ii) If there exists an isostatic well-positioned placement of G in (Rd , k · kP ) then |E| = d |V | − d
and |E(H)| ≤ d |V (H)| − d for each subgraph H.

If a graph satisfies the counting conditions in (ii) then it is said to be (d, d)-tight. The following
characterization provides a converse to Theorem 24.11(ii) in the case of polyhedral norms on R2 .

Theorem 24.12 Let G be a finite simple graph. Then, for all polyhedral norms k · kP on R2 , the
following statements are equivalent.
(i) There exists a well-positioned isostatic placement of G in (R2 , k · kP ).
(ii) G is (2, 2)-tight.

One method of proof for the above theorem is to use a construction scheme for (2, 2)-tight graphs
consisting of four types of graph move: 0-extensions, 1-extensions, vertex splitting and vertex-to-K4
moves. The base graph in this class is K4 . See Chapter 19 for a discussion of inductive constructions.
For all polyhedral norms on R2 the existence of a well-positioned isostatic placement of a graph
is also characterized by the following spanning tree property.

Theorem 24.13 Let G = (V, E) be a finite simple graph. The following statements are equivalent.
(i) There exists a well-positioned isostatic placement of G in (R2 , k · kP ).
(ii) G is expressible as a union of two edge-disjoint spanning trees.

It is not currently known whether the previous two theorems extend to d-dimensional frame-
works. However, it is known that graphs which are (d, d)-tight are precisely those which are ex-
pressible as an edge-disjoint union of d spanning trees. This is a result of Nash-Williams [10]. Thus
if either one of the above theorems extends to d-dimensional frameworks then they must both ex-
tend. Considering Theorem 24.6, it is sufficient for the `∞ norm to show that any decomposition
of G into d edge-disjoint spanning trees is realizable in the sense that the spanning trees are pre-
cisely the maximal monochrome subgraphs induced by some placement of the graph in Rd . This
has recently been proved for d = 2 (see [2]) but remains an open problem for d ≥ 3.

24.3.2 Regular Points

If a graph G admits an infinitesimally rigid placement in a d-dimensional Euclidean space then
almost all placements of G in this space will be rigid. For polyhedral norms, this is no longer the
case and so some care must be taken in adapting any Euclidean notions of genericity for polyhedral
norms.
A point p is a regular point for the rigidity map if fG is differentiable at p and the differential
d fG (x), regarded as a function of x, achieves its maximum rank at p. In this case (G, p) is said to
be a regular framework. If there exists an infinitesimally rigid placement for G then (G, p) must be
rigid for all regular points p and, in the Euclidean setting, these regular points form an open and
dense set in (Rd )|V | .
Rigidity with Polyhedral Norms 537

In the case of a polyhedral norm, the set of regular points of the rigidity map fG is still an
open set but it is no longer dense in (Rd )|V | . To see this note that a small perturbation of any well-
positioned placement of a graph will not alter the induced framework coloring. In particular, the
rigidity matrix will be unchanged. This situation is illustrated in the example below.

Example 55 Let (G, p) be the well-positioned bar-joint framework in (R2 , k · k∞ ) illustrated in

Figure 24.8. The maximal monochrome subgraph GF2 (indicated in gray) is not a spanning subgraph
of G and so (G, p) is flexible. The graph G is (2, 2)-tight and so, by Theorem 24.11, there exists
an isostatic placement of G in (R2 , k · k∞ ). By Theorem 24.2, the set of regular placements of G
is precisely the set of well-positioned rigid placements of G. Thus p, and any sufficiently small
perturbation of p, is a non-regular placement of G.

1 F2

F1

−1 1

−1

Figure 24.8
A non-regular, flexible placement of a (2, 2)-tight graph in (R2 , k · k∞ ).

A stronger notion of genericity sometimes used in Euclidean rigidity theory is to require not only
that (G, p) is regular but that every subframework of the complete framework (KV , p) is regular. Let
us refer to such a placement as completely regular. The set of all completely regular placements of
a graph in d-dimensional Euclidean space is always non-empty (in fact it is dense in (Rd )|V | ). This
is not the case for polyhedral norms as the following example shows.

Example 56 Consider a well-positioned placement p of the complete graph K6 in (R2 , k · k∞ ).

The induced framework coloring of the edges of K6 contains a monochrome subgraph GF which
itself must contain a copy of the complete graph K3 . The rigidity matrix for the subframework
(K3 , p) has rank 2. Since the regular placements of K3 have rigidity matrix with rank 3, (K3 , p) is
not regular. Thus there does not exist a well-positioned placement of K6 in (R2 , k · k∞ ) for which
all subframeworks are regular. More generally, it follows from Ramsey’s theorem that given any
polyhedral norm on Rd there is a complete graph for which no completely regular placements exist.

24.3.3 Symmetric Isostatic Placements

In the presence of a non-trivial symmetry group, a graph which admits an isostatic symmetric place-
ment must satisfy further counting conditions. We present below a general formula from which these
conditions can be derived. The set of vertices (respectively, edges) which are fixed by a symmetry
operation γ ∈ Γ is denoted Vγ (respectively, Eγ ).

Theorem 24.14 If there exists an isostatic, Γ-symmetric, well-positioned placement of G in (Rd , k ·

kP ) then,
|Eγ | = trace(τ(γ)) (|Vγ | − 1),
for each symmetry operation γ ∈ Γ.
538 Handbook of Geometric Constraint Systems Principles

Table 24.1
Necessary counting conditions for graphs which admit a placement as an isostatic framework in
(Rd , k · kP ) where d = 2 or 3.

Dimension
Symmetry
operation 2 3

1 |E| = 2|V | − 2 |E| = 3|V | − 3

s |Es | = 0 |Vs | ≥ 1, and, |Vs | = 1 iff |Es | = 0

i |Vi | = 1 and |Ei | = 0, or, |Vi | = 1 and |Ei | = 0, or,

|Vi | = 0 and |Ei | = 2, |Vi | = 0 and |Ei | = 3

C2 (as above) |VC2 | = 1 and |EC2 | = 0, or,

|VC2 | = 0 and |EC2 | = 1

C3 |VC3 | = 1 and |EC3 | = 0, or, |EC3 | = 0

|VC3 | = 0 and |EC3 | = 1

C4 |VC4 | ≤ 1 and |EC4 | = 0 |VC4 | = 1 and |EC4 | = 0

Cn , n ≥ 5 |VCn | = 1 and |ECn | = 0 |VCn | = 1 and |ECn | = 0

S3 n/a |VS3 | = 1 and |ES3 | = 0

S4 n/a |VS4 | = 1 and |ES4 | = 0, or,

|VS4 | = 0 and |ES4 | = 1

S5 n/a |VS5 | = 1 and |ES5 | = 0

S6 n/a |ES6 | = 0

Sn , n ≥ 7 n/a |VSn | = 1 and |ESn | = 0

Note: The left-hand column lists the possible symmetry operations. The counting conditions refer
to the number of vertices and edges which are necessarily fixed by a given symmetry operation.

In the case of 2- and 3-dimensional frameworks the necessary counts are listed in Table 24.1
for all possible symmetry operations. Standard notation is used: s denotes a reflection, i denotes
an inversion, Cn denotes an n-fold rotation and Sn denotes an improper rotation. In Example 52, a
C2 -symmetric isostatic placement of a graph in (R2 , k · k∞ ) is illustrated. Note that in this case the
graph has no fixed vertices and two fixed edges (one of two possible counting conditions listed in
Table 24.1). In general, these counts alone will not be sufficient to guarantee the existence of a rigid
placement with a particular symmetry group. However, as demonstrated in the next section, they
may be used to establish sufficient conditions.
Rigidity with Polyhedral Norms 539

In the presence of a free action on the vertex set, graphs which admit a symmetrically isostatic
placement with reflectional symmetry can be characterized in terms of sparsity counts on its asso-
ciated gain graph.

Theorem 24.15 Let G be a graph with a group action θ : Z2 → Aut(G) which acts freely on the
vertex set. The following are equivalent.
(i) There exists a well-positioned and symmetrically isostatic placement of G in (R2 , k · k∞ ) with
reflectional symmetry in a coordinate axis.
(ii) The gain graph is (2, 2, 1)-gain-tight.

The (2, 2, 1)-gain-tight condition states that the gain graph G0 satisfies |E(G0 )| = 2|V (G0 )| − 1,
each set F of edges in G0 satisfies |F| ≤ 2|V (F)| − 1 and each balanced set of edges in G0 satisfies
the stronger condition |F| ≤ 2|V (F)| − 2. Such graphs are constructible from a single unbalanced
loop using four types of graph move (see [5]).

24.3.4 Symmetric Tree Decompositions

The isometry group for the Euclidean plane is generated by uncountably many reflections and rota-
tions. The `∞ norm on R2 , in contrast, admits just eight linear isometries. As a result, we can group
the non-trivial symmetry operations for a framework in (R2 , k · k∞ ) into the following four types:
reflection in a coordinate axis, reflection in a diagonal line, half-turn rotation and four-fold rotation.
It is possible to characterize graphs which admit a symmetric isostatic placement (R2 , k · k∞ ) for
each of these types of symmetry operation.

Theorem 24.16 Let G be a finite simple graph. The following statements are equivalent.

(i) There exists a well-positioned isostatic placement of G in (R2 , k · k∞ ) with reflectional sym-
metry in a coordinate axis.
(ii) There exists a group action θ : Z2 → Aut(G) such that G is an edge-disjoint union of two
symmetric spanning trees and every edge orbit contains two distinct edges.

The class of Z2 -symmetric graphs which satisfy the conditions of the above theorem are con-
structible using four graph moves (see [8]). The smallest graph in this class is the wheel graph
W5 .

Example 57 Figure 24.9 illustrates a placement of the wheel graph W5 as an isostatic reflection
framework in (R2 , k · k∞ ). Note that the induced maximal monochrome subgraphs are edge-disjoint
symmetric spanning trees and each edge orbit contains two distinct edges.

1 F2

F1

−1 1

−1

Figure 24.9
A placement of the wheel graph W5 in (R2 , k · k∞ ) as an isostatic reflection framework.
540 References

The main reason that we need to distinguish between reflections in a coordinate axis and reflec-
tions in a diagonal line is that in the former case framework colors are preserved under the reflection
while in the latter case they are reversed. This seemingly minor difference actually results in two
very different classes of graph. In particular, graphs arising in the former case must contain exactly
one fixed vertex while those in the latter may contain any number of fixed vertices. In the following
a pair of edge-disjoint spanning trees is referred to as anti-symmetric if they are interchanged by the
action of the group.

Theorem 24.17 Let G be a finite simple graph. The following statements are equivalent.

(i) There exists a well-positioned isostatic placement of G in (R2 , k · k∞ ) with reflectional sym-
metry in a diagonal line.
(ii) There exists a group action θ : Z2 → Aut(G) such that G is an edge-disjoint union of two
anti-symmetric spanning trees and every edge orbit contains two distinct edges.

Much like a reflection in a coordinate axis, a half-turn rotation applied to a framework will
preserve the framework color of each edge. However, the graphs which admit isostatic placements
with half-turn symmetry may have either zero or two fixed edges and so form a strictly larger class
than that in Theorem 24.16.

Theorem 24.18 Let G be a finite simple graph. The following statements are equivalent.
(i) There exists a well-positioned isostatic placement of G in (R2 , k · k∞ ) with half-turn rotational
symmetry.
(ii) There exists a group action θ : Z2 → Aut(G) such that G is an edge-disjoint union of two
symmetric spanning trees and either all, or, all but two edge orbits contain two distinct edges.

Applying a four-fold rotation to a framework will reverse the induced framework color for each
edges. This is similar to the case of reflection in a diagonal line but again the class of graphs arising
in these two cases are very different.

Theorem 24.19 Let G be a finite simple graph. The following statements are equivalent.
(i) There exists a well-positioned isostatic placement of G in (R2 , k · k∞ ) with four-fold rotational
symmetry.
(ii) There exists a group action θ : Z4 → Aut(G) such that G is an edge-disjoint union of anti-
symmetric spanning trees and either all, or, all but two edge orbits contain four distinct edges.

References
[1] Victor Alexandrov. Flexible polyhedra in Minkowski 3-space. Manuscripta Math.,
111(3):341–356, 2003.
[2] K. Clinch and D. Kitson. Constructing isostatic frameworks for the `∞ -plane. Preprint, 2017.
[3] Steven J. Gortler and Dylan P. Thurston. Generic global rigidity in complex and pseudo-
Euclidean spaces. In Rigidity and symmetry, volume 70 of Fields Inst. Commun., pages 131–
154. Springer, New York, 2014.
References 541

[4] D. Kitson and S. C. Power. Infinitesimal rigidity for non-Euclidean bar-joint frameworks.
Bull. Lond. Math. Soc., 46(4):685–697, 2014.
[5] D. Kitson and B. Schulze. Motions of grid-like reflection frameworks. Journal of Symbolic
Computation, To appear. arxiv.org/abs/1709.09026.
[6] Derek Kitson. Finite and infinitesimal rigidity with polyhedral norms. Discrete Comput.
Geom., 54(2):390–411, 2015.
[7] Derek Kitson and Bernd Schulze. Maxwell-Laman counts for bar-joint frameworks in normed
spaces. Linear Algebra Appl., 481:313–329, 2015.
[8] Derek Kitson and Bernd Schulze. Symmetric isostatic frameworks with `1 or `∞ distance
constraints. Electron. J. Combin., 23(4):Paper 4.23, 23, 2016.
[9] Horst Martini, Konrad J. Swanepoel, and Gunter Weiß. The geometry of Minkowski spaces—
a survey. I. Expo. Math., 19(2):97–142, 2001.
[10] C. St. J. A. Nash-Williams. Decomposition of finite graphs into forests. J. London Math. Soc.,
39:12, 1964.
[11] F.V. Saliola and W. Whiteley. Some notes on the equivalence of first-order rigidity in various
geometries. Preprint, 2007. arxiv.org/abs/0709.3354.
[12] Bernd Schulze and Walter Whiteley. Coning, symmetry and spherical frameworks. Discrete
Comput. Geom., 48(3):622–657, 2012.
[13] Hellmuth Stachel. Flexible octahedra in the hyperbolic space. In Non-Euclidean geometries,
volume 581 of Math. Appl. (N. Y.), pages 209–225. Springer, New York, 2006.
[14] A. C. Thompson. Minkowski geometry, volume 63 of Encyclopedia of Mathematics and its
Applications. Cambridge University Press, Cambridge, 1996.
Chapter 25
Combinatorial Rigidity of Symmetric
and Periodic Frameworks
Bernd Schulze
Lancaster University, Lancaster, U.K.

CONTENTS
25.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 543
25.2 Incidentally Symmetric Isostatic Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 544
25.2.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 544
25.2.2 Symmetry-Adapted Maxwell Counts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 545
25.2.3 Characterizations of Symmetric Isostatic Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 548
25.3 Forced-Symmetric Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 549
25.3.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 549
25.3.2 Symmetric Motions and the Orbit Rigidity Matrix . . . . . . . . . . . . . . . . . . . . . . . . . 551
25.3.3 Characterizations of Forced-Symmetric Rigid Graphs . . . . . . . . . . . . . . . . . . . . . . 553
25.4 Incidentally Symmetric Infinitesimally Rigid Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . 554
25.4.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554
25.4.2 Phase-Symmetric Orbit Rigidity Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 556
25.4.3 Characterizations of Symmetric Infinitesimally Rigid Graphs . . . . . . . . . . . . . . 557
25.5 Periodic Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559
25.5.1 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559
25.5.2 Maxwell Counts for Periodic Rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 560
25.5.3 Characterizations of Periodic Rigid Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 561
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 562

25.1 Introduction
Many structures – be they man-made, such as a building, bridge or mechanical linkage, or found
in nature, such as a biomolecule, protein or crystal – exhibit non-trivial symmetries. It is therefore
important to study the impact of symmetry on the rigidity and flexibility properties of geometric
constraint systems.
In Section 25.2, we present some fundamental methods and results for the detection of
symmetry-induced infinitesimal flexes and self-stresses in frameworks which count to be isostatic
without symmetry.
In Section 25.3, we then study the rigidity of forced-symmetric frameworks. That is, given a
framework with a certain symmetry, we aim to decide whether the framework has a non-trivial

543
544 Handbook of Geometric Constraint Systems Principles

motion that maintains this symmetry. This theory has undergone rapid development in recent years
and is particularly useful for the detection of hidden continuous flexibility in structures.
The question of whether an incidentally symmetric framework is infinitesimally rigid, i.e.,
whether a symmetric framework has any (possibly symmetry-breaking) infinitesimal flex, is more
challenging. However, extensions of the tools from the theory of forced-symmetric rigidity have re-
cently also led to new insights into the infinitesimal rigidity of incidentally symmetric frameworks.
In particular, for a number of symmetry groups, there now exist combinatorial characterizations of
infinitesimally rigid frameworks which are as generic as possible subject to the given symmetry
constraints. These results are summarized in Section 25.4.
Finally, in Section 25.5, we also provide some fundamental results concerning the rigidity of
infinite periodic frameworks, both with a fixed and a flexible lattice representation.

25.2 Incidentally Symmetric Isostatic Frameworks

25.2.1 Glossary
Γ-symmetric graph: For a finite group Γ, a simple graph G = (V, E) for which there exists a group
action θ : Γ → Aut(G), where Aut(G) denotes the automorphism group of G.
Γ-symmetric framework: For a Γ-symmetric graph G (with respect to θ : Γ → Aut(G)), and a
homomorphism τ : Γ → O(Rd ), a framework (G, p) satisfying

τ(γ)(pi ) = pθ (γ)(i) for all γ ∈ Γ and all i ∈ V.

We also refer to (G, p) as a Γ-symmetric realization of G (with respect to θ and τ).

Symmetry group: For a Γ-symmetric framework, the subgroup τ(Γ) = {τ(γ) | γ ∈ Γ} of the or-
thogonal group O(Rd ).

Group representation: For a group Γ and a linear space X, a homomorphism ρ : Γ → GL(X). The
space X is called the representation space of ρ. Two representations are considered equivalent
if they are similar.
Invariant subspace: For a representation ρ : Γ → GL(X), a subspace U ⊆ X with the property that
ρ(γ)(U) ⊆ U for all γ ∈ Γ.
Irreducible representation: A group representation ρ : Γ → GL(X) which has no non-trivial ρ-
invariant subspaces.
Tensor product of representations: For two representations ρ1 and ρ2 of a group Γ, the represen-
tation ρ1 ⊗ ρ2 of Γ given by ρ1 ⊗ ρ2 (γ) = ρ1 (γ) ⊗ ρ2 (γ) for all γ ∈ Γ.
Character: For a representation ρ of a group Γ, the vector χ(ρ) whose ith component is the trace
of ρ(γi ) for some fixed ordering γ1 , . . . , γ|Γ| of the elements of Γ.
Fixed vertex or edge: For a Γ-symmetric graph G (with respect to θ : Γ → Aut(G)) and γ ∈ Γ, a
vertex i is fixed by γ if θ (γ)(i) = i. The number of vertices of G that are fixed by γ is denoted
by |Vγ |. Similarly, an edge e = {i, j} of G is fixed by γ if θ (γ)(e) = e, that is, if either θ (γ)
fixes both i and j or θ (γ)(i) = j and θ (γ)( j) = i. The number of edges of G that are fixed by γ
is denoted by |Eγ |.
Combinatorial Rigidity of Symmetric and Periodic Frameworks 545

Γ-generic framework: A Γ-symmetric framework (G, p) (with respect to θ and τ) whose rigidity
matrix has maximal rank among all Γ-symmetric realizations of G (with respect to θ and τ).
Γ-symmetric isostatic graph: A Γ-symmetric graph G for which some (equivalently, almost all)
Γ-symmetric realizations of G are isostatic.

25.2.2 Symmetry-Adapted Maxwell Counts

A fundamental result for the rigidity analysis of symmetric structures is that the rigidity matrix
of a symmetric framework (G, p) with symmetry group τ(Γ) can be transformed into a block-
diagonalized form via methods from group representation theory [21, 47, 48]. Using this block
structure of the rigidity matrix, the (infinitesimal) rigidity analysis of (G, p) can be broken up into
independent subproblems, one for each irreducible representation ρi of Γ, where each subproblem
considers the relationship between external forces on the joints of (G, p) and the resulting internal
forces in the bars of (G, p) that share the same symmetry properties described by ρi [12, 22, 36, 48].
More precisely, for a Γ-symmetric framework (G, p) (with respect to θ and τ), let PE : Γ →
GL(R|E| ) be the representation which assigns to each γ ∈ Γ the permutation matrix of the permuta-
tion θ (γ) of E, that is, PE (γ) = [δi,θ (γ)( j) )]i, j , where δ denotes the Kronecker delta symbol. Further,
let τ ⊗ PV : Γ → Rd|V | be the tensor product of the representations τ and PV , where (analogous to
PE ) PV : Γ → GL(R|V | ) assigns to each γ ∈ Γ the permutation matrix of the permutation θ (γ) of V .
Then we have the following basic result [48, 55].
Theorem 25.1 (Intertwining Property Of The Rigidity Matrix) Let (G, p) be a Γ-symmetric
framework with respect to θ : Γ → Aut(G) and τ : Γ → O(Rd ). Then, for the rigidity matrix R(G, p),
we have
R(G, p)(τ ⊗ PV )(γ) = PE (γ)R(G, p) for all γ ∈ Γ

From Theorem 25.1 and Schur’s lemma, we obtain the following corollary.
Corollary 25.2 ((Block-Diagonalization Of The Rigidity Matrix)) Let Γ be a group, and let
ρ0 , . . . , ρr be the irreducible representations of Γ. If (G, p) is a Γ-symmetric framework, then there
exist invertible matrices S and T such that the rigidity matrix takes on the block form
 
Re0 (G, p) 0
T > R(G, p)S := R(G, p) = 
 .. .

e .
0 Rr (G, p)
e

Note that for i ∈ {0, . . . , r}, the size of the block matrix Rei (G, p) is dim(Wi ) × dim(Vi ), where
Wi and Vi denote the PE -invariant and (τ ⊗ PV )-invariant subspace corresponding to the irreducible
representation ρi , respectively.
Let T (G, p) denote the space of trivial infinitesimal motions of (G, p). Then it is easy to show
that T (G, p) is a (τ ⊗PV )-invariant subspace of Rd|V | [48]. Thus, we may form the subrepresentation
(τ ⊗ PV )(T ) of τ ⊗ PV with representation space T (G, p), and T (G, p) can be written as a direct
sum of (τ ⊗ PV )-invariant subspaces Ti , i = 0, . . . , r. Clearly, for (G, p) to be isostatic, we must have
dim(Wi ) = dim(Vi )−dim(Ti ) for each i. These conditions can be summarized as follows [12, 36, 48].
Theorem 25.3 (Symmetry-Adapted Maxwell Rule) Let (G, p) be an isostatic framework which
is Γ-symmetric with respect to θ and τ. Then we have
χ(PE ) = χ(τ ⊗ PV ) − χ((τ ⊗ PV )(T ) ). (25.1)
546 Handbook of Geometric Constraint Systems Principles

Each of the characters in Equation (25.1) can easily be computed for any symmetry group τ(Γ)
[11, 47, 48]. The calculations of characters for isostatic frameworks in the plane and in 3-space
are shown in Table 25.2.2 and 25.2.2, respectively (see also Theorem 25.4). In these tables (and
throughout this chapter) we use the Schoenflies notation to denote the various symmetry groups
and their elements, as this is one of the standard notations for symmetric structures [1, 3]. In this
notation, s and Cn denote a reflection in a (d − 1)-dimensional hyperplane, and a rotation by 2π/n
about a (d − 2)-dimensional axis, respectively. Moreover, Cs is a symmetry group generated by a
single reflection, and Cn is a symmetry group generated by a rotation Cn .
Note that we may write each of the characters in Equation (25.1) uniquely as a linear combi-
nation of the characters of the irreducible representations of Γ. Therefore, if the symmetry-adapted
Maxwell rule shows that a framework is not isostatic, then a comparison of coefficients in these
linear combinations may be used to determine the invariant subspaces to which the detected in-
finitesimal flexes or self-stresses belong.

(a) (b)

Figure 25.1
Realizations of the complete bipartite graph K3,3 with reflection symmetry Cs = {Id, s} in the plane.
The framework in (a) is isostatic, whereas the framework in (b) is infinitesimally flexible, as detected
by the symmetry-adapted Maxwell rule.

Example 58 For the framework (G, p) in Figure 25.1 (b), we have χ(PE ) = (9, 3), χ(PV ) = (6, 0),
χ(τ) = (2, 0), and χ(τ ⊗ PV ) = (12, 0). Moreover, χ((τ ⊗ PV )(T ) ) = (3, −1). Thus,

χ(PE ) = (9, 3) 6= (12, 0) − (3, −1) = χ(τ ⊗ PV ) − χ((τ ⊗ PV )(T ) ),

and hence, by Theorem 25.3, (G, p) is not isostatic.

Let ρ0 be the trivial (“fully-symmetric”) irreducible representation of Cs which assigns 1 to
each element of Cs , and let ρ1 be the non-trivial (“anti-symmetric”) irreducible representation of
Cs which assigns 1 to Id and −1 to s. Then we have (9, 3) = 6ρ0 + 3ρ1 and (9, 1) = 5ρ0 + 4ρ1 , and
hence we may conclude that (G, p) has an anti-symmetric infinitesimal flex and a fully-symmetric
self-stress.

By considering the vector equation (25.1) componentwise, we may obtain very simple necessary
conditions for a Γ-symmetric framework (G, p) to be isostatic in terms of the number of vertices
and edges of G that are fixed by the elements of Γ [11].

Theorem 25.4 (Conditions For Individual Symmetry Operations) Let (G, p) be an isostatic
framework which is Γ-symmetric with respect to θ and τ. Then, for every γ ∈ Γ, we have

|Eγ | = trace(τ(γ)) · |Vγ | − trace((τ ⊗ PV )(T ) (γ)).

Combinatorial Rigidity of Symmetric and Periodic Frameworks 547

Table 25.1
Calculations of characters for the symmetry-adapted Maxwell rule in the plane.
Id Cn>2 C2 s
χ(PE ) |E| |ECn | |EC2 | |Es |
χ(τ ⊗ PV ) 2|V | (2 cos 2π
n )|VCn | −2|VC2 | 0
(T
χ((τ ⊗ PV ) ) ) 3 2 cos 2π + 1 −1 −1
n

In particular, it follows from Theorem 25.4 and Table 25.2.2 that if the symmetry group τ(Γ) of
a 2-dimensional isostatic framework contains a reflection s, then we must have |Es | = 1. Similarly,
if τ(Γ) contains a half-turn C2 , then we must have |VC2 | = 0 and |EC2 | = 1, and if τ(Γ) contains a
three-fold rotation C3 , then we must have |VC3 | = 0. Moreover, it follows that there cannot exist any
n-fold rotation in τ(Γ) with n > 3. Thus, there are only 5 non-trivial symmetry groups for which
we can construct an isostatic framework in the plane, namely the rotational groups C2 and C3 , the
reflectional group Cs , and the dihedral groups C2v and C3v of order 4 and 6 [11].

(a) (b) (c) (d) (e)

Figure 25.2
Symmetric isostatic frameworks in the plane (with fixed edges shown in gray color). (a) for C2 , we
have |VC2 | = 0 and |EC2 | = 1; (b) for C3 , we have |VC3 | = 0; (c) for Cs , we have |Es | = 1; (d,e) for the
groups C2v and C3v , the conditions for (a), (b), (c) must be satisfied by all rotations and reflections.

Table 25.2
Calculations of characters for the symmetry-adapted Maxwell rule in 3-space.

Id Cn>2 C2 s i Sn>2
χ(PE ) |E| |ECn | |EC2 | |Es | |Ei | |ESn |
χ(τ ⊗ PV ) 3|V | (2 cos 2π
n + 1)|VCn | −|VC2 | |Vs | −3|Vi | (2 cos 2π
n − 1)|VSn |
χ((τ ⊗ PV )(T ) ) 6 2π
4 cos n + 2 −2 0 0 0

While for 3-dimensional symmetric isostatic frameworks, there are still restrictions on the num-
ber of fixed structural elements (which can be derived from the calculations in Table 25.2.2 [11]),
548 Handbook of Geometric Constraint Systems Principles

we may use Cauchy’s rigidity theorem for triangulated convex polyhedra to construct infinite fam-
ilies of isostatic frameworks for every symmetry group in 3-space [11]. Note that in Table 25.2.2,
i denotes inversion in the origin, and Sn denotes an “improper rotation,” i.e., a rotation by 2π/n,
followed by a reflection in a mirror perpendicular to the rotational axis.
Analogous symmetry-adapted counting rules have also been established for various other types
of geometric constraint systems, such as body-bar and body-hinge frameworks [13, 17, 53], point-
line frameworks [36], periodic frameworks [15] and frameworks in non-Euclidean normed spaces
[24, 25] (see also Chapter 24).

25.2.3 Characterizations of Symmetric Isostatic Graphs

Let G be a Γ-symmetric graph (with respect to θ : Γ → Aut(G)) and let τ : Γ → O(Rd ) be a homo-
morphism. Then all Γ-generic realizations of G (i.e., almost all Γ-symmetric realizations of G) share
the same infinitesimal rigidity properties [48, 49]. Thus, Γ-generic rigidity is a purely combinatorial
property.
In the following we provide combinatorial characterizations of Γ-generic isostatic frameworks
for the groups C2 , C3 , and Cs in the plane, starting with the simplest case, the group C3 [51]. The
analogous conjectures for the dihedral groups remain open [50]. Due to the well-known difficulties
for the non-symmetric situtation (see [61] for example), analogous results for symmetric bar-joint
frameworks in dimension 3 and higher have also not yet been established.
In the following, a graph G is said to satisfy the Laman conditions if |E| = 2|V | − 3, and for
every subgraph (V 0 , E) with |V 0 | ≤ 2 vertices, we have |E 0 | ≤ 2|V 0 | − 3.
Theorem 25.5 (C3 -Generic Isostatic Graphs) Let G be a Z3 -symmetric graph (with respect to
θ : Z3 → Aut(G)) with at least three vertices. Further, let Z3 = hγi and τ : Z3 → O(R2 ) be a
homomorphism so that τ(Z3 ) = C3 . Then the following are equivalent:
(a) there exists a C3 -symmetric framework (G, p), and G is a Z3 -symmetric isostatic graph (with
respect to θ and τ);
(b) G satisfies the Laman conditions and |VC3 | = 0;
(c) G has a proper 3Tree2 partition into three trees T0 , T1 , T2 so that θ (γ)Ti = Ti+1 for i = 0, 1, 2,
where the indices are added modulo 3 (see Figure 25.3 (a)).

T0
T1
T2

(a) (b) (c) (d)

Figure 25.3
Proper 3Tree2 partitions (see Chapter 18, Theorem 18.25) satisfying the conditions in part (c) of
Theorems 25.5, 25.6, and 25.7.

For the half-turn symmetry group C2 in the plane, we have the following result.
Theorem 25.6 (C2 -Generic Isostatic Graphs) Let G be a Z2 -symmetric graph (with respect to
θ : Z2 → Aut(G)) with at least two vertices. Further, let Z2 = hγi and τ : Z2 → O(R2 ) be a ho-
momorphism so that τ(Z2 ) = C2 . Then the following are equivalent:
Combinatorial Rigidity of Symmetric and Periodic Frameworks 549

(a) there exists a C2 -symmetric framework (G, p), and G is a Z2 -symmetric isostatic graph (with
respect to θ and τ);
(b) G satisfies the Laman conditions as well as |VC2 | = 0 and |EC2 | = 1;
(c) G has a proper 3Tree2 partition into three trees T0 , T1 , T2 , where θ (γ)T1 = T2 , and T0 is a
spanning tree with θ (γ)T0 = T0 (see Figure 25.3 (b)).

Finally, we provide the analogous result for the group Cs describing mirror symmetry in the
plane.

Theorem 25.7 (Cs -Generic Isostatic Graphs) Let G be a Z2 -symmetric graph (with respect to
θ : Z2 → Aut(G)) with at least two vertices. Further, let Z2 = hγi and τ : Z2 → O(R2 ) be a ho-
momorphism so that τ(Z2 ) = Cs . Then the following are equivalent:
(a) there exists a Cs -symmetric framework (G, p), and G is a Z2 -symmetric isostatic graph (with
respect to θ and τ);
(b) G satisfies the Laman conditions and |Es | = 1;
(c) G has a proper 3Tree2 partition into three trees T0 , T1 , T2 so that either θ (γ)T1 = T2 , and
θ (γ)T0 = T0 (see Figure 25.3 (c)), or there exists an edge e = {i, j} in T1 whose end-vertices
i and j are both fixed by θ (γ), and θ (γ)(T1 − i) = T2 and θ (γ)T0 = T0 (see Figure 25.3 (d)).

For the symmetry groups τ(Γ) = C2 , C3 , Cs in the plane, there are also characterizations of Γ-
symmetric isostatic graphs in terms of symmetric Henneberg-type construction sequences [50, 51,
55] (see also Chapter 19).
For example, for the group C3 , this construction sequence starts with the complete graph K3
on three vertices and consists of three graph operations, namely a symmetric Henneberg 1- and a
symmetric Henneberg 2-move (which consist of three standard Henneberg 1- and 2-moves, respec-
tively, carried out simultaneously in a symmetric fashion) and a symmetric “’K3 -addition move,”
which symmetrically joins each vertex of a new K3 to a vertex of the previous graph, as illustrated
in Figure 25.2 (b) (see [51] for details).
For the group C2 , appropriate symmetric versions of the Henneberg 1- and 2-moves suffice to
characterize C2 -symmetric isostatic graphs [50]. For the group Cs , however, we also need a symmet-
ric version of the X-replacement [50, 61].
For a number of conjectures, as well as some initial results, regarding sufficient conditons for
Γ-generic body-bar frameworks in 3-space to be isostatic, we refer the reader to [17].

25.3 Forced-Symmetric Frameworks

25.3.1 Glossary
Free action on the vertex set of a graph: For a Γ-symmetric graph G, a group action θ : Γ →
Aut(G) with the property that θ (γ)(i) 6= i for all i ∈ V and all non-trivial γ ∈ Γ. Similarly,
θ is free on the edge set of G if θ (γ)(e) 6= e for all e ∈ E and all non-trivial γ ∈ Γ.
Quotient graph: For a Γ-symmetric graph G, the multigraph G/Γ with vertex set V /Γ = {Γi | i ∈
V } and edge set E/Γ = {Γe | e ∈ E}, where Γi and Γe are the vertex and edge orbits of i and
e = {i, j} defined by {θ (γ)(i) | γ ∈ Γ} and {{θ (γ)(i), θ (γ)( j)} | γ ∈ Γ}, respectively.
550 Handbook of Geometric Constraint Systems Principles

Quotient Γ-gain graph: Let G be a Γ-symmetric graph with respect to the group action θ : Γ →
Aut(G) which is free on the vertex set of G. The quotient Γ-gain graph of G is obtained from
the quotient graph G/Γ of G by orienting the edges of G/Γ and labelling the edges of G/Γ via
the function ψ : E/Γ → Γ as follows. Each edge orbit Γe connecting Γi and Γ j in G/Γ can be
written as {{θ (γ)(i), θ (γ) ◦ θ (α)( j)} | γ ∈ Γ} for a unique α ∈ Γ. For each Γe, orient Γe from
Γi to Γ j in G/Γ and assign to it the group element α. Note that the resulting quotient Γ-gain
graph (G0 , ψ) of G is unique up to choices of representative vertices and that the orientation is
only used as a reference orientation and may be changed, provided that we also modify ψ so
that if Γe is an edge in one direction, and Γe−1 is the same edge in the opposite direction, then
ψ(Γe−1 ) = ψ(Γe)−1 .
In the following, we will denote the vertex and edge set of (G0 , ψ) by V0 and E0 , respectively,
and use the tilde symbol to denote the vertices and edges of (G0 , ψ). We also refer to G as the
covering graph of (G0 , ψ).
Gain of a closed walk: For a quotient Γ-gain graph (G0 , ψ) and a closed walk

W = ĩ1 , ẽ1 , ĩ2 , ẽ2 , ĩ3 , . . . , ĩk , ẽk , ĩ1

of (G0 , ψ), the group element ψ(W ) = Πt=1 k ψ(ẽ )sign(ẽt ) , where sign(ẽ ) = 1 if ẽ is directed
t t t
from ĩt to ĩt+1 , and sign(ẽt ) = −1 otherwise.
Subgroup induced by edge set: For a quotient Γ-gain graph (G0 , ψ), a subset F ⊆ E0 , and a vertex
ĩ of the vertex set V (F) ⊆ V0 induced by F, the subgroup hFiψ,ĩ = {ψ(W )|W ∈ W(F, ĩ)} of Γ,
where W(F, ĩ) is the set of closed walks starting at ĩ using only edges of F.
Balanced edge set: For a quotient Γ-gain graph (G0 , ψ), a connected edge subset F of E0 is called
balanced if hFiψ,ĩ = {id} for some ĩ ∈ V (F) (or equivalently, hFiψ,ĩ = {id} for all ĩ ∈ V (F)). A
disconnected subset of E0 is balanced if all of its connected components are balanced. A subset
of E0 is called unbalanced if it is not balanced (i.e., if it contains a cycle that is not balanced).
(k, `, m)-gain-sparse: For non-negative integers k, `, m with m ≤ `, a quotient Γ-gain graph (G0 , ψ)
satisfying (
k|V (F)| − `, for all non-empty balanced F ⊆ E0 ,
|F| ≤
k|V (F)| − m, for all non-empty F ⊆ E0 .
If we also have |E0 | = k|V0 | − m, then (G0 , ψ) is called (k, `, m)-gain-tight.
Cyclic edge set: For a quotient Γ-gain graph (G0 , ψ), a connected edge subset F of E0 is called
cyclic if hFiψ,ĩ is a cyclic subgroup of Γ for some ĩ ∈ V (F) (or equivalently, for all ĩ ∈ V (F)).
A disconnected subset of E0 is cyclic if all of its connected components are cyclic.
Orbit rigidity matrix: For a d-dimensional Γ-symmetric framework (G, p) (with respect to θ and
τ), where θ : Γ → Aut(G) is free on the vertex set of G, and its quotient Γ-gain graph (G0 , ψ),
the |E0 | × d|V0 | matrix O(G0 , ψ, p) defined as follows. Choose a representative vertex ĩ for each
vertex Γi in V0 . The row corresponding to the edge ẽ = (ĩ, j˜), ĩ 6= j˜, with gain ψ(ẽ) has the form

ĩ j˜

.
z }| { z }|
{
0...0 p(ĩ) − τ(ψ(ẽ))p( j˜) 0 . . . 0 ˜ −1
p( j) − τ(ψ(ẽ)) p(ĩ) 0 . . . 0

If ẽ = (ĩ, ĩ) is a loop at ĩ, then the row corresponding to ẽ has the form
ĩ

.
z }| {
0...0 2p(ĩ) − τ(ψ(ẽ))p(ĩ) − τ(ψ(ẽ))−1 p(ĩ) 0...0 0 0...0
Combinatorial Rigidity of Symmetric and Periodic Frameworks 551

Fully Γ-symmetric infinitesimal motion: For a Γ-symmetric framework (G, p), an infinitesimal
motion u : V → Rd of (G, p) satisfying

τ(γ)ui = uθ (γ)(i) for all γ ∈ Γ and all i ∈ V.

s
p1 p4 p1 p3
1̃ 1̃
id id
2̃ s
p2 p3 s p2 p4 2̃
s s
(a) (b) (c) (d)

Figure 25.4
(a,b) A fully Cs -symmetric infinitesimally rigid framework and its corresponding quotient Cs -gain
graph. The infinitesimal translation in (a) spans the space of fully Cs -symmetric trivial infinitesi-
mal motions. (c,d) A Cs -symmetric framework with a fully Cs -symmetric infinitesimal flex and its
corresponding quotient Cs -gain graph.

Fully Γ-symmetric self-stress: For a Γ-symmetric framework (G, p), a self-stress ω : E → R with
the property that ω(e) = ω( f ) whenever e and f belong to the same edge orbit Γe of G.
Fully Γ-symmetric infinitesimally rigid framework: A Γ-symmetric framework for which every
fully Γ-symmetric infinitesimal motion is trivial. (See also Figure 25.4.)
ρ0 -generic framework: A Γ-symmetric framework (G, p) (with respect to θ and τ) whose orbit
rigidity matrix has maximal rank among all Γ-symmetric realizations of G (with respect to θ
and τ).

25.3.2 Symmetric Motions and the Orbit Rigidity Matrix

In this section we restrict attention to those motions of a symmetric framework which preserve all of
the symmetries of the framework throughout the path. Our first result is a symmetric version of the
Theorem of Asimov-Roth [2], which says that for almost all Γ-symmetric realizations of a graph G,
the existence of a fully Γ-symmetric infinitesimal flex also guarantees the existence of a continuous
symmetry-preserving flex [14, 52]. This result provides one of the key motivations for the study of
forced-symmetric rigidity.

Theorem 25.8 (Symmetry-Preserving Motions) A Γ-symmetric ρ0 -generic framework (G, p) has

a fully Γ-symmetric infinitesimal flex if and only if (G, p) has a continuous flex which preserves the
symmetry of (G, p) throughout the path.

Note that the fully Γ-symmetric rigidity properties of a Γ-symmetric framework are described
by the submatrix block Re0 (G, p) of the block-diagonalized rigidity matrix R(G, e p) (recall Theo-
rem 25.3), which corresponds to the trivial irreducible representation ρ0 of Γ (i.e., the 1-dimensional
representation which assigns 1 to each element of Γ). However, to obtain the explicit entries of this
matrix, we need to go through the laborious process of block-diagonalizing the rigidity matrix. To
simplify the rigidity analysis of forced-symmetric frameworks, the orbit rigidity matrix was intro-
duced in [56]. This matrix is equivalent to the matrix Re0 (G, p) (see Theorem 25.9), but its entries
have the transparent form described in Section 25.3.1.
552 Handbook of Geometric Constraint Systems Principles

Theorem 25.9 (The Orbit Rigidity Matrix) Let (G, p) be a Γ-symmetric framework with respect
to θ : Γ → Aut(G) (which is free on the vertex set of G) and τ : Γ → O(Rd ). The kernel of the orbit
rigidity matrix O(G0 , ψ, p) is isomorphic to the space of fully Γ-symmetric infinitesimal motions of
(G, p), and the kernel of O(G0 , ψ, p)T is isomorphic to the space of fully Γ-symmetric self-stresses
of (G, p).

Let trivτ(Γ) denote the dimension of the space of fully Γ-symmetric trivial infinitesimal mo-
tions of (G, p). Then (G, p) is clearly fully Γ-symmetric infinitesimally rigid if and only if
rank O(G0 , ψ, p) = d|V0 | − trivτ(Γ) , provided that G has at least d vertices. This leads to the fol-
lowing necessary conditions for a framework to be fully Γ-symmetric rigid [20, 56].

Theorem 25.10 (Forced-Symmetric Maxwell Counts) Let G be a graph with at least d vertices
and let (G, p) be a fully Γ-symmetric infinitesimally rigid framework with respect to the action
θ : Γ → Aut(G) which is free on the vertex set of G and τ : Γ → O(Rd ). Then the quotient Γ-gain
graph of G contains a spanning subgraph (H, ψ) with edge set E00 which satisfies
(a) |E00 | = d|V0 | − trivτ(Γ)

(b) |F| ≤ d|V (F)| − trivτ(hFiψ,ĩ ) for all F ⊆ E00 and all ĩ ∈ V (F),

where trivτ(hFiψ,ĩ ) is the dimension of the space of fully (hFiψ,ĩ )-symmetric trivial infinitesimal mo-
tions of the framework induced by the edges in F.

Note that trivτ(Γ) can easily be computed for any symmetry group in any dimension. For all
symmetry groups in the plane and in 3-space, these numbers are also recorded in the corresponding
character tables for these groups [1, 3].
For example, for mirror symmetry in the plane, we have trivCs = 1, as an infinitesimal translation
along the mirror line forms a basis of the space of fully Cs -symmetric trivial infinitesimal motions
(see also Figure 1.4 (a)). Similarly, for half-turn symmetry in the plane, we have trivC2 = 1, as an
infinitesimal rotation about the origin spans the space of fully C2 -symmetric trivial infinitesimal
motions.
Therefore, using Theorems 25.8 and 25.10, we may detect hidden continuous (symmetry-
preserving) flexibility in frameworks by simply counting the number of vertex and edge orbits in
the underlying graph.
A simple example in 3-space is the ρ0 -generic realization of the octahedral graph with half-
turn symmetry C2 shown in Figure 25.5(a). Generic realizations of this graph without symmetry
are isostatic in 3-space. For the numbers of vertex and edge orbits, we have |V0 | = 3 and |E0 | = 6.
Moreover, we have trivC2 = 2 [1, 3, 56]. Thus, |E0 | < 3|V0 | − trivC2 , and we may conclude that there
exists a symmetry-preserving continuous flex.
The definition of the orbit rigidity matrix has also been extended to Γ-symmetric frameworks
where Γ does not act freely on the vertex set [56]. This leads to adjusted necessary conditions for
fully Γ-symmetric infinitesimal rigidity. While the overall counts on the numbers of vertex and
edge orbits can easily be obtained (as demonstrated by the example below), the corresponding gain-
sparsity counts become signifcantly less clear and transparent, and they have not yet been studied
in detail.
Consider the ρ0 -generic realization of the octahedral graph with reflectional symmetry Cs shown
in Figure 25.5 (b). For the numbers of vertex and edge orbits, we have |V0 | = 4 and |E0 | = 6.
Moreover, we have trivCs = 3 [1, 3, 56]. Since there are two vertices which are fixed by the reflection,
and each of these vertices contributes only two columns to the orbit rigidity matrix (as each of them
needs to remain on the mirror plane), the framework has 2 · 3 + 2 · 2 − 3 = 7 fully Z2 -symmetric
non-trivial degrees of freedom. Since |E0 | < 7, it follows that there exists a symmetry-preserving
continuous flex.
Combinatorial Rigidity of Symmetric and Periodic Frameworks 553

Figure 25.5
Flexible octahedra (also known as “Bricard octahedra”) with half-turn symmetry (a) and with mirror
symmetry (b).

25.3.3 Characterizations of Forced-Symmetric Rigid Graphs

All ρ0 -generic realizations of a graph share the same fully Γ-symmetric rigidity properties, and for a
number of symmetry groups in the plane, Laman-type theorems for fully Γ-symmetric rigidity have
been established. For the groups Cs and Cn , we have the following result (see [20, 27, 28, 33, 31] for
various proofs).

Theorem 25.11 (Reflectional or Rotational Symmetry in 2D) Let (G, p) be a Zn≥2 -symmetric
ρ0 -generic framework with respect to the action θ : Zn → Aut(G) which is free on the vertex set
of G and τ : Zn → O(R2 ). Then (G, p) is forced Zn -symmetric infinitesimally rigid if and only if the
quotient Zn -gain graph of G contains a spanning subgraph (H, ψ) which is (2, 3, 1)-gain-tight.

The count 2|V0 | − 1 for the number of edges of (H, ψ) is due to the fact that trivτ(Zn ) = 1 for
all cyclic groups Zn , n ≥ 2. For a dihedral group Cnv , we always have trivCnv = 0. This leads to the
following result [20].

Theorem 25.12 (Dihedral Symmetry in 2D) Let D2n denote the dihedral group of order 2n, where
n ≥ 3 is an odd integer. Further, let (G, p) be a D2n -symmetric ρ0 -generic framework with respect
to the action θ : D2n → Aut(G) which is free on the vertex set of G, and τ : D2n → O(R2 ), where
τ(D2n ) = Cnv . Then G(p) is forced D2n -symmetric infinitesimally rigid if and only if the quotient
D2n -gain graph of G contains a spanning subgraph (H, ψ) with edge set E00 which satisfies
(a) |E00 | = 2|V0 |

0
2|V (F)| − 3 for all non-empty balanced F ⊆ E0 ,

(b) |F| ≤ 2|V (F)| − 1 for all non-empty unbalanced and cyclic F ⊆ E00 ,
for all non-empty F ⊆ E00 .

2|V (F)|


Note that the gain-sparsity counts in Theorems 25.11 and 25.12 can be checked in polynomial
time [5, 20]. In general, if we want to check a quotient Γ-gain graph for (k, `, m)-gain-sparsity,
where 0 ≤ ` ≤ 2k − 1, then we may proceed as follows. We first verify that the quotient Γ-gain gaph
is (k, m)-sparse using a standard “pebble game algorithm” [6, 26]. We then test whether every edge
set violating the (k, `)-sparsity count induces a certain subgroup of Γ. It suffices to test this for every
circuit in the matroid induced by the (k, `)-sparsity count, and these circuits can be enumerated in
polynomial time [58]. As we will see in the following sections, this algorithm may also be used to
check for infinitesimal rigidity of incidentally symmetric frameworks and for rigidity of periodic
frameworks.
554 Handbook of Geometric Constraint Systems Principles

For the groups Cs , Cn , and Cnv (n odd), there also exist alternative characterizations of ρ0 -generic
fully Γ-symmetric infinitesimally rigid frameworks in terms of symmetric Henneberg-type construc-
tion sequences for the corresponding quotient Γ-gain graphs [20].
The only remaining groups are the dihedral groups of “even order,” that is, the groups D2n , where
n is an even integer. For these groups, the counts in Theorem 25.12 are in general not sufficient for
a ρ0 -generic framework to be fully D2n -symmetric infinitesimally rigid. Two counterexamples are
depicted in Figure 25.6. See [20] for further examples, as well as for some results on recursive
constructions which preserve fully D2n -symmetric infinitesimal rigidity.

γ
s0

id γ s s0 id s s id
s
id
(a) (b)

Figure 25.6
Quotient D4 -gain graphs (with the directions of the edges omitted), where D4 consists of the identity
id, the half-turn γ, and the reflections s and s0 : (a) a quotient D4 -gain graph whose covering graph,
the complete bipartite graph K4,4 , is flexible for D4 -generic realizations in the plane (this is also
known as Bottema’s mechanism [10]); (b) another quotient D4 -gain graph whose covering graph is
flexible for D4 -generic realizations in the plane.

For body-bar frameworks, there are combinatorial characterizations for fully Γ-symmetric in-
finitesimal rigidity for all symmetry groups in all dimensions [59]. These counts are analogous
to the counts in Theorem 25.10, and are conjectured to also extend to body-hinge and molecular
frameworks [38].
In [35] combinatorial characterizations for fully Γ-symmetric infinitesimal rigidity have also
been obtained for Euclidean frameworks in 3-space whose vertices are constrained to lie on various
surfaces (such as a cone or a cylinder).
Finally, we note that the fully Γ-symmetric rigidity properties of a Γ-symmetric framework
(including continous symmetry-preserving flexibility) can also be transferred to other metrics, such
as the spherical or hyperbolic metric, using the orbit rigidity matrix and a symmetrized version of
the technique of “coning” [57].

25.4 Incidentally Symmetric Infinitesimally Rigid Frameworks

25.4.1 Glossary
Phase-symmetric infinitesimal motion: Let Γ be the group Zk = {0, 1, . . . , k − 1} and for t =
0, 1, . . . , k − 1, let ρt : Γ → C \ {0} be the irreducible representation of Γ defined by ρt ( j) = ω t j ,
2πi
where ω denotes the root of unity e k . For a Γ-symmetric framework (G, p), an infinitesimal
motion u : V → Rd of (G, p) is called ρt -symmetric if it lies in the (τ ⊗ PV )-invariant subspace
Combinatorial Rigidity of Symmetric and Periodic Frameworks 555

Vt corresponding to ρt , that is, if it satisfies

τ(γ)ui = ω tγ uθ (γ)(i) for all γ ∈ Γ and all i ∈ V.

(See also Figure 25.7.) Similarly, a self-stress of (G, p) is called ρt -symmetric if it lies in the
PE -invariant subspace Wt corresponding to ρt .

(a) (b) (c)

Figure 25.7
ρ1 -symmetric (or anti-symmetric) infinitesimal motions of frameworks with mirror symmetry in the
plane: (a), (b) anti-symmetric infinitesimal flexes; (c) an anti-symmetric trivial infinitesimal motion.

Phase-symmetric infinitesimally rigid framework: For the group Γ = Zk , a Γ-symmetric frame-

work (G, p) is ρt -symmetric infinitesimally rigid if every ρt -symmetric infinitesimal motion of
(G, p) is trivial.
Phase-symmetric orbit rigidity matrix: Let (G, p) be a Γ-symmetric framework with respect to
θ : Γ → Aut(G) and τ : Γ → O(Rd ), where Γ is the group Zk and θ is free on the vertex set of G.
Further, let (G0 , ψ) be the quotient Γ-gain graph of G. The ρt -symmetric orbit rigidity matrix
is the |E0 | × d|V0 | matrix Ot (G0 , ψ, p) defined as follows. Choose a representative vertex ĩ for
each vertex Γi in V0 . The row corresponding to the edge ẽ = (ĩ, j˜), ĩ 6= j˜, with gain ψ(ẽ) has the
form
ĩ j˜

.
z }| { z }| {
tψ(ẽ) −1
0...0 p(ĩ) − τ(ψ(ẽ))p( j˜) 0...0 ω (p( j˜) − τ(ψ(ẽ)) p(ĩ)) 0...0
If ẽ = (ĩ, ĩ) is a loop at ĩ, then the row corresponding to ẽ has the form

ĩ

.
z }| {
tψ(ẽ) −1
0...0 p(ĩ) − τ(ψ(ẽ))p(ĩ) + ω (p(ĩ) − τ(ψ(ẽ)) p(ĩ)) 0...0 0 0...0

ρt -generic framework: A Γ-symmetric framework (G, p) (with respect to θ and τ) whose ρt -symmetric orbit
rigidity matrix has maximal rank among all Γ-symmetric realizations of G (with respect to θ and τ).

Unbalanced circuit: A minimal edge subet F of a quotient Zk -gain graph (G0 , ψ) so that

(a) F is unbalanced;
(b) |F| > 2|V (F)| − 1;
(c) there is a vertex ĩ ∈ V (F), an element γ ∈ Zk , and a labeling function ψ 0 : E0 → Zk equivalent to ψ
such that ψ 0 (ẽ) = id for every ẽ ∈ F not incident to ĩ, and ψ 0 (ẽ) ∈ {id, γ} for every ẽ ∈ F directed
to ĩ (assuming that every edge incident to ĩ is directed to ĩ). See also Figure 25.8.
556 Handbook of Geometric Constraint Systems Principles

e i γ
γ
ĩ
γ ẽ
(a) (b)

Figure 25.8
A Z5 -symmetric graph (a) and its corresponding quotient Z5 -gain graph whose edge set is an un-
balanced circuit (b). The orientation and gain labeling is omitted for all edges with gain id, and γ
denotes rotation by 2π/5.

25.4.2 Phase-Symmetric Orbit Rigidity Matrices

As we have seen in Section 25.2.2, an isostatic Γ-symmetric framework (G, p) must obey certain
restrictions on the numbers of vertices and edges that are fixed by the various elements of Γ. Conse-
quently, an infinitesimally rigid symmetric framework usually does not contain a spanning isostatic
subframework with the same symmetry (see also Figure 25.9). Thus, to decide whether a given
Γ-generic framework is infinitesimally rigid is in general not possible using the combinatorial re-
sults of Section 25.2.3. However, in those cases further insights may be gained by analyzing the
ρt -symmetric infinitesmal rigidity properties of (G, p) for each irreducible representation ρt of Γ.
The basic idea is that if (G, p) is ρt -symmetric infinitesimally rigid for every t, then the block-
diagonalization of the rigidity matrix R(G,
e p) guarantees that (G, p) is infinitesimally rigid. The
ρt -symmetric infinitesimal rigidity properties of (G, p) are described by the block matrix Ret (G, p)
of R(G,
e p). However, the explicit entries of Ret (G, p) are quite arduous to obtain and difficult to work
with. Therefore, analogous to the forced-symmmetric situation (which is associated with the trivial
irreducible representation of Γ), an equivalent but more transparent “orbit rigidity matrix” has been
established in [55] for each block matrix Ret (G, p). While there seem to be no major obstacles in
extending these methods to irreducible representations of arbitrary finite groups, abelian groups are
particularly simple since their irreducible representations are all 1-dimensional (over the complex
numbers). Here we focus on the simplest case where Γ is the cyclic group Zk .

Figure 25.9
Infinitesimally rigid symmetric frameworks in R2 with respective symmetry groups Cs and C2 which
do not contain a spanning isostatic subframework with the same symmetry.

Theorem 25.13 (Phase-Symmetric Orbit Rigidity Matrices) Let Γ = Zk and let (G, p) be a Γ-
symmetric framework (with respect to θ : Γ → Aut(G) and τ : Γ → O(Rd )). The kernel of the ρt -
symmetric orbit rigidity matrix Ot (G0 , ψ, p) is isomorphic to the space of ρt -symmetric infinitesimal
motions of (G, p), and the kernel of O(G0 , ψ, p)T is isomorphic to the space of ρt -symmetric self-
stresses of (G, p).
Combinatorial Rigidity of Symmetric and Periodic Frameworks 557

Analogous to Theorem 25.10, there exist some simple necessary conditions for phase-symmetric
infinitesimal rigidity for all symmetry groups in all dimensions. These conditions are obtained by
replacing each number trivτ(hFiψ,ĩ ) in the counts of Theorem 25.10 by the dimension of the corre-
sponding space of ρt -symmetric trivial infinitesimal motions. In particular, for the cyclic groups in
the plane we have the following result [55].

Theorem 25.14 (Phase-Symmetric Maxwell Counts) Let (G, p) be a Zk -symmetric framework

with respect to the action θ : Zk → Aut(G) which is free on the vertex set of G and τ : Zk → O(R2 ).
If (G, p) is ρt -symmetric infinitesimally rigid, then the quotient Zk -gain graph of G contains a span-
ning subgraph (H, ψ) so that
(a) for k = 2, (H, ψ) is (2, 3, 1)-gain-tight if t = 0 and (2, 3, 2)-gain tight if t = 1;
(b) for k ≥ 3 and t ∈ {0, 1, k − 1}, (H, ψ) is (2, 3, 1)-gain-tight;
(c) for k ≥ 4 and t ∈
/ {0, 1, k − 1}, (H, ψ) is (2, 3, 0)-gain-tight.

Since a Γ-symmetric framework (G, p) can only be infinitesimally rigid if it is ρt -symmetric

infinitesimally rigid for each irreducible representation ρt of Γ, Theorem 25.14 provides necessary
conditions for (G, p) to be infinitesimally rigid. For the groups of order 2 and 3, the conditions
in Theorem 25.14 are also sufficient for Γ-generic infinitesimal rigidity (see Theorems 25.17 and
25.19). However, for Zk , k ≥ 4, there are further necessary conditions [55]. It follows from Theo-
rem 25.14 that for k ≥ 4, a Zk -generic framework can only be infinitesimally rigid if the correspond-
ing Zk -quotient gain graph (G0 , ψ) contains a spanning subgraph (H, ψ) with 2|V0 | edges, which in
turn contains a spanning (2, 3, 1)-gain-tight subgraph. In addition, (H, ψ) cannot contain any unbal-
anced circuit [19, 55]. (Recall Figure 25.8.) Moreover, it follows from the following theorem that if
(H, ψ) contains an edge subset F with hFiψ,ĩ ' Z2 , then F must be (2, 3, 2)-gain sparse [19, 55]. In
particular, (H, ψ) cannot contain a loop with gain k/2.

Theorem 25.15 (Further Necessary Conditions For Infinitesimal Rigidity) Let k ≥ 4, and let
(G, p) be a Zk -symmetric framework with respect to the action θ : Zk → Aut(G) which is free
on the vertex set of G and τ : Zk → O(R2 ). Further, let t be an odd integer with 1 ≤ t ≤ k − 1. If
Ot (G0 , ψ, p) is row independent, then F is (2, 3, 2)-gain sparse for every F ⊆ E0 such that hFiψ,ĩ is
isomorphic to Z2 .

Using Theorem 25.15, we may construct Zk -generic infinitesimally flexible frameworks for even
k ≥ 6 whose underlying graphs are generically rigid without symmetry [55]. The easiest example is
the complete bipartite graph K3,3 with C6 symmetry shown in Figure 25.10 (b).
It was conjectured in [55] that if k is odd, then the existence of a spanning Laman subgraph
(i.e., a spanning subgraph which is generically isostatic without symmetry) in a graph G guarantees
that Zk -generic realizations of G are still infinitesimally rigid. However counterexamples to this
conjecture have recently been constructed in [18]. (See Figure 25.10(c).)

25.4.3 Characterizations of Symmetric Infinitesimally Rigid Graphs

Using the anti-symmetric orbit rigidity matrix and recursive operations on quotient Z2 -gain graphs,
the following result was established in [55].

Theorem 25.16 (Anti-Symmetric Infinitesimal Rigidity for Cs and C2 in 2D) Let

(G, p) be a Z2 -symmetric ρ1 -generic framework with respect to the action θ : Z2 → Aut(G) which is
free on the vertex set of G and τ : Z2 → O(R2 ). Then (G, p) is ρ1 -symmetric infinitesimally rigid if
and only if the quotient Z2 -gain graph of G contains a spanning subgraph (H, ψ) which is (2, 3, 2)-
gain-tight.
558 Handbook of Geometric Constraint Systems Principles
1 2
0 0
2 0 0 1
3
1 2
(a) (b) (c)

Figure 25.10
A quotient Z6 -gain graph (a) and its covering graph K3,3 (b), which is isostatic for generic realiza-
tions, but infinitesimally flexible for Z6 -symmetric realizations; (c) a quotient Zk -gain graph whose
covering graph is generically rigid without symmetry, but becomes infinitesimally flexible for Zk -
symmetric realizations, where k ≥ 7. The directions of edges are omitted in the gain graphs.

Note that the count 2|V0 | − 2 for the number of edges of (H, ψ) is due to the fact that the
dimension of the space of ρ1 -symmetric trivial infinitesimal motions is equal to 2. As a simple
consequence of Theorems 25.11 and 25.16, we obtain the following combinatorial characterizations
for infinitesimal rigidity under reflection or half-turn symmetry in the plane [55].

Theorem 25.17 (Z2 -Generic Infinitesimal Rigidity for Cs and C2 in 2D) Let (G, p) be a Z2 -
generic framework with respect to the action θ : Z2 → Aut(G) which is free on the vertex set of
G and τ : Z2 → O(R2 ). Then (G, p) is infinitesimally rigid if and only if the quotient Z2 -gain graph
of G contains a spanning (2, 3, i)-gain-tight subgraph (Hi , ψi ) for each i = 1, 2.

For the symmetry group C3 , ρt -symmetric infinitesimal rigidity is described by the same (2, 3, 1)-
gain sparsity count for each t = 0, 1, 2.

Theorem 25.18 (Phase-Symmetric Infinitesimal Rigidity for C3 in 2D) Let t ∈ {0, 1, 2} and let
(G, p) be a Z3 -symmetric ρt -generic framework with respect to the action θ : Z3 → Aut(G) which is
free on the vertex set of G and τ : Z3 → O(R2 ). Then (G, p) is ρt -symmetric infinitesimally rigid if
and only if the quotient Z3 -gain graph of G contains a spanning subgraph (H, ψ) which is (2, 3, 1)-
gain-tight.

Theorem 25.18 gives rise to the following characterization for C3 -generic infinitesimal rigidity
(which is equivalent to the characterization given in Theorem 25.5 (b)) [55].

Theorem 25.19 (Z3 -Generic Infinitesimal Rigidity for C3 in 2D) Let (G, p) be a Z3 -generic
framework with respect to the action θ : Z3 → Aut(G) which is free on the vertex set of G and
τ : Z3 → O(R2 ). Then (G, p) is infinitesimally rigid if and only if the quotient Z3 -gain graph of G
contains a spanning (2, 3, 1)-gain-tight subgraph.

As we have seen in the previous section, for k ≥ 4, the necessary counting conditions for a Zk -
symmetric framework to be infinitesimally rigid are more complex. However, for odd k < 1000, a
characterization for Zk -generic infinitesimal rigidity has very recently been obtained in [18, 19]. In
particular, for the cyclic groups Zk of prime order k < 1000, the necessary conditions outlined in
Section 25.4.2 have been shown to be sufficient for Zk -generic infinitesimal rigidity [18, 19].
For dimensions 3 and higher, combinatorial characterizations for infinitesimal rigidity have been
obtained for body-bar and body-hinge frameworks for the groups Z2 × · · · × Z2 . These characteri-
zations are given in terms of packings of bases of signed-graphic matroids on quotient multi-graphs
[54]. Below we provide a sample result for reflection and half-turn symmetry in 3-space. For precise
definitions of symmetric body-bar and body-hinge frameworks and their quotient graphs, see [54].
Combinatorial Rigidity of Symmetric and Periodic Frameworks 559

Theorem 25.20 (Z2 -Generic Infinitesimal Body-Hinge Rigidity in 3D) Let (G, h) be a Z2 -generic
body-hinge framework in R3 with respect to the action θ : Z2 → Aut(G) which is free on the vertex
set and the edge set of G, and τ : Z2 → O(R3 ). Then (G, h) is infinitesimally rigid if and only if the
quotient Z2 -gain graph (G0 , ψ) of G has the following properties.
(a) for τ(Z2 ) = Cs , (G0 , ψ) contains three edge-disjoint spanning trees and three subgraphs with
the property that each connected component contains exactly one cycle, which is unbalanced.

(b) for τ(Z2 ) = C2 , (G0 , ψ) contains two edge-disjoint spanning trees and four subgraphs with
the property that each connected component contains exactly one cycle, which is unbalanced.

It is conjectured that these results also extend to the special class of molecular frameworks [38].

25.5 Periodic Frameworks

25.5.1 Glossary
d-periodic graph: For a group Γ isomorphic to Zd , a simple infinite graph G̃ = (Ṽ , Ẽ) with finite
degree at every vertex for which there exists a group action θ : Γ → Aut(G̃) which is free on
the vertex set of G̃ and such that the quotient graph G̃/Γ is finite.
Quotient Zd -gain graph: For a d-periodic graph G̃, the directed group-labelled multi-graph
(G̃0 , ψ) obtained from the quotient graph of G̃ by orienting and labelling the edges (via the
function ψ : Ẽ/Γ → Γ) in the analogous way as described for the quotient Γ-gain graph of a
finite Γ-symmetric graph in Section 25.3.1. (See also Figure 25.11.)
d-periodic framework: For a d-periodic graph G̃ (with respect to θ ), a d-periodic placement of
G̃ is a function p̃ : Ṽ → Rd and a faithful representation L : Γ → T (Rd ), where T (Rd ) is the
group of translations of Rd (which we identify with the space Rd of translation vectors) such
that
p̃i + L(γ) = p̃θ (γ)(i) for all γ ∈ Γ and all i ∈ Ṽ .
A d-periodic graph G̃ (with respect to θ ) together with its d-periodic placement ( p̃, L) is called
a d-periodic framework, which is denoted by (G̃, p̃, L).
d-periodic infinitesimal motion: Choose an isomorphism Γ ' Zd , and let (G̃, p̃, L) be a d-periodic
framework (with respect to θ : Γ → Aut(G̃)). Let v1 , . . . , va be a set of representatives for the
vertex orbits of G̃, and let the b pairs (vi , θ (γβ )v j ) be a set of representatives for the edge orbits
of G̃. Further, let xi = p̃(vi ) for each i, and let µk = L(γk ), k = 1, . . . , d, be the translation vectors
in Rd which correspond to the standard basis γ1 , . . . , γd of Γ. Let γβ = dk=1 ckβ γk for ckβ ∈ Z,
P
2
and µ(β ) = dk=1 ckβ µk . A vector (y1 , . . . , ya , ν1 , . . . , νd ) ∈ Rda+d is an infinitesimal motion of
P

(G̃, p̃, L) if
h(x j + µ(β )) − xi , (y j + ν(β )) − yi i = 0 for β = 1, . . . , b,

where h·, ·i represents the inner product, and ν(β ) = dk=1 ckβ νk . The matrix corresponding to
P

the linear system above is called the periodic rigidity matrix of (G̃, p̃, L).
d-periodic infinitesimally rigid framework: A d-periodic framework (G̃, p̃, L) for which every
d-periodic infinitesimal motion is trivial.
560 Handbook of Geometric Constraint Systems Principles

d-periodic generic framework: A d-periodic framework (G̃, p̃, L) whose periodic rigidity matrix
has maximal rank among all d-periodic realizations (G̃, p̃0 , L0 ) of G̃ (with any choice of L0 ).
Zd -rank of an edge set: For a quotient Zd -gain graph (G̃0 , ψ), and an edge subset F of (G̃0 , ψ),
the smallest cardinality of a generating set for the subgroup of Zd induced by F, where the
subgroup induced by F is defined in the analogous way as for an edge subset of a quotient
Γ-gain graph for a finite group Γ (see Section 25.3.1).

25.5.2 Maxwell Counts for Periodic Rigidity

Largely motivated by practical applications in fields such as engineering, materials science, and
crystallography, as well as by mathematical applications in areas such as sphere packings, the rigid-
ity and flexibility analysis of periodic frameworks has gained significant attention in recent years.
Most of this work has focused on forced-periodic rigidity, i.e., on the question whether a periodic
framework has a flex which preserves the periodicity of the framework throughout the motion.
Therefore, we will restrict attention to forced-periodic rigidity in this section. (For other types of
rigidity analyses of infinite frameworks, see [4, 32, 37, 39, 41], for example.) It is important to note
that while the periodicity of the framework needs to be maintained, the periodic lattice determined
by the framework does not need to be preserved as the framework moves. In other words, we are
interested in the flexibility of forced-periodic frameworks with a fully flexible lattice representation.
A mathematical formulation of infinitesimal rigidity for such frameworks has recently been given
in [7] (see also [16, 40]). A summary of the key definitions is provided in Section 25.5.1.
An equivalent alternative description of d-periodic infinitesimal rigidity and the periodic rigidity
matrix has been given in [42, 44]. While in [7] a periodic framework is considered as a realization
of an infinite graph with a periodic group action, in [42, 44] a periodic framework is considered as
a realization of a finite graph on a fundamental domain of Rd (i.e., a flat torus). Since in the latter
description, the orientation of the periodic lattice is fixed, rotations are no longer trivial motions. In
other words, frameworks on a torus are equivalence classes of periodic frameworks (G̃, p̃, L) under
rotation. In particular, this approach has been used to extend results to periodic frameworks with a
fixed or partially flexible lattice representation. This is useful to analyze motions at short time scales
as well as slower deformations of material [34, 43, 45].
The following result is stated for d-periodic frameworks with a fully flexible lattice represen-
tation [29]. However, it can also be adapted to other types of lattice flexibility, including the fixed
lattice [42].

Theorem 25.21 (Periodicity-Preserving Motions) A d-periodic generic framework (G̃, p̃, L) has
a d-periodic infinitesimal flex if and only if (G̃, p̃, L) has a continuous flex which preserves the
periodicity throughout the path.

Using the periodic rigidity matrix (and its modified versions for the various partially flexible and
the fixed lattice representations), it is easy to derive some basic necessary counts for a d-periodic
framework to be d-periodic infinitesimally rigid. We summarize these conditions in the following
theorem (see, for example, [29, 42]). Similar counts also exist for other types of lattice flexibility,
such as a distortional (volume-preserving, but shape-changing) or hydrostatic (shape-preserving, but
volume-changing) lattice deformation [42], as well as for crystallographic frameworks [30, 46, 59].

Theorem 25.22 (Periodic Maxwell Counts) Let (G̃, p̃, L) be a d-periodic infinitesimally rigid
framework. Then the quotient Zd -gain graph of G̃ contains a spanning subgraph (H̃, ψ) with |Ṽ0 |
vertices and |Ẽ0 | edges such that

(a) for L(Zd ) fully flexible: |Ẽ0 | = d|Ṽ0 | + d2 ;

Combinatorial Rigidity of Symmetric and Periodic Frameworks 561

(b) for translation vectors of L(Zd ) allowed to scale independently: |Ẽ0 | = d|Ṽ0 |;
(c) for L(Zd ) fixed: |Ẽ0 | = d|Ṽ0 | − d.

We may also derive further necessary conditions for d-periodic infinitesimal rigidity by consid-
ering all edge-induced subgraphs of (H̃, ψ). However, as in the case of finite symmetric frameworks,
these counts are more complex. In particular, for any edge subset F of (H̃, ψ), we need to take into
account the Zd -rank of F (see also Theorems 25.23 and 25.24 for example).

(0, 1)
(1, 0) (1, 0)

(0, 0) (0, 0)
(1, 0)

(a) (b) (c) (d)

Figure 25.11
Parts of infinite periodic frameworks in the plane (a,c) and their respective quotient Z2 -gain graphs
(b,d). The framework in (a) is infinitesimally flexible under a fully flexible lattice representation, and
the framework in (c) is isostatic under a fixed lattice representation (even though it is disconnected).

25.5.3 Characterizations of Periodic Rigid Graphs

As for finite forced-symmetric frameworks, the quotient Γ-gain graph plays a key role in estab-
lishing combinatorial characterizations of periodic generic infinitesimally rigid frameworks. The
following characterization for periodic frameworks in the plane with a fully flexible lattice represen-
tation has been obtained in [29] using periodic direction networks.

Theorem 25.23 (Periodic Rigidity For The Fully Flexible Lattice) Let (G̃, p̃, L) be a 2-periodic
generic framework. Then (G̃, p̃, L) is 2-periodic infinitesimally rigid if and only if the quotient Z2 -
gain graph of G̃ contains a spanning subgraph (H̃, ψ) with edge set Ẽ0 and vertex set Ṽ0 which
satisfies
(a) |Ẽ0 | = 2|Ṽ0 | + 1;
(b) |F| ≤ 2|V (F)| − 3 + 2k(F) − 2(c(F) − 1) for all F ⊆ Ẽ0 ,
where k(F) is the Z2 -rank of F, and c(F) is the number of connected components of the subgraph
induced by F.

For periodic frameworks in the plane with a fixed lattice representation, we have the following
result, which was obtained using a Henneberg-type inductive construction for the quotient Z2 -gain
graphs [45]. See also [29] for an alternative proof.

Theorem 25.24 (Periodic Rigidity For The Fixed Lattice) Let (G̃, p̃, L) be a 2-periodic generic
framework, where L(Z2 ) is non-singular and has to remain fixed. Then (G̃, p̃, L) is 2-periodic in-
finitesimally rigid if and only if the quotient Z2 -gain graph of G̃ contains a spanning subgraph
(H̃, ψ) with edge set Ẽ0 and vertex set Ṽ0 which satisfies
(a) |Ẽ0 | = 2|Ṽ0 | − 2;
562 References

(b) |F| ≤ 2|V (F)| − 3 for all F ⊆ Ẽ0 with Z2 -rank equal to 0;
(c) |F| ≤ 2|V (F)| − 2 for all F ⊆ Ẽ0 .

The analogous result for periodic frameworks in the plane with a partially flexible lattice repre-
sentation (with one degree of freedom) can be found in [34, 42]. Furthermore, combinatorial charac-
terizations of infinitesimally rigid periodic frameworks in the plane with a fixed-area or fixed-angle
fundamental domain are given in [31].
There also exist several initial results regarding the rigidity of periodic frameworks with ad-
ditional symmetry which are forced to maintain the full crystallographic group of the framework
throughout any motion [30, 43, 46, 59]. In particular, combinatorial characterizations of infinitesi-
mally rigid crystallographic frameworks in the plane with a fully flexible lattice representation are
given in [30] for the case where the group is generated by translations and rotations. For crystallo-
graphic body-bar frameworks with a fixed lattice representation in d-dimensional space, complete
combinatorial characterizations for forced-symmetric infinitesimal rigidity are presented in [59].
One may also seek combinatorial characterizations of infinitesimally rigid periodic frameworks
on a more basic level. In one of the first investigations of rigid periodic structures, it was shown in
[60] that a quotient graph G̃/Γ is the quotient graph of an infinitesimally rigid periodic framework
in d-space with a fixed lattice representation for some gain assignment of the edges of G̃/Γ if and
only if G̃/Γ contains a spanning subgraph which is the union of d edge-disjoint spanning trees. This
has been extended to periodic frameworks with a fully flexible lattice representation in [8].

Theorem 25.25 (Periodic Rigidity for Generic Liftings) A quotient graph G̃/Γ is the quotient
graph of an infinitesimally rigid periodic framework in d-space for some gain assignment of the
edges of G̃/Γ if and only if G̃/Γ contains a spanning subgraph H̃ with edge set Ẽ0 and vertex set Ṽ0
so that
(a) |Ẽ0 | = d|Ṽ0 | + d2 ;

(b) H̃ has a spanning subgraph with d|Ṽ0 | − d edges which has the property that every subgraph
with m edges and n vertices satisfies m ≤ dn − d.

The analogous result for periodic body-bar frameworks in d-space has been established in [9].
As part of the effort to gain insights into “incidentally periodic” infinitesimally rigid frameworks,
recent work has also provided algebraic (and for d = 2 even some combinatorial) characterizations
of d-periodic frameworks which are “ultrarigid,” that is, infinitesimally rigid for any choice of the
periodicity lattice [32].
Finally, we note that very recent work has also established necessary conditions for a d-periodic
generic framework to be globally rigid under fixed lattice representations [23]. Analogous to Hen-
drickon’s theorem for finite frameworks, these necessary conditions consist of a graph connectivity
condition and a redundant rigidity condition. In dimension 2, these conditions have also been con-
firmed to be sufficient, giving a combinatorial characterization of globally rigid 2-periodic generic
frameworks under fixed lattice representations [23]. Extensions of this result to periodic frameworks
under flexible lattice representations have not yet been obtained. There are also no analogous results
yet regarding the (forced-symmetric) global rigidity of symmetric finite frameworks.

References
[1] S.L. Altmann and P. Herzig. Point-Group Theory Tables. Clarendon Press, Oxford, 1994.
References 563

[2] L. Asimov and B. Roth. The Rigidity Of Graphs. Transactions of the AMS, 245:279–289,
1978.
[3] P.W. Atkins, M.S. Child, and C.S.G. Phillips. Tables for group theory. Oxford University
Press, 1970.
[4] G. Badri. Rigidity operators and the flexibility of infinite bar-joint frameworks. Ph.D. thesis,
Department of Mathematics and Statistics, Lancaster University, 2015.
[5] M. Berardi, B. Heeringa, J. Malestein, and L. Theran. Rigid components in fixed-lattice and
cone frameworks. Proc. 23rd Canadian Conference on Computational Geometry, pages 1–6,
2011.
[6] A.R. Berg and T. Jordán. Algorithms for graph rigidity and scene analysis. Proc. 11th Annual
European Symposium on Algorithms (ESA), 2832 of LNCS:78–89, 2003.
[7] C. Borcea and I. Streinu. Periodic frameworks and flexibility. Proceedings of the Royal
Society A, 466:2633–2649, 2010.
[8] C. Borcea and I. Streinu. Minimally rigid periodic graphs. Bull. Lond. Math. Soc.,
43(6):1093–1103, 2011.
[9] C. Borcea, I. Streinu, and S. Tanigawa. Periodic body-and-bar frameworks. Proc. 24th ACM
symposium on Computational Geometry (SoCG2012), pages 347–356, 2012.
[10] O. Bottema. Die Bahnkurven eines merkwürdigen Zwölfstabgetriebes. Österr. Ing.-Arch,
14:218–222, 1960.
[11] R. Connelly, P.W. Fowler, S.D. Guest, B. Schulze, and W. Whiteley. When is a symmetric
pin-jointed framework isostatic? International Journal of Solids and Structures, 46:762–773,
2009.
[12] P.W. Fowler and S.D. Guest. A symmetry extension of Maxwell’s rule for rigidity of frames.
International Journal of Solids and Structures, 37:1793–1804, 2000.
[13] S.D. Guest and P.W. Fowler. A symmetry-extended mobility rule. Mechanism and Machine
Theory, 40:1002–1014, 2005.
[14] S.D. Guest and P.W. Fowler. Symmetry conditions and finite mechanisms. Mechanics of
Materials and Structures, 2(6), 2007.
[15] S.D. Guest and P.W. Fowler. Symmetry-extended counting rules for periodic frameworks.
Phil. Trans. of the Royal Society A, 372(2008), 2014.
[16] S.D. Guest and J.W. Hutchinson. On the determinacy of repetitive structures. Journal of the
Mechanics and Physics of Solids, 51(3):383–391, 2003.
[17] S.D. Guest, B. Schulze, and W. Whiteley. When is a symmetric body-bar structure isostatic?
International Journal of Solids and Structures, 47:2745–2754, 2010.
[18] R. Ikeshita. Infinitesimal rigidity of symmetric frameworks. Master’s thesis, Department of
Mathematical Informatics, The University of Tokyo, 2015.
[19] R. Ikeshita and S. Tanigawa. Count matroids of group-labeled graphs. Combinatorica,
doi.org/10.1007/s00493-016-3469-8, 2017.
[20] T. Jordan, V. Kaszanitzky, and S. Tanigawa. Gain-sparsity and symmetry-forced rigidity in
the plane. Discrete & Computational Geometry, 55:314–372, 2016.
[21] R.D. Kangwai and S.D. Guest. Symmetry-adapted equilibrium matrices. International Jour-
nal of Solids and Structures, 37:1525–1548, 2000.
564 References

[22] R.D. Kangwai, S.D. Guest, and S. Pellegrino. An introduction to the analysis of symmetric
structures. Computers and Structures, 71:671–688, 1999.
[23] V. Kaszanitzky, B. Schulze, and S. Tanigawa. Global rigidity of periodic graphs under fixed-
lattice representations. arXiv:1612.01379, 2016.
[24] D. Kitson and B. Schulze. Maxwell-Laman counts for bar-joint frameworks in normed spaces.
Linear Algebra Appl., 481:313–329, 2015.
[25] D. Kitson and B. Schulze. Symmetric isostatic frameworks with `1 or `∞ distance constraints.
The Electronic Journal of Combinatorics, 23(4):1–13, P4.23, 2016.
[26] A. Lee and I. Streinu. Pebble game algorithms and sparse graphs. Discrete Mathematics,
308:1425–1437, 2008.
[27] J. Malestein and L. Theran. Generic rigidity of frameworks with orientation-preserving crys-
tallographic symmetry. arXiv:1108.2518, 2011.
[28] J. Malestein and L. Theran. Generic rigidity of reflection frameworks. arXiv:1203.2276,
2012.
[29] J. Malestein and L. Theran. Generic combinatorial rigidity of periodic frameworks. Advances
in Mathematics, 233(1):291–331, 2013.
[30] J. Malestein and L. Theran. Frameworks with forced symmetry II: orientation-preserving
crystallographic groups. Geometriae Dedicata, 170(1):219–262, 2014.
[31] J. Malestein and L. Theran. Generic rigidity with forced symmetry and sparse colored graphs.
In: Proc. of Fields Workshop on Rigidity and Symmetry, 2014.
[32] J. Malestein and L. Theran. Ultrarigid periodic frameworks. arXiv:1404.2319, 2014.
[33] J. Malestein and L. Theran. Frameworks with forced symmetry I: reflections and rotations.
Discrete & Computational Geometry, 54(2):339–367, 2015.
[34] A. Nixon and E. Ross. Periodic rigidity on a variable torus using inductive constructions. The
Electronic Journal of Combinatorics, 22(1):P1, 2015.
[35] A. Nixon and B. Schulze. Symmetry-forced rigidity of frameworks on surfaces. Geometriae
Dedicata, 182(1):163–201, 2016.
[36] J.C. Owen and S.C. Power. Frameworks, symmetry and rigidity. Int. J. Comput. Geom. Appl.,
20:723–750, 2010.
[37] J.C. Owen and S.C. Power. Infinite bar-joint frameworks, crystals and operator theory. New
York Journal of Mathematics, 17:445–490, 2011.
[38] J. Porta, L. Ros, B. Schulze, A. Sljoka, and W. Whiteley. On the symmetric molecular con-
jectures. Computational Kinematics, Mechanisms and Machine Science, 15:175–184, 2014.
[39] S.C. Power. Crystal frameworks, matrix-valued functions and rigidity operators. In Concrete
Operators, Spectral Theory, Operators in Harmonic Analysis and Approximation, pages 405–
420. Springer, 2014.
[40] S.C. Power. Crystal frameworks, symmetry and affinely periodic flexes. New York Journal of
Mathematics, 20:665–693, 2014.
[41] S.C. Power. Polynomials for crystal frameworks and the rigid unit mode spectrum. Philo-
sophical Transactions of the Royal Society A, 372(2008), 2014.
[42] E. Ross. The Rigidity of Periodic Frameworks as Graphs on a Torus. Ph.D. thesis, Department
of Mathematics and Statistics, York University, 2011.
References 565

[43] E. Ross. The rigidity of periodic body-bar frameworks on the three-dimensional fixed torus.
Phil. Trans. of the Royal Society A, 372(2008), 2014.
[44] E. Ross. The rigidity of periodic frameworks as graphs on a fixed torus. Contributions to
Discrete Mathematics, 9(1), 2014.
[45] E. Ross. Inductive constructions for frameworks on a two-dimensional fixed torus. Discrete
& Computational Geometry, 54(1):78–109, 2015.
[46] E. Ross, B. Schulze, and W. Whiteley. Finite motions from periodic frameworks with added
symmetry. International Journal of Solids and Structures, 48:1711–1729, 2011.
[47] B. Schulze. Combinatorial and Geometric Rigidity with Symmetry Constraints. Ph.D. thesis,
Department of Mathematics and Statistics, York University, 2009.
[48] B. Schulze. Block-diagonalized rigidity matrices of symmetric frameworks and applications.
Beiträge zur Algebra und Geometrie, 51(2):427–466, 2010.
[49] B. Schulze. Injective and non-injective realizations with symmetry. Contributions to Discrete
Mathematics, 5:59–89, 2010.
[50] B. Schulze. Symmetric Laman theorems for the groups c2 and cs . The Electronic Journal of
Combinatorics, 17(1):1–61, R154, 2010.
[51] B. Schulze. Symmetric versions of Laman’s Theorem. Discrete & Computational Geometry,
44(4):946–972, 2010.
[52] B. Schulze. Symmetry as a sufficient condition for a finite flex. SIAM Journal on Discrete
Mathematics, 24(4):1291–1312, 2010.
[53] B. Schulze, P.W. Fowler, and S.D. Guest. When is a symmetric body-hinge structure isostatic?
International Journal of Solids and Structures, 51:2157–2166, 2014.
[54] B. Schulze and S. Tanigawa. Linking rigid bodies symmetrically. European Journal of Com-
binatorics, 42:145–166, 2014.
[55] B. Schulze and S. Tanigawa. Infinitesimal rigidity of symmetric frameworks. SIAM Journal
on Discrete Mathematics, 29(3):1259–1286, 2015.
[56] B. Schulze and W. Whiteley. The orbit rigidity matrix of a symmetric framework. Discrete &
Computational Geometry, 46(3):561–598, 2011.
[57] B. Schulze and W. Whiteley. Coning, symmetry, and spherical frameworks. Discrete &
Computational Geometry, 48(3):622–657, 2012.
[58] P.D. Seymour. A note on hyperplane generation. J. Combin. Theory Ser. B, 61:88–91, 1994.
[59] S. Tanigawa. Matroids of Gain Graphs in Applied Discrete Geometry. Trans. Amer. Math.
Soc., 367:8597–8641, 2015.
[60] W. Whiteley. The union of matroids and the rigidity of frameworks. SIAM J. Discrete Math.,
1(2):237–255, 1988.
[61] W. Whiteley. Some Matroids from Discrete Applied Geometry. Contemporary Mathematics,
197:171–311, 1996.
Index

(2, 2, 1)-gain-tight, 539 k-trees, 235

(2, 3)-circuit, 415 l1 -diversity, 221
(2, 3)-sparse, 415 l1 -metric, 214
(2, 3)-tight, 415 l1 -qusi-semimetric, 223
(2k + 1)-gonal inequality, 217 l p -metric, 214
(d, d)-tight, 536 l p -norm, 213
(k, `, m)-gain-sparse, 550 l p -space, 213
(k, `)-sparse, 415 n-over-constrained, 195
(k, `)-sparsity, 437 n-skeleton, 296
matroid, 453 RDL -circuit, 481
(k, `)-tight, 415 RDL -connected, 482
(k, `)-tightness, 437 Rd -bridge, 467
(k, `, m)-gain-sparse, 424 Rd -circuit, 467
(k, `, m)-gain-tight, 424 Rd -component, 467
0-extension, 394 trivial, 467
1-extendability property, 389, 395 Rd -connected, 467
1-extension, 394 Rd -independent, 467
2-skeleton, 290 0-extension, 386, 415
2-sum, 391 0-extensions, 536
2-tree decomposition, 397, 408 0-reduction, 415
4-horn, 294, 296 1-extension, 389, 415
M-connected, 416 1-extensions, 536
S-gain graph, 365 1-path graph, 248
S-regular, 365 1-reduction, 415
S-symmetric, 364 2-sum, 416
S-symmetric framework, 365 2-sum operation, 420
V -replacement, 415, 417 3-tree decomposition, 398
X-replacement, 415 3D GCS solvers, 169
Γ-generic, 545 6R linkage, see linkage
Γ-symmetric framework, 544
Γ-symmetric graph, 544 abstract rigidity matroid, 386, 395
Γ-gain graph, 424 abstract tensegrity polygon, 309
Zd -rank of an edge set, 560 admissible, 415
T -vanishing polynomial, 71 admissible order, 70
ρ0 -generic, 551 advanced algebraic invariant, 127
ρt -generic, 555 affine motions, 202
d-flattenability, 238 affine representation, 124
d-periodic framework, 559 algebraically independent, 10
d-periodic generic, 560 angle
d-periodic graph, 559 construction axiom, 35
d-periodic infinitesimal motion, 559 full-angle, 30
d-periodic infinitesimally rigid, 559 measure of, 42
k-plate-bar framework, 444 sum of, 42
anti-parallelogram, 262

567
568 Index

antisymmetrization of rational monomial invari- bipartite

ant, 78 naturally, 491
Archimedes’s axiom, 40 block and hole polyhedra, 417
area body pin graph, 408
method, 28 body-and-cad, 506
of a circle, 43 body-bar
assembly mode, 261, 269 framework, 436
associated framing, 336 generic, 437
Assur components, 403 global rigidity, 436
Assur decomposition, 186 graph, 437
Assur Graphs, 186 infinitesimal rigidity, 436
Assur scheme, 402 rigidity, 436
atomic extensor, 101, 103, 104 rigidity matrix, 439
augmentation problem, 483 body-bar framework, 422
axiom body-bar-hinge framework, 442
Archimedes’s axiom, 40 body-hinge
Circle-Circle continuity, 39 framework, 440
congruence axioms, 35 identified, 444
continuity, 39 generic, 440
Dedekind’s axiom, 39 global rigidity, 440
elementary continuity, 39 graph, 440
five-segments axiom, 37 infinitesimal rigidity, 440
Hilbert’s axioms, 30, 35 rigidity, 440
Hilbert’s continuity, 40 skeleton, 451
incidence axioms, 35 body-hinge framework, 422
Line-Circle continuity, 39 body-pin framework, 444
order axioms, 35 bond diagram, 265
parallel postulate, 36, 41 bond theory, 265
Pasch, 37, 44 Boole problem, 214
Playfair’s axiom, 36 Boolean quadric polytope, 220
protractor postulate, 42 boolean quadric polytope, 214
ruler and compass, 39 Borsuk, 34
Segment-Circle continuity, 39 brace, 385
Tarski’s axioms, 30, 37 bracket, 62, 63, 67
bracket algebra, 68
balanced edge set, 550 bracket monomial, 87, 96, 101, 102, 105
balanced gain graph, 424 bracket polynomial, 67
bar, 385 bracket ring, 68
bar framework, 201 brackets, 86, 87, 91, 92, 94, 96, 97, 99, 103
Baranovskii cone, 217 Bricard octahedron, 290, 296
barycentric coordinates, 32 brick, 453
basic algebraic invariant, 127 -partition, 455
basis, 384 Buchberger’s Algorithm, 276
Bell inequality, 220
bellows conjecture, 293 cable, 300
Bennett linkage, see linkage CAD, 428, 505
Bertini’s Theorem, 279 cad constraint, 510
betweenness axioms, 37 Cayley
binary supermetric cone, 226 Cayley configuration space, 235
binomial Cayley expansion, 73 Cayley parameters, 234
binomial proof, 129 low Cayley complexity, 248
Index 569

Cayley algebra, extended, 322 coning, 348

Cayley bracket, 73 connected matroid, 397
Cayley configuration space of G with respect to connectivity, 453
the nonedge set F, 234 Connelly sphere, 293, 296
Cayley configuration spaces, 428 constellation, 214
Cayley expansion, 73 constrainedness, 10
Cayley factorization, 75 consistently over-constrained, 10
Cayley-Menger determinant, 108, 128 generically over-constrained, 10
Ceva’s Theorem, 78 generically under-over-constrained, 10
chiral molecule, 112 generically well-over-constrained, 10
chirality constraint, 110 over-constrained, 10
chordal graphs, 209 under-constrained, 10
Chou, 28 well-constrained, 10
Church-Rosser property, 155 constraint
cleavage unit, 476 angle, 488
Clifford algebra, 126 distance, 488
Clifford bracket, 127 constraint graph, 2
Clifford bracket algebra, 127 constraint graph (CG), 142
closure, 384 constraint-graph, 488
cocycle, 395 constructible, 279
collinear, 63 convex hull, 44
collineation, 66 convex realisation, 309
columnwise reduction, 70 correlation cone, 214
columnwise straightening, 70 correlation polytope, 220
commuting time metric, 223 count matroid, 453
completable graph, 173 countably infinite graphs, 425
complete bipartite graphs, 209 covariance map, 220
completed graph, 173 Crapo’s binomial, 75
completion, 172 cross ratio, 79
completion of rational monomial invariant, 77 cut cone, 214, 215
complex, 349 cut diversity, 221
complex geometric objects, 148, 150 cut diversity cone, 221
complex numbers, 275 cut polytope, 215
compound geometric elements, 148, 150 cut polytope of a graph, 218
compound geometric objects, 148, 150 cut semimetric, 214, 215
computer-aided design, 488 cut subgraph, 215
concurrent, 63 cycle, 384
concyclicity, 79 cycle inequality, 219
cone framework, 352
cone graph, 352 de Sitter space, 364
configuration matrix, 203 decomposition-recombination, 5, 182
configuration space, 378, 385 Dedekind’s continuity axiom, 39
conformal geometric algebra, 126 deductive database method, 24
conformal model, 122 deficit, 140, 144
conformal point at infinity, 122 degree of a variety, 279
conformation, 110 degree of freedom, 385, 386, 392
conformation space, 110 degrees of freedom, 150
conformational ensemble, 115 Dehn invariant, 296
congruence axioms, 37 Delaunay triangulation, 44
congruent frameworks, 201 Delta-Robot, 266, 268
conic at infinity, 305, 344, 367 Denavit-Hartenberg parameters, 262, 263, 264
570 Index

dependence, 384 inverse, 259

Desargues, 488 norm, 259
Desargues’ Theorem, 76 Dulmage-Mendelsohn, 185
Desargues’s theorem, 33, 41 dynamic maintenance, 196
dimension, 279
lower dimension axiom, 37 edge elongation, 382
upper dimension axiom, 37 edge pinch, 422
dimensional rigidity, 202 edge-connectivity, 452
dimensionally rigid, 305, 367 edge-redundant rigidity, 453
direct kinematics, see kinematics edge-splitting, 416
direction-length framework, 480 EMBED algorithm, 115
congruent, 480 end-effector, 255–256, 262–266
equivalent, 480 energy function, 305
globally rigid, 480 equation graph, 185
rigid, 480 equilibriated force, 383
direction-length graph, 425 equilibrium load, 301
direction-length graphs, 428 equilibrium self-stress, 11
distance geometry, 108 equilibrium stress, 202, 301, 353, 383, 463
distance or metric cone equilibrium stress matrix, 343
l pp distance cone, 240 equilibrium stress vector, 343
projections of the cone, 240 equivalent frameworks, 201
strata of the cone, 240 Euclid
Diversity cone, 221 Elements, 40
dominates, 300 fifth postulate, 36, 37
dotted, 92 Euclidean distance constraint system, 233
double 1-extensions, 428 Euclidean distance matrix, 4, 214
double banana, 519 Euclidean metric, 214
double-V conjecture, 417 Euler
DR-plan, 182 angles, 257
canonical DR-plan, 188 parameters, 257
cluster minimality, 183 Euler’s formula, 43
fan-in, 182 extensor, 98–101, 103, 104
home coordinate system, 184 extreme edges, 247
indecomposable system, 183 extreme graphs, 247
irreducible system, 185
optimal, 182 facial reduction, 205
proper vertex-maximality, 183 factorization, 265, 266
pseudosequential DR-plan, 189 feasible, 415
shape recognition, 185 final polynomial, 71
size, 182 finite forbidden minor, 235
DR-planner First Fundamental Theorem of Invariant Theory,
dense, 187 67
Frontier, 187 first order rigid, 385
cluster minimality, 188 flex, 380
dual affine representation, 124 force-load, 320
dual Minkowski representation, 124 forced symmetric infinitesimal flex, 365
dual number, 259 forced symmetrically rigid, 365
inverse, 259 forced symmetry, 425
dual quaternion, 259–260 four color theorem, 43
action on point, 260 four-bar linkage, see linkage
conjugate, 259 four-cycle, 248
Index 571

four-horn, 294, 296 Gelernter, 24

frame, 499 general, 278
fixed, 254, 255, 257, 260, 261, 266 body-and-cad, 517
moving, 254, 255, 257, 260, 261 general Bell inequality, 220
framed cycles in general position, 331 general position, 385
framework, 6, 320, 378, 385, 462 generalized maximum matching, 185
bar-joint framework, 6 generalized network flow, 185
body-and-cad, 506, 507, 509, 510 generic, 278, 341, 385
body-hinge, 8 body-bar framework, 437
body-pin, 473 body-hinge, 440
congruent, 462 generic embedding, 380
degeneracy, 11 generic framework, 207
degenerate, 493 generic property, 12, 462
degree-of-freedom, 11 generic rigidity, 385
dependent, 12 generic rigidity matroid, 13, 395
direction congruent, 479 genericity
direction-length, 503 1-Dof tree-decomposable linkage, 247
equivalent, 462 cluster, 247
fixed-slope, 495 convex Cayley configuration, 242
flexible, 11 generic d-flattenability, 242
generic, 462 generic property, 242
generic framework, 6 geometric algebra, 44, 126
globally rigid, 11, 462 geometric conditions, strong, 327
independent, 12 geometric conditions, system of, 324
infinitesimal flexes, 11 geometric constraint system, 377
infinitesimal motions, 11 Euclidean bar-and-joint, 2
infinitesimally rigid, 11 Euclidean distance constraint, 2
isostatic, 11 realization, 2
minimally rigid, 11 solution, 2
parallel, 479 geometric constraint system (GCS), 140, 142
point-hyperplane, 502 grounded, 2
point-line, 488, 492 pinned, 2
regular valued, 477 geometric constraint systems (GCS), 245
rigid, 10, 462 geometric primitive, 2
strongly rigid, 11 body, 8
tensegrity framework, 7 hinge, 8
tight, 480 global rigidity, 378
framework in general position, 326 body-bar, 436
framework:bar and joint, 377 body-hinge, 440
frameworks on surfaces, 426 globally flexible, 341
Fubini number, 222 globally linked, 474
full-angle method, 30 globally loose, 474
fully Γ-symmetric infinitesimal motion, 551 globally rigid, 305, 341
fully Γ-symmetric infinitesimally rigid, 551 body-and-cad, 511
fully Γ-symmetric self-stress, 551 globally rigid cluster, 474
globally rigid framework, 366
gain, 550 globally rigid graph, 367
gain graph, 539 Gröbner basis, 44, 276
Gale matrix, 203 Grübler count, 408
Gale transform, 203 grafting operation, 423
Gao, 28 graph
572 Index

k-rigid, 470 hemimetric polytope, 226

cad, 508, 510 Hendrickson
globally rigid, 462 theorem, 450
nearly 3-connected, 469 Hendrickson graph, 464
point-line, 488, 490 Henneberg construction, 8
primitive cad, 508 Henneberg moves, 416
random, 472 Henneberg-1 move, 386, 394
realization, 462 Henneberg-2 move, 389, 394
redundantly rigid, 463 Heron’s formula, 42
redundantly rigid component, 469 Hessenberg, 41
trivial, 469 Hilbert, 34
rigid, 462 axioms, 30
rigid component, 469 continuity axiom, 40
sparse Hilbert basis theorem, 275
(k, `)-, 518 hitting time quasi-metric, 223
[a, b]-, 518 homogeneous bracket polynomial, 70
tight, 480 homogeneous coordinates, 62, 87, 88, 90
(k, `)-, 518 homogeneous model, 124
[a, b]-, 518 homogeneous transformation matrix, 255, 257,
vertex-labelled, 487 263
vertex-redundantly rigid, 470 hyper-maps, 43
graph k-sum, 235 hyperbolic, 349
graph completion, 173 Hyperbolic space, 364
graph connectivity hypermetric, 214
k-connected, 463 hypermetric cone, 216
6-mixed connected, 470 hypermetric inequality, 217
cyclically k-edge-connected, 470 hypermetric polytope, 217
graph minor, 235
graph non-edge, 187 ideal, 275
graph operation implied constraint, 384
X-replacement, 466 incenter, 42
0-extension, 475 incidence geometry, 61
1-extension, 464 incidence relation, 62
2-sum, 467 incidence theorem, 63
cone, 464 incidence tree, 192
diamond-split, 466 incidental symmetry, 425
vertex split, 464 independence, 384
Grassmann-Cayley algebra, 44, 63, 85, 87, 97, independent
98, 100, 101 body-and-cad, 517
Grassmann-Plücker ideal, 68 body-and-cad, generically, 517
Grassmann-Plücker relation, 67 induced diversity metric, 221
Grassmannian, 90, 91, 94, 99 inductive construction, 414
gravity center, 42 inductive geometric reasoning, 115
Greenberg, 34 infinitesimal flex, 300, 352, 526
Gröbner basis, 42 infinitesimal motion, 385
Gupta, 37 body-and-cad, 515
infinitesimal rigid matroid, 384
Hales, 43 infinitesimal rigidity, 379
Hartshorne, 34 body-bar, 436
hemimetric, 215 body-hinge, 440
hemimetric cone, 226 infinitesimal rigidity matroid, 395
Index 573

infinitesimally rigid, 300, 352 law of sines, 42

body-and-cad, 515, 516 Leggett-Garg inequality, 221
body-and-cad, generically, 517 linearization, 11
instance solvability, 144 link, 254, 385
intended solution, 161 linkage, 2, 233, 253–256, 260, 261, 263, 266,
interactive GCS solver, 162 385
intersecting chord theorem, 42 d-directed orientation, 186
invariant division, 71 1-Dof tree-decomposable, 247
invariant ratio, 77 anti-parallelogram, 262
invariant remainder, 71 Bennett, 262, 265
invariant top reduction, 71 closed 6R, 265, 266
inverse kinematics, see kinematics degree-of-freedom(Dof), 246
isogonal, 292 deltoid, 261
isometric embeddable, 213 extreme linkages, 247
isostatic framework, 395 four-bar, 255, 256, 260, 261, 265
free, 186
Jessen icosahedron, 295, 296 Goldberg, 262, 265, 266
join, 63, 98, 100 grounded graph, 186
joint, 254, 254, 266, 268, 385 indecomposable system, 191
Cardan, 254 parallel, see mechanism
cylindrical, 254 parallelogram, 261
helical, 254 pin, 186
limit, 264 pinned graph, 186
parameter, 262–264, 267 QRS, 247
planar, 254 serial, see mechanism
prismatic, 254, 263, 265, 267, 268 strongly d-Assur, 187
revolute, 254, 256, 260, 262–267 synthesis, 253, 263, 264–266
spherical, 254, 267, 268 tree decomposable, 247
universal, 254 linkages, 176
Jordan curve theorem, 43 mechanisms, 186
linkgraph, 254, 254, 255, 262, 266
K4 contraction, 427 locally closed, 279
Kepler locally rigid, 300
conjecture, 43
proposition, 43 manipulator, see mechanism
Kimberling, 42 mass point method, 32
kinematics matrix completion, 428
direct, 261, 262, 264, 267–269 matroid
inverse, 263, 264, 267–268 circuit axiom, 497
Knuth’s axioms, 44 component, 467
connected, 467
Laguerre’s formula, 80 direct sum, 467
Laman counts, 14 fixed-slope, 495
Laman graph, 148, 155 point-line, 496
Laman graphs, 7 rigidity, 494
Laman plus one graph, 416 union, 498
Laman’s condition, 389, 395 matroid: connected, 397
Laman’s theorem, 148, 155, 389 maximum-cut problem, 214
last level vertex, 248 Maxwell, 536
lateration graph, 464 Maxwell condition
law of cosines, 42 body-bar, 439
574 Index

body-hinge, 442 uniquely localizable, 463

mechanism, 400, 408 Newton, 43
3-RPR, 261, 267, 267, 268, 269 node, 385
cable driven, 269 non-degeneracy condition, 63
parallel, 254, 254, 266–267 non-degeneracy monomial, 71
redundant, 254, 262, 263 normal form, 69, 92, 93, 97
serial, 254, 254, 262–266 null bracket algebra, 127
singularity, 253, 263, 266–269 null vector, 122
mechanism science, 253–269 Nullstellensatz, 275
mechanisms, 428 numerical algebraic geometry, 277
meet, 63, 97, 98, 100, 101, 126
Menelaus’ Theorem, 78 optimal Cayley modification (OCM), 245
Menger’s Theorem, 109 Optimal Well-Formed Incidence Selection Algo-
metric, 213 rithm, 194
metric affine space, 107 orbit rigidity matrix, 550
metric cone, 214, 215, 218 ordered Bell number, 222
metric polytope, 215 orientations, 196
metric polytope of a graph, 218 oriented l1 -norm, 223
metric space, 107, 213 oriented cut, 215
metric vector space, 107 oriented cut cone, 222
metrization, 115 oriented cut polytope, 223
minimal complete Cayley distance, 248 oriented multicut, 222
minimal complete Cayley vector, 248 oriented multicut cone, 222
minimally rigid oriented multicut polytope, 223
body-and-cad, 511 oriented switching, 223
Minkowski representation, 124 oriented triangle inequality, 222
Minkowski space, 363 Origami constructions, 43
minor closed, 235 orthocenter, 42
minor of a graph, 218 overbraced framework, 384
Miquel’s Theorem, 81, 130, 131 overconstrained, 144
mixed graph, 480 overlap graph, 192
1-extension, 481
direction balanced, 480 panel-hinge framework, 444
direction-length rigid, 480 Pappus’ Theorem, 64
molecular conjecture, 9, 423 parallel postulate, 37, 42
molecular framework, 422 parallelogram, 261, 265
monic Clifford monomial, 126 partial metric, 215
monomial Cayley expansion, 73 partial semimetric, 224
monomial proof, 129 partial semimetric convex body, 224
Morley’s theorem, 41 partition-connectivity, 453
motion, 385 Pascal’s Conic Theorem, 66
motion polynomial, 266 Pasch’s axiom, 35, 37, 44
Moulton’s plane, 41 path planning, 266
multicut, 215 Peano, 37
multicut hemimetric cone, 226 Peaucellier, 490
pebble game, 400, 501
Nash-Williams, 536 [a, b]-, 519
negative column-deglex order, 70 pebble games, 14
negative vector, 122 perimeter inequality, 215
network periodic frameworks, 425
localization problem, 463 phase-symmetric infinitesimal motion, 554
Index 575

phase-symmetric infinitesimally rigid, 555 quantitative extension of geometric theorem,

phase-symmetric orbit rigidity matrix, 555 130
π, 43 quasi generically overconstrained, 143
pinching, 453 quasi generically underconstrained, 143
pinned framework, 381, 385, 401, 408 quasi-generic, 367
pinned framework conditions, 401 quasi-hypermetric cone, 223
pinned isostatic, 401 quasi-metric, 215
pinned vertices, 400 quasi-semimetric, 222
pinning problem quasi-semimetric polytope, 222
R2 -connected, 483 quasi-stress matrix, 205
globally rigid, 482 quaternion, 255, 258–259
Plücker, 87, 89, 90 action on vector, 259
Plücker coordinates, 86, 90, 91 conjugate, 259
Plücker relation, 92, 95, 96 dual, 254, 255, 258
Plücker relations, 94 inverse, 259
Plücker coordinates, 266 norm, 259
placement of geometric elements, 159 quotient Γ-gain graph, 550
Playfair’s axiom, 36, 37
points in general position, 207 r-component, 392
point model, 255 radically solvable, 428
polyhedral frameworks, 6 rank, 384
polyhedral norm, 525 rank function, 501
polyhedral surface, 293, 296 rational Cayley factorization, 75, 76
polyhedron: strictly convex, 289, 297 rational monomial invariant, 77
positive vector, 122 reachability, 175
Positivestellensatz, 283 real algebraic geometry, 280
prestress, 383 real algebraic variety, 280
prestress stable, 301 realization
primitive cad constraint, 508, 511 body-and-cad, 510
angular, 508 realization type, 247
blind, 508 reduced meet product, 126
projective invariance, 66 reduced rigidity matrix, 382
projective rigidity matrix, 361 redundant constraint, 384
projective space, 86, 88, 89, 91, 93 redundant rigidity, 450
projective transformation, 353 vertex-, 451
proper 2-tree decomposition, 397, 408 redundantly rigid, 345, 396, 415
proper configuration, 66 redundant mechanism, see mechanism
proper stress, 301 regular multidistance cone, 227
protractor postulate, 42 regular points, 536
pseudo Euclidean, 349 relaxed symmetry, 223
pseudo-Euclidean semimetric, 109 repartitioning polytope, 217
Pseudo-Euclidean space, 363 representation
Ptolemy’s inequality, 110 linear, 500
Ptolemy’s Theorem, 128 resolution, 301
Ptolemy’s theorem, 42 resolution scheme of a graph, 333
pure condition resolution scheme of a tree, 333
body-and-cad, 517 resolution scheme, completely generic, 334
Pythagoras’s theorem, 42 resolvable force, 383
resolvable stress, 383
quadratically solvable, 148, 156 resultant force, 383
quadratically-radically solvable(QRS), 247 reversal, 223
576 Index

ridge graph, 215 sharp triangle inequality, 224

rigid, 6, 392 shuffle formula, 73
Rigid Body, 148 simple Grassmann-Cayley expression, 101
rigid body, 150, 157 simplex inequality, 226
rigid component, 447 Simson’s Theorem, 132
rigid framework, 378, 385 Simson’s theorem, 31
rigid:edge k-rigid, 396 singularity, 260, 261
rigid:vertex k-rigid, 396 slicing, 367
Rigidity, 148 sliding, 353
rigidity, 155, 158 smooth, 279
body-bar, 436 solution of a GCS, 148
body-hinge, 440 sparsity, 14
infinitesimal rigidity, 10 sparsity matroid, 14
map, 492 special minimally rigid graph, 476
rigidity matrix, 11 sphere, 349
rigidity matroid, 10, 12 Spherical framework, 352
Rigidity function, 378 Spherical infinitesimal flex, 352
rigidity map, 527 Spherical rigidity matrix, 355
rigidity matrix, 352, 380, 385, 467, 528 spider web, 310
body-and-cad, 515 square of a graph, 470
body-and-cad, generic, 517 standard embedding, 62
body-bar, 439 static rigidity, 383
rigidity matroid, 467 statically rigid, 301
robotic formations, 428 Stewart-Gough platform, 266, 267, 268, 269
robust proof, 129 stiffness, 253, 268
root identification, 144, 158, 160–162, 164 straight, 69
root identification problem, 162 straight bracket monomial, 69
root selection, 144, 158, 160–162, 164 straightening algorithm, 69
rooted (k, `)-arc-connectivity, 453 strain, 382
row-deglex order, 70 stress, 300, 383
ruler and compass stress matrix, 203, 305, 367, 463
formalization, 40 maximum rank, 464
strict proper stress, 301
Schoenberg’s quadratic form, 109 strongly connected component, 185
Schoenberg’s Theorem, 109 strongly connected graph, 185
Schwabhauser, 34 strut, 300
scissors congruent, 296 Study
seam matroid, 191 condition, 255, 257, 259, 260
Second Fundamental Theorem of Invariant The- kinematic map, 256–258, 267
ory, 68 parameters, 254, 256, 257–260
second point of intersection, 126 quadric, 256, 257, 260, 264, 266
second-order flex, 301 subgraph
second-order rigid, 301 rigid, 490
self-motion, 267, 268–269 submodular, 398, 408, 496, 497
self-stress, 320 submodular function, 408
semi-algebraic set, 282 subvariety, 278
semimetric, 213 super stable, 305, 367
semimetric space, 107 superbrick, 453
separable, 220 -partition, 455
serializable, 151 supermetric, 215
serializable GCS, 145, 151 supermetric cone, 225
Index 577

surgery, on a graph, 328 trivial subgraph, 188

surgery, on a resolution scheme, 334
switching, 216 unbalanced circuit, 555
symmetry, 527, 531 unbalanced edge set, 550
symmetry group, 544 underconstrained GSC, 171
synthesis underlying bar framework, 301
see linkage synthesis, 253 unit ball, 214
synthetic geometry, 34 unit disk framework, 470
Szmielew, 37 unit disk graph, 470
universal rigidity, 201
tableau, 93, 96, 103 universal tensegrity stratification, 321
tableau form, 69 universally rigid, 305, 367
Tarski, 34
axioms, 30, 37 van der Waerden relation, 69
Tarski-Seidenberg Theorem, 282 Van der Waerden syzygies, 86, 92
tensegrity, 320 variational GCS, 145, 153
tensegrity framework, 300 variety, 275
tensegrity polygon, 309 Varignon’s theorem, 29, 44
tensegrity, non-parallelizable, 326 vertex addition, 416
Thales’s circle theorem, 42 vertex cocycles, 388, 395
theorem vertex displacement, 382
Desargues, 41 vertex split, 394
four color, 43 vertex splitting, 415, 417, 536
Heron’s formula, 42 vertex-redundant rigidity, 421
intersecting chord, 42 vertex-to-K4 move, 536
Jordan curve, 43 vertex-to-4-cycle, 415, 417
Kepler conjecture, 43
Morley, 41 weightable quasi-semimetric, 222
Pappus, 38 weightable quasi-semimetric polytope, 222
Pascal, 41 workspace, 264, 269
Ptolemy, 42 Wu’s method, 42, 44
Pythagoras, 42
Young’s straightening algorithm, 69
segment addition, 35
side-angle-side, 35 Zariski closure, 278, 281
Simson, 31 Zariski topology, 278, 282
Thales’s circle, 42 zeolite, 473
Varignon, 29, 44 Zhang, 28
three-party, 221
tight Bell inequality, 220
transformation
affine, 494
tree-connectivity, 453
Triangle centers, 42
triangle decomposable, 148, 153, 155, 173
triangle exchange, 423
triangle inequality, 213
triangle inequality limit, 111
triangulated surfaces, 417
triangulation, 291
trilateration graphs, 208
trivial flexes, 379