Siggraph2001 Visualizingquaternions

Visualizing Quaternions
Course Notes for SIGGRAPH 2001

Course Organizer
Andrew J. Hanson
Computer Science Department
Indiana University
Bloomington, IN 47405 USA
Email: hanson@cs.indiana.edu
Abstract
This tutorial focuses on establishing an intuitive visual understanding of the relationship between
ordinary 3D rotations and their quaternion representations. We begin building this intuition by
showing how quaternion-like properties appear and can be exploited even in 2D space. Quater-
nions are then introduced in several alternative representations that do not necessarily require ab-
stract mathematical constructs for their visualization. We then proceed to develop visualizations of
quaternion applications such as moving frames and orientation splines. Finally, we briefly discuss
the problem of generalizing quaternion concepts to higher dimensions using Clifford algebras.
Presenter’s Biography
Andrew J. Hanson is a professor of computer science at Indiana University, and has regularly
taught courses in computer graphics, computer vision, and scientific visualization. He received a
BA in chemistry and physics from Harvard College in 1966 and a PhD in theoretical physics from
MIT in 1971. Before coming to Indiana University, he did research in theoretical physics at the
Institute for Advanced Study, Stanford, and Berkeley, and then in computer vision at the SRI Artifi-
cial Intelligence Center in Menlo Park, CA. He has published in IEEE Computer, CG&A, TVCG,
ACM Computing Surveys, and has over a dozen papers in the IEEE Visualization Proceedings.
He has also contributed three articles to the Graphics Gems series dealing with user interfaces for
rotations and with techniques of N-dimensional geometry. Previous experience with conference
tutorials includes a Siggraph ’98 tutorial on N-dimensional graphics, a Visualization ’98 course
on Clifford Algebras and Quaternions, and a tutorial on Visualizing Quaternions presented at both
Siggraph ’99 and Siggraph 2000. Major research interests include scientific visualization, machine
vision, computer graphics, perception, and the design of interactive user interfaces for virtual re-
ality and visualization applications. Particular visualization applications currently being studied
include an astrophysical treatment of the local galactic neighborhood of the sun, the exploitation
of constrained navigation for visualization environments, and applications of graphics in dimen-
sions greater than three to mathematics and theoretical physics.
1
Contents
Abstract 1
Presenter’s Biography 1
Contents 3
General Information 4
1 Overview 5
2 Fundamentals of Quaternions 5
3 Visualizing Quaternion Geometry 5
4 Quaternion Frames 6
5 Clifford Algebras 6
6 MeshView and Mathematica Demonstration Software 7
References 7
Slides: I: Fundamentals of Quaternions
Slides: II:Visualizing Quaternion Geometry
Slides: III: Quaternion Frames
Slides: IV: Clifford Algebras
Paper: “Geometry for N-dimensional Graphics,” Andrew J. Hanson

Paper: “Rotations for N-dimensional Graphics,” Andrew J. Hanson
Paper: “Visualizing Quaternion Rotation,” J. C. Hart, G. K. Francis, and L. H. Kauffman
Paper: “Constrained Optimal Framings of Curves and Surfaces using Quaternion Gauss
Maps,” Andrew J. Hanson
Paper: IUCS Technical Report 518: “Quaternion Gauss Maps and Optimal Framings of
Curves and Surfaces,” Andrew J. Hanson
Paper: “Quaternion Frame Approach to Streamline Visualization,” A.J. Hanson and H. Ma
Paper: “Meshview: Visualizing the Fourth Dimension,” A.J. Hanson, K. Ishkov, and J. Ma
2
General Information on the Tutorial
Course Syllabus
Summary: This tutorial will deal with visualizable representations of quaternions, their fea-
tures, technology, folklore, and applications. The introduction will focus on visually understand-
ing quaternions themselves by exploiting parallels to complex variables and 2D rotations. Starting
from this basis, the tutorial will proceed to give visualizations of advanced quaternion applications.
Prerequisites: Participants should be comfortable with and have an appreciation for conven-
tional mathematical methods of 3D computer graphics and geometry used in graphics transfor-
mations and rendering. The material will be of most interest to those wishing to deepen their
intuitive understanding of quaternion-based animation, moving coordinate frames, and 3D curves
and surfaces appearing in graphics and scientific visualization applications.
Objectives: Participants will learn the basic facts relating quaternions to ordinary 3D rotations,
as well as methods for examining the properties of quaternion constructions using interactive visu-
alization methods. A variety of applications, including quaternion splines and moving coordinate
frames for curves and surfaces, will be examined in this context. Finally, a few facts about the
deeper relationship between quaternions and Clifford algebras in higher dimensions will be pre-
sented.
Outline: This is a two-hour tutorial and the material will be arranged approximately as follows:
I. (45 min) Introduction to Rotation Representations. Develop formulas and techniques

for seeing how 2D rotations, orientation frames, and their time evolution equations can be
visualized and studied using ordinary complex variables. Develop the parallel relationship
between 3D rotations and quaternions.
II. (45 min) Visualization Techniques for Quaternions. Visualizing static and moving quater-
nion frames, along with quaternion splines, as 4D geometric objects.
III. (15 min) Applications of Quaternion Visualization. Extend this intuition into the quater-
nion representation of moving orientation frames.
IV. (15 min) Clifford Algebras: the Bigger Picture. Start to see how it all fits into Clifford
algebras.
3
1 Overview
Practitioners of computer graphics and animation frequently represent 3D rotations using the
quaternion formalism, a mathematical tool that originated with William Rowan Hamilton in the
19th century, and is now an essential part of modern analysis, group theory, differential geometry,
and even quantum physics. Quaternions are in many ways very simple, and yet there are enormous
subtleties to address in the process of fully understanding and exploiting their properties. The
purpose of this Tutorial is to construct an intuitive bridge between our intuitions about 2D and 3D
rotations and the quaternion representation.
The Tutorial will begin with an introduction to rotations in 2D, which will be found to have
surprising richness, and will proceed to the construction of the relation between 3D rotations and
quaternions. Quaternion visualization methods of various sorts will be introduced, followed by
some applications of the quaternion frame representation to problems of interest by graphicists
and visualization scientists. Finally, we will briefly touch on the relationship between Clifford
algebras and quaternion rotation representations. An extensive bibliography of related literature is
included, as well as several relevant reprints and technical reports, a Mathematica implementation
of the Quaternion Frenet Equations, and the Meshview software system for viewing 4D objects.
2 Fundamentals of Quaternions
We will begin with a basic introduction to rotations in general, showing how 2D rotations contain
the seeds for what we need to understand about 3D rotations; see, for example, [39]. We will
then proceed to look at a variety of methods for understanding quaternions and making meaningful
pictures of constructs involving them. These methods will range from some of the ideas introduced
by Hart, Francis, and Kauffman [53] for motivating the need for double-valued parameterizations
of rotations, to theoretical background given in [46, 47, 40, 50].
Traditional treatments of quaternions range from the original works of Hamilton and Tait [35,
81] to a variety of recent studies such as those of Altmann, Pletincks, Juttler, and Kuipers [2, 69,
59, 63]. The 4D frames of the quaternions themselves, in contrast to the relationship between 3D
frames and quaternions, are treated in the German literature, e.g., [12, 65].
In our treatment, we will focus on the use of 2D rotations as a rich but algebraically simple
proving ground in which we can see many of the key features of quaternion geometry in a very
manageable context. The relationship between 3D rotations and quaternions is then introduced as
a natural extension of the 2D systems.
3 Visualizing Quaternion Geometry

In order to clearly understand our options for making graphical visualizations of quaternions, we
next look at the ways in which points on spheres can be viewed in reduced dimensions, discovering
luckily that 3D graphics is just sufficient to make a usable interactive workstation system for look-
ing at quaternions, quaternion curves, and even quaternion surfaces. The basic “trick” involves
4
the observation that if we have a four-vector quaternion q = (q0 ; q) obeying q q = 1, then the
four-vector lies on the three-sphere S3 and has only
p three independent components: if we display
just q, we can in principle infer the value of q0 = 1 , q q. We supply a viewer, the Meshview
system [51, 52] developed by the presenter and his students, which allows the input and interactive
examination of quaternion objects.
4 Quaternion Frames
In this section, we study the nature of quaternions as representations of frames in 3D. Our visu-
alizations again exploit the fact that quaternions are points on the three-sphere embedded in 4D;
the three-sphere (S3 ) is analogous to an ordinary ball or two-sphere (S2 ) embedded in 3D, except
that the three-sphere is a solid object instead of a surface. To manipulate, display, and visualize
rotations in 3D, we may convert 3D rotations to 4D quaternion points and treat the entire problem
in the framework of 4D geometry. The methods in this section follow closely techniques intro-
duced in Hanson and Ma [46, 47] for representing families of coordinate frames on curves in 3D
as curves in the 4D quaternion space. The extensions to coordinate frames on surfaces and the
corresponding induced surfaces in quaternion space are studied in [40, 50].
The same methods extend to the study of quaternion animation splines, introduced to the graph-
ics community originally by Shoemake [73]. We give an overview of the issues of constructing
splines with various desirable continuity properties following the method of Schlag [71] applied
to quaternion Bezier, Catmull-Rom, and uniform B-splines. These methods do not easily permit
long-lasting periodic motion; this problem, as well as some problems with derivative computation,
are resolved by the exponential map approach of Kim, et al. (see, e.g., [61, 32]); The exponential
map and its visualization properties are therefore briefly noted as well. Alternative approaches
such as the rational quaternion spline method of Jüttler [59, 60] and the variational methods of
Barr et al. [10, 70] are mentioned but not treated in detail.
5 Clifford Algebras
The quaternion-based formalism for handling and visualizing rotations works well in dimensions 2,
3, and 4 because in those dimensions the Spin group, the double covering of the orthogonal group,
has simple topology and geometry. Going beyond four dimensions is of course much harder.
Clifford algebras form the basis used in pure mathematics to treat the Spin groups in arbitrary
dimensions (see, e.g., [3, 58]); furthermore, viewed in the context of arbitrary dimensions, studying
the Clifford algebra approach provides additional depth to our understanding of dimensions 2, 3,
and 4 — we can get a better feeling for what properties are accidents of the low dimension and
which are in fact general and extensible concepts.
5
6 MeshView and Mathematica Demonstration Software
The course CDROM should contain a copy of the MeshView 4D visualization package [51, 52],
which runs under SGI-IRIX, SUN-SOLARIS, and, with some effort, Linux. This contains a sample
of quaternion data sets, as well as many other 4D objects, and has facilities specifically designed for
quaternion visualization experiments. There is also a Mathematica notebook qfrmint.nb that
explicitly implements a numerical integration of the Frenet frame equations in quaternion form,
vastly improving the exactly equivalent calculation for the standard Frenet equations implemented
by Gray [33].
Acknowledgments
The slightly edited versions of the author’s papers from Graphics Gems IV and V [38, 39] are in-
cluded in the CDROM and Course Notes with the kind permission of Academic Press. Republished
in the Course Notes are two key papers from IEEE Transactions on Visualization and Computer
Graphics [47], and from the Proceedings of IEEE Visualization [40] of the IEEE Computer Society
Press. We thank Ji-Ping Sha for his patience with Clifford algebras, and John Hart for supplying
the paper by Hart, Francis, and Kauffman included here.
References
[1] B. Alpern, L. Carter, M. Grayson, and C. Pelkie. Orientation maps: Techniques for visual-
izing rotations (a consumer’s guide). In Proceedings of Visualization ’93, pages 183–188.
IEEE Computer Society Press, 1993.
[2] S. L. Altmann. Rotations, Quaternions, and Double Groups. Oxford University Press, 1986.
[3] M.F. Atiyah, R. Bott, and A. Shapiro. Clifford modules. Topology, 3, Suppl. 1:3–38, 1986.
[4] T. Banchoff and J. Werner. Linear Algebra through Geometry. Springer-Verlag, 1983.
[5] T. F. Banchoff. Visualizing two-dimensional phenomena in four-dimensional space: A com-
puter graphics approach. In E. Wegman and D. Priest, editors, Statistical Image Processing
and Computer Graphics, pages 187–202. Marcel Dekker, Inc., New York, 1986.
[6] Thomas F. Banchoff. Beyond the Third Dimension: Geometry, Computer Graphics, and
Higher Dimensions. Scientific American Library, New York, NY, 1990.
[7] David Banks. Interactive display and manipulation of two-dimensional surfaces in four di-
mensional space. In Symposium on Interactive 3D Graphics, pages 197–207, New York,
1992. ACM.
6
[8] David Banks. Interactive manipulation and display of two-dimensional surfaces in four-
dimensional space. In David Zeltzer, editor, Computer Graphics (1992 Symposium on Inter-
active 3D Graphics), volume 25, pages 197–207, March 1992.
[9] David C. Banks. Illumination in diverse codimensions. In Computer Graphics, pages 327–
334, New York, 1994. ACM. Proceedings of SIGGRAPH 1994; Annual Conference Series
1994.
[10] A. Barr, B. Currin, S. Gabriel, and J. Hughes. Smooth interpolation of orientations with
angular velocity constraints using quaternions. In Computer Graphics Proceedings, Annual
Conference Series, pages 313–320, 1992. Proceedings of SIGGRAPH ’92.
[11] Richard L. Bishop. There is more than one way to frame a curve. Amer. Math. Monthly,
82(3):246–251, March 1975.
[12] Wilhelm Blaschke. Kinematik und Quaternionen. VEB Deutscher Verlag der Wis-
senschaften, Berlin, 1960.
[13] Jules Bloomenthal. Calculation of reference frames along a space curve. In Andrew Glassner,
editor, Graphics Gems, pages 567–571. Academic Press, Cambridge, MA, 1990.
[14] Kenneth A. Brakke. The surface evolver. Experimental Mathematics, 1(2):141–165, 1992.
The “Evolver” system, manual, and sample data files are available by anonymous ftp from
geom.umn.edu, The Geometry Center, Minneapolis MN.
[15] D. W. Brisson, editor. Hypergraphics: Visualizing Complex Relationships in Art, Science and
Technology, volume 24. Westview Press, 1978.
[16] S. A. Carey, R. P. Burton, and D. M. Campbell. Shades of a higher dimension. Computer

Graphics World, pages 93–94, October 1987.
[17] Michael Chen, S. Joy Mountford, and Abigail Sellen. A study in interactive 3-d rotation
using 2-d control devices. In Proceedings of Siggraph 88, volume 22, pages 121–130, 1988.
[18] H.S.M. Coxeter. Regular Complex Polytopes. Cambridge University Press, second edition,
1991.
[19] R. A. Cross and A. J. Hanson. Virtual reality performance for virtual geometry. In Proceed-
ings of Visualization ’94, pages 156–163. IEEE Computer Society Press, 1994.
[20] A.R. Edmonds. Angular Momentum in Quantum Mechanics. Princeton University Press,
Princeton, New Jersey, 1957.
[21] N.V. Efimov and E.R. Rozendorn. Linear Algebra and Multi-Dimensional Geometry. Mir
Publishers, Moscow, 1975.
7
[22] T. Eguchi, P. B. Gilkey, and A. J. Hanson. Gravitation, gauge theories and differential geom-
etry. Physics Reports, 66(6):213–393, December 1980.
[23] L. P. Eisenhart. A Treatise on the Differential Geometry of Curves and Surfaces. Dover, New
York, 1909 (1960).
[24] S. Feiner and C. Beshers. Visualizing n-dimensional virtual worlds with n-vision. Computer
Graphics, 24(2):37–38, March 1990.
[25] S. Feiner and C. Beshers. Worlds within worlds: Metaphors for exploring n-dimensional
virtual worlds. In Proceedings of UIST ’90, Snowbird, Utah, pages 76–83, October 1990.
[26] Gerd Fischer. Mathematische Modelle, volume I and II. Friedr. Vieweg & Sohn, Braun-
schweig/Wiesbaden, 1986.
[27] H. Flanders. Differential Forms. Academic Press, New York, 1963.
[28] J.D. Foley, A. van Dam, S.K. Feiner, and J.F. Hughes. Computer Graphics, Principles and
Practice. Addison-Wesley, second edition, 1990. page 227.
[29] A. R. Forsyth. Geometry of Four Dimensions. Cambridge University Press, 1930.
[30] George K. Francis. A Topological Picturebook. Springer Verlag, 1987.
[31] Herbert Goldstein. Classical Mechanics. Addison-Wesley, 1950.
[32] F.S. Grassia. Practical parameterization of rotations using the exponential map. Journal of
Graphics Tools, 3(3):29–48, 1998.
[33] Alfred Gray. Modern Differential Geometry of Curves and Surfaces. CRC Press, Inc., Boca
Raton, FL, second edition, 1998.
[34] Cindy M. Grimm and John F. Hughes. Modeling surfaces with arbitrary topology using
manifolds. In Computer Graphics Proceedings, Annual Conference Series, pages 359–368,
1995. Proceedings of SIGGRAPH ’95.
[35] W.R. Hamilton. Lectures on Quaternions. Cambridge University Press, 1853.
[36] A. J. Hanson. The rolling ball. In David Kirk, editor, Graphics Gems III, pages 51–60.
Academic Press, Cambridge, MA, 1992.
[37] A. J. Hanson. A construction for computer visualization of certain complex curves. Notices
of the Amer.Math.Soc., 41(9):1156–1163, November/December 1994.
[38] A. J. Hanson. Geometry for n-dimensional graphics. In Paul Heckbert, editor, Graphics
Gems IV, pages 149–170. Academic Press, Cambridge, MA, 1994.
8
[39] A. J. Hanson. Rotations for n-dimensional graphics. In Alan Paeth, editor, Graphics Gems
V, pages 55–64. Academic Press, Cambridge, MA, 1995.
[40] A. J. Hanson. Constrained optimal framings of curves and surfaces using quaternion gauss
maps. In Proceedings of Visualization ’98, pages 375–382. IEEE Computer Society Press,
1998.
[41] A. J. Hanson and R. A. Cross. Interactive visualization methods for four dimensions. In
Proceedings of Visualization ’93, pages 196–203. IEEE Computer Society Press, 1993.
[42] A. J. Hanson and P. A. Heng. Visualizing the fourth dimension using geometry and light. In
Proceedings of Visualization ’91, pages 321–328. IEEE Computer Society Press, 1991.
[43] A. J. Hanson and P. A. Heng. Four-dimensional views of 3d scalar fields. In Proceedings of

Visualization ’92, pages 84–91. IEEE Computer Society Press, 1992.
[44] A. J. Hanson and P. A. Heng. Foursight. In Siggraph Video Review, volume 85. ACM
Siggraph, 1992. Scene 11, Presented in the Animation Screening Room at SIGGRAPH ’92,
Chicago, Illinois, July 28–31, 1992.
[45] A. J. Hanson and P. A. Heng. Illuminating the fourth dimension. Computer Graphics and
Applications, 12(4):54–62, July 1992.
[46] A. J. Hanson and H. Ma. Visualizing flow with quaternion frames. In Proceedings of Visual-
ization ’94, pages 108–115. IEEE Computer Society Press, 1994.
[47] A. J. Hanson and H. Ma. Quaternion frame approach to streamline visualization. IEEE
Trans. on Visualiz. and Comp. Graphics, 1(2):164–174, June 1995.
[48] A. J. Hanson and H. Ma. Space walking. In Proceedings of Visualization ’95, pages 126–133.
IEEE Computer Society Press, 1995.
[49] A. J. Hanson, T. Munzner, and G. K. Francis. Interactive methods for visualizable geometry.
IEEE Computer, 27(7):73–83, July 1994.
[50] A.J. Hanson. Quaternion gauss maps and optimal framings of curves and surfaces. Indiana
University Computer Science Department Technical Report 518 (October, 1998).
[51] A.J. Hanson, K. Ishkov, and J. Ma. Meshview. A portable 4D geometry viewer written in
OpenGL/Motif, available by anonymous ftp from ftp.cs.indiana.edu:pub/hanson.
[52] A.J. Hanson, K. Ishkov, and J. Ma. Meshview: Visualizing the fourth dimension. Overview
of the MeshView 4D geometry viewer.
[53] John C. Hart, George K. Francis, and Louis H. Kauffman. Visualizing quaternion rotation.
ACM Trans. on Graphics, 13(3):256–276, 1994.
9
[54] D. Hilbert and S. Cohn-Vossen. Geometry and the Imagination. Chelsea, New York, 1952.
[55] John G. Hocking and Gail S. Young. Topology. Addison-Wesley, 1961.
[56] C. Hoffmann and J. Zhou. Some techniques for visualizing surfaces in four-dimensional
space. Computer-Aided Design, 23:83–91, 1991.
[57] S. Hollasch. Four-space visualization of 4D objects. Master’s thesis, Arizona State Univer-
sity, August 1991.
[58] H.B. Lawson Jr. and M.L. Michelsohn. Spin Geometry. Princeton University Press, 1989.
[59] B. Jüttler. Visualization of moving objects using dual quaternion curves. Computers and
Graphics, 18(3):315–326, 1994.
[60] B. Jüttler and M.G. Wagner. Computer-aided design with saptial rational B-spline motions.
Journal of Mechanical Design, 118:193–201, June 1996.
[61] Myoung-Jun Kim, Myung-Soo Kim, and Sung Yong Shin. A general construction scheme
for unit quaternion curves with simple high order derivatives. In Computer Graphics Pro-
ceedings, Annual Conference Series, pages 369–376, 1995. Proceedings of SIGGRAPH ’95.
[62] F.. Klock. Two moving coordinate frames for sweeping along a 3d trajectory. Computer
Aided Geometric Design, 3, 1986.
[63] J.B. Kuipers. Quaternions and Rotation Sequences. Princeton University Press, 1999.
[64] J. Milnor. Topology from the Differentiable Viewpoint. The University Press of Virginia,
Charlottesville, 1965.
[65] Hans Robert Müller. Sphärische Kinematik. VEB Deutscher Verlag der Wissenschaften,
Berlin, 1962.
[66] G. M. Nielson. Smooth interpolation of orientations. In N.M. Thalman and D. Thalman,

editors, Computer Animation ’93, pages 75–93, Tokyo, June 1993. Springer-Verlag.
[67] Michael A. Noll. A computer technique for displaying n-dimensional hyperobjects. Commu-
nications of the ACM, 10(8):469–473, August 1967.
[68] Mark Phillips, Silvio Levy, and Tamara Munzner. Geomview: An interactive geometry
viewer. Notices of the Amer. Math. Society, 40(8):985–988, October 1993. Available by
anonymous ftp from geom.umn.edu, The Geometry Center, Minneapolis MN.
[69] D. Pletincks. Quaternion calculus as a basic tool in computer graphics. The Visual Computer,
5(1):2–13, 1989.
10
[70] Ravi Ramamoorthi and Alan H. Barr. Fast construction of accurate quaternion splines. In
Turner Whitted, editor, SIGGRAPH 97 Conference Proceedings, Annual Conference Series,
pages 287–292. ACM SIGGRAPH, Addison Wesley, August 1997. ISBN 0-89791-896-7.
[71] John Schlag. Using geometric constructions to interpolate orientation with quaternions. In
James Arvo, editor, Graphics Gems II, pages 377–380. Academic Press, 1991.
[72] Uri Shani and Dana H. Ballard. Splines as embeddings for generalized cylinders. Computer
Vision, Graphics, and Image Processing, 27:129–156, 1984.
[73] K. Shoemake. Animating rotation with quaternion curves. In Computer Graphics, volume 19,
pages 245–254, 1985. Proceedings of SIGGRAPH 1985.
[74] K. Shoemake. Animation with quaternions. Siggraph Course Lecture Notes, 1987.
[75] Ken Shoemake. Arcball rotation control. In Paul Heckbert, editor, Graphics Gems IV, pages
175–192. Academic Press, 1994.
[76] Ken Shoemake. Fiber bundle twist reduction. In Paul Heckbert, editor, Graphics Gems IV,
pages 230–236. Academic Press, 1994.
[77] D.M.Y. Sommerville. An Introduction to the Geometry of N Dimensions. Reprinted by Dover

Press, 1958.
[78] N. Steenrod. The Topology of Fibre Bundles. Princeton University Press, 1951. Princeton
Mathematical Series 14.
[79] K. V. Steiner and R. P. Burton. Hidden volumes: The 4th dimension. Computer Graphics
World, pages 71–74, February 1987.
[80] D. J. Struik. Lectures on Classical Differential Geometry. Addison-Wesley, 1961.
[81] P.G. Tait. An Elementary Treatise on Quaternions. Cambridge University Press, 1890.
[82] J. R. Weeks. The Shape of Space. Marcel Dekker, New York, 1985.
[83] S. Weinberg. Gravitation and Cosmology: Principles and Applications of General Relativity.
John Wiley and Sons, 1972.
[84] E.T. Whittaker. A Treatise on the Analytical Dynamics of Particles and Rigid Bodies. Dover,
New York, New York, 1944.
11
GRAND PLAN
I: Fundamentals of Quaternions
Andrew J. Hanson
Computer Science Department II: Visualizing Quaternion Geometry
Indiana University
III: Quaternion Frames
Siggraph 2001 Tutorial
IV: Clifford Algebras
1 2
II: Visualizing Quaternion Geometry

I: Fundamentals of Quaternions
The Spherical Projection Trick: Visualizing
Motivation unit vectors.
2D Frames: Simple example, complex numbers. Quaternion Frames

3D Frames: Rotations and quaternions. Quaternion Curves
Quaternion Splines
3 4
IV: Clifford Algebras

III: Quaternion Frames
Clifford Algebras: Generalize quaternion struc-
Quaternion Curves: generalize the Frenet Frame ture to N-dimensions
Quaternion Frame Evolution Reflections and Rotations: New ways of look-

Quaternion Curve and Surface Optimization
ing at rotations
Pin(N), Spin(N), O(N), and SO(N)
5 6
Part I: OUTLINE
Part I: Fundamentals
of Motivation
Quaternions 2D Frames: Simple example, complex numbers.
Andrew J. Hanson 3D Frames: Rotations and quaternions.
Indiana University
7 8
Motivation
Quaternion methods are now commonplace in Basic Issues

graphics.
Basic fact number 1: Rotation matrices are Co-
Quaternions are very geometric, but we seldom ordinate Frame Axes.
attempt to visualize their properties geometri- Basic fact number 2: Rotation matrices form
cally. groups, which have geometric properties.
That’s going to be our job today!

9 10
Task vs Strategy Simple Example: 2D Rotations
Our task: Understand Rotations.

2D rotations ) geometric origin for complex numbers.
Rotations don’t just act on geometry, rotations Complex numbers are a special subspace of quater-
are geometry. nions.
Our strategy: the geometry should help us Thus 2D rotations introduce us to quaternions and their
to visualize the properties of rotations.
geometric meaning.
11 12
Frames in 2D Frames in 2D
The tangent and normal to 2D curve move continuously along The tangent and normal to 2D curve move continuously along
the curve: the curve:
^ ^
N N
^
^ T
T
θ θ
13 14
Frames in 2D
Frame Matrix in 2D
The tangent and normal to 2D curve move continuously along
the curve:
This motion is described at each point (or time) by the matrix:
h i
^
N ^
T
R2() = ^ N
T ^
2 3
θ
4 cos , sin 5 :
sin cos
=
15 16
The Belt Trick:

Another 2D Frame
Is There a Problem?
If we did not know about cos
2 + sin
2 = 1, we might
Demonstration: Rotations “want to be doubled” to get
represent the frame differently, e.g., as:
2 3 back where you started.
R2(A; B ) = 4 A B 5 , :
B A
with the constraint A2 + B 2 = 1.
See: Hart, Francis, and Kauffman.
17 18
Half-Angle Transform:
A Fix for the Problem?

A Fix for the Problem? Or, with a = cos(=2), b = sin(=2),

(i.e., A = a2 , b2, B = 2ab),
2 3
66 cos2

2 , sin
2
2 ,
2 cos sin
2

2 7
7
we could parameterize as:
2 3
R2() =
66 77
R2(a; b) = 4 a
2 , b2 ,2ab 5 :
66 sin cos2
, 2 77 2ab a2 , b2
64 2 cos
2 2 2 sin
2 75
where orthonormality implies
a2 + b2)2 = 1
(
which reduces back to a2 + b2 = 1.

19 20
Frame Evolution in 2D
Examine time-evolution of 2D frame (on our way to 3D): First

in (t) coordinates:
2 3
So the pair ( a; b) provides an odd double-valued parame- h i
4 cos , sin 5 :
sin cos
^ N
T ^ =
terization of the frame:
2 3
h a , b2 ,2ab 5 :
i
4
2
2ab a2 , b2
^ N
T ^ =
Differentiate to find frame equations:
where (a; b) is precisely the same frame as (,a; ,b). _

^
T t ^
( ) = + N
_
^
N t
( ) = ,T^ ;
where (t) = d=dt is the curvature.
21 22
Frame Evolution in (a; b):
Frame Evolution in 2D ^; N
Using the basis (T ^ ) we have Four equations with Three
constraints from orthonormality, for One true degree of free-

Rearrange to make a “vector matrix:” dom.
2 3 2 32 3
(t) 7 +(t)
5 4 T(t) 5
_
^ ^
64 T 5 = 4
0
_
^
N ( ) t ,(t) 0 ^( )
N t Major Simplification occurs in (a; b) coordinates!!
2 3 2 32 3
_
= 24
,
aa_ bb_ 5 = 2 4 a b 5 4 a_ 5 ,
^
T
ab_ + ba_ b a b_
23 24
Frame Evolution in (a; b):
2D Quaternion Frames!
_
^
But this formula for T is just N^ , where
_
^ _
^
2 3 2 32 3 Rearranging terms, both T and N eqns reduce to
,
N^ = 4 2 2ab2 5 = 4 a ,b 5 4 ,b 5 2 3 2 3 2 3
a ,b b a a 4 a_ 5 = 1 4 0 5 4a5, b
b_ 2 + 0
or
2 32 32 3
N^ = 4 ab ,ab 5 4 01 ,01 5 4 ab 5 This is the square root of frame equations.
25 26
2D Quaternions . . . Rotation as Complex Multiplication
So one equation in the two “quaternion” variables (a; b) with If we let (a + ib) = exp (i =2) we see that
the constraint a2 + b2 = 1 contains both the frame equations rotation is complex multiplication!
_
^
T = + N^
“Quaternion Frames” in 2D are just complex numbers, with
_
^
N = ,T^
) this is much better for computer implementation, etc. Evolution Eqns = derivative of exp (i =2)!
27 28
Rotation with no matrices!

The Algebra of 2D Rotations
This is the miracle:
The algebra corresponding to 2D rotations is easy: just com-
a + ib = ei=2 plex multiplication!!
represents rotations “more nicely” than the matrices R( ). (a + ib )(a + ib)
(a ; b ) (a; b) =
0 0 0 0
a ib )(a + ib) = ei( +)=2 = A + iB = a a , b b + i(a b + ab )

0
0 0 0 0 0 0
( +

= (a a , b b; a b + ab )
0 0 0 0
where if we want the matrix, we write:

2 3 = (A; B )
R(0)R() = R(0 + ) = 4 A
2 , B 2 ,2AB 5
2AB A2 , B 2
29 30
The Geometry of 2D Rotations The Geometry of 2D Rotations
(a; b) with a2 + b2 = 1 is a point on the unit circle, also (a; b) with a2 + b2 = 1 is a point on the unit circle, also
written S 1. Rotations are just complex multiplication, and written S 1. Rotations are just complex multiplication, and
take you around the unit circle like this: take you around the unit circle like this:
(a’a-b’b, a’b+ab’) (a’, b’) (a’a-b’b, a’b+ab’) (a’, b’)
θ + θ’ θ + θ’
(a,b) 2
2 θ’/2 θ’/2
θ /2 (a,b)
(1,0) θ/2 (1,0)
31 32
The Geometry of 2D Rotations

Quaternion Frames
(a; b) with a2 + b2 = 1 is a point on the unit circle, also
written S 1. Rotations are just complex multiplication, and In 3D, repeat our trick: take square root of the frame, but now
take you around the unit circle like this:
use quaternions:
(a’a-b’b, a’b+ab’) (a’, b’)
θ + θ’
2
θ’/2
Write down the 3D frame.
θ/2 (a,b)
(1,0)
Write as double-valued quadratic form.
Rewrite linearly in the new variables.
33 34
The Geometry of 3D Rotations The Geometry of 3D Rotations
We begin with a basic fact: We begin with a basic fact:
Euler theorem: every 3D frame can be written as a spinning Euler theorem: every 3D frame can be written as a spinning
by about a fixed axis n
^, the eigenvector of the rotation ma- by about a fixed axis n
^, the eigenvector of the rotation ma-
trix: trix:
^
n ^
n
θ θ
35 36
The Geometry of 3D Rotations
Quaternion Frames . . .
We begin with a basic fact:
Euler theorem: every 3D frame can be written as a spinning Matrix giving 3D rotation by about axis n
^:
by about a fixed axis n
^, the eigenvector of the rotation ma-
trix: R3(; n^) =
2 3
^
n
6c n 2
+ ( 1) (1 , c) n1n2(1 , c) , sn3 n3n1(1 , c) + sn27
n n , c sn3 c + (n2)2(1 , c) n3n2(1 , c) , sn1775
6
6
4 1 2(1 )+
θ
n1n3(1 , c) , sn2 n2n3(1 , c) + sn1 c + (n3)2(1 , c)
^n
where c = cos , s = sin , and n ^ = 1.
37 38
Quaternion Frame Parameters
To find and axis n^, given any rotation matrix or frame M , Quaternions and Rotations
we need two steps:

Some set of axes can be chosen as the identity matrix:
TrM = 1 + 2 cos y
) solve for . 1 0 0
M , Mt = = 0 1 0
2
66
0 ,2n3 sin +2n2 sin 37 11111111
00000000 x
0 0 1
4 +2n3 sin 0 ,2n1 sin 75 00000000

11111111
00000000
11111111
,2n2 sin +2n1 sin 0
z
00000000
11111111
) solve for n^ as long as 6= 0.
39 40
Quaternions and Rotations Quaternions and Rotations
Any arbitrary set of axes forms the columns of an orthogonal Any arbitrary set of axes forms the columns of an orthogonal
rotation matrix: rotation matrix:
b
b
1111111
0000000 111111
000000
000000
111111
0000000
1111111
a
0000000
11111110000000000
1111111111 a
x
b
x
c
x 000000
111111 a
000000111111111111
000000000000
111111
a
x
b
x
c
x
0000000
1111111
0000
11110000000000
1111111111
0000000000
1111111111
= a b c
000000111111111111
000000000000
111111
= a b c
0000000
1111111
0000
1111
y y y
0000000
1111111
y y y
0000
1111 a
z
b
z
c
z 0000000
1111111
0000000
1111111
a
z
b
z
c
z
0000
1111 c1111111
0000000
c 1111
0000
0000
1111 0000000
1111111
41 42
Quaternions and Rotations Quaternions and Rotations
Any arbitrary set of axes forms the columns of an orthogonal Any arbitrary set of axes forms the columns of an orthogonal
rotation matrix: rotation matrix:
b
11111
00000 00111100b
00000
11111
00000
11111 a b c 1100 a b c
00000
1111100000000000
11111111111a x x x
11001100 x x x
00000
11111
0000000
1111111
0000011111111111
1111100000000000 = a b c
110000000000000
a = a b c
0000000
1111111 00000000
11111111 11111111111
y y y y y y
0000000
1111111 a
z
b
z
c
z
c11111111
00000000 a
z
b
z
c
z
c 1111111
0000000 00000000
11111111
0000000
1111111
43 44
Quaternions and Rotations Rotations and Quadratic Polynomials

2 3
By Euler’s theorem, that matrix has an eigenvector n
is representable as a single rotation about n
^, and so
Remember R2( ) = 4
a2 , b2 ,2ab 5?
identity:
^ applied to the
2ab a2 , b2
b
1111111
0000000 01
What if we try a 3 3 matrix R3 instead of 2 2?
0000000
1111111
a
0000000
11111110000000000
1111111111 a
x
b
x
c
x 1010
0000000
11111110000000000
1111111111
00001111111111
11110000000000
= a b c
1010 2 3
0000000
1111111
0000
1111
y y y
q 2 2q , q , q 2q1q2 , 2q0q3 2q1q3 + 2q0q2 77
2 2
0000
1111
0000
1111
a
z
b
z
c
z 10 6 0+ 1 2 3
c 1111
0000
0000
1111
6
6
4 qq
2 1 2+2 0 3q q q02 , q12 + q22 , q32 2q2q3 , 2q0q1 75
2q1q3 , 2q0q2 2q2q3 + 2q0 q1 q02 , q12 , q22 + q32
y
1 0 0
^
n
i 6= j ) pair to see a
= 0 1 0
= 0 0 1 Hint: set q1 = q2 = 0 or any other (

11111111
00000000
00000000
11111111
x
familiar sight!
θ
z 00000000
11111111
00000000
11111111
45 46
Quaternions and Rotations

Quaternions and Rotations . . .
Why does this matrix parameterize a rotation? Because Columns
of R3(q0; q1; q2; q3) are orthogonal: So if we require q02 + q12 + q22 + q32 = 1 , orthonormality is
coli colj = 0 for i 6= j assured and R3(q0; q1; q2; q3) is a rotation.
This implies q is a point on 3-sphere in 4D.

What is LENGTH of 3-vector column?

coli coli = ( 0 + q2 q12 + q22 + q32)2 NOTE: q ) ,q gives same R3().
47 48
HOW does q = (q0; q) represent rotations? Quaternions and Rotations . . .
LOOK at WHAT happens if you do TWO rotations?

?
R3(; n^) = R3(q0; q1; q2; q3)
EXAMINE the action of two rotations
NOTICE: Choosing R(q )R(q) = R(Q)

0
) EXPRESS in quadratic forms in q and LOOK FOR an analog

q(; n^) = (cos 2 ; n
^ sin
2
of complex multiplication:
makes the R3 equation an IDENTITY .

49 50
RESULT: the following multiplication rule Algebra of Quaternions

q
0 q = Q yields exactly the correct 3 3 rotation
= 3D Rotations!
matrix R(Q):
2 i 3 2
66 Q0 = q0
h
q 7
0 , 0 , q20 q2 , q30 q3 37
66 q0q0 q1q1
2D rotation matrices are represented
= q0 0 0 + q20 q3 , q30 q2 7777

i0 77
66
66 Q1
h
q 7
i1 777
66
66 q0q1 + q1q0
by complex multiplication
= q0 0 0 + q30 q1 , q10 q3 7777

=
66 Q
64 2
h
q
i2 75
7 66 q q + q q
64 0 2 2 0
= q0 0 0 0 0 5
q0q3 + q3q0 + q1q2 , q2q1
h 3D rotation matrices are represented
Q3 q
3
by quaternion multiplication
This is Quaternion Multiplication.
51 52
Algebraic 2D/3D Rotations
Therefore in 3D, the 2D complex multiplication

Algebra of Quaternions . . .
(a ; b ) (a; b) = (a a , b b; a b + ab )
0 0 0 0 0 0
The is easier to remember by dividing it into the

is replaced by 4D quaternion multiplication:
scalar piece q0 and the vector piece ~q:
q q q q , q1q1 , q2q2 , q3q3; 0q = (q00 q0 , q~0 ~q;
0 0 0 0 0
= ( 0 0
q
q0q1 + q1q0 + q2q3 , q3q2;
0 0 0 0
0 ~0
q + q0 q + q~0 ~q)
q0q2 + q2q0 + q3q1 , q1q3;
0 0 0 0
q0~
q0q3 + q3q0 + q1q2 , q2q1)

0 0 0 0
53 54
Quaternions and Rotations
Another miracle: let us generalize the 2D equation Quaternions and Rotations . . .

a + ib = ei=2
Then if we take i2 = j2 = k2 = ,1, and i j = k (cyclic),
How? We set quaternion multiplication rule is automatic!
q q ; q1; q2; q3)

= ( 0
) q = q0 + iq1 + jq2 + kq3 is the standard representation
= q0 + iq1 + jq2 + kq3
for a quaternion, and we can also use 2 2 Pauli matrices
in place of (i; j; k) if we want.
= e
^=2)
(I n

with q0 = cos(=2) and ~q = n

^ sin(=2)
and I = (i; j; k).
55 56
Summarize Quaternion Properties

Key to Quaternion Intuition
Unit four-vector. Take q = (q0; q1; q2; q3) = (q0;~q) to
Fundamental Intuition: We know obey constraint q q = 1.
q0 = cos(=2); ~q = n^ sin(=2)
We also know that any coordinate frame M can be written Multiplication rule. The quaternion product q and p is
as M = R(; n
^). q p = (q0p0 , ~q ~p; q0~p + p0~q + ~q ~p),
Therefore or, alternatively,
2 3 2
~q points exactly along the axis we have to rotate
6
[q p]0
7 6 ,
q0p0 q1p1 q2p2 , , q3p3 37
around to go from identity I to M , and the length of
6
6
[q p]1 77 6
6 q 0p1 + q1 p0 + q2p3 , q3p2 777
6
4
6 [q p]2 77=6
5 6 q0p2 + q2p0 + q3p1
4 , q1p3 75
~q tells us how much to rotate. [q p]3 q0p3 + q3p0 + q1p2 , q2p1
57 58
Quaternion Summary . . .
Quaternion Summary . . .
Rotation Correspondence. Let
Rotation Correspondence. The unit quaternions q and q = (cos ; n^ sin ) ;
,q correspond to a single 3D rotation R3: 2 2
with n
^ a unit 3-vector, n
^ n
^ = 1. R(; n^) is usual
Then
2
q 2 q , q , q 2q1q2 , 2q0q3 2q1q3 + 2q q
2 2 2
3 ? to n^.
3D rotation by in the plane
66 0 + 1 2 3 0 2 7
qq q q q02 , q12 + q22 , q32 2q2q3 , 2q q
64 2 1 2 + 2 0 3 7
0 1 7
5 Inversion. Any 3 3 matrix R can be inverted for q up
2q1q3 , 2q0 q2 2q2 q3 + 2q0q1 q02 , q12 , q22 + q32 to a sign. Carefully treat singularities! Can choose sign,
e.g., by local consistency, to get continuous frames.
59 60
SUMMARY
Complex numbers represent 2D frames.

Complex multiplication represents 2D rotation.
Quaternions represent 3D frames.
Quaternion multiplication represents 3D rotation.
61
Part II: OUTLINE
The Spherical Projection Trick: Visualizing unit
Part II: Visualizing vectors.
Quaternion Geometry Quaternion Frames
Andrew J. Hanson
Quaternion Curves
Indiana University Quaternion Splines
1 2
The Geometry of Quaternions
Recall (a; b) with a2 + b2 = 1 is a unit-length

complex number The Geometry of Quaternions . . .
or a point on the unit circle S 1 .
(a’a-b’b, a’b+ab’) (a’, b’)
Similarly, q = (q0; ~q) with q02 + ~q2 = 1 is a
θ + θ’
(a,b) unit-length quaternion or a point on the unit 3-
2 θ’/2
θ /2 sphere S 3.
(1,0)
3 4
Visualizing a Quaternion??
The Geometry of Quaternions . . .
Learn how to Visualize a quaternion by starting
Rotations combine by taking the quaternion prod- with a visualization of a point on S 1, the circle:
uct of the geometric values of 4D points on S 3: r
q0 = 1 (q1)2,
q’ * q q (unseen)
0
(q0, q)
(1,0)
−θ
(q0’, q’) cos −θ 2
2 q (seen)
n sin−θ
^
2
5 6
Visualizing a Quaternion?? . . .
Next, visualize a point on S 2, the ordinary sphere

using only the projection ~q:
r
q0 = 1 , (q1)2 (q2)2 ,
Demo: 2D Unit Vector Projection
q 00
11
11
00
0 q
01
10
q 01
7 8
Visualizing a Quaternion?? . . .
Finally, visualize a point on S 3, the quaternion space:

DISPLAY only ~q, but INFER
r
q0 = 1 , ,
(q1)2 (q2)2 (q3)2 ,
Demo: 3D Unit Vector Projection
11
00 0110
q
0110
11
00
00
11
9 10
Visualize Quaternion Rotations
n sin 2 ) is
Each 4D quaternion point q = (cos 2 ; ^

a frame — a 3 3 rotation matrix generated by
Demo: 4D Unit Vector Projection applying R(; ^
n) to the identity frame.
Identity Matrix is the quaternion q = (1; 0; 0; 0).
Visualize q using only the VECTOR part ~q, so

Identity is the zero vector.
11 12
Visualize Quaternion Rotations . . .
Quaternion Interpretation
The quaternion rotation by about ^

n:
Rotation axis is Fundamental 3-vector: We know
q = (q0; q) = (cos(=2); ^
n sin(=2))
q0 = cos(=2); ~q = n^ sin(=2)
represents the matrix R(; ^
n). We also know that any coordinate frame M can be written
as M = R(; n
^).
Action of rotating Identity by about ^

n: Therefore
q (1; 0; 0; 0) gives Vector part: ~q points exactly along the axis we have to rotate
around to go from identity I to M , and the length of
~0 ) ^
n sin(=2) ~q tells us how much to rotate.
13 14
Displaying spherical points

Displaying a point on a sphere is ambiguous:
N
Demo: Growing Quaternion = Rotation about Axis

q
^
q = required Euler theorem axis.
-h
S
The same horizontal projection is shared by the
North vector (h;~q) and the South vector ( h;~q). ,
15 16
Orbit is Front/Back Line

Projective spherical points
2D projection of orbit around sphere has two half
circles that line up in 2D, match at two edge points Polar Projection removes ambiguity, destroys sym-
only in 3D. metry:
,
Projection Center
the North vector (h;~q) and the South vector ( h;~q)

project to different points on the line.
17 18
Displaying S 3
Displaying S 3 . . .
A quaternion point can be displayed in

11
00 01 01
Parallel Projection: so q = (h;~q) lines up with 01
10 01
01
01
0011
0011
q = (,h;~q),
11
00 11
00 10
Polar projection: so only the “north pole” projects (a) (b) (c)
within the unit sphere, and “south pole” is at 1

(a) Usual vector quaternion point. (b) Orbits through
northern and southern hemispheres. (c) Polar pro-
of R3. jection: north pole at origin, south pole at infinity.
19 20
Quaternion Curves
Represent Families of Frames
Each orientation is a 4D point on the 3-sphere rep-

resenting a quaternion.
Demo: Simple Quaternion Curves
Thus families of frames, which are really rotation

matrices, become curves on the 3-sphere.
) treat these curves just like any other curve. . .

21 22
Families of Quaternion Frames, . . . Families of Quaternion Frames, . . .
Example: torus knot and its (twice around) quater- Example: Surface describes two-degree-of-freedom
nion Frenet frame: joint
-1 qy
-0.5
1 0
1 0.5
1
0.5
5 1
0.5
0
0.5
qz 0
-0.5
1
0.5
0
-1 -0.5
0.5 qz
-1 0
-0.5 -1 -0.5
0 -0.5
0.5
-1 -0.5
-1
1 -0.5 -1
0
qx -0.5
0
qy 0.5 0
0.5 qx
0.5
see: Hanson and Ma, “Quaternion Frame Approach to Streamline Visual- 11 1
ization,” IEEE Trans. on Visualiz. and Comp. Graphics, 1, No. 2, pp. 164– Remark: Group theory demands that x and y ro-
174 (June, 1995). tations combine to give some z rotation!
23 24
Interpolating on Spheres
Quaternion Interpolations N -dimensional spherical interpolation employs the “SLERP,”
a constant angular velocity transition between two directions,
Shoemake (Siggraph ’85) proposed using quaternions in- n
^1 and n
^2 :
stead of Euler angles to get smooth frame interpolations
without Gimbal Lock:
^12(t) = Slerp(^
n n1 ; n
^2; t)
BEST CHOICE: Animate objects and cameras using ro- = n
^1
,
sin((1 t) )
+n
^2
sin(t )
sin( ) sin( )
tations represented on S by quaternions
3
where cos = n
^1 n
^2.
25 26
Quaternion Interpolations
Many variations have been proposed since then; simplest is

simply to apply the formula iteratively to give analog of the de
Remark on Spherical Interpolation
Casteljau spline construction:
(The SLERP formula is simply the result of applying a Gram-
final point
Schmidt decomposition to the two vectors while enforcing 234
1234
unit norm in any dimension.)
23 34
123 2 3
12
1
4
27 28
Spline Families Spline Families . . .
Schlag (in Graphics Gems II (1991)) gives recursive form for

For Euclidean space, the interpolator is
several cubic splines:
L(a; b; t) = a(1 , t) + bt
S (x1; x2; x3; x4; t) =
L(L(L(x1; x2; f12(t)); L(x2; x3; f23(t)); f123(t)); while for Spherical space, the interpolator is
L(L(x2; x3; f23(t)); L(x3; x4; f34(t)); f234(t)); L(a; b; t) = a

,
sin((1 t) )
+b
sin(t )
sin sin
f (t))
where a b = cos .
29 30
Spline Families . . . Spline Families . . .
Catmull-Rom Bezier
f12 = t + 1 f23 = t f34 = t , 1 f12 = t f23 = t f34 = t

f123 = (t+1)
2 f t
234 = 2 f123 = t f234 = t
f = t f =t
31 32
Plane Interpolations
Spline Families . . . In Euclidean space, these three basic cubic splines look like
this:
Uniform B-spline 2 2 2
1.5 1.5 1.5
1 1 1
0.5 0.5 0.5
-1 -0.5 0.5 1 -1 -0.5 0.5 1 -1 -0.5 0.5 1
f12 = (t+2)
f23 = (t+1)
f t
-0.5 -0.5 -0.5
3 3 34 = 3 -1 -1 -1
f123 = (t+1)
2 f t
234 = 2
Bezier Catmull-Rom Uniform B
f = t The differences are in the derivatives: Bezier has to start

matching all over at every fourth point; Catmull-Rom matches
the first derivative; and B-spline is the cadillac, matching all
derivatives but no control points.
33 34
Spherical Interpolations Quaternion Interpolations
1 1 1 1
0.5 0.5 0.5 0.5
0 0 0 0
-0.5 -0.5 -0.5 -0.5
-1 -1 1 -1 -1 1
1 1 1 1
0.5 0.5
0.5 0.5 0.5 0.5
0 0 0
0 0 0
-0.5 -0.5 -0.5 -0.5

1 1
-0.5 -0.5
-1 -1 0.5 -1 -1 0.5
-1 -1 0 -1 -1 0
-0.5 -0.5 -1 -0.5 -0.5 -1
0 0 -0.5 0 0 -0.5
-1 -1
-0.5 -0.5
0.5 0.5 0 0.5 0.5 0
0.5 -1 0.5 -1
1 1 1 1 1 1
Bezier Catmull-Rom Uniform B Bezier Catmull-Rom Uniform B
35 36
Optional:
Exponential Quaternion Map
Exponential map (Kim, Kim, and Shin; see also Grassia)

has these advantages:
Derivative Computation Simpler. Initial velocity condi-

Demo: Interpolated Keyframe Quaternion Curves tions work.
Derivative Matching Simpler. Easier to match neigh-
boring derivatives.
No Renormalization Req’d.
4 vars, 3 DOF ) 3 independent vars, constraints auto-
matic.
37 38
Exponential Quaternion Map . . .

METHOD: use only three “vector” degrees of freedom ~
n=
^ =2 : exponentiating gives four-component quaternion obey-
n
Then, using logarithms of quaternion control points, e.g.,
ing unit 3-sphere constraint: !
~n = log q = q ,1(q0) q
q(~n) = e ^ =2)
(I n 2 jqj = cos jqj
= (cos(=2); n
^ sin(=2)) where ~
n is not normalized, we can use splines in the reduced
Euclidean 3D space.
Note (Grassia): the quaternion frame evolution equations
can be then be written directly in terms of the 3-vector ~
n!
39 40

The trick:
Wherever you would see x0 + t(x1 , x0) in a Euclidean Then, with !i = log((qi,1),1 qi), let
spline, replace (x1 , x0) by q(t) = q0
Y
k
exp(!ii(t))
i=1
! = log((q0),1 q1)
) Euclidean spline i(t) methods can be used, and deriva-
(This is easy: quaternion multiplication gives a new Q = tive computations are simplified.
(Q0; Q) , and ! = log Q is computed as usual.)
41 42
Exponential Quaternion Map . . . Even More Quaternion Interpolations . . .
Example of a many-control-point Catmull-Rom spline done
using the exponential method: A number of other approaches can be found in the bibliogra-
-1
-0.5 phy. Other literature includes:
0
0.5
1
1
Barr et al. Global optimization emulating vanishing 4th
0.5 derivative of Euclidean cubic splines, but on quaternion
0
three-sphere.
-0.5
-1
1
0.5
0
Jüttler et al.: Generalized rational splines, including quater-
-0.5
-1 nion component as well as a spatial component.
43 44
SUMMARY
The Spherical Projection Trick: Visualizing

unit vectors.
Quaternion Frames: ^n in quaternion tells how

to make frame.
Quaternion Curves: are like any other curve;

spline applications.
45
Visualizing Quaternions Part III: OUTLINE
Part III: Quaternion Frames Quaternion Curves: generalize the Frenet Frame
Andrew J. Hanson
Quaternion Frame Evolution
Indiana University Quaternion Curve and Surface Optimization
1 2
Motivating Problem: Framing Curves

What are Frames used For?
Move objects and object parts in an animated scene. KA
Move the camera generating the rendered viewpoint of

A
A
A
A
the scene.
Attach tubes and textures to thickened lines, oriented tex- The (3,5) torus knot.
tures to surfaces. Line drawing useless.

Compare shapes of similar curves. Tubing based on parallel transport, not periodic.
Closeup of the non-periodic mismatch.
3 4
Motivating Problems: Surfaces

Motivating Problems: Curves
A
K
A
A
A
A
A
A
A
A
A
A A smooth 3D surface patch: two ways to get bottom frame.
Closeup of the non-periodic mismatch.
Can’t apply texture. No unique orthonormal frame is derivable from the parame-
terization.
5 6
3D Curves: Frenet and PT Frames 3D curve frames, contd
Now give more details of 3D frames: Classic Moving Frame: Frenet frame is locally defined, e.g., by
0 00
2 0 3 2
T (t)
66 0 77 66
0 1
32
k (t) k (t) (t)
7766
3
7 2 T B(t) = kxx0((tt)) xx00((tt))k
N ,1
4 (t)5 = 4 k (t) 0 (t) 54 (t)75 : N
B 0(t) ,2
k (t) (t), 0 (t) B but has problems on the “roof.”
T
Serret-Frenet frame: k2 = 0, k1 = (t) is the curvature,

B N
???
B
and (t) = (t) is the classical torsion. LOCAL. T B
N T
N
B
N
Parallel Transport frame (Bishop): = 0 to get minimal turn-
T
ing. NON-LOCAL = an INTEGRAL.
7 8
3D curve frames, contd 3D curve frames, contd
Bishop’s Parallel Transport frame is integrated over whole

curve, non-local, but no problems on “roof:”
N1
N1 N2
T
N1
N2 N2
T T
N1 N2
T
N1
Geodesic Reference Frame is the frame found by tilting North
T
Pole of “canonical frame” along a great circle until it points in
desired direction (tangent for curves, normal for surfaces).
N2
9 10
Sample Curve Tubings and their Frames 3D Frames to Quaternion Frames
Unit four-vector. Take q = (q0; q1; q2; q3) = (q0; q) to

obey constraint q q = 1.
Multiplication rule. Let q p be the quaternion product

of two quaternions q and p, where
2 3 2 3
6
[q p] 7 0
6 q p q p q0 0, 1 1, 2p2 , q3p3 77
6
6
[q p] 77
1
6
6 q p +q p +q 01 1 0 2p3 , q3p2 77
Tubings based on Frenet, Geodesic Reference, and Parallel
6
4
6 [q p] 77=6
5 2
6q p + q p + q
4 02 2 0 3p1 , q1p3 75
Transport frames.
[q p] 3 q p +q p +q 03 3 0 1p2 , q2p1
Easily see PT has least “Twist,” but lacks periodicity.
11 12
Quaternion Frame Evolution
. . . to Quaternion Frames . . .
3
Just as in 2D, let columns of R (q ) be a frame: ( ; T N; B);
Quaternion Correspondence. The unit quaternions q this system has nine components.
,
and q correspond to a single 3D rotation R (q ): 3
2 2 3
q + q 2 , q22 , q32
64 0 q q 1
2 2
q1 q2 , q0 q3 2 q1 q3 +2
q0 q2
2 +2
1 2 q q
0 3 q 0 1 +
2 , q2 q2 , q2
2 3 2q q
2 3 , 22
q q
7
0 1 5
Derivatives of the i-th column Ri in quaternion coordinates
2q1q3 , 2q0q2 2q2q3 + 2q0q1 2
q0 , q1 , q2 + q32
2
have the form:
Rotation Correspondence.

R_ i = 2Wi [q_(t)] where i = 1; 2; 3 and, e.g.,
2 3
2n n n
Let q = (cos ; ^ sin ), with ^ a unit 3-vector, ^ ^ =
2 nn 6
q 0 1 ,2 ,3
q q q
7
1 . Then R(; ^) is usual 3D rotation by in the plane 1
W = 64 q 3 2 1 0
q q q 75
perpendicular to ^. n ,2 3 ,0 1
q q q q
(rows form mutually orthonormal basis).

13 14
Quaternion Frame Evolution . . . Quaternion Frames . . .
When we simplify by eliminating Wi . . . Properties of Tait’s quaternion frame equations:

we find the square root of the 3D frame eqns!
Antisymmetry ) q (t) q_(t) = 0 as required to keep
Tait (1890) derived the quaternion equation that makes all 9 constant unit radius on 3-sphere.
3D frame equations reduce to: q_ = (1=2)q (0; k) or: Nine equations and six constraints become four equa-
tions and one constraint , keeping quaternion on the 3-
2 3 2 3 2 3
0
66 q_ 77 6 0k2 ,k1 , 0 7 66 q 77 sphere. ) Good for computer implementation.
1 ,2 0 ,1 1
66 q_ 77 = 1 666 k k 77 66 q 77
2
64 q_ 75 1 , 0 , 2
2 64 k 2
7
k 75 64 q 75 MATHEMATICA code implementing this differential equa-
3 q_ k1 k2 3 0 q tion is provided.
15 16
Example of a Quaternion Frame Curve

Left Curve = torus knot tubed with Frenet frame; Right Curve
is projection from 4D of (twice around) quaternion Frenet
Quaternion Frames . . . frames:
Analogous treatment (given in Hanson Tech Note in Course 0.5
Notes) applies also to the Weingarten equations, allowing -0.5

1
a direct quaternion treatment of the classical differential -1

-1 0
0.5
-0.5
0 -0.5
geometry of surfaces as well. 0.5

1
-1
see Notes: Hanson and Ma, “Quaternion Frame Approach to Streamline

Visualization,” IEEE Trans. on Visualiz. and Comp. Graphics, 1, No. 2,
pp. 164–174 (June, 1995).
17 18
Minimizing Quaternion Length Solves Periodic
Tube Minimizing Quaternion Length Works
Quaternion space optimization of the non-periodic parallel Result of Quaternion space optimization of the (3,5) torus
transport frame of the (3,5) torus knot. knot frame.
0 0.5 1 0 0.5 1
-1 -0.5 -1 -0.5
1 1
0.5
5 0.5
5
0 0
-0.5
.5 -0.5
.5
-1 -1 KA
A
1
0.5
1
0.5 A
0
-0.5
0
-0.5
A
-1 -1 A
A
A
A
see Notes: “Constrained Optimal Framings of Curves and Surfaces using A
A
Quaternion Gauss Maps,” Proceedings of Visualization ’98 , pp. 375–382 A
(1998).
19 20
Can also Optimize Quaternion Frames on 3D Frames for Patch

0.4
Patch: 0.4 0.5
-0.2
0
0.2 0.4
0.25 0.4
0.2 0.2 0 0.2
-0.25
.25
1
0 0 0
0.2 0.2 0.2

0.5
0.1 0.1 0.1
0 0 0
0
-0.1 -0.1 -0.1
-0.2 -0.2 -0.2

-0.5
0 0 0
0.2 0.2 0.2
0.4 0.4 0.4
(a) (b) (c) (d)

(a) (b) (c) (d)
Quaternion frames for (a) Geodesic Ref. (b) One edge Par-
allel Transport. (c) Random. (d) Minimal area result.
Quaternion frames for (a) Geodesic Ref. (b) One edge Parallel Transport.
(c) Random. (d) Minimal area result.
21 22
SUMMARY
Quaternions can represent frames.

Curve frames ) quaternion curves.
Surface patch frames ) quaternion surface patches.
Minimizing quaternion length or area finds parallel
transport “minimal turning” set of frames.
23
Part IV: OUTLINE
Visualizing Quaternions Clifford Algebra: Introduction to generalizing

complex numbers and quaternions.
Part IV: Clifford Algebras
Reflections vs Rotations: How to make a ro-
Andrew J. Hanson tation.
Indiana University Pin(N), Spin(N), O(N), and SO(N): Double cov-
erings of N-dimensional rotations.
1 2
Motivation Foundations
N -dimensional space is described formally by
Quaternions are too special. Complex num- basis vectors e , i = 1; : : : ; N and real num-
i
bers and quaternions run out of steam after di- bers v , so a vector looks like
i
mensions 2,3,4.
Search for some generalizable idea. How does V = Xv e i i
the (a2 , b2), etc., form generalize to N -dim ro-

i
tation matrices?
The length is computed from the inner product:
Clifford Algebra: Clifford found the generaliza- kV k2 =< V; V >= X v g v i ij j
ij
tion, but the really interesting relation to spin 1/2
elementary particles came much later. where gij is the identity matrix in Euclidean space.
3 4
Foundations . . . Clifford Algebra . . .
But the basis vectors obey a strange multiplica- How does this odd product concern us??
tion rule:
eei j + ee
j i = ,2gij
1. Because it contains in any dimension a way of
expressing rotations as multiple reflections about
This is the CLIFFORD ALGEBRA. a plane.
(Note: physicists would recognize these formu- 2. Because these expressions of rotations are nat-
las as those obeyed by the Pauli matrices or the ural square roots of the familiar N N orthog-
Dirac matrices.) onal matrix approach to writing rotations.
5 6
Clifford Algebra implements reflections Clifford Algebra – sample calculation
If A =
P
ae
i i is any vector with kAk = 1, then, With A =
P
ae
i i and kAk = 1, compute V 0:
using the Clifford multiplication rule,
V 0 = A V A = V , 2A < A; V > AV A =

X
av a ee e
i j k i j k
i;j;k

=
X
a v a ,e e , 2g e
i j k j i ij k
i;j;k

This is just a reflection of the component of V lying =
X
a v a ,e fe e g , 2g e
i j k j i k ij k
in the direction of A about the plane i;j;k
= ,V a a 21(e e
X
+ e e ) , 2A < A; V >
< A; X >= 0 i;j
i k i k k i
7 8
Clifford Algebra – sample calculation, contd Clifford Algebra reflections
Use symmetrization trick — only one-half of the

symmetrized sum multiplying aiak can survive;
P V 0 = A V A = V , 2A < A; V >
that in turn is just the Clifford rule right hand side!
A V A = V X aiak21( 2gik) 2A < A; V

, , , > ^
V ^ ^ .V)
A(A
,
i;j
= +V < A; A > 2A < A; V >

A
reflection
QED. plane
,
Note: The MINUS sign in 2gij in the original def-
V’(reflected) ^ ^
= V - 2 A(A .V)
inition makes the sign of V positive.
9 10
Clifford Algebra rotations
Let B =
P
be i i be another vector with kB k = 1: Clifford Algebra rotations
Repeating the Clifford multiplication rule, This can be shown (e.g, in Mathematica) to be a
proper rotation of the vector V , that is
V 00
BV0B =
= BAV AB V 00 = B A V A B = X R v e ij j i
= V 0 , 2B < B; V 0 >
ij
= V , 2A < A; V > where Rij is an orthonormal matrix of unit deter-

,2B < B; V > +4B < B; A >< A; V > minant.
11 12
Clifford Algebra rotation graph
Clifford Algebra rotations
V"(rotated)
V
Graphically, we have this:
^
A 01
10
V 00 BAV AB
^
= B 11
00
^ ^.
X
R ve ^ reflection B (B V’)
= ij j i
A
plane
ij V’(reflected)
^ reflection
B plane
13 14
N pairs params constraints freedom

2 1
2 2 3 1
3 1
2 3 3 3
Clifford Algebra rotations 4 2
4 4 10 6
5 2
4 5 10 10
NOTE that in higher dimensions, you may need
more than a single (A; B ) pair to exhaust all pos-
6 3
6 6 21 15
sible rotations: 7 3
6 7 21 21
. . . . .
N p = b 2 c 2p N p(2p + 1)
N (
N N ,1)
2
15 16
Examples of Clifford Algebras . . .
Examples of Clifford Algebras

N = 2: The basis (1; e1; e2; e1e2) exhausts
all possible Clifford products.
N ;e
= 1: The basis (1 1) with ( 1)2 = 1e , ,
Since e1e2e1e2 = 1, we can identify this ba-
is just the complex numbers. But be careful: sis with the quaternions (1; i; j; k)! But be care-
there is only one dimension, so the only possi- ful: there are only two dimensions, so this is not
ble reflection is x ! ,x. This is not enough to enough to do 3D rotations!
do 2D rotations! Where is 2D Rot? True basis of rotations is the
even part of the family of all Clifford products,
or (1; e1e2) !!
17 18
2D Rotations done right 2D Rotations done right . . .
What is i? What we called i = p,1 is really So the half-angle formula is mandatory! Our
i = e1e2 2D transformation was not so silly after all; noth-
How do we rotate in 2D? Let ing else generalizes to N -dimensions. The Clifford
R = a + be1e2, Ry = a , be1e2: algebra for N = 2 automatically produces:
R V Ry = V 0 = v10 e1 + v20 e2 2
a2 , b2 ,2ab 75 ;
3
R2(a; b) = 64
2ab a2 , b2
where V 0 now means a rotation, and
2
0 3 2 2 2 ,2ab 37 26 v1 37
64 v1 75 = 64 a , b where a2 + b2 = 1, and we have the solution
v20 2ab a2 , b2
54
v2 5 a = cos(=2), b = sin(=2).
19 20
3D Rotations done right 3D Rotations done right . . .
3D is of course a little trickier: here the full basis

of all Clifford products is 8-dimensional:
General 3D Rotation: with
; e ; e2; e3; e2e3; e3e1; e1e2; e1e2e3)
(1 1 R = q0 + q1e2e3 + q2e3e1 + q3e1e2
Even part is rotations. To exclude reflections, we have
R V Ry = v0e
we keep only the even part: 4
X
; e e ; e3e1; e1e2)
(1 2 3 i =1
i i
These are the quaternions: identify these with where the coefficients of v i in v0 are precisely
i
(1; ; ; ).
i j k our old quaternion formula.
21 22
Higher dimensions:
Higher dimensions:
But we can do a little counting to see what is going
on in N dimensions, where we know that the num-
As one might expect, higher dimensions are much ber of rotational degrees of freedom is
more complicated, and do not work out so neatly,
except for a convenient accident in N = 4, which N (N , 1)=2
allows a “double-quaternion” form (see “Rotations
for N-dimensional Graphics”).
Degrees of freedom in higher dimensional Spin rep-
resentations:
23 24
N Dim(even Dim(Rotations) Constraints (the Pin(N), Spin(N), O(N), SO(N) and all that . . .
Clifford) difference) Spin representations of orthogonal groups follow from the
1 1 0 1 Clifford Algebra Cl(N): (So do spinors — but another time...)
2 2 1 1
3 4 3 1
Pin(N). G is “Pin” if it’s a general reflection; G includes
4 8 6 2
all elements of Cl(N).
5 16 10 6
6 32 15 17 Spin(N). G is “Spin” if it’s a general rotation; G is only
7 64 21 43 even elements of Cl(N).
8 128 28 100 O(N).

G V G
y is “O” if G is in Pin and result is a vector
. . . . reflection.
2N ,
2N ,1 N (N , 1)=2 2+N
N N
2 SO(N). G V G
y is “SO” if G is in Spin and result is a
25 vector rotation. 26
GRAND CONCLUSION
Rotation Matrices: can be represented by a “square
root” object with simpler geometric properties than rota-
tions (= quaternions for N = 2; 3; 4).
Visualization of Quaternions: is possible using sphere
projection trick and a solid unit sphere.
Quaternion Curves, Surfaces, Volumes: embedded in
that sphere represent animations, flows, curve tubings,
joint ranges, etc.
Clifford Algebras: form the rigorous basis for the whole
set of concepts.
27
A
A II.6
Geometry for N-Dimensional
Graphics
Andrew J. Hanson
Indiana University
Bloomington, IN 47405
hanson@cs.indiana.edu
} Introduction }
Textbook graphics treatments commonly use special notations for the geometry of 2 and 3
dimensions that are not obviously generalizable to higher dimensions. Here we collect a family
of geometric formulas frequently used in graphics that are easily extendible to N dimensions
as well as being helpful alternatives to standard 2D and 3D notations.
What use are such formulas? In mathematical visualization, which commonly must deal
with higher dimensions — 4 real dimensions, 2 complex dimensions, etc. — the utility is self-
evident (see, e.g., (Banchoff 1990, Francis 1987, Hanson and Heng 1992b, Phillips et al. 1993)).
The visualization of statistical data also frequently utilizes techniques of N -dimensional display
(see, e.g., (Noll 1967, Feiner and Beshers 1990a, Feiner and Beshers 1990b, Brun et al. 1989,
Hanson and Heng 1992a)). We hope that publicizing some of the basic techniques will encour-
age further exploitation of N -dimensional graphics in scientific visualization problems.
We classify the formulas we present into the following categories: basic notation and the
N -simplex; rotation formulas; imaging in N -dimensions; N -dimensional hyperplanes and vol-
umes; N -dimensional cross-products and normals; clipping formulas; the point-hyperplane dis-
tance; barycentric coordinates and parametric hyperplanes; N -dimensional ray-tracing meth-
ods. An appendix collects a set of obscure Levi-Civita symbol techniques for computing with
determinants. For additional details and insights, we refer the reader to classic sources such as
(Sommerville 1958, Coxeter 1991, Hocking and Young 1961) and (Banchoff and Werner 1983,
Efimov and Rozendorn 1975).
} Definitions — What is a Simplex, Anyway? }

In a nutshell, an N -simplex is a set of (N + 1) points that together specify the simplest non-
vanishing N -dimensional volume element (e.g., two points delimit a line segment in 1D, 3
points a triangle in 2D, 4 points a tetrahedron in 3D, etc.). From a mathematical point of view,

Copyright c 1994 by Academic Press, Inc.
All rights of reproduction in any form reserved.
149 ISBN 0-12-336156-7
150 }
1 2 3 4
N N
Figure 1. 2D projections of simplexes with dimension 1–4. An -simplex is defined by ( + 1) linearly
independent points and generalizes the concept of a line segment or a triangular surface patch.
there are lots of different N -dimensional spaces: here we will restrict ourselves to ordinary flat,
real Euclidean spaces of N dimensions with global orthogonal coordinates that we can write as
~x x; y; z; : : : ; w
= ( )
or more pedantically as
~x x ;x ;x ;:::;x
= (
(1) (2) (3) (N )
) :
We will use the first, less cumbersome, notation whenever it seems clearer.
Our first type of object in N -dimensions, the 0-dimensional point ~x, may be thought of as
a vector from the origin to the designated set of coordinate values. The next type of object is
the 1-dimensional line, which is determined by giving two points (~x0 ; ~x1 ); the line segment
from ~x0 to ~x1 is called a 1-simplex. If we now take three noncollinear points (~x0 ; ~x1 ; ~x2 ), these
uniquely specify a plane; the triangular area delineated by these points is a 2-simplex. A 3-
simplex is a solid tetrahedron formed by a set of four noncoplanar points, and so on. In figure
1, we show schematic diagrams of the first few simplexes projected to 2D.
Starting with the (N + 1) points (~x0 ; ~x1 ; ~x2 ; : : : ; ~xN ) defining a simplex, one then connects
all possible pairs of points to form edges, all possible triples to form faces, and so on, resulting
in the structure of component “parts” given in table 1. The next higher object uses its predeces-
sor as a building block: a triangular face is built from three edges, a tetrahedron is built from
four triangular faces, a 4-simplex is built from 5 tetrahedra.
The general idea should now be clear: (N + 1) linearly independent points define a hy-
perplane of dimension N and specify the boundaries of an N -dimensional coordinate patch
comprising an N -simplex (Hocking and Young 1961). Just as the surfaces modeling a 3D ob-
ject may be broken up (or tessellated) into triangular patches, N -dimensional objects may be
tessellated into (N , 1)-dimensional simplexes that define their geometry.
II.6 Geometry for N-Dimensional Graphics } 151
N
Table 1. Numbers of component structures making up an -simplex. For example, in 2D, the basic
simplex is the triangle with 3 points, 3 edges, and one 2D face.
Dimension of Space
Type of Simplex N =1 N =2 N =3 N = 4 ...
N +1

N

Points (0D) 2 3 4 5 ... = N +1
1
N +1

Edges (1D simplex) 1 3 6 10 ...
2
N +1

Faces (2D simplex) 0 1 4 10 ...
3
N +1
Volumes (3D simplex) 0 1 5 ...
4
.. .. .. .. .. ..

..
. . . . . . .
, 2)D simplex N +1
,1

(N ...
N
, 1)D simplex N +1

(N ... = N +1
N
N +1
N D simplex 1 ... =1
N +1
} }
Rotations
N
In N Euclidean dimensions, there are N N , = degrees of rotational freedom
2
= ( 1) 2
corresponding to the free parameters of the group SO N . In 2D, that means we only have one
( )
rotational degree of freedom given by the angle used to mix the x and y coordinates. In 3D,
there are 3 parameters, which can be thought of as corresponding either to three Euler angles
or to the three independent quaternion coordinates that remain when we represent rotations in
terms of unit quaternions. In 4D, there are 6 degrees of freedom, and the familiar 3D picture of
“rotating about an axis” is no longer valid; each rotation leaves an entire plane fixed, not just
one axis.
General rotations in N dimensions may be viewed as a sequence of elementary rotations.
Each elementary rotation acts in the plane of a particular pair, say (i; j ), of coordinates, leaving
an (N , 2)-dimensional subspace unchanged; we may write any such rotation in the form
x0 (i)
= x
(i)
x cos (j )
sin
x0 (j )
= x x
(i)
sin +
(j )
cos
x0 (k )
= x k 6 i; j :
(k )
( = )
It is important to remember that order matters when doing a sequence of nested rotations; for
example, two sequences of small 3D rotations, one consisting of a (2; 3)-plane rotation followed
152 }
Image ^ (1)
x
^
u
(1)
Camera
^ (N)
x Origin
f (N)
d (N)
Figure 2. Schematic view of the projection process for an N -dimensional pinhole camera.
by a (3; 1)-plane rotation, and the other with the order reversed, will differ by a rotation in the
(1; 2)-plane. (See any standard reference such as (Edmonds 1957).)
We then have a number of options for controlling rotations in N -dimensional Euclidean

space. Among these are the following:
i; j -space pairs. A brute-force choice would be just to pick a sequence of i; j
( ) ( ) planes
in which to rotate using a series of matrix multiplications.

(i; j; k )-space triples. A more interesting choice for an interactive system is to provide
the user with a family of (i; j; k ) triples having a 2D controller like a mouse coupled to
two of the degrees of freedom, and having the 3rd degree of freedom accessible in some
other way — with a different button, from context using the “virtual sphere” algorithm
of (Chen et al. 1988), or implicitly using a context-free method like the “rolling-ball” al-
gorithm (Hanson 1992). The simplest example is (1; 2; 3) in 3D, with the mouse coupled
to rotations about the x ^ -axis (2; 3) and the y
^ -axis (3; 1), giving ^
z-axis (1; 2) rotations as
a side-effect. In 4D, one would have four copies of such a controller, (1; 2; 3), (2; 3; 4),
(3; 1; 4), and (1; 2; 4), or two copies exploiting the decomposition of SO (4) infinitesimal
N
rotations into two independent copies of ordinary 3D rotations. In N dimensions,
3
sets of these controllers (far too many when N is large!) could in principle be used.
} N -dimensional Imaging }
The general concept of an “image” is a projection of a point ~x x ; x ; : : : ; x from
= (
(1) (2) (N )
)
dimension N to a point ~u of dimension N , along a line. That is, the image of a 2D world
( 1)
far
far
far
near near near

Figure 3. Qualitative results of perspective projection of a wire-frame square, a cube, and a hypercube
in 2D, 3D, and 4D, respectively.
is a projection to 1D film, 3D worlds project to 2D film, 4D worlds project to 3D film, and so

on. Since we can rotate our coordinate system as we please, we lose no generality if we assume
this projection is along the N -th coordinate axis. An orthographic or parallel projection results
if we simply throw out the N -th coordinate x(N ) of each point. A pinhole camera perspective
projection (see figure 2) results when, in addition, we scale the first (N , 1) coordinates by
dividing by (dN , x(N ) )=fN , where dN is the distance along the positive N -th axis to the
camera focal point and fN is the focal length. One may need to project this first image to
successively lower dimensions to make it displayable on a 2D graphics screen; thus a hierarchy
of up to (N , 2) parameter sets f(fN ; dN ); : : : ; (f3 ; d3 )g may be introduced if desired.
In the familiar 3D case, we replace a vertex (x; y; z ) of an object by the 2D coordinates
(xf=(d , z ); yf=(d , z )), so that more distant objects (in the negative z direction) are shrunk
in the 2D image. In 4D, entire solid objects are shrunk, thus giving rise to the familiar wire-
frame hypercube shown in figure 3 that has the more distant cubic hyperfaces actually lying
inside the projection of the nearest cube.
As we will see a bit later when we discuss normals and cross-products, the usual shading
approaches allow only (N , 1)-manifolds to interact uniquely with a light ray. That is, the
generalization of a viewable “object” to N dimensions is a manifold of dimension (N , 1) that
bounds an N -dimensional volume; only this boundary is visible in the projected image if the
object is opaque. For example, curves in 2D reflect light toward the focal point to form images
on a “film line,” surface patches in 3D form area images on a 2D film plane, volume patches in
4D form volume images in the 3D film volume, etc. The image of this (N , 1)-dimensional
patch may be ray traced or scan converted. Objects are typically represented as tessellations
which consist of a collection of (N , 1)-dimensional simplexes; for example, triangular surface
patches form models of the visible parts of 3D objects, while tetrahedral volumes form models
of the visible parts of 4D objects. (An interesting side issue is how to display meaningful
illuminated images of lower dimensional manifolds — lines in 3D, surfaces and lines in 4D,
etc.; see (Hanson and Heng 1992b) for further discussion.)
154 }
x0
xc
x
^
n
c x1
Origin
Figure 4. The line from ~x 0 to ~x 1 whose points obey the equation n
^ ~x , ~x
( 0 ) = 0. The constant c is
just n
^ ~x
0.
} Hyperplanes and Volume Formulas }

Implicit Equation of a Hyperplane. In 2D, a special role is played by the single linear
equation defining a line; in 3D, the analogous single linear equation defines a plane. In N -
dimensions, the following implicit linear equation describes a set of points belonging to an
(N , 1)-dimensional hyperplane:
n^ (~x , ~x0 ) = 0 : (1)

Here ~x0 is any point on the hyperplane and conventionally n ^ n
^ = 1. The geometric interpre-
tation of this equation in 2D is the 1D line shown in figure 4. In general, n ^ is a normalized unit
vector that is perpendicular to the hyperplane, and n ^ ~

x0 = c is simply the (signed) distance
from the origin to the hyperplane. The point ~xc = cn ^ is the point on the hyperplane closest to
the origin; the point closest to some other point P~ is ~xc = P~ + n^ fn x0 , P~ )g.
^ (~
Simplex Volumes and Subvolumes. The volume (by which we always mean the N -
dimensional hypervolume) of an N -simplex is determined in a natural way by a determinant of
its (N + 1) defining points (Sommerville 1958):
2 3
x x x x
66 y y y y 77
1 2 N 0
V
66 .. .. . . .. .. 77 :
1 2 N 0
66 . . . . . 77
1
N =
N !
det
4 w w w w 5
(2)
1 2 0

N
1 1 1 1
The bottom row of 1’s in eq. (2) corresponds to the familiar homogeneous coordinate used with
4 4 projection matrices in 3D graphics. We will attempt to convince the reader in a mo-
ment that disastrous sign inconsistencies result unless the global origin ~x0 of the N -simplex’s
coordinate system is in the last column as shown.
The expression for the volume in eq. (2) is signed, which means that it implicitly defines the
N -dimensional generalization of the Right-Hand Rule typically adopted to determine triangle
orientation in 3D geometry. For example, we observe that if ~x0 = (0; 0; : : : ; 0) is the origin
and we choose ~x1 = (1; 0; : : : ; 0), ~x2 = (0; 1; 0; : : : ; 0), and so on, the value of the determinant
is +1. If we had put ~x0 in the first row in eq. (2), the sign would alternate from dimension to
dimension! We will exploit this signed determinant shortly to define N -dimensional normal
vectors, and again later to formulate N -dimensional clipping.
First, we use the standard column-subtraction identity for determinants to reduce the dimen-
sion of the determinant in eq. (2) by one, expressing it in a form that is manifestly translation-
invariant:
2 3
66 xy ,, xy
( 1 0) ( x ,x
2
y ,y
0) x , x
y , y
( N 0) x
y 777
0
V
66 ( 1 0) ( 2 0) ( N 0) 0
77
66
1 .. .. .. ..
..
N
w 75
N = det . . . . .
!
(4 w ,w
1 0) ( w ,w
2 0) w , w ( 0) 0

N
0 0 0 1
2 x ,x x ,x x , x 3
66 y , y y , y 77
( 1 0) ( 2 0) ( 0)
y ,y
N
64 75 :
1 ( 1 0) ( 2 0) ( N 0)
=
N !
det ..
.
..
.
..
.
..
.
(3)
( w ,w1 0) ( w ,w
2 0) w , w ( N 0)
These formulas for VN can be intuitively understood as generalizations of the familiar 3D triple
scalar product,
~x , ~x ~x , ~x ~x , ~x ;
[( 1 0) ( 2 0 )] ( 3 0)
which gives the volume of the parallelepiped with sides ~x , ~x ; ~x , ~x ; ~x , ~x . The (( 1 0) ( 2 0) ( 3 0 ))

corresponding tetrahedron with vertices at the points ~x ; ~x ; ~x ; ~x has one-sixth the volume
( 0 1 2 3)
of the parallelepiped. The analogous observation in N dimensions is that the factor of =N in 1 !
eq. (3) is the proportionality factor between the volume of the N -simplex and the volume of
the parallelepiped whose edges are given by the matrix columns.
Invariance. The volume determinant is invariant under rotations. To see this explicitly, let
jX j be the matrix in eq. (3) and let jRj be any orthonormal rotation matrix (i.e., one whose
columns are of unit length and are mutually perpendicular, with unit determinant); then, letting
jX 0 j = jRj jX j, we find
det jX 0 j = det( jRj jX j ) = det jRj det jX j = det jX j N V ;
= ! N
156 }
since the determinant of a product is the product of the determinants.
A manifestly translation and rotation invariant form for the square of the volume element is
2
(V N )
2
=
N
1
det jX X j
t
!
2v ; v ; v ;N 3
2 66 v ;
(1 1)
v ;
(1 2)
v ; N 77
(1 )
64 .. 75 ;
1 (2 1) (2 2) (2 )
=
N !
det
.
..
.
..
.
..
.
(4)
v N;( 1) v N;( 2) v N; N
( )
where v (i; j ) = (~xi , ~x0 ) (~xj , ~x0 ).

This invariant form is not presented as an idle observation; we now exploit it to show how to
construct volume forms for subspaces of N -dimensional spaces, for which the defining vertices
of the desired simplex cannot form square matrices!
The trick here is to note that while VK , for K < N , is not expressible in terms of a square
matrix of coordinate differences the way VN is, we may write VK as the determinant of a square
matrix in one particular coordinate frame, and multiply this matrix by its transpose to get a form
like eq. 4, which does not depend on the frame. Since the form is invariant, we can transform
back to an arbitrary frame to find the following expression for VK in terms of its K basis vectors
xk , ~x0) of dimension N :
(~
, ~x 3 2 ~x
, ~x 77 2 66 ~x
1 0

V ( K)
K
2
= ..
1
7
5 !
~
x , ~
x
det ~
x , ~
x64
2
~
x , ~
x 0
1 0 2 0 K 0
.
, ~x ~x
; v ; v ;K 3 2v K 0
; v ; v ; K 77
2 66 v
(1 1) (1 2) (1 )
64
75 :
1 (2 1) (2 2) (2 )
K = ..
. !
..
.
det
..
.
..
.
(5)
v K; v K; v K; K ( 1) ( 2) ( )
That is, to compute a volume of dimension K in N dimensions, find the K independent basis
vectors spanning the subspace, and form a square K K matrix of dot products related to
V by multiplying the N K matrix of column vectors by its transpose on the left. When
2
K , we see that we have simply the squared Euclidean distance in N dimensions, v ;

K
= 1 (1 1) =
(~x , ~x ~x , ~x .
1 0) ( 1 0)
} Normals and the Cross-Product }

A frequently asked question in N -dimensional geometry concerns how to define a normal vec-
tor as a cross-product of edges for use in geometry and shading calculations. To begin with,
you must have an (N , 1)-manifold (a line in 2D, surface in 3D, volume in 4D) in order to
have a well-defined normal vector; otherwise, you may have a normal space (a plane, a vol-
ume, etc.). Suppose you have an ordered set of (N , 1) edge vectors (~xk , ~x0 ) tangent to this
(N , 1)-manifold at a point ~ x0; typically these vectors are the edges of one of the (N , 1)-
simplexes in the tessellation. Then the normal N ~ at the point is a generalized cross-product
whose components are cofactors of the last column in the following (notationally abusive!)
determinant:
N~ = N x2 N
x^ + y y^ + N z^ + + N w
z ^ w
3
66 xy ( 1 ,x 0) ( x ,x
2 0) x ( ,1 , x0 ) x^
77
,y y ,y y ,1 , y0 ) y^
N
66 z ( 1
,z
0) ( 2
z ,z
0)
z
( N
,1 , z0 ) ^z
77
= det
664
( 1 0) ( 2 0) ( N
77 : (6)
..
.
..
.
..
.
..
.
..
. 5
(w ,w
1 0) ( w ,w
2 0) w ( N ,1 , w0 ) w
^
As usual, we can normalize using kN ~ k, the square root of the sum of the squares of the co-
factors, to form the normalized normal n ^ = N=~ kN~ k. A quick check shows that if the vectors
xk , ~x0) are assigned to the first (N , 1) coordinate axes in order, this normal vector points
(~
in the direction of the positive N -th axis. For example, in 2D, we want the normal to the vector
~ = (,(y1 , y0); (x1 , x0)) so that a vector purely in the x direction
(x1 , x0 ; y1 , y0 ) to be N
has a normal in the positive y direction; placing the column of unit vectors (x ^; y
^; z
^; : : : ; w
^ ) in
the first column fails this test. The 3D case can be done either way because an even number
of columns are crossed! It is tempting to move the column of unit vectors to the first column
instead of the last, but one must resist: the choice given here is the one to use for consistent
behavior across different dimensions!
The qualitative interpretation of eq. (6) can now be summarized as follows:
2D: Given two points (~x0 ; ~x1 ) determining a line in 2D, the cross-product of a single vector
is the normal to the line.
3D: Given three points defining a plane in 3D, the cross-product of the two 3D vectors
outlining the resulting triangle is the familiar formula (~x1 , ~x0 ) (~x2 , ~x0 ) for the normal
N~ to the plane.
4D: In four dimensions, we use four points to construct the three vectors (~x1 , ~x0 ); (~x2 ,
~x0); (~x3 , ~x0 ); the cross product of these vectors is a four-vector that is perpendicular
to each vector and thus is interpretable as the normal to the tetrahedron specified by the
original four points.
From this point on, the relationship to standard graphics computations should be evident:
If, in N -dimensional space, the (N , 1)-manifold to be rendered is tessellated into (N , 1)-
simplexes, use eq. (6) to compute the normal of each simplex for flat shading. For interpolated
shading, compute the normal at each vertex (e.g., by averaging the normals of all neighboring
158 }
simplexes and normalizing or by computing the gradient of an implicit function specifying the
vertex). Compute the intensity at a point for which you know the normal by taking the dot
product of the appropriate illumination vector with the normal (e.g, by plugging it into the last
column of eq. (6)). If appropriate, set the dot product to zero if it is negative (pointing away
from the light). Back face culling, to avoid rendering simplexes pointing away from the camera,
is accomplished in exactly the same way: plug the camera view vector into the last column of
eq. (6) and discard the simplex if the result is negative.
Dot Products of Cross Products. We conclude this section with the remark that some-
times computing the dot product between a normal and a simple vector is not enough; if we
need to know the relative orientation of two face normals (e.g., to determine whether a finer
tessellation is required), we must compute the dot products of normals. In principle, this can
be done by brute force directly from eq. (6). Here we note an alternative formulation that is the
N -dimensional generalization of the 3D formula for the decomposition of the dot product of
~ = A~ B~ and
two cross products; in the 3D case, if one normal is given by the cross product X
~ ~ ~
the other by Y = C D , we can write
X~ Y~ = ( A~ B~ C~ D~
) ( ) = ( A~ C~ B~ D~ , A~ D~ B~ C~ :
)( ) ( )( ) (7)
We note that the degenerate case for the square of a cross product is
A~ B~ A~ B~
( ) ( A~ A~ B~ B~ , A~ B~ ;
) = ( )( ) ( )
2
which, if is the angle between A ~ and B~ , reduces to the identity kA~ k kB~ k kA~ k kB~ k , 2 2
sin
2
=
2 2
kA~ k kB~ k
2 2 2
cos .
The generalization of this expression to N dimensions can be derived from the product of
two Levi-Civita symbols (see the Appendix). If X ~ and Y~ are two cross products formed from
the sets of vectors ~x ; ~x ; : : : ; ~x , and ~y ; ~y ; : : : ; ~y , , then
1 2 N 1 1 2 N 1
X
X~ Y~ = x 1 x 2 : : : x N,,1 y 1 y 2 : : : y N,,1
(i )
1
(i )
2
(i
N 1
) (j ) (j )
1 2 N
(j
1
)
all indices
2 1 2 1 N ,1 3
66 12 11 2 N ,1 77
i j i j i j
2 2
det 64 .. i j i j
.. ..
.
.. 75 ;
i j
(8)
. . .
N ,1 1 N ,1 2
i j i j N ,1 N ,1
i j
where the Kronecker delta, ij , is defined as
= 1 i j =
i6 j:
ij
= 0 =
It is easy to verify that for N = 3 this reduces to eq. (7).

More remarkable, however, is the fact that this formula shows that the square magnitude of
the normal N~ of a hyperplane given in eq. (6) is the subvolume of the corresponding paral-
lelepiped specified by eq. (5). That is, not only does the direction of eq. (6) have an important
geometric meaning with respect to the (N , 1)-simplex specifying the hyperplane, but so does
its magnitude! We find
2 v ; v ; v ;N , 3
66 v ; 77
(1 1) (1 2) (1 1)
v ; v ;N ,
N~ N~ 64 75 N, V , :
(2 1) (2 2) (2 1)
2
= det .. .. .. .. = (( 1)! N 1)
. . . .
vN, ;
( 1 1) vN, ;
( 1 2) v N , ;N ,
( 1 1)
} Clipping Tests in N Dimensions }

Now we can exploit the properties of the volume formula to define clipping (“which side”) tests
in any dimension. If we replace (~xN , ~x0 ) by (~x , ~x0 ), eq. (3) becomes a function VN (~x).
Furthermore, this function has the remarkable property that it is an alternative form for the
hyperplane equation, eq. (1), when VN (~x) = 0.
We can furthermore determine on which side of the (N , 1)-dimensional hyperplane deter-
mined by (~x0 ; ~x1 ; : : : ; ~xN ,1 ) an arbitrary point ~x lies simply by checking the sign of VN (~x).
That is,
V ~x
N( ) the point ~x lies on a hyperplane and solves an equation of the form eq. (1).
) = 0
V ~x > ) the point ~x lies above the hyperplane.

N( ) 0
V ~x < ) the point ~x lies below the hyperplane.

N( ) 0
Note: The special case V is of course just the general criterion for discovering linear
N = 0
dependence among a set of N vector variables. This has the following elegant geometric
( + 1)
interpretation: In 2D, we use the formula to compute the area of the triangle formed by 3
points (~x0 ; ~x1 ; ~x); if the area vanishes, the 3 points lie on a single line. In 3D, if the volume
of the tetrahedron formed by 4 points (~x0 ; ~x1 ; ~x2 ; ~x) vanishes, all 4 points are coplanar, and
so on. Vanishing N -volume means the points lie in a hyperplane of dimension no greater than
(N , 1).
These relationships between the sign of VN (~x) and the relative position of ~x are precisely
those we are accustomed to examining when we clip vectors (e.g., edges of a triangle) to lie on
one side of a plane in a viewing frustum or within a projected viewing rectangle. For example,
a 2D clipping line defined by the vector ~x1 , ~x0 = (x1 , x0 ; y1 , y0 ) has a right-handed
(unnormalized) normal N ~ = (,(y1 , y0); (x1 , x0)). Writing the 2D volume as the area A,
eq. (3) becomes
x ,x x,x h~ i
A ~x
( ) =
1
2
det
(
y ,y
( 1
1 0)
0)
(
( y,y
0)
0)
=
1
2
N ~x , ~x
( 0)
160 }
x+
V
x+
N
x
2
N
h x h
A 1
A
L x
0 x
1
x-
x p
0 p
x-
~x
Figure 5. In 2D, the line through 0 to 1 defined by n ^ ~x
( ~x , ~x
0 ) = 0 partitions the plane into two
~x
regions, one where this expression is positive (e.g., for + ) and another where it is negative (e.g., for , ). ~x
~x ; ~x ; ~x
In 3D, the analogous procedure uses the plane defined by ( 0 1 2 ) to divide 3-space into two half
h
spaces. The same pictures serve to show how the distance from a point to a hyperplane is computable
from the ratio of the simplex volume to the lower-dimensional volume of its base, i.e., 2 or 3 . A=L V=A
for some arbitrary point ~x, and so we recover the form of eq. (1) as
n^ (~x , ~x0 ) =
A ; 2
k~x , ~x k
1 0
where n ^ = N=~ kN~ k; the relationship of ~x to the clipping line is determined by the sign.
In 3D, when clipping a line against a plane, everything reduces to the traditional form, namely
the dot product between a 3D cross-product and a vector from a point ~x0 in the clipping plane
to the point ~x being clipped. The normal to the plane through (~x0 ; ~x1 ; ~x2 ) is
N~ =
~x , ~x y ~x, y, ~x y , y
( 1 0) ( 2 0) (9)
z ,z z ,z ;
( 1 0) ( 2 0)
= + det
x ,x x ,x x ,x x , x ;
( 1 0) ( 2 0)
, z ,z
det
z ,z ;
(
( 1
1 0)
0)
(
( 2
2 0)
0)
+ det
(
y ,y
( 1
1 0)
0)
(
y ,y
( 2
2 0)
0)
and we again find the same general form,
n^ (~x , ~x0 ) = ~ ;
6 V
kN k
whose sign determines where ~x falls. Figure 5 summarizes the relationship of the signed vol-
ume to the clipping task in 2D and 3D.
Hyperplanes for clipping applications in any dimension are therefore easily defined and
checked by choosing ~xN to be the test point ~x and checking the sign of eq. (3). If N ~ and a
point ~x0 are easy to determine directly, then the procedure reduces to checking the sign of the
left hand side of eq. (1).
The final step is to find the desired point on the truncated, clipped line. Since the clipped
form of a triangle, tetrahedron, etc., can be determined from the clipped forms of the component
lines, we need only consider the point at which a line straddling the clipping hyperplane inter-
sects this hyperplane. If the line to be clipped is given parametrically as ~x(t) = ~xa + t(~xb , ~xa ),
where ~xa and ~xb are on opposite sides of the clipping hyperplane so 0 t 1, then we simply
plug ~x(t) into V (~x) = 0 and solve for t:
~x , ~x ~x , ~x ~x , ~x
~x , ~x ~x , ~x ~x , ~x n^ (~x , ~x :
t =
det 1 0 2 0 a 0
=
n^ (~x
a
, ~x
0)
(10)
det 1 0 2 0 a b a b)
Here n
^ is of course just the normal to the clipping hyperplane, discussed in detail above.
} Point-Hyperplane Distance }
The general formula for the volume of a parallelepiped is the product of the base and the height,
W = Bh. In N dimensions, if we take WN = N ! VN to be the volume of the parallelepiped
with edges (~x1 , ~x0 ); (~x2 , ~x0 ); : : : ; (~xN ,1 , ~x0 ); (~x , ~x0 ), this generalizes to
hW , ; W N = N 1
where h is the perpendicular distance from the point ~x to the N , -dimensional parallelepiped ( 1)
with volume W , NN , V , and edges ~x , ~x ; ~x , ~x ; : : : ; ~x , , ~x . We

1 = ( 1)! N 1 ( 1 0) ( 2 0) ( N 1 0)
may thus immediately compute the distance h from a point to a hyperplane as
h WW NV
N, V , V ,
=
N NV :
=
! N
(11) =
N
, N 1 ( 1)! N 1 N 1
Note! Here one must use the trick of eq. 4 to express W , in terms of the square root of a N 1
square determinant given by the product of two non-square matrices.
Thus in 2D, the area of a triangle (~x0 ; ~x1 ; ~x) is
x ,x x,x
A V = 2 =
1
2
W 2 =
1
2
det
y ,y
(
( 1
1
y,y
0)
0)
(
(
0)
0)
and the length-squared of the base is L2 = ( 1 0) ( 1 0) ~x , ~x ~x , ~x so, with A

= (1=2)hL,
the height becomes h = 2A=L = W2 =L = W2 =W1 . In 3D, the volume of the tetrahedron
(~
x0 ; ~x1 ; ~x2; ~x) is V = V3 = (1=6)W3 and the area A = (1=2)W2 of the triangular base may be
~x , ~x ~x , ~x ~x , ~x ~x , ~x :
written
(2A )
2
= ( W2)
2
= det
(
(
1
~x , ~x ~x , ~x
2
0)
0)
(
(
1
1
0)
0)
(
(
1
~x , ~x ~x , ~x
2
0)
0)
(
(
2
2
0)
0)
162 }
x2
x1
x
x
x0 x0
x1
Figure 6. Barycentric coordinates in N dimensions.
We know V = (1=3)hA, and so h = 3V=A = 6V=2A = W3 =W2 . (See figure 5.) We note
for reference that, as we showed earlier, the base (N , 1)-volume is related to its normal by
N~ N~ = WN2 ,1.
Here we also typically need to answer one last question, namely where is the point p~ on the
base hyperplane closest to the point ~x whose distance h we just computed? This can be found
by parameterizing the line from ~x to the base hyperplane along the normal n ^ to the hyperplane
as ~x(t) = ~x + tn ^ (~
^ , writing the implicit equation for the hyperplane as n x(t) , ~x0) = 0, and
solving for the mutual solution tp = n ^ (~
x0 , ~x) = ,h. Thus
p~ = ~x n n ~x , ~x
+ ^(^ ( 0 ))
= ~x , hn :
^
} Barycentric Coordinates }
Barycentric coordinates (see, e.g., (Hocking and Young 1961), chapter 5) are a practical way to
parameterize lines, surfaces, etc., for applications that must compute where various geometric
objects intersect. In practice, the barycentric coordinate method reduces to specifying two
points (~x0 ; ~x1 ) on a line, three points (~x0 ; ~x1 ; ~x2 ) on a plane, four points (~x0 ; ~x1 ; ~x2 ; ~x3 ) in
a volume, etc., and parameterizing the line segment, enclosed triangular area, and enclosed
tetrahedral volume, etc., respectively, by
~x t ( ) = ~x t ~x , ~x
0 + ( 1 0) (12)
~x t ; t
( 1 2) = ~x t ~x , ~x
0 + 1( 1 t ~x , ~x
0) + 2( 2 0) (13)
~x t ; t ; t
( 1 2 3) = ~x t ~x , ~x
0 + 1( 1 t ~x , ~x
0) + 2( 2 t ~x , ~x
0) + 3( 3 0) (14)
:
The line and plane geometries are shown in figure 6. The interpolated point then lies within the
N -simplex defined by the specified points provided
t 0 1
t ; t ; ,t ,t
0 1 1 0 2 1 0 (1 1 2) 1
0 t ; t ; t ; ,t ,t ,t
1 1 0 2 1 0 3 1 0 (1 1 2 3) 1
:::
Center of What? However, this is really only half the story of barycentric coordinates. For
the other half, we seek a geometric interpretation of the parameters ti when we are given the
value of ~x.
First let us look at the simple case when ~x lies on the line segment between ~x0 and ~x1 .
Solving eq. (12) for t directly gives
~x , ~x ~x , ~x :
t ( 0) ( 1 0)
~x , ~x ~x , ~x
=
( 1 0) ( 1 0)
That is, t is the fraction of the distance that ~x has traveled along the line, the ratio between the
length from ~x to ~x and the total length. But, since ~x , ~x
0 ~x , ~x ~x , ~x , we easily see 1 0 = 1 + 0
that an alternative parameterization would be to take t t and 1 =
t ~x , ~x ~x , ~x ( 1 ) ( 1 0)
~x , ~x ~x , ~x
0 =
( 1 0) ( 1 0)
so that 1 = t t and eq. (12) for ~x becomes

0 + 1
~x t ; t t ~x t ~x :
( 0 1) = 0 0 + 1 1
If t
0 = 1 , then the entire fraction of the distance from ~x to ~x is assigned to t and ~x ~x . If 1 0 = 0
t
1 = 1 , then the entire fraction of the distance from ~x to ~x is assigned to t and ~x ~x . 0 1 = 1
Next, suppose we know ~x in a plane and wish to compute its barycentric coordinates by
solving eq. (13) for t ; t . Once we realize that ~x , ~x and ~x , ~x form the basis for an
( 1 2) ( 1 0) ( 2 0)
affine coordinate system for the plane specified by ~x ; ~x ; ~x in any dimension, we see that ( 0 1 2)
we may measure the relative barycentric coordinates by taking the dot product with each basis
vector:
~x , ~x ~x , ~x
( 0) ( 1 t k~x , ~x k t ~x , ~x ~x , ~x
0) = 1 1 0
2
+ 2( 2 0) ( 1 0)
~x , ~x ~x , ~x
( 0) ( 2 t ~x , ~x ~x , ~x t k~x , ~x k :
0) = 1( 1 0) ( 2 0) + 2 2 0
2
Extending the previously introduced abbreviation to the form v x; j ~x , ~x ~x , ~x ( ) = ( 0) ( j 0)

and solving this pair of equations by Cramer’s rule, we get
v x; ( 1) v (1 2);
vv x;; v ;
det
t
1 =
(
(1 1)
2)
v
(2 2)
(1 2);
det
v ; (2 1) v (2 2);
164 }
v ; (1 1) v x;
( 1)
vv ;; v x; :
det
t 2 =
(1 2)
(1 1) v ;
(
(1 2)
2)
det
v ; (2 1) v ; (2 2)
The denominator is clearly proportional to the square of the area of the triangle (~x0 ; ~x1 ; ~x2 ),
and the numerators have the form of squared areas as well. In N dimensions, the numerators
reduce to determinants of products of non-square matrices, and so may not be expressed as
two separate determinants! However, if we transform to a coordinate system that contains the
triangle within the plane of two coordinate axes, or if N = 2, an effectively square matrix is
recovered; one factor of area in the denominator then cancels out, giving the intuitively expected
result that the barycentric coordinates are ratios of two areas: t1 = A(~x; ~x0 ; ~x1 )=A(~x0 ; ~x1 ; ~x2 ),
t2 = A(~x; ~x2 ; ~x0 )=A(~x0 ; ~x1; ~x2 ). This leads us to introduce the generalized version of t0 for
the line, namely,
t0 = 1 , t , t AA ~x~x ;;~x~x ;;~x~x

1 2 =
( 1 2 )
~x , ~x ~x , ~x
( ~x , ~x ~x , ~x
( 0 1 2)
1 0) ( 1 ) ( 1 0) ( 2 )
det
~x , ~x ~x , ~x ~x , ~x ~x , ~x :
=
( 2
v ; v ;
0) ( 1
(1 1)
) ( 2
(1 2)
0) ( 2 )
v ; v ;
det
(2 2) (2 2)
Here we used the squaring argument given above to extend t from its special-coordinate- 0
system interpretation as the fraction of the area contributed by the triangle ~x; ~x ; ~x to the ( 1 2)
invariant form. This form obviously has the desired property that t when ~x ~x , and we 0 = 1 = 0
finally have the sought equation (with t t t) 1 = 0 + 1 + 2
~x t ; t ; t
( 0 1 2) = t ~x0 0 + 1 t ~x 1 + 2 t ~x :
2
It is amusing to note that the determinant identity 1 = t0 + t1 + t2 and its higher analogs, which
are nontrivial to derive, generalize the simple identity ~x1 , ~x0 = ~x1 , ~x + ~x , ~x0 that we used
in the 1D case.
Thus we can construct barycentric coordinates in any dimension which intuitively correspond
to fractions of hypervolumes; each barycentric coordinate is the hypervolume of an N -simplex
defined by the point ~x and all but one of the other simplex-defining points divided by the volume
of the whole simplex. The actual computation, however, is best done using the squared-volume
form because only that form is independent of the chosen coordinate system.
Note: The volumes are signed; even if ~x lies outside the N -simplex volume, the ratios remain
correct due to the cancellation between the larger volumes and the negative volumes. We also
remark that the generalized formulas for ti in any dimension, with 1 = N t , give an elegant P
i=0 i
geometric interpretation of Cramer’s rule as ratios of simplex volumes.

P2
x(t) x(t1,t2)
P
C C
P
1
Figure 7. Schematic diagram comparing an ordinary camera ray and a planar “thick ray” used in N-
dimensional ray-tracing methods.
} Ray Tracing }
It is often useful to compute the intersection of a ray passing through two points (typically the
camera focal point C ~ and an image point P~ ) with a geometrical object. In N dimensions, this
object will typically be an (N , 1)-simplex defining an oriented visible “face” with a normal
vector computable as described above. We need to do several things: compute the intersection
of the ray with the hyperplane containing the “face,” check to see whether the point lies within
the simplex’s boundaries (observe that this is a clipping problem), and see whether the normal
vector points in the direction of the ray (making it visible).
We formulate this procedure by first writing
X~ t
( ) = C~ t P~ , C~
+ ( )
for the position of a point on the camera ray, as illustrated in figure 7. Then we consider a single
(N , 1)-simplex of the tessellation to be described either by a known normal or by using the
set of N points giving its vertices to define its normal via eq. (6); in either case, we can write
the equation of any other point ~x lying within the simplex as
n^ (~x , ~x0 ) = 0 :
~ (t) in the simplex that lies on
Plugging in the parametric ray equation, we solve for the point X
the ray:
n^ ~x0 , C~
t :
n^ P~ , C~
=
A useful generalization of ray-tracing to N -dimensions follows from the observation that a

“thick ray” is cast into space by an open-ended simplex that is essentially a barycentric coordi-
nate form with the restriction 0 (1,t1 ,t2 ,: : :) 1 relaxed (see, e.g., (Hanson and Cross 1993)).
166 }
A planar ray such as that shown in figure 7 then has two parameters,
X~ t ; t
( 1 2) = C~ t P~ , C~
+ 1( 1 ) + t P~ , C~ ;
2( 2 )
with obvious generalizations to volume rays, etc. Intersecting such a planar ray with an (N , 2)-
dimensional manifold (describable using (N ,2) barycentric parameters) results in N equations
with N unknown parameters, and thus a unique point is determined as the mutual solution. In
3D, a plane intersects a line in one point, in 4D two planes intersect in a single point, while in
5D a plane intersects a volume in a point. Other generalizations, including rays that intersect
particular geometries in lines and surfaces, can easily be constructed. For example, the inter-
section of a planar ray with the single hyperplane equation for a 3-manifold in 4D leaves one
undetermined parameter, and is therefore a line.
} Conclusion }
Geometry is an essential tool employed in the creation of computer graphics images of every-
day objects. Statistical data analysis, mathematics, and science, on the other hand, provide
many problems where N -dimensional generalizations of the familiar 2D and 3D formulas are
required. The N -dimensional formulas and insights into the nature of geometry that we have
presented here provide a practical guide for extending computer graphics into these higher-
dimensional domains.
} Appendix: Determinants and the Levi-Civita Symbol }

One of the unifying features that has permitted us throughout this treatment to extend formulas
to arbitrary dimensions has been the use of determinants. But what if you encounter an expres-
sion involving determinants that has not been given here and you wish to work out its algebraic
properties for yourself? In this appendix, we outline for the reader a useful mathematical tool
for treating determinants, the Levi-Civita symbol. References for this are hard to locate; the
author learned these techniques by apprenticeship while studying general relativity, but even
classic texts like Møller (Møller 1972) contain only passing mention of the methods; somewhat
more detail is given in hard-to-find sources such as (Efimov and Rozendorn 1975).
First we define two basic objects, the Kronecker delta, ij ,
= 1 i j
=
i6 j
ij
= 0 =
and the Levi-Civita symbol, ijk::: , which is the totally antisymmetric pseudotensor with the
properties
= 1 i; j; k; : : :
in an even permutation of cyclic order
,
ij k:::
= 1 i; j; k; : : :
in an odd permutation of cyclic order
= 0 when any two indices are equal.
All indices are assumed to range from 1 to N , e.g., i = f1; 2; : : : ; (N , 1); N g, so that, for
example, (1234,1342,4132,4321), are even permutations and (1324,2134,1243,4312) are odd
permutations.
We can use the Kronecker delta to write the dot product between two N -dimensional vectors
as a matrix product with the Kronecker delta representing the unit matrix,
0 1
XX X @X X
A~ B~ BA
N N N N N
A B
= i ij j = A AB ;
i ij j = i i (15)
i=1 j =1 i=1 j =1 i=1
and the Levi-Civita symbol to write the determinant of a matrix jM j as

X
det [ M ] = 12
i i :::i N M1 1 M2 2 M
;i ;i N;i N :
all ik indices
The fundamental formula for the product of two Levi-Civita symbols is:
2 3
66 12 11 12 22 12 NN 77
i j i j i j
12
i i :::iN 1 2 Nj j :::j = det 64 .. i j
.. ..
.
i j
..
i j
75 :
. . .
N 1 N 2 N N
i j i j i j
(Note that if we set fj1 j2 : : : jN g = f1; 2; : : : ; N g, the second Levi-Civita symbol reduces to
+1, and the resulting determinant is an explicit realization of the antisymmetry of the Levi-
Civita symbol itself as a determinant of Kronecker deltas!)

With this notation, the generalized cross product N ~ of eq. (6), simplified by setting ~x0 = 0,
can be written
X
N~ = 12 i i :::iN ,1 N x1
i
1
(i )
x 2 x N,,1 x
(i )
2
(i
N 1
)
^
(i N) ;
all indices
where x ^; y
^ N are the unit vectors (x ^; : : : ; w
(i )
^ ) of the coordinate system. The dot product be-
tween the normal and another vector simply becomes

X
N~ L~ = 123
i i i :::i N ,1 N x1
i
1
(i )
x 2 x 3 x N,,1 L
(i )
2
(i )
3
(i
N 1
) (iN) :
all indices
The reader can verify that, in 2D, Nk =

P x 2 (i)
= ( ,y; x , and so on.
+ ) We conclude
i=1 ik
with two examples of applications:

168 }
Rotations of Normals. Is the normal N ~ a vector? Almost.PTo check this, we must rotate
each column vector in the cross product formula using x0(i) = N j =1
Rij x(j) and compute the
~ . Using the identity ((Efimov and Rozendorn 1975), p. 203),
behavior of N
X
12 i i :::i N ,1 N det [R]
i = 12 j j :::j N ,1 N R 1 1 R 2 2 R N ,1 N ,1 R N N
j j i j i j i j i ;
all jk indices
we find
X
N0 (i)
= 12i i :::i N ,1 i R 1 1x 1 R 2 2x
i j
(j )
1 i j
(j )
2
2 R N ,1 N ,1 x
i j
(j
N
N ,1 )
,1
all indices
except i
X N
= R N
ij
(j )
det [ R :
]
j =1
Therefore N~ is a pseudotensor, and behaves as a vector for ordinary rotations (which have
det [ R ] = 1 ), but changes sign if [R] contains an odd number of reflections.
Contraction Formula. The contraction of two partial determinants of (N ,K ) N -dimensional

vectors can expanded in terms of products of Kronecker deltas as follows:
X
12
i i :::i N ,K N ,K+1 N 1 2 N ,K N ,K+1 N
i :::i j j :::j i :::i =
N ,K+1 N
2 1 N ,K 3
i :::i
1 2
66 21 11 2 N ,K 77
i j i j i j
2 2
K ! det 64 .. i j i j
.. ..
.
i j
.. 75 :
. . .
N ,K 1 N ,K 2 N ,K N ,K
i j i j i j
The expression eq. (8) for the dot product of two normals is a special case of this formula.
Acknowledgment. This work was supported in part by NSF grant IRI-91-06389.

Bibliography
(Banchoff and Werner 1983) T. Banchoff and J. Werner. Linear Algebra through Geometry.
Springer-Verlag, 1983.
(Banchoff 1990) Thomas F. Banchoff. Beyond the Third Dimension: Geometry, Computer
Graphics, and Higher Dimensions. Scientific American Library, New York, NY, 1990.
(Brun et al. 1989) R. Brun, O. Couet, C. Vandoni, and P. Zanarini. PAW – Physics Analysis
Workstation, The Complete Reference. CERN, Geneva, Switzerland, October 1989. Ver-
sion 1.07.
(Chen et al. 1988) Michael Chen, S. Joy Mountford, and Abigail Sellen. A study in interactive
3-d rotation using 2-d control devices. In Proceedings of Siggraph 88, volume 22, pages
121–130, 1988.
(Coxeter 1991) H.S.M. Coxeter. Regular Complex Polytopes. Cambridge University Press,
second edition, 1991.
(Edmonds 1957) A.R. Edmonds. Angular Momentum in Quantum Mechanics. Princeton Uni-
versity Press, Princeton, New Jersey, 1957.
(Efimov and Rozendorn 1975) N.V. Efimov and E.R. Rozendorn. Linear Algebra and Multi-
Dimensional Geometry. Mir Publishers, Moscow, 1975.
(Feiner and Beshers 1990a) S. Feiner and C. Beshers. Visualizing n-dimensional virtual worlds
with n-vision. Computer Graphics, 24(2):37–38, March 1990.
(Feiner and Beshers 1990b) S. Feiner and C. Beshers. Worlds within worlds: Metaphors for
exploring n-dimensional virtual worlds. In Proceedings of UIST ’90, Snowbird, Utah,
pages 76–83, October 1990.
(Francis 1987) G.K. Francis. A Topological Picturebook. Springer-Verlag, New York, 1987.
(Hanson and Cross 1993) A.J. Hanson and R.A. Cross. Interactive visualization methods for
four dimensions. In IEEE Visualization ’93, 1993. To be published.

169 ISBN 0-12-336156-7
170 } BIBLIOGRAPHY
(Hanson and Heng 1992a) Andrew J. Hanson and Pheng A. Heng. Four-dimensional views of
3-D scalar fields. In Arie E. Kaufman and Gregory M. Nielson, editors, Proceedings of
Visualization 92, pages 84–91, Los Alamitos, CA, October 1992. IEEE Computer Society
Press.
(Hanson and Heng 1992b) Andrew J. Hanson and Pheng A. Heng. Illuminating the fourth
dimension. IEEE Computer Graphics and Applications, 12(4):54–62, July 1992.
(Hanson 1992) Andrew J. Hanson. The rolling ball. In David Kirk, editor, Graphics Gems III,
pages 51–60. Academic Press, San Diego, CA, 1992.
(Hocking and Young 1961) John G. Hocking and Gail S. Young. Topology. Addison-Wesley,
1961.
(Møller 1972) C. Møller. The Theory of Relativity. Oxford, Clarendon Press, 1972.
(Noll 1967) Michael A. Noll. A computer technique for displaying n-dimensional hyperob-
jects. Communications of the ACM, 10(8):469–473, August 1967.
(Phillips et al. 1993) Mark Phillips, Silvio Levy, and Tamara Munzner. Geomview: An interac-
tive geometry viewer. Notices of the Amer. Math. Society, 40(8):985–988, October 1993.
Available by anonymous ftp from geom.umn.edu, The Geometry Center, Minneapolis
MN.
(Sommerville 1958) D.M.Y. Sommerville. An Introduction to the Geometry of N Dimensions.
Reprinted by Dover Press, 1958.
A
A II.4
Rotations for N-Dimensional
Graphics
Andrew J. Hanson
Indiana University
Bloomington, IN 47405
hanson@cs.indiana.edu
} Introduction }
In a previous Gem (Hanson 1994), “Geometry for N -Dimensional Graphics,” we described a
family of techniques for dealing with the geometry of N -dimensional models in the context
of graphics applications. Here, we build on that framework to look in more detail at rotations
in N -dimensional Euclidean space. In particular, we give a natural N -dimensional extension
of the 3D rolling ball technique described in an earlier Gem (Hanson 1992), along with the
corresponding analog of the Virtual Sphere method (Chen et al. 1988). Next, we touch on
practical methods for specifying and understanding the parameters of N -dimensional rotations.
Finally, we give the explicit 4D extension of 3D quaternion orientation splines.
For additional details and insights, we refer the reader to classic sources such as (Som-
merville 1958,Coxeter 1991,Hocking and Young 1961,Efimov and Rozendorn 1975).
} The Rolling Ball in N Dimensions }

Basic Intuition of the Rolling Ball. The basic intuitive property of a rolling ball (or tan-
v
gent space) rotation algorithm in any dimension is that it takes a unit vector ^ 0 = (0; 0; : : : ; 0; 1)
pointing purely in the N -th direction (towards the “north pole” of the ball) and tips it in the di-
n
rection of an arbitrary unit vector ^ = (n1 ; n2 ; : : : ; nN ,1 ; 0) lying in the (N , 1)-plane tangent
v
to the ball at the north pole, thus producing a new, rotated unit vector ^ , where
v^ = MN v^ 0 n v
= ^ sin + ^ 0 cos ;
as indicated schematically in Figure 1a. (Note: for notational simplicity, we choose to write the
components of column vectors as horizontal lists.)
If we choose the convention that positive rotations are right-handed and progress counter-
clockwise, a positive rotation of the north pole actually tilts it into the negative direction of the
v
remaining axis of the rotation plane. That is, if the 2D “rolling circle” acts on ^ 0 = (0; 1) and

55 ISBN 0-12-543457-X
56 }
^ ^
n ^
v0
n
^ ^
v0 v cos θ
^
v θ
-1 -sin θ 0 +1
(a) (b)
v n
Figure 1. Tilting the “north pole” vector ^ 0 in the direction of the tangent vector ^ , as though rolling a
ball by placing one’s finger directly on the north pole and pulling in the direction ^ . n
n^ = (,1; 0) as shown in Figure 1b, then
v^ = M v^ = n^ sin + v^
2 0 0 cos = ( , sin ; cos ) ;
where the rotation matrix M2 can be written
M2 =

cos , sin
=

c ,s
+ sin cos +s c

c +nx s
=
,nxs c
: (1)
If we choose a right-handed overall coordinate frame, the sign of n^ will automatically generate
the correct sign convention.
Synopsis: Qualitatively speaking, if we imagine looking straight down at the north

pole, the rolling ball pulls the unseen N -th component of a vector along the direc-
n
tion ^ of the (N , 1)-dimensional controller motion, bringing the unseen compo-
nent gradually into view.
Implementation. In practice, we choose a radius R for the ball containing the object or scene
to be rotated and move our controller (slider, 2D mouse, 3D mouse, . . . ) a distance r in the
n
tangent direction ^ , as indicated in Figure 2a. Working from the simplified diagram in Figure
2b, we define D 2 = R2 + r 2 and choose the rotation parameters c = cos = R=D and
s = sin = r=D .
For interactive systems, this choice has the particular advantage that, however rapidly the
user moves the controller, 0 (r=D ) < +1, so 0 < =2. Depending upon the desired
interface behavior, an alternative choice would be to take = r=R. This requires comput-
ing a trigonometric function instead of a square root, and may cause large discontinuities in
orientation for large controller motion.
3D. The explicit 3D rolling ball formula can be derived starting from an arbitrary 2D mouse
r
displacement ~ = (x; y; 0) = (rnx ; rny ; 0), where n2x + n2y = 1. Then one replaces Equation
II.4 Rotations for N-Dimensional Graphics } 57
+ North
r^
n r
D^
v R^
v R^
v0 R
D
θ θ
- Tangent + Tangent
(a) (b)
Figure 2. The notation used in implementing the rolling ball rotation model for N dimensions.
(1) with nx = +1 by the analogous 33 matrix R0 for (x; z ) rotations and encloses this in
a conjugate pair of rotations Rxy that transform the 2D mouse displacement ~ into the strictly r
r
positive x-direction and back. Since even ~ = (,1; 0; 0) is rotated to the positive x-direction
before R0 acts, all signs are correct. With the explicit matrices
2
nx ,ny 0
3 2
c 0 +s
3
Rxy = 4 ny nx 0 5 ; R0 = 4 0 1 0 5 ;
0 0 1 ,s 0 c
we find an alternative derivation of the formula in our earlier Gem (Hanson 1992):
M3 = Rxy R0 (Rxy ),1

2
c + (ny )2 (1 c) , ,nxny (1 , c) nx s
3
= 4 ,
nx ny (1 c) , c + (nx ) (1 , c)
2
ny s 5
nx s , ,ny s c
2
1 , (nx) (1 , c)2
,nxny (1 , c) nx s
3
= 4 ,nxny (1 , c) 1 , (ny ) (1 , c)
2
ny s 5 : (2)
,nxs ,ny s c
r
4D. The 4D case takes as input a 3D mouse motion ~ = (x; y; z; 0) = (rnx ; rny ; rnz ; 0), with
n2x + n2y + n2z = 1. Then one first transforms (ny ; nz ) into a pure y -component, rotates that
result to yield a pure x-component, performs a rotation by in the (x; w)-plane, and reverses
the first two rotations. Defining the required matrices as
2 3
1 0 0 0
2
nx ,ryz 0 0
3 2
c 0 0 +s
3
6 0
6
ny
ryz , rnyzz 0 7
7 6 ryz
6 nx 0 0 7 6
7;R = 6 0 1 0 0 7
7;
Ryz = 6 nz ny 7 ; Rxy = 4 0
4 0 ryz ryz 0 5 0 0 1 0 5 4 0 0 1 0 5
0 0 0 1 0 0 0 1 ,s 0 0 c
(3)
58 }
2
where ryz = n2y + n2z , we find
M4 = Ryz Rxy R0 (Rxy ),1 (Ryz ),1

2
1 , (nx) (1 , c) ,(1 , c)nxny ,(1 , c)nxnz
2
snx
3
6
6 ,(1 , c)nx ny 1 , (ny ) (1 , c) ,(1 , c)ny nz 2
sny 7
7 :
= 4 ,(1 , c)nxnz ,(1 , c)ny nz 1 , (nz ) (1 , c) 2
snz 5
(4)
,snx ,sny ,snz c
ND. The extension of this procedure to any dimension is accomplished by having the con-
troller interface supply an (N , 1)-dimensional vector ~ = (rn1 ; rn2 ; : : : ; rnN ,1 ; 0) with r
r r n n
~ ~ = r 2 and ^ ^ = 1 and applying the rotation
MN = RN ,2;N ,1 RN ,3;N ,2 R ; R (R ; ), (RN , ;N , ), (RN , ;N , ),

12 0 12
1
3 2
1
2 1
1
2 3
1 , (n ) (1 , c)
1
2
,(1 , c)n n ,(1 , c)nN , n
2 1 sn 1 1 1
6
6
6
,(1 , c)n n 1 2 1 , (n ) (1 , c)
2
2
,(1 , c)nN , n sn 1 2
7
7
7
2
= 6
.. .. .. .. ..7 (5)
6 . . . . .7
6
4 ,(1 , c)n nN , ,(1 , c)n nN ,
1 1 2 1 1 , (nN , ) (1 , c)
1
2 7
snN ,1 5
,sn ,sn
1 2 ,snN , 1 c
r n
Recall that the controller input ~ = r ^ that selects the direction to “pull” also determines
c = cos = R=D , s = sin = r=D , with D 2 = R2 + r 2 , or, alternatively, = r=R.
} Controlling the Remaining Rotational Degrees of Freedom }

There are N (N , 1)=2 parameters in a general N -dimensional orthogonal rotation matrix, one
parameter for each possible pair of axes specifying a plane of rotation (the 3D intuition about
“axes of rotation” does not extend simply to higher dimensions). The matrix MN in Equation
(5) has only (N , 1) independent parameters: we must now understand what happened to the
other (N , 1)(N , 2)=2 degrees of freedom needed for arbitrary rotations.
In fact, the non-commutativity of the rotation group allows us to generate all the other ro-
tations by small circular motions of the controller in the (N , 1)-dimensional subspace of
r n
~ = r ^ . Moving the controller in circles in the (1; 2)-plane, (1; 3)-plane, etc., of the (N , 1)-
dimensional controller space exactly generates the missing (N , 1)(N , 2)=2 rotations required
to exhaust the full parameter space. In mathematical terms, the additional motions are generated
by the commutation relations of the SO (N ) Lie algebra for i; j = 1; : : : ; N , 1,
[RiN ; RjN ] = ij RNN , jN RiN + iN RjN , NN Rij
= , Rij :
The minus sign in the above equation means that clockwise controller motions in the (i; j )-plane
inevitably produce counterclockwise rotations of the object, and vice-versa. Thus the philoso-
phy (Hanson 1992) of achieving the full set of context-free rotation group transformations with
a limited set of controller moves extends perfectly to N -dimensions. Implementation Note: In
practice, the effectiveness of this technique varies considerably with the application; the size of
the counter-rotation experienced may be relatively small for parameters that give appropriate
spatial motion sensitivity with current 3D mouse technology.
Alternative Context Philosophies. The rolling ball interface is a context-free interface

that allows the user of a virtual reality application to ignore the absolute position of the con-
troller and requires no supplementary cursor context display; thus one may avoid distractions
that may disturb stereography and immersive effects in a virtual reality environment. However
some applications are better adapted to context-sensitive interfaces like the Arcball method
(Shoemake 1994) or the Virtual Sphere approach (Chen et al. 1988). The Virtual Sphere
approach in particular can be straightforwardly extended to higher dimensions by using the
rolling ball equations inside a displayed spatial context (typically a sphere) and changing over
to an (N , 1)-dimensional rolling ball outside the context; that is, as the controller approaches
and passes the displayed inner domain context sphere, the rotation action changes to one that
leaves the N -th coordinate fixed but changes the remaining (N , 1) coordinates as though an
(N , 1)-dimensional rolling ball controller were attached to the nearest point on the sphere.
Similar flexibility can be achieved by using a different controller state to signal a discrete rather
than a continuous context switch to the (N , 1)-dimensional controller.
} Handy Formulas for N -Dimensional Rotations }

For some applications the incremental orientation control methods described above are not as
useful as knowing a single matrix for the entire N -dimensional orientation frame for an object.
We note three ways to represent such an orientation frame:
Columns are new axes. One straightforward construction simply notes that if the default
x
coordinate frame is represented by the orthonormal set of unit vectors ^ 1 = (1; 0; : : : ; 0),
x x
^ 2 = (0; 1; 0; : : : ; 0), . . . , ^ N = (0; : : : ; 0; 1), and the desired axes of the new (orthonormal)
a (1) (2)
a
(N )
a
coordinate frame are known to be ^ 1 = (a1 ; a1 ; : : : ; a1 ), ^ 2 , . . . , ^ N , then the rotation
matrix that transforms any vector to that frame just has the new axes as its columns:
M =

a^ a^ a^N
1 2

:
The orthonormality constraints give M the required N (N , 1)=2 degrees of freedom.

60 }
Concatenated subplane rotations. Rotations in the plane of a pair of coordinate axes
x x
(^ i ; ^ j ), i; j = 1; : : : ; N can be written as the block matrix
2
1 0 0 0 0 0
3
6 .. .. .. .. .. .. .. .. .. 7
6 . . . . . . . . . 7
6 7
6
6 0 cos ij 0 0 , sin ij 0 7
7
6
6
6
0 0 1 0 0 0
7
7
7
Rij (ij ) = 6
.. .. .. .. .. .. .. .. .. 7 (6)
6 . . . . . . . . . 7
6
6 0 0 0 1 0 0
7
7
6
6
6
0 sin ij 0 0 cos ij 0
7
7
7
6 .. .. .. .. .. .. .. .. .. 7
4 . . . . . . . . . 5
0 0 0 0 0 1
and thus the N (N , 1)=2 distinct Rij (ij ) may be concatenated in some order to produce a
rotation matrix such as Y
M = Rij (ij )
i<j
with N (N , 1)=2 degrees of freedom parametrized by fij g. However, since the matrices
Rij do not commute, different orderings give different results and it is difficult to intuitively
understand the global rotation. In fact, as is the case for 3D Euler angles, one may even repeat
some matrices (with distinct parameters) and omit others, and still not miss any degrees of
freedom.
Quotient Space Decomposition. Another useful decomposition relies on the classic

quotient property of the topological spaces of the orthogonal groups (Helgason 1962),
SO (N )=SO (N , 1) = S N , 1
; (7)
where S K is a K -dimensional topological sphere. In practical terms, this means that the
N (N , 1)=2 parameters of SO (N ), the mathematical group of N -dimensional orthogonal
rotations, can be viewed as a nested family of points on spheres. The 2D form is the matrix (1)
parameterizing the points on the circle S 1 ; the 3D form reduces to the standard matrix
2
c + (n1 )2 (1 c) , , ,
n1 n2 (1 c) sn3 ,
n3 n1 (1 c) + sn2
3
n ,
M3 (; ^ ) = 4 n1 n2 (1 c) + sn3 ,
c + (n2 )2 (1 c) , ,
n3 n2 (1 c) sn1 5 (8)
, ,
n1 n3 (1 c) sn2 ,
n2 n3 (1 c) + sn1 ,
c + (n3 )2 (1 c)
n n
where the two free parameters of ^ ^ = (n1 )2 + (n2 )2 + (n3 )2 = 1 describe a point on
the 2-sphere. These two parameters plus a third from the S 1 described by c2 + s2 = 1 (i.e.,
c = cos ; s = sin ) yield the required total of three free parameters equivalent to the three
Euler angles. The 4D and higher forms are already too unwieldy to be conveniently written as
single matrices.
} Interpolating N-Dimensional Orientation Frames }

To define a uniform-angular-velocity interpolation between two N -dimensional orientation
frames, we might consider independently interpolating each angle in Equation (6), or we might
take the quotient space decomposition given by the hierarchy of points on the spheres (S N ,1 ; : : : ; S 2 ; S 1 )
and apply a constant angular velocity spherical interpolation to each spherical point in each suc-
cessive dimension using the “Slerp”
n^ n n n^ sin((1 , t)) + n^ sin(t )

12 (t) = Slerp(^ 1 ; ^ 2 ; t) = 1 2
sin( ) sin( )
n n
where cos = ^ 1 ^ 2 . (This formula is simply the result of applying a Gram-Schmidt decom-
position while enforcing unit norm in any dimension.)
Either of these often achieves the goal of smooth appearance, but the solutions are neither
unique nor mathematically compelling, since the curve is not guaranteed to be a geodesic in
SO (N ).
The specification of geodesic curves in SO (N ) is a difficult problem in general (Barr et al.
1992); fortunately, the two most important cases for interactive systems, N = 3 and N = 4,
have elegant solutions using the covering or “Spin” groups. For SO (3), geodesic interpolations
and suitable corresponding splines are definable using Shoemake’s quaternion splines (Shoe-
make 1985), which can be simply formulated using Slerps on S 3 as follows: let ^ be a unit n
3-vector, so that
q n
q0 = cos(=2); ~ = ^ sin(=2)
is automatically a point on S 3 due to the constraint (q0 )2 + (q1 )2 + (q2 )2 + (q3 )2 = 1. Then
each point on S 3 corresponds to an SO (3) rotation matrix
2
q02 + q12 q22 q32 , , 2q1 q2 , 2q q
0 3 2q1 q3 + 2q0 q2
3
R3 = 4 2q1 q2 + 2q0 q3 q02 + q22 2

1 ,q ,q 2
3 2q2 q3 , 2q q 0 1
5 (9)
2q1 q3 2q0 q2 , 2q2 q3 + 2q0 q1 q02 + q32 ,q ,q
2
1
2
2
which the reader can verify reduces exactly to the nested-sphere form in Equation (8). Note
that the quaternions q and ,q each correspond to the same 3D rotation. Slerping q generates
sequences of matrices R3 (t) that are geodesic interpolations. Arbitrary splines can be defined
using the method of Schlag (Schlag 1991).
Quaternions in Four Dimensions. In four dimensions, the correspondence between the

rotation group SO (4) and the spin group Spin(4) that is its double covering may be computed
by extending quaternion multiplication to act not just on 3-vectors (“pure” quaternions) v =
V
(0; ~ ), but on full 4-vector quaternions v in the following way:
3
X
R v = q v p,1 :

=0
62 }
We thus find that the general double-quaternion parameterization for 4D rotation matrices takes
the form
2
q0 p0 + q1 p1 + q2 p2 + q3 p3 , ,
q1 p0 q0 p1 q3 p2 + q2 p3
6
6 ,q p
1 0 + q0 p1 ,q3 p2 + q2 p3 , ,
q0 p0 + q1 p1 q2 p2 q3 p3
R4 = 4 ,q p
2 0 + q0 p2 ,q1 p3 + q3 p1 q1 p2 + q2 p1 + q0 p3 + q3 p0
,q p
3 0 + q p
0 3 ,q2 p1 + q1 p2 , ,
q1 p3 + q3 p1 q0 p2 q2 p0
, ,
q2 p0 q0 p2 q1 p3 + q3 p1 , ,
q3 p0 q0 p3 q2 p1 + q1 p2
3
, ,
q1 p2 + p1 q2 p0 q3 q0 p3 q1 p3 + p1 q3 + p0 q2 + q0 p2 7
7 :
, ,
q0 p0 + q2 p2 q1 p1 q3 p3 , ,
q2 p3 + q3 p2 q0 p1 q1 p0 5
(10)
q2 p3 + q3 p2 + q1 p0 + p0 q1 , ,
q0 p0 + q3 p3 q1 p1 q2 p2
One may check that Equation (9) is just the lower right-hand corner of the degenerate p = q
case of Equation (10).
Shoemake-style interpolation between two distinct 4D frames is now achieved by applying
the desired Slerp-based interpolation method independently to a set of quaternion coordinates
q (t) on one three-sphere, and to a separate set of quaternion coordinates p(t) on another. The
resulting matrix R4 (t) gives geodesic interpolations for simple Slerps, and can be used as the
basis for corresponding spline methods (Schlag 1991,Barr et al. 1992). Analogs of the N = 3
and N = 4 approaches for general N involve computing Spin(N ) geodesics and thus are quite
complex.
Controls. As pointed out in (Shoemake 1994), the Arcball controller can be adapted with
complete faithfulness of spirit to the 4D case, since one can pick two points in a three-sphere
to specify an initial 4D frame, and then pick two more points in the three-sphere to define
the current 4D frame. Equation (10) gives the complete form of the effective 4D rotation.
Alternately, one can replace the 4D rolling ball or Virtual Sphere controls described earlier by
a pair (or more) of 3D controllers (Hanson 1992).
Acknowledgment. This work was supported in part by NSF grant IRI-91-06389.
} Bibliography }
(Barr et al. 1992) Alan H. Barr, Bena Currin, Steven Gabriel, and John F. Hughes. Smooth in-
terpolation of orientations with angular velocity constraints using quaternions. Computer
Graphics, 26(2):313–320, 1992.
(Chen et al. 1988) Michael Chen, S. Joy Mountford, and Abigail Sellen. A study in interactive
3-D rotation using 2-D control devices. In Proceedings of Siggraph 88, volume 22, pages
121–130, 1988.
(Coxeter 1991) H.S.M. Coxeter. Regular Complex Polytopes. Cambridge University Press,
second edition, 1991.
(Efimov and Rozendorn 1975) N.V. Efimov and E.R. Rozendorn. Linear Algebra and Multi-
Dimensional Geometry. Mir Publishers, Moscow, 1975.
(Hanson 1992) Andrew J. Hanson. The rolling ball. In David Kirk, editor, Graphics Gems III,
pages 51–60. Academic Press, 1992.
(Hanson 1994) Andrew J. Hanson. Geometry for n-dimensional graphics. In Paul Heckbert,
editor, Graphics Gems IV, pages 149–170. Academic Press, 1994.
(Helgason 1962) Sigurdur Helgason. Differential Geometry and Symmetric Spaces. Academic
Press, New York, 1962.
(Hocking and Young 1961) John G. Hocking and Gail S. Young. Topology. Addison-Wesley,
1961.
(Schlag 1991) John Schlag. Using geometric constructions to interpolate orientations with
quaternions. In James Arvo, editor, Graphics Gems II, pages 377–380. Academic Press,
1991.
(Shoemake 1985) K. Shoemake. Animating rotation with quaternion curves. In Computer
Graphics, volume 19, pages 245–254, 1985. Proceedings of SIGGRAPH 1985.
(Shoemake 1994) Ken Shoemake. Arcball rotation control. In Paul Heckbert, editor, Graphics
Gems IV, pages 172–192. Academic Press, 1994.
(Sommerville 1958) D.M.Y. Sommerville. An Introduction to the Geometry of N Dimensions.
Reprinted by Dover Press, 1958.
Visualizing Quaternion Rotation
JOHN C. HART
Washington State University
GEORGE K. FRANCIS
University oflllinois, Urbana
and
LOUIS H. KAUFFMAN
University of Illinois, Chicago
Quaternions play a vital roIe in the representation of rotations in computer graphics, primarily
for animation and user interfaces. Unfortunately, quaternion rotation is often left as an ad-
vanced topic in computer graphics education due to difficulties in portraying the four-
dimensiona) space of the quaternions. One tool for overcoming these obstacles is the quaternion
demonstrator, a physical visual aid consisting primarily of a belt. Every quaternion used to
specify a rotation can be represented by fixing one end of the belt and rotating the other.
Multiplication of quaternions is demonstrated by the composition of rotations, and the resulting
twists in the belt depict visually how quaternions interpolate rotation.
This article introduces to computer graphics the exponential notation that mathematicians
have used t.a represent unit quaternions. Exponential notation combines the angle and axis of
the rotation into a concise quaternion expression. This notation allows the article to present more
clearly a mechanica[ quaternion demonstrator consisting of a ribbon and a tag, and develop a
computer simulation suitable for interactive educational packages. Local deformations and the
belt trick are used to minimize the ribbon’s twisting and simulate a natural-appearing interac-
tive quatemion demonstrator.
Categoriesand Subject Descriptors: 1.3,5 [Computational Geometry and Object Modeling]:

Hierarchy and Geometric Transformations; 1.3.6 [Methodology and Techniques]: Graphics
Data Structures and Data Types
General Terms: Algorithms, Theory
Additional Key Words and Phrases: Deformation, education, orientation interpolation, quater-
nions, rotation, visualization
J. C. Hart is supported in part by the NSF under a Research Initiation Award CCR-930921O.
Equipment was provided by the Imaging Research Laboratory, which is supported by the NSF
under grant CDA-9121675.
Authors’ addresses: J. C. Hart, School of Electrical Engineering and Computer Science, Washing-
ton State University, Pullman, WA 99164-2752; email: hart@ eecs.wsu.edu; G. K. Francis,
Department of Mathematics, University of Illinois at Urbana-Champaign, Urbana, IL 61801;
email: ,@ancis(iimath.uiuc.edu; L. H. Kauffman, Department of Mathematics, Statistics, and
Computer Science, University of Illinois at Chicago, Chicago, IL 60680; email: u10451@ uicvm.
bitnet,
Permission to copy without fee all or part of this material is granted provided that the copies are
not made or distributed for direct commercial advantage, the ACM copyright notice and the title
of the publication and its date appear, and notice is given that copying is by permission of the
Association for ComputingMachinery.To copy otherwise,or to republish,requiresa fee and/or
specificpermission.
@ 1994 ACM 0730-0301/94/0700-0256 $03.50
ACM Transactions on Graphics, Vol. 13, No. 3, July 1994, Pages 256-276.
Visualizing Quaternion Rotation . 257
1. INTRODUCTION
The method of specifying rotations and orientations of coordinate systems via
unit quaternions was formally introduced to the computer graphics commu-
nity by the publication of Shoemake [ 1985]. Quaternions were used in graph-
ics programming informally mostly by geometers because Sir William Rowan
Hamilton’s [Hamilton 1866] beautiful invention is not regularly taught in
college. Quaternions encode rotations by four real numbers (or two complex
numbers), whereas the linear representation of these transformations as
3 X 3 matrices requires nine. Moreover, Hamilton impressed explicit geomet-
rical meaning into every detail of his algebraic system, which guides intuition
and facilitates implementation [Francis and Kauffman 1994].
Interpolating the quaternionic representation of a sequence of rotations is
more natural than doing so for the familiar Euler angles, such as yaw, pitch,
and roll. The quaternions occupy a smooth, seamless, isotropic space which is
a generalization of the surface of a sphere. Thus, there is no need for special
care to avoid singularities, such as gimbal lock, where two rotation axes
collapse into one and thus make the interpolation irreversible.
Bezier curves were used in Shoemake [ 1985] to spline the quaternions
representing rotations, while Barr et al. [1992] used energy-minimizing
curves for demonstrably smoother motions. Quaternions provide an easy
mechanism for specifying an arbitrary rotation about an arbitrary axis. This
has long been exploited in keyboard user interfaces, and most recently for
specifying 3-dimensional rotations with a 2-dimensional mouse [Hanson 1992;
Shoemake 1992].
1.1 Overview
This article builds on previous work in quaternion rotation to derive an
implementation of the quaternion demonstrator. The first half of the article
summarizes various recent works on the quaternions. Section 2 reviews the
quaternion representation of three-dimensional rotation, based on Shoemake
[ 1985] and Francis and Kauffman [ 1994], and describes the quaternion
demonstrator, as devised originally in Kauffman [ 1987; 1991]. Section 3
describes the belt trick, summarizing Kauffman [ 1991] and Francis and
Kauffman [ 1994] and explaining the mathematics behind the animation “Air
on the Dirac Strings” [Sandin et al. 1993].
These sections form the basis for this article’s original contribution, found
in the second half of the article. Beginning with Section 4, techniques from
differential geometry model the quaternion demonstrator, regulating the
twists and motions of the belt. Section 5 describes the resulting implementa-
tion and outlines directions for further research.
1.2 Background
An object P is assumed to be defined with respect to some canonical coordi-
nate frame. The orientation of @ is represented by a rotation that takes the
object from its canonical coordinate frame to its current state.
ACM Transactions on (lraph,cs, Vol. 13, No. 3, (July 1994
258 . John C. Hart et al
This article relies heavily on the deformation techniques developed in Barr

[1984]. We use globally and locally specified deformations. A globally speci-
fied deformation alters explicitly the positions of points in an object whereas
a locally specified deformation affects the tangent space of an object, and new
positions result only after an integration over the deformed tangent space.
2. THE QUATERNIONS
The four-dimensional space, H, of quaternions is spanned by the real axis,
and three further orthogonal axes, spanned by vectors i, j, k, called the
principal imaginaries, which obey Hamilton’s rules
iz=jz=kz=ijk= –1. (1)

These imaginaries signify the three-dimensional vectors
i = (1,0,0),
j = (0,1,0),
k = (0,0,1).
Multiplication of these imaginaries resembles a cross product
ij=k, jk=i, ki=j,
ji=–k, kj =-i, ik=–j (2)
and is clearly noncommutative. Quaternion multiplication causes rotation:
multiplication on the right by j causes a 90 degree rotation in four-dimen-
sional space, rotating the i axis into the k axis, and rotating the k axis into
the – i axis. Quaternion multiplication differs from the cross product in that
ii=~=kk=–l whereasi Xi=jxj=k Xk=O.
A quaternion q = r + xi + y j + z k consists of a real part r and a pure
part xi + y j + zk [Hamilton 1866]. We will call quaternions with zero real
part (r = O) pure quaternions. Pure quaternions will also be simultaneously
represented as a column vector
x
v=
()Y
z
=xi+yj+zk,
Under this notation, the same symbol can simultaneously represent

(3)
both a
vector and a pure quaternion, depending on its context. For example,
V2 =
—v-v (4)
because the LHS of (4) treats v as a pure quaternion whereas the RHS of (4)
treats v as a vector.
Let q ~ = al + VI and q2 = az + V2 be two quaternions. Their sum is
ql +qz= (al +a2)+(vl +Vz),
and their product is

q1q2 = a1a2 – VI *V2 + alvz + a2v1 + VI X V2.
ACM Transactions on Graphics, Vol. 13, No. 3, July 1994

The quaternion q = a + v also decomposes into a + b u which resembles a

complex number, where the imaginary u is a unit three-vector
a-/b
U = v/b = ‘i + ‘j + ~k,
Ilvl Ilvll Ilvll
i] z/b
such that Ilull= 1, and x, y, z are the same coordinates used in (3). The pure
unit-magnitude quaternion u resembles ‘~- ‘—--: ----- 1: C--—
L1l~lllltl~lllii[y llU1ll ‘~- --—-1JJleX
L1le Gt)lll ---
plane in that Uz = – 1.
Let q = a + bu be a quaternion. Its conjugate is @ = a – bu, and its
magnitude is
Ilqll = qq = @ =
2.1 Quaternion Rotation

Rotations in computer graphics are typically represented by quaternions of
unit magnitude [Shoemake 1985], which we will call unit quaternions. The
unit quaternions {q : Ilqll = 1) form a hypersphere S:) c HI. In particular, for
any unit quaternion q G S“], (5) implies
l_–
q – q. (6)
In other words, to invert a unit quaternion, we simply negate its pure part.
In Shoemake [ 1985], a rotation of 0 about the axis u was represented as
the unit quaternion
1
q=cos~O+sin~Ou
which matches the complex-like form of a quaternion q = a + bu, where a is

the real component and b the imaginary component along the new imaginary
axis specified by the unit vector (pure quaternion) u. The abbreviation
e ‘(’ = cos O + i sin 0, borrowed from complex analysis, has a long history of
use in the engineering sciences. We can similarly represent the aforemen-
tioned unit quaternion q more concisely using exponential notation as
In the same manner that engineers read the expression e‘”, the reader should
likewise understand the notation e{ 1‘ 2)(’” not as e to some imaginary power
but simply as the quaternion that represents a rotation of o about the axis u.
Exponential notation was chosen to represent quaternion rotations berein
to promote consistency between the converging fields of computer graphics
and mathematics. Such quaternion exponential notation has a lengthy his-
tory in mathematics (e.g., as the so-called exponential map in different
geometry [Spivak 1965]). However, since quaternion multiplication is non-
commutative, likewise quaternion exponentiation does not in general follow
the rules of real or complex exponentiation (e.g., e{] 2)’””~e’1 “’”u’ +
ACM Transaction. on ~,raphics, VOI 13, NO 3. July 1994
260 . John C. Hart et al.
e(l\2x@’”’ +““’)). Quaternion exponentiation is formally defined in Francis and

Kauffman [1994], along with a discussion of its properties illustrated by
several examples.
Given a unit quaternion q that represents a rotation, the question remains
of how to apply this rotation to an arbitrary vector (pure quaternion) v = R 3.
From Shoemake [ 1985], we find two results: a function R(q) that returns the
3 x 3 (nonhomogeneous) transformation matrix corresponding to the rotation
represented by q, and a two-to-one correspondence between unit quaternions
S3 and the space of all rotations S0(3) (the group of special-orthogonal 3 x 3
matrices). As a consequence of these two results, we have
R(q)v=qvq-1. (7)
The LHS of(7) treats v as a column vector and yields a new column vector of
the same length by left-multiplying the special-orthogonal matrix returned by
the function R. The RHS of(7) treats v as a pure quaternion and yields a new
pure quaternion of the same magnitude. In fact, both apply the rotation
represented by the unit quaternion q to v. We denote rotations with the
notation on the LHS of(7), but implement rotations more efficiently using the
formula on the RHS of (7). (Since q = S3, q -1 simplifies to @.)
Hence, the otherwise complicated procedure of rotating a vector about an
arbitrary axis simplifies in our notation to
R(e+@u)v
which rotates v ● R!3 about the u G S2 axis by an angle of O [Francis and

Kauffman 1994].
2.2 The Quaternion Demonstrator

The quaternion demonstrator [Kauffman 1987; 1991] is a mechanical unit
quaternion multiplier. It consists primarily of a ribbon, called the belt, with
one end fixed and the other end free. Fastened to the free end of the belt is a
rectangle, called the tag. (An alternative demonstrator was discovered by
Kauffman and E. Oshins [Kauffman 1991] that uses only one human arm.)
The orientations of the tag, along with the twists in the belt, represent the
unit quaternions. The tag is inscribed with labels indicating the quaternion it
currently represents. The top side of the tag is inscribed with a 1 and an
upside-down j. Its bottom side is inscribed with a k and an upside-down i, in
the fashion suggested by Figure 1. (If you wish to follow along using your
right arm, your right hand will be the tag. Your fingerprints are 1; your palm
is j; your fingernails are k; and the back of your hand is i. As a tag, your hand
will represent the quaternion associated with the part of your hand facing up
and toward the same direction you are facing.)
First we orient the quaternion coordinate frame such that the imaginaries
form a right-handed coordinate system where i points right, j up, and k
toward the viewer. The belt of the quaternion demonstrator is embedded in
the i – k plane centered along the k axis with the fixed end at the origin and
the tag at + k. The eyepoint is assumed to be somewhere in the positive j – k
ACMTransactions on Graphics, Vol. 13, No. 3, July 1994.
Fig. 1. The quaternion demonstrator in the 1 state (top and bottom views). This figure uses the
computer simulation of the quaternion demonstrator, which does not use a tag. Instead it
superimposes the 1, i. j, and k labels directly onto the tag end of the belt.
quadrant. (In your arm’s coordinate system, the origin is your right shoulder;
the i axis points toward your left shoulder; the j axis points up; and the k
axis points in the direction you are facing.)
The canonical state of the demonstrator consists of the untwisted belt in
this configuration. This state represents the value 1, and is shown in Figure
1. (Extending your arm out in front of you with your palm up puts your arm
in the 1 state.)
Rotating the tag by 180 degrees with respect to the i axis puts the tag
underneath the belt. This state represents the value i as can be read on the
tag end of the belt (Figure 2 red). Call this rotation a “flip.” (Keeping your
arm and wrist straight, rotate 180 degrees about your shoulder’s axis by
dropping your arm to your waist and raise it back up behind you such that
your palm is facing down. The back of your hand is facing up and toward the
front, so your arm now represents the quaternion i.1
Reset the system to the canonical state 1. Rotating the tag by 180 degrees
with respect to the j axis puts the tag to the right of the fixed end of the belt.
This state represents the value j as can also be read on the tag end of the belt
(Figure 2 green). Call this rotation a “spin.” (From the 1 state, rotate your
arm horizontally until you touch your chest with your fingertips.
Your palm faces up and toward the front, and your arm now represents the
quaternion j.1
Resetting to 1 again and rotating the tag - 180 degrees with respect to the
k axis flips the tag over. This state represents the value k, as indicated by the
upright k on the tag end of the belt (Figure 2 blue). Call this rotation a
“turn.” (From the 1 state, turn your wrist about your arm’s axis until your
ACM Transactions on Graphics. Vol. 13. No 3, July 1994
Fig. 2. The quatemion demonstrator multiplica-

tion by each of i (red), j (green), and k (blue).
palm is facing down. Your fingernails face up and to the front, and your arm
now represents the quaternion k.)
The fact that positive k is obtained by a negative rotation in the right-
handed coordinate system is an artifact of the belt-centered coordinate sys-
tem. Centering a right-handed coordinate system on the tag with k extending
along the belt, j up, and i to the right yields rotations consistent with the
right-handed coordinate system. (For the quaternion demonstrator’s task of
teaching quaternion multiplication the simplicity of the belt-centered coordi-
nate system outweighs the familiarity of the right-handed rotation rules of
the tag-centered coordinate system.)
The negative quaternion imaginaries are likewise produced by the opposite
flips, twists, and turns, respectively. Although the labels on the tag have no
signs, we can tell a positive imaginary from a negative imaginary by the
direction of the twists in the belt. (The quaternion -i is represented from the
1 state by bending your arm at the elbow - 180 degrees about your shoulder’s
axis, with your palm facing down as if to pat yourself on the back. The
quaternion -j is represented by spinning your hand horizontally - 180
degrees about the up axis into the position a waiter would use to hold a tray.
The quaternion -k is physiologically impossible to represent in the arm
coordinate system.)
In this system, multiplication is represented by the composition of corre-
sponding rotations of the tag. For example, to demonstrate ij = k, we find
that a flip followed by a spin is equivalent to a turn (after a little translation
of the tag-moving the tag does not rotate it and does not change the state of
the demonstrator). A reverse turn (multiplication by - kl returns the system
to its original state. (From the 1 state, flip 180 degrees around your shoulder’s
axis by dropping your arm to your waist and raising it up again behind you
with your palm facing down into the i state. Using your shoulder, spin about j

Fig. 3. The belt trick equation: k’ = ( -k? proven visually using a quaternion demonstrator.
The ends of the belt remain parallel through the transformation. Each step of the transformation
represents the value - 1.
180 degrees by swinging your arm out and back to the front with your palm
down. Your arm is now in front of you, palm down, representing k.)
By (11, the - 1 state is achieved by two flips (i’), two spins (i” 1, two turns
(k” 1, or a flip-spin-turn (ijk). Each of these operations returns the tag to its
original orientation but puts a 360 degree twist in the belt. (From the 1 state,
twist your wrist 360 degrees about the arm’s axis until your palm faces up
again (k2). Your fingerprints are facing up and in front, but with a full 360
degree twist in your arm, which now represents - 1.)
We can also create - 1 by two reverse flips (( -i)” 1 which also returns the
tag to its original orientation but puts the opposite ( - 360 degrees) twist in
the belt (more on this in the next section.)
3. THE BELT TRICK

In the quaternion demonstrator, - 1 can be represented as k’ by two turns
which return the tag to its original state but cause a 360 degree twist in the
belt. Two reverse turns represents - 1 also, as ( - kJ2, but cause a - 360
degree twist in the belt. The quaternion demonstrator has (at least) two
distinct representations for - 1.
These two representations are equivalent, however. Consider the demon-
strator in the k2 state, after two turns. Without rotating the tag, move the
tag in a positive 360 degree arc around the fixed end of the belt. This belt
trick changes the 360 degree twist in the belt into a -360 degree twist, and
proves the Belt Trick Equation shown in Figure 3.
(You can perform a variation of the belt trick called the plate trick, shown
live in Sandin et al. 119931, with your arm by representing j4 = 1. From the 1
ACM Transactions on Graphics. Vol 13. No. 3, July 1994

state, multiply by j by touching your chest with your fingertips. Multiply by j

again by moving your hand under your arm and back out, always keeping the
palm up, into the – 1 state. Multiply by j again by swinging your hand into
the –j state, the waiter position, again always keeping your palm up.
Multiply by j one final time back into the 1 state. You have rotated your hand
720 degrees without incurring the associated, and painful, 720 degree twist
in your arm. Section 3.2 explains how this is possible.)
Discussions of the belt trick can be found in Misner et al. [1973], Bolker
[1973], Kauffman [1987], Francis [1987], Kauffman [1991], aqd Francis and
Kauffman [1994]. The rest of this section follows Kauffman [1991] and
Francis and Kauffman [1994], developing a globally specified deformation for
simulating the belt trick topologically, treating the belt more like a rubber
band than a ribbon. Thus we find that the quaternions are neatly represented
by a combination of rotational mechanism and appropriate topology.
3.1 The Belt Trick Deformation

Whereas Figure 3 demonstrates the belt trick using the quaternion demon-
strator, where the belt is fixed at one end and free at the tag, the following
discussion uses an equivalent but alternate construction. This new system
consists of a belt connecting two concentric spheres, and is described quanti-
tatively as follows.
Let
S(r) =rS2 = {X:11X11=r)
specify a sphere of radius r centered at the origin. The hollow ball
M.B = (Js(r),
r = r<)
where O < r. < 1 is the radius of the hollowed-out part, will serve to define
the space in which the belt performs its trick.
The spheres S(l), S(rO ) are called the outer sphere and the inner sphere,
respectively. The outer sphere will represent the fixed end of the belt, and the
inner sphere will represent the tag of the quaternion demonstrator. We can
use intervals to represent a belt connecting the inner sphere to the outer
sphere as
.99= ([-p, p], 0,[0, l])n HEl
where p < r. is half the width of the belt.

The belt-trick can now be illustrated as a global deformation 11~:R’3 ~ R 3
parameterized by time t = [0, 1]. At t = O, B~(&Z’) deforms the belt ~, giving
it a 360 degree twist. Aa t + 1, the belt will continuously deform into a belt
with a – 360 degree twist without rotating the inner or outer spheres—keep-
ing the ends of the belt fixed.
ACMTransactions
on Graphics,Vol. 13, No. 3, July 1994.
Fig. 4. Unit quaternion paths corresponding to various stages of the belt trick deformation. The
sphere S’ is a slice of the hypersphere S”, consisting of unit quaternions whose j components is
zero (similar to the figures at the end of Barr et al. [ 19921).
The belt trick deformation shears the belt, rotating each increasingly larger
spherical “shell” in WB by angles of increasingly larger value about an axis
that changes over time. The rotation angle function is
1 - Ml
O(x) = 27-7 (8)
0
whereas the rotation axis function is
u(t) = e”‘jk = (sinrt,O,cosnt). (9)
The function R, from (71, specifies rotation about an arbitrary axis, and is
used to define the belt trick deformation
B,(x) = R( e+H(x)u(l))X, (10)
Consider the unit quaternions used for rotations in the belt trick deforma-
tion. These are plotted in Figure 4. At t = 0, as 11x/J ranges from p. to 1, the
unit quaternions form an arc from 1 through k to - 1. As t: 0 + 1 this arc
rotates about S3 from one side to the other. At t = l/2 this arc extends from
1 through i to - 1, and at t = 1 this arc extends from 1 through -k to - 1.
Since the arcs all begin at 1, the orientation of the inner sphere, and end at
- 1, the orientation of the outer sphere, the inner and outer spheres do not
rotate during the belt trick.
ACM Transactions on Graphics, Vol. 13. No. ,3. July 1994
3.2 The Unfurling Deformation

Composing the belt trick with a twist along the h axis produces a belt trick
that takes the 720 degree twisted belt into an untwisted belt, and is given in
Francis and Kauffman [1994] as
~,, t(x) = R(e;o(x)u(f))l?(e+’z’’k)x

_— ~(e; Wdu(f)e+s2mk (11)
)x
where s E [ – 1, 1] is used to twist the belt. At s = – 1 and t = O, the

deformation B.,, leaves the belt $5’ unchanged. As s -+ 1 while t = O the
deformation B,, ~ rotates the inner spheres – 720 degrees, which puts a – 720
degree twist in the belt and returns the sphere to its original orientation.
Then, while s = 1 as t + 1, the – 720 degree twist unfurls around the inner
sphere, returning the system to its original state with an untwisted belt. This
process is illustrated by Figure 5. (The Appendix describes the ray-tracing
technique used to render this figure.)
The unit quaternions used for rotations in the unfurling deformation are
plotted in Figure 6. At s = 1, t = O, as 11xIIranges from pO to 1, the unit
quaternions form a circle from 1 through k through – 1 through – k and back
to 1. As t :0 + 1 this circle contracts around S~, always intersecting 1. At
t = 1/2 the circle extends from 1 through i and back to 1, and at t = 1 circle
degenerates to the point 1. Since the circles all begin and end at 1, as with
the belt trick, the inner and outer spheres do not rotate during the unfurling.
4. SIMULATING THE QUATERNION DEMONSTRATOR

The belt trick and unfurling deformations are global deformations. They
maintain the belt’s volume (for the same reason the twist deformation [Barr
1984] preserves volume) but stretch the length of the belt like a rubber band
such that the inner and outer spheres remain centered about the origin. The
belt in the quaternion demonstrator maintains a constant length, but the tag
is free to move about. This suggests that a local deformation should be used
to simulate the belt in the quaternion demonstrator. Furthermore, this local
deformation should minimize the twisting of the belt.
First, the orientation of the tag is represented by a quaternion q. Then a
geodesic (a great arc on S3 ) of unit quaternions interpolates the orientations
along the belt from the orientation of the fixed end of the belt, 1, to the
orientation of the tag, q. Finally, changes to this geodesic resulting from
rotations of the tag must be carefully monitored to prevent drastic changes in
the shape of the belt.
4.1 Simulating the Tag

In the physical quaternion demonstrator, the tag is a small rectangle at-
tached to the free end of the belt. In the simulation, we consider the tag to be
the edge of the free end of the belt.
Let VF represent the orientation of the edge of the belt’s fixed end as a
vector from one corner to the other. Let VT denote the orientation of the tag,
ACMTransactions
on Graphics,Vol. 13,No.3, July 1994.
Visualizing Quaternion Rotation - 267
Fig. 5. The unfurling, using a global defor-

mation to show how to remove a 720 degree
twist from a ribbon without moving either
end.
Fig. 6. Unit quaternion paths corresponding to various stages of the unfurling deformation
as the vector connecting the corresponding corners at the edge of the belt’s
free end. Let q E S” be a unit quaternion denoting the state of the quater-
nion demonstrator. Then the orientation of the tag with respect to the
orientation of the fixed end of the belt is given by
-
V T- R(q)vf..
Multiplication of q by an imaginary rotates the tag 180 degrees discretely.

In the simulation, these rotations are performed incrementally and appear
ACM Transactions on Graphics. Vol 13. No 3. .July 1994
268 - John C. Hart et al.
Fig. 7. The belt as the union of rotated line

segments (left). The orientation of v,, the
fixed edge of the belt, corresponds to one
whereas the orientation of vT, the tag edge of
the belt, corresponds to q. The geodesic on
!S3 (right) connects 1 to q, and specifies the
orientations of the segments interpolating v,
to VT.
continuous. With the physical quaternion demonstrator, multiplication by i is

performed by flipping the tag. In the simulation, multiplication by i is
accomplished by pressing and holding the “i” key and watching the tag slowly
flip.
Let q0 be a unit quaternion denoting the current orientation of the tag.
Then incremental multiplication of the tag is simulated by
q1 = qoecu, (12)
where unit quaternion q1 specifies the new tag orientation; u is one of i, j, or

k, and E is a small rotation angle. In our implementation, setting E = 0.02
radians resulted in a pleasing ribbon animation speed.
4.2 Simulating the Belt

In Section 3, the belt was sheared by a family of rotating concentric spheres.
Here, the belt is best represented by a family of rotating line segments. As
before, let v, represent the vector at the edge of the fixed end of the belt, and
let vr represent the vector at the tag.
The orientation of the fixed end of the belt v, corresponds to the unit
quaternion 1 whereas the orientation of the tag end of the belt vr corre-
sponds to q. Let I c S3 be the geodesic connecting 1 to q. Then the belt
consists of the union of line segments whose orientations interpolate the
orientation of v, into the orientation of vr, specifically the orientations
represented by points along the geodesic I, as shown in Figure 7.
We describe the local deformation of the belt by applying the rotations
represented by the quaternions along the geodesic I to the tangent space of
the belt. The deformed belt is then constructed as an initial-value problem by
integrating the belt over these deformed tangent vectors.
First, decompose q into exponential form as
q=r+v,
0 = 2cos-l r,
V V
u=J(vI1= Sin*8
q = eieu,
ACM Transactions on Graphics, Vol. 13, No. 3, July 1994.

nO
~ b.
n(s)
h,
\ (s)
Fig. 8, The Frenet frame along the spine of the belt,
x(s) t(s)
v
* \\
1
I
A
YQ
where cos -1 always returns values in the range ( – n, n]. This form reveals
the rotation q represents. The geodesic r extending from 1 to q on S3 is
parameterized by the function Y(s) G r for s G [0, l]-as
y(s) = e+sou. (13)
The unit quaternion function y(s) specifies the orientations the belt twists
through on its path from its fixed end to the tag. (Quaternion exponentiation
maps the line segment in R 3 connecting the origin to (1/2)0 u to the geodesic
in S3 connecting 1 to q.)
Let XO = O be the position, and to = k, b. = i, and nO = j be the Frenet
frame (tangent t, binormal b, and normal n) of the center of the fixed end of
the belt, as in Figure 8.
The local description of points along the spine of the belt is given by the
Frenet frame
t(s) = R(y(s))to,
b(s) = R(y(s))bo,
n(s) = R(y(s))no.
Integrating the tangent t(s) produces points along the spine of the belt
x(s) =X. +
J‘t(dck.
o
( 14)
The belt is formed as a ruled surface consisting of line segments connecting

the vertices
x(s) ~ pb(s),
where, as before, p is one-half the width of the belt. The twisting of the belt
visualizes the interpolation of orientations of a line segment from its fixed
end to the tag end.
Since the Frenet frame is just rotated by these functions, the length of the
belt remains unchanged under the deformation. (By the way, if the locally
specified deformation were not simply a rotation, then the normal vector
transformation rule [Barr 1984] would apply. In such a case, the transformed
binormal would be found via the inverse transpose of the Jacobian matrix of
R( ), and the normal would be constructed from the cross product of the
tangent vector with the binormal vector.)
4.3 Limiting Belt Velocity

The orientation interpolation geodesic on S3 is an arc extending from 1 (the
north pole) to q. As q passes by – 1 (the south pole) the geodesic generated
by (13) will move from one side of S3 to the other (since the range of cos -1 is
(–m, n]).
This movement keeps the belt from accumulating unnecessary twists. Since
the geodesic connecting 1 to q is the shortest path on S3, belt tricks occur
naturally as the tag is rotated to avoid twists of greater than 2 T in the belt.
This movement has one disadvantage in that certain small movements of q
near the south pole cause the geodesic to swing around quickly to the other
side of S3 resulting in a belt trick that is too fast for the user to follow. In
fact, if the tag rotates directly through – 1, the geodesic snaps from one side
of S3 to the other, causing an instantaneous belt trick (an instant reverse of
the twist in the belt).
For example, turning the tag about the k axis from the initial 1 state
produces eventually a 2 n twist in the ribbon about the k axis. Turning the
tag slightly causes an instantaneous belt trick, snapping the belt from a 2 m
twist to a – 2 IT twist. Although both configurations are nearly equivalent,
representing nearby quaternion values, their appearance to the user is quite
different.
There are two remedies for handling instantaneous belt tricks. The first
remedy senses when the tag orientation path crosses – 1, or nearly misses it.
When it does, this remedy assumes control of the demonstrator from the user
and performs an explicit animated belt trick to remove the excess twist in the
belt.
The second remedy capitalizes on numerical error to perform belt tricks
automatically, as necessary. As (12) rotates the tag incrementally, small
numerical errors will accumulate in the quaternion representation of the
tag’s orientation. In other words, turning the tag about the k axis will
introduce slight rotations about the i and j axes as well. By the time the tag’s
turning has twisted the belt by 2 n and more, these perturbations will cause
the tag’s orientation quaternion q to miss the south pole. The resulting
geodesics will quickly swing across S3 producing a belt trick, although
possibly at a very fast rate.
We chose to implement the second remedy in the simulation of the quater-
nion demonstrator for its elegance and because it never assumes control of
the demonstrator away from the user. This elegance may give the impression
that the second remeby avoids belt tricks, which are an essential point of the
quaternion demonstrator. To the contrary, belt tricks resulting from near
misses of the south pole are indiscernible from the belt trick required to
simulate the belt continuously through a direct hit. The dependence of this
remedy on numerical noise affects its robustness in that a direct hit on the
ACM Transactions on Graphics, Vol. 13, No. 3, July 1994,
south pole would lock up the quaternion demonstrator simulation. Such a

direct hit is highly unlikely and has never occured in our experience. The only
remaining task is to regulate the rate at which the quaternion demonstrator
performs automatic belt tricks.
The speed of the automatic belt trick is regulated by controlling the speed
at which the geodesic ~ flips around Ss, which, in turn, is controlled from the
rate of rotation of the tag by regulating the ● in (12). The rest of this section
is devoted to deriving the amount of regulation of ~ necessary to control the
speed of the automatic belt tricks when they occur.
Following Misner et al. [ 1973], we can describe a unit quaternion r + xi +
yj + ,zk in spherical coordinates with three angles a, @, and 0 as
x = sin u sin ~COs$,
y = sinasin~sin~,
z = sinocos~,
r = coscY.
The inverse is computed as
a=cos-lr,
+.cosd -
sin a ‘
“z
f)=cos-l
sin a sin 0
Y
= sin 1
sin a sin 0 ‘
Also from Misner et al. [ 1973], the differential of geodesic length is given by
dsz =dcr2 + sin2(a)(d@2 + sin2(@)d02). (15)

Geodesics extending from the north pole to q have fixed d, (-), and an a
that ranges from O at the north pole to a positive value at q.
Specify the original tag orientation qO in polar form as ( aO, 00, flo ) and
likewise with the new tag orientation q ~. The geodesic r. connecting 1 to qO
is of the polar form ([0, aO], #0, 190),and the new geodesic rl is ([0, all, 41, O1).
We are only concerned with geodesics that extend from the north pole to near
the south pole, at least where a., a, > IT/2. We also assume, without loss of
generality, that @o < dl and 00< 6’].
Let corresponding points on r. and rl be points of equal a. By observation,
the maximum distance between corresponding points on rO and rl occurs at
the equator, where a = 1/2. Let the distance geodesic r~ denote the geodesic
between corresponding equatorial points of r., I’l. By regulating the length of
r~, we regulate the rate of the belt trick.
Let lid = #Jl – 40 and A(I = @l – O.. From (15) we can approximate the
length of rd with
s’ = A@2 + sin2(@O)A02 (16)
since the change in a on r~ is zero and sin2 ( 7r/2) = 1

ACM Transactions on Graphics, Vol. 13, No. 3, .July 1994
Let s~,X be the maximum allowed length of rd—the maximum rate of

change between rO and rl. If the length of r~ exceeds the maximum allowed
length s~.X then we must reduce the increment of q. Let
A = sma./s (17)
be the amount rd needs to be scaled back to meet the maximum allowed

length. Then
s&X = A’s’
= A2(A~2 + sin2(@O)A62),
= (A A@)2 + sin2(@O)(AAO)2.
Changing the polar values # and 9 of qO by no more than AA@ and AA8
prevents the geodesic from rotating too fast around S3—prevents the belt
from moving too quickly. Hence, the incremental rotation
q~ = qoe*’”, ( 18)
where A is given in (17) as the quotient of the parameter s~aX and the
“distance” between successive geodesics s, produces a new tag orientation
sufllciently close to the original to limit belt movement properly when neces-
sary. In practice, setting s~,X = 0.1 disciplines the belt into reasonable
behavior.
5. CONCLUSION
Using the methods of Section 4, we have constructed a simulated quaternion
demonstrator, as described in Section 5.1. The simulated quaternion demon-
strator not only demonstrates unit quaternion multiplication, like its physical
counterpart, but also illustrates the quaternion interpolation of orientation
from one end of the belt to the other. The initial success of this prototype has
inspired several ideas for further research in this direction, which are de-
scribed in Section 5.2.
5.1 Results
Our implementation of the quaternion demonstrator, titled “quatdemo,” was
developed on an SGI Indigo Elan, and can be obtained via anonymous ftp to
the Imaging Research Laboratory at irl.eecs.wsu.edu from the directory
/pub/IRL/quatdemo.
It consists of the coordinate axis and a belt. The imaginaries are labeled at
the tag end of the belt, in their corresponding orientations. Pressing and
holding the “i,””’ j, “ or “k key rotates the tag end of the belt, causing minimal
twists in the belt. The current unit quaternion value is represented visually
by the configuration of the demonstrator and is verified by a formatted text
version of the current unit quaternion value.
Our implementation simulates the spine of the belt discretely with 256
samples, using the Euler method to approximate the integral (14). Euler
integration is highly susceptible to accumulated error, but errors in the
position of the tag end of the quaternion demonstrator are of little conse-
quence. The only state where the position of the tag is noticeable is – 1,
where the tag end of the belt should be coincident with the fixed end of the
belt. We are again fortunate in that the numerical noise accumulated in the
incremental rotations of the tag, statistically, prevents this state from being
represented exactly.
Figure 2 displays the various states of the quaternion demonstrator simu-
lated by the methods discussed in Section 4. As expected, belt tricks occur
automatically when necessary to remove excess twists in the belt; the belt
never contains more than a full 360 degree twist in any direction. Figure 3
displays an automatically occurring belt trick, which results at the midway
point when holding the “k key down.
After an automatic belt trick, enough error accumulates to cause the spine
of the belt to return to a position slightly offset from its original state. In our
implementation, pressing the space bar resets the quaternion demonstrator
to the 1 state.
5.2 Further Research

The concept of illustrating the track of a rotation through the use of attached
belts to objects is basic to the quaternion demonstrator. In this article we
have considered the resulting symmetries of a rectangle in three-dimensional
space. The same results apply to the symmetry of any object in three-
space R ‘).
Formally, Let @ c R’:i be a subset of R’:] containing the origin. Let S0(3)
denote the set of rotations about the origin of R ‘3. Let Symm( F ) denote the
subgroup of 50(3) consisting in those rotations g = S0(3) for which g(~) = ~
setwise, Now the three-sphere Ss of unit quaternions covers the S0(3) doubly
via the map
77: S:)+ s0(3) :V+l?(q)v=qvq ‘, (19)
where v is a pure quaternion—hence v is a vector in R:). This is an abstract

description of our representation of rotations by quaternions.
Now, the set of unit quaternions, n 1(Symm( @ )), covering the symmetry
group of the object, is a subgroup of S?, called the binary group of Symm( F).
If the object F“ is the rectangular tag Y, this group is the eight-element
quaternion group,
(20)
This is the abstract description of what the quaternionic demonstrator has

demonstrated.
An extension of the demonstrator (by attaching a belt to another object, P)
can ill ustrate the binary group for the symmetries of any object. These
groups, in the case of regular solids such as the tetrahedron, octahedron, or
icosahedron, are of great interest both practically and mathematically. This
extension of our demonstrator is one of the immediate prospects for further
research.
A(;M Transactions on Graphics, Vol. 1:3,No. 3. July 1994
It is also possible to extend our methods to study the structure of rotations

of four-space and to the study and illustration of properties of octonians
[Cayley 1897], which are an eight-dimensional generalization of the quater-
nions.
APPENDIX
RENDERING
The most straightforward method for rendering a deformed object is to
polygonize its surface, apply the deformation to the polygon vertices, then
render the resulting polygons. Special-purpose hardware can render polygo-
nal objects in real time, permitting interactive modeling. For example, the
local deformation of the belt used in the quaternion demonstrator was
rendered in this fashion. Polygonization can be problematic when investigat-
ing deformations in that the result of deforming vertices produces a polygon
that may be a poor fit when compared to the deformation of the entire
polygon, requiring some form of detection and dynamic subdivision.
Alternatively, we can render the deformed object directly as an implicit
surface, preserving, at least to pixel precision, the detail of the deformed
geometry. Let ~(x) be a function implicitly defining the set A c R3 such that
f(x) <o-x Ez4, (21)

f(x)= O=x GdA, (22)
f(X)> O= XGlR3\A, (23)
and let D: R 3 + R 3 denote the deformation function. Then the deformed set
D(A) is implicitly defined by the function ~ o D l(x).
With few exceptions, ray tracing is the means for direct visualization of
implicit surfaces. Some recent ray intersection methods require the Lipschitz
constant of the function [Kalra and Barr 1989; Hart 1993]. The Lipschitz
constant of a function f A + B from metric space (A, dA ) to metric space
(B, d~ ) is the smallest positive value A such that
dB(f(x), f(y)) < AdA(x, y) (24)
for all x, y = A. The Lipschitz constant bounds the amount a transformation

can expand an object. If f R’ + R, then the Lipschitz constant of f indicates
the steepest slope in the graph of f. One can use the Lipschitz-based
ray-tracing method to investigate the “unfurling,” the global deformation
described in Section 3.2.
By observation, the largest dilation caused by B,,,, occurs when s = 1,
t = O. This deformation winds the segment [r, l]i twice around the origin.
Reducing to R’2, the parametric equation of this curve takes the x axis from
[r, 1] into the twice-winding curve as
f,(x) =X COS26NX), (25)

(Y(x) =xsin20(x) (26)

with derivatives
f~(x) ==cos2#(x) - 20’(x) sin2@(x), (27)
f{,(x) =sin20(x) +20’(x) cos26(x). (28)
21r
o’(x)” –— (29)
1 – r,,
Where (29) corresponds to (8). The arc length of this double twist is found
using
167T2
=1+ (31)
(1 – ro)2°
Since dsz reaches its maximum (over the proper domain) when x = 1, we
have the Lipschitz constant
(32)
The deformation B] ~, dilates more than any other B,,,, for all s E [ – 1, 1]
and t e [0, 1]. Equation (32) is an upper bound of the Lipschitz constant for
B.,,, and under a similar argument, for the belt trick deformation B,. Hence,
(32) is a suitable (though not necessarily optimal for all parameters s and t )
Lipschitz bound for ray-tracing the results of the belt trick and unfurling
deformations.
This Lipschitz constant was used to render the unfurling demonstration in
Sandin et al. [ 1993], excerpted in Figure 5.
ACKNOWLEDGMENTS
We wish to thank the editor and reviewers for very extensive comments and
suggestions on this topic.
REFERENCES
BARR, A. H., CLTRRIN, B., GABRIEL, S., ANO HIJGHIW, J. F. 1992. Smooth interpolation of
orientations with angular velocity constraints using quaternims. Compuf. ~rnph. 26, 2 (July),
313 320.
BARR, A. H. 1984. Global and local deformations of solid primitives. Compuf. Graph. 18, 3
(July), 21 30.
Bt)i.KK~, E. 1973. The spinor spanner. Am. Math. Month. 80, 9 (Nov.), 977 984.
CAw.KY, A. 1897. The (%llected Mathematical Papers of Arth ur C’CI.de.V.carnbridw [Jniwrsity
Press, Cambridge, U.K.
~llAN(lS, (1. K. 1987. A Topologiml Picturebook. Springer-Verlag, New York.
FItANIIS, G. K. ANI}KAIIWMAN, L. H. 1994, Air on the Dirac strings. In Mathcrnaticcl Lq+acy of
Wi(he[m Magnus. AMS, Providence, RI.
HAMIl,T(~~,W. R. 1866. E/cment,s of Quaternion,s, Longmans Green. London.
ACM Transactions on Graphics, Vo. I:1, NO 3, July 1994

HANSON,A, 1992, The rolling ball: Applications of a method for controlling three degrees of
freedom using two-dimensional input devices. In Graphics Gems, D. Kirk, Ed. Vol. 3. Academic
Press, San Diego, Calif,, 51-60.
HART, J. C. 1993. Sphere tracing: Simple robust antialiased rendering of distance-based
implicit surfaces. Tech. Rep. EECS-93-15, School of EECS, Washington State Univ., Pullman,
Wash. Appears in SIGGRAPH ’93, Course Notes #25 “Modeling, Visualizing and Animating
Implicit Surfaces.”
KALRA, D. AND BARR, A. H. 1989. Guaranteed ray intersections with implicit surfaces. Comput.
Graph. 23, 3 (July), 297-306.
KAUFFMAN, L. H. 1991. Knots and Physics. World Scientific, Teaneck, N.J.
KAUFFMAN,L. H. 1987. On Knots. Princeton University Press, Princeton, NJ,
MISNER, C, W,, THORNE, K. S., AND WHEELER, J, A. 1973. Gravitation. Freeman, San Fran-
cisco, Calif.
SANDIN, D. J., KAUFFMAN,L. H., ANrIFRANCM,G. K. 1993. Air on the Dirac strings. SIGGHAPH
Video Reu. 93. Animation.
SHOEMARE,K. 1992. ARCBALL: A user interface for specifying three-dimensional orientation
using a mouse. In Proceedings of Graphics Interface ’92 (May) 151-156.
SHO~MAIO?,K. 1985. Animating rotation with quaternion curves. Comput. Graph. 19, 3 (July),
245-254.
SPIVAK, M. 1965. Calculus on Manifolds. W. A. Benjamin, New York.
Received October 1993; revised July 1994; accepted July 1994
Editor: John Hughes

Constrained Optimal Framings of Curves and Surfaces using Quaternion
Gauss Maps
Andrew J. Hanson

Indiana University
K
A
A
A
A
A
(a) (b) (c)

Figure 1: The (3,5) torus knot, a complex periodic 3D curve. (a) The line drawing is nearly useless as a 3D representation. (b) A
tubing based on parallel transporting an initial reference frame produces an informative visualization, but is not periodic. (c) The
arrow in this closeup exposes the subtle but crucial non-periodic mismatch between the starting and ending parallel-transport
frames; this would invalidate any attempt to texture the tube. The methods of this paper provide robust parameterization-
invariant principles for resolving such problems.
Abstract 1 Introduction
We propose a general paradigm for computing optimal coordinate We propose a general framework for selecting optimal systems
frame fields that may be exploited to visualize curves and surfaces. of coordinate frames that can be applied to visualizing geometric
Parallel-transport framings, which work well for open curves, gen- structures such as curves and surfaces in three-dimensional space.
erally fail to have desirable properties for cyclic curves and for The methods contain “minimal-turning” parallel-transport framings
surfaces. We suggest that minimal quaternion measure provides of curves as a special case, are independent of parameterization, and
an appropriate heuristic generalization of parallel transport. Our extend naturally to situations where parallel transport is not appli-
approach differs from minimal-tangential-acceleration approaches cable.
due to the addition of “sliding ring” constraints that fix one frame
axis, but allow an axial rotational freedom whose value is varied Motivation. Many visualization problems require techniques for
in the optimization process. Our fundamental tool is the quater- effectively displaying the properties of curves and surfaces. The
nion Gauss map, a generalization to quaternion space of the tan- problem of finding appropriate representations can be quite chal-
gent map for curves and of the Gauss map for surfaces. The quater- lenging. Representations of space curves based on single lines are
nion Gauss map takes 3D coordinate frame fields for curves and often inadequate for graphics purposes; significantly better images
surfaces into corresponding curves and surfaces constrained to the result from choosing a “tubing” to display the curve as a graphics
space of possible orientations in quaternion space. Standard opti- object with spatial extent. Vanishing curvature invalidates meth-
mization tools provide application-specific means of choosing op- ods such as the Frenet frame, and alternative approaches to tubing
timal, e.g., length- or area-minimizing, quaternion frame fields in involve heuristics unrelated to parameterization-invariant optimiza-
this constrained space. tion measures in order to achieve such properties as periodicity.
Similar problems occur in the construction of suitable visualiza-
tions of complex surfaces and oriented particle systems on surfaces,
CR Categories: I.3.6 [Computer Graphics]: Methodology and
since the intrinsic orientation properties may be poorly exposed by
Techniques. I.3.8 [Computer Graphics]: Applications.
the original representation. If a surface patch is represented by a
Keywords: Quaternions; Frames; Tubing; Curves; Surfaces rectangular but nonorthogonal mesh, for example, there is no obvi-
ous way to choose among alternative local orthonormal frame as-
signments; if the surface has regions of vanishing curvature, meth-
Email: hanson@cs.indiana.edu ods based on directions of principal curvatures break down as well.
(a) (b) (c)
Figure 2: (a) A smooth 3D surface patch having a non-orthogonal parameterization, along with its geometrically-fixed normals at the four
corners. No unique orthonormal frame is derivable from the parameterization. If we imitate parallel transport for curves to evolve the initial
frame at the top corner to choose the frame at the bottom corner, we find that paths (b) and (c) result in incompatible final frames at the bottom
corner. This paper addresses the problem of systematically choosing a compatible set of surface frames in situations like this.
While we emphasize curves and surfaces in this paper to provide where one direction of the frame is already fixed, and the chosen
intuitive examples, there are several parallel problem domains that functional minimization in quaternion space must obey the addi-
can be addressed with identical techniques. Among these are extru- tional constraint imposed by the fixed family of directions. Addi-
sion methods and generalized cones in geometric modeling, the im- tional references of interest, especially regarding the treatment of
position of constraints on a camera-frame axis in key-frame anima- surfaces, include [14, 20]. Figure 3 provides a visualization of the
tion, and the selection of a 2D array of camera-frame axis choices difference between the general interpolation problem and our con-
as a condition on a constrained-navigation environment (see, e.g., strained problem: a typical spline minimizes the bending energy
Hanson and Wernert [13]). specified by the chosen anchor points; requiring intermediate points
Figure 1 summarizes the basic class of problems involving to slide on constrained paths during the minimization modifies the
curves that will concern us here. The line drawing (a) of a (3,5) problem. In particular, 3D spline curves need not intersect any of
torus knot provides no useful information about the 3D structure. the constraint paths. In addition, we note that we typically have al-
Improving the visualization by creating a tubing involves a subtle ready sampled our curves and surfaces as finely as we need, so that
dilemma that we attempt to expose in the rest of the figure. We can- piecewise linear splines are generally sufficient for the applications
not use a periodic Frenet frame as a basis for this tubing because in- we discuss.
flection points or near-inflection points occur for many nice-looking Our solution to the problem is to transform the intrinsic geo-
torus knot parameterizations, and in such cases the Frenet frame is metric quantities such as the tangent field of a curve and the normal
undefined or twists wildly. The parallel-transport tubing shown in field of a surface to quaternion space and to construct the quaternion
(b) is well-behaved but not periodic; by looking carefully at the manifold corresponding to the one remaining degree of rotational
magnified portion next to the arrow in Figure 1(c), one can see a freedom in the choice of coordinate frame at each point. Paths in
gross mismatch in the tessellation due to the nonperiodicity. While this space of possible frames correspond to specific choices of the
it would be possible in many applications to ignore this mismatch, quaternion Gauss map, a subspace of the space of possible quater-
it has been the subject of a wide variety of previous papers (see, nion frames of the object to be visualized. Mathematically speak-
e.g., [16, 24, 5]), and must obviously be repaired for many other ing, the space of possible frames is the circular Hopf fiber lying
applications such as those requiring textured periodic tubes. above the point in S2 corresponding to each specific curve tangent
Figure 2 illustrates a corresponding problem for surface patches. or surface normal (see, e.g., [26, 3]).
While the normals to the four corners of the patch are always well-
defined (a), one finds two different frames for the bottom corner
depending upon whether one parallel transports the initial frame Parallel Transport and Minimal Measure. Constraining
around the left-hand path (b) or the right-hand path (c). There is each quaternion point (a frame) to its own circular quaternion
no immediately obvious right way to choose a family of frames path (the axial degree of rotational freedom), we then minimize
covering this surface patch. the quaternion length of the frame assignment for curves and the
Our goal is to propose a systematic family of optimization meth- quaternion area of the frame assignment for surfaces to achieve an
ods for resolving problems such as these. optimal frame choice; this choice reduces to the parallel-transport
frame for simple cases. Our justification for choosing minimal
quaternion length for curves is that there is a unique rotation in
Methodology. We focus on unit quaternion representations of the plane of two neighboring tangents that takes each tangent di-
coordinate frames because of the well-known natural structure of rection to its next neighbor along a curve: this is the geodesic arc
unit quaternions as points on the three-sphere S3 , which admits a connecting the two frames in quaternion space, and is therefore the
natural distance measure for defining optimization problems, and minimum distance between the quaternion points representing the
supports in addition a variety of regular frame-interpolation meth- two frames. The choice of minimal area for surface frames is more
ods (see, e.g., [25, 23, 19, 15]). We do not address the related heuristic, basically a plausibility argument that the generalization
question of optimal freely moving frames treated by the minimal- of minimal length is minimal area; no doubt this could be made
tangential-acceleration methods (see, e.g., [2, 22, 8]); we are in- more rigorous.
stead concerned with closely-spaced points on curves and surfaces By imposing other criteria such as endpoint derivative values and
(a) (b) (c)
Figure 3: (a) The camera frame interpolation problem is analogous to the problem of finding a minimal-bending spline curve through a series
of fixed key points. (b) The optimal curve frame assignment problem is analogous to fixing the end points of a curve segment and choosing
in addition a family of lines along which the intermediate points are constrained to slide during the optimization process; in 3D, the spline
path need not pass through the constraint lines. (c) In this paper, our sample points are generally close enough together that we apply the
constraints to piecewise linear splines analogous to those shown here.
minimal bending energy (see Barr et al. [2, 22]), the short straight possible choices of the normal plane orientation: it is just the set of
line segments and polygons that result from the simplest mini- rotation matrices R(; T ^ (t)) (or quaternions q (; T ^ (t))) that leave
mization could be smoothed to become generalized splines passing T^ (t) fixed.
through the required constraint rings; since, in practice, our curve For surfaces, the analogous construction follows from determin-
and surface samplings are arbitrarily dense, this was not pursued in
ing the unit normal N ^ (u; v ) at each point x(u; v ) on the surface
patch. The needed family of rotations R(; N ^ (u; v )) (or quater-
the current investigation.
For space curves, specifying a frame assignment as a quaternion
path leads at once to tubular surfaces that provide a “thickened” rep- nions q (; N ^ (u; v ))) now leaves N ^ (u; v ) fixed and parameterizes
resentation of the curve that interacts well with texturing, lighting, the space of possible tangent directions completing a frame defini-
and rendering models. For surface patches, the approach results in tion at each point x(u; v ).
a structure equivalent to that of an anisotropic oriented particle sys- We now define f (; v ^ ) = (f0 ; f1 ; f2 ; f3 ) to be a quaternion de-
tem (also a species of texture) whose pairs of tangent vector fields scribing the family of frames for which the direction v ^ is a preferred
in the surface produce natural flow fields that characterize the local fixed axis of the frame, such as the tangent or normal vectors. The
surface properties and are easy to display. orthonormal triad of 3-vectors describing the desired frame is
F (; v^ ) =
"
f02 + f12 , f22 , f32 22f1 f22 , 22f0 f32 2f1 f3 + 2f0 f2 #
Background. General questions involving the specification of
curve framings have been investigated in many contexts; for a rep-
resentative selection of approaches, see, e.g., [16, 24, 5]. The 2f1 f2 + 2f0 f3 f0 , f1 + f2 , f3 22f2 f32 , 22f0 f12 ;(1)
quaternion Gauss map is a logical extension of the quaternion frame 2f1 f3 , 2f0 f2 2f2 f3 + 2f0 f1 f0 , f1 , f2 + f3
approach to visualizing space curves introduced by Hanson and Ma
[11, 12]. For basic information on orientation spaces and their rela- where one column, typically the 3rd column, must be v ^.
The standard rotation matrix R(; v^ ) leaves v
^ fixed but does not

tionship to quaternions, see, e.g., [1, 21, 19].
Background on the differential geometry of curves and surfaces have v
^ as one column of the 3 3 rotation matrix, and so we have
may be found in sources such as the classical treatise of Eisenhart more work to do. To compute f (; v^ ), we need the following:
[7] and in Gray’s M ATHEMATICA-based text [9], which inspired a
number of the illustrations in this paper. The classical Frenet frame
A base reference frame b(^ v) that is guaranteed to have, say,
the 3rd column exactly aligned with a chosen vector v ^ , which
is defined and studied in these texts. The frame we refer to as the
is either the tangent to a curve or the normal to a surface.
parallel-transport frame was first described carefully by Bishop [4],
and has been commonly used in graphics (see, e.g., [5, 24, 17]). A A one-parameter family of rotations that leaves a fixed direc-
significant difference between these two methods is that the Frenet tion v
^ invariant.
frame is locally defined but possibly discontinuous, whereas the
parallel-transport frame is continuous but non-local, corresponding The latter family of rotations is given simply by the standard
to the solution of a differential equation. quaternion
q(; v^ ) = (cos 2 ; v^ sin 2 ) ; (2)
2 The Space of Frames < 4, while the base frame can be chosen as
for 0
We begin by introducing the key concept of the space of possible
frames.
b(^v) = q(arccos(^z v^ ); (^z v^ )=k^z v^ k) : (3)
Suppose at each sample point x(t) of a curve, we are given a We refer hereafter to the frame b(^ v) as the Geodesic Reference
^ (t), computed by whatever method one likes
unit tangent vector, T Frame because it tilts the reference vector ^ z along a geodesic arc
(two-point sampling, five-point sampling, analytic, etc.). Then one until it is aligned with v
^ ; see Figure 4. If v
^ = z ^, there is no prob-
can immediately write down a one-parameter family describing all lem, since we just take b(^ v) to be the quaternion (1; 0); if v^ = ^z, ,
Closed Curve Example. In Figure 6, we show a simple closed
curve, the trefoil knot, the quaternion plot of its periodic Frenet
frame, and, just to show we can do it, the entire constraint surface
in which the Frenet frame and all other possible quaternion fram-
ings of the trefoil must lie. In the next section, we show the results
of optimizing a continuous family of frames lying within this re-
markable surface.
3 Minimal Frames
We have computed a wide selection of examples using the Evolver
of K. Brakke [6] as our optimization tool. The Evolver is a
public-domain, extensively documented system with a huge range
of constraint-solving capabilities, widely used in mathematics and
certain engineering problems. It has a very simple interface for han-
dling parametric constraints like our sliding ring constraints, and
can also handle a wide variety of energy functionals and boundary
specifications. Most of the examples shown here take only a few
seconds to stabilize in the Evolver; more complex geometries will
of course take longer.
Two enhancements to the Evolver handle the specific is-
sues related to quaternion optimization; the symmetry specifi-
Figure 4: Example of the Geodesic Reference Frame: on the cation symmetry_group "central_symmetry" identifies
northern hemisphere of a 2-sphere, the Geodesic Reference
Frame tilts the z^ axis of the north pole’s identity frame along
,
the quaternion q with q if desired during the variation to prevent
reflected double traversals from varying independently, and the sys-
the shortest arc to align with a specified reference direction. tem is able to use the pullback metric on the sphere
X
we may choose any compatible quaternion such as (0; 1; 0; 0). We
ds2 = dx dx r,4 (r2 , x x )
i j i;j i j
i;j
escape the classic difficulty of being unable to assign a global frame
to all of S2 because we need a parameterization of all possible to compute distances directly on the quaternion three-sphere. Com-
frames, not any one particular global frame. If one wants to use putation using this metric, however, is very slow, and so in practice
a reference frame that is not the identity frame, one must premulti- we have used the Euclidean R4 chord approximation, which works
ply b(^v) on the right by a quaternion rotating from the identity into quite well for closely spaced samples and is much faster. (There are
that reference frame; this is important when constructing a nonstan- other choices of three-sphere parameterizations and quaternion dis-
dard Geodesic Reference Frame such as that required to smoothly tance measures that we have not yet attempted that could be more
describe a neighborhood of the southern hemisphere of S2 . efficient still.) The energy functional that we chose to specify for
We can thus write the full family of possible quaternion frames the Evolver (or that would be implemented in a dedicated system) is
keeping v ^ as a fixed element of the frame triad to be the quaternion thus simply the sum of the Euclidean lengths of each line segment
product in R4 : X

f (; v^ ) = q(; v^ ) b(^v) ; (4) d= jx , x j
i j

where denotes quaternion multiplication and all possible frames i;j

are described twice since 0 < 4 . To summarize, if we specify
jj p p
where q = q q = q0 q0 + q1 q1 + q2 q2 + q3 q3 . For surface
^ to be fixed, then the variable in f (; v
a frame axis v ^ ) serves to
areas, the Evolver breaks polygons into triangles, computes their
parameterize a ring in quaternion space, each point of which corre-
areas, and minimizes the total sum as the vertex positions vary.
sponds to a particular 3D frame, and each frame has a diametrically
Our own use of the Evolver required only changing the pa-
opposite twin.
rameter “#define BDRYMAX 20” in skeleton.h to the
desired (large) value corresponding to the number of desired
Surface Patch Example. Figure 5 shows how the frame sliding rings and recompiling. Then, remembering to set
choice problem of Figure 2 can be visualized in the quaternion “space_dimension 4” when working in R4 , one needs in ad-
space of frames. We choose a quaternion projection that shows dition a piece of code similar to the following M ATHEMATICA frag-
only the 3-vector part of the quaternion, dropping q0 . A frame ment to translate Eq. (4) into the boundary constraints for each fixed
choice is achieved by moving a point around the sliding ring con- vector (tangent or normal) and the chosen initial quaternion refer-
straint defined by Eq. (4) to the desired position. The constraint ence frame:
rings in Figure 5 are the generalizations to quaternion space of the
constraint lines symbolized in Figure 3(b). The vertex A admits Do[ring = Qprod[makeQfromVec[vlist[[i]],P1],
a family of frames f (; z
^) that is a circle in quaternion space, but qreflist[[i]]]//Chop;
projects “edge-on” to a vertical bar in our default projection. The Write[file," boundary ",i," parame-
spaces of frames at the other vertices project as ellipses. The outer ters 1"];
ring in Figure 5(b) is touched by two paths, corresponding to the Write[file, "x1: ", CForm[ ring[[2]]]];
clockwise and counterclockwise parallel transport routes in Figure Write[file, "x2: ", CForm[ ring[[3]]]];
5(a); the gap between the intercepts in the outer ring corresponds to Write[file, "x3: ", CForm[ ring[[4]]]];
the inequivalence of the two frames at the bottom vertex of Figure Write[file, "x4: ", CForm[ ring[[1]]]],
5(a). {i,1,Length[vlist]}]
A
B’
B
B C
B
A
B’ A
C’
C’
C
C = C’
B’
(a) (b) (c)
Figure 5: A different viewpoint of the mismatch problem of Figure 2. (a) Choosing different routes to determine the frame at the bottom
point results in the incompatible frames shown here in 3D space. (b) The same information is presented here in the quaternion space-of-
frames picture. We use throughout a quaternion projection that shows only the 3-vector part of the quaternion, dropping q0 ; this is much like
projecting away z in a polar projection of the 2-sphere. Each heavy black curve is a ring of possible frame choices that keep fixed the normals
in (a); the labels mark the point in quaternion space corresponding to the frames at the corners in (a), so the gap between the labels C and C ’
represents the frame mismatch in quaternion space on the same constraint ring. (The apparent vertical line is the result of drawing a squashed
circle of frames at vertex A in this projection.) (c) The method proposed in this paper to resolve this conflict is to fix one point, say A, divide
the polygon ABCB 0 into triangles, and slide B , C , and B 0 along the constraint rings until the total triangle areas are minimized, and some
compromise with C = C 0 is reached.
0.5
-0.5
1
0.5
-1
-1 0
-0.5
0 -0.5
0.5
-1
1
(a) (b) (c)
Figure 6: (a) A trefoil torus knot. (b) Its quaternion Frenet frame projected to 3D. For this trefoil knot, the frame does not close on itself in
quaternion space unless the curve is traversed twice, corresponding to the double-valued “mirror” image of the rotation space that can occur
in the quaternion representation. The longer segments in (b) correspond to the three high-torsion segments observable in (a). (c) The full
constraint space for the trefoil knot is a very complex surface swept out by the constraint rings. All quaternions are projected to 3D using
only the vector part.
Here Qprod and makeQfromVec perform the quaternion
product and create the quaternion corresponding to Eq. (4) with P1
replacing . Note that, since the Evolver displays only the first three
coordinates, we have moved the scalar quaternion to the end; then
the Evolver will display our preferred projection automatically.
With these preliminaries, the Evolver can easily be used to min-
imize the length of the total piecewise linear path among sliding
ring constraints for selected curves, and the total area spanned by
analogous sliding rings for surfaces. We made no attempt to go be-
yond piecewise linear curves. One interesting result is that there
appear to be families of topologically distinct minima: depending
on the conditions imposed, one may find either two disjoint curves
(surfaces), one the q ! ,( q ) image of the other, or a single quater-
nion curve (surface) that contains its own reflection, such as that in Figure 7: Helix (left) and its evolving quaternion frames (right).
Figure 6(b). The families of frame manifolds containing their own Starting from the Geodesic Reference quaternion frame for a sin-
reflected images have minima distinct from the disjoint families. gle turn of the helix, the very dark gray circle, the Evolver pro-
We now present some simple examples to give a feeling for the duces these intermediate steps while minimizing the total quater-
process. nion curve length subject to the constraints in the space of frames.
The final result is the white curve, which is identical to several deci-
mal points with the parallel transport quaternion frame for the same
Minimal Quaternion Frames for Space Curves. The helix helix; note that the quaternion length of the white curve is the short-
provides a good initial example of the procedure we have formu- est, even though in this projection that is not obvious. The numeri-
lated. We know that we can always find an initial framing of a cal energies of the four curves, from dark to light in color, are 3.03,
curve based on the Geodesic Reference algorithm; however, sup- 2.91, 2.82, and 2.66 for the Parallel Transport frame. The individ-
pose we wish to impose minimal length in quaternion space on the ual tubings used to display these curves are in fact created using the
framing we select, and we do not know whether this frame is op- parallel transport frame for each individual curve.
timal with respect to that measure. Then, as illustrated in Figure
7, we can compute the ring constraints on the possible quaternion
frames at each sample point and let the Evolver automatically find
the optimal framing. The results for several stages of this evolution Manifolds. For general manifolds, one must treat patches one
are shown in the Figure; the final configuration is indistinguishable at a time in any event, since global frames may not exist at all.
from the parallel-transport frame, confirming experimentally our Although the locally optimal patches cannot be globally joined to
theoretical expectation that parallel transport produces the minimal one another, we conjecture that some applications might benefit
possible twisting. from the next best thing: matching boundary frames of neighbor-
ing patches using transitional rotations (see, e.g., [18, 10]). We
In Figure 1, we introduced the question of finding an opti-
have carried this out explicitly for simple cases, but omit it here for
mal framing of a particular (3,5) torus knot whose almost-optimal
brevity.
parallel-transport framing was not periodic. In Figure 8, we show
the solution to this problem achieved by clamping the initial and fi-
nal quaternion frames to coincide, then letting the Evolver pick the Extensions to Other Domains. We have focussed for expos-
shortest quaternion path for all the other frames. itory purposes in this paper on frames with intrinsic natural con-
The types of solutions we find are essentially the same for all straints imposed by the tangents to curves and normals to surfaces.
reparameterizations of the curve; regardless of the spacing of the However, the method extends almost trivially to applications in-
sampling, the continuous surface of possible frames is geometri- volving externally specified constraints on frames. Geometric con-
cally the same in quaternion space, so paths that are minimal for struction algorithms based on extrusions reduce to the tubing prob-
one sampling should be approximately identical to paths for any lem. For ordinary camera control interpolation, one could constrain
reasonable sampling. On the other hand, if we want special con- any direction of the camera frame to be fixed by calculating its ap-
ditions for certain parameter values, it is easy to fix any number of propriate constraint ring in the quaternion Gauss map, and then ex-
particular orientations at other points on the curve, just as we fixed tend a method like that of Barr et al. [2, 22]) to smoothly compute
the starting points above; derivative values and smoothness con- intermediate frames subject to the constraints. For more general
straints leading to generalized splines could be similarly specified constrained navigation methods like those described by Hanson and
(see, e.g., Barr et al. [2, 22]). Wernert [13]), the camera vertical direction could be fixed at chosen
points over the entire constraint manifold, and the remaining frame
parameters determined by optimization within the manifold of ring
Surface Patch Framings. A classic simple example of a sur- constraints, possibly subject to fixing entire key-frames at selected
face patch framing problem was presented in the discussion of Fig- locations or boundaries.
ures 2 and 5. This problem can also be handled naturally by the
Evolver: we choose an initial quaternion frame for the mesh and
minimize the area in quaternion space subject to the constraints that 4 Conclusion
the normals remain unchanged. That is, the frame choices may only
slide around constraint rings such as those depicted in Figure 5(b) We have introduced a general framework derived from the quater-
for the frames at the corners. The results are shown in Figures 10 nion Gauss map for studying and selecting appropriate families of
and 9. As a test, we started one case in a random initial state with coordinate frames for curves and surface patches in 3D space. Min-
a range of 2 in the starting values. All converged to the same op- imizing length for quaternion curve maps and area for surfaces is
timal final framing. While more complex examples could be given, proposed as the appropriate generalization of parallel transport for
all the essential features of the method short of dealing rigorously the selection of optimal frame fields. These smooth frames can
with non-trivial topological manifolds are illustrated by this surface be used to generate tubular surfaces based on the space curves,
patch example. thus allowing their effective display on polygon-based graphics en-
gines with texturing. The analogous results for surface patches al- [11] A. J. Hanson and H. Ma. Quaternion frame approach
low the selection of optimal local coordinate systems that may be to streamline visualization. IEEE Trans. on Visualiz. and
adapted for display purposes and related applications such as tex- Comp. Graphics, 1(2):164–174, June 1995.
turing based on oriented particle systems. Our principal new tool
is the space of all possible frames, a manifold of constraints im- [12] A. J. Hanson and Hui Ma. Visualizing flow with quaternion
mersed in the space of quaternion frames. By defining energies frames. In Proceedings of Visualization ’94, pages 108–115.
and boundary conditions in this space one can produce a rich va- IEEE Computer Society Press, 1994.
riety of application-adapted criteria for specifying optimal families
of frames. Work remaining to be done in the future includes apply- [13] A. J. Hanson and E. Wernert. Constrained 3d navigation with
ing the method to other domains such as geometric modeling and 2d controllers. In Proceedings of Visualization ’97, pages
viewpoint interpolation, studying more carefully the topologically 175–182. IEEE Computer Society Press, 1997.
distinct minimal quaternion area solutions found for certain surface [14] J. T. Kajiya. Anisotropic reflection models. In Computer
framings, and studying more challenging problems in the surface Graphics, volume 19, pages 15–21, 1985. Proceedings of
domain, e.g., topological tori with various numerical bumps and SIGGRAPH ’85.
deformations are known to admit global frames, but little is known
about how to compute good ones, and this method is a logical can- [15] Myoung-Jun Kim, Myung-Soo Kim, and Sung Yong Shin. A
didate. general construction scheme for unit quaternion curves with
simple high order derivatives. In Computer Graphics Pro-
ceedings, Annual Conference Series, pages 369–376, 1995.
Acknowledgments Proceedings of SIGGRAPH ’95.
The author gratefully acknowledges the cordial hospitality of [16] F.. Klock. Two moving coordinate frames for sweeping along
Claude Puech and the members of the iMAGIS laboratory, a a 3d trajectory. Computer Aided Geometric Design, 3, 1986.
joint project of CNRS, INRIA, Institut National Polytechnique de
Grenoble, and Université Joseph Fourier, where this research was [17] N. L. Max. Computer representation of molecular surfaces.
initiated. Thanks are also due to Ken Brakke for help with the IEEE Computer Graphics and Applications, 3(5):21–29, Aug
Evolver. This research was made possible in part by NSF infras- 1983.
tructure grant CDA 93-03189.
[18] J. Milnor. Topology from the Differentiable Viewpoint. The
University Press of Virginia, Charlottesville, 1965.
References
[19] G. M. Nielson. Smooth interpolation of orientations. In N.M.
[1] S. L. Altmann. Rotations, Quaternions, and Double Groups. Thalman and D. Thalman, editors, Computer Animation ’93,
Oxford University Press, 1986. pages 75–93, Tokyo, June 1993. Springer-Verlag.
[2] Alan H. Barr, Bena Currin, Steven Gabriel, and John F. [20] John C. Platt and Alan H. Barr. Constraint methods for flex-
Hughes. Smooth interpolation of orientations with angular ve- ible models. In John Dill, editor, Computer Graphics (SIG-
locity constraints using quaternions. In Edwin E. Catmull, ed- GRAPH ’88 Proceedings), volume 22, pages 279–288, Au-
itor, Computer Graphics (SIGGRAPH ’92 Proceedings), vol- gust 1988.
ume 26, pages 313–320, July 1992.
[21] D. Pletincks. Quaternion calculus as a basic tool in computer
[3] M. Berger. Geometry I, II. Springer Verlag, Berlin, 1987. graphics. The Visual Computer, 5(1):2–13, 1989.
[4] Richard L. Bishop. There is more than one way to frame a [22] Ravi Ramamoorthi and Alan H. Barr. Fast construction of
curve. Amer. Math. Monthly, 82(3):246–251, March 1975. accurate quaternion splines. In Turner Whitted, editor, SIG-
GRAPH 97 Conference Proceedings, Annual Conference Se-
[5] Jules Bloomenthal. Calculation of reference frames along a
ries, pages 287–292. ACM SIGGRAPH, Addison Wesley,
space curve. In Andrew Glassner, editor, Graphics Gems,
August 1997. ISBN 0-89791-896-7.
pages 567–571. Academic Press, Cambridge, MA, 1990.
[6] Kenneth A. Brakke. The surface evolver. Experimental Math- [23] John Schlag. Using geometric constructions to interpolate ori-
ematics, 1(2):141–165, 1992. The “Evolver” system, manual, entation with quaternions. In James Arvo, editor, Graphics
and sample data files are available by anonymous ftp from Gems II, pages 377–380. Academic Press, 1991.
geom.umn.edu, The Geometry Center, Minneapolis MN. [24] Uri Shani and Dana H. Ballard. Splines as embeddings for
[7] L. P. Eisenhart. A Treatise on the Differential Geometry of generalized cylinders. Computer Vision, Graphics, and Image
Curves and Surfaces. Dover, New York, 1909 (1960). Processing, 27:129–156, 1984.
[8] S. Gabriel and James T. Kajiya. Spline interpolation in curved [25] K. Shoemake. Animating rotation with quaternion curves. In
space. In State of the Art Image Synthesis, 1985. Siggraph ’85 Computer Graphics, volume 19, pages 245–254, 1985. Pro-
Course notes. ceedings of SIGGRAPH 1985.
[9] Alfred Gray. Modern Differential Geometry of Curves and [26] Ken Shoemake. Fiber bundle twist reduction. In Paul Heck-
Surfaces with Mathematica. CRC Press, Inc., Boca Raton, bert, editor, Graphics Gems IV, pages 230–236. Academic
FL, second edition, 1998. Press, 1994.
[10] Cindy M. Grimm and John F. Hughes. Modeling surfaces

with arbitrary topology using manifolds. In Computer Graph-
ics Proceedings, Annual Conference Series, pages 359–368,
0 0.5 1 0 0.5 1
-1 -0.5 -1 -0.5
1 1
0.5
5 0.5
5
A
K
0 0
A AA
K
A A
-0.5
.5 -0.5
.5
KA KA
A A A A
A A A A
-1 -1
1 1
A A
0.5 0.5
0 0
-0.5 -0.5
A A
-1 -1
(a) (b) (c) (d)
Figure 8: Optimization of the non-periodic parallel transport frame of the (3,5) torus knot introduced in Figure 1 to produce a nearby periodic
framing. (a) The original quaternion parallel transport frame used to produce the tubing in Figure 1(b,c). (b) The frame mismatch, repeated
for completeness. (c) The result of fixing the final frame to coincide with the initial frame, leaving the other frames free to move on the
constraint rings, and minimizing the resulting total length in quaternion space. The length of the original curve was 13.777 and that of the
final was 13.700, not a large difference, but noticeable enough in the tube and the quaternion space plot. (d) Closeup of the corresponding
framing of the knot in ordinary 3D space, showing that the mismatch problem has been successfully resolved. This tube can now be textured,
since the frames match exactly.
(a) (b) (c) (d)
Figure 9: Study of possible and optimal reference frames on a surface patch; the corresponding quaternion fields are given in Figure 10. (a)
The Geodesic Reference frames for the small patch of Figure 2. (b) Two-step parallel transport frames. (c) Random frames. (d) The unique
frame configuration resulting from minimizing area in quaternion space with the upper corner fixed.
0.4 0.4 0.4

-0.2
0.5 0
0.2 0.2 0.2
0.4
0.2
0.25
0
0 0 -0.25
.25 0
1
0.2 0.2 0.2
0.1 0.1 0.5 0.1
0 0 0
-0.1 -0.1 0 -0.1
-0.2 -0.2 -0.2

-0.5
0 0 0
0.2 0.2 0.2
0.4 0.4 0.4
(a) (b) (c) (d)
Figure 10: Quaternion areas corresponding to the frame assignments in Figure 9. (a) The initial Geodesic Reference quaternions for the small
patch shown in Figure 2. (b) Initial quaternions from parallel transporting the vertex frame down one edge, and then across line by line. (c)
A random starting configuration with the single same fixed corner point as (a) and (b) and a range of to + relative to the Geodesic ,
Reference frame. (d) The result of minimization of the quaternion area is the same for all starting points. The relative areas are: 0.147, 0.154,
0.296, and 0.141, respectively. Thus the Geodesic Reference is very close to optimal, but is distinct.
TECHNICAL REPORT NO. 518
Quaternion Gauss Maps

and
Optimal Framings of Curves and Surfaces
by
Andrew J. Hanson
October 1998
COMPUTER SCIENCE DEPARTMENT

INDIANA UNIVERSITY
Bloomington, Indiana 47405-4101
Quaternion Gauss Maps
and Optimal Framings of Curves and Surfaces
Andrew J. Hanson
Department of Computer Science
Lindley Hall 215
Indiana University
October 1998
Abstract
We propose a general paradigm for generating optimal coordinate frame fields that may be ex-
ploited to annotate and display curves and surfaces. Parallel-transport framings, which work well
for open curves, generally fail to have desirable properties for cyclic curves and for surfaces. We
suggest that minimal quaternion measure provides an appropriate generalization of parallel trans-
port. Our fundamental tool is the “quaternion Gauss map,” a generalization to quaternion space
of the tangent map for curves and of the Gauss map for surfaces. The quaternion Gauss map
takes 3D coordinate frame fields for curves and surfaces into corresponding curves and surfaces
constrained to the space of possible orientations in quaternion space. Standard optimization tools
provide application-specific means of choosing optimal, e.g., length- or area-minimizing, quater-
nion frame fields in this constrained space. We observe that some structures may have distinct
classes of minimal quaternion framings, e.g, one disconnected from its quaternion reflection, and
another that continuously includes its own quaternion reflection. We suggest an effective method
for visualizing the geometry of quaternion maps that is used throughout. Quaternion derivations
of the general moving-frame equations for both curves and surfaces are given; these equations are
the quaternion analogs of the Frenet and Weingarten equations, respectively. We present examples
of results of the suggested optimization procedures and the corresponding tubings of space curves
and sets of frames for surfaces and surface patches.
1 Introduction
We propose a general framework for selecting optimal systems of coordinate frames that can be
applied to the study of geometric structures such as curves and surfaces in three-dimensional space.
The methods contain “minimal-turning” parallel-transport framings of curves as a special case,
1
A
K
A
A
A
A
(a) (b) (c)
Figure 1: The (3,5) torus knot, a complex periodic 3D curve. (a) The line drawing is nearly
useless as a 3D representation. (b) A tubing based on parallel transporting an initial reference
frame produces an informative visualization, but is not periodic. (c) The arrow in this closeup
exposes the subtle but crucial non-periodic mismatch between the starting and ending parallel-
transport frames; this would invalidate any attempt to texture the tube. The methods of this
paper provide robust parameterization-invariant principles for resolving such problems.
and extend naturally to situations where parallel-transport is not applicable. This article presents
additional details of the IEEE Visualization ’98 paper by the author [16].
Motivation. Many graphics problems require techniques for effectively displaying the properties
of curves and surfaces. The problem of finding appropriate representations can be quite challeng-
ing. Representations of space curves based on single lines are often inadequate for graphics pur-
poses; significantly better images result from choosing a “tubing” to display the curve as a graph-
ics object with spatial extent. Vanishing curvature invalidates methods such as the Frenet frame,
and alternative approaches such as those based on parallel transport involve arbitrary heuristics
to achieve such properties as periodicity. Similar problems occur in the construction of suitable
visualizations of complex surfaces and oriented particle systems on surfaces. If a surface patch
is represented by a rectangular but nonorthogonal mesh, for example, there is no obvious local
orthonormal frame assignment; if the surface has regions of vanishing curvature, methods based
on directions of principal curvatures break down as well.
While we emphasize curves and surfaces in this paper to provide intuitive examples, there are
several parallel problem domains that can be addressed with identical techniques. Among these are
extrusion methods and generalized cones in geometric modeling, the imposition of constraints on
a camera-frame axis in key-frame animation, and the selection of a 2D array of camera-frame axis
choices as a condition on a constrained-navigation environment (see, e.g., Hanson and Wernert
[20]).
Figure 1 summarizes the basic class of problems involving curves that will concern us here.
The line drawing (a) of a (3,5) torus knot provides no useful information about the 3D structure.
2
(a) (b) (c)
Figure 2: (a) A smooth 3D surface patch having a non-orthogonal parameterization, along with
its geometrically-fixed normals at the four corners. No unique orthonormal frame is derivable
from the parameterization. If we imitate parallel transport for curves to evolve the initial frame
at the top corner to choose the frame at the bottom corner, we find that paths (b) and (c) re-
sult in incompatible final frames at the bottom corner. This paper addresses the problem of
systematically choosing a compatible set of surface frames in situations like this.
Improving the visualization by creating a tubing involves a subtle dilemma that we attempt to ex-
pose in the rest of the figure. We cannot use a periodic Frenet frame as a basis for this tubing
because inflection points or near-inflection points occur for many nice-looking torus knot parame-
terizations, and in such cases the Frenet frame is undefined or twists wildly. The parallel-transport
tubing shown in (b) is well-behaved but not periodic; by looking carefully at the magnified portion
next to the arrow in Figure 1(c), one can see a gross mismatch in the tessellation due to the non-
periodicity which would, for example, preclude the assignment of a consistent texture. While it
would be possible in many applications to ignore this mismatch, it has been the subject of a wide
variety of previous papers (see, e.g., [24, 36, 5]), and must obviously be repaired for many other
applications such as those requiring textured periodic tubes.
Figure 2 illustrates a corresponding problem for surface patches. While the normals to the four
corners of the patch are always well-defined (a), one finds two different frames for the bottom
corner depending upon whether one parallel transports the initial frame around the left-hand path
(b) or the right-hand path (c). There is no immediately obvious right way to choose a family of
frames covering this surface patch.
Our goal is to propose a systematic family of optimization methods for resolving problems
such as these.
Methodology. We focus on unit quaternion representations of coordinate frames because of the

well-known natural structure of unit quaternions as points on the three-sphere S3 , which admits
a natural distance measure for defining optimization problems, and supports in addition a variety
of regular frame-interpolation methods (see, e.g., [37, 35, 31, 23]). We do not address the related
3
(a) (b) (c)
Figure 3: (a) The camera frame interpolation problem is analogous to the problem of finding a
minimal-bending spline curve through a series of fixed key points. (b) The optimal curve frame
assignment problem is analogous to fixing the end points of a curve segment and choosing in
addition a family of lines along which the intermediate points are constrained to slide during the
optimization process; in 3D, the spline path need not pass through the constraint lines. (c) In
this paper, our sample points are generally close enough together that we apply the constraints to
piecewise linear splines analogous to those shown here.
question of optimal freely moving frames treated by the minimal-tangential-acceleration methods

(see, e.g., [2, 34, 11]); we are instead concerned with closely-spaced points on curves and sur-
faces where one direction of the frame is already fixed, and the chosen functional minimization
in quaternion space must obey the additional constraint imposed by the fixed family of directions.
Additional references of interest, especially regarding the treatment of surfaces, include [22, 32].
Figure 3 provides a visualization of the difference between the general interpolation problem and
our constrained problem: a typical spline minimizes the bending energy specified by the chosen
anchor points; requiring intermediate points to slide on constrained paths during the minimization
modifies the problem. In particular, 3D spline curves need not intersect any of the constraint paths.
In addition, we note that we typically have already sampled our curves and surfaces as finely as we
need, so that piecewise linear splines are generally sufficient for the applications we discuss.
Our solution to the problem is to transform the intrinsic geometric quantities such as the tangent
field of a curve and the normal field of a surface to quaternion space and to construct the quater-
nion manifold corresponding to the one remaining degree of rotational freedom in the choice of
coordinate frame at each point. Curves and surfaces in these spaces of possible frames correspond
to specific choices of the quaternion Gauss map, a subspace of the space of possible quaternion
frames of the object to be visualized. Mathematically speaking, the space of possible frames is
the circular Hopf fiber lying above the point in S2 corresponding to each specific curve tangent or
surface normal (see, e.g., [39, 3]).
For space curves, specifying a frame assignment as a quaternion path leads at once to tubular
surfaces that provide a “thickened” representation of the curve that interacts well with lighting and
rendering models. For surface patches, the approach results in a structure equivalent to that of
4
an anisotropic oriented particle system whose pairs of tangent vector fields in the surface produce
natural flow fields that characterize the local surface properties and are easy to display. We will
see that certain complex features of surfaces that are well-known in manifold theory arise naturally
and can be clearly visualized using the quaternion Gauss map.
In the course of the discussion, we introduce a useful method of visualizing the geometry of
the space of quaternions in which quaternion Gauss maps and the spaces of possible quaternion
frames are represented. We show how to compute the required subspaces of frames in practice,
and how to express this information in a form that can be used to optimize an energy measure,
thereby leading to optimal frame choices. We also outline in the appendix a treatment of the
curve and surface frame differential equations expressed directly in quaternion coordinates using
the quaternion Lie algebra; these methods expose essential fundamental features of the quaternion
frame methodology that are analogous to the Frenet and Weingarten equations.
Parallel Transport and Minimal Measure. Constraining each quaternion point (a frame) to its
own circular quaternion path (the axial degree of rotational freedom), we then minimize the quater-
nion length of the frame assignment for curves and the quaternion area of the frame assignment
for surfaces to achieve an optimal frame choice; this choice reduces to the parallel-transport frame
for simple cases. Our justification for choosing minimal quaternion length for curves is that there
is a unique rotation in the plane of two neighboring tangents that takes each tangent direction to
its next neighbor along a curve: this is the geodesic arc connecting the two frames in quaternion
space, and is therefore the minimum distance between the quaternion points representing the two
frames. The choice of minimal area for surface frames is more heuristic, basically a plausibility
argument that the generalization of minimal length is minimal area; no doubt this could be made
more rigorous.
By imposing other criteria such as endpoint derivative values and minimal bending energy (see
Barr et al. [2, 34]), the short straight line segments and polygons that result from the simplest
minimization could be smoothed to become generalized splines passing through the required con-
straint rings; since, in practice, our curve and surface samplings are arbitrarily dense, this was not
pursued in the current investigation.
For space curves, specifying a frame assignment as a quaternion path leads at once to tubular
surfaces that provide a “thickened” representation of the curve that interacts well with texturing,
lighting, and rendering models. For surface patches, the approach results in a structure equivalent
to that of an anisotropic oriented particle system (also a species of texture) whose pairs of tangent
vector fields in the surface produce natural flow fields that characterize the local surface properties
and are easy to display.
Background. General questions involving the specification of curve framings have been inves-
tigated in many contexts; for a representative selection of approaches, see, e.g., [24, 36, 5, 28].
The quaternion Gauss map is a logical extension of the quaternion frame approach to visualizing
space curves introduced by Hanson and Ma [19, 18]. The formulation of the quaternion form of
the differential equations for frame evolution was introduced as early as the 1890’s by Tait [41].
5
^
N ^
^ T
T h
^
g N
f ^
T
^ a e ^
N d c N
^
T
b ^
^ T
N
Figure 4: A smooth 2D curve with its normal and tangent frame fields. The segments d and
f are intended to be straight.
For basic information on orientation spaces and their relationship to quaternions, see, e.g., [1,
33, 25]. Our own conventions are summarized in the appendix. The task of visualizing quaternions
is also important, and we will describe our own approach below; for an interesting alternative, see
Hart, Francis, and Kauffman [21]. Additional background on the differential geometry of curves
and surfaces may be found in sources such as the classical treatise of Eisenhart [8] and in Gray’s
M ATHEMATICA-based text [12], which inspired a number of the illustrations in this paper.
2 The Differential Geometry of Coordinate Frames

Our first goal is to define moving coordinate frames that are attached to curves and surfaces in
3D space. We will assume that our curves and surfaces are defined in practice by a discrete set of
sample points connected by straight line segments, so that numerical derivatives can be defined at
each point if analytic derivatives are not available. We begin with a pedagogical presentation of
the properties of 2D curves, and then extend the surprisingly rich concepts that arise to 3D curves
and surfaces.
2.1 Orientation Maps of 2D Curves

Suppose we have a smooth, arbitrarily differentiable 2D curve x(t) = x(t)^ x + y (t)^y. The curve
itself generates a continuous set of changing tangents and normals of the form
T(t) = dx(t)=dt = x0 x^ + y0y^ (1)
N(t) = y0x^ , x0y^ : (2)
We choose this relative orientation convention so that in any dimension the tangent vector is ex-
pressible as the positive-signed cross-product of the normal(s); see [15] for further details. Unit
6
d g
h e
c
a ^
N
g ec d ^
T b f
h
f
b a
(a) (b)
Figure 5: 2D Gauss map sketches of (a) the normal directions and (b) the tangent directions
corresponding to the U-shaped curve in Figure 4. All these vectors lie on the unit circle in 2D.
The straight line segments along d and f in Figure 4 correspond to single points in both maps.
^ = v=kvk, so the
length vectors will hereafter be distinguished with the conventional notation v
normalized tangent and normal directions are denoted by T ^ and N^.
In Figure 4, we show an example of a 2D curve with its tangent and normal fields. The nor-
malized tangent and normal fields have only one degree of freedom, which we denote by the angle
(t); the column vectors N^ and T^ then represent a moving orthonormal coordinate frame that may
be expressed in the form
h i " #
N
^ T
^
cos , sin :
sin cos
= (3)
We may derive a 2D version of the frame equations by differentiating the frame to get
^ 0 (t)
N = +vT^ (4)
^ 0 (t)
T = ,vN^ ; (5)
where (t) is the curvature and v (t) is the “velocity” relative to the infinitesimal measure of curve
length ds2 = dx(t) dx(t), that is, d(t)=dt = (ds=dt)(d(s)=ds) = v (t)(t).
Note: we will find sign choices to be a subtle exercise throughout this paper. In Figure 4, the
fact that the normal N
^ is chosen to point to the outside of a curve encircling an enclosed area in
the right-hand sense makes the system inequivalent to the Frenet frame of the corresponding 3D
curve, which would have N ^ pointing inwards everywhere except around the point e, and would be
undefined along the straight segments d and f .
7
2D Tangent Map and Gauss Map. A 2D version of the Gauss map [8, 12] used in the classical
differential geometry of surfaces follows when we discard the original curve in Figure 4 and restrict
our view to show only the path of the normalized normals, as in Figure 5(a), or the normalized
tangents, as in Figure 5(b); both vector fields take values only in the unit circle. We note that any
sufficiently small open neighborhood of the curve has unique tangent and normal directions, up
to the possibility of a shared limit point for straight segments such as d and f in Figure 4; over
the whole curve, however, particular neighborhoods of directions may be repeated many times,
resulting in an overlapping, non-unique 2D map, as indicated schematically in Figure 5. We will
accept this as a feature, not necessarily a deficiency, of the construction.
2D “Quaternions.” In the appendix, we present the details of a derivation of a quaternion-like

approach to the representation of 2D frames that may be informative to some readers. A brief
summary begins by noting that the normal and tangent vectors can be parameterized by a quadratic
form in the two variables a and b as
h i " cos , sin # " a , b ,2ab #
2 2
N T a ,b ;
^ ^ =
sin cos 2ab
= 2 2 (6)
where imposing the constraint a2 + b2 = 1 guarantees orthonormality of the frame.

By taking derivatives and extracting common factors, we find that the single matrix equation
" # " # " #
a0 1
= v (t)
, a
0
b0 2 + 0 b (7)
in the two variables with one constraint contains both the frame equations T^ 0 = ,vN ^ and N^0 =
+vT ^ . The N^ 0 and T
^ 0 equations are superficially a more complex set of two vector equations
in four variables with three constraints. Equation (7) is effectively the square root of the frame
equations. Rotations may be realized as complex multiplication in (a + ib), and the pair (a; b) =
(cos(=2); sin(=2)) parameterizes any rotation. Since (a; b) (,a; ,b), the variables give a
double covering of the space of rotations if we take the angular range from 0 ! 4 instead of 2 .
These are precisely the properties we expect of quaternion representations of rotations.
2.2 3D Space Curves

We now move on to three-dimensional space curves. The fundamental difference in 3D is that,
while the tangent direction is still determinable directly from the space curve, there is an additional
degree of rotational freedom in the normal plane, the portion of the frame perpendicular to the
tangent vector. This is indicated schematically in Figure 6.
Tangent Map. The tangent direction of a 3D curve at each point is given simply by taking the
algebraic or numerical derivative of the curve at each sample point and normalizing the result. Each
tangent direction thus has two degrees of freedom and lies on the surface of the two-sphere S2 . The
curve resulting from joining the ends of neighboring tangents is the tangent map of the curve. As
8
^
T
^
N2
x(t)
^
N1
Figure 6: General form of a moving frame for a 3D curve x(t), with the tangent direction T ^
determined directly from the curve derivative, and the exact orientation of the basis (N1 ; N2 ) for
^ ^
the normal plane determined only up to an axial rotation about T ^.
in the 2D case treated above, the tangent map of a 3D curve is not necessarily single-valued except
in local neighborhoods, and may have limit points (e.g., if there are straight segments). In Figure
7(a,b), we show examples of two classic 3D curves, one a closed knot, the (2,3) trefoil knot lying
on the surface of a torus, and the other the open helix:
xtorus(p; q )(a; b; c)(t) = a b cos(qt)) cos(pt) x^ + (a + b cos(qt)) sin(pt) y^ + c sin(qt) ^z

( +
xhelix (a; b; c)(t) = a cos(t) x^ + b sin(t) y^ + ct ^z :
Differentiating these curves yields the tangent maps in Figure 7(c).
General Form of Curve Framings in 3D. The evolution properties of all possible frames for
a 3D curve x(t) can be written in a unified framework. The basic idea is to consider an arbitrary
frame to be represented in the form of columns of a 3 3 orthonormal rotation matrix,
h i
Curve Frame = N
^ N
1
^ T
2
^ : (8)
Here T ^ (t) = x0 (t)=kx0 (t)k is the normalized tangent vector determined directly by the curve
geometry, and which is thus unalterable; (N^ (t); N^ (t)) is a pair of orthonormal vectors spanning
1 2
the plane perpendicular to the tangent vector at each point of the curve. Since kT ^ k2 = kN^ k2 =
1
kN^ 2k = 1 and all other inner products vanish by definition, any change in a basis vector must be
2
orthogonal to itself and thereby expressible in terms of the other two basis vectors. Thus the most
general possible form for the frame evolution equations is
2 ^0 3 2 32 ^ (t) 3
N (t) +kz (t) ,ky (t) N
66 ^ 0 77 6 0
76 7
4 N0 (t) 5 = v(t) 4,kz (t) 0 +kx(t)564 N^ (t) 75 ;
1 1
2 (9)
+ky (t) ,kx (t)
2
^ (t)
T 0 ^ (t )
T
where v (t) = kx0 (t)k is the velocity of the curve if we are not using a unit speed parameterization.
9
(a) (b) (c)
Figure 7: Tangent maps. (a) The (2,3) torus knot and the helix as 3D line drawings. (b) Illus-
trating an application of tubing to make the 3D curves more interpretable. (c) The corresponding
normalized tangent maps determined directly from the curve geometry. These are curves on the
two-sphere, and have also been tubed to improve visibility.
10
The particular choice of notation and signs for the curvatures k in Eq. (9) is compellingly
motivated by the quaternion Lie algebra treatment in the appendix, and its natural properties are
also exposed using the Darboux form of the equations,
^0
N = v(t) F N^
1 1
N02
^ = v(t) F N^2 (10)
^0
T = v(t) F T^ ;
where F generalizes the Darboux vector field (see, e.g., Gray [12], p. 205):
F = kx N
^ +k N
1 y 2 + kz T :
^ ^ (11)
The square magnitude of the total “force” acting on the frame is kFk2 = kx2 + ky2 + kz2 , and we will
see below that this is a minimum for the parallel-transport frame.
The arbitrariness of the basis (N^ (t); N^ (t)) for the plane perpendicular to T ^ (t) can be ex-
1 2
ploited as desired to eliminate any one of the (kx; ky ; kz ) (see, e.g., [4]). For example, if
M
^ = ^ cos , N
N ^ sin
1 1 2
M2
^ = N1 sin + N2 cos ;
^ ^ (12)
differentiating and substituting Eq. (9) yields

^0
M = ^ (k , 0 ) , T
M ^ (k sin + k cos ) (13)
1 2 z x y
M02
^ = ,M1 (kz , ) + T(kx cos , ky sin ) :
^ 0 ^ (14)
Thus the angle (t) may be chosen to cancel the angular velocity kz in the (N
^ (t); N
1
^ (t)) plane.
2
The same argument holds for any other pair. Attempting to eliminate additional components pro-
duces new mixing, leaving at least two independent components in the evolution matrix.
Tubing. For completeness, we note that to generate a ribbon or tube such as those used to display
curves throughout this paper, one simply sweeps the chosen set of frames through each curve point
p(t) to produce a connected tube,
x(t; ) = p(t) + cos N
^ (t) + sin N
1
^ (t) :
2
The resulting structure is sampled in t and over one full 2 period in to produce a tessellated tube.
Arbitrary functions of (t; ) can be introduced instead of the cosine and sine to produce ribbons
and general linear structures.
Classical Frames. We now note a variety of approaches to assigning frames to an entire 3D

space curve, each with its own peculiar advantages. Figure 8 compares the tubings of the (2,3)
trefoil knot and the helix for each of the three frames described below.
11
(a) (b) (c)
Figure 8: Curve framings for the (2,3) torus knot and the helix based on (a) Frenet frame, (b)
Geodesic Reference frame (minimal tilt from North pole), and (c) Parallel Transport frame, which
is not periodic like the other frames.
12
Frenet-Serret Frame. This classical frame is determined by local conditions at each point of
the curve, but is undefined whenever the curvature vanishes (e.g., when the curve straightens
out or has an inflection point). For the Frenet frame, kx = 0, ky is the inverse radius of
curvature, i.e., the curvature (t), and kz (t) is the torsion (t), which mixes the two normal
vectors in their local plane. This choice produces the usual equations
2 ^0 3 2 32 T^ (t) 3
T (t) (t)
66 ^ 0 77 6 0 0
76 7
4 N0 (t) 5 = v(t) 4,(t) 0 (t)564 N^ (t) 75 : (15)
^ (t)
B 0 , (t) 0 B^ (t)
Note that the squared Darboux vector is thus kFk2 = 2
+ .
2 2
If x(t) is any thrice-differentiable space curve, we can identify the triad of normalized Frenet
frame vectors directly with the local derivatives of the curve,
^ (t)
x0 (t)
T =
kx0(t)k
N ^ (t)
^ =N = ^ (t) T
B ^ (t) (16)
1
^ (t)
x0 (t) x00 (t)
N
^ =B
2 =
kx0(t) x00(t)k ;
with = kx0 (t) x00 (t)k=kx0 (t)k3 , = x0 (t) x00 (t) x000 (t)=kx0 (t) x00 (t)k2 . For further
details, see [8, 12].
Parallel-Transport Frame. This frame is equivalent to a heuristic approach that has been
frequently used in graphics applications (see, e.g., [24, 36, 5, 28]). A careful mathemati-
cal treatment by Bishop [4] presents its differential properties in a form that can be easily
compared with the standard features of the Frenet frame. The parallel transport frame is
distinguished by the fact that it uses the smallest possible rotation at each curve sample to
align the current tangent vector with the next tangent vector. The current orientation of
the plane normal to the tangent vector depends on the history of the curve, starting with
an arbitrary initial frame, and so one is essentially integrating a differential equation for
the frame change around the curve. The frame depends on the initial conditions, and can-
not be determined locally on the curve like the Frenet frame. The algorithm with the best
limiting properties [27] for computing this frame involves determining the normal direction
^ = T T
N i i+1 =kTi Ti+1 k to the plane of two successive tangents to the curve, finding
the angle = arccos(T ^ T^
i i+1 ), and rotating the current frame to the next frame using the
3 3 matrix R(; N) or its corresponding quaternion (see appendix)
^
q(; N^ ) = q(arccos(T^ i T^ i ); Ti Ti =kTi Ti k) :

+1 +1 +1 (17)
If the successive tangents are collinear, one leaves the frame unchanged; if the tangents are
anti-collinear, a result can be returned, but it is not uniquely determined.
13
To identify the parallel transport frame with Eq. (9), we set ky ) k1 , ,kx ) k2 , and kz = 0
to avoid unnecessary mixing between the normal components (effectively the definition of
parallel transport); this choice produces Bishop’s frame equations,
2 ^0 3 2 32 ^ (t) 3
N (t)
66 ^ 0 77 , k (t) N
,k (t)75664 N^ (t) 775 :
0 0
1
6 1 1
4 N0 (t) 5 = v(t) 4 0
2
0 2 2 (18)
^ (t)
T k (t) k (t) 0
1 2 ^ (t)
T
Since kT
^ 0 k2 = (k )2 + (k )2 is an invariant independent of the choice of the normal frame,
1 2
Bishop identifies the curvature, orientation, and angular velocity
=
(t) = ( k) 2
+(k) 2
1 2
1
! 2
(t) = arctan
k 2
k 1
!(t) =
d(t) :
dt
k 1 and k2 thus correspond to a Cartesian coordinate system for the “curvature polar coordi-
R
nates” (; ) with = 0 + ! (t) dt; ! (t) is effectively the classical torsion (t) appearing in
the Frenet equations. Note that the squared Darboux vector kFk2 = kT ^ 0 k2 = k 2 + k 2 = 2
1 2
is now a frame invariant. It is missing the torsion component present for the Frenet frame,
and thus assumes its minimal value.
Geodesic Reference Frame. In this paper, we will often need a frame that is guaranteed to
have a particular axis in one direction, but we will not care about the remaining axes because
they will be considered as a space of possibilities. A convenient frame with these properties
can always be constructed starting from the assumption that there exists a canonical reference
frame in which, say, the ^z axis corresponds to the preferred direction. Thus if v ^ is the
desired direction of the new axis, we can simply tilt the reference axis ^z into v ^ along a
minimal, geodesic curve using an ordinary rotation R(; n ^ ) or its corresponding quaternion
(see appendix):
q(; n^ ) = q(arccos(^z v^ ); ^z v^ =kz^ v^ k) : (19)
Clearly any reference frame, including frames related to the viewing parameters of a moving
observer, could be used instead of ^z. This frame has the drawback that it is ambiguous
whenever v ^ = ,z ^; sequences of frames passing through this point will not necessarily be
smoothly varying since only a single instance of a one-parameter family of frames can be
returned automatically by a context-free algorithm. Luckily, this is of no consequence for our
application. As we will discuss later in the quaternion framework, this property is directly
related to the absence of a global vector field on the two-sphere.
General Frames. We will henceforth work with the general framework for coordinate
frames of arbitrary generality, rather than choosing conventional frames or hybrids of the
14
frames described so far (see, e.g., Klock [24]). While the classical frames have many funda-
mentally appealing mathematical properties, we are not in fact restricted to use any one of
them. Keeping the tangent vector field intact, we may modify the angle of rotation about the
tangent vector at will to produce an application-dependent frame assignment. An example
of such an application is a closed curve with inflection points: the Frenet frame is periodic
but not globally defined, the parallel transport frame will not be periodic in general, and
the Geodesic Reference frame will be periodic but may have discontinuities for antipodal
orientations. Thus, to get a satisfactory smooth global frame, we need something close to
a parallel transport frame but with a periodic boundary condition; an example of an ad hoc
solution is to take the Parallel Transport frame and impose periodicity by adding to each ver-
tex’s axial rotation a fraction of the angular deficit of the parallel transport frame after one
circuit. But this is highly heuristic and depends strongly on the chosen parameterization. In
the following sections, we introduce a more comprehensive approach.
2.3 3D Surfaces
If we are given a surface patch x(u; v ) with some set of non-degenerate coordinates (u; v ), we may
determine the normals at each point by computing
N(u; v ) = xu xv ; (20)
where xu = @ x=@u and xv = @ x=@v . For surfaces defined numerically in terms of vertices
and triangles, we would choose a standard procedure such as averaging the normals of the faces
surrounding each vertex to determine the vertex normal. Alternatively, if we have an implicit
surface described by the level-set function f (x) = 0, the normals may be computed directly from
the gradient at any point x satisfying the level set equation:
N(x) = rf (x) :
The normalized normal is defined as usual by N ^ = N=kNk.
For 3D curves, the geometry of the curve determined the tangent vector T ^ and left a pair of
normal vectors (N1 ; N2 ) with one extra degree of freedom to be determined in the total frame
^ ^
[N
^ N
1
^ T
2
^ ]. The analogous observation for surfaces is that the geometry fixes the normal at each
^ ;T
surface point, leaving a pair of tangent vectors (T ^ ) with one extra degree of freedom to be
1 2
determined in the total surface frame,
h i
Surface Frame = T
^ T
1
^ N
2
^ : (21)
When a (u; v ) surface parameterizationhis available,

i the surface partial derivatives xu and xv
can in principle be used to assign a frame T1 T2 N (using Gram-Schmidt if xu xv 6= 0), but
^ ^ ^
there is no reason to believe that this frame has any special properties in general. In practice, it is
extremely convenient to define a rectangular mesh on the surface patch, and a grid parameterized
by (u; v ) typically serves this purpose.
15
(c) (d)
(a) (b)
Figure 9: Classical Gauss maps of surfaces. (a) An ellipsoid and (b) a portion of a torus. (c,d)
The corresponding standard Gauss maps of the normal vectors onto the sphere. Patches with
coincident normals (e.g., for the full torus) would overlap in this representation.
16
Classical Gauss Map. The surface analog of the tangent map of a curve is the Gauss map,
which takes a selection of points on the surface, typically connected by a mesh of some sort, and
associates to each point its normalized surface normal. The Gauss map is then the plot of each of
these normals in the coordinate system of a unit sphere S2 in R3 . The Gauss map is guaranteed
to be unique in some sufficiently small open set of each point of a regular surface, but may be
arbitrarily multiple valued for the entire surface; note also that many nearby surface points can be
mapped to a single point in the Gauss map, e.g., for certain types of planar curves in the surface or
a planar area patch.
In Figure 9, we show a coordinate mesh on an elliptical surface and its single-valued Gauss
map, as well as a quarter of a torus and its Gauss map; the Gauss map of the entire torus would
cover the sphere twice, and there are two entire circles on the torus that correspond to single points,
the North and South poles, in the Gauss map.
Surface Frame Evolution. The equations for the evolution of a surface frame follow the same
basic structure as those of a space curve, except the derivatives are now directional, with two lin-
early independent degrees of freedom corresponding to the tangent basis (T ^ ; T ^ ) in the surface.
1 2
Typically (see [8, 12]), one assumes a not-necessarily-orthogonal parameterization (u; v ) that per-
mits one to express the tangent space in terms of the partial derivatives (xu ; xv ), giving the normals
^ (u; v ) of Eq. (20). Then one can express the local curvatures in terms of any linearly independent
N
pair of vector fields (U; V) as
DUN^ DV N^ = K (U V) (22)
DU N^ V + U DV N^ = 2 H (U V ) : (23)
With U = xu r and V = xv r, we get the classical expressions. As Gray succinctly notes,
since all the derivatives of N
^ are perpendicular to N
^ , the whole apparatus amounts to constructing
the tangent map of the Gauss map.
If we try to build the geometry of surfaces from a parametric representation, then each direc-
tional derivative has a a vector equation of the form of Eq. (9). Thus we may write equations of
the general form
2^ 3 2 32 ^ (u; v) 3
T (u; v ) +az (u; v ) ,ay (u; v ) T
@ 66 T^ (u; v) 77 = 64,a (u; v)
1 0
76 T^ (u; v) 77
+ax (u; v )56
1
4
@u N^ (u; v)
2 5 z 0 4 25 (24)
+ay (u; v ) ,ax (u; v ) 0 ^ (u; v )
N
and 2^ 3 2 32 ^ (u; v) 3
T (u; v ) +bz (u; v ) ,by (u; v ) T
@ 66 T^ (u; v) 77 = 64,b (u; v)
1 0
76 T^ (u; v) 77 :
+bx (u; v )56
1
4
@v N^ (u; v)
2 5 z 0 4 2 5 (25)
+by (u; v ) ,bx (u; v ) 0 ^ (u; v )
N
The last lines of each of Eqs. (24) and (25) are typically combined in textbook treatments to
give 2 3 "^ #
@N
T1 (u; v )
^ (u;v )
4 @N
@u 5 = [K] ^ (u; v )
: (26)
^ (u;v )
@v
T2
17
(a) (b) (c)
Figure 10: Examples of frame choices for the upper portion of an ordinary sphere. (a) Frames
derived from standard polar coordinates on sphere. (b) Geodesic Reference frame for the sphere;
each frame is as close as possible to the canonical coordinate axes at the North pole. (c) Frames
derived from projective coordinates on the sphere, which turn out to be the same frame field as
the Geodesic Reference frame.
where the matrix [K] has eigenvalues that are the principal curvatures k1 and k2 , and thus
K = det [K] = k k 1 2 (27)
is the Gaussian curvature and

H = 12 tr [K] = 21 (k 1 + k) 2 (28)
is the mean curvature.
Examples of Surface Framings. If we are given a description of a surface, we can compute

normals and choices of the corresponding frames by various means. In Figure 10, we illustrate
three of these for the sphere. The first is derived from the standard orthonormal polar coordinate
system, and the second is the extension to surfaces of the Geodesic Reference frame, which assigns
the frame closest to a standard reference axis at the North Pole. The third is a frame based on polar
projective coordinates for the sphere,
x(u; v) =
2 u
1+ u2
+ v 2
y(u; v) =
2 v
u +v
1+ 2 2
(29)
z(u; v) = 11 , u ,v ; 2 2
+u +v 2 2
which map the real plane into the unit sphere with x + y + z2 2 2
= 1
except for the point at
infinity corresponding to the South pole. In fact, the polar projective coordinates generate the
18
same assignments as the Geodesic Reference frame does, so, except for the difference in locations
of the grid sampling, these are the same framings.
Note: Do not be confused by alternate samplings of the same framings; if a parameterization
x(u; v ) gives a frame with T1 = @ x(u; v )=@u and T2 = @ x(u; v )=@v , we can change to a polar
sampled mesh, (r = (u2 + v 2 )1=2 ; = arctan(v; u)), yet still retain the same frames at the same
points x(r; ) = x(u = r cos ; v = r sin ).
3 Quaternion Frames
In Section 2.1, we discussed the nature of 2D frames and noted a means of re-expressing the four
equations with three constraints of the conventional frame system more efficiently; we showed a
transformation into an equivalent set of two equations involving a single pair of variables obeying
a unit length constraint and whose rotation transformation properties were realized by complex
multiplication. Quaternions accomplish exactly this same transformation for 3D rotations: they
permit the nine coupled frame equations with six orthonormality constraints in 3D to be succinctly
summarized in terms of four quaternion equations with the single constraint of unit length. De-
tailed derivations along with other basic properties of quaternions are provided for reference in the
appendix. A brief summary is given below.
Quaternion Frame Equations. Our task is now to rephrase the general properties of curve and
surface frames in quaternion language so that, for example, we have a sensible space in which to
consider optimizing frame assignments.
We begin with the standard definition for the correspondence between 3 3 matrices Rij and
quaternions q : X
Rq (V)i = Rij V j = q (0; V i) q,1 : (30)
j
Henceforth, we will use the notation “” to distinguish quaternion multiplication, and will use “”
when necessary to denote ordinary Euclidean inner products. Next, we express each orthonormal
frame component as a column of Rij by using an arbitrary quaternion to rotate each of the three
Cartesian reference axes to a new, arbitrary, orientation:
N
^ or T
1
^
1 = q (0; x^ ) q,
1
N2 or T2
^ ^ = q (0; y^ ) q,
1
(31)
T^ or N
^ = q (0; ^z) q, :
1
(Technically speaking, in the above equation T ^ really means the quaternion (0; T ^ ) with only a
vector part, etc.) All this can be transformed into the following explicit representation of the frame
vectors as columns of a matrix of quaternion quadratic forms:
h i
[N
^ ] [N
^ ] [T
^] =
h 1 2
i
[T
^ ]
1 [T
^ ]
2 [N
^] =
19
2 3
q 2
+q ,q ,q 2q q , 2q q
2 2
2q q + 2q q
2
64 2q q + 2q q q , q + q , q 2q q , 2q q 75 :
0
1 2
1 2
0 3
3
2
1 2
2 2
0 3
2
1 3
2 3
0 2
0 1 (32)
0 1 2 3
2q q , 2q q
1 3 2q q + 2q q q ,q ,q +q
0 2 2 3 0 1
2
0
2
1
2
2
2
3
Note: M ATHEMATICA users should remind themselves that matrices are stored as lists of rows in
M ATHEMATICA, so one must transpose a standard matrix to easily retrieve column vectors from
Eq. (32) and avoid mysterious sign errors.
Taking differentials of Eq. (31), we generate expressions of the form
dq = q (q, dq) = q 21 (0; k)

1
(33)
dq, 1
= (dq
, q) q,
1 1
= ,(q, dq) q,
1 1
= , 21 (0; k) q, 1
(34)
where
k = 2(q0 dq , q dq0 , q dq) :
Substituting these expressions into the the calculation for the first column, we immediately find the
expected commutators of quaternion multiplication:
dN^ 1 = dq (0; x^ ) q, + q (0; x^ ) dq,

1 1
=
1
2
q ((0; k) (0; x^ ) , (0; x^ ) (0; k)) q, 1
= q (0; k x^ ) q, : 1
The rest of the columns are computed similarly, and a straightforward expansion of the components
of the cross products proves the correspondence between Eq. (33) and Eq. (9).
To relate the derivative to a specific curve coordinate system, for example, we would introduce
the curve velocity normalization v (t) = kx0 (t)k and write
q0 = v(t) 12 q (0; k) : (35)
One of our favorite ways of rewriting this equation follows directly from the full form for the
quaternion multiplication rule given in the appendix; since this multiplication can be written as an
orthogonal matrix multiplication on the 4D quaternion space, we could equally well write
2 q0 3 2 0 ,k ,k ,k 3 2 q 3
66 q0 0
77 16
x
+kz ,ky 7
y z
6 0
77
64 q0 1
75 = v(t) 2 64 +k ,k 0 +k 775 664 qq
6 +kx 0 1
75 : (36)
2 y z x 2
q0 3 +kz +ky ,kx 0 q 3
This is the 3D analog of Eq. (7).
20
At this point, there are many other directions we could carry this basic structure, but we will
not pursue the general theory of quaternion differential geometry further here. We will conclude
with a short summary of the quaternion treatment of the classical surface equations. Starting from
Eq. (33), we are led immediately to the quaternion analogs of Eqs. (24) and (25):
qu @q=@u =
1
2
q (0; a) (37)
qv @q=@v =
1
2
q (0; b) : (38)
But how shall we express the curvatures in a way similar to the classical formula in Eq. (26)?
An elegant form follows by pursuing the quaternion analog of the vector field equations given in
Eqs. (22,23). We write
qu q,v 1
= , 14 q (0; a) (0; b) q, 1
= , 14 q (,a b; a b) q, 1
h i
= , 14 ,a b Î + (a b)x T^ 1 + (a b)y T^ 2 + (a b)z N^ ; (39)
where we use the quaternion forms in Eq. (31) with the addition of the quaternion identity element
I = (1
^ ; 0) = q (1; 0)q,1 for the frame vectors. We see that the projection to the normal direction
gives precisely the determinant (a b)z = K identified in Eq. (26) as the scalar curvature. The
mean curvature follows from an expression similar to Eq. (23),
q (0; x^ ) q,u
1
+ q (0; y^ ) q,v 1
= , 21 q (,x^ a , y^ b; x^ a + y^ b) q, 1
h i
= , 12 ,(ax + by )Î + bz T^ , az T^
1 2 a , bx )N^ ;(40)
+( y
where again the coefficient of the normal, (ay , bx ) = tr [K] = 2H , is the desired expression. Sim-
ilar equations can be phrased directly in the 4D quaternion manifold using the forms of Eq. (36).
3.1 Visualizing Quaternion Frames

Seeing the parameters of a single quaternion. Any (unit) quaternion is a point on S3 and there-
fore is described by three parameters incorporated in the standard parameterization
q(; n^ ) = (cos 2 ; n^ sin 2 ) ; (41)
where 0 < 4 , and the eigenvector of the rotation matrix (unchanged by the rotation), is a
^ = (cos cos ; sin cos ; sin ) with 0 < 2
point on the two-sphere S2 representable as n
and ,=2 =2.
21
1
0.5
q (unseen)
0
−θ
-0.5
1
cos −θ 2 0.5
2 q (seen) 0
-1
-0.5
n sin−θ
^ -1 -0.5
2
0 0.5 -1
1
(a) (b)
Figure 11: Illustration of how the q0 part of a quaternion is “known” if we have a 3D image of the
vector part q = n^ sin 2 of the quaternion. (b) Schematic representation of the concentric-sphere
uniform distance scales needed to form a mental model of the metric distances in quaternion
space between two points in the parallel 3D projection. Distances are roughly Euclidean near
the origin (q 0 in (a)) and equal-length lines appear increasingly compressed as the radius
approaches unity.
An informative visualization of quaternions can be constructed by examining their properties

carefully. If we simply make a 3D display of the vector part of the quaternion, n
^ sin 2 , we see that
the scalar element of the quaternion is redundant, since, for each ,
0 1 =

1 2
q = cos 2 = @1 , n^ sin 2 A :
2
0 (42)
That is, q0 is just the implicitly known height of the 4D unit vector in the unseen projection di-
rection, as illustrated in Figure 11(a). In Figure 11(b), we schematize the mental model of metric
distance required to complete the interpretation of the visualization. If we imagine dividing the
arc of the semi-circle in Figure 11(a) into equal angular segments, the arc lengths are all the same
distance apart in spherical coordinates. Projected onto the q plane, however, the projected spacing
is non-uniformly scaled by a a factor of sin . Thus to keep our vision of distance consistent, we
imagine the space to be like 3D graph paper with concentric spheres drawn at equal distances in the
special scale space; such 3D graph paper would look like Figure 11(b). Distances are essentially
Euclidean near the 3D origin, for small 3D radii, and are magnified as the radius approaches unity
to make the marked spheres equidistant in conceptual space.
If we assume the positive root is always taken for q0 , then we effectively restrict ourselves to
a single hemisphere of S3 and eliminate the two-fold redundancy in the correspondence between
22
0110
11
00 11
00
11
00 0110 0110
01 01 11
00
00
11
(a) (b) (c)
Figure 12: (a) This image represents the 3D vector part of the quaternion q = (cos 2 ; ^t sin 2 )
representing a single instance of the one-parameter family of possible rotations leaving in-
variant the tangent vector ^t at one point of a space curve. (b) A representation of the entire
one-parameter space of possible frames having the same tangent vector ^t. The vector part of
the quaternion must lie on the diameter of the two-sphere in 3D depicted here. We depict the
diameter as a very skinny ellipse, because it is in fact a degenerate projection of a circle in 4D,
which could be exposed as shown by making a small 4D rotation before projecting to 3D. (c)
A polar projection of the same object removes the doubling by projecting the circle to a line
through infinity in R3 .
quaternions and the rotation group. Alternatively, despite the fact that quaternions with both signs
of q0 map to the same point in this projection, we can indicate the simultaneous presence of both
hemispheres using graphical cues; one possible method is to use saturated colors in the “front”
hemisphere, and faded colors (suggesting distance) for objects in the “back” hemisphere.
Hemispheres in S3 . To clarify the terminology, we note that a projected hemisphere for S2 is a

filled disk (a “two-ball”) in the plane, and the full surface of the sphere consists of two such disks
joined at the outer circular boundary curve; for S3 , we use the word hemisphere to indicate a filled
solid two-sphere (technically a “three-ball”), and imagine the full volume of the three-sphere to
consist of two such spherical solids joined on the skin (a two-sphere) of the surface enclosing both.
The family of possible values of Eq. (41) projects to a double-valued line (actually an “edge-
on” projection of a circle) which is a directed diameter of the unit two-sphere, in the direction of n
^;
in a polar projection, this circle becomes a line to infinity through the origin. These representations
of a unit quaternion as a vector from the origin to a point inside the solid two-sphere (the three-ball)
are illustrated schematically in Figure 12.
Any particular 3D rotation is represented twice, since the quaternion circle is parameterized by
0 < 4 . A simple parallel projection thus produces two solid balls on top of each other in the
3D projection, one the analog of the “North pole disk” of a two-sphere parallel projected from 3D
to a screen, the other the analog of the “South pole disk” of a two-sphere. The analog of a polar
projection, which for a two-sphere sends the North pole to infinity of R2 , flattens the three-sphere
out to fill R3 , as shown in Figure 12(c), and eliminates the double-valued properties of the parallel
projection.
23
CURVE LENGTHS (2,3) Torus Knot Helix
Frenet Frame 14.3168 6.18501
Geodesic Reference Frame 14.6468 7.82897
Parallel Transport Frame 10.1865 6.06301
Table 1: Relative lengths (in radians) of the quaternion frame maps for various frame choices
describing the (2,3) torus knot and the helix. The Parallel Transport frame is the shortest possible
frame map.
3.2 Quaternion Frames for Curves

We now can produce quaternion frames for space curves directly by several techniques.
Quaternions from Local 3D Frames. In the case of the Frenet frame, we have no choice but to
consider each frame as totally independent of the others. Each is locally computable, and there is
in principle no relation between them, since the curvature could vanish at any point. In this case,
we compute the frames directly from Eq. (16), thus deriving a 3 3 orthogonal matrix R(t) at
each point of the curve. We then apply standard inversion algorithms [37, 35, 31] to obtain the
corresponding quaternion up to a sign. Finally, we apply a simple operator that checks the local
continuity of the corresponding frames. If two quaternion vectors representing neighboring frames
have a dot product near negative one, we change the sign of one to keep it near its neighbors. If
two neighbors are excessively far apart in terms of the 4D angle between them, and are not simply
near-negatives of one another, then the Frenet frame probably is poorly defined and should be
tagged as such until continuity resumes. Figure 13 shows the Frenet frame tubing of a torus knot
and the corresponding trajectory of these frames in the vector subspace of quaternion space.
Note: Forcing close quaternion Frenet frames on closed curves such as torus knots results in
a very interesting phenomenon. Depending on the parameters of the curve, the path in quaternion
space may close after a single traversal of the curve, or it may require two or more traversals, as
in the case shown in Figure 13. We have checked this feature on a wide range of torus knots,
and found that there are generally “jumps” between needing different numbers of circuits at those
parameter values that imply inflection points (zero curvature) in the curve.
Direct Quaternion Frames. The Geodesic Reference frame and the Parallel Transport frame, in
contrast to the Frenet frame, can be defined directly in terms of quaternions if desired, as indicated
in Section 2.2; all that is needed is an initial quaternion reference frame, and then the geometry of
the curve specifies enough at each point to express the needed rotation in quaternion form.
Comparison of Tubings and Quaternion Frames. Previously, in Figure 8, we compared the

tubings for the (2,3) torus knot and for the helix based on the Frenet, Geodesic Reference, and
Parallel Transport frames. The corresponding quaternion paths are illustrated together in Figure
14. The Parallel Transport frame shown uses the initial Frenet frame as a starting point; we could
24
1
0.5
-0.5
1
0.5
-1
-1 0
-0.5
0 -0.5
0.5
-1
1
(a) (b)
Figure 13: (a) A trefoil torus knot. (b) Its quaternion Frenet frame projected to 3D. For this
trefoil knot, the frame does not close on itself in quaternion space unless the curve is traversed
twice, corresponding to the double-valued “mirror” image of the rotation space that can occur in
the quaternion representation. Observe the longer segments in (b): these correspond to the three
high-torsion segments observable in (a).
25
(a) (b)
Figure 14: (a) Quaternion frames in “standard” 3D vector visualization projection for the (2,3)
torus knot: Red—Geodesic Reference: this is planar by construction, since all 3D points must
lie in the plane perpendicular to the reference axis; the 3D origin is at the centroid of the red
curve. Green—Frenet: the Frenet frame is actually cyclic, but to see this easily for this 2,3
torus knot, the mirror image of the current frame must be added, giving effectively a double
traversal of the curve as shown in Figure 13. Cyan—Parallel Transport: the PT frame must
be given a starting value, which here is seen at the top center of the image to coincide with
the (green) Frenet frame. The PT frame is not cyclic, but is the shortest path, with three very
noticeable tight loops. (b) The same selection of quaternion frames for the helix. Again, the red
Geodesic Reference curve is planar (and cycles back on itself twice for this helix); the green
Frenet frame takes a longer path that will return to its original orientation, and the cyan Parallel
Transport frame, seen starting at the same orientation as the Frenet, will not ordinarily return
to the same orientation, but will have the shortest 4D path length. (The hidden double circuit
of the Geodesic Reference frame for this helix in fact makes it longer.)
26
use any starting quaternion with the correct tangent vector. The relative path lengths of the curves
in Figure 14 are summarized in Table 1.
We note the following properties:
Frenet. Periodic for periodic non-singular curves, has a tendency to twist a bit too much
(where the torsion is high), leaving long jumps between neighboring samples in quaternion
space; undefined at inflection points and zero curvature segments.
Geodesic Reference. Also guaranteed to be periodic for periodic curves, but has the odd
property that it always lies in a plane perpendicular to the reference axis in our preferred 3D
quaternion projection. Ambiguous and therefore potentially not smooth for frames opposing
the reference frame direction.
Parallel Transport. This is the quaternion frame with minimal 4D length, though it may
be difficult to see this feature immediately in our standard projection. It is not in general a
periodic path. Different choices of starting frame produce curves of identical length differing
by rigid (possibly reflecting) 4D motions (see Eq. (56)).
3.3 Quaternion Gauss Map for Surfaces

The quaternion Gauss map extends the Gauss map to include a representation of the entire co-
ordinate frame at each surface point, introducing a number of new issues. In particular there is
a useful, but mathematically suspect, approach that we might call an “engineering” approach to
the quaternion Gauss map that lets us quickly get informative visualizations for those special cases
where we are given a locally orthogonal parameterization of the surface except perhaps for isolated
singularities of the coordinate system.
For these cases, we may construct the precise quaternion analog of the Gauss map by lifting
the surface’s coordinate mesh into the space of quaternions at each value of the orthonormal co-
ordinatization (u; v ) of the surface or surface patch. The correspondence of this map to the Gauss
map is not directly visible, since (see Eq. (32)) the normal directions of the Gauss map are non-
trivial quadratic forms constructed from all the quaternion components; however, a projection to a
subspace of the quaternion space based on the bilinear action of quaternions on pure vectors may
be constructed by imitating the projection of the Hopf fibration of S3 (see, e.g., Shoemake [39, 3]).
In Figure 15, we show two such cases, an ellipsoid with orthonormal polar coordinates singular
at the poles and a torus with global, nonsingular, coordinates, using our now-standard projections
of the quaternion Gauss map to 3D. In each of these cases, a single circuit of the surface gener-
ates only one-half of the quaternion surface shown; the symmetric quaternion figure results from
traversing the surface twice to adjoin the reflected image of the single-circuit quaternion surface.
That is, each point on the 3D surfaces appears twice, once at q , and once at ,q , in these periodic
quaternion Gauss maps.
We see that the singular coordinate system typically used for the ellipsoid is topologically a
cylinder; the circles corresponding to the singularities of the coordinate system (circles of normal
directions) at the North and South poles correspond to boundaries of the quaternion Gauss map.
27
(c) (d)
(a) (b)
Figure 15: Examples of quaternion Gauss maps for surfaces. (a) The ellipsoid and (b) the torus.
(c,d) The corresponding Quaternion Gauss maps, projected from the three-sphere in 4D. The
equatorial direction has been traversed twice in order to get a closed path in the map; the singular
poles in the ellipsoid coordinate system correspond to the edges or boundaries of the quaternion-
space ribbon.
28
(a) (b) (c)
Figure 16: Examples of quaternion Gauss maps for the frame choices for the upper portion of an
ordinary sphere given originally in Figure 10. (a) Frames derived from standard polar coordinates
on sphere. (b) Geodesic reference frame for the sphere; each frame is as close as possible to the
canonical coordinate axes at the North pole. (c) Frames derived from projective coordinates on
the sphere.
PATCH AREAS Hemispherical patch

Polar Coordinates 2.1546
Geodesic Reference Frame 1.9548
Table 2: Areas (in steradians) of the quaternion frame maps for the polar coordinate and Geodesic
Reference frame choices on the hemispherical patches of Figure 16.
The torus, which has the extremely unusual feature that it possesses a global regular coordinate
system, has a (reflection doubled) quaternion Gauss map which is another, four-dimensional, torus
embedded in the quaternion S3 space.
Quaternion Maps of Alternative Sphere Frames. In Figure 10, we showed three alternate sets
of frames for the upper half of an ordinary sphere. The assigned coordinate systems may be
converted directly into quaternion frames and coerced into consistency in the usual manner. In
Figure 16, we show the results. The Geodesic Reference frames and the projective coordinates are
in fact the same space of frames computed in different ways: both are planes perpendicular to the
z^ axis. The coordinate systems used to compute the quaternion Gauss maps in parts (a) and (b)
of the figure are commensurate, so we may compare the areas, computed using solid angle on the
three-sphere in units of steradians; the results are shown in Table 2.
Covering the Sphere and the Geodesic Reference Frame South Pole Singularity. The Geodesic
Reference frame for a surface patch has the peculiarity that it has an ambiguity whenever the vector
to be assigned is exactly opposite the reference frame. As we show in Figure 17, the tilting from the
29
1 1
0.5 0.5
0 0
-0.5 -0.5
-1 -1
1 1
0.5 0.5
0 0
-0.5 -0.5
-1 -1
-1 -1
-0.5 -0.5
0 0
0.5 0.5
1 1
(a) (b)
Figure 17: The Geodesic reference frame tilts to an ambiguous result as the tilt angle approaches
, the inverted direction of the chosen reference frame. We see two different 3D projections
of the quaternion surface, (a) giving the vector coordinates (q1 ; q2 ; q3 ), and (b) the coordinates
(q0 ; q1 ; q2 ). The center is the North pole, the middle ring is the equator, and outer circle is in
fact the space of possible frames at the South pole of the sphere: there is no unique way to tilt
the North pole to the orientation of the South pole, as there is a full circle of arbitrariness in the
choice.
30
reference frame in quaternion space (easily seen in ordinary 3D space as well) eventually reaches
a quaternion circle representing the ambiguous orientation of the frame with reference direction
along the ,z^ axis. This phenomenon is a practical consequence of the fact that the two-sphere
does not admit a global vector field: according to classical manifold theory (see, e.g., Milnor or
Grimm and Hughes [29, 13]), one needs at least two separate patches, one for the North pole and
one for the South pole, to place a complete set of coordinates (or equivalently, for our problem, a
set of frames) on a sphere.
The more mathematical approach requires that interesting surfaces be defined as a collection of
patches [13, 7], and the spaces of frames for each patch must be matched up and sewn together by
assigning a transition function along the boundaries. There are a variety of ways one can approach
the problem of taking a manifold and associating fiber bundles with it; the most relevant fiber
bundle for the context of the current problem is the space of moving frames of the space R3 in
which the surface is embedded [7, 40]. We in fact move as usual from the space of frames to
the space of associated quaternions. Then at each point x of a patch we have frames that are
matrix-valued functions from the patch into the group SU(2) of quaternion frames (which we treat
as the topological space S3 ). We can express the relationship between the frames q and q 0 of
two neighboring patches U and U 0 , represented as quaternions, via quaternion multiplication by a
transition function t:
q0 = t q :
We may in fact explicitly construct the transition functions between the two patches as quater-
nion maps, giving a quaternion version of one of the classical procedures of manifold theory. In
Figure 18(a), we show the the projective coordinates on the sphere that produced the set of co-
ordinate frames in Figure 10(c), which are essentially equivalent to those in (b) sampled at polar
coordinate values. Using the polar coordinate sampling, so that we can easily identify the equa-
tor, we show in Figure 18(b) the quaternion maps corresponding to the coordinate frames derived
from this orthonormal coordinate system covering the North pole (disk in center) and the South
pole (smashed side view of a hemisphere in the 3D projection with its q ! (,q ) partner). These
coordinate systems agree at exactly one point on the equator, which is (almost) evident from the
figure; note that we have chosen to display the coordinate systems only up to the equator, unlike
the patches of Figure 17, which cover the entire sphere except for one pole.
In order to establish a mapping covering the complete sphere, we must write down an explicit
correspondence between the quaternion frames for each patch at each shared point on the equator.
In Figure 19(a), we show the geodesic arcs on S3 symbolizing the transition rotation
t() = qsouth () qnorth

, ()
1
at each point on the equatorial circle parameterized by . Note carefully the order of quaternion
multiplication; with our conventions a different order will not work. The arcs themselves are
actually segments of the space of possible frames, since the simplest rotation between two frames
with the same normal (at the same point on the equator) is a geodesic rotation about that normal.
Figure 19(b) and (c) shows the transition functions q () sampled at regular intervals in and
referred to the origin (1; 0; 0; 0) in quaternion space. Each quaternion point at the end of an arc
31
-1 -1
1 -0.5 1 -0.5
0.5 0 0.5 0
0.5 0.5
0 1 0 1
-0.5 -0.5
-1 -1
1 1
0.5 0.5
0 0
-0.5 -0.5
-1 -1
(a) (b)
0.5
1
-0.5
0.5
0 -1
1
-0.5 0.5
0
-1 -0.5
-1
(c)
Figure 18: (a) The North pole projective coordinatization of the sphere; (b) a similar regular patch
for the South pole. Because of the “no-hair” theorem, no single regular patch can cover the entire
sphere. (c) The quaternion mappings of the systems of frames given by the North and South pole
coordinate patches, sampled in polar coordinates. The q ! (,q ) reflected images are included,
though the North pole’s images both have the same projection and are thus indistinguishable here.
The maps in (c) extend only to the equator, unlike the patches given in Figure 17.
represents a rotation to be applied to a point on the North pole patch equator to obtain the coordinate
frame at the corresponding point on the South pole patch equator. One point is in fact the identity,
and there is some degeneracy due to reflection symmetry across the equator.
32
(a) (b) (c)
Figure 19: (a) The transition functions from the North pole frame to the South pole frame as arcs
in the three-sphere. These arcs are pieces of the space of possible frames with a given normal
on the equatorial point. (b) A representation of the transition functions as arcs from the origin in
rotation space (the pole (1; 0; 0; 0) in quaternion space) common to all the arcs here. The ends of
the arcs thus represent the actual rotation needed to match the coordinate systems at each point
on the equator. (c) A different projection from 4D to 3D, showing more details of the structure
of the transition function arcs, which have a two-fold degeneracy in the standard projection (b).
33
4 The Space of Frames
We at last ready to introduce the key concept of the space of possible frames.
Suppose at each sample point x(t) of a curve, we are given a unit tangent vector, T ^ (t), com-
puted by whatever method one likes (two-point sampling, five-point sampling, analytic, etc.). Then
one can immediately write down a one-parameter family describing all possible choices of the nor-
mal plane orientation: it is just the set of rotation matrices R(; T ^ (t)) (or quaternions q (; T ^ (t)))
^ (t) fixed.
that leave T
For surfaces, the analogous construction follows from determining the unit normal N ^ (u; v ) at
each point x(u; v ) on the surface patch. The needed family of rotations R(; N ^ (u; v )) (or quater-
nions q (; N
^ (u; v ))) now leaves N ^ (u; v ) fixed and parameterizes the space of possible tangent
directions completing a frame definition at each point x(u; v ).
However, there is one slight complication: the family of frames R(; v ^ ) leaving v ^ fixed does
not have v^ as one column of the 3 3 rotation matrix, and so does not actually describe the desired
family of frames. Therefore we proceed as follows:
We define f (; v ^ ) = (f0 ; f1 ; f2 ; f3 ) to be a quaternion describing the family of frames for
which the direction v ^ is a preferred fixed axis of the frame, such as the tangent or normal vectors.
The orthonormal triad of 3-vectors describing the desired frame is
2F (; v^ ) = 3
f2
+f ,f ,f
2 2
2f f , 2f f
2
2f f + 2f f
64 2f f + 2f f f , f + f , f 2f f , 2f f 75;
0 1
1 2
2
0 3
3
2
1 2
2 2
0 3
2
1 3
2 3
0 2
0 1 (43)
0 1 2 3
2f f , 2f f
1 3 2f f + 2f f
0 2 f ,f ,f +f
2 3 0 1
2
0 1
2 2
2
2
3
where one column, typically the 3rd column, must be v ^.

The standard rotation matrix R(; v^ ) leaves v
^ fixed but does not have v^ as one column of the
3 3 rotation matrix, and so we have more work to do. To compute f (; v^ ), we need the following:
A base reference frame b(^v) that is guaranteed to have, say, the 3rd column exactly aligned
with a chosen vector v
^ , which is either the tangent to a curve or the normal to a surface.
A one-parameter family of rotations that leaves a fixed direction v

^ invariant.
The latter family of rotations is given simply by the standard quaternion
q(; v^ ) = (cos 2 ; v^ sin 2 ) ; (44)
for 0 < 4, while the base frame can be chosen as

b(^v) = q(arccos(^z v^ ); (^z v^ )=kz^ v^ k) : (45)
We refer hereafter to the frame b(^v) as the Geodesic Reference Frame because it tilts the reference
vector ^z along a geodesic arc until it is aligned with v
^ ; see Figure 20. If v
^ = z
^, there is no
problem, since we just take b(^v) to be the quaternion (1; 0); if v^ = ,z^, we may choose any
34
Figure 20: Example of the Geodesic Reference Frame: on the northern hemisphere of a 2-sphere,
the Geodesic Reference Frame tilts the ^z axis of the north pole’s identity frame along the shortest
arc to align with a specified reference direction.
compatible quaternion such as (0; 1; 0; 0). We escape the classic difficulty of being unable to
assign a global frame to all of S2 because we need a parameterization of all possible frames, not
any one particular global frame. If one wants to use a reference frame that is not the identity frame,
one must premultiply b(^ v) on the right by a quaternion rotating from the identity into that reference
frame; this is important when constructing a nonstandard Geodesic Reference Frame such as that
required to smoothly describe a neighborhood of the southern hemisphere of S2 .
We can thus write the full family of possible quaternion frames keeping v ^ as a fixed element
of the frame triad to be the quaternion product
f (; v^ ) = q(; v^ ) b(^v) ; (46)
where denotes quaternion multiplication and all possible frames are described twice since 0
< 4. To summarize, if we specify a frame axis v^ to be fixed, then the variable in f (; v^ )
serves to parameterize a ring in quaternion space, each point of which corresponds to a particular
3D frame, and each frame has a diametrically opposite twin.
We argue that, since optimization will typically be done in the full quaternion space, the fact
that two opposite-sign quaternions map to the same physical three-space rotation is not a detriment;
in fact, it potentially permits an additional stability in the variational process since rotations by +
35
and , are not close to each other in quaternion space as they are in ordinary rotation matrices. In
principle, any quaternion Gauss map can be replaced by its exact negative, and the variational pro-
cess could converge from an ambiguous starting point to either one; the frames would be the same.
In our standard projection, the two reflection-equivalent maps are inversions of one another about
the 3D origin; their unseen opposite q0 values can of course cause an additional large separation of
the maps in 4D space.
4.1 Full Space of Curve Frames

We can now construct the space of frames step by step using the method above. In Figure 21, we
illustrate various views of the construction of the space of frames for the trefoil knot, beginning
with a few tangent vectors and the quaternion basis frames corresponding to quaternions that tilt
the reference axis into this tangent direction. The circular curve of quaternions representing the
space of normal frames is drawn for each tangent; each basis frame touches this curve once. Then
the family of these circular curves sweeps out a cylindrical two-manifold, the full space of invariant
frames for a 3D curve.
This space has several nontrivial properties. One is that, given one circular ring of frames,
a neighboring ring that is a parallel-transported version of the first ring is a so-called “Clifford
parallel” of the first ring: the distance from any point on one ring to the nearest point on the second
ring is the same. This is nontrivial to visualize and is a feature of the 4D space we are working
in. Another property is that the intervals between rings in the quaternion space directly indicate
the curvature. This comes about because the magnitude of T ^ 0 is related to the parallel transport
transition between any two sample points, given by Eq. (17); since the parallel transport frames
are legal frames, and since the starting frame is arbitrary, each full ring is a parallel transport of
its predecessor, with the angular distance of the transition rotation providing a measure of the
curvature relative to the sampling interval.
4.2 Full Space of Surface Maps

The full space of frames for a surface patch is even more complex to visualize, since it is a “hyper-
cylindrical” 3-manifold, formed by the direct product of patches of surface area with the rings of
possible frames through each surface point.
As a very simple case of a surface, consider the patch introduced at the beginning of the paper
in Figure 2(a). The coordinate system used does not provide a unique tangent frame, and so one
cannot immediately determine a logical frame choice.
In Figure 22, we show spaces of possible frames for the four corners as four rings of quaternion
values compatible with the normals at the patch corners. Parallel transporting the initial frame
along the two different routes in Figure 2(b,c) produces incompatible frames at the final corner; we
represent this situation in Figure 22 by drawing the routes in quaternion space between the initial
frame (the degenerate circle appearing as a central vertical line) and the final frame; the mismatch
between the two final frames is illustrated by the fact that the two paths meet at different points on
the final ring specifying the frame freedom for the bottom corner’s frame.
36
1
0.5
z
0
-0.5
-1 1
-1 0.5
-0.5
0
0
0.5 -0.5 y
x
-1
1
(a) (b) (c)
Figure 21: (a) The first several pieces of the construction of the invariant quaternion space for the
frames of the trefoil knot. The red fan of vectors shows the first several elements of the tangent
map, represented as vectors from the origin to the surface of the two-sphere and connected by a line.
Each green vector points from the origin to the Geodesic Reference element of the quaternion space
q(arccos(^t z^); ^t z^=k^t z^k) guaranteed to produce a frame with the tangent ^t. The black curves are
the first several elements of the one-parameter space of quaternions representing all possible quaternion
frames with the tangent ^t. (b) This piece of the space of possible frames represented as a continuous
surface, where a circle on the surface corresponds to the space of frames for one point on the curve. (c)
The rest of the full constraint space for the trefoil knot. All quaternions are projected to 3D using only
the vector part.
37
A
B’
B C B
B
A
C’ B’ A
C C’
C = C’
B’
(a) (b) (c)
Figure 22: A different viewpoint of the mismatch problem of Figure 2. (a) Choosing different
routes to determine the frame at the bottom point results in the incompatible frames shown here in
3D space. (b) The same information is presented here in the quaternion space-of-frames picture.
We use throughout a quaternion projection that shows only the 3-vector part of the quaternion,
dropping q0 ; this is much like projecting away z in a polar projection of the 2-sphere. Each heavy
black curve is a ring of possible frame choices that keep fixed the normals in (a); the labels mark
the point in quaternion space corresponding to the frames at the corners in (a), so the gap between
the labels C and C ’ represents the frame mismatch in quaternion space on the same constraint
ring. (The apparent vertical line is the result of drawing a squashed circle of frames at vertex A in
this projection.) (c) The method proposed in this paper to resolve this conflict is to fix one point,
say A, divide the polygon ABCB 0 into triangles, and slide B , C , and B 0 along the constraint
rings until the total triangle areas are minimized, and some compromise with C = C 0 is reached.
38
0.4 0.2
0 -0.2
-0.4
1
1
0.5 0.5
0 0
-0.5 -0.5
-0.5
-0.4 -0.25
-0.2 0
0 -1
-1 0.25
0.2 0.5 0.25
0 -0.25 0.5
0.4 -0.5
(a) (b)
Figure 23: (a) A more complete picture of the space of frames for this surface patch; the surface
shown is a sparse quaternion frame choice for the surface, and we show a subset of the rings of
constraints. Each ring passes through one quaternion point on the frame map, the point specifying
the current frame choice. Variations must keep each vertex on its ring. (b) An equivalent set of
frames is formed by applying a rotation to the entire set of frames. All points follow their own ring
of constraints to keep the same normal, These pictures represent the three-manifold in quaternion
space swept out by the possible variations.
39
Sliding and Overall Rotational Freedom. In Figure 23(a), We go one step further, and first
show how the quaternion Gauss map of an entire patch is situated relative to the ring space; keeping
one corner fixed and sliding the rest of the frames around the circular rings takes us to distinct
families of frames, which obviously have different areas in the quaternion space. Finally, in Figure
23(b), we keep the fundamental space of frames the same, but exercise the freedom to choose
the single parameter describing the basis for the overall orientation; rotating the basis sweeps out
both the three-manifold describing the space of frames for this patch, and the family of equivalent
frames differing by an insignificant orientation change in the basis vector.
In order to resolve the frame choice ambiguity, one needs a systematic approach; we propose
in the next section to accomplish this by optimizing appropriate quantities, e.g., by minimizing the
area of the quaternion Gauss map in quaternion space.
We remark that the general features of the surface curvature can in principle be noted from the
space of possible frames in a similar manner to that for curves. The family of curves through any
point spanning the surfaces tangent space at that point possesses a family of rings parallel to the
space of frames at the point, allowing estimates of the rates of change in different directions; the
principal curvatures then correspond to the maxima and minima.
5 Choosing Paths in Quaternion Space

We have now seen that the space of possible frames at any point of a curve or surface thus takes the
form of a great circle on the unit three-sphere representing the unit quaternions in 4D Euclidean
space. While diametrically opposite points on this circle represent the same frame in 3D space,
the two-fold redundancy can actually be an advantage, since it helps avoid certain types of wrap-
around problems encountered when trying to find paths in the space. Our task then is to select a
set of values of the parameter on each of these great circles.
The advantage of looking at this entire problem in the space of quaternions is that one can
clearly compare the intrinsic properties of the various choices by examining such properties as
length and smoothness in the three-sphere. We note the following issues:
Frame-frame distance. Suppose we are given two neighboring tangents, ^t1 and ^t2 , and two
corresponding candidate frame choices parameterized by 1 and 2 . What is the “distance”
in frame space between these? The simplest way to see how we should define the distance is
by observing that, by Euler’s fundamental theorem, there is a single rotation matrix R(; n ^)
That takes one frame to the other; if R1 (1 ; ^t1) and R2 (2 ; ^t2 ) are the two frames, then one
can write R = (R2 (R1 ),1 ) and solve for and n ^ . Clearly the value of gives a sensible
measure of the closeness of the two frames.
Quaternion distance. We remark that essentially the same procedure is required to obtain
the parameters of R directly or to find the value of the equivalent quaternion. If we work in
quaternion space, we compute q1 (1 ; ^t1 ) and q2 (2 ; ^t2 ), and then find rather more straightfor-
wardly an equivalent result by noting that the zeroth component of q = q2 (q1 ),1 is identical
to the rotation-invariant scalar product of the two quaternions, q1 q2 , and thus provides the
40
needed angle at once:
= 2 arccos(q q ) :
1 2
Approximation by Euclidean distance. One can in principle compute quaternions in polar

coordinates and use the induced metric on the sphere to compute precise arc-length dis-
tance integrals. However, one generally can expect to be dealing with fine tessellations of
smoothly varying geometric objects; in this case, it may be sufficient for numerical purposes
to estimate frame-to-frame distances using the Euclidean distance in R4 , since the chord of
an arc approximates the arc length well for small angles.
Optimal Path Choice Strategies. Why would one want to choose one particular set of values of
the frame parameters over another? The most obvious is to keep a tubing from making wild twists
such as those that occur for the Frenet frame of a curve with inflection points. In general, one can
imagine wanting to minimize the total twisting, the aggregate angular acceleration, etc., subject to
a variety of boundary conditions. A bewildering variety of energy functions to minimize can be
found in the literature (see, e.g., [6]). We summarize a selection of such criteria for choosing a
space of frames below, with the caveat that one certainly may think of others!
Minimal Length and Area. The most obvious criterion is to minimize the total turning an-
gle experienced by the curve frames. Fixing the frames at the ends of a curve may be required
by periodicity or external conditions, so one good solution would be one that minimizes the
sum total of the turning angles needed to get from the starting to the ending frame. The
length to minimize is just the sum of the angles rotated between successive frame choices, as
noted above, either exact or approximate. Similar arguments apply to the area of a surface’s
quaternion Gauss map.
Parallel Transport along Geodesics. Given a particular initial frame, and no further bound-
ary constraints, one may also choose the frame that uses the minimum local distance to get
between each neighboring frame. Since the parallel transport algorithm corresponding to the
Bishop frame uses precisely the smallest possible rotation to get from one frame to the next,
this gives the minimal free path that could be computed frame-by-frame. On a surface, the
resulting paths are essentially geodesics, but, as noted in Figure 2, there is no obvious analog
of a global parallel transport approach to surface framing.
Minimal Acceleration. Barr, Currin, Gabriel, and Hughes [2] proposed a direct general-
ization of the “no-acceleration” criterion of cubic Euclidean splines for quaternion curves
constrained to the three-sphere; the basic concept was to globally minimize the squared tan-
gential acceleration experienced by a curve of unit quaternions. Though the main application
of that paper was animation, the principles are entirely valid for numerically computing op-
timal frames for curves and surfaces in our context.
Keyframe splines and constraints. If for some reason one must pass through or near certain
specified frames with possible derivative constraints, then a direct spline construction in the
quaternion space may actually be preferred; see, e.g., [37, 35, 31, 39, 23]. Most splines can
41
be viewed in some sense as solving an optimization problem with various constraints and
conditions, and so the keyframe problem essentially reverts again to an optimization.
General Remarks. For both curves and surfaces, there is a single degree of freedom in the
frame choice at each point where we have sampled the tangent or normal direction, respectively.
This degree of freedom corresponds to a relatively common “sliding ring” constraint that occurs
frequently in minimization problems. General packages for solving constraints are mentioned in
Barr, et al. [2], who chose MINOS [30]. For our own experiments, we have chosen Brakke’s
Surface Evolver package [6], which has a very simple interface for handling parametric constraints
as “boundary” conditions, and can be used for a wide variety of general optimization problems.
Two enhancements to the Evolver have recently been added to handle the specific issues related to
quaternion optimization; a new symmetry “ symmetry_group "central_symmetry" ”
identifies the quaternion q with ,q if desired during the variation to prevent reflected double traver-
sals like that in Figure 13 from varying independently, and the system is now able to use the pull-
back metric on the sphere
X
ds2
= dxi dxj r, (r i;j , xi xj )
4 2
i;j
to compute distances directly on the three-sphere. Computation using the metric, however, is very
slow, and so in practice we have used the Euclidean R4 chord approximation, which works quite
well for closely spaced samples and is much faster. Yet another alternative proposed by Brakke
(private communication) is to use periodic coordinates on S3 of the form
(x 1 = sin r cos s; x 2 r sin s; x = cos r cos t; x = cos r sin t) ;
= sin 3 4
and to vary directly on an R3 space with (x = r; y = s; z = t) and the metric

2 3
1 0 0
64 0 sin x 0 75 :
2
0 0 cos x 2
Our own use of the Evolver required only changing the parameter “#define BDRYMAX 20”
in skeleton.h to the desired (large) value and recompiling. Then, remembering to set
“space_dimension 4” when working in R4 , one needs in addition a piece of code similar
to the following M ATHEMATICA fragment to define the boundary constraints for each fixed vector
(tangent or normal) and the chosen initial quaternion frame:
Do[ringeqn = Qprod[makeQfromVec[veclist[[i]],P1],
q0list[[i]]]//Chop;
Write[file, " boundary ",i," parameters 1"];
Write[file, "x1: ", CForm[ ringeqn[[2]]]];
Write[file, "x4: ", CForm[ ringeqn[[1]]]],
{i,1,Length[veclist]}]
42
Note that, since Evolver displays only the first three coordinates, we have moved the scalar quater-
nion to the end; then the Evolver will display our preferred projection automatically.
General Remarks on Optimization in Quaternion Space. Numerical optimization remains a

bit of an art, requiring patience and resourcefulness on the part of the investigator. We found,
for example, that curve optimization was relatively more stable than surface optimization because
single curve outliers add huge amounts to the length, whereas single surface points stuck in a
far away crevice may contribute only a tiny amount to the area of a large surface. Although
the Evolver in principle handles spherical distances, we used the default 4D Euclidean distance
measure as an approximation; this generally corresponded well to explicit area calculations using
solid angle performed on the same data sets. However, we did find that extremely random initial
conditions (unrealistic for most applications), could produce isolated points stuck in local minima
diametrically across quaternion space, at q ! ,q , from where they should be. This type of
problem can be largely avoided simply by running a consistency preprocessor to force nearby
neighbors to be on the same side of the three-sphere. Another useful technique is to organize the
data into hierarchies and optimize from the coarse scale down to the fine scale. In other cases when
things seem unreasonably stuck, a manual “simulated annealing” procedure like that afforded by
the Evolver’s jiggle option often helps.
6 Examples
We now present some examples of frame choices computed using the Evolver to minimize the
length of the total path among sliding ring constraints for selected curves, and the total area spanned
by analogous sliding rings for surfaces. One interesting result is that there appear to be families of
distinct minima: if the initial data for a periodic surface, for example, are set up to return directly
to the same point in quaternion space after one period, one has two disjoint surfaces, one the
q ! (,q) image of the other; if the data do not naturally repeat after one cycle, they must after
two, since there are only two quaternion values that map to the same frame. The family of frame
surfaces containing their own reflected images have a minimum distinct from the disjoint family.
Minimal Quaternion Frames for Space Curves. The helix provides a good initial example of
the procedure we have formulated. We know that we can always find an initial framing of a curve
based on the Geodesic Reference algorithm; however, suppose we wish to impose minimal length
in quaternion space on the framing we select, and we do not know whether this frame is optimal
with respect to that measure. Then, as illustrated in Figure 24, we can send the ring constraints on
the possible quaternion frames at each sample point to the Surface Evolver and let it automatically
find the optimal framing. The results and energies for several stages of this evolution are shown in
the figure; the final configuration is indistinguishable from the Parallel Transport frame, confirming
experimentally our theoretical expectation that parallel transport produces the minimal possible
twisting.
43
Figure 24: Starting from the Geodesic Reference quaternion frame for a single turn of the helix,
the very dark gray circle, the Evolver produces these intermediate steps while minimizing the
total quaternion curve length subject to the constraints in the space of frames. The final result is
the white curve, which is identical to several decimal points with the Parallel Transport quaternion
frame for the same helix. The numerical energies of the curves, from dark to light in color, are
3.03, 2.91, 2.82, and 2.66 for the Parallel Transport frame. The individual tubings used to display
these curves are in fact created using the Parallel Transport frame for each individual curve.
In Figure 1, we introduced the question of finding an optimal framing of a particular (3,5)

torus knot whose almost-optimal Parallel Transport framing was not periodic. In Figure 25, we
show the solution to this problem achieved by clamping the initial and final quaternion frames
to coincide, then letting the Evolver pick the shortest quaternion path for all the other frames.
It would be possible, as in the case of the (2,3) torus knot framing shown in Figure 13, to have
different conditions produce a framing solution containing its own reflected image rather than
having a distinct reflected image as is the case for Figure 25.
The types of solutions we find are remarkable in that they should be essentially the same for
all reparameterizations of the curve; regardless of the spacing of the sampling, the continuous
surface of possible frames is geometrically the same in quaternion space, so paths that are minimal
for one sampling should be approximately identical to paths for any reasonable sampling. On
the other hand, if we want special conditions for certain parameter values, it is easy to fix any
number of particular orientations at other points on the curve, just as we fixed the starting points
above; derivative values and smoothness constraints leading to generalized splines can be similarly
specified (see [2]).
Surface Patch Framings. A classic simple example of a surface patch framing problem was
presented in the discussion of Figures 2 and 22. This problem can also be handled by the Evolver:
44
0 0.5 1
-1 -0.5
1
0.5
5
0
KA
A
A
A
-0.5
.5
A A
K
A
A A
A A
-1 A A
1 A A
0.5
0
-0.5 A
-1 A
A
(a) (b)
0 0.5 1
-1 -0.5
1
0.5
5
0
KAA
A
A
-0.5
.5
A A
K
A
A A
A A
-1 A A
1 A A
0.5
0
-0.5
A
-1 A
A
(c) (d)
Figure 25: Optimization of the non-periodic parallel transport frame of the (3,5) torus knot in-
troduced in Figure 1 to produce a nearby periodic framing. (a) The original quaternion parallel
transport frame used to produce the tubing in Figure 1(b,c). (b) The frame mismatch, repeated for
completeness. (c) The result of fixing the final frame to coincide with the initial frame, leaving
the other frames free to move on the constraint rings, and minimizing the resulting total length in
quaternion space. The length of the original curve was 13.777 and that of the final was 13.700, not
a large difference, but noticeable enough in the tube and the quaternion space plot. (d) Closeup of
the corresponding framing of the knot in ordinary 3D space, showing that the mismatch problem
has been successfully resolved. This tube can now be textured, since the frames match exactly.
45
we choose an initial quaternion frame for the mesh corresponding to one of the arbitrary choices
noted, and minimize the area in quaternion space subject to the constraints that the normals remain
unchanged, and hence the frame choices may only slide around the rings depicted in Figure 22(b).
The results are shown in Figures 26, and 27. As a test, we started one case with a random initial
state with a range of 2 in the starting values. All converged to the same optimal final framing.
A basic observation is that while none of the standard guesses appeared optimal, the Geodesic
Reference frame is very close to optimal for patches that do not bend too much.
Minimal Surfaces. Minimal surfaces possess many special properties following from the fact
that the mean curvature is everywhere the vanishing sum of two canceling local principal curvatures
[12]. We present a family of classic examples here that is remarkable for the fact that the usual
framings are already very close or exactly optimal; thus we do not have much work to do except to
admire the results, though there may be some interesting theorems implicit that would be beyond
the scope of this paper to pursue.
In Figure 28(a,b,c), we present the following classical minimal surfaces:
xcatenoid (u; v ) = cos u cosh v x^ + sin u cosh v y^ + v ^z (47)

xhelicoid(u; v ) = v cos u x^ + v sin u y^ + u ^z (48)
xenneper (u; v ) (u , u =3 + uv ) x ^ + (v , v =3 + vu ) y
^ + (u , v ) ^
z
3 2 3 2 2 2
= (49)
The quaternion Gauss map choices determined by these parameterizations and by the Geodesic
Reference algorithm are shown in Figure 29. The coordinate-based catenoid map and helicoid
map are 4 double coverings, while Enneper’s surface curiously has a coordinate system map that
is exactly identical to the Geodesic Reference framing. For the periodic framings of the catenoid
and helicoid, we find the noteworthy result that the Geodesic Reference frame, which has a disjoint
quaternion reflected image, is a minimum under variations of the surface that is distinct from the
quaternion frames derived from the coordinate systems which are also minima, but contain their
own q ! (,q ) reflected images. The Enneper surface quaternion frames, which are the same,
appear to move very slightly around the borders under minimization, but it is not clear to what
extent this is significant as opposed to a numerical border effect in the variation. The resulting
3D frame triads are shown in Figure 30 for comparison. A theoretical analysis of the general
properties of quaternion Gauss maps for minimal surfaces is beyond the scope of this paper, but
experimentally we see that there could be very interesting general properties.
Manifolds. For general manifolds, one must treat patches one at a time in any event, since global
frames may not exist at all. Although the locally optimal patches cannot be globally joined to one
another, we conjecture that some applications might benefit from the next best thing: matching
boundary frames of neighboring patches using transitional rotations (see, e.g., [29, 13]). We have
carried this out explicitly for simple cases, but omit it here for brevity.
Extensions to Other Domains. We have focussed for expository purposes in this paper on
frames with intrinsic natural constraints imposed by the tangents to curves and normals to sur-
46
0.4 0.4
0.2 0.2
0 0
0.2 0.2
0.1 0.1
0 0
-0.1 -0.1
-0.2 -0.2
0 0
0.2 0.2
0.4 0.4
(a) (b)
-0.2
0.5 0 0.4
0.2
0.25 0.4
0 0.2
-0.25
.25
1
0
0.2
0.5
0.1
0
0
-0.1
-0.2
-0.5
0.2
0.4
(c) (d)
Figure 26: (a) The initial Geodesic Reference quaternions for the small patch shown in Figure
2. (b) Initial quaternions from parallel transporting the vertex frame down one edge, and then
across line by line. (c) A random starting configuration with the single same fixed corner point
as (a) and (b) and a range of , to + relative to the Geodesic Reference frame. (d) The result
of minimization of the quaternion area is the same for all starting points. The relative areas
are: 0.147, 0.154, 0.296, and 0.141, respectively. Thus the Geodesic Reference is very close to
optimal.
47
(a) (b)
(c) (d)
Figure 27: The 3D frame configurations corresponding to the quaternion fields in Figure 26. (a)
The Geodesic Reference frame. (b) Two-step parallel transport frame. (c) Random frames. (d)
The frame configuration resulting from minimizing area in quaternion space.
48
faces. However, the method extends almost trivially to applications involving externally specified
constraints on frames. Geometric construction algorithms based on extrusions reduce to the tubing
problem. For ordinary camera control interpolation, one could constrain any direction of the cam-
era frame to be fixed by calculating its appropriate constraint ring in the quaternion Gauss map,
and then extend a method like that of Barr et al. [2, 34]) to smoothly compute intermediate frames
subject to the constraints. For more general constrained navigation methods like those described
by Hanson and Wernert [20]), the camera vertical direction could be fixed at chosen points over the
entire constraint manifold, and the remaining frame parameters determined by optimization within
the manifold of ring constraints, possibly subject to fixing entire key-frames at selected locations
or boundaries.
7 Conclusion and Future Directions

We have introduced a general framework derived from the quaternion Gauss map for studying and
selecting appropriate families of coordinate frames for curves and surface patches in 3D space.
Minimizing length for quaternion curve maps and area for surfaces is proposed as the appropriate
generalization of parallel transport for the selection of optimal frame fields. These smooth frames
can be used to generate tubular surfaces based on the space curves, thus allowing their effective
display on polygon-based graphics engines. The analogous results for surface patches allow the
selection of optimal local coordinate systems that may be adapted for display purposes and related
applications such as oriented particle systems. Our principal new tool is the space of all possible
frames, a manifold of constraints immersed in the space of quaternion frames. By defining energies
and boundary conditions in this space one can produce a rich variety of application-adapted criteria
for specifying optimal families of frames.
Topics for future investigation include the treatment of manifolds in higher dimensional spaces,
improved interfaces for visualizing the quaternion optimization process and its results, and further
analysis of the pure mathematics implied by the general framework. N -dimensional generaliza-
tions of the Frenet frame equations have been studied in the literature (see, e.g., Forsyth [10]),
but the analogs of quaternions in higher dimensions are much more complex and involve Clif-
ford algebras and the corresponding Spin groups (see, e.g., Lawson and Michelson [26]). Special
simplifications do occur for the 4D case, however, allowing a treatment in terms of pairs of unit
quaternions (see, e.g., [15]); this case must in fact be investigated to produce a more rigorous
formulation of the 4D surface tubings proposed in [17]. Among other applications that may be ap-
proached by the quaternion formulation of coordinate frames we note the description of anisotropic
surfaces (see, e.g., Kajiya [22]), the quaternion generalization of bump-mapping, and the dynamics
of anisotropic particle systems. Another possible application could be the determination of optimal
configurations for long-chain molecules and similar 1D and 2D structures. There are thus ample
challenges for future work.
49
Acknowledgments
The author gratefully acknowledges the cordial hospitality of Claude Puech and the members of
the iMAGIS project, a joint project of CNRS, INRIA, Institut National Polytechnique de Grenoble,
and Université Joseph Fourier. We were also fortunate to have the collaboration of Hui Ma in
developing the early foundations of this work. Finally, we are deeply indebted to Ken Brakke for
his readiness to provide a wealth of mathematical insight as well as assistance in effectively using
the Evolver system. This research was made possible in part by NSF infrastructure grant CDA
93-03189.
50
(a) (b) (c)
Figure 28: (a) The catenoid, a classic minimal surface in 3D space with a natural orthonormal
parameterization. (b) The helicoid. (c) Enneper’s surface.
51
1 1 1
0.5 0.5 0.5
0 0 0
-0.5 -0.5 -0.5

1 1 1
-1 0.5 -1 0.5 -1 0.5

-1 0 -1 0 -1 0
-0.5 -0.5 -0.5
0 -0.5 0 -0.5 0 -0.5
0.5 0.5 0.5
-1 -1 -1
1 1 1
(d) (e) (f)
1 1 1
0.5 0.5 0.5
0 0 0
-0.5 -0.5 -0.5

1 1 1
-1 0.5 -1 0.5 -1 0.5

-1 0 -1 0 -1 0
-0.5 -0.5 -0.5
0 -0.5 0 -0.5 0 -0.5
0.5 0.5 0.5
-1 -1 -1
1 1 1
(a) (b) (c)
Figure 29: The Geodesic Reference quaternion frames of (a) the catenoid, (b) the helicoid, and (c)
Enneper’s surface. (d, e, f) The corresponding quaternion Gauss maps determined directly from
the parameterization. Both the catenoid and the helicoid fail to be cyclic in quaternion space
without a 4 turn around the repeating direction, so these are doubled maps. The Enneper’s
surface framing turns out to be identical to its Geodesic Reference frame.
52
(d) (e) (f)
(a) (b) (c)
Figure 30: The 3D Geodesic Reference frames displayed directly on the surfaces of (a) the
catenoid, (b) the helicoid, and (c) Enneper’s surface. (d,e,f) The 3D frames computed directly
from the standard parameterizations; since Enneper’s surface is the same, we show in (f) a differ-
ent viewpoint of the same frames.
53
Appendix: Quaternion Frames
2D “Quaternion” Frames
We provide below an exercise that may give some insight into the quaternion world. We show that
we may express the 2D frame equations in terms of a new set of variables exactly analogous to
quaternions in 3D. We begin by guessing a double-valued quadratic form for the frame:
h i " cos , sin # " a , b ,2ab # 2 2
N a ,b :
^ T ^ =
sin cos 2ab
= (50) 2 2
We can easily verify that if a + b = 1, this is an orthonormal parameterization of the frame, and
2 2
that is related to (a; b) by the half-angle formulas:

a = cos(=2); b = sin(=2) :
If desired, the redundant parameter can be eliminated locally by using projective coordinates such
as c = b=a = tan(=2) to get the form (compare Eq. (30))
h i " #
N
^ T^ =
,2c : 1 1 ,c 2
1+c 2c 1,c 2 2 (51)

" # " #
If we now define W =
a , b , b , a
b a and W = a ,b , then we may write
1 2
" # " #
2W
a0 ^ 0 ; 2W a
=N
0
^0
=T
1
b0 b0 2
and we may also express the right-hand side of the 2D frame equations as
" # " #
W + ,0 ab = +vT^ :
1
0
^ 0 = ,vN
The analogous expression for W yields T ^ . Matching terms and multiplying by W T =
2 i
W ,i , we find that the equation
1
" # " # " #

a0
= v
, a 1 0
b0
2 + 0 b (52)
contains both the frame equations T ^ 0 = ,N^ and N^ 0 = +T

^ , but now in 2D “quaternion” space!
If we take the angular range from 0 ! 4 instead of 2 , we have a 2 : 1 quadratic mapping from
(a; b) to (N ^ ) because (a; b) (,a; ,b) in Eq. (50).
^; T
Equation (52) is the square root of the frame equations (note the factor of (1=2)). The curvature
matrix is basically g ,1 dg , an element of the Lie algebra for the 2D rotation “spin group,” and takes
the explicit form,
" #" # " #
a b a0 ,b0 aa0 + bb0 ,ab0 + ba0 :
,b a b0 a0 =
ab0 , ba0 aa0 + bb0
54
Here aa0 + bb0 =0 due to the constraint a2 + b2 = 1, and
0 " " 0 # #
ab0 , ba0 = cos
cos

, sin 2 , 2 sin 2
2 2 2
=
0 ;
2
giving the identification v = 0 when we pull out the factor of 1=2 as in Eq. (52). The actual group
properties in (a; b) space follow from the multiplication rule (easily deduced from the formulas for
the trigonometric functions of sums of angles)
a; b) (~a; ~b) = (aa~ , b~b; a~b + ba~) ;

(
which is in turn isomorphic to complex multiplication with (a; b) = a + ib.This is no surprise,

since SO(2) and its double covering spin group are subgroups of the corresponding 3D rotation
groups, and complex numbers are a subset of the quaternions.
3D Quaternion Frames
We next outline the basic features of quaternion frames; see, e.g., [1] for a nice textbook treatment
of quaternions and their properties.
A quaternion frame is a four-vector q = (q0 ; q1 ; q2 ; q3 ) = (q0 ; ~q) with the following features:
Unit Norm. If we define the inner product of two quaternions as
qp=q p 0 0 + qp
1 1 + qp 2 2 + qp ;
3 3 (53)
then the components of a quaternion frame obey the constraint
q q = (q ) 0
2
+( q)
1
2
+( q)
2
2
+( q) 3
2
=1 ; (54)
and therefore lie on S3 , the three-sphere embedded in four-dimensional Euclidean space R4 .

Multiplication rule. The quaternion product of two quaternions p and q is defined to give a
positive cross-product in the vector part, and may be written as
p q = (p q , p q; p q + q p + p q) ;
0 0 0 0
or more explicitly in component form as

2 [p q] 3 2p q ,p q ,p q ,p q 3
6 77 66 p q + p q + p q , p q
0 0 1 1 2 2 3 3
77
p q = 664 [[pp qq]]
0
1
75 = 64 p q + p q + p q , p q
1 0 0 1 2 3 3 2
75 : (55)
2 0 0 2 3 1 1 3
[p q ] p q +p q +p q ,p q
2
3 3 0 0 3 1 2 2 1
This rule is isomorphic to left multiplication in the group SU(2), the double covering of the
ordinary 3D rotation group SO(3). What is more useful for our purposes is the fact that it
55
is also isomorphic to multiplication by a member of the group of (possibly sign-reversing)
orthogonal transformations in R4 :
2 p ,p ,p ,p 32 q 3
6 0 1 2 3
77 66 q 0
77
p q = [P ] q = 664 pp pp ,pp ,pp
1 0 3 2
75 64 q1
75 ; (56)
2 3 0 1 2
p ,p p p 3 2 1 0 q3
where [P ] is an orthogonal matrix, [P ]t [P ] = I4 ; since [P ] has only 3 free parameters, it does
not itself include all 4D rotations. However, we may recover the remaining 3 parameters by
considering transformation by right multiplication to be an independent operation, resulting
in a similar matrix but with the signs in the lower right-hand off-diagonal 3 3 section
reversed. (This corresponds to the well-known decomposition of the 4D rotation group into
two 3D rotations; see, e.g., [14].)
If two quaternions a and b are transformed by multiplying them by the same quaternion p,
their inner product a b transforms as
( p a) (p b) = (a b)(p p) (57)
and so is invariant if p is a unit quaternion frame representing a rotation. This also follows
trivially from the fact that [P ] is orthogonal.
The inverse of a unit quaternion satisfies q q ,1 = (1; 0) and is easily shown to take the form
q,1 = (q0; ,q). The relative quaternion rotation t transforming between two quaternions
may be represented using the product
t = p q, 1
=(pq 0 0 +p q; q p , p q , p q ) :
0 0
This has the convenient property that the zeroth component is the invariant 4D inner product
p q = cos(=2), where is the angle of the rotation in 3D space needed to rotate along
a geodesic from the frame denoted by q to that given by p. In fact, the 4D inner product
reduces to
p q, 1
+ q p, 1
= (2 q p; 0) ;
while the 3D dot product and cross product arise from the symmetric and antisymmetric
sums of quaternions containing only a 3-vector part:
pq (0 ; p) (0; q) = (,p q; p q)

pq+qp = ,2 p q
pq,qp = 2p q
Mapping to 3D rotations. Every possible 3D rotation R (a 3 3 orthogonal matrix)

can be constructed from either of two related quaternions, q = (q ; q ; q ; q ) or ,q =
0 1 2 3
56
(,q ; ,q ; ,q ; ,q ), using the quadratic relationship Rq (V) = q (0; V) q, , written
0 1 2 3
1
explicitly as
2 3
q +q ,q ,q
2
2q q , 2q q
2 2
2q q + 2q q
2
R = 64 2q q + 2q q q , q + q , q 2q q , 2q q 75 :
0 1 2 3 1 2 0 3 1 3 0 2
2 2 2 2
1 2 0 3 0 1(58) 2 3 2 3 0 1
2q q , 2q q 2q q + 2 q q
1 3 q ,q ,q +q
0 2 2 3 0 1
2
0
2
1
2
2
2
3
The signs here result from choosing the left multiplication convention Rp Rq (V) = Rpq (V) =
(p q ) (0; V) (p q ), . Algorithms for the inverse mapping from R to q require careful
1
singularity checking, and are detailed, e.g., in [31, 38].

The analog for Eq. (58) of the projective coordinates for 2D rotations noted in Eq. (51) is
obtained by converting to the projective variable c = q=q0 = tan(=2) n
^ and factoring out
q) ( q) 2
1 1
kck :
2 0
( 0 =
( q)
0
2
+q q =
1+q q=(q ) 0
2
=
1+ 2
We then find
2 3
1+c ,c ,c 2 2 2
c c , 2c
2 c c + 2c 7
2
1 6
R = 1 + kck 4 2c c + 2c
1
1 2
2
3
3
1,c +c ,c
1 2
2 2
3
2
1 3
2c c , 2c 5:
2 3
2
1 (59)
1 2 3
2c c , 2c 2c c + 2c 1,c ,c +c
2
2 2 2
1 3 2 2 3 1 1 2 3
Rotation Correspondence. When we substitute q (; n ^ ) = (cos 2 ; n

^ sin 2 ) into Eq. (58),
where n ^ n
^ = 1 is a unit three-vector lying on the two-sphere S , R(; n
2
^ ) becomes the
standard matrix for a rotation by in the plane perpendicular to n ^ . The quadratic form
ensures that the two distinct unit quaternions q and ,q in S correspond to the same SO(3)
3
rotation. For reference, the explicit form of R(; n

^ ) is [9]
2 3
c + (n ) (1 , c) n n (1 , c) , sn
2
n n (1 , c) + sn
R(; n^ ) = 64 n n (1 , c) + sn c + (n ) (1 , c) n n (1 , c) , sn 75 ;
1 1 2 3 3 1 2
2
1 2 3 (60) 2 3 2 1
n n (1 , c) , sn n n (1 , c) + sn c + (n ) (1 , c)
1 3 2 2 3 1 3
2
where c = cos , s = sin , and n ^ n^ = 1. For example, choosing the quaternion q =
(cos ; 0; 0; sin ) yields the rotation matrix

2 2
2 3
cos , sin 0
R = 64 sin cos 0 75 ;
0 0 1
^ = (1; 0; 0) and y
producing a right-handed rotation of the basis vectors x ^ = (0; 1; 0) around
the ^z axis.
Quaternion Frame Evolution. All 3D coordinate frames can be expressed in the form of
^ ;N
quaternions using Eq. (58). If we assume the columns of Eq. (58) are the vectors (N ^ ;T
^ ),
1 2
57
respectively, one can explicitly express each vector in terms of the following matrices
2
64 qq 0 q 1 ,q ,q 37
2 3
[W] 1 = 3 q 2 q q 5
1 0 (61)
,q q ,q q
2 2 3 0
3 1
[W] 2 =
64 ,qq
0
3 q
,q
2
1
q ,q 7
1
q ,q 5
2
0
3 (62)
q q q q
2 1 0 3 2
3
[W] 3 =
64 ,qq
2
1
q
,q
3
0
q
q
0
3
q 71
q 5;
2 (63)
q
0 ,q 1 ,q 2 q 3
with the result that N^ = [W ] [q ], N ^ = [W ] [q ], and T ^ = [W ] [q ], where [q ] is the

1 1 2 2 3
column vector with components (q0 ; q1 ; q2 ; q3 ). Differentiating each of these expressions and
substituting Eq. (9), one finds that factors of the matrices [Wi ] can be pulled out and a single
universal equation linear in the quaternions remains:
2 q0 3 2 0 ,k ,k ,k 3 2q 3
66 q0 0
77 v 66 +kx 0 x ,kyz +kyz 77 66 q 0
77
64 q01
75 = 2 64 +k +k 0 ,k 75 64 q1
75 : (64)
2 y z x 2
q03 +kz ,ky +kx 0 q3
The first occurrence of this equation that we are aware of is in the works of Tait [41]. Here
v(t) = kx0 (t)k is the scalar magnitude of the curve derivative if a unit-speed parameter-
ization is not being used for the curve. One may consider Eq. (64) to be in some sense
the square root of the 3D frame equations. Alternatively, we can deduce directly from
Rq (V) = q (0; V) q,1, dq = q (q,1 dq), and (q,1 dq) = ,(dq,1 q), that the
3D vector equations are equivalent to the quaternion form
q0 =
1
2
v q (0; kx ; ky ; kz ) = v q (0; k)
1
2
(65)
(q, )0
1
= , 21 v (0; k) q, ; 1
(66)
where k = 2(q0 dq , q dq0 , q dq), or, explicitly,
k0 = 2(dqxqx + dqy qy + dqz qz + dq q ) = 0 0 0
kx = 2(q dqx , qx dq , qy dqz + qz dqy )

0 0
ky = 2(q dqy , qy dq , qz dqx + qx dqz )

0 0
kz = 2(q dqz , qz dq , qx dqy + qy dqx ) :

0 0
Here k0 =0 is the diagonal value in Eq. (64).

The quaternion approach to the frame equations exemplified by Eq. (64) (or Eq. (65)) has
the following key properties:
58
– q(t) q0(t) = 0 by construction. Thus all unit quaternions remain unit quaternions as
they evolve by this equation.
– The number of equations has been reduced from nine coupled equations with six or-
thonormality constraints to four coupled equations incorporating a single constraint
that keeps the solution vector confined to the three-sphere.
Quaternion Surface Evolution. The same set of equations can be considered to work on
curves that are paths in a surface, thus permitting a quaternion equivalent to the Weingarten
equations for the classical differential geometry of surfaces as well. Explicit forms permit-
ting the recovery of the classical equations follow from re-expressing Eqs. (22) and (23) in
quaternion form. The curvature equation is essentially the cross-product of two derivatives
of the form of Eqs. (65,66), and thus obtainable by a quaternion multiplication:
qu q,v1
= , 14 q (0; a) (0; b) q, 1
= , 14 q (,a b; a b) q, 1
h i
= , 14 ,a b Î + (a b)x T^ 1 + (a b)y T^ 2 + (a b)z N^ : (67)
^ (0; N
Here Eq. (31) defines the quaternion frame vectors, N ^ ) = q (0; ^z) q ,1 , etc., and
we have introduced the identity element Î = (1; 0) = q (1; 0) q ,1 as a fourth quaternion
basis vector. The mean curvature equation has only one derivative and a free vector field; an
expression producing the right combination of terms is
^ q q ,1 + T
T ^ q q ,1 = , 21 q (,x^ a , y^ b; x^ a + y^ b) q, 1
1 u 2 v
h i
= , 12 ,(ax + by )Î + bz T^ , az T^
1 2 a , bx )N^ (:68)
+( y
Projecting out the N

^ component of these equations recovers the scalar and mean curvatures:
" #
K a u; v)
+ y( ,ax(u; v) = ax by , ay bx
= det
"
b u; v)
+ y( ,bx (u; v) #
H 1 +ay (u; v ) ,ax(u; v) = 1 (ay , bx ) :
=
2
tr
+by (u; v ) ,bx (u; v) 2
References
[1] A LTMANN , S. L. Rotations, Quaternions, and Double Groups. Oxford University Press,
1986.
[2] BARR , A. H., C URRIN , B., G ABRIEL , S., AND H UGHES , J. F. Smooth interpolation of ori-
entations with angular velocity constraints using quaternions. In Computer Graphics (SIG-
GRAPH ’92 Proceedings) (July 1992), E. E. Catmull, Ed., vol. 26, pp. 313–320.
59
[3] B ERGER , M. Geometry I, II. Springer Verlag, Berlin, 1987.
[4] B ISHOP, R. L. There is more than one way to frame a curve. Amer. Math. Monthly 82, 3
(March 1975), 246–251.
[5] B LOOMENTHAL , J. Calculation of reference frames along a space curve. In Graphics Gems,
A. Glassner, Ed. Academic Press, Cambridge, MA, 1990, pp. 567–571.
[6] B RAKKE , K. A. The surface evolver. Experimental Mathematics 1, 2 (1992), 141–165.

The “Evolver” system, manual, and sample data files are available by anonymous ftp from
geom.umn.edu, The Geometry Center, Minneapolis MN.
[7] E GUCHI , T., G ILKEY, P., AND H ANSON , A. Gravitation, gauge theories and differential
geometry. Physics Reports 66, 6 (December 1980), 213–393.
[8] E ISENHART, L. P. A Treatise on the Differential Geometry of Curves and Surfaces. Dover,
New York, 1909 (1960).
[9] F OLEY, J., VAN DAM , A., F EINER , S., AND H UGHES , J. Computer Graphics, Principles
and Practice, second ed. Addison-Wesley, 1990. page 227.
[10] F ORSYTH , A. R. Geometry of Four Dimensions. Cambridge University Press, 1930.
[11] G ABRIEL , S., AND K AJIYA , J. T. Spline interpolation in curved space. In State of the Art
Image Synthesis (1985). Siggraph ’85 Course notes.
[12] G RAY, A. Modern Differential Geometry of Curves and Surfaces with Mathematica, sec-
ond ed. CRC Press, Inc., Boca Raton, FL, 1998.
[13] G RIMM , C. M., AND H UGHES , J. F. Modeling surfaces with arbitrary topology using man-
ifolds. In Computer Graphics Proceedings, Annual Conference Series (1995), pp. 359–368.
Proceedings of SIGGRAPH ’95.
[14] H ANSON , A. J. The rolling ball. In Graphics Gems III, D. Kirk, Ed. Academic Press,
Cambridge, MA, 1992, pp. 51–60.
[15] H ANSON , A. J. Rotations for n-dimensional graphics. In Graphics Gems V, A. Paeth, Ed.
Academic Press, Cambridge, MA, 1995, pp. 55–64.
[16] H ANSON , A. J. Constrained optimal framings of curves and surfaces using

quaternion gauss maps. In Proceedings of Visualization ’98 (1998), IEEE Com-
puter Society Press, pp. 375–382. Missing pages 2 and 4 in original Proceed-
ings; correct version in second printing of Procedings, on conference CDROM, or
ftp://ftp.cs.indiana.edu/pub/hanson/Vis98/vis98.quat.pdf.
[17] H ANSON , A. J., AND H ENG , P. A. Illuminating the fourth dimension. Computer Graphics
and Applications 12, 4 (July 1992), 54–62.
60
[18] H ANSON , A. J., AND M A , H. Visualizing flow with quaternion frames. In Proceedings of
Visualization ’94 (1994), IEEE Computer Society Press, pp. 108–115.
[19] H ANSON , A. J., AND M A , H. Quaternion frame approach to streamline visualization. IEEE
Trans. on Visualiz. and Comp. Graphics 1, 2 (June 1995), 164–174.
[20] H ANSON , A. J., AND W ERNERT, E. A. Constrained 3d navigation with 2d controllers. In

Proceedings of Visualization ’97 (1997), IEEE Computer Society Press, pp. 175–182.
[21] H ART, J. C., F RANCIS , G. K., AND K AUFFMAN , L. H. Visualizing quaternion rotation.
ACM Trans. on Graphics 13, 3 (1994), 256–276.
[22] K AJIYA , J. T. Anisotropic reflection models. In Computer Graphics (1985), vol. 19, pp. 15–
[23] K IM , M.-J., K IM , M.-S., AND S HIN , S. Y. A general construction scheme for unit quater-
nion curves with simple high order derivatives. In Computer Graphics Proceedings, Annual
Conference Series (1995), pp. 369–376. Proceedings of SIGGRAPH ’95.
[24] K LOCK , F. Two moving coordinate frames for sweeping along a 3d trajectory. Computer
Aided Geometric Design 3 (1986).
[25] K UIPERS , J. Quaternions and Rotation Sequences. Princeton University Press, 1999.
[26] L AWSON , H. B., AND M ICHELSOHN , M.-L. Spin Geometry. Princeton University Press,
1989.
[27] M A , H. Curve and Surface Framing for Scientific Visualization and Domain Dependent
Navigation. PhD thesis, Indiana University, February 1996.
[28] M AX , N. L. DNA animation, from atom to chromosome. Journal of Molecular Graphics 3,

2 (1985), 69–71.
[29] M ILNOR , J. Topology from the Differentiable Viewpoint. The University Press of Virginia,
Charlottesville, 1965.
[30] M URTAGH , B. A., AND S AUNDER , M. A. MINOS 5.0 user’s guide. Technical Report SOL
83-20, Dept. of Operations Research, Stanford University, 1983.
[31] N IELSON , G. M. Smooth interpolation of orientations. In Computer Animation ’93 (Tokyo,

June 1993), N. Thalman and D. Thalman, Eds., Springer-Verlag, pp. 75–93.
[32] P LATT, J. C., AND BARR , A. H. Constraint methods for flexible models. In Computer
Graphics (SIGGRAPH ’88 Proceedings) (Aug. 1988), J. Dill, Ed., vol. 22, pp. 279–288.
[33] P LETINCKS , D. Quaternion calculus as a basic tool in computer graphics. The Visual Com-
puter 5, 1 (1989), 2–13.
61
[34] R AMAMOORTHI , R., AND BARR , A. H. Fast construction of accurate quaternion splines.
In SIGGRAPH 97 Conference Proceedings (Aug. 1997), T. Whitted, Ed., Annual Conference
Series, ACM SIGGRAPH, Addison Wesley, pp. 287–292. ISBN 0-89791-896-7.
[35] S CHLAG , J. Using geometric constructions to interpolate orientation with quaternions. In

Graphics Gems II, J. Arvo, Ed. Academic Press, 1991, pp. 377–380.
[36] S HANI , U., AND BALLARD , D. H. Splines as embeddings for generalized cylinders. Com-
puter Vision, Graphics, and Image Processing 27 (1984), 129–156.
[37] S HOEMAKE , K. Animating rotation with quaternion curves. In Computer Graphics (1985),
vol. 19, pp. 245–254. Proceedings of SIGGRAPH 1985.
[38] S HOEMAKE , K. Animation with quaternions. Siggraph Course Lecture Notes, 1987.
[39] S HOEMAKE , K. Fiber bundle twist reduction. In Graphics Gems IV, P. Heckbert, Ed. Aca-
demic Press, 1994, pp. 230–236.
[40] S TEENROD , N. The Topology of Fibre Bundles. Princeton University Press, 1951. Princeton
Mathematical Series 14.
[41] TAIT, P. An Elementary Treatise on Quaternions. Cambridge University Press, 1890.
62
Meshview: Visualizing the Fourth Dimension
Andrew J. Hanson Konstantine I. Ishkov Jeff H. May

Indiana University
Figure 1: Meshview’s interface window with a four-torus drawn Figure 2: Meshview’s key-frame animation interface controlling a set
using edges, vertices, and negative screen door transparency. of animated polygons; the texture coordinates are also key-framed.
Abstract device-independent as possible within the given framework. It has

been successfully compiled under Linux 2.2.2, SGI IRIX 5.2 to
Meshview is an interactive visualization system for viewing points, IRIX 6.5, and SUN SOLARIS 2.6. Meshview should in principle
curves, and two-dimensional manifolds embedded in 3D or 4D, be portable to any workstation that supports OpenGL and X/Motif
with the emphasis on handling 4D objects. All rigid motions in or appropriate simulators such as Mesa and LessTif.
3D and 4D can be performed under mouse (or 3D mouse) control, Our basic motivations for developing yet another 4D viewer in-
while key-frame animations support motions and deformations of stead of using an existing system such as, e.g., Geomview [17],
such objects. Meshview is written in C, OpenGL, and X/Motif with were twofold: (1) the available user interfaces, particularly for free-
the objective of being as compact, portable, and device-independent form rotational exploration, were poorly suited to both our perfor-
as possible within the given framework. The system has been used mance needs and our preferences for interfaces that allow heads-up,
successfully to do research on a variety of problems such as 4D context-free manipulation. We in fact conceived and implemented
viewing interfaces, mathematical visualization of classical higher a particularly interesting context-free interface, the “rolling ball”
dimensional geometry, Riemann surfaces, functions of two com- in 4D (described below) that permits complete exploration of the
plex variables, and 4D quaternion representations of 3D coordinate 6 degree-of-freedom orientation space in 4D with only 3 controller
frames. parameters. (2) Our need for robust high-performance custom in-
Keywords: four dimensions; curve and surface visualization terfaces for perceptual psychology measurements and research into
objects like quaternions and knotted spheres in 4D.
The basic design features of Meshview have evolved over a pe-
1 Introduction riod of several years, beginning with the first release of version 1.0
in July of 1994, and continuing with a number of refinements, in-
Meshview is an interactive 4D viewing system that fluidly dis- cluding screen-door transparency, animation, and texture, that were
plays points, curves, and two-dimensional manifolds embedded in added during 1998–1999 to version 1.2. In the next sections, we
3D or 4D, as well as key-frame animations representing motions outline the design philosophy and features of Meshview, point out
and deformations of such objects. It is written in C, OpenGL, its particular strong points, discuss the mathematical underpinnings
and X/Motif with the objective of being as compact, portable, and of its unique interface features, and present a selection of appli-
cations, focusing in particular on the suitability of Meshview for
Current address: Lucent Technologies, Holmdel, NJ building intuitions in classical mathematics and for the quaternion
y Current address: Intel Corporation, Santa Clara, CA methods used routinely in computer graphics.
2 Design Features of Meshview 3D context can be disentangled. Meshview supports a
choice between applying the 4D rolling ball to the current
The adjectives giving the overall goals of the design include: screen coordinates of the object’s 3D projection (“context-
Fast and general. Use display lists in OpenGL. Enhance
free,” the default), or applying to the object’s local 3D coor-
dinate system context (using the “axes” display to help show
the Geomview/OOGL MESH and OFF file formats. Support the context). The latter is useful for looking at different sides
color per object or color per vertex. of the object’s 3D projection while performing a 4D rotation.
Small. Keep as simple and independent as possible. This is especially useful for the 2D mouse interface.
Portable. Restrict to C, OpenGL, and X/Motif.

Sample data. The release includes a selection of example ge-
ometry files, including the 4D flat torus, Steiner surface (RP2
Freely distributable. Non-proprietary. embedded in 4D), 4D Fermat surfaces and much more. (See
the README file in the data directory for details, and see the
Support document generation. A straightforward image file color page of this article for examples.) A selection of short
generator is provided, and the state of any screen is easily programs for generating such files is also available.

saveable as a “setting” file for later restoration or refinement
of a view. Help. A simple online help file to remind the user of keyboard
shortcuts and interface options is provided.
The principal features of the design are:
Flexible file format. Reads extensions of the Ge- 3 Fundamental Methods.
omview/OOGL MESH, OFF and LIST file formats, plus its
own enhancements FRAMES, DOT, and LINE. 3D/4D Rolling Ball. The 4D rolling ball formula was derived in
[4], and this is the method implemented in Meshview for both for
Interactive examination support. Rotates/translates/scales the 2D mouse and the 3D mouse on the desktop. The remarkable
objects in 3D and 4D interactively under mouse control using property of this algorithm is that 4D orientation control requires
the 3D and 4D rolling ball models for rotations. exactly three control parameters, thus making it usable for a stan-
Momentum. The momentum option is available on all mo-
dard mouse with two buttons switching from (x; y )-plane control
to (x; z )-plane control, and making it ideally suited to the “flying
tions.
mouse” or CAVE “wand” 3-degree-of-freedom user interface de-
Drawing options. Optionally draws faces, edges, vertices, vices. Let X ~ = (X; Y; Z ) be a displacement obtained from the
normals, palette, unit sphere, the lighting vector, and a refer- 3-degree-of-freedom input device, and define r2 = X 2 + Y 2 + Z 2 .
ence set of 4D axes. Take a constant R with units 10 or 20 times larger than the average
value of r, compute D2 = R2 + r2 , compute the fundamental ro-
Pseudocolor palettes. A wide range of color palette options tation coefficients c = cos = R=D, s = sin = r=D, and then
for 4D depth color coding a provided based on the NCSA take x = X=r; y = Y =r; z = Z=r, so x2 + y 2 + z 2 = 1. Finally,
palette library. rotate each 4-vector by the following matrix before reprojecting to
Geometry locator panel. An interactive parametric space
the 3D volume image:
2 3
1 , x (1 , c) ,(1 , c)xy ,(1 , c)xz
2
“picker” (or point locator) is supplied for any MESH file or sx
list of MESH files. Any individual mesh in a set can be se- 6
4 , (1 , c)xy 1 , y (1 , c)
2
,(1 , c)yz sy 7
5
lected in turn. ,(1 , c)xz ,(1 , c)yz 1 , z (1 , c)
2
sz
Quaternion rotation panel. Quaternion multiplication is iso- ,sx ,sy ,sz c
morphic to multiplying a unit vector in the three-sphere S 3 by

an orthogonal 4 4 matrix that can be derived directly from
The 3D rolling ball method is correspondingly used for 3D orien-
tation control; it is basically the same formula except simplified by
rotations acting on the 3D coordinate frame. This panel vi-
sualizes the change in orientation of the 3D frame, the unit
setting z = 0 and reducing the matrix to 3 3.
quaternion to which it corresponds, and the action of the cor-
responding quaternion multiplication on the 4D object in the 4 Geometry
main window (which is not simply related to a 3D rotation).
Preservation of state. System state is saved for later recov-
Meshview data formats are strongly influenced by the OOGL (Ob-
ject Oriented Graphics Language) file format used by Geomview
ery, including current 3D and 4D viewing matrices, the cam- [17], but circumstances and practical experience with complex ge-
era setting, background color, light direction, and rendered ometries led us to deviate from strict adherence to the OOGL for-
ppm image of the current scene. This is useful for recon- mat.
structing the state of an illustration for a publication. Meshview 1.2 now supports MESH, OFF, LIST, FRAMES,
Restoration. Loads palettes and saved system states.
DOT, and LINE formats. MESH, OFF and LIST are very similar
to the OOGL file formats. The OOGL formats are not fully imple-
Face shading options. Surface facets can be flat shaded, mented (e.g., there is currently no support for files with more than
smoothly interpolated, depicted with one color on both sides, 4 dimensions), but on the other hand several enhancements have
depicted with two different colors for front and back surfaces, been added. The FRAMES, DOT and LINE formats are specific
or textured. In addition, any arbitrary palette can be used to to Meshview, where the FRAMES format is used for key frame
color code the 4D depth of each point in the current 3D pro- animation, and DOT and LINE are used to display dots and lines
jection: this is useful when rotating objects in 4D. respectively.
The overall syntax is quite straightforward, and is documented
Projection options. Both 3D and 4D permit perspective (po- in an accompanying README file that we cannot describe in de-
lar projection) and orthogonal projection. tail here due to space limitations. The data files are composed of
lists of points in 3D or 4D, various color attachments, and texture that begged to be explored interactively. Meshview in some sense
coordinates, all of which get translated in the implementation into was originally motivated by our need for our own customizable sys-
the obvious OpenGL representations. tem for this purpose. As a result, some of the earliest models cre-
ated for Meshview came from the film; a typical 4D “spun knot,”
that is not even knotted is shown in Figure 6. Many other “classic”
5 Selected Controls 4D mathematical figures are in the Meshview geometry library, in-
cluding the 4-torus (just the product of two circles) in Figures (1,3),
Below we present a selection of the possible controls in Meshview; and the crosscap/Steiner Roman Surface in Figures (4,5), which can
(u; v ) denotes the incremental mouse coordinates. in fact be rotated into one another in Meshview (a fact uncovered
3D viewing: during the early work by Banchoff on 4D visualization — see [1]);
leftbutton 3D rotation (3D rolling ball) R3(u,v) the equations for both of these figures can be found in the classic
middlebutton 3D translation in x-y plane T3(u,v,0) book Geometry and the Imagination [16].
rightbutton 3D translation along z axis T3(0,0,-v)
Shift+right 3D rotation around z axis R2(u)
3D lighting:
6.2 Complex Functions
Ctrl+left 3D rotation (3D rolling ball) R3(u,v) Some of the first author’s earliest work on mathematical visu-
Ctrl+middle 3D rotation around z axis R2(u) alization started with attempts to visualize the Fermat surfaces
4D viewing: [9, 10, 14], which are extensions of the Fermat-theorem equations
Shift+left xyw rotation (4D rolling ball) R4(u,v,0) to two complex variables of the form
Shift+middle xzw rotation (4D rolling ball) R4(u,0,-v)
n
(z1 ) + (z2 )
n=1
<Key>r Reset the 4D and 3D matrices and 3D
light direction, stop momentum.
<Key>F3 (3D mouse) Toggles 3D mouse. which are in effect the n-fold Riemann surfaces of the complex
* Left 3D mouse: 4D rolling ball equation
* Middle 3D mouse: 3D orientation and f (z ) = (1 z n )(1=n) :
,
position
Meshview provides any number of ways of exploring the 2D man-
* Right 3D mouse: reset
ifolds that result from solving these two real equations in four real
Appearances:
variables and looking at projections of the natural embedding in
<Key>1 both sides of face use same color
four real dimensions. In Figure 9, we show such a surface color-
<Key>2 two sides of face use different colors
coded by the phase transformations from the fundamental domain;
<Key>3 4D depth color coding
Figure 10 shows the same object with pseudocolor coded 4D depth.
<Key>4 texture coding
Figure 11 shows the two complex planes z1 = 0 and z2 = 0 su-
<Key>5,6,7 screen-door off, positive, negative
perimposed on the n = 3 Fermat surface. Solutions of the closely
related equations (z1 )m (z2 )n = 1 are shown in Figures 7 and 8.
Utilities:
<Key>f toggle face drawing (default on)
<Key>e toggle edges
<Key>v toggle vertices 6.3 Quaternion Visualization
<Key>n toggle normals
<Key>u toggle unit quaternion sphere The relation of 4D unit quaternions to rotations, orientations, and
<Key>p toggle palette camera frame interpolation has been familiar to computer graphi-
<Key>l toggle light ray cists since their relevance was pointed out by Shoemake in 1985
<Key>a toggle 4D orientation axes [18]. Meshview was used extensively to create the figures and ani-
Viewing: mations accompanying our research on mapping streamline orienta-
Ctrl+p 3D perspective projection (default) tion frames to quaternion spaces [11]. Subsequent research [5] em-
Ctrl+o 3D orthogonal projection ployed Meshview as well to visualize the nature of optimal quater-
<Key>w,x,y,z 4D projection along w(default), x, y, or nion curves and surfaces corresponding to frame assignments for
z-axis 3D curves and surfaces. Figures 12 and 13 show the quaternion
Shift+o 4D orthogonal projection (default) form of several possible tangent frame assignments for a (2,3) torus
Shift+p 4D polar projection knot; Figure 14 adds an actual quaternion surface representing the
space of all possible such frames.
6 Applications 6.4 Context for Perceptual Experiments

Meshview has been used in our laboratory to examine objects and Meshview’s basic facilities have been adapted to a series of exper-
create imagery for journal articles since its conception in 1994. iments currently underway in the Perception/Action laboratory at
With the addition of further features such as stereo, key-frame ani- the Indiana University Department of Psychology. The recently
mation, texture, and screen door transparency during the last year, added key-frame animation and deformation capabilities are essen-
many additional applications are possible. Among the specific ap- tial here; future work on the perceptual nature of 3D and 4D rigid
plications for which Meshview has been employed in its bare or versus elastic motion is planned in this context.
task-enhanced forms, we note the following:
6.5 Virtual Reality Features
6.1 4D Mathematical Visualization
Several features of the current Meshview support desktop virtual
The production of the video animation “knotˆ4” [15], which con- reality functionality. On a stereo-equipped SGI, the system will
cerned the visualization of knotted spheres embedded in four Eu- bring up a stereo screen that may be viewed with Stereographics
clidean dimensions, generated a family of very interesting objects CrystalEyes equipment. Various parameters can be adjusted to the
user’s taste. We generally assume a non-moving user so that head- [3] Carolina Cruz-Neira, Daniel J. Sandin, and Thomas A. De-
tracking, while feasible, is not a high priority in the unenhanced Fanti. Surround-screen projection-based virtual reality: The
desktop system. design and implementation of the CAVE. In James T. Ka-
Perhaps of more interest is the support (implemented under IRIX jiya, editor, Computer Graphics (SIGGRAPH ’93 Proceed-
6.x, but not difficult in general) for the Logitech 3D mouse, which ings), volume 27, pages 135–142, August 1993.
is a full six-degree-of-freedom device with four buttons. By em-
ploying the 4D rolling ball algorithm [4] in its purest form, the 3D [4] A. J. Hanson. Rotations for n-dimensional graphics. In
position alone of the 3D mouse can naturally control all six rotation Alan Paeth, editor, Graphics Gems V, pages 55–64. Academic
planes of a 4D mathematical object. This is accomplished by hav- Press, Cambridge, MA, 1995.
ing (x; y; z )-motions rotate in the (x; w),(y; w), and (z; w) planes,
[5] A. J. Hanson. Constrained optimal framings of curves and
respectively, while “rowing” circular motions of the mouse position
surfaces using quaternion gauss maps. In Proceedings of Visu-
in the (y; z ),(z; x), and (x; y ) planes, respectively, produce 4D ro-
alization ’98, pages 375–382. IEEE Computer Society Press,
tations in the (y; z ),(z; x), and (x; y ) planes themselves, exhausting
1998.
the entire 4D orientation space. The motion of a single 3D point at
the tip of the 3D mouse can thus be used to seek out any possible [6] A. J. Hanson and R. A. Cross. Interactive visualization meth-
projection from 4D into 3D. ods for four dimensions. In Proceedings of Visualization ’93,
pages 196–203. IEEE Computer Society Press, 1993.
7 Conclusion and Future Work [7] A. J. Hanson and P. A. Heng. Four-dimensional views of 3D
scalar fields. In Proceedings of Visualization ’92, pages 84–
The Meshview system is a minimalist approach to a very flexible 91. IEEE Computer Society Press, 1992.
and full-featured utility for examining and building intuition about
4D structures, e.g., 4D geometry and topology, two complex vari- [8] A. J. Hanson and P. A. Heng. Illuminating the fourth dimen-
ables, and quaternions. Its implementation strategy is to use only sion. Computer Graphics and Applications, 12(4):54–62, July
C, Motif, and OpenGL, thereby facilitating portability, maintain- 1992.
ability, extensibility, and compactness of design. The supported
data formats conform very closely to the OOGL formats imple- [9] A. J. Hanson, P. A. Heng, and B. C. Kaplan. Techniques for
mented by Geomview [17], with a handful of extensions. A vari- visualizing Fermat’s last theorem: A case study. In Proceed-
ety of desktop virtual reality techniques are incorporated, including ings of Visualization 90, pages 97–106, San Francisco, Octo-
the 4D rolling ball method for manipulating 4D displays, switched- ber 1990. IEEE Computer Society Press.
field Stereographics stereography, and the Logitech flying mouse. [10] A. J. Hanson, P. A. Heng, and B. C. Kaplan. Visualizing Fer-
We anticipate a limited CAVETM [3] implementation in the near mat’s last theorem. SIGGRAPH Video Review, 61(4), 1990.
future. 3:37 minute video animation.
Future plans include a number of ambitious extensions to cre-
ate robust and publicly available implementations of related high- [11] A. J. Hanson and H. Ma. Quaternion frame approach
dimensional visualization techniques, including interactive 4D rota- to streamline visualization. IEEE Trans. on Visualiz. and
tion and and volume rendering of 3-manifolds embedded in 4D [8], Comp. Graphics, 1(2):164–174, June 1995.
4D renderings of 3D scalar fields [7], and automatic generation and
fast interactive rendering of “thickened” 2-manifolds in 4D [6, 2]. [12] A. J. Hanson and H. Ma. Space walking. In Proceedings
Specific extensions involving quaternion visualization and quater- of Visualization ’95, pages 126–133. IEEE Computer Society
nion frame optimization are also envisioned, including quaternion Press, 1995.
maps of streamlines and stream surfaces in the manner of [11], and
[13] A. J. Hanson and E. Wernert. Constrained 3D navigation with
the automatic generation of optimal tubings of curves and framings
2D controllers. In Proceedings of Visualization ’97, pages
of surfaces as described in [5]. The general approaches to more
175–182. IEEE Computer Society Press, 1997.
sophisticated data navigation strategies such as those proposed in
[12, 13] would also be appropriate extensions to the Meshview fam- [14] A.J. Hanson. A construction for computer visualization
ily of interaction modes. of certain complex curves. Notices of the Amer.Math.Soc.,
The URL for Meshview is ftp://ftp.cs.indiana.edu/ 41(9):1156–1163, November/December 1994.
and the file is pub/hanson/Meshview.1.2.tar.gz.
[15] Andrew J. Hanson. knotˆ4. Video animation of knotted
Acknowledgments spheres in four dimensions. Published in Siggraph Video Re-
view 93, Scene 1 (1993).
This research was made possible in part by NSF infrastructure grant
[16] D. Hilbert and S. Cohn-Vossen. Geometry and the Imagina-
CDA 93-03189. Thanks are due to John N. Huffman for his assis-
tion. Chelsea, New York, 1952.
tance with the Linux port.
[17] Mark Phillips, Silvio Levy, and Tamara Munzner. Geomview:
An interactive geometry viewer. Notices of the Amer. Math.
References Society, 40(8):985–988, October 1993. Available by anony-
mous ftp from geom.umn.edu, The Geometry Center, Min-
[1] T. F. Banchoff. Beyond the third dimension: Geometry, com- neapolis MN.
puter graphics, and higher dimensions. Scientific American
Library, 1990. [18] K. Shoemake. Animating rotation with quaternion curves. In
Computer Graphics, volume 19, pages 245–254, 1985. Pro-
[2] R. A. Cross and A. J. Hanson. Virtual reality performance for ceedings of SIGGRAPH 1985.
virtual geometry. In Proceedings of Visualization ’94, pages
156–163. IEEE Computer Society Press, 1994.
Figure 3: 4D depth colored 4-torus Figure 4: Crosscap = 4D rotated Roman surface Figure 5: Steiner Roman surface
Figure 6: Twist-spun trefoil knot Figure 7: z1 z2 = 1 Figure 8: z1 (z2 )2 = 1
Figure 9: N = 4 Fermat surface coded by 2D Figure 10: N = 4 Fermat surface with color Figure 11: N = 3 Fermat surface with z1 = 0
complex phase transform coded 4D depth and z2 = 0 complex planes
Figure 12: Quaternion Frenet frame of (2,3) Figure 13: Selection of alternate quaternion tan- Figure 14: Quaternion manifold of allowed (2,3)
torus knot with color coded 4D depth gent frames for (2,3) torus knot torus knot tangent frames

Siggraph2001 Visualizingquaternions

Uploaded by

Copyright:

Available Formats

Siggraph2001 Visualizingquaternions

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Siggraph2001 Visualizingquaternions

Uploaded by

Copyright:

Available Formats

Visualizing Quaternions

Course Notes for SIGGRAPH 2001

3 Visualizing Quaternion Geometry 5

6 MeshView and Mathematica Demonstration Software 7

Paper: “Geometry for N-dimensional Graphics,” Andrew J. Hanson

Paper: “Quaternion Frame Approach to Streamline Visualization,” A.J. Hanson and H. Ma

I. (45 min) Introduction to Rotation Representations. Develop formulas and techniques

3 Visualizing Quaternion Geometry

[16] S. A. Carey, R. P. Burton, and D. M. Campbell. Shades of a higher dimension. Computer

[27] H. Flanders. Differential Forms. Academic Press, New York, 1963.

[29] A. R. Forsyth. Geometry of Four Dimensions. Cambridge University Press, 1930.

[30] George K. Francis. A Topological Picturebook. Springer Verlag, 1987.

[31] Herbert Goldstein. Classical Mechanics. Addison-Wesley, 1950.

[35] W.R. Hamilton. Lectures on Quaternions. Cambridge University Press, 1853.

[43] A. J. Hanson and P. A. Heng. Four-dimensional views of 3d scalar fields. In Proceedings of

[55] John G. Hocking and Gail S. Young. Topology. Addison-Wesley, 1961.

[66] G. M. Nielson. Smooth interpolation of orientations. In N.M. Thalman and D. Thalman,

[77] D.M.Y. Sommerville. An Introduction to the Geometry of N Dimensions. Reprinted by Dover

[80] D. J. Struik. Lectures on Classical Differential Geometry. Addison-Wesley, 1961.

II: Visualizing Quaternion Geometry

 2D Frames: Simple example, complex numbers.  Quaternion Frames

IV: Clifford Algebras

 Quaternion Frame Evolution  Reflections and Rotations: New ways of look-

 Pin(N), Spin(N), O(N), and SO(N)

 Quaternion methods are now commonplace in Basic Issues

 That’s going to be our job today!

Task vs Strategy Simple Example: 2D Rotations

 Our task: Understand Rotations.

The Belt Trick:

A Fix for the Problem?

A Fix for the Problem? Or, with a = cos(=2), b = sin(=2),

which reduces back to a2 + b2 = 1.

Examine time-evolution of 2D frame (on our way to 3D): First

where (a; b) is precisely the same frame as (,a; ,b). _

Frame Evolution in (a; b):

constraints from orthonormality, for One true degree of free-

2D Quaternions . . . Rotation as Complex Multiplication

Rotation with no matrices!

a ib )(a + ib) = ei( +)=2 = A + iB = a a , b b + i(a b + ab )

where if we want the matrix, we write:

(a’a-b’b, a’b+ab’) (a’, b’) (a’a-b’b, a’b+ab’) (a’, b’)

The Geometry of 2D Rotations

The Geometry of 3D Rotations The Geometry of 3D Rotations

We begin with a basic fact: We begin with a basic fact:

Quaternion Frame Parameters

we need two steps:

4 +2n3 sin  0 ,2n1 sin  75 00000000

Quaternions and Rotations Quaternions and Rotations

Quaternions and Rotations Rotations and Quadratic Polynomials

= 0 0 1 Hint: set q1 = q2 = 0 or any other (

Quaternions and Rotations

This implies q is a point on 3-sphere in 4D.

HOW does q = (q0; q) represent rotations? Quaternions and Rotations . . .

LOOK at WHAT happens if you do TWO rotations?

NOTICE: Choosing R(q )R(q) = R(Q)

) EXPRESS in quadratic forms in q and LOOK FOR an analog

makes the R3 equation an IDENTITY .

Quaternions and Rotations . . .

RESULT: the following multiplication rule Algebra of Quaternions

2D Frames: Simple example, complex numbers. Quaternion Frames

Quaternion Frame Evolution Reflections and Rotations: New ways of look-

Pin(N), Spin(N), O(N), and SO(N)

Quaternion methods are now commonplace in Basic Issues

That’s going to be our job today!

Our task: Understand Rotations.

A Fix for the Problem? Or, with a = cos(=2), b = sin(=2),

a ib )(a + ib) = ei( +)=2 = A + iB = a a , b b + i(a b + ab )

4 +2n3 sin 0 ,2n1 sin 75 00000000

) EXPRESS in quadratic forms in q and LOOK FOR an analog

with q0 = cos(=2) and ~q = n

Complex numbers represent 2D frames.

Quaternion Geometry Quaternion Frames

The quaternion rotation by about ^

Action of rotating Identity by about ^

Derivative Computation Simpler. Initial velocity condi-

The Spherical Projection Trick: Visualizing

Quaternion Frames: ^n in quaternion tells how

Quaternion Curves: are like any other curve;

Move objects and object parts in an animated scene. KA

Move the camera generating the rendered viewpoint of

tures to surfaces. Line drawing useless.

Serret-Frenet frame: k2 = 0, k1 = (t) is the curvature,

Unit four-vector. Take q = (q0; q1; q2; q3) = (q0; q) to

Multiplication rule. Let q p be the quaternion product

Analogous treatment (given in Hanson Tech Note in Course 0.5