Parallel Coordinates

Visual Multidimensional Geometry

and Its Applications
 To

Hadassah, Dona and Haim, Avigail, Ofra, Yuval, Akiva

... a wonderful family

and

Bernie Dismdale, Lew Leeburg, Alex Hurwitz,

Misha Boz, Baxter Armstrong, Ralph Alterowitz
... great friends and colleagues

and

Heinz von Foerster

... Mischiefologist, Magician,
Cybernetician, and Teacher
 Parallel Coordinates

Visual Multidimensional Geometry

and Its Applications

with 230 color illustrations

Alfred Inselberg
Foreword by Ben Shneiderman

CD-ROM
Alfred Inselberg
Tel Aviv University
School of Mathematical Sciences
69 978 Tel Aviv
Israel
[email protected]

ISBN 978-0-387-21507-5 e-ISBN 978-0-387-68628-8

DOI 10.1007/978-0-387-68628-8
Springer Dordrecht Heidelberg London New York

Library of Congress Control Number: 2008927731

Mathematics Subject Classification (2000): 15-xx, 51-xx, 62-xx, 62-09, 68-xx, 14J26, 14J29, 14J70: 15-01, 15A04, 15A06;
51-01; 51A05, 54-01; 68-01, 68P01, 68W01; 90-01, 90-08, 90C05, 90C29, 90C59

© Springer Science+Business Media, LLC 2009

All rights reserved. This work may not be translated or copied in whole or in part without the written permission of
the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief
excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage
and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter
developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as
such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights.

Printed on acid-free paper

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

0

Foreword

The breakthrough idea of parallel coordinates has enchanted many people with its
cleverness and power. Al Inselberg, the source of that cleverness and power, finally
shares the depth and breadth of his invention in this historically important book.
I’ve been waiting impatiently for this exposition since I was captivated by
Inselberg’s lecture at the University of Maryland in 1979. He had already understood
the potential parallel coordinates has for generating insights into high-dimensional
analysis and information visualization. In the following decades he polished the
arguments, built effective software, and demonstrated value in important applica-
tions. Now a broad community of readers can benefit from his insights and effective
presentation.
I believe that Inselberg’s parallel coordinates is a transformational ideal that
matches the importance of René Descartes’ (1596–1650) invention of Cartesian
coordinates. Just as Cartesian coordinates help us understand 2D and 3D geometry,
parallel coordinates offer fresh ways of thinking about and proving theorems in
higher-dimensional geometries. At the same time they will lead to more powerful
tools for solving practical problems in a wide variety of applications. It is rare to
encounter such a mind-shattering idea with such historic importance.
While Inselberg’s insight and exposition opens the door to many discoveries,
there is much work to be done for generations of mathematicians, computer scien-
tists, programmers, and domain experts who will need to build on these innovative
ideas.
To understand and apply parallel coordinates, many further innovations are
needed. Those who can readily reshape their perceptual skills and realign their
cognitive frames will be well-equipped for this nontrivial task. I still struggle with
these novelties, which Inselberg has masterfully accomplished. In each of my dozen
meetings with Inselberg, he relieved me of my confusion and steered me closer to
clarity. There is a chance that those who learn parallel coordinates early in life
may be able to more easily see higher-dimensional spaces. More likely, master
teachers, inspired interface designers, and gifted domain experts will enable future
generations to grasp these profound ideas as if they were entirely natural, just as
we see Cartesian coordinates.
I encourage readers to read each paragraph carefully and make sure they fully
grasp the ideas. Inselberg has made this possible by his lucid and often charming
vi Foreword

prose, filled with intriguing historical references and clever quips. The computer
scientists, programmers, dataminers, and information visualization experts may
want to jump ahead to Chapter 10 which has practical examples and numerous
screen shots. The recent additional insights from Inselberg’s students and followers
give a hint of the thousands of papers that are likely to be inspired by this work. More
importantly, the applications of parallel coordinates could contribute to solving
some of the complex multidimensional problems of our time. I hope every reader
will strive to put these ideas to work in making a better world.

Ben Shneiderman
University of Maryland, College Park
May 2008
0

Contents

Foreword . . . . . . . . . . . . . . . . . . . . . . . . . v
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . xv
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . xxiii
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Multidimensional Visualization . . . . . . . . . . . 1

1.2 ♣ FT-1 How? . . . . . . . . . . . . . . . . . . 2
2 Geometry Background . . . . . . . . . . . . . . . . . . 7

2.1 Preliminaries . . . . . . . . . . . . . . . . . . . 7
2.1.1 In the Beginning ... . . . . . . . . . . . . 7
2.2 ♣ FT-1 Why Projective Geometry? . . . . . . . . . 9
2.2.1 Projective Plane Model . . . . . . . . . . . 10
2.2.2 Axioms for Projective Geometry . . . . . . . 13
2.2.3 Principle of Duality . . . . . . . . . . . . 16
2.3 ** Finite Geometries . . . . . . . . . . . . . . . . 19
2.3.1 Seven-Point Geometry . . . . . . . . . . . 19
2.3.2 Finite Geometry with 13 Points . . . . . . . . 21
2.4 ♣ FT-2 Analytic Projective Geometry . . . . . . . . 22
♣ FT-2e . . . . . . . . . . . . . . . . . . . . 24
2.4.1 Homogeneous Coordinates on a Line . . . . . 25
2.5 The Fundamental Transformations of Projective Geometry . 26
2.6 More on Homogeneous Coordinates . . . . . . . . . . 29
2.6.1 Linearizability of Projective Transformations . . 29
2.7 ** The Linearized Projective Transformations . . . . . . 31
2.7.1 Projective Transformations in the Plane P2 . . . 32
2.7.2 Projective Transformations in P3 . . . . . . . 33
2.8 ** Coordinate System Rotations . . . . . . . . . . . 36
2.9 ** Analytic Proofs . . . . . . . . . . . . . . . . . 39
viii Contents

2.9.1 Computational Proof . . . . . . . . . . . . 39

2.9.2 Algebraic Proof . . . . . . . . . . . . . . 41
2.10 ** Conics on the Projective Plane . . . . . . . . . . . 43
2.10.1 In the Projective Plane All Conics Are Ellipses . 43
2.10.2 Generalizing Pappus’s Theorem . . . . . . . 44
2.11 ** Geometries and Invariants . . . . . . . . . . . . 45
3 ♣ FT-1 The Plane with Parallel Coordinates . . . . . . . . . 49

3.1 The Fundamental Duality . . . . . . . . . . . . . . 49

3.2 Transformations under the Duality . . . . . . . . . . 57
3.2.1 Rotations and Translations . . . . . . . . . 57
3.2.2 ** Recognizing Orthogonality . . . . . . . . 58
3.3 A Preview . . . . . . . . . . . . . . . . . . . . 61
4 Multidimensional Lines . . . . . . . . . . . . . . . . . . 63

4.1 ♣ FT-1 Representing Lines in RN . . . . . . . . . . 63

4.1.1 Elementary Approach . . . . . . . . . . . 63
♣ FT-1e . . . . . . . . . . . . . . . . . . . . 68
4.1.2 ♣ FT-2 Some Properties of the Indexed Points . 70
♣ FT-2e . . . . . . . . . . . . . . . . . . . . 73
4.1.3 Representation Mapping I . . . . . . . . . . 73
4.1.4 ** The General Case . . . . . . . . . . . . 75
4.1.5 ** Construction Algorithms . . . . . . . . . 79
4.1.6 Convenient Display of Multidimensional Lines . 83
4.1.7 ♣ FT-3 Rotations and Translations . . . . . 87
4.2 Distance and Proximity Properties . . . . . . . . . . 88
4.2.1 Intersecting Lines . . . . . . . . . . . . . 88
4.2.2 Nonintersections . . . . . . . . . . . . . 89
♣ FT-3e . . . . . . . . . . . . . . . . . . . . 91
4.2.3 Minimum Distance between Two Lines in RN . . 91
4.2.4 ** Air Traffic Control . . . . . . . . . . . 100
5 Planes, p-Flats, and Hyperplanes . . . . . . . . . . . . . . 115

5.1 ♣ FT-1 Planes in R3 . . . . . . . . . . . . . . . 115

5.1.1 Vertical Line Representation . . . . . . . . . 115
5.1.2 ** Planar Coordinates . . . . . . . . . . . 118
♣ FT-1e . . . . . . . . . . . . . . . . . . . . 124
 Contents ix

5.2 ♣ FT-2 Representation by Indexed Points . . . . . . . 124

5.2.1 The family of “Superplanes” E . . . . . . . . 125
5.2.2 The Triply Indexed Points . . . . . . . . . . 127
♣ FT-2e . . . . . . . . . . . . . . . . . . . . 135
5.3 ♣ FT-3 Construction Algorithms . . . . . . . . . . 136
5.3.1 Planes and Lines . . . . . . . . . . . . . 136
5.3.2 The Four Indexed Points . . . . . . . . . . 138
♣ FT-3e . . . . . . . . . . . . . . . . . . . . 141
5.3.3 Special Planes . . . . . . . . . . . . . . 141
♣ FT-4 . . . . . . . . . . . . . . . . . . . . 144
5.3.4 Intersecting a Plane with a Line . . . . . . . 145
♣ FT-4e . . . . . . . . . . . . . . . . . . . . 146
5.3.5 Points and Planes: Separation in R3 . . . . . . 146
5.3.6 ** Separation in R3 : An Old Approach . . . . 149
5.3.7 ♣ FT-5 Rotation of a Plane about a Line and the
Dual Translation . . . . . . . . . . . . . 151
5.4 Hyperplanes and p-Flats in RN . . . . . . . . . . . 159
5.4.1 The Higher-Dimensional Superplanes . . . . . 159
5.4.2 Indexed Points in RN . . . . . . . . . . . 164
♣ FT-6 . . . . . . . . . . . . . . . . . . . . 168
♣ FT-6e . . . . . . . . . . . . . . . . . . . . 168
5.4.3 Collinearity Property . . . . . . . . . . . . 169
♣ FT-7 . . . . . . . . . . . . . . . . . . . . 170
♣ FT-7e . . . . . . . . . . . . . . . . . . . . 170
5.5 ** Construction Algorithms in R4 . . . . . . . . . . 174
5.5.1 The Five Indexed Points . . . . . . . . . . 174
5.5.2 Intersecting Hyperplanes in R4 . . . . . . . . 175
5.6 Detecting Near Coplanarity . . . . . . . . . . . . . 179
5.7 Representation Mapping, Version II . . . . . . . . . . 180
6 ** Envelopes . . . . . . . . . . . . . . . . . . . . . . 185

6.1 The Basic Idea . . . . . . . . . . . . . . . . . . 185

6.2 Formulation . . . . . . . . . . . . . . . . . . . 186
6.3 Necessary and Sufficient Conditions . . . . . . . . . . 188
6.3.1 Singular Points . . . . . . . . . . . . . . 189
6.4 Examples: Envelopes of Families of Curves . . . . . . . 190
x Contents

7 Curves . . . . . . . . . . . . . . . . . . . . . . . . . 195

7.1 ♣ FT-1 Point-Curves and Line-Curves . . . . . . . . 195

7.1.1 Separation in the xy Plane . . . . . . . . . . 196
7.2 Duality between Cusps and Inflection Points . . . . . . 199
♣ FT-2 . . . . . . . . . . . . . . . . . . . . 200
7.3 ♣ FT-3 Point-Curves → Point-Curves . . . . . . . . 204
7.4 ♣ FT-4 Curve Plotting . . . . . . . . . . . . . . . 207
7.4.1 Space-Curves . . . . . . . . . . . . . . . 213
7.5 ♣ FT-5 Transforms of Conics . . . . . . . . . . . . 216
7.5.1 ** Proving the Duality Conics ↔ Conics . . . . 217
7.5.2 ♣ FT-6 Classification of the Conic Transforms . 220
♣ FT-6e . . . . . . . . . . . . . . . . . . . . 225
7.5.3 ** The Conic Transform’s Ideal Points . . . . . 226
7.6 ♣ FT-7 Transforms of Algebraic Curves . . . . . . . 229
7.7 ♣ FT-8 Convex Sets and Their Relatives . . . . . . . 231
7.7.1 Gconics and Their Transforms . . . . . . . . 232
♣ FT-8e . . . . . . . . . . . . . . . . . . . . 232
7.7.2 ** Operational Dualities with Gconics . . . . . 238
8 Proximity of Lines, Planes, and Flats . . . . . . . . . . . . . 243

8.1 ♣ FT-1 Motivation and a Topology for Proximity . . . . 243

8.2 Proximity of Lines and Line Neighborhoods . . . . . . 244
♣ FT-1e . . . . . . . . . . . . . . . . . . . . 250
8.3 ♣ FT-2 Proximity of Hyperplanes . . . . . . . . . . 251
8.3.1 Formulation of the Problem in RN . . . . . . 251
8.3.2 Outline of the Development . . . . . . . . . 253
♣ FT-2e . . . . . . . . . . . . . . . . . . . . 254
8.3.3 Properties of fN . . . . . . . . . . . . . 254
8.3.4 The Region  . . . . . . . . . . . . . . 257
♣ FT-3 . . . . . . . . . . . . . . . . . . . . 259
8.3.5 ** Construction of  . . . . . . . . . . . . 263
♣ FT-4 . . . . . . . . . . . . . . . . . . . . 264
♣ FT-4e . . . . . . . . . . . . . . . . . . . . 275
8.3.6 ** The Full Image NH . . . . . . . . . . . 275
♣ FT-5 . . . . . . . . . . . . . . . . . . . . 283
 Contents xi

8.3.7 Some Examples . . . . . . . . . . . . . . 283

♣ FT-6 . . . . . . . . . . . . . . . . . . . . 285
8.3.8 Matching Points and Navigation within NH . . . 287
9 Hypersurfaces in RN . . . . . . . . . . . . . . . . . . . 297

9.1 ♣ FT-1 Preliminaries . . . . . . . . . . . . . . . 297

9.2 Formulation . . . . . . . . . . . . . . . . . . . 301
9.3 ♣ FT-2 Boundary Contours . . . . . . . . . . . . 304
♣ FT-2e . . . . . . . . . . . . . . . . . . . . 306
9.4 ♣ FT-3 Developable Surfaces . . . . . . . . . . . . 310
♣ FT-3e . . . . . . . . . . . . . . . . . . . . 314
9.4.1 ♣ FT-4 Ambiguities and Uniqueness . . . . . 316
♣ FT-4e . . . . . . . . . . . . . . . . . . . . 317
9.4.2 Reconstruction of Developable Surfaces . . . . 320
9.4.3 Specific Developables . . . . . . . . . . . 325
♣ FT-5 . . . . . . . . . . . . . . . . . . . . 326
♣ FT-6 . . . . . . . . . . . . . . . . . . . . 333
♣ FT-7 . . . . . . . . . . . . . . . . . . . . 336
♣ FT-7e . . . . . . . . . . . . . . . . . . . . 336
9.4.4 ♣ FT-8 General Developable Surfaces . . . . 342
9.4.5 ** Higher Dimensions . . . . . . . . . . . 344
9.5 ♣ FT-9 Ruled Surfaces . . . . . . . . . . . . . . 346
♣ FT-9e . . . . . . . . . . . . . . . . . . . . 348
9.6 ** Approximate Quadric Hypersurfaces . . . . . . . . 359
9.6.1 The Hyperellipsoids . . . . . . . . . . . . 360
9.6.2 The Hyperellipse . . . . . . . . . . . . . 362
9.6.3 Intelligent Process Control . . . . . . . . . 363
9.7 ♣ FT-10 More General Surfaces . . . . . . . . . . . 365
♣ FT-10e . . . . . . . . . . . . . . . . . . . . 372
9.7.1 ** A More General Matching Algorithm . . . . 373
10 ♣ FT Data Mining and Other Applications . . . . . . . . . . 379

10.1 Introduction . . . . . . . . . . . . . . . . . . . 379

10.1.1 Origins . . . . . . . . . . . . . . . . . 380
10.1.2 The Case for Visualization . . . . . . . . . 380
xii Contents

10.2 Exploratory Data Analysis with -coords . . . . . . . . 382

10.2.1 Multidimensional Detective . . . . . . . . . 382
10.2.2 An Easy Case Study: Satellite Data . . . . . . 383
10.2.3 Quasars Dataset . . . . . . . . . . . . . . 390
10.2.4 Compound Queries: Financial Data . . . . . . 392
10.2.5 Hundreds of Variables . . . . . . . . . . . 400
10.2.6 Production of VLSI (Chips) . . . . . . . . . 401
10.3 Classification . . . . . . . . . . . . . . . . . . . 406
10.3.1 Case Study: Noise Signature Recognition . . . . 412
10.4 Visual and Computational Models . . . . . . . . . . 416
10.5 Parallel Coordinates: The Bare Essentials . . . . . . . 418
10.5.1 Lines . . . . . . . . . . . . . . . . . . 418
10.5.2 Planes and Hyperplanes . . . . . . . . . . . 419
10.5.3 Nonlinear Multivariate Relations: Hypersurfaces . 423
10.6 Future . . . . . . . . . . . . . . . . . . . . . 425
11 Recent Results . . . . . . . . . . . . . . . . . . . . . 429

11.1 Displaying Several Lines Efficiently . . . . . . . . . . 429

11.1.1 Abstract . . . . . . . . . . . . . . . . . 429
11.1.2 Introduction . . . . . . . . . . . . . . . 429
11.1.3 General Approach . . . . . . . . . . . . . 431
11.1.4 Flipping Axes . . . . . . . . . . . . . . 432
11.1.5 Rotating Axes . . . . . . . . . . . . . . 435
11.1.6 Transforming Axes . . . . . . . . . . . . 438
11.1.7 Back to Rotating Axes . . . . . . . . . . . 441
11.2 Separating Point Clusters on Different Planes . . . . . . 443
11.2.1 Overview . . . . . . . . . . . . . . . . 443
11.2.2 The Algorithm . . . . . . . . . . . . . . 443
11.2.3 Mean and Variance Estimation . . . . . . . . 445
11.2.4 Error Probability Approximation . . . . . . . 448
11.2.5 Complexity . . . . . . . . . . . . . . . 450
11.3 Surface Representation Decomposition and Developable
Quadrics . . . . . . . . . . . . . . . . . . . . . 451
11.3.1 The Hyperplane Representation as a Linear Transfor-
mation on the Coefficients . . . . . . . . . 452
11.3.2 A Brief Review of Duality in Projective Space . . 452
11.3.3 Extending the Projective Duality to Degenerate
Quadrics . . . . . . . . . . . . . . . . 453
 Contents xiii

11.3.4 The Decomposition and Its Application to Develo-

pable Quadrics . . . . . . . . . . . . . . 456
11.3.5 A Numerical Example: An Elliptical Cone in 5-Space 457
11.4 Network Visualization and Analysis with Parallel Coordinates 460
11.4.1 Introduction . . . . . . . . . . . . . . . 460
11.4.2 The NEVIS Transformation . . . . . . . . . 461
11.4.3 Properties of NEVIS . . . . . . . . . . . . 464
11.4.4 Large-Scale Networks . . . . . . . . . . . 469
11.5 To See C2 . . . . . . . . . . . . . . . . . . . . 476
11.5.1 The Complex Field C . . . . . . . . . . . 476
11.5.2 The Space C2 . . . . . . . . . . . . . . . 477
11.5.3 Representing Points on a Line in C2 . . . . . . 479
11.5.4 Holomorphic Functions . . . . . . . . . . . 483
12 Solutions to Selected Exercises . . . . . . . . . . . . . . . 489

12.1 Solutions to Exercises in Chapter 2 . . . . . . . . . . 489

12.2 Solutions to Exercises in Chapter 3 . . . . . . . . . . 509
12.3 Solutions to Exercises in Chapter 4 . . . . . . . . . . 514
12.4 Solutions to Exercises in Chapter 5 . . . . . . . . . . 520
Notation and List of Symbols . . . . . . . . . . . . . . . . . 531
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . 535
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . 547
This page intentionally left blank
0

Preface

What and Why

In the late 1980s, researchers in computer graphics recognized a specific stream of

applications as an emerging field and called it visualization. By now it has become
a discipline with its own journals, conferences, and community of active workers
applying it to science, city planning, entertainment, and much more. Even our reality
is becoming more virtual. Yet since time immemorial people have expressed them-
selves visually, realizing in due course insight can also be derived from images [63].
In 1854, a cholera epidemic raging in a London neighborhood prompted Dr. John
Snow to search for remedies. He examined the data consisting of a table with the
addresses of those who had died from the disease. On a map of the neighborhood,
which fortunately had the positions of the water wells, he placed dots at the locations
of the recorded deaths. The concentration of dots in the vicinity of just one well
was visually striking. He became convinced that there was a connection, and had
the handle of the suspect well replaced. The epidemic stopped! Apparently the
disease was being transmitted by contact with the handle. This true story is widely
considered an early success of visualization [172]; the picture provided insight no
one had gleaned from the table of the data.
Legend has it that Archimedes was absorbed in a diagram when a Roman
soldier killed him; the first recorded death in defense of visualization! “Do not
disturb my circles” he pleaded as he was being struck by the sword. Visualization
flourished in geometry, where visual interaction with diagrams is interwoven in the
testing of conjectures and construction of proofs. Our tremendous ability in pattern
recognition enables us to extract insight from images. This essence of visualization
is abstracted and adapted in the more general problem-solving process to the extent
that we form a mental image of a problem we are trying to solve and at times we
say see when we mean understand.
My interest in visualization was sparked and nourished while learning Euclidean
geometry. Later, while studying multidimensional geometries I became frustrated
by the absence of visualization. Basically, we were working with equations, inter-
preting them geometrically, without the enjoyment and benefit of pictures. I kept
wondering about ways to draw “pictures” of multidimensional objects. Of course,
it would also be great fun doing synthetic high-dimensional constructions and have
xvi Preface

tools for accurate visualization of multivariate problems. These thoughts, best

attributed to the impetuousity of youth, absorbed me. What coalesced from this
struggle is parallel coordinates.
A superficial resemblance to nomography [19] was noted early in the develop-
ment [97]. “Nomograms are drawings made on the plane in order to replace
cumbersome numerical calculations” ([144] p. 9), a technique that declined with
the advent of modern computation. I was unaware until recently2 of d’Ocagne’s
marvelous monograph on parallel coordinates in two dimensions [141] with a point
↔ line correspondence. D’Ocagne was interested in the computational applications
of nomography [142] rather than the development of a multidimensional coordinate
system, which is where we come in. Invented independently, parallel coordinates
is a general-purpose VISUAL multidimensional coordinate system. There are theo-
rems on the unique representation of multidimensional objects, and geometrical
algorithms for intersections, containment, minimal distances, proximities, etc.,
with a wide variety of applications.
The writing is undertaken with trepidation, knowing what kind of book it should
not be: not a text in mathematics, though it should contain the foundations and
development, nor computer science nor the other components of the methodology,
nor a source-book for the variety of included applications. Rather it is an amalgam
of all these, comprehensible and appealing to mathematicians, computer scien-
tists, statisticians, engineers, scientists, and all those wanting to acquire deeper
understanding of their multivariate problems by way of visualization.
In due course, it is hoped that the methodology will not only find its way into
textbooks on various applications, but also be adopted for visual experimentation
and explanations in teaching mathematics and the sciences at the university, high-
school, and even elementary-school levels.

Style and Structure of the Book

This is a textbook with ample material for a one-semester course, easily adapted to
shorter periods (see the following section), and containing advanced seminar topics.
The writing evolved during a decade of teaching “visualization of multidimensional
geometry and its applications,” a popular elective course in the applied mathematics/
computer science curricula. It is also a reference for self-study or a companion text
for courses on information visualization (particularly the chapter on data mining),
visualization in general, data mining, GIS, management, statistics, linear algebra,
analytic geometry, complex variables, and fields dealing with multivariate problems
2
I am indebted to M. Friendly for pointing this out in [63].
 Preface xvii

(economics, engineering, physics, psychology, management, arts, and social sci-

ences). The parallel coordinates methodology provides a multidisciplinary visual
language and paradigm.
The mantra is to let intuition guide the formalism starting with the motivation for
the emergence of parallel coordinates (-coords). Then the 2-dimensional plane is
viewed with -coords as compared to Cartesian coordinates, leading to the discovery
of surprising and pleasing properties. The chapters on the representation of lines,
hyperplanes, curves, and topologies for proximity of lines and hyperplanes are
interlaced with applications on collision-avoidance algorithms for air traffic con-
trol, geometric modeling, computer vision, statistics, and more. On this foundation
the representation of surfaces is built, yielding exciting results such as viewing con-
vexity in any dimension , recognizing nonorientability (as for the Möbius strip),
and applications to intelligent instrumentation, process control, and decision sup-
port. There follows an easy-to-read ch apter on dat a mining a nd informatio n
visualization with many fun applications (exploratory data analysis (EDA) on
satellite, financial, vehicle recognition from noise-signature, astronomical and
industrial datasets, automatic classification, decision support, nonlinear models).
Results too recent to be incorporated in the text are included in the appendix at
the end. They complement the representation of lines, developable surfaces,
visualization and analysis of large networks, detecting and separating clusters of
coplanar points, and the visualization of complex-valued functions: to see C2.
The emphasis is on visualization, though computational and analytical aspects
are treated rigorously, and where needed, extensively. Each chapter contains a
variety of exercises, easy ones at first to check the understanding of the basics
and more with increasing difficulty. Some exercises have served well as semester
project topics. Where appropriate, open questions and research directions are
pointed out. In places, details are deferred to references when technicalities
may obscure the underlying ideas or unnecessarily intimidate the reader. I hope
that the more mathematically sophisticated readers will forgive the occasional
‘‘sleights of hand.’’
To make the book reasonably self-contained, after the introduction, there is
a chapter on the needed geometrical background. It covers projective geometry,
explaining the notions of duality and homogeneous coordinates as used in the
subsequent rigorous development. The application to perspective, as in computer
graphics, is developed with the realization that the “points at infinity” on the
perspective plane are images of ideal points (directions). There is also a short
intuitive chapter on the theory of envelopes; helpful background for the ensuing
chapters on curves and surfaces. The attached CD contains the interactive learning
module (ILM) for self-study and class demonstrations. There are pointers and
xviii Preface

guidance for its use at the relevant sections in the text. Projective geometry theorems,
dualities, transformations, and properties of representations are understood quickly
and enjoyably.

♣ FT Fast Track: How to Skip Reading Most of the Book

In each chapter the key portions are marked successively by ♣ FT-5 . The numeral,
here 5, indicates the fifth fast-track marker within the chapter. The essentials can be
swiftly grasped by following these road signs and returning for additional details
later if needed. Typically the ♣ FT-5 symbol is placed at the beginning of a section
to be read to the end or up to an “end marker” such as ♣ FT-5e . The landmarks3
may also be figures. Covering these highlights is also a good way to review the
material. The fast-track entries appear in the table of contents for sections or sub-
sections. Portions marked with ** contain in-depth treatment of specific topics
and are not needed for the basic ideas and results. Those interested in the applica-
tions can read the chapter on data mining and then scan the text for more details.
Otherwise, choose your goal and follow the illuminated path:

• the essentials can be mastered quickly with the FT option;

• for extensive and sound foundations, include unmarked portions;
• to reach expert level, add some ** sections.

Course Organization
Much has been learned from the teaching experience. Reasonably well prepared
and motivated students enjoyed and absorbed the material readily, contributing
new results and ideas. After some experimentation, the course outline that proved
most successful is the presentation of a course overview in the first two to three
hours, including some demonstrations. This motivates the subject with problems
and examples and also shows why, due to the 2-D point ↔ line duality, parallel
coordinates need to be considered in the projective rather than the Euclidean plane.
The development as given in the chapter sequence can then be safely followed.
Demonstrations with the ILM in the CD as well as several applications are always
well received. The instructor can organize group projects that provide such demon-
strations. Some of the ones already given are available over the Internet. Also,
there is an increasing collection of software commercially or freely available that
is suitable for class demonstrations and use.
3
I am indebted to Chao-Kuei Hung for this wonderful idea.
 Preface xix

Instead of exams, the students, usually in groups of two or three, do projects.

They are encouraged to choose topics in some area in which they have a per-
sonal interest and apply what they have learned in the course. Such projects have
involved expanding the mathematical foundations, development and implementa-
tion of algorithms, writing programs particularly for data mining and modeling, and
more. Examples of actual projects include “visualization of the Internet,” “collision
detection and intersection in billiards,” “visualizing the ‘learning’ process in neural
networks,” “topologies for proximity of flats (higher-dimensional planes),” several
geometrical construction algorithms, “characterizing painters (i.e., impressionists,
etc.),” “characterizing music,” data mining on a great variety of multivariate datasets
(i.e., pollutants, bioinformatics, weather data, computer dating, constituents of
various medicines and foods, etc.), differentiation and integration, fractals and linear
programming with -coords, large network visualization, visualizing complex-
valued functions, and other topics. A wide variety of software on representations of
surfaces, curve-plotters, implementation of results (i.e., translations ↔ rotations),
etc. were produced. Some projects were subsequently presented at conferences,
and others led to master’s theses. Student are encouraged to pursue self-discovery,
including writing programs for experimentation and testing of conjectures. Solu-
tions to many exercises are included at the end. The recommended course syllabus
follows.

** Syllabus

• Introduction to scientific information and multidimensional visualization, com-

plete course overview. One lecture with demos on the foundations and numerous
applications.
• Projective geometry: foundations, dualities, collinearity theorems, homoge-
neous coordinates, linearization of projective transformations, conics in P2 ,
applications to computer graphics: transformations and their matrices; perspec-
tive: points at infinity are images of ideal points, geometrical invariants. Two
lectures with demos.
• Parallel coordinates in the plane: dualities: point ↔ line, translation ↔ rotation,
other transformations, parallelism and orthogonality, visual and automatic data
mining. One lecture with demos.
• Lines in N-space; two lectures

– Representing lines by (N −1) points with two indices, parallel lines, general
representation, the 3-point collinearity property; demos.
xx Preface

– Transformations, improving the display of multidimensional lines, inter-

sections, minimum L1 distance between pairs of lines, upper and lower
proximity bounds.
– Application: collision-avoidance algorithms for air traffic control; demos.
• Planes, hyperplanes, and flats in N-Space; three lectures.

– Detecting coplanarity of points on a grid, representing hyperplanes by

(N − 1) parallel lines and a point (polygonal line), duality: rotation of a
plane about a line ↔ translation of a point on a line, manufacturing data;
demos.
– Recognizing M-dimensional objects from their (M − 1)-dimensional
subsets, recursive construction mapping.
– Detecting coplanarity of randomly chosen points, representing hyperplanes
by (N − 1) points with N indices, the family of superplanes and their
special properties, reading the equation of a hyperplane from the picture,
duality: rotation of a plane about a line ↔ translation of (N + 1) points on
lines, designing rotations and transformations from their “planar signature,”
synthetic constructions: parallel planes, intersections, above/below relation
between points and planes and more, “near coplanarity,” separating point-
clusters on several planes; demos.
• Envelopes of one-parameter families of curves, background for the curves and
surfaces chapters. One-half lecture (discretionary).
• Curves; one and one-half lectures.

– Point-curves and line-curves, convex down/up curves and their transforms,

duality: cusp ↔ inflectio point, point-curve ↔ line-curve.
– conics ↔ conics via the Möbius transformation; six cases, algrebraic curves:
the Plücker formulas and dualities, curve plotters, examples of curve trans-
forms, analysis of important points and portions of curves, periodicities, and
symmetries.
– Convex curves and their relatives, generalized conics and their dualities; six
cases, operational transformations.

• Families of proximate hyperplanes; one and one-half lectures.

– Topology for proximity of flats.

– Proximity of lines and line neighborhoods.
– Proximity of hyperplanes, formulation of the problem in N-space, hyper-
plane neighborhoods with (N − 1) regions, the three cases: bounded
 Preface xxi

convex polygons, unbounded convex polygons, generalized hyperbolas,

constructing neighborhoods with complexity O(N ), the general case.
– Navigation within neighborhoods, finding the (N − 1) points representing
specific hyperplanes, matching algorithm; demos.
– Applications, a central problem in many fields, statistics (regression),
geometric modeling, computer vision.

• Hypersurfaces in N-space; two lectures.

– Formulation of the problem, hypersurfaces represented by (N − 1) linked

regions, boundary contours, quadric and general algebraic hypersurfaces.
– Developable hypersurfaces represented by (N − 1) linked curves, ambigui-
ties and uniqueness, reconstruction, classes of developables, duality:
space curves ↔ developables, examples: developable helicoids and more;
demos.
– Ruled hypersurfaces, several examples of their representation including
Möbius strip, visualizing nonorientability in N dimensions; demos.
– More general hypersurfaces, convex and nonconvex, and how to distinguish
them in N dimensions, searching for matching algorithms, interior points.
– An important approximate representation, interior points, applications to
decision support and intelligent process control; demos.

• Data mining and other applications (self-contained chapter); two lectures.

– Origins, the case for visualization.

– exploratory data analysis (EDA); visual data mining

* Multidimensional detective, atomic queries, an easy case study: GIS

data, Boolean operators and compound queries; financial data, hundreds
of variables; demos.
– Automatic classification, feature selection and ordering, comparisons, and
two case studies; demos.
– Visual and computational models, decision support: interelationships,
sensitivies, constraint and trade-off analysis; demos.
• Recent developments (appendix); two lectures (discretionary).

– Representing several lines efficiently; Shlomi Cohen-Ganor.

– Separating points on several planes; Nir Shahaf.
– Surface representation decomposition and developable surfaces; Chao-Kuei
Hung.
xxii Preface

– Network visualization and analysis with -coords; Yaron Singer and Ohad
Greenspan.
– To see C2 : visualizing complex-valued functions; Yoav Yaari.
• Time permitting, students can choose to lecture on various topics in visualiza-
tion, e.g., Internet visualization, fluid flow, visualizing the function of neural
networks, detecting network intrusions, etc.

Prerequisites
The course attracted students from mathematics, computer science, engineering,
statistics, management, data mining, and a wide variety of sciences (physics,
chemistry, biology, etc.), geography, social sciences, and even linguistics. They
ranged from second year to advanced doctorate students. Background in linear
algebra is important for those who want to master the subject. Programming skills
including some computer graphics are helpful though not necessary.
0

Acknowledgments

I first proposed parallel coordinates in 1959 when I was a graduate student at the
University of Illinois (Champaign-Urbana). Two topologists, Professors D. Bourgin
and S.S. Cairns, encouraged me to pursue the idea, but it took me many years to
take their advice seriously. Later, I had the pleasure and priviledge to collaborate
with Bernie Dimsdale4 at the IBM Los Angeles Science Center (LASC), who made
many important contributions and left his imprint throughout this work. Also at
LASC, Alex Hurwitz’s numerous contributions, suggestions, and ideas as well as
Juan Rivero’s application to the phase space for nonlinear differential equations
benefited the early development.
From the students, starting in 1978 Moti Reif was bravely the first to do
his graduate research on parallel coordinates (convexity algorithms) [110]; John
Eickemeyer discovered among other things the “superplanes” [47]; Avijit Chatterjee
developed Eickemeyer’s results further for the visualization of polytopes in N-space
[23]; Misha Boz made many contributions over the years, and together with
Bernie Dimsdale, we did the first successful implementation of the collision avoid-
ance algorithms [104]; Steve Cohan [31] and Paolo Fiorini developed applications
to robotics [60]; Jim Adams (circa 1981), Alex Hurwitz, and later Tuval Chomut
pioneered the first successful implementation of parallel coordinates for exploratory
data analysis (EDA) [30] as well as contributing to the convexity algorithms with
Moti Reif; Chao Kuei Hung derived an elegant characterization of developable
surfaces in terms of their tangent planes [93], [108] (see [94] for updated results);
Lou Ann Colby investigated dualities between rigid transformations; Richard
Addison later improved the implementation of the collision avoidance algorithms;
and Greg Wong together with John Eickemeyer under Alex Hurwitz’s leadership
developed and implemented further substantial improvements; Tova and Shlomo
Avidan made innovative improvements on the data mining application [5], [103];
Tanya Matskewich and Yaakov Brenner did a very sophisticated and useful char-
acterization of proximity for flats [131] that developed into a full chapter; Nadav
Helfman [81] further developed the application to decision support; Adi Shlomi
designed and wrote much of the software in the ILM. Liat Cohen solved the exercises
and also contributed many ideas and improvements. David Adjiashvili wrote the

4
A long-term associate of John von Neuman.
xxiv Acknowledgments

beautiful software displaying surfaces in -coords. In the spirit of geometry, these

pictures led to key results that were subsequently proved. Several other student
projects made important innovations and are acknowledged in the text.
Many researchers contributed to the development and applications of parallel
coordinates (in alphabetical order). The Andrienkos [2], J. Dykes et al. [44],
R. Edsall [45], and A. MacEachren et al. [43] introduced -coords to GIS and
geovisualization, Kim Esbensen [53] contributed to data analysis and dualities,
H. Carter, C. Gennings, and K. Dawson on response surfaces in statistics [67];
Amit Goel [71] on aircraft design; Chris Jones in optimization [115]; John Helly
[82] on early application to data analysis; Hans Hinterberger’s [156] contributions
in comparative visualization and data density analysis; Matt Ward et al. introduced
hierarchical parallel coordinates [65] and added much to the field [181], Helwig
Hauser’s innovative contributions included parallel sets for categorical data [79];
Antony Unwin et al. made wide-ranging numerous contributions [177]; Li Yang
[187] applied -coords to the visualization of association rules; H. Choi and Heejo
Lee’s [29] and P. Hertzog [83] are important contributions to detecting network intru-
sion, other important contributors to the field include Luc Girardin and Dominigue
Brodbeck [70], Xiaxin Yen, et al. [189], T. Itoh, et al. [111], and Stefan Axelsson [6].
G. Conti [32] produced a ground-breaking book on security data visualization; there
are exciting recent works by H. Ye and Z. Lin’s [188] with an innovative contribution
to optimization (simulated annealing); T. Kipouros et al. [120] with a sophisticated
optimization for turbomachinery design [120]; S. El Medjani et al. [49] proposed
a novel application using -coords as straight-line detectors; R. Rosenbaum and
H. Schumann [151] on progression in visualization; F. Rossi [153] on visual data
mining and machine learning; Huamin Qu et al. [89] on air pollution analysis
and clustering; M. Tory et al. on -coords interfaces [171]; J. Johanson et al. [113];
G. Ellis and A. Dix [50] on clustering and clutter reduction; H. Siirtola and K.J. Räihä
on interaction with -coords; B. Pham, Y. Cai and R. Brown on traditional Chinese
medicine [147], and there are proposals to enhance -coords using curves [129],
place them in 3-D [139], [114], or modify them as starplots [51]. C.B. Hurley and
R.W. Olford contributed to the definitive study on axes permutations [95]. This list
is by no means exhaustive.
There have been a number of noteworthy visualization contributions since the
1980s: the grand tour by Asimov [4], dimensional inference and more by Buja et al.
[66], multidimensional scaling and convergence of correlation matrices by Chu-Hou
Chen [27], [26], a chain of valuable and varied ideas (most notably RadViz) by
Georges Grinstein et al. [87] [88] and Daniel Keim et al. [119] [118], hierarchical
visualization of multivariate functions by Mihalisin et al. [135], [84]. Eminent are
the seminal contributions of Jacob Zaavi and Gregory Piatetsky-Shapiro et al. [57]
in all aspects of data mining and the role of visualization.
 Acknowledgments xxv

Bertin’s book [14] is the early classic in information visualization. The field
has been blossoming since then as attested by the volumes [22] and [11] containing
organized collections of important papers, the seminal source on data mining [57],
and the wonderful texts of Colin Ware emphasizing perception and design [182],
Chaomei Chen’s [25] work on a range of cognitive social and collaborative activities,
and the comprehensive expository collection of methodologies by Robert Spence
[164].
I have been supported and introduced to a variety of good problems by several
friends and colleagues, especially (listed in alphabetical order) by George Bekey,
Ronald Coiffman, Julia Chung, Alon Efrat, Marti Hearst, Peter Jones, Pei Ling Lai
[109], [124], Khosrow Motamedi, Karl Müller, David McIlroy, Niko Shlamberger
and Ana Trejak, Randall Smith of GM Research, and Fred Warner. I am especially
grateful to John Nacerino of EPA in Las Vegas for his guidance and steadfast
encouragement. His initiative and efforts enabled the distribution of the parallel
coordinates software to academic researchers. Since 1990 the IEEE annual visua-
lization conferences have been an excellent forum in which to learn and present
ideas as well as receive valuable suggestions and criticism. In this community the
following colleagues were especially helpful: Dan Bergeron [185], Ed Chi [28],
Georges Grinstein, Beth Hetzler, Helwig Hauser, Daniel Keim, Chaim Lefkovits,
Ted Mihalisin, Tamara Munzer, Alex Pang, Catherine Plaisant [160], Bill Ribarsky,
Bernice Rogowitz, Holly Rushmeier, Lloyd Treinish, Bob Spence and Lisa Tweedie
[175], John Stasko, Theresa-Marie Thyne, Matt Ward, and Pak C. Wong, who
together with R.D. Bergeron wrote the first survey of the field [185] among their
other contributions. I learned a great deal from Ben Shneiderman’s [159] [160],
[11] fundamental insights and encyclopedic contributions to visualization, and also
from Marti Hearst’s [112], [80] courses, varied interests and stimulating ideas in
information visualization, user interfaces, and text mining. Of great benefit has been
the interaction with the statistics and statistical graphics communities, especially
with Daniel Asimov, Andreas Buja, Jogesh Babu, Dianne Cook [34], Larry Cox,
Daniela Di Benedetto [33], Michael Friendly [62], Heike Hoffman, Catherine
Hurley, Moon Yul Huh [91], Sigbert Klinke, Carlo Lauro [125], Jungi Nakano
[139], Francisco Palumbo et al. [125], David Scott, Valentino Tontodonato [170],
Ritei Shibata et al. [123], Deborah F. Swayne [168], Martin Theus, Luke Tierney,
Simon Urbanek, Antony Unwin [176], Bill Venables, Ed Wegman [184], Adalbert
Wilhelm, Lee Wilkinson and Graham Wills.
Special thanks for their help and encouraement are due to several of my col-
leagues in Israel: Yoav Benjamini, Iddo Carmon, Daniel Cohen-Or, Ronen Feldman,
Camil Fuchs, Dony Gat, Hillel Gauchman, David Gilat, David Levin, Mark Last,
Leonid Polterovich, Eugenii Shustin and Jacob Zahavi.
xxvi Acknowledgments

Over the years, many managers at IBM, among them Lew Leeburg, Baxter
Armstrong, Homer Givin, Jim Jordan, John Kepler, Kevin McAuliffe, Abe Peled,
Bernie Rudin, Arvid Shmaltz, and John Wolff, supported, encouraged and in many
ways helped with this effort.
With patience Ann Kostant, Laura Held, Alicia de los Reyes, and their staff
from Springer most capably encouraged, cajoled, corrected and produced this book.
Special thanks are due Chao-Kuei Hung, who read the whole manuscript and pro-
vided invaluable ideas and corrections. To all these friends and others whom I may
have inadvertently omitted, I offer my profound gratitude. I am especially grateful
to my family for their unswerving multidimensional encouragement, inspiration,
patience, and precious sense of humor; without them I would have never finished.
1

Introduction

1.1 Multidimensional Visualization

About half of our sensory neurons are dedicated to vision,5 endowing us with a
remarkable pattern-recognition ability. Visualization, an emerging field with huge
potential, aims to incorporate this tremendous capability into our problem-solving
process. Rather than dictionary definitions we consider visualization as a collection
of application-specific mappings:

P roblem Domain −→ V isual Range.

For various classes of problems, visual models are constructed that provide insight
through images [132] about the problem. This is an extension of Hamming’s 1962
adage that “the purpose of computing is insight, not numbers.” While the develop-
ment of software and hardware tools has been impressive, concepts broadening
the applicability of visualization have lagged behind. Scientifi visualization for
applications having a physical or geometric basis (e.g., electromagnetic fields,
geography, fluid flow) has blossomed. Yet progress has been slower in information
visualization for fields not endowed with such structure, as in finance, process con-
trol, and communications, where many variables are involved, further complicating
their visual expression.
5
If one takes as an informal but meaningful definition of visual as “affected by light stimulation to
the eyes,” and if the quantity of neurons in the cortex is used as a baseline, then the 50% figure is
the one often quoted by experts. I am grateful to Prof. Enio Mingolla, of Boston University, and
Prof. Simo Vanni, of Helsinski University of Technology, who together with other experts confirmed
this information.

A. Inselberg, Parallel Coordinates, DOI 10.1007/978-0-387-68628-8_1,

© Springer Science+Business Media, LLC 2009
2 1. Introduction

Understanding the underlying geometry of multidimensional problems can pro-

vide insights into what is possible and what is not. In an astonishing but not well
known paper [46] the physicist Ehrenfest showed in 1917 that planetary orbits are
stable only in space of dimension 3. Higher-dimensional planetary systems, if they
ever existed, would have a short career due to the orbits’ instability, which offers an
interesting hypothesis for the dimensionality of our habitat. Another dimensionality
curiosity is that rotating rigid bodies have an axis of rotation only in odd-integer
dimensions [169, p. 189], Goldstein [72, pp. 158–166]. Had our dimensionality
been 4, the Earth would not have a fixed axis, resulting in the absence of seasons,6
among other things. Clearly dimensionality can have important implications. For a
collection of anecdotes and opinions of great scienstists on the role of visualization
in science see Ron Kriz’s website [122]. For “dimensional adventures,” the classic
Flatland [1] by Abbot, Flatterland by Stewart [165] and Banchoff’s journey beyond
the third dimension [8] are warmly recommended.
The fascination and exploration of dimensionality may predate Aristotle and
Ptolemy, who argued that space has only three dimensions. It was not until the
nineteenth century that Riemann, Lobachevsky, Bolyai, and Gauss unshackled our
imagination, limited by the experience and perception7 of our three-dimensional
habitation, with their non-Euclidean and higher-dimensional geometries. Would
a way be found to visualize these wonders? People have certainly tried, yet the
available multidimensional visualizations can display only finite multivariate point
sets (see the beautiful collection in [63] [172], [173], [174] and bibliographies in
[185] and [106]).

1.2 ♣ FT-1 How?

For the visualization of multivariate relations a conceptual breakthrough emerged

that is, in the spirit of Descartes, a multidimensional coordinate system. In geometry,
parallelism, which does not require a notion of angle, rather than orthogonality,
is the more fundamental concept. Coupled with the fact that orthogonality “uses
up” the plane very fast, this inspired me to experiment and then present8 Parallel
Coordinates in 1959. The systematic development began in 1977 [97] (short reviews

6
Resulting from the axis’ tilt.
7
As an interesting aside, there is strong evidence suggesting that even 3-D perception depends greatly
on our 2-D cognition (see [20], for example).
8
In his combinatorial topology course, Prof. S.S. Cairns lamented the absence of multidimentional
visualization tools, prompting me to share the idea with the class.
 1.2 ♣ FT-1 How? 3

Figure 1.1. The polygonal line C̄ represents the point C = (c1 , c2 , c3 , c4 , c5 ).

in [105], [100]). The goal was the systematic visualization of multidimensional

problems without loss of information having the properties listed below.

1. It should have low representational complexity, i.e., O(N ) for N variables,

be linear in the number of dimensions (variables); compare with the common
“scatterplot matrix,” whose pairwise projections have O(N 2 ) complexity.
2. It should apply to any dimension, treating every variable uniformly, unlike
Chernoff faces or glyphs and other multivariate visualizations [172].
3. The displayed objects should be recognizable under projective transformations
(i.e., rotation, translation, scaling, perspective).
4. Properties of the N-dimensional objects should be apparent from the display.
5. It should be mathematically rigorous.

These points are amplified in Chapter 10 on data mining.

In the Euclidean plane R2 with xy Cartesian coordinates, N copies of the real
line R labeled X̄1 , X̄2 , . . . , X̄N are placed equidistant and perpendicular to the
x-axis. They are the axes of the parallel coordinates system for the Euclidean
N -dimensional space RN , all having the same positive orientation as the y-axis.
A point C ∈ RN with coordinates (c1 , c2 , . . . , cN ) is represented by the complete
polygonal line C̄ (i.e., the complete lines containing the segments between the axes)
whose N vertices are at (i − 1, ci ) on the X̄i -axis for i = 1, . . . , N, as shown in
4 1. Introduction

Fig. 1.1. In this way, a one-to-one correspondence between points in RN and planar
polygonal lines with vertices on the parallel axes is established.
In principle, a large number of axes can be placed and be seen parallel to each
other. The representation of points is deceptively simple, and much development
with additional ideas is needed to enable the visualization of multivariate relations
or equivalently multidimensional objects. Specifically, the representation of an
M-dimensional object, 2 ≤ M, will be obtained from its (M − 1)-dimensional
subsets. For example, the representation (i.e., image) of a line is obtained from
the polygonal lines representing points on the line, and this turns out to be their
envelope.9 Next, the representation of a plane in R3 is obtained from the represen-
tation of the lines, rather than the points, it contains. This leads to a recursion that
turns out to work spendidly, but we are getting ahead of ourselves. The recursive
construction algorithm is properly considered in Chapter 5 on hyperplanes.
Parallel coordinates (abbreviated -coords) transforms the search for multivari-
ate relations in a data set into a pattern-recognition problem. This is the core reason
for the widespread use in exploratory data analysis (EDA) (also called “visual data
mining”) and the inclusion in software packages starting with EDA (Chomut [30]),
[59], (Hinterberger [156]), Spotfire (Christopher Ahlberg and Ben Shneiderman
[3]), ExplorN (Bolorforoush, Carr et al.), WinViZ (Eickemeyer), VisDB ([119]),
Xmdv ([181]), XGobi ([167]), Strata ([155]), PVE (Adams, Hurwitz, Inselberg,
Chatterjee [9]), and many others. This type of application [101] hinges on:

• an informative display of the data,

• good choice of queries, and
• skillful interaction between the user and display in searching for patterns
corresponding to relationships among the variables.
Part of the discovery process (classification) has been automated [103], and the
Parallax software contains both visual and automatic data mining modules based on
-coords. The self-contained Chapter 10 is devoted to the data mining applications,
the design of queries, guidelines, strategies, and “tricks” for knowledge discovery
and classification. They are illustrated with several case studies on real multivariate
(one with hundreds of parameters) data sets garnering surprising and at times “spicy”
information. There are other applications to statistics [67], air traffic control (col-
lision avoidance algorithms, presented in Chapter 4 on lines), geometric modeling
and computer vision (Chapter 8 is on the visualization of families of “close” lines
and hyperplanes), approximations, decision support (including process control)
[81], optimization ([24], [115]), recent different innovative applications like [178]

9
Chapter 6 contains a short intuitive introduction to the theory of envelopes.
 1.2 ♣ FT-1 How? 5

and others acknowledged earlier. For data mining, the grand vision is the creation
of algorithms whose input is a massive data set and output 2-dimensional patterns;
the “graph” of multivariate relations discovered in the data and, as described in
Chapter 10, this goal is now within our grasp.
The next chapter provides background in geometry. Readers may skim through
it, referring to specific topics as the need arises. Prior to that, some exploration on
the Internet is worth doing with queries such as:
• Visualization
• Multivariate + Visualization
• Multidimensional + Visualization
• “Scientific Visualization”
• “Information Visualization” or Infovis
• “Parallel Coordinates”
• “Parallel Coordinates + Applications”
• Visualization + Applications
• “Parallel Coordinates” + a specific application.
This page intentionally left blank
2

Geometry Background

“Let no one ignorant of Geometry enter” – At entrance to Plato’s Academy

2.1 Preliminaries
2.1.1 In the Beginning ...
The Nile’s flooding in Egypt used to destroy landmarks and position markers, raising
the need for reliable measurements and originating the development of geometry
Geo-ω (for land) metry (for measure-μτρια).10 The Rhind Papyrus, which is
in the British Museum, was written by an Egyptian priest named Ahmes about
1,800 B.C.E. It treats the measurement of planar regions and solids. It contains the
 2
remarkable approximation of π ≈ 16 9 = 3.1604. The contemporary Babylo-
nians, as well as Chinese, used 3 for the value of π. The value of 3 for π is also
mentioned in the Hebrew scriptures (see I Kings 7:23). A rabbi called Nehemia (p. 5
[74]) proposed the value of 22 7 for π, but apparently it was not accepted. By the
way, only in 1789 was it first proved by J. Lambert that π is an irrational number
and in 1882, F. Lindemann proved that π is transcendental (i.e., it is not the root of
any algebraic equation with rational coefficients). For more enjoyable folklore on
π the reader is referred to [12].
It is fair to consider the Egyptians more as surveyors, who used a collection of
geometrical formulas, mostly approximations, rather than geometers. For proper
attribution, deductive geometry was accomplished by the Greeks, probably begin-
ning with Thales of Miletus (≈ 640–545 B.C.E.), who was familiar with Egyptian

10
This story was reported by Proclus in the fifth century C.E., who wrote the early history of geometry.

A. Inselberg, Parallel Coordinates, DOI 10.1007/978-0-387-68628-8_2,

© Springer Science+Business Media, LLC 2009
8 2. Geometry Background

and Babylonian mathematics. Pythagoras, Hippocrates (not the physician of the

same name), Menaechmus, Archimedes, Apollonius, Plato, Archytas, Eudoxus and
Theaetetus continued the development until it was codified by Euclid in his Elements
τ oιχια (≈ 300 B.C.E.). Entirely novel, and characteristic of Greek mathematics,
is the axiomatic method whereby theorems are established by proof. Particularly
relevant to computer science are the constructions of Euclidean geometry. They
consist of a finite set of well-defined operations and provide an unambiguous pro-
cedure that terminates for all inputs! Truly the first algorithms ... and with proofs.
The constructions of Euclidean geometry involved ruler and compass. A remarkable
discovery was made in 1672 by the Danish geometer G. Mohr (1640–1697) and inde-
pendently in 1797 by the Italian L. Mascheroni (1750–1800).11 They showed that
all of these constructions can be accomplished by compass alone! [38]. The proofs
were, in the true spirit of computer science, by simulation. It was demonstrated
that any operation involving a ruler can be replaced by a finite number of compass
operations; a real equivalence proof like those in automata theory. Alternatively,
J. Steiner and J.V. Poncelet showed that all constructions in Euclidean geometry
can be accomplished with a straightedge when in addition, a single circle and its
center are also given. Continuing in this vein, it is natural to ask what is left if the
compass is discarded and only the ruler is left. What remains is projective geometry,
an elegant branch of geometry simpler than Euclid’s but rich and interesting in its
own right. It deals with points, lines, planes, and spaces where distance between
two points or angles between two lines are not measured and there are no parallel
lines in the plane! For a modern treatment of projective geometry see [154].

Exercises
1. To get a feel for the Mohr–Mascheroni result, using compass only write (and
prove) an algorithm for:
(a) constructing 3 collinear points,
(b) given a segment OA construct a point B such that OB is twice as long as
OA,
(c) bisecting a segment,
(d) trisecting a segment,
(e) given a circle C and two points determining a line l, construct the point
of intersection C ∩ l (be careful here, the line l is not given, but only two
points on the line l are. This is what makes the problem hard).
(f) constructing the point of intersection between two lines (hard).

For a short exposition on the geometry of compasses see [145, (pp. 23–25)] and [36, (pp. 145–151)],
11

quoted in [38].
 2.2 ♣ FT-1 Why Projective Geometry? 9

2. Given an angle A, provide a construction for finding an angle = 1/2 A using

the Steiner–Poncelet constraint, i.e., with a straightedge, a single circle, and its
center.
To recall and rejuvenate some Euclidean geometry skills try some of the
following12 :
3. Provide a Euclidean geometry construction (i.e., using circle and ruler only)
that transforms a rectangle into a square of equal area.
4. One of the three famous unsolvable problems in Euclidean geometry is “squar-
ing the circle.” Given a square with area A construct, using ruler and compass
only, a square having area 2A. What other shapes can this be generalized to?
5. Prove that a triangle ABC is isosceles if the bisectors AD and BE of the two
angles A and B are equal. Here D and E are the points of intersection of the
angle bisectors with the sides BC and AC respectively. This is a deceptively
hard problem.

2.2 ♣ FT-1 Why Projective Geometry?

We begin the formal aspects with an excursion into projective geometry, but why?
There are two important concepts and tools that we will need and that are also
widely used in many other fields and applications. They are:

• duality and
• homogeneous coordinates (see [146] for applications of homogeneous coordi-
nates to computer graphics).
Both of these are projective concepts and can be properly understood only in
this context. The experience is quite broadening and will get us “unglued” from the
narrow mindset of Euclidean geometry. For example, entirely new notions, such
as finite geometries, arise naturally. Projective geometry was initially motivated by
the fine arts. The Italian architect Brunelleschi in 1425 first proposed a geometrical
theory of perspective that was formalized by Alberti a few years later. It included
the idea of “points at infinity” and influenced art in a profound way, as shown by
contrasting prior “flat” paintings such as Martini’s Annunciation with those where
“depth” is perceived, such as Leonardo’s Last Supper and Raphael’s Athens, shown
in the chapter on projective geometry by Morris Kline [140, vol. 1, 622–641].
In the absence of photography, realism in drawings had not only artistic but also
great practical value.
12
See also the delightful short paper by Hudson [90].
10 2. Geometry Background

Figure 2.1. Projections and Sections.

2.2.1 Projective Plane Model

Projective geometry deals with
properties that are invariant under
projections and sections, see Fig. 2.1,
such as “incidence” and “collinear-
ity.” A beautiful projective theorem,
discovered by the French architect
Desargue in 1640, states that two
triangles, ABC and A
B
C
, are in
perspective from a point O if and
only if they are in perspective from Figure 2.2. Points and lines in Perspective.
a line TSR; see Fig. 2.4 (left) in the
Euclidean plane and Fig. 2.4 (right), where O is a “point at infinity” in the vertical
direction. Note the role of duality in the interchange between the words point and
line. Of course, we have not explained what is meant by “perspective”; see Fig. 2.2
for an indication, but we are on the way.
It will help our intuition to have a model for visualizing the projective space,
and only afterward do we present the more formal axiomatic treatment. Our “build-
ing blocks,” otherwise called primitives, are concepts and objects with which we
are already familiar. The notation commonly used in geometry appears in the
list below; points and lines are denoted by capital and lowercase Latin letters
(preferably in italics) respectively, while planes are denoted by lowercase Greek
letters.
 2.2 ♣ FT-1 Why Projective Geometry? 11

Primitives
• points - P, Q, A
• lines - , r, s
• planes - π, α, β
• incidence relation - P is incident to , or P is on .
Of course, incidence and
containment are not equi-
valent, since P is incident
to if and only if is inci-
dent to P ; P ∈ is true and
∈ P is not.
We start with the Eucli-
dean space R3 and add Figure 2.3. Ideal points and augmented lines.
something new. Let π ⊂ R 3
be a plane. To each line ⊂ π we associate an ideal point (“point at infinity”) L∞
such that distinct intersecting lines have distinct ideal points while parallel lines
have the same ideal point. Effectively, then, ideal points signify direction. Since no

Figure 2.4. Desargue’s theorem. The three points T , S, R, T = B

C
∩ BC, S = A
C
∩ AC,
R = A
B
∩ AB, are always collinear for when the triangles are in perspective with respect to an
ordinary point (left) or an ideal point (“point at infinity”) (right).
12 2. Geometry Background

distinction is made between the “+” and “−” directions for each line , one and not
two L∞ need to be appended.
On occasion, the notation Pm∞ is used for ideal points, with m being the slope
of the corresponding direction. The points of R3 are called regular to distinguish
them from the “ideal” ones. Proceeding, we call the pair ( , L∞ ) an augmented
line. Then for r, s ⊂ π,

P ordinary point ⇐⇒ r not  s,
(r, R∞ ) ∩ (s, S∞ ) =
R∞ ⇐⇒ r  s.

For a given plane π ⊂ R3 , let p∞ = {L∞ | ⊂ π, ( , L∞ )} be the collection

of all ideal points (directions) on π. We will see that p∞ satisfies the axioms for
a line, and for this reason it is called an ideal line. The pair (π, p∞ ) is called an
augmented plane for, as will be seen, it satisfies the axioms for a plane. Proceeding,
the ideal lines for each plane π ⊂ R3 are collected and denoted by π∞ = {p∞ |π ⊂
R3 , (π, p∞ )}; we refer to π∞ as the ideal plane and the pair (R3 , π∞ ) as the
augmented space. All in all, then, “points” are either ordinary or ideal, “lines” are
either augmented or ideal, and “planes” are either augmented planes or the ideal
plane.
Hilbert [85] proposed an elegant model of the projective plane whose construc-
tion (actually a modified version) is shown in Fig. 2.6. A hemisphere S together with
the disk D capping it is placed in a Euclidean plane π, with its south pole tangent
to the origin. A stereographic projection is established such that a point P ∈ π is
connected to the center C of S. The point P
= CP∩S is the image of P . All points of
π (i.e., “ordinary points”) are mapped uniquely onto the surface points of S. As the
point P = (a1 , a2 ) goes farther and farther away from the origin in the constant
direction having slope m = a2 /a1 , it approaches the ideal point13 Pm∞ . Then the
line CP approaches the line parallel to π that includes the diameter dm of the disk D
having slope m. This establishes a one-to-one correspondence dm ←→ Pm∞ . The
points on the hemisphere’s surface are the images of the ordinary points, while the
diameters of D are the images of the ideal points providing the image of all of P2
as described above.14 It is a very helpful visual model. For example, a line ⊂ π
is mapped into a great semicircle (see exercise below), and the augmented line
∞ ) includes the diameter d . Parallel augmented lines, as shown in Fig. 2.7,
(l, Pm m
“intersect” at their common ideal point. Two points on a projective line specify two

13
The subscript m, in the notation of an ideal point, indicates the corresponding direction, and it is
included only when the direction is relevant.
14
We use homogeneous coordinates only with real values so there is no need to use RP2 as seen
elsewhere, the R distinguishing the real projective plane from others.
 2.2 ♣ FT-1 Why Projective Geometry? 13

Figure 2.5. On the left, figure for Exercise 4; and on the right, figure for Exercise 5.

intervals; one containing the ideal point and its complement being the “ordinary”
interval of the Euclidean line.
A very basic though often unfamiliar property of P2 is that it is not orientable.
That is, it is not possible to assign a consistent sense of left and right ([18] p. 372).
This matter arises in various topics in our development, and we shall learn how to
cope with it when the need arises.

Exercises
1. Show that lines are mapped into great semicircles by stereographic projection.
2. Show that stereographic projection preserves angles (i.e., is isogonal)
3. How are circles mapped by stereographic projection?
4. On a piece of paper, as shown in Fig. 2.5 (left), there is a point R and portions of
two lines intersecting at some point T outside the paper. Using only a straight-
edge (ruler), draw the line joining the points R and T (Hint: Use Desargues’s
theorem). From [13, vol. I, p. 119].
5. Given two points R and T and a short ruler, as shown in Fig. 2.5 (right), not
long enough to reach both points, draw the line joining R and T (Hints: Use the
result of the previous exercise and slide the ruler along a line segment.) From
[13, vol. I, p. 119].

2.2.2 Axioms for Projective Geometry

The time has come to describe projective 3-space more formally in terms of the
axioms, i.e., the properties of the primitives: points, lines, planes, and the incidence
(or “on”) relation.
14 2. Geometry Background

Figure 2.6. Model of the projective plane.

Figure 2.7. Parallel lines on the projective plane model.

 2.2 ♣ FT-1 Why Projective Geometry? 15

Axioms for 2-D and 3-D

1. (2-D) Two distinct points are on (or incident to) a
unique line. That is, if P = Q then P , Q ∈ 1 and P , Q ∈
2 ⇒ 1 = 2 .
2. (3-D) Three distinct noncollinear points are on a
unique plane. That is, if P = Q, Q = R, P = R, and
there is no line  {P , Q, R} ⊂ , then {P , Q, R} ⊂ πi ,
i = 1, 2 ⇒ π1 = π2 .
3. (* 2-D) Two distinct coplanar lines are on (i.e.,
intersect or determine) a unique point. With the “ideal”
elements we always have 1 ∩ l2 = ∅ and P1 , P2 ∈
1 ∩ 2 ⇒ P1 = P2 .
This axiom and the remaining axioms are
marked with an asterisk (∗), for they differ from
their corresponding Euclidean cousins. Specifi-
cally, this axiom corresponds to Euclid’s famous
fifth postulate, which postulated the existence of
parallel lines.
4. (* 3-D) A plane π and a line not
contained in π are on (i.e., intersect at)
a unique point ∩ π = ∅ always and
P1 , P2 ∈ ∩ π ⇒ P1 = P2 .
5. (* 3-D) Two distinct planes are on
(i.e., intersect in) a unique line π1 ∩π2 = ∅
and π1 ∩ π2 = , where is unique.
It is not difficult to prove
that the Euclidean plane
together with the ideal and
augmented elements satis-
fies the projective plane
axioms. Let us assume that
ideal points and augmented
lines satisfy the axioms
(see exercises) and study an
example.
Theorem 2.2.1. The ideal
line p∞ of an augmented
plane (π, p∞ ) is a line.
16 2. Geometry Background

Figure 2.8. Axiom 1 is illustrated in (a) Figure 2.9. Self-dual: point

and its dual axiom (3) in (b). on line and line on point.

Proof. We need to show that p∞ satisfies the axioms for lines, namely axioms 3
and 4.
Let ( , L∞ ) be any augmented line on (π, p∞ ). Since p∞ is the collection
of all ideal points on the augmented plane, ( , L∞ ) ∩ p∞ = L∞ , so axiom 3 is
satisfied.
Let (π
, p
∞ ) be an augmented plane not containing p∞ ; therefore π
is not 
to π. Hence π ∩ π
= ⇒ (π, p∞ ) ∩ (π
, p∞
) = ( , L ), where L is the ideal
∞ ∞
point associated with . 
Two excellent references for projective geometry are [7] and [39].

Exercises
1. Show that the ideal points satisfy the axioms of points.
2. Show that augmented lines satisfy the axioms of lines.
3. Show that augmented planes satisfy the axioms of planes.
4. Show that the ideal plane π∞ satisfies the axioms of a plane.

2.2.3 Principle of Duality

The axioms yield a principle not found in Euclidean geometry. In 2-D there is
a point ↔ line duality stemming from (axiom 1) two points are on a line and
(axiom 3) two lines are on a point L. As a result, planar figures can be “dualized”
by switching the words point and line. For example, three distinct noncollinear (i.e.,
not on the same line) points P, Q, and R determine a triangle. Dually, three distinct
nonintersecting (i.e., not on the same point) lines p, q, r also determine a triangle.
That is, triangles are self-dual. Note also the projective cases in Fig. 2.11: (left)
 2.2 ♣ FT-1 Why Projective Geometry? 17

Figure 2.10. Self-dual 2: Triangle (a) and “triline” in (b).

Figure 2.11. Projective triangles (left) and quadrangle (right).

triangles and (right) the four points P , Q, R∞ , S∞ , which form a quadrilateral

(“quadrangle”) with edges PQ, PR∞ , QS∞ , ∞ .
The 3-D duality is point ↔ plane and results from axiom 2, whose dual is that
three noncollinear planes π1 , π2 , π3 are on a point P ; Fig. 2.12. By the way, this
is the reason why point ↔ line dualities do not generalize to higher dimensions.
Other examples of the point ↔ plane duality are that two planes π1 , π2 are on a
line, whose dual is axiom 1, and two points P1 , P2 are on a line. Also axiom 4: a
plane π and a line l not on π are on (determine) a point is dualized to a point P and
a line l not on P are on (determine) a plane.
The implications of duality are far deeper than those just indicated. For if a
(projective) theorem involves points and lines in 2-D or points and planes in 3-D,
its dual is also true. We illustrate this with the earliest known projective theorem,
18 2. Geometry Background

Figure 2.12. Three noncollinear planes on a point; dual of axiom 2.

Figure 2.13. Pappus’s theorem (left) and its dual (right).

proved by Pappus in the third century C.E. Using the notation ·, instead of ∩,
for intersection Pappus’s theorem states that for any distinct points Ai on a line
r and any distinct points Bi on a line s, where i = 1, 2, 3, the points Ci , where
Ci = Aj Bk · Ak Bj for j = k = i, are always on a line p. The dual theorem
is obtained simply by interchanging the words point ↔ line. In our notation it is
 2.2 ** Finite Geometries 19

equivalent to switching lowercase ↔ capital letters as follows: for any distinct lines
ai on a point R and any distinct lines bi on a point S, where i = 1, 2, 3, the lines ci ,
where ci = aj bk · ak bj for j = k = i, are always on a point P . No further proof
is needed!

Fun and Games with the ILM

Open the ILM: set Chapter = 2D Point Line duality.

The display shows two sets of axes.
On the right, click to draw a point and see the corresponding line on the left
Move the point and watch the line.

Next set Subject = Point ↔ Line.

Draw a point on the left and see the corresponding line on the right.

2.3 ** Finite Geometries

Let a finit set of points and a nonempty collection of point subsets called lines
satisfy the following axioms:
1. Any two distinct points A and B are on a unique line .
2. If p and q are distinct lines, there is a unique point on both.
3. There are at least three distinct points on any line.
4. Not all points are on the same line.
Such a set is called a finit geometry. It exemplifies the stark contrast with Euclidean
geometry. Point, line, and the on relation are primitives specified by the above
axioms only. Additional properties can be introduced to enrich the geometry.

2.3.1 Seven-Point Geometry

From the third axiom there are n ≥ 3 points on a line. The smallest geometry,
with n = 3 points on each line and a total of n2 − n + 1 = 7 (see Exercise 3
below) points or lines is illustrated in Fig. 2.14. The lines are a = [A, D, B],
b = [A, E, C], c = [C, F, B], d = [A, G, F ], e = [B, G, E], f = [C, G, D],
g = [D, E, F ]. Observe that the projective plane axioms are satisfied (i.e., two
distinct points are on a unique line and vice versa) for ∀ lines x ∈ {a, . . . , g}
20 2. Geometry Background

Figure 2.14. A “picture” of the 7-point geometry.

Figure 2.15. “Pictures” for 13-point geometry. There are four “lines” on each point and vice versa.
On the right, the points are indicated in bold numbers and lines in italics.

and ∀ points X ∈ {A, . . . G}. The lines/curves joining the • in the picture have
no significance other than showing the association of the points belonging to a
particular line.
 2.3 ** Finite Geometries 21

↓ lines 1 2 3 4 5 6 7 8 9 10 11 12 13 ← points
1 1 2 3 4
2 1 5 6 7
3 1 8 9 10
4 1 11 12 13
5 2 5 8 11
6 2 6 9 12
7 2 7 10 13
8 3 5 9 13
9 3 6 10 11
10 3 7 8 12
11 4 5 10 12
12 4 6 8 13
13 4 7 9 11
Figure 2.16. Incidence matrix for the 13-point geometry.

2.3.2 Finite Geometry with 13 Points

Finite geometries can be pictured in different ways, as in Figure 2.15, and specified
by an incidence matrix for points and lines as for the 13-point geometry. Here, it
is more convenient to denote the points and lines by numbers rather than letters.
Those interested in pursuing the fascinating subject of finite geometries are referred
to [116] and [117].

Exercises
1. Prove the 3-D dual shown in Fig. 2.12, i.e., that three noncollinear planes are
on a point, using only the projective 3-D space axioms.
2. The second axiom may be weakened to 2a: “If p and q are distinct lines, there
is at least one point on on both lines.” Prove that with axiom 2 replaced by 2a
above, the resulting set of points and lines also satisfies the projective plane
axioms.
3. Prove that in a finite geometry with n points on any line there are n2 − n + 1
points (or lines) altogether.
4. Construct diagrams for the geometries with n = 5 and 6 points and try to
generalize. This is hard: you may want to look up a topic in combinatorics
called “Latin squares.”
22 2. Geometry Background

Figure 2.17. Homogeneous coordinates. Correspondence between points P in the x1 x2 -plane and
lines p through the origin in 3-D.

2.4 ♣ FT-2 Analytic Projective Geometry

Even though there is no notion of distance in projective space, it is possible to

introduce a very useful coordinate system. The geometric development here is for
the plane and can be generalized to projective N-space PN .
We consider the x1 x2 plane in R3 with Cartesian coordinates as shown in
Fig. 2.17 and the plane π : x3 = 1. The point P = (a1 , a2 ) in the x1 x2 -plane is
identified with the point P
= (a1 , a2 , 1) in the plane π . The line p through the
points (0, 0, 0) and P
is constructed and is described by the equations
⎧
⎨ x2 = aa2 x1 ,
1
p:
⎩ x = 1x .
3 a2 2
 2.4 ♣ FT-2 Analytic Projective Geometry 23

In this way, the point-to-line correspondence

P = (a1 , a2 ) ∈ x1 x2 -plane

←→ p ⊂ x1 x2 x3 -space, p on (a1 , a2 , 1), (0, 0, 0),

is established. Since p can be identified by any one of its points (ka1 , ka2 , k) for
k = 0, the correspondence can be alternatively written as

P : (a1 , a2 ) ←→ (ka1 , ka2 , k), 0 = k ∈ R.

That is, (k1 a1 , k1 a2 , k1 ) and any other multiple (k2 a1 , k2 a2 , k2 ) for k1 = k2 rep-
resent the same point, since they both lie on the same line p. These triples are the
point coordinates of P .
Let us return to Fig. 2.17 and imagine the point P moving farther away from
the origin but always in the direction with slope m = aa21 approaching the ideal
point in this direction. The point coordinates of P are (a1 , a2 , 1) or (1, m, a11 ), and
as a1 → ∞ the ideal point is represented by the triples (ka1 , ka2 , 0) for k = 0.
The important realization is that the 0 in the third component identifies ideal points.
This system is called homogeneous coordinates, allowing the representation both
ordinary and ideal points, something that cannot be done with Cartesian coordinates;
it has other nice properties as well.
Appealing to duality, a line x can be represented as an ordered triple of real
numbers that is distinguished from a triple representing points using capitals for the
coordinates themselves and “[ ]” instead of “( )” as

x : [X1 , X2 , X3 ] = [0, 0, 0],

again with the convention that [kX1 , kX2 , kX3 ], ∀k = 0, represents the same line.
These triples are called line coordinates. So (1, 2, 3) represents a point and [1, 2, 3]
a line. Actually, this is quite reasonable, for consider a line described in Cartesian
coordinates by

A1 x1 + A2 x2 + A3 = 0. (2.1)

For A3 = 0 the same line is also given by

A1 A2
x1 + x2 + 1 = 0. (2.2)
A3 A3
In either case, the three constants A1 , A2 , A3 completely specify the line. So taking
the triple [A1 , A2 , A3 ] with the understanding that Ai is the coefficient of xi for
24 2. Geometry Background

i = 1, 2 and A3 is the constant completely describes the line in (2.1). Equivalently,

the triple [ A1 A2
A3 , A3 , 1] describes the line as given by (2.2), and hence these two triples
and their multiples identify the same line.
There remains to define the on relation. We say that a point X = (x1 , x2 , x3 ) is
on the line x = [X1 , X2 , X3 ], and dually the line x is on the point if and only if

X · x = x · X = x1 X1 + x2 X2 + x3 X3 = 0.

Consider now the three points Y = (y1 , y2 , y3 ), Z = (z1 , z2 , z3 ), W =

(w1 , w2 , w3 ), all on the line x = [X1 , X2 , X3 ]. Then

y1 X1 + y2 X2 + y3 X3 = 0,
z1 X1 + z2 X2 + z3 X3 = 0,
w1 X1 + w2 X2 + w3 X3 = 0.

This system of linear homogeneous equations has a nontrivial solution [X1 , X2 , X3 ]

if and only if the deteterminant of the coefficients is zero. That is, Y, Z, W are on
a line x if and only if
 
 y1 y2 y3 
 
 z1 z2 z3  = 0. (2.3)
 
 w1 w2 w3 

This condition provides the line coordinates of x as

     

 y2 y3   y3 y1   y1 y2 
 ,   (2.4)
 z2 z3   z3 z1  ,  z1 z2 
.

A direct consequence is that the points X, Y , and λX + μY for λ, μ ∈ R are

collinear.
What are the line coordinates of the ideal line? We can find out by substituting
two distinct ideal points in (2.4), i.e., z3 = y3 = 0. This yields [0, 0, y1 z2 − y2 z1 ]
or [0, 0, 1]. So far, it is all falling into place: we have a coordinate system that
provides coordinates for every regular point (from which its Cartesian coordinates
can be recovered) and also coordinates for the ideal elements (points and line).
In addition, homogeneous coordinates can be extended to projective N -space and
finite geometries. ♣ FT-2e
Consider now the 7-point geometry (i.e., n = 3) described earlier and take the
algebra J2 , the integers modulo 2, with elements {1, 0}. The binary operations +
and × are defined by
 2.4 ♣ FT-2 Analytic Projective Geometry 25

+ 0 1 × 0 1
0 0 1 0 0 0 .
1 1 0 1 0 1
Now we collect all triples (x1 , x2 , x3 ) except (0, 0, 0) and label them:

A = (0, 0, 1), B = (0, 1, 0), C = (1, 0, 0),

D = (0, 1, 1), E = (1, 0, 1), F = (1, 1, 0), G = (1, 1, 1).

A pair of points is chosen specifying a line x = [X1 , X2 , X3 ] whose coordinates

are computed from (2.4) for Xi ∈ J2 , yielding

a = [1, 0, 0] b = [0, 1, 0] c = [0, 0, 1]

d = [1, 1, 0] e = [1, 0, 1] f = [0, 1, 1] g = [1, 1, 1].

Note that each point is on two lines; for example A is on the lines having X3 = 0.
Similarly, other finite geometries can be coordinatized using the appropriate algebra
Jp of integers modulo p, that is, the set of integers {1, 2, . . . , p − 1} with addition
and multiplication analogously defined (see a book on modern algebra). To clarify,
let us use J3 to coordinatize the 13-point geometry. Excluding (0, 0, 0), there are
26 triples (a1 , a2 , a3 ), ai ∈ J3 , with the multiples k(a1 , a2 , a3 ), k = 0 ∈ J3 , also
representing the same point. The only multiple of interest is 2(a1 , a2 , a3 ), which
equals (a1 , a2 , a3 ) for ai ∈ J3 , yielding 13 distinct triples. Similarly, with J4 there
63 nonzero triples whose multiples by 2 or 3 are the same in that algebra, yielding
21 distinct triples, which can therefore serve as the homogeneous coordinates of
the 21-point geometry (Exercise 3 below).

2.4.1 Homogeneous Coordinates on a Line

Let us specialize homogeneous coordinates to a line o, which for convenience we
take as being on the two points

O = (0, 1, 1), U = (1, 1, 1).

The coordinates of a point X = (x1 , x2 , x2 ) ∈ o together with those of O and U

satisfy the collinearity condition of (2.3); therefore x2 = x3 , so X = (x1 /x2 , 1, 1),
with X = (x1 , 0, 0) being the line’s ideal point if x2 = 0. Since there is no
variation in the third component, the coordinates of regular points can be simplified
to O = (0, 1), U = (1, 1) and in general, X = (x1 , x2 ), as indicated in Fig. 2.18.
We can write

X : (x1 , x2 ) = a2 (1, 1) + a1 (0, 1) = a2 U + a1 O

26 2. Geometry Background

Figure 2.18. Homogeneous coordinates on the line.

Figure 2.19. Two lines p, p

in perspective from the point P .

for x1 = a2 , x2 = a1 + a2 , and O, U acting as a basis. The point O can be thought

of as the “origin,” and U establishes the “unit”; recall that we do not have a metric
in the projective plane. There is a further simplification obtained by dividing by x2
to represent the point X by the single value x = x1 /x2 .

2.5 The Fundamental Transformations

of Projective Geometry

A short diversion into the basic transformations of projective geometry is needed in

order to prove a fundamental property of homogeneous coordinates. Consider two
directed (in the Euclidean plane) lines p, p
and a point P not on the lines, as shown
in Fig. 2.19. A line on P intersecting p at the point A and p
at A
establishes a
one-to-one correspondence called perspectivity between the points of p and p
. It is
denoted by
 2.5 The Fundamental Transformations of Projective Geometry 27

p(A, B, C, D, . . . ) ∧ p
(A
, B
, C
, D
, . . . ).
P

The two lines are said to be in perspective with respect to the point P , which
may be indicated above the ∧ symbol. Of course, there is the dual concept, shown
in Fig. 2.20, of two points P , P
being in perspective with respect to a line p
denoted by
P (a, b, c, d, . . . ) ∧ P
(a
, b
, c
, d
, . . . ).

Figure 2.20. Two points P , P

in perspective from the line p.

Consider now a perspectivity p(A, B, C, D, . . . ) ∧ p

(A
, B
, C
, D
, . . . ) as
shown in Fig. 2.21 and the cross-ratio
AC/BC
(A, B; C, D) = .
AD/BD

Theorem 2.5.1. The cross-ratio is invariant under perspectivity.

Proof. We give an elementary proof in the Euclidean plane, leaving the projective
aspects for the exercises. With reference to Fig. 2.21,
IP·AC areaAPC AP·CP sin(APC)
IP·BC areaBPC BP·CP sin(BPC)
(A, B; C, D) = IP·AD
= areaAPD
= AP·DP sin(APD)
,
IP·BD areaBPD BP·DP sin(BPD)
or
sin(APC) · sin(BPD)
(A, B; C, D) = ± . (2.5)
sin(BPC) · sin(APD)
28 2. Geometry Background

Figure 2.21. Diagram for the cross-ratio property.

Similarly, it can be shown that (A

, B
; C
, D
) also equals the right-hand-side of
(2.5). There remains to show that the signs match (see exercises below). As stated,
this definition of the cross-ratio applies only to the Euclidean plane, but it can
be appropriately generalized to the augmented plane (π, p∞ ), the cross-section
invariance being a projective property. 
The set of all points on a line is called a pencil of points, and dually, the set
of all lines on a point is called a pencil of lines. A pencil of points may undergo
successive perspectivities, as in

P P Pn−1
p1 ∧1 p2 ∧2 . . . ∧
pn .

A sequence of perspectivities is called a projectivity and is denoted by p ∧ pn . In pro-

jective geometry we study the properties that remain invariant under projectivities.
We state without proof a couple of important results.
Theorem 2.5.2. A projectivity is determined by the correspondence between two
triples of distinct points. That is, if the triples (A, B, C) and (A
, B
, C
) correspond
under a projectivity p ∧ pn , then the correspondence between any X ∈ p and
X
∈ p
is uniquely determined (i.e., projectivity is determined by just three points).
This does not determine the actual sequence of perspectivities from which the
projectivity resulted. Rather, it just determines the correspondence between the
original and final pencils of points.
Theorem 2.5.3. A projectivity can be expressed by at most two perspectivities.
 2.6 More on Homogeneous Coordinates 29

Figure 2.22. A projectivity composed of four perspectivities.

Exercises
1. Show that the 24 distinct permutations of four symbols A, B, C, and D result
in exactly six different values of the cross-ratios as formed from these four
symbols (corresponding to four points on a line)
2. Prove Theorem 2.5.1 when the perspectivity is with respect to an ideal point.
3. Prove Theorem 2.5.2 in the Euclidean plane and also with a perspectivity with
respect to an ideal point.
4. Prove Theorem 2.5.3 in the Euclidean plane and also with a perspectivity with
respect to an ideal point.

2.6 More on Homogeneous Coordinates

2.6.1 Linearizability of Projective Transformations
It is easy to see that the cross-ratio remains invariant under a projectivity and in
fact, completely characterizes (determines) the projectivity. This invariance leads
to a very important result and the reason homogeneous coordinates are so widely
used in fields such as computer graphics. It is proved first for the plane P2 and then
generalized to higher dimensions. For a projectivity s ∧ s
between two directed lines
30 2. Geometry Background

s and s
(as in Fig. 2.21, see lines p and p
), four points and their corresponding
single homogeneous coordinates indicated by (), P : (p), Q : (q), R : (r), X : (x)
are chosen to form the cross-ratio
(p − r)(q − x)
(P , Q; R, X) = .
(p − x)(q − r)
Similarly, the cross-ratio
(p
− r
)(q
− x
)
(P
, Q
; R
, X
) =
(p
− x
)(q
− r
)
is formed with the corresponding points P
: (p
), Q
: (q
), R
: (r
), X
: (x
) on
s
. By the cross-ratio invariance, under a projectivity, we have

(P , Q; R, X) = (P
, Q
; R
, X
), (2.6)

yielding
ax + b
x
= . (2.7)
cx + d
The a, b, c, d are given in terms of the coordinates p, q, r, p
, q
, and r
, and their
actual values are immaterial to the development. This is a linear fractional transfor-
mation, also called a Möbius transformation, completely specifying the projectivity
s ∧ s
. Reverting back to the ratio obtained from the homogeneous coordinates, i.e.,
x = x1 /x2 and x
= x1
/x2
, we obtain from (2.7),
x1
ax1 + bx2

= . (2.8)
x2 cx1 + dx2
Then rewriting the homogeneous coordinates of a point on a line in terms of two
components as X = (x1 , x2 ), see Fig. 2.18, we get
  
x
1 ax1 + bx2
k ,1 = ,1 .
x2
cx1 + dx2

Here k is an arbitrary proportionality constant, which is needed, since multiples of

a triple (the 1 in the third position is not shown; see Fig. 2.18) represent the same
point. Eliminating the fractions yields

k(x1
, x2
) = (ax1 + bx, cx1 + dx2 ),

from which we obtain the linear relations

kx1
= ax1 + bx2 , kx2
= cx1 + dx2 (2.9)
 2.6 ** The Linearized Projective Transformations 31

between the coordinates on the directed lines s and s

. The transformation (2.8) is
called projective, and this is the most general class of nonlinear transformations
that can be linearized by homogeneous coordinates. It is worth examining this in
greater detail. Rewrite the transformation as
   
ax + b bc − ad 1 a
T (x) = = · + ,
cx + d c2 x + d/c c
which provides the decomposition of T into

T = T4 ◦ T3 ◦ T2 ◦ T1 ,

where ◦ denotes the transformation composition and

d 1 bc − ad a
T1 (x) = x + , T2 (x) = , T3 (x) = x, T4 (x) = x + .
c x c2 c
That is, T is composed of two translations T1 , T4 , a scaling T3 , and an inversion T2 .
Inversions and translations, being nonlinear, are linearized, and this property makes
homogeneous coordinates especially useful in applications, since the projective
transformations, translation, perspective, rotation, and scaling, can all be described
in terms of matrices. Hence, a sequence of such transformations can be expressed as
a matrix product and computed in a convenient and efficient way (e.g., see [146]).
For n-dimensional projective space PN and projectivities there, the linear
fractional transformation (2.7) generalizes to
a1i x1
+ a2i x2
+ · · · + aN,i xN

+a
(N +1),i
xi

= , i = 1, 2, . . . , n, (2.10)
b1i x1
+ b2i x2
+ · · · + bN,i xN

+b
(N +1),i

where we have reverted to homogeneous coordinates by

x
xj
xi

=
i , xj
= , j = 1, 2, . . . , (n + 1).
x(N+1) x(N +1)

Here the xi are Cartesian coordinates, while the xi

and xi

are homogeneous
coordinates.

2.7 ** The Linearized Projective Transformations

To recap, the projective transformation

T (x) = ax (2.11)
32 2. Geometry Background

is linear and is called scaling when a > 0, a = 1; then it is a contraction for a < 1
and a stretching, or magnificatio , when a > 1. For complex a = r(cos θ + i sin θ),
the result is a rotation of x by an angle θ about the origin as well as a scaling by r.
By contrast, translation
T (x) = x + a (2.12)
and inversion or perspective
1
T (x) = (2.13)
x
are not linear, and for these the linearization, accomplished via homogeneous
coordinates, enables the transformations to be expressed in matrix form. This is
of importance in such applications as computer graphics and geometric modeling.

2.7.1 Projective Transformations in the Plane P2

For notational convenience, a point P = (x1 , x2 , x3 ) and its transform P
=
(x1
, x2
, x3
), expressed in homogeneous coordinates, are taken as row vectors, while
the transformations are denoted by two capital letters. We take x3
= c
and x3 = c,
such that for ordinary points, c, c
= 0, and c, c
= 0 for ideal points. Many
references in computer graphics, geometric modeling, robotics, and other fields
consider only the case x3 = 1, which unnecessarily excludes the ideal points.
The scaling transformation P
= PS(s1 , s2 ), where xi
= si xi and x3
= x3 = c,
in matrix form is
⎛ ⎞
s1 0 0
PS : (x1
, x2
, c) = (x1 , x2 , c)⎝ 0 s2 0 ⎠. (2.14)
0 0 1

The rotation P
= PR(θ) by an angle θ about the origin is
⎛ ⎞
cos θ sin θ 0
PR : (x1
, x2
, c) = (x1 , x2 , c)⎝ − sin θ cos θ 0 ⎠. (2.15)
0 0 1

For the translation P

= PT(t1 , t2 ), with xi
= xi + ti , we can set c = 1, since
translation of ideal points is not meaningful, to obtain
⎛ ⎞
1 0 0
PT : (x1
, x2
, 1) = (x1 , x2 , 1)⎝ 0 1 0 ⎠. (2.16)
t1 t2 1
 2.7 ** The Linearized Projective Transformations 33

Successive scaling or rotation transformations result in successive multiplications

of the corresponding matrices, as they should, since these are linear transformations.
The fun and beauty is that it also works with translations. For P
= PT1 (t11 , t21 ) and
P

= P
T2 (t12 , t22 ), the result is P

= PT1 T2 = PT2 T1 so that xi

= xi + ti1 + ti2 .
The matrices commute with respect to multiplication, as they should, since the
result of successive translations does not depend on their order.
The composition of scalings, rotations, and translations is obtained by the
multiplication of the corresponding matrices. In addition, the matrices are sparse
(i.e., they have many zeros), and the computational advantage is tremendous.

2.7.2 Projective Transformations in P3

The generalizations of these transformations to 3-D are straightforward, with the
previous remarks regarding x3 = c now pertaining to x4 = c. For scaling,
P
= PS(s1 , s2 , s3 ),
⎛ ⎞
s1 0 0 0
⎜ 0 s2 0 0 ⎟
PS : (x1
, x2
, x3
, c) = (x1 , x2 , x3 , c)⎜ ⎟
⎝ 0 0 s3 0 ⎠ . (2.17)
0 0 0 1

Let the matrices for the rotation about the xi -axes by an angle θ be denoted by
PRi (θ), i = 1, 2, 3. The rotation about x3 is given by
⎛ ⎞
cos θ sin θ 0 0
⎜ − sin θ cos θ 0 0 ⎟
PR3 : (x1
, x2
, x3
, c) = (x1 , x2 , x3 , c)⎜
⎝
⎟, (2.18)
0 0 1 0 ⎠
0 0 0 1

and is a direct extension of the 2-D rotation. The two other rotation matrices are
⎛ ⎞
1 0 0 0
⎜ 0 cos θ sin θ 0 ⎟
PR1 (θ) = ⎜ ⎝ 0 − sin θ cos θ 0 ⎠ ,
⎟

0 0 0 1
⎛ ⎞ (2.19)
cos θ 0 − sin θ 0
⎜ 0 1 0 0 ⎟
PR2 (θ) = ⎜ ⎝ sin θ 0 cos θ
⎟.
0 ⎠
0 0 0 1
34 2. Geometry Background

Figure 2.23. Perspective with respect to the x3 -axis.

For 3-D translations P

= PT(t1 , t2 , t3 ), c = 1 as in 2-D, and
⎛ ⎞
1 0 0 0
⎜ 0 1 0 0 ⎟
PT : (x1
, x2
, x3
, 1) = (x1 , x2 , x3 , 1)⎜
⎝ 0
⎟. (2.20)
0 1 0 ⎠
t1 t2 t3 1

In 3-D, the perspective transformation is important, for it provides images with a

sense of depth and realism. It is convenient to choose the center of the perspective
transformation on the negative x3 -axis at a distance d behind the perspective plane,
as shown in Fig. 2.23. The transformation equations, obtained by similar triangles,
are
x1 x2
x1
= , x2
= , x3
= 0. (2.21)
x3 /d + 1 x3 /d + 1

Denote the perspective transformation matrix by V [πper , Cper ], where πper and
Cper are the perspective plane and center respectively. For the setup in Fig. 2.23,
P
= PV[x3 = 0, (0, 0, −d)], and explicitly,
 2.7 ** The Linearized Projective Transformations 35

⎛ ⎞
1 0 0 0
⎜ 0 1 0 0 ⎟
(x1
, x2
, 0, w) = (x1 , x2 , x3 , c)⎜
⎝ 0
⎟, (2.22)
0 0 1/d ⎠
0 0 0 1
with the relation between c and w emerging next. The perspective image of an
ideal point P = (1, x2 /x1 , x3 /x1 , 0), called a vanishing point, is the ordinary point
P
= (x1
, x2
, 0, x3 /d) = (dx1 /x3 , dx2 /x3 , 0, 1). Hence
x3 x1 x2
w= + c, x1
= , x2
= , x3
= 0.
d w w
All projective transformations can now be performed successively by matrix mul-
tiplication. The matrix of the perspective transformation, unlike those of other
projective transformations, is singular, confirming that an object cannot be uniquely
retrieved from its perspective image.

Figure 2.24. The perspective image of an object (partially shown).

In the figure, the perspective transformation of three vertices A, B, C is traced, showing their images
A
, B
, C
. This is sometimes called a 2-point perspective, having two vanishing points V1 and V2 , the
images of ideal points, giving the image “3-dimensionality.” Actually, the ideal point in the vertical
direction is the “third vanishing point.” It would have been a 3-point perspective had the perspective
plane not been parallel to the front (vertical) face of the object. All lines joining the object’s points
with their images intersect at the center of perspective.
36 2. Geometry Background

2.8 ** Coordinate System Rotations

This section, complementing the discussion of rotations about one axis above, is
adapted from an old gem [161, pp. 240–241] (the textbook from which I learned
analytic geometry). In R3 with a positive direction assigned to a line , translate
the coordinate system to any point Q ∈ , denoting by α, β, γ the direction
angles between the positive direction of and the translates of the axes x1 , x2 , x3
respectively. Their cosines, called direction cosines, have the property

cos2 α + cos2 β + cos2 γ = 1. (2.23)

Denoting by êi the unit vectors in the xi directions, we obtain that

ˆ = ê1 cos α + ê2 cos β + ê3 cos γ (2.24)

is the unit vector in the positive direction of , with (2.23) equivalent to ˆ · ˆ = | |

ˆ 2.
The cosine of the angle φ between two directed lines 1 , 2 is given in terms of
their corresponding unit vectors ˆ1 , ˆ2 , and their corresponding direction angles,
subscripted by 1 and 2, respectively, are given by

cos φ = ˆ1 · ˆ2 = cos α1 cos α2 + cos β1 cos β2 + cos γ1 cos γ2 , (2.25)

with cos φ = 0 when the lines are perpendicular.

For a plane π, let n be the directed line through the origin O and perpendicular
to π with direction angles α, β, γ . The normal form of the equation of π is

π : x1 cos α + x2 cos β + x3 cos γ = d, (2.26)

with d the directed distance of π from O. It is an easy matter to verify that the
directed distance p of π from the point P = (p1 , p2 , p3 ) is given by

p = p1 cos α + p2 cos β + p3 cos γ − d. (2.27)

With reference to Fig. 2.25, the coordinate system x1 , x2 , x3 is rotated about the
origin O to the system x1
, x2
, x3
, all points remaining stationary with a point P
having two sets of coordinates (x1 , x2 , x3 ), (x1
, x2
, x3
). Our task is to find the
relationship between them. Letting the direction cosines of the xi
axes with respect
to xi be cos αi , cos βi , cos γi for i = 1, 2, 3 respectively, we obtain

cos2 αi + cos2 βi + cos2 γi = 1,

cos αj cos αk + cos βj cos βk + cos γj cos γk = 0, (2.28)

 2.8 ** Coordinate System Rotations 37

Figure 2.25. Rotation of the coordinate system x1 x2 x3 about the origin.

where j = k i, j, k = 1, 2, 3, the first part corresponding to (2.23) for the xi

and the
second due to the orthogonality between the pairs xj
, xk
of axes. The plane x2
x3
,
which passes through the origin O, has equation

x2
x3
: x1 cos α1 + x2 cos β1 + x3 cos γ1 = 0,

with coefficients the direction cosines of the axis x1

, which is orthogonal to it.
Similarly, for all the 2-planes,

xj
xk
: x1 cos αi + x2 cos βi + x3 cos γi = 0, i = j = k. (2.29)

The directed distance of the point P from the plane xj

xk
is its xi
coordinate,
i = j = k; hence by (2.27), (2.29),

xi
= x1 cos αi + x2 cos βi + x3 cos γi . (2.30)

These describe the coordinate transformation {x1 , x2 , x3 } → {x1

, x2
, x3
}. Revers-
ing the roles of xi and xi
in the previous reasoning, we obtain the inverse
transformation {x1
, x2
, x3
} → {x1 , x2 , x3 }:

xi = x1 cos αi + x2 cos βj + x3 cos γk , i = j = k. (2.31)

38 2. Geometry Background

For consistency with Section 2.7, (2.30) and (2.31) are rewritten in homogeneous
coordinates and in matrix form:
⎛ ⎞
cos α1 cos β1 cos γ1 0
⎜ cos α2 cos β2 cos γ2 0 ⎟
(x1
, x2
, x3
, 1) = (x1 , x2 , x3 , 1)⎜ ⎟
⎝ cos α3 cos β3 cos γ3 0 ⎠ , (2.32)
0 0 0 1
⎛ ⎞
cos α1 cos α2 cos α3 0
⎜ cos β1 cos β2 cos β3 0 ⎟
(x1 , x2 , x3 , 1) = (x1
, x2
, x3
, 1)⎜ ⎟
⎝ cos γ1 cos γ2 cos γ3 0 ⎠ . (2.33)
0 0 0 1

The matrix A of the transformation in (2.30) has determinant |A| = 1, and

the matrix of the inverse transformation (2.31) is the transpose AT . These, and
the matrices corresponding to the rotations about one axis in Section 2.7, are
examples of orthogonal matrices (transformations), which also appear in the pre-
vious section. As a reminder, the above equations together with the six conditions
of (9.47) specify the rotation of the coordinate system about the origin. By the
way, such a rotation about an arbitrary point can be obtained by translating the
origin first and then rotating, or equivalently, multiplying the matrix A in (9.48)
by that in (2.20), with resulting matrix describing the combined transformation.
That is,
⎛ ⎞
cos α1 cos β1 cos γ1 0
⎜ cos α2 cos β2 cos γ2 0 ⎟
(x1
, x2
, x3
, 1) = (x1 , x2 , x3 , 1)⎜ ⎟
⎝ cos α3 cos β3 cos γ3 0 ⎠ . (2.34)
t1 t2 t3 1

From the vector and matrix form, the generalization to RN is direct, with a
slight change in notation denoting the direction angle between the xi axis and a
directed line ⊂ RN by αi and the unit vector in the direction of :


N 
N
ˆ = êi cos αi , cos2 αi = 1. (2.35)
i=1 i=1

Proceeding as above, the corresponding (N + 1) × (N + 1) transformation matrix is

obtained together with the (N 2 + N)/2 additional conditions analogous to (9.47):
N conditions from the above equation and N(N −1)/2 more due to the orthogonality
of all pairs of axes.
 2.8 ** Analytic Proofs 39

2.9 ** Analytic Proofs

Desargues’s theorem will be helpful in subsequent chapters. This is an oppurtune
time to prove it and learn yet another nice way to utilize homogeneous coordinates.

2.9.1 Computational Proof

Three noncollinear points Pi , specified in homogeneous coordinates by triples
Ai , i = 1, 2, 3, are chosen on the projective plane P2 . Another point P , with
triple A, is selected such that A = c1 A1 + c2 A2 + c3 A3 with all ci = 0.
Each point Pi is also represented by the triple ci Ai = A∗i , which are fixed from
now on, leaving A = A∗1 + A∗2 + A∗3 . Any other point Q can be described
as Q = a1 A∗1 + a2 A∗2 + a3 A∗3 with the A∗i of the multiples now fixed. The
coefficients can be used as a new set of coordinates, i.e., Q = (a1 , a2 , a3 ) and
P1 = (1, 0, 0), P2 = (0, 1, 0), P3 = (0, 0, 1), P = (1, 1, 1). In this coordinate
system it is easy to see that Q = (a1 , a2 , a3 ) and Q
= k(a1 , a2 , a3 ), k = 0, repre-
sent the same point, allowing us to write Q = Q
. This is similar to homogeneous
coordinates with 0 in the third position no longer implying that the point is ideal.
This turns out to be a very useful tool.
To illustrate, we provide an analytic proof of Desargues’s theorem. It is conve-
nient to take the triangle’s vertices A, B, C as the basis triples and P = A + B + C
or P = (1, 1, 1) as the point of perspectivity, as shown in Fig. 2.26. The three
lines from P through the vertices of the triangle are denoted by a
, b
, and c
. With
A = (1, 0, 0) and P = (1, 1, 1) on a
, from (2.4) we find that
     

 1 1   1 1   1 1 
a = 

, ,  = [0, 1, −1].
0 0   0 1   1 0 
For a point A
= (a, a1 , a2 ) ∈ a
, where A = A
, the collinearity of P , A, A
and
(2.3) yield
 
 a a1 a2 
 
 1 1 1  = 0,

 1 0 0 
which shows that a1 = a2 . Since these values are arbitrary, we can take A
=
(a, 1, 1). Similarly, we obtain the line coordinates b
= [1, 0, −1], c
= [1, −1, 0]
and the point coordinates B
= (1, b, 1), C
= (1, 1, c). Let r be the line determined
by A and B. Then
     

 0 0   0 1   1 0 
r=    ,   ,   = [0, 0, 1].
1 0   0 0   0 1 
40 2. Geometry Background

Figure 2.26. Diagram for analytic proof of Desargue’s theorem.

Similarly, for the line r

determined by A
and B
,
     

 1 1   a 1   a 1 
r = 

, ,  = [1 − b, a − 1, ab − 1].
b 1   1 1   1 b 

To find the intersection r ∩ r

= R, we can use the dual of (2.4). That is, in its
original form, (2.4) yields the coordinates of the line determined by the coordinates
of two of its points. In the dual, the same equation yields the coordinates of the
point on the two lines from their coordinates. Namely,
      
 0 1   1 0   0 0 
R =  ,
 
,
 

a − 1 ab − 1 ab − 1 1 − b 1−b a−1 

= (1 − a, b − 1, 0).
 2.9 ** Analytic Proofs 41

Continuing in the same way, we obtain

S = BC · B
C
= s ∩ s
= (0, 1 − b, c − 1),

T = AC · A
C
= t ∩ t
= (a − 1, 0, 1 − c).

Applying again (2.3) with the coordinates found for R, S, T yields

 
 1−a b−1
 0 
 0
 1 − b c − 1  = 0,
 a−1 0 1−c 

showing that R, S, and T are always collinear, proving Desargues’s theorem in the
projective plane.

2.9.2 Algebraic Proof

When the coordinates of the points and lines are not explicitly needed, their compu-
tation can be avoided with an algebraic proof. In Fig. 2.26, let all points be identified
by their homogeneous coordinate triples. By the collinearity of P , A, A
there exist
nonzero constants c1 , c2 such that

P = c1 A + c2 A
. (2.36)

The multiple c1 A identifies the same point as A, and similarly, c2 A

identifies −A
.
Without loss of generality we can take the triples c1 A, c2 A
for the representation
of the points A, −A
, simplifying (2.37) to

P = A − A
. (2.37)

The new triples for A, A

then remain fixe as in the previous section. Similarly,
there exist constant multiples of B, B
, C, C
such that

P = B − B
, P = C − C
. (2.38)

Then

A − B = A
− B
. (2.39)

The point A − B is on the line AB, and the point A

− B
is on A
B
. Therefore,

A − B = A
− B
= R = AB ∩ A
B
. (2.40)
42 2. Geometry Background

Continuing, we obtain

C − A = C
− A
= T = AC ∩ A
C
, B − C = B
− C
= S = BC ∩ B
C
.
(2.41)

Altogether,

R + S + T = A − B + C − A + B − C = 0, (2.42)

showing that the intersection points R, S, and T are collinear.

It is worth contrasting these proofs with the synthetic proofs of Desargues’s
theorem, as illustrated in Fig. 2.27.

Figure 2.27. Diagram for the synthetic proof of Desargues theorem.

This proof involves embedding the 2-D perspectivity in 3-D. The theorem statement is that two
triangles in perspective from a point P are in perspective from a line p.
 2.9 ** Conics on the Projective Plane 43

Fun and Games with the ILM

Open the ILM: set Chapter = Geometry and click on “Lock Guides.”
The display shows Desargues’s theorem with the three intersection
points R, S, T collinear.
Click and move any of the red triangle vertices along the lines.
The three points R, S, T remain collinear.

Next set Example = Pappus and click on “Lock Guides.”

The display shows Pappus’s theorem.
Again click on and move any of the six points Ai , Bi along the lines.
The three intersection points Ci remain collinear.

Exercises
1. Provide an analytic proof of Pappus’s theorem.
2. Generalize and find the representation of the ideal elements of (R3 , p∞ ) in
terms of ordered quadruples.
3. Generalize the representation of the ideal elements of PN in terms of ordered
(N + 1)-tuples.

2.10 ** Conics on the Projective Plane

Curves can be considered as a collection of points and dually as a collection of lines
whose envelope forms a curve, (see Fig. 2.28.) To distinguish them, they are referred
to as point-curves and line-curves respectively. The duality of the projective plane
enables us to treat them equivalently and leads to some nice results, illustrated here
with conic point- and line-curves.

2.10.1 In the Projective Plane All Conics Are Ellipses

That all conics are ellipses in the projective plane can be observed on the right of
Fig. 2.28, which relates to the position of an ellipse with respect to the ideal line
l∞ . That is, an ordinary (Euclidean) ellipse has no ideal points, and a Euclidean
parabola corresponds to an ellipse tangent to the ideal line and hence having one
ideal point corresponding to the direction in which the parabola “opens toward
44 2. Geometry Background

Figure 2.28. Ellipse (point-curve) and hyperbola (line-curve). Projective conics are ellipses.

Figure 2.29. Ellipse (left), parabola (center), and hyperbola (right) in the projective plane model.

infinity.” A Euclidean hyperbola is a projective ellipse intersecting l∞ at two ideal

points corresponding to the directions of its asymptotes. The three cases in Fig. 2.29
are shown in the projective plane model.

2.10.2 Generalizing Pappus’s Theorem

In 1640, Pascal generalized Pappus’s theorem to conics.
Theorem 2.10.1 (Pascal) If a simple hexagon A1 A2 A3 B1 B2 B3 is inscribed in
a conic, the intersections Ci = Aj Bk · Ak Bj , i = j = k, are collinear.
 2.10 ** Geometries and Invariants 45

Figure 2.30. Pascal’s theorem (a) and its dual, Briachnon’s theorem (b).

Briachnon proved the result depicted in Fig. 2.30 in 1806, which turned out to be
the dual of Pascal’s theorem.

2.11 ** Geometries and Invariants

As already mentioned in connection with (2.10), homogeneous coordinates can be
extended to higher dimensions. Specifically, a point P ∈ PN is represented by an
(N+1)-tuple, the point coordinates

P = (x1 , x2 , . . . , xN , xN+1 ) or (kx1 , kx2 , . . . , kxN , kxN +1 ), k = 0,

where not all components are zero. Dually, a hyperplane π is represented by the
(N + 1)-tuple

π = [X1 , X2 , . . . , XN , XN+1 ], or [kX1 , kX2 , . . . , kXN , kXN +1 ],

where again not all components are zero, these being the hyperplane coordinates
of π. The on relation as before is defined by the inner product of the coordinates,
i.e.,


N +1
P on π ⇐⇒ xi Xi = 0.
i=1
46 2. Geometry Background

A projective transformation in PN is expressed by (2.10) and can be linearized

by homogeneous coordinates. Projective transformations leave the cross-ratio
invariant, and they are the most general transformations employed in projective
geometry.
Various geometries were developed during the nineteenth century in addition
to the Euclidean: affine, projective, spherical and inversive. Felix Klein proposed a
way of classifying geometries that became known as the Erlangen Program, since
he was a professor at the University of Erlangen. A geometry is regarded as a space
together with a group of transformations acting on it. Geometrical properties are
those remaining invariant under the group of transformations. As a reminder, a
group is a set of elements G and a binary operation ◦ with

a ◦ b = c ∈ G ∀a, b ∈ G (closure property),

(a ◦ b) ◦ c = a ◦ (b ◦ c) (associativity)
∃ e ∈ G  a ◦ e = e ◦ a = a, ∀a ∈ G (identity)
∀a ∈ G ∃a −1  a −1 ◦ a = a ◦ a −1 = e (inverse).

Exercises
Show that the set of all linear fractional transformations (i.e., the projective
transformations) with the composition operation form a group.
The group of projective transformations has important special cases that form sub-
groups of transformations, i.e., subsets that are groups. With slight changes in the
notation, consider the case of translations, for which the b’s in the denominator of
(2.10) are zero except for the last one, i.e.,
x
= xA + c,
with row vectors denoted in bold, e.g., x, and A is an N × N matrix. In general, a
mapping
L:D→D
is a linear transformation if it preserves linear combinations, i.e.,
L(ax + by) = aL(x) + bL(y).
A general affin transformation can be written as
T (u) = L(u) + c,
 2.11 ** Geometries and Invariants 47

where c is a constant vector and L is a linear transformation. In this case,

T (ax + by) = aL(x) + bL(y) + c = aT (x) + bT (y) = aL(x) + bL(y) + 2c.

In short, translations are

 not linear transformations.
  Affine transformations leave
convex combinations ai xi , where a1 = 1 invariant and also preserve
incidence (i.e., the on relation).
Let us return to the representation

x
= xA + c

and consider the case where the matrix has the property AAT = λ2 I , where λ is
real and T denotes the transpose of a matrix. This is the similarity group. First take
the case c = 0. The square of the length, |x|, of x is given by x · x = xxT and is
called the norm of x. For the similarity transformation x
= xA, we have

x
· x
= (xA) · (xA) = λ2 x · x.

Effectively, |x| has been amplified by ±λ. For two vectors x, y, under similarity we
obtain

x
· y
= λ2 x · y x · y = |x| · |y| cos(x, y),

where (x, y) is the angle between the two vectors. So

|x
| · |y
| cos(x
, y
) = λ2 |x| · |y| cos(x, y).

Therefore, x⊥y ⇐⇒ x
⊥y
, and in general, cos(x
, y
) = cos(x, y). So under
similarity, the ratio of distances and angles between vectors is preserved, or

|x
| |x|

= .
|y | |y|

Specializing further to orthogonal matrices with |A| = ±1 or |A||AT | = |A|2 = 1,

where |A| denotes the determinant of A, we obtain the important group of orthogonal
transformations.
The Euclidean distance between two vectors x, y is given by d(x, y) =
[(x − y) · (x − y)T ]1/2 :

d(x
, y
) = [(x
− y
) · (x
− y
)T ]1/2 = [(x − y)AAT (x − y)T ]1/2

= [(x − y) · (x − y)T ]1/2 = d(x, y),

48 2. Geometry Background

showing that distance is also invariant under orthogonal transformations. Altogether

then, areas, angles, and orthogonality are also preserved. All these together with
the “pure” translation

x
= x + c

form the rigid motions, which are the invariants of Euclidean geometry.
To recapitulate, as pointed out by Klein, the above geometries are projective
geometries with additional conditions. For example, affine geometry is obtained by
selecting any line in a projective geometry and considering only the transformations
leaving that line invariant. This line plays the role of the ideal line and can be
safely ignored. The reader is referred to [163], [157] for background and history
of N-dimensional geometry, and to [152], [18] for learning more about the various
geometries. Aspects of differential geometry, starting with a short chapter on the
theory of envelopes, are covered later in the study of curves and surfaces.
3

♣ FT-1 The Plane with

Parallel Coordinates

3.1 The Fundamental Duality

A point in the x1 x2 plane is represented in parallel
coordinates (abbreviated -coords) by a line in the
xy-plane (Fig. 3.1). To find out what a line “looks
like” in -coords, a set of collinear points is selected
(Fig. 3.2, top right), and the lines representing them
intersect (top left) at a point! Remarkably, this is true
in general, and for the line

: x2 = mx1 + b, (3.1)

the corresponding point is


d b Figure 3.1. Points on
¯
: , , m = 1, (3.2)
1−m 1−m the plane are represented by
lines.
where d is the distance between the parallel axes.
As will be appreciated later, the interaxis distance is directed, so that interchanging
the axes makes the sign of d negative. The point ¯ is said to represent the line .
It is evident that a point ↔ line duality15 is at work here, which must be properly
considered in the projective P2 rather than the Euclidean R2 plane. In (3.2), as

15
See section on duality in Chapter 2.

A. Inselberg, Parallel Coordinates, DOI 10.1007/978-0-387-68628-8_3,

© Springer Science+Business Media, LLC 2009
50 3. ♣ FT-1 The Plane with Parallel Coordinates

Figure 3.2. In the plane, parallel coordinates induce a point ←→ line duality.

m → 1, ¯ → ∞ in a constant direc-
tion. This is further clarified in Fig. 3.3,
showing that the lines representing points
of a line with m = 1 are parallel with
slope b/d. This is why a line with m = 1
is represented by the ideal point in the
direction with slope b/d. Use of the pro-
jective plane allows the inclusion of lines
with m = 1 to complete the one-to-one
correspondence between lines and points
↔ ¯ as given by (3.1) and (3.2).
From now on, the x1 x2 and xy Figure 3.3. Lines representing points
planes are considered as two copies of the on the line x2 = x1 + b.
 3.1 The Fundamental Duality 51

projective plane P2 . An object S in the x1 x2 plane is represented in the xy plane by

its image, denoted by S̄.

Fun and Games with the ILM

Open the ILM: set Chapter = 2D Point Line duality.

The display shows two sets of axes.
On the right, click to draw a point and see the corresponding line on the left.
Move the point and watch the line.

Next set Subject = Point ↔ Line.

Draw points on the left in various x-positions and see the dual line
on the right.

Expressing (3.2) in homogeneous coordinates with triples within [ . . . ] and

( . . . ) denoting line and point coordinates respectively, we obtain

: [m, −1, b] −→ ¯ : (d, b, 1 − m). (3.3)

It is clear that lines with m = 1 are represented by the ideal points (d, b, 0). The
horizontal position of ¯ depends only on the slope m, as illustrated in Fig. 7.11.
Specifically, for vertical lines : x1 = c, m = ∞ and ¯ is at y = c on the y-axis,
while for horizontal lines, : x2 = a (i.e., m = 0), ¯ is at x = a (on the x-axis).
Lines with negative slope are represented by points between the parallel axes, lines
with 0 < m < 1 by points to the right of the X̄2 -axis, and lines with 1 < m by
points to the left of the X̄1 -axis. This explains why the representation of a point
P is taken as the whole line P̄ , rather than a segment, for the intersection of the
line P̄ with points P ∈ may lie outside the strip between the axes. It is clear that
parallel lines having the same slope are represented by points on the same vertical
line, a property that turns out to be very useful. Therefore, the ideal point Pm∞ is
represented by the vertical line x = 1−m 1 shown in Fig. 3.5 and represents all the
lines with slope m.
For complete generality, we use the line description

: a1 x1 + a2 x2 + a3 = 0, (3.4)
52 3. ♣ FT-1 The Plane with Parallel Coordinates

Figure 3.4. The horizontal position of ¯ depends only on the slope m of . See (3.4) and (3.5) for
the case m = ±∞.

Figure 3.5. Under the duality, parallel lines map into points on the same vertical line. In the
projective plane model, the great semicircles representing the lines share the same diameter, since
the lines have the same ideal point.
 3.1 The Fundamental Duality 53

and for a2 = 0, we have m = − aa21 , and b = − aa23 , providing the correspondence

: [a1 , a2 , a3 ] −→ ¯ : (da2 , −a3 , a1 + a2 ). (3.5)

¯
In turn, this specifies a linear transformation between the triples and :

¯ = A , = A−1 ,
¯

where and ¯ are considered as column vectors. The 3 × 3 matrix is

⎡ ⎤ ⎡ ⎤
0 d 0 −1/d 0 1
A = ⎣ 0 0 −1 ⎦ , A−1 = ⎣ 1/d 0 0 ⎦, (3.6)
1 1 0 0 −1 0

which can be easily computed by taking three simple triples, such as

[1, 0, 0], [0, 1, 0], and [0, 0, 1] for .
Since → ¯ must be considered in the projective plane, we can verify that
the ideal line l∞ = [0, 0, 1] is mapped into ¯∞ = (0, 1, 0), the ideal point in the
vertical direction, this being the “intersection” of all the lines representing the ideal
points, which we have already seen are vertical. It is clear, then, that all lines of
the projective plane P2 are mapped into points of P2 as given by (3.6). Below we
summarize the various cases:
⎧ ⎫
⎪ → ¯
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ [a1 , 0, a2 ], m = ∞ → (0, −a2 , a1 ), on the y-axis ⎪
⎪
⎪
⎪ ⎪
⎪
⎨ ⎬
[0, a2 , a3 ], m = 0 → (d, −a3 /a2 , 1), on the x-axis .
⎪
⎪ ⎪
⎪
⎪
⎪   
a3 m ⎪  ⎪
⎪
⎪ [−a2 , a2 , a3 ], m = 1 → 1, − da a3
, 0 , ideal point slope − da ⎪
⎪
⎪
⎪ 2 2 d ⎪
⎪
⎪
⎩ ⎪
⎭
[0, 0, 1], ∞ → (0, 1, 0), ¯∞ vertical direction

For the other half of the duality, using homogeneous coordinates we look carefully
into the point–line P → P̄ correspondence illustrated in Fig. 3.6. The point P =
(p1 , p2 , p3 ) = (p1 /p3 , p2 /p3 , 1), p3 = 0, is represented in the xy plane by the
line P̄ with slope (p2 − p1 )/dp3 . Hence

P̄ : (dp3 )y = (p2 − p1 )x + d p1 ,

so that

P = (p1 , p2 , p3 ) −→ P̄ = [(p1 − p2 ), dp3 , −dp1 ]. (3.7)

54 3. ♣ FT-1 The Plane with Parallel Coordinates

Figure 3.6. A point P = (p1 , p2 , p3 ) = (p1 /p3 , p2 /p3 , 1), p3 = 0, is represented by a line P̄ .

The ideal point Pm∞ = (p1 , p2 , 0) = (1, p2 /p1 , 0) has direction with slope m =
p2 /p1 . Therefore, P̄m∞ = [(p1 − p2 ), 0, −dp1 ] = [1 − m, 0, −d], which is the
vertical line at x = d/(1 − m), as expected (see also Fig. 7.11). Considering
P and P̄ as column vectors, we can write the above correspondence as a linear
transformation:
P̄ = BP, P = B −1 P̄ ,
with
⎡ ⎤ ⎡ ⎤
−1 1 0 0 0 1/d
B=⎣ 0 0 −d ⎦ , B −1 = ⎣ 1 0 1/d ⎦ . (3.8)
d 0 0 0 −1/d 0
So, in -coords there is a fundamental point ↔ line duality, formalized by (3.5)
and (3.7). Curiously, for d = 1, AT = B −1 . Altogether, then, we have the
A B
transformations line −→ ¯ point and point P −→ P̄ line. We can continue
A B ¯¯ ¯¯ That
with −→ ¯ −→ , but in general, the line is not the same as the line .
is, B(A ) = , and likewise, A(BP) = P (see Exercises 2, 3). By the way, the
matrices A, B are quite sparse (i.e., have many zeros), and since usually d = 1, the
computation to and from -coords is not too costly.
 3.1 The Fundamental Duality 55

Figure 3.7. Motivation for the notation X̄i of the axes; note the position of the x̄i . Two lines
x1 = a, x2 = b on the point P (left) whose images are two points ā, b̄ on the line P̄ (right). The
axis X̄1 represents the ideal point in the vertical direction (hence x̄2 is on it) and X̄2 in the horizontal
(with x̄1 on it).

A few words about the notation

used in Fig. 3.7 are in order. Each of
the axes X̄i , i = 1, 2, represents a
family of parallel lines. Specifically,
X̄i represents the parallel lines xi =
constant or the ideal point in the direc-
tion of xi . Note the representation of
x1 and x2 above.
The image (representation) r̄ of
a curve r can be obtained as the
envelope16 of the lines representing
the points of r. This is distinguished Figure 3.8. Hyperbola (point-curve) →
by referring to the original curves as Ellipse (line-curve).
point-curves and their images as line-
curves as in the example in Fig. 3.8. Line-curves are constructed from the envelope
of the family of curves (for an old but very nice treatment of envelopes, see [17]).
In fact, the point ¯ can also be found in this way to be the envelope of the infinite
family of lines Ā. In Chapter 7 on curves it is seen that the mapping

point-curve ↔ point-curve
16
As explained in the chapter on envelopes.
56 3. ♣ FT-1 The Plane with Parallel Coordinates

is achieved directly by a transformation between the x1 x2 and xy projective planes

[99]. Additional related topics can be found in [97].

Exercises
1. Verify (3.2). Hint: the lines Āi , i = 1, 2, are on the points (0, ai ), (d, mai + b)
of the xy plane and ¯ = Ā1 ∩ Ā2 .
2. Apply → ¯ → ¯¯ → · · · . Does the pattern repeat? That is, do we have
A B

→ · · · → after a finite number of transformations? What is meant here is

the application of the line-to-point transformation expressed by the matrix A,
then the point-to-line transformation expressed by the matrix B, etc.
3. Repeat for P → P̄ → P̄¯ → · · · . Does this pattern repeat?
B A

Fun and Games with the ILM

Open the ILM: set Chapter = 2D Point Line duality,

Subject = Point ↔ Line, Example = Horizontal Slope,
check Lock Horizontal.
On the left, move a point horizontally, while watching the rotating
dual line on the right.

Change to Example = Vertical Slope, check Lock Vertical.

Next set Subject = Point ↔ Line.
On the left, translate a point vertically while watching the horizontal
translation of the dual line.

Change to Example = Parallel Projective, check Lock Vertical.

There are parallel lines in Cartesian and  coordinates
and on the Projective Plane Model (PPM).
On the lower left translate a point vertically while watching the
action in the other windows.
 3.2 Transformations under the Duality 57

3.2 Transformations under the Duality

3.2.1 Rotations and Translations
With reference to Fig. 3.9, a line originally horizontal, shown in position o , is
rotated counterclockwise about one of its points, which for convenience is taken
as the origin O. Under the duality, the corresponding path of ¯ is along the line
Ō, since in any of its rotated positions (say 1 , 2 , 3 etc.), still contains the
point O. To better understand the corresponding motion of the point , ¯ recall that
in the projective plane a line behaves like a “closed curve” (see Fig. 3.5) “joined”
at its ideal point. Hence due to the counterclockwise rotation, the point ¯ moves

Figure 3.9. Duality between rotations and translations.

58 3. ♣ FT-1 The Plane with Parallel Coordinates

in the direction of increasing x (i.e., to the right) until the line’s slope is 1 when
¯ passes through the ideal point and “returns” to the Euclidean plane now, to the
left of the X 1 -axis but still moving to the right. Dually, a translation of point A
in the positive direction along a line (taken for convenience coincident with the
x1 -axis) corresponds to the rotation of the line Ā in the clockwise direction. The
corresponding dualities in higher dimensions, which we will meet in due course,
are especially useful.

Fun and Games with the ILM

Open the ILM: set Chapter = 2D Point Line duality,

Subject = Point ↔ Line, Example = Translation ↔ Rotation,
check Lock Horizontal.
On the left move a point horizontally while watching the rotating
dual line on the right.

Change to Example = Rotation ↔ Translation, check Lock Horizontal.

On the right translate a point horizontally while watching the rotation
on the left.

3.2.2 ** Recognizing Orthogonality

Parallel lines (Fig. 3.5) can be recognized from their dual point images, and so can
orthogonal lines, by means of the following two transformations:
1. Reflectio about the line x = 21 , denoted by R 1 . In the xy plane the reflection
2
of the vertical line P̄m∞ (the set of points representing the lines with slope m)
about the line x = 21 is the image, as shown in Fig. 3.10, of the lines with slope
the reciprocal 1/m. That is,
R 1 (P̄m∞ ) = P̄ ∞
1 .
2 m

 
2. Circle inversion, denoted by C 1 , with respect to the circle centered at 21 , 0
2
and radius 21 . We know that the vertical line P̄m∞ at x = 1/(1 − m) represents
all lines with slope m. The tangent from the point ¯o = P̄m∞ ∩x-axis, Fig. 3.11,
touches the circle at the point ¯1 whose x-coordinate is 1/(1 + m). So the
 3.2 Transformations under the Duality 59

Figure 3.10. Reflection about the line x = 21 .

Figure 3.11. Circle inversion.

∞ , representing all lines with slope −m.

vertical17 line at x = 1/(1 + m) is P̄−m
∞ ∞
We write C 1 (P̄m ) = P̄−m .
2
17
I am indebted to M. Mostrel for this result.
60 3. ♣ FT-1 The Plane with Parallel Coordinates

Figure 3.12. (a) Square, (b) 3-D cube, (c) 5-D hypercube, all with edges of unit length.

The family of lines orthogonal to those with slope m is found with the composition

R 1 C 1 (P̄m∞ ) = P̄ ∞1 .
2 2 −m

∞ from 1 , 0

The x-distances a, b (shown in Figures 3.10 and 3.11) of P̄m∞ and P̄−m 2
are
(1 − m) 1+m
a=− , b=− ,
2(1 + m) 2(1 − m)
 3.3 A Preview 61

respectively. Therefore, ab = 1/4 is the invariance for mutually orthogonal families

of lines. All this shows that orthogonality information is preserved in the -coords
representation.

3.3 A Preview
A glimpse of what lies ahead is seen in Fig. 9.1. In part (a) we see a square with unit
side in Cartesian and dually in parallel coordinates. Note that the edge AB is the point
represented by the intersection of the lines Ā with B̄ representing the corresponding
two vertices. The situation becomes more interesting in part (b), where a cube with
unit side and its image are shown. The collection of all vertices has coordinates that
are all possible ordered triples of 0’s and 1’s. Therefore the cube’s image consists
of two adjacent copies of the square’s image. Note the polygonal line indicated as
the image of the vertex E = (1, 0, 0), for example. It is clear that all vertices are
included in the image. Also, the connection with the square and the symmetry of
the cube are suggested, but not clearly understood, in the “double-square” image.
In due course, we will see that this image contains all the information about the cube,
namely, the edges and faces as well as the interior and exterior. Whereas we do not
know how to show a hypercube in 5-D or higher in Cartesian coordinates, there is no
problem doing so in parallel coordinates, as seen in part (c) of the figure, and with,
as we shall eventually show, all the information present. We will also see that the
images with repetitive patterns represent objects with high symmetry more general
than the hypercube. The reason this works is that our duality has nice generalizations
to higher dimensions, which is our next topic. Though the ensuing development is
carried out in the projective plane, it works out that we can still use all the Euclidean
notions of angle, distance, etc. as long as ideal elements are not involved.

Exercises
1. Draw in Cartesian and parallel coordinates a point P = (p1 , p2 ) and a line
before and after:
(a) the translation of the origin to the point (a1 , a2 ),
(b) the reflections about the x1 and x2 axes,
(c) the rotation of the axes about the origin by an angle θ,
(d) the composition of a translation followed by a rotation.
2. Given a family of lines L having slope m, generalize the transformations
in Section 3.2.2 to construct the family of lines Lθ at an angle θ with the
lines in L.
This page intentionally left blank
4

Multidimensional Lines

4.1 ♣ FT-1 Representing Lines in RN

There are several point ↔ line dualities (such as the very useful Hough transform) in the
plane that do not generalize well for higher dimensions. This is because the natural duality
is point ↔ hyperplane in projective N-space PN with point ↔ line when N = 2. However,
in -coords there is a useful and direct generalization, which is the subject of this chapter.
At first the basic idea for lines in RN , rather than PN , is derived intuitively, paving the way
for the general case, which is treated subsequently (as in [106] and [107]).

4.1.1 Elementary Approach

Adjacent Variables Form
What is meant by “a line in RN ”? In R3 , a line is the intersection of two planes.
Generalizing, for RN , a line is the intersection of N − 1 nonparallel hyperplanes.
Equivalently, it is the set of points (specified by N -tuples) satisfying a set of N − 1
linearly independent linear equations. After some manipulation, with the exception
of some special cases to be treated later, the set of equations can be put in the
convenient form
⎧
⎪ 1,2 : x2 = m2 x1 + b2 ,
⎪
⎪
⎪
⎪ 2,3 : x3 = m3 x2 + b3 ,
⎪
⎪
⎨
: ··· (4.1)
⎪
⎪ i−1,i : xi = mi xi−1 + bi ,
⎪
⎪
⎪
⎪
⎩ ···
⎪
N−1,N : xN = mN xN −1 + bN ,

A. Inselberg, Parallel Coordinates, DOI 10.1007/978-0-387-68628-8_4,

© Springer Science+Business Media, LLC 2009
64 4. Multidimensional Lines

Figure 4.1. A point on a line in 5-D.

each equation containing a pair of adjacently indexed variables. In the xi−1 xi plane,
the relation labeled i−1,i is a line, and by our point ↔ line duality of Chapter 3, it
can be represented by a point ¯i−1,i . To clarify matters, the situation is illustrated
in Fig. 4.1 with an example in R5 . A straightforward extension of the 2-D duality
is that a point A = (a1 , a2 , a3 , a4 , a5 ) is in when the polygonal line Ā is on
each of the points ¯12 , ¯23 , ¯34 , ¯45 representing the four linear relations of (10.5)
for N = 5. Let Ā(i−1),i denote the (i − 1, i) portion (line) of Ā on the points
(ai−1 , 0), (ai , 0). Then Ā12 on ¯12 ⇒ (a1 , a2 ) satisfies the 12 relation, and suc-
cessively (a2 , a3 ), (a3 , a4 ), (a4 , a5 ) satisfy the 23 , 34 , 45 relations respectively.
Hence in -coords we can see a line ⊂ RN , as represented by N − 1 points ¯ with
two indices, and its points A ∈ represented by the Ā on all the . ¯
Unless otherwise stated, from now on, the distance between each pair of adja-
cent axes is taken as one unit, as shown in Fig. 4.2. Since the y-axis is not coincident
with the X̄i−1 -axis, the x-coordinate of the point representing the line i−1,i (see
(3.3) and Fig. 3.2) needs to be translated by (i − 2). That is,

1 bi
¯i−1,i = + (i − 2), ,
(1 − mi ) (1 − mi )
 4.1 ♣ FT-1 Representing Lines in RN 65

Figure 4.2. Spacing between adjacent axes is 1 unit.

and in homogeneous coordinates:

¯i−1,i = ((i − 2)(1 − mi ) + 1, bi , 1 − mi ). (4.2)
There are N − 1 such points for i = 2, . . . , N representing a line . This is the first
instance in which the indexing arises in the representation and plays a key role in the
subsequent development. For now it suffices to observe that without the indexing,
the representation of the line is ambiguous. The exact number of lines that can be
represented by the same points indexed differently and other related questions are
studied later, in Section 4.1.4. Since the display space is at a premium, the indexing
is usually not included in the picture, but it must always be accessible.
In Fig. 4.3 are several polygonal lines representing points on an interval of a
line ⊂ R10 . The intersection of the polygonal lines between the X̄2

and X̄3 -axes

is the point ¯2,3 . All nine points representing that 10-dimensional line can be seen
(or constructed) with their horizontal positions depending on the first coordinate
of (4.2). Do not be misled by the fact that here, all of the ’s¯ except ¯1,2 and ¯6,7
are between their corresponding axes. This is due to the choice of mi ≤ 0 made
for display convenience only. As for 2-D, the x-coordinate of ¯i−1,i is between the
X̄i−1 , X̄i axes when mi < 0, to the left of X̄i−1 when mi > 1, and to the right of
X̄i when 0 < mi < 1.

Base Variable Form

Another common way of describing a line ⊂ RN is in terms of one variable,
sometimes called the base, which after appropriate relabeling may be taken as x1 .
66 4. Multidimensional Lines

Figure 4.3. Line interval in R10 .

The thicker polygonal lines represent its endpoints. The adjacent-variables representation, consisting
of nine properly indexed points, is obtained by the sequential intersections of the polygonal lines.
Note that ¯1,2 is to the right of the X2 -axis, and ¯6,7 is an ideal point. The remaining points are
between the corresponding pairs of axes.

Then
⎧
⎪
⎪ 1,2 : x2 = m12 x1 + b21 ,
⎪
⎪
⎪
⎪ 1,3 : x3 = m13 x1 + b31 ,
⎪
⎪
⎨
: ··· (4.3)
⎪
⎪ 1,i : xi = m1i x1 + bi1 ,
⎪
⎪
⎪
⎪
⎪ ···
⎪
⎩
1,N : xN = m1N x1 + bN 1,

and the N − 1 points representing the line are

¯1,i = (i − 1, bi1 , 1 − m1i ). (4.4)

With the axes placed 1 unit apart, the distance between X̄i and X̄1 (which is
coincident with the y-axis) is i − 1. This brings the point ¯1,i between the axes
X̄1 , X̄i when m1i < 0, to the left of X̄1 when m1i > 1, and to the right of X̄i when
0 < m1i < 1. All the remarks about indexing pertain equally well to this or any
other representation of .
 4.1 ♣ FT-1 Representing Lines in RN 67

What about changing the parametrization? Let us consider an example in which

a line in ⊂ R5 is specified by four indexed points of the form ¯i−1,i and it is
required to find the linear relation between x2 and x5 :
2,5 : x5 = m25 x2 + b25 . (4.5)
There is no problem returning to the system of four linear equations i−1,i ,
i = 2, . . . , 5, specifying the line , from which the points ¯i−1,i were obtained,
doing the necessary algebraic manipulations to obtain (4.5). But this can be done
equivalently by the geometrical construction shown in Fig. 4.4. By assigning any
two values to x1 , the polygonal lines representing two points on are constructed.
Joining the two pairs of x2 and x5 values provides two lines whose intersection
must, due to the line → point mapping in 2-D, be the point ¯2,5 with coordinates

¯ 3 b25
25 = + 1, , (4.6)
1 − m25 1 − m25
providing the values of m25 and b25 . By the way, the 3 is for the distance between
the X̄2 - and X̄5 -axes, and the translation by 1 is for the distance between the X̄2 - and
y-axes. Incidentally, we see that all the information about a line is really contained
in its N − 1 indexed points. For the remainder, all sections marked by ** can be
safely omitted at first reading.

Figure 4.4. Algorithm for constructing a pairwise linear relation.

Here ¯25 is constructed from the four points ¯i−1,i representing the line.
68 4. Multidimensional Lines

Fun and Games with the ILM

set Chapter = Multidimensional Lines, Subject = Multidimensional,

Example = Indexed Points.
The two points (polygonal lines) in 6-D specify a line.
The adjacent variables indexed points for the line are shown.

For some experimentation:

change any of the points’ coordinates by moving it along its axis and
observe the change in two of the indexed points. Repeat with other variables.
Try to make two, three or more of the indexed points coincident.

Next click adjacent variables off and base variable on.

Repeat the above experimentation for the base variable indexed points.
In anticipation of the result in Section 4.1.2,
experiment with the collinearity option by changing the index triples.

♣ FT-1e

Two-Point Form
The representation of a line independent of parametrization can best be constructed
from two of its points. Starting with

Pr : (p1,r , p2,r , . . . , pN,r ), r = 1, 2,

for any other point P ∈ , we have P = λP1 + (1 − λ)P2 , λ ∈ R. When both

points are ideal, they define an ideal line. Otherwise, if one of the points is ideal,
it is a simple matter to replace it by a second ordinary point distinct from the first
to obtain the same line . So without loss of generality, only lines defined by two
ordinary points need to be dealt with.
Two pairs

(pi,r , pj,r ), i, j ∈ 1, 2, . . . , N, r ∈ 1, 2, (4.7)

specify the projection ij of on the xi xj coordinate plane, which is represented

by a point ¯i,j . It is essential to identify it with the proper indices as in Fig. 4.5.
 4.1 ♣ FT-1 Representing Lines in RN 69

Figure 4.5. Construction of ¯i,j .

For the data given in (4.7), we have

ij : j xi − i xj + Di,j = 0, (4.8)

with line coordinates [j , −i , Di,j ] for

⎧
⎪
⎪  = pi,2 − pi,1 ,
⎨ i
j = pj,2 − pj,1 , (4.9)
⎪
⎪
⎩
Di,j = pj,2 i − pi,2 j = pj,1 i − pi,1 j ,

for all i, j ≤ N. When i and j are not both zero, (4.8) actually defines a line
that is represented by the point

¯i,j : ((i − 1)j − (j − 1)i , −Di,j , j − i ), (4.10)

where as usual, the distance between adjacent coordinate axes is 1. The relevant
transformations here are

¯i,j = Ai,j i,j , i,j = A−1 ¯

i,j i,j , (4.11)

where
⎛ ⎞
(i − 1) (j − 1) 0
Aij = ⎝ 0 0 −1 ⎠
1 1 0
70 4. Multidimensional Lines

and
⎛ ⎞
1 0 (1 − j )
(i − j )A−1
ij = ⎝ −1 0 (i − 1) ⎠ .
0 (j − 1) 0

There are still some important considerations missing, which are completed in
Section 4.1.4.

4.1.2 ♣ FT-2 Some Properties of the Indexed Points

The 3-point Collinearity
The representation of a line in terms of N − 1 points with two indices can be
given for any of the various line descriptions by linear equations each involving
a pair of variables. Once a sufficient representation in terms of N − 1 points is
given, any other point ¯i,j can be obtained from the representation by a geometric
construction as in Fig. 4.4. It is important to clarify that by the representation S̄ of an
object S is meant the minimal subset of the xy plane from which the representative
P̄ of any point P ∈ S can be constructed. For a line ⊂ PN we have seen that
N(N −1) indexed points in the xy plane, corresponding to all distinct pairs of the N
variables, can be constructed. However only N −1 of these, corresponding to N − 1
linearly independent equations in pairs of these variables, are needed to represent the
line .
The indexed points representing the linear relation between pairs of vari-
ables for an N-dimensional line have a striking and very useful property. For
i = j = k ∈ [1, 2, . . . , N], the three points ¯i,j , ¯j,k , ¯i,k are always collinear.
This can be seen by considering again two points Pr = (p1,r , . . . , pN,r ), r = 1, 2,
on and their projections on the xi , xj , xk three-space as shown in Fig. 4.6.
The projected portions of the points are the vertices of two triangles, with the
collinearity of ¯i,j , ¯j,k , ¯i,k being a consequence of Desargues’s theorem.18 Here
the two triangles are in perspective with respect to the ideal point in the vertical
direction. This property, as will be seen, is the “backbone” of the construction
for the representation of higher-dimensional p-flats in N-space (that is, planes of
dimension 2 ≤ p ≤ N − 1), and it is referred to as the the 3-point collinearity
property.

18
Two triangles in perspective from a point are in perspective from a line.
 4.1 ♣ FT-1 Representing Lines in RN 71

Figure 4.6. The collinearity for the three points ¯i,j , ¯j,k , ¯i,k . The two indicated triangles (blue
and red) are in perspective with respect to the ideal point in the vertical direction. The y-axis is off
scale.

Fun and Games with the ILM

set Chapter = Multidimensional Lines, Subject = Multidimensional,

Example = Collinearity.
The two points (polygonal lines) in 6-D specifying a line appear.
Three indexed points are collinear.

Change any point’s coordinates by moving it along its axis.

Observe that the collinearity is preserved.
Choose other triples of variable indices and follow the collinearity property.

Lines in a Plane
An important special case is that for any line in R3 , the three points
¯1,2 , ¯1,3 , ¯2,3 are always collinear. In turn, this provides an important corollary.
72 4. Multidimensional Lines

Figure 4.7. Three lines ,

,

in a plane π ⊂ R3 .
In -coords, the three-point collinearity property applied to the three points ¯12 , ¯13 , ¯23 determines
a line L̄, and similarly,
determines the line L̄
. Denoting the intersection point by π̄123 = L̄ ∩ L̄
,
the line L̄

, determined by any other line

, is on the point π̄123 . The picture hints that the points

L, L
, L

are on a line in the plane π.

Consider the plane

π : c1 x1 + c2 x2 + c3 x3 = c0 ,

and choose two lines ,

⊂ π, denoting by L̄, L̄
the lines determined by
¯1,2 , ¯2,3 , ¯1,3 and ¯
1,2 , ¯
2,3 , ¯
1,3 respectively; see Fig. 4.7. The coordinates
of π̄123 = L̄ ∩ L̄
turn out to be

π̄123 = (c2 + 2c3 , co , c1 + c2 + c3 ). (4.12)

They depend on the coefficients of π and not on the choice of and

. Hence, for
any other line
⊂ π, as in Fig. 4.7, the corresponding L¯

must also be on π̄123 .

This is a preview of the next chapter, where it is proved that two such points each
with three indices represent a plane π in R3 .
 4.1 ♣ FT-1 Representing Lines in RN 73

Exercises
1. Consider a line ⊂ RN whose description, by a set of linear equations, contains
xi = ci , a constant, for some i.
(a) How would such a line be represented?
(b) Can such a line be represented by N − 1 indexed points?
2. Prove the property illustrated in Fig. 4.7
3. Given a line and the N − 1 points ¯i−1,i , i = 2, . . . , N, provide an algorithm
for determining the point ¯ij , for any distinct pair i, j , and state its complexity;
see Fig. 4.4.
4. For lines such that some of the m’s are close to 1, their representation becomes
problematic. Find a transformation that can always convert points outside the
axes to equivalent points between the axes. Write an algorithm to perform this
transformation so that all points representing a line are between the axes. What
is the algorithm’s complexity?
♣ FT-2e

4.1.3 Representation Mapping I

The 3-point collinearity property and its higher-dimensional relatives are utilized
in the representation of p-flats, planes of dimension 0 ≤ p ≤ N − 1, in RN by
means of indexed points. The 0 lower bound pertains to points P ∈ PN .
The method’s goal is the construction of a mapping J for the unique repre-
sentation of objects (i.e., subsets) of PN by subsets of P2 . The object B ⊂ PN is
represented by its image J (B) = B̄. Starting with the representation of a point
by a polygonal line P ∈ PN , it is clear that J is not a point-to-point mapping.
Formally,
N 2
J : 2P → 2P × 2[1,2,... ,N] , (4.13)

where 2A = [B ⊆ A] denotes the power set of a set A, the set of all subsets of A,
and RN is embedded in the projective N-space PN . The 2[1,2,... ,N] in the product
pertains to the indexing by subsets of [1, 2, . . . , N].
The answer to “how can RN be mapped into the plane R2 without loss of
information?” is found by examining the cardinality19 of the domain and range
19
Only elementary considerations are needed. For example, the sets of rational numbers and real num-
bers R have cardinalities ℵ0 and ℵ1 respectively. The comparison ℵ0 < ℵ1 signifies the nonexistence
of a one-to-one and onto mapping from the rationals to the reals. For more on the fascinating topic of
infinite sets and cardinal numbers see any good text on set theory.
74 4. Multidimensional Lines

Figure 4.8. A point P i = (p1i , p2i , . . . , pji , . . . , pNi ) represented by N points pji with one index.

of J . Let |A| denote the cardinality of a set A and recall that by taking the
finite product of a set with itself the cardinality remains invariant. For our specific
case,

|2R | = |2R | = ℵ2 .
2 N
|R| = |R2 | = |RN | = ℵ1 ⇒

With the domain and range having the same cardinality, the construction of a one-
to-one mapping J is possible; what is left is to actually do it. “The devil is in the
details,” as sages have told. The image J (B) of a set B ⊂ PN is given in terms
of the xy Cartesian coordinates superimposed on the parallel axes with J (B) =
J (C) ⇔ B = C. At this stage we can get a glimpse of the mapping’s recursive
construction algorithm, the recursion being on the dimensionality of the object
being represented.

1. The nonrecursive step is the alternate representation of a point

P i = (p1i , p2i , . . . , pji , . . . , pN

i
).

Rather than using a polygonal line for P̄ i , for consistency with what follows,
the representation is also given in terms of indexed points in the xy plane.
Specifically, P i ∈ PN is mapped into (or represented by) N points with
one index, as shown in Fig. 4.8. Of course, this is completely equivalent to
its polygonal line representation, where the indexing sequence provides the
straight-line connectivity when needed.
 4.1 ♣ FT-1 Representing Lines in RN 75

If Nr denotes the number of points and nr the number of indices appearing in

the representation, for a (point) 0-fla ,

Nr + nr = N + 1.

2. The fi st recursive step is the construction of a 1-fla (line) ⊂ PN consisting

of N − 1 points with two indices. The steps are:
(a) for each of two distinct 0-flat P1 , P2 ⊂ the N representating points
are connected via the indexing sequence to provide two polygonal lines,
(b) intersecting adjacent portions of the polygonal lines yields the N − 1
points ¯i−1,i , i = 2, . . . , N.
Alternative but equivalent representations correspond to the intersection of
N − 1 pairs of lines joining the appropriate alternate vertices of the polygonal
lines.
Checking the previous sum for 1-flat yields

Nr + nr = (N − 1) + 2 = N + 1.

3. The construction procedure turns out to be generic for the representation of

p-flats of all dimensions.
The emerging relation for Nr + nr holds with the caveat that the subset of points
in the representation is minimal. It is revisited as new representations are obtained.

4.1.4 ** The General Case

Indexed Points for Arbitrary Parametrization
Though the basics have been clarified, there remain representational issues due to
the orientation of some lines with respect to the principal axes. Specifically, a line
may be perpendicular to a principal 2-plane xi xj . Then both xi and xj are constant,
and the projection of i,j on the xi xj plane is a point (pi , pj ) represented in parallel
coordinates by a line on the two points (i − 1, pi ), (j − 1, pj ), which plays the
role of ¯i,j . These contingencies, though conceptually straightforward, require that
the foundations be revisited in a thorough way.
In RN there are N(N − 1) different linear equations connecting the variables
pairwise, and therefore, as N increases, the number of sets of equations available
for describing a line increases rapidly. A precise count as a function of N is given
shortly. The reasons for choosing one description rather than another are irrelevant
here. Rather, the class of such descriptions is of interest, for which purpose a graph-
theoretic approach is appropriate. Let the N variables be represented by N vertices
76 4. Multidimensional Lines

of a graph. An edge connecting a pair of vertices indicates that a linear relation is

given between the corresponding pair of variables. A set of equations describing
a line corresponds to a subgraph with N vertices and N − 1 edges, which must
be a spanning tree of the complete graph of N vertices. For if the subgraph is not
connected, the corresponding system of equations contains at least two independent
sets of variables such that their values are independent of values taken on by the
variables of other sets. Also, if there is a closed loop in the graph, then at least one
vertex is disconnected, there being only N − 1 edges. Once a particular variable
is selected and designated as the root of the tree, the number of different ways in
which N vertices and N − 1 edges can be assembled into a spanning tree is N (N −1)
(Cayley’s formula) [150]. In [54] there is a nice proof20 that the number of spanning
trees with N distinct vertices is N (N−2) . From this result, Cayley’s formula can be
immediately obtained by choosing any one of the N vertices as the root. From our
viewpoint this is the number of distinct N − 1 linearly independent equations, with
the chosen variable as the root, that can be selected out of the possible N(N − 1)
linear equations between pairs of variables. Assigning a value to the start variable,
corresponding to the root, initiates the computation and suffices to determine the
values of all the remaining variables for the specification of a point on the line.
Since variables that are constant are not involved in linear relations with another
variable, there is no edge connecting their vertex to another vertex. That is to say,
such vertices are isolated. If there are N − M such variables, then the remaining
M vertices together with M − 1 edges again form a spanning tree as before, rooted
once a start variable is chosen. Of course, in this case there are M (M−1) ways for
selecting the M − 1 independent linear equations, between pairs of variables, with
the designated start variable. Let S be the set of indices for which the variables are
not constant, where

S = (i1 , i2 , . . . , iM )

and 1 ≤ i1 < i2 < · · · < iM ≤ N , and let S̃ be the complementary set of indices.
For the example shown in Fig. 4.9, S = (1, 3, 4, 5, 8, 9) and S̃ = (2, 6, 7). Note
that two lines 1 and 2 are parallel if and only if
• S1 ≡ S2 = S and
• for all ir = is in S1 the ¯ir ,is have the same abscissa for both lines.
In other words, the ¯ir ,is for each pair ir is ∈ S lie on a vertical line as in the 2-D
case discussed in Chapter 3. Clearly, consideration of S˜i , i = 1, 2, is not needed,
since xi and xj constant implies that is orthogonal to the xi xj plane.
20
Also cited in [54] is [137], which has a collection of different proofs for the same result.
 4.1 ♣ FT-1 Representing Lines in RN 77

Figure 4.9. Tree with 5 as the root corresponds to x5 being the start variable. The isolated vertices
have indices S̃ = (2, 6, 7).

We have shown that any line ⊂ RN is given, after an appropriate relabeling

of the variables, by the N − 1 equations of the form
j j
i,j : xj = mi xi + bi , i < j ∈ S = (i1 , i2 , . . . , iM ), (4.14)

o,k : xk = bk , k ∈ S̃ = (iM+1 , . . . , iN ). (4.15)

The “dummy” index o is introduced to preserve uniformity in indexing each relation

by two indices rather than sometimes by two and at other times (i.e., the constant
variables) by one. From this and the considerations provided in Section 4.1.1, the
following theorem is immediate:
Theorem 4.1.1 (Representation of a line ⊂ RN for arbitrary parametriza-
tion) Any line ⊂ RN can be represented by N − 1 points each of which has two
indices, where M − 1 points are given by
⎧
⎨ ¯i,j = (j − i, bj 1 − mj ) for i, j ∈ S = (i1 , i2 , . . . , iM ), i = j, and
i i
⎩ ¯ = (k − 1, b , 1) for k ∈ S̃ = (i
o,k k M+1 , . . . , iN ).
(4.16)

As in the previous cases, it is clear that the horizontal position of the point ¯i,j is
determined by the slope m . Specifically, for i < j , ¯i,j is to the left of the X̄i -axis
j
i
j j
for mi > 1, between the X̄i - and X̄j -axes for mi < 0, and to the right of the
X̄j -axis for 0 < mi < 1. The construction of a point on given ¯ is illustrated
j
78 4. Multidimensional Lines

in Section 4.1.5. Once a value is assigned to any variable with index in S via the
¯i,j ’s, all other variables with indices in S are known. Since variables in S̃ are also
known, all the N coordinates of a point on are known.

Two-Point Form
The general two-point form representation, obtained in a completely analogous way,
is given by the following theorem:
Theorem 4.1.2 (Representation of a line ⊂ RN : Two-Point Form) Any line
in RN can be represented by N − 1 points, where M − 1 points (some possibly
ideal), one for each edge connecting variable i and variable j , are given by
¯i,j = (xi,j , yi,j ),
where for i = j the point coordinates are
(j − i)i − (i − 1)(i − j ) pi,1 pj,2 − pi,2 pj,1
xi,j = , yi,j = , (4.17)
i − j j − i
and for i = j , the ideal point (direction) is
xi,j = j − i, yi,j = pj,1 − pi,1 . (4.18)
In this case, pj,2 − pi,2 = pj,1 − pi,1 , and there N − M points given by
(k − 1, xk ), (4.19)
one for each k ∈ S̃. By the definition of ¯i,j , the three points (i − 1, xi ), ¯i,j ,
(j − 1, xj ) lie on a straight line, so that as a consequence of (4.17),
(i − 1 − xi,j )xj + (j − i)yi,j − (j − 1 − xi,j )xi = 0, (4.20)
and as a consequence of (4.18),
xi,j xj − xi,j xi = (j − i)yi,j . (4.21)
Once a value is assigned to any variable with index in S, then via the ¯i,j ’s all
other variables with indices in S are found. But variables in S̃ being constant are
also known, and so are all the N coordinates of a point. Given xi let λ satisfy the
equation
xi = λpi,1 + (1 − λ)pi,2 , (4.22)
which is always possible, since by hypothesis, pi,1 = pi,2 . It follows directly from
(4.17), (4.18), (4.20), and (4.21) that
xj = λpj,1 + (1 − λ)pj,2 . (4.23)
That is, any point so defined satisfies the definition and belongs to the straight line.
 4.1 ♣ FT-1 Representing Lines in RN 79

4.1.5 ** Construction Algorithms

And now, after all this preparation, we provide “pencil and paper” construction
algorithms for viewing the objects that we have been discussing. This entails the
representation of the line (in terms of indexed points), as well as the construction
of points on the line. Together with the rooted tree representation for the desired
parametrization (i.e., the i,j ) and the isolated vertices corresponding to the vari-
ables with constant values, an array D with N − 1 rows and five columns is used as
data structure. Essentially, the order of rows is unimportant, although programming
considerations may dictate otherwise. There are two distinct types of rows in the
structure, i.e.,
i j xi,j yi,j hi,j
0 k 0 bk 1
The first type provides the indices of variables connected by an edge in the first two
columns. The next three columns are the homogeneous coordinates (xi,j , yi,j , hi,j )
of the point ¯i,j , hi,j = 1 for regular (i.e., finite) and hi,j = 0 when it is ideal, in
which case the coordinates (xi,j , yi,j ) define the slope of the direction. The second
row type pertains to xk = bk a constant, with xi,j = 0, yi,j = bk , and hi,j = 1.
In other words, the variable’s index is in column two, and the variable’s value is in
column four. Columns three and five have only 0’s and 1’s respectively.
In the complete data structure there are M − 1 rows of the first type and N − M
rows of the second type. Of course, for fixed variables there is some waste in the
structure. But since the occurrence of such variables will certainly be a rare event,
it is not useful to maintain this part of the data in a separate structure. For the time
being it is assumed that no i is zero, that is, all variables are unconstrained. It is
apparent that there is sufficient data in the data structure to compute a point on the
line. To guide the computation of such a point, a framework, called adjacency data,
is provided. It consists of a set of arrays denoted by Ai , one for each i ∈ S. The
array Ai has two rows and as many columns as there are edges connecting the ith
variable to other variables having the form

i1 j2 · · · ij
Ai = .
r1 r2 · · · rj
The first column of the array signifies that there is an edge connecting variables
with indices i and i1 and that ¯i,i1 is found in columns 3, 4, 5 of row r1 of the
data structure, and likewise for the remaining columns of the array. Note that
reordering rows of the data structure requires a like reordering of the second rows
of the adjacency arrays. The algorithm constructing a point on a line can now be
stated.
80 4. Multidimensional Lines

Input: the adjacency data, the data structure, and an index k for which xk , the
start variable, is to be assigned a value vk .
Output: a list of N values, the coordinates of a point on a line.
1. for i = 1 to N mark all x’s new
2. put k on queue (Q), for vk on xk ← old
3. while Q not empty do
4. for adjacency array Ai for first i in Q
(a) while adjacency array not empty do
(b) for xj not old, j first in row 1
i. find the value of xj
ii. append j to Q
iii. xj ← old
iv. delete first column of A
(c) delete first index in Q
5. stop
The treatment of the queue Q is such that each variable is dealt with at most
once. The graph is connected, and via the adjacency data, each variable enters
the queue. Thus each variable is dealt with exactly once. Given the value of one
variable, there are N − 1 calculations required for the remaining variables. Now
each edge (i, j ) appears twice in the adjacency data, once in Ai and once in Aj .
Since there are N − 1 edges, this means that there are 2(N − 1) edge occurrences
in the algorithm. But once a variable is calculated, it is not calculated again. Hence
there are N − 1 calculations of variables with N − 1 bypasses, and the number of
calculations is O(N). The extension to the general case is obvious. It is also worth
noting that for the case (4.14), the adjacency set is implicit in the data structure, as
is of course the special cases of (4.1) and (4.3). The various aspects and nuances of
the above are illustrated and clarified with the following example.

An Example
To clarify matters, consider a line ⊂ R9 specified by two of its points whose
coordinates appear in the first three columns of Table 4.1. As already discussed, a
number of descriptions of the line in terms of linear equations involving pairs of
variables is possible. The detailed construction of the chosen set of ¯i,j is shown in
Fig. 4.10. First the polygonal lines representing the two points P1 and P2 are drawn.
Then the ¯i,j , corresponding to the choice of x5 for start variable, are obtained by
 4.1 ♣ FT-1 Representing Lines in RN 81

Table 4.1. In the left three columns are the coordinates of two points specifying
a line from which a set of eight ¯i,j is constructed. The data structure used in
the point construction algorithm, given in the next five columns, provides the ¯
and their locations. Note the values hi,j = 1 indicating finite and hi,j = 0 ideal
points. The points designating xk = constant (i.e., k ∈ S̃) have i = xi,j = 0.

Coords Point 1 Point 2 i j xi,j yi,j hi,j

x1 −4.0 2.0 1 5 2.4 −0.7 1

x2 2.0 2.0 3 5 3.2 −0.3 1
x3 −3.0 3.0 4 1 7.5 2.0 1
x4 −1.6 2.0 8 5 5.5 0.25 1
x5 1.5 −2.5 9 3 6.0 −1.0 0
x6 −2.0 −2.0 0 2 1 2.0 1
x7 −0.5 −0.5 0 6 5 −2.0 1
x8 −1.0 3.0 0 7 6 −0.5 1
x9 −4.0 2.0

Figure 4.10. Construction of the ¯i,j for the rooted tree in Fig. 4.9 and data in Table 4.1.

the intersection of pairs of thin dashed lines seen in the figure. For example, in order
to obtain ¯1,5 , the x1 and x5 coordinates of each point are joined. The intersection of
82 4. Multidimensional Lines

Figure 4.11. The ¯i,j and the construction of a new point P ∈ for x5 = −0.75.

these two lines is ¯1,5 . The corresponding rooted tree for the start variable x5 is the
one already shown in Fig. 4.9. Altogether, for this tree the points required are ¯1,5 ,
¯3,5 , ¯1,4 , ¯5,8 , and ¯3,9 . The last is an ideal point, since x9 = x3 = −1; hence
the slope is 1, and it is indicated in the figure as a direction. The corresponding
data structure appears in the rightmost five columns of the table. Here the number
of unconstrained variables is M = 5, and there are N − M − 1 = 3 constant
variables.
The ¯i,j together with the points ¯0,j , on the X̄2 -, X̄6 -, and X̄7 -axes, for the
constant xj are also shown in Fig. 4.11 together with the construction of a point
P ∈ for a given value of x5 . Recall that x5 , the start variable, corresponds to the
root of the tree. The construction can be easily modified for another choice of start
variable, possibly M = 5.

Identifying the ¯i,j

There are certain practical problems that must be dealt with. One difficulty arises
in labeling these indexed points. The following example illustrates the nature of the
problem. Given the two points

P1 = (a, a + b, . . . , a + (N − 1)b),
P2 = (c, c + d, . . . , c + (N − 1)d),
 4.1 ♣ FT-1 Representing Lines in RN 83

where b = d for any pair i, j , we have


a − c ad − bc
¯i,j = , . (4.24)
d −b d −b

That is, all the ¯i,j may be congruent (a situation that is revisited in Chapter 5).
Even without such unusual contigencies, the labeling of the points may cause
overcrowding in the figure. This suggests that a separate tabulation of the
¯i,j may be preferable to attempting to label these points directly on the
diagram.

Fun and Games with the ILM

set Chapter = Multidimensional Lines, Subject = Multidimensional,

Example = Indexed points.
The two points (polygonal lines) in 6-D specifying a line appear.
Make some of the coordinates of the two points equal.
Observe the resulting indexed points.

Again change the coordinates so that each polygonal line

becomes a straight line.
Observe the coinciding index points.

4.1.6 Convenient Display of Multidimensional Lines

Lines with Positive Slopes

Often, the ¯ are out of the scope of the picture. For such cases, two line segments
(portions of the polygonal lines representing two points on the line) may be shown
whose intersection (when extended) defines the . ¯ The question arises whether it
is possible to recast the problem of line representation in such a way that “distant”
points are eliminated. A scheme is proposed that not only eliminates ideal points
from the representation but also has a number of other advantages.
Let us transform the variables by

xi
= ei xi , (4.25)
84 4. Multidimensional Lines

Figure 4.12. The ¯

i,j for the example with the points P̄1
and P̄2
are drawn first.
Then the two points ¯
1,4 , ¯
3,9 , marked by rectangles, are constructed as indicated. All other points
are unchanged, though a prime (
) is added for notational consistency. Note the absence of ideal
points.

where ei = +1 for index i ∈ S̃. For the variables in S suppose that there is an edge
on the tree joining the vertices xi and xj ; the vertex for xj is either the root or lies
between the vertex for xi and the root. In this case, the linear equation is written as
j
xi = mi xj + bj .

The remaining ei are developed recursively. Starting at the root index i, set ei = +1
and visit the remaining nodes in order, setting e1 = +1 or −1 to ensure that the

j
slope m i of the line

j
xi = m i xj + bj
(4.26)

j j
j j
is negative, where m i = ei ej mi and b i = ei bi . Since the slope is negative, it
follows immediately that the x-coordinate of ¯
i,j is between i − 1 and j − 1, and
j
that the y-coordinate is less in absolute value than the absolute value of bi . With
everything nicely bounded, all the ¯i,j ’s can be placed within the frame of a picture.
Calculation of a point on the line proceeds as in the preceding algorithm for the
 4.1 ♣ FT-1 Representing Lines in RN 85

Figure 4.13. The polygonal line for the point P now denoted by P
in the new system X̄i
.

X̄-system, followed by appropriate changes of sign and reflection in the X̄i -axis of
those coordinates for which ei is negative. Thus the algorithm is still of order O(N ).
Likewise, construction proceeds with the X̄
-system, followed by reflection in the
X̄i -axis of those coordinates for which ei is negative. However, while adjacency
data depend only on the underlying graph structure, the numbers ei are determined
by the signs of the slopes, and hence must be recalculated for each case. Evidently,
this does not alter the fact that the algorithm is of order O(N ). For the above example
it turns out that the numbers ei are successively 1, 1, 1, −1, 1, 1, 1, 1, −1. The data
structure is given in Table 4.2, where now no fifth column is needed. It is really a
simple matter to notice by inspecting Fig. 4.10 that only the points ¯1,4 and ¯3,9 do
not lie between their corresponding two axes. Hence only for the indices 4 and 9 is
ei = −1, as indicated.
Whereas the transformation (4.25) xi → xi
works well for a single line, its
application to several lines simultaneously can become problematic. Allowing dif-
ferent pairwise linear relations for each line greatly increases the prospect that their
index-point representation can be brought within the framework of the picture (see
exercises below). An elegant algorithm for the convenient display of several lines
was obtained by S. Cohen-Ganor for his course project and is presented in the last
chapter, entitled Recent Results.
86 4. Multidimensional Lines

Table 4.2. The ¯

i,j ’s where the location of xi,j is
now between the corresponding axes.
The ¯
are constructed from the transformed points
P̄i
, i = 1, 2, and shown in Fig. 4.12. The previously
constructed point P for x5 = 0.75 is shown in the
X̄i -coordinate system in Fig. 4.13.

i j xi,j yi,j

1 5 2.4 −0.7
3 5 3.2 −0.3
4 1 1.875 −0.5
8 5 5.5 0.25
9 3 5.0 0.5
0 2 2.0 0
0 6 −2.0 0
0 7 −0.5 0

Exercises
1. For the example of Section 4.1.5, choose a start variable other than x5 and
repeat the constructions. Namely,

(a) exhibit the rooted tree,

(b) construct and show the corresponding ,¯
(c)

construct and show the point P ,

(d) construct and show the ¯
,
(e) construct and show the point P in the xi
system.
2. This is an example of the tree enumeration for N = 4. Show that,

(a) only four distinct kinds of trees are possible, see Fig. 4.14, and
(b) the numbers of distinct rooted trees (allowing permutation of vertex
labels) are those shown below each tree.
How many distinct (not rooted) trees are there?
3. The questions below pertain to the line representational problem of Sec-
tion 4.1.6 and the transformation (4.25).

(a) For N = 3 let and

be two lines such m21 , m32 < 0 and
m
2
3
1 , m2 > 0. Is it possible via the transformation of (4.25) to bring
 4.1 ♣ FT-1 Representing Lines in RN 87

Figure 4.14. The four kinds of distinct rooted trees with labeled vertices (i.e., 1, 2, 3, 4).
The number of trees allowing permutations of the vertex labels is listed below each tree type.

all the four ¯

between the corresponding axes? If not, give a counter-
example.
(b) For N = 3 and three lines ,
,

state and prove necessary and sufficient

conditions for the existence of a transformation xi → xi
such that the
¯i,j , ¯
r,s , ¯

u,v are between the corresponding axes. Provide a specific

example.
(c) Answer the previous question for specific numbers N ≥ 4 and the number
of lines L ≥ 4.
(d) Generalize the result of the previous question for arbitrary N and L; this
is an open question, and likely to be hard.

4.1.7 ♣ FT-3 Rotations and Translations

The projective transformation dualities between parallel and Cartesian coordinates
extend [23] to RN . Recall the duality between the rotation of a line about one
of its points O corresponding to the translation of the point ¯ on the line Ō in
R2 . Figure 4.15 shows a line ⊂ R5 rotated about one of its points P to a
new position
. This corresponds to the simultaneous linear translations of the
¯ to new positions ¯
on P̄ . In fact, the pencil of all lines on P is represented
by the collection of all quadruples ¯i,i+1 , i = 1, 2, 3, 4, with ¯i,i+1 ∈ P̄i,i ,
where P̄i,i+1 denotes the straight line of P̄ intersecting the X̄i and X̄i+1 axes.
Clearly both ¯i,i+1 and ¯
i,i+1 are on P̄i,i+1 , and this leads nicely to the next
topic.
88 4. Multidimensional Lines

Figure 4.15. Two intersecting lines in 5-D.

Another interpretation of this picture is that the line rotates about one of its points shown here by
the polygonal line P̄ . This corresponds to a linear translation of the ¯ to new positions ¯
on P̄ .

4.2 Distance and Proximity Properties

4.2.1 Intersecting Lines
It is clear from Fig. 4.15 that P = ∩
, suggesting an easy intersection algorithm
for the adjacent-variables description of lines. Given two lines ,
we test for
intersection by joining the ¯i,i+1 and ¯
i,i+1 with a line denoted by P̄i,i+1 . Then
and
intersect if and only if the line P̄i,i+1 intersect the xi+1 -axis at the same
point as the line P̄i+1,i+2 ∀i. The polygonal line so constructed represents the point
where the two lines intersect. Formally, for two lines and
, described in terms
of adjacent variables,

∃P = ∩
⇔ P̄i,i+1 ∩ P̄i+1,i+2 = (i, pi+1 ) ∀i = 1, . . . , N − 1, (4.27)

where xi+1 (P ) = pi+1 the ith coordinate.

 4.2 Distance and Proximity Properties 89

Figure 4.16. Intersection of two lines in 4-D for the base-variable line description.
This shows the collision, space (x1 , x2 , x3 ), and time T coordinates of two particles moving with
constant velocity whose trajectories ,
are described by the ¯T i , ¯
T i .

Actually, P̄1,i = (p1 , pi ) is the projection on the x1 xi plane of the intersection

point P = ∩
for the base-variable description; Fig. 4.16. The intersection
conditions for the base-variable description of the lines as given by
!
, 1i : xi = mi x1 + bi ,
(4.28)

,
1i : xi
= m
i x1 + bi
,
are
b
− bi bi
∃P = ∩
⇔ αi = −
i =− = p1 , ∀i = 2, . . . , N, (4.29)
mi − mi mi

where x1 (P ) = p1 . Obtaining an intersection algorithm with O(N ) time complex-

ity and generalizations to other line parametrizations are straightforward. There
are special cases not covered by these conditions, whose treatment is simply not
warranted at this stage (see [107] for a complete treatment and [97] for an earlier
exposition).

4.2.2 Nonintersections
It is convenient to illustrate the situation in 4-D with the base-variable description
of a line
90 4. Multidimensional Lines

xi = vi T + so,i , i = 1, 2, 3. (4.30)

For an interesting interpretation, let T denote elapsed time and (x1 , x2 , x3 ) the space
coordinates of a particle moving with constant velocity V = (v1 , v2 , v3 ) and initial
position So = (so,1 , so,2 , so,3 ). The complete trajectory information of the particle
is provided by (4.30) and equivalently its three-point representation. The two sets
of triple points ¯T i and ¯
T i describe the trajectories of two moving particles, and
the construction in Fig. 4.16 shows an instance in which two particles collide; they
go through the same point in space at the same time; a time–space intersection.
Perhaps some of the power of the -coords representation can be appreciated from
this simple example.
Clearly, intersections in space (3-D and certainly higher) are very rare, so it
behooves us to study nonintersections. The intersection algorithm is also useful
to verify nonintersections as in illustrated in the sequence of Figures 4.17, 4.18,
and 4.19 with the representing points of two lines in 4-D. Pairs of ¯T i and ¯
T i are
connected with the three lines so formed not intersecting the T -axis at the same

Figure 4.17. Nonintersection between two lines in 4-D.

Note the maximum gap on the T̄ -axis formed by the lines joining the ’s ¯ with the same subscript.
The polygonal lines represent the points P , P
, one on each trajectory, where the minimum distance
of D = 20.5 occurs at time T = 4; Theorem 4.2.1.
 4.2 Distance and Proximity Properties 91

Figure 4.18. Another nonintersection between two lines in 4-D.

Here the minimum distance between the lines is 10, and this occurs at time = 1.6. Note the diminishing
maximum gap on the T̄ -axis compared to that in Fig. 4.17. The polygonal lines represent the points
where the minimum distance occurs.

value. The minimum distance D between the lines is recorded. Denote the line on
the points ¯T i , ¯
T i by C̄i , its intersection with the T̄ -axis by

αi = T̄ ∩ C̄i , (4.31)

¯ ¯
| → 0 as D → 0,
¯ ¯
= [max αi , min αi ]. It appears that |I ,
and the interval I ,
which is reasonable, for we already know that |I , ¯ ¯
| = 0 when the two lines
intersect (D = 0). ♣ FT-3e

4.2.3 Minimum Distance between Two Lines in RN

Now we consider the more general problem of finding and, if possible, visualizing
the minimum distance between two lines and the points on each line where the
minimum occurs. In many problems what is required is to minimize the distance
when one or more of the variables are constrained to the same value for both lines.
An example is in air traffic control and motion planning in general, where of interest
92 4. Multidimensional Lines

Figure 4.19. Near intersection between two lines in 4-D.

is knowing the time, and position, when two aircraft are the closest. Therefore, it
is appropriate for such applications to constrain the time and find the minimum
distance and position under that constraint. In that case, the minimum distance is
zero when there is a time–space collision. For if time is not constrained, there will
be two points Pm = (tm , x1 , . . . , xN−1 ) ∈ , Pm
= (tm
, x
, . . . , x

1 N −1 ) ∈
where the minimum distance occurs with tm = tm
that would not be the closest
distance between the two moving objects at the same time.

Constrained L1 Metric Minimum Distance

The following result suggests that in a way, the “natural” metric for -coords is
the L1 (or “Manhattan”) metric, where the various components can actually be
seen; see Fig. 4.20. The L1 distance between two points P = (x1 , x2 , . . . , xN ) and
P
= (x1
, x2
, . . . , xN

) is given by


N
L1 (P , P
) = |xi − xi
|.
i=1

Now if P ∈ and P
∈
for the lines given by (4.28) and the base variable x1
is constrained to be the same for both points, then the distance is given by
 4.2 Distance and Proximity Properties 93

Figure 4.20. L1 distance between the points P = (x1 , . . . , xi , . . . , xN ), P

= (x1
, . . . , xi
,

).
. . . , xN

  

N 
N  b  N
|xi
|mi | x1 + =
i
L1 (x1 ) = − xi | =  |mi ||x1 − αi |,
i=2 i=2
m i i=2
(4.32)
where the intercepts αi are defined in (4.29).
Theorem 4.2.1. The unique minimum value of L1 (x1 ) is attained at x1 = αi for at
least one i = 2, . . . , N.

Proof. Step 1. On the interval αk ≤ x1 ≤ αk+1 ,


k 
N
L1 (x1 ) = |mi |(x1 − αi ) + |mi |(αi − x1 )
i=2 k+1
 
k 
N
= |mi | − |mi | (x1 − αi ),
i=2 i=k+1
94 4. Multidimensional Lines

Figure 4.21. Monotone increasing portions of L1 (x1 ).

is a linear function of x1 with slope


k 
N
Tk = |mi | − |mi |.
i=2 i=k+1
It attains its minimum at one of its endpoints unless Tk = 0, in which case L1 (x1 )
is constant over the interval [αk , αk+1 ].
Step 2. Reorder the index i if necessary so that αi ≤ αi+1 ∀i. First we consider
the case αi < αi+1 ∀i and no mi = 0. It is clear that for x1 ≤ α1 , L1 (x1 )
increases monotonically as x1 decreases, since the |x1 − αi | increase ∀i. Similarly,
for αN ≤ x1 , again L1 (x1 ) increases monotonically as x1 increases, since the
|x1 − αi | increase ∀i. So if L1 (x1 ) has a minimum, it occurs at some x1 ∈ [α2 , αN ].
However, this together with the conclusion of Step 1 implies that the minimum
occurs at some x1 = αi for at least one value of i.

Step 3. Let SI = L1 (αI +1 ) − L1 (αI ), and by the reordering in Step 2,


N 
N
SI = |mi ||αI +1 − αi | − |mi ||αI − αi |
i=2 i=2


I 
N
= |mi |(αI +1 − αi ) − |mi |(αI +1 − αi )
i=2 i=I +1


I 
N
− |mi |(αI − αi ) + |mi |(αi − αI )
i=2 i=I +1
 4.2 Distance and Proximity Properties 95


I 
N
= |mi |(αI +1 − αI ) − |mi |(αI +1 − αI )
i=2 i=I +1
 
I 
N
= (αI +1 − αI ) |mi | − |mi | .
i=2 i=I +1

Therefore,

SI = (αI +1 − αI )TI . (4.33)

Since αI +1 − αI > 0, the sign of SI is determined by the slope of the line segment
over the interval (αI , αI +1 ).

Step 4. By Step 1,


I 
N +1
I 
N
TI = |mi | − |mi |, TI +1 = |mi | − |mi |.
i=2 i=I +1 i=2 i=I +2

That is, TI +1 is found by moving the term |mI +1 | from the negative to the positive
sum. Therefore, TI is monotone increasing with I . Hence from (4.33), if TI = 0,
the minimum of SI is attained at a single point. Further, for TI = 0, the minimum
is attained over an entire interval (αI , αI +1 ).

Step 5. We claim that the smallest value of I such that


I 
N
|mi | ≥ 0.5 |mi | (4.34)
i=2 i=2

is the value of i for αi where L1 (x1 ) attains its minimum. For this is the first value of
i for which the slope TI can change sign. This observation provides an algorithm for
finding the α∗i by construction. This provides a graphical construction for finding
the minimum described in Fig. 4.22.

Step 6. When some of the αi are equal, then by (4.33), there will be stationary
values of L1 (i.e., regions where it is constant) and that may or may not be at the
minimum; see Fig. 4.23.

Step 7. If there are |mi | = 0 for i in a set E, then

 
L1 (x1 ) = |bi | + |mi ||x1 − αi |, (4.35)
E Ẽ
96 4. Multidimensional Lines

Figure 4.22. Constructing the x1 = αI minimizing the L1 distance between two lines.
For comparison, the minimum L2 distance occurs at x1 = α ∗ . The bar chart on the right of the
X̄6 -axis shows the construction of the slopes TI . There the |mi | are added in the order 6, 2, 4, 3, 5
obtained from the order of increasing αi (as shown in the X̄1 -axis). The index I of the interval with
the median value of the N i=2 |mi | provides the correct x1 = αI . Here |m4 | dominates the sum,
yielding I = 4, so the minimum occurs at α4 .

Figure 4.23. Here the L1 and L2 minimum distances are identical occurring at the same value of
the constrained variable.

where E denotes the complement of E. This can be carried out by a modification

of the above procedure. It turns out that for small |mi | and |bi | not small,
the corresponding αi are very large, and if not all the |mi |’s are small, then the
procedure simply ignores such values. 
 4.2 Distance and Proximity Properties 97

** On the L2 Minimum Distance between Two Lines: Monotonicity with N

Rather than the standard minimization approach, consider fitting a sphere with
center at a point P ∈ and tangent to
at a point P
. The radius r of this sphere
is the (minimum) distance between the two lines with P and P
the points where it
occurs. To accomplish this fit, for any number α let P ∈ , where x1 (P ) = α and
S is the sphere with radius r centered at P , i.e.,


N
S: (xi − mi α − bi )2 = r 2 ,
i=1

and for P
∈
with x1 (P
) = α
and xi = xi
for the xi
of
(see 4.28), so we
obtain

N
S ∩
: (m
i α
+ bi
− mi α − bi )2 = r 2 .
i=1

We now “shrink” the sphere so that it is just tangent to

by finding the minimum
r 2 . Expressing this in matrix form yields
⎛ ⎞⎛ ⎞
M −C −D α
r 2 = (α α
1) ⎝ −C M
D
⎠ ⎝ α
⎠ (4.36)
−D D
B 1

= Mα 2 + M
α
2 − 2Cαα
+ B + 2(D
α
− Dα). (4.37)

a quadratic relationship between the three unknowns α, α

, r where


N 
N 
N
M
= m
2
i , M= m2i , C= mi m
i ,
i=1 i=1 i=1


N
B= bi2 , bi = bi
− bi ,
i=1


N 
N
D= bi mi , D
= bi m
i . (4.38)
i=1 i=1

The minimum r is found from

∂r 2 ∂r 2
= 0 ⇒ Mα − Cα
= D, = 0 ⇒ Cα − M
α
= D
. (4.39)
∂α ∂α
98 4. Multidimensional Lines

The minimum is reached when

M
D − CD
CD − MD

r 2 = B + D
α
− Dα, α= , α
= , (4.40)
Q Q

where Q = MM
− C 2 ≥ 0 by the Cauchy–Schwarz inequality. When Q = 0
and B = 0, the lines are parallel and the constant square distance between them is
r 2 = B − D 2 /M. In R2 with m2 = m
2 , the minimum is r ≡ 0, since the lines
always intersect.

One Constrained Variable and Monotonicity with N

The minimization of the distance now follows for x1 = α ∗ constrained to be the
same for both ,
. By a similar argument, s is the minimum distance in this case,
and the α ∗ that minimizes it is found to be
N
∗ (D − D
) W i=2 bi mi
α = =− =−  N
, s 2 = B + (D − D
)α ∗ ,
V V 2
i=2 mi
(4.41)
N
N
where V = i=2 (mi − m i )2 , W = i=2 bi mi , or
 N 2

N
2 2 i=2 bi mi
s = bi − N . (4.42)
2
i=2 i=2 mi

Applying Lagrange’s identity yields the alternative form


N 
s2 mi 2 = (bj mk − bk mj )2 . (4.43)
i=2 2≤k<j ≤N

How does the minimum distance s vary as a function of the dimension N? Let
and
be two lines in RN+1 with P and P
respectively their closest points and
consider the projections N ,
N on RN with closest points PN , PN
respectively.
In (4.42), denote the distance s for RN by sN and the corresponding α ∗ by αN ∗,
bi = Bi , mi = Mi , and


N 
N
DN = Mi2 , CN = Bi Mi .
i=2 i=2
 4.2 Distance and Proximity Properties 99

Subscript the analogous variables by N + 1 for RN +1 . Then


N 2
CN CN
2 ∗
sN = Bi2 − , αN =− . (4.44)
i=2
DN DN

It is easily found that

2 2 (BN+1 DN − CN MN +1 )2
sN+1 − sN = ≥ 0,
DN (DN + MN
2
+1 )

and that sN ↑ N (i.e., monotone increasing) with sN = sN +1 ⇐⇒

CN M
∗ DN + BN+1 DNN+1 BN +1 ∗
αN+1 =− 2
=− = αN ,
MN +1 MN +1
1+ DN

with xN+1 (P ) = xN+1 (P

). All this shows that not only does sN increase mono-
tonically with N but also that in RN , the closest points between N and
N are not
in general the projections of the closest points P , P
∈ RN +1 .

** L1 versus L2 Constrained Minimum Distance

For the constrained L2 distance on x1 we consider the square


N
L2 (x1 )2 = (mi )2 (x1 − αi )2 .
i=2

It is left as an exercise to prove for the unconstrained case that in RN ,

L1 (x1 )
L1 (x1 ) ≥ L2 (x1 ) ≥ √ . (4.45)
N
Therefore, for the constrained case where the dimensionality is N − 1,

L1 (x1 )2 ≤ (N − 1)L2 (x1 )2 .

Hence if α minimizes L1 on a particular interval I , i.e.,

L1 (α) ≤ L1 (x1 ) ∀x1 ∈ I,

and β minimizes L2 2 on I , i.e.,

L2 (β)2 ≤ L2 (x1 )2 ∀x1 ∈ I,

100 4. Multidimensional Lines

then
L1 2 (β) L1 2 (α)
L2 (α)2 ≥ L2 (β)2 ≥ ≥ . (4.46)
(N − 1) (N − 1)
From (4.45) and (4.46) we obtain upper and lower bounds for the minimum L2 (x1 )
in terms of the minimum of L1 (x1 ). It turns out that these bounds are tight, since
there exist cases in which these minima are equal for the same x1 . Still, it seems that
the L2 minimum distance does not lend itself easily to visualization in -coords as
the L1 distance does. Consider this a challenge!
Finding the minimum distance between lines raises the question of how to
measure the “closeness” — according to some criterion — between lines. This is a
matter of considerable interest in various applications such as computer vision and
geometric modeling. Gauging the “closeness” between things of the same kind is
really the subject matter of topology, whose original name was analysis situs. Later,
in Chapter 8, various topologies for lines, p-flats, and hyperplanes are studied in
depth.

Exercises
1. Write an algorithm for constructing a point on a line given the ¯ and state its
complexity.
2. Write an algorithm for constructing the representation of a line given two points
P1 , P2 on the line. Be careful and delimit the special cases.
3. Write an algorithm for constructing the intersection of two lines or verifying
nonintersection and state its complexity. Clarify the special cases.
4. Write an algorithm for constructing the constrained L1 minimum distance and
state its complexity. List all special cases in which the algorithm fails and why.
5. Prove (4.45). Hint: use induction on the dimension N.
6. Construct a case in which L1 (x1 ) has stationary portions. State conditions for
this to occur in general.
7. Construct a case in which the L2 (x1 ) and L1 (x1 ) minima are equal for the same
x1 .
8. Write an algorithm for minimizing (4.35) and state its complexity.

4.2.4 ** Air Traffic Control

Around 1985, the interest generated by the impending design and construction of a
new air traffic control (ATC) system in the United States led, among other things, to
exploration for new information displays. One of them was based on the -coords
methodology (US patent # 4,823,272), and it is briefly illustrated here in this,
 4.2 Distance and Proximity Properties 101

Figure 4.24. Path (left) and the full trajectory information (right) of an aircraft.
In -coords, the aircraft’s position can be found and displayed at any time.

the concluding section of the chapter. The reader interested on some nontechnical
background on ATC is referred to [76] and for an early description of the then
projected new system AAAS to [96].

Displaying Trajectory Information

The trajectory of an aircraft (considered as a point) with constant velocity, as occurs
for much of the time in civil aviation, is a line segment in 4-D, which may be
described by
⎧
⎪
⎪ x = v1 T + s1 ,
⎨ 1
xi+1 = v2 xi + s2 , (4.47)
⎪
⎪
⎩
xi+1 = v3 xi + s3 ,

where the xi are the space coordinates, T is time, V = (v1 , v2 , v3 ) is the velocity
vector, and S = (s1 , s2 , s3 ) is the initial position at T = 0. We already know that in
parallel coordinates the complete trajectory information is given by three stationary
points, which may be labeled21 T:1, 1:2, and 2:3. Using the T̄ (time)-axis as a clock,
the present position of the aircraft as well as where it was and where it will be at
any given time can be found and displayed as in Fig. 4.24. Recall that the horizontal
coordinate of the points representing the trajectory given by (4.47) is determined
only from the quantities mi , with m1 the velocity component in the x1 direction.

21
For brevity we omit the ¯ part from the notation whenever it is clear from the context.
102 4. Multidimensional Lines

Figure 4.25. Two aircraft flying on the same path, since their 1:2 and 2:3 points coincide.
They have a constant separation, with their velocities the same, since T:1 points have the same
x-coordinate.

With x3 the altitude, m2 is the tangent of what in aviation parlance is called the
heading, and the horizontal position of T:3 provides the vertical velocity.
Certain properties of the trajectories are immediately evident from the display.
Consider the situation in which aircraft are flying on the same path, as is often the
case in civil aviation. An example is shown in Fig. 4.25, where the points 1:2 and
2:3 are shared, since these are the points describing the path in -coords. Further,
it is evident that they have the same velocity, since the points T:1 have the same
horizontal coordinate. Otherwise, if one point is between the T̄ , X̄1 and the other
outside axes, their velocities must have opposite signs, and since they have the same
path, they are flying in opposite directions. If they are both between the axes or both
outside the axes, then the leftmost point corresponds to the greater velocity. Other
properties are also “visible.”
The time and positions at which the distance between two aircraft is mini-
mum is displayed in Fig. 4.26. This critical situation is accurately portrayed in
-coords, while in 3-D displays, not to speak of 2-D projections, it can appear
misleading. The possibility of transforming angular to lateral deviations is illustrated
in Fig. 4.27. This is important since our eyes can distinguish lateral deviations
better rather than angular deviations, such as those that may occur in the assigned
path. This is a “human factors” advantage which may be partially offset by the
nonlinear scale, a result of the denominator (1 − mi ) of the points that represent the
path.
 4.2 Distance and Proximity Properties 103

Figure 4.26. Closest approach of two aircraft.

The time and closest positions are clearly seen in -coords. Appearances can be misleading in a
3-D (near perspective) display, where the aircraft appear to be nearly colliding. It is even more
uninformative in 2-D, where only a projection is displayed.

Figure 4.27. Transforming deviations in heading (angle) to lateral deviations.

Particle Fields, Scrapes, and Conflic Intervals

The goal of ATC is to direct aircraft safely to their destination while minimally
interfering with their planned trajectories. This involves maintaining a minimum
104 4. Multidimensional Lines

Figure 4.28. Protected Space in 3-D is a thin cylinder.

separation between the aircraft; detecting impending conflict , which are violations
of the minimum separation; and resolving the conflicts by changing the trajectories
subject to certain constraints and priorities. The typical safety requirements in 3-D
(sometimes referred to as “separation standard”) involve horizontal and vertical
separations, as shown in Fig. 4.28. The protected airspace is a thin cylinder usually
2,000 feet high and diameter of 5 nautical miles. This shape is sometimes called
a “puck,” due to its similarity with the hockey puck. Conflict resolution is an
instance of the theoretical asteroid avoidance problem (AAP), which turns out
in general to be NP-hard [21]. In ATC, for a number of good reasons it is preferred
to resolve conflicts if possible without altitude changes. An algorithm for planar
(i.e., no altitude change) conflict detection and resolution in ATC (USA patent
# 5,058,024) was proposed in conjunction with the design competition for the new
ATC system (AAAS). It was tested on some complex groupings of flying aircraft
(“complex scenarios” for short) that were provided by the USA Federal Aviation
 4.2 Distance and Proximity Properties 105

Figure 4.29. Relation between maneuver speed and turn angle.

An equal-speed maneuver results with turn angle α, whereas turn angles greater than or less than α
require speeds slower or faster than |Vk | respectively.

Administration [127]. We shall give the rudiments of the algorithm and then apply
it to one of the complex scenarios.
Given a set of aircraft considered as points moving with constant velocity, if the
minimum separation to be maintained is denoted by d, then the protected airspaces,
for the planar case, are circles with radius d/2 centered at each aircraft and moving
with the same velocity. The ensuing conflict and resolution analysis is done first
for a pair of circles and then for an arbitrary number of circles. There is no loss of
generality in reducing this to the point-circle configuration above (by shrinking one
circle to) a point (the aircraft itself) ACk with velocity Vk , while the circle Ci for the
aircraft ACi has double the original radius and velocity Vi , as shown in Fig. 4.29.
Here the velocities are assumed to be constant, as when the aircraft are flying after
having reached their cruising altitude. A line , taken here for convenience to be
vertical, going through ACk is shown at the initial time t = 0. On a vector fiel is
106 4. Multidimensional Lines

defined with the points on the line now being considered as “particles” endowed
with the same velocity Vk as ACk for t ≥ 0. It is clear that there exists exactly one
such particle, initially in position (k10 , bik
0 ), that just “scrapes” (i.e., is tangent to)
the circle Ci at time tbik . Similarly, there exists a unique particle initially in position
(k10 , fik
0 ) that exactly scrapes C from the front. The particles starting above b0
i ik
pass the circle safely from the back, particles below fik 0 pass in the front, while

those in the interval Iki = [fik 0 , b0 ] eventually intersect the circle. This provides
ik
the conflict detection, since for any aircraft on , only those on Iki will be in conflict
with ACi . For lack of a better name, Iki is called the conflic interval of i with respect
to k. There is more information in Fig. 4.29 worth pointing out. The parallelogram
enclosing Ci has two sides parallel to Vi and two parallel to Vki , the velocity of
k relative to i. That is, an observer sitting on Ci will see the particles passing by
with velocity Vki . Only the particles entering this parallelogram will intersect Ci ,
and these are the particles on Iki . So the particle scraping the circle from the back
will be seen by such an observer traveling on the top (back) tangent to Ci and
touching it at the point Bj k . That, in fact, is the point where the back scrape will
occur. The intersection of this side with occurs precisely at the point (k10 , bik 0 )
(why?), providing the particle responsible for the back scrape. Since this particle
has velocity Vk , the intersection of a line in this direction with the path of Bj k
(which has velocity Vi ) must be the position where the back scrape occurs, from
which, in turn, the time tbj k can be found. All this is easily seen by noticing that
the triangle so formed is similar to the triangle showing vector subtraction in the
upper left. The situation is exactly the same vis-a-vis the front scrape. In this way,
running through all the indices i = 1, . . . , N, i = k, the corresponding conflict
intervals can be found.
In -coords the paths of the particles on Iki are transformed into points. These
paths all being parallel, they are represented by points and will, in turn, correspond
to an interval, say Iki ¯ on a vertical line at x = 1 with m the slope of Vk . The
1−m
situation becomes more interesting when this is considered ∀i = 1, . . . , N, i = k.
Since m is still the same, all the corresponding Iki ¯ fall on the same vertical line.
Let us see how this works on one of the complex scenarios (# 8) from [127];
where 10 aircraft are involved. What makes it “interesting” is that 6 of these,
shown in Fig. 4.30, are flying at the same altitude and somewhat menacingly
toward each other; several impending conflicts are seen in Fig. 4.31. The corre-
sponding conflict intervals are shown in Fig. 4.32, showing the paths for the Back
scrapes and Front scrapes of the aircraft # 2, . . . , 6 and the times when they occur.
The times are computed from (4.47) or directly from its point representation in
-coords.
 4.2 Distance and Proximity Properties 107

Figure 4.30. Six aircraft flying at the same altitude.

These positions are at a certain time (taken as 0 seconds and shown on the left-hand-corner). Circles
centered at each aircraft are the protected airspaces, with the diameter the minimum allowable
separation. The arrows represent the velocities.

Figure 4.31. Conflicts, indicated by overlaping circles, within the next five minutes.
108 4. Multidimensional Lines

Figure 4.32. Conflict Intervals I1k , k = 2, 3, 4, 5, 6.

These are with respect to aircraft # 1 for the complex scenario shown in Fig. 4.30 and the conflicts
shown in Fig. 4.31. The path of aircraft # 1 is shown as a point on the vertical line. Lowest point
shown is 3B and represents the path of Back scrapes # 3, and next to it is the time when this occurs.
In general, B or F denotes back or front scrapes for the indicated aircraft number and the time at
which the scrape occurs.

Conflic Resolution
Now the fun begins! Let us denote by k̄ the point representing the path of ACk .
Clearly, ACk is conflict with aircraft # i ⇔ k̄ ∈ Iki . For example, from Fig. 4.32
it is easily seen that aircraft # 1 is in conflict with # 2 and # 3 and no others. The
 4.2 Distance and Proximity Properties 109

challenge is to resolve the conflicts with the allowable maneuvers, which in our case
are parallel offsets, which are commonly used en route. Such a maneuver consists
of:

1. a turn (left or right) that is bounded above, say, by a θMAX , that is, a turn that
is relatively “small,”
2. followed by a relatively short straight path; this is the “offset,”
3. then a second turn returning the aircraft to a path parallel to the original one
(i.e., the same heading). The new path is not allowed to be “too far” from the
original, i.e., the distance between the path is bounded above, say, by s.
4. At all times, the aircraft’s speed, i.e., |V |, is constant, that is, no slowdowns or
speedups are allowed even during the maneuvers.
The turn here is idealized in the sense that it is considered “instantaneous” in
time and occurs at a point in space. All this can be adjusted to more realistic models
based on the capabilities of various airplanes, policies for various airlines, and
air sectors, as well as many other constraints and priorities such as time-to-conflict
(“validation time”), all of which are beyond our scope here. The strategy for conflict
resolution then is to use parallel offset manuevers subject to the given constraints
in order to redirect the aircraft to conflict-free paths. The conflict intervals are well
suited for this task. Since eventually the new path is parallel to the original path, the
idea, when # j is in conflict with some aircraft, is to find the nearest conflict-free
path, say j¯
, and check to see whether under the constraints, the maneuvers allow
# j to reach this path prior to the scrape. That is, we want # j to behave like the
particle originally with path j
in a neighborhood of the scrape. Specifically, for
# 1 we see in Fig. 4.32 that a good candidate is the path of the particle causing the
back scrape 3B. This is found by taking the union of all conflict intervals containing
1̄, which is also an interval (from a basic theorem in topology) and examining
the closest endpoint. In general, the new interval’s endpoint may lie on conflict
intervals not included in the union, such as 3F ∈ I14 , in which case these intervals
are added to the union until a conflict-free path is found. This process terminates,
since there is a finite number of intervals. Then the closest endpoint is checked to
see whether it is feasible, that is, whether the aircraft in conflict, such as # 1 here,
can be redirected to the new path subject to the constraints. It is not difficult to apply
the maximum-offset constraint.
To see how equal-speed maneuvers can be achieved, refer again to Fig. 4.32,
where the turn-angle-to-maneuver-speed relation is illustrated. When ACk makes a
turn with angle α, an isosceles triangle is formed such that after the turn, the aircraft
traveling with speed |Vk | arrives at J at the same time as the particle originally at
fik0 , since they both travel with the same speed over an equal distance. Hence after
110 4. Multidimensional Lines

the point J, ACk , now in its new path, mimics the behavior of the particle, namely
the scrape, in the neighborhood of some circle. From this picture we can also see
that turns of more or less than α will reach the path of fik 0 sooner or later than
the particle, necessitating a slowdown or speedup by ACk to accomplish the same
scrape as the particle. Of course, a choice of lines other than the vertical provides
more and different options, and this is exploited in the implementation.
The general AAP is NP-hard, so even its specializations may have very high
complexity. Since conflict resolution requires a real-time solution, one way to handle
the need for fast response is by cascading the algorithm in various levels, which we
briefly outline below, of increasing complexity and power. Starting with the first
level (having the lowest complexity), the set of aircraft is processed by successive
levels until either a resolution is found or no resolution is found by the highest level,
though, in our experience, satisfactory resolutions were found at the lowest level.
It is, of course, even better to run the algorithms for the various levels in parallel.
The algorithm is greedy, and in testing, we found that different resolutions are found
by the various levels. Different resolutions (or no resolutions) may also be found
by reordering aircraft on input. Listing the aircraft in order of decreasing number
of conflicts they are involved in turns out to be a “good” ordering. Other ordering
heuristics, which also consider the time-to-conflict, have been tried with varying
success. When several resolutions for a conflict scenario are available, rating them
as to their “desirability” has turned out to be a very elusive and very interesting
problem.
In the simplest case (“simple rules”), an aircraft in conflict, say ACj , is placed
on the path found, as indicated above, satisfying all the constraints. This is repeated
with the updated (due to the repositioning of ACj ) conflict intervals. If this works
out for all the aircraft in the input list, a resolution is found; otherwise, the next level
is invoked. The computational worst-case complexity for this level is O(N 2 log N),
where N is the number of aircraft.
When the simple rules level finds no resolution for a particular aircraft ACk , a
sweep of the conflict intervals determines which aircraft prevented it (i.e., blocked
it) from being resolved, and there may be more than one. A criterion is applied
in order to choose the “most troublesome” aircraft blocking the resolution, tem-
porarily removing it and resolving the conflicts, if possible, using the simple rules.
Such potential conflict resolution is provisionally taken, and a resolution for the
remaining aircraft, including the one that was removed, is attempted. If successful,
the resolution spawns additional maneuvers for the blocking aircraft one at a time.
This backtracking can be nicely implemented with recursion.
The worst-case complexity of the recursion is exponential. In order to termi-
nate, the maximum number of times R that the removal of a blocking aircraft is
 4.2 Distance and Proximity Properties 111

Figure 4.33. Conflict resolution with parallel-offset maneuvers. Three pairs of tangent circles.

allowed to take place is fixed a priori, yielding that the maximum complexity is
O(N R+1 log N). It was also found profitable to experiment with different reorder-
ing heuristics rather than invoke backtracking. Though frequently successful, no
clear scenario-independent criteria for the reordering were found. However, using
several particle lines rather than one led to resolutions for the vast majority of
scenarios tried using only the simple rules. For equal-speed maneuvers, every
particle line corresponds to a specific maneuver turn. In addition to the fact that
“idealized” turns are used here, we must also consider that a pilot cannot execute a
turn with absolute precision. So a certain tolerance, say θ, for the error is involved.
So placing the particle lines in multiples of θ and covering in this way the allowable
turn range (i.e. ±θmax ) is a way to check for all implementable resolutions without
an increase in the time complexity.
Let us illustrate all this with the chosen scenario shown in Fig. 4.30. To estimate
its difficulty in a practical way, an expert air traffic controller was asked to resolve
it. He could resolve only four of the aircraft without altitude change. Even that
required a great deal of time, and the solution was certainly not “real-time.” Our
algorithm was able to resolve all conflicts without altitude changes using only two
particle lines, the resolution being shown at two time frames in Figs. 4.33 and 4.34.
There are several instances of tangent circles. This is because the algorithm is built
to utilize the scrapes, which, in turn, result in minimally disturbing the aircraft from
their original paths. The figures should illustrate the difficulty in finding constrained
112 4. Multidimensional Lines

Figure 4.34. A triple tangency.

Figure 4.35. Resolution in 3-D.

maneuvers that allow the circles to possibly touch but always exquisitely avoiding
overlapping.
As an interesting aside, notice from Fig. 4.32 that ∩6k=2 I1k = [4F, 6F]. So any
point on the interval [4F, 6F] represents a path that eventually intersects each one
 4.2 Distance and Proximity Properties 113

of the circles. This is an instance of the one-shot problem (OSP), which in general
is NP-complete [133]. In 1989, another algorithm using particle fields on the whole
plane, rather than just particle lines, was derived (US patent # 5,173,861). It also
accomplished some resolutions with altitude changes; Fig. 4.35. It was reported22
that the US Air Force is testing the use of these algorithms for positioning several
aircraft flying in formation [64] and for conflict detection and resolution [162].

Exercises
1. In Fig. 4.32 most of the associated times with the scrapes increase, since the
point representing the scrape path is higher on the vertical line. Provide the
conditions needed for these times to increase or decrease monotonically.
2. Provide the formulation for using several line fields (θ) at different angles
θ with the horizontal axis. Describe how you would efficiently construct the
corresponding conflic intervals and their functional dependence on θ.
3. Define a particle field on the whole plane, as was done only for . Find the
allowable particle regions (i.e., the set of particles satisfying the constraints)
associated with:
(a) Maximum offset,
(b) Maximum angle,
(c) Conditions where the approach may fail.
4. Find the conditions for which the approach of “particle lines” fails.
5. Formulate the “one-shot problem” carefully. Describe conditions for which the
conflict intervals can be used interactively to find one-shot solutions.

22
I am grateful to D. Sobiski for this information.
This page intentionally left blank
This page intentionally left blank
5

Planes, p-Flats, and Hyperplanes

5.1 ♣ FT-1 Planes in R 3

5.1.1 Vertical Line Representation

A hyperplane in RN can be translated to one that contains the origin, which is an
(N − 1)-dimensional linear subspace of RN . Since RN −1 can be represented in
-coords by N − 1 vertical lines and a polygonal line representing the origin, it is
reasonable to expect a similar representation for hyperplanes in RN . An intuitive
discussion for R3 clarifies matters. Consider a plane π as shown in Fig. 5.1,
intersecting the x1 x2 plane at the line y 1 and the x2 x3 plane at the line y 2 with
A = y 1 ∩ y 2 . The lines y i , i = 1, 2, being lines in R3 , are represented by two
i , ȳ i each as shown. Next we construct a nonorthogonal coordinate
points, ȳ12 23
system on π using the y i as axes, consisting of the lines parallel to y 1 and the
lines parallel to y 2 . Any point P ∈ π can be specified as the intersection of two
lines, one parallel to y 1 and the other parallel to y 2 . The family of lines parallel
to y 1 is represented by a vertical line Ȳ1 containing the point ȳ12 1 . Similarly, the
2 represents the lines parallel to y 2 . Strictly
vertical line Ȳ2 containing the point ȳ23
speaking, the vertical lines Ȳi represent the projections on the x1 x2 and x2 x3 planes,
respectively, of the two families of parallel lines. Therefore, the vertical lines
represent two such families of parallel lines on any plane parallel to π. By choosing
a point, say A, as the origin, we obtain a coordinate system specific to π. So
the plane can be represented by two vertical lines and a polygonal line. Clearly,
the same argument applies in any dimension, and therefore a hyperplane in RN
can be represented by N − 1 vertical lines and a polygonal line representing one
of its points [97]. Conversely, a set of coplanar points chosen on a grid such as

A. Inselberg, Parallel Coordinates, DOI 10.1007/978-0-387-68628-8_5,

© Springer Science+Business Media, LLC 2009
116 5. Planes, p-Flats, and Hyperplanes

Figure 5.1. A plane π in R3 represented by two vertical lines and a polygonal line.
This is a planar coordinate system with the polygonal line representing the origin.

Figure 5.2. A set of coplanar points in R3 with the two vertical lines pattern.

the one formed by lines parallel to a coordinate system, as shown in Fig. 5.1, is
represented by polygonal lines with a pattern specifying two vertical lines as shown
in Fig. 5.2.
About the time these patterns were being discovered, word got around, and we
were requested to visually explore for relations that may exist in a set of industrial
 5.1 ♣ FT-1 Planes in R3 117

Figure 5.3. Industrial data with a “vertical line” pattern between R111 and R112 axes.

Figure 5.4. The portion between the R111 and R112 axes is magnified (left).
This suggest that there is linear relation between R111, R112, and an unknown parameter, as is also
indicated in the Cartesian coordinates plot on the right.

data consisting of several thousand records with eight variables. Part of this data
set is plotted in parallel coordinates and shown in Fig. 5.3. The pattern between
the R111 and R112 axes, which for clarity is magnified and shown in Fig. 5.4,
118 5. Planes, p-Flats, and Hyperplanes

resembles one of vertical lines formed in Fig. 5.2. Yet for this pattern to represent
a set of coplanar points, at least two, not one, vertical lines are needed, even for
the plane with minimal dimensionality in R 3 . Still, the variables R111 and R112
are linearly interrelated, and there must be another variable, say X, also linearly
related to R111 and R112, as is clear from Fig. 5.4 (see also the first exercise below).
All this suggested that an important variable was not being measured. With these
hints, a search was mounted and the missing variable was found. This, of course,
was a stroke of good fortune and an instance of a favorite stochastic theorem, i.e.,
“You can’t be unlucky all the time.” Let us return from our digression. Before this
is done, it is worth distinguishing the vertical line pattern between the R8 and R9
axes with that between the R111 and R112 axes. On the R8 axis and lower part of
the R9 axis, the data are equally spaced, while this is not the case with the last two
axes.

5.1.2 ** Planar Coordinates

The vertical lines representation of planes can also be obtained using the theory of
envelopes,23 starting with the description

π : c1 x1 + c2 x2 + c3 x3 = c0 . (5.1)

At first the envelope of the family of lines parallel to y 1 is found, as seen in Fig. 5.1.
Two coordinates are needed to determine a point P ∈ π . One coordinate, say
x3 = α, specifies a plane π
parallel to the x1 x2 -plane. The line
(α) = π ∩ π
is
in fact one of the lines parallel to y 1 ; see Fig. 5.7. It has the representation:

¯
23 (α) = (2, α),
¯
(α) :  2 c0 −c3 α  (5.2)
¯
12 (α) = c1c+c 2
, c1 +c2 ,

since


23 (α) : x3 = α,
(α) : (5.3)

12 (α) : c1 x1 + c2 x2 = c0 − c3 α,

where as usual, the distance between the parallel axes is one unit.
Returning to Fig. 5.2, each vertical line is formed by a set of points, and on the
Ȳ1 -axis these are the ¯
12 (α). The polygonal lines intersecting at each such point
represent points on the corresponding line
12 (α). Together, these intersections
provide one of the two vertical lines in the representation of π. As we know from
23
See the chapter on envelopes.
 5.1 ♣ FT-1 Planes in R3 119

Chapter 3, for each α, ¯

12 (α) is the envelope of a set of polygonal lines representing
points on a line. As α varies, the ¯
12 (α) form the vertical line Ȳ1 , since, as one can
see from (5.2), their x-coordinate is a constant (independent of α). Strictly speaking,
the points ¯
23 (α) also vary with α and fall on the X̄3 -axis. Since this would also
occur for any other plane and the X̄3 -axis already exists, it does not contribute any
information about the specific plane π, so typically it is not considered as part of
the representation of a particular plane. This takes care of the lines on π parallel to
the y 1 -axis. In exactly the same way it is found that the vertical line Ȳ2 represents
the lines parallel to the y 2 -axis.
What has effectively happened is that a coordinate system is constructed in the
plane π consisting of two vertical lines and a polygonal line as an origin. The plane
π is after all a 2-fla , i.e., it can be determined by three points. In general, an n-fla
is a linear manifold (translate of a linear space) that is determined by n + 1 points.
For an excellent short book on N-dimensional geometry see [163]. Now we would
like to use this coordinate system to specify points on π. With reference to Fig. 5.5,
consider a point P ∈ π. From its (x1 , x2 , x3 ) coordinates we want to find its y 1 and
y 2 coordinates or equivalently to express P = 1 ∩ 2 , where 1 , 2 are lines on π
and parallel to the y 1 - and y 2 -axes respectively. So we look at P12 , the projection
of P on the x1 x2 plane and construct the line through P12 parallel to y 1 . This line
is represented by the point ¯112 = P̄12 ∩ Ȳ1 on the Ȳ1 -axis. Similarly, we obtain
¯223 = P̄23 ∩ Ȳ2 on the Ȳ2 -axis, where P23 is the projection of P on the x2 x3 plane.
Notice how careful tracking of the indices is already helpful in the construction.
The pair ( ¯112 , ¯223 ) are the planar coordinates of P̄ in terms of the Ȳ1 - and Ȳ2 -axes.
An interesting consequence is that a line ⊂ π can be represented by a single point
η̄ in, terms of the Ȳ1 , Ȳ2 coordinate system. Of course, it is still represented by two
¯ in terms of the original X̄i , i = 1, 2, 3, Fig. 5.5.
points, the s,
Theorem 5.1.2. For a line ⊂ π ⊂ R3 , the points ¯12 , ¯23 , and η̄ are collinear.

Proof. Step 1. With reference to the figure below, let P1 , P2 ∈ . The two lines AB
and A
B
joining the planar coordinates of P̄1 and P̄2 intersect at η̄ = AB ∩ A
B
.
Step 2. The two triangles ABC and A
B
C
formed between Ȳ1 and Ȳ2 are in
perspective with respect to the ideal point in the vertical direction.

Step 3. The sides AB and CB are portions of P̄2 and correspond to A

B
and C
B
,
which are portions of P̄1 .
Step 4. By Step 3, ¯12 = AC ∩ A
C
and ¯23 = BC ∩ B
C
.
Step 5. By Step 2 through Step 4 and Desargues’s theorem, ¯12 , ¯23 , and η̄ are all
on the same line L̄. 
120 5. Planes, p-Flats, and Hyperplanes

Figure 5.5. A line in a plane π is represented by one point η̄12 .

This is in terms of the planar coordinates Ȳ1 and Ȳ2 . The point η̄12 is collinear with the two points
¯12 and ¯23 .

Of course, by the 3-point collinearity property of Chapter 4, the point ¯13 is also
on L̄.
Corollary 5.1.3. The rotation of a plane in R3 about a line corresponds to a
translation of a point on a line.

Proof. With reference to the Fig. 5.6, let ⊂ π ⊂ R3 . Rotate π about to a new
position π∗. Let ¯12 , ¯23 represent in the X̄1 , X̄2 , X̄3 coordinates and η̄, η̄∗ in the
Ȳ1 , Ȳ2 and Ȳ1 ∗, Ȳ2 ∗ coordinate systems respectively. By the theorem, ¯12 , ¯23 , and
η̄ are on a line L̄. Also, ¯12 , ¯23 , and η̄∗ are on the line L̄. That is, the rotation of π
about corresponds in -coords to a translation of the point η̄ on L̄. 
This is the 3-D analogue of (rotation of a line around a point) → (translation
of a point along a line) in 2-D and corresponds to the point ↔ plane duality in P3 .
Let us explore the computation involved, where is given by

12 : x2 = m2 x1 + b2 ,
: (5.4)
23 : x3 = m3 x2 + b3 .
 5.1 ♣ FT-1 Planes in R3 121

Figure 5.6. Rotation of a plane about a line ↔ translation of a point along a line.

Figure 5.7. The construction of Ȳ1 (left), the first, and Ȳ2 (right), the second, vertical axes.
Start with a line on π parallel to the y 1 and y 2 axes respectively.
122 5. Planes, p-Flats, and Hyperplanes

There exists a one-parameter, say k, family24 of planes containing . In fact,

the equation of any one of the planes in the family is given by

π : (x3 − m3 x2 − b3 ) + k(x2 − m2 x1 − b2 ) = 0. (5.5)

It can be verified that in Cartesian coordinates,

 
m23 − 2m3 − k 2 b2 k 2 + m3 b3
η̄12 = ,− . (5.6)
m23 − m3 + k 2 (m2 − 1) m23 − m3 + k 2 (m2 − 1)

Every value of k corresponds to a position (of the rotated plane) π and, in turn, about
a position (of the translated point). η along the line formed by the points ¯12 and ¯23 .
The generalization for RN is straightforward and is not covered here (see [97]).

Fun and Games with the ILM

Open the ILM2. On the right is a plane π containing a line .

The plane is represented by two vertical lines Ȳ1 , Ȳ2 on the left.
The line is specifie by two points P1 , P2 (right) and represented by
¯12 , ¯23 on the left. The point η̄12 represents the line with
respect to the planar axes Ȳ1 , Ȳ2 . All three points ¯12 , ¯23 , η̄12
are on a line L̄.

For some experimentation: slide the point π̄123 along L̄;

notice the translation of η̄12 and corresponding rotation of the plane.
Change the line and/or plane and repeat.
Click other buttons and experiment.

Exercises
1. The pattern between the R111 and R112 axes in Figures 5.3 and 5.4 raises a
question. What if a different permutation of the axes were chosen? Whereas
the coplanarity information would still be preserved, the “vertical line pattern”
would not be seen. Was it fortuitous that the “right” axes permutation was
chosen to plot the data? It is not difficult to show that for N variables, there
are O(N/2) “cleverly” chosen permutations that provide all possible adjacent
24
In the language of projective geometry this is called a pencil of planes on the line.
 5.1 ♣ FT-1 Planes in R3 123

Figure 5.8. Constructing a point in a plane.

pairs of axes. For example, for N = 6, three such permutations are (just
showing the subscripts of the X̄i ) 126354, 231465, 342516 [184]. Returning to
the example, there are four such permutations, one of which has the coplanarity
pattern.
(a) Prove this permutation result. Hint: construct an undirected graph whose
vertices are variables, and edges between a pair of vertices designate that
the corresponding vertices are adjacent in a particular permutation.
(b) Find an algorithm for computing these special O(N/2) permutations.
(c) Generalize this result for adjacent triples.
(d) Generalize this result for adjacent p-tuples.
2. Given three noncollinear points in R3 , find an algorithm for constructing in
-coords a set of planar coordinates for the plane determined by the three
points.
3. Construct a set of planar coordinates for a plane π perpendicular to one of the
principal 2-planes (i.e., π being parallel to one of the orthogonal axes).
4. Given a set of planar coordinates for a plane π, find an algorithm for
constructing a plane π
perpendicular to it.
5. In -coords find an algorithm for determining the intersection of two planes.
124 5. Planes, p-Flats, and Hyperplanes

6. Given a plane π and a line , provide an algorithm that in -coords determines

whether ⊂ π .
7. Given a plane π and a point P , provide an algorithm for determining whether
P is below, on, or above the plane. For the notion of above/below, consider π
partitioning R3 in to two half spaces that need to be distinguished. All points
in one half space are “above,” etc.
8. Provide an algorithm in -coords for intersecting a line with a plane.
9. Verify,

(a) (5.5),
(b) (5.6).
♣ FT-1e

5.2 ♣ FT-2 Representation by Indexed Points

The representation of a plane in terms of vertical lines is basically the representation

of a specific coordinate system in the plane. With this representation, the coplanarity
of a set of points can be checked visually only if the points are on a rectangular grid
as in Fig. 5.2. In Fig. 5.9 the polygonal lines shown on the X̄1 , X̄2 , X̄3 -axes represent

Figure 5.9. Randomly selected coplanar points in R3 represented by the polygonal lines on the
first three axes.
 5.2 ♣ FT-2 Representation by Indexed Points 125

Figure 5.10. Coplanarity!

A pencil of lines on a point is formed by joining the pairs of points representing lines on a plane.

randomly sampled points in a plane π in R3 . There is no discernible pattern. Is there

a -coords “pattern” associated with coplanarity of randomly selected points?
A new approach is called for [47]. Describing even a simple 3-dimensional
object such as a room in terms of its points is just not intuitive. The description of it in
terms of cross-sections including the boundary planes helps us visualize the object.
Pursuing the analogy, it is proposed to describe p-dimensional objects in terms of
their (p − 1)-dimensional subsets rather than their points. So the representation of
a plane π is attempted in terms of the lines it contains. From each pair of polygonal
lines shown in Fig. 5.9, the representation of the corresponding line ⊂ π is
constructed. The result in Fig. 5.10 is stunning: the lines joining the pairs of points
¯12 , ¯23 in turn determine a pencil of lines on a point, and this is characteristic of
coplanarity. Let us explore it.

5.2.1 The family of “Superplanes” E

Behind the striking pattern in Fig. 5.10 lurks a very special subspace of RN . Until
now, the -coords axes were taken equidistant, with the y- and X̄1 -axes coincident.
This was a matter of convenience, which has served us well until now. The time
has come to look at the general setting shown in Fig. 5.11, the position of the X̄i
specified by the directed (i.e., signed) distance di , with the stipulation that for some
i = j , di = dj . We consider the set of points P ∈ RN whose representation in
-coords collapses to a straight line. That is, P̄ : y = mx + b, and for a specific
choice of (m, b), the corresponding point is
126 5. Planes, p-Flats, and Hyperplanes

P = (md1 + b, . . . , mdN + b) = m(d1 , d2 , . . . , dN ) + b(1, . . . , 1). (5.7)

For m, b ∈ R this collection of points forms the subspace π s of RN spanned by

the two N-tuples (d1 , d2 , . . . , dN ), (1, . . . , 1). It contains the points (0, 0, . . . , 0)
and (1, 1, . . . , 1) and the complete line u on these points. For future reference, the
unit vector on u anchored at the origin and pointing towards (1, 1 . . . , 1) is denoted
by u. The two-dimensional subspaces on the line u are called superplanes and
play a crucial role in our development. The family of superplanes generated by all
possible axes spanning (i.e., all (d1 , d2 , . . . , dN ), (1, . . . , 1)) is denoted by E , after
J. Eickenmeyer, who first discovered them. The situation is especially interesting
in R3 where the superplanes are

π s = {m(d1 , d2 , d3 ) + b(1, 1, 1)|m, b ∈ R},

and whose equation

π s : (d3 − d2 )x1 + (d1 − d3 )x2 + (d2 − d1 )x3 = 0 (5.8)

can be obtained by using numerically convenient values (i.e., m = b = 0,

m = 0, b = 1, m = 1, b = 0). It describes the pencil of planes, on the line u
each specified by the spacing of the axes. As for (5.5), it can be rewritten as a
one-parameter family:

Figure 5.11. Points in RN represented by straight (rather than polygonal) lines.

 5.2 ♣ FT-2 Representation by Indexed Points 127

d2 − d3
π s : (x3 − x2 ) + k(x2 − x1 ) = 0, k = , (5.9)
d2 − d1

with the ratio k determining the particular plane. With the y- and X̄1 -axes coincident
and the standard axes spacing d1 = 0, d2 = 1, d3 = 2, the superplane (sometimes
referred to as the “first superplane”) is

π1s : x1 − 2x2 + x3 = 0. (5.10)

The subscript 1 distinguishes it from the others formed with different spacing of the
axes. The important property from Chapter 4 can now be restated in terms of π1s .
Theorem 5.2.1 (3-point collinearity) For any line ⊂ RN , the points
¯ij , ¯j k , ¯ik , where the i, j, k are distinct integers ∈ [1, 2, . . . , N], are on a line L̄,
where L = ∩ π1s .
This is the backbone of the recursive (in the dimension) construction algorithm that
follows.

5.2.2 The Triply Indexed Points

Let us now review the construction leading to the single intersection in Fig. 5.10.
For any line ⊂ π, the three points ¯12 , ¯23 , ¯13 are:
• on a line L̄ by the 3-point-collinearity property, and conversely, since L̄ lies on
the three points, the point L is on .
• Further, since L̄ is a straight line, L must also be a point of the “superplane”
π1s .
Therefore

∩ π1s = L. (5.11)

This is true for every line ⊂ π. Specifically, for a line

⊂ π, ∃L
∈

∩ π s = L
.

Now L and L
specify a line π ⊂ π1s represented by a single point ¯π . Altogether,
!
L, L
∈ π ⇒ π ⊂ π

⇒ for π = π1s , π ∩ π1s = π ,

L, L ∈ π1 ⇒ π ⊂ π1
s s

showing that the point where all the lines intersect in Fig. 5.10 is ¯π . For reasons
that are clarified shortly, we refer to ¯π by π̄123 .
128 5. Planes, p-Flats, and Hyperplanes

Figure 5.12. The axes spacing for the second superplane π1s
.

Figure 5.13. Transferring the values from the X̄1 -axis to the X̄1
-axis.

A plane π ⊂ R3 is determined by two intersecting lines. It is advantageous to

use two lines belonging to “superplanes,” since their representation requires only
one point. One such line is π = π ∩ π1s . To determine the second line we revisit
(5.8) and choose another convenient spacing of the axes. As indicated in Fig. 5.9,
the X̄1 -axis is translated (recall that this corresponds to a rotation) to X̄1
, one unit
to the right of X̄3 , as shown in Fig. 5.12 with d2 = 1, d3 = 2, d1
= 3. The new
superplane is given by
π1s
: x1 + x2 − 2x3 = 0. (5.12)
 5.2 ♣ FT-2 Representation by Indexed Points 129

Transferring the x1 coordinate of each point P to the X̄1

-axis as shown in
Fig. 5.13 and repeating the construction, we obtain the second point, denoted
by π̄231
and shown in Fig. 5.15. The corresponding line is none other than the
intersection
π = π ∩ π1s
. What about the line H̄ joining the points π̄123 and
π̄231
? Clearly, H ∈ π , since π̄123 represents π and similarly H ∈
π . Further,
since H̄ is a line in the X̄1 , X̄2 , X̄3 coordinate system and also in the X̄2 , X̄3 , X̄1

system,
H = π ∩
π = π1s ∩ π1s
∩ π, (5.13)
as illustrated in Fig. 5.14. So the intersecting lines π ,
π determine the plane π ,
and therefore the plane can be represented by the two points π̄123 and π̄231
. But
what is going on? In Chapter 4 we saw that a line in R3 is also represented by
two points. To distinguish the two points representing a plane, three indices are
attached, while only two indices are used for the point representation of a line. The
deeper reason is, of course, that the linear relations specifying lines involve (or can

Figure 5.14. The intersections of a plane π with the two superplanes π s 1 and π1s
.
These are the two lines π ,
π that specify the plane and provide its representation.
130 5. Planes, p-Flats, and Hyperplanes

Figure 5.15. The plane π represented by two points.

be reduced to) two variables, whereas those involving planes necessarily involve
three variables. Specifically, for the plane,

π : c1 x1 + c2 x2 + c3 x3 = c0 , (5.14)
⎧ −c3
⎨ π12 : x2 = − cc1+2c c0
x + c +2c ,
s 2 3 1 2 3
π = π ∩ π 1 : (5.15)
⎩ : x = − 2c1 +c2 x + c0 .
π23 3 c3 −c1 2 c3 −c1

Therefore, in homogeneous coordinates, recalling that 1, the distance of X̄2 from

the y-axis, must be added to the first coordinate of ¯
π23 ,

π̄123 = ¯π12 = ¯π23 = (c2 + 2c3 , c0 , c1 + c2 + c3 ). (5.16)

Continuing,
⎧
⎨
π12 : 2c1 +c3
x2 = − 2c x + 2c2c+c
0
,
2 +c3 1

π ∩ π1s

2 3
=π : (5.17)
⎩
c2 −c1
x3 = − 2c x + 2c c+c
0
π23 : 1 +c3 2 1 3
,

and from the second equation it is immediate that

¯
π23 = (3c1 + c2 + 2c3 , c0 , c1 + c2 + c3 ). (5.18)
 5.2 ♣ FT-2 Representation by Indexed Points 131

Figure 5.16. The location of the ¯12 and ¯1

2 points.

Since
π is a line in the superplane π1s
, it must be that ¯
π12 = ¯
π23 . Yet a direct
computation from the first equation of (5.17) yields
¯
π12 = (2c2 + c3 , c0 , 2(c1 + c2 + c3 )). (5.19)
What is going on? Figure 5.16 explains the “riddle,” reminding us that π1s
is a
superplane in the X̄2 , X̄3 , X̄1
coordinate system, where its points such as P , Q
are represented by straight lines intersecting at ¯
π1
2 = ¯
π23 . To obtain ¯
π12 , the
x1 values must be transferred to the X̄1 -axis and then the corresponding polygonal
lines constructed to obtain their intersection at ¯
π12 = ¯
π1
2 . Recall from Chapter 3
the line → ¯ point correspondence

d b
: x2 = mx1 + b → ¯ : , , m = 1, (5.20)
1−m 1−m
where d is the directed interaxis distance. The distances from X̄1
to X̄2 and the
y-axis are two and three units respectively. Then together with the first equation in
(5.17),

−2 + c0
¯

π1
2 = 2c +c 3, 2c +c , (5.21)
1+ 1 3 2c2 +c3 1+ 1 3 2c2 +c3
132 5. Planes, p-Flats, and Hyperplanes

which in homogeneous coordinates matches (5.19), and in analogy to (5.16), we

record25 the result as

π̄231
= ¯
π1
2 = ¯
π23 = (3c1 + c2 + 2c3 , c0 , c1 + c2 + c3 ). (5.22)

To simplify the notation, we also write π̄1

= π̄1
23 = π̄231
and π̄0
= π̄123 . The
coordinates of the two points π̄0
and π̄1
contain three independent parameters and
suffice to determine the coefficients of π. Let S = c1 + c2 + c3 and denote by
x0
, x1
the x Cartesian coordinates of π̄123 and π̄1
23 respectively when S = 0.
Then
c1
x1
− x0
= 3 = 3c1
, (5.23)
S
where the ci
= ci /S, i = 0, 1, 2, 3, are the normalized coefficients. Exploring
this further, the X̄2 -axis is translated to the right by three units, the construction is
repeated, and a third point is obtained:

π̄31
2
= π̄1
2
3 = (3c1 + 4c2 + 2c3 , c0 , c1 + c2 + c3 ), (5.24)

which is also denoted by π̄2

and its x coordinate by x2
. Then

x2
− x1
= 3c2
. (5.25)

Finally, by translating the X̄3 -axis three units to the right and repeating the
construction as shown in Fig. 5.17, a fourth point is obtained,

π̄1
2
3
= π̄3
= (3c1 + 4c2 + 5c3 , c0 , c1 + c2 + c3 ), (5.26)

and for x3
, its x coordinate

x3
− x2
= 3c3
. (5.27)

Clearly, the third and fourth points are dependent on the first two, and by an easy
calculation it is found that

x2
= 6 − (x0
+ x1
), x3
= 3 + x0
. (5.28)

So with the normalization c1

+c2
+c3
= 1, the distance between adjacent (according
to the indexing) points of the plane’s representation is proportional to the corre-
sponding coefficient. The proportionality constant equals the dimensionality of the
space (see Exercise 5). Their ordinate is the constant c0
, so the equation of the plane
25
Another detailed example is computed later for the principal 2-planes.
 5.2 ♣ FT-2 Representation by Indexed Points 133

Figure 5.17. The plane π intersected with four superplanes.

Each point represents one of the resulting lines.

can effectively be read from the picture as shown in Fig. 5.18. Occasionally, we say
that the distance between adjacent indexed points is the corresponding coefficient
when the dimensionality and hence the proportionality constant are clear. When
the coefficients’ sum is zero, the plane is an sp represented by ideal points. All
the superplanes are on the line u, so what is the angle between the four different
superplanes we generated? To answer this, it is convenient to introduce and work
with vector notation. Letting x = (x1 , x2 , x3 ), the coefficients of a plane π j are
j j j
cj = (c1 , c2 , c3 ), so the plane’s equation is π j : cj · x = c0 , where “·” stands for
j

the inner (or dot) product. From analytic geometry, the angle φ between two planes
π 1 , π 2 is found by

c1 · c2
cos φ = ± 1
. (5.29)
[(c1 · c1 )(c2 · c2 )] 2

Fixing the angle by adopting the + sign above, we find that π1s
is obtained from
π1s by a clockwise rotation of 120◦ about u. Checking for the third superplane

π1s
2
3 : −2x1 + x2 + x3 = 0 (5.30)
134 5. Planes, p-Flats, and Hyperplanes

Figure 5.18. The distances between adjacent points are proportional to the coefficients.
For π : c1 x1 + c2 x2 + c3 x3 = c0 with the normalization c1 + c2 + c3 = 1, the proportionality
constant is the dimensionality of the space. The plane’s equation can be read from the picture!

resulting from the translation of the X̄2 -axis, the dihedral angle with π1s
is also
120◦ . Not surprisingly, the fourth superplane π1s
2
3
, generated by the translation
of the X̄3 -axis, coincides with π1s .
To clarify, the rotations about u and the superplanes they generate do not affect
the plane π, which remains stationary, intersecting u at H given in (5.13) and shown
in Fig. 5.14. This is the reason that all points π̄123 , π̄1
23 , π̄1
2
3 , π̄1
2
3
have the
same y-coordinate c0 /(c1 + c2 + c3 ) on the horizontal line H̄ . This useful property
is also true for the points representing lines and is put to good use in the construction
algorithms discussed next. For consistency with this chapter’s format the equations
of a line are rewritten as

1,2 : c11 x1 + c12 x2 = c10 ,
: (5.31)
2,3 : c22 x2 + c23 x3 = c20 ,

emphasizing that each of the 1,2 , 2,3 is a projecting plane in R3 orthogonal to

the x1 x2 , x2 x3 planes respectively with c13 = c21 = 0. It is clear, then, that the
points ¯12 , ¯1
2 , ¯1
2
all have the same y-coordinate and are on a horizontal line;
let us call it 12. Similarly, the triple ¯23 , ¯2
3 , ¯2
3
are on a horizontal line 23, and
 5.2 ♣ FT-3 Construction Algorithms 135

¯13 , ¯1
3 , ¯1
3
are on another horizontal line 13. All this is seen in Fig. 5.25 of the
next section.

Exercises
1. Derive the equation equivalent to (5.8) for RN and state its properties.
2. What planes are represented when the points π̄123 and π̄231
are ideal?
3. How are the planes in class E represented?
4. Is the line intersection shown in Fig. 5.10 a necessary and sufficient condition
for coplanarity? Prove or give a counterexample.
5. Show that in RN , the distance between adjacent indexed points is N times the
corresponding coefficient, as in R3 it is 3; equations (5.23), (5.25), and (5.27).
6. (a) Find the equations of the three superplanes containing the x1 -, x2 - and
x3 -axes respectively.
(b) What are the three dihedral angles between pairs of these superplanes?
7. Provide an affine transformation resulting in the four points π̄123 , π̄231
, π̄31
2
,
π̄1
2
3
being collinear and with distance ci between pairs rather than 3ci .
8. The vertical line representation of planes actually requires four rather than the
two vertical lines shown. Explore the case in which none of the four vertical
lines coincides with the coordinate axis X̄i .
9. Provide algorithms based on the indexed-points representation for finding the
intersection between:
(a) a pair of 2-flats, and
(b) a line and a 2-flat.
Delimit carefully the special cases in which the algorithms fail.
10. Given the representation of a plane in terms of vertical lines,
(a) how would one check whether the vertical lines represent two orthogonal
families of parallel lines?
(b) How could the vertical axes be transformed to vertical axes representing
orthogonal families of parallel lines?
11. Construct a vertical line representation Ȳi for a plane π ⊂ R3 with π̄123 ∈ Ȳ1 ,
and π̄231
∈ Ȳ2 .
12. Given a point P ∈ R3 and a plane π ⊂ R 3 , provide an algorithm for deter-
mining whether P ∈ π or in which of the half spaces (“sides”) of π (partitions
R3 ) P lies.
13. Generalize the 3-point-collinearity property for N ≥ 4.
♣ FT-2e
136 5. Planes, p-Flats, and Hyperplanes

5.3 ♣ FT-3 Construction Algorithms

A dream mentioned at the outset is to do multidimensional synthetic constructions

with this new coordinate system as has already been done for lines and now with
planes and flats starting in R3 . Also, this is an additional opportunity to better under-
stand the properties of the indexed points by indulging in the easy constructions
they enable.

5.3.1 Planes and Lines

Half Spaces
With reference to Fig. 5.14, a plane ρ parallel
to π intersects each of the two superplanes
π1s , π1s
at lines ρ ,
ρ parallel to π and π

respectively. Hence the points π̄123 , ρ̄123 are

on a vertical line, and similarly, the points
ρ̄231
, π̄231
are also on a vertical line. Of course,
this is entirely analogous to the conditions for
parallel lines, as it should be, for after all,
the triply indexed points represent lines (i.e.,
π , ρ , and
π ,
ρ ). From the representation
of two parallel planes π, ρ, we agree that the
H̄ (H ∈ u) with the higher y-coordinate iden-
tifies the higher (above) plane. This clarifies
how to distinguish and recognize half spaces
as illustrated in Fig. 5.19, which is the corne-
stone for the study of convexity. The two vertical
lines together with H̄ also provide a coordinate
system of planar coordinates as described in Figure 5.19. The parallel
Section 5.1.1 based on the intersecting lines π planes ρ, τ are above and below
and π
. Usually there is no reason to distin- respectively the plane π .
guish between index permutations like 231
and The upper half-space of π is
1
23, which are henceforth considered the same marked by the two dashed half
unless indicated otherwise. lines.

Line Contained in a Plane

Recognition that a line is contained in a plane π is immediate from the construction
of the points π̄123 . Specifically, a line ⊂ π if and only if ¯12 , ¯23 , π̄123 are
 5.3 ♣ FT-3 Construction Algorithms 137

Figure 5.20. A line is contained in a plane π ⇔ the points ¯12 , ¯13 , ¯23 , π̄123 are on a line P̄ .
Alternatively if ¯1
2 , ¯1
3 , ¯23 , π̄231
are on a line P¯
, then P = π ∩ π and P
= π
∩ π .

collinear. This property is the equivalent of the collinearity in Fig. 5.5 for the two-
vertical-lines planar representation. As illustrated in Fig. 5.20, the line ⊂ π
intersects the superplanes at the two points P = ∩ π , P
= ∩
π with P̄ , P¯

on π̄123 , π̄231
respectively, since these triply indexed points are actually ¯π and
¯
π . Hence ¯23 = P̄ ∩ P¯
, since the line P¯
represents P
on the X̄2 -, X̄3 -, X̄1
-axes.
By transferring the value of its first coordinate p1
to the X̄1 -axis, we obtain the
P¯12
portion of the polygonal line P¯
in the X̄ -, X̄ -, X̄ -axes with ¯ = P̄ ∩ P¯
.
1 2 3 12 12
Therefore

⊂ π ⇔ ¯12 , ¯13 , ¯23 , π̄123 ∈ P̄, P = π ∩ π. (5.32)

The proof for the ⇒ direction is similar. Such a collinearity must also hold for the line
P¯
on ¯23 , π̄1
23 for the X̄2 -, X̄3 -, X̄1
-axes; that is, ⊂ π ⇔ ¯1
2 , ¯1
3 , ¯23 , π̄1
23
are on a line P¯
where P
=
π ∩ π. This leads to a beautiful rotation ↔ translation
duality, shown in Fig. 5.38 and presented shortly, analogous to that obtained with
the vertical lines planar representation seen in Fig. 5.6.
138 5. Planes, p-Flats, and Hyperplanes

Figure 5.21. Two planes π and ρ intersecting at a line .

Intersecting Planes
To realize our goal of doing higher-dimensional constructions, that is, for N > 3,
we gradually wean the reader from the need to use 3-D pictures. In the process we
discover that many constructions even for 3-D are simpler to do and certainly to
draw in -coords. Such is the construction shown in Fig. 5.21 for the intersection
of two planes ρ ∩ π = . Finding the intersection π ∩ ρ ∩ π1s = P is trivial, for
P̄ is the line on the points π̄123 , ρ̄123 . Similarly for π ∩ ρ ∩ π1s
= P
, P̄ is the
line on the points π̄1
23 , ρ̄1
23 . Therefore P , P
∈ = π ∩ ρ and ¯23 = P̄ ∩ P¯
.
Transferring the p1
coordinate of P
to the X̄1 -axis provides the 12 portion of P

and its intersection with P̄ is the point ¯12 . The points ¯12 , ¯23 are both on π̄123
and ρ̄123 , confirming that (5.32) the line is on both planes, simplicity itself!

5.3.2 The Four Indexed Points

The laborious construction for the four points π̄123 , π̄231
, π̄31
2
, π̄1
2
3
seen in
Figure 5.17 can be greatly simplified. A 2-flat is determined by any three points
it contains, and this is the basis in the ensuing construction algorithm. On each of
the points π̄123 , π̄231
, any two lines P̄ and P¯
respectively are chosen as shown
in Fig. 5.22. We know that P ∈ π1s ∩ π, P
∈ π1s
∩ π, and ¯23 = P̄ ∩ P¯
,
where ⊂ π is the line on P and P
. The p1
value (i.e., P̄
∩ X̄i
) is transferred
 5.3 ♣ FT-3 Construction Algorithms 139

Figure 5.22. Representation of a plane π ⊂ R3 by two indexed points.

The first step in the construction of the points π̄32
1
, π̄1
2
3
from π̄123 , π̄231
. A (any) line ⊂ π is
constructed as in Fig. 5.20. The points ¯12 , ¯13 are constructed and the horizontal lines 12, 13, 23 are
¯ with the corresponding indices.
drawn on the ’s

from the X̄1

to the X̄1 -axis and joined to the p2
coordinate intersecting P̄ at ¯12 ;
the point ¯13 is similarly obtained. Horizontal lines 12, 13, 23 are drawn through
the three ¯ points as shown, providing at their intersections with the line P¯
the
points ¯1
2 , ¯1
3 specifying on the X̄2 -, X̄3 -, X̄1
-axes. Moving on to Fig. 5.23,
the value p2
is transferred to the X̄2
axis and joined to the p1
coordinate on X̄1

intersecting the 12 line at the point ¯1

2
, which together with ¯1
3 specifies on
the X̄3 , X̄1
, X̄2
axes. Therefore, as for P and P
, the line P̄

on these two points

represents the point P

∈ π1s

∩π =

π , where π1s

is the superplane corresponding

to this axes spacing (i.e., d1
= 3, d2
= 4, d3 = 2). Hence π̄123 = P̄

∩ H̄ and
¯23 = P̄

∩ 23. Proceeding as in Fig. 5.24, we place a line P̄

parallel to P̄ on
the point ¯1
2
. In fact, P

is the image of P in the X̄1

, X̄2
, X̄3
axes, where, as
we saw earlier, the superplane π1s
2
3
is the image of π123 s . The intersections of

P¯

with H̄ , 13, 23 complete the construction. The four points are referred to as
the augmented representation of a plane. Use of the relations in (8.60) between
the x-coordinates of the four indexed points simplifies the construction. By the
way, Fig. 5.25, showing the ¯ points corresponding to a line in the translated
-coords systems, is obtained with a similar construction. The intersections of
140 5. Planes, p-Flats, and Hyperplanes

Figure 5.23. Construction of the third point π̄1

2
3 .

Figure 5.24. Construction of the fourth point π̄1

2
3
.
Note that P̄

on ¯1
2
is parallel to P̄ , since P

is the point P = π ∩ π on the X̄1

, X̄2
, X̄3
axes.
 5.3 ♣ FT-3 Construction Algorithms 141

Figure 5.25. The collinear triples of points .¯

They represent a line in the four coordinate systems X̄1 X̄2 X̄3 , X̄2 X̄3 X̄1
, X̄3 X̄1
X̄2
, and X̄1
X̄2
X̄3
.

with the superplanes are the points P = ∩ π1s, P

= ∩ π1s
, P

= ∩ π1s
2
,
P

= ∩ π1s
2
3
. The further simplification of the construction via

x2
= 6 − (x0
+ x1
), x3
= 3 + x0
, (5.33)

obtained earlier, where x2

, x3
are the x-coordinates of π1s
2
, π1s
2
3
, respectively,
is left as an exercise.
♣ FT-3e

5.3.3 Special Planes

For future reference, we record the indexed points of some planes that are
encountered on several occasions.

** The Principal 2-Planes

Our first examples are for the x1 x2 plane α : x3 = 0, x2 x3 plane β : x1 = 0, and the
x1 x3 plane γ : x2 = 0. To obtain further practice in computing, the indexed points
142 5. Planes, p-Flats, and Hyperplanes

of α are obtained from the fundamentals (i.e., as indexed points of lines) rather than
from (5.16), (5.22), (5.24), (5.26). It is preferable to use the general relation
: [a1 , a2 , a3 ] −→ ¯ : (da2 , −a3 , a1 + a2 ) ,
rather than eq. (5.20) to obtain
⎧   ⎫
⎨ α12 : x2 = 21 x1  ¯α12 = 1 1 , 0, 1 ⎪ ⎬
 1− 2
s
α = α ∩ π1 : ⇒ = (2, 0, 1) = ᾱ123 .
⎩  ⎪
α23 : x3 = 0  ⎭
¯α23 = (1 + 1, 0, 1)
Proceeding as for (5.21), substituting the interaxis distance d = −2 between X̄1

and X̄2 , and a translation by 3 from the y-axis yields

⎧
  −2 ⎫
⎨ α1
2 : x2 = −x1  ¯
α1
2 = 1+1 + 3, 0, 1 ⎬

α = α ∩ π1s
: ⇒
⎩
 ⎭
α23 : x3 = 0  ¯
α23 = (1 + 1, 0, 1)
= (2, 0, 1) = α231
¯
.
For the third point, using the superplane given in (5.30) yields
⎧

  1  ⎫
⎨ α1
2
: x2 = 2x1  ¯

α1
2
= 1−2 + 3, 0, 1 ⎬

α = α ∩ π1s
2
: ⇒
⎩

 ⎭
α2
3 : x3 = 0  ¯

α2
3 = (−2 + 4, 0, 1)
= (2, 0, 1) = ᾱ31
2
.
Hence the first three points are congruent, as they should be due to the zero
coefficients of x1 and x2 in α’s equation. As has already been pointed out, the
fourth superplane π1s
2
3
coincides with the first π1s , though for the computation of
the corresponding indexed point, the values based on the X̄1
, X̄2
, X̄3
coordinate
system are used, yielding
⎧

   ⎫
⎨ α1
2
: x2 = 21 x1  ¯

α1
2
= 1 1 + 3, 0, 1 ⎪ ⎬
1− 2

= α ∩ π s
: ⇒
α

123 ⎩
 ⎪

α2
3
: x3 = 0  ¯

α2
3
= (1 + 4, 0, 1)
⎭

= (5, 0, 1) = ᾱ1
2
3
.
This point’s distance of 3 from the others is, of course, due to the only nonzero
coefficient in α’s equation. The representation is shown in Fig. 5.26 (left).
Alternatively, the location of the indexed points can be found graphically, by the
intersection of two lines representing two points in the plane. The representations
of β and γ are shown in Fig. 5.26 (right) and Fig. 5.27 respectively.
 5.3 ♣ FT-3 Construction Algorithms 143

Figure 5.26. Indexed points corresponding to α : x3 = 0 for the x1 x2 principal 2-plane on the left
and for the x2 x3 principal 2-plane β : x1 = 0 on the right.

Figure 5.27. Indexed points corresponding to γ : x2 = 0, the principal 2-plane x1 x3 .

** The Constant Planes

Next are the more general planes κ : xi = k0 , i = 1, 2, 3, with k0 a constant. For

κ : x1 = k0 (5.34)

via (5.16), (5.22), (5.24), (5.26), we obtain

κ̄123 = (0, k0 , 1), κ̄231

= κ̄31
2
= κ̄1
2
3
= (3, k0 , 1), (5.35)

providing the representation shown in Fig. 5.28 (left). For

κ : x2 = k0 , (5.36)
144 5. Planes, p-Flats, and Hyperplanes

the indexed points are

κ̄123 = κ̄231
= (1, k0 , 1), κ̄31
2
= κ̄1
2
3
= (4, k0 , 1), (5.37)

for the representation in Fig. 5.28 (right). Continuing,

κ : x2 = k0 (5.38)

has the indexed points

κ̄123 = κ̄231
= κ̄31
2
= (2, k0 , 1), κ̄1
2
3
= (5, k0 , 1), (5.39)

with the representation in Fig. 5.29. Observe that the distance between adjacent
points is three times the value of the corresponding coefficient.

Figure 5.28. Indexed points representing constant planes.

For κ : x1 = k0 (left) and κ : x2 = k0 (right).

Figure 5.29. Representation of the plane κ : x3 = k0 .

 5.3 ♣ FT-3 Construction Algorithms 145

♣ FT-4 Projecting Planes and Lines

An important class of planes

is the projecting planes, which
are perpendicular to one of the
principal 2-planes. Our first
encounter with them was in
Chapter 4, where in RN a
line is defined as the inter-
section of N − 1 projecting
planes. In general, the equa-
tion lj for the pairwise lin-
ear relation between xl and
xj describes a projecting plane
perpendicular to the princi-
pal 2-plane xl xj . In Fig. 5.30
we see two projecting planes
in R3 intersecting at a line
. Where it is important to
avoid an inconsistency, the
Figure 5.30. A line as the intersection of the
projecting planes on α, β, and
projecting planes α ⊥ α and β ⊥ β.
γ of a line are denoted
by α, β instead of 12 , 23
respectively. Figure 5.31 is in fact Fig. 5.25 reincarnated in terms of the projecting
planes of whose representation is directly discerned from the ’s. ¯ For example,
α 31
2
= α 1
2
3
, and these also coincide with 1
2
, since α’s x3 coefficient is
zero.

5.3.4 Intersecting a Plane with a Line

The description of a line in terms of projecting planes motivates an easy and
intuitive algorithm, shown in Fig. 5.32, for finding its intersection with a plane
π . The line r = α ∩ π is found, and then R = r ∩ = π ∩ . This is
how it works. The algorithm’s input is the initial data, consisting of the point
pairs π̄123 , π̄231
, and ¯12 , ¯23 , specifying the plane π and line respectively. The
point ¯1
2 is easily constructed as in Fig. 5.22 from the coordinates x1 , x2 of any
point on and transferring the x1 value to the X̄1
axis. The formality of writing
α 123 = ¯12 and α 231
= ¯1
2 clarifies the use of the planes intersection con-
struction in Section 5.3.1 to obtain r = π ∩ α, actually r̄23 , as in Fig. 5.3.3 with
146 5. Planes, p-Flats, and Hyperplanes

Figure 5.31. The projecting planes α, β, γ .

This is the line whose ¯ points are shown in Fig. 5.25.

r̄12 = ¯12 , since r ⊂ α. Then R = ∩ r = ∩ π. In effect, the system of linear

equations
⎧
⎪
⎨π : c1 x1 + c2 x2 + c3 x3 = c0
1,2 : c11 x1 + c12 x2 = c10 (5.40)
⎪
⎩
2,3 : c22 x2 + c23 x3 = c20
is solved geometrically. The construction is simple, and is easier to draw than in
Cartesian coordinates. Also, it generalizes nicely to N dimensions.
♣ FT-4e

5.3.5 Points and Planes: Separation in R3

Let us return to the matter of half spaces, Section 5.3.1, now considering a plane
π together with its normal vector n partitioning R3 into two oriented half-spaces.
 5.3 ♣ FT-3 Construction Algorithms 147

Figure 5.32. Finding ∩ π using the projecting plane α.

Figure 5.33. Intersection of a line with a plane.

On the left are the initial data, namely points specifying the plane π and line , for the intersection
construction on the right. First r = π ∩ α is constructed (we need only r̄23 , since r̄12 = ¯12 ). Then
R = (r1 , r2 , r3 ) = r ∩ = π ∩ .
148 5. Planes, p-Flats, and Hyperplanes

Figure 5.34. A point P on a plane π. The plane π

is constructed parallel to π and on a point P

which is below π .

Recall that a point P is “above” the plane π if it is on a plane πP parallel to

and above π as shown in Fig. 5.19. The normal vector n points in the “above”
direction.
The construction above instigates a “naı̈ve” algorithm, benefiting from the
special properties of -coords, for determining the relation of a point and a plane
(i.e., on, above or below) rather than laboriously imitating the situation in Cartesian
coordinates (as in the next section). The idea is simple and direct: given a plane π and
a point P , take any line on P and intersect it with π. Then P ∈ π ⇔ P = ∩ π.
This is shown in Fig. 5.34. A plane perpendicular to the x1 x2 plane is chosen
specifying the intersection of line 12 with π , 23 . The construction reveals that
P = ∩ π. Taking another point P
∈ π with the same x1 x2 coordinates as P ,
the plane π
parallel to π and on P
is easily constructed. The point ¯
23 is placed
at the intersection of P̄
and, together with ¯12 , specifies a line
parallel to .
 5.3 ♣ FT-3 Construction Algorithms 149

Joining ¯12 to ¯
23 and extending until it intersects the 0
vertical line locates the
point π̄0

, and similarly the point π̄1

is found specifying π
—really simple! From
the -coords display it is immediate that P
is below. One really has to struggle in
order to understand that from the Cartesian coordinates, and a lot of care is needed
to show this in that picture.
In Linear Programming the object is to move a plane parallel to another plane
π (objective function) on the highest vertex of a convex polytope CP (constraints).
It is worth trying to adapt this construction for many points to discover this top
vertex.

5.3.6 ** Separation in R 3 : An Old Approach

For26 many applications it is useful to introduce orientation and discuss oriented
half spaces, as shown in Fig. 8.11 (left) with respect to lines. A point P : (p1 , p2 )
is on, above, or below the line : x2 = mx1 + b if the expression (p2 − p1 m) is
equal to, greater than or less than b, as shown on the right-hand part of Fig. 8.11.
Since we are working in the projective plane P2 , we consider any regular (i.e.,
Euclidean) point as being below the ideal line ∞ . This situation is more “fun” in
-coords, for there the separation criterion “flips” at m = 1. Correspondingly, P̄
is the line y = (p2 − p1 )x + p1 , where to simplify matters we set d = 1. This is
due to the nonorientability of P2 as already mentioned in Chapter 2 on geometry.
For horizontal lines, “+” is the positive y direction, and for vertical lines, the
positive x direction. The various cases for = ∞ are summarized in the following
lemma:
Lemma 5.3.1. P is on, below, above a line whose slope m < 1(m ≥ 1) ⇐⇒ P̄
¯
is on, below (above), above (below) .
In R3 , for a plane π and a point P , for any plane ρ not parallel to π with P ∈ ρ,
the intersection = π ∩ ρ is found. Within the plane ρ, Lemma 5.3.1 is applied
to determine whether P is on, below, or above π ⇔ it is on, below, or above the
line . The construction for this determination is shown in Fig. 5.36, where a point
T = (t1 , t2 , t3 ) and plane π are shown. The constant plane κ : x1 = t1 , which
contains T and = κ ∩ π, is chosen as well as the point P = (t1 , t2 , p3 ) ∈ (and
hence in π). Viewing the picture in -coords, is found as the intersection of the
two planes κ, π (by the construction in Fig. 5.21). Actually, only ¯23 is needed,
since ¯12 coincides with κ̄123 , so that κ is ’s projecting plane. The T̄12 and P̄12
portions of T̄ , P̄ coincide, and ¯23 is below T̄23 and between the X̄2 and X̄3 axes,
so the slope of 23 is negative, so by Lemma 5.3.1, T is above π .
26
This section is only left for comparing the new and old approaches for solving this problem
150 5. Planes, p-Flats, and Hyperplanes

Figure 5.35. Oriented half spaces and orientation in -coords.

Left: Orienting half spaces on each side of a line : ax+by = c for c ≥ 0. Points in “−

: ax+by < c
are “below” and “above” with the reverse inequality. Right: In -coords the above–below relation
“flips” for lines with m = 1 due to the nonorientability of P2 .

Figure 5.36. Point T = (t1 , t2 , t3 ) is above the plane π .

Line = π ∩ κ where κ : x1 = t1 with P = (t1 , t2 , p3 ) ∈ ∩ π. With ¯23 between the X̄2 , X̄1 , i.e.,
the slope of 23 is negative, and below the portion T̄23 of the polygonal line T̄ , T is above in the
plane κ and also π. This is also clear from the picture since T , P ∈ κ have the same x1 , x2 coords
and p3 < t3 .
 5.3 ♣ FT-3 Construction Algorithms 151

Figure 5.37. Point T = (t1 , t2 , t3 ) is above the plane π .

Line = π ∩ κ, where κ : x1 = t1 with P = (t1 , t2 , p3 ) ∈ ∩ π .

While we are at it, we might as well find the plane ρ parallel to π and containing
the point T . Through T the line r parallel to is determined (see Fig. 5.37) by the
point r̄23 at the intersection of the vertical line on ¯23 and T̄23 . The line R̄ on r̄23
and r̄12 = ¯12 represents the point R = r ∩ π1s by Theorem 5.2.1 and is on the
sought-after plane ρ, since r ⊂ ρ. The point ρ̄123 must be on R̄ at the intersection
of the vertical line through π̄123 , since ρ is to be parallel to π. All the rest is now
determined; H̄ρ is the horizontal line through ρ̄123 . It intersects the vertical line
through π̄231
at ρ̄231
.
With -coords, then, we have the means to do easy synthetic constructions,
such as those we enjoyed doing in elementary geometry, the lower-dimensional
ones honing our intuition for their multidimensional counterparts, which we shortly
meet.

5.3.7 ♣ FT-5 Rotation of a Plane about a Line and the Dual Translation
The analogue to the translation ↔ rotation duality in Fig. 5.6, based on the index-
point representation, brings into play many of the constructions we just learned.
Starting with augmented representation of a 2-flat with four rather than two points,
as derived in Figs. 5.22, 5.23 and 5.24, the intent is to show the rotation of a plane
π 2 : c1 x1 + c2 x2 + c3 x3 = c0 about a line π 1 . With only one line and plane
152 5. Planes, p-Flats, and Hyperplanes

Figure 5.38. Rotation of a 2-flat (plane) about a 1-flat (line) in R3 .

This corresponds to a translation of the points with three indices on the horizontal line H̄ along the
lines L̄, L̄
, L̄

, L̄

joining the points with two indices.

appearing, the superscripts denoting the flats’ dimensionality are not needed, for
the flat’s dimensionality can be determined from the number of subscripts. With
reference to Fig. 5.38, a position of π 2 is shown by the points π̄123 , π̄231
, π̄31
2
,
π̄1
2
3
, representing π 2 on the horizontal line H̄ , while the line’s π 1 representing
points are π̄12 , π̄13 , π̄23 on the line L̄, which, as in Fig. 5.25, also appears in terms
of the other triples of axes with L = π 1 ∩ π1s . Similarly, L
, L

, L

are the line’s

intersections with the second, third and fourth superplanes. To clarify,

π̄1
2 π̄1
3 π̄2
3 on line L̄
, π̄12
π̄13
π̄2
3
on line L̄

, π̄12
π̄13
π̄2
3
on line L̄

With π 1 ⊂ π 2 , the picture is that shown in Fig. 5.20, where the point π̄123 is
on the intersection of H̄ with the line L̄, and similarly,

π̄231
= H̄ ∩ L̄
, π̄31
2
= H̄ ∩ L̄

, π̄1
2
3
= H̄ ∩ L̄

. (5.41)

The distance between adjacent points being proportional to the coefficient of the
plane’s equation (see Fig. 5.18), the intersections of the L̄-lines mark the posi-
tions where coefficients are zero. There, the plane π 2 is perpendicular to the
corresponding principal 2-plane. Specifically, at

H̄ ∩ L̄ ∩ L̄
, c1 = 0 and π 2 ⊥ x2 x3 plane,
 5.3 ♣ FT-3 Construction Algorithms 153

H̄ ∩ L̄
∩ L̄

, c2 = 0 and π 2 ⊥ x1 x3 plane, (5.42)

H̄ ∩ L¯

∩ L̄

, c3 = 0 and π 2 ⊥ x1 x2 plane.

Now translate H̄ vertically with the four triply indexed π̄ points moving along the
four lines L̄, . . . , L̄

. The conditions (5.41) hold, so that at any such translated posi-

tion of H̄ , the containment property of Fig. 5.20 holds. Hence, the corresponding
transformation must be the rotation of π 2 about π 1 passing along the way through
the positions specified by (5.42) with all points on π 1 being invariant under the trans-
formation. The variation of the coefficients can be followed throughout the transit.
An important phenomenon occurs when the horizontal line H̄ crosses the inter-
section points of the L̄i lines with i = 0, 1, 2 denoting the number of primes
(
,

). The coefficients ci are nonzero and, in this case, all positive when H̄ is in
the position shown in Fig. 5.39. When H̄ is on the point L̄
∩ L̄

, the π̄231
, π̄31
2

coincide. Hence c2 = 0 (see Fig. 5.18) and the plane π is perpendicular to the x1 x3
plane. Pursuing this closely, in Fig. 5.40, H̄ is just above this point, and in Fig. 5.41,
just below the point, significantly showing that π has flippe . Consider the oriented
plane together with its normal vector N , whose components are the coefficients ci .
We show next that H̄ just traversing all three points L̄i ∩ L̄i+1 corresponds to the
plane making a 180◦ rotation about its axis, resulting in the normal N → −N.

Fun and Games with the ILM

Open ILM2 and click the fift button with the -coords icon.
The plane and line in Cartesian coordinates are on the right.
A picture similar to Fig. 5.38 appears on the left.

For some experimentation:

translate vertically the horizontal line H̄ on the four π̄ points and observe
that the π̄ slide along the L̄, . . . , L̄

lines.
Note the distance between the adjacent π̄
and the coefficients of the plane’s equation at the bottom.

In preparation for the next discussion,

place the line H̄ on each of the line intersections, i.e., L̄ ∩ L̄
, etc.
Note that then the plane is perpendicular to a principal 2-plane.

Change the line and/or plane and repeat.

154 5. Planes, p-Flats, and Hyperplanes

Figure 5.39. Rotation of a plane about a line in Cartesian and -coords.

In this instance, all the coefficients ci of the plane’s equation are less than or equal to zero.

Figure 5.40. Here c2 is small and positive.

The line H̄ being just above the point L̄
∩ L̄

indicates that the plane is nearly perpendicular to the

2-plane x1 x3 .

Let
L̄i : y = mi x + bi , i = 0, 1, 2, 3, (5.43)


(b2 − b1 ) (m2 b1 − m1 b2 )
(xI , yI ) = L̄ ∩ L̄ = − , . (5.44)
(m2 − m1 ) (m2 − m1 )
 5.3 ♣ FT-3 Construction Algorithms 155

Figure 5.41. Here c2 is small and negative.

The line H̄ is just below the point L̄
∩ L̄

. The plane is nearly perpendicular to the principal plane

x1 x3 but has “flipped” compared to the previous figure.

For a value  > 0, with the superscripts + , − denoting values above or below yI
respectively,

y + = yI +  = mi xi+ + bi , y − = yI +  = mi xi− + bi ,

xi+ = xI + /mi , xi− = xI − /mi .

Then

c2+ = k(x2+ − x1+ ) = (m1 − m2 )/m1 m2 ,

c2− = k(x2− − x1− ) = −(m1 − m2 )/m1 m2 ,

⇒ c2+ /c2− = −1, (5.45)

where k is a proportionality constant. It is clear from Fig. 5.39 or from an easy

calculation that c1+ /c1− ≈ 1 ≈ c3+ /c3− for  small. Therefore, as  → 0,

c1+ /c1− → 1, c3+ /c3− → 1, and c2+ /c2− → −1. (5.46)

As H̄ traverses the point L̄

∩ L̄

, the plane π remains perpendicular to the x1 x3

plane, with the x2 component of the normal N flippin its orientation by 180◦ .
With c1 and c3 unchanged, this amounts to the plane π “flipping” with respect to
the principal plane x1 x3 . That is what is seen in Figs. 5.40, 5.41. Of course, a
156 5. Planes, p-Flats, and Hyperplanes

Figure 5.42. Rotation of a plane π 2 about a line π 1 with c1 constant.

similar flip occurs when H̄ traverses either L̄ ∩ L̄

or L̄

∩ L̄

, as stated below;
the superscripts +, − refer to the above/below position of H̄ with respect to the
intersection point.
Corollary 5.3.2 (Flip) For an oriented plane π having normal vector N =
(c1 , c2 , c3 ) rotating about a line , the traversal of the point L̄i ∩ L̄i+1 , i = 0, 1, 2,
by H̄ corresponds to ci+ /ci− → −1, cj+ /cj− → 1, j = i.
This observation is useful in the study of nonorientability, as for the Möbius strip
in Chapter 9, on surfaces.

Specializations and Generalizations

We have seen that the rotation of a plane about an axis is completely determined by
the translation of H̄ with the π̄123 , π̄231
along the two lines L̄, L̄
. Pictorializing the
rotation facilitates its customization to specific cases. For example, what rotation
 5.3 ♣ FT-3 Construction Algorithms 157

leaves the coefficient c1 of the plane’s equation invariant? The geometric require-
ment is for L̄ and L̄
to be parallel, that is, their point of intersection π̄23 is ideal. The
rotation then leaves the first coefficient c1 invariant. The direction of L̄ and hence
of π̄23 is determined by b3 . All of this is seen in Fig. 5.42 with the correspond-
ing translation of H̄ . The π̄123 , π̄231
, π̄31
2
, π̄1
2
3
rolling along L̄, L̄
, L¯

, L¯

respectively show the variation of the other coefficients of the plane’s equation.
The construction starts by choosing L̄ and L̄
in the direction given by b3 , while
the horizontal distance is 3c1 . As in Fig. 5.22, the coordinates of L̄
determine the
points π̄12 , π̄13 and the lines 12, 13. Or conversely, these are obtained from the
line’s equation, either way proceeding as outlined in Section 5.3.2, where only L¯

is not parallel to L̄. The pencil of planes obtained by the vertical translation of H̄
is now

π 2 : (x3 − x2 − b3 ) + k(x2 − m2 x1 − b2 ) = 0. (5.47)

Here it is assumed that the axis of rotation π 1 lies on the plane π 2 (see Exercise 8).
Then for π 2 : c1 x1 + c2 x2 + c3 x3 = c0 ,
1 :x =x +b ,
π23 3 2 3
(5.48)
2 : x = c1 x + c0 −c3 b3 .
π12 2 c2 +c3 1 c2 +c3

The translation of H̄ is not the essential “ingredient,” for there is no verti-

cal movement when the axis of a plane’s rotation is on the point H . For our
purposes this means that the lines L̄, L̄
are on H̄ and the points π̄123 , π̄231

move only horizontally, and there are some interesting subcases. For example,
rotation of a plane π about its π results in π̄123 being stationary (since it is
π ) while π̄231
is translated along H̄ , together, of course, with the other two
points. Similarly, rotation about
π leaves π̄231
stationary. Rotation about a general
axis π 1 ⊂ π 2 on H results in all triply indexed points moving simultaneously
along H̄ .
The rotation of a plane π viewed in -coords as the movement of points opens
up new vistas for generalization. For parallel translations of π , the correspond-
ing movement in -coords is vertical, with the “pencil” of planes so generated
being the family represented in Fig. 5.19. It is useful to design the transforma-
tion of planes defining the corresponding pair of loci of the points π̄123 , π̄231

on the plane. For example, in Chapter 9 it is shown that any developable sur-
face σ is represented by a pair of curves with the same minimum and maximum
y-coordinate, σ being the envelope of a one-parameter family of tangent planes
(Exercise 9). So in RN , a developable hypersurface is represented by N − 1 such
curves.
158 5. Planes, p-Flats, and Hyperplanes

With some care, the composition of transformations can be specified and visual-
ized as the union of the corresponding loci. Consider, for example, the representation
of “wobbling,” as in the motion of a spinning top, as the composition of a rotation
about an axis rotating about its tip while moving along a path in a plane (Exer-
cise 10). The interesting generalizations to RN are an excellent research topic; see,
for example, [15, pp. 403–409], and Sections 2.7 and 2.8.

Exercises
1. Given three noncollinear points P ∈ π1s , P
∈ π1s
, P

∈ π1s
2
, provide an
algorithm that finds the representation of the 2-flat π 2 that contains the three
points. Even though a plane is uniquely specified by three noncollinear points,
in this case the algorithm should take advantage of the fact that three given
points belong to the three principal sp.
2. For the construction of the four indexed points in Section 5.3.2, prove that any
choice of the three lines L̄, L̄
, L̄

gives the same result.

3. Following the representation of a line as the intersection of two projection
planes,

(a) express a point P as the intersection of three constant planes (as

determined by its coordinates), and
(b) for any plane π and point P , determine if P is above or on the plane π .
4. Draw the figures corresponding to Fig. 5.42 but for rotations leaving the second
and separately the third coefficients of the plane’s equation invariant.
5. Find k in (5.47) as a function of the corresponding y-coordinate of H̄ and then
specialize it to the subcases given.
6. Draw the figure corresponding to Fig. 5.38 for an axis of rotation through the
point H of the plane. In this case L̄ and H̄ coincide.
7. Find the angle of rotation of a 2-flat π about a 1-flat π 1 as a function of the
motion of the triply indexed π̄ points. Specialize to the case that π 1 is on the
point H .
8. Draw the figure showing the rotation of a plane π 2 about a line π 1 not on π 2
and include π 2 ∩ π 1 .
9. Describe the transformation of a plane in which the points π̄123 and π̄1
23
each trace two different ellipses having the same minimum and maximum
y-coordinate.
10. Provide the geometric locus for the representation of a plane’s wobbling (last
sentence of this section).
11. Construct the constant-planes representation in higher dimensions.
 5.3 Hyperplanes and p-Flats in RN 159

12. Prove the correctness of the algorithm for

(a) The intersection of two planes.

(b) The intersection of a line with a plane.

This means that you should show that for any input, the algorithm terminates
with the correct output.

5.4 Hyperplanes and p-Flats in RN

5.4.1 The Higher-Dimensional Superplanes
Just as for R3 , we expect the superplanes in RN consisting of points of the form (5.7)
to play a fundamental role. They are denoted by π N s to mark their dimensionality
N, abbreviated to sp, and are on the line

u : x2 = x1 , x3 = x2 , . . . , xi = xi−1 , . . . , xN = xN −1 . (5.49)

The axes spacing d = (d1 , d2 , . . . , di , . . . , dN ) notation is used, designating the

X̄i axes placed at xi = di . Being 2-flats, the sp are described by N − 2 linearly
independent equations, which after some matrix manipulations can be written with
two variables each as
⎧ Ns
⎪
⎪ π123 : (x2 − x1 ) + k1 (x3 − x2 ) = 0,
⎪
⎪
⎪
⎪ Ns : (x − x ) + k (x − x ) = 0,
⎪
⎪ π234 3 2 2 4 3
⎪
⎪
⎪
⎨ ···
π Ns :
⎪
⎪ Ns
πi(i+1)(i+2) : (xi+i − xi ) + ki (xi+2 − xi+1 ) = 0,
⎪
⎪
⎪
⎪
⎪
⎪ ···
⎪
⎪
⎪
⎩ Ns
π(N−2)(N−1)N : (xN−i − xN −i−1 ) + kN −1 (xN − xN −1 ) = 0.
(5.50)

The form of the equations shows that in the 3-dimensional subspace xi−2 xi−1 xi ,
Ns
the π(i−2)(i−1)i are the pencil of 2-flats on the 1-flat (line) xi−1 = xi−2 , xi = xi−1 .
Noting that each such 2-flat contains the points (1, 1, 1), (di−2 , di−1 , di ) enables
the elimination of the parameter ki in the equations π N s . Rewriting in terms of the
axes spacing yields
160 5. Planes, p-Flats, and Hyperplanes

π Ns (d1 , d2 , d3 ) : (d3 − d2 )x1 + (d1 − d3 )x2 + (d2 − d1 )x3 = 0,

π Ns (d2 , d3 , d4 ) : (d4 − d3 )x2 + (d2 − d4 )x3 + (d3 − d2 )x4 = 0,
...
π Ns (di , di+1 , di+2 ) : (di+2 − di+1 )xi + (di − di+2 )xi+1
(5.51)
+(di+3 − di )xi+2 = 0,
...
π Ns (dN−2 , dN−1 , dN ) : (dN − dN −1 )xN −2 + (dN −2 − dN )xN −1
+(dN−1 − dN−2 )xN = 0.

For example, in R4 the sp are given by

⎧
⎨π1234s : (d − d )x + (d − d )x + (d − d )x = 0,
3 2 1 1 3 2 2 1 3
4s
π : (5.52)
⎩π 4s : (d − d )x + (d − d )x + (d − d )x = 0.
234 4 3 2 2 4 3 3 2 4

This is a good time to define a recursive notational convention27 for the axis spacing
resulting from the successive translations of the X̄i to the new positions X̄i
. For
the initial N-axes system X̄1 , X̄2 , . . . , X̄N , we write
i
" #$ %
0
dN = (0, 1, 2, . . . , i − 1, . . . , N − 1),

stating that the first axis X̄1 is placed at x = 0, X̄2 is placed at x = 1, . . . and X̄N at
x = N − 1. After translating the X̄1 axis one unit to the right of X̄N and renaming
it X̄1
, the axis spacing N-tuple is
i
" #$ %
1 0
dN = dN + (N, 0, . . . , 0) = (N, 1, 2, . . . , i − 1, . . . , N − 1).

And after i such successive unit translations with the X̄i axis in position X̄i
, the
axis spacing for i = 0, . . . , N, k = 1, . . . , N is given by
i
" #$ %
i 0
dN = dN + (N, . . . , N, 0, . . . , 0) (5.53)
i
" #$ %
= (N, N + 1, . . . , i − 1 + N, i, . . . , N − 1) = (dik ).
27
I am indebted to Liat Cohen for proposing this notation.
 5.4 Hyperplanes and p-Flats in RN 161

To clarify the indices, i refers to the number of the axis translations, and k is the
position (component) within the vector dN i . By means of the step function


1 i ≥ k,
Si (k) = (5.54)
0 i < k,

the axis spacing after the ith translation is

i
dN = (dik ) = (k − 1 + NSi (k)). (5.55)

When the dimensionality is clear from the context, the subscript N can be omitted.
i spacing, the points π̄ p
For a flat π p expressed in terms of the dN 1
,... ,i
,i+1,... ,N of its
p p
representation are denoted compactly by π̄i
, and it is consistent to write π p = π0
,
that is, π p described in terms of the axis spacing dN 0.

Returning now to the higher-dimensional sp, in R4 for the standard spacing d0 ,

 s
π123 : x1 − 2x2 + x3 = 0,
4s
π0
: (5.56)
s : x − 2x + x = 0.
π234 2 3 4

The axes are translated, as in R3 , to generate different sp corresponding to rotations

about u, which are summarized in Fig. 5.43. First, the axis X̄1 is translated to position

Figure 5.43. The rotations of π 4 s about the line u.

4s , π 4s are
This is a projection on a plane perpendicular to u, and so projections of the 3-flats π123 234
lines.
162 5. Planes, p-Flats, and Hyperplanes

X̄1
one unit to the right of X̄4 with the resulting axis spacing d1 = (4, 1, 2, 3),
yielding
 s
π1
23 : x1 + 2x2 − 3x3 = 0,
π14s
: (5.57)
s
π234 : x2 − 2x3 + x4 = 0.

The angle of rotation between π123 4s and π 4s computed via (5.29) is

√ √ 1 23
cos (− 3/7) = 180 − φ, φ = cos ( 3/7) ≈ 49.1◦ . Note that since
−1 ◦ −1
s remains unchanged, this is not a complete rotation of π 4s about the line u.
π234 1234
Proceeding with the translation of X̄2 to position X̄2
one unit to the right of X̄1

provides the axis spacing d2 = (4, 5, 2, 3) and the sp

 s
π1
2
3 : 3x1 − 2x2 − x3 = 0,
π24s
: (5.58)
π2s
34 : x2 + 2x3 − 3x4 = 0.

The angle between π1s

23 and π1s
2
3 is cos−1 (−1/7) = 2φ, while the angle
◦
2
34 is 180 − φ. With the translation of X̄3 to position X̄3 one
between π234s and π s

3
unit to the right of X̄2 , d = (4, 5, 6, 3) and
 s
π1
2
3
: x1 − 2x2 + x3 = 0,
π34s
: (5.59)
π2s
3
4 : 3x2 − 2x3 − x4 = 0,

returning π1s
2
3
to its original position π123 s while bringing π s
2
3
4 to an angle 2φ
from π2
34 . Again this is not a complete rotation of the whole sp about the line u. The
s

final translation of X̄4 to X̄4

one unit to the right of X̄3 provides d4 = (4, 5, 6, 7)
and
 s
π1
2
3
: x1 − 2x2 + x3 = 0,
4s
π4
: (5.60)
π2s
3
4
: x2 − 2x3 + x4 = 0,

s
which is identical to π1234 . Unlike the case in R3 , here the rotations’ angles are
not all equal, though the sum is a full circle. The “anomaly” suggests that φ(N )
is a function of the dimensionality N with φ(3) = 60◦ and φ(4) ≈ 49.1◦ , and it
N s is the sp for the standard axes
is interesting to investigate further. For RN , π123
spacing d0 = (0, 1, 2, . . . , N − 2, N − 1). The sp, after the translation of X̄1 to
position X̄1
with axis spacing d1 = (N, 1, 2, . . . , N − 2, N − 1), is π123N s . The
equations derived from (5.51) are:
 5.4 Hyperplanes and p-Flats in RN 163

⎧
⎨π Ns : x1 − 2x2 + x3 = 0,
123
(5.61)
⎩π Ns : x + (N − 2)x + (1 − N)x = 0.
1
23 1 2 3

The angle function is

√ 
φ(N) = cos−1 √ 3(N−2) , (5.62)
2 3−3N+N 2

some of whose values are φ(5) ≈ 47.88◦ , φ(6) = 45◦ , and the limN →∞ φ(N ) =
30◦ .
Next let us compute the 1-flat (line) intersection π = π 4s ∩ π where the plane
π ⊂ R4 to find the index-points representation for π, first for a general axis spacing
d = (d1 , d2 , d3 , d4 ) described by
⎧
⎪
⎪ π : c1 x1 + c2 x2 + c3 x3 + c4 x4 = c0 ,
⎪
⎨
4s : (d − d )x + (d − d )x + (d − d )x = 0,
π123 (5.63)
3 2 1 1 3 2 2 1 3
⎪
⎪
⎪
⎩ 4s
π234 : (d4 − d3 )x2 + (d2 − d4 )x3 + (d3 − d2 )x4 = 0.

With the notation


a = c1 (d2 − d1 ) + c3 (d2 − d3 ) + c4 (d2 − d4 ),
(5.64)
b = c2 (d2 − d1 ) + c3 (d3 − d1 ) + c4 (d4 − d1 ),

the line intersection is given by

a (d2 − d1 )
π : x2 = − x1 + c0 , (5.65)
b b
4s , ¯
since π ⊂ π1234 πij = ¯πkr , for distinct indices i, j, k, r ∈ (1, 2, 3, 4), so that
in homogeneous coordinates, we obtain
 
4 
4
¯π12 = d2 − d1 )b + d1 ci , (d2 − d1 )c0 , ci (5.66)
i=1 i=1
 
4 
4
= ci di , c0 , ci .
i=1 i=1

Substituting the axis spacing dN i in the above, the indexed points for the

representation of π are obtained below, where S = 4i=1 ci :
164 5. Planes, p-Flats, and Hyperplanes

⎧
⎪
⎪ π̄0
= π̄1234 = (c2 + 2c3 + 3c4 , c0 , S),
⎪
⎪
⎪
⎪
⎪
⎪ π̄
= π̄1
234 = (4c1 + c2 + 2c3 + 3c4 , c0 , S)
⎨ 1
π̄2
= π̄1
2
34 = (4c1 + 5c2 + 2c3 + 3c4 , c0 , S), (5.67)
⎪
⎪
⎪
⎪
⎪
⎪ π̄3
= π̄1
2
3
4 = (4c1 + 5c2 + 6c3 + 3c4 , c0 , S),
⎪
⎪
⎩ π̄
= π̄



= (4c1 + 5c2 + 6c3 + 7c4 , c0 , S).
4 1234

We wrote this out in full detail to show that again, the distance between adja-
cently indexed points, appearing on the left above in simplified notation, is as
for R3 proportional (equal to the dimension) to the corresponding coefficient.
Specifically,
⎧
⎪
⎪ π̄1
− π̄0
= (4c1 , 0, 0),
⎪
⎪
⎪
⎨ π̄2
− π̄1
= (4c2 , 0, 0),
(5.68)
⎪
⎪ π̄3
− π̄2
= (4c3 , 0, 0),
⎪
⎪
⎪
⎩
π̄4
− π̄3
= (4c4 , 0, 0).
From this the relations analogous to (8.60) are immediately found, i.e.,

x(π̄3
) = 12 − [x(π̄0
) + x(π̄1
) + x(π̄2
)], x(π̄4
) = 4 + x(π̄0
). (5.69)

To reduce the cumbersome notation, the sp equations (5.56), (5.57), (5.58), (5.59),
(5.60) are referred to by π04s
, π14s
, π24s
, π34s
, π44s
respectively. Next we show that
the indexed points can be easily obtained in general and the property in (5.68) holds
for all N.

Exercises
1. State explicitly the equations for the first sp in R5 .
2. Perform the standard translations of the five coordinate axes in R5 to obtain
explicitly the corresponding rotated five sp.
3. Obtain the coordinates of the six points arising in the representation of a 4-flat
in R5 . This is the generalization of (5.67).

5.4.2 Indexed Points in RN

Remarkably, the collinearity property (as in Theorem 5.2.1) generalizes to higher
dimensions, enabling the recursive (on the dimensionality) construction of the
representation of p-flats for 2 ≤ p ≤ N −1. To achieve this, some intermediate steps
are needed. The indexed point corresponding to the axis spacing dN i (i.e., obtained
 5.4 Hyperplanes and p-Flats in RN 165

from the translation of the axes X̄1 , . . . , X̄i , to the positions X̄1
, . . . , X̄i
, see
(9.1), is denoted by π̄i
.
Theorem 5.4.1 (B. Dimsdale) The 1-fla π ∩ πiN
s , where i = 0, 1, . . . , N and

N
π: ck xk = c0 , (5.70)
k=1

is a hyperplane in RN ; an (N − 1)-fla is represented by the point

 

N 
N
π̄i
= dik ck , c0 , ck , (5.71)
k=1 k=1

where the dik are the interaxis distances for the spacing dN i as given in (9.1).
Explicitly, using (9.3),
 

N 
N
π̄i
= (k − 1 + NSi (k))ck , c0 , ck . (5.72)
k=1 k=1

Proof. To determine the 1-flat and obtain its representation, it suffices to find two
points in the intersection π ∩ πiNs

.

Step 1. Points Pj = (p1j , p2j , . . . , pNj ) ∈ πiN

s are such that
∃mj , bj ∈ R  pkj = mj dk + bj . (5.73)

Step 2. So for Pj ∈ π ∩ πiNs

, in addition to (5.73), the coordinates of Pj must also
satisfy (5.70), i.e.,
N
ck (mj dk + bj ) = c0 .
k=1

Step 3. By Step 1, P¯j is the straight line y = mj x +bj in the xy plane, and together
with Step 2 we have
N
ck {mj dk + (y − mj x)} = c0 ,
k=1
or      

N 
N 
N
−mj ck x+ ck y = −mj dk ck + c0 ; (5.74)
k=1 k=1 k=1

this is an alternative form of the equation for the line P¯j .

166 5. Planes, p-Flats, and Hyperplanes

Step 4. Since xy is the projective plane P2 , for any two distinct points P1 , P2 the
corresponding lines P¯1 and P¯2 intersect at
   
N N
(−m1 + m2 ) ck x = (−m1 + m2 ) dk ck
k=1 k=1
   

N 
N
⇒ ck x= dk ck .
k=1 k=1

Substitution in (5.74) provides the remaining coordinate:

 

N
ck y = c0 .
k=1

The point (x, y) is independent of the particular lines P̄1 and P̄2 used, so all
the lines P̄j for the P̄j in Step 2 must intersect at the same (x, y). Converting to
homogeneous coordinates yields (5.71). 
Corollary 5.4.2 (Hyperplane Representation; J. Eickemeyer) The hyper-
plane π given by (5.70) is represented by the N − 1 points π̄i
with N indices
given by (5.72), for i = 0, 1, 2, . . . , (N − 2).
Specifically,
⎧
⎪
⎪ π̄0
= π̄12...N = (c2 + 2c3 + · · · + (k − 1)ck + · · · + (N − 1)cN , c0 , S),
⎪
⎪
⎪
⎪
⎪
⎪π̄1
= (Nc1 + c2 + · · · + (k − 1)ck + · · · + (N − 1)cN , c0 , S),
⎪
⎪
⎪
⎪
⎪
⎪ π̄2
= (Nc1 + (N + 1)c2 + · · · + (k − 1)ck + · · · + (N − 1)cN , c0 , S),
⎪
⎪
⎪
⎪
⎪
⎪ ···
⎪
⎨
π̄i
= (Nc1 + · · · + (N + i − 1)ci + ici+1 + · · · + (N − 1)cN , c0 , S), i ≥ 1,
⎪
⎪
⎪
⎪
⎪
⎪ ···
⎪
⎪
⎪
⎪ π̄(N−2)
= (Nc1 + · · · + (2N − 3)cN −2 + (N − 2)cN −1 + (N − 1)cN , c0 , S),
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ π̄(N−1)
= (Nc1 + · · · + (2N − 3)cN −2 + (2N − 2)cN −1 + (N − 1)cN , c0 , S),
⎪
⎪
⎪
⎩
π̄N
= (Nc1 + · · · + (2N − 3)cN−2 + (2N − 2)cN −1 + (2N − 1)cN , c0 , S),
(5.75)

where S = N k=1 ck . The first N − 1 points suffice for the representation. As was
done for R and R4 , it is useful to generate the two additional points π̄N
3

−1 , π̄N ,
 5.4 Hyperplanes and p-Flats in RN 167

N−1 N
which here, based on the axis spacing dN , dN , with the their x coordinates
(analogous to (5.69)) given by


N −2

x(N−1)
= N(N − 1) − xi
, xN = N + x0
, (5.76)
i=0

with the compact notation xi

= x(π̄i
).
A p-flat in RN is specified by N − p linearly independent linear equations,
which, without loss of generality, can be of the form
⎧ p
⎪
⎪ π12···(p+1) : c11 x1 + · · · + c(p−1)1 xp + cp1 xp+1 = c10 ,
⎪
⎪
⎪
⎪
⎪
⎪
p
π23···(p+2) : c22 x2 + · · · + cp2 xp+1 + c(p+1)2 xp+2 = c20 ,
⎪
⎪
⎪
⎪
⎪
⎪ ···
⎪
⎨
p
π : πp
⎪ j ···(p+j ) : cjj xj + · · · + c(p+j −1)j xp+j −1 + c(p+j )j xp+j = cj 0 ,
⎪
⎪
⎪
⎪
⎪
⎪ ···
⎪
⎪
⎪
⎪ p
⎪
⎪ π(N−p)···N : c(N−p)(N−p) xN −p + · · ·
⎪
⎪
⎩
+c(N−1)(N−p) xN−1 + cN (N −p) xN = c(N −p)0 ,

and is rewritten compactly as

 !
p

p+j
πp : πj ...(p+j ) : cj k xk = cj 0 , j = 1, 2, . . . , (N − p) . (5.77)
k=i

A p-flat π p ⊂ RN is the intersection of N − p hyperplanes, and (5.4.2) is

the analogue of the “adjacent-variable” description for lines in Chapter 4 with
analogous indexing. Unless otherwise specified, a p-flat is described by (5.4.2)
0.
with the standard spacing dN
Theorem 5.4.3 (J. Eickemeyer) A p-fla in RN given by (5.77) is represented
by the (N − p)p points with p + 1 indices
p+1 
p
 
p+1
π̄{j ...(p+j )}
= dik cj k , cj 0 , cj k , (5.78)
i
k=1 k=1

where j = 1, 2, . . . , N −p, i = 1, 2, . . . , p, and the dik are the distances specifie

by the axis spacing dN i .
168 5. Planes, p-Flats, and Hyperplanes

p
Proof. Each πj ···(p+j ) in (5.77) can be considered as a hyperplane in R(p+1) :
xj . . . xj +p+1 , whose representation, according to Corollary 5.4.2, consists of the
p
p points π̄{j ... ,(p+j )}
, i = 1, . . . , p. They are prescribed by the axis spacing
i
i = (d , . . . , d
dN i1 i(p+1) , . . . , N) as Theorem 5.4.1. There are (N −p) hyperplanes
described by (5.77); therefore there are (N − p)p such points altogether. 
♣ FT-6 To clarify, a hyperplane π in R4 π (i.e., 3-flat) can be represented by
the three points

π̄1234 , π̄1
234 , π̄1
2
34 , (5.79)

while for a 2-flat π 2 , p = 2, N = 4, p(N − p) = 4, and it is represented by the

four points

2 2
π123 : π̄123 , π̄12
23 ; 2
π234 2
: π̄234 , π̄22
34 . (5.80)

Similarly in R5 , a hyperplane π is represented by four points, a 3-flat π 3 and a

2-flat π 2 by six points each as, indicated below:
⎧
⎪
⎪ π : π̄12345 , π̄1
2345 , π̄1
2
345 , π̄1
2
3
45 ,
⎪
⎪
⎪
⎪ ⎧
⎪
⎪ ⎨ π3 3 3 3
⎪
⎪ 1234 : π̄1234 , π̄1
234 , π̄1
2
34 ,
⎪
⎪ π3 :
⎪
⎪ ⎩ π3
⎨ 3 3 3
2345 : π̄2345 , π̄2
345 , π̄2
3
45 ,
⎧ (5.81)
⎪
⎪
⎪
⎪ ⎪
⎪
2 : π̄ 2 , π̄ 2 ,
π123 123 1
23
⎪
⎪ ⎪
⎨
⎪
⎪
⎪
⎪ π2 : 2 : π̄ 2 , π̄ 2 ,
π234
⎪
⎪ ⎪ 234 2
34
⎪
⎪ ⎪
⎪
⎩ ⎩ 2 2 , π̄ 2 .
π345 : π̄345 3
45

In many instances it is possible to use simplified notation by just retain-

ing the subscripts so that the three points in (5.79) are referred to by
1234, 1
234, 1
2
34. The dimensionality is one less than the number of indices.
Continuing, the points representing pi 3 in (5.81) are denoted simply by
1234, 1
234, 1
2
34; 2345, 2
345, 2
3
45. Since there two sets of three points with
five different indices altogether, we can conclude that this is a 2-flat in R5 . The
simplified and the more formal notation are used interchangeably as in Section 5.5.
This theorem unifies all previous results for p-flats π p with 0 ≤ p < N .
♣ FT-6e
 5.4 Hyperplanes and p-Flats in RN 169

5.4.3 Collinearity Property

The underpinning of the construction algorithm for the point representation of a
2-flat π 2 ⊂ R3 , as we saw, is the collinearity property. Namely, for any π 1 ⊂ π 2 ,
1 , π̄ 1 , π̄ 1 are collinear with π̄
the points π̄12 13 23 123 . For the generalization to p-flats,
let
p p p
L̄j k = π̄j ···(p+j
k
) • π̄(j +1)···(p+j +1)
k
(5.82)
p
denote the line L̄j k on the indicated two points. The topic of this section is the proof
that π (p−1)1 , π (p−1)2 ⊂ π p ⊂ RN ,
p (p−1)1 (p−1)2
π̄j ···(p+j +1) = L̄j ∩ L̄j . (5.83)

As an example, for j = 1, p = 2, N = 3, our old friend from Section 5.2.2 is

recast as
π 1k 1 1 2 π 11 π 12
L̄1 = π̄12k • π̄23k , k = 1, 2, π̄123 = L̄1 ∩ L̄1 .

The pair (5.82) and (5.83) state the basic recursive construction implied in
the representation mapping stated formally below. The recursion is on the dimen-
sionality, increased by one at each stage, of the flat whose representative points are
constructed. Though the notation may seem intimidating, the idea is straightforward,
and to clarify it we illustrate it for a hyperplane π 3 ⊂ R4 in Figs. 5.44 and 5.46
starting from the four points π 01 , π 02 , π 03 , π 04 with → pointing to the construction
result. Diagrammatically, the sequence of steps is
& ⎫
π 01 , π 02 → π 11 21 ⎪
→ π (π̄123 , π̄1
23 )⎪
2 1 21
⎪
⎪
π 02 , π 03 → π 12 ⎬
& → π 3 (π̄1234
3 , π̄13
234 , π̄13
2
34 ).
⎪
⎪
π 02 , π 03 → π 12 ⎪
→ π22 (π̄234 22
, π̄22
234 ) ⎪
⎭
π ,π → π
03 04 13

(5.84)

For the construction of a regular (i.e., not ideal) flat, the flats π p ⊂ π 3 with
dimensionality 3 > p ≥ 1 in the construction must have a nonempty intersec-
tion. From the polygonal lines representing π 01 , π 02 , π 03 , the two 1-flats π 11 , π 12
11 11 11
are constructed with π 11 ∩ π 12 = π 02 , yielding the points π̄12 , π̄23 , π̄34 , as
shown in Fig. 5.44 (left). The portion of the construction involving x1 , x2 , x3 is
shown in Cartesian coordinates in Fig. 5.45. The three representing points for each
1-flat are joined by two lines to form polygonal lines having three vertices (the
170 5. Planes, p-Flats, and Hyperplanes

Figure 5.44. ♣ FT-7 Recursive construction in R4 .

A pair of points π 01 , π 02 determines a line (1-flat) π 11 represented by the three constructed points
11
π̄i−1,i , i = 2, 3, 4 (left). Another 1-flat π 12 is determined by one of these points π 02 and an additional
point π 03 as represented by the three black points. Since π 11 ∩ π 12 = π 02 , the two 1-flats determine
21 21
a 2-flat π 21 , and two of its representing points π̄123 , π̄234 are seen on the right. They are the
intersections of the two polygonal lines joining the points previously obtained representing the two
1-flats.

21 21
points). From the intersection of these new polygonal lines the points π̄123 , π̄234 ,
3
representing a 2-flat contained on π , are constructed as shown on the right of
Fig. 5.44. Similarly, π 22 is constructed from the three points π 02 , π 03 , π 04 in the
same way.
At any stage, a point representing π̄ r , where the superscript is the flat’s dimen-
sion, is obtained by the intersection of any pair of lines joining points representing
flats of dimension r − 1 contained in π r .
The axes X̄1 , X̄2 are each translated four units to the right, and construction
3 , π̄ 3
proceeds until all three representing points π̄1234 3
1
234 , π̄1
2
34 are obtained. The
first translation is shown in Fig. 5.46 on the right. As a reminder, these points
represent the lines that are the intersections of π 3 with the first superplane, due to
the standard spacing with d1 = 0, d2 = 1, d3 = 2, d4 = 3, followed by the second
sp with d1
= 4 and then the third sp with d2
= 5 in R4 .
Theorem 5.4.4 (Collinearity Construction Algorithm) For any π (p−2) ⊂
(p−2) (p−2) (p−1)
π (p−1) ⊂ RN , the points π̄1...(p−1) , π̄2...(p−1)p , π̄1...(p−1)p are collinear.
 5.4 Hyperplanes and p-Flats in RN 171

Figure 5.45. The construction of the x1 x2 x3 part of the 3-flat π 3 from four points (0-flats).

Proof. Step 1. Let the (p − 1)- and (p − 2)-flats be given by

⎧ p+i−1
⎪ (p−1)
⎨π (p−1) : πi...(p+i−1) : cik xk = c0i , i = 1, . . . , N − p + 1,
k=i
⎪
⎩π (p−2) : π (p−2) p+j −2
j ...(p+j −2) : k=j aj k xk = a0j , j = 1, . . . , N − p + 2.
(5.85)

Step 2. Consider two distinct points Ar = (α1r , . . . , αN

r ), r = 1, 2 ∈ π (p−2) , and
(p−1)
substitute their first p-components in the equation for π12···p in (5.85) to obtain

c11 α1r + · · · + c(p−1)1 α(p−1)

r
+ cp1 αpr = c01 . (5.86)

(p−2) (p−2)
Substitution in the first two equations for π (p−2) , i.e., π12...(p−2) , π2...(p−2)(p−1) ,
yields
⎧ (p−2)
⎨π12...(p−2) : a11 α1r + · · · + a(p−2)1 α(p−2)
r + a(p−1)1 α(p−1)r = a01 ,
⎩ (p−2)
π2...(p−2)(p−1) : a22 α2r + · · · + a(p−1)2 α(p−1)
r + ap2 αpr = a02 ,
172 5. Planes, p-Flats, and Hyperplanes

Figure 5.46. Recursive construction in R4 continued.

Two 2-flats π 21 , previously constructed, and another π 22 represented by the two black points (left),
determine a 3-flat π 3 . Pairs of points representing the same 2-flat are joined, and their intersection
3 . This is one of the three points representing the 3-flat. The “debris” from the
is the point π̄1234
previous constructions, points with fewer than four indices, can now be discarded. A new axis (right)
X̄1
is placed one unit to the right of X̄3 , and the x1 values are transferred to it from the X̄1 axis.
Points are now represented by new polygonal lines between the X̄2 and X̄1
axes, and one of the
11 11 on the new triple of -coordinate axes, is constructed as in the
points π̄41
, representing the 1-flat π

first step. ♣ FT-7e

whose sum is
a11 α1r + (a21 + a22 )α2r + · · · + (a(p−1)1 + a(p−1)2 )α(p−1)
r
+ ap2 αpr

= a01 + a02 . (5.87)

Step 3. Equations (5.86) and (5.87) are the same, since subtracting one from the
other yields
α1r (c11 − a11 ) + α2r {c21 − (a21 + a22 )} + · · · + αpr (cp1 − ap2 )

= c01 − (a01 + a02 ). (5.88)

Letting r = 1, 2 successively in (5.88) provides two equations whose difference is
[α11 − α12 ]b1 + [α21 − α22 ] + b2 + · · · + [αp1 − αp2 ]bp = 0, (5.89)
the bi being the coefficients of (5.88). This is effectively an identity for every pair
of distinct points in π (p−2) . Hence the coefficients and hence the right-hand-side
of (5.88) must vanish:
 5.4 Hyperplanes and p-Flats in RN 173

c01 = (a01 + a02 ), c11 = a11 , cp1 = ap2 , ck1 = (ak1 + ak2 ),

k = 2, . . . , p − 1. (5.90)

Step 4. The homogeneous coordinates of the three points in question as obtained

from Theorem 5.4.3 are
p−1 
(p−2)
 
p−1
π̄1...(p−1) = d1k a1k , a01 , a1k ,
k=1 k=1
 p 
(p−2)
 
p
π̄2...(p−1)p = d2k a2k , a02 , a2k ,
k=2 k=2

and
 p 
(p−1)
 
p
π̄12...(p−1)p = d1k c2k , c01 , c1k .
k=1 k=1

Note that for this portion of the construction, the axis spacing, is d1k = d2k ,
k = 1, 2, . . . , p. Forming the determinant from these homogeneous coordinates
yields
 p−1 p−1 
 
 k=1 d 1k a 1k a 01 k=1 a 1k 
 
  p  p 
 d a a a 
 k=2 2k 2k 02 k=2 2k 
 p p 
 d 1k c2k c 01 c2k

k=1 k=1 .
From Step 3 and the observation on the axis spacing, the last row is the sum of the
first and second rows; thus the value of the determinant is zero, showing that the
three points are collinear. 

(p−2)
Corollary 5.4.5. For any π (p−2) ⊂ π (p−1) ⊂ RN , the points π̄{j ...(p+j −2)}
,
i
(p−2) (p−1)
π̄{(j +1)...(p+j −1)}
, π̄{(j ...(p+j −1)}
are collinear.
i i

The proof is the same, taking proper care to keep track of the indices and the
corresponding axis spacing. The recursive construction is illustrated for a 5-flat in
R6 from Figs. 5.47 through 5.49.

Exercises
1. What points represent the sp in RN ?
174 5. Planes, p-Flats, and Hyperplanes

2. State the coordinates of the five points representing a 5-flat π ⊂ R6 .

3. State the points representing a 4-flat π 4 , 3-flat π 3 , and 2-flat π 2 in R6 as in
(5.81).
4. Prove (5.76) for RN .

5.5 ** Construction Algorithms in R 4

The generalization of the 3-D construction algorithms to 4-D is direct and is an
opportune time to introduce simplified notation, which is used when the context is
clear.

5.5.1 The Five Indexed Points

The first example is the equivalent of the four-indexed-point algorithm in
Section 5.3.2 for 4-D. That is, given three points π̄1234 , π̄1
234 , π̄1
2
34 specifying
a 3-flat π 3 : c1 x1 + c2 x2 + c3 x3 + c4 x4 = c0 in P4 , we wish to construct the
other two, π̄1
2
3
4 and π̄1
2
3
4
. We start by revisiting the last step in the recursive
construction Fig. 5.46 and show it again in Fig. 5.50 (left). This is the stage where
two lines P̄1 , P̄2 each on the 123 and 234 points of a 2-flat determine the first point
π̄1234 representing a 3-flat π 3 , a situation completely analogous to that in Fig. 5.10,
where here the dimensionality of the objects involved is raised by one. Of course,
the lines represent points P1 ∈ π 4 s1 ∩ π 21 and P2 ∈ π 4 s1 ∩ π 22 for two 2-flats
π 21 , π 22 contained in π 3 . As a reminder, containment is seen by the on relation.
For example, line P̄ is on point π̄123 2 ⇔ P ∈ π 2 , and here in particular the point

P is on the line π123 2 = π 2 ∩ π s . As for the 3-D case, the point π̄

1 1234 represents a
line on an sp. Specifically, π̄1234 = π14s ∩ π 3 . From a line P̄1 on the points π̄2341
(which we assume has been previously constructed) and π̄1
231 , the second point
π̄1
234 is determined, as shown on the right of Fig. 5.50.
The algorithm’s input consists of three points π̄1234 3 , π̄ 3 3
1
234 , π̄1
2
34 , shown
in simplified notation 1234, 1
234, 1
2
34 on the left of Fig. 5.51. Four points
determine a hyperplane π 3 ⊂ R4 . On the right, three lines P̄ , P̄
, P̄

are chosen on
the points 1234, 1
234, 1
2
34 respectively. Clearly P ∈ π 3 ∩ π14s , P
∈ π 3 ∩ π14s
,
P

∈ π 3 ∩ π14s

, and these three points determine a 2-flat π 2 ⊂ π 3 , which can be

described by
⎧
⎨π123 2 : a x +a x +a x =a ,
1 1 2 2 3 3 0
2
π : (5.91)
⎩π 2 : b x + b x + b x = b .
234 2 2 3 3 4 4 0
 5.5 ** Construction Algorithms in R4 175

As for a 1-flat in a 2-flat, see Fig. 5.20; also, for a 2-flat in a 3-flat P̄ ∩ P̄
, π̄234 2 is

one of the points representing π 2 . It is denoted by 234 and is on the the horizontal
line 234, so that 2
34 is 234 ∩ P̄

. These two points represent π234 2 given by

(5.91), a 2-flat in the x2 x3 x4 subspace of R4 . The points π̄22

3
4 , π̄22
3
4
, denoted by
2
3
4, 2
3
4
, are needed next and are determined from 234, 2
34 via the construction
algorithm in Section 5.3.2. For the representation of the 2-flat π123 2 as in (5.91), any
2 can be chosen on P̄ , such as the one denoted by 123 in Fig. 5.52. The
point π̄123
intersections of the horizontal line 123 on 123 and P̄
, P̄

determine the points

1
23, 1
2
3. Four points determine a hyperplane π 3 ⊂ R4 , and so far we have used
three. Proceeding as shown in Fig. 5.52, the line P̄

is drawn on the point 2

3
4 and
parallel to P̄ ; the point P

is the point P in the X̄1

, X̄2
, X̄3
, X̄4
axes, just as for
the 3-D case but here P

∈ π 3 ∩ π14s
2
3
4
with π14s
2
3
4
coinciding with π1234
4s . The

intersections of P̄ with the lines 123, 1234 determine the points 1 2 3 and 1
2
3
4.

The construction is completed by drawing the line P̄ iv on the points 2

3
4
, 1
2
3

whose intersection with 1234 provides the the fifth point of the representation of
π 3 . The four independent points P , P
, P

, P iv determine π 3 . The algorithm’s

output is invariant with respect to the choice of points as well as the selection of π 2
(see Exercise 3). Other equivalent determinations for the 3-flat are one point P and
a 2-flat π 2 , or three lines, which are the 3-flat’s intersection with the sp providing
the three indexed points’ representation.
By the way, it is immediately clear that a 2-flat π 2 ⊂ π 3 ⇔ π¯2 123 , π¯2 234 ,
¯
π 3 1234 are collinear and equivalently for the remaining points representing π 3 .
This is the generalization of the ⊂ π 2 condition shown in Fig. 5.20 and of
course a particular instance of the recursion on the dimensionality collinearity
construction.

5.5.2 Intersecting Hyperplanes in R4

Next we provide an algorithm that constructs the intersection of two 3-flats π 31 , π 32 ,
the result being a 2-flat π 2 like the one given in (5.91). The algorithm’s input
is shown on the left of Fig. 5.53, consisting of three points in -coords each
representing the two 3-flats. On the right, the lines P̄ , P̄
, P̄

joining the pairs

1234, 1
234, 1
2
34 represent three points that specify the 2-flat π 2 = π 31 ∩ π 32
whose representation is constructed in the next steps. As seen, the point π̄234 2 =

P̄ ∩ P̄
, 234 also establishes the 234 horizontal line with π̄22
34 = P¯

∩ 234 marked
by 2
34. The points 2
3
4, 2
3
4
and, using the previous algorithm, also 1
2
3
4 are
constructed and for easy reference are marked by dotted circles.
176 5. Planes, p-Flats, and Hyperplanes

The remaining steps are illustrated in Fig. 5.54:

1. Draw the line P̄ iv on the point 2
3
4
parallel to P̄ , as in the 3-D con-
struction, since the point P iv is the point P but in the rotated coordinates
designated by 1
, 2
, 3
, 4
in the simplifed notation. Necessarily, P̄ iv is on the
1
2
3
4
coordinate system and on the two horizontal lines 1234 ¯ i , i = 1, 2.

The point 2 3 4 was constructed previously for this purpose. Alternatively,

any of the two points 1
2
3
4
can be used, but its construction is more
laborious.
2. The line P̄

on the points 2
3
4, 1
2
3
41 is constructed, providing 1
2
3
42 =
P̄

∩ 12342 and 1
2
3
= P̄

∩ P̄ iv , which establishes the 123 horizontal

line.
3. The points 123 = P̄ ∩ 123, 1
23 = P¯
∩ 123, 1
2
3 = P¯

∩ 123 are now

available. The coincidence of the points 123 and 1
23 here is due to the second
coefficient a2 = 0 of π123 2 , see (5.91), and is not true in general.

2 is a 3-flat in R4 parallel to the x axis, and hence its equation has

Note that π123 4
zero coefficient for x4 . The locations of the points 123, 1
23, 1
2
3, 1
2
3
, which
are now available together with the first equation of (5.67) and equation (5.68)
2 coefficients. Similarly, the coefficients of
enable us to obtain the values of the π123
2 , a 3-flat parallel to the x axis (hence zero coefficient of x is
the equation of π234 1 1
zero), are found. The two equations describe explicitly the algorithm’s output π 2 .
Further higher-dimensional constructions are not presented. They offer
wonderful opportunities for exercises (below), projects, and research topics.

Exercises
1. Simplify the above constructions using the relations in (5.69).
2. Given four points each in a successive sp, provide an algorithm to construct the
3-flat containing them.
3. For the algorithm in Section 5.5.1, show that the result is invariant for any choice
of points P ∈ π 3 ∩ π14s , P
∈ π 3 ∩ π14s
, P

∈ π 3 ∩ π14s

P iv ∈ π 3 ∩ π14s
2
3
4 ,
2 .
and π123
4. Generalize the algorithm in Section 5.3.6 for a point P and 3-flat π 3 in R4 .
5. For a given 3-flat π 3 and a 1-flat (line) , provide conditions for containment
⊂ π 3 and ∩ π 3 = ∅. What is ∩ π 3 ?
6. Generalize the algorithm of Section 5.5.1 to R5 and construct the six points
arising in the representation of a 4-flat given the first four points.
7. Generalize the algorithm of Section 5.5.2 to construct the intersection of two
4-flats in R5 .
 5.5 ** Construction Algorithms in R4 177

Figure 5.47. Randomly selected points on a hyperplane in R6 .

Polygonal lines (left) on the X̄1 , . . . , X̄6 axes representing randomly selected points on a 5-flat
1i 1i
π 5 ⊂ R6 . The π̄12 , π̄23 portions of the 1-flats ⊂ π 5 constructed (right) from the polygonal lines
lines. No pattern is evident.

Figure 5.48. The π̄123

2i 2i
, π̄234 (left) points for the 2-flats ⊂ π 5 .
1i 1i 1i 2i 2i
They are constructed from the polygonal lines joining π̄12 , π̄23 , π̄34 . The π̄123 , π̄234 (right) portions
1i 1i 1i
of the 2-flats ⊂ π 5 constructed from the polygonal lines joining π̄12 , π̄23 , π̄34 .

8. Generalize the rotation of a 2-flat in R3 to display in -coords the rotation of

a 3-flat in R4 about an axis (i.e., a 1-flat). If it is not possible to perform this
rotation, show graphically the constraints involved; hard.
178 5. Planes, p-Flats, and Hyperplanes

Figure 5.49. This is it!

4i 4i
On the left are the π̄12345 , π̄23456 of the 4-flats ⊂ π 5 constructed from the polygonal lines joining
3i 3i 3i
π̄1234 , π̄2345 , π̄3456 . This shows that the original points whose representation is in Fig. 5.47 (left)
are on a 5-flat in R6 . The remaining points of the representation are obtained in the same way, and
all seven points of the representation of π 5 are seen on the right. The coefficients of its equation are
equal to six times the distance between sequentially indexed points as in Fig. 5.18 for R3 .

Figure 5.50. Continuing the construction of indexed points for a 3-flat π 3 in 4D.
On the left is the point π̄1234 previously constructed in Fig. 5.46. Using simplified notation, the
construction is continued on the right to obtain the point π̄1
234 , marked by 1
234. The points
1231 , 1
231 , 2341 are points of the representation of a 2-flat π12 contained in π 3 . The lines P̄1 , P̄1
on
1234 and 1
234 share the indices 234 and necessarily intersect at the point 2341 .
 5.6 Detecting Near Coplanarity 179

Figure 5.51. On the left are the initial data for a 3-flat π 3 in R4 .
Three points (right) P ∈ π 3 ∩π14s , P
∈ π 3 ∩π14s , P

∈ π 3 ∩π 4s are chosen. In the x2 x3 x4 subspace

of R4 they determine a 2-flat π234
2 .

Figure 5.52. Completing the construction of the five points representing a hyperplane in R4 .

5.6 Detecting Near Coplanarity

The coplanarity of a set of points S ⊂ π can be visually verified. If the points are
perturbed, staying close to but no longer being on the plane π , can “near coplanarity”
still be detected? Let us formulate this question more specifically by perturbing the
coefficients ci of a plane’s equation by a small amount i . This generates a family
of “proximate” planes forming a surface resembling a “multiply twisted slab.” Now
the experiment is performed by selecting a random set of points from such a twisted
180 5. Planes, p-Flats, and Hyperplanes

Figure 5.53. On the left are the initial data specifying two 3-flats π 31 , π 32 in R4 .
On the right, the lines P̄ , P̄
, P̄

are drawn on the pair of points 1234, 1

234, 1
2
34 respectively,
providing the 234 and 2
34 points for the 2-flat π 2 = π 31 ∩ π 32 . Then the points 2
3
4, 2
3
4
for
2 and 1
2
3
4 , shown within the dotted circles, are constructed.
π234 1

slab, and repeating the construction for the representation of planes. As shown
in Figs. 5.55, 5.56, there is a remarkable resemblance to the coplanarity pattern.
The construction also works for any N . It is also possible to obtain error bounds
measuring the “near coplanarity” [131]. This topic is covered in Chapter 8 on
proximity.
Experiments performing similar constructions with points selected from several
twisted slabs simultaneously showed that it is possible to determine the slabs
from which the points were obtained or conversely can be fitted to. All this has
important and interesting applications (USA patent # 5,631,982). Given a set of
points composed of point clusters each from a different plane (or hyperplane),
the determination of these planes can be made with very high accuracy and low
computational complexity, as shown in the appendix on recent results.

5.7 Representation Mapping, Version II

Let us revisit and update the representation mapping
N 2
J : 2P → 2P × 2[1,2,... ,N] , (5.92)
in view of this chapter’s results. The recursive construction starts with the non-
recursive step for the representation of a point P ∈ RN , a 0-flat π 0 , in terms of N
points each with one index, the values of the coordinates pi , i ∈ [1, . . . , N], of P
on the X̄i axes. For consistency, a point is also specified by the N equations
xi = pi , i = 1, . . . , N. (5.93)
 5.7 Representation Mapping, Version II 181

Figure 5.54. Intersection of two 3-flats in 4-D.

The result is the 2-flat π 2 given by the points 123, 1
2
3
, etc. representing π123
2 , and 234, 2
34, etc.
2
for π234 .

The construction of a p-flat from its points (0-flats) to lines (1-flats) and so on
proceeds according to Theorem 5.4.4, the dimensionality being raised by 1 until it
reaches p, providing the indexed points needed for the representation.
Let Nr be the number of points and nr the number of indices appearing in the
representation of a flat in RN . Then for a p-flat, Nr = (N − p)p, nr = p + 1, so
that

Nr + nr = (N − p)p + (p + 1) = −p 2 + p(N + 1) + 1. (5.94)

In RN some examples are

1. p = 0: points π 0 : Nr + nr = N + 1 ,
2. p = 1: lines π 1 : Nr + nr = (N − 1) + 2 = N + 1 ,
3. p = 2: 2-flats (2-planes) π 2 : Nr + nr = (N − 2)2 + 3 = 2N − 1 ,
4. p = N −1: hyperplanes Nr = N −1, nr = N : Nr +nr = N −1+N = 2N −1.
182 5. Planes, p-Flats, and Hyperplanes

Figure 5.55. Polygonal lines representing a randomly selected set of “nearly” coplanar points.

Figure 5.56. The “near-coplanarity” pattern.

The picture on the left is very similar to that obtained for coplanarity, with the points of intersection
forming two clusters (right).

Note that (5.94) does not include the case for p = 0 (points). In summary, the
representation by points reveals that the object being represented is a p-flat whose
dimensionality is one less than the number of indices used. The dimensionality of
the space where the p-flat resides is, of course, equal to the number of parallel axes,
or it can also be found from the number of points (N − p)p if one knows the value
of p.
 5.7 Representation Mapping, Version II 183

We have seen that the image I (U ) = Ū of a U ⊂ PN consists of several points

each indexed by a subset of [1, 2, . . . , N], the indexing being an essential part of
the representation. The mapping I provides a unique representation, since it is one-
to-one with U¯1 = U¯2 ⇔ U1 = U2 . By representation is meant a minimal subset of
points that uniquely identify the object in question, for as we have seen, there are
many redundant points. The final version of the J is given in the conclusion after
the representation of surfaces.
6

** Envelopes

Note: Reading of this chapter is discretionary. It provides foundational background

for the chapters on curves and surfaces.

6.1 The Basic Idea

Envelopes are encountered in many areas
of mathematics as well as in several appli-
cations. Also, they are useful background
for the construction of representations in
parallel coordinates. In our era, envelopes
are no longer a “fashionable” subject and
are often not covered in coursework. For
good references one needs to look at
older sources. For these reasons, a quick
overview is included here. An extensive
treatise is available in [17], while good Figure 6.1. Envelope is a circle.
shorter sections can be found in [35], [73], The circle C is tangent at each point
and [143]. to a specific line (x, y, a), which is
rotated at a constant distance r from the
First let us build some intuition. Con-
point O.
sider a line (x, y, a) = 0 in the xy plane,
as shown in Fig. 6.1, rotated at a constant
distance r from a point O. Its position is specified by the angle a. This rotation
traces a circle C of radius r and centered at O. At each point of C there is exactly
one line, as specified by a, that is tangent to C at that point. We say that C is the

A. Inselberg, Parallel Coordinates, DOI 10.1007/978-0-387-68628-8_6,

© Springer Science+Business Media, LLC 2009
186 6. ** Envelopes

envelope of the family of curves (x, y, a) = 0.

For another example, consider a rod with length
d constrained in the vertical plane xy sliding
so that it always touches the wall and floor as
shown in Fig. 6.2. Formally, we have a family of
segments d(x, y, a) all having constant length d
whose position can be prescribed by the angle a
between the floor and the segment. These rods
are tangent to a curve called A, which is one of
the four arcs of an astroid given by Figure 6.2. Envelope is one
quarter of an astroid.
x 2/3 + y 2/3 = d. A rod d constrained to slide in
the vertical plane while touch-
This, of course, needs to be shown. ing both the vertical wall and
In each example there is a one-parameter the floor traces (i.e., is always
family, F , of curves F (x, y, a) = 0, each curve tangent to) a curve called an
being specified by a value of the parameter a. astroid.
Also, in each case we saw that there is yet
another curve E that is tangent at each point to a curve in the family F . In the
first case this was the circle C, and in the second, the astroid A. It happened that the
families of curves in both cases were linear in x and y, but this does not detract from
the generality. For example, the astroid, with all four arcs, turns out to be tangent
at each one of its points to a curve of the family

x2 y2
F (x, y, a) = 2 + − 1 = 0,
a (1 − a)2
as shown in Fig. 6.3. These are the ellipses centered at the origin with semiaxes of
length a and a − 1. That is, the sum of the length of the semiaxes is always 1; so
the circle with diameter 1 is included in this family.

6.2 Formulation
A one-parameter family of curves in the xy plane can be described by a function

F (x, y, a) = 0, (6.1)

where a ∈ R. There may exist a curve E that touches28 at each of its points a member
of (6.1). The curve E is called the envelope of the family F . There remains to see
28
In the ensuing it will become clear why the word “touches” rather than “tangent” is used here.
 6.2 Formulation 187

Figure 6.3. Envelope is a full astroid.

The envelope of the family of ellipses with sum of their semiaxes constant.

Figure 6.4. The envelope E and two intersecting neighboring curves of F .

how E , when it exists, can be found. Keeping with the mantra “let intuition guide the
formalism,” we will add the conditions as the need is encountered and explained.
At first we assume that the curves of F are tangent to E , and that neighboring
curves of the family intersect at points “close” to E and lie on the same side. This
is the purpose of the illustration in Fig. 6.4, where the curves F (x, y, a) = 0 and
F (x, y, a + h) = 0 are shown intersecting at P . The idea is to find the point of
intersection P of these two curves and let h → 0. Then, given sufficient smoothness
188 6. ** Envelopes

conditions, P must approach the point of contact of F (x, y, a) = 0 with E , since

at P ,

F (x, y, a) = F (x, y, a + h) = 0,

and therefore
F (x, y, a + h) − F (x, y, a)
= 0.
h
When the function F (x, y, a) = 0 is differentiable with respect to a ∀a ∈ R
and h → 0 above, we obtain that the point of contact with E has coordinates (x, y),
which are found by eliminating the parameter a from the two equations:

F (x, y, a) = 0,
(6.2)
∂F (x, y, a) = 0.
∂a

6.3 Necessary and Sufficient Conditions

Under what conditions does (6.2) yield the envelope, and what other possibilities
may exist? Let us assume that E is parametrizable and can be given in terms of two
continuously differentiable functions

x = x(a), y = y(a), (6.3)

with
 2
dx 2 dy
+ = 0, (6.4)
da da
and is tangent to the curve F (x, y, a) = 0 with the same value of the parameter
a at the point of contact, as shown in Fig. 6.5. The underlying reason why such a
parametrization is possible here is that at each point (x, y) of E there exists an a
specifying F (x, y, a) at that point. Hence it is possible to parametrize E in terms
of a with

F (x(a), y(a), a) = 0 (6.5)

valid at the point of contact. Differentiating (6.5) with respect to a, we obtain

dx dy
Fx + Fy + Fa = 0. (6.6)
da da
 6.3 Necessary and Sufficient Conditions 189

Figure 6.5. Point of contact between the envelope E and a curve in F .

dy
The derivatives dx
da , da are proportional to the direction cosines of the tangent
t of E , while the quantities Fx , Fy are proportional to the direction cosines of the
normal n to the curve F (x, y, a) = 0 of the family. These two vectors t, n must be
orthogonal to each other. Therefore, the inner product of the corresponding direction
cosines satisfies
dx dy
Fx + Fy = 0, (6.7)
da da
so that (6.6) and (6.7) together imply

Fa (x, y, a) = 0.

Hence the conditions of (6.2), which were previously found sufficien , have now
been shown to be necessary for the existence of E .

6.3.1 Singular Points

When

Fx (x, y, a) = Fy (x, y, a) = 0,

the curve F (x, y, a) = 0 of the family may have singular points (where F is not
continuously differentiable there), and no conlusions can be drawn about the contact
of the curves. For algebraic curves (i.e., polynomials), the singular points may be
• points of self-intersection: points where the curve crosses itself,
• points of self-tangency: where the derivative is continuous but the curve has
portions above and below the tangent at each point,
• cusps: where the derivative is not continuous,
• isolated points: points disjoint from the curve.
190 6. ** Envelopes

An algebraic curve can have only a finite number of singular points, and when
it has none, it is called smooth. When the envelope is found using (6.2), E may also
contain singular points.

6.4 Examples: Envelopes of Families of Curves

The following examples are intended to clarify the method and indicate various
possibilities.
1. The family
F (x, y, a) = x 2 + y 2 − 2x cos a − 2y sin a = 0 (6.8)
describes a family of circles that have unit radius and pass through the origin.
This becomes a bit clearer when (6.8) is rewritten as
(x − cos a)2 + (y − sin a)2 = 1,
indicating that the circles are centered on the circumference of the circle
C1 : x = cos a, y = sin a,
which has unit radius and center at the origin, as shown in Fig. 6.6.
To find the envelope, we differentiate to obtain
Fa = 0 #⇒ x sin a − y cos a = 0. (6.9)

Figure 6.6. Family of circles centered on the unit circle C1 .

Note the envelope consiting of the outer circle and the isolated point at the origin.
 6.4 Examples: Envelopes of Families of Curves 191

Figure 6.7. Family of self-tangent curves and their envelope.

The two equations (6.8) and (6.9) are satisfied by x = 0 and y = 0.

Hence (0, 0) ∈ E . Further, when x 2 + y 2 = 0, the solution (do this as an
exercise) is
y x
sin a = , cos a = ,
2 2
and eliminating a yields

x 2 + y 2 = 4.

That is, using (6.2), we found that E consists of

(a) the circle of radius 2 centered at the origin, and
(b) the isolated point at the origin.
2. Now we consider the family of curves

F (x, y, a) = y 2 − (x − a)4 = 0. (6.10)

Each curve in this family consists of a parabola and its reflection about the x
axis, so there is a point of self-tangency at (a, 0). Using (6.2), the envelope
is found to consist of all the points (called the locus) of self-tangency, as
illustrated in Fig. 6.7. This example shows that it is not essential, as assumed
in the formulation, for the curves to be on one side of E , and it suffices that
there be sets of points of intersection of neighboring curves on either side that
enable the limit as h → 0 in the derivation of (6.2) to exist.
3. Applying (6.2) to the family of cubical parabolas

F (x, y, a) = (x − a)2 − y 3 = 0 (6.11)

192 6. ** Envelopes

Figure 6.8. Family of curves with cusps.

gives the x axis, which is the locus of the cusps, as a result. From Fig. 6.8 it is
clear, however, that this is not the envelope, since it is not tangent to the curves
of the family at the point of contact. It is instructive to review the derivation
of (6.2) and find which of the assumptions failed to be true here.
4. The next example illustrates that the points of the envelope need not be the
limit of neighboring intersecting points of the curves in the family. Consider
the family of cubical parabolas

F (x, y, a) = y − (x − a)3 = 0, (6.12)

which are identical except for the translation by a and do not intersect each
other. Applying (6.2) provides the condition Fa = 3(x − a)2 = 0, which in
turn shows that E is the x axis. We accept this as the envelope of the family,
since each F is tangent to it, as shown in Fig. 6.9.
5. For our last example we consider the family of strophoids

F (x, y, a) = [(y − a)2 + x 2 ](x − 2) + x = 0, (6.13)

Figure 6.9. Family of curves with inflection points.

 6.4 Examples: Envelopes of Families of Curves 193

Figure 6.10. Family of self-intersecting curves.

which as Fig. 6.10 shows are self-intersecting (the point of intersection is also
called a double point). Each curve can be translated to any other curve. We find
that
Fa = −2(y − a)(x − 2) = 0 #⇒ x = 2 or y = a.
Since no finite finite value of y corresponds to x = 2, we take y = a, so that
x 2 (x −2)+2 = 0. This suggests that E consists of the two vertical lines x = 0
and x = 1. Yet it is clear from the figure that only x = 0 can be considered as
the envelope, while the line x = 1 is the locus of the self-intersections. Again
it is a good idea to run through the derivation to see which assumptions failed
to be true.

Exercises
1. Prove that one arc of the astroid is the envelope of the falling stick as described
at the beginning of this section.
2. Prove that all four arcs of the astroid are the envelopes of the ellipses whose
semiaxes have a constant sum.
194 6. ** Envelopes

3. Find the envelope of the family of ellipses whose semiaxes are collinear and
whose areas are constant. Hint: Area of ellipse = πab, where a, b are the
lengths of the semiaxes.
In connection with parallel coordinates,
4. Prove the basic point ↔ line duality using envelopes.
5. Derive the vertical-line representation of planes using envelopes.
6. Derive the 2-point representation of planes in R 3 using envelopes.
7. Show using envelopes that in P 2 , conics are mapped into conics. This is an
excellent but challenging project topic.
7

Curves

7.1 ♣ FT-1 Point-Curves and Line-Curves

We revisit the fundamental duality in the plane, Chapter 3, starting with the point-
to-line correspondence

P : (p1 , p2 , p3 ) −→ P̄ : [(p1 − p2 ), dp3 , −dp1 ], (7.1)

where d is the distance between the X̄1 and X̄2 axes and the triples within [· · · ] and
(· · · ) are the line and point homogeneous coordinates respectively. Unless otherwise
stated, P is a regular (i.e., Euclidean) point, so without loss of generality, p3 = 1,
whence

P : (p1 , p2 , 1) −→ P̄ : [(p1 − p2 ), d, −dp1 ]. (7.2)

The line-to-point correspondence is

: [a1 , a2 , a3 ] −→ ¯ : (da2 , −a3 , a1 + a2 ), (7.3)

for : a1 x1 + a2 x2 + a3 = 0. When a2 = 0, the slope of is m = − aa21 , the

intercept is b = − aa23 , and

: [m, −1, b] −→ ¯ : (d, b, 1 − m). (7.4)

The image of each point P on a curve c is a line obtained from (7.1), as shown in
Fig. 7.1 (left), and the envelope of all such lines, when it exists, is the curve c̄, the
image of c. So we need to consider two types of curves: the point-curves we have
known and loved all our lives, and which are made up of points, and line-curves,

A. Inselberg, Parallel Coordinates, DOI 10.1007/978-0-387-68628-8_7,

© Springer Science+Business Media, LLC 2009
196 7. Curves

Figure 7.1. Point c and line curves c̄. (Right) Vertex in r maps into a straight line in r̄.

which are formed by the envelope of a family of lines, usually but not always their
tangents.
The point ↔ line duality in the projective plane P2 causes the duality

point-curve ↔ line-curve.

¯ the point image of the line (curve) of given by (7.3), is the envelope of the
In fact, ,
lines P̄ for P ∈ . In this chapter we learn to construct the representation of curves
in -coords efficiently and study properties of the image curves. In the ensuing, we
consider piecewise differentiable curves r, i.e., r = r1 ∪ r2 ∪ · · · ∪ ri ∪ ri+1 · · · ∪
rn , where the ri are differentiable. The endpoints of r may be ideal. At a vertex
Vi = ri ∩ rr+1 where there are two tangents, r̄ will contain a straight-line segment
V̄i separating the two line-curves r̄1 , r̄i+1 ; see Fig. 7.1 (right).

7.1.1 Separation in the xy Plane

Recall the basic idea of sepa-
rating points by a line. In the
Euclidean x1 x2 plane, a point
P : (p1 , p2 ) is said to be on,
above, or below the line :
x2 = mx1 + b if the expres-
sion (p2 −p1 m) =, <, > b; see
the right-hand part of Fig. 7.2.
This is equivalent to placing
a normal vector n on the line
, providing two half-planes on
either side of the line as well as
a direction (that of the normal). Figure 7.2. The “above” and “below” relations
The above or below relation of between points and lines switch at m = 1.
 7.1 ♣ FT-1 Point-Curves and Line-Curves 197

a point P with respect to is then specified in terms of the half-plane on which P

lies. For a vertical line, say x1 = c, the direction still determines unambiguously
the half-spaces and hence the above/below relation, which would otherwise be
ambiguous.
Alternatively, we can revert to the line (7.3) with a1 = 0 and the convention
that the direction of increasing x1 specifies the “above” half-plane. Since we are
working in the projective plane P2 , we consider any regular (i.e., Euclidean) point
as being below the ideal line ∞ . As indicated on the left-hand side of Fig. 7.2, this
situation is more fun in -coords, for there the separation criterion “flips” at m = 1.
Correspondingly, P̄ is the line y = (p2 − p1 )x + p1 , where to simplify matters
we set d = 1 from now on. The various cases for = ∞ are summarized29 in the
following lemma.
Lemma 7.1.1. 1. P is on, below (above) a line whose slope m < 1 ⇐⇒ P̄ is
¯
on, below (above) .
2. P is on, below (above) a line whose slope m ≥ 1 ⇐⇒ P̄ is on, above
¯
(below) .
Let

Mc (I ) = max{mc (P )|P (x1 ) ∈ c, x1 ∈ I }, (7.5)

where mc (P ) is the slope of the tangent of the curve c : x2 = r(x1 ) at the point
P . Further, by considering at each vertex the supporting line with maximum slope
instead of the tangent in the above definition, Mc (I ) can be defined for a complete
piecewise smooth curve. Clearly, then, tracing the above/below relation in the
image c̄ of a curve c depends on whether Mc (I ) goes through the value 1. There
are some straight forward consequences of the lemma, which for convenience are
listed separately. They are true because the statements are true pointwise.
Corollary 7.1.2. 1. For Mc < 1, c is below (above) a point P ⇐⇒ c̄ is below
(above) P̄ .
2. For Mc ≥ 1, c is below (above) a point P ⇐⇒ c̄ is above (below) P̄ .

Definitio 7.1.3. A curve segment r is convex downward (cd) (also sometimes called
“concave”) if it is on or below any one of its tangents (or more generally, supporting
lines). It is convex upward (cu) if it is on or above any one of its tangents; see Fig. 7.3.
When for the curve r : x2 = r(x1 ), r
and r

exist in the interval of interest,

then where r is cd, r
is monotone increasing; hence r

> 0. Further, where r is cu,

29
Fortunately, consideration of the “slope subcases” is not required for higher dimensions; see
section 5.3.5 in chapter 5 on planes.
198 7. Curves

Figure 7.3. Curves r flip cu ↔ cd for Mr < 1 and do not flip for Mr ≥ 1; Corollary 7.1.4.

r
is monotone decreasing, so r

< 0. If there is a transition from cu to cd (or vice

versa), then necessarily r

= 0 there; more on this shortly.

Corollary 7.1.4. 1. If r is a cd (cu) curve segment with Mr < 1, then r̄ is cu
(cd).
2. If r is a cd (cu) curve segment with Mr ≥ 1, then r̄ is cd (cu).

Proof. Since r̄ is on or above any of its tangents, it follows that r̄ cu, as shown in
Fig. 7.3. Pay attention that even though the above/below relation between , P , , ¯
and P̄ is preserved, the cu/cd property between r and r̄ “flips,” since r is a point-
curve and r̄ is a line-curve. The proof for the remaining cases (two of which are
illustrated in Fig. 7.3) is completely analogous. 
Certain properties are not intrinsic to a curve. Rather, they are incidental to
the curve’s position or orientation with respect to a specific coordinate system. For
example, vertical or horizontal tangents do not persist under a rotation, and for
that matter, cu ↔ cd with appropriate rotations. Properties such as shape (circle,
triangle) that are unaltered by rotations, translations, and scaling are geometrical.
The properties of curves and surfaces varying pointwise are studied in differential
geometry using the techniques of calculus and vector analysis for the more elemen-
tary aspects, and tensor methods for the more advanced. In keeping with our aim of
making this text widely accessible, only the essentials of this subject are employed.
For a more rigorous and thorough background there are excellent references such
 7.2 Duality between Cusps and Inflection Points 199

as [166], [75], [183], [48], with [121] recommended for a most enjoyable overview
interlaced with the historical development of differential geometry (Chapters 23
and 37).

7.2 Duality between Cusps and Inflection Points

The description of inflectio
points (ip) turns out to be a bit
stickier than that given at the ele-
mentary calculus level (see, for
example, [55] and [180]). The
tangent to a curve c touches it
at the point of contact P =
(x1 , r(x1 )). Where r

(x1 ) = 0,
the circle with the largest radius
R(x1 ) having the same tangent
at P as c is found, and it mea-
sures the local curvature = 1/R
of c (though “circularity” would
be a more accurate term). The
circle’s arc is the best “circular” Figure 7.4. Types of points on a curve.
The coordinate system here is the tangent and
approximation of the curve at P .
normal at the point.
Clearly this circle’s center, called
the center of curvature, lies on the
concave side of c. The directed line perpendicular to the tangent at P and pointing
toward the center of curvature is the normal. Directing the tangent so that together
with the normal as axes they form a right-handed orthogonal coordinate system
provides a convenient frame of reference for the study of the curve’s behavior and
characterization of various types of points, described next and illustrated in Fig. 7.4.

1. Regular30 : where the normal but not the tangent crosses the curve. A line
crosses a curve when parts of the curve (consisting of more than a point) lie
on both sides of the curve as in (a) of Fig. 7.4.
2. Inflectio point (ip): where both the tangent and the normal cross the curve as
in (b) of Fig. 7.4.
3. Cusp of the first kind: where the tangent but not the normal crosses the curve
as in (d) of Fig. 7.4.

30
This name here is not to be confused with a regular (i.e., Euclidean) point of the projective plane.
200 7. Curves

4. Cusp of the second kind: where neither the tangent nor the normal crosses the
curve as in (c) of Fig. 7.4.
This by no means exhausts the possibilities. We will not deal with other kinds
of singularities, which have more complicated rigorous descriptions, though some
are easy enough to illustrate. Points that are not regular are called singular. In fact,
we will consider only cusps of the first kind and refer to them simply as cusps.
It is important to note that curves having cusps and inflection points have both
a cu and cd portion in a neighborhood of these points. Strikingly, our point–line
duality induces an ip–cusp duality. This result, which we obtain next with straight
forward geometric argument, is more general than that shown in algebraic geometry
specifically for algebraic (polynomial) curves [61, p. 88].
Theorem 7.2.1 (ip ↔ cusp duality) ♣ FT-2
1. I ip of a point-curve r with tangent i ↔ ī cusp of the line-curve r̄ with tan-
gent I¯.
2. C cusp of a point-curve r with tangent c ↔ c̄ ip of the line-curve r̄ with tan-
gent C̄.
Proof. Step 1. Consider a curve r having an inflection point I with tangent i in the
tangent (x1 )–normal (x2 ) coordinate system shown in Fig. 7.5. Since the slope of
the tangent i at I is zero, ∃ an interval S on the x1 (tangent) axis containing I and
where Mr < 1. It is easy to see that in a neighborhood S of an inflection point the
curve is necessarily monotone (increasing or decreasing). Further, with S, r has a

Figure 7.5. Inflection point (ip) to cusp.

 7.2 Duality between Cusps and Inflection Points 201

cu part ru and a cd part rd with the transition occurring at the inflection point I (see
exercises below).
Step 2. By Step 1, for each a ∈ I ∃a1
also ∈ S such that r
(a) = r
(a1
).
Step 3. By Step 2 and (7.3) for the tangents t, t
to r at the points P = r(a),
P = (a1
) respectively, t¯, t¯
the corresponding points on r̄ occur at the same
x = 1−m 1 , m being the common slope of t and t
, of the xy plane. Hence the
curve r is double-valued.
Step 4. By Step 1 and Corollary 7.1.4, r̄ must consist of a cu and cd portion.
Specifically, since Mr < 1 on I , r̄u is cd and r̄d is cu, and I¯ is the tangent separating
them at ī.
Step 5. Now let us consider the case that the tangent at the point of inflection has
slope other than 0. Specifically, as shown in Fig. 7.6, r has inflection point I1
with tangent i1 having slope less than 1 and not necessarily 0. By exactly the same
argument as above, r̄ has a cusp at ī1 with tangent I¯1 . The portions ru and rd become
cd r¯u and cu r¯d respectively.

Figure 7.6. The duality inflection points ↔ cusps is slope-independent m = 1.

202 7. Curves

Figure 7.7. The image of a cusp is an inflection point.

Step 6. The counterpart for a curve s with Ms > 1 having an inflection point (see
Fig. 7.6) is still the same kind of argument, where now the cu and cd parts stay cu
and cd in the image r̄.

Step 7. The proof of the second assertion for cusps is similar. Again we start with a
point-curve r shown in Fig. 7.7 with the normal–tangent coordinate system, having
a cusp C with slope c. It is clear that the image c̄ equals x ∩ X¯2 . Since C is a cusp,
∃ an interval S on the x1 (tangent) axis with C ∈ S, and on this interval, Mr < 1.
Further, ∀aS∃a
∈ S such that the derivatives r
(a), −r
(a
) are equal, with t and t

being the tangents to r at a and a

respectively. This is due to the double-valuedness
of r in some neighborhood.

Step 8. Step 7 implies that t¯ and t¯ will be on opposite sides of the X¯2 axis where c̄
resides. Hence ru and rd will be on opposite sides of the X¯2 axis, since this is true
∀a ∈ S. Since Mr < 1 there, r¯u is cd and rd is cu.

Step 9. The image r̄ = r¯u ∪ r¯d by Step 8 has an inflection point at c̄ with
tangent C̄.

Step 10. The proof when the tangent at the cusp is not horizontal is the exact
counterpart of the proof for the inflection given above and illustrated in Fig. 7.8.

Step 11. Finally a word about the slope of the curve going through 1. The case for
inflection points is illustrated in Fig. 7.9. The case for cusps involves a subtlety, and
the reader is encouraged to work it out in detail (see exercises below). 
 7.2 Duality between Cusps and Inflection Points 203

Figure 7.8. The duality cusp ↔ inflection point is independent of slope.

Figure 7.9. Inflection points to cusps when the curve goes through m = 1.

Exercises
1. Show that in a neighborhood of an inflection point I of a curve r:
(a) the curve r has a cu portion ru and cd portion rc separated by I , and
(b) the curve r is monotone there.
2. Show that in a neighborhood of a cusp C of a curve r:
(a) the curve r has a cu and cd portion separated by the tangent at C, and
(b) the curve r has a monotone increasing and monotone decreasing portion
also separated by the tangent at C.
204 7. Curves

3. Find the envelope of each of the following curves in a neighborhood of its

singular points (most of them at (0, 0)) and identify the kind of singularity:
(a) y = x 3.
(b) (x 2 + y 2 )2 − 2a 2 (x 2 − y 2 ) = 0.
(c) the astroid: x 2/3 + y 2/3 = a 2/3 .
(d) y = x 1/3 .
(e) y 3 − x 2 = 0.
(f) the cissoid: x 3 + xy 2 − 2ay 2 = 0.
(g) y = x 3 , x < 0 and = x 2 , x ≥ 0.
(h) the arclight curve: x 2 y 2 + a 2 x 2 − a 2 y 2 = 0.
(i) y = x 2/3 .
(j) y 3 − x 4 = 0.
(k) the trisectrix: x 3 + xy 2 + ay 2 − 3ax 2 = 0.
4. Carefully study the image of cusps whose tangent has slope 1.
5. What happens when a point-curve has vertical slope at its cusp, and what is the
point-curve whose line-curve image has a cusp with vertical slope?

7.3 ♣ FT-3 Point-Curves → Point-Curves

As posed, the construction of a curve’s image involves the sequence of operations

{point-curve} → {line-curve} → {point-curve = envelope of line-curve}.

Unlike the example shown earlier in Fig. 7.1, where the line-curve image (i.e., the
underlying envelope) is clear, in general the overlapping lines obscure portions of the
resulting line-curve (see Exercise 1 below). In what follows, the intermediate steps
are eliminated, resulting in a one-step equivalent {point-curve} → {point-curve)}
transformation having a natural generalization, as we see later on, to surfaces in
RN .
Consider a planar curve c, which for the sake of generality is given implicitly
by

c : F (x1 , x2 ) = 0. (7.6)

Rather than plot lines for the images of its points, the better way is to deal directly
with the tangents at each point and convert them via (7.4) to points. Then, as
illustrated on the right part of Fig. 7.10, determine the relation c̄ : f (x, y) = 0. This
is not only a tremendous labor-saving device, but it also avoids completely plotting
 7.3 ♣ FT-3 Point-Curves → Point-Curves 205

Figure 7.10. Obtaining the point-curve c̄ directly from the point-curve c.

and working with the line-curves. Proceeding, we obtain the total differential of F
and the slope, i.e.,
dx2 Fx
dF = Fx1 dx1 + Fx2 dx2 = 0 ⇒ = − 1.
dx1 Fx2

For a point (x1 , x2 ) on the the tangent o at the point Po = (x1o , x2o ) ∈ c, its slope
is given by

(x2 − x2o )
mo = −Fxo1 /Fxo2 = ,
(x1 − x1o )

and the intercept is found by setting x1 = 0 above, with the x2 axis

x o Fxo + x o Fxo
bo = 2 2 o 1 1 .
Fx2

Substituting in (7.3) yields the point coordinates

∂F /∂x2 (x1 ∂F /∂x1 + x2 ∂F /∂x2 )
x=d , y= , (7.7)
(∂F /∂x1 + ∂F /∂x2 ) (∂F /∂x1 + ∂F /∂x2 )

where d is the interaxis distance and the superscript o is dropped, since the result is
true for any point (x1 , x2 ). Ideal points of c̄ occur at the zeros of the denominator.
206 7. Curves

The smoothness of c ensures the continuity of c̄. The result, (7.7), is known as the
“Inselberg transformation” [41], [42].
Tracing the analogous steps for the transformation in the reverse direction
f (x, y) = 0 → F (x1 , x2 ) = 0, we start with
dy fx
df = fx dx + fy dy = 0 ⇒ =− .
dy fy

The line P̄o that is tangent to the curve c̄ (see the left part of Fig. 7.10) at the point
¯ = (xo , yo ) has the equation
fo y − yo
P̄o : − xo = .
fy x − xo

Since it represents the point Po = (x1o , x2o ), the intersections with the X̄1 , X̄2 axes
are at x1o and x2o respectively. In turn, these are obtained from the equation at x = 0
and x = d, providing
xfx + yfy (x − d)fx + yfx
x1 = , x2 = , (7.8)
fy fy

where again, reference to the original point (x1o , x2o ) is dropped. In this way we have
obtained the transformation:

c : point-curve(x1 x2 plane) ↔ c̄ : point-curve(xy plane),

where we have forsaken the messy intermediate step of having to deal with the
line-curve. When the original point-curve is given explicitly by x2 = g(x1 ), then
(7.7) reduces to
d x2 + x1 g
(x1 )
x= , y= . (7.9)
1 − g
(x1 ) 1 − g
(x1 )
The value d = 1 is used throughout this chapter. The significance of varying d is
better appreciated in Chapter 9, on surfaces. From now on, the images are shown
as point-curve images, and occasionally, when there are no overlapping problems,
as line-curves.

Exercises
1. Plot the line-curve corresponding to x2 = sin x1 . Hint: the inflection-point ↔
cusp duality is helpful in understanding the curve.
2. Find and plot the image of x2 = sin x1 directly as a point-curve using (7.9).
 7.3 ♣ FT-4 Curve Plotting 207

3. Derive the analogue of (7.8) for x2 = g(x1 ).

4. From (7.9), find the image of g(x1 ) = a0 x1 n + x1 .
5. Derive equations (7.7) and (7.8) directly as the envelope of the family of lines
corresponding to the points of the curve c given by (7.6).
6. Write a program that uses the above results to plot the point-curves c̄ for both
implicitly and explicitly defined point-curves c using the above results. This
can be a nice project topic.

7.4 ♣ FT-4 Curve Plotting

The image of a piecewise smooth curve can be computed and plotted via (7.7).
Qualitatively, we can learn quite a bit from some elementary considerations of the
duality by reviewing Fig. 7.11, which we saw earlier in Chapter 3, helping us sketch
the curve’s image. The portofolio of accompanying curves and their images, chosen
to hone our experience, is drawn using a “curve-plotter” based on MATLAB.31
Due to the notational limitations of MATLAB, the bars “ ¯ ” are not included in the
figures.

Figure 7.11. Horizontal position of ¯ depends only on the slope m of .

31
This beautiful curve-plotter (Exercise 6) was written by Tal Tzur and Gilat Schwartz for their course
project in 2004.
208 7. Curves

Figure 7.12. The algebraic curve c : x2 = x13 + x12 − x1 − 1 (left) has image (right) of degree 6.

In Section 7.6 it is pointed out that the image c̄ of an algebraic (i.e., described
by a polynomial) curve c of degree n is also algebraic with degree n∗ as given by
the Plücker class formula (7.40), where s, d are the number of cusps and double-
crossing points respectively. For the curve c on the left of Fig. 7.12, n = 3,
d = 0, s = 0, hence for the image curve (right), we have n∗ = 6. The analysis
is facilitated by Fig. 7.13. The curve c has slope m = 1 at the points AL and AR ,
causing the image curve c̄ to split; the points of tangency there are mapped to ideal
points. The portion cI of c between AL and AR has the the right branch of c̄ as its
image. According to Theorem 7.2.1, the inflection point I is mapped into the cusp ī
of c̄ in between the two axes, since its slope mi is negative. On cI , the curve’s slope
m satisfies mi ≤ m < 1, and for this reason, c̄I opens to the right, intersecting
X̄2 at two points: the images of the tangents at the extrema where m = 0. The
curve’s intersection with the X̄1 axis at two points is due to the zero points at the
extrema; the higher intersection corresponds to Ma . By Corollary 7.1.4, the left
portion of cI , being cd, maps into the upper portion of the of c̄I , which is cu and
approaches ĀL asymptotically. Similarly, the right portion of cI , being cu, maps
into the lower portion, which is cd, approaching ĀR asymptotically. The left portion
of c̄ approaches the X̄1 axis asymptotically as |x1 | → ∞, and the curve’s slope m
goes to ∞. The upper portion of the left branch is cu, being the image of the portion
−∞ < x1 < x1 (AL ), which is also cu. Note that the symmetry of c with respect
 7.4 ♣ FT-4 Curve Plotting 209

Figure 7.13. Analysis of the image curve.

The portions of the original curve c above are specified in terms of the slope (left) at the cubic curves’
important points.

to the tangent i at I is transformed to symmetry with respect to the line I¯ through

the cusp ī, the point image of the tangent i (see exercises below).
The curve c shown in Fig. 7.14 is prescribed by an implicit polynomial with
n = 3, s = d = 0. Again c̄ has degree n∗ = 6 but with two cusps stemming
from the 2 ip of c. The image of the right branch c̄R of c is the portion containing
the cusps. The slope m goes to ∞ as x1 → ∞ with c̄R approaching the X̄1 axis
asymptotically. The (two) points of c where the slope m is equal to 1 are on the oval,
splitting its image to the hyperbola-like part. Proceeding with another curve c (left)
(Fig. 7.15) also prescribed by an implicit polynomial, again n = 3, s = d = 0,
and the image curve c̄ has degree n∗ = 6. There are two points of c where m = 1
responsible for the split of c̄. The cusp to the right of the X̄2 axis is the image of the
lower ip of c where the 0 < m < 1, and also, since m → ∞ as x1 → ∞, this part
of c̄ approaches the X̄1 axis asymptotically. In the next example, shown in Fig. 7.16,
the image curve has degree n∗ = 4, since n = 3, d = 1, s = 0. Both points of
c where m = 1 are to the right of the double-crossing point, with the part cR to
the right of these points mapped into the upper portion of c. The remaining part cL
of c containing the double point is mapped into two branches, each approaching
asymptotically the negative X̄1 axis. The two tangents at the double point map into
two points: the one on X̄2 is the image of the horizontal tangent. These two points
210 7. Curves

Figure 7.14. The image of the curve c : F (x1 , x2 ) = x13 − x1 − x22 = 0 has degree 6.

Figure 7.15. The image of the curve c : F (x1 , x2 ) = x13 + x12 − x1 − x22 = 0 also has degree 6.

have the same tangent, the x axis. A tangent at two points of a curve is called a
bitangent. The algebraic curve c shown in Fig. 7.17 is specified parametrically, and
has n = 3, s = 1, d = 0, so that c̄ has degree n∗ = 3. The two branches of c are
tangent to the x1 axis at the cusp, and therefore, it maps to the inflection point where
 7.4 ♣ FT-4 Curve Plotting 211

Figure 7.16. The image of the curve c : F (x1 , x2 ) = x13 + x12 − 3x1 x2 = 0 has degree 4.

Figure 7.17. The image of the parametric polynomial x(t) = t 2 , x2 (t) = t 3 has degree 3.

c̄ crosses the X̄2 axis. As in the previous examples, due to m → ∞ as x1 → ∞,

c̄ approaches the positive X̄1 axis asymptotically on either side. It is very easy to
sketch the image of the exponential curve seen in Fig. 7.18. There is one point with
slope m = 1 spliting the image curve c̄ into a part left of X̄1 , corresponding to the
part on the right, where 1 < m < ∞, and the left part of c, where 0 < m < 1, whose
212 7. Curves

Figure 7.18. The image of the exponential curve c : x2 = ex1 .

The image does not have a portion between the axes, since the curve does not have negative slopes.

Figure 7.19. The image of c : x2 = sin(x1 ) in the interval [0, 2π ] is a nonoscillating curve.

image is to the right of the X̄2 axis. In the absence of negative slopes of c, c̄ does not
have a portion between the -axes. It is interesting to trace the image of oscillatory
curves starting with the trigonometric function x2 = sin(x1 ) plotted in Fig. 7.19 for
the interval x1 ∈ [−π, 0]. The ip at x1 = π maps into the cusp that is in between
 7.4 ♣ FT-4 Curve Plotting 213

Figure 7.20. The image of c : x2 = sin(x1 ) for x1 ∈ [−2π, 2π ].

It is symmetric about the x axis. The ip at x1 = ±π are mapped into the two cusps, and the ip at the
origin having slope m = 1 is mapped to the ideal point along the x axis.

the axes, since the tangents’ slopes m are negative in its vicinity. In the remainder,
the slope satisfies 0 < m(x1 ) = cos(x1 ) ≤ 1, and hence c̄ opens to the right, the
upper portion approaching the x axis asymptotically due to the ideal point from
x1 = 0 where the slope is 1. The graph for the interval x1 ∈ [0, 2π] is the mirror
image of the curve above (Fig. 7.20), since the slopes satisfy m(x1 ) = cos(−x1 ).
Altogether, then, the ip at x1 = ±π with m = 1 are mapped into the two cusps
with x = 1/2 and the ip at the origin, having slope m = 1, is mapped to the ideal
point along the x axis. The mirror symmetry is preserved in the image of the curve
c : x2 = x1 cos(x1 ) in Fig. 7.21 and for analogous reasons.

7.4.1 Space-Curves
Curves and their images are an essential component in the representation of surfaces
(Chapter 9) in terms of planar regions whose boundaries are the images of certain
space-curves. It will be seen that via the wonderful properties of the superplanes, the
boundaries can be obtained from images of plane-curves as studied in this chapter.
Still, a few words about space-curves are obligatory. Let us start with a convenient
parametric definition with sc denoting a space-curve:
214 7. Curves

Figure 7.21. The image of the oscillatory curve x2 = x1 cos(x1 ).

It is a nonoscillating curve symmetric with respect to the x axis.

sc : xi = xi (s), s ∈ I ⊂ R, i = 1, 2, 3, (7.10)

where I is an interval and the parameter s is often taken as the arc length. As for
plane-curves, the tangent t at a point P is found as the limiting chord joining P to
an adjoining point. Two other adjoining points Q, R together with P form a plane,
which in the limit as (Q, R) → P is the osculating plane—a rather endearing
appellation—containing the tangent t. Within the osculating plane, the local center
of curvature of sc at P is found as for plane-curves and the principal normal n at P
directed towards the center is chosen. It is called principal to distinguish it from the
infinity of normals on the plane orthogonal to t at P , or just normal when the context
is clear. The directed tangent t is chosen so that together with n as axes they form a
right-handed system. Next, the binormal b is included orthogonal to the osculating
plane and directed so that t → n → b has the same sense as x1 → x2 → x3 , i.e.,

b = t × n, (7.11)

the three usually taken as unit vectors. The rate of change

db
= −τ n (7.12)
ds
introduces the torsion τ , which measures the “nonplanarity,” that is, the change of
the osculating plane along the curve with τ = 0 for plane curves. The derivatives
 7.4 ♣ FT-4 Curve Plotting 215

of b, t, n provide the important equations of Frenet relating the three unit vectors
to the two fundamental properties: torsion and curvature. Prescribing the curvature
and torsion as functions of the arc length completely determines a space-curve up
to its position and orientation. We leave (see Exercise 13) this interesting topic after
this very short introduction in order to return to plane curves for the remainder of
the chapter.

Exercises
1. Prove that the symmetry about the tangent i of the ip I corresponds to the
symmetry about the line I¯ through ī, or find portions of c̄ where the symmetry
fails.
2. What points correspond to the max and min of the left-hand branch in Fig. 7.12?
3. Is the double -point ↔ bitangent duality seen in Fig. 7.14 true in general?
Prove or give a counterexample.
4. For the curve in Fig. 7.14, provide an analysis of its image and draw a figure
of the important points with the tangents similar to Fig. 7.13.
5. What portions of c̄ in Fig. 7.14 are symmetric about a line and why?
6. Analyze the image of the curve in Fig. 7.15, sketching a figure similar to
Fig. 7.13.
7. Is there an algebraic curve of degree 3 whose image has degree 5? Give an
example of an algebraic curve of any degree having image with degree 5.
8. Rotate the exponential curve in Fig. 7.18 to obtain an image having points for
all values of x.
9. In Fig. 7.21, mark the portions of c and their corresponding parts in the image
c̄.
10. Prove that the image c̄ in Fig. 7.19 consists of monotone increasing and
monotone decreasing portions, and find the portions of c corresponding to
each.
11. Let c : x2 = f (x1 ) be an odd function (i.e., f (x1 ) = −f (x1 )). Is c̄ necessarily
symmetric about the x-axis?
12. For a general curve c given by (7.6), the image of the tangent at each point
P ∈ c is the point given by (7.7). Find
(a) the image of the normal at P .
(b) the formula for the curvature of the image curve c̄; hard.
13. Space curves
(a) Find the image cs of a smooth space-curve cs : xi = xi (s) ⊂ R3 ,
i = 1, 2, 3. Hint: recall from Fig. 7.10 that the image of a curve is
216 7. Curves

Table 7.1. Equalities between invariants of Plücker curves and their duals.

Point(s) on curve c → map into points of the curve c̄ ⇒ relation

The two points of c on a bitangent map into a double point of c̄ b = d∗

A double point of c maps into two points on a bitangent of c̄ d = b∗
An inflection point of c maps into a cusp of c̄ ip = s ∗
A cusp of c maps into an inflection point of c̄ s = ip∗

constructed from the tangents at each point. For a space-curve the tangent
at each point is a line in R3 and is represented by two points ¯12 , ¯23 ,
rather than one point as for R2 . Hence the image cs consists of two curves,
say cs 12 and cs 23 , in the xy plane.
(b) For the space-curve c above, using the property that the three points
¯12 , ¯23 , ¯13 are collinear, how can the curve cs 12 be determined from
the other two curves?
(c) Obtain generalizations of the dualities listed in Table 7.1.
(d) Find the image cs of a smooth space-curve cs : xi = xi (s) ⊂ RN ,
i = 1, . . . , N, consisting of N − 1 curves in the xy plane and generalize
the dualities in Table 7.1.
This is an excellent project topic.

7.5 ♣ FT-5 Transforms of Conics

The conic sections are described by the quadratic function

F (x1 , x2 ) = A1 x12 + 2A4 x1 x2 + A2 x22 + 2A5 x1 + 2A6 x2 + A3 = 0
⎛ ⎞⎛ ⎞
A1 A4 A5 x1
= (x1 , x2 , 1) ⎝ A4 A2 A6 ⎠ ⎝ x2 ⎠ , (7.13)
A5 A6 A3 1
where the type of conic is determined by the sign of the discriminant
 = (A24 − A1 A2 ). The coefficient matrix is denoted by A, and its determinant,
which plays an important role in the development, is
det A = A3 (A1 A2 − A24 ) − A1 A26 − A2 A25 + 2A4 A5 A6 . (7.14)
In this section it is first shown that conics ↔ conics, and then criteria for the
determination of specific conic images are developed.
 7.5 ♣ FT-5 Transforms of Conics 217

7.5.1 ** Proving the Duality Conics ↔ Conics

For conics, using the identity that for a polynomial F of degree n, F (x) = 0 ⇒
∇F · x = ∇F · x − nF , the second expression being linear, (7.7) becomes

A4 x1 + A2 x2 + A6
x=d ,
[(A1 + A4 )x1 + (A2 + A4 )x2 + (A5 + A6 )]
A5 x1 + A6 x2 + A3
y=− . (7.15)
[(A1 + A4 )x1 + (A2 + A4 )x2 + (A5 + A6 )]

Note that the interaxis distance d enters as a multiplier only in the x coordinate
due to the line → point mapping. These are Möbius32 transformations, which form
a group (see any good book in modern algebra or [10]). This observation enables
substantial simplifications of the earlier treatment of conics and their transforms
(see [41] and [42]).33 The inverse, expressing x1 and x2 in terms of x and y, is a
Möbius transformation of the form
a11 x + a12 y + a13 a21 x + a22 y + a23
x1 = , x2 = , (7.16)
a31 x + a32 y + a33 a31 x + a32 y + a33

which can also be written as

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
x1 x ' a11 a12 a13
⎝ x2 ⎠ = α ⎝ y ⎠ (a31 x + a32 y + a33 ), α = ⎝ a21 a22 a23 ⎠ .
1 1 a31 a32 a33
(7.17)

It is also known that the Möbius transformations are homomorphic to the group of
nonsingular matrices of appropriate size. Without entering into unnecessary details,
the Möbius transformation
b11 x + b12 y + b13 b21 x + b22 y + b23
u= , v= , (7.18)
b31 x + b32 y + b33 b31 x + b32 y + b33

corresponds to the matrix M1 = [bij ], where it is required that det(M1 ) = 1,

normalizing if necessary. That does not pose a problem if M1 is nonsingular, since
the numerators and denominators of the fractions in (7.18) can be multiplied by
the same constant without altering the transformation. In our case, the matrix
corresponding to the Möbius transformation (7.15) is
32
Also called linear rational transformations.
33
The result conics ↔ conics was first proved by B. Dimsdale.
218 7. Curves

⎛ ⎞ ⎛ ⎞
dA4 dA2 dA6 0 d 0
C
= ⎝ −A5 −A6 −A3 ⎠ = ⎝ 0 0 −1 ⎠ A = RA.
(A1 + A4 ) (A2 + A4 ) (A5 + A6 ) 1 1 0
(7.19)
The matrix R is obtained by applying row operations on C
to achieve the
indicated reduction. Noting that det(C
) = − det(A) = H , we normalize by
C = C
/H = RA/H to obtain det(C) = 1 so as to satisfy the requirement of
the aforementioned homomorphism. This corresponds to multiplying both numer-
ator and denominator of the Möbius transformation (7.15), or equivalently the
elements of the corresponding matrix, by −(det A)−1/3 . The coeffiecients, and
in turn the corresponding matrix α = [aij ], in (7.16), which is the inverse Möbius
transformation of (7.15), are simply the elements of C −1 . In terms of the original
coefficients,
(C
)−1 = HC−1 = HA−1 R −1 = −adj(A)R −1 = α, (7.20)

where adj(A) = [Aij ]T is the adjoint, with the Aij being the cofactors of the matrix
A. Proceeding with the evaluation,
⎛ ⎞⎛ ⎞
A11 A21 A31 −1/d 0 1
(C
)−1 = − ⎝ A12 A22 A32 ⎠ ⎝ 1/d 0 0 ⎠
A13 A23 A33 0 −1 0
⎛ ⎞
(A11 − A21 )/d A31 −A11
= ⎝ (A12 − A22 )/d A32 −A12 ⎠ . (7.21)
(A13 − A23 )/d A33 −A13
The individual aij are
⎫
a11 = (A11 − A21 )/d = [A3 (A2 + A4 ) − (A6 2 + A5 A6 )]/d, ⎪
⎪
⎪
⎪
a12 = A31 = A4 A6 − A2 A5 , ⎪
⎪
⎪
⎪
⎪
⎪
a13 = −A11 = A6 − A2 A3 ,
2 ⎪
⎪
⎪
⎪
⎪
a21 = (A12 − A22 )/d = [(A5 + A5 A6 ) − A3 (A1 + A4 )]/d, ⎪
2 ⎪
⎪
⎬
a22 = A32 = A4 A5 − A1 A6 , (7.22)
⎪
⎪
a23 = −A12 = A3 A4 − A5 A6 , ⎪
⎪
⎪
⎪
⎪
⎪
a31 = (A13 − A23 )/d = [A6 (A1 + A4 ) − A5 (A2 + A4 )]/d, ⎪ ⎪
⎪
⎪
⎪
⎪
⎪
a32 = A33 = A1 A2 − A4 ,
2
⎪
⎪
⎪
⎭
a33 = −A13 = A2 A5 − A4 A6 .
 7.5 ♣ FT-5 Transforms of Conics 219

Recapitulating,
⎛ ⎞
x1
F (x1 , x2 ) = (x1 x2 1)A ⎝ x2 ⎠
1
⎛ ⎞
x '
T
= (x y 1)α Aα ⎝ y ⎠ (a31 x + a32 y + a33 )2 = 0, (7.23)
1

for (a31 x + a32 y + a33 )2 = 0. Or

⎛ ⎞ ⎛ ⎞
x a1 a4 a5
1 T
f (x, y) = H (x y 1)a ⎝ y ⎠ = 0, a = (α Aα) = ⎝ a4 a2 a6 ⎠ .
H
1 a5 a6 a3
(7.24)

Further,

a = (R −1 )T (A−1 )T AA−1 R −1 H = (R −1 )T α, (7.25)

where (A−1 )T = (AT )−1 = A−1 , since A is symmetric. Therefore,

⎛ ⎞
a1 a4 a3
a = ⎝ a4 a2 a6 ⎠
a5 a6 a3
⎛ ⎞⎛ ⎞
−1/d 1/d 0 (A11 − A21 )/d A31 −A11
=⎝ 0 0 −1 ⎠ ⎝ (A12 − A22 )/d A32 −A12 ⎠
1 0 0 (A13 − A23 )/d A33 −A13
⎛ ⎞
(2A12 − A22 − A11 )/d 2 (A23 − A13 )/d (A11 − A12 )/d
=⎝ (A23 − A13 )/d −A33 A13 ⎠.
(A11 − A12 )/d A13 −A11
(7.26)

The reduction is due to the symmetry of A and hence of its cofactors, i.e., Aij = Aj i ,
confirming that a is symmetric. The individual ai are
220 7. Curves

⎫
a1 = [A3 (A1 + A2 + 2A4 ) − (A5 + A6 )2 ]/d 2 , ⎪
⎪
⎪
⎪
⎪
⎪
a2 = A1 A2 − A4 2 , ⎪
⎪
⎪
⎪
⎪
⎬
a3 = A2 A3 − A6 2 ,
(7.27)
a4 = [A13 − A23 ]/d = [A6 (A1 + A4 ) − A5 (A2 + A4 )]/d, ⎪
⎪
⎪
⎪
⎪
⎪
⎪
a5 = [A6 + A5 A6 − A3 (A2 + A4 )]/d,
2 ⎪
⎪
⎪
⎪
⎭
a6 = A2 A5 − A4 A6 .

The corresponding results for the inverse transform, i.e.,

conics in the xy plane &→ conics in the x1 x2 plane

are left as an exercise. The general results obtained here enable the study and
complete classification of the conic transforms. The transforms of specific conics
are obtained from the criteria developed next.

7.5.2 ♣ FT-6 Classification of the Conic Transforms

Projectively, conics are ellipses having 0 (ellipse), 1 (parabola), or 2 (hyperbola)
ideal points. The conic’s ideal point(s) can be found by taking limx1 →∞ (x2 /x1 ) =
m, where m is a constant (i.e., the direction of the ideal point), from (7.13), yielding
the relation

A2 m2 + 2A4 m + A1 = 0. (7.28)

To include the vertical direction (i.e., where the slope is ± = ∞), evaluating the
reciprocal limx2 →∞ (x1 /x2 ) = M provides

A1 M 2 + 2A4 M + A2 = 0. (7.29)

The conic’s indicator I is defined as the sum of the coefficients

I = A1 + A2 + 2A4 . (7.30)

The conic transforms are given by

f (x, y) = a1 x 2 + 2a4 xy + a2 y 2 + 2a5 x + 2a6 y + a3 = 0, (7.31)

the conic type being determined from the sign of

δ = a42 − a1 a2 = H (A1 + A2 + 2A4 ) = HI. (7.32)

 7.5 ♣ FT-5 Transforms of Conics 221

Figure 7.22. Ellipses always map into hyperbolas.

Each assymptote is the image of a point where the tangent has slope 1.

When H < 0, the conic has no real points.34 When H = 0, the conic is a point.
The other two degenerate cases occur for conics consisting either of a single line
or a pair of intersecting lines. In the ensuing it is assumed that H > 0, which fixes
the signs of the quadratic’s coefficients. In view of (7.32), the indicator I together
with the sign of  tells the story. For a particular conic c, only the coefficients
A1 , A2 , and A4 determine the type of conic c̄. So rotations of c, as reflecte by A4 ,
may change the type of its transform, while translations, which are specifie by the
linear and constant terms cannot. The analysis below is listed by conic types.
Ellipse,  < 00. Then A1 A2 > A24 ≥ 0; hence A1 and A2 have the same sign.
If they are positive, then A1 > A24 /A2 and

I = (A1 + A2 + 2A4 ) > (A24 + A22 + 2A4 A2 )/A2 = (A2 + A4 )2 /A2 ≥ 0,

so the transform is a hyperbola (Fig. 7.22), since δ > 0. The same inequality
holds if A1 and A2 are both negative. For an ellipse, of course, the mi are complex
conjugates.

34
The quadratic (7.13) may be solved for x2 in terms of x1 and the coefficients Ai . There are no real values
of x2 if the discriminant is negative. This provides a quadratic inequality in x1 whose discriminant
turns out to be H .
222 7. Curves

Figure 7.23. A parabola whose ideal point does not have direction with slope 1.
It always transforms to a hyperbola with a vertical asymptote. The other asymptote is the image of
the point where the parabola has tangent with slope 1.

Parabola,  = 00. If any one of A1 , A2 , A4 is 0, then all three are 0, and (7.13)
does not describe a quadratic function. So A2 = 0 and A1 = A24 /A2 . Then
⎫
(A1 + A2 + 2A4 ) = (A2 + A4 )2 /A2 , ⎪
⎪
⎪
⎪
(A1 + A4 ) = (A2 + A4 )A4 /A2 , ⎪
⎪
⎪
⎬
H = A1 A26 + A2 A25 − 2A4 A5 A6 = (A4 A6 − A2 A5 )2 /A2 , (7.33)
⎪
⎪
a4 = (A2 + A4 )(A4 A6 − A2 A5 )/A2 , ⎪
⎪
⎪
⎪
⎪
⎭
a1 = [A3 (A2 + A4 )2 − A2 (A5 + A6 )2 ]/A2 .

Hence if I = 0, then A2 = −A4 , δ = HI > 0, and the transform is a hyperbola,

as shown in Fig. 7.23. Otherwise, when I = 0, from (7.34), the parabola’s ideal
point has slope 1, and the transform is a parabola. This self-dual case is shown in
Fig. 7.24.
Hyperbola,  > 00. If I < 0, then the transform is an ellipse, as shown in
Fig. 7.25. This is the dual of the ellipse &→ hyperbola transform.
Otherwise, if I = 0 (see (7.28)), then one of the hyperbola’s ideal points has
slope 1 and the transform is a parabola (Fig. 7.26). This is the dual of the
parabola &→ hyperbola transform.
Lastly, if I > 0, the transform is a hyperbola. This self-dual case is shown in
Fig. 7.27.
 7.5 ♣ FT-5 Transforms of Conics 223

Figure 7.24. A parabola whose ideal point has slope 1 transforms to a parabola; self-dual.

Figure 7.25. Hyperbola to ellipse; dual of the case shown in Fig. 7.22.
224 7. Curves

Figure 7.26. If the asympote has slope 1, then the hyperbola → parabola; dual of Fig. 7.23.

Figure 7.27. Hyperbola to hyperpola; self-dual case.

Recasting the indicator I in terms of the conic’s ideal points provides not only an
intuitive way of understanding the transforms but also, as shown in the next section,
a pointer to generalizations. For A2 = 0 let mi , i = 1, 2, be the roots of (7.28) with
 7.5 ♣ FT-5 Transforms of Conics 225

m1 ≤ m2 when they are real. From the coefficients expressed in terms of the roots
by A1 = A2 m1 m2 , 2A4 = −A2 (m1 + m2 ),

I = A2 (1 − m1 )(1 − m2 ), (7.34)

with one of the mi equal to 0 when A1 = 0. For A2 = 0 the roots of (7.29) are
M1 = 1/m1 = A4 /A1 , the only one also obtained from (7.28), and M2 = 0 (i.e.,
m2 = ±∞). Finally, when both A1 = A2 are zero, the two roots, obtained one
from each of the equations, are m1 = 0 and M2 = 0 (m2 = ±∞).

Whence Come the Hyperbolas?

We have seen that hyperbolas occur as images of ellipses, parabolas, and also
hyperbolas. Is there a way to decide from the hyperbola what kind of conic it is the
image of? To clarify this, it is necessary to trace the image of interior points. The
interior points of an ellipse are mapped via the 2-D point ←→ line into lines that
are between the hyperbola’s two branches. This is also true for the two cases (see
exercises below). Even without this observation, the easiest case is for a hyperbola
that has a vertical asymptote. From the foregoing analysis, this can only be the
image of a parabola, where the vertical asymptote represents the parabola’s ideal
point. This is seen well in Fig. 7.23, where the parabola opens in the direction with
slope m = −1 and the vertical asymptote is at x = 1/2. The other asymptote is, of
course, determined by the point where the parabola has a tangent with slope m = 1
and is discussed below, providing (7.38).
To distinguish the remaining two cases, it is worth scrutinizing the left portions
of Figures 7.22 and 7.27. The essential difference and clue is the horizontal sepa-
ration along the x axis of the two left and right branches in Fig. 7.27, as contrasted
with the vertical separation, along the y axis, of the top and bottom branches. Note
that the hyperbola on the right side of Fig. 7.27 contains the ideal points whose
directions are between the upper left asymptote and the upper right asymptote.
That is all the directions of the lines that cannot be tangents to the hyperbola. All
these ideal points are mapped into the vertical lines between the two branches and
hence the horizontal separation. Of course, there are also regular points between
the branches of the original hyperbola, and these are mapped into nonvertical lines
between the branches of the image hyperbola. By contrast, the ellipse has no ideal
points, and hence there cannot be any horizontal separation between the branches
of its image. Rather, all interior points appear as lines between the two branches.
♣ FT-6e
226 7. Curves

Alternative Analysis Based on the Indicator

Now let us revisit the results above in view of (7.34).
• When the mi are complex conjugates, the conic is an ellipse with I > 0.
• Otherwise, when there is a single real root m1 , the conic is a parabola, and the
transform is
– a hyperbola with one vertical asymptote when mi = 1, and
– a parabola when mi = 1.
• Finally, when there are two distinct real roots, the conic is a hyperbola whose
asymptotes have slopes mi . There are two such hyperbolas: one whose tangents
have slope m ∈ [m1 , m2 ] and the other with slope m ∈ [m1 , m2 ]. For easy refer-
ence we refer to the two hyperbolas having the same asymptotes as conjugate.
The conic’s transform is determined by whether 1 ∈ [m1 , m2 ]. Recalling (7.4):
– when m1 < 1 < m2 , the points of the conic’s transform have x =
d/(1 − m), whose range is the finit interval [1/(1 − m2 ), 1/(1 − m1 )],
hence the transform must be an ellipse, I < 0.
– For the conjugate hyperbola whose tangents have slopes m ∈ (m1 , m2 ) (i.e.,
two semi-infinite intervals), the transform must be a hyperbola, I > 0.
– For m1 < 1 = m2 , one hyperbola has tangents with slope m ∈ (m1 , 1).
The conjugate hyperbola’s tangents have slope m ∈ (m1 , 1) as shown in
Fig. 7.26. Since I = 0, the images of both hyperbolas are parabolas tangent
at their vertices and the reflection of each other.

7.5.3 ** The Conic Transform’s Ideal Points

It is instructive to investigate the conic transforms from their ideal points. Again
we have to consider two limits to obtain the corresponding equations. To wit,
limx→∞ y/x in (9.21) yields
√
2 −a4 ± H I
a2 R + 2a4 R + a1 = 0 ⇒ R = , (7.35)
a2
and limy→∞ (x/y) = r for
√
2 −a4 ± H I
a1 r + 2a4 r + a2 = 0 ⇒ r = . (7.36)
a1
Let us now investigate the conic transform’s ideal points methodically.
 7.5 ♣ FT-5 Transforms of Conics 227

• When I < 0, the conic is a hyperbola with m1 < 1 < m2 , whose transform,
being an ellipse, has no ideal points.
• For I = 0, we have A1 + A4 = −(A2 + A4 ), and the conic’s transform is a
parabola. If the original conic is a parabola, then − = a2 = 0, r = −a4 /a1 ,
and from (7.33), A2 + A4 = 0, a4 = 0 ⇒ r = 0. Hence the transform’s ideal
point is in the vertical direction as seen in Fig. 7.24. Otherwise, the original
conic is a hyperbola. Then
A2 + A4
r=− , (7.37)
A5 + A 6
and the transform’s ideal point has slope 1/r = R. Note that this is due to
δ = 0 ⇒ a4 /a1 = a2 /a4 .
• Finally, for I > 0, r has two real values, and the transform is a hyperbola. If the
original conic is an ellipse, the transform has asymptotes with finite slopes 1/r,
as shown in Fig. 7.22.
If it is a parabola, then the roots are r = 0, which accounts for the vertical
asymptote, and
a4
r = −2 , (7.38)
a1
which provides the other asymptote shown in Fig. 7.23.
If it is a hyperbola, this is self-dual, and the transform has asymptotes with finite
slopes 1/r, as seen in Fig. 7.27.
Observe that the transforms of a parabola always have an ideal point in the vertical
direction (r = 0). In turn, this is because each parabola’s ideal points are mapped
into a vertical line by the basic point ↔ line. A hyperbola’s two ideal points are also
mapped to vertical lines, which turn out to be vertical tangents of the corresponding
transform, i.e., the vertical tangents at the ellipses’ endpoints in Fig. 7.25, the
parabola’s vertex in Fig. 7.26, and the two vertices of the hyperbola in Fig. 7.27.
To summarize, the transform’s ideal points are either
• the images of the conic’s ideal points r = 0 for parabolas as shown in Fig. 7.24
and 7.27, or
• the images of points where the conic has tangents with slope 1. They correspond
to the real nonzero roots of (7.28).

** Eliminating the Conics’ Linear Terms by Translation

The conic transform’s ideal points are very sensitive to changes in the original
conic’s orientation. It is worthwhile to study the effects of translation. For this
228 7. Curves

reason, the linear terms in (7.13) are omitted (i.e., A5 = A6 = 0), since they can
be eliminated with a translation. Then
⎫
a1 = A3 (A1 + A2 + 2A4 ), ⎪ ⎪
⎪
⎪
a2 = A1 A2 − A4 ,2 ⎪
⎪
⎪
⎪
⎪
⎪
a3 = A2 A3 , ⎬
(7.39)
a4 = 0, ⎪
⎪
⎪
⎪
⎪
⎪
a5 = A3 (A2 + A4 ), ⎪
⎪
⎪
⎪
⎭
a6 = 0,

with δ = HI and H = A3 . It is striking that the cross term vanishes (a4 = 0).
The resulting conic transforms are in standard position with their symmetry axes
parallel to either the x or y axis. For illustration, the parabola in Fig. 7.23 is not
centered, and its transform is not in standard position. The reason for the parabola’s
horizontal ideal point (i.e., 1/r = 0) in Fig. 7.26 is seen from (7.37). This is due to
the duality
( & ( &
translation of a rotation of a
⇐⇒ .
point along a line line about a point

This observation deserves further analysis, since there is a translation in the x but
not in the y direction.
Note that all the transforms in the figures, except Fig. 7.23, are translated in the
x direction. Further, the original conic is centered, and the resulting conic transform
is not rotated from the standard position. A surprising observation is that only five of
the six cases are found from the results above when the linear terms are eliminated.
This is because when A5 and A6 are zero, a parabola’s H vanishes and the two roots
of (7.28) collapse to M = 0 (see (7.38)). For this reason, the parabola in Fig. 7.23
is carefully not centered in order to avoid this. Recall that the normalization of the
determinant matrix’s C
, (7.19), involved division by H , and hence the method
breaks down for H = 0 (see exercises below).

Exercises
1. Using the theory of envelopes, find the transforms of the three different kinds
of conics in
(a) standard position,
(b) translated in the x and separately in the y direction,
(c) with the cross term not zero.
 7.5 ♣ FT-7 Transforms of Algebraic Curves 229

2. What is the conic transform of the hyperbola x2 = 1/x1 and its conjugate?
3. Find the points of a conic having tangents with slope 1 and the corresponding
ideal points on their transform.
4. Find the conic transform for the case A5 = A6 = 0 directly to avoid the
problem with H = 0. Is there a parabola &→ hyperbola transformation in this
case?
5. Using (7.34), construct a table listing the conditions for the Aj when I > 0.
For example, I > 0 and A2 > 0 ⇒ 1 > ri , or 1 < ri , i = 1, 2, and proceed
as for 1 > ri > 0 ⇒ A2 > A1 > 0 > A4 , etc.
6. Find the conic transform of the conjugate hyperpola of that shown in
(a) Fig. 7.25.
(b) Fig. 7.26.
(c) Fig. 7.27.
7. Which conic has as transform the conjugate of the hyperbola shown in the xy
plane in Fig. 7.27?
8. What is the image of a hyperbola having a vertical asymptote like the one shown
on the left part of Fig. 7.23?
9. Prove that interior points of the conics mapped into hyperbolas are mapped
into lines between the two branches. Here the “interior” points of a hyperbola
are those between the two branches.
10. Consider hyperbolas having the property that if two lines are completely
between the two branches, then the point on the two lines is also between
the two branches. Delimit the class of hyperbolas having this property and
provide a counterexample of a hyperbola without this property.
11. Obtain the main result in the reverse direction, that is, show that

line-conics in the xy plane &→ point-conics in the x1 x2 plane

and delimit the various cases.

7.6 ♣ FT-7 Transforms of Algebraic Curves

Conic transforms are studied for two reasons. One is the ease of use of the Möbius
transformations, which is completely general for quadrics, the surfaces prescribed
by quadratic equations, in any dimension. The other is that their detailed analysis
serves as a model and guide in the study of the related images of far more complex
curves, regions, and their ramifications in the next section.
230 7. Curves

The class of algebraic curves, those described by polynomial equations, is

studied in the field of algebraic geometry ([37], [78], [179], and [86]). An algebraic
curve c has many invariants that are properties independent of the particular coor-
dinate system used. Examples are the number of components in its graph, degree
n, number of double points d (points where the curve crosses itself once), cusps s,
inflectio points ip, and bitangents b (i.e., tangents at two points). The results here
apply to any curve c and its dual curve c∗ for any point ↔ line duality. The same
symbols with the ∗ superscript denote the dual’s invariants. Algebraic curves and
their duals have been studied extensively, starting in 1830 by the mathematician
and physicist Julius Plücker, who also made other important contributions to the
field. His results apply to the class of Plücker curves [61, pp. 73–94] having the
following properties:
1. c is irreducible and of degree n ≥ 2, and
2. the singularities of c, c̄ are at most simple double points (i.e., points where the
curve crosses itself once) and cusps.
Whereas all irreducible quadratics and cubics are Plücker curves, there exist
quartics, n = 4, that are not.
Of interest here are the relations between a curve’s invariants and those of its
dual. As indicated in the equalities tabulated below, in addition to the ip ↔ cusp
duality, there is a bitangent ↔ double point duality, which we already met in
Fig. 7.15. This is reasonable, for the two tangents at a double point map into the
two points on a bitangent, which is the double point’s image, and in fact, it is true
for general and not just algebraic curves. The dual of c is an algebraic curve whose
degree n∗ depends on n and the invariants d, s as given by the Plücker class formula

n∗ = n(n − 1) − 2d − 3s. (7.40)

For n = 2 the Plücker class formula yields n∗ = 2 and s ∗ = 0, confirming the

conclusions in Section 7.5.
The polynomial describing the dual c∗ can be found for any point ↔ line
duality by two different methods [61, p. 74]. However, with n∗ = O(n2 ) and
the complexity increasing rapidly with the number of nonzero coefficients of the
polynomial specifying c, the process is very tedious. All this applies to c̄, which
is the image under the particular -coords duality, when c is an algebraic curve c.
As pointed out in Section 7.4, the properties of c̄ are immediately available from
Plücker’s results. Together with the qualitative considerations discussed and a good
curve plotter, a complete grasp of c̄ and its properties can be obtained. This avoids the
laborious process of obtaining the polynomial for c̄ and its subsequent computation
for plotting. There are numerous generalizations of the Plücker formulas.
 7.7 ♣ FT-8 Convex Sets and Their Relatives 231

7.7 ♣ FT-8 Convex Sets and Their Relatives

Consider a double cone, as shown in Fig. 7.28, whose base is a bounded convex
set rather than a circle. The three type of sections shown are generalizations of the
conics and are conveniently called gconics. They are one of the following:
a bounded convex set is abbreviated by bc,
an unbounded convex set, denoted by uc, containing a nonempty set of ideal
points whose slope m is in an interval m ∈ [m1 , m2 ], or
a generalized hyperbola, denoted by gh, consisting of two full (not segments)
lines u , , called asymptotes, and two infinite chains: one convex-upward chain
cu above both asymptotes and another convex-downward chain c below both
asymptotes.
In the past, the corresponding regions were also referred to as estars, pstars, and
hstars [99], [110]. Here the bc or uc are considered as consisting of their interior
points and boundaries. A gh consists of the points between the upper and lower
chains and the boundaries. As for the hyperbola, this is best understood in the
setting of the projective plane, where the gh is a “bc” intersecting the ideal line
at two points, the asymptotes. In the event that the upper and lower chains have
straight-line edges at the endpoints, then these two infinite lines are the asymptotes:
u joins the leftmost point (i.e., with minimum x) of the upper chain cu to the

Figure 7.28. Gconics: three types of sections.

Bounded convex bc (left), unbounded convex uc (right), gh regions (two ideal points) (middle).
232 7. Curves

rightmost point (i.e., with maximum x) of the lower chain c and, joins the
leftmost point of c to the rightmost point of cu .
It turns out that the conic transforms are special cases of the more general and
completely analogous results for gconics [99], [110]. In order to study the gconic
transforms, it is assumed that their boundaries are point-curves that transform nicely
to line-curves, as discussed at the beginning of this chapter. Since the treatment is in
P2 , the understanding is that convexity is interpreted completely as in the Euclidean
R2 plane with the line segment (not including the ideal point) joining a pair of points
in the set also contained in the convex set.

7.7.1 Gconics and Their Transforms

The insight obtained from the indicator in Section 7.5.2 points the way to the
corresponding analysis for the gconics, though there is no explicit relation like
(7.34). The discussion pertains to the supporting lines of a set’s boundary at points
where the tangent is not defined. The key convexity property used here is that for
any given slope m, a convex set has a unique pair of supporting lines with that slope.
As for the conics, finite and infinite line segments are the degenerate cases of a bc
and uc respectively, while a pair of intersecting lines is a degenerate gh. These are
easily handled directly, so they are excluded from the ensuing treatment. Unless
otherwise stated, a gconic is the set consisting of the boundary and its interior.
Theorem 7.7.1. The images of gconics are gconics.
1. BC is a bc ⇒ BC is a gh.
** Proof. The boundary ∂(BC) of a bounded convex BC has precisely two
points Au , A with supporting lines au above a having slope m = 1. The line
w on Au , A divides the boundary into two chains cu above and c below it
respectively. The image ∂(BC) is split by the two “asymptotes” A and Au into
two chains. There exists a point V on ∂(BC) such that cu = c1 ∪ c2 , where
c1 = A V is a cu chain with M1 ≥ 1 and c2 = V Au a cd chain with M2 < 1.
Hence, by Corollary 7.1.4, cu = c1 ∪ c2 is a cu chain. Similarly, it can be
shown that c is a cd chain. Therefore ∂(BC) is the boundary of a gh. 
An example is shown in Fig. 7.29. Further, as seen in Fig. 7.34, details of the
mapping of the bc into a gh can be easily tracked. That is, A ∩ Au = w̄, c1
and c2 are the parts of cu from the left (which corresponds to m ≥ 1) and the
right of the x axis (for m > −∞) respectively. Similarly for c . As for ellipses,
interior points of the bc are mapped into lines between the two branches of
the gh. This observation has useful consequences discussed at the end of this
section.
 7.7 ♣ FT-8 Convex Sets and Their Relatives 233

Figure 7.29. A bounded convex set bc always transforms to a gh (generalized hyperbola).

This is the generalization of the ellipse → hyperbola transform of Fig. 7.22.

2. UC is a uc ⇒
(a) UC is a gh if UC does not have an ideal point with slope m = 1, i.e.,
1 ∈ [m1 , m2 ].
(b) UC is a uc if UC has an ideal point with slope m = 1, i.e., 1 ∈ [m1 , m2 ].

** Proof. (a) UC has one supporting line with slope m = 1 at a point B ∈

∂(UC). If m1 < 0, ∃V ∈ ∂(UC) with a vertical supporting line v there,35 and if
m2 < 0, ∃H ∈ ∂(UC) with a horizontal supporting line H there. Let c1 be
the boundary portions from V upward, c2 from V to B, c3 from B to H , and
c4 the remainder from H onward. Note that c1 is cu, and c2 , c3 , c4 are cd with
M1 < 0, M2 ≥ 1, M3 ≤ 1, and M4 < 1. Therefore, by Corollary 7.1.4, c1 ∪ c2 is
a cd above B and the vertical line Pm∞2 forming the upper branch of a gh in the xy
plane, while c3 ∪ c4 is a cu providing the lower branch of the gh below B and Pm∞1 .
The argument still holds if some of the ci are not empty, that is, when there does
not exist B or H . It is easy to prove that the image of an interior point A of UC
is a line Ā between the two branches. Hence the strip with endpoints 1/(1 − m2 ),
1/(1 − m1 ) contains the vertical lines that are the images of the ideal points in UC;
this strip is also between the upper and lower branches. 
In the example shown in Fig. 7.30 there is only one ideal point.

35
Note that m1 here corresponds to the upper portion of ∂(UC).
234 7. Curves

Figure 7.30. An unbounded convex set uc with ideal point slope m = 1 transforms to a gh.
This is the generalization of the conic case in Fig. 7.23.

Figure 7.31. A uc with ideal point slope m = 1 transforms to a uc; self-dual case.
This is the generalization of the conics case in Fig. 7.24.

** Proof. (b) UC has supporting lines with slope m ∈ [m1 , m2 ], where m1 < 1 <
m2 . Hence ∃V ∈ ∂(UC) with a vertical supporting line v there. The portion c1 of
∂(UC) from V downward has M1 < 1 and is cd, while the portion c2 upward from
V has M2 ≥ 1 and is cu. Again appealing to the corollary, ∂(UC) = c1 ∪ c2 is
cu. It is formed among other things by the collection of all supporting lines having
slopes m ∈ [m1 , m2 ]. So as shown in Fig. 7.31, the transform contains at least one
ideal point, corresponding to (7.37) for conic parabolas, and is therefore a uc.

 7.7 ♣ FT-8 Convex Sets and Their Relatives 235

Figure 7.32. A gh → to a bc.

Here the asymptotes slopes are m1 < 1 < m2 , with the supporting line’s slope m ∈ [m1 , m2 ]. This
is the generalization of the conic case in Fig. 7.25.

** Proof. (a) Since m1 < 1 < m2 , the slope m of the supporting lines of GH is
not equal to 1. Hence GH is not split and does not have any ideal points ⇒ it is a
bc whose x ranges in an interval having endpoints 1−m1 and 1−m1 , the order, of
1 2
course, being determined by the signs of the mi . 
It is worth studying the example shown in Fig. 7.32 below, tracking the images
of the various portions of the gh.
** Proof. (b) GH has at least one ideal point, which is the image of the point of
GH where the supporting line has slope m = 1, and since GH cannot have any other
supporting lines with m = 1, it has only one component ⇒ GH is a uc. 

** Proof. (c) As for the conics, when a hyperbola →

& ellipse, its conjugate
hyperbola &→ hyperbola. This case corresponds to a gh “conjugate” to that
considered in case (a) as shown in Fig. 7.33. 
Referring again to Fig. 7.34, consider the bc set shown; let us call it SC, and
distinguish it from its subset NSC, where edges AC and DF of SC have been replaced
by AB, BC and DE, EF respectively. The gh we see in -coords is SC. It is also the
image of the convex hull C(NSC) of the nonconvex set NSC. The images of the
points B, E and edges AB, BC, ED, EF are lines and points, respectively, and are
between the upper and lower branches SC. This is the smallest gh that contains
NDS. In effect, one gets the image of the convex hull of a nonconvex set “for free”
in -coords. The precise statement for this result is the following:
Corollary 7.7.2. If NSC is a nonconvex set and gh with OSC ⊃ NSC, then OSC ⊇
C (NSC).
236 7. Curves

Figure 7.33. ♣ FT-8e A gh → gh; self dual.

Here 1 ∈ [m1 , m2 ], mi are the asymptotes’ slopes. Generalization of Fig. 7.27.

Figure 7.34. Interior points of bc → lines between two brances of the corresponding gh.

This result has been used in convex hull construction algorithms [110].36 Here
the edges of the bc may be curves, and hence this approach lends itself to the
construction of convex hulls where some of the “edges” are curved arches (see [92]).

36
The algorithm was improved by Nitzan Giles in 2004 in his course project.
 7.7 ♣ FT-8 Convex Sets and Their Relatives 237

Indications are presented in the next chapter on the feasibility of generalizations

to higher dimensions of the result on the appearance of bounded convex sets in
-coords as generalized hyperbolas.
The generalization of the remarks in Section 7.5.2 as they pertain to ghs is fairly
direct. We have seen that the interior points of a bc are mapped into lines that are
between the two branches of its image, whereas a gh having vertical asymptotes can
only be the image of a uc, as seen in Fig. 7.30. As for the conics, ghs that represent
bcs have branches that can be separated by horizontal lines and ghs that represent
ghs have branches that can be separated by vertical lines.
The proof of the converse result, GC is a gconic ⇒ GC is a gconic, is left as
an exercise.

Exercises

1. Prove that an unbounded convex set having more than one ideal point has ideal
points with direction having slope m ∈ [m1 , m2 ]. That is, either m has a single
value or it ranges over a complete interval.
2. What is the image of a square centered at the origin with

(a) axis-parallel sides,

(b) rotated by 45◦ ?

3. What is the image of the region in the second quadrant above the line 1 : x2 =
−3/4x1 and below the line 2 : x2 = −1/4x1 ?
4. What is the image of the region in the fourth quadrant below the line 1 : x2 =
−3/4x1 and above the line 2 : x2 = −1/4x1 ?
5. What is the image of the regions in the second and fourth quadrants between
the two lines 1 , 2 ?
6. Study the images of the gconics obtained from a double tetrahedron.
7. Examine the images GC of a gconic GC and match the corresponding chains
ci with the corresponding chains ci in GC for the case:

(a) GC is a gh. Distinguish the three cases.

(b) GC is a uc. How can the image of a uc and a gh be distinguished?
(c) GC is a bc.

8. Delimit the class of ghs with the property that if two lines are completely
between the two branches, then the point on the two lines is also between the
two branches. Give a counterexample of a gh without this property.
9. Prove that GC is a gconic ⇒ GC is a gconic.
238 7. Curves

7.7.2 ** Operational Dualities with Gconics

While studying the image of a bc as in Fig. 7.34 and experimenting with set
operations involving gconics and their images, some interesting De Morgan-like
dualities emerged, motivating the following definitions of set operation on two
ghs GHj , j = 1, 2, whose two branches have vertical separation and no vertical
asymptotes, i.e., they can only be images of bcs. Such ghs can be thought of as
consisting of complete lines between their two branches.
Definitio 7.7.3. The gh such that GH ⊇ GH1 ∪ GH2 and such that for any other
gh, GH
, where GH
⊇ GH1 ∪ GH2 ⇒ GH
⊇ GH, is called the outer union37 of
the two ghs. The operation is denoted by GH1 ) GH2 .
An example is shown on the left part of Fig. 7.35. Note that the upper branch
of the outer union consists of all arches (i.e., lines or curves) belonging to the upper
branches of the GHj on or above all points of the GHj . Likewise, the lower branch
consists of all arches belonging to the lower branches that are on or below all points
of the GHj . Where the endpoints are not ideal, the corresponding asymptote is the
line joining the diametrically opposite endpoints of the upper and lower branches.
Note that the “innermost” asymptotes of the ghs become the asymptotes of the outer
union. In general, GH1 ) GH2 = GH1 ∪ GH2 , the outer union being the smallest
gh containing the union of the two ghs.

Figure 7.35. The convex union (also called “convex merge”) of bcs.
It corresponds to the outer union of their images (ghs).

37
This differs from the nomeclature in [110].
 7.7 ♣ FT-8 Convex Sets and Their Relatives 239

Figure 7.36. Inner intersection and intersections are dual.

Definitio 7.7.4. The gh such that GH ⊆ GH1 ∩ GH2 and such that for any other
gh, GH
, where GH
⊆ GH1 ∩ GH2 ⇒ GH
⊆ GH, is called the inner intersection
of the two ghs. The operation is denoted by GH1 * GH2 .
An example is shown on the left part of Fig. 7.36. The inner intersection is the
gh whose upper branch is the cd curve consisting of all vertices of the ghs and the
arches (i.e., lines or curves) joining them on either of the upper branches. Likewise,
the lower branch is the cu curve consisting of all vertices and arches joining them
that are above all points of the lower branches. The asymptotes are the ideal points,
if any. Otherwise, they are the lines joining the diametrically opposite endpoints of
the upper and lower chains. Here the asymptotes may not belong to any of the ghs.
In general, GH1 * GH2 = GH1 ∩ GH2 . It is clear from the definition that the inner
intersection is the largest gh contained in the intersection of the two ghs. Further,
whereas all vertices of the inner intersection are vertices of the GHj , some of the
edges joining such vertices are not edges of the GHj . Roughly speaking, the outer
union of two ghs yields a gh whose branches contain the outermost portions of the
two ghs and any additional edges needed to form the gh, whose significance will
be shortly understood. This is also the reason why it does not always coincide with
the set union of the ghs. The inner intersection, by contrast, contains the innermost
portions of the two ghs. The results below, stated in terms of two bcs BCi , i = 1, 2,
where C (S) denotes the convex hull of the set S, clarify the motivation for these
definitions.
240 7. Curves

Theorem 7.7.5. C (BC1 ∪ BC2 ) = BC1 ) BC2 .38

Proof. The proof is immediate. The convex merge C (BC1 ∪ BC2 ) is the smallest
convex set containing BC1 ∪ BC2 ⇒ C (BC1 ∪ BC2 ) is the smallest gh containing
BC1 ∪ BC2 , whence the result. 
Notice in Fig. 7.35 that all lines in ∂(BC1 ) BC2 ) represent vertices of the ghs,
whereas any additional vertices represent the additional edges needed to form the
convex merge. As expected, the asymptotes represent the vertices of C (BC1 ∪ BC2 )
where there are supporting lines with slope m = 1.
Theorem 7.7.6. BC1 ∩ BC2 = BC1 * BC2 .

Proof. By definition, the inner intersection is the largest gh contained in BC1 ∪ BC2 .
Hence BC1 * BC2 is the largest convex set contained in the inverse image of
BC1 ∩ BC2 , which is also convex. So the two convex sets must be equal and
therefore their images are equal. 
As seen in Fig. 7.36, the edges of BC1 * BC2 that are not edges of any of the
ghs represent the new vertices formed by the intersection of the corresponding bcs.
Note that the asymptotes here represent the two new vertices of the convex merge
and where the supporting lines have slope m = 1. An easy induction proves the
following theorem.
Theorem 7.7.7. C (∪ni=1 BCi ) = )ni=1 BCi .
Since the union of a countable collection of bcs may not be a bc, this theorem
is in general true for finite collections. The corresponding result for outer unions is
true for countable collections of bcs.
Theorem 7.7.8. ∩∞ ∞
i=1 BCi = *i=1 BCi .
These interesting operational dualities have some algorithmic applications
[110].

Exercises
1. Give an example where the outer union of two ghs is not the union of the ghs.
2. Give an example where the inner intersection is not the intersection of the ghs.
3. Find the image of the intersection of the two ellipses ax12 + bx22 = 1 and
x1
b2
+ ax22 = 1.
4. Find the image of the convex merge of the two ellipses above.

38
This operation is also called the convex merge of the two sets BCi .
 7.7 ♣ FT-8 Convex Sets and Their Relatives 241

5. Find the image of the convex merge of two disjoint squares. Investigate various
options concerning position and orientation and the effects on the images.
6. Prove
(a) Theorem 7.7.7.
(b) Theorem 7.7.8.
3 i j
7. Consider the general third-degree curve F (x1 , x2 ) = i+j =0 aij x1 x2 and
(a) show that it can be written in matrix form as
⎛ ⎞
x22
⎜ x1 x2 ⎟
⎜ ⎟
⎜ x12 ⎟
F (x1 , x2 ) = (x2 , x1 x2 , x1 , x2 , x1 , 1)A ⎜
2 2
⎜
⎟,
⎟ (7.41)
⎜ x2 ⎟
⎝ x1 ⎠
1

where A is a symmetric 6 × 6 nonsingular matrix.

(b) Deduce the degree of the image curve from this equation.
(c) (Hard) Apply the transformations of (7.7) to (7.41) to obtain the image
curve.
This page intentionally left blank
8

Proximity of Lines, Planes, and Flats

8.1 ♣ FT-1 Motivation and a Topology for Proximity

In order to apply the results of the representation of flats by indexed points, their
behavior in the presence of errors needs to be understood. This was briefly touched
on in Chapter 5, where we saw that small variations in the coefficients of a plane’s
equation yield a pair of closely spaced point clusters representing the corresponding
“close” planes. The encouraging message is that “closeness” is visually evident.
To pursue this in depth, we need some concepts discussed in Chapter 7. For this
reason, we did not continue with this treatment earlier. While there are many
sources of errors in the applications, from our viewpoint it suffices to consider
the accumulated errors in terms of the resulting variations ci ∈ [ci− , ci+ ] in the
coefficients ci of the plane’s equations. This in turn generates a whole family F
of “close” flats. Even in R 3 , the direct visualization of such a family of planes is
difficult. The challenge is not only the visualization but also the construction of
tools for appraising the proximity, referred to as “error tolerancing,” of the flats in
terms of the errors’ ranges |ci− − ci+ |.
Let us repeat the earlier experiment more systematically and examine the family
of “close” planes

{π : c1 x1 + c2 x2 + c3 x3 = c0 , ci ∈ [ci− , ci+ ], ci− < ci+ }.

Computing the two-point representation in -coords of some of these planes, we

see in Fig. 10.43 the corresponding pair of point clusters. Closeness is apparent, and
more significantly, the distribution of the points is not “chaotic.” The outline of two
polygonal patterns can be discerned. Not only is the family of planes “visualizable,”

A. Inselberg, Parallel Coordinates, DOI 10.1007/978-0-387-68628-8_8,

© Springer Science+Business Media, LLC 2009
244 8. Proximity of Lines, Planes, and Flats

Figure 8.1. Pair of point clusters representing close planes. Note the hexagonal patterns.

but also the variations in several directions. Might it be possible to see, estimate,
and compare errors?
The proper setting for treating the proximity of objects in a family F is by
means of a topology [138]. Every object f ∈ F is endowed with a collection of
subsets Nf = {N|N ⊂ F }, called neighborhoods, such that
1. each f ∈ F has an Nf ,
2. f ∈ N∀N ∈ Nf ,
3. if N1 ∈ Nf and N ⊃ N1 , then N ∈ Nf ,
4. if N1 , N2 ∈ Nf , then N1 ∩ N2 ∈ Nf .
Then proximity is gauged with respect to the topology in the sense that f1 and f2
are “close” if ∃N ∈ Nf1 such that f2 ∈ N and hence also N ∈ Nf2 . Intuitively, we
would like a neighborhood of a flat f to be like a “template,” a region covered by all
flats close to f . In dealing with such regions, the results on gconics from Chapter 7
turn to be helpful, and for this reason this subject is not included, in Chapter 5 on
planes.

8.2 Proximity of Lines and Line Neighborhoods

We start the development by looking into families of lines. Is it reasonable to
consider lines in RN being “close” if the coefficients of their equations are close?39
In R2 consider the collection of lines
39
This differs from the proximity studied in Chapter 4 in terms of the minimum distance between pairs
of lines.
 8.2 Proximity of Lines and Line Neighborhoods 245

F = { | : c1 x1 + c2 x2 = 1, c1 , c2 ∈ R}.

The value of the constant c0 on the right-hand side here is taken as 1 but is allowed
to vary later. Of interest is the collection of subsets N = {NL|NL ⊂ F }, where

NL = { | : c1 x1 + c2 x2 = 1, ci ∈ [ci− , ci+ ] i = 1, 2}. (8.1)

It is convenient to denote the endpoints of the interval [ci−, ci+ ] by signi , the sign
of the superscript of ci± ; that is, signi = − and signi = + stand for the lower and
upper bounds respectively of ci . The extreme lines obtained by the four different
combinations of (sign1 , sign2 ) are
⎫
(−, −) : c1− x1 + c2− x2 = 1, ⎪
⎪
⎪
⎪
− + ⎪
(−, +) : c1 x1 + c2 x2 = 1, ⎬
(8.2)
(+, −) : c1+ x1 + c2− x2 = 1, ⎪
⎪
⎪
⎪
⎪
+ + ⎭
(+, +) : c1 x1 + c2 x2 = 1.

These lines bound the region RNL ⊂ R2 covered by all the lines of NL, a “template”
referred to earlier. An example is shown on the left part of Fig. 8.2, where the extreme
lines in (8.2) are constructed from the points
       
1 1 1 1
P1+ = , 0 , P1− = , 0 , P2+ = 0, + , P2− = 0, − .
c1+ c1− c2 c2

The region RNL resembles a hyperbola and is in fact a generalized hyperbola, a gh

in the terminology of Chapter 7. All lines in NL are enclosed between two opposite
convex unbounded regions with two “asymptotes,” the lines (+, −), (−, +), joining
them at two ideal points.
Clearly, as |ci− − ci+ | → 0 the corresponding lines become more coincident,
and it is fair to consider them as “close.” It is legitimate, then, to consider the RN L
as candidates for “line-neighborhoods” of close lines. Yet no matter how small
the |ci − − ci + | are, the region RNL is unbounded and becomes wider as the xi
increase. Perhaps it is more manageable to study the situation in the parametric
“dual” space with the point (c1 , c2 ) → : c1 x1 + c2 x2 = 1 line correspondence.
Then the rectangles [c1− , c1+ ] × [c2− , c2+ ] ⊂ R2 are alternative candidates for “line-
neighborhoods.” Pursuing the idea further, in some applications, points are given,
together with some lines. The task includes the construction of the “best” (accord-
ing to some criterion) line from these points. Skipping the details, the important
observation here is that we need to work with and view both points and lines. This
246 8. Proximity of Lines, Planes, and Flats

Figure 8.2. A Line neighborhood.

On the left is a region covered by lines “close” to , and on the right are the points in -coords
representing the lines in this region. With reference to the nomenclature and results in Chapter 7,
this is an instance of a gh (a “generalized hyperbola”) on the left that is mapped into a bc, here a
bounded convex quadrilateral (right)

can be accomplished using RN ⊂ R2 but is not possible using the parameter space
in which only the lines, represented by points, are available. Checking the situation
in ||-coords, shown on the right part of Fig. 8.2, the unbounded region RN L is
transformed into a simple convex quadrilateral; let us call it NL. From the point ↔
line duality, both points and lines are available to work with.
The intention is to form a topology; therefore the possible intersections between
different quadrilaterals NL need to be examined. Let
− +
NLj = { | : c1,j x1 + c2,j x2 = 1, ci,j ∈ [ci,j , ci,j ] i = 1, 2j = 1, 2},
(8.3)
and the intersection operation N = N1 ∩ N2 is defined by
NL = { | : c1 x1 + c2 x2 = 1, ci ∈ [ci− , ci+ ] = [ci,1
− +
, ci,1 −
] ∩ [ci,2 +
, ci,2 ] i = 1, 2}.
(8.4)
 8.2 Proximity of Lines and Line Neighborhoods 247

The different intersection outcomes are these:

1. if [ci− , ci+ ] = ∅ for any i, then NL = ∅,

2. if [ci− , ci+ ] = ci for i = 1, 2, then NL = { } for : c1 x1 + c2 x2 = 1;
3. if [c1− , c1+ ] = c1 and c2− < c2+ , then NL is a “fan” of lines on the point
P1 = ( c11 , 0) ⇒ NL is a segment on P¯1 . Similarly, if [c2− , c2+ ] = c2 and
 
c− < c+ , NL is a segment on P¯2 where P2 = 1 , 0 ; or
1 1 c2
4. NL is a quadrilateral contained in both NL1 and NL2 .

Though the set intersection of two quadrilaterals may be a triangle, the intersection
operation defined by (8.4) excludes that possibility (see the exercises below). With
each of these cases a legitimate NL, it is clear that the family NL = {NL|NL ⊂ F }
is a topology of line-neighborhoods.
By contrast, the intersection of two RN L is not in general such a region. Four
distinct RNL are shown in Fig. 8.3, and though some pairs have a nonempty
intersection, it does not contain any complete lines. This is clearly seen from the
corresponding NL, which are disjoint. So the regions RN L considered as point sets
cannot serve as neighborhoods for a topology for lines. The alternative is to consider

Figure 8.3. Several line neighborhoods.

The regions RNLi , i = 1, 2, 3, 4, covered by four families of lines in orthogonal coordinates and
their images NLi in || coordinates.
248 8. Proximity of Lines, Planes, and Flats

RNL as consisting of the complete lines covering it, greatly complicating matters.
Instead, we choose to work with their images NL, which satisfy all the requirements
for a useful topology of lines. Additionally, with ||-coords we are not limited by
the dimensionality of the space either for lines or more general flats. This leads to
some surprising, elegant, and very useful generalizations.
Before getting to that, we enlarge the class of lines in N replacing the 1 by c0
and allowing it to vary within an interval

NL = { | : c1 x1 + c2 x2 = c0 , ci ∈ [ci− , ci+ ] i = 0, 1, 2}. (8.5)

Each such line is represented by the point


¯ c2 c0
= , . (8.6)
c1 + c2 c1 + c2

In order to study the resulting neighborhood NL, first we take c0 = c0− . This is
basically the case already studied, and the corresponding neighborhood, whose
construction is shown in the lower part of Fig. 8.4, is formed using the points
with values of c0− in the numerator. The vertices of the quadrilateral are labeled by
(sign1 , sign2 , −), since they are the points representing the four extreme lines as
before. Next the value c0 = c0+ is chosen and the corresponding quadrilateral is
constructed now using the points with values of c0+ in the numerator with vertices
(sign1 , sign2 , +). Inspecting the y-coordinate of the point given by (8.6), it is clear
that for the intermediate values c0 = [c0− , c0+ ] between the two endpoints,

c0− c0 c0+
≤ ≤ ,
c1 + c 2 c1 + c 2 c1 + c 2
the vertices of the quadrilaterals are translated upward from the quadrilateral
obtained from c0− until the vertices for c0+ with the same x-coordinate. The
resulting NL is now a hexagon with the leftmost and rightmost edges verti-
cal. The vertices going counterclockwise are (−, +, −), (+, +, −), (+, −, −),
(+, −, +), (−, −, +), (−, +, +). For clarity the three quadrilaterals corresponding
to c0− , c0+ , and c0 ∈ [c0− , c0+ ] are also shown separately in the upper portion of
Fig. 8.5 with fewer details.
We reexamine the prospects for a topology using the hexagons as line-
neighborhoods, starting with
− +
NLj,0 = { | : c1,j x1 + c2,j x2 = c0 , ci,j ∈ [ci,j , ci,j ], i = 0, 1, 2, j = 1, 2},
(8.7)
 8.2 Proximity of Lines and Line Neighborhoods 249

Figure 8.4. Construction of a neighborhood NL for NL = { | : c1 x1 + c2 x2 = c0 , ci ∈

[ci− , ci+ ] i = 0, 1, 2}. The “exploded” view also shows the quadrilaterals NL− for c0 = c0− and
NL+ and c0 = c0+ , whose vertices are marked with black and blank oval vertices respectively. The
complete NL is a hexagon having two vertical edges, three vertices from NL− , the lowest ones, and
three vertices from NL+ .

Figure 8.5. The hexagon of the previous figure without the construction details.
The arrow shows that the three diagonals intersect at a point (Exercise 6).
250 8. Proximity of Lines, Planes, and Flats

for the intersection NL = NL1 ∩ NL2 defined by

NL0 = { | : c1 x1 + c2 x2 = c0 , ci ∈ [ci− , ci+ ]

− + − +
= [ci,1 , ci,1 ] ∩ [ci,2 , ci,2 ], i = 0, 1, 2}. (8.8)

♣ FT-1e
The different outcomes of the intersection are then:
1. If [ci− , ci+ ] = ∅ for any i, then NL = ∅.
2. If [c0− , c0+ ] = 0 and [ci− , ci+ ] = 0 for either i = 1 or i = 2, then is on the
x2 -axis if i = 1 and on the x1 -axis if i = 2.
3. If [c0− , c0+ ] = 0 and [ci− , ci+ ] = 0 for both i = 1, 2, then RN L consists
of two overlapping “fans” of lines at the origin (0, 0), so NL consists of two
overlapping intervals on the x-axis. In effect, this is the case where the hexagon
(see Fig. 8.5) collapses to a horizontal line.
4. If [c0− , c0+ ] = c0 = 0, then this yields a quadrilateral NL with the cases
enumerated earlier.
5. If c0− < c0+ and ci is constant for either i = 1, 2, then NL is a quadrilateral with
two vertical sides and RNL consists of two strips each having constant width.
6. NL is a hexagon contained in both NL1,0 and NL2,0 .
Again we can conclude that the family of hexagons NL0 constitutes a topology
of line-neighborhoods, since all the enumerated situations are also specific instances
of NLj,0 . Allowing the variation of c0 ∈ [c0− , c0+ ] is the most general case for a
family of lines. To complete the study there still remains to examine the situation in
which the sum c1 + c2 = 0 with some of the corresponding lines, having slope 1,
is represented by ideal points.
The generalization to RN is direct, a family of close lines is represented by N −1
regions, and it comes out as a special case in what follows (see also Exercise 8
below). For earlier treatments of line-neighborhoods and topologies for flats in
||-coords see [107] and [23]. The exposition here is adapted from [131].

Exercises
1. Show that the intersection, as defined by (8.4), of two quadrilaterals NLj
cannot be a triangle. Draw the RNL corresponding to a triangular NL contained
between the X̄1 and X¯2 axes and describe the family of lines NL.
2. Draw examples of the quadrilaterals NLj for j = 1, 2 and the RN Lj for all the
different cases of the intersections NL = NL1 ∩ NL2 .
3. Show that when NL, as given by (8.1), is a quadrilateral, it is also convex.
 8.2 ♣ FT-2 Proximity of Hyperplanes 251

4. Construct an example in which NL is not convex and describe the region and
that of RNL .
5. Show that when NL0 , as given by (8.7), is a hexagon, it is also convex.
6. Show that the three diagonals of the hexagon having exactly opposite signs
(i.e., (-,+,+) is joined to (+,-,-)) intersect at the same point; see also Fig. 8.5.
7. Draw examples of the hexagons NLj,0 , for j = 1, 2 and the RN Lj,0 for all the
different cases of the intersections NL0 = NL1,0 ∩ NL2,0 .
8. Generalize to RN the representation of a family of close lines and provide an
algorithm for constructing the line neighborhoods; this is a good topic for a
course project.

8.3 ♣ FT-2 Proximity of Hyperplanes

8.3.1 Formulation of the Problem in RN

In R3 , for the plane

π : c1 x1 + c2 x2 + c3 x3 = c0 , (8.9)

c2 + 2c3 + 3c1 c0
π̄1
= π̄1
23 = ,
c1 + c 2 + c 3 c1 + c 2 + c 3

c3 + 2c1 c0
= (1, 0) + , , (8.10)
c1 + c2 + c3 c1 + c2 + c3
an observation showing that π̄1
= π̄1
23 can be obtained from

c2 + 2c3 c0
π̄0
= π̄123 = ,
c1 + c2 + c3 c1 + c2 + c3
by a shift and the counterclockwise cyclic permutation c1 → c3 , c3 → c2 , c2 → c1 .
This combinatorial relation allows for a reduction in the number of indexed points
that need to be considered. As a reminder, any two points in the xy plane labeled
π̄123, π̄1
23 with the same y-coordinate are required for a valid representation of
π , the horizontal distance between them being proportional to c1 . Tracing the
permutation between the ck also clarifies that the value of π̄1
23 can not be obtained
from π̄123 without additionally specifying a value for c1 . Continuing,

c1 + 2c2 c0
π̄2
= π̄1
2
3 = (2, 0) + , , (8.11)
c1 + c2 + c3 c1 + c2 + c3
252 8. Proximity of Lines, Planes, and Flats

as obtained from π̄0

= π̄123 via a shift and the clockwise cyclic permutation
c1 → c2 , c2 → c3 , c3 → c1 . Finally,

c2 + 2c3 c0
π̄3
= π̄1
2
3
= (3, 0) + , = (3, 0) + π̄0
. (8.12)
c1 + c2 + c3 c1 + c2 + c3

The analogous relations hold for RN . For a hyperplane π : c1 x1 + c2 x2 + · · · +

cN xN = c0 let c = (c1 , c2 , . . . , cN ) be the N -tuple of the coefficients
N and for an
N -tuple Z = (z1 , z2 , . . . , zN ), denote the sum by S(Z) = j =1 zj . According
to Corollary 5.4.2, the first and second points representing π are

c2 + · · · + (k − 1)ck + · · · + (N − 1)cN c0
π̄12...N = π̄0
= , , (8.13)
S S

Nc1 + c2 + · · · + (k − 1)ck + · · · + (N − 1)cN c0
π̄1
= , , (8.14)
S S
N
where S = S(c) = j =1 cj . Here

π̄1
= (1, 0)

c3 + 2c4 + · · · + (k − 2)ck−1 + · · · + (N − 2)cN + (N − 1)c1 c0
+ ,
S S
(8.15)

is obtained from π̄0

by a shift and the counterclockwise cyclic permutation c1 →
cN → cN−1 → · · · → ck → ck−1 → · · · → c2 → c1 generalizing (8.10), a
process that can be continued inductively for the remaining

π̄i


Nc1 + (N + 1)c2 +· · · +(N +i −1)ci + ic(i+1) + · · · +(N −1)cN c0
= , ,
S S
(8.16)

obtained from π̄0

by a shift of (i, 0) and an appropriate permutation of the ck . This
is applied, amplified, and better understood in Section 8.3.6.
We proceed now with the study of the class of all hyperplanes

F = {π|π : c1 x1 + c2 x2 + · · · + cN xN = 1, ci ∈ R for i = 1, 2, . . . , N},

 8.3 ♣ FT-2 Proximity of Hyperplanes 253

and the collection of neighborhoods Nπ = {NH|NH ⊂ F }, where

NH = {π|π : c1 x1 + c2 x2 + · · · + cN xN = 1, ci ∈ [ci− , ci+ ] i = 1, 2, . . . , N}.
(8.17)
Note that for convenience, the value of c0 = 1 is used and is adjusted later to arbitrary
values as for line neighborhoods. In the study of the N − 1 regions composing NH,
in view of the previous reduction, it suffices to investigate the region  = NH0

containing the points π̄0

. These are the function values of
fN : RN &→ R2
in x, y coordinates,
 N 
j =1 (j − 1)cj 1
fN (c) = (x(c), y(c)) = , , (8.18)
S S

reverting to homogeneous coordinates to accommodate ideal points. The subscript

“0
” is dropped when the context is clear. The cj range over the N-dimensional
“box” is
B = [c1− , c1+ ] × · · · × [cN
− +
, cN ] ⊂ R N .
The stage is now set for the task ahead: to understand the properties of fN , its action
on B,

fN
B &−→  = fN (B) ⊂ P2 ,

and the structure of  representing the family of hyperplanes in N .

The previously studied line neighborhoods serve as a guide and prototype. The
region of “close” lines, Fig. 8.2, is a gh (generalized hyperbola), and from the
previous chapter we know that its -coords image can be any bc (bounded convex
set), the case covered so far, uc (unbounded convex set), or gh. The region in RN
covered by “close” hyperplanes is a complex N-dimensional gh whose image in
-coords consists of N − 1 regions in R2 . As will be seen, these are also bc, uc, gh.
From these regions the properties of “close” hyperplanes can be ascertained and
visualized without loss of information.

8.3.2 Outline of the Development

The key steps in the ensuing are outlined here so that the development will not be
obscured by the details.
254 8. Proximity of Lines, Planes, and Flats

1. At first, only the region  containing the points π̄0

need be considered with
the constant on the right-hand side c0 = 1.
2. Analysis of the function fN in Section 8.3.3 enables the construction of  as
the intersection of N “fan” regions shown in Fig. 8.10. The result is a polygonal
region whose vertices are the values of fN at 2N key vertices of the hyperbox
B given by (8.45) and traversed along the path (8.46).
3. The polygonal regions are either convex bounded or unbounded or gh with
2N vertices depending on the intersection of B with the hyperplane πc : c1 +
c2 + · · · + cN = 0 shown in Fig. 8.24. The computational complexity of the
construction is O(N).
4. The remaining N − 2 regions are then easily obtained by variations in the
traversal of the path in (8.46).
5. Allowing c0 = 1 is dealt with following (8.55), adding two vertical edges to
the previous polygonal region.
6. In the last section, points within the regions are matched, so that together they
represent a valid element (i.e., plane or hyperplane) of the family. This serves
as a prototype for the equivalent problem in the surface representation.
♣ FT-2e

8.3.3 Properties of fN
The coordinate x of fN can be found in terms of the second coordinate y = 1/S
via the decomposition below, for k = 1, 2, . . . , N, as
N 
j =1 (j − 1)cj (k − 1)S (1 − k)S + N j =1 (j − 1)cj
x= = +
S S S
N N N
S − j =1 kcj − S + j =1 j cj j =1 (j − k)cj
= (k − 1) + = (k − 1) + ,
S S
yielding the linear relation

P̄ k : x = C k y + (k − 1), (8.19)

with coefficient

k−1 
N
C k = C k (c1 , c2 , . . . , ck−1 , ck+1 , . . . , cN ) = (j − k) cj + (j − k) cj ,
$ %" # $ %" #
j =1 <0 j =k+1 >0
(8.20)
 8.3 ♣ FT-2 Proximity of Hyperplanes 255

and
⎧ 
⎨ C 1 = C 1 (c2 , c3 , . . . , cN ) = N j =2 (j − 1)cj ,
⎩ C N = C N (c , c , . . . , c N −1 (8.21)
1 2 N −1 ) = j =1 (j − N)cj .

In the above decomposition, the first sum has negative terms and the second
positive terms. Further, it is clear that C k is independent of ck . The line P̄ k
having slope 1/C k is on the point (k − 1, 0). Each point on P̄ k is the π̄0
for a
hyperplane π.
The properties and crucial role of the P k in the development are better under-
stood in terms of the underlying geometry. For a plane π ⊂ R3 as in Fig. 8.6, note
that P 1 , P 2 , P 3 ∈ π = π ∩ π1s with x1 (P 1 ) = 0, x2 (P 2 ) = 0, x3 (P 3 ) = 0,
and ¯π = π̄123 . Recall that the coefficient vector c = (c1 , c2 , c3 ) is normal to the
plane π . Translation of ¯π along P̄ 1 corresponds to a rotation that leaves c2 and
c3 invariant, since C 1 is a constant, while c1 varies. The situation is illustrated in
Fig. 8.7. On P1 , the normal vector ranges from c = (c1− , c2 , c3 ) to c = (c1+ , c2 , c3 ),
tracing a plane πc1 orthogonal to all the planes with such c. These planes π rotate
about the axis n1 through P1 and normal to πc1 with π intersecting the superplane

Figure 8.6. A plane π and its intersection π = π̄123 with the first superplane π1s .
The points P k = π ∩ xi xj plane for k = i, j .
256 8. Proximity of Lines, Planes, and Flats

Figure 8.7. The vector c = (c1 , c2 , c3 ) is normal to π .

On P1 , varying just c1 ∈ [c1− , c1+ ] generates the plane πc1 where c ranges. The translation of
¯π = π̄123 ∈ P̄ 1 (right) corresponds to the rotation of π about the axis n1 ⊥ πc1 through P1 . The
translations along P̄ 2 , P̄ 3 leaving c2 , c3 invariant respectively correspond to similar rotations about
lines on P 2 and P 3 .

π1s at different lines π corresponding to different points ¯π = π̄123 translated

along P¯1 ; clearly n1 ⊂ π1s . The same considerations apply for P2 and P3 . The
generalization to RN is direct, with the translation of ¯π along P̄ k corresponding
to an equivalent rotation with C k , which determines P̄ k (and hence P k ), being
independent of ck .
Let
⎧ k
⎪
⎪ Cm = C k (c1+ , . . . , ck−1
+ −
, ck+1 −
, . . . , cN ),
⎪
⎪
⎪
⎪ k = C k (c− , . . . , c− , c+ , . . . , c+ ),
⎨ CM 1 k−1 k+1 N
(8.22)
⎪
⎪ 1 = C 1 (c− , . . . , c− ), C N = C N (c+ , . . . , c+ ),
⎪ m
⎪
C 2 N m 1 N −1
⎪
⎪
⎩ 1 + + N = C N (c , . . . , c− ).
−
CM = C 1 (c2 , . . . , cN ), CM 1 N −1
 8.3 ♣ FT-2 Proximity of Hyperplanes 257

Since cj− ≤ cj ≤ cj+ ,

k−1 − k−1 k−1 +
⎫
j =1 (j − k)cj ≥ j =1 (j − k)cj ≥ j =1 (j − k)W icj ⎬ k
N ⇒ Cm ≤ C k ≤ CMk
.
+ N N −⎭
j =k+1 (j − k)c j ≥ j =k+1 (i − k)cj ≥ j =k+1 (j − k)c j
(8.23)

That is, Cmk and C k are the minimum and maximum values of C k .
M
The multivariate function fN can be analyzed one variable at a time. Specifi-
cally, for ck , the region of interest is swept by the rotation of the line P̄ k about the
point (k − 1, 0) between the lines
 k
P̄m : x = Cmk y + (k − 1),
(8.24)
k : x = C k y + (k − 1),
P̄M M

from the minimum 1/CM k to maximum 1/C k slopes. Caution: the slopes are the
m
reciprocals of the C k . The variation of ck is due to this rotation, whose dual is a
translation of P k but within the superplane π1s , so that its representation P̄ k remains
a line (rather than a polygonal line).
For R3 , translations of the three P k within the segments [P k , Phk ] ⊂ π1s are
indicated on the left part of Fig. 8.8. The subscripts , h are used for the endpoints
Pmk or PM k instead of m, M, since their relative positions depend on the signs of

Cmk , C k , and this is irrelevant to the discussion now. In -coords (see the right part
M
of Fig. 8.8) the intersection of the three sectorial regions swept by the dual rotations
about the points (k − 1, 0) contains the points π̄0
corresponding to the planes
generated by the simultaneous translations of the P k , giving us a first glimpse
at a region  and explaining the hexagonal patterns seen in Fig. 10.43. All this
begs the question: are there equivalences between translations – rotations? This is
answered with Fig. 8.9 showing translation – rotation duals. Starting then from a
plane π, a family of “close” planes in a neighborhood N0
can be generated from
π by the aforementioned rotations and/or the translations of the P k shown in the
picture. In due course, vertical translations for an arbitrary c0 in π ’s equation will
be added. The technical details that follow do not obscure our understanding when
the underlying geometry is kept clearly mind.

8.3.4 The Region 

Components and Structure of 
After the short and incomplete preview in R3 , we proceed with the generalization
to RN . Let us denote the “fan” regions swept between P̄Mk and P̄ k by  . Now
m k
258 8. Proximity of Lines, Planes, and Flats

Figure 8.8. Rotation of P¯k about (k − 1, 0).

It is dual to the translation of P k within a segment [P k , Phk ] ⊂ π1s , so that P¯k remains a straight
(rather than polygonal) line. The intersection of the three sectorial regions on the right is the
region .

it is imperative, in order to cover all possibilities, to examine the positions of Pmk

relative to PMk.

Case 1. 0 < Cm k ; then 1/C k ≥ 1/C k ≥ 1/C k , and P̄ k ranges between

m M
k
the lines P̄M with minimum and P̄mk maximum slopes respectively; first part of
Fig. 8.10.

Case 2. Cm k < 0 < C k ; then 1/C k < 0 < 1/C k , and P̄ k ranges between the
M m M
lines P̄mk with minimum and P̄M k maximum slopes respectively. Due to the sign

change, C k goes through zero when P̄ is vertical at (k − 1, 0), corresponding to

P being an ideal point and therefore π being parallel to the xk axis; second part of
Fig. 8.10.

Case 3. Finally, when CM k < 0, 0 > 1/C k > 1/C k > 1/C k , and P̄ k ranges
m M
between P̄M k with the minimum and P̄ k with maximum slopes respectively; third
m
part of Fig. 8.10.
 8.3 ♣ FT-2 Proximity of Hyperplanes 259

Figure 8.9. Within  a translation along P̄ 2 of π̄0

to a new position.
It can also be achieved by two translations, first along a P̄s3 then along a P̄f1 , or equivalently by two
rotations P̄s1 &→ P̄f1 followed by P̄s3 &→ P̄f3 .

For the case S(c) = 0, this analysis breaks down, since y = 1/S is no
longer defined, necessitating the consideration of fN (8.49) in homogeneous
coordinates.

Lemma 8.3.9. ♣ FT-3  = ∩N

k=1 k . Either

(a)  is a convex set strictly above or below the x axis, or

(b)  consists of a gh region, the upper branch above and lower branch below
the x axis.

** Proof. Each k consists of two sectorial (and of course convex) regions sym-
metric about the point (k − 1, 0). When the intersection of the portions below the
x axis is empty (as in Fig. 8.8, for example), then  is above the x axis and is
convex, since it is the intersection of convex regions. Similarly, it may happen that
all the k intersect below the x axis. Otherwise, some of the k have a nonempty
intersection that is convex above, and others may have a convex intersection below
the x axis. 
In the second case, it will be seen that there exists a combination of the cj for
which S(c) = 0, resulting in  being a gh. For the construction of  we take our
260 8. Proximity of Lines, Planes, and Flats

Figure 8.10. “Fan” regions k swept by the rotations of P̄ k about the point (k − 1, 0).

k ∩ P̄ k+1
cue from Fig. 8.8 and investigate the intersections P̄mk ∩ P̄mk+1 and P̄M M
k and C k :
starting with the relations between the successive Cm M


k−2 
N
(j − k + 1)cj+ + (j − k + 1)cj−
(k−1)
Cm =
j =1 j =k
⎛ ⎞

k−2 
N 
k−2 
N
= (j − k)cj+ + (j − k)cj− + ⎝ cj+ + cj− ⎠
j =1 j =k j =1 j =k
⎛ ⎞

k−1 
N 
k−1 
N
= (j − k)cj+ + (j − k)cj− + ⎝ cj+ + cj− ⎠ .
j =1 j =k+1 j =1 j =k
 8.3 ♣ FT-2 Proximity of Hyperplanes 261

k and rewriting,
Carrying out similarly for CM
⎧  
⎪
⎨ Cm
(k−1) k + k−1 c+ + N c− ,
= Cm j =1 j j =k j
 N  (8.25)
⎪
⎩ C (k−1) = C k + k−1 − +
M M c
j =1 j + c
j =k j ,

which points out the key vertices on the boundary ∂B:

⎧
⎨Ak = (c1+ , c2+ , . . . , ck−1
+
, ck− , ck+1
− −
, . . . , cN ),
with
⎩B = (c− , c− , . . . , c− , c+ , c+ , . . . , c+ ),
k 1 2 k−1 k k+1 N
⎫
A1 = (c− , . . . , c− ) = BN ,⎬
1 N
(8.26)
+
B1 = (c1+ , . . . , cN ) = AN ,⎭
so that (8.25) becomes
⎧
⎨ Cm
(k−1)
= Cm
k + S(A ),
k
(8.27)
⎩ (k−1)
CM = k
CM + S(Bk ).
These 2N vertices are characterized by a single change of sign in the superscripts
at the kth position. Pursuing the cue in Fig. 8.8, we obtain the successive line
(k−1)
intersections P̄mk ∩ P̄m by
k (k−1) 1
Cm y + (k − 1) = Cm y + (k − 2) ⇒ y= , (8.28)
S(Ak )
k , so that for k = 2, . . . , N − 1,
and correspondingly for P̄M
 
Cmk 1
k k (k−1)
ām = P̄m ∩ P̄m = fN (Ak ) = (x(Ak ), y(Ak )) = + (k−1), ,
S(Ak ) S(Ak )
(8.29)
 
C k 1
k k (k−1) m
āM = P̄M ∩ P̄M = fN (Bk ) = (x(Bk ), y(Bk )) = + (k−1), .
S(Bk ) S(Bk )
(8.30)
For k = 1 and k = N,
P̄m1 ∩ P̄M
N
= f (A1 ) = f (BN ), 1
P̄mN ∩ P̄M = f (B1 ) = f (AN ). (8.31)
262 8. Proximity of Lines, Planes, and Flats

Other nice and useful relations for 1 ≤ r < k are

(r+1) (r+1) (k−1)


k
r k r k
Cm − Cm = Cm − Cm + Cm − · · · + Cm − Cm = S(Aj ),
j =r+1
(8.32)

(r+1) (r+1) (k−1)


k
r k r k
CM − CM = CM − CM + CM − · · · + CM − CM = S(Bj ).
j =r+1
(8.33)

From these we can find the intersection points

rk (k−1) (k − r)
ām = P̄mr ∩ P̄m , rk
y(ām ) = k , (8.34)
j =r+1 S(Aj )

rk k (k−1) rk (k − r)
āM = P̄M ∩ P̄M , y(b̄M ) = k , (8.35)
j =r+1 S(Bj )

and specifically for r = 1,

1k (k−1) (k − 1)
ām = P̄m1 ∩ P̄m , 1k
y(ām ) = k , (8.36)
j =2 S(Aj )

rk k (k−1) rk (k − 1)
b̄M = P̄M ∩ P̄M , y(b̄M ) = k . (8.37)
j =2 S(Bj )

There are also some important monotonicities with S(B1 ) ≥ S(A1 ) and
 + N −
⎫
S(Ak ) = k−1j =1 cj + j =k cj ⎬
 N ⇒ S(Ak−1 ) ≤ S(Ak ), (8.38)
+ − ⎭
S(Ak−1 ) = k−2 c
j =1 j + j =k−1 c j

showing that
1 1
= y(AN ) ≤ y(AN −1 ) ≤ · · · ≤ y(A1 ) = .
S(AN ) S(A1 )
Similarly, S(Bk−1 ) ≥ S(Bk ) ⇒
1 1
= y(BN ) ≥ y(BN−1 ) ≥ · · · ≥ y(B1 ) = .
S(BN ) S(B1 )
 8.3 ♣ FT-2 Proximity of Hyperplanes 263

If there is a combination of the coefficients c such that S(c) = c1 +c2 +· · ·+cN = 0,

while these ck are in [ck− , ck+ ], then y(Ak ) and y(Ak+1 ) have different signs, so
that the above monotonicities become

y(Ak ) ≤ y(Ak−1 ) ≤ · · · ≤ y(A2 ) ≤ y(A1 ) < 0 < y(AN )

≤ y(AN−1 ) ≤ · · · ≤ y(Ak+1 ); (8.39)

showing that at most only one change of sign in the sequence of the y(Ak ) can
occur. Similarly, at most one change of sign can occur in the sequence of the y(Bk ),

y(Br ) ≤ y(Br+1 ) ≤ · · · ≤ y(BN ) = y(A1 ) < 0 < y(B1 )

= y(AN ) ≤ · · · ≤ y(Br−1 ). (8.40)

8.3.5 ** Construction of 
Consider the first part of the lemma and specifically when  is above the x-axis.
All values of y(B) are greater than 0 for any point B ∈ B, and hence y(Ak ) > 0,
(k−1) k and C (k−1) > C k from (8.25). There are three
y(BK ) > 0, so that Cm > Cm M M
k:
possible variations for the slopes of P̄mk , P̄M
1. all slopes positive,
1 1 1 1 1 1
N
≥ (N−1)
≥ ··· ≥ 1
> 0, N
≥ (N −1)
≥ ··· ≥ 1
> 0;
Cm Cm Cm CM CM CM
(8.41)

1 ≥ · · · ≥ Ck > 0 > C (k+1)

2. slopes change sign i.e. Cm m m ≥ · · · ≥ Cm
N ; hence

1 1 1 1 1
k
≥ ··· ≥ 1
> 0 > N ≥ · · · ≥ (k+1) , k
Cm Cm Cm Cm CM
1 1 1
≥ ··· ≥ 1
>0> N
≥ ··· ≥ (k+1)
; (8.42)
CM CM CM

3. all slopes are negative,

1 1 1 1 1 1
0> N
≥ (N−1)
≥ ··· ≥ 1
, 0> N
≥ (N −1)
≥ ··· ≥ 1
.
Cm Cm Cm CM CM CM
(8.43)
264 8. Proximity of Lines, Planes, and Flats

Figure 8.11. Orienting the half-spaces.

For each side of a line : ax + by = c for c ≥ 0. Points in the portion “−” : ax + by < c are “below”
and “above” with the reverse inequality. For horizontal lines, “+” is the positive y direction, and
for vertical lines, the positive x direction.

Figure 8.12. Starting the construction of .

k are all positive;  is enclosed within the quadrilateral  ∩  .
Here the slopes of the lines P̄mk , P̄M 1 N

Figure 8.11 recalls the “above”/“below” relation for a point P and a line ,
which is helpful in what follows. Recall from Lemma 8.3.9 that  ⊂ k . The
intersection  = 1 ∩ N = 1N , where 1 is the sectorial region between
 8.3 ♣ FT-2 Proximity of Hyperplanes 265

Figure 8.13. The slopes of the lines P̄mk , P̄Mk change sign.
 is enclosed within the quadrilateral 1 ∩ N .

1 , P̄ 1 , and  between P̄ N , P̄ N , is a quadrilateral with f (A ) = f (B ),

P̄M m N M m N 1 N N
fN (AN ) = fN (B1 ) its highest and lowest vertices respectively. For case 1, shown
in Fig. 8.12, all of  is to the right of the X̄k axis and is of course contained in
the quadrilateral 1N . For 2, shown in Fig. 8.13,  is between the X̄1 and X̄N
axes, the slopes of the lines changing sign at some intermediate index according
to (8.41). In the third case, (8.43),  is on the left of the X̄1 axis, as seen in
Fig. 8.14. For  below the x axis, all the values of y(B) are negative for any point
(k−1)
B ∈ B, and hence y(Ak ) < 0, y(BK ) < 0. Then from (8.25), Cm < Cmk and
(k−1) k , and the three possibilities for the slopes of P̄ k , P̄ k analogous to
CM < CM m M
(8.41), (8.42), (8.43) result in  being the mirror images, with respect to the x
axis, of the previous situation, with fN (A1 ) still higher than fN (AN ) as in (8.39),
(8.40).
It is clear that the vertices fN (A1 ), fN (AN ) together with a segment from each
of P̄m1 , P̄MN belong to ∂. Therefore, f (A ), f (A ) ∈  for any k. Hence,
N 1 N N k
as illustrated in Fig. 8.15, neither of the lines P̄mk , P̄Mk intersects the line joining
fN (A1 ), fN (AN ), for if one of them did, then fN (A1 ), fN (AN ) would lie on
opposite sides of that line, so that both of them cannot belong to ∂. At this stage,
k is immaterial, but for the sake of the discussion we take the
the order of the P̄mk , P̄M
266 8. Proximity of Lines, Planes, and Flats

Figure 8.14. The slopes of the lines P̄mk , P̄Mk are all negative.
 is enclosed within the quadrilateral 1 ∩ N .

arrangement shown with n̄ik = P̄M k ∩ P̄ 1 higher in y than y(A ). Hence n̄ ∈ ∂,
m 1 ik
meaning that the segment of P̄m1 belonging to ∂ has an enpoint āik = P̄m1 ∩ P̄mk for
some k. By the same argument, there is an endpoint b̄ik = P̄M 1 ∩ P̄ k ∈ ∂. From
M
(8.36) and (8.39), a comparison of the y coordinates of the points ā12 = fN (A2 )
and ā1k yields
(k − 1) (k − 1) 1
y(ā1k ) = k ≤ k = = y(A2 ), (8.44)
S(A 2)
j =2 S(Aj ) j =2 S(A2 )
illustrated in Fig. 8.16, showing that the point ā12 is fN (A2 ) ∈ ∂ with ā1k ∈ 
for k ≥ 3. So it is the segment with endpoints fN (A1 ), fN A2 that belongs to ∂,
and similarly for the vertices fN (BN ), fN (BN −1 ) ∈ ∂ together with the segment
joining them.
Let Ak , Ak+1 be the endpoints of the edge ak , and Bk , Bk+1 the endpoints of
the edge bk , respectively, on ∂B. Then
⎧
⎨ ak = (c1+ , c2+ , . . . , ck−1
+ −
, ck , ck+1 −
, . . . , cN ),
c ∈ [ck− , ck+ ]. (8.45)
⎩ b = (c− , c− , . . . , c− , c , c+ , . . . , c+ ), k
k 1 2 k−1 k k+1 N
 8.3 ♣ FT-2 Proximity of Hyperplanes 267

Figure 8.15. The segment connecting fN (A1 ) and fN (AN ) is contained in k for all k .
The point n̄ik is equal to P̄m1 ∩ P̄M
k ∈ ∂ for k ≥ 2, but the point ā equals P̄ 1 ∩ P̄ k ∈ ∂ for
1k m m
some k.

Restating, we have established that the image fN (a1 ) of the edge a1 joining
A1 , A2 , and similarly the edge fN (b1 ) with endpoints fN (BN ), fN (BN −1 ),
belongs to  (recall that A1 = BN ). Inductively, the same argument shows that
fN (ak ), fN (bk ) ⊂ ∂  ∀ k = 1, . . . , N. We have shown that the construction
of  is accomplished by sweeping along the x axis to the points (0, 0), . . . ,
(k −1, 0), . . . , (N −1, 0) where the sectorial regions k are anchored, intersecting
them sequentially, (1 ∩ 2 ) ∩ · · · ∩ k , terminating with N to generate , as
illustrated in Fig. 8.18.
The three possibilities for the segments forming ∂ are shown in Fig. 8.17,
the third case (on the right) still to be discussed. In due course it will be seen that
the segments can be either semi-infinite lines or projective complements of finite
segments. Equivalently, the construction of  corresponds to the judicious traversal
of the portion of ∂B consisting of the 2N vertices Ak , Bk , specified in (8.45), on
the path P along the 2N edges a1 , a2 , . . . , aN , b1 , b2 , . . . , bN . At each vertex,
say Ak , the point fN (Ak ) in the xy plane is computed and joined by a straight line
segment to the previous one fN (Ak−1 ). To emphasize, out of a total of 2N vertices
(and edges) on ∂B, only 2N are needed for the construction of . The traversal
268 8. Proximity of Lines, Planes, and Flats

Figure 8.16. The segment of P̄m1 with endpoints fN (A1 ), fN (A2 ).

It is included in ∂, with the triangular region shown excluded from .

Figure 8.17. The segments fN (ak ) ⊂ P̄mk .

This is for the case where  is completely above or below the x axis (left and center) and the split on
the right, which can occur at most for one k. The situation for the edges fN (bk ) ⊂ ∂ is the same.
The segments in the first two cases can also be semi-infinite lines, as in Fig. 8.22.
 8.3 ♣ FT-2 Proximity of Hyperplanes 269

Figure 8.18. ♣ FT-4 Construction of .

The kth region k contributes the vertices fN (Ak ), fN (Bk ) to ∂, and together with the region k+1 ,
the vertices fN (Ak+1 ), fN (Bk+1 ) and edges fN (ak ), fN (bk ), as highlighted.

direction for ak is ck+ → ck− and ck− → ck+ , for bk , shown in Fig. 8.19 for 3-D and
Fig. 8.20 for N -D:
, A2 −→ A3 → · · · → Ak → · · · → AN −1 -
P : A1 = BN+1 B1 = AN +1
. BN−1 ← BN−2 ← · · · ← Br ← · · · ← B2 /
(8.46)
The construction of  is akin to a central problem in computational geometry:
the construction of the convex hull of a planar set of N points. This problem is
equivalent to sorting and can be accomplished with O(N log N) worst-case com-
putational complexity [148]. In our case, the sweep shown in Fig. 8.18 provides the
2N vertices on ∂ already sorted according to their connectivity (order in which
270 8. Proximity of Lines, Planes, and Flats

Figure 8.19. 3D-box B in the space of coefficients.

The vertices and edges, three each Ak , Bk , ak , bk along the path P are shown. The notation ± indicates
the sign of ck± at the kth component.

Figure 8.20. Image B̄ of domain B.

This is an N-dimensional box in the coefficient space c1 × c2 × · · · × cN . The dotted lines are the
polygonal lines representing the box’s vertices. The solid line shows the vertex Āk , and the dashed
portion (together with the remaining solid line) shows one of the points on the edge āk the arrow
on the C̄k axis is the direction of traversal from ck+ → ck− . Each full traversal of ck ∈ [ck− , ck+ ]
corresponds to an edge ak , one of the N edges on the vertex Ak . The full path P can be traced in this
manner.
 8.3 ♣ FT-2 Proximity of Hyperplanes 271

Figure 8.21. The box B in the 3-D space of coefficients c1 , c2 , c3 , (its positions with respect to
the plane πc : c1 + c2 + c3 = 0.) Namely, πc ∩ B = ∅, or if πc ∩ B = ∅, πc is a supporting plane at
either vertex A1 or B1 , or πc intersects two edges of the path P.

they are joined), that is, the path in (8.46), thus lowering the required complexity
to O(N). 
We have been carefully skirting the prospect of the sum S(c) = N j =1 cj = 0

Nsome c ∈ B, and this is dealt with now. In c ∈ R , the coefficient space πc :

for N

j =1 cj = 0 is a hyperplane, and its relative position to the box B has important

ramifications with respect to the corresponding . So for the determination of ,
the procedure is to first check

+ ⇒ πc ∩ B = ∅ or
sign(S(A1 )S(B1 )) =
− ⇒ πc ∩ B = ∅ with S(A1 ) < 0, S(B1 ) > 0.
(8.47)

Figure 8.21 illustrates these possibilities in R3 with πc away from B, intersecting

some edges of B or containing only one of its vertices. With n = (1, 1, . . . , 1) ⊥ πc ,
clearly the only vertices for which πc can be a supporting plane of B are A1 or B1 .
For if πc intersects the diagonal joining A1 or B1 , then it must also intersect some
edges and hence cannot be a supporting plane. Drawing these three situations in RN
with πc and the N-dimensional box B in Fig. 8.20 is left as an exercise (Exercise 4).
To treat the case B ∩ πc = ∅, it is necessary to consider the function fN of (8.49)
in homogoneous coordinates:
272 8. Proximity of Lines, Planes, and Flats

⎛ ⎞

N
fN (c) = (x(c), y(c)) = ⎝ (j − 1)cj , 1, S(c)⎠ . (8.48)
j =1

Consider the case that πc  ∩ B is either A1 or B1 . The image of that vertex is an

ideal point with slope 1/ N j =1 (j − 1)cj , and  is a uc, an unbounded convex
set in the nomenclature of Chapter 7. The construction of ∂ proceeds exactly as
outlined above, the exception being that  opens in the direction of its one ideal
point, and the edges on fN (B2 ) and fN (AN −1 ) are half-lines. Such an example
with  opening downward and fN (B1 ) the ideal point is shown in Fig. 8.22. Thus
far, we have seen that the image of ak is a finite segment fN (ak ) ⊂ P̄mk when
πc ∩ B = ∅ or a semi-infinite line when πc ∩ B = A1 orB1 , and the same holds
for the images of bk . The remaining possibility is that πc intersects several edges of
∂. There are N edges at each vertex, but inspection of the path P in (8.46) as well

Figure 8.22.  can be a uc. Here fN (B1 ) is an ideal point.

 8.3 ♣ FT-2 Proximity of Hyperplanes 273

Figure 8.23.  can be a gh.

as the monotonicities in (8.39), (8.40) shows that at most two edges of that path are
intersected by πc ; this is also seen in Fig. 8.21. The image of such an edge, as shown
on the right of Fig. 8.17, is the projective complement of the finite segment with
fN (Ak+1 ) now higher than fN (Ak ). Then  is a gh, a generalized hyperbola in
the nomenclature of Chapter 7 on curves as shown in Fig. 8.23, whose construction
proceeds according to the order given by (8.39), (8.40). A variation occurs when
πc intersects ∂B at one of the vertices V not A1 or B1 . Then  is a gh with one
asymptote corresponding to the ideal point V̄ being vertical, as shown in Fig. 7.23
and Fig. 7.30 (see Exercise 5).
For easy reference, the construction by which the region  is obtained is called
the “scissors algorithm,” that is, starting with the “fan” regions as shown in Fig. 8.10,
their intersection described in Lemma 8.3.9, and the construction as described in
this section, in particular (8.41), (8.42), (8.43). Clearly this construction algorithm
outputs specific types of regions, and not any 2N-gon region can be its output.
The conditions resulting in the three types of regions are brought “under one roof”
visually in Fig. 8.24 and formally below.
Theorem 8.3.10 (T. Matskewich) The region  is 2N-gon above or below (but
not intersecting) the x-axis, which is
1. bc if πc ∩ B = ∅,
2. uc if πc ∩ B = {A1 } or πc ∩ B = {B1 } ,
274 8. Proximity of Lines, Planes, and Flats

Figure 8.24. The intersection of the hyperplane πc with the path P ⊂ B determines the type of
the region . When πc ∩ B = ∅ (top),  is a bc (bounded convex) 2N -gon as in Fig. 8.18; otherwise,
a uc (unbounded convex) when πc is a supporting hyperplane to B at either vertex A1 or B1 (middle)
(Fig. 8.22), or as in the bottom part a gh (generalized hyperbola) (Fig. 8.23) when πc intersects more
than one segment B. In addition, if πc cuts through one of the vertices V , then V̄ is an ideal point
and is a vertical asymptote of  (Exercise 5).

3. gh if πc intersects B at more than one edge, and has a vertical asymptote if in

addition, πc contains a vertex of B.
The construction of the remaining regions, ’s companions, forming the image
of the neighborhood NH given by (8.17) is basically the same and is the next
section’s topic. Incidentally, NH is a complex hypersurface, the analogue of an
N -Dimensional gh, and as we see, it is represented by N − 1 planar regions,
which in turn are bc, uc, or gh. This lays the groundwork for the representa-
tion of hypersurfaces in the next chapter and shows the generality that has been
achieved.
 8.3 ♣ FT-2 Proximity of Hyperplanes 275

Figure 8.25. Pathway for the computation of ∂ when N = 3.

The label ci indicates the only coefficient varying along the edge. Starting from A1 = (−, −, −),
the next vertex A2 = (+, −, −) is found by the variation of c1 between its extreme values. All the
remaining vertices are found in the same way.

After all this preparation, let us show an example of the theorem’s Case 1 for
N = 3. What this amounts to is knowing the vertices A1 , A2 , A3 , B1 , B2 , B3 for
the computation of the boundary of  = NH0
. These are among the vertices of
the 3-D box in Fig. 8.19. On the path shown in Fig. 8.25, the label ck indicates the
only coefficient varying along that edge. The resulting hexagon for c1 ∈ [1, 2],
c2 ∈ [2, 3], c3 ∈ [3, 4] is drawn on the left side of Fig. 8.26 using the formula
for π̄0
. ♣ FT-4e

8.3.6 ** The Full Image NH

To generate the regions NHi
of NH for i = 0, . . . , N it suffices to find the vertices
of ∂B, which in turn provide the vertices of ∂NHi
, analogous to the traversal (8.46)
for i = 0. The 2N -gon NH1
contains the points π̄1
, which are the function values
of
fN1
: RN &→ R2
in x, y coordinates, where now,
 N 
Nc1 + j =1 (j − 1)cj 1
fN1
(c) = (x1
(c), y(c)) = , . (8.49)
S S
The analysis as in Section 8.3.3 involves finding the coordinate x in terms of y = 1/S
via the decomposition below, for k = 1, 2, . . . , N:
 
Nc1 + N j =1 (j − 1)cj Nc1 + N j =1 (j − k − 1)cj
x= =k+ .
S S
276 8. Proximity of Lines, Planes, and Flats

Figure 8.26. The hexagonal regions  = NH0

(left) and NH1
(right).
Here the family of planes is specified by c1 ∈ [1/3, 1.5], c2 ∈ [1/3, 2.5], c3 ∈ [1/3, 1]. Compare
this diagram with Fig. 10.43 at the beginning of the chapter.

And for (8.57), we obtain the linear relation

P̄1k
: x = C1k
y + k. (8.50)
The subscript “1
” distinguishes the above from P0k
, which until now was denoted
by P k . The coefficient here is
C1k
= C k (c1 , c2 , . . . , ck , ck+2 , . . . , cN )


k 
N
= Nc1 + (j − k − 1) c + (j − k − 1) c , (8.51)
$ %" # j $ %" # j
j =1 <0 j =k+2 >0
with
⎧ 
⎨ C 1
= C 1 (c1 , c3 , . . . , . . . , cN ) = (N − 1)c1 + Nj =3 (j − 2)cj ,
1
⎩ C N = C N (c , . . . , c  (8.52)
1
2 N−1 , c ) = N (j − N − 1)c .
N j =2 j
 8.3 ♣ FT-2 Proximity of Hyperplanes 277

Figure 8.27. Pathway for the construction of ∂NH1

for N = 3 (left).
These are the π̄1
points based on the axis triple X̄2 , X̄3 , X̄1
The labels ci indicate the only coefficient
varying along the edge. This traversal is found by applying the shift ck → ck+1 mod 3 to the previous
diagram in Fig. 8.25 starting with A1 = (−, −, −) and c2 varrying between its extreme values along
the edge connecting it to the next vertex A1
,2 = (−, +, −) (while c1 , c3 are constant). All the
remaining vertices are determined in the same way. The axis shift by 1 for the construction of NH1

is shown on the right. The top and bottom vertices are indicated by the vertices A1 , B1 of B, rather
than the cumbersome notation fN1
(A1 ), fN1
(B1 ) used to compute them, while the lines are labeled
by the coefficients varying along them.

The key observation is that C1k

, is independent of ck+1 , whereas C k is independent
of ck . So the shift ck → ck+1 mod N applied to all the previous formulas found
from C k (henceforth denoted by C0k
) yields the correct relations for C1k
, namely,
⎧
⎪
⎪ C k = C1k
(c1+ , . . . , ck+ , ck+2− −
, . . . , cN ),
⎪ 1
,m
⎪
⎪
⎪
⎨ C k
= C k (c− , . . . , ck− , c+ , . . . , cN +
),
1 ,M 1 k+2
(8.53)
⎪
⎪ C11
,m = C11
(c1− , . . . , cN
−
), C1N
,m = C1N
(c2+ , . . . , cN
+
),
⎪
⎪
⎪
⎪
⎩ C1 1 + + N − −
1
,M = C1
(c1 , . . . , cN ), C1
,M = C1
(c2 , . . . , cN ).
N

As before, the subscripts m and M indicate that C1k

,m and C1k
,M are the minimum
and maximum values of C1k
.
By applying the shift ck → ck+1 mod N to the previous results, we can avoid
the laborious calculations to obtain the needed vertices of B. The y-range of all the
NHi
is the same, so the vertices A1 and B1 , which yield top and bottom vertices,
are included for all i. It is helpful to return to the 3D example and find the vertices
traversed in ∂B for the construction of NH1
, as shown on the left part of Fig. 8.27.
From these vertices the second hexagon shown in Fig. 8.26 is constructed using
278 8. Proximity of Lines, Planes, and Flats

Figure 8.28. Generalizing the shift by i, k → k + i mod N , units in RN .

Figure 8.29. The shift by 2 units, k → k + 2 mod 3 and the resulting path on ∂B.

the formula for π̄1

. The shift and its results are shown pictorially on the right part
of Fig. 8.27. The matter easily generalizes to the shift k → k + i mod N units
in RN to successively construct the system of N axes X̄i
, . . . , X̄(N +i−1)
for the
construction of the NHi
as shown in Fig. 8.28. Note that the distance between the

X̄i and X̄i−1 axes is N − 1. The specific case for N = 3 and i = 2 is shown next in
Fig. 8.31. The region NH2
is computed using the vertices shown with the formula
 8.3 ♣ FT-2 Proximity of Hyperplanes 279

for π̄2
. For N = 3, i = 3, the region NH3
is identical to NH0
but translated three
units to the right. A more interesting example is in Fig. 8.30 for N = 4, i = 3; it is
helpful for understanding general situation.
For small N, the vertices can be conveniently found by the traversal diagram.
For the sake of completeness, it is useful to know for large k the sequence of vertices
Ai
,k , Bi
,k ∈ ∂B, and the cr varying along some of the edges resulting from the ith
shift are listed below.
⎧ ci
⎪
⎪
⎪ A1 = (c1− , . . . , cN −
) −→,
⎪
⎪
⎪
⎪ − − + − −
⎪ Ai
,2 = (c1 , . . . , ci−1 , ci , ci+1 , . . . , cN ),
⎪
⎪
⎪
⎪
⎪ A
= (c− , . . . , c− , c+ , c+ , c− , . . . , c− ),
⎪
⎪
⎪
⎪
i ,3 1 i−1 i i+1 i+2 N
⎪
⎪ ...
⎪
⎪
⎪
⎪
⎪
⎪ Ai
,k = (c1− , . . . , ci−1
−
, ci+ , . . . , ci+k−1
+ −
), k < N − i + 1,
⎪
⎪ , . . . , cN
⎪
⎪
⎪
⎪ ...
⎪
⎪
⎪
⎪
⎪ cN + cN +1
⎪
⎪
⎪ −→ Ai
,N −i+1 = (c1− , . . . , ci−1 −
, ci+ , . . . , cN +
−1 , cN ) −→,
⎪
⎪
⎪
⎪
⎪
⎨ Ai
,N −i = (c+ , c− , . . . , c− , c+ , . . . , c+ , c+ ) −→,
cN +2
1 2 i−1 i N −1 N
(8.54)
⎪
⎪ ...
⎪
⎪
⎪
⎪ ci−2 + + + − + + ci−1
⎪
⎪ −→ Ai
,N = (c1 , c2 , . . . , ci−2 , ci−1 , ci , . . . , cN ) −→
⎪
⎪
⎪
⎪
⎪
⎪ B1 = (c1+ , . . . , cN +
) −→
ci
⎪
⎪
⎪
⎪
⎪
⎪ + ci+1
⎪
⎪
⎪ Bi
,2 = (c1+ , . . . , ci−1
+
, ci− , ci+1
+
, . . . , cN ) −→
⎪
⎪
⎪
⎪
⎪
⎪ ...
⎪
⎪
⎪
⎪ cN + + + − − cN +1
⎪
⎪ −→ B
,N −i+1 = (c , c , . . . , c , c , . . . , c ) −→
⎪
⎪
i 1 2 i−1 i N
⎪
⎪
⎪
⎪ ...
⎪
⎪
⎪
⎩ ci−2 − ci−1 ci
−→ Bi
,N = (c1− , c2− , . . . , ci−2 − +
, ci−1 , ci− , . . . , cN ) −→ A1 −→ .

Corollary 8.3.11 (T. Matskewich) Out of a total of 2N vertices of ∂B 2N(N −1),

those given in (8.54) are needed for the construction of the complete neighborhood
NH consisting of N − 1 polygonal regions (2N-gons).
280 8. Proximity of Lines, Planes, and Flats

Figure 8.30. For N = 4 the shift is i = 3, i.e., k → k + 3 mod 4.

This is the resulting path and vertices on ∂B.

Figure 8.31. The shift by 2 units, k → k + 2 mod 3 and the resulting path on ∂B.

Adding c0
As was done for lines, there remains the task of enlarging the family of “close”
hyperplanes by allowing c0 to vary in the equation of π. The neighborhoods are
now defined by
NH = {π|π : c1 x1 + c2 x2 + · · · + cN xN = c0 ,

ci ∈ [ci− , ci+ ], i = 0, 1, 2, . . . , N}. (8.55)

When necessary, the notation NH0 for a general c0 is used to distinguish it from the
previous neighborhoods with c0 = 1. As for line neighborhoods, the y coordinates
 8.3 ♣ FT-2 Proximity of Hyperplanes 281

Figure 8.32. Construction of the vertices when c0 = 0.

Instead of the original vertices shown in black circles, there are now two for c0+ , higher, and for c0− ,
both with the same x coordinate. The corresponding construction lines contributing to the new edges
are shown.

of the vertices change according to c0 , and two vertical edges are added, so that the
polygons have 2(N + 1) vertices and edges. The polygon still consists of the points
π̄0
, which are now the function values of
 N 
N+1 2 j =1 (j− 1)cj c0
fN : R &→ R , fN (c) = (x(c), y(c)) = , , (8.56)
S S

c ∈ B0 = [c0− , c0+ ] × [c1− , c1+ ] × · · · × [cN

− +
, cN ] ⊂ RN +1 ,

an (N +1)-dimensional box. The insertion of vertical edges poses a bit of a challenge

for uc and gh regions, and the interesting details are left as exercises. Other than
that, the addition of c0 poses no difficulties. Repeating the previous decomposition
N N
j =1 (j − 1)cj j =1 (j − k)cj
x= = (k − 1) + ,
S S
282 8. Proximity of Lines, Planes, and Flats

Figure 8.33. ♣ FT-5 The four hexagonal regions.

Here the family of planes has c1 ∈ [1/3, 1.5], c2 ∈ [1/3, 2.5], c3 ∈ [1/3, 1]. Note that the last
region NH3
is identical to the first NH0
translated three units to the right. Note the overlap in the
last two regions, which suggests that there may be planes π in this family with c3 = 0 (when
π̄2
= π̄3
), which is not possible from the definition of the ck intervals. However, there cannot be
points π̄2
= π̄3
∈ NH2
∩ NH3
with π̄0
∈ NH0
and π̄1
∈ NH1
. For example, the two 0 with
vertical arrows are the locations of π̄0
, π̄3
in NH0
and NH3
for (−, +, −), while the  with a
horizontal arrow is the location of (+, −, −) in NH2
.

the x coordinate of fN , for k = 1, 2, . . . , N, is found in terms of the second

coordinate, which is now y = c0 /S, just adding c0 to the previous linear relation
with the C k defined as before:
Ck
P̄ k : x = y + (k − 1). (8.57)
c0
The net result in the construction of the polygons is that now for each k there are
two lines
Cmk Cmk
k k
P̄m,− :x= − y + (k − 1), P̄m,+ :x= + y + (k − 1) (8.58)
c0 c0
k
CM k
CM
k k
P̄M,− :x= y + (k − 1), P̄M,+ :x= y + (k − 1). (8.59)
c0− c0+
Figure 8.4 shows the situation for line neighborhoods, which basically remains the
same here. As shown in Fig. 8.32, for each k there are now two points correspond-
ing to c0+ and c0− with the same x coordinate obtained by the intersection of the
 8.3 ♣ FT-2 Proximity of Hyperplanes 283

Figure 8.34. ♣ FT-6 The octagons  = NH0

(left) and NH1
(right) for the family of planes
with c0 ∈ [0.85, 1.15], c1 ∈ [1/3, 1.5], c2 ∈ [1/3, 2.5], c3 ∈ [1/3, 1]. The superscipt of c0 is the
first entry of the quadruples (±, ., ., ., ) designating the vertices. Vertices with c0+ are marked by 0
and those for c0− with ellipses.

corresponding P̄ ’s with ± subscripts respectively. Though c0 does not affect the

type of NH, which is determined by S(c) = 0 if 0 ∈ [c0− , c0+ ], the x axis intersects
the region if c0− < 0, requiring an amendment of Theorem 8.3.10. All this becomes
clearer as we consider some examples.
Theorem 8.3.12. For c0 ∈ [c0− , c0+ ], the region  is a 2(N + 1)-gon, which may
have its leftmost or rightmost (or both) edges vertical and which is

1. bc if πc ∩ B = ∅,
2. uc if πc ∩ B = {A1 } or πc ∩ B = {B1 } ,
3. gh if πc intersects B at more than one edge, and has a vertical asymptote if in
addition, πc contains a vertex of B.

8.3.7 Some Examples

A few concrete examples illustrate some special cases in the process, showing that
the constructions are basically easy. The relations

x(π̄1
2
3 ) = 6 − x(π̄123 ) − x(π̄1
23 ), x(π̄1
2
3
) = 3 + x(π̄123 ), (8.60)
284 8. Proximity of Lines, Planes, and Flats

Figure 8.35. The octagon NH0

obtained from the hexagon in Fig. 8.26 with c0 ∈ [−0.5, 1].
The x-axis is crossed, since c0 = 0 is included in the range.

and their higher-dimensional generalizations given in Chapter 5 turn out to be very

useful in the computations. We complete the previous 3D example in Fig. 8.26,
constructing the hexagon NH2
using the vertices for the shift k → k + 2 mod 3
shown in Fig. 8.31. The fourth region NH3
, see (8.60), is simply a translation of
NH0
three units to the right. Here we encounter an unexpected overlap between
these two hexagons, suggesting that there may be planes π in this family with
π̄2
= π̄3
⇒ c3 = 0, which is not possible, since here c3 ∈ [1/3, 1]. For the fun of
it, for a couple of coincident points on this overlap, their corresponding π̄0
, π̄1
are
 8.3 ♣ FT-2 Proximity of Hyperplanes 285

Figure 8.36. A uc in 3D for c1 ∈ [−2/3, 1.50], c2 ∈ [1/3, 2.50], c2 ∈ [1/3, 1.0].

Compare with the coefficient ranges of the hexagon(s) in Fig. 8.26 and note that they differ only for
c1− . Here the sum c1− + c2− + c3− is equal to 0.

found, and they differ, showing that these belong to two different planes. All this is
detailed in the caption of Fig. 8.33.
Still working with Fig. 8.26 but now with variable c0 ∈ [0.85, 1.15], two
octagons are produced with two vertical edges as shown in Fig. 8.34. For c0 < 0
the resulting region crosses the x axis; Fig. 8.35. By the way, the appearance of
vertical edges does not necessarily imply that c0 is varying. For instance, in the
family of planes in Fig. 8.26, just changing c1− = 1, which is also the value of
c3+ , produces the hexagon NH0 , whose leftmost edge of NH0 , on (−, −, +) and
(−, +, +), is vertical (see Exercise 9). Minimally varying this family of planes
286 8. Proximity of Lines, Planes, and Flats

Figure 8.37. A uc in 3D for c1 ∈ [−1, 1.50], c2 ∈ [−1, 2.50], c2 ∈ [1/3, 1.0].

Compare with the coefficient ranges of the hexagon(s) in Fig. 8.26 and note that they differ only for
c1− , c2− . Here the plane πc : c1 + c2 + c3 = 0 intersects two edges of the box B with the vertices
(−, −, −), (−, −, +) below πc .
 8.3 ♣ FT-2 Proximity of Hyperplanes 287

Figure 8.38. A gh in 3D for c1 ∈ [−3.5, 1.50], c2 ∈ [1/3, 2.50], c2 ∈ [1/3, 1.0].

Compare with the coefficient ranges of the hexagon(s) in Fig. 8.26 to note that they differ only in c1− .
Here the plane πc : c1 + c2 + c3 = 0 contains the vertex (−, +, +) of ∂B, which becomes an ideal
point of the gh.

shows how the regions evolve from bc to uc (Fig. 8.36), then gh (Fig. 8.37) and
gh with an ideal point (Fig. 8.38). The ideal point of the uc is computed directly
from the function (8.49). We then move on to 4D with the sequence of Figures 8.40,
8.41, 8.42, 8.43 having already pointed out that the ideal point of a uc is computed
directly from the function fN .

8.3.8 Matching Points and Navigation within NH

The NHi contain all the points representing the  family of planes given by (8.17).
Though not all the vertices of ∂B are used in the construction of the 2N-vertex
regions NHi , their images play an important role in delimiting which N − 1 points
with the same y, one in each NHi
, i = 0, 2, . . . , N − 2, represent a plane of
the family . After all, there are linear relations, those defining the family ,
which introduce additional constraints connecting these points. The problem and its
288 8. Proximity of Lines, Planes, and Flats

Figure 8.39. The gh in Fig. 8.37 with the range c0 ∈ [0.5, 1] added.
In the quintuplet (., ., ., ., .), the first position is for the superscript of c0± . The previous vertices shown
in Fig. 8.37, marked with •, belong to the gh with c0+ = 1.

solution are illustrated using the previous family of 3D planes whose four hexagons
are shown in Fig. 8.44. The images of all vertices of ∂B are shown, with each
hexagon containing two additional interior points.
Two points with the same y, shown in Fig. 8.44, are in NH0 and NH1 , but
they cannot represent a plane π in , since π̄2
∈ NH3 . From the position
of the points the represented plane π is found having coefficients c1 = 0.5,
c2 = 0.65, c3 = 0.18; hence π ∈ , since π̄2
∈ [1/3, 1]. With this example and
upon reflection, it is clear that any N − 1 points with the same y, one in each NHi
,
 8.3 ♣ FT-2 Proximity of Hyperplanes 289

Figure 8.40. A NH3

uc in 4D constructed according to the traversal diagram in Fig. 8.30.

Figure 8.41. The previous uc NH3

region in 4D with c0 ∈ [0.5, 1].
290 8. Proximity of Lines, Planes, and Flats

Figure 8.42. The two extreme regions for the previous uc NH3
in 4D with c0 ∈ [−0.5, 1].
Above is the NH3
for c0 = −0.5 and below for c0 = 1.

i = 0, 1, . . . , N −2, do not necessarily represent a plane in the family . Exploring

further, the same point π0
marked with • on the edge (−, −, −), (+, −, −) of NH0

together with a companion point (π) ¯ 1

with the same y coordinate on the edge
(−, −, −, ), (+, −, −) of NH1
(not shown) represent the plane with coefficients
c1 = 0.44, c2 = 0.76 and c3 = 0.13, which is also not a plane in . Next, consider
the pair π̄0
marked with 2 and the previous π̄1
∈ NH1
, the point marked with
both 2 and •, and π̄2
∈ NH2
, marked with a 2. These represent the plane π in 
having coefficients c1 = 0.33, c2 = 0.55, c3 = 0.45.
So what is going on? The dotted line joining the vertices (−, −, −), (−, +, −)
together with the edges on (−, −, −), (−, −, +) and on (−, −, +), (−, +, −)
 8.3 ♣ FT-2 Proximity of Hyperplanes 291

Figure 8.43. The full uc NH3

in 4D with c0 ∈ [−0.5, 1]. Note the vertical edge on the left.
292 8. Proximity of Lines, Planes, and Flats

Figure 8.44. The four hexagons previously seen in Fig. 8.33.

For a family of planes  with c1 ∈ [1/3, 1.5], c2 ∈ [1/3, 2.5], c3 ∈ [1/3, 1] now showing the images
of all vertices of ∂B, two being interior points. Though the first two points with the same y, marked
by •, are in NH1 and NH2 respectively, they do not represent a plane in , since π̄2
∈ NH3 .

Figure 8.45. Finding interior points.

 8.3 ♣ FT-2 Proximity of Hyperplanes 293

Figure 8.46. Triangulation in orthogonal coordinates.

Two sections of the family of planes  perpendicular to the x3 and x1 axes. Focus on the intersection
point P of the lines (−, −, −), (−, −, +) shown in the x1 x2 and x2 x3 planes. Choose the x coordinates
of π̄0
, π̄1 to fix a π̄ ∈  and on P . The slopes in the x1 x2 and x2 x3 planes are determined from
the x-coordinates of π̄0
, π̄1 respectively. Generate this family of planes parallel to π and in  by
allowing x2 to vary until it hits the constraint; here the lowest point. The range of x2 differs in the
two sections, and the effective range is the minimum. This corresponds to the triangulation procedure
shown above in -coords.

form a triangle in NH0

, and also in other NHi
, as shown in Fig. 8.45. Choose
a point π̄0
on the edge (−, −, −), (−, −, +) on NH0
, marked with 2, and on
the same edge NH1
its companion point π̄1
with the same y. Clearly, this π
is in  with coefficients c2 = c2− = c3 = c3− = 1/3. Any pair of points,
one on each of the vertical lines, through these two points represents a plane π

parallel to π and hence with coefficients c2

= c3
by the previous construction for
parallel planes. The range of y for which π
∈  is determined by the highest
intersection of the verticals with an edge of the triangle, in this case with the
triangle NH1
, as shown. The corresponding situation in Cartesian coordinates is
illustrated in Fig. 8.46. To find further interior points with the same x coordinates,
variation of the y range may be possible in some cases (see Fig. 8.47) but not
others (Fig. 8.48) by further triangulation. It is best to study these pictures to
understand what is happening. There is an important caveat to avoid vertical edges
in the triangulation, and when one is encountered, then use its image in another
NHi
, as was done in Fig. 8.47. The remaining details are left for the exercises
and future generations. Suffice it to say that this procedure generalizes directly
to RN .
294 8. Proximity of Lines, Planes, and Flats

Figure 8.47. Continuing the interior point construction from triangle 1 to triangle 2.

Figure 8.48. Continuation along the vertical by extending the triangulation.

Here it is not possible due to gap between the two triangles in NH2
.
 8.3 ♣ FT-2 Proximity of Hyperplanes 295

Exercises
1. Investigate the neighborhood family Nπ of hyperplanes given by (8.17).
Do they form a topology with respect to the criteria set in Section 8.1? This is
a good topic for a course project.
2. Find the relations corresponding to (8.10), (8.11), (8.12) but for the indexed
points π̄1234 , π̄1
234 , π̄1
2
34 , π̄1
2
3
4 , π̄1
2
3
4
of a hyperplane π ⊂ R4 .
3. Prove analytically the equivalences translation ↔ rotation, rotation ↔ transla-
tion indicated in Fig. 8.9.
4. Draw the equivalent of Fig. 8.21 in the space of coefficients RN using the box
B in Fig. 8.20.
5. Draw the construction of  when πc intersects B at a vertex As and an edge
br . Hint: A¯k is a vertical asymptote of .
6. Construct and draw the uc, gh cases for the line neighborhoods  in RN . Give
specific examples for N = 3, 4.
7. For the neighborhoods given by (8.55) with c0 = 1, carry out the analysis of
the resulting shape of the regions , and construct and draw examples of the
possible options. State the result formally as a corollary to Theorem 8.3.10. This
together with a software implementation where the user specifies the intervals
[ck− , ck+ ], and the region  is drawn, is an excellent topic for a course project.
8. Give examples of polygonal regions with 2N vertices/edges and prove that they
can not be the output of the “scissors algorithm.”
9. Draw some NH ¯ 0 for 2N-gons uc and gh with vertical edges and c0 .
10. Write an algorithm for finding interior points in RN for N = 3, 4. This together
with a software implementation is an excellent topic for a course project.
This page intentionally left blank
9

Hypersurfaces in RN

9.1 ♣ FT-1 Preliminaries

Early in the development of -coords, the representation of surfaces was attempted

by simply plotting the surface’s points as polygonal lines. Except for very special
surfaces (i.e., spheres, ellipsoids), this turned out to be even more hopeless than what
happened with line-curves. The next improvement was the surface representation
by the envelope of these polygonal lines. This is akin to using the projections in
-coords of some surface sections. Since projections do not preserve information,
this approach also does not work, though it is useful (but not accurate) at least for
ellipsoids ([98], [81]), and for this reason it is discussed later on. However, it is
a mess (worse than cusps and other special points in line-curves) for nonconvex
regions. The way out turns out to be the natural continuation of curve images as
treated in Chapter 7.
We start by revisiting Fig. 9.1, which was shown earlier, in Chapter 3. The
top shows the hypercube representation and serves as an excellent paradigm for
what is to follow. In (a) is a square in Cartesian and -coords. Note the role of
duality where a vertex, say A, is mapped into a line Ā, and conversely, an edge,
say AB, is mapped into a point AB. Proceeding, in (b) the image of the 3-D cube is
twice the pattern of the square, where the edges can be identified by two points and
faces by two vertical lines and a polygonal line. This is a good example, showing
where the general representation of lines in Section 4.1.4 [106] is needed for the
edges with two of the three variables constant (in effect, this simplifies matters).
The representation of the cube’s faces stems from the vertical-line representation
of hyperplanes discussed in Chapter 5. The pattern’s repetition pleasingly reveals
the cube’s symmetry and its relation to the square. Whereas we do not know
how to show a hypercube in Cartesian coordinates, it is no problem doing so in

A. Inselberg, Parallel Coordinates, DOI 10.1007/978-0-387-68628-8_9,

© Springer Science+Business Media, LLC 2009
298 9. Hypersurfaces in RN

Figure 9.1. (a) Square (b) Cube in R3 (c) Hypercube in R5 – all edges have unit length. (top)
Showing the vertices and (bottom) representation based on the tangent planes as developed in this
chapter.
 9.1 ♣ FT-1 Preliminaries 299

-coords. In (c) is the representation of a hypercube in R5 consisting of four times

the square’s pattern. Again all vertices, edges, and faces can be determined from
the picture. Interior points are represented by polygonal lines within the pattern.
Polygonal lines that include a boundary segment represent boundary points. The
bottom part of Fig. 9.1 depicts the new representation developed in this chapter.
This suggests how the representation of higher-dimensional versions of a surface
may be obtained from its 3-D representation. All this generalizes nicely to the
representation of polytopes in RN [23]. A rotated or translated hypercube can
still be easily represented and recognized (exercises). The complete description of
polyhedral objects is conveniently given in terms of incidence matrices, which are
tables listing the vertices, the edges, and faces of all orders that they are on (i.e.,
their incidence) (see [146, pp. 139–140]).
To lay the groundwork, recall the recursive notation for the axis spacing (con-
sidered as an N -tuple) developed in Chapter 5 on planes. For the initial N-axis
system X̄1 , X̄2 , . . . , X̄N , we write
i
" #$ %
0
dN = (0, 1, 2, . . . , i − 1, . . . , N − 1).

This states that the first axis X̄1 is placed at x = 0, X̄2 is placed at x = 1, . . . until
X̄N , which is at x = N − 1. After translating the X̄1 axis one unit to the right of
X̄N , renaming it X̄1
, the axis spacing N-tuple is
i
" #$ %
1 0
dN = dN + (N, 0, . . . , 0) = (N, 1, 2, . . . , i − 1, . . . , N − 1).

And after i such successive unit translations with the X̄i axis in position X̄i
, the
axis spacing for i = 0, . . . , N, k = 1, . . . , N is given by
i i
" #$ % " #$ %
i 0
dN = dN + (N, . . . , N, 0, . . . , 0) = (N, N + 1, . . . , N + i − 1, i, . . . , N − 1)

= (dik ). (9.1)
To clarify the indices, i is the number of axis translations and k is the position
(component) within the vector dN i . Using the step function

1 i ≥ k,
Si (k) = (9.2)
0 i < k,

the axis spacing after the ith translation can be conveniently written as
i
dN = (dik ) = (k − 1 + NSi (k)). (9.3)
300 9. Hypersurfaces in RN

When the dimensionality is clear from the context, the subscript N can be omitted.
i spacing, the points π̄ p
For a flat π p expressed in terms of the dN 1
,... ,i
,i+1,... ,N
p
of its representation are denoted compactly by π̄i
. Consistent with this notation,
p
π p = π0
for π p described in terms of the axis spacing dN 0 . So the common spacings

are d30 = (0, 1, 2), d31 = (3, 1, 2), and so on.

It is useful to recast the representation of planes in vector form. For
π : c1 x1 + c2 x2 + c3 x3 = c0 , (9.4)
the coefficient vector c = (c1 , c2 , c3 ), and u = (1, 1, 1), the representing points of
π given in inner-product (denoted by “·”) form are:
π̄ = (c · d3i , c0 , c · u) = (c · d3i , c0 , c1 + c2 + c3 ). (9.5)
The first coordinates of π̄, with the axis spacing given by (9.3), are
⎧
⎪
⎪ c · d0 = c · (0, 1, 2) = c2 + 2c3 ,
⎪
⎪
⎨ c · d1 = c · (3, 1, 2) = 3c + c + 2c ,
1 2 3
2
(9.6)
⎪
⎪ c · d = c · (3, 4, 2) = 3c1 + 4c2 + 2c3 ,
⎪
⎪
⎩
c · d3 = c · (3, 4, 5) = 3c1 + 4c2 + 5c3 .

For xi = c · di /S with S = N−1 i=0 ci , The relations
x2 = 6 − [x0 + x1 ], x3 = 3 + x0 ,
i /S and S = N c ,
and in general for RN with xi = c · dN i=1 i

N −2
x(N−1) = N(N − 1) − xi , xN = N + x0 ,
i=0
provide the connection between the N − 1 independent representing points and the
remaining two.

Exercises
1. Show the 5-D hypercube in -coords after a
(a) translation,
(b) rotation, and
(c) both translated and rotated.
2. Construct incidence matrices for the 3-D and 5-D hypercube shown in Fig. 9.1.
3. Construct the representation in -coords and incidence matrices for an object
(e.g., a pyramid) of your choice in 3-D and 5-D; see also [23].
 9.2 Formulation 301

Figure 9.2. A surface σ ∈ E is represented by two linked planar regions σ̄123 , σ̄231
.
They consist of the pairs of points representing all its tangent planes.

9.2 Formulation
We now pursue the representation of smooth surfaces in RN and in particular the
family E of surfaces40 that are the envelopes of their tangent planes. The discussion is
confined to R3 when the higher-dimensional generalizations are clear. Occasionally,
as for curves, we use terminology and results that can be found in the references
on differential geometry. Also recommended is the beautiful and comprehensive
website on surfaces [130]. The basic idea for the representation of a surface σ ∈ E
is illustrated in Fig. 9.2. The tangent plane π at each point P ∈ σ is mapped to
the two planar points π̄123 , π̄231
= π̄1
23 . Collecting the points by their indices
for all the tangent planes yields two planar regions σ̄123 (x, y), σ̄1
23 (x, y), one
each for the index triples. The regions must be linked by a matching algorithm
that selects the pairs of points (one from each region) representing valid tangent
planes of σ . The two linked regions form the representation σ̄ of the surface.
An example of the linking was already seen in Chapter 8 between the polygonal
regions representing families of proximate planes. The intent is to reconstruct the
surface from its representation. The manner and extent to which this is possible,
i.e., when the -coords mapping is invertible, are among the issues studied in this
chapter.
40
Not to be confused with the family of superplanes also denoted by E.
302 9. Hypersurfaces in RN

Formally, a surface σ ∈ E given by

σ : F (x) = 0, x = (x1 , x2 , x3 ), (9.7)
is represented in -coords by
σ &→ σ̄ = (σ̄123 (x, y), σ̄1
23 (x, y)) ⊂ P2 × P2 , (9.8)

where use of the projective plane P2 allows for the presence of ideal points. In the
notation, the link between the two regions is indicated by placing them within the
( , ), which are omitted when the discussion pertains to just the region(s). The
functions used are assumed to be continuous in all their variables together with
such of its derivatives as are involved in the discussion.

∂F , ∂F , ∂F  ,
The gradient vector of F , ∇F = ∂x 1 ∂x2 ∂x3 P
at the point P is normal
to the surface σ at P ⇒ the tangent plane π of σ at the point P0 (x0 ), (x0 ) =
(x10 , x20 , x30 ) is given by


3
∂F 0 0 0
π : ∇F · (x − x0 ) = (xi − xi0 ) (x , x , x ) = 0. (9.9)
i=1
∂xi 1 2 3

Therefore the representing points of π are

π̄i
(s, t) = (∇F · di , ∇F · (x0 ), ∇F · u), i = 0, 1. (9.10)
For our purposes it is often preferable to describe a surface in terms of two para-
meters as
σ : F (s, t) = F (x) = 0, x = x(s, t), s ∈ Is , t ∈ It , (9.11)

where Is , It are intervals of R. The equivalent description for hypersurfaces in RN

requires N − 1 parameters. This form is due to the great mathematician Gauss.
The vector xs (s, k2 ) is tangent to the surface curve t = k2 , and xt (k1 , t) is tangent
to the surface curve s = k1 , the ki = constant, and the subscripts denote partial
derivatives with respect to the indicated variables. Therefore, the vector xs × xt is
normal to the point P = F (x0 ) ∈ σ , where x0 = x(k1 , k2 ), and the tangent plane
π at P0 (x0 ) is conveniently given by
 
 x1 − x 0 x2 − x 0 x3 − x 0 
 1 2 3 
π(s, t) : (x − x0 ) · xs × xt =  x1t x2t x3 t  = 0. (9.12)

 x x2s x3 s 
1s

Alternatively, σ ∈ E can be considered as the envelope of the 2-parameter

family π(s, t) of its tangent planes, each contributing a pair of points, one each for
 9.2 Formulation 303

the planar regions σ̄123 , σ̄1

23 , constituting the image σ̄ . The representing points of
the tangent planes are equivalently given by

π̄i
(s, t) = (n̂ · di , n̂ · (x0 ), n̂ · u), i = 0, 1, (9.13)

where
xs × xt
n̂ =
|xs × xt |

is the unit normal vector to σ at the point P0 . Stated explicitly the point &→ pair-of-
points mapping, from either (9.10) or (9.12), is

x ∈ σ &→ π &→ (π̄0

, π̄1
) = (π̄123 , π̄1
23 ) = ((x, y), (x
, y)), (9.14)

with a slight change in notation, x = x0

and x
= x1
for the x-coordinates of
π̄123 = π̄0
, π̄1
23 = π̄1
respectively41 . The y is the same for both points. In terms
of the gradient’s components, Fi = ∂F /∂xi ,
⎧ 2 +2F3
⎪
⎪ x = F1F+F 2 +F3
,
⎪
⎨
y = x1 FF1 +x 2 F2 +x3 F3
1 +F2 +F3
, (9.15)
⎪
⎪
⎪
⎩
1 +F2 +2F3
x = 2 3FF1 +F2 +F3 .

These transformations are the direct extension of the 2-D point ↔ point curve
transformations derived in Chapter 7. A word of caution: when the interaxis distance
d is not equal to 1, the right-hand sides of x and x
above need to be multiplied by
d and 2d respectively (see (7.7)); an example is given further; see (9.22).
The generalization to the hypersurfaces E of RN is direct. The image of the
tangent hyperplane at a point P ∈ σ ∈ E consists of N − 1 points determined from
(9.10) by using the appropriate axis spacing dNi , i = 0, . . . , N −2, the x coordinate
being the xi
as given above. The resulting transformation, the N-D extension of
(9.15) with N terms in the numerator and denominator, determines the point (N −1)-
tuples mapping the surface σ ∈ E into σ̄ consisting of (N − 1) planar regions. Next
the construction of the regions is undertaken and subsequently the link or matching
algorithm between them is found.

41
The short-subscripts 0
or 1
notation is used interchangeably with that showing the full index list,
i.e., 123 or 1
23.
304 9. Hypersurfaces in RN

Figure 9.3. Intersection of a surface σ with the two superplanes π0s

, π1s
.
The points of the boundary ∂ σ̄ are the images of the tangent planes at the points of the two curves
σ ∩ π0s
and σ ∩ π1s
. The surface above is a hyperboloid of one sheet.

9.3 ♣ FT-2 Boundary Contours

The development is presented for R3 , where the generalization to RN , via vector

notation, is straightforward. The construction of the regions σ̄ and in particular
their boundary ∂ σ̄ is greatly facilitated by an observation that is a substantial labor-
saving device compared to the computation of the envelope of the tangent planes.
In the ensuing, the surface σ in question is translated and rotated so that the first
two hyperplanes are not tangent to the surface and their intersections are curves
(i.e., not empty or isolated points), for example in Fig. 9.3. There may be many
such positions and orientations, referred to as a standard position, and for our
purposes any one will do. Once the representation of σ in a standard position is
found, its representation in the original position and orientation may be found via
the translation ↔ rotation duality discussed later on.
The following lemma is not true for all surfaces σ ∈ E , as we will discover
toward the end of the chapter, but it applies to a large subclass (including convex
hypersurfaces), which we call E G. The situation is reminiscent of the conditions
leading to (6.2) for obtaining the envelope of families of curves, which is not
completely general but still very useful. Roughly speaking, if σ ∈ E G for every
tangent P̄ at a point π̄i
of ∂ σ̄i
, there exists a neighborhood of π̄i
where all points
of σ̄i
are on the same side of P̄ . Henceforth, unless otherwise stated, when the
boundaries of the representing regions of a surface σ are found using the lemma, it
is assumed that σ ⊂ E G.
 9.3 ♣ FT-2 Boundary Contours 305

Figure 9.4. Formation of boundary contours.

Lemma 9.3.1 (Boundary of σ̄ ) For σ a surface in R3 , ∂ σ̄123 is the image of

σ ∩ π0s
. That is, ∂ σ̄123 = σ ∩ π0s
and similarly ∂ σ̄1
23 = σ ∩ π1s
.

Proof. With reference to Fig. 9.4, each point ρ̄123 ∈ σ̄123 is a representing point of
a plane ρ tangent to the surface σ at a point R by the definition of the region σ̄123 .
In fact, ρ̄123 = ¯ρ , where ρ ∩ π0s
= ρ . Note that for a plane π tangent at point
P ∈ σ, π̄123 ∈ ∂ σ̄123 , π = π ∩ π0s
and P ∈ π . The image P̄ is a line tangent to
∂ σ̄123 at π̄123 = ¯123 ⇒ P ∈ π0s
: duality point-of-tangency ↔ line-of-tangency.
In short, every point of σ̄123 represents a line of π0s
due to the smoothness of ∂σ and,
therefore, by continuity of ∂ σ̄123 , every tangent to ∂ σ̄123 represents a point of π0s

and ∂σ . Hence ∂ σ̄123 = σ ∩ π0s

and similarly ∂ σ̄1
23 = σ ∩ π1s
. The smoothness
of ∂σ ensures the continuity of ∂ σ̄123 . 
This basic result tremendously simplifies matters. The boundary ∂ σ̄ is simply
the image of two curves: the intersections of σ with the first two superplanes π0s
, π1s
.
Note further that for a plane π tangent at a point P ∈ σ , π̄123 ∈ ∂ σ̄123 and π̄1
23 ∈
∂ σ̄1
23 ⇔ P ∈ σ ∩ π0s
∩ π1s
. Finding the boundary of surface representations
is one of the major reasons for studying the transform of curves to -coords in
Chapter 7.
Lemma 9.3.2 (Boundary of σ̄ in RN ) For a σ ⊂ RN , ∂ σ̄ is composed of N − 1
curves that are the images of the intersections of σ with the fi st N − 1 superplanes.
306 9. Hypersurfaces in RN

Corollary 9.3.3 (Ideal points) An ideal point on the boundary ∂ σ̄ is the image
of a tangent plane at a point of σ , parallel to u.

Proof. It is an immediate consequence of (9.10) that ideal points occur in the image
when the third component is zero. That is, the tangent plane’s normal is normal to
u or the tangent plane is parallel to u. 
♣ FT-2e
An algebraic surface is one described by a polynomial equation, providing an
important special case of Lemma 9.3.2.
Corollary 9.3.4 (Boundary of an algebraic surface) The boundary σ̄ of an
algebraic surface σ of degree n is composed of N − 1 algebraic curves each of
degree n(n − 1) or less.
The corresponding result in algebraic geometry42 is that the dual of a nonsingular
algebraic surface of degree n has degree n(n − 1)(N −1) [52], [186]. Here the
boundary representations can be found with the aid of Plücker’s results, presented
in Section 7.6. For F in (9.15) a quadratic polynomial, the corresponding surface
is called a quadric, and in R3 ,

σ : F (x1 , x2 , x3 ) = (x1 , x2 , x3 , 1)A(x1 , x2 , x3 , 1)T , (9.16)

A being a symmetric 4 × 4 matrix for R3 and (N + 1) × (N + 1) for RN .

Corollary 9.3.5 (Boundary of σ̄ for σ a quadric) A quadric surface σ ⊂ RN
is represented by N − 1 linked planar regions with conic boundaries.
This is the direct extension from conics in 2-D to quadrics in N-D, and as for conics,
their type depends on the type of quadric and its orientation, as pointed out next.
Incidentally, quadrics may be classified according to their matrix as in [126,
pp. 455–7]. A matrix is called positive or negative definit if all its eigenvalues are
positive or negative, respectively, and indefinit if it has both positive and negative
eigenvalues. The corresponding surface has a local minimum/maximum if A is
positive/negative definite and a saddle if A is indefinite. Numerically, a fast test
for the positive definiteness of a symmetric matrix A is to attempt to factor it
as A = R T R, where R is upper triangular with positive diagonal entries. For a
nice treatment of quadrics and their properties see also [183, Chap. VII]. Another
beautiful but nonelementary classification of quadrics is in [13, vol. II, Chap. 15].
The boundary ∂ σ̄ is the image of space curves contained in the superplanes.
A space curve (see also Section 7.4.1) c ⊂ RN can also be represented as the image
of its tangent lines. In turn, the tangent t at a point P ∈ c can be represented by
42
I am indebted to Prof. Evgeni Shustin for these references.
 9.3 ♣ FT-2 Boundary Contours 307

Figure 9.5. The image c̄ of a space curve c contained in π1s is formed by two coincident curves.

N − 1 doubly indexed points, say t¯i−1,i , such that c̄ consists of N − 1 planar curves
c̄i−1,i formed by the t¯i−1,i , two such curves for R3 . Each c̄i−1,i is the image of
the projection of c on the 2-plane xi−1,i . The fun part is that in our case, these
two curves are coincident, since they are the images of a curve in a superplane, say
c ⊂ π0s
, where each tangent t at a point P ∈ c is a line in the sp π0s
, and hence
t¯12 = t¯23 . An example is in order. Consider the cylinder
CY : x22 + x32 = 1, (9.17)
which is centered along the x1 axis. For a moment we anticipate a result from
the next subsection that the regions representing a cylinder are curves and find
the representation curves starting with CY12 . Based on the aforementioned, the
intersection c with the first superplane
π0s
: x1 − 2x2 + x3 = 0 (9.18)
must be found. The intent is to show that the projections of c = π0s
∩ CY on the
principal planes have identical images. Substituting
1
x3 = 2x2 − x1 , x2 = (x1 + x3 ),
2
successively into (9.17), we obtain
c12 = x12 − 4x1 x2 + 5x22 − 1 = 0, c13 = x12 + 2x1 x3 + 5x22 − 4 = 0, (9.19)
308 9. Hypersurfaces in RN

which are the projections of c on the x1 x2 , x1 x3 planes, respectively; c23 is avail-

able directly from (9.17). Of course, all these are planar ellipses, and we refer to
Section 7.5, on the conic transforms, to find their images. For convenience the
appropriate notation and formulae are restated. The image of the quadratic

F (x1 , x2 ) = A1 x12 + 2A4 x1 x2 + A2 x22 + 2A5 x1 + 2A6 x2 + A3 = 0 (9.20)

is the quadratic

f (x, y) = a1 x 2 + 2a4 xy + a2 y 2 + 2a5 x + 2a6 y + a3 = 0 (9.21)

with the coefficients

⎫
a1 = [A3 (A1 + A2 + 2A4 ) − (A5 + A6 )2 ]/d 2 , ⎪
⎪
⎪
⎪
a2 = A1 A2 − A24 , ⎪
⎪
⎪
⎪
⎪
⎪
a3 = A2 A3 − A26 , ⎬
(9.22)
a4 = [A13 − A23 ]/d = [A6 (A1 + A4 ) − A5 (A2 + A4 )]/d, ⎪
⎪
⎪
⎪
⎪
a5 = [A6 + A5 A6 − A3 (A2 + A4 )]/d,
2 ⎪
⎪
⎪
⎪
⎪
⎭
a6 = A2 A5 − A4 A6 ,

with d the interaxis distance and  = (A24 − A1 A2 ) the discriminant; the indicator
I = A1 + A2 + 2A4 plays an important role in the transformation. For c12 , the
interaxis distance is d = 1, and the nonzero coefficients are

A1 = 1, A4 = −2, A2 = 5, A3 = −1.

So from (9.22), the coefficients of its image are

a1 = −2, a2 = 1, a3 = −5, a4 = 0, a5 = 3, a6 = 0

for the image curve

c̄12 : −2x 2 + y 2 + 6x − 5 = 0. (9.23)

Continuing for c13 , the distance between the X̄1 and X̄3 axes is d = 2, the nonzero
coefficients (where now the second variable is x3 ) and its image’s coefficients are

A1 = 1, A4 = 1, A2 = 5, A3 = −4, a1 = −8, a2 = 4,
a3 = −20, a4 = 0, a5 = 24, a6 = 0,

which, upon division by 4, define the same curve as in (9.23). Note that without
the insertion of the interaxis distance d in the formulas, the computation would
 9.3 ♣ FT-2 Boundary Contours 309

not be correct. The image c̄23 is found next to verify the assertion. The nonzero
coefficients of c23 and its image are

A1 = 1, A2 = 1, A3 = −1, a1 = −8, a2 = 1, a3 = −1,

a4 = 0, a5 = 1, a6 = 0,

with the distance between the X̄2 and X̄3 axes d = 1. We write

c̄23 : −2x
2 + y 2 + 2x
− 1 = 0, x
= x − 1, (9.24)

since the X̄2 axis is one unit to the right of the y axis. In the xy coordinate system,

c̄23 : −2x 2 + y 2 + 6x − 5 = 0. (9.25)

Hence images of all the projections of c, which are ellipses, are coincident
hyperbolas c̄12 = c̄13 = c̄23 = c̄ = CY12 .
Proceeding, the image CY23 is the image of the curve c
= CY ∩ π1s
, where

π1s
: x1 + x2 − 2x3 = 0 (9.26)

is the second sp based on the coordinate axis X̄1

, X̄2 , X̄3 . Actually CY23 = c̄23
has already been obtained above, since the X̄2 , X̄3 axes are shared by the two
coordinate systems. This means that the two parts CY23 , CY12 of CY are coincident,
an occurrence that turns out to be characteristic of cylinders whose rulings are
parallel to the x1 axis. It is instructive to verify this in terms of the other two
projections of c
found by the substitutions from (9.26),
1
x2 = 2x3 − x1 , x3 = (x1 + x2 ),
2
into (9.17), which yields

c1

3 : x12 − 4x1 x3 + 5x32 − 1 = 0, c1

2 : x12 + 2x1 x2 + 5x22 − 4 = 0. (9.27)

The nonzero coefficients of c1

3 and the ai are

A1 = 1, A4 = −2, A2 = 5, A3 = −1, a1 = −2, a2 = 1,

a3 = −5, a4 = 0, a5 = −3, a6 = 0,

with d = −1 the (directed) distance between the X̄1

, X̄3 axes, yielding

c̄
1
3 : −2x
2 + y 2 − 6x
− 5 = 0.
310 9. Hypersurfaces in RN

Here x
= x − 3 due to the distance of three units of X̄1
from the y axis, so

c̄
1
3 : −2(x − 3)2 + y 2 − 6(x − 3) − 5 = −2x 2 + y 2 + 6x − 5 = 0, (9.28)

which matches c̄23 . Finally, d = −2 and x

= x − 1 for c1

2 , whose image is also
the above hyperbola.
This extended presentation illustrates the detail involved and emphasizes the
simplification achieved. For each region, the image of only one projection needs to
be computed, so we may as well pick the simplest one, where the two parallel axes
are adjacent with d = 1. The conclusion, a refinement of Lemma 9.3.2, is stated in
general.
Theorem 9.3.6 (Representation of σ̄ in RN ) A smooth hypersurface σ ⊂ RN
can be represented by N − 1 linked (i.e., with a matching algorithm) regions
σ̄i,i+1 ⊂ P2 , i = 1, . . . , N − 1, with ∂ σ̄i,i+1 = c̄(i
−1),i
= (σ ∩ πiN
s )(i
−1),i
.
To clarify the notation, πiNs
is the (i + 1)th sp (recall that the first one is also
denoted by π0Ns
). The image of the space curve πiN
s ∩ σ is the image of only
one projection; we choose (πiNs
∩ σ )(i
−1),i
corresponding to the adjacent axes
X̄i
−1 , X̄i
for which d = 1, with the convention that for i
= 1, X̄(i
−1) = X̄N .
A horizontal translation by N + i − 1 units to the left as shown in Fig. 9.6 is needed
to obtain the correct x coordinate. For an object, a point P , for example, the notation
P̄i
refers to the representation of P with respect to the N coordinate axes after the
ith shift.
For R3 the preferred projections are (π1s ∩ σ )12 , (π1s
∩ σ )1
3 , and for R4 ,
(π14s ∩ σ )12 , (π14s
∩ σ )1
4 , (π14s
∩ σ )1
2
, and so on. Next, the regions of σ̄ must
be linked with a matching algorithm, as for the regions in Chapter 8 representing
families of proximate flats, which select the valid N − 1 points representing a
tangent plane of σ . This is dealt with incrementally for surface types of increasing
complexity.

9.4 ♣ FT-3 Developable Surfaces

The sphere cannot be cut and then flattened without distortion. Motivated by
mapmaking, surfaces were sought whose shape is “close” to spherical and can
be unrolled into a plane without streching or contracting. Euler first considered this
problem, and subsequently, starting in 1771, Monge made major contributions to the
subject of developable surfaces. Monge pointed out potential applications especially
to architecture, paving the way for the modern contoured architectural marvels.
Gauss and others followed with the development of the differential geometry of
 9.4 ♣ FT-3 Developable Surfaces 311

Figure 9.6. Notation and distances for the shift by i units in RN .

Due to the distance of the X̄1
from the y axis, the horizontal translation by N − i + 1 is needed to
obtain the correct x coordinate.

more genenal surfaces. Developable surfaces (“developables” for short) are the
class D ⊂ E that are the envelope of a one-parameter family of planes; they serve
as an excellent starting point for our study of surface representation. Finding their
image, matching, and reconstruction algorithms is straightforward, and the results
offer crucial guides for coping with the more general representation problems. The
pioneering work, and the basis for the exposition here on their representation, is due
to C.K. Hung [93], [92], and the recent more complete exposition [94]. Formally,
let the family of tangent planes describing a developable σ be specified by

π(t) : c(t) · x = c0 (t), t ∈ It ⊂ R, (9.29)

where It is an interval of R. The conditions of (9.35) are added, as proved below,

to guarantee that the corresponding developable is well defined. In other words, it
is not a single plane, a family of parallel planes, nor a pencil of planes intersecting
on a line.
Consider the two planes π(t) : c(t) · x = c0 (t), π(t + t) : c(t + t) · x =
c0 (t + t) intersecting at a line r, which is necessarily also on the plane

[c(t + t) − c(t)] · x = [c0 (t + t) − c0 (t)].

312 9. Hypersurfaces in RN

Dividing by t and taking the limit as t → 0, the limiting position of r is

r(t) : c(t) · x = c0 (t), c

(t) · x = c0
(t), (9.30)

the prime
denoting differentiation with respect to the parameter t. This line is
called a ruling, generator, or characteristic of the plane π(t) that contains it. Three
planes π(t), π(t + 1 t), π(t + 1 t + 2 t) intersect at a point C, which is also on
the plane

[c(t + 1 t + 2 t) − c(t + 2 t)] · x = [c0 (t + 1 t + 2 t) − c0 (t + 1 t)].

As above, dividing by 2 t and taking the limit as 2 t → 0, we obtain the first of

the planes

πd (t + 1 t) : c
(t + 1 t) · x = c0
(t + 1 t), πd (t) : c
(t)) · x = c0
(t),

with the second one obtained similarly. Forming their difference, dividing by 1 t,
and taking the limit as 1 t → 0, we find the limiting position of the point C as the
intersection of the three planes

c(t) · x = c0 (t), c
(t) · x = c0
(t), c

(t) · x = c0

(t). (9.31)

This is the characteristic point of the plane π(t); it lies on the plane’s ruling r(t).
When all the π(t) enveloping a developable σ have the same characteristic point,
σ is a cylinder, whose rulings are parallel if the point is ideal, or a cone when all
rulings lie on a regular point.
We state without proof that in general, the locus of characteristic points is a curve
cc, called the edge of regression, and at each point C(t) ∈ cc the ruling r(t) of the
same plane π(t) is tangent to cc. The family of planes (9.29) satisfying the conditions
in (9.35) has an envelope that is a cylinder, a cone, or a tangential developable having
two sheets generated by the rulings tangent to the curve cc. It is the envelope of
its osculating planes. The intersection of the tangential developable with the plane
normal to the ruling r(t) at a point C(t) has a cusp at that point. Induced by the point
↔ plane duality, there is an interesting duality between space curves, considered as
the locus of a one-parameter family of points and developable surfaces formed by
a one-parameter family of tangent planes [166, p. 72]. This can serve as a blueprint
for an analogous duality in RN , which we do not pursue here. Without entering into
further detail, we mention that a necessary and sufficient condition for a smooth
surface σ ∈ E to be developable is that its Gaussian curvature vanishes [166,
pp. 73–91]. A specific example, the circular double-cone, helps clarify matters.
The parametric equations describing the surface are derived with the aid of Fig. 9.7,
showing a cone CC symmetrical about the x3 axis with apex angle α. Let P ∈ CC
 9.4 ♣ FT-3 Developable Surfaces 313

Figure 9.7. Deriving the parametric equations of a circular cone.

be a point at a distance t from the apex on the circular slice of CC with the plane
x3 = t cos α. The radius of P from the x3 axis is t sin α and at an angle s from the
x1 axis. The coordinates of P in terms of the two parameters (s, t) are
CC : x(s, t) = (x1 = t sin α cos s, x2 = t sin α sin s, x3 = t cos α). (9.32)
Next we derive the family of tangent planes of CC using (9.12):
 
 x1 − x10 x2 − x20 x3 − x30 

 
π :  −t sin α sin s t sin α cos s 0  = 0. (9.33)
 
 sin α cos s sin α sin s cos α 
The parameter t in the second row cancels out, and so the developable CC has a
one-parameter family of tanget planes
π(s) : (x1 − x10 ) cos α cos s + (x2 − x20 ) cos α sin s − (x3 − x30 ) sin α = 0.
(9.34)
314 9. Hypersurfaces in RN

A one-parameter family of planes is mapped into two sets of points, a pair for each
parameter value, so the two regions representing developables are curves. Though
this is intuitively clear, the rigorous proof below, needed to exclude pathological
cases, is involved, and its reading can be omitted.
Theorem 9.4.1 (C.K. Hung; Developable Surfaces) Let a surface σ ∈ D with
tangent planes given by (9.29), and a neighborhood U ⊂ It ⊂ R, where the two
conditions
  
dx(t) ∂ c(t) · di dy(t) ∂ c0 (t)
= = 0, = = 0, (9.35)
dt ∂t c(t) · u dt ∂t c(t) · u

are not simultaneously satisfie for all t ∈ U . Then the set of points {π̄(t)i
|t ∈ U }
representing the tangent planes (9.29) are curves.
♣ FT-3e
** Proof. The points representing the planes (9.29) specifying the developable
surface σ are
π̄(t)i
= (c(t) · di , c0 (t), c(t) · u) ∀t ∈ It . (9.36)
These points are the image of a curve defined by the parameter t ∈ It , provided that
the functions x(t), y(t) are single-valued and the conditions (9.35) are satisfied.
These conditions fail if and only if there exist constants k1 , k2 such that
 
c(t) · di ≡ k1 c(t) · u, c(t) · (di − k1 u) ≡ 0,
⇒ (9.37)
c0 (t) ≡ k2 c(t) · u, c(t) · (k2 u) ≡ c0 (t),
identically for t in the neighborhood U ⊂ It . Consider first the case k2 = 0. Then
necessarily c0 (t) ≡ 0 ⇒ that all the planes π(t) contain the origin. Let ê1 and 1
denote the unit vector and line respectively through the origin in the direction of
(di − k1 u), i.e.,
(di − k1 u)
ê1 = i
= K1 (di − k1 u).
|d − k1 u|
Then the first identity implies that ∀t ∈ U, ê1 is orthogonal to the vector c(t) which
is normal to the plane π(t). Unless c
(t) ≡ 0 for all t ∈ U , then
)
π(t) = 1 , (9.38)
t∈U
showing that the family of planes π(t) given by (9.29), rather than being the envelope
of a developable surface, is either a single plane (and hence σ itself) or a pencil of
planes on the line 1 .
 9.4 ♣ FT-3 Developable Surfaces 315

Next consider the case k2 = 0. From the first identity, ê1 ⊥ c(t), and from the
second, c(t) ⊥ k2 u ⇒ ê1 and k2 u must be linearly independent; for if not, then
they would be in the same direction and hence both would be orthogonal to c(t).
So every linear combination with nonzero coefficients of the two vectors is not zero,
and in particular, u − (u · ê1 )ê1 = 0. Let us denote the unit vector in this direction
by ê2 , i.e.,
(u − (u · ê1 )ê1 )
ê2 = = K2 [u − (u · ê1 )ê1 ].
|u − (u · ê1 )ê1 |
The two unit vectors are orthogonal, since ê1 · ê2 = K2 ê1 · (u − (u · ê1 )ê1 ) = 0.
Further, c(t) · ê2 = K2 c(t) · u. Then (9.37) reduces to
(
c(t) · ê1 ≡ 0,
(9.39)
c(t) · ê2 ≡ Kc0 (t),

with K constant. Recall the vector identities


(a × b · c) = (a · b × c),
(9.40)
(a×b)×c) = (c · a)b − (a · b)a.

We verify by direct substitution that the vector

c(t) = Kc0 (t)ê2 + g(t)(ê1 × ê2 ) (9.41)

satisfies (9.39) for an arbitrary function g(t) ∈ C 1 . To wit,


c(t) · ê1 = Kc0 (t)ê2 · ê1 + g(t)(ê1 × ê2 ) · ê1 = 0,
(9.42)
c(t) · ê2 = Kc0 (t)ê2 · ê1 + g(t)(ê1 × ê2 ) × ê2 = Kc0 (t),

from the orthogonality of ê1 and ê2 and the vector identities. It consists of a vector
in the direction of ê2 and another perpendicular to the direction of ê1 × ê2 . The
family of tangent planes, (9.29), can be rewritten as

[Kc0 (t)ê2 + g(t)(ê1 × ê2 )] · x = c0 (t). (9.43)

From

(Kc0 (t)ê2 + g(t)(ê1 × ê2 )) · ê1 ≡ 0,
(Kc0 (t)ê2 + g(t)(ê1 × ê2 )) · (ê2 + s ê1 ) = Kc0 (t),

it is clear that the line 2 : ê2 + s ê1 , parametrized by s, is contained in all the
tangent planes (9.43) of σ . Again, provided that c
(t) = 0, these planes, rather than
316 9. Hypersurfaces in RN

being the envelope of a surface, are a pencil of planes on the line 2 . In short, the
requirement that the tangent planes of a surface σ ∈ D not satisfy the conditions
in (9.35) ensures that (9.36) defines two (piecewise) C 1 curves in R3 that represent
a developable not simultaneously satisfying the two conditions of (9.35). Use of
vector notation throughout allows us to conclude that the N − 1 curves

σ̄i
(t) = {π̄i
(t) = (c(t) · di , c0 (t), c(t) · u)}, i = 0, . . . , N − 2, (9.44)

represent such a developable in RN . Since each point set π̄i

(t) depends on only
one parameter, it is clear that for each i, it is a curve σ̄i
(t) rather than a region (i.e.,
it has no interior with points). That is, for σ a developable,

σ̄ = {∂ σ̄i
|i = 0, . . . , N − 2}.


9.4.1 ♣ FT-4 Ambiguities and Uniqueness

Can a pair of curves represent different developable surfaces? Moreover, can a
developable surface be reconstructed from its representation? These questions are
interrelated, and a proper answer requires an algorithm for identifying every tangent
plane of σ from the two points, one on each of the representing curves. The solution
is immediate when the curves are available in parametric form σ̄123 (t) and σ̄231
(t);
each value of t provides two unique points identifying each tangent plane π(t).
The challenge arises when the representing curves σ̄123 , σ̄231
are given without
parametrization. Recall that a plane π is uniquely identified by two points having
the same y-coordinate. Therefore, if one of the representing curves is monotonic
in the y-direction, so is the other. The intersection of the horizontal line through a
point π̄123 ∈ σ̄123 with the curve σ̄231
is on the second point π̄231
corresponding
to the tangent plane π. In effect, then the curves can be parametrized by either x or
y. Otherwise, for each monotone portion of the σ̄123 curve there must be a corre-
sponding monotone portion of σ̄1
23 in the same range of y. If not, the curves cannot
represent a developable surface. A unique correspondence can be established by a
pair of start points π̄123 ∈ σ̄123 , π̄231
∈ σ̄231
, on a horizontal line H̄ identifying
a specific tangent plane π. Tracking the continuous translation of the line H̄ in the
direction of increasing or decreasing y as in Fig. 9.8 provides the correspondence of
point pairs, one on each curve, identifying the tangent planes. Between the curves’
extrema, H̄ may intersect each curve at several points. For example, there are two
points of intersection for each ellipse in Fig. 9.8, allowing for four possible pairings
of start points, providing, in general, four different developables; and there are
other problems. In the examples below, though the curves may be traversable with
 9.4 ♣ FT-3 Developable Surfaces 317

Figure 9.8. The pair of points π̄123 , π̄231

represents an unambiguous plane.
By contrast the ᾱ and β̄ points represent planes which are a1, ambiguous in the first sense.

a proper partition and matching, they may not always represent a valid developable.
This issue is examined later on.
♣ FT-4e
Definition (Ambiguous plane) A tangent plane π of a surface σ ∈ D is ambigu-
ous in the first sense, denoted by a1, if either π̄123 or π̄1
23 is an extreme point in the
y-direction of its corresponding curve. It is ambiguous in the second sense, denoted
by a2, if either of the points is a crossing point on its curve. All other planes are
unambiguous.
The notation a1 or a2 is used interchangeably either for ambiguous planes or the
points representing them. Returning to Fig. 9.8, it is seen that the planes α and β are
a1, since they are represented by extreme points of the curves. They cannot serve as
start points, for in the absence of other information, the curve’s traversal can proceed
either to the left or right of the points. Instances of different representing points on
one of the curves are shown in Fig. 9.9. Such points are necessarily at an extremum.
Figure 9.10 illustrates the ambiguity, due to the multiple choice of directions, in
attempting to traverse the curve starting from a crossing point. Otherwise, choosing a
start point, say γ̄123 , limits the choice of γ̄231
to be on a matching monotone portion
inducing the traversal in the indicated direction. Reaching the crossing point43
confronts us with a choice of directions. Here the stipulation that the curves be C 1
eliminates the ambiguity, forcing the traversal to occur along the path preserving
the continuity of the direction. That is, the path of traversal has no “kinks.” Another

43
In Fig. 9.24 it is seen that crossing points represent a bitangent plane. I am grateful to Y. Perlmutter
for proving that in the representation of developables, crossing points occur in pairs; Exercise 3.
318 9. Hypersurfaces in RN

Figure 9.9. ** Points representing different tangent planes may overlap at extreme points.
The four planes α, β, γ , δ are a1.

Figure 9.10. Crossings points a2 cannot serve as start points (unless they are also va, see Fig. 9.11.)
There is no problem traversing them once a proper start point is chosen.

choice of start points, say δ̄123 , δ̄231

, induces an opposite traversal direction. The
example illustrates the need for an additional requirement: there exists a continuous
tracing in the x direction of both curves such that the pair of extrema, say at y = ymi ,
is traversed simultaneously on both curves. In other words, matching extrema occur
in the same sequence as x increases. A pair of such curves is called well matched or
simply matched. So far, we have seen that while a1 and a2 cannot be used as starting
points, they pose no difficulty in traversing once a proper start point is chosen.
There is a subcase of a2 that defeats the traversal even when good start points are
available.
Definition (Very Ambiguous plane) Let π be an a2 plane such that at least one
of its representing points, say π̄123 , is a crossing point. It is called very ambiguous,
and denoted by va, if in addition either
 9.4 ♣ FT-3 Developable Surfaces 319

Figure 9.11. One of the curves representing a developable surface σ has four a1 points.
They are marked by “x.” There is also an a2 point marked by “o,” which, being a crossing point
where the tangents of the curve’s two parts coincide, is also a va.

1. π̄123 (and hence π̄231

) is not an ideal point and at least one of the curves has
a pair of branches sharing the same tangent direction at π̄123 (or π̄231
), or
2. π̄123 (and hence π̄231
) is an ideal point shared by a pair of curves’ branches.

When a va point (see Figs. 9.11 and 9.12) is reached, the traversal cannot
proceed, since either choice of branch is possible. Of course, a way to cope with
this ambiguity is to partition the curve(s) at va points into separate branches without
va points and then match the pairs as outlined earlier. The strategy is to proceed
with the curves’ traversal, avoiding previously traced portions of the curve. The
discussion is basically the proof of the result formally stated below.
Lemma 9.4.2 (Unique Representation of Developables) Let σ be a smooth
developable with no va tangents represented by two smooth curves σ̄123 , σ̄1
23 and
π (referred to as the start plane) any unambiguous tangent plane of π . Then the rep-
resenting curves σ̄123 , σ̄231
together with the two points π̄123 ∈ σ̄123 , π̄231
∈ σ̄231

uniquely represent σ .
The suggested partitioning above at va points, let us call it patching, can be
applied at points with more than two self-intersections, again resolving ambiguities.
Allowing patching, there always exists a unique developable surface represented
by each pair of the resulting patched curves and good start points.
320 9. Hypersurfaces in RN

Figure 9.12. ** Two branches whose asymptotes are all horizontal.

The ideal point in the horizontal direction is a va.

9.4.2 Reconstruction of Developable Surfaces

Henceforth, when we refer to the representation σ̄ of a developable, the curve por-
tions under discussion are assumed smooth and monotone. That is, σ̄ is partitioned
and matched into portions that can be uniquely traversed such as σ̄123 and σ̄1
23
seen in Fig. 9.13, with which we illustrate the reconstruction of the developable σ ;
reference also to Fig. 9.4 is helpful. The points π̄123 , π̄1
23 with the same y represent
a tangent plane of the developable σ . The tangent lines P̄ , Q̄ represent the points

P ∈ π ∩ π0s
∩ σ, Q ∈ π ∩ π1s
∩ σ,

so that the line r = π ∩ σ on P , Q is a ruling of the developable with

r̄23 = P̄ ∩ Q̄, r̄12 = Q̄0 ∩ P̄ , (9.45)

where Q̄0 denotes the representation of Q in the first coordinate system X̄1 , X̄2 , X̄3
(recall the notation after Theorem 9.3.6).
The representation of r is based on the construction discussed in Section 5.3.1
for determining that a line is contained in a plane. To repeat, the point P on the
ruling r is on σ and π0s
and Q ∈ r ∩ π1s
. All points on this ruling (and σ ) are
constructed in the usual way for lines, i.e., choosing a value of one variable and
constructing the remaining coordinates from the points r̄12 , r̄23 .
Theorem 9.4.3. A developable with a well-matched representation is (piecewise)
reconstructible.
 9.4 ♣ FT-3 Developable Surfaces 321

Figure 9.13. Reconstructing a developable.

The ruling r is represented by r̄23 = P̄ ∩ Q̄, shown by a 0, and r̄12 = P̄ ∩ Q̄0
, shown by a 2,
where Q̄0
is the 1, 2 part of the representation of Q with respect to the coordinate axes X̄1 , X̄2 , X̄3 .

Other rulings are constructed in the same way: choosing a value y that deter-
mines the pair of points representing a tangent plane, finding the tangents, etc.
Conversely, given the two representing curves of a developable σ , a line is one
of its rulings if and only if there are two lines on ¯23 tangent to σ̄123 and σ̄1
23
at points with the same y. An example of a line that is not a ruling is seen in
Fig. 9.14.
What is the locus rr 23 of all points r̄23 where r is a ruling of a developable?
This is obtained as the result of performing the above construction for all y in the
representation of σ . The intersection of the two tangents at the pair of points with the
same y representing a tangent plane for all y needs to found (Exercise 7). Conversely,
given the two loci, construct the developable from its rulings (see Exercise 8). There
follows a pleasing result on the rulings of a developable with reflected representing
curves.
Corollary 9.4.4 (Atariah, Biniaminy, Itai, Kosloff (ABIK); Rulings of reflec
ted representing curves) If a developable σ is represented by two curves σ̄0
, σ̄1
,
one the reflectio of the other about a line x = a, then the rulings’ representing
curves are two vertical lines with rr 23 at x = a and rr 12 at x = 3 − 2a.
The proof is left as an exercise (see Exercise 6 below).
322 9. Hypersurfaces in RN

Figure 9.14. The line with the ¯23 shown cannot be a ruling of the developable σ .
This is because the lines on ¯23 have points of tangency to σ̄123 and σ̄1
23 with different y.

Let us look at a couple of examples44 that also illustrate the reflected represent-
ing curves property. Figure 9.16 (left) shows two cubic curves, one the reflection
of the other about a vertical line, representing a developable σ whose rulings
are represented by two vertical lines (right). The developable σ is reconstructed
from its rulings and seen in Fig. 9.17. The line of cusps is the dual of the inflec-
tion points in the representing curves. It is seen and discussed further later on.
Another example is shown in Fig. 9.18, where the representing curves (right) are
two exponentials where there is no symmetry with respect to a vertical line. The
curves representing the rulings are shown on the (right) and the developable in
Fig. 9.19.
Points on the rulings are necessarily surface points of the developable. This
raises the question of how to recognize on which “side” of the developable an
arbitrary point lies, in particular, when the developable is closed, how to con-
struct interior points. This is answered partially, starting with cylinders, in the next
section.

44
I am grateful to my students D. Atariah, M. Biniaminy, T. Itai, and Z. Kosloff for providing these
examples and results, including the corollary above from their course project.
 9.4 ♣ FT-3 Developable Surfaces 323

Figure 9.15. For the rulings r the locus of points r̄23 , constructed as in Fig. 9.13.
This is a curve rr 23 . At the pair of points marked by • the tangents are vertical, resulting in an ideal
point spliting rr 23 . The arrow’s direction indicates the y coordinate of r̄23 increases monotonically
with the y of the corresponding points representing to the tangent plane.

Exercises
1. Which of the following ambiguities are “resolvable” and how?
(a) representation curves have one horizontal segment.
(b) representation curves have multiple crossings at the same point.
2. For an ambiguous plane π can only one of π̄123 , π̄231
be an extremum but not
the other?
3. Give an example of a pair of curves that are not C 1 and cannot uniquely represent
a developable, and another such pair that uniquely represent a developable
surface.
4. In the representation σ̄123 , σ̄1
23 of a developable σ prove that for a ruling r,
r̄12 = r̄23 , where at the corresponding pair π̄123 , π̄1
23 one of the tangents is
horizontal. Hint: see the highest pair of points on the two curves in Fig. 9.15.
5. Prove that the point of intersection σ̄0
∩ σ̄1
is the r̄23 of a ruling r of the
developable σ , e.g., the intersection of the two curves in Fig. 9.18 is such a
point.
324 9. Hypersurfaces in RN

Figure 9.16. Two cubic curves symmetric about a line x = a representing a developable.
On the right are the loci representing the developable’s rulings; two vertical lines with rr 23 at x = a
and rr 12 at x = 3 − 2a.

Figure 9.17. The reconstructed developable from two viewing angles.

6. Prove that if a developable σ is represented by two curves σ̄0

, σ̄1
, one the
reflection of the other about a line x = a, then the representing curves of the
rulings are two vertical lines with rr 23 at x = a and rr 12 at x = 3 − 2a. This
is corollary 9.4.4.
7. Write and implement the algorithm whose input is the representation of
σ̄123 , σ̄1
23 , as a function of y and partitioned in monotone pairs, and outputs
the plots of the loci rr 23 , rr 12 . This is a good course project.
 9.4 ♣ FT-3 Developable Surfaces 325

Figure 9.18. Reconstruction of developable from two exponential curves representing it.
On the right are the rulings’ representing curves: solid rr 12 and dashed lines rr 23 respectively.

Figure 9.19. The reconstructed developable from two viewing angles.

8. Write and implement the algorithm whose input is the two curves rr 23 , rr 12
(above) and that draws the rulings of the corresponding developable. This is a
good course project.

9.4.3 Specific Developables

We look now at specific developables, starting with the circular cylinders and cones,
and study their representation, rulings, interior points, symmetry properties, and
generalizations. Then we continue with more general developables.
326 9. Hypersurfaces in RN

Figure 9.20. ♣ FT-5 A pair of hyperbolas representing the cylinder shown on the left.
The • are the “handles” for changing orientation in the software used and have no significance here.

Cylinders
Corollary 9.4.5 (C.K. Hung; Cylinders) Elliptic cylinders in R3 are repre-
sented by a pair of hyperbolas.
Being quadrics by Corollary 9.3.5, they are represented by a pair of conic
curves. The cylinder’s orientation determines the number √ of tangent planes it has
parallel to the line u with direction unit vector u = 1/ 3(ê1 + ê2 + ê3 ), which, in
turn, by Corollary 9.3.3 determines the representing curve’s ideal points and hence
the type of conics representing the cylinder. For every plane tangent to a circular
cylinder there is another tangent plane parallel to it at a distance of D, the cylinder’s
diameter, yielding a pair of hyperbolas. So parabolas cannot be representing curves
of a circular cylinder. This observation is originally due to A. Chatterjee [93].
Further, the representing curves cannot be ellipses that, according to the section
on conic transforms in Chapter 7 on curves, are the images of hyperbolas. The
intersection of a plane (including the superplanes) with a cylinder is an ellipse or a
pair of lines in the degenerate case. A circular cylinder is easily transformed to one
having elliptical sections perpendicular to its symmetry axis, and therefore these
observations are equally true for elliptical cylinders.
A cylinder’s rulings are parallel; hence its rr 12 and rr 23 must be two vertical
segments. Figures 9.20 and 9.21 are two examples of the cylinders’ representation
by hyperbolas. As noted earlier (Exercise 5), the points of intersection of σ̄0
∩ σ̄1

represent the r̄23 of a ruling. Hence the interval between the two points of intersection
is the rr 23 of the cylinder’s rulings. This shows an interesting property of hyperbolic
curves. Namely, at any point of rr 23 there are two lines tangent to hyperbolas (above
 9.4 ♣ FT-3 Developable Surfaces 327

Figure 9.21. Two hyperbolas, one coincident with the X̄3 axis, representing the cylinder.

or below) at points with the same y-coordinate, since these are the representing
points of a tangent plane (Exercise 2 below).
Let us look at this more carefully with the aid of Fig. 9.22 (left), starting
with the top point of rr 23 , the vertical interval between the two intersections.
The construction is shown for determining the top point of rr 12 by converting
the straight-line tangent on σ̄1
23 = σ̄1
to a polygonal line in the X̄1 , X̄2 , X̄3 axis
system and intersecting it with the straight line tangent to σ̄123 = σ̄0
, and similarly
for the bottom point, and thus all of rr 12 is determined. If needed, the interval rr 13
is found in the same way. By the way, once the representation of a cylinder’s ruling
is found, the cylinder’s orientation, being the same, is also known.
Next, the pair of points representing a specific ruling need to be matched, as
shown in Fig. 9.22 (right). Choosing a point r̄12 ∈ rr 12 , the tangents T̄ , T̄
from it to
the two branches of σ̄0
intersect the line rr 23 at two points r̄23 , r̄23
corresponding to
the two r, r
rulings with the same r̄12 . The points of tangency are the representing
points π̄0
, π̄0

of two tangent planes. The points r̄12 , r̄23 , π̄0
are on the tangent
line T̄ , and it is easy to see that also r̄1
2 , r̄23 , π̄1
are collinear (not shown). We can
conclude that the line r is on the tangent plane π and hence is a ruling, and similarly
for r
. Hence T , T
are points on r, r
, and in fact, on the cylinder’s intersection
with π0s
, the first sp forming the curve σ0
. Any point ¯23 between r̄23 , r̄23
together
with r̄12 represents a line parallel to the rulings but interior to the cylinder. So any
point on is in the cylinder’s interior. Conversely, to check whether a point P is
328 9. Hypersurfaces in RN

Figure 9.22. Portions of two hyperbolas σ̄0

= σ̄123 , σ̄1
= σ̄1
23 representing a cylinder.
(Left) The vertical interval between the two intersections, which necessarily have the same
x-coordinate, is the locus rr 23 of the cylinder’s rulings. The top point of the vertical interval rr 12 is
constructed from the top intersection, and the lowest point is at the intersection of the vertical line
and the tangent to σ̄0
at the low intersection point. (Right) For each r̄12 ∈ rr 12 there are two rulings

are at the intersections of the tangents to σ̄
from r̄ ∈ rr . Between these two
r, r
. The r̄23 , r̄23 0 12 12
points, any other point ¯23 ∈ rr 23 together with r̄12 represents a line internal to the cylinder parallel
to its rulings.

interior, it suffices to check whether there is an interior line parallel to the rulings
with P ∈ (Exercise 5).
** A cylinder’s representation as a function of its orientation. The examination
of the dependence of a cylinder’s representation on its orientation may be safely
omitted (up to and including (9.59)) by readers not interested in these details.
To obtain the equation of a circular cylinder with general orientation, we start
with the cylinder studied earlier,

CY : x22 + x32 = r 2 , (9.46)

which is symmetric about the x1 axis, with radius r. The rotation of the coordinate
system about the origin is applied using the transformation derived in Section 2.8
 9.4 ♣ FT-3 Developable Surfaces 329

and is restated for convenience. The coordinate system x1 , x2 , x3 is rotated about

the origin O to the system x1
, x2
, x3
, all points remaining stationary, so that a point
P has two sets of coordinates (x1 , x2 , x3 ), (x1
, x2
, x3
). Letting the direction cosines
of the xi
axes with respect to xi be cos αi , cos βi , cos γi for i = 1, 2, 3, respectively,
we obtain

cos2 αi + cos2 βi + cos2 γi = 1, cos αj cos αk + cos βj βk + cos γj γk ,

j = k, i, j, k = 1, 2, 3. (9.47)

The first is due to the property of direction cosines, and the second due to the
orthogonality between all pairs of axes xj , xk , i = j = k. The transformation
relating the two coordinate systems is
⎛ ⎞
cos α1 cos β1 cos γ1 0
⎜ cos α2 cos β2 cos γ2 0 ⎟
(x1
, x2
, x3
, 1) = (x1 , x2 , x3 , 1) ⎜
⎝ cos α3 cos β3
⎟
cos γ3 0 ⎠
0 0 0 1

= (x1 , x2 , x3 , 1)A, (9.48)

⎛ ⎞
cos α1 cos α2 cos α3 0
⎜ cos β1 cos β2 cos β3 0 ⎟
(x1 , x2 , x3 , 1) = (x1
, x2
, x3
, 1) ⎜
⎝ cos γ1
⎟
cos γ2 cos γ3 0 ⎠
0 0 0 1

= (x1
, x2
, x3
, 1)AT . (9.49)

The matrix A is orthogonal with determinant |A| = 1 and inverse its transpose AT ,
which is the matrix of the inverse transformation {x1
, x2
, x3
} → {x1 , x2 , x3 }.
To apply the rotation transformation, CY is expressed in matrix form:
⎛ ⎞⎛ ⎞
0 0 0 0 x1
⎜ 0 1 0 0 ⎟ ⎜ ⎟
CY : (x1 , x2 , x3 , 1) ⎜ ⎟ ⎜ x2 ⎟
⎝ 0 0 1 0 ⎠ ⎝ x3 ⎠
0 0 0 −r 2 1

= (x1 , x2 , x3 , 1)C(x1 , x2 , x3 , 1)T . (9.50)

Applying the transformation yields the cylinder

CG : (x1
, x2
, x3
, 1)AT CA(x1
, x2
, x3
, 1)T (9.51)
330 9. Hypersurfaces in RN

with the matrix AT CA written below:

⎛
cos2 α2 + cos2 α3 + cos α2 cos β2 + α3 cos β3
⎜ cos α cos β + α cos β cos2 β2 + cos2 γ2
⎜ 2 2 3 3
⎜
⎝ cos α2 cos γ2 + cos α3 cos γ3 cos β2 cos γ2 + cos β3 cos γ3
0 0
⎞
cos α2 cos γ2 + cos α3 cos γ3 0
cos β2 cos γ2 + cos β3 cos γ3 ⎟
0 ⎟
⎟
cos2 γ2 + cos2 γ3 0 ⎠
0 −r 2
Applying the conditions of (9.47) results in
⎛ ⎞
1 − cos2 α1 − cos α1 cos β1 − cos α1 cos γ1 0
⎜ − cos α cos β 1 − cos2 β1 − cos β1 cos γ1 ⎟
⎜ 1 1 0 ⎟
AT CA = ⎜ ⎟.
⎝ − cos α1 cos γ1 − cos β1 cos γ1 1 − cos2 γ1 0 ⎠
0 0 0 −r 2
Completing the evaluation, with ai = cos αi , yields


3
CG : (1 + ai2 )xi2 − {(a1 x1 + a2 x2 )2 + (a1 x1 + a3 x3 )2 + (a2 x2 + a3 x3 )2 }
i=1

= r 2, (9.52)

where the superscript

is dropped. Note that the cylinder is given only in terms of
the αi , which are the direction angles of its axis of symmetry, which we call s. For
example, for a1 = 1, a2 = a3 =√0, s is the x1 axis, a2 = 1, a1 = a3 = 0, s is the
x2 axis. For a1 = a2 = a3 = 1/ 3, the axis √ of symmetry s is the line earlier called
u having direction the unit vector û = 1/ 3(ê1 + ê2 + ê3 ), or explicitly, yielding
3 2
CG : x12 + x22 + x32 − (x1 x2 + x1 x3 + x2 x3 ) = r . (9.53)
2
From Lemma 9.3.1, the representation of CG is the image of two curves c0 , c1 ,
which are its intersections with the first two superplanes (see discussion of previous
example (9.17) to (9.28)). Proceeding, for the curve (CG)12 = CG ∩ π0s
= c̄0 it
suffices to find (CG ∩ π0s
)23 = c̄23
0 . With π s : x − 2x + x = 0, substituting
0
1 2 3
x1 = 2x2 − x3 into (9.52) yields the c0 curve’s projection on the x2 x3 plane:
 9.4 ♣ FT-3 Developable Surfaces 331

0
(CG ∩ π0s
)23 = c23 : (5 − B12 )x22 + (2 − B22 )x32 − 2(2 + B1 B2 )x2 x3 = r 2 ,
(9.54)

where B1 = (2a1 + a2 ) and B2 = (a3 − a1 ). The discriminant (see (9.22)) is

 = A24 − A1 A2 = −6 + B12 + B22 + (B1 + 2B2 )2 ,

since the intersection of a cylinder with a plane is an ellipse  < 0 except possibly
for a degenerate case. Recall that the superplanes are the pencils of planes on u such
that any superplane, and so also π0s
, π1s
, intersects the cylinder given by (9.53) in
two parallel lines (rulings), and since  = 0 in this case, they are considered
a degenerate parabola. Otherwise, the image CG23 = (CG ∩ π0s
)23 is always a
hyperbola and can be found via (9.22) using the interaxis distance d = 1 and
afterward translating x by 1 (the distance of the X̄2 from the y axis). It can also
be computed without the translation of x from the image of (CG ∩ π0s
)12 = c12 0 if
available, and in either case the result is the same:

(CG)12 = CG0
: I x 2 + (/r 2 )y 2 − [3 − B2 (B1 + B2 )]x − (2 − B22 ) = 0,
(9.55)

where I = 3 − (a1 + a2 + a3 )2 . As mentioned, the cylinder’s axis of symmetry

passes through the origin. Translating the origin does not change the coefficients of
the quadratic terms in (9.65) and therefore has no effect in determining the type of
conic obtained for the image (CG)12 .
In order to make a comparison later we also find the image (CG)23 . Proceed-
ing as before for the intersection (CG)23 = CG ∩ π1s
= c̄1 , it suffices to find
(CG ∩ π1s
)23 = c̄23
1 with π s : x + 2x − x = 0, substituting x = 2x − x into
1
1 2 3 1 3 2
1
(9.65) to obtain the c curve’s projection on the x2 x3 -plane:
1
(CG ∩ π1s
)23 = c23 : (2 − C12 )x22 + (5 − C22 )x32 − 2(2 + C1 C2 )x2 x3 = r 2 ,
(9.56)

where C1 = (a2 − a1 ) and C2 = (2a1 + a3 ). Its transform is

CG23 = CG1
: I (x + 1)2 + (/r 2 )y 2 − [3 − C2 (C1 + C2 )](x + 1) − (2 − C22 )
= 0. (9.57)

The translation (x + 1) is due to the X̄2 axis being one unit from the y axis. This
together with (9.54) describes the pair of hyperbolas in the representation. Returning
to the representation of the cylinder CD given by (9.53) it is found from (9.55) by
332 9. Hypersurfaces in RN

√
substituting a1 = a2 = a3 = 1/ 3, together with x1 = 2x2 − x3 , and is the pair
of lines
 √
x3 = x2 ± r/ 2,
√ (9.58)
x2 = x1 ± r/ 2.

Each of these lines is represented by two vertical lines. So the degenerate cylinder
representation consists of the four vertical lines ¯12 , ¯23 , two for each index pair:
√ √
¯12 : x = ± r/ 2, ¯23 : x = 1 ± r/ 2. (9.59)

In vector notation, a circular cylinder is given by

x(t, v) = b + r(ŝ1 cos t + ŝ2 sin t) + vŝ, (9.60)

where ŝ is the unit vector in the axis direction and the two other unit vectors are
such that s = ŝ1 × ŝ2 , which are implicitly given in terms of the basic unit vectors
êi , i = 1, 2, 3. The two parameters v, t are real numbers with 0 ≤ t ≤ 2π. The
constant vector b and the cylinder’s axis of symmetry s share a point, which is the
center of a circle of radius r, and a point on its circumference is given by the vector
p = r(ŝ1 cos t + ŝ2 sin t). This vector is normal to the tangent plane

π(t) : (ŝ1 cos t + ŝ2 sin t) · x = (ŝ1 cos t + ŝ2 sin t) · b + r (9.61)

containing the ruling parallel to s given by a fixed value of t in (9.60). The rep-
resentation can be found directly from Theorem 9.4.1 and the family of tangent
planes above as t ∈ [0, 2π] rather than from the intersections with the superplanes
π0s
, π1s
. To do this, (9.61) is rewritten as


3
π(t) : (A
i sin t + Bi
cos t)xi = (A0 sin t + B0 cos t + C0 ), (9.62)
i=1

where the constants A

i , Bi
, Ci
derive from the decomposition of the unit vectors
ŝ, ŝ1 , ŝ2 in terms of the êi , i = 1, 2, 3. The parallel-coordinate representation of the
planes π(t) are the points with homogeneous coordinates (x
(t), y
(t), S(t)):
⎧

⎨ x (t) = A1 sin t + B1 cos t,

⎪
y
(t) = A0 sin t + B0 cos t + C0 , (9.63)
⎪
⎩
S(t) = A4 sin t + B4 cos t,
where S(t) is the sum of the xi coefficients. As t varies, a curve on the projective
plane is traced having ideal points where S(t) = 0. The detailed form of the
 9.4 ♣ FT-3 Developable Surfaces 333

constants is immaterial to the subsequent calculation. Solving for cos t and sin t
and substituting in the trigonometric identity sin2 t + cos2 t = 1 yields
 
A4 x
− A1 S 2 B1 (y
− C0 ) − B0 x
2
+ = 1. (9.64)
A4 B1 − A1 B4 A1 B0 − A0 B1
With the replacements x = x
/S, y = y
/S, this equation describes a regular or
degenerate conic, in this case, one of the representing hyperbolic curves of the
cylinder in (9.60).
Generalized cylinders. A surface’s description in vector form is convenient and
also clarifies its structure. It was proposed by the mathematical physicist W.J. Gibbs,
who was also an early advocate for visualization in science45 [68], [69].
A general cylinder is specified by the equation
x(s, v) = c(s) + vr, (9.65)
where c(s) is a space curve given in terms of its arc length s, v is the second
parameter, and r a constant vector. Its rulings are given by vr with specific values
of s and are therefore parallel. In (9.60), c(s) is the circle with radius r centered
at b, and the rulings are vs. We can conclude that the intersections of π0s
, π1s
are
perspectives (in the sense of perspective transformations in projective geometry)
of curves whose images provide the representing curves. By the way, this form
is dimension-invariant, and we can consider it as defining a surface in RN , a
developable being a hypersurface with a one-parameter family of tangent planes
represented by N − 1 curves. From now on, unless otherwise specified, cylinder
refers to a general cylinder. In Fig. 9.23 a cylinder is shown having cusps along one
of its rulings. Its two representing curves, each with an inflection point, remind us of
the known developable ↔ curve duality. In fact, we met an example of this duality
earlier, in Fig. 9.17, also showing a generalized cylinder having two representing
curves with inflection points, Fig. 9.16, with rr 12 , rr 23 two vertical lines due to
its parallel rulings. There is a further pleasing duality here: the representing curves
seen in Fig. 9.24 indicate that a plane tangent to two rulings (i.e., bitangent) exists
and is represented by crossing points, one on each curve, akin to the bitangent
↔ crossing point for curves. From our previous discussions, it is clear that the
vertical semi-infinite line rr 23 goes through the crossing point that after all, is the
common r̄23 of the two rulings on the bitangent plane. This also points out that the
crossing point and cusps in the representing curves seen earlier in Figs. 9.9 and 9.10
correspond to bitangent planes and rulings that are the loci of inflection points in
the developables.
45
For the importance of visual thinking in science see [122].
334 9. Hypersurfaces in RN

Figure 9.23. ♣ FT-6 A general cylinder illustrating the developable ↔ curve duality.
The ruling formed by cusps is transformed to an inflectio point in each of the representing curves.

Figure 9.24. ♣ FT-6e The two leaves of the surface in the previous figure are extended.
Here the surface, in addition to the line of cusps, has a bitangent plane, which is represented by the
two representing curves crossing at the same point. Such a hypersurface in RN is represented by
N − 1 curves with crossing points and inflection points at the same y coordinate (Exercise 16).

Symmetry and coincidence. A symmetric pair of representing curves necessarily

represents a cylinder, as stated in Corollary 9.4.4. An example of the developable’s
reconstruction from its rulings is shown in Fig. 9.16. Earlier, we also saw that
a cylinder with (9.17) having rulings parallel to the x1 axis has two coincident
 9.4 ♣ FT-3 Developable Surfaces 335

Figure 9.25. Coinciding curves represent a cylinder with rulings parallel to the x1 axis.
¯ 12 is on the X̄2 axis and rr
The rr ¯ 13 on the X̄3 axis. A ruling r on a point P and its representation in
-coords is shown.

hyperbolic representing curves. These are instances of general results characteristic

of cylinders, which, together with variants of Corollary 9.4.4, were discovered by
C.K. Hung [93].
Corollary 9.4.6 (C.K. Hung; Coincident representing curves) A developable
σ in R3 is a cylinder with rulings parallel to the x1 axis if and only its two
representing curves σ̄12 , σ̄23 coincide.
Proof. An instance of the two curves σ̄12 , σ̄23 coinciding, shown in Fig. 9.25, clari-
fies the proof. A tangent P̄ at any point represents a point on σ0
with coordinates
(p1 , p2 , p3 ) and also a point on σ1
with coordinates P
= (p1
, p2 , p3 ). With
respect to the axes X̄1
, X̄2 , X̄3 , P̄
is a straight line (coincident with P̄ ) and a
polygonal line with respect to X̄1 , X̄2 , X̄3 . The intersection P̄ ∩ P̄
contains the
interval of P̄ between the X̄2 and X̄3 axes. This shows that the line r on P , P
is
perpendicular to the x2 x3 plane at the point (0, p2 , p3 ) and therefore parallel to the
x1 axis. Note that r̄12 is on the X̄2 axis at the value p2 . Since this is true for any
point, the r̄12 are on a vertical line rr 12 on the X̄2 axis. For the second point in
representation of the r we can use the r̄13 , which are all on a vertical line rr 13 on
the X̄3 axis. So the lines r formed by the points P , P
are all parallel, so they are
rulings and σ must be a cylinder. The rr 12 , rr 13 may be complete or semi-infinite
lines or finite intervals.
For the proof in the opposite direction assume that σ is a cylinder with rulings
parallel to the x1 axis. Then rr 12 , rr 13 are the X̄2 , X̄3 axes respectively. Choose
a pair r̄12 , r̄13 representing a ruling r. The points are the intersections of a line
P̄ tangent to σ̄0
and another line Q̄ tangent to σ̄1
. Because r̄12 , r̄13 ∈ P̄ ∩ Q̄,
336 9. Hypersurfaces in RN

the whole interval from r̄12 to r̄13 is contained in P̄ ∩ Q̄. Two lines sharing an
interval must be identical, that is, P̄ = Q̄. So these must be tangents to the same
curve possibly at two different points, which must have the same y value. The curve
representing a developable is locally monotonic; hence the points of tangency must
be identical if their y values are equal. We can conclude that the two representing
curves coincide. 

♣ FT-7 Cones

A developable all of whose rulings intersect on a single characteristic point, recall

(9.31), is a cone whose representing curves and rulings are studied next.
Corollary 9.4.7 (Cones; C.K. Hung) Circular Cones in R3 are represented by
a pair of conic curves.
Circular cones, being quadrics, are, by Corollary 9.3.5, represented by a pair of
conic curves whose type is determined, via Corollary 9.3.3, by the number of
tangent√ planes the cone has parallel to the line u in the direction of the unit vector
û = 1/ 3(ê1 + ê2 + ê3 ). Unlike cylinders, all three types of conics can represent
curves, as stated in Table 9.1. The angle θ between the cone’s ruling and the axis
of symmetry s is constant. So for û to be parallel to πt , one of the cone’s tangent
planes, there must exist a plane π containing û such that the angle π ∩ s equals
θ. Consider two tangent planes πt1 , πt2 intersecting at the line . If u is parallel to
, then it is also parallel to the two tangent planes, and it is clear that u cannot be
parallel to more than two tangent planes.
Concerning the rulings from Figs. 9.27, 9.29, 9.31 where one of conics collapses
to a line or line segment, we learn that at least one of the curves rr 23 , rr 12 repre-
senting the cone’s ruling is that line (or segment). Further discussion of the rulings,
interior lines, and points of a cone are left to as Exercises 10 and 11. By the way,
since rr 12 , rr 23 cannot be vertical, as for cylinders the hyperbolas, representing
cones differ from those representing cylinders (Exercise 13).

Table 9.1. Cones are represented by conic curves.

Number of tangent planes parallel to u Represented by a pair of

0 ellipses
1 parabolas
2 hyperbolas
 9.4 ♣ FT-3 Developable Surfaces 337

Figure 9.26. A circular (double) cone without tangent planes parallel to the line u.
The cone is represented by two ellipses. Two points, one on each ellipse, represent one of the tangent
planes.

Figure 9.27. Another (double) cone without tangent planes parallel to u.

It is represented by two ellipses, one collapsing to a line segment. This is the representation of the
double cone, though here and subsequently only one cone is shown.

** The representation of cones as a function of the orientation and apex angle.

As for cylinders, we study a right circular cone CC with vertex at a from its equation
obtained from the observation [128, p. 32] that any ruling has the same angle θ ,
with the symmetry axis having direction the unit vector ŝ = s1 ê1 + s2 ê2 + s3 ê3 .
For x ∈ CC and k = cos θ, the angle between x − a (a ruling on CC) and ŝ is θ or
π − θ, letting

CC : |(x − a) · ŝ| = k|(x − a)|.

338 9. Hypersurfaces in RN

Figure 9.28. A circular (double) cone with one tangent plane parallel to the line u.
The cone is represented by two parabolas. Two points, one on each parabola, represent one of the
tangent planes.

Figure 9.29. Another circular (double) cone with one tangent plane parallel to the line u.
The cone is represented by two parabolas, one collapsing to a half-line.

It is convenient to square both sides and write the cone’s equation as

CC : [(x − a) · ŝ]2 = k 2 (x − a) · (x − a). (9.66)

For the sake of simplicity, the cone’s vertex is placed at the origin (0, 0, 0), for the
translation to a does not change the type of representing conic curve. We find the
 9.4 ♣ FT-3 Developable Surfaces 339

Figure 9.30. A circular cone with two tangent planes parallel to the line u.
The cone is represented by two hyperbolas. The two points, one on each hyperbola, represent one of
the tangent planes.

intersection with the first superplane π0s

, in fact, the projection on the x2 x3 plane,
namely

0
(CC ∩ π0s
)23 = c23 : (−5k 2 +D 2 )x22 + (−2k 2 +C 2 )x32 + (2k 2 − CD)x2 x3 = 0,
(9.67)

where C = (s1 − s3 ), D = (2s1 + s2 ). We recall that the signs of the indicator I

and discriminant  determine the type of conic curve. They are

I = −3k 2 + (s1 + s2 + s3 )2 ,  = −6k 4 + k 2 [(2C − D)2 + C 2 + D 2 ]

= −6k 4 + k 2 [3(s1 − s3 )2 + 2],

3 
using 2
i=1 si = 1, j =k sj sk = 0 in the last step:

*
|s1 + s2 + s3 | (3(s1 − s3 )2 + 2
I =0⇒k=± √ , =0⇒k=± . (9.68)
3 6

Elliptic representing curves, Figs. 9.26, 9.27, are produced when I < 0, 0 < 
+ there are no tangent planes parallel to u.√ The vertex angle constraints are
and
[3(s1 − s3 )2 + 2]/6 > |k| > |s1 + s2 + s3 |/ 3.
There exist an orientation and a cone, specified by k, simultaneously for
I = 0 =  for
340 9. Hypersurfaces in RN

2(s1 + s2 + s3 )2 = (2s3 + s2 )2 + (s1 − s3 )2 + (2s1 + s2 )2 ⇒ s1 = s3 = s,

1
|k| = √ , (9.69)
3
the value of k obtained+from the second condition (for  = 0) in (9.68). From
3
i=1 si = 1, s2 = 1 − 2s 2 , and the first condition in (9.68), s = s1 =
2
√
s3 = 0, s2 = 1 or s = s1 = s√ 3 = 2/3, s2 = ±1/ 3 are the orientations when
the double cone with k = 1/ 3 has one tangent plane parallel to the line u,
yielding a pair of parabolic representing curves, Figs. 9.28, 9.29. When  > 0 and
I = 0, the representing conics are also parabolas, + and an easy calculation yields
the orientation
√ constraint s1 =
 s 3 . This implies that [2(s1 − s3 )2 + 2]/6 > |k| or
|k| ∈ (1/3, 5/6).
When  < 0, the curve given by (9.67) is an ellipse, and irrespective of
the +
sign of I , its image is
√ a hyperbola, Figs. 9.30 and √ 9.31.
√ From  < 0, we
get 3(s1 − s3 ) + 2 < 6|k|, and in turn |k| ∈ (1/ 3, 5/6) without further
2
restrictions on the orientation (i.e., the si ). Also of interest here is the combination
 = 0, the conic CC ∩ π0s
being a parabola and I = 0, so that s1 = s3 , with
√
|k| > 1/ 3, is the image curve is a hyperbola with one vertical asymptote.
+ For the combination  > 0, I > 0, adding the inequalities yields
[3(s1 − s3 )2 + 4]/3 < |k|; the corresponding cones have two tangent planes
parallel to u, producing a pair of hyperbolas, Figs. 9.30, 9.31, as representing

Figure 9.31. Another circular cone with two tangent planes parallel to the line u.
The cone is represented by two hyperbolas, one of them collapsing to a line.
 9.4 ♣ FT-3 Developable Surfaces 341

Table 9.2. Summary of representation results for cones.

Pair of curves representing Cone Orientation & vertex angle θ = cos−1 k

Ellipses – Figs. 9.26, 9.27 I < 0,  > 0, s1 = s3

+ √
[3(s1 − s3 )2 + 2]/6 > |k| > |s1 + s2 + s3 |/ 3
√
Parabolas – Figs. 9.28, 9.29 I = 0,  = 0, |k| = ±1/ 3
s1 = s3 = 0, s2 = 1,
√
s1 = s3 = 2/3, s2 = ±1/ 3.

Parabolas opening vertically I = 0,  > 0, s1 = s3

+ √
[3(s1 − s3 )2 + 2]/6 > |k| = |s1 + s2 + s3 |/ 3
+
Hyperbolas – Figs. 9.30, 9.31 all I,  < 0, [3(s1 − s3 )2 + 2]/6 < |k|

Hyperbolas I > 0,  > 0

+
[3(s1 − s3 )2 + 4]/3 < |k|
+
Hyperbolas with vertical asymptote I = 0,  = 0, [3(s1 − s3 )2 + 2]/6 = |k|
|k| = |s1 + s2 + s3 |

curves. This is also the case for  = 0, I = 0, constraining the vertex angle by
+
3(s1 − s3 )2 + 2]/6 = |k| and |k| = |s1 + s2 + s3 |; here each hyperbola has
one vertical asymptote. These conditions, which are combinations of the cone’s
size, measured by the vertex angle, and the orientation, producing 0, 1, or 2 tangent
planes parallel to û, are summarized in Table 9.2. A cone in 5-D whose representing
curves are ellipses is shown in Fig. 11.13 in Section 11.3, wherein an innovative
approach to surface representation as applied to quadrics is presented. A general
cone can be described by

x(s, v) = a + vt(s), (9.70)

where t(s) is a unit vector field described on a curve, such as a circle for circular
cones, all rulings passing through the vertex at a and a point of the curve t(s) [166,
pp. 90–91].
342 9. Hypersurfaces in RN

9.4.4 ♣ FT-8 General Developable Surfaces

In general, a developable surface can be described by

x(s, v) = y(s) + vg(s), |g(s)| = 1, (y

+ vg
)×g = 0. (9.71)

The equation describes the more general ruled surfaces, and it is the additional
condition on y
, v, g, g
that specializes it to developables. An example of an
intricate developable is the helicoid

x1 = a cos s − av sin s, x2 = a sin s + av cos s, x3 = b(s + v). (9.72)

The representing curves at two orientations, shown in Figs. 9.32, 9.33, are instances
of interactive exploration, reminding us of interactivity’s importance in visualiza-
tion. Proving that the helicoid is developable and finding the curves representing
its rulings (Exercise 9) are left to the reader. The class of developables consists of
planes, cylinders, cones, tangent surfaces of a curve, and combinations thereof.
Quite a few, e.g., helicoid, polar, tangential, and rectifying developables, which
are examples of tangent surfaces of a curve [166, pp. 66–72], are difficult to recog-
nize visually as developables. Yet in -coords they are immediately recognizable,
being represented by curves, and some properties are apparent from the duality
developables ↔ space curves, i.e., cusp ↔ inflection-poin , Fig. 9.23; bitangent
plane ↔ crossing-point, Fig. 9.24. It is significant that for developables in the
above examples, the two representing curves are similar. These curves are images

Figure 9.32. Developable helicoid and its two representing curves.

The two points on the right (one on each curve) represent the tangent plane shown on the left
determined by the two line intersections with the first π0s
and second π1s
superplanes.
 9.4 ♣ FT-3 Developable Surfaces 343

Figure 9.33. At this orientation, the representating curves have several ideal points.
They correspond to the helicoid’s tangent planes parallel to u (making equal angles with all coordinate
axes).

of the surfaces’ section with the first two superplanes, so they are projectively
equivalent (i.e., one can be obtained from the other by a projective transformation)
for the cylinders and cones. For the helicoid, whose rulings are a pencil of lines
on its axis, the reason is not so obvious. Still for other developables this may
not be true. It would be nice to settle the question whether two very different
curves such as those in Fig. 9.10 represent a developable (see Exercise 16). So for
developables with similar representing curves, the transition to RN is seamless:
they are represented by N − 1 similar planar curves, properties are revealed by
duality, and the reconstruction is analogous, i.e., three linked points on three curves
determine each tangent hyperplane for N = 4 as discussed in the next section.
In our surroundings there are many objects that look “nearly” developable,
and many parts are manufactured by deforming flat sheets. This and the simple
-coordinate representation motivate the study of “approximate developables”
(Exercise 18), as for the proximate planes in Chapter 8, by introducing pertur-
bations in the developables’ equation, allowing |(y
+ vg
)×g| < c = 0 but small
(see also the following section on ruled surfaces), and there may be other useful
definitions. The conjecture is that families of “nearly developables” are represented
in -coords by curved “strips” with properties “close” to those discovered by duality
(of developables) with representing curves and are amenable to an intuitive treat-
ment. New classes of surfaces (patches) and their N-D generalizations suitable for
applications such as geometric modeling may emerge.
344 9. Hypersurfaces in RN

9.4.5 ** Higher Dimensions

A developable hypersurface σ in higher dimensions is the envelope of a one-
parameter family of hyperplanes. Each of these hyperplanes is one of its tangent
hyperplanes. All the results in this section are generalizable to developable hyper-
surfaces in dimensions N > 3. These results are stated without proof below for
readers who would like to turn these theories into practical implementations.
Denote the N − 1 representing curves by σ̄j , where j ∈ I is one of the indexing
sets {1, 2, . . . , N}, {2, 3, . . . , N, 1
}, . . . , {N − 1, N, 1
, 2
, . . . , (N − 2)
}. If the
σ̄j ’s are monotonic with respect to the y axis, then σ can be reconstructed as
follows:

1. Draw a horizontal line to intersect the σ̄j ’s at N − 1 points π̄(t0 )j that together
specify a tangent hyperplane π(t0 ).
2. Draw N − 1 tangent lines lj at each π̄(t0 )j to each σ̄j . Each of these lines
specifies the intersecting point of the “ruling” (an (N − 2)-flat) of σ at π(t0 )
with the superplane πjs of the corresponding index.
3. Find the N intercepts of each of the N − 1 lj with the corresponding axes to
obtain the coordinates of these points of intersection.
4. Draw horizontal lines to bring all numbers to the first set of axes
{X̄1 , X̄2 , . . . , X̄N }.
5. Follow the construction procedure for 3-D to obtain the (N − 2)-dimensional
ruling.
In general, when the representing curves are not monotonic with respect to
the y axis, the curves can be partitioned into k traceable pieces at self-tangential
points. The representation is unique if one (N − 1)-tuple of points of equal
y-value is given for matching one set of pieces, for a total of k such tuples. The
reconstruction algorithm still works most of the time by producing a maximum of
O(k N−1 ) possible alternative hypersurface even if the matching tuples are not given
a priori. It fails only when the representing curves have an infinite number of self-
tangencies that break the curves into similar pieces, an infinite number of which have
the same number of extremal points and the same extremal y-coordinates occurring
in exactly the same sequence. Curve dualities have analogues for developables in
R3 , as illustrated in Figs. 9.23, 9.24 and in higher dimensions.

Exercises

1. Derive the representation and rulings’ curves for quadric cylinders (i.e.,
parabolic and hyperbolic). This is an easy course project.
 9.4 ♣ FT-3 Developable Surfaces 345

2. For two hyperbolas representing a circular cylinder, prove that for any point on
the vertical line rr 23 , the tangents to the two hyperbolic branches (above the
point) have the same y-coordinate.
3. Prove that if one of the representing curves of a developable has a crossing
point, so does the other representing curve. Do crossing points always coincide
as in Fig. 9.24?
4. For the representing hyperbolas in Fig. 9.20 construct the vertical intervals:
(a) rr 1
2 consisting of the representing points r̄1
2 and
(b) rr 13 consisting of the representing points r̄13 .
Do the same for the representing hyperbolas in Fig. 9.21.
5. Given a circular cylinder CG and a point P , write an algorithm to decide
whether P is interior to CG.
6. Generalize Corollary 9.4.6 to 4-D and higher dimensions.
7. What kind of curves can the representing curves of a cone’s rulings be?
8. Show that (9.71) defines a developable surface. Hint: find the normal vector
and show that it depends on one parameter.
9. Show that the helicoid given by (9.72) is developable and find its representing
curves. Hint: check that (y
+ vg
)×g = 0 as in (9.71).
10. Write and implement an algorithm that inputs a cone’s representation and
outputs the representing curves rr 12 , rr 23 of its rulings.
11. Give an algorithm that constructs half-lines and points interior to a cone.
12. Find the curves representing the rulings of the helicoid given by (9.72).
13. How do the hyperbolas representing cones differ from those representing
cylinders?
14. Generalize the representation of circular cones to N-D (see also Section 11.3).
In particular, are there cases in which different conic curves appear simultane-
ously in the representation of a cone in N-D?
15. Prove that the representing curve σ0
of a developable σ is composed of the
envelopes of the family of lines R̄ on the points r̄12 , r̄13 , r̄23 of the rulings r of
the developable. Similarly, the curve σ1
is the envelope of the family of lines
R̄
on r̄1
2 , r̄1
3 , r̄23 .
16. Prove or give a counterexample that the two representing curves of a develop-
able in R3 are projectively equivalent (i.e., one can be obtained from the other
via a planar projective transformation) and generalize to RN ; this is a research
problem.
17. Dualities for developables in RN (see Figs. 9.23, 9.24). Prove that
(a) a line of cusps is represented by N inflection points having the same y
coordinate and
346 9. Hypersurfaces in RN

Figure 9.34. Generation of a ruled surface.

Notice the rulings successively “twisting” about the base curve C.

(b) a bitangent plane is represented by crossing points. Find the number and
positions of the crossing points.
18. By allowing “small variations” in the equation of developables (like the
proximate planes in Chapter 8), define and study families of “approximate
developables.” This is a serious research problem.

9.5 ♣ FT-9 Ruled Surfaces

Developable surfaces are a subset of the more general class of ruled surfaces R
that are generated by a one-parameter family of lines. A ruled surface, also called
a scroll, can be created by the motion of a line called a generating line, generator,
or ruling, as shown in Fig. 9.34. Let C : y = y(s) be a curve and g(s) a unit vector
in the direction of a ruling passing through a point y(s) of C. A point P : x(s, v)
on the surface is then given by

x(s, v) = y(s) + vg(s) (y

+ vg
)×g = 0 ∀(s, v). (9.73)

The curve C : y = y(s) is called the base curve or directrix. The rulings are the
curves s = constant. When y is a constant, the surface is a cone, and when g(s) is
a constant, the surface is a cylinder. The difference here is that y
+ vg
× g = 0,
allowing ruled surfaces to “twist,” unlike developables. The tangent plane πP at a
 9.5 ♣ FT-9 Ruled Surfaces 347

Figure 9.35. The saddle (left) is a doubly ruled surface.

One of the two regions representing it on the left. Note the conic (parabolic) boundary.

point P on a ruled surface ρ contains the whole ruling rP on P . For another point
Q ∈ ρ, the tangent plane πQ still contains rp . Whereas all points on a ruling of a
developable have the same tangent plane, moving along points on a ruling r of a ruled
surface causes the corresponding tangent planes to rotate about r, this being the
essential difference and the basis for the -coords representation of ruled surfaces.
There is a wonderful collection of ruled and many other kinds of surfaces in
[130]. Ruled surfaces used as architectural elements are not only beautiful but also
have great structural strength. A surface can be doubly ruled in the sense that any
one of its points is on two lines completely contained in the surface (i.e., it can be
generated by either one of two sets of moving lines). An elegant result [18, p. 42]
is that in R3 , the only doubly ruled surfaces are quadrics: the hyperboloid of one
sheet and saddle (hyperbolic hyperboloid).
It has been noted, in Exercise 15 in the previous section, that the representing
curve σ̄0
of a developable σ can be obtained as the envelope of the family of lines R̄0

on the points r̄12 , r̄13 , r̄23 for each ruling r of σ . Similarly, σ̄1
is the envelope of the
lines R̄1
on r̄1
2 , r̄1
3 , r̄23 . A matched pair of points π̄0
∈ σ̄0
, π̄1
∈ σ̄1
represents
the single plane containing r ⊂ σ and tangent to every one of its points. By contrast,
a plane tangent to a ruled surface ρ at a point contains a full ruling r, but as the
point of tangency is translated continuously along r, the tangent plane continuously
rotates about r. In -coords the points π̄0
, π̄1 continuously translate in tandem
(with the same y coordinate) along the corresponding lines R̄0
, R̄1
representing
the rotating tangent planes along r (see Fig. 5.6). An example is the saddle, Fig. 9.35
(left). By Corollary 9.3.5, it is represented by two regions whose boundaries are
conic curves. The first region and the lines R̄0
tangent to the parabolic boundary
(their envelope) are shown on the right.
348 9. Hypersurfaces in RN

Theorem 9.5.1 (Representation of Ruled Surfaces, A) A ruled surface ρ is

represented by the regions ρ̄j
, j = 0, 1, containing the families of lines Rj = {R̄j }
whose envelopes are the boundaries ∂ ρ̄j
specifie by Lemma 9.3.1. ♣ FT-9e
To emphasize, a developable is represented by two curves that are the envelopes
of the families of lines Rj
(formed from the representation of its rulings as described
above). A ruled surface ρ is represented by two regions ρ̄j
, j = 0, 1, whose
boundaries ∂ ρ̄j
are also the envelopes of the lines Rj
(obtained in the same way
from the representation of its rulings), together with the lines Rj
. Equivalently,
it is helpful to consider ∂ ρ̄j
as a line-curve and ρ̄j
the region covered by its
tangents. The region’s structure enables the construction of the matching algorithm
for choosing pairs of points representing the surface’s tangent planes. By the way,
the boundary curves deriving from the rulings of a developable differ from those
obtained from the rulings of a ruled surface. These points are clarified with the
examples starting with the representation of the SA,

SA : x12 − x22 = x3 . (9.74)

We know that ∂(SA)0

= (SA ∩ π0s
)12 , and also that one projection of the space
curve SA ∩ π0s
suffices. As indicated by the subscripts 12, here the projection on
the x1 x2 principal 2-plane is used. From π0s
: x1 − 2x2 + x3 = 0, x3 = 2x2 − x1
and upon substitution in (9.74) the desired curve is obtained:

(SA ∩ π0s
)12 : x12 − x22 + x1 − 2x2 = 0,

a hyperbola whose representation in -coords (see (9.20) , (9.21), (9.22)) is as

already mentioned, Fig. 9.35 (right), a parabola given by

(SA)0
: −x 2 − 4y 2 − 4xy + 4x − 4y − 1 = 0. (9.75)

The transformation of the hyperbola (9.5) to the parabola in (9.75) is nicely found,
as shown in Fig. 9.36 (left), using a curve plotter.46 The second region is found by
repeating the procedure for π1s
: x1 + x2 − 2x3 = 0, x2 = 2x3 − x1 , yielding

(SA ∩ π1s
)1
3 : 2x12 − 4x32 + 4x1 x3 − x3 = 0.

This is a hyperbola, shown in Fig. 9.36 (right), whose transform in -coords is the
hyperbola

(SA)1
: x 2 + 48y 2 − 16xy + 2x − 8y + 1 = 0, (9.76)
46
Thanks to E. Rodity and V. Gurvich, who wrote the curve plotter for their course project
 9.5 ♣ FT-9 Ruled Surfaces 349

Figure 9.36. The intersection of the saddle SA and the first superplane π0s
is a hyperbola.
(Left) The hyperbola transforms to a parabola, the boundary of the region (SA)0
. (Right) The
intersection of the saddle SA and the second superplane π1s
is a hyperbola that transforms to another
hyperbola, the boundary of the region (SA)1
.

seen on the right. Here, as explained for (9.27), the (directed) interaxis distance
d = −1 from the X̄1
to the X̄3 axis is used. These two regions are combined to
provide the complete representation of SA. Prior to that, we rotate the x1 , x2 axes
about the x3 axis by −45◦ , simplifying the saddle’s equation to

SR : 2x1 x2 = x3 (9.77)

and changing the planes tangent to the line u, which determine the presence of ideal
points on the boundary curves, yielding the saddle’s representing regions: ∂(SR)0

is a hyperbola and ∂(SR)1

an ellipse seen on the right of Fig. 9.38 after the −45◦
rotation.
To keep the subsequent discussion intuitive and reasonably precise without
getting bogged down in technical details, we state without proof that the regions
encountered have boundaries consisting of a finite number of simple (i.e., non-self-
crossing) curves, each of which can be partitioned into a finite number of convex
(upward or downward) curves. Jordan’s theorem states that a simple closed curve
separates the plane R2 into two regions: the interior and exterior. For our needs,
we generalize this result to P2 by allowing a finite number of ideal points on the
boundary which we consider “projectively” as a simple closed curve. Of course, the
presence of ideal points needs to be treated judiciously. Relative to a region , a
point is exterior if there is a tangent from it to the boundary. Otherwise, it is interior.
For ∂ simple and closed, there is a tangent in each direction, except those of the
ideal points it contains. Hence, other than these, all other ideal points are exterior
to . Two points are on opposite sides of ∂, one interior and the other exterior,
if the line on the points crosses the boundary an odd number of times. To avoid
350 9. Hypersurfaces in RN

Figure 9.37. The saddle’s representation is the complement of the shaded regions.
The saddle SA is represented by the complements of the two shaded regions ((SA)0
and (SA)1
)
having parabolic and hyperbolic boundaries respectively. The points π̄0
, π̄1
representing a tangent
plane and the ruling r ⊂ π are constructed with the matching algorithm. A different ruling on π can
be constructed suggesting that SA is doubly ruled.

complicating this rule, the line is not allowed to cross at points where two portions
(one convex upward and the other downward) of ∂ are joined. For ∂ a convex
curve (not closed), it is helpful to consider it as a line-curve to recognize that its
exterior is still the region covered by its tangents. It is clear from Theorem 9.5.1
that for each point π̄j
∈ ∂ ρ̄j
, the region ρ̄j
must contain at least a segment of the
tangent at π̄j
, and hence all its points are on ∂ ρ̄j
or its exterior.
Corollary 9.5.2 (Representation of ruled surfaces; B) A ruled surface ρ ⊂ R3
is represented by ∂ ρ̄j
, j = 0, 1, and the regions exterior to it.
Basically, this is a restatement of Theorem 9.5.1. Keep in mind that no points
π̄j
(representing tangent planes of ρ) can be in the interior of the regions whose
boundaries are ∂ ρ̄j
. Returning to SA, on the xy plane its representing regions
(SA)0
, (SA)1
consist of the boundaries and the exterior of the convex portions
shown in Fig. 9.37 covered by their tangent lines. The two regions are superimposed
 9.5 ♣ FT-9 Ruled Surfaces 351

Figure 9.38. Conics representing the saddle.

Representation of the saddle showing an elliptical boundary and the tips of the two hyperbolic
branches. For ruled surfaces, the regions are covered by the line tangents to the two boundaries. The
two linked points represent a tangent plane.

in the sense that two indexed points, each belonging to a different region as dis-
tinguished from the index 0
or 1
, may be in the same position (x, y). As for
developables, the boundaries ∂(SA)0
, ∂(SA)1
are the envelopes of the family
of lines derived from their rulings. Here the two boundary curves do not have
the same y-range, something that distinguishes them from representing curves of
developables. For the representation of a ruled surface ρ, the two regions ρ̄j
need
to be linked with a “matching algorithm” enabling the construction of valid pairs
of points representing tangent planes. The algorithm is described in terms of the
representation of SA in Fig. 9.37. On a line y = c, a point π̄0
∈ ρ̄0
is chosen
and lines R̄, S̄ from it tangent to each of the ∂ ρ̄j
are constructed. In general,
there are several tangents from a point to each region, though here there are two
tangents to each conic region. Choose any other point π̄1
∈ ρ̄1
on the line y = c,
fixing the value of the first coefficient in π’s equation, also with two tangents
R̄
, S¯
to ∂ ρ̄j
on it. Then π, represented by π̄0
, π̄1
, is a tangent plane of SA,
since the tangent lines generate the regions’ boundaries (envelope). A ruling r on
π is found by r̄23 = R̄ ∩ R̄
and another by s̄23 = S̄ ∩ S¯
. This is significant,
for as previously mentioned, SA is doubly ruled. However, there are several such
intersections, suggesting that for developables, a convention is needed for the unique
reconstruction such as the possible generalization of the start points convention for
developables. Given ∂ ρ̄j
, a line y = c and a link-function g(x0 , x1 ) is provided,
ensuring that for all values of xj where the points are π̄j
= (xj , c) are exterior to
352 9. Hypersurfaces in RN

or on ∂ ρ̄j
, the pairs π̄0
, π̄1
represent tangent planes of ρ. Referring to Fig. 9.37,
an additional rule h(x0 , x1 ) is required to match the tangents such as R̄ on π̄0
with
a tangent R̄
on π̄1
. Then reconstruction of ρ’s tangent planes proceeds from the
pairs π̄0
∈ R̄, π̄1
∈ R̄
found by the vertical translation of the line y = constant
within the range of allowable values of y. The choice of the start line y = c is
important for its intersections with the ∂ ρ̄j
; the points π̄j
must be exterior to or on
boundaries. Filling out the necessary details, providing a proof or refutation with
counterexamples or alternative and better conditions, is left for Exercises 10 and 9.
The vertical translation of the line y = constant, while the points π̄j
, j = 0, 1,
are constrained to slide along their tangents R̄, R̄
, corresponds to the rotation of
the tangent plane about the ruling r. The tangent to ∂ ρ̄j
from π̄j
represents a point
πjs
= ρ ∩ π for either j = 0 or 1 so it is a point of tangency of π with ρ. The
construction for finding the point of tangency of π with ρ when the π̄j
are elsewhere
on R̄ or R̄
is deferred to the reader (Exercise 11). Avoiding tedious technicalities
at the expense of some rigor, we outlined above the steps required to obtain the
representation of a ruled surface ρ, the regions ρ̄j
, and the matching algorithm
linking them. The “straight-lines pattern” within the regions ρ̄j
, characteristic of
ruled surfaces, is easily discovered and explored interactively, as shown in Fig. 9.38,
by rotating and translating the surface and viewing its resulting representation.
A few words about the computation are in order. A rectangular grid is constructed
in the x1 x2 plane (or whatever parameters occur in the surface’s description) and
the tangent planes are computed at the grid points. The grid point density is also
interactively adjusted to reduce clutter.
Conoids are ruled surfaces whose rulings are parallel to a plane and are per-
pendicular to an axis normal to the plane such as those shown in Fig. 9.39.47 Spiral
staircases are shaped after the right conoids. Another specimen is the parabolic
conoid seen in Fig. 9.40 (center), which, like the saddle, has a parabolic directrix.
It is given by
PC : x1 x32 = x2 , (9.78)
and is defined in the first and third quadrant of the x1 x2 plane. The boundaries of
its representing regions are the images of
∂(PC)0
= (PC ∩ π0s
)12 : x13 − 4x12 x2 + 4x1 x22 − x2 = 0,

∂(PC)1
= (PC ∩ π1s
)1
3 : x1 x32 − 2x3 + x1 = 0. (9.79)
These are cubic curves, and from Plücker’s formula, Section 7.6, we find that their
transforms are algebraic curves of degree 6. The first is a cardioid (left) in Fig. 9.40,
47
The figures were obtained using the beautiful software written by M. Spivak for her course project.
 9.5 ♣ FT-9 Ruled Surfaces 353

Figure 9.39. Conoids.

The rulings of Whitney’s umbrella (left) and Plüker’s conoid (right) are perpendicular to a line, which
here is coincident with the vertical axis. They are both conoids.

Figure 9.40. (Center) The parabolic conoid PC is a ruled algebraic surface of degree 3.
The first representing region (PC)0
, a cardioid, is on the left, and the second (PC)1
on the right.
The cusps are the images of inflection points.

and the second is on the right. Both are symmetric with respect to the x axis, the cusps
being the transforms of inflection points. The region (PC)0
consists of the cardioid’s
exterior, the region covered by its tangents, and (PC)1
is the exterior of the six
convex curves forming its boundary. The superposition of the two representing
354 9. Hypersurfaces in RN

Figure 9.41. Representation of the parabolic conoid and one of its tangent planes.

regions is shown in Fig. 9.41, where some tangents on the left cusp are drawn
illustrating subregions where the exterior of some boundary curves overlap. The
regions are linked with the matching algorithm described above. On a line y = c,
a point π̄0
∈ (PC)0
is chosen having two tangents to the cardioid’s boundary. The
second point π̄1
∈ (PC)1
, also on y = c, may have several tangents to its region’s
boundaries, reminding us of the need, mentioned earlier, for a convention enabling
the unique reconstruction of a ruled surface from its representation (Exercise 10).
The representation picture is like a landscape map pointing out interesting features
and pathways inviting exploration and study of the surface. The goal is, as we
gain experience and understanding, the discovery of invariants corresponding to
distinct surface properties such as dualities for the developables. In Fig. 9.42 are
the representations for two orientations with the straight lines tangent to the regions’
boundaries.
A famous ruled surface is the Möbius strip, described by

1 1
x = y(θ) + vg(θ), − < v < , 0 ≤ θ ≤ 2π,
2 2
y(θ) = (cos θ )ê1 + (sin θ )ê2 , (9.80)
  
1 1 1
g(θ) = sin θ cos θ ê1 + sin θ sin θ ê2 + cos θ ê3 ,
2 2 2
 9.5 ♣ FT-9 Ruled Surfaces 355

Figure 9.42. Representations (left and center) of a parabolic conoid for two orientations.
Representation of the ruled cubic (Exercise 4) on the right. The straight lines reveal that the surfaces
are ruled and partially outline the regions’ boundaries. A tangent plane is represented by the two
linked points.

It is nonorientable in the sense that tracing the normal vector at a moving point
along a loop, specified by its directrix the circle y(θ) = cos θ ê1 + sin θ ê2 [128,
p. 170], the normal flips by 180◦ from its original direction when the traversal is
completed. In other words, this surface has only one side. The strip’s structure is
elucidated in Fig. 9.43 (right), showing that the ruling moving along the directrix
twists at an angle θ/2 from the vertical, where θ is the angle swept along the directrix.
The ruling intersects the central axis in three positions, inverting its orientation by
180◦ by the time it completes the circuit to create the twisted wonder we see in
Fig. 9.43 (left).
How does this surface appear in -coords? Dazzling patterns represent it at
various orientations, Fig. 9.44. Of particular interest is the search for a pattern
characterizing nonorientability. In Fig. 9.45 we see a closed circuit on the strip
and the corresponding pair of representing curves. This is repeated for a different
orientation, producing a pair of representing curves, that are easier to study, Fig. 9.46
(right). Each intersection point represents a tangent plane π for which π̄0
and π̄1

coincide, that is, π is perpendicular to the x2 x3 principal plane with its first coeffi-
cient c1 = 0. A nice way to understand the twisting is suggested in Fig. 9.46 (left),
where the strip is placed on a plane, in this case the x2 x3 plane, creating three “folds”
that do not have to be symmetrically positioned like those in the picture. At each of
the folds, the tangent plane is perpendicular to the x2 x3 plane, where by the “flip”
corollary, Corollary 5.3.2, the first coefficient c1 of the tangent plane’s equation
changes sign, say + → − → + → −, from the start to the completion of the circuit.
356 9. Hypersurfaces in RN

Figure 9.43. Möbius strip surface and its structure (left).

A ruling traversing the circular directrix intersects the central axis three times in one complete circuit.

Figure 9.44. The two cusps on the left (dual of inflection points, see Fig. 9.23) show that the Möbius
strip has an “inflection point” in three dimensions. The curves going towards infinity upwards and
downwards show that it is a closed surface. The cusps and loops may merge as shown on the right.
A tangent plane is represented by the two linked points.

At each of the three intersections of the representing curves in Fig. 9.46 (right),
c1 = 0. By the same argument, the remaining coefficients c2 , c3 vanish three times
 9.5 ♣ FT-9 Ruled Surfaces 357

Figure 9.45. Traversing one circuit on a Möbius strip.

Note the intersections between the pair of representing curves.

Figure 9.46. Visualizing nonorientability.

(Left) Möbius strip (thanks and acknowlegment to [130]) placed on one of the principal planes, say
x2 x3 . Note the three folds resulting in three tangents planes perpendicular to the x2 x3 plane. (Right)
Traversal of a closed circuit on the Möbius strip represented in -coords by two curves. There are
three crossing points, each corresponding to a tangent plane perpendicular to the x2 x3 plane.
358 9. Hypersurfaces in RN

and change their sign. Hence the tangent plane’s normal vector c = (c1 , c2 , c3 )
becomes −c and flips direction. If the triple intersection of the representing curves
is characteristic of nonorientability, what about the more numerous intersections in
the pair of curves in Fig. 9.45? It is in such situations that interactivity is essential.
Choosing a start position with a pair of linked points representing a tangent plane, the
curves are traversed, where we observe that the two points do not coincide at every
intersection point but only at three of them. With a bit of patience, the reader can
actually verify this from the picture, starting the traversal from the indicated position
with the two linked points and tracing their path, paying attention to the three ideal
points in each representing curve, returning to the start, observing that π̄0
and π̄1

coincide at only three of the intersection points. The three intersections together
with the ensuing flip are clearly a necessary and sufficient condition for nonori-
entability. For a closed circuit on a doubly twisted Möbius strip, the representing
curves π̄0
and π̄1
coincide at six intersection points corresponding to six folds, and
so on.
It appears, then, that a pattern characteristic of nonorientability has been found
whose verification relies on interactively tracing a closed circuit on the representing
curves. The picture, of course, is not proof, but it provides insights that may lead
to a proof; this is an objective of visualization. This points the way to researching
nonorientability in R4 and higher dimensions as well as to knot theory. Here is
another thought: twisting a curve about its tangent at a point creates an inflection
point. So is the cusp seen in Fig. 9.44 (right), which is the dual of an inflection
point, together with the surrounding pattern in the second curve, an inkling of an
nonorientable twist, like a “3-D inflection point” produced by twisting and then
joining a curved strip? Though imprecise, this depiction is in the spirit of visualiza-
tion, being intuitive and stimulating ideas. Does the birdlike pattern represent the
Möbius strip’s nonorientability? These observations and questions may stimulate
further research on the representation of nonorientability and ruled surfaces. The
representation results thus far can be summarized in the following theorem.
Theorem 9.5.3 (Representation of developable and ruled surfaces in RN ) A
surface σ ⊂ RN is represented by N − 1 regions σ̄ ⊂ P2 and is

1. ruled but not developable ⇔ all points of σ̄ are exterior with respect to the
boundary of at least one of its N − 1 regions,
2. developable ⇔ σ̄ consists of N − 1 curves, the boundaries of σ̄ , all having
the same y-range.

Prior to treating more general surfaces, an older inexact but very useful surface
representation is included, illustrated with an application to instrumentation and
 9.5 ** Approximate Quadric Hypersurfaces 359

process control. It is of specialized interest and independent, so that it may be

skipped without impairing other chapter topics.

Exercises
1. Prove, or give a counterexample, that a ruled surface that is not developable
cannot be convex.
2. Prove theorem 9.5.3 for RN .
3. Construct the -coords representation of the hyperboloid of one sheet, Fig. 9.3,
given by x12 + x22 − x32 = 1.
4. Construct the representation of the ruled cubic x3 = x12 x2 .
5. Construct the representation of Zindler’s conoid given by 2x1 x2 = x3 (x12 −x22 ).
6. Construct the representation of Plücker’s conoid given by x3 (x12 + x22 ) =
x1 2 − x2 2 ; Fig. 9.39 (right).
7. Construct the representation of Whitney’s umbrella given by x32 = x22 ; Fig. 9.39
(left).
8. Explore the representation of the parabolic conoid in Fig. 9.41. List and justify
your findings.
9. Determine from its representation ρ̄ whether the surface ρ is doubly ruled.
10. Provide a convention, like the “start points” for developables, that enables the
unique reconstruction of a ruled surface ρ from its representation; this is a
research problem.
11. Provide an algorithm for constructing the point of tangency of a plane π tangent
to a ruled surface ρ from the representations ρ̄ and π̄; this is a research problem.
12. Prove or provide a counterexample of the nonorientability pattern conjecture
for R3 and RN ; this is a research problem.

9.6 ** Approximate Quadric Hypersurfaces

The problem of finding interior points to hypersurfaces is important in applications
such as process control and decision support [81]. Here an algorithm for finding and
visualizing such points is given. In the interest of brevity, the algorithm is discussed
concurrently with the representation of some of the hypersurfaces to which it applies.
The point is constructed by successively defining its coordinates in some order. The
following steps are required:
1. Initially find the range of the variable whose value is going to be selected.
2. Select a value of the variable in this range. This reduces the dimensionality of
the point selection problem.
360 9. Hypersurfaces in RN

3. Find the range of the next variable for the reduced convex hypersurface and
select a value in that range.
4. Repeat the previous step until the last variable is selected.
The convexity and boundedness of the convex hypersurface guarantee that in each
case, the range is an interval. The success of the algorithm hinges on finding the
intervals efficiently.
As we will see, for bounded convex hypersurfaces the intervals are found by
taking the projections of the hypersurface’s sections. The boundary of these sections
is transformed into parallel coordinates, resulting in the stepwise display of the point
selection process. Such a display provides:
1. information on the local properties of the hypersurface and
2. proximity information for the selected point, i.e., a measure of distance of the
point from the boundary.
These points are illustrated subsequently when the algorithm is applied to some
specific kinds of hypersurfaces. The algorithm also applies to certain nonconvex
hypersurfaces where the available ranges are a finit union of intervals at every stage
of the point selection process.

9.6.1 The Hyperellipsoids

Let A be a positive definite symmetric matrix and consider the function
 T 
A eT X
F (X) = (X, 1) , (9.81)
e α 1

where X = (X1 , X2 , . . . , XN ) and e are row vectors, and α is a constant. The

equation

F (X) = 0

defines a hyperellipsoid with real points X, provided that

e(A−1 )T − α > 0.

A point is interior, on, or exterior to the surface according as F (X) is less than,
equal to, or greater than 0. Given a point X, it is straightforward to determine its
location relative to the hyperellipsoid. The object here is to find, component by
component, an interior point.
There are two kinds of processes required for this purpose. Fixing the value
of one of the variables produces a planar section of the hyperellipsoid that is
 9.6 ** Approximate Quadric Hypersurfaces 361

itself a hyperellipsoid of dimension one less than the original; this necessitates
the computation of a new array for the resulting reduced hyperellpsoid. The other
requirement is to calculate the available range of the remaining free variables by
taking the projection of the reduced hyperellipsoid on a coordinate plane. Now let

A1 B T
A= , e = (e1 , e2 ), (9.82)
B C

where A1 consists of the first 2 × 2 submatrix of A, C is A with the first two rows
and columns deleted, B and B T account for the remainder of A, with e1 consisting
of the first two elements of e and e2 , which is the remainder of e. It is a relatively
straightforward to prove the following.
Theorem (B. Dimsdale). Let X ∗ = (X1 , X2 ). The projection of F (X ∗ ) on the
X1 , X2 plane is the ellipse
 ∗T 
∗ ∗ Ā ēT X
f (X ) = 0 = (X , 1) (9.83)
ē a 1

and
⎫
Ā = A − B T C −1 B, ⎬ ⎪
ē = −1
e1 − e2 C B, (9.84)
⎪
⎭
ᾱ = α − e2 C −1 e2T .

Given the projection, the maximum ranges of X1 and X2 are easily calculated and
so is the available range for X2 , given X1 .
The hyperbola that is the transform of the ellipse into parallel coordinates is
essential for the display of the selection process. For one thing, the available range
of X1 is the interval on the X1 axis between the two branches of the hyperbola, and
similarly for X2 . Furthermore, the range of X2 , once X1 is selected, is exhibited by
drawing tangents to the hyperbola from the point on the X1 axis, the chosen value
of X1 . As the process is iterated through all the variables, the entire sequence of
selections and ranges is presented in a parallel coordinate display. Furthermore, the
hyperbolas themselves provide indications as to the sensitivity of various choices.
For an example of the display, see Fig. 9.47. In that picture the outer hyperbolic seg-
ments are the transforms of the original hyperellipsoid projections on the coordinate
planes. The inner segments come from the successive slices of the hyperellipsoid
and show the successive ranges. Note that a point is exterior to the hyperellipsoid
if and only if the intermediate envelopes, representing successive slices of the
hypersurface, are crossed by the polygons, is interior if it nowhere touches, and
is on the surface otherwise. The computational complexity is O(N 3 ) [42].
362 9. Hypersurfaces in RN

Figure 9.47. Interior point construction for a convex 20-dimensional hyperellipsoid.

This can be the model of a process depending on 20 parameters. An interior point is constructed and
represented by the polygonal line shown. The values of the 20 parameters specify a feasible state
of the system. At this state the critical parameters are x13 , x14 , x15 where the available ranges (seen
between the two inner curves) are narrowest, i.e., the point is closest to the boundary.

9.6.2 The Hyperellipse

The procedure for hyperellipsoids calls for repeated projections on coordinate
planes, at a considerable cost in computational complexity. The obvious sug-
gestion is to define an object whose projections do not change as the iteration
proceeds.
Consider a collection of hypercylinders whose projections on the coordinate
planes are ellipses. The intersection of such a set of hypercylinders, if it exists, is
clearly a convex object. With some attention to detail, such objects can be defined so
that they exist, are bounded, and have a property called consistency, which, briefly,
means that a valid value for any variable is consistent with a valid value for any
other variable. A notable property of this kind of surface is that the computational
complexity is O(N), since the surface is defined by N − 1 ellipses, whence N − 1
quadratic equations must be solved [42].
 9.6 ** Approximate Quadric Hypersurfaces 363

9.6.3 Intelligent Process Control

Representation of the Domain
The following points are worth noting at the outset of this section:
1. Controlling a process requires considerable human intelligence. A proposed
system, or any other, is called intelligent if it can replace—to a considerable
extent—the human operator for this task, or if an operator is present, the system
can render intelligent help by choosing the critical parameters, at any given
time, with explanations as to why they are critical. Further, there is the ability
to do what if scenarios and see what options there exist for getting out of “tight
spots.”
2. It is known that any dynamical system with N parameters can be considered
as a static system in N + 1 parameters; the original N parameters and time.
This applies also to processes.
An approximate but useful representation of a general convex hypersurface can be
obtained by computing the envelope of the collection of polygonal lines represent-
ing its points. For multidimensional polyhedra, such a representation is unique.
Otherwise, the description of the hypersurface needs to be retained in addition to its
parallel coordinate representation. Such a representation of a convex hypersurface in
R20 is shown in Fig. 9.47. It corresponds to a particular (nonlinear) relation, which
for our purpose describes a process involving 20 parameters labeled x1 , . . . , x20 .
An N-tuple is a feasible point for the process iff it is interior to the hypersurface. This
is the geometric equivalent of the statement that a specific combination of parameter
values satisfies all the constraints imposed on the process. When a mathematical
description for the process is not available, an approximation (a best fit according to
some criterion) for the envelope can be obtained from available data. Geometrically,
the relation between adjacent axes is indicated by the shape of the envelope, and here,
due to the hypersurface’s convexity, is a gh (a generalized hyperbola). To economize
on-display space, only the strip between the axes is shown. Further, the envelope’s
shape clearly depends on the particular order in which the variables appear (as axis
labels). As we will see, rather than a problem, this provides an opportunity for
choosing preferred orderings of the variables subject to certain criteria.

The Rule
Controlling the process is equivalent geometrically to staying in the interior of
the hypersurface. We apply the algorithm for doing so. The interval on the x1
axis between the upper and lower portions of the envelope indicates the range of
parameter x1 . When a value for x1 = a1 is chosen in the available range of x1 , the
364 9. Hypersurfaces in RN

number of variables in the relation is reduced by one (i.e., from 20 to 19 variables

in our example). Again with reference to Fig. 9.47, from the fixed value of x1 ,
tangents to the upper and lower portions of the envelope are drawn (here the points
of tangency are not seen, since they lie outside the strip between the x1 and x2
axes). The envelope of the 19-dimensional hypersurface is then obtained from the
(description) of the original hypersurface with the value x1 = a1 . Again for display-
space economy, only the strip between x2 and x3 of the new envelope is shown and
is always contained between the upper and lower portions of the original envelope.
The restricted available range of x2 , due to the constraint of fixing the value of one
of the variables, is the interval on the x2 axis between the upper and lower tangents.
Proceeding, a value x2 = a2 is chosen within the current available range of x2 ,
reducing the dimensionality by one (e.g., 18), and the whole process is repeated.
A polygonal line is shown that always lies between the intermediary envelopes.
It represents an interior point to the hypersurface, and all interior points can be
found in this way. So this provides us with a way of finding and displaying feasible
points for the process. Also the algorithm is very fast and suitable for real-time
operation.

The Display as an Instrument Panel

When the value of a variable is fixed, the complete envelope of the 19-dimensional
surface may be drawn, showing at a glance not only the current value of any of
the parameters (in this case x1 ) but, unlike a conventional “intrument panel,” the
current available range of all the parameters; this is due to the interrelationship
among them. There are other salient features of the display. Note that for the
parameters x13 , x14 , x15 the available ranges are the narrowest. This indicates that
the feasible point is closest to the boundary with respect to these critical parameters.
Hence, when there is need for rapid decisions (such as in piloting very fast aircraft),
significant information reduction can be achieved by controlling the critical param-
eters whose priority is decided by their proximity to the boundary, adjusting the
remaining variables subsequently. This also brings out the possibility of displaying
or highlighting only the information desired (possibly specific subsystems) at any
given time. When a point is not feasible, the display shows the corresponding
polygonal line crossing some of the intermediate envelopes. It is then evident which
parameters must be changed and by how much in order to bring the polygonal
line inside the envelope and the process in control. In short, this instrument panel
provides a variety of information that an operator would somehow need to know
(at least implicitly derive or guess based on experience) in order to control the
process.
 9.6 ♣ FT-10 More General Surfaces 365

Order, Learning, and Explanations

Ideally, one would like to have a display in which the relation between any pair
of variables, say xi and xj , is apparent. There are N(N − 1)/2 such pairs for
N variables, which suggests that O(N 2 ) displays with different orderings may be
needed. Fortunately, this is not the case. It turns out [184] that for N = 2n, exactly n
“well chosen” displays suffice, while for N = 2n + 1, n + 1 displays suffice. This is
due to the fact that a complete graph of N vertices can be obtained as a union of that
many “properly chosen” spanning trees. For example, for N = 6 the permutations
126354, 231465, 342516 circularly contain every possible pair (independent of the
order) of adjacent subscripts from 1 to 6. Concerning the ordering of the variables, an
advantageous ordering is according to sensitivity. In fact, it is possible to implement
a dynamic ordering in which the order at each step is decided according to some
rule, for example, that the variable displayed and fixed next is the one with the
smallest available range.
When an approximation of the hypersurface is obtained from data, the repre-
sentation can be altered in view of new information. When a new feasible point is
found that is not obtained via the algorithm from the approximate representation of
the hypersurface, using the approximation technique again would enlarge the hyper-
surface in some region. On the other hand, an incorrect feasible point obtained from
the approximation would yield a restriction of the hypersurface. So a continuously
more accurate representation of the process can be obtained as a result of successive
trials. Hence, the system has the ability to learn.
At any stage of the algorithm an explanation of why a specific value of a variable
results in a feasible or infeasible point is available in terms of whether the polygonal
line intersects the intermediate envelopes. Also, having arrived at a partial choice
of parameter values, it is possible to conduct what if scenarios—an important look-
ahead feature—before actually fixing the value of the remaining parameters. What
is proposed then is a new technique for process control, which also suggests a
generalization of the concept of expert system to include those whose knowledge
representation is a hypersurface together with a rule for “navigating” in the interior.

9.7 ♣ FT-10 More General Surfaces

The formulation in (9.15) applies to any sufficiently smooth surface. We open this
last section by examining the representation of the venerable sphere in R3 ,

SP : x12 + x22 + x32 = r 2 . (9.85)

366 9. Hypersurfaces in RN

Figure 9.48. Representation of a sphere.

Representation of a sphere centered at the origin (left) and after a translation along the x1 axis (right),
causing the two hyperbolas to rotate in opposite directions.

Its intersections with the first superplane π0s

, a circle C0, and its hyperbolic image
C0, are

C0 : 2x12 − 4x1 x2 + 5x12 − r 2 = 0,

C0 = ∂(SP)0
: 3r 2 x 2 − 6y 2 − 6r 2 x + 5r 2 = 0. (9.86)
√
The hyperbola’s vertices (i.e., min and max) are at the points (1, ± 2/3r). The
second boundary ∂(SP)1
is the same hyperbola translated to the right by one unit,
as shown on the left of Fig. 9.48. Due to its symmetry, a hypersphere in RN , like a
hypercube, Fig. 9.1, is represented by N − 1 identical patterns (hyperbolas) spaced
successively one unit apart. Translation of the sphere along the x1 axis rotates
∂(SP)0
clockwise and ∂(SP)1
counter-clockwise, Fig. 9.48 (right), an interesting
manifestation of the translation ↔ rotation duality we have encountered on previous
occasions. It is useful to compare this representation with the inexact one given in
the previous section. Consider the 5-D sphere

5SP; x12 + x22 + x13 + x42 + x52 = r 2 , (9.87)

whose projection on the principal x1 x2 plane is the circle x12 + x22 = r 2 repre-
sented by the hyperbola 2r 2 x 2 − y 2 − 2r 2 x + r 2 = 0 going through the points
 9.7 ♣ FT-10 More General Surfaces 367

Figure 9.49. Interior point construction of a sphere in R5 .

(x = 0, y = ±1) and (x = 1, y = ±1). The other projections are represented by

translations of this hyperbola successively by one unit. All four are superimposed
as in Fig. 9.49, showing that the exact representation, also with hyperbolic bound-
aries, is an “upgrade” though the hyperbolas differ. The picture helps in following
the construction of a point interior to 5SP. Let us start by choosing a value of
x1 = a1 ∈ [−r, r], placing this on the X̄1 axis, and drawing the tangents from
this point (0, a1 ) to the upper and lower branches of the first hyperbola. This is
equivalent to intersecting 5SP with the plane x1 = a1 , resulting in the 4-D sphere

4SP : x22 + x13 + x42 + x52 = r 2 − a12 , (9.88)

,
with radius r 2 − a12 , seen in the figure on the X̄2 axis at its intersection with
each of the two tangents. Also shown is the hyperbola
, representing
, the projection
of 4SP on the x2 x3 plane. A choice x2 = a2 ∈ [− r 2 − a12 , r 2 − a12 ] is made,
reducing the dimension, providing a 3-D sphere, and so on. As long as the successive
choices are in the allowable range (i.e., radius of reduced sphere), the polygonal
line joining the ai does not intersect the intermediate curves like the one shown,
and it represents an interior point of 5SP. If the polygonal line is tangent to any of
the intermediate curves, it represents a point on the surface, and if it intersects, then
it represents an exterior point. It is clear, then, that all interior and surface points of
a sphere SPN ⊂ RN are represented by the polygonal lines between the hyperbolas
that represent the projections on the principal planes. Therefore, these hyperbolas
368 9. Hypersurfaces in RN

√
slightly overlap the hyperbolas ∂(SPi
) that intersect the X̄i axes at y = ± 2/3r
(Exercise 4).
All the points representing tangent planes are in the interior √ of the upper and
lower branches (Exercise 5). Recall the unit vector û = 1/ 3(ê1 + ê2 + ê3 )
along the line u. Let πu be the plane through the origin with unit normal û. The
intersection πu ∩ SP = CU is a great circle, which plays an important role in the
representation of SP. All tangent planes of SP along CU have the sum of their
coefficients S = c1 + c2 + c3 = 0 vanishing, for N · û = 0, where N = (c1 , c2 , c3 ),
since being parallel to the line u, they are represented by ideal points, thus causing
the split into two hyperbolic branches. The representing points of the tangent planes
above CU are on the branch, and those below form the lower branch. Note that at the
north and south poles the tangent planes are πp : x3 = ±1 with representing points
π̄0
= (2, ±r, 1) on the X̄2 axis but not on the boundary SP0
, which is formed by
the representing points of tangent planes on π0s
∩ SP and does not include the poles.
Similarly, the π̄1
are on the X̄1
axis and the same y = ±r.
Whereas the interior points of the hyperbolic regions properly matched (Exer-
cise 6) represent the sphere’s tangent planes, there is another interesting matching
that gives a feel about the sphere’s interior. Consider the representation of concentric
spheres (centered at the origin) with radii r ranging between a minimum rm and
maximum rM . As r increases, the hyperbolas’ vertices and the rest of the curve
rise linearly with r, enlarging the region between the upper and lower branches.
Let a pair of points at y = ya be given on the hyperbolic boundaries representing
a tangent plane πa of one of the spheres. Its √ radius a is found from the vertices of
the hyperbolas with extreme values at y = ± 2/3a. Draw the vertical lines on the
pair of given points. Any other pair of points at y = yc represents a tangent plane
πc parallel to πa on the sphere with radius c = yc /ya a.
The intersection of the superplanes with ellipsoids and paraboloids are ellipses,
whence their representation by hyperbolic regions, Fig. 9.50. For the paraboloid, the
intersections may also be parabolas, in which case the representing regions would
be either hyperbolas, each with one vertical asymptote, or vertical parabolas; see
Section 7 (Exercise 7). In general, the intersection of the superplanes with bounded
and convex surface σ is a bounded convex set bc whose -coords representation is
a generalized hyperbola gh, and such an example is shown on the left in Fig. 9.51
(Exercise 8).
With the stipulation that all representing points are in the interior or boundary
points, consider a bounded convex hypersurface whose representation at a specific
orientation consists only of gh regions. By the translation ↔ rotation duality evident
in Fig. 9.48, the regions remain gh for all other orientations (i.e., translations and/or
rotations). The representation of an unbounded convex hypersurface also consists
 9.7 ♣ FT-10 More General Surfaces 369

Figure 9.50. Representation of an ellipsoid on the left and a paraboloid on the right.

Figure 9.51. Representations: non-quadric convex surface (left) and a surface with a small bump
(right).

of gh regions unless it has some regions (say p) in the direction of u, in which

case p of its representing regions are parabola-like gp (generalized parabolas)48
having one ideal point and N − p − 1 gh regions. Without entering into details, we
conjecture that the gp regions have ideal points in the vertical direction.

48
Discussed in the chapter on curves.
370 9. Hypersurfaces in RN

Figure 9.52. Representation of a nonconvex surface with two pointed depressions.

The depressions are mapped into pairs of “swirls.” Each planar section of the depression is a cusp
and by the cusp – inflection-poin duality (see Fig. 9.23 and 9.24) is mapped into two curves with
inflection-points, which together form the pair of swirling patterns. Both depressions are visible in
the representation though one is hidden in the perspective on the left.

Theorem 9.7.1 (Convexity) A bounded surface σ ⊂ RN is convex ⇔ for one

orientation it is represented by N − 1 gh regions whose points are interior or on
the boundary. If the representation has some gp and the rest gh regions, then the
hypersurface is unbounded and convex.
This is a powerful result, important in many applications, enabling recognition
of convexity in RN . Next we examine the representation of surfaces having small
deviations from convexity. In Fig. 9.51 on the right is a surface σb with a small
“bump.” Its orientation was interactively varied in every imaginable way, and signifi-
cantly, its representations, like the one shown, were all markedly different from a gh.
In certain orientations the intersections of σb with the superplanes are bc, bounded
and convex, yet the representation regions were not gh. This important observation
is explored further. No less significant is that the representing regions are not chaotic,
having interesting similar shapes varying with changing orientation (Exercise 9a).
Proceeding, a surface σd with two small “dimples” or depressions, considered here
as a perturbation of convexity “opposite” to that in σb , shown in Fig. 9.52, is explored
next. Interesting representation patterns are seen with varying orientations; one (on
the right) contains two hyperbolic regions and other patterns outside them, evidence
of “near convexity.” The two pairs of like “swirls” (center) correspond to the two
dimples (Exercise 9b). Again it is rather remarkable that for a multitude of orienta-
tions, none of the representations consists of just two hyperbolas, as may naively be
expected when the intersection with the superplanes is a bounded convex set bc.
 9.7 ♣ FT-10 More General Surfaces 371

Figure 9.53. Torus representation showing two bitangent planes (left) and hyperbolic slices (right).

Figure 9.54. Astroidal surface and its representations for two orientations.
There are two pairs of crossing points (one pair is overlapping in the middle) which are present for
all orientations.

Many hyperbolas appear in the representation of the torus, Fig. 9.53 (right), but
unlike the representing hyperbolas of convex surfaces, here the hyperbolas are not
contained within an outermost hyperbola but are spread in the horizontal direction.
Curves from the two representing regions intersect at two points with the same x
coordinate. This is reminiscent of the crossing points corresponding to bitangent
planes. We surmise that the crossing points represent the two parallel planes tangent
at the top and bottom of the torus (Exercise 10). The representation of the asteroidal
surface [130] in Fig. 9.54 has a pattern with three crossing points that persists in
several orientations. Upon careful scrutiny we see that only at the midpoint do
372 9. Hypersurfaces in RN

curves from the two regions intersect, resembling the crossing points in the torus’s
representation. Could this point represent a tangent plane approaching one of the
surface’s vertices and becoming a bitangent plane in the limit? Another association
that comes to mind is of an exterior supporting plane touching the surface at exactly
three vertices.
The significant finding from these examples is that the -coords representation
even of complex nonconvex surfaces in various orientations is not chaotic but con-
sists of discernible patterns associated with their properties. All this may lead to new
surface classifications (Exercise 11). The surface representation seems preferable
to the standard surface representation (in terms of projections) even for R3 ; surface
properties like convexity and features like dimples, bumps, coiling and more are
not hidden in various orientations. The reason is that in -coords representation, a
surface is not only described by its points but also by the normal vector at each
point (review Fig. 9.2).
Arising from this discussion is the question, what happened to the claim that
the representing boundaries are the images of the surface’s intersection with the
superplanes? Let us return to the proof of Lemma 9.3.1 and note that the argument
is based on the assumption that the curve is on one side of its tangent. Recall that in
Section 6.2, on envelopes, the analysis was also based on the family being on one
side of the envelope, leading to necessary and sufficient conditions for the existence
and a method for obtaining the envelope. As was seen in the subsequent examples,
the method works even when there are singularities and when this condition fails.
In such cases the envelope has several components, each consisting of special points
on the family of curves, e.g., inflection points, cusps, extrema, crossing points, and
others. For the problem now at hand, the boundary of a representing region is an
envelope of lines “touching” it (but not necessarily tangent to it) and the curves
formed by the representing points. We saw that the method of Lemma 9.3.1 applies
well to developable, convex, and even ruled nonconvex surfaces; the representing
regions lie on the interior or exterior respectively and on one side of the boundary.
We found that for nonconvex surfaces, parts of the representing regions lie on both
sides of the boundary, so the condition of Lemma 9.3.1 fails, and the boundary is
not necessarily only the image of the surfaces’ intersection with the superplanes.
As pointed out below, this has important ramifications and is basically good news.
The refinement of Lemma 9.3.1 for nonconvex surfaces is a difficult research prob-
lem, related to that for the envelopes of general families of curves and surfaces.
These observations, even if partially true, may lead to a new approach for the
construction of (also approximate) convex hulls in RN as for R2 [110].
♣ FT-10e
 9.7 ♣ FT-10 More General Surfaces 373

Figure 9.55. Hexagons representing a family  of proximate planes.

The coefficients’ range is c1 ∈ [1.5, 0.33], c2 ∈ [2.5, 0.33], c3 ∈ [1, 0.33], c0 = 1 in R3 .
The two points interior to the first two hexagons represent the plane with (see lower left)
c1 = 0.43, c2 = 0.49, c3 = 0.31, so it is not in the family  even though the third point is
interior to the third hexagon. The software used allows for the interactive variation of the coefficients
and positions of the representing points in the first two hexagons.

9.7.1 ** A More General Matching Algorithm

Recall the family of curves gh (generalized hyperbola), uc (unbounded convex),
and bc (bounded convex) in Section 7.7. For our purposes here, consider their
generalization as the class CC of surfaces in RN such that for a σ ∈ CC, in a given
orientation, the boundaries of the N − 1 regions of σ̄ are either gh, uc, or bc, with
the points representing their tangent planes interior to their respective regions. The
matching algorithm scheme in Section 8.3.8, on proximate flats, can be used also
for this class of surfaces. For this reason we recall the procedure, featuring another
student course project.49 As Fig. 9.55 shows, it does not suffice that π̄0
and π̄1

be in the interior (or boundary) of the first and second hexagons respectively for
the plane π ∈ . The hexagons are bc bounded convex sets, and so the collection
of planes in  cannot be convex; for then it would be represented by gh regions.

49
Acknowledgment and thanks to Y. Lifshits and I. Makarov, who wrote this wonderful program for their
course project in 2005: navigation within the hexagonal regions representing families of proximate
planes.
374 9. Hypersurfaces in RN

Figure 9.56. A plane in the family .

Here the two indicated points represent the plane with c1 = 0.458, c2 = 0.389, c3 = 0.542, which is
in the family . Three vertices are chosen representing a simplex in R3 and represented by triangles
within the hexagons. Two representing points within the triangles in the first and second hexagons
represent a plane in .

A simplex within  is chosen by specifying three vertices in the hexagons. For π̄0

and π̄1
in the triangles within the first and second hexagons, the pair represent a
valid plane π ∈ , as shown in Fig. 9.56. What is helpful here is that a triangle in
R2 is also represented by a triangle in -coords. So the triangular faces of the 3-D
simplex remain triangular in their representation. The process can be continued for
higher dimensions,50 allowing also variations in c0 ; Fig. So  ⊂ R3 is represented
by octagons with two vertical edges, Fig. 9.57 and  ⊂ R4 by decagons with two
vertical edges.
In view of what we have learned, we can surmise that the intersections of
the hyperplanes in a family  ⊂ RN of proximate hyperplanes form a gh if it is
represented by convex 2N-gons, a uc if it is represented by uc’s, and a polyhedron
if it is represented by gh’s. Care is required to state this precisely (Exercise 3).
The reason for bringing up this approach again is that it may apply to surfaces in
50
Acknowledgment and thanks to A. Feigenbaum, T. Kafri, and S. Kamhin for extending the Y. Lifshits
and I. Makarov software to R4 .
 9.7 ♣ FT-10 More General Surfaces 375

Figure 9.57. Octagons representing a family of proximate planes  ⊂ R3 .

The two vertical edges are due to the variation of c0 ; here c0 = [1, 2].

Figure 9.58. Decagons with two vertical edges representing proximate hyperplanes in R4 .

CC, whose representation consists of gh, uc, or bc regions with all the representing
points in their interior, and the regions can be partitioned into convex portions
(Exercise 12).
Tantalizing as it has been, we wind up for now our glimpse into the challenges,
beauty, and promise of surface representation.

Chapter acknowledgments. While writing this chapter I recognized even more

the crucial contributions of my colleagues and students. In the early 1980s, Bernie
Dimsdale, my long-time friend and colleague, and I started studying the repre-
sentation of surfaces summarized in Section 9.6. Chao-Kuei Hung, in the early
376 9. Hypersurfaces in RN

1990s, pioneered the exact representation of surfaces in terms of their tangent

planes, obtaining results on developables and initiating the study of ruled surfaces.
Many insights and beautiful figures were obtained with the wonderful interactive
software written by David Adjiashvili for his course project. Years earlier, Michael
Spivak was capable and brave enough to write the first interactive software on
surface representation. Liat Cohen’s clever notation for the axis spacing together
with many other ideas and suggestions were very helpful. There are others already
acknowledged and those who contributed their software in the attached CD. To all
I am profoundly grateful.

Exercises
1. (a) Prove that the hyperbolic boundaries of the sphere SP rotate in opposite
angular directions with a translation of SP along the x1 axis. Find the
exact relation between the translation by x1 and angle of rotation.
(b) What changes in the representation with translations along the remaining
axes?
(c) Generalize for the representation a sphere in RN translated along an axis
xi , i = 1, . . . , N.
2. With reference to Fig. 9.1, find the representing boundaries of the cube in R3
using Lemma 9.3.1 and show that they are gh or uc. Generalize this for the
hypercube in RN .
3. For a family of proximate hyperplanes  ⊂ RN describe precisely (with
proofs) the boundary of  when it is represented by N − 1 convex 2N-gons,
ub, or bc regions.
4. Provide an algorithm for constructing interior points of a sphere directly from
the hyperbolas of its exact representation; hard.
5. Prove that the points representing tangent planes of a sphere are each on the
boundary or interior to its corresponding hyperbolic region.
6. Provide the matching algorithm for the sphere’s representation; this is a
research problem.
7. Under what conditions is an elliptic paraboloid represented by vertical parabolic
regions or by hyperbolas each with one vertical asymptote?
8. Provide criteria for distinguishing

(a) the convex quadrics (cylinder, ellipsoid, paraboloid) from each other
by their hyperbolic representing regions. Hint: the section on conics in
Chapter 7 on curves;
(b) the orientation of the convex quadrics from their representation.
 9.7 ♣ FT-10 More General Surfaces 377

9. (a) Take a convex body such as a sphere in R3 , introduce a small bump on

its surface, and derive the representing shapes of the resulting surface σb
;
hard.
(b) Take a convex body such as a sphere in R3 , introduce a small dimple, and
derive the representing shapes of the resulting surface σd
; hard.
(c) Compare the representations of σb
and σd
. Provide a criterion for
distinguishing between the two kinds of nonconvexity.
10. Prove or disprove that the two tangent planes at the top and bottom of the torus
are represented by two points at the same x coordinate where curves from the
two representing regions cross; see Fig. 9.53, center.
11. Study the representation of the class of surfaces of revolution. Provide criteria
for recognizing a surface of revolution from its representation; this is a research
problem.
12. Apply the method in the last section to the representation of a sphere in R3 .
13. Prove or give a counter example of the dimple → swirl duality seen in Fig. 9.52.
This page intentionally left blank
10

♣ FT Data Mining and Other Applications

10.1 Introduction

The first, and still most popular, application of parallel coordinates is in exploratory
data analysis (EDA) to discover data subsets (relations) that fulfill certain objectives
and guide the formulation of hypotheses. A data set with M items has 2M subsets,
any one of which may be the one we really want. With a good data display, our
fantastic pattern-recognition ability cannot only cut great swaths in our search
through this combinatorial explosion, but also extract insights from the visual
patterns. These are the core reasons for data visualization. With parallel coordinates
(-coords), the search for relations in multivariate data sets is transformed into a
2-D pattern-recognition problem. Guidelines and strategies for knowledge dis-
covery are illustrated on several real data sets, one with hundreds of variables.
A geometric classification algorithm is presented and applied to complex datasets.
It has low computational complexity, providing the classification rule explicitly
and visually. The minimal set of variables required to state the rule is found
and ordered by their predictive value. Multivariate relations can be modeled as
hypersurfaces and used for decision support. A model of a country’s economy
reveals sensitivities, impact of constraints, trade-offs, and sectors unknowingly
competing for the same resources. This chapter can be read independently of the
rest, with foundational background provided where needed. Readers are encouraged
to browse through the book, looking into specific topics when the motivation arises.
Collision-avoidance algorithms for air traffic control are discussed separately in
Chapter 4, on multidimensional lines.

A. Inselberg, Parallel Coordinates, DOI 10.1007/978-0-387-68628-8_10,

© Springer Science+Business Media, LLC 2009
380 10. ♣ FT Data Mining and Other Applications

10.1.1 Origins
For the visualization of multivariate problems, numerous mappings encoding
multidimensional information visually into 2-D or 3-D (see [63] and [172], [173],
[174]) have been invented to augment our perception, which is limited by our
3-dimensional habitation. Wonderful successes such as Minard’s “Napoleon’s
March to Moscow,” Snow’s “dot map,” and others are ad hoc and exceptional.
Succinct multivariate relations are rarely apparent from static displays; interactivity
is essential. In turn, this raises the issues of effective GUI (graphic user interface),
queries, exploration strategies, and information-preserving displays, but we are
getting ahead of ourselves.

10.1.2 The Case for Visualization

Searching a data-set with M items for interesting, depending on the objectives,
properties is inherently hard. There are 2M possible subsets, any one of which may
satisfy the objectives. The visual cues that our eyes can pick from a good data display
help us to navigate through this combinatorial explosion. How this is done is part
of the story. Clearly, if the transformation data → picture clobbers information, a
great deal is lost right at the start. We postulate that a display of data sets with N
variables suitable for exploration satisfies the following requirements:
1. should preserve information – the dataset can be completely reconstructed from
the picture,
2. has low representational complexity – the computational cost of constructing
the display is low,
3. works for any N – not limited by the dimension,
4. treats every variable uniformly,
5. has invariance under projective transformations – the data set can be recognized
after rotations, translations, scalings, and perspective transformations,
6. reveals multivariate relations in the dataset – the most important and
controversial single criterion,
7. is based on a rigorous mathematical and algorithmic methodology – to
eliminate ambiguity in the results.
No completeness or uniqueness is claimed for this list, which should invite
criticism and changes. Commentary for each item, illustrated by examples and
comparisons, is listed in the same order.
1. The numerical values for each N -tuple (i.e., each data point) are recoverable
from the scatterplot matrix (SM) and the -coords display. This may not be
necessary or desirable for presentation graphics (e.g., pie charts, histograms).
 10.1 Introduction 381

2. In the pairwise scatterplot matrix, the N variables appear N(N − 1)/2 times.
By contrast, there are only N axes in -coords, though there is an additional
preprocessing cost, pointed out in the ensuing. For N ≥ 10, the practical
barriers due to the required display space and visual difficulty limit the use of
SM. These barriers are less stringent for -coords.
3. Even with a nice perspective, orthogonal coordinates are inherently limi-
ted to N = 3, due to the dimensionality of our existence. With added
aspects, the illusion of a few more dimensions can be added. By contrast, for
-coords, implementation capabilities rather than conceptual barriers determine
the maximum useable N.
4. An example is the “Chernoff Faces” display, where each variable corresponds
to a specific facial feature and treated accordingly. The correspondence facial
feature → variable is arbitrary. Choosing a different correspondence gives a
different display, and the fun part is that there is no way to know that the two
different displays portray the same data set. Of course, this is true for general
glyphs displays.
5. This is true for SM and for -coords, though it has not been implemented in
general. Incidentally, this is a wonderful M.Sc. thesis topic.
6. The real value of visualization, in my opinion, is not seeing “zillions of objects”
but rather recognizing relations among them. We know that projections lose
information that can possibly be recovered with interactivity. Nevertheless,
important clues that can guide the exploration are lost. So it is preferable to
start with a display containing all the information, though it may be difficult
to uncover. The visual cues available and how they guide the exploration are
crucial deciding factors.
7. The value of rigor is self-evident.

These and additional issues that constitute the discovery process are better
appreciated via the exploration of real data sets. The basic queries are introduced
first with the example of a satellite and then a much larger astronomical data
set. Subsequently, they are combined with Boolean operators to form complex
queries applied to financial data. An example with several hundred variables
is discussed briefly before we move to automatic classification. Visualization
and -coords play key roles in the geometric algorithm’s conception, internal
function, and visual presentation of the classification rule. The minimal set of
variables needed to state the rule is found and ordered according to their predictive
value.
Mathematical background is interspersed to provide a deeper understanding
and wiser use of -coords and its applications. Specifically,
382 10. ♣ FT Data Mining and Other Applications

1. learning the patterns corresponding to the basic relations and seeking them out
for exploratory data analysis,
2. understanding the design and use of queries,
3. motivating further sophisticated applications to statistics such as response
surfaces [67], and
4. applications to regression; see the chapter on proximate planes,
5. understanding that the relational information resides in the crossings,
6. concentrating the relational information in the data in clear patterns elimi-
nating the polygonal lines altogether as with the “proximate planes” is feasible.
Before entering into the nitty gritty, we pose a visualization challenge that we
ask the reader to ponder. For a plane

π : c1 x1 + c2 x2 + c3 x3 = c0 , (10.1)

allow the coefficients to vary within a small interval. This generates a family of
“close” (let us call them proximate) planes

 = {π : c1 x1 + c2 x2 + c3 x3 = c0 , ci ∈ [ci− , ci+ ], i = 0, 1, 2, 3}. (10.2)

These are the planes generated by small rotations and translations of π with respect
to the three coordinate axes. Altogether they form a “twisted slab,” which even in
3-D with orthogonal axes is difficult to visualize. Conversely, given many points
in 3-D, how can it be discovered, using any general visual method you like, that
they lie on a twisted slab and how can such a creature be visualized and described
precisely, for N = 3 and then for any N? In the meantime, you can project, pursue,
scale, reduce, scintillate, regress, corollate, or tribulate to your heart’s content, but
please do not peek at the answer given at the end.

10.2 Exploratory Data Analysis with -coords

10.2.1 Multidimensional Detective
Parallel coordinates transform multivariate relations into 2-D patterns suitable for
exploration and analysis. For this reason they are included in many software tools.
The queries “parallel coordinates + software” on Google returned about 31,000
“hits” and “scatterplot matrix + software” about 15,000. Irrespective of the appar-
ent 2:1 relative ratio, the comparable numbers for the two astounded me, having
heard the appellations “esoteric,” “unnatural,” “difficult,” “squiggly,” and more for
-coords after their introduction.
 10.2 Exploratory Data Analysis with -coords 383

The exploration51 paradigm is that of a detective, starting from the data, search-
ing for clues leading to conjectures, testing, backtracking until voilá, the “culprit” is
discovered! The task is especially intricate when many variables (i.e., dimensions)
are involved, calling for the employment of a multidimensional detective (MD). As if
there were any doubt, our display of choice is -coords, where the data appears in
the by now familiar squiggly blotches and which, by means of queries, the MD
skilfully dissects to find precious hidden secrets.
During the ensuing interaction, think, dear reader, how similar queries can
be stated using other exploration methodologies, including the ubiquitous spread-
sheets. More important, what visual clues are available that would prompt the use
of such queries. A few basics (covered in Chapter 3) are recalled. In -coords
due to the point ↔ line and other dualities, some but not all actions are best
performed in the dual. The queries, which are the “cutting tools,” operate on the
display, i.e., dual. Their design should exploit the methodology’s strengths and
avoid its weaknesses, rather than mimic the action of queries operating on standard
“nondual” displays. Like a surgeon’s many specialized cutting tools, one of our
early software versions had many specialized queries. Not only was it hard to
classify and remember them but they still could not handle all situations encoun-
tered. After experimentation, I chose a few (three) intuitive queries called atomic,
which can be combined via Boolean operations to form complex intricate cuts.
Even for relatively small data sets, the -coords display can look uninformative
and intimidating. Lack of understanding of the basics of the underlying geometry
and poor choice of queries limits the use of -coords to unrealistically small
data sets. Summarizing, the requirements for successful exploratory data analysis
are:
• an informative data display,
• good choice of queries, and
• skillful interaction with the display.

10.2.2 An Easy Case Study: Satellite Data

The first admonition is
• do not let the picture intimidate you,
as can easily happen by taking an uninformed look at Fig. 10.2 showing the dataset
to be explored. It consists of over 9,000 measurements with 9 variables, the first
51
The venerable name “exploratory data analysis” (EDA) is used interchangeably with the currently
more fashionable “visual data mining.”
384 10. ♣ FT Data Mining and Other Applications

Figure 10.1. Seven types of ground emissions were measured in this region of Slovenia.
Measurements recorded by the LandSat Thematic Mapper are shown in subsequent figures. Thanks
to Dr. Ana Tretjak and Dr. Niko Schlamberger, Statistics Office of Slovenia, for providing the data.
(Right) The display is the map’s rectangular region; the dot marks the position where the 7-tuple
shown in the next figure was measured.

two (X, Y ) specify the location on the map in Fig. 10.1 (left), a portion of Slovenia,
where 7 types of ground emissions are measured by satellite. The ground location,
(X, Y ), of one data item is shown in Fig. 10.1 (right), which corresponds to the map’s
region and remains open during the exploration. The query, shown in Fig. 10.2, used
to select the data item is called Pinch. It is activated by the button P on the tool bar.
By means of this query, a set of polygonal lines (i.e., data items) can be chosen by
being “pinched” between the axes. The cursor’s movement changes the position of
the selected arrowhead, which is the larger of the two shown. In due course, various
parts of the GUI are illustrated (Parallax).52 Aside from starting the exploration
without biases, it is essential

• to understand the objectives.

Here the task is the detection and location of various ground features (built-up
areas, vegetation, water, etc.) on the map. There is a prominent lake in the lower-
left corner with an unusual shape of an upward-pointing “finger.” This brings up
the next admonition, that no matter how messy it looks,

52
MDG’s Ltd proprietary software; all rights reserved, is used by permission.
 10.2 Exploratory Data Analysis with -coords 385

Figure 10.2. Query on parallax showing a single data item.

The X, Y (position, also shown on the right of Fig. 10.1), and values of the 7-tuple (B1, B2, B3,
B4, B5, B6, B7) at that point.

• carefully scrutinize the data display for clues and patterns.

Follow up on anything that catches the eye: gaps, regularities, holes, twists, peaks
and valleys, density contrasts like the one at the lower values of B3 through B7.
Using the Interval query, activated by the I button, starting at the minimum we grab
the low range of B4 (between the arrowheads) stopping at the dense part as shown
in Fig. 10.3. The result, on the left of Fig. 10.4, is amazing. We found the water!
The lake is clearly visible together with two other regions, which in the map turn
out to be small streams. Our scrutiny having been rewarded, we recall the adage,
• a good thing may be worth repeating.
Now examining for density variations within the selected lower interval of B4, we
notice another. The lowest part is much denser. Experimenting a bit, appreciating
the importance of interactivity, we select the sparse portion, Fig. 10.5, which defines
the water’s edge (right) Fig. 10.4 and in fact more. By dropping the lower arrow, we
see the lake filling up, starting from the edge, i.e., shallow water first. So the lower
values of B4 reveal the water, and the lowest “measure,” the water’s depth; not bad
for few minutes of playing around. But all this pertains to a single variable, when
we are supposed to be demonstrating multivariate exploration. This is a valid point,
386 10. ♣ FT Data Mining and Other Applications

Figure 10.3. Finding water regions.

The contrast due to density differences around the lower values of B4 is the visual cue that prompts
this query.

Figure 10.4. (Left) The lake; result of query shown in Fig. 10.3.
On the right is just the lake’s edge. It is the result of the query shown in Fig. 10.5.

but we did pick B4 among several variables. Further, this is a nice “warm-up” for
the subsequent more involved examples, enabling us to show two of the queries.
The astute observer must have already noticed the regularity, the vertical bands,
between the B1, B2, and B3 axes. This is where the angle query, activated by the
 10.2 Exploratory Data Analysis with -coords 387

Figure 10.5. Query finding the water’s edge.

A button, comes into play. As the name implies, it selects groups of lines within
a user-specified angle range. A data subset is selected between the B2 and B3
axes as shown, with enlarged interaxis distance better showing the vertical bands,
in Fig. 10.6 (left) to select a data subset that corresponds on the map to regions
with high vegetation. Clicking the A button and placing the cursor in the middle
of one axis opens an angle, with vertex on the midrange of the previous (left) axis,
whose range is controlled by the arrow movements on the right axis. Actually, this
“rule” (i.e., relation among some parameters) for finding vegetation can be refined
by tweaking a couple of more parameters. This raises the topic of rule finding in
general, classificatio , which is taken up in Section 10.3.
The angle and pinch queries are motivated by the line → point ¯ duality

d b
: x2 = mx1 + b ↔ ¯ = , (10.3)
1−m 1−m
in -coords, illustrated in Fig. 10.7, where the interaxis distance is d. As seen from
its x coordinate, the point ¯ lies between the parallel axes when the line’s slope m
is negative, to the right of the X̄2 axis for 0 < m < 1, and to the left of X̄1 for
m > 1. Lines with m = 1 are mapped to the direction with slope b/d in the xy
plane, with d the interaxis distance and b the constant (intercept) in the equation of
. This points out that dualities properly reside in the projective plane, the directions
388 10. ♣ FT Data Mining and Other Applications

Figure 10.6. Finding regions with vegetation.

Here the angle query is used (left) between the B2 and B3 axes. Note the arrowheads on the B3 axis,
which specify the angle range for the selected lines.

being the ideal points, rather than the Euclidean plane. For sets of points having a
“overall” direction with negative slope, i.e., are “negatively correlated,” the lines
representing them in -coords cross each other roughly between the axes, and they
can be selected with the pinch query. For positively correlated sets of points, their
corresponding lines cross outside the axes and can be selected with the angle query.
All this exemplifies the need to understand some of the basic geometry so as to work
effectively with the queries and of course design them properly. The three atomic
queries having been introduced, there remains to learn how they can be combined
to construct complex queries.
 10.2 Exploratory Data Analysis with -coords 389

Figure 10.7. Parallel coordinates induce a point ¯ ↔ line duality (left).

(Right) The horizontal position of the point ¯ representing the line is determined only by the line’s
slope m. The vertical line : x1 = a1 is represented by the point ¯ at the value a1 on the X̄1 axis.

Prior to that, Fig. 10.6 (left) raises the question, what if the B2 and B3 axes were
not adjacent? Then the pattern and hence their pairwise relation would be missed.
Hence the axis permutation used for the exploration is important. In particular, what
is the minimum number of permutations among N axes containing the adjacencies
for all pairs of axes? It turns out [95], [184] that M permutations are needed for
even N = 2M, and M + 1 for odd N = 2M + 1. It is fun to see why. Label the
N vertices of a graph with the index of the variables Xi , i = 1, . . . , N, as shown
in Fig. 10.8 for N = 6. An edge joining vertex i with j signifies that the axes
indexed by i, j are adjacent. The graph on the left is a Hamiltonian path, for it
contains all the vertices. Such paths have been studied beginning with Euler in the
eighteenth century with modern applications to the “traveling salesman” problem
and elsewhere [77, pp. 66], [16, pp. 12]. The graph corresponds to the axis index
permutation 126354. On the right, the union with the additional two Hamiltonian
paths, starting at vertices 2 and 3, forms the complete graph, which contains all
possible edges. Hence the three permutations 126354, 231465, 342516 contain all
possible adjacent pairs; just try it. The remaining permutations are obtained from
the first by successively adding 1 mod 6, and this works for general N .
390 10. ♣ FT Data Mining and Other Applications

Figure 10.8. (Left) First Hamiltonian path on vertices 1, . . . , 6.

It corresponds to the (axes) index permutation 126354. (Right) The complete graph as the union of
the three distinct Hamiltonian paths starting successively at the vertices 1, 2, 3.

Returning to EDA, the icon with the Rubik’s Cube on Parallax’s toolbar activates
a permutation editor, which automatically generates the Hamiltonian permutations
(HP). After scrutinizing the data set display, the recommended next step is to run
through the O(N/2) HP. This is how all nice adjacencies such as the one in Fig. 10.6
are discovered. Then using the editor, patch your own custom-made permutation
containing all the parts you like in the HP. With this preprocessing cost, referred to
earlier in list item 2 of the introduction, the user sets her own best permutation to
work with. Of course, there is nothing to prevent one from including axes several
times in different positions and experimenting with different permutations in the
course of the exploration.

10.2.3 Quasars Dataset

Much larger datasets53 can be explored, and to make this point convincingly,
an example from the Quasars catalog having more than 46,000 data items and
21 parameters as displayed in Fig. 10.9 is used.
There are two clusters and a prominent outlier with high value in the second
variable. It is evident that several parameters behave similarly and can be omitted,
at least initially, uncluttering the display. The reduced dataset is shown on the left
of Fig. 10.10, with the two clusters colored differently, enhancing the differences
and similarities between them. The Pinch query used here is indicated with the
oppossing arrows; recall that it operates between the axes, as contrasted with that
in Fig. 10.4, called Interval, which operates on a single-variable axis.
53
I am grateful to Don Schneider, Jogesh Babu, and Erik Feigelson, who suggested this data set, and
Hyunsook Lee, who downloaded the data set from
http : //astrostatistics.psu.edu/datasets/SDSS− quasar.dat, and assisted with the exploration.
 10.2 Exploratory Data Analysis with -coords 391

Figure 10.9. The full Quasars dataset with 46, 000+ data items and 21 parameters.
There are two major clusters and an outlier with high value in the second variable. Several parameters
having similar values can be eliminated, uncluttering the display, as shown below in Fig. 10.10.

Figure 10.10. The two clusters are colored differently.

(Left) The high range between the sixth and seventh variables is selected, revealing patterns involving
the intermediate range of the variable labeled Radio (appearing as fifth from the right). (Right)
On the scatterplot of X-Ray vs. Radio, the only cluster without vertical or horizontal segments is
selected.
392 10. ♣ FT Data Mining and Other Applications

Figure 10.11. The selected cluster consists of 234 data items.

(Left) Relation between X-Ray and Radio. (Right) Relation between Radio and sig-K.

Note the patterns involving the intermediate range of values of the parameter
labeled Radio. On the right within the scatterplot of the variables Radio vs. X-Ray,
the only subset without horizontal or vertical segments (for missing values) is
selected using a query called Polygon. It has 234 data items with interesting rela-
tions, as shown in Fig. 10.11, between Radio vs. X-Ray and Radio vs. sig-K. The
explorations here are relatively straightforward. Much more complex multivariate
relations were discovered in complex datasets, and some are shown next. For such a
discovery process, more queries are needed whose design requires a deeper under-
standing of the methodology’s foundations. They are summarized next, preparing
the ground for some sophisticated applications described later.

10.2.4 Compound Queries: Financial Data

To be explored next is the financial data set shown in Fig. 10.12, the goal being to
discover relations useful for investments and trading. The data for the years 1986
(second tick on the third axis) and 1992 are selected and compared. In 1986, the
yen had the greatest volatility among the three currencies, interests varied in the
midrange, gold had a price gap, while SP500 was uniformly low. By comparison
in 1992, the yen was stable while the pound sterling was very volatile (possibly
due to Soros’s speculation); interests and gold prices were low, and the SP500 was
uniformly high. Two interval queries are combined with the OR Boolean operator
(i.e., union) to obtain this picture. We continue
• “looking for the gold” by checking out patterns that caught our attention.
The data for 1986 is isolated in Fig. 10.13, and the lower range in the gold price
gap is selected. Gold prices were low until the second week in August, when they
 10.2 Exploratory Data Analysis with -coords 393

Figure 10.12. Financial data.

Quotes by week on Monday, month, year, first three axes fix the date; Sterling (BPS) Dmark, yen
rates per dollar; fourth, fifth, sixth axes; 3MTB, 30YTB interest rates in percent for the first three
months and 30-year bonds, respectively; seventh, eighth axes; gold in $/ounce, ninth, SP500 index
values on tenth axis.

jumped and stayed higher. The exploration was carried out in the presence of four
financial experts, who carefully recorded the relation between low yen, high 3MTB
rates, and low gold prices. By the way, the low yen rate of exchange means that the
yen had a high value relative to the US dollar.
There are two bunches of crossing lines between sixth and seventh axes in
Fig. 10.12, which together comprise more than 80% of the data set. This and
recalling the previous discussion on the line ↔ point mapping in Fig. 10.7 points
out the strong negative correlation between yen and 3MTB rates. The smaller cluster
in Fig. 10.14 is selected. Moving from the top range of any of the two axes, with
the I query, and lowering the range causes the other variable’s range to rise and is
a nice way to show negative correlation interactively. For the contrarians among
us, we check also for positive correlation, Fig. 10.15. We find that it exists when
gold prices are low to midrange, as happened for a period in the 1990s. This is
a free investment tip for bucking the main trend shown in Fig. 10.14. It is also
a nice opportunity for showing the inversion feature activated by the icon with
two cyclical arrows. A variable is selected and the min/max values on that axis
are inverted. Diverging lines (as for positive correlation) now intersect Fig. 10.16,
making it easier visually to spot the crossing and hence the correlation. Actually,
394 10. ♣ FT Data Mining and Other Applications

Figure 10.13. Gold prices in 1986.

Gold prices jumped in the second week of August. Note the correlation between the low yen, high
3MTB rates, and low gold price range.

Figure 10.14. Negative correlation.

The crossing lines between the sixth and seventh axes in Fig. 10.12 show strong negative correlation
between yen and 3MTB rates. One cluster is selected with the Pinch query and combined with the
high and low ranges on the yen axis. Data for the years 1986 and 1992 are selected.
 10.2 Exploratory Data Analysis with -coords 395

Figure 10.15. Positive correlation.

A positive correlation where the yen and 3MTB rates move together when gold prices are low to
midrange.

Figure 10.16. Inverting the 3MTB axis.

Now the lines between the yen-3MTB and 3MTB-30MTB axes in Fig. 10.15 cross.
396 10. ♣ FT Data Mining and Other Applications

Figure 10.17. Variations in currency exchange rates.

Variations in the rate of exchange of the currencies correlate with movements in the price of gold.

the recommendation is to work with the A query, experimenting with various angle
ranges using inversion to check out or confirm special clusters. When stuck, don’t
just stand there, but
• vary one of the variables, watching for interesting variations in the other
variables.
Doing this on the yen axis, Fig. 10.17, we strike another gold connection. The
(rough) intersection of a bunch of lines joining yen to the Dmark corresponds, by
the duality, to their rate of exchange. When the rate of exchange changes, so does
the intersection and the price of gold! That is, movements in currency exchange
rates and the price range of gold go together. Are there any indications that are
associated with the high range of gold? The top price range is selected, Fig. 10.18,
and prompted by the result of the previous query, we check out the exchange rate
between Sterling (BPS) and Dmark (or yen), and the result is stunning: a perfect
straight line. The slope is the rate of exchange, which is constant when gold tops
out. The relation between Sterling (BPS) and the Dmark is checked for different
price ranges of gold, Fig. 10.19, and the only regularity found is the one straight
line above. Aside from the trading guideline it establishes, it suggests “behind-the-
scenes manipulation of the gold market.” We could have said that, but we won’t.
We perish this thought and proceed with the Boolean complement, Fig. 10.20, of an
I (or any other) query. Not finding anything, we select a narrow but dense range on
 10.2 Exploratory Data Analysis with -coords 397

Figure 10.18. High gold.

Note the perfect straight line in the sterling vs. Dmark plot. The slope is the rate of exchange between
them, which remains constant when gold prices peak.

the yen, Fig. 10.21, and notice an interesting relation between the Dmark, interest
rates, and gold.
There is an exploratory step akin to “multidimensional contouring,” which we
fondly call Zebra, activated by the last icon button on the right with the appropriate
skin color. A variable axis is selected, the SP500 axis in Fig. 10.22, and divided
into a number (user specified) of intervals (here it is 4) and colored differently. This
shows the connections (influence) of the intervals with the remaining variables,
which here is richly structured especially for the highest range. So what does it
take for the SP500 to rise? This is a good question and helps introduce Parallax’s
classifier, described in 10.3. The result, shown in Fig. 10.23, confirms the investment
community’s experience that low 3MTB and gold predict high SP500. A comparison
with the results obtained on this data set with other visualization tools would be
instructive, though unfortunately one is not available. Still let us consider such an
analysis done by the scatterplot matrix. There are 10 variables (axes), which requires
45 pairwise scatterplots each, even with a large monitor screen being no larger than
about 2.5 × 2.5 cm2 . Varying one, two, or more variables in tandem and observing
the effects simultaneously over all the variables in the 45 squares may be possible but
quite challenging. By contrast, the effects of varying the Dmark, conditionally for
398 10. ♣ FT Data Mining and Other Applications

Figure 10.19. Two price ranges of gold.

The associated sterling vs. Dmark plots show no regularity.

Figure 10.20. The complement of an I query.

stable yen, are easily seen on the two interest rates, gold as well as the remaining
variables in one, Fig. 10.21. This example illustrates the difficulties due to high
 10.2 Exploratory Data Analysis with -coords 399

Figure 10.21. Yen stable.

For the yen trading in a narrow range, high Dmark goes with low 3MTB rates, low Dmark goes with
high 3MTB rates, while mid 3MTB rates go with high gold.

Figure 10.22. The zebra query.

It partitions and colors the segments of values differently. A variable, here the SP500 axis, is divided
it into equal (here 4) intervals. This quickly reveals interelationships. Note especially those for the
highest SP500 range and see the next figure.
400 10. ♣ FT Data Mining and Other Applications

Figure 10.23. The rule for high SP500.

Both 3MTB (the “short-bond” as it is called) and gold are low and in this order of importance.

representational complexity (see Section 10.1.2 item #2), which is O(N 2 ) for the
scatterplot matrix but O(N) for -coords and made even clearer with the next
data set.

10.2.5 Hundreds of Variables

An important question frequently asked is, how many variables can be handled with
-coords? The largest data set that I have effectively worked with had about 800
variables and 10,000 data entries. With various techniques developed over the years
and the automatic classifier discussed in the next section, much larger data sets can
be handled. Still the relevant admonition is
• be skeptical about the quality of data sets with a large number of variables.
When hundreds or more variables are involved, it is unlikely that there are many
people around who have a good feel for what is happening, as confirmed by my
experience. A case in point is the data set shown in Fig. 10.24, consisting of
instrumentation measurements for a complex process. The immediate observation is
that many instruments recorded 0 for the duration, something which was unnoticed
by the process monitors. Another curiosity was the repetitive patterns on the right.
It turns that several variables were measured in more than one location using
 10.2 Exploratory Data Analysis with -coords 401

Figure 10.24. Manufacturing process measurements; 400 variables.

different names. When the data set was cleaned up of superfluous information,
it was reduced to about 90 variables, as shown in Fig. 10.25 and eventually to
about 30, which contained the information of real interest. By my tracking, the
phenomenon of repetitive measurements is widespread, with at least 10% of the
variables occurring in large data sets being duplicates or near duplicates, possibly
due to instrument nonuniformities, as suggested in the 2-variable scatterplot2 in
Fig. 10.25. Here the repetitive observations were easily detected due to the fortuitous
variable permutation in the display. Since repetitive measurements occur frequently,
it may be worth adding to the software an automated feature to detect and exhibit the
suspect variables. This brief exposure is just an indication that large (in dimension,
i.e., in number of variables) data sets can still be gainfully explored in -coords.
There follows a different example of EDA on a process-control data set [102],
where compound queries turned out to be very useful and where we learn to add to
the list of exploration guidelines arguably the most important one:
• test the assumptions and especially the “I am really sure of”s.

10.2.6 Production of VLSI (Chips)

The data set displayed in Fig. 10.26 consists of production data of several batches
of a specific VLSI (computer chip) with measurements of 16 parameters involved
402 10. ♣ FT Data Mining and Other Applications

Figure 10.25. The above data set after “cleanup” with about 90 variables left.

Figure 10.26. Dataset; VLSI production with 16 parameters.

 10.2 Exploratory Data Analysis with -coords 403

Figure 10.27. The batches high in Yield, X1, and Quality, X2.

in the process. The parameters are denoted by X1, X2, . . . , X16. The yield, as
the percentage of useful chips produced in the batch, is denoted by X1, and X2
is a measure of the quality (given in terms of speed performance) of the batch.
Ten different categories of defects are monitored, and the variables’ scales of X3
through X12 are 100% pure (zero defect) so that the amounts of defect appear at
the top and increasing amounts appear proportionately lower. The remaining X13
through X16 denote some physical parameters.
Since the goal here is to raise the yield, X1, while maintaining high quality, X2,
we have the problem of multiobjective optimization due to the presence of more
than one objective. The production specialists believed that it was the presence of
defects that prevented high yields and qualities. So their purpose in life was to keep
on pushing — at considerable cost and effort — for zero defects.
With this in mind, the result of our first query is shown in Fig. 10.27, where
the batches having the highest X1 and X2 have been isolated. This in an attempt
to obtain clues; and two real good ones came forth. Notice the resulting range of
X15, where there is a significant separation into two clusters. As it turns out, this
gap yielded important insight into the physics of the problem. The other clue is
almost hidden. A careful comparison — and here interactivity of the software is
404 10. ♣ FT Data Mining and Other Applications

Figure 10.28. The batches with zero in nine out of ten defect types.

essential — between Fig. 10.26 and Fig. 10.27 shows that some batches that were
high in X3 (i.e., low in that defect) were not included in the selected subset. That
casts some doubt into the belief that zero defects are the panacea and motivates
the next query, where we search for batches having zero defects in at least nine
(excluding X3, where we saw that there are problems) out of the ten categories.
The result is shown in Fig. 10.28 and is a shocker. There are nine such batches,
and all of them have poor yields and for the most part also low quality! That this
was not questioned and discovered earlier is surprising. We scrutinize the original
picture Fig. 10.26 for visual cues relevant to our objectives and our findings so far.
And there is one staring us in the face! Among the ten defects X3 through X12,
whatever X6 is, its graph is very different from those of the others. It shows that
the process is much more sensitive to variations in X6 than the others. For this
reason, we chose to treat X6 differently and remove its zero-defect constraint. This
query (not shown) showed that the very best batch (i.e., highest yield with very high
quality) does not have zeros (or the lowest values) for X3 and X6; a most heretical
finding. It was confirmed by the next query, which isolated the cluster of batches
with the top yields (note the gap in X1 between them and the remaining batches).
 10.2 Exploratory Data Analysis with -coords 405

Figure 10.29. The batches with the highest yields.

They do not have the lowest defects of type X3 and X6.

These are shown in Fig. 10.29, and they confirm that small amounts (the ranges can
be clearly delimited) of X3- and X6-type defects are essential for high yields and
quality.
Returning to the subset of data that best satisfied the objectives, Fig. 10.27, in
order to explore the gap in the range of X15, we found that the cluster with the
high range of X15 gives the lowest (of the high) yields X1, and worse, it does not
give consistently high quality X2, whereas the cluster corresponding to the lower
range has the higher qualities and the full range of the high yield. It is evident that
the small ranges of X3, X6 close to (but not equal to) zero, together with the short
(lower) range of X15, provide necessary conditions for obtaining high yields and
quality. This is also indicated in Fig. 10.29. By a stroke of good luck, these three can
also be checked early in the process, avoiding the need of “throwing good money
after bad” (i.e., by continuing the production of a batch whose values of X3, X6,
and X15 are not in the small “good” ranges we have found).
These findings were significant and differed from those found with other
methods for statistical process control [9]. This approach has been successfully used
406 10. ♣ FT Data Mining and Other Applications

Figure 10.30. Only the lower range of X15 is associated with the highest yields and quality.

in a wide variety of applications, including the manufacture of printed circuit boards,

PVC and manganese production, financial data, and determining skill profiles (i.e.,
as in drivers, pilots).

10.3 Classification
Though it is fun to undertake this kind of exploration, the level of skill and patience
required tends to discourage some users. It is not surprising, then, that the most
persistent requests and admonitions have been for tools that, at least partially,
automate the knowledge discovery process [103]. Classification is a basic task
in data analysis and pattern recognition, and an algorithm accomplishing it is called
a classifie [149], [57], [136]. The input is a dataset P and a designated subset S.
The output is a characterization, a set of conditions or rules, to distinguish elements
of S from all other members of P , the “global” dataset. The output may also be that
there is insufficient information to provide the desired distinction.
With parallel coordinates, a dataset P with N variables is transformed into
a set of points in N-dimensional space. In this setting, the designated subset S
can be described by means of a hypersurface that encloses just the points of S.
 10.3 Classification 407

In practical situations, the strict enclosure requirement is dropped, and some points
of S may be omitted (“false negatives”), and some points of P − S are allowed
(“false positives”) in the hypersurface. The description of such a hypersurface is
equivalent to the rule for identifying, within some acceptable error, the elements of
S. Casting the problem in a geometric setting leads us to visualize how such a rule
may work. This entails:
1. Use of an efficient “wrapping” (a convex-hull approximation) algorithm to
enclose the points of S in a hypersurface S1 containing S and in general also
some points of P − S; so S ⊂ S1 .
2. The points in (P −S) ∩S1 are isolated and the wrapping algorithm is applied to
enclose them, and usually also some points of S1 , producing a new hypersurface
S2 with S ⊃ (S1 − S2 ).
3. The points in S not included in S1 −S2 are next marked for input to the wrapping
algorithm, a new hypersurface S3 is produced containing these points as well
as some other points in P − (S1 − S2 ), resulting in S ⊂ (S1 − S2 ) ∪ S3 .
4. The process is repeated, alternatively producing upper and lower contain-
ment bounds for S; termination occurs when an error criterion (which can
be user-specified) is satisfied or when convergence is not achieved. After
termination is obtained, two error measures are available to estimate the rule’s
precision:
• Train & Test. A portion of the dataset (usually 2/3) selected at random is used
to derive the classification rule, which is then tested on the remaining 1/3 of the
data.
• Cross-Correlation.
It can and does happen that the process does not converge when P does not
contain sufficient information to characterize S. It may also happen that S is so
“porous” (i.e., spongelike) that an inordinate number of iterations are required.
On convergence, say at step 2n, the description of S is provided as

S ≈ (S1 − S2 ) ∪ (S3 − S4 ) ∪ · · · ∪ (S2n−1 − S2n ), (10.4)

this being the terminating expression resulting from the algorithm that we call nested
cavities (NC).
The user can select a subset of the available variables and restrict the rule
generation to these variables. In certain applications, as in process control, not all
variables can be controlled, and hence it is useful to have a rule involving only the
accessible (i.e., controllable) variables. An important additional benefit is that the
minimal set of variables needed to state the rule is found and ordered according
408 10. ♣ FT Data Mining and Other Applications

to their predictive value. These variables may be considered as the best features to
identify the selected subset. The algorithm is display-independent and there is no
inherent limitation as to the size and number of variables in the dataset. Summarizing
for NC,

• an approximate convex-hull boundary for each cavity is obtained,

• utilizing properties of the representation of multidimensional objects in
-coords, a very low polynomial worst-case complexity of O(N 2 |P |2 ) in the
number of variables N and dataset size |P | is obtained; it is worth contrasting this
with the often unknown, or unstated, or very high (even exponential) complexity
of other classifiers,
• an intriguing prospect, due to the low complexity, is that the rule can be derived
in near real time, making the classifier adaptive to changing conditions,
• the minimal subset of variables needed for classification is found,
• the rule is given explicitly in terms of conditions on these variables, i.e., included
and excluded intervals, and provides a “picture” showing the complex distribu-
tions with regions where there are data and “holes” with no data providing
insights to the domain experts.
During 1990–1993, Michie, Spiegelhalter, and Taylor [134], on behalf of
the ESPRIT program of the European Union, made extensive studies of several
classifiers applied to diverse data sets. About 23 different classifiers were applied to
about 20 data sets for comparative trials in the StatLog project. This was designed
to test classification procedures on large-scale commercially important problems
in order to determine suitability of the various techniques to industry. The results
obtained by NC are compared with those obtained by other well-known classi-
fiers used in Statlog on two benchmark datasets and shown in the accompanying
tables.

1. Satellite image dataset has over 6,000 items and NC’s classification error was
9%, k-NN was next with 9.4%, the remainder had errors as high as 30%, and
one was unable to be classified at all, as shown in Table 10.1.
2. Vowel recognition data with about 1,000 data items. The data collection process
involves digital sampling of speech with acoustic signal processing, followed
by recognition of the phonemes, groups of phonemes, and words. The goal here
is a speaker-independent rule based on 10 variables of 11 vowels occurring in
various words spoken (recorded and processed) by 15 British male and female
speakers. Deterding [40] collected this dataset of vowels, which can be found
in the CMU benchmark repository on the Internet. There are 528 entries for
training and 462 for testing. Three other types of classifiers were also applied
 10.3 Classification 409

Table 10.1. Summary of the Statlog results and com-

parison with the nested cavities (NC) classifier for the
satellite image data.

Error rate %
Rank Classifie Train Test

1 NC 4.3 9.0
2 k-NN 8.9 9.4
3 LVQ 4.8 10.5
4 DIPOL92 5.1 11.1
5 RBF 11.1 12.1
6 ALLOC80 3.6 13.2
7 IndCART 2.3 13.8
8 CART 7.9 13.8
9 Backprop 11.2 13.9
10 Baytree 2.0 14.7
11 CN2 1.0 15.0
12 C4.5 4.0 15.0
13 NewID 6.7 15.0
14 Cal5 12.5 15.1
15 Quadisc 10.6 15.5
16 AC 2 11.3 15.7
17 SMART 12.3 15.9
18 Cascade 11.2 16.3
19 Logdisc 11.9 16.3
20 Discrim 14.9 17.1
21 Kohonen 10.1 17.9
22 CASTLE 18.6 19.4
23 NaiveBay 30.8 28.7
24 ITrule Failed Failed

to this dataset: neural networks and k-NN by Robinson and Fallside [56], and
decision trees by Shang and Breiman [158]. Again NC was top with 7.9%;
next was CART-DB at 10%; the rest were down to 66%, with many unable to
provide a classification rule. The results are shown in Table 10.2.
410 10. ♣ FT Data Mining and Other Applications

Table 10.2. Summary of classification results for the vowel dataset.

Rank Classifie Testing Mode Test Error Rate %

1 nested cavities (NC) cross validation 7.9

2 CART-DB cross validation 10.0
3 nested cavities (NC) train & test 10.5
4 CART cross validation 21.8
5 k-NN train & test 44.0
6 RBF train & test 46.5
7 multilayer perceptron train & test 49.4
8 single-layer perceptron train & test 66.7

3. A neural-pulse dataset has interesting and unusual features. There are two
classes of neurons whose outputs to stimuli are to be distinguished. They consist
of 32 different pulses measured in a monkey’s brain (poor thing!). There are
600 samples with 32 variables (the pulses).54 Various classification methods
were unable to obtain a rule. With NC, convergence is obtained requiring only
9 of the 32 parameters for the classification rule for category 1. The resulting
ordering shows a striking separation. In Fig. 10.31, the first pair of variables
x1 , x2 in the original order is plotted on the left. On the right, the best pair
x11 , x14 , as chosen by the classifier’s ordering speaks for itself. By the way,
the discovery of this manually would require constructing a scatterplot matrix
with 496 pairs, then carefully inspecting and comparing the individual plots.
The implementation provides all the next-best sections, some of which are
shown in Fig. 10.31, to aid the visualization of the rule. The data set consists
of two “pretzel-like” clusters wrapping closely in 8-D, one enclosing the other,
showing that the classifier can actually “carve” highly complex regions with the
cavity shown. One can understand why separation of clusters by hyperplanes
or nearest-neighbor techniques can fail badly for such data sets. The rule has
4% error, some of which are shown in Fig. 10.32.
The rules are explicit, “visualizable,” optimally ordering the minimal set of
variables needed to state the rule without loss of information. There are variations
that apply in some situations where the NC classifier fails, such as the presence of

54
I am grateful to Prof. R. Coiffman and his group at the Departments of Mathematics and Computer
Science at Yale University for giving me this data set.
 10.3 Classification 411

Figure 10.31. The neural-pulses dataset with 32 parameters and two categories.
The data set is shown in the background. On the left plot are the first two parameters in the original
order. The classifier found the nine parameters needed to state the rule with 4% error and ordered
them according to their predictive value. The best two parameters are plotted on the right, showing
the separation achieved.

several large “holes” (see [103]). Further, keeping in mind that the classification
rule is the result of several iterations suggests heuristics for dealing with the pesky
problem of overfittin . The iterations can be stopped just where the corrections in
(10.4) become very small, i.e., the Si consist of a small number of points. The
number of iterations is user defined, and the resulting rule yields an error in the test
stage more stable under variations in the number of points of the test set. In addition,
the user can exclude variables from being used in the description of the rule; those
ordered last are the ones providing the smaller corrections and hence more liable to
overcorrect.
412 10. ♣ FT Data Mining and Other Applications

Figure 10.32. Neural dataset classification.

Further cross-sections of the hypersurface corresponding to the classification rule.

10.3.1 Case Study: Noise Signature Recognition

It is worthwhile going through another classification example in greater detail.
About 1000 vehicles (trucks) were run on a road test tract embedded with sound
sensors at various locations and distances. The resulting noise was measured sepa-
rately within seven frequency ranges. For each range, the measurements were
subsequently transformed (via wavelets) into a single number. The resulting seven
numbers v1 , v2 , . . . , v7 provide the noise signature for each specific vehicle. The
truck data is divided into 11 classes (by type, manufacturer). Each data item (for a
truck) consists of nine parameters: the truck ID, class #, and noise signature. The
data set is shown in Fig. 10.33. There are trucks manufactured in one country A
and belonging to the army of one country I ; the remainder are made in country R
and belong to the army of a hostile country S.
The army of country I would like to detect the enemy’s vehicles from a distance
by their noise signature (i.e., the way they sound), something analogous to “land
sonar.” The vehicles of country S are the classes #7 – 11, which are selected and
shown in Fig. 10.34. The values of v4 show that there are two vehicle clusters.
 10.3 Classification 413

Figure 10.33. Eleven categories of vehicles (trucks).

The noise signature defined by the values of v1 , v2 , . . . , v7 is given for each vehicle.

Figure 10.34. Top five classes consist of vehicles manufactured in country R.

The values of v4 indicate two clusters. The values in the top dense range turn out to be for trucks
with automatic transmission, while the remaining lower values correspond to trucks with manual
transmission.
414 10. ♣ FT Data Mining and Other Applications

Figure 10.35. Automatic transmissions.

Cluster of vehicles S1 manufactured by country R having automatic transmissions. This is the
classifier’s input.

Examining the vehicles’ files from the IDs, a surprising separation is found: those
with v4 values in the top (dense range) have — let’s call them S1 (shown in
Fig. 10.35) — automatic transmission, and the remainder (denoted by S2 ) with
sparse v4 values have manual transmission! Classifying each of the S1 , S2 separately
and then taking the union of the rules provides a much more accurate overall rule
for S.
Proceeding, the classification rule for subset S1 is obtained with an error of
2%; Fig. 10.36. Variable v6 is redundant and can be omitted. The two variables
with the highest predictive value are v4 , v2 , and their plot indicates that they are
three clusters in the data set: a dense one with low v2 and v4 lower than S1 , and a
sparse one with higher values of v2 . Even though the rule’s error is very low, this is a
situation in which information on the misclassified trucks is very important. For this
purpose, we stop the iteration after the first step. The output is the subset shown in
Fig. 10.37. It consists of data from the top five categories, i.e., the subset S and some
from category #2. Knowing the classifier’s final output is obtained by subsequent
iterations refining this subset; we surmise that the misclassifieds are likely to be
from class #2. Independently, it was found that the class #2 trucks, manufactured
 10.3 Classification 415

Figure 10.36. Output: the classification rule with a 2% error is found.

Variable v6 is redundant, and the two best predictors v4 , v2 plot shows the three clusters composing
the data set.

Figure 10.37. The output after the first iteration.

In addition to trucks in the top five classes, the data set contains some from class # 2, so these are the
ones likely to be misclassified.
416 10. ♣ FT Data Mining and Other Applications

in country A, have been copied by a manufacturer in country R. No wonder, then,

that their sound is similar and can be confused with those in S.

10.4 Visual and Computational Models

Finally, we illustrate the methodology’s ability to model multivariate relations in
terms of hypersurfaces, just as we model a relation between two variables by a planar
region. Then by using the interior point algorithm, as shown in Fig. 10.45 of the next
section, with the model we can do trade-off analyses, discover sensitivities, under-
stand the impact of constraints, and in some cases do optimization. For this purpose
we shall use a data set consisting of the outputs of various economic sectors and other
expenditures of a particular (and real) country. It consists of the monetary values over
several years for the agricultural, fis ing, and mining sector outputs, manufacturing
and construction industries, together with government, miscellaneous spending, and
resulting GNP; eight variables altogether. We will not take up the full ramifications
of constructing a model from data. Rather, we want to illustrate how -coords may
be used as a modeling tool. Using the least squares technique, we “fit” a function
to this data set and are not concerned at this stage whether the choice is a “good”
one. The function obtained bounds a region in R8 , 8-dimensional Euclidean space,
and is represented by the upper and lower curves shown in Fig. 10.38.

Figure 10.38. Model of a country’s economy.

Choosing high agricultural and high fis ing output forces low mining output.
 10.4 Visual and Computational Models 417

Figure 10.39. Competition for labor between the fis ing and mining sectors.

The picture is in effect a simple visual model of the country’s economy, incorpo-
rating its capabilities, limitations, and interelationships among the sectors. A point
interior to the region satisfies all the constraints simultaneously, and therefore
represents a feasible economic policy for that country (i.e., the 8-tuple of values).
Using the interior point algorithm we can construct such points. It can be done
interactively by sequentially choosing values of the variables, and we see the result
of one such choice in Fig. 10.38. Once a value of the first variable is chosen (in this
case the agricultural output) within it’s range, the dimensionality of the region is
reduced by one. In fact, the upper and lower curves between the second and third
axes correspond to the resulting 7-dimensional hypersurface and show the available
range of the second variable (fis ing) reduced by the constraint. This can be seen
(but not shown here) for the rest of the variables. That is, due to the relationship
between the eight variables, a constraint on one of them impacts all the remaining
ones and restricts their range. The display allows us to experiment and actually
see the impact of such decisions downstream. By interactively varying the chosen
value for the first variable, we found that it not possible to have a policy that favors
agriculture without also favoring fis ing and vice versa.
Proceeding, a very high value from the available range of fis ing is chosen, and
it corresponds to very low values of the mining sector. By contrast, in Fig. 10.38 we
see that a low value in fis ing yields high values for the mining sector. This inverse
correlation was examined, and it was found that the country in question has a large
number of migrating semiskilled workers. When the fishing industry is doing well,
418 10. ♣ FT Data Mining and Other Applications

most of them are attracted to it, leaving few available to work in the mines and
vice versa. The comparison between the two figures shows the competition for the
same resource between mining and fis ing. It is especially instructive to discover
this interactively. The construction of the interior point proceeds in the same way.
In the next section, in the discussion on surfaces this construction is shown for
higher-dimensional hypersurfaces.

10.5 Parallel Coordinates: The Bare Essentials

The following short review of -coords and the discussion on duality provide the
essential background to make this chapter self-contained.

10.5.1 Lines
An N-dimensional line can be described by the N − 1 linear equations
⎧
⎪
⎪ : x2 = m2 x1 + b2 ,
⎪ 1,2
⎪
⎪
⎪ 2,3 : x3 = m3 x2 + b3 ,
⎪
⎪
⎪
⎨ ···
: (10.5)
⎪
⎪ i−1,i : xi = mi xi−1 + bi ,
⎪
⎪
⎪
⎪ ···
⎪
⎪
⎪
⎩
N−1,N : xN = mN xN −1 + bN ,

each with a pair of adjacently indexed variables. In the xi−1 xi plane, the relation
labeled i−1,i , N = 2, . . . , N, is a line, and by the line ↔ point duality (10.3), it
can be represented by the point

1 b
¯i−1,i =
i
+ (i − 2), . (10.6)
(1 − mi ) (1 − mi )

Here the interaxis distance is 1, so that i − 2 is distance between the y (or X̄1 ) and
X̄i−1 axes. Actually, any N − 1 independent equations such as

i,j : xi = mi,j xj + bi,j , (10.7)

can equivalently specify the line , for (10.7) is the projection of on the xi xj 2-D
plane, and N − 1 such independent projections completely describe . There is a
beautiful and very important relationship illustrated in Fig. 10.40 (left). For a line
in 3-D, the three points ¯12 , ¯13 , ¯23 are collinear (this line is denoted by L̄), and
 10.5 Parallel Coordinates: The Bare Essentials 419

Figure 10.40. Properties of multidimensional lines.

(Left) The three points ¯i,j , ¯j,k , ¯i,k are collinear for i = j = k. (Right) A line interval in 10-D.

any two points represent . It is easy to see that a polygonal line on all N − 1 points
given by (10.6) or their equivalent represents a point on the line . Conversely, two
points determine a line . Starting with the two polygonal lines representing the
points, the N − 1 intersections of their X̄i−1 , X̄i portions are the ¯i−1,i points for
the line . A line interval in 10-D and several of its points are seen on the (right) of
Fig. 10.40. By the way, the indexing of the points ¯ is essential. It is usually omitted
to save display space but it must always be available for it is needed for algorithms
that involve lines (e.g., intersections and minimum distance; see Chapter 4).

10.5.2 Planes and Hyperplanes

While a line can be determined from its projections, a plane even in 3-D cannot.
A new approach is called for [47]. Rather than discerning a p-dimensional object
from its points, it is described in terms of its (p−1)-dimensional subsets constructed
from the points. Let us see how this works. In Fig. 10.41 (left), polygonal lines
representing a set of coplanar points in 3-D are seen. From this picture even the
most persistent pattern-seeker cannot detect any clues hinting at a relation among
the three variables, much less a linear one. The plane has dimension p = 2, so
we look at lines (having dimension p − 1 = 1) on the plane constructed so that
from each pair of polygonal lines, the lines L̄ of the 3-point collinearity shown in
Fig. 10.40 (left) are obtained. The result, shown on the right, is stunning. All the
L̄ lines intersect at a point, which turns out to be characteristic of coplanarity but
not enough to specify the plane. Translating the first axis X̄1 to the position X̄1
,
one unit to the right of the X̄3 axis, and repeating the construction, based on the
420 10. ♣ FT Data Mining and Other Applications

Figure 10.41. Coplanarity.

(Left) The polygonal lines on the first three axes represent a set of coplanar points in 3-D. (Right)
Coplanarity! Forming lines on the plane, with the 3-point collinearity, the resulting lines intersect at
a point.

axis triple X̄2 , X̄3 , X̄1

, yields a second point shown in Fig. 10.42 (left). For a plane
described by

π : c1 x1 + c2 x2 + c3 x3 = c0 , (10.8)

the two points, in the order they are constructed, are respectively
 
c2 + 2c3 c0 3c1 + c2 + 2c3 c0
π̄123 = , , π̄1 23 =

, , (10.9)
S S S S

for S = c1 +c2 +c3 . Three subscripts correspond to the three variables appearing in
the plane’s equation and the axis triple used for their construction, and distinguish
them from the points with two subscripts, representing lines. The second and third
axes can also be consecutively translated, as indicated in Fig. 10.41 (left), repeating
the construction to generate two more points, denoted by π̄1
2
3 , π̄1
2
3
. These points
can also be found in an easier way (see Chapter 5). The gist of all this is shown in
Fig. 10.42 (right). The distance between successive points is 3ci . The equation of
the plane π can actually be read from the picture!
In general, a hyperplane in N dimensions is represented uniquely by N − 1
points, each with N indices. There is an algorithm that constructs these points
recursively, raising the dimensionality by one at each step, as is done here starting
from points (0-dimensional) and constructing lines (1-dimensional). By the way,
all the nice higher-dimensional projective dualities such as point ↔ hyperplane
 10.5 Parallel Coordinates: The Bare Essentials 421

Figure 10.42. Plane representation.

(Left) The two points where the lines intersect uniquely determine a plane π in 3-D. (Right)
From four points, similarly constructed by consecutive axis translation, the coefficients of
π : c1 x1 + c2 x2 + c3 x3 = c0 , where the coefficients are normalized (c1 + c2 + c3 = 1), can
be read from the picture!

and rotation ↔ translation, hold. Further, a multidimensional object represented in

-coords can still be recognized after it has been acted on by projective transforma-
tion (i.e., translation, rotation, scaling, and perspective). The recursive construction
and its properties are at the heart of the -coords method for the visualization of
multivariate relations.

Challenge: Visualizing Families of Proximate Planes

Returning to 3-D, it turns out that for points as in Fig. 10.41 that are “nearly”
coplanar (i.e., have small errors), the construction produces a pattern very similar
to that in Fig. 10.42 (left). A little experiment is in order. Let us return to the family
of proximate (i.e., close) planes generated by
 = {π : c1 x1 + c2 x2 + c3 x3 = c0 , ci ∈ [ci− , ci+ ], i = 0, 1, 2, 3}, (10.10)
randomly chosing values of the ci within the allowed intervals to determine several
planes π ∈ , keeping at first c0 = 1, and plotting the two points π̄123 , π̄1
23 as
shown in Fig. 10.43 (left). Not only is closeness apparent, but more significantly, the
distribution of the points is not chaotic. The outline of two hexagonal patterns can
be discerned. The family of “close” planes is visualizable, but also the variations in
several directions. It is possible to see, estimate, and compare errors or proximity.
422 10. ♣ FT Data Mining and Other Applications

Figure 10.43. A family of close planes.

(Left) A pair of point clusters representing close planes. (Right) The hexagonal regions (interior)
are the regions containing the points π̄123 (left) and π̄1
23 for the family of planes with c0 = 1 and
c1 ∈ [1/3, 1.5], c2 ∈ [1/3, 2.5], c3 ∈ [1/3, 1]. For c0 varying, here c0 ∈ [0.85, 1.15], the regions
(exterior) are octagonal with two vertical edges.

It can be proved that in 3-D, the set of pairs of points representing the family of
proximate planes forms two convex hexagons when c0 = 1, with an example shown
in Fig. 10.43 (right), and are contained in octagons, each with two vertical edges
for varying c0 . In general, a family of proximate hyperplanes in N-D is represented
by N − 1 convex 2N-gons when c0 = 1, or 2(N + 1)-gons for c0 varying (for an
extensive exposition, see Chapter 8). These polygonal regions can be constructed
with O(N) computational complexity. Choosing a point in one of the polygonal
regions, an algorithm matches the remaining possible N − 2 points, one each from
the remaining convex polygons, which represent and identify hyperplanes in the
family by N − 1 points.
We pose the thesis that visualization is not about seeing many things but
rather discovering relations among them. While the display of randomly sam-
pled points from a family of proximate hyperplanes is utterly chaotic (the mess
in Fig. 10.41 (right) from points in just one plane), their proximate coplanarity
relation corresponds to a clear and compact pattern. With -coords we can focus
and concentrate the relational information rather than wallowing in the details,
ergo the remark “without loss of information” when referring to -coords. This
is the methodology’s real strength and where the future lies. Here, then, is
a visualization challenge: how else can proximate coplanarity be detected and
seen?
 10.5 Parallel Coordinates: The Bare Essentials 423

10.5.3 Nonlinear Multivariate Relations: Hypersurfaces

A relation among two real variables is represented geometrically by a unique region
in 2-D. Analogously, a relation between N variables corresponds to a hypersurface
in N -D, hence the need to say something about the representation of hypersurfaces
in -coords. A smooth surface in 3-D (and also N-D) can be described as the
envelope (discussed in Chapter 6) of all its tangent planes. This is the basis for
the representation shown in Fig. 10.44 (left). Every point of the surface is mapped
into the two points representing its tangent plane at the point. This generates two
planar regions, and for N-D there are N − 1 such regions. These regions are linked,
just like the polygons above, to provide the proper N − 1 points representing each
tangent hyperplane and from which the hypersurface can be reconstructed. Classes
of surfaces can be immediately distinguished from their -coords display (see the
chapter on surfaces for extensive treatment). For developable surfaces the regions
consist of boundary curves only with no interior points; regions for ruled surfaces
have grids consisting of straight lines; quadric surfaces have regions with conic
boundaries. These are some examples.
There is a simpler but inexact surface representation that is quite useful when
used judiciously. The polygonal lines representing points on the boundary are
plotted and their envelope “represents” the surface; the “ ” are a reminder that
this is not a unique representation. In Fig. 10.45 (left) are the upper and lower
envelopes for a sphere in 5-D consisting of four overlapping hyperbolas, which must
be distinguished from those in Fig. 10.44 (right), which is exact and interestingly

Figure 10.44. Surface representation.

(Left) A smooth surface σ is represented by two planar regions σ̄123 , σ̄231
consisting of pairs of points
representing its tangent planes. (Right) One of the two hyperbolic regions representing a sphere in
3-D.
424 10. ♣ FT Data Mining and Other Applications

Figure 10.45. Interior point construction.

(Left) A sphere in 5-D showing the construction of an interior point (polygonal line). (Right) The
general interior point (polygonal line) construction algorithm shown for a convex hypersurface in
20-D.
 10.6 Future 425

enough, are also hyperbolas, the curves determined by points representing the
sphere’s tangent planes. Retaining the exact surface description (i.e., its equation)
internally, interior points can be constructed and displayed as shown for the 5-D
sphere in Fig. 10.45 (left). On the right the same construction is shown, but for
a more complex 20-dimensional convex hypersurface (“model”). The intermedi-
ate curves (upper and lower) also provide valuable information and previews of
coming attractions. They indicate a neighborhood of the point (represented by the
polygonal line) and provide a feel for the local curvature. Note the narrow strips
around X13, X14, X15 (as compared to the surrounding ones), indicating that at
this state, these are the critical variables where the point is bumping the boundary.
A theorem guarantees that a polygonal line that is between all the intermediate
curves/envelopes represents an interior point of the hypersurface, and all interior
points can be found in this way. If the polygonal line is tangent at any one of the
intermediate curves, then it represents a boundary point, while if it crosses any one
of the intermediate curves, it represents an exterior point. The latter enables us to
see, in an application, the first variable for which the construction failed and what is
needed to make corrections. By varying the choice of value over the available range
of the variable interactively, sensitive regions (where small changes produce large
changes downstream) and other properties of the model can be easily discovered.
Once the construction of a point is completed, it is possible to vary the values of
each variable and see how this affects the remaining variables. So one can do trade-
off analysis in this way and provide a powerful tool for, decision support, process
control, and other applications. As new data becomes available the model can be
updated, with decisions being made based on the most recent information. This
algorithm is used in the earlier example on a model for a country’s economy shown
in Figs. 10.38, 10.39.

10.6 Future
Searching for patterns in a -coords display is what skillful exploration is about.
If there are multivariate relations in the data set, the patterns are there, though they
may be covered by the overlapping polygonal lines, and that is not all. Our vision
is not multidimensional. We do not perceive a room that is 3-dimensional from its
points, which are 0-dimensional, but from the 2-dimensional planes that enclose and
define it. The recursive construction algorithm does exactly that for the visualization
of p-dimensional objects from their (p − 1)-dimensional subsets, one dimension
less. We advocate including this algorithm within our armory of interactive analysis
tools. Whatever p-dimensional relations exist are revealed by the pattern from the
representation of the tangent hyperplanes of the corresponding hypersurface. The
426 10. ♣ FT Data Mining and Other Applications

Figure 10.46. A 3-D helicoid in Cartesian and -coords.

The two intersecting lines (left) specify one of the helicoid’s tangent planes represented by a pair of
points one on each of the curves (right). A helicoid in N -D is represented by N − 1 such curves.

Figure 10.47. A 3-D Möbius strip in Cartesian and -coords.

The two intersecting lines (left) specify one of the strip’s tangent planes represented by a pair of
points, one on each of the curves (right). A path on a Möbius strip in N -D is represented by N − 1
such curves. Also see figures in Chapter 9 on Surfaces.
 10.6 Future 427

polygonal lines are completely discarded, for the relation is concentrated in the
pattern: linear relations into points, proximate coplanarity into convex polygons,
quadrics into conics, and so on. Note further, again with reference to Figs. 10.40 and
10.41, that relational information resides at the crossings. What can be achieved for
the representation of complex relations by patterns is exemplified by the pictures of
a helicoid and Möbius strip in Cartesian and -coords in Figs. 10.46, 10.47. These
are state-of-the-art results, showing what is achievable and how easily it generalizes
to N-D. Can one imagine a higher-dimensional helicoid, much less a nonorientable
surface like the Möbius strip? It is possible to do such a process on a data set, though
at present it is computationally slow. The challenge is to speed up the algorithm for
real-time response and thus break the gridlock of multidimensional visualization.
There will still be work and fun for the multidimensional detectives visually separa-
ting and classifying the no longer hidden regions, identifying complex multivariate
relations that have evaded us until now.
This page intentionally left blank
11

Recent Results

11.1 Displaying Several Lines Efficiently

Shlomi Cohen-Ganor
Department of Computer Sciences
Tel Aviv University
[email protected]

11.1.1 Abstract
This chapter suggests ways to improve the visualization of several lines in parallel
coordinates, specifically how to overcome the inconvenience caused by the fact
that the indexed points are not necessarily arranged “neatly” along the horizontal
axis. Several approaches are presented by which the input data (or alternatively, the
axes) may be transformed so as to reduce or eliminate the problem, with minimal
information loss.

11.1.2 Introduction
Recall that in order to visualize a line ∈ RN , we first represent it as a set of
equations of the following form:55
55
It is also possible to represent the linear relations between the components in different notation, for
example,
xi = mi x1 + bi
for i ∈ [2, . . . , N], where x1 corresponds to some “base” variable. Such representation may convey
important properties of the data, but the corresponding graphic representation inherently lacks the
“neatness” we are seeking in this chapter.

A. Inselberg, Parallel Coordinates, DOI 10.1007/978-0-387-68628-8_11,

© Springer Science+Business Media, LLC 2009
430 11. Recent Results

xi = mi xi−1 + bi

for i ∈ [2, . . . , N]. The line is then visualized in parallel coordinates by N − 1

indexed points as follows (in homogeneous coordinates):

¯i,i+1 : ((i − 2)(1 − mi ) + 1, bi , 1 − mi ).

If has constant components, they are first plotted independently, using any
convenient graphic notation; then the above approach is used for the nonconstant
components.
Some basic properties of the indexed points:
• For any pair of points on , the lines connecting their xi and xi+1 components
intersect at ¯i,i+1 .
• The horizontal position of a point ¯i,i+1 is determined by the corresponding
slope mi .
• The vertical position of a point is determined by the slope as well as the
corresponding displacement bi .
Let us refine the relation between the slope mi and the horizontal position of ¯i,i+1 :
• If mi < 0, the point lies between the xi and xi+1 axes.
• If mi = 0, the point lies on the xi+1 axis.
• If 0 < mi < 1, the point lies to the right of the xi+1 axis.
• If mi = 1, the point is ideal.
• If mi > 1, the point lies to the left of the xi axis.
Several problems are apparent:
1. A visualization of ideal points is required for some lines.
2. The indexed point ¯i,i+1 may fall anywhere in the 2D display, regardless of
i or of any other indexed point. Indeed, different indexed points may fall on
the same point in 2D. In other words, there is no visual separation between
dimension components.
3. If we wish to plot multiple lines in this manner, the display quickly becomes
very confusing.
4. There is a discontinuity along the horizontal display axis as the slope approaches
1: A line with slope 1 + ε would appear at the extreme left side, while a line
with slope 1 − ε would appear at the extreme right side.
These problems are demonstrated in Figures 11.1 and 11.2. Different ways to
overcome these problems are suggested in what follows.
 11.1 Displaying Several Lines Efficiently 431

Figure 11.1. A line in R9 (9-dimensional Euclidean space).

Note that the first four indexed points are easy to find and understand, while the last four are “mixed
up” and confusing. This is because the slopes corresponding to the latter points are positive.

11.1.3 General Approach

Apparently, the easiest way to improve the representation of lines is to invent a new
and simpler way of plotting the line data. Specifically, one could easily map the pair
(mi , bi ) onto the area bounded by the axes xi and xi+1 using any continuous and
one-to-one function f : R2 → (0, 1] × R. This approach would avoid the problems
mentioned.
However, an arbitrary choice of the function f would also lose many important
properties of the indexed point approach, perhaps most importantly the ability to
detect n-flats from a set of points. We would like to construct a mapping that is
essentially the same as that described in the previous section, but without having
indexed points outside the area bounded by their corresponding axes. As mentioned,
this occurs naturally when the slope mi is negative, indicating that as one variable
(dimension) increases, the other decreases and vice versa. Let us seek ways to
“make the slope negative.”
432 11. Recent Results

Figure 11.2. Five randomly chosen lines in R9 . The display is very hard to understand.

11.1.4 Flipping Axes

As described in an earlier chapter, the up–down directions of the axes are arbitrary,
and may be flipped for display purposes, as long as we provide some indication of
which vertical direction is positive and which is negative. Flipping an axis means
negating the corresponding dimension of the input data, so the slopes from and to
that dimension are also negated.
This idea leads to the “axis flipping” algorithm for one line, described in
Chapter 4. Let us recall the essential steps:
• For each i ∈ [2, N ]:
– if mi > 0 flip axis i, negate mi , bi , and mi+1
– plot ¯i,i+1
This works perfectly for a single line, as shown in Figure 11.3. What if we
have multiple lines? We could run the algorithm separately for each line, and if
we were lucky, the axes would be flipped the same way for all of them, in which
case we would have a solution. However, this is not generally the case. Let us
investigate the possibilities, first introducing some simple definitions.
 11.1 Displaying Several Lines Efficiently 433

Figure 11.3. Displaying the line shown in Figure 11.1 after flipping axes x6 and x9 .

Agreement and Disagreement of Lines and Dimensions

Let us consider a set L of lines in RN . For each pair of dimensions (i, j ) ∈ N × N,
and for each line ∈ L, we can find the slope m ij of the line’s projection on the
xi , xj plane.
Definition We say that lines , k ∈ L agree on (i, j ) if m ij mkij > 0 (meaning
m ij and mkij have the same sign).
Intuitively, if each line describes a linear relation between the variables, then
the direction of change of j as i increases/decreases is the same in both relations.
The agreement relation of lines has the following properties:
434 11. Recent Results

1. For any given (i, j ) ∈ N × N, line agreement is an equivalence relation

on L.
2. Flipping any axis does not change line agreement (if i or j is flipped, both mlij
and mkij are negated).
3. If lines , k agree on i, j , then the indexed point ¯i,j is between the axes i, j if
and only if the indexed point k̄i,j is between the axes i, j .
Recall that we would like all indexed points to be between the axes, and we can
achieve this for any one line by flipping axes. It is evident that if all lines in L agree
on i, j , we can simultaneously get all the indexed points corresponding to i, j where
we want them.
Definition We say that dimensions i, j agree on L if all lines in L agree on them.
The agreement relation of dimensions has the following properties:
1. If dimensions i, j agree on L then for any line ∈ L, ¯i,j is between the axes
i, j if and only if all the indexed points corresponding to i, j are between the
axes i, j . This derives directly from property 3 of line agreement.
2. If dimensions i, j do not agree on L, there is at least one line such that ¯i,j
is not between the axes i, j (Otherwise, all lines agree on i, j by property 3 of
line agreement, and so the dimensions agree by definition).
3. Flipping any axis does not change dimensions agreement (since it does not
change lines agreement).
This proves the following conclusion:
• If every pair of adjacent dimensions i, j agree for L, then running the “axis
flipping” algorithm on any line in L yields a solution for all lines in L.
• Otherwise, no solution based on flipping exists.
One may wonder whether the order of the axes is important, since the conclusion
only mentions adjacent dimensions. Perhaps we could try running the algorithm on
different permutations of the axes. However, all of the properties presented above
hold for any pair of dimensions and not just adjacent ones. Furthermore, let us
redefine agreement of lines as a relation between dimensions:
Definition We say that dimensions (i, j ) ∈ N × N agree on lines , k ∈ L if
m ij mkij > 0 (meaning m ij and mkij have the same sign).
This is an equivalence relation on N × N for any , k ∈ L. Specifically, it is
transitive, since the sign of m ij is determined by the signs of m id and m dj for any
dimension d.
 11.1 Displaying Several Lines Efficiently 435

It follows that dimension agreement is also transitive and therefore is also an

equivalence relation on N × N. So, we can restate our conclusion more generally:

• If all dimensions agree for L, then running the axis flipping algorithm on any
line in L, for any permutation of the axes, will give a solution for all lines in L.
• Otherwise, no solution based on flipping or permuting the axes exists.
This suggests a method for finding a partial solution in the general case (not
necessarily the best partial solution, but a simple one nonetheless). We can easily
find the equivalence classes of the dimensions and reorder the axes by the classes.
Running the “axis flipping” algorithm on these axes (on any line) will give a solution
such that indexed points corresponding to dimensions from the same class will be
between their appropriate axes. This solution maximizes the number of dimensions
without any “bad” indexed points, though not necessarily the total number of “bad”
indexed points.
To conclude this section, we have shown that the simple “axis flipping” algo-
rithm finds a full solution if one exists, or a partial one otherwise. Reordering the
axes may improve the partial solution, but will never uncover a full solution. The
basic algorithm is of order O(N), and checking that the solution is correct for all
lines is of order O(L × N), which is the lower bound for the problem. We have
also shown a necessary and sufficient criterion for the existence of a solution.
So far we have allowed flipping axes and reordering them. Next we will
investigate other operations that can be used.

11.1.5 Rotating Axes

We have proved that the flipping/reordering approach is sufficient if all dimensions
agree on the set of lines, meaning that for each pair of lines l, k and each pair of
dimensions i, j , the slopes mlij and mkij have the same sign. For brevity we will
simply say that the lines agree (on any pair of dimensions), which means the same
thing by definition.
Let us find a geometric interpretation of this criterion. We are interested only
in slopes, so let us translate all lines so that they pass through the origin. Ignore
lines parallel to one of the Cartesian axes (see note on constant components in the
introduction).
In 2D, agreement of lines means that all lines after translation are contained in
exactly two quadrants. More specifically, either they all pass through quadrants I
and III, or through quadrants II and IV; see Figure 11.4.
This seems somewhat arbitrary. Consider a group of lines in 2D that have some
positive and some negative slopes, but all have a very small absolute slope (as shown
436 11. Recent Results

Figure 11.4. Agreeing lines in 2D.

in Figure 11.5). These lines do not agree, but they can be made to agree by slightly
rotating the axes.
More generally, a set of lines in 2D can be made to agree by rotation if and
only if they can be bounded between two orthogonal lines. An appropriate rotation
angle can be found by taking the angle between one of the bounding lines and one
of the axes.
Checking this new criterion is easy, because if a pair of bounding lines exists,
then a pair exists such that the slope of one of the lines is just below56 one of the
given lines. We can look for solutions as follows:

• For each given line, position a bounding line just below it.
• Construct the orthogonal bounding line and count the given lines bounded
between them.
• Choose the bounds that contain the most given lines.

56
“Just below” means “smaller by some infinitesimal value.”
 11.1 Displaying Several Lines Efficiently 437

Figure 11.5. Disagreeing lines in 2D.

The green lines have positive slopes, while the red lines have negative slopes. The diamonds show
one way to bound all lines between two orthogonal lines.

This method yields a rotation angle such that the greatest number of lines have
no “bad” indexed points at all. It also allows us to choose some lines that must not
have “bad” indexed points, or to assign priorities to the lines, etc.
It is worth noting that orthogonal lines never agree, and they cannot be made
to agree by any rotation (since they will always remain orthogonal).
All that was said is true for 2D. In 3D, the agreement criterion means that all
lines are contained in exactly two octants. We would like to find a 3D rotation, if
one exists, such that a given set of lines will be contained in exactly two octants
after the rotation.
As in 2D, a set of lines in 3D can be made to agree by rotation if and only if
they can be bounded between three orthogonal lines (this is not an exact definition).
In this case, an appropriate rotation matrix can be constructed by inverting the
matrix that maps the Cartesian axes onto unit vectors on the bounding lines.
Unlike the 2D case, it is not clear how one can find the bounding lines, or check
whether they exist. This question is left open.
The subject of rotations in 3D and higher dimensions will be revisited later on.
438 11. Recent Results

11.1.6 Transforming Axes

We have seen evidence that applying a rotation to the data is useful in some cases
for achieving a better visualization. While this may not be acceptable in some
applications, rotated data preserves many properties of the original data. Let us now
take a step further in this direction and allow a broader class of transformations.
Since we are interested in representing lines, we will require that the transformation
preserve lines. We do not want to lose any information, so we will also require that
the transformation be invertible. The class of invertible linear transformations seems
a natural direction to investigate.
Let us represent lines as vectors: for each given line, construct a vector from
the origin in the same direction as the line. For convenience and without loss
of generality, we will construct the vector so that the first nonzero component
is positive. A set of m lines in RN can now be represented by an m × n matrix V ,
where each column holds the vector representation of one line. The transformation
is represented by an n × n matrix T that is nonsingular (this guarantees an invertible
transformation).
For the lines to agree, the matrix TV should have the same sign for all compo-
nents in the same row. Without loss of generality, we will require the matrix TV to
have only positive cells. Geometrically, this is equivalent to requiring all lines to
pass through two specific quadrants/octants etc., which is more than we need, but
the notation becomes simpler and no genuine solutions are lost.
Now we must find a nonsingular T such that TV > 0 (cell by cell). Equivalently,
V T T T > 0 (cell by cell). To find T T we must find N linearly independent vectors
x ∈ R N satisfying V T x > 0 (cell by cell). These vectors are the rows of T . We will
now describe an algorithm for finding such vectors.
Recall that the first column of V T is nonnegative because we took vectors whose
first nonzero component is positive. For now, we assume that the first column is
positive:
• Assign x1 = e1
• Choose any ε > 0
• For j ∈ [2, . . . , N] do:
- ε−V T .
– Find tj = max T
ij
: 1 ≤ i ≤ m ∪ {0}
Vi1
– Assign xj = tj ∗ e1 + ej
Proof of correctness:
The assumption guarantees that V T x1 > 0 (cell by cell).
 11.1 Displaying Several Lines Efficiently 439

From the construction,


N ε − VijT
T T T
(V xj )i = Vik xkj = tj Vi1 + VijT ≥ T
T
Vi1 + VijT = ε > 0.
k=1 Vi1

It is clear that the matrix T T is upper triangular with determinant 1, so T is

nonsingular.
If the assumption on the first column does not hold, i.e., there are zeros, it is
always possible to find a rotation matrix R such that (RV)T satisfies the assumption
(this can be shown by constructing an infinitesimal rotation such that the zeros
become nonzeros, but the other cells remain nonzeros). The rotation can then be
considered part of the linear transformation. Alternatively, if some other column
does not contain zeros, it can be made the first column by reordering the dimensions,
and it can be made positive by reversing some of the unit vectors as necessary.
We have found a solution for any given set of lines! In conclusion, if invertible
linear transformations of the data are allowed, we can transform any set of lines
into agreeing lines, which can then be perfectly visualized by the “axis flipping”

Figure 11.6. Five randomly chosen lines in R9 (one ideal point is not displayed).
440 11. Recent Results

algorithm. Actually, we could modify the matrix so that no flipping is required at

all, since this is just a special case of invertible linear transformations.
The method will be exemplified using five random lines in R9 , shown in
Figure 11.6.
The corresponding vector matrix is
⎛ ⎞
.51 .19 .34 .49 .12
⎜ −.38 −.62 −.18
⎜ .12 −.04 ⎟ ⎟
⎜ −.19 .14 −.22 .29 .51 ⎟
⎜ ⎟
⎜ −.44 −.34
⎜ .37 −.52 −.53 ⎟ ⎟
V =⎜ ⎜ .16 −.07 .65 −.57 .43 ⎟
⎟.
⎜ .59
⎜ .44 −.18 −.25 −.17 ⎟ ⎟
⎜ .05
⎜ .8 0.0 .04 −.44 ⎟ ⎟
⎝ −.92 .26 .07 −.17 .25 ⎠
−.09 .22 .29 −.52 −.06
Applying the construction algorithm with ε = 0.5 yields
⎛ ⎞
1 0 0 0 0 0 0 0 0
⎜ 5.8947 1 0 0 0 0 0 0 0 ⎟
⎜ ⎟
⎜ 2.1176 0 1 0 0 0 0 0 0 ⎟
⎜ ⎟
⎜ 8.5833 0 0 1 0 0 0 0 0 ⎟
⎜ ⎟
T =⎜ ⎟.
⎜ 3.0 0 0 0 1 0 0 0 0 ⎟
⎜ 5.5833 0 0 0 0 1 0 0 0 ⎟
⎜ ⎟
⎜ 7.8333 0 0 0 0 0 1 0 0 ⎟
⎜ ⎟
⎝ 2.7843 0 0 0 0 0 0 1 0 ⎠
4.6667 0 0 0 0 0 0 0 1
And the transformed vectors are
⎛ ⎞
.51 .19 .34 .49 .12
⎜ 2.6263 0.5 1.8242 3.0084 0.6674 ⎟
⎜ ⎟
⎜ 0.89 0.5424 0.5 1.3276 0.7641 ⎟
⎜ ⎟
⎜ 3.9375 1.2908 3.2883 3.6858 0.5 ⎟
⎜ ⎟
V =⎜ ⎟.
⎜ 1.69 0.5 1.67 0.9 0.79 ⎟
⎜ 3.4375 1.5008 1.7183 2.4858 0.5 ⎟
⎜ ⎟
⎜ 4.045 2.2883 2.6633 3.8783 0.5 ⎟
⎜ ⎟
⎝ 0.5 0.789 1.0167 1.1943 0.5841 ⎠
2.29 1.1067 1.8767 1.7667 0.5
Indeed, all cells are positive, as expected. Notice that the first row remains
unchanged, and that every other row contains ε at least once. A simple way to
 11.1 Displaying Several Lines Efficiently 441

Figure 11.7. The same five lines after transformation. The visual display is much clearer than
before.

get the transformed lines is to choose a pair of points on each line, apply the
transformation T to the points, and construct the lines passing through the points
correspondingly. The result can be seen in Figure 11.7. The visual display is much
clearer than before.
We can do even better by noticing that the indexes are now redundant: all
points between axes Xi , Xi+1 have indexes i, i + 1. The final result is shown in
Figure 11.8.

11.1.7 Back to Rotating Axes

Now that we have found a general solution, we would still like to find a solution
based only on rotation and flipping, since this preserves more properties of the data
(particularly angles).
442 11. Recent Results

Figure 11.8. The same five lines after transformation, without the indexes.

In terms of linear algebra, we still want to find N linearly independent vectors

x ∈ RN satisfying V T x > 0 cell by cell, but we also want them to be orthogonal.57
This would guarantee that the resulting transformation matrix T is not only inver-
tible but also orthogonal, which is the same as saying that it is a rotation matrix.
Geometrically, the criterion V T x > 0 defines an open polyhedron. We know that
the hyperplanes defining it all pass through the origin, so the origin must be an
extreme point. We also know that the polyhedron is unbounded; multiplying a
point inside the polyhedron by any positive number yields another point inside
the polyhedron. We are seeking a set of N points inside the polyhedron such that
the lines connecting them with the origin are orthogonal. Such a set exists if and
only if it is possible to place an N-cube so that one of the cube’s vertices is at
the origin and all other vertices are inside the polyhedron. This problem is left
open.

57
More precisely, we want them to be orthonormal, but the additional constraint does not affect the
problem significantly.
 11.2 Separating Point Clusters on Different Planes 443

11.2 Separating Point Clusters on Different Planes

Nir Shahaf
School of Mathematical Sciences
Tel Aviv University, Israel
[email protected]

11.2.1 Overview
We start by describing a family of problems. Suppose that we have m planes in RN
and that we have a sample of n random points taken from each of those planes. Our
purpose is to be able to recover the planes from the sample data (without knowing
which sample point came from which plane). This family of problems depends on
two parameters: m, the number of planes, and N, the dimension. We solve the
problem in a certain setting where m = 3 and N = 3 (extension to other problems
from the family is immediate). We show a probabilistic algorithm that solves the
problem with high probabilty; the complexity of that algorithm is O(n2 ).

11.2.2 The Algorithm

• All possible pairs of the input points are taken, and for each pair (except for the
special case that the polygonal lines representing the two points are parallel),
the line joining the two points is found (represented by two points).
• The input for the next step is the lines found in the previous one. The lines are
randomly divided into pairs; this is the only random element in the algorithm.
For each pair, the intersection point is taken. If the four points from which the
two lines originated came from the same plane, the found point will be the point
representing the intersection line between the plane and the first superplane.
We refer to those points as “superlines”.
• In this step we look for clusters of points. All the “superlines” that were created
by points coming from a mutual plane will consolidate on the point representing
the intersection line between the plane and the first superplane. At the beginning,
our workspace is a portion of the plane containing all the “superlines.” At each
step of the computation, the workspace is divided into 100 equally sized portions.
For each portion, the number of “superlines” in it is counted. For each portion
containing a number of points that is larger than some threshold (which is
determined in a way described later), we recursively look for consolidating
points in it (taking the portion as the new workspace). Once the total number
of points in the workspace does not change for a certain number of recursion
444 11. Recent Results

Figure 11.9. Sample run of the algorithm. “Interesting” rectangles are shown.

Figure 11.10. A cluster of 21 points is found.

steps (depending on the precision of the calculations), we declare that a point

has been found. In this way, three points are found, one for each plane.
The process is performed four times, once for each of the coordinate systems:
(X1 , X2 , X3 ), (X2 , X3 , X1
), (X3 , X1
, X2
), and eventually (X1
, X2
, X3
).
• After the previous steps we have the four intersection points of the planes with
the four first superplanes; after sorting the point triplets, the planes’ equations
can be easily found.

One question remains open: how do we determine the best threshold? For
obvious efficiency reasons, we would like to choose a threshold that is as high as
possible. However, choosing too high a threshold might cause us to miss some of
 11.2 Separating Point Clusters on Different Planes 445

the intersection clusters. The next subsection deals with the question of determining
a threshold that is not too high and not too low.

11.2.3 Mean and Variance Estimation

Let n be the number of random points taken on each of the three planes. The total
number of points is 3n. In the first phase of the computation we have taken all
possible pairs; each pair yielded a line. Denote the number of lines by α:

3n 3n(3n − 1)
α= = .
2 2
Suppose that π is one of the three planes. Since we have taken n points from each
plane, the number of lines on the plane π (out of the lines created by taking all pairs
of points) would be

n n(n − 1)
β= = .
2 2
In the second phase of the computation we have paired all the α lines and created
α “superlines.” Let Y , Y , . . . , Y be random variables; Y indicates the contri-
2 1 2 α i
bution of the ith line to the total number of superlines consolidating on the point
representing the plane π. If the ith line was not on π, then Yi ≡ 0 (since this line
is not on π, in any possible pairing, a superline that was created by this line would
not be the point representing π). We can therefore observe only the subset of those
random variables corresponding to the β lines on the plane π. Indicate them by
X1 , X2 , . . . , Xβ . Observing the ith line, this line will contribute 21 to the number
of consolidating superlines (its pair would contribute the other 21 , and together they
contribute their mutual superline) if and only if its pair is one of the β lines as well.
After choosing the first element of the pair, there are α − 1 elements left in all, and
 
β − 1 “good” elements left. Therefore P Xi = 21 = β−1 α−1 . Denote that probability
by p:
⎧
⎨ 1 with probability p,
2
Xi =
⎩0 with probability 1 − p.

Let X be a random variable indicating the total number of superlines consolidating

β
on the point representing the plane π. Then X = i=1 Xi . Since EXi = p2 , we
have that

β 
β 
β
p pβ
EX = E Xi = EXi = = .
i=1 i=1 i=1
2 2
446 11. Recent Results

In order to calculate the variance of X, we use the fact that

 β 
 β 
Var(X) = Var Xi = Var(Xi ) + 2 cov(Xi , Xj ).
i=1 i=1 1≤i<j ≤β

Since 2Xi is Bernoulli distributed, we have that

β −1 α−β
4Var(Xi ) = Var(2Xi ) = p · (1 − p) = · ,
α−1 α−1
(α − β)(β − 1)
Var(Xi ) = .
4(α − 1)2
For i = j we have that cov(Xi , Xj ) = E(Xi Xj ) − E(Xi )E(Xj ), and therefore we
need to find the distribution of Xi Xj . Since both Xi and Xj have only two possible
values (0 and 21 ), Xi Xj has the same property and may have only the following
two values, 0 and 41 :
 
1 1 1
P Xi Xj = = P Xi = , Xj =
4 2 2
 
1 1  1
= P Xi = · P X j = Xi = .
2 2 2
    
We already know that P Xi = 21 = p. We need to evaluate P Xj = 21 Xi = 21 .
Assuming that the ith line was paired with a “good” line, then the j th line will be
paired with a “good” line as well if and only if one of the following occurs: the j th
line was paired with the ith line, or it was paired with a different “good” line. The
1 (we want to choose a specific element out of
probability for the first event is α−1
the α − 1 available elements left). The probability for the second is β−3α−3 (we have
already excluded three elements: the ith line, its pair, and the j th line). Therefore
 
1 1 β −3
P Xi Xj = =p· +
4 α−1 α−3
β − 1 (α − 3) + (β − 3)(α − 1)
= ·
α−1 (α − 1)(α − 3)
β − 1 α − 3 + αβ − β − 3α + 3
= ·
α−1 (α − 1)(α − 3)
(β − 1)(αβ − β − 2α)
= ,
(α − 1)2 (α − 3)
 11.2 Separating Point Clusters on Different Planes 447

and therefore
(β − 1)(αβ − β − 2α)
E(Xi Xj ) = .
4 · (α − 1)2 (α − 3)
Obviously,

p2 (β − 1)2
E(Xi ) · E(Xj ) = 2 = .
2 4 · (α − 1)2
Hence

cov(Xi , Xj )

= E(Xi Xj ) − E(Xi ) · E(Xj )

(β − 1)(αβ − β − 2α) (β − 1)2

= −
4 · (α − 1)2 (α − 3) 4 · (α − 1)2

αβ 2 − β 2 − 2αβ − αβ + β + 2α − (β − 1)2 (α − 3)
=
4 · (α − 1)2 (α − 3)

αβ 2 − β 2 − 2αβ − αβ + β + 2α − (β 2 − 2β + 1)(α − 3)
=
4 · (α − 1)2 (α − 3)

αβ 2 − β 2 − 2αβ − αβ + β + 2α − αβ 2 + 3β 2 + 2αβ − 6β − α + 3
=
4 · (α − 1)2 (α − 3)

2β 2 − αβ − 5β + α + 3
= ,
4 · (α − 1)2 (α − 3)
and eventually,


β 
Var(X) = Var(Xi ) + 2 cov(Xi , Xj )
i=1 1≤i<j ≤β

β(α − β)(β − 1) β(β − 1) 2β 2 − αβ − 5β + α + 3

= + 2 · · .
4(α − 1)2 2 4 · (α − 1)2 (α − 3)

Corollary 11.2.1. Let X be a random variable indicating the number of superlines

consolidating on the point representing the intersection between the plane π and a
superplane. Then
448 11. Recent Results

pβ
EX = , (11.1)
2
β(α − β)(β − 1) β(β − 1) 2β 2 − αβ − 5β + α + 3
Var(X) = + 2 · · . (11.2)
4(α − 1)2 2 4 · (α − 1)2 (α − 3)

11.2.4 Error Probability Approximation

In the second phase of the computation, the pairing phase, we pair the α lines
to obtain α2 “superlines.” Let X be the random variable defined in the previous
subsection. We are now interested in the following question: what is the probability
that X is exactly k?
A different approach can be taken in order to describe the pairing of the α lines.
Let γ be the number of all possible pairs of lines taken from the α lines:

α α(α − 1)
γ = = .
2 2
Let δ be the number of all possible “good” pairs, where both of the lines come from
the “good” β elements:

β β(β − 1)
δ= = .
2 2

Consider the following process: for α2 rounds, we pick pairs from the γ possible
pairs. Our choice might be impossible, since two distinct pairs we choose might
have a line in common (that would not be a legal pairing). However, every possible
pairing can be described as a choice of its pairs out of all the possible pairs. Let Y
be a random variable indicating the number of “good” pairs chosen (out of the δ
“good” pairs). We indicate the sample space by  and define the following events:

A = {ω ∈  | ω is a legal pairing},

B = {ω ∈  | X(ω) = k},

C = {ω ∈  | Y (ω) = k}.

Notice that when X = k, our choice is a legal pairing, and in addition, we have
chosen exactly k “good” pairs: B = A ∩ C. Therefore: B  C.
Corollary 11.2.2. For X and Y define as above,

P (X = k) < P (Y = k).
 11.2 Separating Point Clusters on Different Planes 449

In order to obtain a bound on the probability that X = k, we simply need to estimate

the probability that Y = k. The process we have described is actually taking α2
samples from a population of γ elements, of which δ are considered “good.” Then
Y is exactly k if when choosing the α2 elements, we choose exactly k “good” ones
and α2 − k “bad” ones (hypergeometric distribution):
δ  γ −δ 
k α
−k
P (Y = k) =  γ2  .
α
2
√
If we take our threshold to be EX − 4σ , where σ = Var(X), then for two
sample cases (where n = 25 and n = 100), the bound probabilities are listed in
Table 11.1 (see also Figures 11.11 and 11.12).
Considering the fact that those probabilities are just bounds, and that the real
ones are much lower (by eliminating all the illegal pairings), the method would

Table 11.1. Error probability bounds.

√
Points Per Plane EX Var(X) EX − 4σ P (Y < EX − 4σ )

25 16.16798 2.55499 5.94799 0.0011748

100 273.11144 10.47581 231.20816 0.004782

Figure 11.11. P (Y = k), 25 points per plane.

450 11. Recent Results

Figure 11.12. P (Y = k), 100 points per plane.

almost never miss the points. Even if it does miss, executing the process once more
will succeed with even higher probability.

11.2.5 Complexity
In this subsection we will estimate the complexity of the process described in the
previous subsections. In the described algorithm, we first take all possible pairs of
the n points. We have α(n) of those, which is of order O(n2 ). At the next stage
we randomly devide the α(n) lines into pairs, find their intersection point, and get
l(n) := α(n)
2 “superlines.” Obviously, the complexity of that stage is of order O(n )
2
as well. The final stage, the recursive consolidation points lookup, requires a more
delicate analysis. Denote by M(n) the following constant:
l(n)
M(n) = ,
t (n)
where t (n) is the threshold chosen as described in the previous subsections
(EX − 4σ ). This constant is a bound on the maximal number of distinct portions of
the workspace that can contain a number of points that is more than the threshold
at the same time. However, for our purposes, we need a bound that does not depend
on n. Fortunately, this is the case. A short computation shows that the sequence
M(n) is strictly decreasing; hence there exists a constant M such that for all n,

M(n) < M.
 11.2 Surface Representation Decomposition and Developable Quadrics 451

Another interesting point (which will not be used here) is that sequence actually
converges:
lim M(n) = 34 ,
n→∞

where the number 3 in the limit is due to the number of planes we started with.
During the recursion process, we take care of portions of the original workspace
that get smaller: first we start with 102 of them; on the second recursion step we
already have 104 of them, and so on. However, on every recursion level, there cannot
be more than M(n) portions that contain more than t (n) points. We now define
one more constant, C, which is dictated by the precision of the computation. The
constant C is the maximal depth of the recursion that still improves the computation.
This is a universal constant, which does not depend on n (it depends only on the
capabilities of the computer performing the calculations). Since on every level
we can inspect at most M(n) portions of the workspace, and since the maximal
recursion level is C, we have that the maximal number of totally inspected portions is
bounded by
C · M(n) < C · M.
Note that this is a very loose bound: it is obviously impossible to be simulta-
neously in M(n) portions in the Cth level. The meaning of that scenario would be
finding M(n) points, while we expect to find only 3. While processing a portion we
count a number of points that is less than O(n2 ) (we observe only the points that
belong to the portion). Therefore the complexity of the consolidation points lookup
is of order O(n2 ), and this is also the complexity of the entire algorithm, as claimed.

11.3 Surface Representation Decomposition and

Developable Quadrics
Chao-Kuei Hung
Department of Information Management
Chaoyang University of Technology
Taichung, Taiwan
[email protected]
In the parallel coordinate system, a smooth hypersurface σ in RN is identified
as the envelope of its tangent hyperplanes. Each of σ ’s tangent hyperplanes π
gives an indexed set of points (π̄12...N , π̄1
2...N , . . . , π̄1
2
...N ) in the projective
plane P2 . The hypersurface σ is thus represented as an indexed set of regions
(σ̄12...N , σ̄1
2...N , . . . , σ̄1
2
...N ) composed of these points.
452 11. Recent Results

In this paper, we show that this representation is the composition of an extended

version of projective duality and a linear transformation. For a quadric, the com-
putations can be stated explicitly. On the other hand, as pointed out in Chapter 9,
each representing region reduces to a curve if the hypersurface is developable.
For a hypersurface that is both developable and quadric, the decomposition the-
orem gives rise to an algorithm for computing the equations of the representing
curves.

11.3.1 The Hyperplane Representation as a Linear Transformation

on the Coefficients
Consider a hyperplane π ∈ RN as embedded in PN with the equation c0 x0 +c1 x1 +
· · · + cN xN = 0 written in projective coordinates. We will focus on just one of its
representing points π̄12...N ∈ P 2 , knowing that all other representing points can be
analyzed similarly.
Let c = [c0 , c1 , . . . , cN ] ∈ P N , u = [1, 1, . . . , 1] ∈ P N , and d =
[d1 , d2 , . . . , dN ] ∈ P N . We recall (9.5) that π̄12...N = [c · u, c · d, −1]. Here
the positions of the coordinates are slightly rearranged, so that the variable corre-
sponding to the inhomogeneous constant in the Euclidean space is placed first. Now
let us define the linear transformation

II : v &→ MII v,

where the matrix MII is directly constructed from the above equations:
⎛ ⎞
0 1 1 ... 1
MII = ⎝ 0 d1 d2 . . . dN ⎠ .
−1 0 0 ... 0

The -coords representation of a hyperplane can thus be seen as an indexed set of

such linear transformations applied to the coefficients of the hyperplane. When we
want to emphasize that a different index gives rise to a different matrix and thus a
different linear transformation, we will index the map by, for example, II 123···N ,
indicating that the d1 , d2 , d3 , . . . , dN entries in the matrix MII 123···N are replaced
with appropriate alternatives.

11.3.2 A Brief Review of Duality in Projective Space

In the proper terms of projective geometry, a point in the projective space is tech-
nically a 1-dimensional vector subspace of RN +1 , while a line is technically a
1-dimensional collection of such vector subspaces, and so on. We will be slightly
 11.3 Surface Representation Decomposition and Developable Quadrics 453

sloppy here and refer to a point as a 0-flat, a line as a 1-flat, and so on, as if they
were embedded in RN .
Let us first recall the well-known “point ←→ hyperplane” duality in projective
geometry:

[c0 , c1 , . . . , cN ] ←→ {[x0 , x1 , . . . , xN ] : c0 x0 + c1 x1 + · · · + cN xN = 0}.

More generally, an m-flat dualizes to an (N − m − 1)-flat. We will not make direct

use of this general version of duality, and therefore a precise formula is not of our
concern. However, we will apply the idea of this generalization to an analogous
situation.
Now let us move on to one of the simplest classes of nonflat geometric objects:
the quadrics, or quadratic hypersurfaces. The equation of a quadric σ is of the
form xT Qx = 0, where x ∈ PN and Q is an (N + 1) × (N + 1) real symmetric
matrix. For N = 2, a nondegenerate quadric is a conic section and may be either
an ellipse, a parabola, or a hyperbola. Ellipsoids, hyperbolas, cones, and cylinders
are all examples of quadrics with N = 3.
If Q is nonsingular, then xT Q−1 x = 0 also defines a quadric σ̄Q . In fact, points
on Q dualize to tangent hyperplanes of σ̄Q , and vice versa. Thus we can understand
duality between quadrics σ and σ̄Q = σQ−1 in terms of the duality between points
and hyperplanes for any σ with nonsingular Q.
These are all familiar results from projective geometry up to this point.

11.3.3 Extending the Projective Duality to Degenerate Quadrics

What if Q is singular? What is a reasonable definition of σ̄Q , the dual geometric
object of σ in this case? Take the following surface in P3 as an example:
⎛ ⎞⎡ ⎤
−1 0 0 0 x0
⎜ 0 4 0 0 ⎟ ⎢ x1 ⎥
σ : [x0 x1 x2 x3 ] ⎜ ⎟⎢
⎝ 0 0 9 0 ⎠ ⎣ x2 ⎦ = 0.
⎥

0 0 0 0 x3

This is the elliptical cylinder 4x12 + 9x22 = 1 in an inhomogeneous equation. In the

same spirit as the nondegenerate case, we will find its tangent hyperplanes, and
dualize these to become points that collectively form the “dual geometric object”
we seek.
Let ξ = [ξ0 , ξ1 , ξ2 , ξ3 ] ∈ σ . The hyperplane πξ passing through ξ and tangent
to σ is

−(x0 − ξ0 )ξ0 + 4(x1 − ξ1 )ξ1 + 9(x2 − ξ2 )ξ2 = 0,

454 11. Recent Results

which reduces to
−ξ0 x0 + 4ξ1 x1 + 9ξ2 x2 = 0,
since ξ ∈ σ . Its dual ξ̄ is therefore [−ξ0 , 4ξ1 , 9ξ2 , 0]. It is easily verified that the
collection of all such points lies on the intersection of the plane x3 = 0 and the
x2 x2
quadric −x02 + 41 + 92 = 0. It is also straightforward to verify that conversely,
every point on the intersection dualizes to a tangent plane of σ . The dual of this
particular elliptical cylinder is therefore an ellipse (a curve) in P3 . Let us call this
curve σ̄ .
Are we justified in calling this curve the “dual” of σ ? If we “dualize” this
ellipse, do we get the elliptical cylinder back? For a brief moment it is easier to
think in R3 rather than in P3 . Let η = (η1 , η2 , 0) be a point on σ̄ , where we
write in inhomogeneous coordinates and η0 is taken as 1. There are many tangent
planes to σ̄ at η. In fact, they are precisely the pencil of planes containing the
tangent straight line to σ̄ at η. The equation of any plane in this family looks like
η1 η2
4 (x1 −η1 )+ 9 (x2 −η2 )+tx3 = 0, where t ∈ R. Indeed, these planes collectively
dualize to one ruling of σ .
The above example can be generalized in two ways.
First let us extend this to higher dimensions. Suppose that Q is an (N + 1) ×
(N + 1) diagonal matrix of rank c + 1 representing a hypersurface σ in PN . Without
loss of generality, we will assume that the nonzero entries are in the upper left
(c + 1) × (c + 1) block, and we will write this block as Q(c+1)(c+1) . That is,
 
Q(c+1)(c+1) 0
Q= 0 0
. The hypersurface σ can be seen as composed of “rulings”
each of which is an (N − c)-flat of the form {[ξ0 , ξ1 , . . . , ξc , xc+1 , . . . , xN ] :
xc+1 , xc+2 , . . . , xN ∈ R}. All points on one “ruling” share the same tangent
hyperplane. Moreover, it is a result from projective geometry that the dual of any
hyperplane of this form lies in the c-flat φ that is the intersection of the N − c
hyperplanes xc+1 = 0, xr+2 = 0, . . . , xN = 0. (Think of the plane x3 = 0 in the
third example.)
We might as well restrict our attention on σ to the confine of φ, a c-dimensional
subspace of PN , since this “cross section” of σ gives all nontrivial features of
σ . The rest is just “linear extension” along the N − c uninteresting coordinate
axes. Denote [x0 x1 . . . xc ]T by xφ . The equation of σ restricted to φ is thus
xφT Rxφ = 0. Since Q(c+1)(c+1) is nonsingular, the dual of σ (in this restriction
space) is xφT Q−1
(c+1)(c+1) xφ = 0. Stepping back into P , the dual of σ is seen as a
N

(c − 1)-dimensional object embedded in a c-dimensional subspace of PN .

The second generalization invokes results from linear algebra. Let Q be a real
symmetric matrix of rank c + 1. It is unitarily diagonalizable as Q = U U T for
 11.3 Surface Representation Decomposition and Developable Quadrics 455

some orthogonal matrix U and some diagonal matrix . In fact, column vectors of U
are the unit eigenvectors of Q, and the diagonal elements of  are the corresponding
eigenvalues. Thus the equation of an arbitrary quadric σ : xT Qx = 0 can be written
as xT U U T x = 0. Perform a change of basis y = U T x to obtain yT y = 0,
and the previous results can then be applied to σ in this new coordinate system,
since  is diagonal. We note that both tangency and duality of flats are invariant
under change of basis. The c + 1 column vectors of U corresponding to the nonzero
eigenvalues span φ, the intersection c-flat. The remaining N − c column vectors
correspond to the directions that form the “rulings.”
The result in this section can be summarized as follows. Let σ be a geometric
object in PN that
• lies in the intersection of e hyperplanes, which is a copy of RN −e that we shall
call φ,
• is composed of “rulings” each of which is a copy of Pf , where L̄ ≡ e,
• has an equation of the form xT Qx = 0 in φ, where the rank of Q is c + 2 ≡
N + 1 − e − f.
Then the duality map D maps σ to a geometric object σ̄ in PN that
• lies in the intersection of ē ≡ f hyperplanes, which is a copy of RN −f that we
shall call ψ,
¯
• is composed of “rulings” each of which is a copy of Pf , where f¯ ≡ e,
• has an equation of the form xT Q̄x = 0 in ψ, where the rank of Q̄ is again
c + 2 = N + 1 − ē − f¯.
This is actually a slightly generalized version of what we have discussed above.
However, we will make use only of the results from the prior discussions and hence
will omit the tedious but straightforward proof of this more general result. On the
other hand, this general result is so interesting that such digressions and some further
comments are warranted.
We choose the mnemonics e, f , and c to mean, respectively,
• emptiness: the difference of dimensions between N and that of the minimal
subspace that contains σ ,
• flatness: the “degree of degeneracy” of σ , or more precisely, the dimension of
each flat “ruling” of σ ,
• curvedness: the number of nonvanishing terms in the diagonalized first
fundamental form, properly generalized to objects in higher dimensions.
In view of the identity e + f + c = N − 1 and the above observation, one can say
that emptiness dualizes to flatness and vice versa, while curvedness is unchanged
456 11. Recent Results

by the duality D. Consider again the above elliptical cylinder 4x12 + 9x22 = x02 as
an example. Here e = 0, f = 1, c = 1, and N = 3. For its dual, a circle in P3 ,
ē = 1, f¯ = 0, and c̄ = 1.
As a special case, c = 0 means f + e = N − 1. This corresponds to the
classical duality that a k-flat in PN dualizes to an (N − k − 1)-flat. As another
special case, c = N − 1 means f = e = 0. This corresponds to the classical
duality for nondegenerate quadrics, whose dual is found by inverting its representing
symmetric matrix. The flat duality and the quadric duality in projective geometry
are thus unified in this general duality map D.
We will make use of only the special case e = 0 in later sections, and will
be particularly interested in the most special case e = 0 and c = 1, the class of
developable hypersurfaces.

11.3.4 The Decomposition and Its Application to Developable

Quadrics
The decomposition theorem can now be easily stated.
Theorem 11.3.1. The parallel-coordinate representation of hypersurfaces using
tangent hyperplanes is equivalent to an indexed set of compositions: {II I ◦ D :
I ∈ {123 · · · N, 1
23 · · · N, 1
2
3 · · · N, . . . , 1
2
3
· · · N}.
This is a direct result of the fact that in parallel-coordinate representations, we
take tangent hyperplanes to objects and then draw their representations in the
-coords plane to represent the original hypersurface. Again for convenience, in
the following discussions we will focus on just one of the compositions and omit the
indices.
This representation typically gives rise to a patch in the -coords plane, since
typically a smooth hypersurface in PN has an (N − 1)-parameter family of tangent
hyperplanes. In the special case of developable hypersurfaces, it has a 1-parameter
family of tangent hyperplanes, which means that D maps the hypersurface to a
curve in PN . Subsequently, II maps the space curve onto a plane curve in P2 . For
a developable quadric, this plane curve can be explicitly computed.
Let σ be a developable quadric in P N defined by the homogeneous equation
x Qx = 0, where Q is an (N + 1) × (N + 1) real symmetric matrix. Its flatness
T
is N − 2, and Q has rank 3.
First, unitarily diagonalize Q as U U T , where U ’s column vectors are the unit
eigenvectors of Q and the diagonal elements of  are the corresponding eigenvalues.
For convenience, we will assume that the only three nonvanishing eigenvalues are
arranged to appear at the upper left corner. As a notational convenience, for any
matrix A we will
 11.3 Surface Representation Decomposition and Developable Quadrics 457

• denote by A∗3 the submatrix composed of the leftmost three columns of A,

• denote by A3∗ the submatrix composed of the topmost three rows of A,
• and denote by A33 the upper left 3 × 3 submatrix of A.
Thus 33 is the only nonvanishing part of , and U∗3 is the submatrix composed
of the corresponding eigenvectors.
Perform a change of coordinate system y = U T x. In the new coordinate
system, the equation of σ simplifies to yT y = 0, which is nothing more than
y3∗ T 33 y3∗ = 0. The equation of the dual is y3∗ T −1
33 y3∗ = 0, y4 = y5 = · · · =
yN+1 = 0.
Let x ∈ D(σ ) be a point in the dual (curve), written in the original coordinate
system. Then z = MII x is a point in σ̄ , the -coords representation of σ . That
is, z = MII U y, since U is unitary (actually orthogonal). Since all but the first
three coordinates of y vanish, we can just focus on z = MII U∗3 y3∗ . Substituting
y3∗ = (MII U∗3 )−1 z into the equation of the dual, we obtain the equation of the
-coords representation of σ in homogeneous coordinates:
σ̄ : zT (MII U∗3 )−T −1 −1
33 (MII U∗3 ) z = 0.

11.3.5 A Numerical Example: An Elliptical Cone in 5-Space

It will be very helpful to have a matrix calculator such as octave or rlab at hand for
verifying the computations in this section. Consider the hypersurface σ : xT Qx = 0
in 5-space, where
⎛ ⎞
32 0 −16 −80 −64 8
⎜
⎜ 0 −1 −1 2 1 −2 ⎟ ⎟
⎜ −16 −1 −17 114 81 6 ⎟
Q=⎜ ⎟.
⎜ −80
⎜ 2 114 −20 14 −52 ⎟ ⎟
⎝ −64 1 81 14 31 −38 ⎠
8 −2 6 −52 −38 −8
First let us verify that the rank of Q is 3 (using the calculator). In the notation
of the duality section, here we have N = 5, e = 0, f = 3, c = 1.
Now diagonalize Q as U U T , where
⎛ ⎞
0.479 −0.282 0.094 0.136 −0.039 0.813
⎜ −0.007
⎜ 0.009 −0.399 0.205 0.891 0.063 ⎟ ⎟
⎜ −0.475
⎜ 0.589 −0.292 −0.176 −0.138 0.541 ⎟ ⎟
U =⎜ ⎟
⎜ −0.501 −0.629 0.098 −0.541 0.159 0.163 ⎟
⎝ −0.496 −0.322 −0.099 0.766 −0.225 0.053 ⎠
0.219 −0.272 −0.853 −0.172 −0.330 −0.112
458 11. Recent Results

and
⎛ ⎞
201.3 0 0 0 0 0
⎜ 0 −178.1 0 0 0 0 ⎟
⎜ ⎟
⎜ 0 0 −6.241 0 0 0 ⎟
=⎜ ⎟.
⎜ 0 0 0 0 0 0 ⎟
⎜ ⎟
⎝ 0 0 0 0 0 0 ⎠
0 0 0 0 0 0

In octave, calling the eig() function to compute the eigenvectors and eigenvalues
of Q gives the desired result, except that the order has to be rearranged so that the
nonvanishing eigenvalues appear at the upper left corner.
For the σ̄12345 curve, the transformation matrix for II is
⎛ ⎞
0 1 1 1 1 1
MII = ⎝ 0 0 1 2 3 4 ⎠ .
−1 0 0 0 0 0

To find σ̄ : [z, x, y]R[z, x, y]T = 0, the equation of the -coords representation of

σ , we compute Q̄ = (MII U∗3 )−T −1 33 (MII U∗3 )
−1 and find that

⎛ ⎞
−10.573 3.662 11.776
Q̄ = ⎝ 3.662 −1.268 −4.076 ⎠ .
11.776 −4.076 −13.112

We need to find a simple parametric equation for this curve in order to draw its
picture. To this end, we try to write Q̄ as
  
1 cT f 0 1 0
,
0 V 0 L c VT

where V is a 2 × 2 rotation matrix and c is the translation vector. This can be done
by first diagonalizing the lower-right 2 × 2 block of Q̄:

Q̄ = V LV T
  
0.297 −0.955 −14.379 0 0.297 0.955
= .
0.955 0.297 0 −0.00133 −0.955 0.297

Then c = L−1 V T (3.662, 11.776)


T = (−0.858, 0.58)T . Finally, f = −10.573 −

cT Lc = 0.00426. The matrix c V0T gives the necessary rotation and translation
1

for simplifying the equation of σ̄ in inhomogeneous equations. After a little further

computation, we obtain the parametric equation as
 11.3 Surface Representation Decomposition and Developable Quadrics 459

  
0.297 −0.955 0.017 cos(t) 0.809
x(t) = + .
0.955 0.297 1.789 sin(t) 0.647
The same technique can be used to find a rigid motion (rotation + translation)
that centers σ at the origin of the new coordinate system and places the axes
of σ aligned with the new axes. Specifically, by decomposing MII as AT BA,
where
⎛ ⎞
1 0 0 0 0 0
⎜ 0.340 −0.010 −0.663
⎜ 0.616 0.294 0.307 ⎟⎟
⎜ 0.137 −0.403 −0.347
⎜ 0.092 −0.119 −0.833 ⎟ ⎟
A=⎜ ⎟
⎜ −1.250 −0.831 0.054 −0.281 0.375 0.295 ⎟
⎝ 3.438 0.384 −0.281 −0.511 0.681 −0.222 ⎠
0.577 −0.010 −0.599 −0.521 −0.543 0.274

Figure 11.13. The -coord representations of a cone in 5-space.

The indicated four points on the curves, with the same y coordinate, represent a plane tangent to the
cone.
460 11. Recent Results

and
⎛ ⎞
0 0 0 0 0 0
⎜ −161.8 ⎟
⎜ 0 0 0 0 0 ⎟
⎜ −6.16 ⎟
⎜ 0 0 0 0 0 ⎟
B=⎜ ⎟,
⎜ 0 0 0 0 0 0 ⎟
⎜ ⎟
⎝ 0 0 0 0 0 0 ⎠
0 0 0 0 0 153

we find that σ is an elliptical cone of the form 161.8x12 + 6.16x22 = 153x52 .

Figure 11.13 shows all four of the indexed curves representing σ , along with a
set of indexed points representing π0 : x1 + 4x2 + 4x3 − 4x4 − 2.3258x6 + 2 = 0,
a randomly chosen tangent plane of σ .

11.4 Network Visualization and Analysis with Parallel

Coordinates

Yaron Singer
School of Computer Science,
Tel Aviv University, Israel
[email protected]
and
Ohad Greenshpan
IBM Research Center
Haifa, Israel
[email protected]

11.4.1 Introduction
Large networks impose a challenge in many areas of research, and visualization
plays a major role in analyzing topological traits and large amounts of (associ-
ated) information. Methods such as SeeNet [1], Genius [2], SpectralNet [3] and
approaches suggested by Bolser et al. [4] and McGrath et al. [5] all display a
variety of network visualization techniques in different research areas, dealing with
complexity and data of large-scale networks. Although the intersection between
the disciplines of network topological analysis and multivariate visualization is
not immediate, with the wide spectrum of applications that we have thus far wit-
nessed [6], [7], [8], it is clear that topological network analysis could benefit from
multivariate visualization in general, and -coords in particular. In this chapter
 11.4 Network Visualization and Analysis with Parallel Coordinates 461

we introduce the Node Extraction Visualization (NEVIS) method, which enables

network topological visualization using -coords.
Since networks are subject to changes that affect their function and performance,
understanding the effects due to changing conditions is vital to their analysis. Here,
we use changing conditions of node extractions to obtain topological insight about
the network. Though the likelihood of such an event varies from one network to
another, in instances where network functionality is determined largely by path
distances and connectivity of nodes, node extraction is a crucial event. NEVIS
transforms a given network into an equivalent representation in -coords. Under
this transformation, networks under changing conditions are produced whose func-
tionality is visualized. Since NEVIS transforms nodes into multidimensional points,
we will see how it can be utilized to study key traits such as node stability, backup,
inter-relationship between nodes, and to identify topological patterns in the network
and other properties of interest in network research. We will also see that NEVIS’s
implementation is simple and modular.

11.4.2 The NEVIS Transformation

We begin by introducing and discussing extracted networks and the χ function,
which are the basic elements of NEVIS. We will then see how to use these elements
in the construction of the NEVIS algorithm.

Extracted Networks

Informally, dynamic networks are networks that are subject to topological change.
We focus on a unique subset of dynamic networks, called extracted networks, which
are defined as follows.
For the network modeled by a directed graph S = (V , E) with nodes
V = {v1 , v2 , . . . , vn } and edges E = {(vi , vj ) ∈ V × V |vi is connected to vj },
let the ith extracted network, denoted by Si , be the original network S without the
node vi and without any edges connected to it. That is,

Si := (V \{vi }, E\{(vj , vk ) ∈ V × V |j = i ∨ k = i}). (11.3)

For example, in Fig. 11.14, the network S and its third extracted network, S3 , are
shown. Transformation of a network with n nodes as above requires producing a set
of extracted networks {S1 , S2 , . . . , Sn }. This set produces an n-dimensional space
as explained in detail below.
462 11. Recent Results

Figure 11.14. Network S and the third extracted network, S3 .

Function χ : Measuring Node Centrality

Topological properties are frequently used to measure a network’s functionality,
which in turn depends on shortest paths and connectivity. To apply an appropriate
measure that enables classification of nodes by these criteria, the χ function below
is used to determine a node’s topological centrality. For a network S = (V , E),
with V and E as above, and some node u ∈ / V , χ : V ∪ {u} −→ R is defined by
 n
1
j =1,j =i d(vi ,vj ) , vi ∈ V ,
χ(vi ) =
−∞, vi ∈
/ V,
with the following conventions:
1. d(vi , vi ) = 0 ∀i ∈ {1, 2, . . . , n},
2. d(vi , vj ) = ∞ ⇐⇒ vj cannot be reached from vi ,
3. if d(vi , vj ) = ∞ then d(v1,v ) = 0.
i j

Clearly, 0 ≤ χ(vi ) ≤ n − 1 ∀vi ∈ V .

To further clarify, for the network shown in Figure 11.15, the χ function values
for each node are given in Table 11.2.

Figure 11.15. Network S whose χ function values are given in Table 11.2.
 11.4 Network Visualization and Analysis with Parallel Coordinates 463

Table 11.2.

vi d(vi , v1) d(vi , v2 ) d(vi , v3) d(vi , v4) d(vi , v5) χ value
n
j =1,j =1 d(v1 ,vj ) = 4
1
v1 0 1 1 1 1
n
j =1,j =2 d(v2 ,vj ) = 3
1
v2 1 0 1 2 2
n
j =1,j =3 d(v3 ,vj ) = 3
1
v3 1 1 0 2 2
n
j =1,j =4 d(v4 ,vj ) = 2.5
1
v4 1 2 2 0 2
n
j =1,j =5 d(v5 ,vj ) = 2.5
1
v5 1 2 2 2 0

Why χ? For a node vi ∈ V , note that χ (vi ) depends on the number of nodes
connected to vi and their distances from it; χ monotonically increases with respect
to both centrality and connectivity of the node. Thus, in relation to other nodes in
the network, for a node vi that is connected to a large number of nodes at short
distances, χ(vi ) is high; symmetrically, for a node vj that is connected to a small
number of nodes at large distances, χ (vj ) is low. These properties make the χ
function a favorable candidate for measuring a node’s centrality in the network.

The NEVIS Algorithm

Having introduced its elements, we now present the NEVIS transformation and the
algorithm used to implement it.
For a network S as described above, we produce n extracted networks
S1 , S2 , . . . , Sn . By extracting the node vi from S, we produce the ith extracted
network Si , which consists of n − 1 nodes. For each extracted network, using the
breadth-first search (BFS) algorithm for graphs to obtain edge distances between
nodes, we calculate the χ function values of all n nodes (by definition, χ (vi ) = 0
in Si ∀i ∈ {1, . . . , n}). This process produces a matrix of order n × n:
CreateMatrix
for all vi ∈ V do
Si ← ExtractNode(vi , S)
for all vj ∈ V \{vi } do
BF S(Si , vj )
χ(vj ) ← 0
for all vk reachable from vj do
χ(vj ) ← χ(vj ) + d(v 1,v )
j k
464 11. Recent Results

end for
Mij ← χ(vj )
end for
end for
return M
The matrix is then transformed to its equivalent multidimensional representation
by considering each row as an n-dimensional point. With parallel coordinates,
each point is represented as a polygonal line. We obtain visualization with the
networks S1 , S2 , . . . , Sn represented by the multidimensional coordinates system
with first, second, . . . , nth axes respectively. In turn, the nodes v1 , v2 , . . . , vn
are represented by n polygonal lines. Figure 11.16 describes a flow of this
implementation.
NEVIS transforms a network with n nodes into n points in an n-dimensional
space. The n extracted networks serve as the components of the nth dimensional
space, and the nodes’ χ values are transformed into points in that space. The
following section discusses the properties of this transformation and their use in
network analysis.

General Network Transformations

Note that we can generalize the idea of network transformation beyond NEVIS.
We can formally categorize network transformations into parallel coordinates with
the following functions:
1. γ function, which is defined on the set of all possible directed graphs and
returns a set of networks.
2. ν function, which is defined on the set of all possible directed graphs and returns
a real value.
In NEVIS we use the γ function, which produces the set of all possible extracted
networks. That is, γ (S) = {S1 , S2 , . . . , Sn } for a network with n nodes. The ν
function used can be seen as a combination of ψ, which is defined by ψ(S) = V ,
and χ, as shown above. Thus selection of γ , ν, ψ, χ can yield different network
transformations as desired.

11.4.3 Properties of NEVIS

We can now begin exploring some of the properties of NEVIS. The principal pro-
perties discussed here facilitate the analysis of large-scale networks, as presented
later.
 11.4 Network Visualization and Analysis with Parallel Coordinates 465

Figure 11.16. The NEVIS transformation network → n points in n-dimensional space.

466 11. Recent Results

Figure 11.17. Visualizing topological significance.

The drops in the sixth axis as opposed to the consistency in the third axis suggest that v3 is backed
up in the network.

Topological Significanc

After creating the n extracted networks S1 , S2 , . . . , Sn , the effects of each extraction

are independently measured by χ(vj ) ∀vj ∈ V \{vi }. By studying the ith component
of the n-tuple produced by NEVIS, one can comprehend the effect of extracting
vi from the network. If vi has a significant effect on the network’s centrality and
connectivity, this property is manifested in a drop of polygonal lines representing
nodes depending on vi in the ith axis in NEVIS. It thus follows that the topological
significance of vi correlates with the number of polygonal lines dropping in NEVIS,
and the sharpness of their drop.
Contrary to our intuition, a high χ does not necessarily guarantee topological
significance. Figure 11.17 displays an example in which v3 has a high χ value in
the network (this can be seen in NEVIS by the fact that its corresponding polygonal
line is the top outlier), though observing the third axis in NEVIS, it is clear that v3 ’s
extraction is not significant to the network. On the other hand, there is a dramatic
drop in the sixth component, despite the fact that χ (v6 ) is not among the highest
in the network. This seemingly counterintuitive example reveals one of the strong
properties in NEVIS. A closer study of the above example shows that the extraction
of v6 is more significant to the network than that of v3 ’s despite χ (v6 ) < χ(v3 ),
since each node connected to v3 is also connected to v10 .
Intuitively, the idea of a backup of a node is the existence of alternative paths
connecting a pair of its neighbors without passing through the node. In Figure 11.17,
v3 ’s extraction is not significant to the network’s stability, since v10 serves as a full
 11.4 Network Visualization and Analysis with Parallel Coordinates 467

backup. We can think of v3 as having a backup of order 4 (v3 ’s degree) in the sense
that there is another node (v10 ) connected to all of v3 ’s four neighbors ({v6 , v7 , v9 ,
v10 }). Formally, suppose that vi is connected to {v1 , . . . , vr }. We say that vi has a
backup of order k, k ≤ r, if there is at least one node in the network connected to k
of vi ’s neighbors. Clearly a node’s significance in the network highly depends on
its backup order. Study of the drops in the ith axis in NEVIS provides a measure of
vi ’s global significance which considers its backup order.
The ability to measure node significance considering both centrality and its
backup order enables evaluating a node’s impact on the network’s topology in the
network. Since a node can be partially backed up by a set of nodes that are reached
at different distances, the ability to consider a complicated factor such as a node’s
backup makes NEVIS a useful tool.

Relationship between Nodes

An immediate conclusion from the above is that relationship between nodes vj and
vk can be studied by examining the drop of vk in the j th axis, and symmetrically
examining the drop of vj in the kth axis. Although the latter is one of its fundamental
properties, NEVIS’s representation of nodes as polygonal lines enables greater study
of the relationship between nodes.
Since each node’s χ value throughout n extractions is represented as a polygonal
line, studying the relationship between the polygonal lines representing vj and vk
provides insight into the relationship between these nodes. Consider the network
portrayed in Figure 11.18, composed of three separate connected components.
To achieve clarity, nodes in the same connected component are given an identical
color. Assume that the axes are ordered in the manner displayed (we refer to this

Figure 11.18. Network composed of strongly connected components.

The figure displays the interrelationship between nodes.
468 11. Recent Results

point below). In this example it is evident that polygonal lines of the same color have
a specified range in which they are subject to change and a complementary range
where they run horizontally. It is easy to see that this property directly derives from
the fact that nodes in the same connected component react similarly to extractions
in the network, influenced only by extractions in the connected component to which
they belong, and insensitive to extractions from other connected components. In this
example we have chosen a specific axis ordering obtained by sorting according to
χ value.

Stability of Nodes
Since the nodes’ χ values are represented by polygonal lines in the NEVIS rep-
resentation, peaks of polygonal lines disclose the corresponding nodes’ centrality
values in instances in which the node extracted has minimal effect. Note that for
a network S = (V , E) as above, it is possible to implement NEVIS as n points
in an (n + 1)-dimensional space by adding the 0th extracted network to S and
thus obtaining visualization of the χ value in S itself. In Figure 11.19, the top
outlier is shown by a red polygonal line, which represents v15 in the network,
with noticeably higher values than the rest of the nodes in the network. This can
be easily verified by observing the node-link diagram. NEVIS orders the poly-
gonal lines, starting from the top, in order of decreasing centrality throughout all
extractions.
Since the axes represent the extracted networks, it follows that polygonal lines
free of dramatic inclines and declines represent nodes that are relatively stable

Figure 11.19. Analyzing stability from the behavior of polygonal lines.

The green polygonal line remains relatively steady, whereas the red polygonal line has sharp drops.
 11.4 Network Visualization and Analysis with Parallel Coordinates 469

and immune to node extractions in the network. Note that node stability does not
necessarily correlate with high centrality value. In Figure 11.19, the sharp inclines
and declines of the red polygonal line testify that its corresponding node v15 is
unstable in the sense that its centrality is highly affected by extractions. In contrast,
it can be inferred that v3 is relatively stable by inspecting the yellow polygonal
line by which it is represented in NEVIS. In this case, it is not clear how this can
be deduced by studying the node-link diagram alone. The ability to analyze node
stability in the network can be widely used in network analysis, since it discloses
weak links in the network.

Identifying Topological Patterns

Identifying unique topological patterns can be used in network analysis to deduce
relevant information in a first encounter with a given network. To identify unique
topological patterns, we use a slight variation of NEVIS by replacing the χ function
with χ0 defined by
 n
1
j =1,j =i d(vi ,vj ) , vj ∈ V ,
χ0 (vi ) =
0, vj ∈
/ V.

NEVIS reveals the network’s distinct topological properties by mapping them one-
to-one into distinct patterns. Figure 11.20A displays a clique, a graph in which all
the nodes are connected between one another, and Figure 11.20B shows a star, a
set of independent nodes all connected to a single node.
Following a one-to-one transformation between topological properties and pat-
terns in -coords, networks with topological patterns that are almost symmetric are
represented as their almost symmetrical equivalents in NEVIS. Figures 11.21A and
11.21B display networks that are almost a clique and almost a star, respectively.
Comparing the NEVIS representation of the perfect graphs with the almost perfect
ones illustrates this point. This property is very useful in the study of topological
properties in networks. Perfect topological patterns are rare; it is their almost coun-
terparts which are encountered in the real world. We now show how this property
can be used in network study.

11.4.4 Large-Scale Networks

In this section we discuss applications of NEVIS on large-scale networks. Since it
synthesizes network analysis and multivariate data visualization, NEVIS benefits
from both worlds; preprocessing algorithms that enable combinatoric reduction of
the network can be applied as well as any methods that enable dealing with large
470 11. Recent Results

Figure 11.20. Clique with 15 nodes is in A, and a star with 15 nodes is in B.

data sets in -coords. We display examples of both in this section. Our interest here
is not to offer novel methods for re-scaling networks or reducing large data sets for
visualization, but rather to show how existing methods of these ideas can be applied
in NEVIS.

Matrix Manipulations
We can now view possible simple solutions to deal with large-scale networks uti-
lizing -coords. The need to analyze large networks raises the issue of scalability,
which can be accomplished with -coords. Figure 11.22A displays a node-link
diagram of a network that consists of 100 nodes, and its NEVIS presentation is
represented in Figure 11.22B. Due to NEVIS’s simple construction, selecting a
subset of nodes to be visualized is easy. Recall from Section 11.4.2 that imple-
menting NEVIS requires producing a matrix that represents the nodes’ χ values
throughout extractions. To ignore polygonal lines, we simply replace the values of
their corresponding rows in the produced matrix with −∞. That is, in order to ignore
 11.4 Network Visualization and Analysis with Parallel Coordinates 471

Figure 11.21. In A, a network that is almost a clique, and in B an almost star topology.

the vj in the NEVIS visualization, replace the j th row in the matrix returned by
CreateMatrix:

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
x11 ... x1n 0 0 0 x11 ... x1n
⎜ .. .. .. ⎟ ⎜ .. .. ..⎟ ⎜ .. .. .. ⎟
⎜ . . ⎟ ⎜ ⎟ ⎜ . . ⎟
⎜ . ⎟ ⎜ . . .⎟ ⎜ . ⎟
⎜ xj 1 xj n ⎟ + ⎜
⎟ −∞ ⎟ ⎜ −∞ ⎟
⎜ ... ⎜ −∞ ... ⎟ = ⎜ −∞ ... ⎟.
⎜ .. .. .. ⎟ ⎜ .. .. .. ⎟ ⎜ . .. .. ⎟
⎝ . . . ⎠ ⎝ . . . ⎠ ⎝ .. . . ⎠
xn1 ... xnn 0 ... 0 xn1 ... xnn

The results of this procedure are shown in Figure 11.22C. Similarly, it is easy to
reduce a subset of axes from NEVIS by reducing the matrix to represent the desired
axes only. An example of such reduction is shown below. The result of reducing
both polygonal lines and axes is displayed in Figure 11.22D and corresponds to the
matrix
472 11. Recent Results

⎛ ⎞ ⎛ ⎞
x11 ... x1j ... x1n 0 ... −∞ ... 0
⎜ x21 ... x2j ... x2n ⎟ ⎜ 0 ... −∞ ... 0 ⎟
⎜ ⎟ ⎜ ⎟
⎜ .. .. .. ⎟+⎜ .. .. .. ⎟
⎝ . ... . ... . ⎠ ⎝ . ... . ... . ⎠
xn1 . . . xnj ... xnn 0 ... −∞ ... 0
⎛ ⎞
x11 ... −∞ ... x1n
⎜ x21 ... −∞ ... x2n ⎟
⎜ ⎟
= ⎜ .. .. .. ⎟.
⎝ . ... . ... . ⎠
xn1 ... −∞ ... xnn

The running time required to implement NEVIS depends on the CreateM-

atrix procedure, and visualization is thereafter achieved in linear time. This
implementation allows performance of such manipulations efficiently. Moreover,

Figure 11.22. Examples of possible solutions for large networks.

Shown are polygonal line reductions that represent 10 nodes in a 100-dimensional space in C, and D
shows the 10 nodes in their corresponding 10-dimensional space.
 11.4 Network Visualization and Analysis with Parallel Coordinates 473

Figure 11.23. NEVIS applied to a genetic regulatory network with over 3000 nodes.
Only one-fifth of the nodes are shown.

due to this implementation, a wide range of solutions to dealing with large data sets
in -coords can be applied in a simple manner.

Combinatoric Manipulations
Figure 11.23 displays a node-link diagram of a genetic regulatory network of over
3000 nodes. The network shown is a protein regulatory network, modeled by a
directed graph, where proteins are represented by nodes and regulatory interactions
are represented by edges. Clearly, the size of the network raises serious obstacles
in attempts for visual analysis. To obtain visualization of the above network, we
apply NEVIS to the network and show nodes with out-degree greater than zero.
The reasoning here is that nodes with zero out-degree have little effect on χ values
(why?), and visualization in NEVIS does not provide any relevant information.
We thus obtain a visualization of 142 nodes of the network in NEVIS, as shown in
Figure 11.24.
We can now use the properties of NEVIS from the section above to analyze
the network. In this instance a few immediate conclusions can be drawn about the
network. Generally, apart from a few specific instances, the polygonal lines run
relatively horizontally, which testifies that the network is relatively stable. Note
also that there is an area that is clear of polygonal lines. We conclude that no
node receives χ values in that area, and that it may testify to segmentation of
two separate connected components in the network. The horizontal polygonal lines
in the lower values support this hypothesis, similar to the connected-components
474 11. Recent Results

Figure 11.24. Visualization of the 142 nodes in the regulatory network of Figure 11.23 with degree
greater than zero.

example presented in the previous section. Examining the top blue polygonal line,
it is evident that despite its high centrality value, its extraction does not affect
the network’s functionality. Using an interactive tool that enables pointing on
the polygonal line and obtaining its number, we can inspect its extraction effect
on other nodes. Examining the 127th axis, we learn that its extraction does not
affect functionality in the network. Closer study reveals that there are, in fact, two
polygonal lines of maximal χ values that race between one another throughout all
extraction. We assume that this node serves as its backup.
There are four nodes whose extraction has significant effect on the network.
This is manifested in the drops of polygonal lines in the area of the 9th, 15th, 34th,
82nd, 124th axes. This signifies that these nodes are have relatively poor backup in
the network, and their extraction affects the network’s functionality.
This example is applicable to biological networks, the Internet and elsewhere.
Other works on combinatoric reduction of the problem can also be applied as
preprocessing algorithms in a similar manner to reduce the combinatoric problem.

Exercises
1. What is the running time of the CreateMatrix algorithm?
2. Give an example of a connected graph that has horizonal portions in NEVIS.
3. Suppose that for S = (V , E) and some node u ∈ / V we implement NEVIS,
replacing the χ function with the function deg : V ∪ {u} → N defined by

|{w : w is connected to v}|, v ∈ V ,
deg(v) =
−∞, v∈/ V.
 11.4 To See C2 475

(a) Apply NEVIS with the deg and χ (not χ0 ) functions on a five-node star
and five-node clique. Any difference? Why?
(b) What would be the motivation to replace deg with χ for topological pattern
identification?
(c) Think of other topological functions under which a star would look the
same as in NEVIS.

4. This question discusses backup order, which has been defined above.

(a) Prove that the relation Rk = {3u, v4| u is a backup of order k of v} is

symmetric. That is, prove that u is a backup of order k of v ⇐⇒ v is a
backup of order k of u.
(b) Is it possible that u has a backup of order k, where k = degree(u), and
there will still be a significant drop in its axis?

References
[1] Becker, R.A., Eick, S.G., Wilks, A.R., Visualizing network data. IEEE Trans-
actions on Visualization and Computer Graphics (1995), Vol. 1–1; 16–21.
[2] Ciccarese, P., Mazzocchi, S., Ferrazzi, F., Sacchi, L., Genius: A new tool
for gene networks visualization. Intelligent Data Analysis in Medicine and
Pharmacology (IDAMAP) Proceedings (2004).
[3] Forman, J.J., Clemons, P.A., Schreiber, S.L., Haggarty, S.J., SpectralNET–an
application for spectral graph analysis and visualization. BMC Bioinformatics.
2005 October 19; 6:260.
[4] Bolser, D., Dafas, P., Harrington, R., Park, J., Schroeder, M., Visualisation
and graph-theoretic analysis of a large-scale protein structural interactome.
BMC Bioinformatics 2003; 4:45.
[5] McGrath, C., Hardt, D.C., Blythe, J., Visualizing Complexity in networks:
Seeing both the forest and the trees. Connections 25(1):37–41 2002 INSNA.
[6] Keim, D.A., Information Visualization and Visual Data Mining. IEEE TVCG,
7–1, 2002, pp. 100–107.
[7] Shneiderman, B., The Eyes Have It: A Task by Data Type Taxonomy for
Information Visualizations. Proceedings of the 1996 IEEE Symposium on
Visual Languages.
[8] Wegenkittl, R., Loffelmann, H., Groller, E., Visualizing the Behaviour of
Higher Dimensional dynamical systems. Proceedings of IEEE VIS 1997
pp. 119–126.
476 11. Recent Results

11.5 To See C2
Yoav Yaari
School of Mathematical Sciences
Tel Aviv University
Tel Aviv, Israel
[email protected]

11.5.1 The Complex Field C

The set of complex numbers x + iy, where x and y are real numbers, together with
the binary operations (+, ×), forms the field C. Using (1, i) as a basis, C can be
identified with the two-dimensional plane R2 , the horizontal and vertical axes being
the real and imaginary axes, respectively. Addition corresponds to translation, while
multiplication involves scaling (of the absolute values) and rotation (Fig. 11.25).

Figure 11.25. Translation, scaling, and rotation.

 11.5 To See C2 477

The goal here is the visualization of C2 , the space of complex-valued functions,

which ostensibly is four-dimensional. Yet trying to represent C2 with four parallel
axis does not seem possible even for a simple function like f = az + b. No axis
assignment (i.e., choosing which axis should represent the real and imaginary parts)
or subsequent axis ordering worked.

11.5.2 The Space C2

Though we are dealing with four-dimensional information, rather than four inde-
pendent coordinates, what is involved are pairs of pairs. This suggests using two
copies of C, which we take as parallel planes with a horizontal one-dimensional axis
w passing through their origins. Of course, this arrangement collapses to -coords
when the imaginary parts are zero.
The two copies of C are denoted by Zi , i = 1, 2, with Xi , Yi as the real and
imaginary axis, respectively. The origin of Z1 is at w = 0 and the origin of Z2
at w = 1; the distance between the parallel planes is one unit (Fig. 11.26). The

Figure 11.26. The complex parallel planes coordinate system.

478 11. Recent Results

coordinates (x, y, w) designate the position of a point in the three-dimensional

space R3 . For example, (2, 1, 0) stands for the point 2i + 1 on Z1 , and (1, −2, 1)
for the point i − 2 on Z2 . To visualize (i.e., represent) the pair 3z1 , z2 4 ∈ C2 we
propose joining the point z1 ∈ Z1 to z2 ∈ Z2 with a straight line. Taking our cue
from the -coords point–line duality, we seek a corresponding relation resulting
from the visualization of C2 .
Let’s see what the representation of a simple linear function, say f (z) = 2z,
looks like. The collection of lines joining the points z ∈ Z1 to 2z ∈ Z2 is given
parametrically by (5(z + tz), 6(z + tz), t)t ∈ R. They intersect at (0, 0, −1)
(Fig. 11.27), which is an auspicious result.
 1Recall b  for f (x) = ax + b a, b ∈ R,
that
the representing point in -coords is 1−a , 1−a . Proceeding in the same way
for the function
  f (z) = az + b with a ∈ R and b ∈ C, the representing point
obtained is 5 1−ab , 6 b , 1 , Just as for -coords, the representing point is
1−a 1−a
undefined in R2 when a = 1, prompting the extension of the representation from R2
to P2 .

Figure 11.27. The lines joining the points (z, 2z) intersect a point.
 11.5 To See C2 479

Figure 11.28. The lines joining the points (z, (1 + i)z) do not intersect.

For the representation of f (z) = az, where a is a complex number (but not real),
the lines joining the pairs 3z, az4 are (5(z + taz − tz), 6(z + taz − tz), t). To find
their intersection, we look for points p, q ∈ Z1 (p = q) such that p + tap − tp =
q + taq − tq, with t being their common w coordinate. That is, 1 + ta − t = 0 or
t = 1−a1 , but t must be real; therefore, these lines do not intersect for complex a,
as shown in Figure 11.28.

11.5.3 Representing Points on a Line in C2

Showing the association between z in Z1 with f (z) in Z2 with straight lines is
abandoned, since it does not yield a duality for a linear function with complex
coefficients. Can a duality be obtained if z and f (z) are joined by a curve? Let
us return to the function f (z) = az + b, and observe that it consists of a scaling
480 11. Recent Results

by |a|, rotation of a radians, and translation by b. Rewriting the function in the

form

1−a b b
f (z) = az + b = az + b =a z− + ,
1−a 1−a 1−a
  b   b 
reveals that the point 5 1−a , 6 1−a , which we call the pivot, plays an
important role. We seek curves that reduce to straight lines for a ∈ R, to pre-
serve the duality which holds in this case. Let us describe the curves by the
function
b
h(z, w) = g(a, w)(z − p) + p, p = ,
1−a
which is essentially the same as f (z) = az + b. The coefficient a is replaced by
g(a, w) to distribute the rotation and scaling along the w axis. In time we will
remove the dependence of h on z. The additional constraints are:

1. g(a, 0) = 1, so the curve contains (5(z), 6(z), 0), and

2. g(a, 1) = a, so the curve contains (5(az + b), 6(az + b), 1). Then the curve
joins the points z in Z1 to az + b in Z2 .
We would like the distance from the pivot to points on the curve given by h to
vary linearly with w. By the way, this also allows us to stay “close” to -coords.
Incorporating the two conditions above we obtain:

|g(a, w)| = 1 + (|a| − 1)w. (11.4)

In addition, the first two boundary values suggest an exponential function g(a, w) =
a L(a,w) , and computing L(a, w) from (11.4) yields
log(1 + (|a| − 1)w)
L(a, w) = , |a| = 1.
log(|a|)
Although the function is undefined when |a| = 1, it is easy to see that it has a limit
as |a| → 1, and we can set L(1, w) = w. This function satisfies the criteria though
it does not have a value for every w. As we know, the log function is defined only
for positive arguments. We conclude that
1
|a| > 1 ⇒ L(a, w) is defined only for w > ,
1 − |a|
1
|a| < 1 ⇒ L(a, w) is defined only for w < ,
1 − |a|
 11.5 To See C2 481

|a| = 1 ⇒ L(a, w) is defined for all w.

These conditions divide the w axis in three portions. As will be seen in the
subsequent examples, the open interval (0, 1) between the parallel planes will be
excluded from the representation of C2 . Altogether, then, the curve we obtained is
given parametrically by:
log(1+(|a|−1)w)  
b b
5 a log(|a|) z− + ,
1−a 1−a
log(1+(|a|−1)w)   
b b
6 a log(|a|) z− + ,w . (11.5)
1−a 1−a
A visual examination of the curve reveals that it is a spiral advancing over the
w-axis exponentially. It can also be described as a logarithmic spiral
 b advancing
  b over
the w axis linearly. It originates at a pivot point on the line 5 1−a , 6 1−a , w
and it is on the surface of a cone whose apex is at the same pivot. To obtain this
result we need find a w such that
log(1+(|a|−1)w)
a log(|a|) = 0.
This occurs only when the exponent is −∞ for |a| > 1 or ∞ for |a| < 1. Both of
1 . Hence, the cone’s apex, which is also
these conditions are satisfied as w → 1−|a|
this limit point of the spiral, is
  
b b 1
5 ,6 , . (11.6)
1−a 1−a 1 − |a|
This point is independent of z, so it must be the intersection of the all the curves
representing points on C2 for the function f (z) = az+b. To complete the argument
we need to verify that it is the only point of intersection. Recall that the w variation
is only along the w axis and so the value of w is the same for both r, q which are on
Z1 . The two curves, one joining the pair 3r, ar +b4 and the other joining 3q, aq +b4
for r = q, are subtracted to obtain
h(r, w) − h(q, w) = g(a, w)(r − q) = a L(a,w) (r − q),

which is zero only when a L(a,w) = 0. This is the intersection point we have already
found. A duality between points and linear complex-valued functions exists when
the pairs (z, f (z)) are joined by the curves (also called concho spirals) specified by
(11.5), as shown in Figure 11.29.
The position of the cone’s apex has properties reminiscent of -coords for R2 ,
where the horizontal position of the point representing a line reveals the slope of the
482 11. Recent Results

Figure 11.29. Representing the function f (z) = (1 + i)z by joining the pairs z, f (z) with concho
spirals given by (11.5).

line. Here, the w coordinate of the cone’s apex reveals the absolute value of the slope
rather than the slope itself. The w axis can be divided into three sections, similar to
-coords. A point whose w coordinate is less than 0 corresponds to a slope whose
absolute value is larger than 1, while a w with coordinate greater than 1 corresponds
to a slope with absolute value in the interval (0, 1). It is clear from (11.5) that points
in the interval (0, 1) must be excluded; if they are not, then 0 < 1−|a| 1 < 1 leads to
a contradiction: a < 0.
It was pointed out earlier that duality is obtained as for -coords when the
imaginary parts are zero. On the real axis the origin can be approached in only
two directions. On the complex plane the origin can be reached via an infinite
number of angles, and hence the conditions above pertain to absolute values. A more
 11.5 To See C2 483

visual explanation is that a spiral whose cone has its apex between the parallel
planes (where 0 < w < 1) only intersects one of the planes and therefore, it can
not represent a pair in C2 point. For lines with slope having absolute value 1,
the corresponding cone becomes a cylinder with a spiral that looks like a spring
coiling from −∞ to ∞ over w. To clarify further, a line in C2 is represented by a
point in R3 . The four parts (real and imaginary) of (z1 , z2 ) are mapped into three
coordinates of the representing point in R3 . Hence we get information about |a|
rather a. In short, we cannot have a full point-line duality. But what we have allows
us to visualize complex-valued functions.

11.5.4 Holomorphic Functions

Next we represent holomorphic (also called analytic) functions f (z), that is, those
whose derivative exists at every point. At each point (z, f (z)), the “complex
tangent” is computed and then mapped into its representing point, just as for
-coords. The collection of these points gives us the representing surface in R3 .
The complex tangent at a point (q, f (q)) is described by the equation g(z) =
f
(q)z − f
(q)q + f (q). In this equation a = f
(q) and b = −f
(q)q + f (q).
So by (11.6), the representing point is

(5(p), 6(p), w),

where
−f
(q)q + f (q) 1
p= ,w = .
1 − f
(q) 1 − |f
(q)|
If f (z) is linear then it is its own complex tangent and therefore is represented
by the same point.

Inversion
Inversion is a common mapping and our first holomorphic function visualization
example. It is defined by
1
f (z) = .
z
In polar coordinates, this is
1 −iθ
f (ρeiθ ) = (ρeiθ )−1 = e ,
ρ
484 11. Recent Results

showing that vector’s absolute value is inverted and the angle is reflected. Actually,
the following variation of inversion is far more interesting:

r2
f (z) = , r ∈ C. (11.7)
z
In order to visualize it, p and w are computed:

r2
f
(z) = − 2
z

q+ q r2 r2
−f
(q)q + f (q) q2
p(q) = =
1 − f
(q) 2
1 + qr 2

2r 2
q 2r 2 q
= = 2
q +r 2
2
q + r2
q2

1 1
w(q) = =  .

1 − |f (q)|  2
1 −  qr 2 

Happily, the resulting surface reveals some properties of the inversion mapping.
The symmetry between z and f (z) in (11.7) is seen in Figure 11.30. It is clear that
2 2
when f (z) = rz , z = fr(z) .
The second observation is that the points (0, 0, 0) and (0, 0, 1) are not part of
the surface but are in the closure of it. These points represent the limit of the surface
tangents as q → 0 and q → ∞. Third, the vertical section of the curve, which
2teiθ for w =
is r 1+t 1 , t ∈ R, consists of two nearly circular arcs. Fourth,
2 e2iθ 1−t −1
a complex r adds a rotation factor to the inversion, causing the surface to rotate
around the w-axis. The r 2 term doubles the angle.

Möbius Transformation

This also known as linear fractional or holomorphic transformation. It is defined

by
az + b
f (z) = .
cz + d
 11.5 To See C2 485

Figure 11.30. Inversion mapping f (z) = 1z .

The computation

ad − bc
f
(z) =
(cz + d)2
ad−bc q + aq+b
− (cq+d)2 cq+d bcq − adq + (aq + b)(cq + d)
p(q) = =
1− ad−bc (cq + d)2 − ad + bc
(cq+d)2

acq 2 + 2bcq + bd
=
(cq + d)2 − ad + bc
486 11. Recent Results

1
w(q) =  
 ad−bc 
1 −  (cq+d)2

provides the image in Figure 11.31. The ratios of the ovals’ diameters to their
intersecting curves have analytical significance that would take too long to explain
here. Observe that inversion is a specific instance of the Möbius mapping, where
a = 0; b = r 2 ; c = 1; d = 0.

Power Function
The power function f (z) = zn , n ∈ N, is yet another holomorphic mapping for
which we can calculate p and w:

Figure 11.31. Möbius mapping f (z) = 1z+1+i

iz+i .
 11.5 To See C2 487

f
(z) = nzn−1

−f
(q)q + f (q) −nq n−1 q + q n (1 − n)q n
p(q) = = =
1 − f
(q) 1 − nq n−1 1 − nq n−1
1 1 1
w(q) =

= = .
1 − |f (q)| 1 − |nq n−1 | 1 − n|q|n−1
We consider at first the cases where n ≥ 2. As can be seen in Figure 11.32, the
surfaces are not symmetric. For w ≥ 1 we get n − 1 cones that all share the same
apex (0, 0, 1). At this point, the derivative of the function is zero. On the opposite
side, where w ≤ 0, the number of cones is the same, but instead of a single apex
they merge into an almost flat disc. This disc gets bigger as w → 0, corresponding
to the function values approaching infinity. The number of cones is actually |n − 1|,
for it is determined by the derivative f
(z) used to construct the picture. Hence, for
inversion, where n = −1, there are two cones. Furthermore, the image of the inverse
function f −1 can actually be obtained from the mirror image of the function, but
this requires some care. Notice that in p(q) there is a q n term, and for 0 < n < 1,
there are n choices for this root. To include all these choices we need to include all
z for 0 < θ < 2nπ. Specifically, in Figure 11.33, n = 1/2 results in a “half-cone,”
which is exactly half the image of its inverse z2 .

Figure 11.32. The function f (z) = z4 .

488 11. Recent Results

Figure 11.33. z2 (left) versus z1/2 (right).

Actually, this introductory portion is only indicative of a much richer collec-

tion of functions that have been studied. Visualization of C2 reveals properties on
symmetries, inverses, nature of singularities and function structure in general, and
this is just the beginning.
12

Solutions to Selected Exercises

12.1 Solutions to Exercises in Chapter 2

Exercises on Page 8
1. Using compass only
(a) Construction of collinear points:
Draw a circle with any center B and any radius r. Choose a point A on
the circle. Draw a circle with center A of radius r. Both circles intersect
at C. With C as the center draw an arc on circle B to obtain the point
D. With D as the center draw an arc to obtain the point E. The points
A, B, E are collinear.

Figure 12.1. Construction of collinear points.

Proof. We shall prove that ∠ABE = 180◦ :

ABC, BCD, BDE are equilateral triangles
⇒ ∠ABC = ∠CBD = ∠DBE = 60◦
⇒ ∠ABE = 180◦ . 

A. Inselberg, Parallel Coordinates, DOI 10.1007/978-0-387-68628-8_12,

© Springer Science+Business Media, LLC 2009
490 12. Solutions to Selected Exercises

(b) Given a segment AB construct a point E such that AB = 2 · AE.

It is the same construction as the previous solution. It is obvious that if
AB = r then AE = 2 r.
(c) Bisecting a segment:
Construct point C such that AC = 2 AB (as in Exercise 1a). Construct
an arc of radius CA with center C. This will intersect the circle of radius
AB centered at A at points D and E. Draw two arcs of radius DA and EA
centered at D and E, respectively. Besides A, the two arcs intersect at
another point F , which is the midpoint of the segment AB.

Figure 12.2. Bisecting a segment.

Proof. AEF ∼ ECA (∼ is “similar”); they are both isosceles with

∠EAC
⇒ AC AE
AE = AF . We also know AC = 2 · AB, AE = AB ⇒ AC = 2 · AE
⇒ AC
AE = 2 ⇒ AB = AE = 2 · AF. 

(d) Trisecting a segment:

Construct point D such that AD = 3 · AB. Construct an arc of radius DA
with center D. It will intersect the circle of radius AB centered at A at
points M and E. Draw two arcs of radius MA and EA centered at M and
E, respectively. Besides A, the two arcs will intersect at another point F ,
which is AB
3 .
 12.1 Solutions to Exercises in Chapter 2 491

Figure 12.3. Trisecting a segment.

Proof. AEF ∼ EDA; they are both isosceles with ∠EAD

AB = AD = AE = AB
⇒ 3 = 3AB FE FA FA
⇒ AB = 3 · AF. 

(e) Constructing the point of intersection between circle C and line l (given
by two points A and B):
With centers A and B and radius AO and BO, respectively, draw two
arcs.

Case 1. If the two arcs intersect only at point O, the center O lies on
the line AB. The arc centered at A intersects the circle at P and Q, and
the arc centered at B intersects the circle at M and N. Now we have to
find the midpoints of those two arcs PQ and MN on the circle.
Let us find the midpoint of PQ with center O:
From P and Q as centers draw two arcs of radius PO. From O draw
arcs OT and OK equal to PQ, then draw two arcs with TP and KQ as
radii and with T and K as centers intersecting at R. Finally, with OR
as radius draw an arc from T ; it intersects the arc PQ at the midpoint
X, which is also the intersection point of the circle and the line AB. The
same construction has to be done on the arc MN in order to find the other
intersection point X
.

Case 2. If those two arcs intersect at O and at another point H , draw

another circle with center H with the same radius of circle O. Both
circles intersect at X and X
as required.
492 12. Solutions to Selected Exercises

Figure 12.4. Intersection between a circle and a line.

Remark. Finding the inverse point of a given point A with respect to

a given circle C with center O and radius r:

Case 1. OA > 21 · r.
Draw an arc with center A of radius AO; the arc intersects the circle at
two points C and D. Draw two arcs centered at C and D of radius CO;
they meet again at A
, which is the inverse point of A.

Case 2. OA ≤ 21 · r.
Find the point A1 such that OA1 = 2 · OA (as in Exercise 1a).
If OA1 > 21 · r find the inverse A
1 of A1 by Case 1. If OA1 ≤ 21 · r,
we must go on doubling, finding in succession A2 , A3 , . . . , Ap in OA,
where OA2 = 2·OA1 , OA3 = 2·OA2 , . . . , OAp = 2·OAp−1 = 2p ·OA
until OAp > 21 · r. We then find the inverse A
p of Ap , and again double
p times, constructing in succession the points A
p−1 , . . . , A
2 , A
1 , A
,
where OA
= 2 · OA
1 = 2p · OA
p .
Then A
is the inverse of A.

Remark. Finding the inverse circle of a given circle C with respect to

a segment AB:
Draw two arcs centered at A and B with radii AO and BO, respectively.
They intersect again at point H .
From H as the center draw an arc with radius HO that intersects the
circle C at points M and N.
 12.1 Solutions to Exercises in Chapter 2 493

Figure 12.5. Inverse point with respect to the circle.

Figure 12.6. Inverse circle with respect to segment AB.

From M and N as centers draw two arcs with radius MO; they again
intersect at point P . The circle with center P and with radius OP is the
inverse circle of C with respect to AB. (Figure 12.6)
(f) Let the lines be given by AB and A
B
. Draw any circle C and by the
preceding method find the inverse circles (with centers T1 , T2 ) of C
with respect to AB and A
B
. These circles intersect at O and Y . The
point X inverse to Y is the required point of intersection, and can be
constructed by the process explained before. That X is the required point
is evident from the fact that Y is the only point that is inverse to a point
of both AB and A
B
. Hence the point X inverse to Y must lie on both AB
and A
B
.
494 12. Solutions to Selected Exercises

Figure 12.7. Intersection between two lines.

2. Using a single circle and a straightedge

Given an angle A, provide a construction for finding an angle = 21 A:
Put the center of the circle at point A; call the intersection points between them
B and D. Continue AB until it intersects the circle at C. Then ∠BCD = ∠BAD 2 .

Figure 12.8. Half-angle.

3. Transform a rectangle to a square of equal area: +

We have rectangle ABCD; AB = b, AD = a ⇒ AC = DB = a 2 + b2 .
Draw a circle centered at B of radius a. Continue AB until it intersects the circle
at point E; then EA = a + b. Bisect the segment EA and call the midpoint F
(Exercise 1c). Draw a circle with center F of radius FE. Draw an arc centered
√ AC intersecting the circle at M. Then EAM is right-angled
at A of radius
⇒ EM = 2ab.
Now we have to construct a square with diagonal EM:
Draw two circles centered at M and E of radius EM; they intersect at V and
U . Using the ruler construct the segment UV; it intersects segment EM on W .
 12.1 Solutions to Exercises in Chapter 2 495

Figure 12.9. The rectangle and the square have the same area.

Draw a circle with center W of radius WE; it intersects segment UV at P and

Q. We have a square EQMP of area = 2 · ab.
√ area 2 · A:
4. Given a square with area A, construct a square with
We know that the diagonal of the square FBCD is 2A. Draw a circle at center
C of radius DC intersecting points M and N. The square BDMN has area 2 · A.

Figure 12.10. Doubling the area of a square.

5. To prove that a triangle having two equal angle bisectors is isosceles, it suffices
to prove the following theorem.
Theorem. In a triangle having two different angles, the smaller angle has
the longer bisector.
Proof. In ABC, AE and BD are the angle bisectors with angles ∠B > ∠A.
Draw a circle through the three points A, D, and B that intersects AE at a point
F . The angles ∠DAF and ∠DBF must be equal (i.e., 21 · A), since they subtend
496 12. Solutions to Selected Exercises

the same arc DF. By construction, angles ∠DAB < ∠FBA ⇒ that the arcs
DB < AF ⇒ the chords DB < AF < AE. 
The above result is sometimes called the “Steiner–Lehmus theorem.” The
problem was first posed in 1840 by Lehmus to Sturm, who passed it on to
the famous geometer Steiner, who apparently gave the first satisfactory proof.
Papers appeared in various journals in 1842, 1844, 1848, 1854–1864, and many
others since. Henderson wrote surveys in Scripta Mathematica, Vol. 21, 1955,
223–232, 309–312, and there are others. The proofs given are indirect, i.e.,
based on the contrapositive, which has upset some people, though Euclid him-
self used this method of proof extensively. There exist several (60+) different
proofs of this result. The above, after some modifications, is mine. Enjoy!

Figure 12.11. “Steiner–Lehmus theorem.”

Exercises on Page 13
1. The line l and point C determine a plane πl in R3 that intersects the hemisphere
in a great circle, c . This semicircle is the image of the line.
2. Let θ be the angle between two lines, l1 , and l2 , on the plane. The angle
between planes πl1 and πl2 is also θ, and therefore, so is the angle between the
semicircles c 1 and c 2 , which are the images of the two lines on the hemisphere.
3. Circles are mapped to circles. An intersection of a cone with a hemisphere is a
circle.
4. The upper line is called l1 and the other is l2 . Using the ruler draw any other line
l3 . Choose two points, point A on line l1 , point K on line l3 . Draw the segment
AK. The intersection point with l2 is C. Choose another point A
on l1 . Draw
KA
. The intersection with l2 is C
. Draw A
R and continue until it intersects
 12.1 Solutions to Exercises in Chapter 2 497

the line l3 at M. Continue RC

until intersection with l3 at Q. QC intersects
AM at R
. On the line RR
there is also the point T . (Desargues’s theorem)
5. The same as before, but the lines should be very close so that all the points can
be drawn with our short ruler. Note to student: Try to give a better construction
than this!

Exercises on Page 16
1. Ideal points satisfy the axioms of points (Axioms 1 + 2):

• Two distinct ideal points are on a unique ideal line.

• Three distinct noncollinear ideal points are in a unique ideal plane.

Figure 12.12. Exercise 4.

2. Augmented lines satisfy the axioms of lines (Axioms 3 + 4):

• Let (l, L∞ ), (p, P∞ ) be two coplanar augmented lines. If l  p they have

the same direction ⇒ the same ideal point ⇒ L∞ = P∞ . If l  p they
intersect at an ordinary unique point.
• Let (l, L∞ ), (π, p∞ ) augmented line and augmented plane such that
l ⊂ π. If l  π ⇒ L∞ ∈ p∞ ⇒ (l, L∞ ) ∩ (π, p∞ ) = L∞ . If l  π they
intersect at a unique point.
3. Augmented planes satisfy the axioms of planes (Axioms 4 + 5):

• Let (π1 , p∞ ) = (π2 , l∞ ). If π1  π2 ⇒ for every line on π1 there is a

parallel line on π2 ⇒ π1 and π2 have the same directions ⇒ (π1 , p∞ ) ∩
(π2 , l∞ ) = p∞ . If π1  π2 they intersect at a unique line, as known.
• The line-on-plane relation is already proved above.
498 12. Solutions to Selected Exercises

4. Ideal plane π∞ satisfies the axioms of a plane:

For this we have to add two more axioms:
(a) two distinct spaces are on a unique plane.
(b) a space and a plane are on a unique line.
• Let (R, π∞ ) be a space with ideal plane and (α, l∞ ) ⊂ (R, π∞ ).
(R, π∞ ) ∩ (α, l∞ ) = (p, A∞ ) by axiom (b) ⇒ π∞ ∩ l∞ = A∞ .
• Let (π∞ = α∞ ) be two distinct ideal planes ⇒ they are associated to
two distinct spaces (not parallel). By Axiom (a) they intersect at a unique
plane, and this plane has a unique ideal line l∞ = π∞ ∩ α∞ .

Exercises on Page 21
A proof for the duals using only the projective 3-space axioms:
1. Two planes are on a line; if they are distinct they are on a unique line according
to Axiom 5.
The dual is: Two points are on a line (Axiom 1).
2. Axiom 4: A plane π and a line not on π are on a point. Its dual is: A point P
and a line are on a plane; choose two distinct points P1 , P2 on ; P , P1 , P2
are noncollinear; according to Axiom 2, they are in a plane.
3. The sets of points and lines satisfy the projective plane axioms.
Proof.
1. Any two distinct points A, B on a unique line satisfy Axiom 1.
2. If p, q are distinct lines, there is at least one point on both; there is one
point on both; this satisfies Axiom 3.
3. There are at least three distinct points on any line ⇒ there are two distinct
points on a line; satisfies Axiom 1.
4. Not all points are on the same line. This means that every three
noncollinear points are on a plane; satisfies Axiom 2. 

4. Let P be any point. By Axiom 1, every point on π determines a line together

with P , which obviously passes through P . Hence all the points of π lie on the
n lines incident with P . Each of these lines is incident with n − 1 points other
than P , so there are n(n − 1) of them, and including P , we get n2 − n + 1
points in π.
5. The algorithm:
Step (kk = 11): Construct a matrix Mn×(n−1) : the first row R1 contains the
numbers 2, . . . , n, every other row Ri contains the next (n − 1) numbers.
The first n lines connect the number 1 with the numbers of Ri (i = 1, . . . , n).
 12.1 Solutions to Exercises in Chapter 2 499

Step (kk = 22): Construct a matrix A(n−1)×(n−1) : A = [M\R1 ]T . Define the

columns of A : C1 , . . . , Cn−1 . The next (n − 1) lines connect the number 2
with every row of A.
Step (kk ≥ 33): Define A
: the numbers on the diagonal of A will be the first
row of A
. For each of the remaining Ci , the columns are permuted by one
element in cyclic order. The next (n − 1) lines will connect k with the rows
of A
.
Do this step until k = n. The process is clarified with the example below.

Figure 12.13. Finite geometry of five points.

500 12. Solutions to Selected Exercises

6. The 5-point Geometry (21 points):

2 3 4 5
6 10 14 18 6 11 16 21
6 7 8 9
7 11 15 19 7 12 17 18
1 −→ 10 11 12 13 2 −→ 3 −→
8 12 16 20 8 13 14 19
14 15 16 17
9 13 17 21 9 10 15 20
18 19 20 21

6 12 14 20 6 13 16 19
7 13 15 21 7 10 17 20
4 −→ 5 −→
8 10 16 18 8 11 14 21
9 11 17 19 9 12 15 18
The 6-Point Geometry (31 points):
2 3 4 5 6
7 12 17 22 27 7 13 19 25 31
7 8 9 10 11
8 13 18 23 28 8 14 20 26 27
12 13 14 15 16
1 −→ 2 −→ 9 14 19 24 29 3 −→ 9 15 21 22 28
17 18 19 20 21
10 15 20 25 30 10 16 17 23 29
22 23 24 25 26
11 16 21 26 31 11 12 18 24 30
27 28 29 30 31

7 14 21 23 30 7 15 18 26 29 7 16 20 24 28
8 15 17 24 31 8 16 19 22 30 8 12 21 25 29
4 −→ 9 16 18 25 27 5 −→ 9 12 20 23 31 6 −→ 9 13 17 26 30
10 12 19 26 28 10 13 21 24 27 10 14 18 22 31
11 13 20 22 29 11 14 17 25 28 11 15 19 23 27

Exercises on Page 29
1. Create 24 distinct permutations of four symbols A, B, C, D. According to the
cross ratio,
AC BD
(A, B; C, D) = (B, A; D, C) = (C, D; A, B) = (D, C; B, A) = · ,
BC AD
AD BC
(A, B; D, C) = (B, A; C, D) = (D, C; A, B) = (C, D; B, A) = · ,
BD AC
AB CD
(A, C; B, D) = (C, A; D, B) = (B, D; A, C) = (D, B; C, A) = · ,
CB AD
CA BD
(C, B; A, D) = (B, C; D, A) = (A, D; C, B) = (D, A; B, C) = · ,
BA CD
 12.1 Solutions to Exercises in Chapter 2 501

Figure 12.14. Finite geometry of six points.

DA BC
(D, B; A, C) = (B, D; C, A) = (A, C; D, B) = (C, A; B, D) = · ,
BA DC
BA CD
(B, C; A, D) = (C, B; D, A) = (A, D; B, C) = (D, A; C, B) = ·
CA BD
⇒ There are six sets of permutations having the same cross ratio values. 
2. The cross ratio is invariant under a prespectivity with respect to an ideal point
P∞ :
Given two lines

l1 (A, B, C, D) ∧∞ l2 (A
, B
, C
, D
),
P

P∞ is an ideal point ⇒ AA
 BB
 CC
 DD
.
502 12. Solutions to Selected Exercises

Figure 12.15. Exercise 2.

Construct a line l3  l1 ; then AB = A

K, BC = KM, CD = MN
⇒ A
B
K ∼ A
C
M

A
B
AB
⇒ =
A
C
AC
A
C
− B
C
AC − BC
⇒

=
AC AC
B
C
BC
⇒1− =1−
A
C
AC
B
C
BC
⇒

= .
AC AC
In the same way, BD B
D

AD = A
D
, so the cross ratio is invariant:
⇒ (A, B; C, D) = (A
, B
; C
, D
).

3. Theorem 2. A projectivity is determined by the correspondence between two

triples of distinct points.
Lemma 1. It is possible, by a sequence of not more than three prespectivities,
to relate any three distinct collinear points to any other three distinct collinear
points.

Proof of Lemma 1. If the two triads ABC and A

B
C
are on distinct lines, let
R, S, C0 denote the points where the respective lines AA
, BB
, BA
meet CC
.
Then ABC and A
B
C
are related by the sequence of two perspectivities
 12.1 Solutions to Exercises in Chapter 2 503

ABC ∧ A
BC0 ∧ A
B
C
.
R S

(If A and A
coincide, we merely use the perspectivity from S.)
If the two triads are on one line, we use a quite arbitrary perspectivity
ABC∧ A1 B1 C1 to obtain a triad on another line and then relate A1 B1 C1 to
A
B
C
by the above construction. 

Figure 12.16. Lemma 1.

Lemma 2. A projectivity having more than two invariant points can only be
the identity.58
Proof of Theorem 2. Suppose we are given three points A, B, C of one range
and corresponding points A
, B
, C
of the other. We proceed, as in the proof
of Lemma 1, to construct a projectivity X ∧ X
such that ABC ∧ A
B
C
.
To establish the uniqueness of this projectivity, we have to prove that a different
construction (e.g., by joining AB
instead of BA
) would yield the same X
for
a given X. Suppose one construction gives ABCX ∧ A
B
C
X
, while another
gives ABCX ∧ A
B
C
X1
. Then by combining the two constructions we obtain

A
B
C
X
∧ A
B
C
X1
.

This combined projectivity has three invariant points A

, B
, C
.
Hence by Lemma 2, X1
must coincide with X
. 

58
H.S.M. Coxeter. The Real Projective Plane, pp. 40–42.
504 12. Solutions to Selected Exercises

Figure 12.17. Theorem 2.

Lemma 3. Let l ∧ m ∧ n, with l = n. Then there are a line m

and points
O P

O
P

O
∈ n, P
∈ l such that l ∧ m
∧ n gives the projectivity l ∧ n.

Proof of Lemma 3. Let l, m, n, O, P be given. Let A, A

be two points on l,
and let

AA
∧ BB
∧ CC
.
O P

Let OP intersect n at O
. Since we assumed O, P , l ∩ n = X are not collinear,
O
= X, so O
∈ l. Draw O
A, O
A
, and let them intersect PC, PC

at D, D
, respectively. Now corresponding sides of the triangles ABD and
A
B
D
intersect at O, P , O
, respectively, which are collinear; hence the
lines joining corresponding vertices are concurrent. Thus m1 passes through
the point Y = l ∩ m.
Thus m1 is determined by D and Y , so as A
varies, D
varies along the line
m1 . Thus our original projectivity is equal to the projectivity
O
P
l ∧ m1 ∧ n.
 12.1 Solutions to Exercises in Chapter 2 505

Performing the same argument again, we can move P to P

= OP ∩ l, and
find a new line m
such that
O
P

l ∧ m
∧ n

gives the original projectivity. 

4. Theorem 3. A projectivity can be expressed by at most two perspectivities.

Proof. A projectivity was defined as a composition of an arbitrary chain of
perspectivities. Thus it will be sufficient to show, by induction, that a chain of
length n > 2 can be reduced to a chain of length n − 1, i.e., it will be sufficient
to prove that a chain of three perspectivities can be reduced to a composition
of two perspectivities.
The argument of Lemma 3 actually shows that the line m can be moved so as
to avoid any given point. Thus one can see easily that it is sufficient to prove
the following:
Let
P Q R
∧m∧n∧o

be a chain of three perspectivities, with = o. Then the resulting projectivity

∧ o can be expressed as a product of at most two perspectivities.
If two of the four lines l, m, n, o coincide, then there are only two perspectives.
Second, we have either m∧ o, in which case we are done, or n can be moved
(Lemma 3) so that the centers of the perspectivities m∧ n and n∧ o are at o, m
respectively. So we have
P Q R
∧m∧n∧o

with , m, n, o all distinct, Q ∈ o, and R ∈ m. Let X = ∩ m, Z = n ∩ o,

and draw h = XZ. We may assume that X ∈ o (indeed, we could have moved
Q
m, by Lemma 3 to make X ∈ o). Therefore Q ∈ h. Project m ∧ h, and let
Q
BB
∧ HH
.
Now CDH ∧ C
D
H
. Corresponding sides intersect at Q, R; hence by
Z

Desargues’s theorem, the remaining corresponding sides intersect at a point

N
N on QR. Thus N is determined by DH alone, and h ∧ o.
506 12. Solutions to Selected Exercises

Figure 12.18. Theorem 3.

Figure 12.19. Pappus’s theorem.

Similarly, ABH ∧ A
B
H
, so using Desargues’s theorem again, we find that
X

AH ∩ A
H
= M ∈ PQ. Hence, ∧ h. So we have the original projectivity
M

represented as a composition of two perspectivities

M N
∧ h ∧ o. 
 12.1 Solutions to Exercises in Chapter 2 507

Exercises on Page 43
1. Analytic proof for Pappus’s theorem:
Assign homogeneous coordinates as follows:
A1 = (1, 0, 0), A2 = (0, 1, 0), C3 = (1, 1, 1), B3 = (0, 0, 1).
On line A2 C3 choose point B1 = (1, p, 1); on line A1 C3 choose point
B2 = (m, 1, 1).
C1 = A2 B3 ∩ A1 B1 = (0, p, 1),

C2 = A1 B3 ∩ A2 B2 = (m, 0, 1),
     

 1 p   1 1   p 1 
B1 C2 =   ,  , = [−pm, 1 − m, p],
m 0   m 1   0 1 
     

 0 p   0 1   p 1 
B2 C1 =   ,  , = [−pm, −m, p − 1],
m 1   m 1   1 1 

A3 = B2 C1 ∩ B1 C2
     
 −pm 1 − m   −pm p   1 − m p 
=  ,
  −pm , ,
−pm −m p − 1   −m p−1 

A3 = (pm, pm, p + m − 1).

 1 
 1 1 
Checking the determinant we find that  0 0 1  =0
pm pm p+m−1
⇒ The points A3 , B3 , C3 are collinear. 
2. Ideal plane of (R 3 , p∞ )
We already know that every ideal point is represented in homogenous coordi-
nates as (P1i , P2i , P3i , 0). A plane π ∈ (R 3 , p∞ ) is C1 X1 + C2 X2 + C3 X3 =
C0 . In homogenous coordinates we can write it as C1 X1 + C2 X2 + C3 X3 +
C4 X4 = 0.
Three ideal points are collinear if the determinant = 0:
 
 X1 X2 X3 X4 
 
 P11 P21 P31 0 
 =0
 P 
 12 P22 P32 0 
 P 0 
13 P23 P33
⇒ C · X4 = 0 ⇐⇒ X4 = 0.
Hence, an ideal plane is represented as [C1 , C2 , C3 , 0].
508 12. Solutions to Selected Exercises

Exercise on Page 46
The set of all linear fractional transformations
ax + b
T (x) =
cx + d
with the composition operation form a group.
Proof. Closure property:
 y+b1 
a ac11y+d +b
T (x) ◦ T (y) =  a y+b 1 
c c11y+d11 + d

(aa1 + bc1 )y + (ab1 + bd1 ) My + N

= = ∈ G.
(ca1 + dc1 )y + (cb1 + dd1 ) Vy + U
Associativity:

(T (x) ◦ T (y)) ◦ T (z)


(aa1 + bc1 y + (ab1 + bd1 )
= ◦ T (z)
(ca1 + dc1 )y + (cb1 + dd1 )
 z+b2 
(aa1 + bc1 ) ac22z+d + (ab1 + bd1 )
= 2
 a2 z+b2 
(ca1 + dc1 ) c2 z+d2 + (cb1 + dd1 )

(aa1 a2 + bc1 a2 + ab1 c2 + bd1 c2 )z + aa1 b2 + bb2 c1 + ab1 d2 + bd1 d2

= ,
(ca2 a2 + dc1 a2 + cc2 b1 + dd1 c2 )z + ca1 b2 + dc1 b2 + cb1 d2 + dd1 d2
T (x) ◦ (T (y) ◦ T (z))
⎛   1 a2 +b1 c2 )z+a1 b2 +b1 d2  ⎞
z+b2 
a1 ac22z+d + b 1 a (a
(c a +d c )z+c b +d d + b
= T (x) ◦ ⎝  a z+b2  =  (a 1a 2+b 1c 2)z+a1b2+b1d2  ⎠
c1 c2 z+d2 + d1
2 2
c (c1 a2 +d1 c2 )z+c 1b 2+d 1d 2 + d
1 2 1 2 1 2 1 2

(aa1 a2 + bc1 a2 + ab1 c2 + bd1 c2 )z + aa1 b2 + bb2 c1 + ab1 d2 + bd1 d2

=
(ca2 a2 + dc1 a2 + cc2 b1 + dd1 c2 )z + ca1 b2 + dc1 b2 + cb1 d2 + dd1 d2
⇒ (T (x) ◦ T (y)) ◦ T (z) = T (x) ◦ (T (y) ◦ T (z)).

Identity:
1x + 1
∃e= ∈ G : ∀ a ∈ G e ◦ a = a ◦ e = a.
1x + 1
 12.2 Solutions to Exercises in Chapter 3 509

Inverse:
a1 x + b1 c1 x + d1
∀a= ∈ G ∃ a −1 = ∈ G : a ◦ a −1 = a −1 ◦ a = e
c1 x + d 1 a1 x + b1
⇒ the projective transformations form a group. 

12.2 Solutions to Exercises in Chapter 3

Exercises on Page 56
1. Given a line

: x2 = mx1 + b,

we shall verify the corresponding point of ,


d b
¯ : , , m = 1.
1−m 1−m

Choose two arbitrary points on l: A1 = (a1 , ma1 +b) and A2 = (a2 , ma2 +b).
The point A1 is represented in -coords as the line
ma1 + b − a1
Ā1 : x2 = x1 + a 1 .
d
The point A2 is represented in -coords as the line
ma2 + b − a2
Ā2 : x2 = x1 + a2 .
d
The intersection point of these lines is

d b
¯ : , , m = 1.
1−m 1−m

This point is not dependent on a1 and a2 . Therefore, every two points on are
¯
represented as two lines in the dual plane that always intersect at .
−1
2. Let us check whether [B A] = I :
k

⎡ ⎤
0 −d −1
B −1 A = ⎣ −d −d 0 ⎦ ∀k ∈ N : [B −1 A]k = I.
0 d 2 0
510 12. Solutions to Selected Exercises

3. Let us check whether [A−1 B]k = I :

⎡ ⎤
0 − d1 − d12
⎢ ⎥
A−1 B = ⎣ 0 0 1
⎦ ∀k ∈ N : [A−1 B]k = I.
d2
−1 0 − d1

There are some dualities that do not repeat themselves, and this is one of them.

Exercises on Page 61
1. Let us look at the point P (−1, 1) and the line : y = 2x + 4. The translation of
the origin O to the point (2, 5) is represented in -coords as a rotation. We can
see that the coordinates are moved and rotated to become the dotted ones. The
point P (−1, 1) −→ P
(−3, −4).
The line : y = 2x + 4 −→
y = 2x + 3.
In -coords ¯ = (−6, −4) −→ ¯
= (−6, −3).
The distance d becomes d
:
,
d
= (y1 − x1 )2 + d 2 .

2. Without moving the coordinates, moving only the objects P and , the reflection
about the x1 and x2 axes: is reflected to 1 : y = −2x − 4 about  the x axis
and to 2 : y = −2x + 4 about the y axis. In -coords, ¯1 = 2, − 43 and
 
l¯2 = 2, 43 . The point P is reflected to P1 (−1, −1) about the x axis, and to
P2 (1, 1) about the y axis. For -coords this is shown in Figure 12.21.
12.2 Solutions to Exercises in Chapter 3 511

Figure 12.20. Translation ↔ rotation.

512 12. Solutions to Selected Exercises

Figure 12.21. Reflection about x axis.

3. The rotation of the axes about the origin by an angle: θ(m = tan(θ))
In -coords,
 

d
d
X1 = , 0 , X2 = ,0 .
1 − tan(90 + θ) 1 − tan(θ)

Cartesian coordinates -coords

 
X-axis −→ X
: y = tan θx X̄1 −→ X̄1
= 1−tan
d
θ , 0
 
−1 x
Y -axis −→ Y
: y = tan X̄2 −→ X̄2
= d ,0
θ 1
1+ tan θ
The distance d becomes d
:
 
 d(1 + tan2 θ) 
 
θ = 45, d
=  .
 1 − tan θ 
2

Choose θ = 30◦ . Then, X̄1 −→ X̄1

(14.19, 0) and X̄2 −→ X̄2
(2.19, 0). If we
don’t move the coordinates but only the objects P and , then
 

cos θ sin θ −1
P = ·
− sin θ cos θ 1
  
cos 30 sin 30 −1 −0.366
= · = ,
− sin 30 cos 30 1 1.366

: y = 0.66x + 2.142 −→ ¯
= (17.67, 6.309).
12.2 Solutions to Exercises in Chapter 3 513
514 12. Solutions to Selected Exercises

4. The composition of the translation of the origin to the point (2, 5) followed by
the rotation of the axes by angle 30◦ yields
P = (−1, 1) −→ P
= (−3, −4) −→ P

= (−4.59, −1.96),

: y = 2x + 4 −→
: y = 2x + 3 −→

: y = 0.659x + 1.6,

¯ = (−6, −4) −→ ¯
= (−6, −3) −→ ¯

= (17.59, 4.69).

12.3 Solutions to Exercises in Chapter 4

Exercises on Page 73
1. Consider a line ⊂ RN whose description, by a set of linear equations, contains
xi = ci , a constant, for some i. Such a line can be represented by N −1 indexed
points. Let S be the set of indices for which the variables are not constant, where
S = (i1 , . . . , iM ), and let S̃ be the complementary set of indices, so
j j
ij : xj = mi xi + bi ∀i < j ∈ S = (i1 , . . . , iM ),
0k : xk = bk ∀k ∈ S̃ = (iM+1 , . . . , iN ).
In homogenous coordinates,

¯ij = (j − i, bi , 1 − mi )
j j
∀i, j ∈ S = (i1 , . . . , iM ); i = j,
¯0k = (k − 1, bk , 1) ∀k ∈ S̃ = (iM+1 , . . . , iN ).
2. Consider the plane
π123 : c1 x1 + c2 x2 + c3 x3 = c0 .
 12.3 Solutions to Exercises in Chapter 4 515

Let us prove that

π̄123 = (c2 + 2c3 , c0 , c1 + c2 + c3 ).

The plane π123 intersects at least two of the planes π1 : x1 = 0, π2 : x2 = 0,

and π3 : x3 = 0.
Let us find the representation of the intersection lines in homogeneous
coordinates:
c2 c0
L1 = π123 ∩ π1 #⇒ c2 x2 + c3 x3 = c0 #⇒ 23 : x3 = − x2 + ,
c3 c3
¯23 = (2c3 + c2 , c0 , c3 + c2 ) and ¯01 = (0, 0, 1) (const.),
     

 c0 c3 + c2   c3 + c2 2c3 + c2   2c3 + c2 c0 
L1 =     ,
0 1 , 1 0   0 0 

= [−c0 , 2c3 + c2 , 0],

c1 c0
L2 = π123 ∩ π2 #⇒ c1 x1 + c3 x3 = c0 #⇒ 13 : x3 = − x1 + ,
c3 c3
¯13 = (2c3 , c0 , c3 + c1 ) and ¯02 = (1, 0, 1) (const.),
     

 c0 c3 + c1   c3 + c1 2c3   2c3 c0 
L2 =     , 
0 1 , 1 1   1 0 

= [c0 , c1 − c3 , −c0 ],
c1 c0
L3 = π123 ∩ π3 #⇒ c1 x1 + c2 x2 = c0 #⇒ 12 : x2 = − x1 + ,
c2 c2
¯12 = (c2 , c0 , c2 + c1 ) and ¯03 = (2, 0, 1) (const.),
     

 c0 c2 + c1   c2 + c1 c2   c2 c0 
L3 =     ,  ,
0 1 , 1 2   2 0 

= [c0 , 2c1 + c2 , −2c0 ],

L1 ∩ L2 = π̄123
     
 2c + c2 0   0 −c0   −c0 2c3 + c2 
=  3 , ,
c 1 − c3 −c0   −c0 c0   c0 c1 − c3 
516 12. Solutions to Selected Exercises

= (−c0 (2c3 + c2 ), −c02 , −c0 (c1 + c2 + c3 ))

= (2c3 + c2 , c0 , c1 + c2 + c3 ),

L1 ∩ L3 = π̄123
     
 2c + c2 0   0 −c0   −c0 2c3 + c2 
=  3 ,
  −2c0
,
  c0
2c1 + c2 −2c0 c0 2c1 + c2 

= (−2c0 (2c3 + c2 ), −2c02 , −2c0 (c1 + c2 + c3 ))

= (2c3 + c2 , c0 , c1 + c2 + c3 ),

L2 ∩ L3 = π̄123
     
 c − c3 −c0   −c0 c0   c0 c1 − c3 
=  1 , ,
2c1 + c2 −2c0   −2c0 c0   c0 2c1 + c2 

= (c0 (2c3 + c2 ), c02 , c0 (c1 + c2 + c3 ))

= (2c3 + c2 , c0 , c1 + c2 + c3 ) 
3. Algorithm for determining the point ¯ij , for any distinct pair i, j of a line
given all the points ¯i−1,i , i = 1, 2, . . . , N:

(a) Choose a point P1,1 on the X̄1 axis.

(b) Construct the line through Pk,1 and ¯k,k+1 for k = 1. Define the
intersection point with the X̄k+1 axis as Pk+1,1 . Repeat this step for
k = 2, . . . , N. The polygonal line through the set of points Pk,1 defines
a point on the original line .
(c) Choose another point P1,2 on the X̄1 axis. Repeat step (b) for all the
points Pk,2 , k = 1, . . . , N. We get another polygonal line through Pk,2 ,
k = 1, . . . , N.
(d) Construct the line L1 through Pi,1 , Pj,1 and the line L2 through Pi,2 , Pj,2 .
The intersection between L1 and L2 is the required point l¯i,j .

Complexity: O(N).
4. For lines where some of the m’s are close to 1 we make a transformation of
variables xi
= ei xi , where ei = 1 ∀i ∈ S̃ (for the constants). For the variables
in S the linear equation is

j
ij : xi = mi xj + bj .
 12.3 Solutions to Exercises in Chapter 4 517

The remaining ei are developed recursively. Starting at the root index i, set
ei = 1 and visit the remaining nodes in order, setting ej = +1 to ensure that
the slope m
i of the line
j

ij : xi = m
i xj + bj

is negative. Here m
i = ei ej mi and bj
= ei bj . Since the slope is negative, it
j j

follows that ¯
ij is between X̄i−1 and X̄j −1 .

Exercises on Page 86
1. (a) For the example, choose as a start variable x1 .
518 12. Solutions to Selected Exercises

(b) The following table represents the lines ij :

Coordinates Point 1 Point 2 i j xi,j yi,j hi,j Point 3

X1 −4 2 1 3 2 1 0 0
X2 2 2 1 4 7.5 2 1 2
X3 −3 3 1 5 2.4 −0.7 1 1
X4 −1.6 2 3 9 6 −1 0 0.8
X5 1.5 −2.5 5 8 5.5 0.25 1 − 76
X6 −2 −2 0 2 1 2 1 −2
X7 −0.5 −0.5 0 6 5 −2 1 −0.5
X8 −1 3 0 7 6 −0.5 1 2
3
X9 −4 2 0

X13 and X39 are ideal points.

(c) Find a point P3 on the same line segment between P1 and P2 .

 12.3 Solutions to Exercises in Chapter 4 519

(d)+(e) Find ¯
and the points P1
and P2
in the x
system:

Coordinates Point 1 Point 2 i j xi,j yi,j

x1
−4 2 1 3 1 −0.5
x2
2 2 1 4 1.87 −0.25
x3
3 −3 1 5 2.4 −0.7
x4
2 −1.6 3 9 3 1.5
x5
1.5 −2.5 5 8 5.5 0.25
x6
−2 −2 0 2 1 2
x7
−0.5 −0.5 0 6 5 −2
x8
−1 3 0 7 6 −0.5
x9
−4 2
520 12. Solutions to Selected Exercises

2. This is an example of the tree enumeration for N = 4.

(a) Only four distinct kinds of trees are possible.
(b) The number of distinct rooted trees is shown below each tree. There are
64 distinct rooted trees: N N −1 −→ 44−1 = 64.

Exercise on Page 100

1. Let us show the translation of the point P to a new position P
and the line
unrotated on P
:

12.4 Solutions to Exercises in Chapter 5

Exercises on Page 122
1. (a) For example, look at the undirected graph of N = 6.
We have three permutations whose every two vertices are adjacent:
⎧
⎨ P1 → (1, 2, 6, 3, 5, 4)
⎪
P2 → (2, 3, 1, 4, 6, 5)
⎪
⎩
P3 → (3, 4, 2, 5, 1, 6)
Notice that the relationship between the permutations is that each time,
we added 1 (mod N) to the previous permutation. Notice that 6 + 1
≡ 1(mod 6).  
(b) An algorithm for computing the O N2 permutations:
Input: a two-way queue of N vertices (by order).
 12.4 Solutions to Exercises in Chapter 5 521

 
Output: if N is even, N2 × (N) matrix (called “Permut [i] [j ]”), if
 
N is odd, N2 + 1 × (N) matrix; every line in the matrix is
a permutation.

k=1
Permut[1][k]=dequeue-head
Permut[1][k++]=dequeue-head
while the queue is not empty do
Permut[1][k++]=dequeue-head
Permut[1][k++]=dequeue-tail
if N is even
for i=2 to N2
for j =1 to N
Permut[i][j ]=Permut[i − 1][j ]+1 (mod N)
if N is odd
for i=2 to N2 + 1
for j =1 to N
Permut[i][j ]=Permut[i − 1][j ]+1 (mod N)
(c) Research problem.
(d) Research problem.
2. Given three noncollinear points in R3 , an algorithm for constructing a set of
planar coordinates for the plane determined by the three points A, B, C is as
follows:

i. First check whether one of the lines AB, BC, AC lies on one of the principal
2-planes; if so, this line determines a planar axis y 1 . In order to determine
522 12. Solutions to Selected Exercises

the other planar axis, assume that AB lies on x2 x3 and determine y 1 , and
AC intersects plane x1 x2 at a point D; AB intersects x1 at E, so the other
planar axis y 2 is determined by DE.
ii. If two lines already lie on the principal 2-planes, they define both planar
axes.
iii. If the three points have one common coordinate, for example, they have
the same x3 , it means that the plane is parallel to the x1 x2 plane; find the
point of intersection of the lines AB, AC, BC with the planes x1 x3 and
x2 x3 to obtain four points: G and F on x1 x3 determine y 1 , and D and M
on x2 x3 determine y 2 .
iv. Otherwise, if the points A, C determine the line and B, C determine
the line
, the points are not collinear; therefore ¯ıj and ¯
ıj are not
coincident. Find the intersecting points of and
with the x1 x2 plane
(points E and D) and with the x2 x3 plane (points F and G). The lines DE
and FG determine the y 1 and y 2 axes respectively. In parallel coordinates
1 and Ȳ 2 , which complete the representation of the
we get the points Ȳ12 23
plane determined by A, B, and C.
3. Given a plane π parallel to one of the orthogonal axes, for example to the axis
x3 , an algorithm for constructing planar coordinates for the plane π, assuming
π  x3 , is as follows:

i. If π intersects x1 and x2 , find the intersecting points with x1 and x2 (points

A and B); they determine planar axes y 1 and Ȳ 2 = X̄3 .
ii. If π  x1 as well, then find only the intersection point with x2 ; this point
will be the origin of the coordinate system, in -coords the planar axes
coincident the ideal points X̄1 and X̄3 .
4. Given a set of planar coordinates y 1 and y 2 for the plane π, here is an algorithm
for constructing a plane π
⊥ π: rotate y 2 about y 1 by 90◦ . Next let

x3 = m3 x2 + b3 ,
y2 :
x1 = 0.

After rotating by 90◦ ,

 b3
x3 = − m13 x2 − m ,
y
2 : 3
x1 = 0.

Then π
is determined by y 1 and y
2 .
 12.4 Solutions to Exercises in Chapter 5 523

5. Here is an algorithm for determining the intersection line of two planes: given
two planes π and π
with their planar coordinates y 1 , y 2 and y
1 , y
2 , find the
intersecting points P1 = y 1 ∩ y
1 and P2 = y 2 ∩ y
2 ; the intersection line is
determined by P1 and P2 . For a different construction, see Section 5.3.4.
6. Given a plane π and a line , here is an algorithm for determining whether
⊂ π: in -coords π is represented by two axes Ȳ 1 and Ȳ 2 ; is represented by
two points ¯12 and ¯23 . If ⊂ π, then can be represented, relative to the planar
coordinates y 1 y 2 , by a single point η12 . Let P1 = ∩ x1 x2 , P2 = ∩ x2 x3 ; in
-coords connect the intersections of P1 with Ȳ 1 and Ȳ 2 with a straight line;
the same is done with the intersections of P̄2 with Ȳ 1 and Ȳ 2 . The intersection
of these two straight lines is the point η12 . If the three points ¯12 , ¯23 , η12 are
collinear, then ⊂ π.
7. Given a plane π and a point P , here is an algorithm for determining whether
P is below or above π:

i. If π is not parallel to any of the orthogonal axes, let P

be the projection
of P on the x1 x2 plane. Let P

be the projection of P on the plane π .

In -coords construct the polygonal line Ē through ȳ231 having the same x
1

coordinate as P̄ and x3 = 0. Next construct the polygonal line P̄ having

the same x1 and x2 as P̄ and the same y 2 as Ē. If the x3 coordinate of P̄
is higher than the x3 coordinate of P̄

, then P is above π .
524 12. Solutions to Selected Exercises

ii. If π is parallel to one of the orthogonal axes, assume π  x3 ; let P

be
the projection of P on x1 x2 plane and let P

be the point on y 1 with the

same x1 coordinate as P
. If the x2 coordinate of P̄ is higher than the x2
coordinate of P̄

, then P is above π.
iii. If π is parallel to one of the principal 2-planes, assume π  x1 x3 ; let
A be the origin of y 1 y 2 . If the x2 coordinate of P̄ is higher than the x2
coordinate of Ā, then P is above π.

8. An algorithm for intersecting a line with a plane: it is convenient to describe

by its projecting planes. Refer to the x1 x2 and x2 x3 projective planes by π1
and π2 respectively, and call their intersections with π PA and PB respectively.
 12.4 Solutions to Exercises in Chapter 5 525

We see that

P = PA ∩ = PB ∩ = PA ∩ PB ∩ .

It is easy to see that

¯12 = PA12 , ¯23 = PB23

define the planar coordinates of π1 and π2 as z1 , z2 and w1 , w2 respectively.

We can see that w1 is coincident with the x1 axis, z2 is coincident with the
526 12. Solutions to Selected Exercises

1 = ¯ , and w̄ 2 = ¯ . We next find the point A = y 1 ∩ z1 and

x3 axis, z̄12 12 23 23
B = y ∩ w 2 ; the polygonal line P̄ intersects Ā on Ȳ 2 (because AP  y 2 ) and
2
intersects B̄ on Ȳ 1 (because PB  y 1 ).
9. (a) i. A pencil of planes containing

12 : x2 = m2 x1 + b2 ,
:
23 : x3 = m3 x2 + b3

is represented by

π : (x3 − m3 x2 − b3 ) + k(x2 − m2 x1 − b2 ) = 0.
 12.4 Solutions to Exercises in Chapter 5 527

Proof.
 Every  containing contains the points (0, b2 , m3 b2 + b3 )
plane
b2
and − m2 , 0, b3 and another general point (e, f, g). We find that π is
represented by
 
f m3 + b 3 − g m3 b2 − g + b3 + em2 m3
m2 x1 − x2 +x3
b2 − f + em2 b2 − f + em2

m3 b2 − g + b3 + em2 m3
+ b2 − m3 b2 − b3 = 0.
b2 − f + em2
Let
g − f m3 − b3
K= .
b2 − f + em2
Then

−m2 Kx1 + (K − m3 )x2 + x3 − b3 − Kb2 = 0,

which is the required relation.

ii. Every plane represented by π contains .

Proof. Clearly, when x2 = m2 x1 +b2 , then x3 = m3 x2 +b3 and conversely.

So ⊂ π.
528 12. Solutions to Selected Exercises

(b) Given π : (x3 − m3 x2 − b3 ) + k(x2 − m2 x1 − b2 ) = 0, therefore every

line ⊂ π is represented by two linear equations:

12 : x2 = m2 x1 + b2 ,
:
23 : x3 = m3 x2 + b3 .

Let y 1 = π ∩ {x3 = 0}. Thus


1 : x = km2 x + kb2 +b3 ,
y12
1 2 k−m3 1 k−m3
y :
y03 : x3 = 0
1

 1
1
ȳ12 = (k − m3 , kb2 + b3 , k − m3 − km2 ),
⇒ Ȳ :
1 = (2, 0, 1).
ȳ03

Let y 2 = π ∩ {x1 = 0}. Thus

 2
y23 : x3 = (m3 − k)x2 + b3 + kb2 ,
y2 :
2 :x =0
y01 1
 2
ȳ23 = (2 + k − m3 , b3 + kb2 , 1 + k − m3 ),
⇒ Ȳ 2 :
2 = (0, 0, 1).
ȳ01
Let

N1 = P̄1 ∩ Ȳ 1 = ȳ12
1
= (k − m3 , kb2 + b3 , k − m3 − km2 ),

N2 = P̄1 ∩ Ȳ 2 = (m3 (2 + k − m3 ), b3 m3 − b3 k, m3 (1 + k − m3 )).

Construct the line through N1 and N2 :

     

 (N1 )2 (N1 )3   (N1 )3 (N1 )1   (N1 )1 (N1 )2 

line1 =   , ,
(N2 )2 (N2 )3   (N2 )3 (N2 )1   (N2 )1 (N2 )2 

= [kb2 m3 + m3 k 2 b2 − kb2 m23 + b3 m3 − m3 b3 k + b3 k 2

+ km2 b3 m3 − k 2 m2 b3 ,

m3 (−m3 + km2 m3 + k − 2km2 − k 2 m2 ),

m3 b3 k − b3 k 2 − 2kb2 m3 − m3 k 2 b2 + kb2 m23 − 2b3 m3 ].

 12.4 Solutions to Exercises in Chapter 5 529

Let

M1 = P̄2 ∩ Ȳ 1 = (k − m3 , b2 k − b2 m3 , k − m3 − km2 ),

M2 = P̄2 ∩ Ȳ 2 = ȳ23
2
= (2 + k − m3 , b3 + kb2 , 1 + k − m3 ).

Construct the line through M1 and M2 :

     

(M1 )2 (M1 )3  (M1 )3 (M1 )1  (M1 )1 (M1 )2 

line2 =  , ,
(M2 )2 (M2 )3  (M2 )3 (M2 )1  (M2 )1 (M2 )2 

= [kb2 − kb2 m3 − b2 m3 + b2 m23 − b3 k + b3 m3

+ k 2 m2 b2 + km2 b3 ,

− m3 + km2 m3 + k − 2km2 − k 2 m2 ,

− (m3 − k)(b2 m3 + b3 − 2b2 )],

η̄ = line1∩line2 (in homogeneous coordinates),

η̄ = [−2m3 + m23 − k 2 , −b3 m3 − k 2 b2 , (m2 − 1)k 2 + m23 − m3 ]


This page intentionally left blank
12

Notation and List of Symbols

Alphabetic

ℵ0 , ℵ1 , ℵ2 , . . . cardinality of infinite sets

, r, s, . . . lines are usually denoted by lowercase italics Latin letters
A, B, C, . . . points are usually denoted by uppercase Latin letters
π, ρ, σ, . . . planes and surfaces are usually denoted by lowercase Greek letters
N positive integers; natural numbers
R real numbers
C complex numbers
5 real part of a complex number
6 imaginary part of a complex number
R2 Euclidean 2-dimensional plane (or space)
P2 projective 2-dimensional plane (or space)
RN Euclidean N-dimensional space
PN projective N-dimensional space

Symbols

General

∈ belongs to
 such that
∃ there exists
∀ for all
⊂ subset of
∪ set union
∩ set intersection

A. Inselberg, Parallel Coordinates, DOI 10.1007/978-0-387-68628-8,

© Springer Science+Business Media, LLC 2009
532 Notation and List of Symbols

∅ empty st (null set)

2A power set of A, the set of all subsets of A
ip inflection point of a curve
n principal normal vector to a space curve
n̂ unit normal vector to a surface
t vector tangent to a space curve
b binormal vector
τ torsion (measures “non-planarity” of a space curve)
× cross product
→ mapping symbol
A→B correspondence (also used for duality)
⇒ implies
⇔ if and only if
<, > less than, greater than
|a| absolute value of a
|b| length of vector b
 norm, length
L1 L1 norm
L2 L2 norm, Euclidean distance

summation
Jp set of integers modulo p with the operations + and ×
≈ approximately equal
= not equal
 parallel
⊥ perpendicular, orthogonal
( , , ) homogeneous coordinates — ordered triple representing a
point in P2 ; point coordinates
[ , , ] homogeneous coordinates — ordered triple representing a
line P2 ; line coordinates
P
p(A, B, C, D, . . . )  perspectivity between the lines p and p
with respect to a
p
(A
, B
, C
, D
, . . . ) point P
 Notation and List of Symbols 533

p  p
projectivity between the lines p and p

df
dx derivative of f with respect to x, also written as f
(x)
∂ boundary of a region 
∂F partial derivative of F with respect to xi , also denoted Fxi
∂xi
and when context is clear, Fi
∇F gradient of F

Specific to Parallel Coordinates (In the order occuring in the text)

-coords abbreviation for parallel coordinate system, parallel
coordinates
Ā image of A in -coords, representation of A
X̄1 , . . . , X̄N axes for parallel coordinate system for N-dimensional space (either
Euclidean or Projective)
¯i,j point representing the linear relation between xi and xj ; (N − 1)
points with two indices appear in the representation of a line in RN
(or PN )
πp p-flat in RN , a plane of dimension 1 ≤ p ≤ (N − 1) in RN
π1 a one-flat is a line, also denoted by
sp abbreviation for superplane, a 2-flat whose points are represented
by straight lines in -coords
πs denotes a superplane
π Ns denotes a superplane in RN
E family of superplanes, named after J. Eickemeyer, who first
discovered them
E also denotes the class of surfaces who are the envelope of their
tangent (hyper) planes
EG class of surfaces contained in E whose representing contours can
be obtained by Lemma 9.3.1
di directed distance of X̄i from the y-axis
d0N (see next), standard axis spacing, when the X̄1 is coincident with
the y-axis and inter-axis distance is one unit for all axes
diN axes spacing where the X̄1 axis is placed one unit after the X̄N and
all axes up to and including X̄i are placed successively one unit
apart and the rest of axes remain in the standard axes spacing.
534 Notation and List of Symbols

π1s first superplane, with the standard axes spacing

π1s
second superplane, the X̄1 placed one unit to the right of X̄N
the remaining axes are in standard axes spacing.
πis
is the ith superplane with axes spacing diN
π̄123 , π̄1
123 two triple indexed points representing the plane (2-flat) π ⊂ R3 ,
also denoted by π̄123 = π̄0
, π̄1
23 = π1
or 123, 1
23 when π is
clear from the context; π̄1
23 is based on the axes spacing d13
π̄123,... ,N = π̄0
point with (N + 1) indices representing a hyperplane (of dimen-
sion N − 1)), and similarly π̄i
for the representing points based
on axes spacing diN
gconics generalized conics, sections of a double cone in R3 whose 2
bases are the same bounded convex set in R2
bc, ubc, gh like conics there are 3 kinds of gconics: bounded convex set (bc),
unbounded convex set (ubc) and generalized hyperbola (gh)
) outer union
* inner intersection
12

Bibliography

[1] A.E. Abbot. A Romance of Many Dimensions. Dover, New York (originally
published in 1884 as ‘Flatland by A Square’, Seeley & Co., London), 1992.
[2] N. Adrienko and G. Adrienko. Constructing Parallel Coordinates Plots for
Problem Solving. in Proc. 1st Inter. Symp. on Smart Graphics, A. Butz,
A. Krueger, P. Oliver and M. Zhou (eds.), ACM Press 9–14, New York,
2001.
[3] C. Ahlberg and B. Shneiderman. Visual Information Seeking: Tight Cou-
pling of Dynamic Query Filters with Starfiel Displays. Proc. CH94 Conf.,
313–317, ACM Press, New York, 1994.
[4] D. Asimov. The grand tour: A tool for viewing multidimensional data.
SIAM J. Sci. & Stat. Comp., 6:128–143, 1985.
[5] T. Avidan and S. Avidan. Parallax – a data mining tool based on parallel
coordinates. Comput. Statit., 13–1:65, 1998.
[6] S. Axelson. Understanding Intrusion Detection Through Visualization.
Ph.D. Thesis, CS Dept. Chalmers Univ., Gotesborg, Sweden, 2005.
[7] F. Ayres Jr. Projective Geometry. Shaum’s Outlines on Math., McGraw-
Hill, New York, 1967.
[8] T.F. Banchoff. Beyond the Third Dimension. Scientific American Library,
New York, 1990.
[9] E.W. Bassett. Ibm’s ibm fix. Industrial Computing, 14(41):23–25, 1995.
[10] A.F. Beardon. The Geometry of Discrete Groups. Springer-Verlag,
New York, 1983.
[11] B. Bederson and B. Shneiderman. (eds.) The Craft of Information Visu-
alization: Readings and Reflection . Morgan-Kaufman, San Francisco,
2003.
[12] J. Bergen, J. Borwein and P. Borwein. Pi: A Source Book. Springer-Verlag,
New York, 1997.
[13] M. Berger. Geometry I & II. Corrected second printing, Springer-Verlag,
New York, 1994.
[14] J. Bertin. Sémiologie Graphique. First pub. 1967, Ecole des Hautes Etudes
en Science, Paris, 1999.
536 Bibliography

[15] D.M. Bloom. Linear Algebra and Geometry. Cambridge University Press,
Cambridge, 1979.
[16] B. Bollobas. Graph Theory. Springer-Verlag, New York, 1979.
[17] V.G. Boltyanskii. Envelopes, R.B. Brown, translator (original in Russian).
Pergamon Press, New York, 1964.
[18] M.F. Brannan, D.A. Esplen and J.J. Gray. Geometry. Cambridge University
Press, New York, 1999.
[19] O.S. Brodetsky. A First Course in Nomography (fi st published in 1920).
G. Bell and Sons, London, 1949.
[20] S. Bulthoff and H.H. Edelman. Psychophysical support for a two-
dimensional view interpolation theory of object recognition. Proc. Nat.
Acad. Sci., 89-Psychology: 60–4, 1992.
[21] J. Canny and J. Reif. New Lower Bound Techniques for Robot Motion
Planning Problems, in Proc. of 28th Symp. on Found. of Comp. Sci. IEEE
Comp. Soc., Washington, D.C., 1987.
[22] S. Card, J. MacKinlay and B. Shneidermanm. (eds.) Readings in
Information Visualization. Morgan Kaufman, San Francisco, 1999.
[23] A. Chatterjee. Visualizing Multidimensional Polytopes and Topologies for
Tolerances. Ph.D. thesis, Dept. Comp. Sci., Univ. of S. Calif., 1995.
[24] A. Chatterjee, P.P. Das and S. Bhattacharya. Visualization in linear pro-
gramming using parallel coordinates. Pattern Recognition, 26–11:1725–36,
1993.
[25] C. Chen. Information Visualization: Beyond the Horizon, 2nd ed. Springer,
New York, 2004.
[26] C.H. Chen. Generalized association plots for information visualization: The
application of the convergence of iteratively formed correlation matrices.
Statistica Sica, 12:1–23, 2002.
[27] C.H. Chen and J.A. Chen. Interactive diagnosis plots for multidimensional
scaling with applications in psychosis disorder data analysis. Statistica
Sica, 10:665–691, 2000.
[28] E.H. Chi. A Framework for Visualizing Information. Kluwer, Amsterdam,
2002.
[29] H. Choi and H. Lee. PCAV: Internet Attack Visualization in Parallel
Coordinates, LNCS 3783, 454–466. Springer-Verlag, New York, 2005.
[30] T. Chomut. Exploratory Data Analysis in Parallel Coordinates. M.Sc.
thesis, Dept. Comp. Sci., UCLA, 1987.
[31] S. M. Cohan and D.C.H. Yang. Mobility analysis in parallel coordinates.
J. Mech. & Mach., 21:63–71, 1986.
 Bibliography 537

[32] G. Conti. Security Data Visualization. No Starch Press, San Francisco,

2007.
[33] D. Conversano, D. Di Benedetto and R. Siciliano. The Clockworktree:
Visual Multivariating Splitting. Cladag 2003, Book of Short Papers 113–7,
Clueb Publisher, Bologna, 2003.
[34] D. Cook and A. Buja. Manual controls for high-dimensional data
projections. J. of Comp. & Graph. Stat., 6–4, 464–480, 1997.
[35] R. Courant. Differential and Integral Calculus vol. 2. Blackie & Son,
London 171–183, 1946.
[36] R. Courant and H.E. Robins. What is Mathematics? Oxford University
Press, New York, 1953.
[37] D. Cox, J. Little and D. O’Shea. Ideals, Varieties, and Algorithms. 2nd
edition, Springer-Verlag, New York, 1997.
[38] H.S.M. Coxeter. Introduction to Geometry. John Wiley, New York, 1969.
[39] H.S.M. Coxeter. Projective Geometry. Second Ed., John Wiley, New York,
1974.
[40] D.H. Deterding. Speaker Normalization for Automatic Speech Recognition.
Ph.D. thesis, Cambridge University, 1989.
[41] B. Dimsdale. Conic Transformations – IBM LASC Tech. Rep. G320-2713.
IBM LA Scientific Center, 1981.
[42] B. Dimsdale. Conic Transformations and Projectivities – IBM LASC Tech.
Rep. G320-2753. IBM LA Scientific Center, 1984.
[43] J.A. Dykes, A.M. MacEachren and M.J. Kraak. (eds). Exploring Geo-
visualization. Elsevier, Amsterdam, 2005.
[44] J.A. Dykes and D.M. Mountain. Seeking structure in records of spatio-
temporal behaviour: visualization issues, efforts and applications. Comput.
Statit. & Data Anal., 43 (Data Visualization II special ed.):581–603, 2003.
[45] R. Edsall. The parallel coordinate plot in action: Design and use for
geographic visualization. Comput. Statit. & Data Anal., 43–4:605–619,
2003.
[46] P. Ehrenfest. In What Way Does it Become Manifest in the Fundamental
Laws of Physics that Space has Three Dimensions? – Appeared in 1917 –
in Collected Scientifi Papers of P. Ehrenfest, Klein, M.J. ed., 400–409.
North-Holland Publishing Company, Interscience Publishers, New York,
1959.
[47] J. Eickemeyer. Visualizing p-flat in N-space using Parallel Coordinates.
Ph.D. thesis, Dept. Comp. Sci., UCLA, 1992.
538 Bibliography

[48] P.L. Eisenhart. A Treatise on the Differential Geometry of Curves and

Surfaces. Dover Publications, New York, 2004.
[49] S. El Mejdani, R. Egli and F. Dubeau. Old & new straight-line detectors,
descriptions & comparisons. Pattern Recognition, 41:1845–66, 2008.
[50] A. Ellis and G. Dix. Enabling automatic clutter reduction in parallel
coordinates. Trans. on Vis. & Comp. Grap., 12–5:717–724, 2006.
[51] N. Elmquist, J. Stasko and P. Tsingas. Datameadow: A visual canvas for
analysis of large-scale multivariate data. Proc. IEEE Vast Conf., to appear,
2007.
[52] L. Ernstrom. A plücker formula for the singular projective varieties.
Commut. Algebra., 5–9:2897–2901, 1997.
[53] K. Esbensen, et al. Multivariate Analysis, 1994, Camo AS, Trondheim,
Norway.
[54] S. Even. Graph Algorithms. Computer Science Press, Rockville MD, 1979.
[55] M.G. Ewing. On the definition of inflection point. Amer. Math. Monthly, 45:
681–683, 1938.
[56] F. Fallside and A.J. Robinson. A Dynamic Connectionist Model of Phoneme
Recognition. Proc. of 1st European Neural Network Conf. (nEURO), 1988.
[57] U.M. Fayad, G. Piatesky-Shapiro, P. Smyth and R. Uthurusamy. Advances
in Knowledge Discovery and Data Mining. AAAI/MIT Press, Cambridge
Mass., 1996.
[58] U.M. Fayad, G. Grinstein and A. Wierse. Information Visualization in Data
Mining and Knowledge Discovery. Morgan Kaufmann, San Francisco,
2001.
[59] R. Finsterwalder. A Parallel Coordinate Editor as a Visual Decision Aid
in Multi-Objective Concurrent Control Engineering Environment 119–122.
IFAC CAD Contr. Sys., Swansea, UK, 1991.
[60] P. Fiorini and A. Inselberg. Configu ation Space Representation in Parallel
Coordinates. IEEE Conf. Rob. & Aut., 1215–1220, 1989.
[61] G. Fischer. Plane Algebraic Curves. Amer. Math. Soc. Press, Providence,
2001.
[62] M. Friendly. Extending mosaic displays: Marginal, conditional, and partial
views of categorical data. J. of Comp. & Graph. Stat., 8–3:373–95, 1999.
[63] M. Friendly and al. Milestones in Thematic Cartography.
www.math.yorku.ca/scs/SCS/Gallery/milestones/, 2005.
[64] LORAL FSD. Test Report for the Quick-Look Flight Demonstration of the
IntraFormation Positioning System, Document Nos. 93-A37-002, Contract
f33615-90-C-3609. LORAL Federal Systems Co, Oswego, 1994.
 Bibliography 539

[65] M. O. Fua, Y.H. Ward and E.A. Rundensteiner. Hierarchical Parallel

Coordinates for Exploration of Large Datasets – Proc.of Conf. on Vis.,
43–50. IEEE Comp. Soc. Press, Los Alamitos, CA, 1999.
[66] G.W. Furnas and A. Buja. Prosection views: Dimensional inference through
sections and projections. J. Comp. & Graph. Stat., 3:323–85, 1994.
[67] C. Gennings, K.S. Dawson, W.H. Carter and R.H. Myers. Interpreting
plots of a multidimensional dose-response surface in parallel coordinates.
Biometrics, 46:719–35, 1990.
[68] W.J. Gibbs. Graphical methods in the thermodynamics of fluids.
Conn. Acad., II-1:309–342, 1873.
[69] W.J. Gibbs. A method of geometrical representation of the thermodynamic
properties of substances by means of surfaces. Conn. Acad., II-2:382–404,
1873.
[70] L. Girardin and D. Brodbeck. Visual Approach for Monitoring Logs. 12th
Systems Adm. Conf., Boston, 1998.
[71] A. Goel. Vizcraft: A Multidimensional Visualization Tool for Aircraft
Configu ation Design. IEEE Vis Conf., 425–8, 1999.
[72] H. Goldstein. Classical Mechanics, 2nd ed. Addition-Wesley, Reading
Mass., 1980.
[73] E. Goursat. A Course in Mathematical Analysis vol. 1. Ginn & Co,
New York, 19xx.
[74] M.J. Greenberg. Euclidean and Non-Euclidean Geometries, Development
and History. W.H. Freeman and Co., 1974.
[75] W.H. Guggenheimer. Differential Geometry. McGraw-Hill, New York,
1963.
[76] H.S. Guyford and J.J. Haggerty. Flight – Life Science Library. Time Inc.,
New York, 1965.
[77] F. Harary. Graph Theory. Addison-Wesley, Reading, Mass., 1969.
[78] J. Harris. Algebraic Geometry. Springer-Verlag, New York, 1992.
[79] H. Hauser. Parallel Sets: Visual Analysis of Categorical Data. Proc. IEEE
Infovis, 2005.
[80] M.A. Hearst. Clustering versus faceted categories for information explo-
ration. Comm. ACM, 49–4:59–61, 2006.
[81] N. Helfman. Navigating in Visual Models of Decision Support Systems.
M.Sc. thesis, Dept. Comp. Sci., Bar Ilan University, Israel, 2000.
[82] J. Helly. Applications of Parallel Coordinates to Complex System Design
and Operation. Proc. Nat. Comp. Grap. Assoc. vol. III, 541–546, 1987.
540 Bibliography

[83] P. Hertzog. Visualizations to Improve Reactivity towards Security Inci-

dents inside Corporate Networks, SIGSAC Conf. Proc., 95–102. ACM,
New York, 2006.
[84] B. Hetzler, P. Whitney, L. Martucci and J. Thomas. Multi-faceted Insight
Through Interoperable Visual Information Analysis Paradigms in Proc.
IEEE on Info Vis pp. 137–144. IEEE Comp. Soc., Los Alamitos, CA,
1998.
[85] D. Hilbert and S. Cohn-Vossen. Geometry and the Imagination. Chelsea,
New York, 1983.
[86] W. Hodge and D. Pedoe. Methods of Algebraic Geometry, Vol. II.
Cambridge Univ. Press, Cambridge, 1952.
[87] P. Hoffman and G. Grinstein. Multidimensional Information Vis. for Data
Mining with Applications for Machine Learning Classifie s. In Info. Vis.
for KDD, Morgan-Kaufman, 2000.
[88] P. Hoffman, G. Grinstein, K. Marx and I. Grosse. DNA Visual and Analytic
Data Mining. Proc. IEEE Infovis, pp. 437–441, 1997.
[89] Q. Huanin, C. Wing-Yi, X. Anbang, C. Kai-Lun, L. Kai-Hon and G. Ping.
Visual analysis of the air pollution problem in hong kong. Trans. on Vis. &
Comp. Grap., 13–6:1408–15, 2007.
[90] H.P. Hudson. Ruler & Compass in ‘Squaring the Circle’ by Hopson et al.
Chelsea, New York, 1953.
[91] M.Y. Huh and K.Y. Song. DAVIS: A java based data visualization system.
Comp. Stat., 17–3:411–423, 2002.
[92] C.K. Hung and D. Ierardi. Constructing Convex Hulls of Piecewise Smooth
Objects. Univ. of S. Calif. CS Dept. Tech R. 94–583, 1994.
[93] C.K. Hung and A. Inselberg. Parallel Coordinate Representation of Smooth
Hypersurfaces. USC Tech. Report #CS-92-531, Los Angeles, 1992.
[94] C.K. Hung and A. Inselberg. Description of Surfaces in Parallel Coordi-
nates by Linked Planar Regions, Mathematics of Surfaces XII, 177–208,
LNCS 4647. Springer-Verlag, New York, 2007.
[95] C.B. Hurley and R.W. Olford. Pairwise Display of High Dimensional
Informatin via Eulerial Tours and Hamiltonian Decompositions. to appear,
2008.
[96] IEEE. The faa’s advanced automation program — special issue. IEEE
Computer J., 20–2, 1987.
[97] A. Inselberg. N-Dimensional Graphics, LASC Tech. Rep. G320-2711,
140 pages. IBM, 1981.
 Bibliography 541

[98] A. Inselberg. Intelligent Instrumentation and Process Control, Proc. 2nd

IEEE Conf. on AI Appl., 302–307. IEEE Comp. Soc., Los Alamitos, CA,
1985.
[99] A. Inselberg. The plane with parallel coordinates. Visual Computer,
1:69–97, 1985.
[100] A. Inselberg. Parallel Coordinates: A Guide for the Perplexed, in Hot
Topics Proc. of IEEE Conf. on Visualization, 35–38. IEEE Comp. Soc.,
Los Alamitos, CA, 1996.
[101] A. Inselberg. Multidimensional Detective, in Proc. of IEEE Information
Visualization ’97, 100–107. IEEE Comp. Soc., Los Alamitos, CA, 1997.
[102] A. Inselberg. Visual data mining with parallel coordinates. Comput. Statit.,
13–1:47–64, 1998.
[103] A. Inselberg and T. Avidan. The Automated Multidimensional Detective, in
Proc. of IEEE Information Visualization ’99, 112–119. IEEE Comp. Soc.,
Los Alamitos, CA, 1999.
[104] A. Inselberg, M. Boz and B. Dimsdale. Planar Conflic Resolution Algo-
rithm for Air-Traffi Control and the One-Shot Problem, in IBM PASC Tech.
Rep. G320-3559. IBM Palo Alto Scientific Center, 1991.
[105] A. Inselberg and B. Dimsdale. Parallel Coordinates: A Tool For Visualizing
Multidimensional Geometry, in Proc. of IEEE Conf. on Vis. ’90, 361–378.
IEEE Comp. Soc., Los Alamitos, CA, 1990.
[106] A. Inselberg and B. Dimsdale. Multidimensional lines i: Representation.
SIAM J. of Applied Math., 54–2:559–577, 1994.
[107] A. Inselberg and B. Dimsdale. Multidimensional lines ii: Proximity and
applications. SIAM J. of Applied Math., 54–2:578–596, 1994.
[108] A. Inselberg and C.K. Hung. A New Representation of Hypersurfaces Based
on Their Tangent Hyperplanes. In Course Notes for Tutorial #16 on Multi-
dimensional Visualization 105–112, SIGGRAPH94, ACM Publications,
1994.
[109] A. Inselberg and P.L. Lai. Visualizing Families of Close Planes, 66. Proc.
5th Asian Conf. on Stat., Hong Kong, 2005.
[110] A. Inselberg, M. Reif and T. Chomut. Convexity algorithms in parallel
coordinates. J. ACM, 34:765–801, 1987.
[111] T. Itoh, et al. Hierarchical visualization of network intrusion detection data.
IEEE Comp. Grap. & Appl., 26–2:40–47, 2006.
[112] M.Y. Ivory and M.A. Hearst. Statistical profiles of highly-rated web sites.
IEEE Inter. Comp., 6–2:55–63, 2002.
542 Bibliography

[113] J. Johansson, P. Ljung, M. Jern and M. Cooper. Revealing Structure within

Clustered Parallel Coordinates Displays, Proc. IEEE Infovis, 125–132.
IEEE Comp. Soc., Los Alamitos, CA, 2005.
[114] J. Johansson, P. Ljung, M. Jern and M. Cooper. Revealing structure in
visualizations of dense 2d and 3d parallel coordinates. Inf. Vis. Pal-Grave-
Mcmillan Journ., 5:125–136, 2006.
[115] C. Jones. Visualization and Optimization. Kluwer Academic Publishers,
Boston, 1996.
[116] M.W. Kantor, A.R. Liebler, E.S. Payne and E.E. Shult. (eds.) Finite Geome-
tries, Buildings, and Related Topics. Clarendon Press Oxford, Oxford,
1990.
[117] F. Karteszi. Introduction to Finite Geometries. North-Holland, Amsterdam-
Oxford, 1976.
[118] D.A. Keim. Designing pixel-oriented visualization techniques: Theory and
applications. Comp. Graphics & Applic., 6–1:59–78, 2000.
[119] D.A. Keim and H.P. Kriegel. Visualization techniques for mining large
databases: A comparison. Trans. Knowl. and Data Engr., 8–6:923–938,
1996.
[120] M. Kipouros, T. Mleczko and M. Savill. Use of Parallel Coordinates
for Post-Analyses of Multiobjective Aerodynamic Design Optimization in
Turbomachinery (to appear). Proc. 4th AIAA Meet., 2008.
[121] M. Kline. Mathematical Thought from Ancient to Modern Times. Oxford
Univ. Press, New York, 1972.
[122] R. Kriz. Visual Thinking by Scientists.
www.sv.vt.edu/classes/ESM4714/Gen-Prin/vizthink.html, 2007.
[123] N. Kumasaka and R. Shibata. High Dimensional Data Visualization the
Textile Plot. Proc. Compstat 2006 Psysica-Verlag, 581–9, 2006.
[124] P.L. Lai and C. Fyfe. A family of canonical correlation analysis networks.
Neural Proc. Let., 14–2:93–105, 2001.
[125] F. Lauro, N.C. Palumbo and A. Iodice D’Enza. New Graphical Symbolic
Objects Representation in Parallel Coordinates, in Between Data Science
and Applications (M. Schader et al. eds). Springer, Heidelberg, 2003.
[126] C.L. Lay. Linear Algebra and its Applications. Updated Edition, Addison-
Wesley, New York, 2000.
[127] R.O. Lejeune. Government Provided Complex Scenarios for the Advanced
Automated System Design Competition Phase, MTR-85W240. MITRE Co,
McLean, Virginia, 1985.
 Bibliography 543

[128] M.M. Lipschutz. Differential Geometry. Shaum’s Outlines on Math.,

McGraw-Hill, New York, 1969.
[129] J. Martin and G. Kennedy. Using Curves to Enhance Parallel Coordi-
nates Visualization. Proc. IEEE Infovis, London, IEEE Comp. Soc., Los
Alamitos, CA, 10–16, 2003.
[130] Mathcurve.com. Great Website for Surfaces.
www.mathcurve.com/surfaces/surfaces.shtml, 2006.
[131] T. Matskewich, A. Inselberg and M. Bercovier. Approximated Planes in
Parallel Coordinates. Proc. of Geom. Model. Conf., St. Malo, Vanderbilt
Univ. Press, 257–266, 2000.
[132] B.H. McCormick, T.A. Defanti and M.D. Brown. Visualization in Scientifi
Computing. Computer Graphics 21–6, ACM SIGGRAPH, New York, 1987.
[133] N. Megido. On the Complexity of some Geometric Problems in Unbounded
Dimension, in IBM Res. Rep. RJ5744(58033). IBM Research, 1987.
[134] D. Michie, D.J. Spiegelhalter and C.C. Taylor. Machine Learning, Neural
and Statistical Classificatio . Ellis Horwood series in AI, 1994.
[135] T. Mihalisin, J. Timlin and J. Schwegler. Visualizing multivariate functions,
data and distributions. IEEE Comp. Graph. and Appl., 1991.
[136] T.M. Mitchell. Machine Learning. McGraw-Hill, 1997.
[137] J.W. Moon. Various Proofs of Cayley’s Formula for Counting Trees,
Seminar in Graph Theory, F. Harary (ed.). Holt, Rinehart & Winston,
1967.
[138] J.R. Munkries. Topology, 2nd. ed. Prentice-Hall, New Jersey, 2000.
[139] J. Nakano and K. Honda. Three-dimensional parallel coordinates plot
used for variable selection. Proc. COMPSTAT 2006, (eds. A. Rizzi and
M. Vichi), Springer, 187–195, 2006.
[140] J.R. Newman. World of Mathematics. Simon and Schuster, New York,
1956.
[141] M. d’Ocagne. Coordonnees parallels et axiale. Gautier-Villars, Paris,
1885.
[142] M. d’Ocagne. Traite de Nomographie. Gautier-Villars, Paris, 1899.
[143] W.F. Osgood. Advanced Calculus. Macmillan, New York, 186–194, 1925.
[144] E. Otto. Nomography. Translated from 1956 Polish ed. by J. Smolska.
Macmillan, New York, 1963.
[145] D. Pedoe. Circles. Pergamon, New York, 1957.
[146] P.A. Penna and R.R. Patterson. Projective Geometry and its Application to
Computer Graphics. Prentice-Hall, New Jersey, 1986.
544 Bibliography

[147] Y. Pham, B. Cai and R. Brown. Visualization Techniques for Tongue

Analysis in Traditional Chinese Medicine. Proc. Med. Imag. SPIE 14–19,
2004.
[148] F.P. Preparata and M.I. Shamos. Computational Geometry. Springer-
Verlag, New York, 1985.
[149] J.R. Quinlan. C4.5: Programs for Machine Learning. Morgan Kaufman,
1993.
[150] J. Riordan. An Introduction to Combinatorial Analysis. John Wiley,
New York, 1958.
[151] H. Rosenbaum and R. Schumann. Chances and Limits of Progression in
Visualization, in SimVis. Proc. Sim. & Vis., Magdeburg, Germany, 2007.
[152] B.A. Rosenfeld. A History of Non-Euclidean Geometry. Springer-Verlag,
New York, 1988.
[153] F. Rossi. Visual Data Mining and Machine Learning, in 14th Europ. ESANN
Proc., 251–264. ESANN, Bruges Belgium, 2006.
[154] P. Samuel. Projective Geometry. Springer-Verlag, New York, 1988.
[155] M. Schall. Diamond and ice: Visual exploratory data analysis tools.
Perspective, J. of OAC at UCLA, 18(2):15–24, 1994.
[156] C. Schmid and H. Hinterberger. Comparative Multivariate Vis. Across
Conceptually Different Graphic Displays, in Proc. of 7th SSDBM. IEEE
Comp. Soc., Los Alamitos, CA, 1994.
[157] O. Schreier and Sperner. Projective Geometry of n Dimensions. Chelsea
Publishing Company, New York, 1985.
[158] N. Shang and L. Breiman. Distribution based trees are more accurate.
Proc. of 1st Inter. Conf. on Neural Info. Proc. (ICONIP96), 133, 1996.
[159] B. Shneiderman. Leonardo’s Laptop. MIT Press, Boston, 2002.
[160] B. Shneiderman and C. Plaisant. Designing the User Interface (4th Edition).
Addison Wesley, Reading MA, 2005.
[161] C.H. Sisam and W.F. Atchison. Analytic Geometry, third ed. Holt, Rinehart
and Winston, New York, 1955.
[162] D.J. Sobiski. Collision Detection and Collision Avoidance, Document Nos.
196A798. LORAL Federal Systems Co, Oswego, NY, 1995.
[163] D.M.Y. Sommerville. An Introduction to the Geometry of N Dimensions.
(First publ. 1929), Dover Publications, New York, 1958.
[164] R. Spence. Information Visualization: Design for Interaction, 2nd ed.
Prentice Hall, 2007.
[165] I. Stewart. Flatterland. Perseus, New York, 2001.
 Bibliography 545

[166] D.J. Struik. Lectures on Classical Differential Geometry 2nd. Ed. Dover
(Addison-Wesley 1961), New York, 1988.
[167] D.F. Swayne, D. Cook and A. Buja. XGobi: Interactive Dynamic Graphics
in the X Window System. JCGS, 7–1, 113–130, 1998.
[168] D.F. Swayne, (T. Lang) Duncan, A. Buja and D. Cook. Ggobi: evolving
from Xgobi into an extensible framework for interactive data visualization.
J. of Comp. & Graph. Stat., 43:423–444, 2003.
[169] J.L. Synge and A. Schild. Tensor Calculus, 2nd ed. University of Toronto
Press, Toronto, 1956.
[170] V. Tontodonato. Symbolic Objects Representation, Ph.D. thesis. Dept. of
Math and Stat., Univ. Federico II, Napoli, 2002.
[171] M. Tory, S. Potts and T. Moller. A parallel coordinates style interface for
exploratory volume visualization. Trans. on Vis. & Comp. Grap., 11–1:
71–80, 2005.
[172] E.R. Tufte. The Visual Display of Quantitative Information. Graphic Press,
Connecticut, 1983.
[173] E.R. Tufte. Envisioning Information. Graphic Press, Connecticut, 1990.
[174] E.R. Tufte. Visual Explanation. Graphic Press, Connecticut, 1996.
[175] L.A. Tweedie, R. Spence, H. Dawkes and Su H. Externalizing Abstract
Mathematical Models, Proc. CHI, Vancouver, Can., 406–412. ACM Press,
1996.
[176] A.R. Unwin. Requirements for interactive graphics software for exploratory
data analysis. Comp. Stat., 14:7–22, 1999.
[177] A.R. Unwin, M. Theus and H. Hofmann. (Eds.) Graphics of Large Datasets.
Springer, New York, 2006.
[178] F.B. Viegas, M. Wattenberg and D. Kushal. Studying Cooperation and Con-
flic between Authors with history flo Visualizations in Proc. of SIGHCI on
Human Factors in Computing Systems, 575–582. ACM Press, New York,
2004.
[179] R.J. Walker. Algebraic Curves. Springer, New York, 1978.
[180] W.A. Walker. What is a point of inflection? Amer. Math. Monthly, 63:
182–183, 1956.
[181] M.O. Ward. XmdvTool: Integrating multiple methods for visualizing multi-
variate data, Proc. IEEE Conf. on Vis., CA, 326–333. IEEE Comp. Soc.,
Los Alamitos, CA, 1994.
[182] C. Ware. Information Visualization: Perception and Design, 2nd ed.
Morgan Kaufman, San Francisco, 2004.
546 Bibliography

[183] C.E. Weatherburn. Differential Geometry of Three Dimensions. Cambridge

Univ. Press, Cambridge, 1955.
[184] E. Wegman. Hyperdimensional data analysis using parallel coordinates.
J. Amer. Stat. Assoc., 85:664–675, 1990.
[185] P.C. Wong and R.D. Bergeron. 30 Years of Multidimensional Multivari-
ate Visualization in Scientifi Visualization: Overviews, Methodologies &
Techniques, G.M. Nelson, H. Mueller and H. Hagen (eds.), 3–33. IEEE
Comp. Soc., Los Alamitos, CA, 1997.
[186] V.O. Ya. Some integral calculus based on euler characteristic, lect. notes.
Amer. Math. Monthly, 1346:127–138, 1988.
[187] L. Yang. Pruning & visualizing generalized association rules in parallel
coordinates. IEEE Know. & Data Engr., 15(1):60–70, 2005.
[188] H. Ye and Z. Lin. Speed-up simulated annealing by parallel coordinates.
Euro J. of OR, 173:59–71, 2006.
[189] X. Yin, et al. VisFlowConnect. Comp. Sec., ACM Workshop on Visua-
lization, Washington, D.C. 26–34, 2004.
12

Index

Abbot A.E., 2 computer graphics

Adjiashvili D., xxiii, 376 transformations and matrices, 31–35
air traffic control conoids, 352
angle between superplanes, 133 parabolic conoid, 352
collision avoidance, 103–106 Plücker’s conoid, 352
conflict intervals, 106 Whitney’s umbrella, 352
conflict resolution, 108–113 construction algorithms
information display, 101 in R3 , 136–159
particle fields, 103–113 half spaces, 136
angle between two planes, 133 intersecting planes, 138
approximate quadric hypersurfaces, 359– line on a plane, 136
365 parallel plane on point, 151
Archimedes, xv plane line intersection, 145
astroidal surface, 370 rotation ↔ translation, 151
Atariah, D., 322 special planes, 141
automatic classification, 406 the four indexed points, 138
Avidan S. and T., xxiii in R4 , 174–179
axes intersecting hyperplanes, 175
axis translation → rotation, 133 the five indexed points, 174
permutations, 122 convex surfaces, 367
spacing, 125 course
prerequisites, xxii
projects, xix
Banchoff T.F., 2 syllabus, xviii–xxii
Bertin J., xxv curves, 195–241
Binniaminy M., 322 algebraic curves, 208
Boz M., xxiii Plücker class formula, 230
bump, 370 Plücker curves, 229–230
examples, 208–211
conics
Chatterjee A., xxiii, 250, 326 classification of transforms, 220–
Chen C., xxv 226
Chen C.C., xxv ideal points, 226–227
Cohen L., xxiii, 376 the Indicator, 226
Cohen S., 429 transforms via Möbius transforma-
Cohen-Ganor S., 85 tion, 217
displaying several lines, 429–442 conics ↔ conics, 216–229
collinearity construction algorithm convex sets and their relatives, 231–
example in R6 , 173 241
collinearity construction algorithm RN , convex upward and downward, 197
170 curve plotting, 216
collinearity property p-flats in RN , 169 cusps and inflection points, 199–204
548 Index

duality symmetry and coincindence, 333

bitangent ↔ double −point, 230 tangent planes, 311
envelope, 43 dimensionality, 2
exponential, 212 dimple, 370
gconics – generalized conics, 231–241 Dimsdale B., xxiii, 361, 375
convex hull with curved edges, 235 conics, 217
convex-hull construction, 235 hyperplanes, 165
gconics ↔ gconics, 232–241 duality, 10
inner intersection, 238 R2 – slope and x-coordinate relation,
operational dualities, 238 51
outer union, 238 bitangent ↔ double point, 230
outer union ↔ inner intersection, cusp ↔ inflection point, 199–204
240 gconics – generalized conics
Inselberg transformation, 206 operational dualities, 238
line, 43 outer union ↔ inner intersection,
oscillatory, 212 240–241
plotting, 207 in 2-D point ↔ line, 15
point, 43
in 3-D point ↔ plane, 15
point-curves ↔ point-curves, 204–
point ↔ line in P2 , 49–62
241
point↔ line
point-curves and line-curves, 195– ¯ 51
horizontal position of ,
196, 204
line → point linear transformation,
space curves, 216
53
trigonometric, 212
point → line linear transformation,
54
d’Ocagne, xvi point-curves and line-curves, 55
data mining, xvi, xvii, xxi, 378–427 projective plane, 49
classification, xvii, xxi recognizing orthogonality, 58
grand vision, 5 rotation ↔ translations, 57
guidelines and strategies, 4 transformations, 56–60
introduction, 379–382
Piatesky-Shapiro G., xxv
Zaavi J., xxv
decision support, xxi, 359, 416–418 Ehrenfest P., 2
Desargue Eickemeyer
theorem, 10 representation of p-flats, 167
developable surfaces representation of hyperplanes, 166
characteristic point, 312 superplanes – E, 125
duality with space curves, 312 Eickemeyer J., xxiii, 125
general cone, 341 ellipsoid, 367
helicoid, 342 envelopes, 55, 118, 185–194
reconstruction, 319 formulation and computation, 186
reconstruction from its rulings, 321 necessary and sufficient conditions,
representation, 314 188
ruling, 312 of families of curves, 190–193
rulings of cylinders, 326 one-parameter family of curves, 185
start points, 319 singural points, 189
 Index 549

Euclid half plane

elements, 8 oriented, 197
parallels axiom, 15 half spaces, 136
exploratory data analysis – EDA, xvii, xxi, Hearst M., xxv
382–383 Helfman N.
data display, 380 decision support, xxiii
hundreds of variables, 400 helicoid (developable), 342
interactivity, 380 Hinterberger H.
pattern recognition, 379 comparative visualization, xxiv
simple case study, 383–390 data density analysis, xxiv
software, 4 homogeneous coordinates, 22–43
software requirements, 4 analytical proofs, 39–43
testing assumptions, 401 for finite geometries, 24
exploratory data analysis – EDA, xxiii on a line, 25
exterior point, 349 Hung C.K., xxiii, 314, 326, 335, 336, 375
convex-hull with curved edges, 235
surface representation decomposition,
financial dataset, 392, 401 451–460
Fiorini P., xxiii Hurwitz A., xxiii
hypercube, 297
hyperellipse, 362
hyperellipsoids, 360
geometry, 7–48
affine, 48
constructions ILM
compass alone, 8 duality in P2
first algorithms, 8 point ↔ line, 51, 56, 58
Poncelet and Steiner conditions, 8 geometry
Egyptians, 7 Desargues’s theorem, 42
Euclid Pappus’s theorem, 42
elements, 8 point ↔ line duality, 19
Greeks, 7 multidimensional lines
origins, 7 general case, 83
parallelism, 2 indexed points – adjacent and base
projective, 8–45 variable, 67
transformations three-point collinearity, 71
invariants, 45–48 planes
geovisualization, xxiv rotation ↔ translation for indexed
Gibbs W.J., 333 points, 122, 153
Giles N., 236 incidence matrix, 300
GIS, xxiv intelligent process control, 363–365
gold, high price domain, 363
correlation with currencies, 397 instrument panel, 364
grand tour learning, 365
Asimov D., xxiv rule, 363
Greenshpan O., 460–475 interaxis distances
Grinstein G., xxv recursive notation, 299
group, 46 Itai T., 322
550 Index

Keim D., xxiv rotations ↔ translations duality, 87

Kosloff Z., 322 separation in the xy plane
Kriz R., 2 above and below relations, 149

Lifshits Y., 373 nearly developable surfaces, 343

network
visualization and analysis, 460
Möbius strip, 354 nomography, xvi
Makarov I., 373 nonlinear models, 416–418
Matskewich T., xxiii, 250, 273, 279 nonorientability
proximate hyperplanes construction, in R2 , 149
279 projective plane P2 , 13
proximate hyperplanes representation, visualization, 358
273 nonorientable surface, 354
multidimensional lines
3-point collinearity, 127
air traffic control, 100–113 paraboloid, 367
construction algorithms, 79–82 parallel coordinates
example, 80 P2 , 49–62
displaying several lines, 429–442 bitangent ↔ double-point duality,
distance and proximity, 88–100 230
L2 minimum distance – monotonic- air traffic control, 100–113
ity with dimension, 97–100 algebraic curves, 229–230
L2 versus L1 minimum distance, 99 applications, xxiv, 5
constrained L1 distance, 92–96 automatic classification, 406
intersecting lines, 88 automatic classifiers, 4
minimum distance between lines, bare essentials, 427
91–100 complex-valued functions, 476–488
nonintersection, 89–100 conics, 216–229
indexed points curves, 55, 195–241
3-point collinearity, 70–72 curve plotting, 207–216
arbitrary parametrization, 75 cusp ↔ inflection-point duality, 199–
dummy index, 77 204
identification, 82 data mining, 378–427
representation, 63–88 decision support, 416–418
adjacent variables, 63–65 definition, 3
base variable, 65 enhancing with curves, xxiv
convenient display of a line with exploratory data analysis – EDA, 382–
positive slopes, 83 383
convenient display of several lines, atomic queries, 385
85 guidelines, 383–390
general case, 75–78 multidimensional detective, 382–
line on a plane, 119 383
transforming positive to negative requirements, 381
slopes, 83–85 first proposed, 2
two-point form, 68–70 gconics – generalized conics
two-point form continued, 78 operational dualities, 238
 Index 551

hypercube, 61 line on a plane, 119

in 3-D, xxiv planes in R3 , 115
Inselberg transformation, 206 planes on a line, 122
multidimensional lines, 63–113 point on a plane, 118
representation, 70 recursive construction, 164
network visualization and analysis, recursive construction algorithm RN ,
460–475 170
notation in R2 representation
bar, point, line, 55 p-flat in RN , 167
orthogonality, 58 augmented, 132
parallax, 4 by two indexed points in R3 , 128
parallelism, 51 collinearity property for p-flats, 169
permutations for adjacencies, 388 hyperplane in RN , 166
planes, p-flats, and hyperplanes, 115– hyperplanes and p-flats in R4 , R5 ,
183 168
point ↔ line duality indexed points, 124
queries, 387 indexed points in RN , 164
properties, 3 planar coordinates, 118
proximate planes and hyperplanes, recursive construction, 164
243–295 recursive construction algorithm,
recursive construction algorithm, 4 127
representation mapping I, 73–75 unification for p-flats 0 ≤ p < N,
separation in R2 and R3 , 149 168
separation in the xy plane, 196 vertical lines, 115
software, 4 rotation ↔ translation indexed points,
starplots, xxiv 151
surfaces in RN , 297–377 rotation ↔ translation vertical lines,
visual data mining – EDA, 4 120
Pascal, 44 separating points clusters on different
pattern recognition, 1 planes, 443–451
pi (π), 7 separation in R3 – points and planes,
planes, p-flats, and hyperplanes 149
approximate planes and flats, 179 superplanes sp, 125
axis spacing, 125 in R4 , 159
axis translation → rotation uper-planes sp
rotation angle in R3 , 133 in R3 , 126
coefficients and distance between point ↔ line duality in P2 , 49–62
adjacent points, 134 projective geometry, 9–45
collinearity construction algorithm 3-point collinearity, 24
example in R6 , 173 analytic, 22–43
collinearity construction algorithm proofs, 39–43
RN , 170 augmented lines, planes, space, 12
coplanarity axioms, 10, 13–16
in industrial data, 117 conics, 43
indexed points, 125 cross-ratio, 27
near coplanarity, 179 invariant under perspectivity, 27
vertical lines pattern, 116 Desargue’s theorem, 12
determining multiple planes, 180 duality, 9, 16–19
552 Index

finite geometries, 19–21 matching points and navigation,

fundamental transformations, 26–29 287–295
homogeneous coordinates, 9, 22–43 properties of fN , 254–257
analytical proofs, 39–43 region , 257–275
for finite geometries, 24 region crossing x axis, 287
on a line, 25 some examples, 283–295
ideal elements, 11 proximity of lines and line neighbor-
incidence relation, 10 hoods, 244–251
line coordinates, 23 topology for proximity, 243–245
linearizability of projective transfor-
mations, 29–31
quasars data set, 390–392
Pappus’s theorem and its dual, 18
queries
Pappus’s theorem generalizations, 44 atomic, 385
Brianchon’s theorem dual of Pas- compound, 401
cal’s theorem, 45 zebra, 397
Pascal theorem for conics, 44
perpectivity and projectivity, 28
perspectivity, 26 recursive construction algorithm, 127
primitives, 10 general in RN , 164
projections and sections, 10 Reif M., xxiii
projective plane model, 10–13 representation
Hilbert, 12 astroidal representation, 370
projective plane nonorientability, 13 bump, 370
projective transformations convex surfaces, 369
in P2 , 32 dimple, 370
in P3 , 33 Möbius strip, 358
linearized, 31–35 sphere, 367
perspective, 32, 34 torus, 370
rotation, 32, 33 representation mapping
scaling, 32 version II, 180
translation, 32, 33 Rivero J., xxiii
proximate planes and hyperplanes, 243– rotation, 158
295 rotation ↔ translation points, 151–158
rotation ↔ translation vertical lines, 120
Chatterjee, A., 250
rotations
Matskewich, T., 250
coordinate system in R3 , 36
neighborhood, 244
in R2 , 32
outline of development, 253
in R3 , 33
proximity of hyperplanes, 251–295
ruled surfaces
adding the c0 , 280–285
conoids, 352
components and structure of , Möbius strip, 354
257–263 representation, 349
conditions for  being bc, uc, or gh, saddle, 347
267–274
costruction of , 263–275
formulation of the problem, 251– saddle, 347
254 Satellite data set, 383–390
full image of NH, 275–279 separation in R2 – points and lines, 149
 Index 553

separation in R3 – points and planes (new hypercube, 297

approach), 146 matching algorithm, 301, 303, 311,
separation in R3 – points and planes (old 348, 351, 352, 372
approach), 149 more general surfaces, 365–377
separation in the xy plane nearly developable surfaces, 343
above and below relations, 196 normal vector, 302
convex upward or downward, 197 preliminaries, 297–300
oriented half plane, 196 representation, 301
Shahaf N., 443–451 algebraic surface, 306
Shneiderman B., xxv astroidal surface, 370
Singer Y., 460–475 boundaries, 305
Snow J., xv boundaries - example, 307
Spence R., xxv boundaries - the class EG, 305
sphere, 367 boundaries of quadrics, 306
Spivak M., 376 boundaries of ruled surfaces, 348,
surfaces 351
representation decomposition and devel- boundaries, ideal points, 305
opable quadrics, 451–460 convex surface boundaries, 366
surfaces in RN , 297–377 convex surfaces, 367, 369
approximate quadrics, 359–365 coordinates of representing points,
decision support, 359 303
hyperellipsoids, 360 Möbius strip, 358
intelligent process control, 362–365 ruled surfaces, 347
boundary contours, 304–310 saddle, 347
developable surfaces, 310–346 sphere, 367
ambiguities and uniqueness, 316– torus, 370
319
representation of developable and
characteristic point, 312 ruled, 358
duality with space curves, 312
representing points relations, 300
general, 342–346
ruled surfaces, 346–359
general cone, 341
conoids, 352
helicoid, 342
Möbius strip, 354
Hung C.K., 311
representation, 349
reconstruction, 319–325, 341
saddle, 347
reconstruction from its rulings, 321
representation, 314
ruling, 312
rulings of cylinders, 326 to see C2 , 476–488
specific developables, 325–341 torus, 370
specific developables, cones, 336– transformations
341 invariants, 45
specific developables, cylinders, linear, 46
326–336 linear fractional, 30
symmetry and coincidence, 333 linearizability of projective transfor-
tangent planes, 311 mations, 31
formulation, 301–304 Möbius, 30
general homomorphic to group of nonsingu-
matching algorithm, 372–377 lar matrices, 217
554 Index

visualization Ward M., xxv

Archimedes, xv Ware C., xxv
early success of, xv
emergence of, xv
information, 1
Yaari Y., 476–488
multidimensional, xvi, 1
science, 2, 333
scientific, 1
VLSI data set, 401–406 zero defects, 401–405