Special Orthogonal Groups and Rotations

Christopher Triola

Submitted in partial fulllment of the requirements for Honors in Mathematics at the University of Mary Washington

Fredericksburg, Virginia

April 2009

This thesis by Christopher Triola is accepted in its present form as satisfying the thesis requirement for Honors in Mathematics.

Date

Approved

Randall D. Helmstutler, Ph.D. thesis advisor

Manning G. Collier, Ph.D. committee member

J. Larry Lehman, Ph.D. committee member

Contents
1 Introduction to Rotation Groups 1.1 Linear Isometries . . . . . . . . . . 1.2 Orthogonal Groups . . . . . . . . . 1.3 Group Properties of SO(2) . . . . 1.4 Group Properties of SO(n), n 3 2 Eigenvalue Structure 2.1 First Results . . . 2.2 SO(2) . . . . . . . 2.3 SO(3) . . . . . . . 2.4 SO(4) . . . . . . . 2.5 R2 , R3 , and R4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1 2 4 5

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

6 . 6 . 7 . 9 . 12 . 15 17 17 19 22 24

3 Other Representations 3.1 C and SO(2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 H and SO(3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 H H and SO(4) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References

Abstract The rotational geometry of n-dimensional Euclidean space is characterized by the special orthogonal groups SO(n). Elements of these groups are length-preserving linear transformations whose matrix representations possess determinant +1. In this paper we investigate some of the group properties of SO(n). We also use linear algebra to study algebraic and geometric properties of SO(2), SO(3), and SO(4). Through this work we nd quantiable dierences between rotations in R2 , R3 , and R4 . Our focus then shifts to alternative representations of these groups, with emphasis on the ring H of quaternions.

Introduction to Rotation Groups

When we think of rotations many examples may come to mind, e.g. the wheels on a car or the Earth spinning as it orbits the Sun. These are all examples of rotations in three dimensions. This set contains all two-dimensional rotations, and as we will later show, every rotation in three dimensions can be reduced to something like a two-dimensional rotation in the right basis. What of higher dimensions? It turns out that the set of rotations on Rn and the operation of composition on these rotations constitute a group. It is from this perspective that we will explore rotations in two, three, and four dimensions. Notice that one of the most fundamental properties of rotations is that they do not change distances. For instance, rotate a cube about a single axis and the distances between any two points on or within the cube will remain the same as before the rotation. This property makes rotations on Rn a subset of the isometries on Rn . In fact, the set of rotations on Rn is a normal subgroup of the group of linear isometries on Rn . Naturally, we will start our journey with a discussion of isometries.

1.1

Linear Isometries

Denition 1.1. An isometry of Rn is a function f : Rn Rn such that, for any two vectors x, y Rn , we have |f (x) f (y)| = |x y|. That is, f preserves distances between points in Rn . Lemma 1.2. Suppose f : Rn Rn is a function that xes the origin. If f is an isometry then f preserves lengths of all vectors in Rn . The converse holds if f is a linear transformation. Proof. Assume f is an isometry. Then it preserves distances between vectors, that is, |f (x)f (y)| = |x y|. Consider the distance between a vector f (x) and the origin. Since f (0) = 0, we have |f (x)| = |f (x) f (0)| = |x 0| = |x|. Hence f preserves lengths. Suppose that f preserves lengths. Given vectors x, y Rn , if f is linear we have |f (x) f (y)| = |f (x y)| = |x y|. Thus f is an isometry. The origin-xing isometries on Rn are called the linear isometries on Rn . This lemma will be very important when we prove that linear isometries on Rn are in fact linear transformations on Rn . The next lemma is important for closure purposes. 1

Lemma 1.3. The composition of two isometries is an isometry. Furthermore, if they are both linear isometries, then so is the composite. Proof. Let f and g be two isometries on Rn . We want to show that f g is another isometry on Rn . For all x,y in Rn , we have |f (g (x)) f (g (y))| = |g (x) g (y)| = |x y|. Hence f g is an isometry. Furthermore, suppose f and g are both linear isometries. Then we know that f (g (0)) = f (0) = 0. So if f and g are both linear, then so is f g . Lemma 1.4. Suppose the function f : Rn Rn is an isometry that moves the origin. Then the function g : Rn Rn given by g (x) = f (x) f (0) is a linear isometry. Proof. Consider two vectors x, y in Rn . By denition of g , |g (x) g (y)| = |f (x) f (0) f (y) + f (0)| = |f (x) f (y)| = |x y|. Hence g is an isometry on Rn . Now, note that g (0) = f (0) f (0) = 0. So g xes the origin and is hence a linear isometry. This allows us to construct a linear isometry g : Rn Rn given any non-linear isometry f : Rn Rn by the formula g (x) = f (x) f (0). Now that we only need to consider linear isometries, we will show that all of these isometries are linear transformations. Before we proceed with the proof, we must rst introduce the orthogonal groups O(n).

1.2

Orthogonal Groups

Consider the following subset of n n matrices with real entries: O(n) = {A GLn | A1 = AT }. This set is known as the orthogonal group of n n matrices. Theorem 1.5. The set O(n) is a group under matrix multiplication. Proof. We know that O(n) possesses an identity element I . It is clear that since AT = A1 every element of O(n) possesses an inverse. It is also clear that matrix multiplication is by its very nature associative, hence O(n) is associative under matrix multiplication. To show that O(n) is closed, consider two arbitrary elements A, B O(n) and note the following: (AB )(AB )T = ABB T AT = ABB T AT = AAT = I. Hence (AB )T = (AB )1 , which makes AB another element of O(n). So, O(n) is closed under matrix multiplication. Thus we have that it is a group. Note. Since AT A = I , if we consider the columns of A to be vectors, then they must be orthonormal vectors. Rn . An alternative denition of O(n) is as the set of n n matrices that preserve inner products on Because we did not use this denition we must prove this using our denition of O(n). 2

Lemma 1.6. Let A be an element of O(n). The transformation associated with A preserves dot products. Proof. If we consider vectors in Rn to be column matrices then we can dene the dot product of u with v in Rn as u v = uT v. Consider the dot product of u, v Rn after the transformation A: Au Av = (Au)T (Av) = uT AT Av = uT v = u v. Hence, elements of O(n) preserve dot products. Lemma 1.7. If A O(n) then A is a linear isometry. Proof. Let A be an element of O(n). Since A preserves dot products, this means it must also preserve lengths in Rn , since the length of a vector v Rn may be dened as |v| = v v. Furthermore, it is clear that the origin is xed since A0 = 0. Thus, by Lemma 1.2, A is a linear isometry. So, we have shown that O(n) is at least a subset of the set of linear isometries. Now, we will show containment in the other direction. Proposition 1.8. Every linear isometry is a linear transformation whose matrix is in O(n). Proof. If we can show that for every origin-xing isometry f : Rn Rn there exists an n n matrix A such that f (x) = Ax for all x Rn , then f must be a linear transformation. We will now construct such a matrix. We begin by letting the ith column of the matrix be given by the vector f (ei ), where ei is the th i standard basis vector for Rn . Since f preserves dot products, the columns of A are orthonormal and thus A O(n). Now we will show that f = A by showing that g = A1 f is the identity. First, it is clear that g is an isometry and that g (0) = 0 so that g preserves length and dot products. Also, we can see that g (ei ) = ei for all i {1, 2, . . . , n}. Hence for a vector x Rn we have the following (as from [4]):
n n n

g (x) =
i=1

[g (x) ei ]ei =
i=1

[g (x) g (ei )]ei =

[x ei ]ei = x.

Thus f = A and f is a linear transformation in O(n). Recall that, in general, det(A) = det(AT ) and det(AB ) = det(A)det(B ). So, for A O(n) we nd that the square of its determinant is det(A)2 = det(A)det(AT ) = det(AAT ) = det(I ) = 1. Hence all orthogonal matrices must have a determinant of 1. 3

Note. The set of elements in O(n) with determinant +1 is the set of all proper rotations on Rn . As we will now prove, this set is a subgroup of O(n) and it is called the special orthogonal group, denoted SO(n). Theorem 1.9. The subset SO(n) = {A O(n) | det(A) = 1} is a subgroup of O(n). Proof. It is clear that the identity is in SO(n). Also, since det(A) = det(AT ), every element of SO(n) inherits its inverse from O(n). We need only check closure. Consider the two elements A, B SO(n). Since detA = detB = 1 we know that det(AB ) = 1. Hence we have shown that SO(n) is closed and thus is a subgroup of O(n). Theorem 1.10. The group SO(n) is a normal subgroup of O(n). Furthermore, O(n)/SO(n) = Z2 . Proof. Let {1} denote the group of these elements under multiplication. Dene the mapping f : O(n) {1} by f (A) = det(A). Clearly f is surjective. Also, it is obvious from the properties of determinants that this is a homomorphism. Furthermore, the kernel of this homomorphism is SO(n) since these are the only elements of O(n) with determinant +1. Therefore, by the First Isomorphism Theorem we have that O(n)/SO(n) = Z2 .

1.3

Group Properties of SO(2)

In this section we will discuss several properties of the group SO(2). We know that this is the group of rotations in the plane. We want a representation for general elements of this group to gain a better understanding of its properties. Let us start by choosing an orthonormal basis for R2 , = {(1, 0), (0, 1)}. Consider the rotation by an arbitrary angle , denoted R . We can see that the vector (1, 0) will transform in the following manner: R (1, 0) = (cos , sin ). This is basic trigonometry. Now we will rotate the other basis vector (0, 1) to nd that R (0, 1) = ( sin , cos ). Now we can construct a matrix representation of R with respect to by using the transformed vectors as columns in the matrix A= cos sin sin cos .

Note that detA = cos2 + sin2 = +1 and that AAT = = = cos sin sin cos cos sin sin cos

cos2 + sin2 cos sin cos sin sin cos cos sin sin2 + cos2 1 0 0 1 .

Hence, ordinary rotations in the plane are indeed in SO(2), as we expected. Now, for completeness, we will show the converse. The columns of any arbitrary element of SO(2) must constitute an orthonormal basis. Therefore, we can think of the columns of an element 4

of SO(2) as two orthogonal vectors on the unit circle centered at the origin in the xy -plane. It is not hard to see that the vectors u = (cos , sin ) parameterize the unit circle centered at the origin. There are only two vectors on the unit circle that are orthogonal to this vector and they are v1 = ( sin , cos ) and v2 = (sin , cos ). We construct a matrix using the vectors u and v1 as the columns: cos sin . sin cos Notice that this is the same matrix that we constructed before so we already know it lies in SO(2). Now, we will examine the matrix constructed using u and v2 : cos sin sin cos .

See that the determinant of this matrix is cos2 sin2 = 1. Hence, this does not belong to SO(2) and we have shown that all elements of SO(2) are of the form cos sin sin cos .

Is SO(2) abelian? We will explore this question using the representation we have just derived. Consider two elements of SO(2): A= cos sin sin cos , B= cos sin sin cos .

Now we check commutativity of A and B : AB = = = cos sin sin cos cos sin sin cos

cos cos sin sin cos sin sin cos sin cos + cos sin sin sin + cos cos cos ( + ) sin ( + ) sin ( + ) cos ( + ) .

If we swap these elements, we get BA = = cos sin sin cos cos sin sin cos .

cos ( + ) sin ( + ) sin ( + ) cos ( + )

We have just shown that for any A, B SO(2), AB = BA, hence SO(2) is abelian. This should not be too surprising to anyone who has experience with merry-go-rounds, wheels, or arc length. Now we will consider the question of commutativity of rotations in higher dimensions.

1.4

Group Properties of SO(n), n 3

We were able to construct simple representations for elements of SO(2). We would like simple representations of SO(3) and SO(4), but as we will discover, the determinants get to be rather complicated which makes it harder to check that a matrix is in SO(n) for larger n. However, we 5

may use our intuition about rotations and matrices to make our work easier. Since we know what elements of SO(2) look like, let us take one and try to insert it in a 3 3 matrix. Consider the matrices cos sin 0 1 0 0 A = sin cos 0 , B = 0 cos sin . 0 0 1 0 sin cos Notice that AT A = B T B = I and that detA = detB = 1, hence these matrices exist in SO(3). Furthermore, notice that they possess block elements of SO(2). This is our trick for inserting matrices from SO(2) into higher dimensions and this will be used frequently when we must construct elements of SO(n) just to give examples of properties. With these two elements of SO(3) let us consider the question: is SO(3) abelian? Consider the product AB : cos sin cos sin2 cos2 sin cos . AB = sin 0 sin cos Now consider the product BA: cos sin 0 cos2 sin . BA = sin cos 2 sin cos cos sin Note that in general AB = BA and hence SO(3) is not abelian. Furthermore, since elements of SO(3) may be embedded in SO(n) for n > 3, we know that SO(2) is the only special orthogonal group that is abelian. However, it is worth noting that there do exist commutative subgroups of SO(n) for all n. An example of a commutative subgroup in SO(3) is the set of matrices of the form cos sin 0 sin cos 0 . 0 0 1 Matrices of this form are commutative since these are rotations in the xy -plane and we already showed that SO(2) is commutative. We have established some basic properties of the special orthogonal groups and can use a parameterization of SO(2) to manufacture some elements of SO(n). We will now proceed to delve into the properties of SO(n) as operations on Rn by investigating their eigenvalue structures.

2
2.1

Eigenvalue Structure
First Results

The objective of this section is to understand the action of elements of SO(n) on Rn . This will give us a more mathematical intuition for what a rotation actually is. But before we can investigate the eigenvalues of elements of SO(n) we must rst understand the tools we will use in our investigation. In this section, we will show that the following properties hold for each A SO(n) and the roots of its characteristic polynomial: detA = 1 = 1 2 3 n , where each i is a root, counted with multiplicity if is a root, so is 6

 if i is real then |i | = 1. Recall that we dene eigenvalues for an n n matrix A to be all R for which there exists a non-zero vector r such that Ar = r. The vectors r for which this is true are called the eigenvectors of A. We call the set of all such eigenvectors for a given eigenvalue , along with the zero vector, an eigenspace of A. Before we can begin our investigation of how eigenvalues relate to our discussion of rotations we must establish the properties listed above. Theorem 2.1. If A is an n n matrix, the determinant of A is equal to the product of the n roots (counted with multiplicity) of the characteristic polynomial of A. Proof. Let A be an n n matrix with characteristic polynomial p() = det[I A] = n + an1 n1 + + a1 + a0 . By the fundamental theorem of algebra, we have p() = ( 1 )( 2 ) ( n ) where i is the ith root of the polynomial. Since p() = det[I A], p(0) = det[A] = (1)n detA = a0 . Hence a0 = detA for even n and a0 = detA for odd n. Since p(0) = a0 we have that a0 = (0 1 )(0 2 ) (0 n ) = (1 )(2 ) (n ) = (1)n 1 2 n . Hence, regardless of the value of n, detA = 1 2 n . We now turn our attention to the second property, that if is a root of the characteristic polynomial of A, so is its conjugate. Note that the characteristic polynomial for a given element A SO(n) is a polynomial with real coecients and hence complex roots of the characteristic polynomial of A come in conjugate pairs. Thus, if is a root of the characteristic polynomial of A then is also a root, though not distinct in the case of real roots. We now move on to the third property. Recall that any transformation A SO(n) is a lengthpreserving isometry, so for all r Rn we have |Ar| = |r|. It follows that if is a real eigenvalue with eigenvector r then |Ar| = |r| = |r| which implies that || = 1. Since all real roots of the characteristic polynomial must have absolute value of 1 and all roots must multiply to 1, the complex roots must all multiply to 1.

2.2

SO(2)

In this section we will nd and discuss all the possible roots of the characteristic polynomials of elements of SO(2). Because we are in SO(2) the characteristic polynomial of each element will have two roots. Consider an arbitrary element A SO(2). We know that detA = 1 = 1 2 . From the previous section we know that |i | = 1 when i is real. It should be clear that the only possible roots are those appearing in the following table.

Roots 1, 1 1, 1 ,

Rotation Angle 0 = 0,

We are fortunate enough to have a simple representation for this group so we will utilize it here. Let us nd the roots of the characteristic polynomial of A= cos sin sin cos SO(2).

Setting the characteristic polynomial equal to 0 gives 2 2 cos + 1 = 0. By the quadratic formula, we have 2 cos 4 cos2 4 2 cos 2 cos2 1 = = = cos sin2 . 2 2 Hence the roots of the characteristic polynomial for an element of SO(2) all come in the form = cos i sin . We can see that the only real value for occurs when sin = 0, which will happen only when = 0, . Thus, we have shown that all of the aforementioned cases do in fact arise from the following scenarios: if = 0 then the roots are 1, 1 if = then the roots are 1, 1 if = 0, then the roots are complex. Note that the formula cos i sin gives the roots for the characteristic polynomial associated with a rotation by an angle . We could just as easily use this process in reverse and calculate the angle of rotation of a transformation based on the roots of its characteristic polynomial. Now we will calculate the eigenspaces for the real eigenvalues that we found and explain them geometrically. We will accomplish this by noting that when = 0 the rotation matrix is I= 1 0 0 1 .

Hence the eigenvalue is 1 with multiplicity 2. Since all vectors r R2 will be solutions to I r = r, all of R2 comprises the eigenspace of this matrix. The dimension of this eigenspace is thus 2. For = , we note that the matrix representation would be I = 1 0 0 1 .

Note that the eigenvalue is 1 with multiplicity 2, and that this is a rotation by . Therefore, once again, all vectors r R2 are solutions to I r = r. Hence the eigenspace for this case is again all of R2 . For complex roots we do not consider the eigenspaces since this would take us outside of the real numbers. Hence, we have now described the eigenvalue structure for all elements of SO(2). It was what our intuition would tell us: only the identity preserves the whole plane, the negative of the identity swaps all vectors about the origin, and except for these two cases, no rotation of the plane maps a vector to a scalar multiple of itself. But we noticed the interesting fact that the roots of the characteristic polynomial come in the form cos i sin where is the angle of rotation, giving a geometric link even to complex roots. Now let us explore higher dimensions. 8

2.3

SO(3)

Consider a matrix A SO(3). By our previously established facts and theorems, the following table displays all possible roots for the characteristic polynomial of A. Roots 1, 1, 1 1, 1, 1 1, , Rotation Angle 0 = 0,

Note. When discussing rotations in R2 an angle was enough to describe a rotation. Now that we are dealing with R3 , we must specify an angle and a plane. Because we are restricted to R3 , this plane may be specied by a vector orthogonal to the plane. This vector may be thought of as what we traditionally call the axis of rotation. Case 1. The only example of the rst case is the identity. This example clearly preserves all of 3-space and hence its eigenspace has dimension 3. Case 2. For this case we shall consider the matrix 1 0 0 A = 0 1 0 . 0 0 1 We can see that once again its transpose is itself and that AAT = I . We can also see that the determinant is +1, hence A SO(3). Because this is diagonal, we can see that the eigenvalues are 1, 1, 1. Now we will compute the eigenspaces for A using the following augmented matrix for =1 11 0 0 0 0 0 0 0 0 1+1 0 0 = 0 2 0 0 0 0 1+1 0 0 0 2 0 and the following for = 1: 2 0 0 0 1 1 0 0 0 0 1 + 1 0 0 = 0 0 0 0 . 0 0 0 0 0 0 1 + 1 0 From this we can tell that for = 1 the eigenspace is generated by 1 0 0 and for = 1 the generators are 0 0 1 , 0 . 0 1 Hence for = 1 we have a one-dimensional eigenspace. That is, all eigenvectors lie along an axis. For = 1 we see that the eigenvectors make up a plane. Notice however that for = 1 the vectors are invariant while for = 1 the vectors are ipped about the origin. This tells us that for this rotation, an axis stays the same and the yz -plane is rotated by . 9

Case 3. We will now deal with the most general case of = 1, , . Consider the matrix 0 1 0 A = 1 0 0 . 0 0 1 Note that AAT = I so AT = A1 and that det(A) = 1, so A SO(3). Now we will solve for its eigenvalues. Setting det(I A) = 0 gives 3 2 + 1 = 0. Since we know from before that 1 must be an eigenvalue, we divide this polynomial by 1 to obtain 2 + 1 = 0 which has solutions = i. Hence we have a rotation whose characteristic polynomial has roots = 1, i, i. Now we will nd the eigenspace for = 1. We start with the following augmented matrix 1 1 0 0 1 1 0 0 1 1 0 0 = 1 1 0 0 0 0 11 0 0 0 0 0 which row reduces to 1 0 0 0 0 1 0 0 . 0 0 0 0 0 0 . 1 So the entire z -axis is preserved but the rest of 3-space is not. If we inspect what this does to a vector in the xy -plane, e.g. (1, 1, 0), we nd the transformed vector is (1, 1, 0). Using the dot product we nd that the cosine of the angle between this vector and its transformed counterpart is equal to 0. This tells us that this is a rotation by 2 . This corresponds to sin = 1. Recall that in SO(2) we could nd all roots using the formula cos i sin , where is the angle of rotation. Recall also that the complex roots for this rotation were i. These complex roots are exactly what we would expect if the same formula applied to roots in SO(3). We will need more evidence to support this claim before we attempt a proof. Though we have given examples of each possible combination of roots for SO(3), we will now demonstrate the eigenvalue properties of SO(3) using a more complex example. Example. Consider the matrix A=
5 8

Thus, the eigenspace is generated by

3 3 8 43

383 1 8
3 4

3 4 3 4 1 2

10

Note that AAT = I and that detA = +1 so that A is in SO(3). Now let us compute its eigenvalues. We will solve the usual equation for determining eigenvalues: 3 3 3 5 + 8 4 8 1 1 3 = 3 + 2 1 = 0. det 3 8 3 + 1 8 4 4 4 3 3 1 4 4 2 As we have shown, +1 must be an eigenvalue so we will divide this polynomial by 1 to obtain 5 2 + + 1 = 0 . 4 We will now use the quadratic formula to solve for the remaining roots. We nd that 5 39 5 i 39 = = . 8 8 Hence the characteristic polynomial of A has two complex roots that are conjugates of each other. Now we will compute the eigenspace for = 1 using the augmented matrix 3 3 3 3 3 3 5 13 1+ 0 0 8 4 8 8 4 38 3 3 1 9 3 = 8 3 1 + 8 3 0 0 4 8 4 8 3 3 3 1 3 1 4 1 2 0 4 0 4 4 2 which row reduces to 1 0 0 0 0 3 2 Hence, the eigenspace is generated by 0 1 .
3 2

0 0 0 0 . 1 0

Note that this transformation still leaves a line invariant. Leonhard Euler also noticed this fact and proved his famous theorem about three-dimensional rotations that, put in our language, is: Eulers Rotation Theorem. If A is an element of SO(3) where A = I then A has a onedimensional eigenspace. This is not to say that A does not also have a two-dimensional eigenspace but that A must have a one-dimensional eigenspace. It is this eigenspace that is known as the axis of rotation. The work we have done in this section should be enough to convince us that this is true. However, as we will see in SO(4), this way of viewing rotations does not generalize at all. In fact, having an axis of rotation is a special case of how we may dene planes in 3-space that only applies to R3 . It is for this reason that we have used the term axis sparingly. Before we proceed with SO(4), we noted earlier that there might be some geometric signicance to the complex roots but needed another example to motivate a proof. We would like to see what angle of rotation corresponds to the matrix example above. The generator of the eigenspace, 3 ), can be seen as the axis of rotation, so we will see what this matrix does to an element of (0, 1, 2 the orthogonal complement of this eigenspace. We can see from the dot product that v = (0, 3 2 , 1) 11

is orthogonal to the axis of rotation. Also note that after being transformed this vector becomes 13 3 15 v = ( 16 , 16 , 5 8 ). Now we will determine the cosine of the angle between these two vectors: vv = 13 65 cos = . 4 32

Solving for the cosine term we see that cos = 5 8 , which is the real part of the complex roots. This tells us that even a rotation about a more general axis still follows this tendency. In a later section we will prove that this same geometric signicance exists for all roots of characteristic polynomials of matrices in SO(3).

2.4

SO(4)

To begin our discussion of SO(4) we will mimic the procedure we used to discuss SO(3). We begin by noting the possible eigenvalues and other complex roots of the characteristic polynomials. We proceed by giving examples of each case and discussing their geometric signicance. Note. In R2 there is only one plane, so we only needed to specify an angle. In R3 we needed to specify a plane and an angle. In R4 we must specify two planes and an angle for each plane. The reasoning behind this will become apparent in this section. Roots 1, 1; 1, 1 1, 1; 1, 1 1, 1; 1, 1 1, 1; , 1, 1; , , ; z, z Rotation Angles 0; 0 ; 0; 0; ; ;

Notice that there are twice the number of cases for SO(4) roots than for SO(3) or SO(2). This comes from the fact that we may rotate in, say, the x1 x2 -plane and then rotate in the x3 x4 -plane without disturbing our previous rotation. As we will see, R4 is an interesting place. Case 1. As usual, the identity handles the rst case and it is clear that the eigenspace for I SO(4) is all of R4 . Case 2. It is simple to construct an example by considering I . Since I x = x for all x R4 , we know that the eigenspace is once again all of R4 . In this case, each vector in R4 is swapped about the origin. Does this rotation take place about an axis? In three dimensions we had a similar case in which a plane of vectors was rotated by . However that had an axis because only a plane was involved and there was an extra degree of freedom which we called the axis. In this case, however, all vectors are included in the eigenspace so there is no favored axis of rotation here: the entire space is ipped. In three-space we think of rotations as occurring about one-dimensional axes, but what is really going on is that a plane is being rotated and any vector with a non-trivial projection in that plane will also be rotated. However, only the component in the direction of the projection will be rotated: the component orthogonal to the plane of rotation will remain unchanged. If we remember, this is what happened in two-space as well. In fact, it is by embedding the two-dimensional rotations into the higher dimensions that we have obtained most of the examples in this section. So, looking at it from this perspective, the rst case would be rotation by 0 radians. The second would be a 12

rotation by of, say, the x1 x2 -plane followed by that this acts exactly as we claim it does: cos sin 0 0 1 sin cos 0 0 0 0 0 1 0 0 0 0 0 1 0 1 0 0 1 = 0 0 0 0 1 0 0 1 = 0 0 0 0 0 0 1 0 0 0 1 0

a rotation of the x3 x4 -plane by . We can see 0 0 0 x1 x2 1 0 0 0 cos sin x3 0 sin cos x4 0 0 x1 x2 0 0 1 0 x3 0 1 x4 x1 x2 = x3 . x4

0 1 0 0 1 0 0 0 0 1 0 0 x1 0 0 x2 0 x3 x4 1

Thinking about rotations in this way may help to alleviate the confusion that is common when trying to visualize higher dimensions. Case 3. For this case, we can use the matrix 1 0 A= 0 0

0 0 0 1 0 0 . 0 1 0 0 0 1

It should be clear that the eigenvalues for this matrix are 1, 1, 1, 1. Solving for the eigenspaces, we will need to use two augmented matrices: (I A|0) and (I A|0). In the rst case we nd an eigenspace generated by 1 0 0 1 , . 0 0 0 0 This eigenspace is simply a copy of the x1 x2 -plane. For the second augmented matrix we obtain an eigenspace generated by 0 0 0 0 , . 1 0 0 1 This eigenspace is a copy of the x3 x4 -plane. However, if we once again use this matrix to see where an arbitrary vector is sent we nd 1 0 0 0 x1 x1 0 1 0 0 x2 = x2 . Ax = 0 0 1 0 x3 x3 0 0 0 1 x4 x4 Hence we have a rotation by of the x3 x4 -plane and a rotation by 0 of the x1 x2 -plane. 13

Case 4. We will show that the following matrix satises this case and then examine the implications of such a rotation: 1 0 0 0 0 1 0 0 B= 0 0 0 1 . 0 0 1 0 It is not hard to see that BB T = I and det(B ) = 1 and hence B SO(4). Now we can move on to its eigenvalues which we will nd in the usual way: 1 0 0 0 0 1 0 0 = ( 1)( 1)(2 + 1) = 0. det(I B ) = det 0 0 1 0 0 1 We see that = 1 appears as a solution with multiplicity 2 and we have = i as the other two solutions. We have thus shown that this is an example of Case 4. Now we must nd the eigenspace for = 1 using the augmented matrix 11 0 0 0 0 0 0 0 0 0 0 11 0 0 0 = 0 0 0 0 0 0 0 1 1 0 0 0 1 1 0 0 0 1 1 0 0 0 1 1 0 which row reduces to 0 0 0 0 Hence, the eigenspace is generated by 1 0 0 1 , 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 1 0 0 . 0 0

which generates a copy of the x1 x2 -plane. To avoid confusion and to see what happens to the other two parameters, let us see where B maps an arbitrary vector: x1 x1 1 0 0 0 0 1 0 0 x2 x2 Bx = 0 0 0 1 x3 = x4 . 0 0 1 0 x4 x3 This rotation leaves the x1 x2 -plane invariant (as we said before) and it rotates the x3 x4 -plane by 2. Case 5. For the case where = 1, 1, , , we will use our previous example as inspiration to construct the following matrix: 1 0 0 0 0 1 0 0 . B = 0 0 0 1 0 0 1 0 14

Once again, it is easy to show that this is an element of SO(4). When we compute its eigenvalues we nd that they are = 1, 1, i, i, so this is in fact an example of Case 5. Watching what this matrix does to an arbitrary vector, however, we see that: 1 0 0 0 x1 x1 0 1 0 0 x2 x2 . = Bx= 0 0 0 1 x3 x4 0 0 1 0 x3 x4 This still rotates the x3 x4 -plane by
2

but it also rotates the x1 x2 -plane by .

Case 6. This is the case where there are no real eigenvalues. We might think that such an example would have to be a very exotic matrix. However, using what weve learned about rotations, all we need is a matrix that doesnt map any vector in R4 to a scalar multiple of itself. So, if we rotate the x1 x2 -plane by 2 and do the same to the x3 x4 -plane, that action should satisfy this case. Lets do that using the matrix 0 1 0 0 1 0 0 0 C= 0 0 0 1 . 0 0 1 0 We now compute the eigenvalues of this matrix: 1 1 det(I C ) = det 0 0 0 0 This factors into (2 + 1)2 = 0. Hence, the eigenvalues are = i, each with multiplicity two, which gives us an example of Case 6. For completeness, we will now verify that it transforms vectors the way we constructed it to: x1 0 1 0 0 x2 1 0 0 0 x2 x1 Cx = 0 0 0 1 x3 = x4 . 0 0 1 0 x4 x3 As we can see this is a rotation of the x1 x2 -plane by just as we claimed during its construction.
2

0 0 1

0 0 = 4 + 2 2 + 1 = 0 . 1

and then a rotation of the x3 x4 -plane by 2,

2.5

R2 , R3 , and R4

The following are important properties to note about the relationships between rotations in two dimensions and those in three and four dimensions. With these properties we will be able to simplify our work in the nal section while remaining mathematically rigorous in the theorems to come. Theorem 2.2. Any arbitrary element of SO(3) may be written as the composition of rotations in the planes generated by the three standard orthogonal basis vectors of R3 .

15

Proof. It is easy to show that the following matrices represent rotations by 1 , 2 , and 3 in the xy -, yz - and xz -planes respectively: cos 3 0 sin 3 cos 1 sin 1 0 1 0 0 . 1 0 Rz = sin 1 cos 1 0 , Rx = 0 cos 2 sin 2 , Ry = 0 sin 3 0 cos 3 0 0 1 0 sin 2 cos 2 As we know, all rotations in R3 aside from the identityhave a one-dimensional eigenspace about which we rotate by an angle . A unit vector n in this eigenspace may be parameterized as n = (cos sin , sin sin , cos ). This is because the one-dimensional eigenspace, or axis of rotation, must intersect the unit sphere centered at the origin and the above can be shown to parameterize the unit sphere. We may align the axis of rotation with the z -axis as follows: cos 0 sin cos sin 0 cos sin sin cos 0 sin sin 1 0 Ry ( )Rz ()T n = 0 sin 0 cos 0 0 1 cos 0 = 0 . 1 We can thus perform the rotation by in the xy -plane and then transform back. All we need to know is the angle of rotation and the angles , which specify the rotations one-dimensional eigenspace to write any element of SO(3) as Rz ()Ry ( )T Rz ()Ry ( )Rz ()T . This is not unique since we could just as easily transform the eigenspace to align with either of the other two axes. Hence, as we claimed, we may write any rotation in SO(3) as the product of rotations in the orthogonal planes of R3 . Recall earlier we noticed a pattern indicative of a possible geometric interpretation of complex eigenvalues. Now we will state and prove this fact in the following theorem. Theorem 2.3. For any element A SO(3) with rotation angle , the roots of the characteristic polynomial of A are 1 and cos i sin . Proof. Let A be an element of SO(3) with axis generated by n = (cos sin , sin sin , cos ). By Theorem 2.2, we may represent A as the product A = Rz ()Ry ( )T Rz ()Ry ( )Rz ()T = (Rz ()Ry ( )T )Rz ()(Rz ()Ry ( )T )T = P Rz ()P T . This means that the matrix A is similar to the matrix cos sin 0 Rz () = sin cos 0 . 0 0 1 16

The characteristic polynomial of this matrix is ( 1)(2 2 cos + 1). Dividing by 1 we nd that the roots that are not equal to 1 are cos i sin . Since A and Rz () are similar they share the same characteristic polynomial. Hence for any rotation by in R3 the roots of the characteristic polynomial are 1, cos + i sin , and cos i sin . Proposition 2.4. Any arbitrary element of SO(4) may be written as the composition of rotations in the planes generated by the four orthogonal basis vectors of R4 . The proof involves changing the basis of an arbitrary element of SO(4) such that the two planes of rotation are made to coincide with the x1 x2 -plane and the x3 x4 -plane. We would change the basis using products of the matrices cos 1 sin 1 0 0 cos 2 0 sin 2 0 cos 3 0 0 sin 3 sin 1 cos 1 0 0 0 1 0 0 1 0 0 , , 0 0 0 1 0 sin 2 0 cos 2 0 0 0 1 0 0 0 0 1 0 0 0 1 sin 3 0 0 cos 3 1 0 0 0 cos 4 sin 4 0 sin 4 cos 4 0 0 0 0 1 0 0 0 cos 5 , 0 0 0 1 0 sin 5 0 0 1 0 sin 5 0 , 0 1 0 0 cos 5 0 0 0 0 1 0 0 . 0 cos 6 sin 6 0 sin 6 cos 6

We would then perform our rotation using cos 1 sin 1 0 sin 1 cos 1 0 0 0 1 0 0 0

the matrices 0 1 0 0 0 0 1 0 0 0 , 0 0 0 cos 2 sin 2 1 0 0 sin 2 cos 2

and then change back using the appropriate transposes of the matrices above. Without an axis it is more dicult to see how to change the basis appropriately, so an actual proof will be omitted. However, this is a known fact that will be assumed later.

Other Representations

Recall that the reason we have been using matrices was that rotations turned out to be linear transformations and matrices are natural representations of linear transformations. However, we must not believe that these are the only possible ways of expressing rotations. After all, we can easily represent a rotation in R3 by using just an axis and an angle. In this section we will turn our attention to alternative representations of the special orthogonal groups. Specically, we will consider those representations given by complex numbers and quaternions.

3.1

C and SO(2)

A complex number is a number with a real and an imaginary component, written z = a + bi. For all elements z1 = a + bi, z2 = c + di C we dene addition as z1 + z2 = (a + c) + (b + d)i. It is simple to check that with this addition operation C is a two-dimensional vector space over the real numbers which makes it isomorphic to R2 as a vector space. If z = x + yi then we dene 17

the conjugate to be z = x yi. Also, we have a multiplication operation dened for every pair of elements as above by z1 z2 = (ac bd) + (ad + bc)i. With this knowledge we may dene the norm of z C as |z | = zz. Recall that the norm is the length of the vector. Another useful property to note is that the norm is multiplicative, that is, |z1 z2 | = |z1 ||z2 |. With this knowledge we are able to express the unit circle, or one-sphere S 1 , as S 1 = {z C : |z | = 1}. Lemma 3.1. The one-sphere is an abelian group under complex multiplication. Proof. Recall that the set of non-zero complex numbers is a group under multiplication. This means we only need to check that S 1 is a subgroup of C {0}. We start by checking closure. Let z1 , z2 be elements of S 1 . Then we have that |z1 z2 | = |z1 ||z2 | = 1 1 = 1. Hence S 1 is closed under the multiplication operation of C. It is clear that |1| = 1 and it should also be clear that this is the identity under multiplication. Hence S 1 possesses an identity. Also, every element z S 1 has an inverse in C such that zz 1 = 1. Then |zz 1 | = |1| = 1 and |z | = |z 1 | = 1, hence if z is in S 1 so is its inverse. Associativity is inherited from the complex numbers and we have shown that S 1 is a group under complex multiplication. Complex multiplication is commutative, so S 1 is abelian. Recall that the set of unit complex numbers may be written in the form ei = cos + i sin . This new form will aid us in the proof of the main theorem for this section. Theorem 3.2. The group SO(2) is isomorphic to S 1 . Proof. Recall that any element in SO(2) may be represented as a matrix A() = cos sin sin cos

where is the angle of rotation, chosen to be in [0, 2 ). Consider the mapping f : S 1 SO(2), where f (ei ) = A(). See that f (ei )f (ei ) = = = Now see that f (ei ei ) = f (ei(+) ) = cos ( + ) sin ( + ) sin ( + ) cos ( + ) cos sin sin cos cos sin sin cos

cos cos sin sin (cos sin + sin cos ) sin cos + cos sin cos cos sin sin cos ( + ) sin ( + ) sin ( + ) cos ( + ) .

= f (ei )f (ei ). 18

Thus f is a homomorphism. By our representation, every rotation in SO(2) is of the form A() for some angle , and in this case f (ei ) = A(). Hence f is surjective. Since is chosen to be in [0, 2 ) it is uniquely determined, and hence f is injective. Therefore S 1 is isomorphic to SO(2).

3.2

H and SO(3)

We have shown that S 1 is a group under complex multiplication and that it is isomorphic to SO(2). We would now like to nd representations for higher-dimensional rotations. Since the elements of SO(3) can be seen as invariant points on the unit sphere S 2 in R3 , we turn our attention to S 2 . Note that an element of SO(3) is not just a point on a sphere: it contains another piece of information, the angle of rotation. Therefore, we look to the next dimension up, the three-sphere S 3 . But before we can hope to nd a representation using elements of the three-sphere, we must establish its multiplication rules. To begin we would like to nd a set that is a four-dimensional vector space and whose non-zero elements form a group under multiplication. Thus we must introduce Hamiltons quaternions. The set of quaternions H is the set of generalized complex numbers q = a0 + a1 i + a2 j + a3 k , where i, j, k are imaginary numbers satisfying the properties: i2 = j 2 = k 2 = ijk = 1 ij = k jk = i ki = j ji = k kj = i ik = j. We dene the addition of quaternions similarly to the way we dened complex addition. Explicitly, for every pair of elements q1 = a0 + a1 i + a2 j + a3 k , q2 = b0 + b1 i + b2 j + b3 k H we dene their addition component-wise: q1 + q2 = (a0 + b0 ) + (a1 + b1 )i + (a2 + b2 )j + (a3 + b3 )k where all coecients are real numbers. Thus we can see that addition is commutative in H. In fact, it is not hard to see that H makes up a four-dimensional vector space and hence is isomorphic to R4 as a vector space. Just like in the complex numbers we have a conjugate and it is dened in a similar fashion. If q = a0 + a1 i + a2 j + a3 k is a quaternion we say that the conjugate of q , denoted q , is given by q = a0 a1 i a2 j a3 k . Another similarity between the quaternions and the complex numbers is that we may dene a multiplication operation on H. This quaternionic multiplication works using the regular distribution laws combined with the identities above, so that for every pair of elements q1 = a0 + a1 i + a2 j + a3 k , q2 = b0 + b1 i + b2 j + b3 k H we dene q1 q2 by q1 q2 = (a0 b0 a1 b1 a2 b2 a3 b3 ) + (a0 b1 + a1 b0 + a2 b3 a3 b2 )i + (a0 b2 + a2 b0 + a3 b1 a1 b3 )j + (a0 b3 + a3 b0 + a1 b2 a2 b1 )k.

19

Notice that this does not always commute. Now let us see what happens when q2 = q1 :
2 2 2 2 2 2 2 q1 q1 = (a2 0 + a1 + a2 + a3 ) + (0)i + (0)j + (0)k = a0 + a1 + a2 + a3 .

This expression happens to be the square of the norm for a vector in R4 , hence we dene the norm of a quaternion q to be |q | = qq. It is not hard to show that the following property holds for all quaternions q1 , q2 : |q1 q2 | = |q1 ||q2 |. Now we are prepared to dene the three-sphere using quaternions. Since the three-sphere S 3 is the set of all unit length vectors in R4 (or H) we may dene S 3 as follows: S 3 = {q H : |q | = 1}. We can now prove the following result. Lemma 3.3. The three-sphere is a non-abelian group under quaternionic multiplication. Proof. Consider the norm of the multiplication of two elements q1 , q2 S 3 : |q1 q2 | = |q1 ||q2 | = 1. Hence S 3 is closed under quaternionic multiplication. Next, we can see that the set contains an identity, namely the number 1. We know this acts as the identity because of how real numbers distribute over the quaternions. Also note that for all q S 3 we have |q |2 = qq = 1, hence q is the inverse of q . It is clear that q 1 S 3 . Finally, S 3 inherits associativity from H. Thus S 3 is a group under quaternionic multiplication. Furthermore, this group is not abelian, since quaternions rarely commute. Now that we know that this is a group we are in a position to prove the following theorem. Theorem 3.4. There exists a two-to-one homomorphism from S 3 onto SO(3). Proof. To show this we must demonstrate that every quaternion of unit length may be mapped to a rotation in R3 and that every rotation in R3 has two quaternion representations. Let n = a1 i + a2 j + a3 k be such that |n| = 1. Consider a unit quaternion written in the form q = cos + n sin . It is not hard to show that by varying both the direction of n and the angle we may represent any element of S 3 in this form. We will show that we can use quaternionic multiplication to construct rotations in R3 . Let r = (x, y, z ) be a vector in R3 . We could just as well write this vector as r = (0, x, y, z ) in R4 . Since R4 is isomorphic to H as a vector space we could also write this vector as r = xi + yj + zk . If we multiply by a unit quaternion q then we see that |q r| = |r|, so that multiplication by q is length-preserving. However, this multiplication will in general yield a non-zero real component in q r, which will prevent us from mapping the result back to R3 . So we need to nd a multiplication whose result does not contain a real part. We will try the following: q rq = (cos + n sin )r(cos n sin ). 20

When we expand this we see that it does not contain a real part:
q rq = [cos2 x + 2 cos sin (za2 ya3 ) sin2 (a1 (xa1 ya2 za3 ) + a2 (xa2 ya1 ) a3 (za1 xa3 ))]i + [cos2 y + 2 cos sin (xa3 za1 ) sin2 (a2 (xa1 ya2 za3 ) + a3 (ya3 za2 ) a1 (xa2 ya1 ))]j + [cos2 z + 2 cos sin (ya1 xa2 ) sin2 (a3 (xa1 ya2 za3 ) + a1 (za1 xa3 ) a2 (ya3 za2 ))]k.

This is a step in the right direction because this vector has only three components. We will now construct the matrix representation of this linear transformation on R3 . By evaluating q rq as r ranges over the standard basis vectors, we see that the matrix representation of this operation is
2 2 cos2 + sin2 (a2 1 a2 a3 ) 2 A= 2 cos sin a3 + 2 sin a1 a2 2 cos sin a2 + 2 sin2 a1 a3

2 cos sin a3 + 2 sin2 a1 a2 2 2 cos2 + sin2 (a2 2 a1 a3 ) 2 2 cos sin a1 + 2 sin a2 a3

2 cos sin a2 + 2 sin2 a1 a3 2 cos sin a1 + 2 sin2 a2 a3 . 2 2 cos2 + sin2 (a2 3 a2 a1 )

Furthermore, it can be shown that detA = 1 and that AT A = I . Given a unit length quaternion q , we dene a transformation fq on R3 by fq (r) = q rq . The argument above proves that fq SO(3). Dene a map f : S 3 SO(3) by f (q ) = fq . We claim that f is a two-to-one surjective homomorphism. Consider two elements q1 , q2 S 3 . Note the following: f (q1 q2 )(r) = q1 q2 rq1 q2 = q1 q2 r(q1 q2 )1
1 1 q1 = q1 q2 rq2

= (f (q1 ) f (q2 ))(r). Thus, f is a homomorphism. Now we will show that this map is onto. Consider the matrix associated with the quaternion q = cos + n sin when n = (1, 0, 0). We nd this by setting a1 = 1 and a2 = a3 = 0 in the matrix A above. We nd that this matrix is 1 0 0 1 0 0 0 cos2 sin2 2 cos sin = 0 cos 2 sin 2 . 0 2 cos sin cos2 sin2 0 sin 2 cos 2 Note that it is not a rotation by but instead a rotation by 2 in the yz -plane. Notice that if n had coincided with the y - or z -axis we would have obtained a similar matrix, except it would represent a rotation of 2 about the chosen axis. Since we can generate rotations about the standard orthogonal basis vectors using quaternions, we may invoke Theorem 2.2 to conclude that this homomorphism is onto. To see that f is two-to-one, we will show that the kernel of f contains only two elements. It is clear that f1 and f1 are both the identity rotation. We claim that 1 are the only elements in the kernel. Suppose that q ker(f ), so that fq is the identity rotation. Then q rq = r for all r = xi + yj + zk . This would imply that q r = rq for all such r. The only quaternions that commute with all pure quaternions are the reals, so q must be real. Since q must be unit length, we conclude that q = 1. Hence f is two-to-one. While this is not an isomorphism because it is two-to-one, by the First Isomorphism Theorem we see that S 3 /Z2 = SO(3). This result is important because it allows us to prove a famous result from topology using only linear and abstract algebra. First, note that in S 3 every point (x1 , x2 , x3 , x4 ) has an antipode (x1 , x2 , x3 , x4 ). This relation leads to a normal subgroup of order 2 which is isomorphic to Z2 . If we take S 3 modulo this subgroup we obtain real projective space RP 3 . There is a famous result from topology that states RP 3 = SO(3). However, we have arrived at this result using mostly linear algebra. 21

3.3

H H and SO(4)

We showed that S 1 is isomorphic to SO(2). We then showed that there is a two-to-one homomorphism from S 3 onto SO(3). We will now attempt to nd a similar representation of SO(4). Intuition might suggest we attempt to nd another sphere to represent SO(4). However, from our previous work with SO(4), we know there are fundamental dierences in the behavior of rotations in R4 . We understand from earlier that SO(3) is like SO(2) except we may specify the plane in which to rotate by giving the orthogonal complement of that plane, namely the axis. This had the eect of taking away commutativity for SO(3). However, elements of SO(4) are not just single elements of SO(2) oriented in odd directions, they are more like pairs of noncommutative elements of SO(2). Elements of SO(3) can be seen as noncommutative elements of SO(2) so perhaps we should investigate S 3 S 3 . As we will soon see, this choice is, in fact, correct. Theorem 3.5. There exists a two-to-one, surjective homomorphism f : S 3 S 3 SO(4). Proof. Let x = (x1 , x2 , x3 , x4 ) be a vector in R4 . Since H is a vector space isomorphic to R4 we may represent this vector as the quaternion x = x1 + x2 i + x3 j + x4 k . Given two unit length quaternions q1 , q2 , we dene a transformation fq1 ,q2 on R4 by fq1 ,q2 (x) = q1 xq2 . We claim that fq1 ,q2 SO(4). We know that |q1 xq2 | = |x| so the transformation fq1 ,q2 preserves length. Also note that if we let x = 0 then q1 xq2 = 0. Thus such a transformation is also originpreserving and hence linear. Since fq1 ,q2 is a length-preserving linear transformation on R4 , we know that it is an element of O(4). We will now construct the matrix representation of a general transformation fq1 ,q2 . Let q1 = a0 + a1 i + a2 j + a3 k and q2 = b0 b1 i b2 j b3 k be two unit quaternions. Then fq1 ,q2 (x) = q1 xq2 where x is an arbitrary vector in R4 . Notice that |xq2 | = |x| and that 0q2 = 0, so that f1,q2 alone is a length-preserving linear transformation on R4 . Hence it is in O(4) and may be represented with a matrix, specically b0 b1 b2 b3 b1 b0 b3 b2 . b2 b3 b0 b1 b3 b2 b1 b0 Solving for the determinant we nd that it is 1, so the linear transformation f1,q2 is an element of SO(4). Along a similar argument we can see that fq1 ,1 is a linear transformation that preserves length. Hence we may nd that its matrix representation is a0 a1 a2 a3 a1 a0 a3 a2 . a2 a3 a0 a1 a3 a2 a1 a0 The determinant of this matrix is also 1, so fq1 ,1 is in SO(4). Now a general transformation fq1 ,q2 is simply the composition of these two transformations. Hence its matrix representation is the product of the matrix representations of f1,q2 and fq1 ,1 . When we multiply these two matrices together we nd that the matrix representation for a general transformation fq1 ,q2 is

0 B B @

a0 b0 a1 b1 a2 b2 a3 b3 a1 b0 + a0 b1 a3 b2 + a2 b3 a2 b0 + a3 b1 + a0 b2 a1 b3 a3 b0 a2 b1 + a1 b2 + a0 b3

a0 b1 a1 b0 + a2 b3 a3 b2 a1 b1 + a0 b0 + a3 b3 + a2 b2 a2 b1 + a3 b0 a0 b3 a1 b2 a3 b1 a2 b0 a1 b3 + a0 b2

a0 b2 a1 b3 a2 b0 + a3 b1 a1 b2 + a0 b3 a3 b0 a2 b1 a2 b2 + a3 b3 + a0 b0 + a1 b1 a3 b2 a2 b3 + a1 b0 a0 b1

a0 b3 + a1 b2 a2 b1 a3 b0 a1 b3 a0 b2 a3 b1 + a2 b0 a2 b3 a3 b2 + a0 b1 a1 b0 a3 b3 + a2 b2 + a1 b1 + a0 b0

1 C C A.

22

Since this matrix is the product of two elements of SO(4) it must also be an element of SO(4). We dene f : S 3 S 3 SO(4) by f (q1 , q2 ) = fq1 ,q2 . We claim that f is a surjective, two-to-one homomorphism. Now consider quaternions q1 , q2 , q3 , q4 S 3 . We check that f is a homomorphism: f (q1 q3 , q2 q4 )(x) = q1 q3 xq2 q4 = q1 q3 x(q2 q4 )1
1 1 = q1 q3 xq4 q2

= (f (q1 , q2 ) f (q3 , q4 ))(x). Hence f is a homomorphism, as claimed. Now, letting a0 = b0 = cos and a3 = b3 cos 2 0 0 sin 2

= sin in the matrix above we obtain 0 0 sin 2 1 0 0 . 0 1 0 0 0 cos 2

Notice that this is a rotation by 2 in the x1 x4 -plane. We can similarly select rotations to occur in any of the six coordinate planes of R4 . Recall that Proposition 2.4 stated that every rotation in SO(4) may be expressed as a product of the resulting six matrices. Thus, by Proposition 2.4 we may conclude that f is onto. Furthermore, notice that f (1, 1) is the identity and so is f (1, 1). Assume that f (q1 , q2 ) is the identity. Then it must be true that q1 xq2 = x for all x R4 . Let x = 1 to see that q1 must equal q2 . Now we have q1 xq1 = x and so q1 commutes with every quaternion x. Hence q1 is real. Since q1 is unit length, it must be 1. Thus the kernel of f contains only two elements, so f is two-to-one.

23

References
[1] Simon Altmann, Rotations, Quaternions, and Double Groups, Dover Publications, Mineola, NY, 1986. [2] Andrew Hanson, Visualizing Quaternions, Morgan Kaufmann, San Francisco, CA, 2006. [3] Ron Larson, Bruce Edwards, and David Falvo, Elementary Linear Algebra, 5 ed., Houghton Miin, Boston, MA, 2004. [4] Kristopher Tapp, Matrix Groups for Undergraduates, Student Mathematical Library, vol. 29, American Mathematical Society, Providence, RI, 2005.

24