Quaternions and Rotations in $E^4$

Author(s): Joel L. Weiner and George R. Wilkens

Source: The American Mathematical Monthly, Vol. 112, No. 1 (Jan., 2005), pp. 69-76
Published by: Mathematical Association of America
Stable URL: http://www.jstor.org/stable/30037387
Accessed: 22/02/2009 13:08

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. 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/action/showPublisher?publisherCode=maa.

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 a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the
scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that
promotes the discovery and use of these resources. For more information about JSTOR, please contact [email protected].

Mathematical Association of America is collaborating with JSTOR to digitize, preserve and extend access to
The American Mathematical Monthly.

http://www.jstor.org
 2. S. V. Kerov, Interlacing measures, in Kirillov's Seminar on Representation Theory, Amer. Math. Soc.
Transl.Ser.2, no. 181,AmericanMathematical
Society,Providence,1998,pp.35-83.
3. D. Romik,Explicitformulasfor hookwalkson continualYoungdiagrams,Adv.Appl.Math.32 (2004)
625-654.

Laboratoirede Probabilitis, Universite Paris 6

Currentaddress:
Departmentof Mathematics, WeizmannInstituteof Science, Rehovot 76100, Israel
[email protected]

Quaternionsand Rotationsin E4
Joel L. Weiner and George R. Wilkens

1. INTRODUCTION. In 1843, Sir WilliamRowanHamiltoninventedthe quater-

nion algebra,which is customarilydenotedH in his honor.Soon after,people rec-
ognized thatquaternionscould be used to representrotationsin E3. In 1855, Arthur
Cayley discoveredthat quaternionscould also be used to representrotationsin E4.
Thisnote exploresCayley'srepresentation. Ultimatelywe use it to showthatanyrota-
tion in E4 is a productof rotationsin a pairof orthogonaltwo-dimensionalsubspaces,
a resultfirstprovedby EdouardGoursat[3].
In section2 we reviewthe algebraicstructureof H and show thatH has a natural
innerproductthatallowsus to identifyit withfour-dimensional EuclideanspaceE4. In
section3 we showthata pairp andq of unitvectors(also calledunitquaternions) in H
determines a rotation Cp,q: H -> H. According to Goursat's result, Cp,qis a product
of rotationsin a pairof orthogonalplanes.By this we meanthe following:thereexist
rotationsR1, R2 : H -* H and a pairof orthogonalplanes Vi and V2in H, such that
the restrictionsR1Iv2and R2Iv areidentitieson theirrespectiveplanesand

Cp,q = R o R2 = R2 O R1.

Thus,H = V1e V2,where V1I V2,and Cp,qrotatesvectorsin the plane V1through

a determinedangleat andvectorsin the plane V2througha determinedanglea2.
The principalgoals of this note are to proveTheorems1 and 2, which are stated
preciselyin section 5. Theorem1 not only provesGoursat'sresultfor Cp,q,but also
showsthatone can easily determinethe planes V1and V2andthe anglesat anda2 in
termsof p andq. Theorem2 establishesthateveryrotationin E4 canbe representedby
some Cp,q.Together,thesetheoremsproveGoursat'sresultfor everyfour-dimensional
rotation.
The observationthat Cp,q(V)= Vi (i = 1, 2) motivatesthe methodof proof.The
Vi are known as invariant subspaces for Cp,q.If we wish to see that Cp,qis indeed a
productof rotations,it is naturalto look firstfor invariantsubspacesof thattransfor-
mation.In section 4 we recall some elementaryresultsfrom the theoryof ordinary
differentialequationsthatarerelatedto subspacesandtwo-dimensionalrotations.Fi-
nally,in section5, we applythese resultsto findthe Cp,q-invariant subspacesandthe
rotationangles.

January2005] NOTES 69
2. THE QUATERNION ALGEBRA. Let H denote

{al + bi + cj + dk: a, b, c, d e R},

anddefineadditionandmultiplicationby a real scalarin component-wisefashion.In

so doing, H becomes a four-dimensionalreal vector space and {1, i, j, k} is a basis
for H. To define the productof two points in H, one simply assertsthe following:
multiplicationdistributesoveraddition;1 is the multiplicativeidentity;and

i2 j2 = k2 _1,
ij = -ji = k, jk = -kj = i, ki = -ik =j.

These operationsof additionandmultiplicationon Hlsatisfyall the axiomsfor a field,

exceptthe commutativityof multiplication.
It is convenientto decomposea quaternionintotwo partsthataretraditionallycalled
its scalarandvectorparts.If q = qol + qli + qj + q3k,thenwe write

q=qo+q,
where qo = qol and q = qli + qj + q3k. We call qo the scalar part and q the vector
part of q. It is straightforward
to checkthatthe product

pq = (po + P)(qo+ q) = (Poqo- P q) + (poq + qoP + p x q), (1)

wherep q andp x q are,respectively,the usualinnerproductandvectorcrossprod-
uct in E3.
Thereis anotherimportantoperationon H; it is called conjugation.If q = qo + q
belongs to H then q = qo - q is called the conjugate of q. Conjugation has several
nice properties,the mostimportantof whichis the following:

pq = qp

(notethe changein order).

Fromequation(1) it also follows that

qq = q= + q + q2 + q3 = q.

Thus, if we identifyH with Euclideanfour-spaceE4 by associatingq with the vec-

tor (qo, ql, q2, q3) and denote the Euclidean inner product of p and q by (p, q), then
qq = (q, q). Usingthefactthatqq and(q, q) arequadraticforms(i.e., eachis R-linear
in the two "slots"thatappearin theseexpressions),it is a simplematterto verifythat

pq + qp = 2(p, q). (2)

Note, in particular,thatp is orthogonalto q if andonly if pq + qp = 0.
Wedenotethe Euclideannormof a quaternionq by Iql.Since scalarscommutewith
everyquaternion,

Ipql2 = (pq, pq) = pqpq = pqqp = plql2p = Ip21q12.

This gives the followingimportantresult:

IPql= Ipllql. (3)

70 © THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 112

 Oncethe normof a quaternionis available,we can obtaina formulafor the inverse
of a quaternionthatis reminiscentof whatoccurswith complexnumbers.It is easy to
showthatif q # 0, then
q-1 q
q1 12

If q in H has Iq| = 1, then we call q a unit quaternion.For a unit quaternionq

we see that q-1 = q. If, in particular, u is a pure unit quaternion (i.e., if u = u),
then u-1 = -u. Thus every pure unit quaternionis a squareroot of -1. Also, by
equation(2), two pureunit quaternionsu andv are orthogonalif andonly if it is the
case that uv + vu = 0.
If q = qo + q is a unit quaternion,thenqo + Iq12= 1. Hence thereis a real num-
ber0 anda pureunitquaternionu suchthatq = 1 cos 0 + u sin0. Since u2 = -1, the
powerseriesexpansionof et leadsto
oo (u0)n

eun = n! =lcos0+usin0, (4)

n=o0

providingequivalentrepresentationsfor a unit quaternionq = qo + q = 1 cos 0 +

u sin0 = e". Note thatneitheru nor 0 is uniquelydeterminedby q. Whenq ±1,
sin = ±| | andu = +q/|q|; whenq = ±1, u can be anypureunitquatemion.
Wenote thateu acts like the usualexponentialas a functionof a complexvariable.
However,since the multiplicationin H is not commutative,if u and v are linearly
independentpureunit quaternions,it is not the case thateuee'Vis the same as either
eveuOor eue+v. However,since each componentof eu0is a differentiablefunctionof
0, it is not difficultto verifythat
d
-eu = -1 sin0 + u cos 0 = ueue= enuu.
dO
3. ROTATIONSIN E3. We introducethe R-lineartransformations representingleft
and right multiplicationin H. Let q be a quaternion.Then Lq : H -+ H and Rq :
H -- H are defined as follows:

Lq(x) = qx, Rq(x) = xq (x e H).

If q is a unit quaternion, then both Lq and Rq are orthogonal transformationsof H.

This is an easy consequence of equation (3). Specifically, when Iq| = 1

ILq(x)I = IqxI = Iq||x| = Ixi.

Thus,for unitquatemionsp andq, the mappingCp,q: H -+ H definedby

Cp,q = Lp 0 Rq = Rq o L

is also an orthogonaltransformation
of H. It is worthnotingfor laterapplications(in
Theorems2 and3) that

Cp1,91 0 CP2,q2 = PP2,q2ql'

We examine briefly the transformationCq,q,where q is a unit quatemion.For

the time being, we simply denote it by C. If we write q = e", where u is a pure

January2005] NOTES 71
unit quaternion, then C(x) = qxq = en"xe-ue. First, observe that C preserves scalar
quaternionsbecause C(1) = qlq = qql = 11 = 1. Since C is an orthogonal transfor-
mation of H, it must also preserve the orthogonal complement to the scalars, the space
of pure quaternions that we henceforth denote by E3. We restrict C to E3 and call the
resulting map C as well. Note that u = u is a member of E3.

Proposition 1. If q is a unit quaternion, then there exist a pure unit quaternion u

and a real scalar 0 such that q = eue. The transformation C : E3 - E3 defined by
C(x) = qxq is a rotation in the plane orthogonal to u through an angle 20.

Proof We have already shown that every unit quaternionhas an exponential represen-
tation. Choose u and 0 so that q = enu. Observe that C(u) = eUoue-u = i, because
u commutes with eu&.Thus C fixes the one-dimensional subspace L spanned by M,
hence fixes its orthogonal complement L' (C E3) as well. Let v = v be a unit vector
in L' and set w = uv = u x v. Notice that uv = -vu, since u and v are orthogonal.
This implies that ve-uo = eUv. Accordingly,

C(i) = eUeve-u° = e2u0v= cos(20)0 + sin(20)7. U

Notice that we can represent every rotation (i.e., every proper orthogonal transfor-
mation of E3) as Cq,qfor an appropriateunit quaternionq. A transformationis proper
if it is orientation-preservingor, in other words, if it has positive determinant.
For an elaboration of the topics presented so far the reader can refer to chapters 17
and 18 of the text by Michael Henle [4].

4. SOME FACTS ABOUT ORDINARY DIFFERENTIAL EQUATIONS. We

now turn our attention to Cp,qfor arbitraryunit quaternions p and q. It was Cayley
who first noticed that these are proper orthogonal transformationsof E4. As has been
known for some time (see [3]), such transformationsmust be the product of two ro-
tations in a pair of orthogonal two-dimensional subspaces of E4. We would like to
see how these rotations and subspaces are related to p and q. Coxeter elucidated this
relation in an earlier paper in this journal [2]. We intend to do the same using distinctly
different methods and, in fact, we will show from first principles that Cp,qis a product.
To do that we call upon one tool from the theory of ordinarydifferential equations.

Proposition 2. Let x : R -- Rn satisfy a kth-order linear homogeneous differential

equation, where 1 < k <_n. Then the image of x lies in a k-dimensional subspace
of R.

Proof We present the proof for the case k = 2, which suits our application. The reader
should be able to generalize this to any k.
Let xo = x(0) and xo = di/dt(0) be the initial position and initial velocity for the
given curve x. Additionally, suppose that x satisfies the second-order linear homoge-
neous differential equation

d2x dx
+ a- + x = 0,
dt2 dt

where a and / are differentiablereal-valued functions of t. From standardODE theory,

we know that when two solutions of this differential equation have the same initial
position and the same initial velocity, the two solutions are identical.

72 © THE MATHEMATICALASSOCIATIONOF AMERICA [Monthly 112

 Now supposethatf, (i = 0, 1) arereal-valuedfunctionsthatsatisfythe differential
equationf" + af' + fif = 0 and,in addition,fo(0) = 1, fo(0) = 0, fi(0) = 0, and
fi(0) = 1. Thenthe curvex(t) = fo(t)xo + fl(t)x satisfiesthe same ODE as x and
has the same initial positionand initial velocity.Thus x = fo(t)xo + fi(t)xo, so we
see thatthe imageof x lies in the subspaceof Wnspannedby xo andxo. I

Thatthe functionsfo and fi of the precedingproofexist is guaranteedby the stan-

dardtheoryfor linearordinarydifferentialequations[1]. The followingpropositionis
an easy consequenceof the proofof Proposition2. It will proveto be quiteuseful.

Proposition3. Supposethat x : IR- H satisfiesthe differentialequation

d2
+ W2X= 0,
dt2
where co > 0 is a constant,and that the initial position vector i(0) and the ini-
tial velocity vector di/dt(0) satisfy the conditions I|(0) = w-lldx/dt(O) and
(i(0), di/dt(0)) = 0. Theni(1) = R(i(0)), whereR is a rotationin theplane of the
imageof i throughan angle cwin the directionthatturnsi(0) towardw-'d*/dt(0).

Proof Followingthe constructionin the proof of Proposition2, we choose fo(t) =

cos(cot) and fi(t) = co-1 sin(wt). Then

dx(O)
x(t) = x(0) cos(cot) + d- -) sin(wt),
dt
whichshowsthat

x(1) = x(0) cos(co)+ cw-1-() sin(cw).

dt
ThattherotationR existsfollows fromthe assumptionsthat*(0) andco-di/dt (0) are
orthogonalvectorsandhavethe samelength. U

5. PROPER ORTHOGONALTRANSFORMATIONSOF E4. Let C be short-

handfor Cp,q,where for suitablechoices of pureunit quaternionsu and v and cor-
responding real numbers 0 and 4, p = eUOand q = ev'. We seek two-dimensional
invariantsubspacesfor C. If x in HIlies in some C-invariantsubspaceS, so does C (x)
for all integersn. Moreover,if Ct madesense for arbitraryrealt, we wouldexpectthe
sameto be trueof Ct(x). It is this observationthatmotivateswhatwe do next.
First,noticethatCt does makesense;in fact, for anyrealt let Ct be definedby

Ct(X) = eUtxert.

To each quaternion x we associate a curve x : IR-+ H defined by *(t) = Ct(x). We

will computetwo derivativesof x. Note thatthe usual formulasfor differentiatinga
productor a compositionapply,as the readercan checkby examiningthe components
in theseformulasor by consideringequation(1) with the realconstantsreplacedwith
real-valuedfunctions.Keep in mind, however,thatthe orderof termsin productsis
important.As the firstderivativeof x we obtain
dr
-(t) = uO eOt x eVrt + eUOtx evrt v =- u0 x(t) + i(t) v4. (5)
dt

January2005] NOTES 73
Differentiatingthe left-handandright-handsides of (5), at the sametime using (5) to
eliminatefirst-orderderivativesof x(t), we get
d2X
d(t) = -(02 + q2)i(t) + 20ui(t)v. (6)
dt2
Now if it happenedthat

ux(t)v = A(t)x(t), (7)

whereXis a real-valuedfunction,thenx would satisfya linearhomogeneoussecond-
orderordinarydifferentialequationwithrealcoefficients.By Proposition2, the image
of x would lie in a two-dimensionalsubspaceand necessarilythe span of x = i(0)
andC(x) = x(1) wouldbe an invariantsubspace.Note howeverthatl|u(t)vl = x(t)|,
becauseu andv areunitvectors.Thusif equation(7) wereto hold, thenX wouldhave
to be eitherthe constantfunction1 or the constantfunction-1 (since all the functions
we considerare continuous).In fact, we can simplify the conditionof equation(7)
withX = ±1.

Lemma 1. Forx in H, uxv = +x if and only if ui(t)v = +x(t) holdsfor all t.

Proof Since x = i(0), the if directionis obvious.To provethe only if direction,as-

sumethatuxv = ix. Apply Ct to bothsides of this equationto get
+Ct(x) = Ct(uxv) = eUtuxveet = uCt(x)v

(i.e., +i(t) = ui(t)v). U

Thuswe look for thosex in H thatsatisfyone of the linearequationsux ± xv = 0.

To do this we introducea basis for H. A naturalchoice is the set consistingof 1, u,
v, and uv. Of course,this is not a basis if u = iv, but it is otherwise.(Note: when
p = ±1 or q = +l, at leastone of u or v maybe chosenarbitrarily. Whenthis occurs,
we alwayschoose u andv to be orthogonal.)
Wefirstconsiderthe case whereu = ±v andlook for solutionsto ux ± xu = 0. It is
easy to see that1 andu aresolutionsto ux - xu = 0; in fact,theyforman orthonormal
basisfor the solutionsto thisequation.Onthe otherhand,the solutionsto ux + xu = 0
arethepurequaternionsx thatareorthogonalto u. In a minorabuseof notationwe use
UI (recallthatu is the vectorpartof u) to signifythis set of solutions.Since any pure
quaternionis necessarilyorthogonalto 1, the spaceIuIis the orthogonalcomplement
of the space spannedby 1 andu. Thusthe solutionspacesto the two equationsux +
xu = 0 give a decompositionof H into the sum of two two-dimensionalorthogonal
subspaces.
Now assumethatu 7 +v. One can then introducethe basis {1, u, v, uv} and in a
straightforward fashion find solutions.However,it is easier to guess solutions.For
example, x = + v satisfies ux - xv = 0. For this x, the curve x lies in a two-
u
dimensionalinvariantsubspace;hence di/dt(0) is also in that subspace.Using (5),
we see that for x = u + v
di
-(0) = uO (u + v) + (u + v) vq = (0 + 4)(uv - 1).
dt
A direct calculation confirms that x = uv - 1 is another solution of ux - xv = 0.
Moreover,since Ct is a rotationfor every t, Ix(t)I = IC (u + v) = |u + vl is a con-
stant. It follows that x(0) = u + v is orthogonal to i'(0) and thus to uv - 1. Observe

74 © THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 112

also that, since uv - 1 = (u + v)v, equation(3) implies that these two quaternions
have the same norm.A directcalculationgives the commonvalue of these norms:
2(1 + cos a), wherea (0 < a < 7) is the anglebetweenu andv. In a similarfash-
ion, we can showthatv - u anduv + 1 areorthogonalsolutionsof ux + xv = 0, each
havingnorm /2(1 - cos a). Finally,it is easy to checkthateachpairof vectorsis or-
thogonalto the otherpair.Thusthe vectorsu + v, uv - 1, v - u, anduv + 1 constitute
an orthogonalbasis for H.
Assume that u 7 -v, let x = u + v, and recall that this x satisfies ux - xv = 0 (or
equivalently uxv = -x). Here lu + vl = luv - 1| = 2(1 + cos a) even when a = 0.
Forthis x, (6) becomes
d2,
dt2(t) + (0 + 2(t) = 0. (8)

Also, by (5) evaluated at t = 0,

di
-(0) = 0ux + Cxv = (0 + ¢)(ux) = (0 + q)(uv - 1).
dt

Hence |0 + - Idxi/dt(0)I= I (0)I and (i(0), di/dt(0)) = 0. Thus we can invoke

Proposition3 andconcludethatC restrictedto the planespannedby u + v anduv - 1
is a rotationthroughthe angle |0 + ¢| in the directionthatturnsu + v towardsign(0 +
¢)(uv - 1). Statedmoresimply,C restrictedto theplanespannedby u + v anduv - 1
is a rotationthrough0 + ¢ in the directionthatturnsu + v towarduv - 1. In a similar
fashion,whenu 7 v we can showthatC restrictedto the planespannedby v - u and
uv + 1 is a rotationby 0 - ¢ in the directionthatturnsv - u towardsuv + 1. The
samekindof resultshold in the remainingcases whenu = ±v.
We consolidatewhatwe havelearnedinto ourfirsttheorem:

Theorem 1. Let p = euOand q = ev, where u and v are pure unit quaternions.
The orthogonal transformation Cp,q of HI is a product of two rotations in orthog-
onal planes. If u I ±v, then Cp,q rotates the plane spanned by u + v and uv - 1
through the angle \0 + 0I and the plane spanned by v - u and uv + 1 through the
angle 10 - I. If u = ±v, then the invariantplanes are the span of 1 and u and its or-
thogonal complement, and the rotation angles in appropriate planes are still \0 + |
and |0 - I|.

As an aside,we notethe followingspecialcase of Theorem1. When04 = 0, which

impliesthatp = 1 or q = 1, the differentialequation(8) is equivalentto (6). Thenfor
everyx the associatedcurvex satisfies(8), and we see thateverynonzerox lies in a
C-invariantplaneandis rotatedthroughthe sameangle,namely,l0 + I|.
We can now prove that every properorthogonaltransformationof H is of the
form Cp,qand so can be describedby Theorem1 as well. Firstnote that C = Cp,q
is propersince its determinantis 1. This follows fromthe fact thatdet(Ct) is a con-
tinuousfunctionof t and can only take on the values ±1 (each Ct is orthogonal).
However,COis the identity,whichhas determinant1.

Theorem 2. If A is a proper orthogonal transformation of H, then there exist unit

quaternions p and q such that A = Cp,q.

Proof Let A be a properorthogonaltransformation of H, andlet p = A(1). Clearly,

p is a unit quaternion.ObservethatCp,1maps 1 to p. Then Cp, o A fixes 1 andthus

January2005] NOTES 75
definesa properorthogonaltransformation
in E3. By Proposition1, thereexists a unit
quaternionq such that C,1 o A = Cq,q.It follows that A = Cp,1o Cq,q= Cpq,q. U

It is not the case that each A is uniquelyrepresentedas Cp,q.Our final theorem

showspreciselyto whatextentthis representation is not unique.

Theorem 3. Let Pi, P2, q1, and q2 be unit quaternions. The transformations Cpl,q1
and Cp2,q2are equal if and only if P2/Pi = q2/ql = -1.

Proof The theoremfollows fromthe observationthatCp,qis the identitytransforma-

tion if andonly if p = q = +1. One directionof this equivalenceis obvious,andthe
proofof the otherdirectionis an easy applicationof Theorem1, whichwe leaveto the
reader. 1

REFERENCES

1. G. Birkhoff and G.-C. Rota, OrdinaryDifferential Equations, 4th ed., John Wiley & Sons, New York,
1989.
2. H. S. M. Coxeter, Quaternionsand reflections, this MONTHLY 53 (1946) 136-146.
3. E. Goursat, Sur les substitutionsorthogonaleset les divisions r6gulibresde l'espace, Annales Scientifiques
de l'Ecole Normale Supdrieure6 (1889) 2-102.
4. M. Henle, Modem Geometries: Non-Euclidean, Projective, and Discrete, 2nd ed., Prentice-Hall, Upper
Saddle River, NJ, 2001.

Departmentof Mathematics, Universityof Hawaii at Manoa, 2565 McCarthyMall, Honolulu, HI 96822-2273

joel@ math.hawaii.edu

Departmentof Mathematics, Universityof Hawaii at Manoa, 2565 McCarthyMall, Honolulu, HI 96822-2273

grw @math.hawaii.edu

Another Proof of the Fundamental

Theorem of Algebra
Jose Carlos de Sousa Oliveira Santos

The goal of this note is to prove the fundamentaltheoremof algebra.To be more

precise,we show thatthe degreeof an irreduciblepolynomialin R[X] is either1 or 2.
The samemethodcan be used to provethatthe degreeof an irreduciblepolynomialin
C[X] is always 1.
Let n be an integerlargerthan 1, and let P be an irreduciblepolynomialin R[X]
of degree n. We assertthat n = 2. Denote by (P) the ideal generatedby P in the
ringIR[X].Since P is irreducible,the quotientof the ringIR[X]by (P) is a field.If we
define # : n"- R[X]/(P) by

(ao, al ..., an-_) H ao + alX + +anlXn-1 + (P),

then l is a group isomorphism from (IRn,+) onto (IR[X]/(P), +). This isomorphism
inducesin the obviousway a fieldstructurein "Rn,
the additionbeingtheusualone. The

76 © THE MATHEMATICALASSOCIATION OF AMERICA [Monthly 112