On the Parametrization of the Three-Dimensional Rotation Group

John Stuelpnagel

SIAM Review, Vol. 6, No. 4. (Oct., 1964), pp. 422-430.

Stable URL:
http://links.jstor.org/sici?sici=0036-1445%28196410%296%3A4%3C422%3AOTPOTT%3E2.0.CO%3B2-8

SIAM Review is currently published by Society for Industrial and Applied Mathematics.

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained
prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in
the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/journals/siam.html.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

JSTOR is an independent not-for-profit organization dedicated to and preserving a digital archive of scholarly journals. For
more information regarding JSTOR, please contact [email protected].

http://www.jstor.org
Sun Apr 8 09:51:05 2007
 SI.1M R E T I E W
Vol. b,No. 4, October, 1964

ON THE PARAMETRIZATION OF THE THREE-DIMENSIONAL

ROTATION GROUP*
JOHN STVELPIYAGELt

1. Introduction. The problem of parainetrizirig the group of rotations of

Euclidean 3-space has been of interest since 1776, when Euler first showed that
this group is itself a 3-dimensional manifold. h primary application of such a
parametrization occurs in the integration of the equations of inotion of a rigid
body. To describe the orientation of the body relative to its center of mass, we
assume given two sets of mutually orthogonal unit vectors, or franles, one frame
being attached to the body and illoving with it, the other being constant and
coiiicidirig mith the nloving frame at time t = 0. The illoving frame at time t
is obtained by applying a rotation X ( t) to the fixed frame, and X ( t ) satisfies the
differential equation z ( t ) = fl(t)X(t), mith X ( 0 ) = I, the identity matrix;
fl(t) is defined by the relation fl(t)v = v X w ( t ) for all 3-vectors v, where w(t)
is the angular velocity vector. We assume Q ( t ) is known, so it is necessary to
integrate the matrix differential equation, or equivalently, a systein of nine scalar
equations, to obtain X ( t ) . However, if it is possible to represent X ( t ) by a set of
less than nine parameters, then the given system is equivalent to a system mith
fewer than nine scalar equations, and the problem may be simplified.
In this paper we show why it is topologically inlpossible to have a global
3-dimensional parametrization without singular points for the rotation group.
This is a special case of a corollary to Brouwer's theorem on the invariance of
domain. We also point out that, although Hopf showed in 1940 that five is the
mininlulll riunlber of paranleters which suffices to represent the rotation group
in a 1-1 global manner, the so-called "quaternion method" of paranletrizing the
group in a 1-2 way, using 4 parameters, is sufficient for practical purposes. I n
addition, three 3-dimensional parametrizations, as well as Hopf's method of using
5 parameters, are examined.
This paper is aimed primarily a t those who have been led by their involvelnent
mith the practical applications of this problem to wonder if there were not a may
to iillprove the present methods of parametrizing rotations without adding
redundarlt parameters; while the answer is negative, it is possible, by adding only
one redundant parameter, to obtain a illethod of representing unrestricted rota-
tions, which leads to simpler differential equations than any of the other inethods
presented.
* Received by the editors RIarch 9, 1964
This research was supported in part by the r n i t e d States Air Force through the Air
Force Office of Scientific Kesearch, Office of Aerospace Research, under Contract S o .
AF 49(638)-1242, in part by the Sational Aeronautics and Space ad mini st ratio^^ under
Contract S o . NASw-845 and in part by the Office of Xaval Research under Contract Xo.
Sonr-3693 (00).
t Research Institute for Advanced Studies (RIAS), i212 Bellona Avenue, Baltimore.
Rlaryland. You at Air Arm Ilivision, f estinghouse Electric Corporation, Box 746, Balti-
more. LI~tryland
 THE THREE-DIMENSIOXAL ROTATIOX GHOUP 423

The author is grateful to Dr. IC. E. I<alnian, of ItIAiS,for suggesting that this
paper be written, as well a5 for illany helpful conversations.
2. Notation and preliminaries. \Ye shall use K to denote the set of coiilples
2 x 2 matrices of the fonn ti = (-Oa - a') ; taking as basis

we see that I< is a 4-diriiensiollal associative no~lconli~lutativedivision algebra

over the real numbers, and the n ~ a psending cvltil +am + +
a3ti3onto
+
$((YO ali +a 2 j +a3k) is an ison~orphisinof K with the yuaternions, where
i, j, k here represent the usual basis for the quaternions. The determinarit of ti
is the square of the cluaternion norin. U will stand for the subset of Ii of deter-
niinant 1, and KOthe subset of trace zero. The elenlents of U arc just the 2 X 2
conlplex unitary matrices of detern~inant1, so 17 is a group, and U is topologically
eyui~alentto the 3-sphere (the unit sphere in E ~ ) since
, any element of I' is of
the form

with ~ u , ' = 1. KOis spariried by the set Q = {ti1 , ti? , ti3/.I:or fixed u E C', a E K,, ,
define linear niaps of KOinto itself by F,,(ti) = utizr-', and A,(ti) = tix - zti,
for any K t KO. The inatrices of I?,, and A, with respect to Q will be denoted by
y ( u ) and 6 (.r ), respectively.
The rotation group will be denoted by R ; it consists of those orthogonal
3 x 3 matrices with deternli~iant1. I,, and 0, will designate the identity and zero
niatrices, respectively, of dinlension n; the subscript will be omitted except when
confusion is possible. The transpose of a vector or nlatrix will be indicated by a
prime, e.g., x' is the row vector obtained by transposing the column vector .T.
For any matrix A , Tr A denotes the trace of rl.
The theoreill of Brouwer on the invariance of donlain, to which we will appeal
in a later sectioil, is stated as follows, and is proved in Hurewicz arid Wallnlan
[ 2 ] :If and B are honieolllorphic subsets of a Euclidean space En and A is
open, then B is open.
3. The topology of R. The nlatrix y(u) of r,, with respect to Q is easily seen
to be

which is orthogonal for all u E C-, and has deteriniriant 1. Also, for u, v E li,
y ( u )y (v) = y (uv), so y is a group hoi~~omorphisi~l.
Since y is continuous, and
U is conlpact and connected, r ( U ) is a conlpact comlected subgroup of R. The
424 JOHN STUELPSAGEL

orily coillpact connected subgroups of R are known to be I, R, and the groups

of rotations about a fixed axis. Sirice r ( U ) leaves no axis fixed, y ( U ) = R.
Note that y ( u ) = r ( v ) if and only if u = -v, so y is a two-to-one map of ['
onto R. Recalling that U is topologically a 3-sphere, we see that R is topologically
equivalent to the sphere with antipodal points identified, that is, projective
3-space.
To find a 1-1 global parametrization of the rotation group using li parameters,
it is necessary to embed the rotation group R in the Euclidean space E" that is,
to find a differentiable 1-1 map with differentiable inverse which carries R into
E', and use the inlage points as representatives of the rotation matrices.
Since R is a 3-dimensional manifold, each point )* has a neighborhood U , which
is hoilleomorphic to an open subset of E3. If there were a honleonlorphism h of
R into E ~ then, h(U,) would be open in by Brouwer's theorem, so h ( R ) ,
being the union of all h( U,) for )* R, would be open in E'. But R is conipact,
arid h(R), being the contiriuous image of a conlpact space, would be compact.
I t is a well-kno~vnfact that no Euclidean space coritairis an open co~llpactsubset,
so there call exist no such homcomorphisin.
The inlpossibility of embedding R topologically in E~ was first proved by H.
Hopf in 1940 [I]. The proof is based on a knowledge of the hoillology ring of
projective 3-space, arid will not be iricluded here. I t is possible, however, to
embed R in @, as Hopf showed, and we shall exanline this embedding in the next
section.
4. Five- and six-dimensional parametrizations. -411 element of R is determined
when its first two colui~lnsare specified, since the third colunln is the cross-
product of these two. Thus the six-vector obtained by vertical juxtaposition of
these two columrls serves to parametrize the group in a 1-1 global manner.
So if X E R, and Xo denotes the 3 X 2 matrix obtained by deleting the last
colun~riof X, then the differential equation ~ o ( t = ) Q(t)Xo(t)is equivalent to
the equation X ( t ) = Q(t)X(t), but contairis only 6 scalar variables.
Let x = (xl , zz , za, z4 , x6 , x6)' be the column vector represeriting the first
two colunlris of the inatrix (1/ 4 2 ) ~ where
, X E R. Then we have the identities
x'x = 1, x ' J , ~ = 0, i = 1, 2, where

Since x'x = 1, the set -11 of all points satisfying the above conditioris is contained
in the unit sphere S j in E! If a is any point of s5not in M , we can project s5- ( a ]
stereographically orito the hyperplane orthogo~ialto a, and thus obtain an ein-
bedding of d l , which is topologically equivalerit to R, in E5.
To do this explicitly, let V be a 3 X 6 matrix with Va = 0, VV' = I g . Then
V'V is the projection alorig a onto the subspace of Eb orthogonal to a. For
x E ilI, let y = Vx/ ( 1 - a'x). This represents the point which is the intersectiori
of the line joiriing x and a with the hyperplane orthogonal to a. I t is defined for
all x E A//, since the denominator vsu~ishcsonly if z = a, but a f l f . The cor-
 THE THREE-DIMESSIOSAL ROTATIOS GROUP 425

resporlderice is 1-1, for if y is a 3-vector, then I."y is orthogorlal to a, and the

line joiriirig V'y and a intersects the unit sphere in a siilgle poi~ltx, where

If .c E JI, then y satisfies the equations a?,a(y'y - I)' +

4 y ' V ~ , a ( ~-' ~1 )
+ 3 y ' ~ . , ~ '=y 0, i = 1, 2. We now have the :-vector
g satisfying the two equa-
tioiis above representirig a point of R in a 1-1 manner, and we wish to find the
differerltial equation satisfied by y, if 2 = Ax, where

Diflerclitiating the equation for y above with respect to t, we ohtairi

7j =
V?
(1 - a'.~.) +(a'?) V.I. 1
- = - (yty + 1 ) ( V + gal)?
(1 - a ' ~ ) ~ 2

The resulting equation for y is clearly not as simple as the origirial lineal
equation for x, and there is no apparent advantage in the reductioil in the
liunlber of scalar variables by this method. I t is possible that an embedding in
E" inay be obtained I\-hich leads to a sinlpler equation for y. This parainetrization
is priinarily of interest because it uses the slllallest possible number of scalar
variables for an everywhere defined, 1-1, corltirluous representatioi~of R, and
because the given embedding is the most obvious and probably the simplest
which can be obtained with five parameters.
5. The quaternion method. As we saw in $3, there is a 2-1 correspondence
y between the quaterilions of unit nor111 and the elements of R. Given the dif-
z(
ferential equation t ) = R( t ) X ( t ) in R, we can determine a differential equation
) 6-such that ~ ( z ~ () t=
zi(t) = ~ ( t ) u ( ti11 ) X(t ), and we now indicate how this
is done.
+
Let a = a1~1 C Y ~ K Zf 0 1 3 ~ 3E K O ,and consider A, acting on KO. The matrix
of A, with respect to the hasis Q is
426 J O H S hTt-ELPNAGEL

and

then A, = - 9 . If ~ ( t is) the solution of ii(t) = u(t)u(t) such that ~ ( 0 = ) I,

it is easily seen that u ( t ) 6 G, since u(t) E KO. -Also, for any fixed K E K O ,
) (u;zC1) = Z~KZL-I- Z L K U - ~ Z ~ Z L=
6 ? , ( ~= - ~ ~ ( Z L K U-
- ' )(UKZL-')a
= -&(r,((~)).
, matris of I',, with respect to &, then^ = QX,
I t follo~vsthat if X = ~ ( u )the
and 7 thus maps solutions of u = UZL onto solutions of x = Q X . If u ( 0 ) = I ,
then X ( 0 ) = y (u(0) ) = I ,so the desired particular solution is obtained.
In terms of the real parameters ul , us , u? , u4 appearing in u, the differential
equation 7 i = au becon1c.s

I t should be noted that the original linear equation is trarisfornled into a linear
equation; this was riot the case with the 5-dimensional method, so this method
is obviously far superior to the previous one. Although the parainetrizatioll is
not 1-1, no difficulties arise, since y is a local homeomorphism.
I t would be reasonable to ask whether it might be possible to obtain a repre-
sentation of this form, that is, one-to-many, using orily three parameters, but
still possessing the property of being a local homeomorphism, and having no
singular points. The answer is no, for this would force the parameter set to be a
"covering space" of R, and it is k1101vr.n that the 3-sphere, which callnot be
represented topologically by less than 3 parameters, is the orily covering space
of R, except for R itself.
6. Three-dimensional parametrizations. As we showed earlier, no X-dimen-
sional parametrizatiori can be both global arid nonsingular; however, there are
a t least three such parametrizations in commori use, each of which has certain
advantages, and we present them here.
The Euler angles are defined in rnany different ways, depending on the prob-
lem to be solved. The definition adopted here is convenient for probleiils irirrolving
orientation of aircraft, etc., since the Euler angles 4, 8, fi correspond to the com-
rilonly used paranleters of roll, pitch, and yaw, respectively.
If

is in R, we define the Euler angles for X as follo~vs:Lct f = xI2 + .r.P, ,$ 2 0. If

,$ Z 0, define 4, 8, 9 by
 TIIE TIIIZEE-DIAIESSIONAL
KOTATIOY GROUP 427

If = 0, SO :c. = 1, then 8 = -zir/2, but 4 and # are not uriiyuely determined,

+
being subject only to the conditions cos (x74 #) = x j , sin (274 #) = -22. +
I n particular, we may, if we wish, always choose # = 0 if 8 = f x / 2 . This de-
termilies 4 uniquely, but the resulting parameters are riot contiriuous functions of
X at 8 = f x/2. The Euler angles enable us to factor X into a product of rota-
tions about the vertical, transverse, arid lorigitudinal axes of the moving rigid
body; in fact,
0

I t is clear that the Euler angles give a parametrization of the rotation group
except at the points 5 = 0, or 8 = & ~ , / 2 .
I ~ X= fix,where Q is as in $5,then it is seen by direct colnputation that

i
1 0 -sin 0

0 -sin 4 cos 4 cos 8,

Since the detelanlinant of the matrix on the left is cos 8, it is clearly singular if
6' = f~ , $ , 8 , $ are determined only if 8 does not take on these values. If it is
/ 2 so
known in advance that certain orientations of the rigid body cannot be assumed,
then we may be able to choose the original coordinate system in such a way that
these orientations correspond to the singular points. In this case, the Euler angles
furnish a satisfactory method for representing the necessary subset of R.
A second inethod of obtaining a 3-dimensional parametrization of the rotation
group is based on the facts that ( I ) for any 3 X 3 skew-symmetric matrix S,
exp S is orthogonal, and ( 2 ) any rotation matrix is the exponerltial of some skew-
sylnmetric matrix.
Let S be a 3 X 3 skew-symmetric matrix and a2 = -3 Tr S2,u 2 0. Then the
characteristic polyrloinial of S is X3 + U'X, so "
S -u's. The power series for
X = exp S may corlsequently be sinlplified, using the relatio~is8'" = (-I)"-'
. a2"-2~2,S2n+l = ( - )n u2 1 AS,~r
and collecting terms, to
sin u 1 - cos u
X = I+-S+ S2.
a a2
The characteristic roots of X are 1, cos u i i sin a. I t is not hard to see that
exp S1 = exp 8 2 if and only if 8 2 = 0 and ul2 = 21ix, or S1 = Sz +
( ~ / c K / u ~ ),S Z
for some integer 16, where U, = - $ Tr s:. 111 particular, if we restrict our at-
tention to those skew-symmetric matrices S for which a 5 A, then exp S1= exp 8 2
if and only if S1 = iS?and al = 03 = T.
Conversely, let X E R , and let a = Tr X, so x3- ax' +
a X - I = 0, and
- 1 5 a 5 3. Then - 1 5 ( a - 1 ) I2 5 1; hence there is a unique angle a,
0 5 a < x, with cos a = ( a - 1),/2. If a # - 1, 3, let
S'= -a(l+2cosa)
sir1 u
I + a ( l2+sinc o us a ) X - - fJ

2 sin a
x?.
'
428 JOHS STUELPNAGEL

if a =3, let S = 0. Then exp S = X, and S is skew-symmetric. If a = - 1, then

s2= (n2/2) ( X - I ) has two skew-symmetric solutions f8 , and exp S
= exp ( - 8 ) = X.
Using the correspondence above, we call parametrize the rotation group by the
set of skew-syn~metricmatrices S with a 5 a ; every rotation matrix corresponds
to at least one skew-symmetric matrix, and those rotations which are involutions
( X is an involution if X' = I ) correspond to two skew-symmetric matrices. If
we identify

with the vector s = (sl , s2 , sg)', then a = / s 1, and R is seen to be topologically

equivalent to the ball 1 s / 5 n with boundary points identified.
The original differential equation x = QX is transfornled by this substitution
into the equation

1
S = n - -2 ( n s - s n ) + (' - a 2a2 (a'2)) (s2n + ns2 - 2 s n s ) .
The derivation of this equation requires some lengthy computations, which are
omitted. Although the number of parameters has been reduced to three, it is
clear that the form of the transformed differential equation is considerably more
conlplex than that of the original. Also, the transformed equation has a pole at
a = 2n, just as we would expect fro111 the nature of the map X -+ S , since the set
of S for which -3 Tr s2= 4n2 is collapsed by the exponential inap into the
identity.
The final 3-din~ensionalparametrization we shall consider is kno~vrlas the
Cayley parametrization (not to be confused with the Cayley-Klein parameters),
and also uses 3 X 3 ske~v-symmetricmatrices to represent rotations. If S is skew-
symmetric, we again let a2 = - $ Tr s2, and now set

Then X is orthogonal, and the characteristic equation of X is A" aA2

+ a A - 1 = 0, with a = [ ( 3 - a 2 ) / ( 1 +a2)],SO the characteristic roots of X
+
are I, [ I / ( 1 a')] ( 1 - a' =t2ia). These roots are real only if a = 0, in which case
all roots are + I . Thus no rotation matrix having - 1 as an eigenvalue may be ob-
tained from a ske~v-symmetricmatrix in this manner.
Conversely, if X R, and a = Tr X, then, for a # - 1, set
 T H E THREE-DIMEYSIOSAL KOTATIOS GROUP 429

S is then ske~v-symn~etric, and this is the inverse of the abovc correspondence.

Differentiating this last equation, substitutiilg 2 = QX, and simplifying, we ob-
+
tain s = $(Sf28 - SQ QS- Q ) , a I<iccati matrix equation for 8.
In this case, if it is kno~vnbeforehand that Tr X is never - 1, this paranletriza-
tion serves to represent all allo~vedorientations.
7. Conclusion. In evaluating the usefulrless of a paranlet1 ization of R, several
factors illust be considered. Anlong these are (1) the rlunlber of parameters
needed, (2) the form of the transformed differential equations, (3) the suscepti-
bility to error of the new equations in machine integration of these ecluations,
and (4) the ease with which a desired output can be obtained when these
equations are integrated.
As we have seen, the 6-dimensional parametrization, using the first two columns
of a rotation nlatrix to describe it, leads to linear equations, and the output is in
a readily usable form, since X is very simply obtained from the given six pa-
ra~neters.
The 5-dinlensional parametrization leads to riorlli~learequations, and ail un-
desirable anlount of computation is necessary to obtain X as an output. This
nlethod, while using one less parameter than the previous method, does not
appear to have anything in particular to reconlmend it, arid it is included only
because it uses the sn~allestpossible ilunlber of parameters in a 1-1 global
paranletrization.
The 4-dimensional or quaternion method has the advantages of leading to
linear equations while using only one redundant parameter, and representing
the inost general possible nlotion of the body. At the same time, the coefficients
of Zi are obtained as quadratic functions of the coefficients of u.
As we showed in §3, no 3-dimensional paranletrizatiorl can be both global
and nonsingular. If the parametrization is global, i.e., every rotation ulatrix
deternlines sonle finite values of the parameters, then there nlust be points
where the parameter values are not uniquely defined, and in this case the de-
rivatives of the paranleters are obviously not defined, so the transformed differ-
ential equations becoille singular at these points, that is, the derivatives beconle
infinite. This occurs, for example, with the Euler angles and the exponential
paranletrization. On the other hand, the Cayley parametrization leads to a well-
defined differential equation, being nonsingular, but it does not represent any
rotation nlatrices of trace - 1, which is a distinct disadvantage, since this will
not eve11 allow 180" rotations about a fixed axis.
The only corn~~lorlly used methods among those presented here are the 6- and
4-parameter nlethods, arid the Euler angles. A coinparison of the advantages
and disadvantages of these nlethods is made by Itobinson in [3]; he concludes
that the quaternion method is the best, a t any rate from the standpoint of
analog computation, for handlirlg unrestricted rotations, although the Euler
angles are useful because of their simple interpretations as roll, pitch and yaw.
That is, the Euler angles themselves provide a usable output, whereas with the
quaternioll method, it is necessary to transforn~the solutiorl to the rotation
group after integrating.
430 J O H S STUELPXAGEL

IIEFERENCES
[ l ] H . HOPF,S y s t e ~ n esymtnetrischer H i l i n e a , ~ ' o r n l e nu n d euklidische Jlodelle der p ~ o j e k t i u e n
Raurne, 1-ierteljschr. Naturforsch. Ges. Zurich, 1940.
[2] IV. HUKEWICZ A N D H . W.ILL~I.IN, 1limen.sion T h e o r g , Princeton Cniversity Press, Prince-
ton, 1948.
[3] A. C. ROBINSON, O n the u s e of q u a t e i ~ n i o n si n s i t n u l a t i o n of rigid-body m o t i o n , Wright
Air Development Center Technical Report 58-17, D a y t o n , Ohio.