View metadata, citation and similar papers at core.ac.

uk brought to you by CORE

provided by Elsevier - Publisher Connector

Quaternions and Matrices of Quaternions*

Fuzhen Zhang+
Department of Mathematical Science
Nova Southeastern University
Fort Lauderdale, Florida 33314

Dedicated to Robert C. Thompson

Submitted by Richard A. Brualdi

ABSTRACT

We give a brief survey on quatemions and matrices of quatemions, present new

proofs for certain known results, and discuss the quatemionic analogues of complex
matrices. The methods of converting a quatemion matrix to a pair of complex matrices
and homotopy theory are emphasized. 0 Elsevier Science Inc., 1997

1. INTRODUCTION

The family of quatemions plays a role in quantum physics [l, 18, 191. It
often appears in mathematics as an algebraic system-a skew field or
noncommutative division algebra [7]. While matrices over commutative rings
have gained much attention [6], the literature on matrices with quatemion
entries, though dating back to 1936 [39], is fragmentary. Renewed interest
has been wnnessed recently [3, 4, 9, 11-13, 31, 36, 35, 47, 481. In hopes that
it will be useful to a wide audience, a concise survey on matrices of
quatemion entries with new proofs for some known results is presented here.

* Part of this paper was presented at the AMS-MAA joint meeting, San Antonio, January
1993, under the title “Everything about the matrices of quatemions.”
’ The main work was done while the author was a student of R. C. Thompson’s at the
University of California, Santa Barbara. E-mail: zhang@alpha .nova. edu.

LINEAR ALGEBRA AND ITS APPLICATIONS 251:21-57 (1997)

0 Elsevier Science Inc., 1997 0024.3795/97/$17.00

655 Avenue of the Americas, New York, NY 10010 SSDI 0024-3795(95)00543-Z
22 FUZHEN ZHANG

The main obstacles in the study of quatemion matrices, as expected come

from the noncommutative multiplication of quatemions. One will find that
working on a quatemion matrix problem is often equivalent to dealing with a
pair of complex matrices. Homotopy theory has also been applied in the
study.
We begin by considering the following three basic questions whose
answers are well known in the case of complex matrices. Suppose that A and
B are n x n matrices with quaternion entries.

QUESTION1. If AB = I, the n X n identity matrix, is it true that

BA = Z?

QUESTION
2. Does every quatemion matrix A have an eigenvalue?

QUESTION 3. Is the numerical range of a quatemion matrix A always

convex, as is the classical numerical range of a complex matrix?

After a moment’s consideration, one will realize that answering these

questions is unexpectedly difficult.
Recall from elementary linear algebra that there are several ways to
answer the first question when A and B are complex matrices, one of which
is to utilize the fact that if det A # 0 then A is invertible; consequently B is
the inverse of A and BA = I. This approach apparently does not apply in our
case, because the determinant of a quatemion matrix makes no sense at this
point. We will give an affirmative answer to the question in Section 4 by
converting each of the quatemion matrices A and B into a pair of complex
matrices.
For Question 2, since the quatemions do not commute, it is necessary to
treat the linear systems Ax = Ax and Ax = xA separately. Rewrite the first
one as (AI - A)x = 0. In the complex case, the fact that det(hZ - A) = 0
has a solution guarantees an eigenvalue for A. This idea does not go through
for quatemion matrices. To avoid determinants, a topological approach is
needed. As one will see in Section 5, the latter system has been well studied,
while the former one is not easy to handle, and few results have been
obtained.
The answer to Question 3 is negative. The quatemionic numerical range
of a matrix is not convex in general, even for a complex normal matrix. In
contrast, the classical numerical range of a complex normal matrix, as is well
known, is the convex hull of the eigenvalues of the matrix. We will give in
Section 9 a complete characterization of the quatemionic numerical range of
a normal matrix with quatemion entries and discuss the convexity of the
QUATERNIONS AND MATRICES OF QUATERNIONS 23

upper complex plane part of the quatemionic numerical range of a matrix

with quatemion entries.
With terminologies of complex matrices similarly defined for quatemion
matrices, we will consider the above problems further, investigate other
aspects such as similarity, rank, determinant, and canonical forms, and
present certain new features that are hard to foresee and to prove. Some
results on quatemion matrices are analogous to those on complex matrices,
some are not.
We will focus only on matrices of quatemions. There is no attempt to
cover everything related to quatemions and review the historical develop-
ment of quatemions. For other aspects such as quatemion algebra and
analysis, see, e.g., [14], [20], [28], [31], [37], and [38].

2. QUATERNIONS AND EQUIVALENCE CLASSES

As usual, let C and [w denote the fields of the complex and real numbers
respectively. Let Q be a four-dimensional vector space over [w with an
ordered basis, denoted by e, i, j, and k.
A renl quaternion, simply called quaternion, is a vector

x = xoe + x,i + x2j + x,k E Q

with real coefficients x,,, x1, xz, xs.

Besides the addition and the scalar multiplication of the vector space Q
over (w, the product of any two of the quatemions e, i, j, k is defined by the
requirement that e act as a identity and by the table

.z=j2=k’= -1,
1

ij = -ji = k, jk= -kj=i, h = -& = j.

If a and b are any (real) scalars, while u, v are any two of e, i, j, k, then
the product (au)(bv) is defined as (abXuv). These rules, along with the
distribution law, determine the product of any two quaternions.
Real numbers and complex numbers can be thought of as quatemions in
the natural way. Thus xae + x,i + x2j + x,k is simply written as xg +
x,i + x2j + x,k.
24 FUZHEN ZHANG

For any x = x0 + x,i + x2j + x,k E Q, we define Re x = x0, the real

part of x; Co x = x1 + x,i, the complex part of x; Im x = x,i + x2j + x,k,
the imaginary part of x; x = x* = x - x,i - x2j - x,k, the conjugate of
x; and 1x1 = a= dm , th e norm of x. x is said to be a
unit quaternion if its norm is 1.
The following facts then follow from the definitions.

THEOREM2.1. Let x, y, and 2 be quaternions. Then

1. x*x = xx*, i.e., 1x1 = 1x*1;

2. 1* 1 is a norm on Cl, i.e.,

1x1 = 0 ifand only if x = 0;

Ix + yl < 1x1+ lyl;

IxyI = lyxl = 1x1lyl;

3. lx? + ly12 = +(1x + y12 + Ix - yl’);

4. x = IxIufor some unit quaternion u;
5. jc = Cj or jcj* = Efor any complex number c;
6. x*ix = <xi + xf - xi - x,z>i + 2(-x,x, + xlxe)j + 2(x,x, +
x,x,)k if x =x,, + x,i + x2j + x,k;
7. x = i<x + x*> + i<x + ix*i) + i<x + jx*j> + i<x + kx*k), and
x* = - i(x + ixi + jxj + kxk);
8. x2 = IRe xl2 - IIln xl2 + 2Re x Im x.
9. (xy)* = y*x*;
10. (xy)z = x(yz);
11. (x + y)2 # x2 + 2xy + y2 in general;
12. x* = x if and only if x E R;
13. ax = xu for euey x E Q, if and only if a E R;
14. x*/l x 1’ is the inverse of x if x z 0, and if x-l denotes the inverse of
x, then lx_11 = 1/1x1;
15. x2 = - 1 has infinitely many of quaternions x as solutions;
16. x and x* are solutions of t2 - 2(Re x)t + /xl2 = 0;
17. every quaternion q can be uniquely expressed as q = cl + c2 j,
where cl and c2 are complex numbers.
QUATERNIONS AND MATRICES OF QUATERNIONS 25

Two quatemions x and y are said to be similar if there exists a nonzero

quatemion tc such that uP1xu = y; this is written as x - y. Obviously, x and
y are similar if and only if there is a unit quatemion o such that u - ’ xu = y,
and two similar quatemions have the same norm. It is routine to check that
- is an equivalence relation on the quatemions. We denote by [xl the
equivalence class containing x.

LEMMA 2.1. Zf q = q. + q,i + q.,j + q3k, then q and q0 +

J-i are similar, namely, q E [qO + J-i].

Proof. Consider the equation of quatemions

qx = x(qO + 4-i). (1)

It is easy to verify that x = Cd- + ql) - qSj + q,k is a

solution to Equation (1) if qf + qi # 0. For the case where q is a complex
number, Theorem 2.1, part 5, may be used. n

One can also prove the lemma by changing (1) to a linear homogeneous
system of real variables.
It is readily seen that 1 x] contains a single element if and only if x E R. If
r G R, then [xl contains infinitely many quatemions, among which there are
only two complex numbers that are a conjugate pair; moreover x - x* for
every quatemion x.
This lemma yields the following theorem.

THEOREM 2.2 (Brenner, 1951; Au-Yeung, 1984). Let x = x0 + x,i +

xzj + x,k and y = y,, + y,i + yzj + y,k be quaternions. Then x and y
are similar if and only if x0 = y0 and x1” + xi + x3” = yt + y: + yi, i.e.,
Re x = Re y and IIm 1~1= [Im y 1.

REMARK 2.1. An alternative way [24] to define quatemions is to consider

the subset of the ring M,(C) of 2 X 2 matrices with complex number entries:
26 FUZHEN ZHANG

Q’ is a subring of M,(C) under the operations of M,(C). Q and Cl! are

essentially the same. In fact, let

l I
aI a2
A:q=a,+a,jEQ+q’= - - EQ’.
- a2 a,

Then _& is
. b’j1 ec ti ve and preserves the operations. Furthermore, 191’ = det 9’,
and the eigenvalues of 9’ are Re 9 f IIm q/i.

3. THE FUNDAMENTAL THEOREM OF ALGEBRA

FOR QUATERNIONS

The fundamental theorem of algebra over the quatemion skew field has
gained little attention. It is of less importance than that over the complex
number field, though some problems may reduce to it.
Let f(r) = a,x” + a,,_,x”-’ + ... +a,x + a, be a polynomial in x
over Q and a, z 0. It is natural to ask whether f(z) = 0 always has a
solution and how many solutions f(x) = 0 has (as an example, x2 + 1 = 0
has infinitely many solutions in Q). More generally, let F(X) be a polynomial
in x with general terms a,, xa, x 1.1 xa,, , where the a,‘s are quatemions. Does
F(x) = 0 have a solution in C2??
This question was partially answered in [16].

THEOREM 3.1 (Eilenberg and Niven, 1944). Let

f(r) = agxal x ... xa,, + 4(X)> (2)

where a,, a,, . . . , a,, are nonzero quaternions, x is a quaternion indetermi-

nant, and 4(x) is a sum of‘afinite number oj-similar monomials b, xb,x ...
xb,, k < n. Then f(x)= 0 h as at least one quaternion solution.

The proof in [16] 1sa topological one and is accomplished by showing that
f(x) and g(x) = x” are homotopic mappings and the latter has degree n.
QUATERNIONS AND MATRICES OF QUATERNIONS 27

Note that f<~> has only one term with “degree” n. As a consequence, we
have the following result which was first shown in [32].

COROLLARY 3.1 (Niven, 1941). Let a, be quaternions, a,, # 0 and

n > 1. Then a,,~” + an_l~npl + 1.. +a, x + a, = 0 has at least one solu-
tion in Q.

Another result worth mentioning is the discussion of the solutions of the

equation xa + br = c over an algebraic division ring [26].

THEOREM 3.2 (Johnson, 1944). Ifa, b, c E Q, and a and b are not

similar, then xa + bx = c has a unique solution.

4. QUATERNION MATRICES AND THEIR ADJOINTS

As has been noticed, there is no theory of eigenvalues, similarity, and

triangular forms for matrices with entries in a general ring; and a field can be
thought of as a “biggest” algebraic system in which the classical theory of
eigenvalues, etc., can be carried through. The question of the extent to which
the properties of matrices over fields can be generalized if the commutativity
of the fields is removed may be of interest and of potential applications to
different research fields. The rest of the paper is devoted to this topic.
Let M,,. .(a), simply M,(Q) w h en m = n, denote the collection of all
m X n matrices with quaternion entries. Besides the ordinary addition and
multiplication, the left [right] scalar multiplication is defined as follows: For
A = (n,,) E M,,,(Q), q +z Q,

4A = (4%) [A4 = (at4)1-

It is easy to see that for A E M,,.(Q), B,,,,(Q), and P, q E Q

(4A)B = Q(AB),

(A41B = A(4B),

(P4)A = P(4A)-

Moreover M,, x .(C.l) is a left [right] vector space over Q.

28 FUZHEN ZHANG

All operators for complex matrices can be performed except the ones such
as

(qA)B = A(@)

in which commutativity is involved.

Justas for complex matrices, we associate to A = (a,$,) E M,,, x ,,(Q) with
z = (a,,) = (&>, the conjugate of A; A?’ = (cz,,~) E M,,.,,,(Q), the transpose
of A; and A* = (x)r E M,,.,,,(Q), the conjugate transpose of A.
A square matrix A E M,,(Q) is said to be normal if AA* = A*A; Hermi-
tian if A* = A; unitary if A*A = 1, the identity matrix; and invertible if
AB = BA = Z for some B E M,,(Q).
Just as with complex matrices, one can define elementary row (column)
operations for quatemion matrices and the corresponding elementary quater-
nion matrices. It is easy to see that applying an elementary row (column)
operation to A is equivalent to multiplying A by the corresponding elemen-
tary quatemion matrix from left (right) and that any square quatemion matrix
can be brought to a diagonal matrix by elementary quatemion matrices.
A list of facts follows immediately, some of which are unexpected.

THEOREM 4.1. Let A E M,,,x,,<U2>, B E M,,x,,(Q). Then

1. (&T = (AT);
2. CAB)*
- --= B*A*;
3. AB # AB in general;
4 ( AB)T # B?‘A’ in general;
5: CAB)-’ = B-‘A-’ ifs and B are invertible;
6. (A*)-’ = (A-‘)* ifA is invertible;
7. (A)-’ # A- ’ in general;
8. (A’)-’ # (A-‘)r in general.

For part 7 or 8, take A = (; ;), A-’ = ( ii 1;).

One of the effective approaches id studying matrices of quatemions may

be by means of converting a matrix of quaternions into a pair of complex
matrices. This was first noticed by Lee [29]. We use his idea to show a
proposition below which has appeared in [13, 301 with different proofs.

PROPOSITION4.1. Let A, B E M,(Q). IfAB = I, then BA = 1.

Proof. First note that the proposition is true for complex matrices.
Let A = A, + A, j, B = B, + B,j, where A,, A,, B,, and B, are
QUATERNIONS AND MATRICES OF QUATERNIONS 29

n X n complex matrices. Then

-
AB=Za (A,B,-A,B,)+(A,B,+A,B,)j=Z

-(n,,x,)(_; ;) =(Z>O)

-
*B,A,-B,A, =I, B,A,+B,A, =0

-
= (B,A, - B,A,) + (B,A, + B,x)j = I

* BA =I. n

For A = A, + A, j E M,,(Q), we shall call the 2n X 2n complex matrix

uniquely determined by A, the complex adjoint matrix or adjoint of the

quatemion matrix A, symbolized xA. Note that if A is a complex matrx, then

The adjoint of a quatemion matrix has been exploited in the study of its
quatemionic numerical range [3, 4, 27, 361; this we will enlarge on later.
If
30 FUZHEN ZHANG

for instance, is a 2 X 2 elementary quaternion matrix, where 9 = 9” + q,i +

qzj + q3k = (90 + y,i) + (cl2 + q3i)j, then

‘1 90 + 9,i 0 yz + y3i\
0 1 0 0
xe= 0 -9; + 9G 1 90 - 9,i *
\o 0 0 1 I

Notice that det( ,yp) = I xpl = 1. More generally,

THEOREM 4.2 (Lee, 1949). Let A, B E M,,(Q). Then

1. XI = zzn;
2. XiB = XA X8;

3. XA+B = XA t XBJ
4. xA* = ( &I*;
5. ~~-1 = (xA)-’ fA_’ exists;
6. xA is unitaq, Hermitian, or normal if and only if A is unitary,
Hermitian, or normal, respect&y;
7. if IA is a Jordan block with A on the diagonal, then

The proof of Theorem 4.2 is by direct verification.

PROPOSITION 4.2. Let A and B be any two n X n complex matrices.

Then

IA B>O
-13
I A’. (3)
QUATERNIONS AND MATRICES OF QUATERNIONS 31

In particular,

1. I xc I 2 0 for any C E M;(Q);

2. IZ+xAI>o;
3. 1fi + BBI 2 0, whenever A and B commute.

Proof. It is known [22,p. 2531 that for every n X n complex matrix A

the nonreal eigenvalues of AA occur in conjugate pairs, and every negative
eigenvalue of AA has even algebraic multiplicity. It follows that 1Z + _%4 > 0,
that is.

I I
I A
> 0,
-A z

for any complex matrix A.

Now we assume, without loss of generality (using continuity), that A is
invertible. Since

we get

I I
A B
> 0.
-B A

When A and B commute,

THEOREM 4.3. Let A E M,(Q). Then the following are equivalent:

1. A is invertible;
2. Ax = 0 has a unique solution 0.
3. ( xAl # 0, i.e., xA is invertible.
4. A has no zero eigenvalue. More precisely, if Ax = hx or Ax = xh for
some quaternion A and some quaternion vector x # 0, then A # 0.
5. A is the product of elementary quaternion matrices.
32 FUZHEN ZHANG

Proof.1 = 2: This is trivial.

2 j 3: Let A = A, + A, j, x = x1 + x2j, where A, and A, are complex
matrices, x1 and x2 are complex column vectors. Then

AX = (Al + ki)( x1 + xi)

= A,x, + A,x,j + A,jx, + A,jx2j

= ( A,x, - A,<) + (A,x, + A,G)j.

Thus Ax = 0 is equivalent to

A,x, -A,< = 0

and

A,x, + A,r, = 0.

Rewriting the above equations as

A,x, +A,(-x,) =0

and
- _
(-A,)x, + A,(-x,) =O,

we have that AX = 0 if and only if

that is, ,yA(xl, - -x,)r = 0. It is immediate that Ax = 0 has a unique solution

if and only if xA(xl, - G)’ = 0 does as well; namely, 1xAl # 0.
3 = 4: This can be seen from 2 * 3.
3 q 1: Suppose that xA is invertible and that
QUATERNIONS AND MATRICES OF QUATERNIONS 33

Let B = B, + B, j. It is easy to check that BA = I and A is invertible by

Proposition 4.1.
1 - 5: This is because if A is invertible, then A can be brought to the
identity matrix by elementary row and column operations. n

REMARK 4.1. Let A be an n X n quatemion matrix. Applying 1 in the

Remark 2.1 to each entry of A, one obtains a complex matrix, another
representation of A, denoted by MA, which has been used in the studies of
quatemionic numerical ranges [27, 361 and of similarity [39]. It is not difficult
to see that there exists a permutation matrix P such that P’x~ P =-HA.

5. EIGENVALUES

We now turn attention to the eigenvalues of quatemion matrices. Since

left and right scalar multiplications are different, we need to treat AX = Ax
and Ax = x A separately.
A quatemion A is said to be a lef (right) eigenvalue provided that
AX = Ax (AX = xA.). The set (A E Q] AX = A, for some x # 0} is called the
lef spectrum of A, denoted by u,( A). The right spectrum is similarly defined
and is denoted by u~,(A).

EXAMPLE 5.1. Let

Then the left eigenvalues of A are 1 and i, whereas the right eigenvalues of
A are 1 and all the quatemions in [il.

EXAMPLE 5.2. Let

A=

Then A has two left eigenvalues + l/(fixi + j>, and infinitely many right
eigenvalues which are the quatemions satisfying A4 + 1 = 0. Note that the
c~,( A) is not discrete and that al(A) n u,.,(A) = 4.
34 FUZHEN ZHANG

EXAMPLE 5.3. Let

A= 0 i
i -i 0' 1

Then k is a left eigenvalue but not a right one. Note that A is Hermitian.

THEOREM 5.1. Let A E M,(C) be an upper triangular matrix. Then a

quaternion A is a lef eigenvalue ofA if and only if A is a diagonal entry.

The proof is basically the same as in the complex case. The situation of
general matrices is much more complicated. Generally speaking, there is no
very close relation between left and right eigenvalues. For real matrices,
however, we have the following theorem.

THEOREM 5.2. Zf A is a real n X n matrix, then the lef and right

eigenvalues of A coincide; that is, al(A) = a,( A).

Proof. Let A be a left eigenvalue of A, i.e., AX = Ax for some x # 0.

For any quatemion q # 0, we have

(qAr/-‘)qx = (qhq-‘)qx

and

Aqx = (qhq-l)qx,

since A is real. Taking 0 # q E Q such that q Aq - ’ is a complex number and

writing qx = y = yI + y2j, we have Ay, = y,qAq-l and Ay, = y2qAq-‘.
It follows that A is a right eigenvalue of A. Similarly one can prove that every
right eigenvalue is also a left eigenvalue. n

For any given matrix A E M,(Q), does there always exist A E Q and
nonzero column vector x of quatemions such that AX = Ax? This is of
course a very basic question. It was raised in [I4, p. 2171 and was later proved
by Wood [411. Woods p roof, adopted below, is a purely topological one.
QUATERNIONS AND MATRICES OF QUATERNIONS 35

THEOREM 5.3 (Wood, 1985). Every n X n quaternion matrix has at least

one lef eigenvalue in Cl.

Proof. Write Ax = Ax as (Al - A)x = 0, and assume that Al - A is

invertible for all A E Q.
Consider the general linear group GL(n, Q), the collection of all invert-
ible n x n matrices of quatemions.
Let

f(t, A) := ft( A) = thl - A, O<t < 1, tAl= 1,

and

g(t, A) := gt( A) = AZ - tA, o<t<1, IAl=l.

Then f and g are homotopies on GL(n, Q). Note that

f,,(A) = -A,

fi( A) = AZ - A = g,(A),

and

g,(A) = AZ.

Thus g,, is homotopic to fo. On the other hand, f0 and g,, viewed as the
maps from the S-sphere S 3 into GL(n, Q), correspond to integers 0 and n,
respectively, in n,GL(n, Q), the third homotopy group (isomorphic to the
integers [53]) of GL(n, Q). This is a contradiction. n

An elementary proof for the cases of 2 X 2 and 3 X 3 matrices has been

recently obtained [34].
By contrast, the right eigenvalues have been well studied, and this sort of
eigenvalues are more useful. We shall simply call the right eigenvalues
eigenvalues for the rest of this paper.

LEMMA 5.1. Zf A E M,,x,,(Q), m<n, then Ax=0 has a nonzero

solution.
36 FUZHEN ZHANG

Proof. Let A = A, + A, j, x = x, + x2j. Then An: = 0 becomes

which has nonzero complex solutions, since 2m < 2n. W

Using this lemma repeatedly, we have the following

LEMMA 5.2. Let u1 be a unit column vector of n quaternion components.

Then there exist n - 1 unit column vectors u2, . . . , u, of n quaternion
components such that {ul, u2,. . . , UJ is an orthogonal set, i.e., uzu, = 0,
s # t.

This lemma may also be interpreted as follows: if u1 is a unit vector, then

one can construct an n x n unitary matrix U with u, as its first column.
Since AX = xh * A(xq) = (Ax)q = xhq = (xqxq-‘Aq), it follows that
if A is an eigenvalue of A, then so is q-‘Aq for any nonzero quatemion q.
Thus if A is a nonreal eigenvalue of A, so is any element in [A]. Therefore A
has finite eigenvalues if and only if all eigenvalues of A are real.

THEOREM 5.4 (Brenner, 1951; Lee, 1949). Any n X n quaternion ma-

trix A has exactly n (right) eigenvalues which are complex numbers with
nonnegative imagina y parts.

Those eigenvalues are said to be the standard eigenvalues of A.

Proof. As before, we write A and x as A = A, + A, j, x = x, + x2j,

where A,, A, are n X n complex matrices, and x1, x2 are complex column
vectors. Then Ax = xh is equivalent to

(4)

or

(5)

where A is a complex number.

QUATERNIONS AND MATRICES OF QUATERNIONS 37

Since

is a 2n x 2n complex matrix, it has exactly 2n complex eigenvalues (includ-

ing multiplicity).
Notice that if a complex matrix X is similar to its conjugate, then the
nonreal eigenvalues of X occur in conjugate pairs (this can be seen by
considering /AZ - XI = IAZ - xl = IAl - XI for real A). xA is similar toL,
so the nonreal eigenvalues of xA appear in conjugate pairs (with the same
multiplicity). With real eigenvalues, we show by induction that every real
eigenvalue of xA occurs an even number of times.
It is trivial when n = I. Let n z 2.
Suppose Ax = xh = Ax, where A is real and x # 0 is a unit vector. Let
uz>..., U, be such that U = (x, u2,. . . , u,) is a unitary matrix, and let

U*AU =
( 1
; B” ,

where B is an (n - 1) X (n - 1) quatemion matrix and (Y is a quatemion

row vector of n - 1 components. It is easy to see that there is a 2n X 2n
invertible matrix T such that

T-‘x”:xA x”T = ( xaTJP1)x.+,( XUT) =

By induction, each real eigenvalue of xB appears an even number of times.

Therefore xA has exactly 2n eigenvalues symmetrically located on the
complex plane, and A has exactly n complex eigenvalues on the upper half
complex plane (including the real axis). The conclusion on the real eigenval-
ues can also be observed by a continuity argument with A - t iZ substituting
for A. n

COROLLARY 5.1 (Lee, 1949). Let A and B be n X n complex matrices.

Then every real eigenualue (if any) of the matrix
38 FUZHEN ZHANG

appears an even number of times, c~ncl the complex eigewalues ofthat matrix
appear in conjugate pairs.

Note that Proposition 4.2 follows from Corollary 5.1 immediately. The
structure of the Jordan canonical form of the block matrix in the above
corollary will be given in the next section.
The following illustrates why this idea does not come through for the (left
eigenvahie) system An- = As.
Let A = A, + h2j. Then

Ax = (A, + A2j)( x, + x2j)

= A,?;,+ A,x,j + A,jx, + A2jx,j

= (A,s, -A?G) + (A,s, +A,c)j,

and

As = (A,r, - A,G) + (A,*,+ A,c)j.

Thus the quatcrnion system Ax = AX is equivalent to the complex system

which is virtually different from (4) and (5).

The above equation can be rewritten as

A, - A,1
0.
-A, + A,1

Therefore the square quaternion matrix A = A, + A, j has a left quaternion

eigenvahle if and only if there exist complex numbers A, and A2 such that

A, - A,1 A, -A,1
- - 0. (6)
-A, +\,I A, - A,1 =
QUATERNIONS AND MATRICES OF QUATERNIONS 39

Note that, by Proposition 4.2, the left-hand side of the above identity is
nonnegative. Can this function, regarded as a function on [w4, vanish? It is
unknown if this approach can go through without using Woods result.

COROLLARY 5.2. Let A be an n x n matrix with quarternion entries.

Then A has exactly n (right) eigenvalues up to equivalent classes.

QUESTION 5.1. How many left eigenvalues does a square quatemion

matrix have?

QUESTION5.2. Is there an elementary proof for the existence of the left

eigenvalues?

QUESTION5.3. Does there exist a matrix B such that xA = BB?

QUESTION5.4. What conditions can be imposed on A and B when

I I
A
= O?
-E Pi

QUESTION5.5. Investigate v~I(A).

REMARK5.1. The term “left eigenvalue” has appeared in literature with

different meanings. Over a general skew field, it is defined [I4, 151 to be the
element A such that xA = Ax for some x # 0. As noticed in [15], a matrix
may have no left eigenvalue in this sense. In the quatemion case, however,
right and left eigenvalues make no difference in the study, since xA = Ax if
and only if A*x* = x*A*. The elements A satisfying Ax = hx for some
x # 0 have been called singular eigenvalues of A [14, 411, and are the same
as the left eigenvalues in the present paper.

6. CANONICAL FORMS

There are three sorts of canonical forms which are of fundamental

importance in linear algebra: rational forms, Schur forms, and Jordan forms.
Every square matrix can be reduced under similarity to a direct sum of block
matrices over a (commutative) field, and to the Schur and Jordan forms over
40 FUZHEN ZHANG

an algebraically closed field. We discuss the Schur and Jordan canonical

forms for quatemion matrices. For the rational canonical form for a matrix
over a general skew field, see [I5].

THEOREM 6.1 (Brenner, 1949). Zf A E M,(Q), then there exists a uni-

ta y matrix U such that U*AU is in upper triangular form.

This can be shown by the mathematical induction and Lemmas 5.1 and
5.2.
As noticed earlier, every diagonal entry of a triangular matrix is a left
eigenvalue. The following says that they are also right eigenvalues.

THEOREM 6.2 (Brenner, 1951). Zf A E M,,(Q) is in triangular form,

then every diagonal element is a (right) eigenvalue of A. Conversely, every
(right) eigenvalue of A is similar to a diagonal entry of A.

Proof. Let A be an n X n upper triangular quaternion matrix with

diagonal entries A,, A,, . . . , A,,. If n = 1, there is nothing to show.
Suppose it is true for n - 1. Let A be of order n and be partitioned as

A=

We see that A, is an eigenvalue of A, since

A(x,,O ,..., o)~= (x,,o ,..., o)?‘A,, x, # 0.

A, is of order n - 1, and has eigenvalues A,, . . . , A, by induction. All we

need to show is that the eigenvalues of A, are eigenvalues of A.
Suppose A is one of h,, . . . , A,,. We may assume that A and A, are not
similar, and let A, y = yh, y # 0. Applying Theorem 3.2, we have a quater-
nion x such that

h,x + ‘yy = xh.

Then

A(;)=(: ;,)(j)=(h’rA;l:ly)=(;;)=(;)h.
QUATERNIONS AND MATRICES OF QUATERNIONS

So A is an eigenvalue of A. Conversely, if A is an eigenvalue of A, then the

complex number contained in [h] is an eigenvalue of x~. On the other hand,
the eigenvalues of xA are also the eigenvalues of A. The conclusion follows.
n

It turns out that if a matrix A of quaternions is brought to a triangular

form T, then the only eigenvalues of A are the diagonal elements of T and
the quatemions similar to them.
The following is the generalization of the Schur canonical form of
complex matrices to quatemion matrices, which has been used in the study of
numerical ranges of quatemion matrices [3, 4, 36, 481.

THEOREM 6.3 (Brenner, 1951; Lee, 1949). Let A he an n X n matrix of

quaternions. Then there exists a unitary matrix U such that U*AU is an upper
triangular matrix with diagonal entries h, + k ,i, . . . , h, + k,i, where the
h, + k,i’s are the standard eigenvalues of A, i.e., k, 2 0, t = 1,2, . . . , n.

For the proof, A is first brought to an upper triangular matrix with

quatemions on the main diagonal by a unitary matrix, then each diagonal
quatemion is changed to a complex number with nonnegative imaginary part
by a pair of elementary quaternion matrices.
Note that Corollary 5.1 foll ows from Theorem 6.3 immediately.

COROLLARY 6.1. Let A he an n X n quaternion matrix with standard

eigenvalues h, + k,i, . . . . h, + k,i. Then

q(A) = [h, + k,i] U *** U [h,, + k,,i].

COROLLARY6.2. A is normal if and only if there exists a unita y matrix

U such that

U*AU = diag{ h, + k,i, . . . , h,, + k,i}

and A is Hermitian if and only if k , = ... = k,, = 0.

We now consider the Jordan canonical form of an n x n matrix with

quatemion entries. Let A E M,(Q). If 1 ’ p ossible for A to have the Jordan
‘t IS
form J with the standard eigenvalues of A on the diagonal of J; that is,
F’AS = J for some quatemion invertible matrix S, then

XSS’XAXS = x] =
42 FUZHEN ZHANG

This means that x, is the Jordan form of xA. Conversely, let j be the Jordan
form of xA, and let P be an invertible matrix such that P-‘/YAP = J. If it can
be shown that P and J take the forms

respectively, then A is similar to a complex n X n Jordan form Jr by a

direct verification. Following this line, a thorough analysis of the eigenvalues
and the corresponding eigenvectors of x~, result in the existence of P and J
with the desired form [do].

THEOREM 6.4 (Wiegmann, 1954). Every n X n quaternion matrix A has

the Jordan canonical form with the standard eigenvalues of A on the diagonal.

COROLLARY 6.3. For any II X n complex m&rices A and B, the block

m&-Lx

has the Jordan canonical f&m

where J is a Jordan form of some n X n complex matrix. Consequently, all the

Jordan blocks are paired.

REMARK 6.1. As we have noted, many terms for describing complex

matrices can similarly be defined for matrices of quatemions, and the parallel
conclusions may be derived for the quaternion cases. For instance, an n X n
Hermitian matrix A is said to be positive (semi-Idefinite if x*Ax > ( > > 0 for
all nonzero column vectors x of n quaternion components. It is seen that a
Hermitian matrix A of quatemion is positive semidefinite if and only
if A has only nonnegative eigenvalues, and if and only if xA is positive
semidefinite.
QUATERNIONS AND MATRICES OF QUATERNIONS 43

7. RANK, SIMILARITY, AND DECOMPOSITIONS

Many aspects of quaternion matrices, such as rank, linear independence,

similarity, characteristic matrix, Gram matrix, and determinantal expansion
theorem, have been discussed in [ll-131, [14, 151, [39], [40], and [42-461.
One can define, in the usual sense, the left and right linear independence
over Q for a set of vectors of quaternions. Note that two linearly dependent
vectors over Q may be linearly independent over c. One can also easily find
an example of two vectors which are left linearly dependent but right linearly
independent. The G ran-Schmidt process (see, e.g., [22, p. 151) is still
effective.
The rank of a quaternion matrix A is defined to be the maximum
number of columns of A which are right linearly independent. It is easy to
see that for any invertible matrices P and Q of suitable sizes, A and PAQ
have the same rank. Thus the rank of A is equal to the number of positive
singular values of A (see Theorem 7.2). If a matrix A is of rank r, then r is
also the maximum number of rows of A that are left linearly independent,
and A is nonsingular (or invertible, meaning BA = AR = 1 for some B) if
and only if A is of (full) rank n.
Let 0: denote the collection of column vectors with II components of
quatemions. Q!’ is a right vector space over Q under the addition and the
right scalar multiplication. If A is an m X n quatemion matrix, then the
solutions of Ax = 0 form a subspace of Q:!, and the subspace has dimension
r if and only if A has rank n - r.
We now give the polar and the singular-value decompositions for quater-
nion matrices.
The proof of the following theorem is based on [40].

THEOKEM 7.1 (Polar decomposition). Let A E M,,(Q). Then there exist a

quaternion unitary matrix U and a quatemion positioe semidefinite Hermitian
matrix H such that A = HU. Moreooer H and U are unique when A is of
rank n.

Proof. Let X be an n X n quatemion unitary matrix such that X *AA*

X = D is a diagonal matrix with the squares of the singular values of A on
the diagonal. Let x.4 = KY be a polar decomposition of the complex matrix
,Y~, where K is a 2 n X 2 n positive semidefinite Hermitian (complex) matrix
and Y is a 2n X 2n unitary (complex) matrix. We show that K = x,, for
some positive semidefinite Hermitian matrix H, and Y = xr, for some
unitary quatemion matrix U.
44 FUZHEN ZHANG

First note that

XA&,’ = K2, XX*XAXA*XX= XD.

Thus

xx* K2xX= XD
and therefore

so

Take H = XD’12X*. Then x,, = xnY. We assume that A is nonsingular

(a little more work is needed for the singular case). Then H is nonsingular
and

Y = ( XI{) -’ XA = x11-1 XA = XH-‘A.

Set U = H- ‘A. Then A = HU, and H and U are as desired. n

The following theorem follows immediately.

THEOREM 7.2 (Singular-value decomposition). Let A E M,,,, ,,(a> be of

rank r. Then there exist unitary quaternion matrice.s U E M,,,(Q) and V E
M,(Q) such that

UAV =

where D, = diag(d,, . . . , cl,) and the d’s are the positive singular values of A.

It is always desirable to convert a quaternion matrix problem into a

complex one. The next two theorems, following from Theorem 7.2 and
Theorem 6.4, are in this direction.
QUATERNIONS AND MATRICES OF QUATERNIONS 45

THEOREM 7.3 (Wolf, 1936). The rank of a quaternion matrix A is r if

and only if A has r nonzero singular zjalues, and if and only if the rank of its
complex adjoint xA is 2r.

It is readily seen that A, A*, AA*, and A*A are all of the same rank.
Two square matrices of quatemions A and B are said to be similar if
there exists an invertible quatemion matrix S of the same size such that
S-‘AS = B. It is immediate that similar quatemion matrices have the same
(right) eigenvalues. This is not true for left eigenvalues. As one sees below,
even the trace is not preserved under similarity. Recall that the trace of a
square matrix is the sum of the main diagonal entries of the matrix, which is
equal to the sum of the eigenvalues of the matrix in the complex case.

EXAMPLE 7.1. Let

Then A and B are similar, but they have different traces and different left
eigenvalues.

EXAMPLE 7.2. Let

U= , and B = U*AU.
1
- -*
\ v5J di I

Then U is unitary and the main diagonal entries of B are i = ik and

-i - $k. However, tr A = 0, tr B = -k. Note that tr A and tr B have the
same real part, but they are neither necessarily equal nor even similar (as
quatemions).

EXAMPLE 7.3. Let

46 FUZHEN ZHANG

Then the column vectors of A arc left linearly dependent and right linearly
independent. A is of rank 2, a11d is invertible.
It is easy to show that an II X )I qllaternion matrix A is diagonalizable if
and only if A has II right linearly independent eigenvectors belonging to
right eigenvahles. The following example says that the eigenvectors belonging
to different eigenavhles are not necessarily (right or left) linearly indepen-
dent.

EXAMPLE 7.4. Let

Then the eigrnvectors ( 1, 0)” ad (i + j, 0)” helonging to the (right) eigenval-

ws i and j, respecti\~rly, are not linrarlv independent. In addition, A is not
diagonalizable. Note that if an )L X II matrix has n distinct right eigenvalues,
no pair of which arc similar, then the matrix is diagonalizable.

COKOI.I,~IW ‘i. 1. A E Jr,,(Q) is rlirqpuzliznble if and only if x,, is

rliclfiot~~~li,c~l~lc.

~lltXl-ION 7.1. ~qqmc~ A E L%f,,(Qe)1I i15 II distinct left eigenvalues, any

two of which are not similar. Is A diugonalizal)le?

KEhlARK ‘i. I. Some parallel properties of normal plte~iOn IKhC!eS to

tliosc of ~01nplex normal niatriccs are ol)tained in [40]. For example, two
normal cpaternion matrices A ad B are commutative if and only if they can
be diagonalized by the same unitary transformation.

8. DETERMINANTS AND THE CAYLEY-IIAMILTON THEOREM

We now discuss the determinants of quaternion matrices. For this pur-

pose, we first define the determinant of a square quatemion matrix A to be
that of its complex adjoint x,,, then derive some results such as the Cayley-
QUATERNIONS AND MATRICES OF QUATERNIONS 47

Hamilton theorem, which have been obtained by several authors in the last
decade through extremely difficult approaches.
Let A be an n x n quatemion matrix, and let xA be the complex adjoint
matrix of A. We define the q-determinant of A to be 1~~1,simply called the
determinant of A, denoted by lAlq, i.e., l Al, = l xAl by definition.
It is immediate that l Al, = l All Al = (det Al” when A is a complex
matrix.
We have the following results on determinants of quaternion matrices.

THEOREM 8.1. Let A and B be n X n quaternion matrices. Then:

1. A is invertible * I Al, # 0.
2. IABI, = IAI,IB(,, consequently IA-‘I, = lAl,I’, ifA_’ exists.
3. IPAQI, = l Al,, for any elementary matrices P and Q.
4. I Al, = n:= Jh,12 > 0, where the A,‘s are the standard eigenvalues of
A.
5. Cayley-Hamilton theorem: Let A be a square matrix of quaternions,
and FA(A) = lAZ,, - ,yAl, called the characteristic polynomial of A, where A
is a complex indeterminant. Then F,(A) = 0, and F,(A,) = 0 if and only if
A, is an eigenvalue.
6. Zf A and B are similar, then l Al, = (Bl, and F,(A) = F,(A).
7. Hadamard theorem: Let A = (a,,) be positive semi&finite Hermi-
tian. Then 141, Q 17;=,q,!t, with equality if and only if A is real diagonal.
Furthermore, if A is partitioned as

A=

then 141, < IBI,lDl,, with equality if and only if C = 0 or I BI,IDI, = 0.

Proof. 1: See Theorem 4.3.

2: This is because xAB = xA xs.
3: It is sufficient to observe that ( xpI = 1 when P is an elementary
quatemion matrix.
4: )Al, > 0 was proved in Proposition 4.2 (notice that the proof did not
involve quatemion matrices).
We know from Theorem 6.3 that for any n X n quatemion matrix A there
exists a unitary matrix U such that A = U*DU, where D is an upper
triangular matrix with the standard eigenvalues A,, . . . , A,, of A on the
diagonal. It is easy to compute 141, = I xAl = l x01 = n~=llA,12.
48 FUZHEN ZHANG

5: It follows from Corollary 5.1 that F,(A) is a real coefficient polyno-

mial. And xfc A) = f( xn) for any real coefficient polynomial f.
Noting that FA( xA) = 0 by the Cayley-Hamilton theorem for complex matri-
ces, we have x~,,(~) = FA( ,yA) = 0, hence I$( A) = 0, as desired.
The second conclusion of this part follows from the fact that every
quatemion is similar to a complex number and that AX = xh is equivalent to
x*(x,, - <)“ = h(x,, - Q”, where x = x1 + zz j and A is a complex num-
ber.
6: This is due to the fact that xA and xH are similar.
7: If A is positive semidefinite Hermitian, we write A = G*G. Then
x,, = ( ,~o)*xo; therefore xA is a complex positive semidefinite Hermitian
matrix (the converse is also true). By the Hadamard theorem for complex
positive semidefinite Hermitian matrices, we have

with equality if and only if A is (real) diagonal.

The latter conclusion follows from that

where P is a product of some permutation matrices. n

From the point of view of how the theory of determinants is algebraically

developed for skew fields, it may be worthwhile defining determinants of
quatemion matrices by their entries. Several versions of the definition of
determinants for quatemion matrices have appeared [ll-13, 30, 42-461. The
definition introduced below, of a rlouhle determinant, is the one that gives us
many properties resembling ordinary ones.
Let S,, be the symmetric group on {1,2, . . . , n). For any (+ E S,, we write
u as the product of disjoint cycles:

u = (n,i,i,, ... i,y)( n2 j, j, 1.. j,) ... (n,k,k, *** k,),

with the convention that each n,, t = 1,2, . . . , r, is the largest number in its
cycle and that n = n, > n2 > ... > n, > 1. For example, if u is the
identity, then IT = (n)(n - 1) . ..(2)(l).
QUATERNIONS AND MATRICES OF QUATERNIONS 49

Let EC(T) be the sign of permutation cr. For A E M,(Q), we associate to

A the quatemion

x . ..
an,k,ak,k3 “’ ak~_,k,ak,n,

= c E((T)cz,.
CTE S”

It is obvious that det A = 1 if A is the identity matrix, and det A is the

same as the usual determinant when all a,,‘s commute. det A so defined
carries few properties of the determinants of complex matrices. For instance,
the column vectors of

A=

are right linearly independent, but det A = i = jk = 0.

Denote ] AId = det( A*A), called the double determinant of A. Based on
this definition, Chen [11-131 and Xie [42-461 obtained the results in the
previous theorem in their series of work, with extraordinary effort.
We next draw the conclusion that 1A], and ] AId are indeed the same. As
we have seen, 1Al, = n:= IIA,1’.To show that IAId is the same quantity, we
need a result of Chen [13, Theorem 4 and 51.

LEMMA8.1 (Chen, 1991). For any n X n quaternion matrices A and B,

IAId = IMId = ITAh, t ABki = IAldBlc~,

where T is any matrix obtained from the n X n identity matrix by adding a

quaternion scalar multiple of one row to another. In particular, 1P-‘l,l =
I PI,I ’ when P is an invertible matrix.

It follows from the definition and the above lemma that if D is an upper
triangular matrix with diagonal entries A,, . . . , A,, then IDl,, = n:= II h,12.
Applying Theorem 6.3, we have
50 FUZHEN ZHANG

THEOREM 8.2. For any yzuztwnion matrix A, 1Al, = 1Ald.

Noticing that A*A is a positive definite Hermitian matrix for every

A E M,,(Q) and using Theorem 8.1, part 7, one obtains

THEOREM 8.3 (&en, 1991; Xie, 1979). Let A be an n X n quaternion

matrix. Then
-
I Al,, < fi i n,j "ij >

j=li=l

and ecpality holds if ancl only if the colunm vectors of A are mutually
orthogonal.

9. NUMERICAL RANGES OF MATRICES WITH

QUATERNION ENTRIES

Numerical ranges of complex matrices have been a very popular topic in

linear algebra (see, e.g., [22, 231). Th e p u rp ose of this section is to study the
numerical range of a quaternion matrix. Particular attention is paid to the
convexity of the quaternionic numerical range of a matrix lying in the closed
upper half complex plane.
Let H be a Hilbert space o\rer a=, and let A be an operator on H. The
classical complex numerical range of the operator A, or the field of values of
A [21], comprises all complex numbers (A( .2-j, x) as x runs over the unit
sphere {x ( x E H, llxll = 1) in H. If A is an n X n complex matrix, the
complex nt~merical r+qe or the field of values [23] of A is the subset
(x-*Ar I x E C”, x* I = l} of the complex plane.
A celebrated result known as Toplitz-Hausdorff theorem ensures the
conversity of the complex numerical range of an operator (a complex matrix).
The quatemionic numerical range of an operator A on a (left) quater-
nionic Hilbert space over Q [37] or a square matrix with quatemion entries
[23, p. 861 is similarly defined, and denoted by W(,( A). To avoid ambiguity,
we designate by W(.( A) the classical numerical range, i.e., the complex
numerical range of a square matrix A with complex entries.
The study of the convexity of WC,(A) as a subset of 6ZJ was begun by
Kippenhahn [37]. S’mce W(,( A) is a subset of the real four-dimensional space
Q, it is not easy to visualize or calculate. The first natural question is what the
part of W,,(A) lying in [w or c looks like; that is, what is W,(A) n [w or
WC(A) f~ @? This problem has been investigated in [27, 25, 3-5, 36, 35, 47,
48i. We briefly state the existing results, then reduce the convexity of W,( A)
QUATERNIONS AND MATRICES OF QUATERNIONS 51

to that of its intersection with the complex plane, and further reduce the
general matrix case to 2-by-2 matrices. For normal quatemion matrices, we
will give explicit characterizations of Wg< A) n [wand W,(A) n @.
Let B(A) = WY<A) n @. We call B(A), as Kippenhahn did, the bild (or
representative) of A, and the subset of B(A) lying in the closed upper half
plane the upper bild of A, denoted by B+(A). Note that a complex number
c E B(A) if and only if C E B(A), by Theorem 2.1, part 5. Thus I?( A) =
B + ( A) U B _ ( A), where B_(A) is the reflection of I? + ( A) about the x-axis
in the complex plane. As we will see shortly, B(A) is not convex in general,
while W,( ,yA), symmetrically suited about the x-axis since xA and L are
similar, is convex, Note that if x E W7( A), then any quatemion similar to x
is contained in W,(A). In particular, x* E W,(A).
A comparison of B(A) and W,( xA) is given below.

THEOREM 9.1 (Kippenhahn, 1951). Let A be a quaternion matrix. Then:

1. B(A) c W,( xA).

2. The upper envelope of B(A) coincides with the upper envelope of
\v~( xA). More precisely, suppose h, + k,i, h, + k,i E B+(A) with the
properties that h, < h,, k,, k, > 0, and h, < Re z < h, for all z E B+(A).
Then the portions of B(A) and WC<xA) above the line segment joining
h, + k,i and h, + k,i coincide.

To prove part 1, let c = u*Au E B( A). Then x, = ,Q xA x,, . Write

u = 11~ + uzj. It is easy to check that c = 4*x.$ E W,( x,.,1, where tZ =
(uI> - G)’ and 2 is a unit vector when u is a unit vector. For (2), let
c = x*x,x E W&xA), where x = (xi, x2,. . . , xZnjT E @‘“. Set u =
(x, - FJ, xg - x‘j )..., XZtl_i - X,Jj>‘. The conclusion then follows from
a careful analysis [36] of the quatemion q = u*Au.
With this theorem one can see that the numerical radius [25] and the
spectral norm [lo] of a quatemion matrix A, defined in the usual sense, are
equal to those of the complex adjoint x., of A, respectively.
Kippenhahn asserted in [27] that the B(A) coincided with W,( xA), and
consequently I?( A) was convex. But his proof of this part is false [25], as the
following example due to Au-Yeung [5] illustrates.

EXAMPLE 8.1. Take

A=
52 FUZHEN ZHANG

Then i, -i E B(A) c Wq< A), but 0 = ii + i( -i) e B(A). Thus B(A),

and therefore W4< A), is not convex. Notice that A is a complex matrix and its
complex numerical range is convex. Some properties of quatemionic numeri-
cal ranges are not parallel to those of complex numerical ranges.

The connectedness of some subsets has been observed.

THEOREM 9.2 (Jamison, 1972). For any operator A on a quaternionic

Hilbeti space, W,(A) n R is either empty or a closed interval of II&

The proof by Jamison is a computational one. Considering an operator on

a left Hilbert space H over Q, Au-Yeung [S] showed that the set {u E
H I( Au, u) = 0, [lull = 1) is connected when A is skew-Hermitian (i.e.,
A* = -A) and the set {u E H [(Au, u) = a, llull = l} is connected when a
is real and A is Hermitian. Based on these observations, he gave another
proof of Jamison’s result and showed the following theorem.
For convenience, we denote the projections of W,(A) on [w and @ by

Wy( A: R) = {Re q I q E Wy( A)}

and

Wy( A: a=) = {Co q I q E W<,( A)}

respectively.

THEOREM 9.3 (Au-Yeung, 1984). Let A be an operator on a quaternionic

Hilbert space. Then

1. Wy( A) is convex if and only if WC,<A) f~ Iw = WY<A : R);

2. WY<A) is convex if and only zf WC,<A) n @ = W,< A : C);
3. conviW,( A)) = {a + pla E [w, Re p = 0 and a + ] pli = Co q jk
some q E W,(A)}, where conv means “convex hull of “;
4. ifa+p,,a+p,~W,,(A)wherea~[W,Rep,=Rep,=O, then
~,uny~$‘Re p = 0, one has a + p E W(,( A) whenever 1pl lies between
Pl an 2’

The last part of the previous theorem ensures that the intersection of
B+(A) with any vertical line is an either empty set or a line segment. The
same is true for horizontal lines [4].
QUATERNIONS AND MATRICES OF QUATERNIONS 53

The projection of W,< A) on @ is always convex [3], since

as can be seen by verifying that

Co(u*Au) = (uy,u;)
(; q(::)
where A = A, + A, j and u = U, + u2 j, as before.
Note that if 9 = 9” + 9,i + q2j + q,k E W,(A), then

90 +dmi~W~(A)n@

and

Thus 9() E W,< A) fl R’ if W,(A) I? @ 1s convex. It turns out, by Theorem

9.3, part 1, that W,< A) is convex if and only if B(A) = W<,(A) n @ is
convex.
It is well known that the complex numerical range of a complex normal
matrix is the convex hull of its eigenvalues. The quatemionic analogue has
been studied recently [3, 4, 36, 47, 481.
As seen in Example 8.1,the quatemionic numerical range of a normal
matrix is not a convex hull of eigenvalues, and not even convex in general. Yet
a somewhat hidden convexity that can be described in terms of eigenvalues
has been observed by So, Thompson, and Zhang [36, 47, 481.

THEOREM 9.4(So,Thompson, and Zhang, 1994). Let A be a normal

matrix of quaternions with standard eigenavlues h, + k,i, . . . , h, + k,i. Then
B+(A) is the convex hull of these eigenvalues of A, (1, 01, and (r, O), where

and
54 FUZHEN ZHANC

under the convention that (h,7k, + h,k,)/(k,5 + k,) means a pair {h,, h,]
when k, = k, = 0. Consequently B+(A) is convex.

The main idea of the proof in [36] or [47] is to convert the upper-bild
investigation into an extremal problem of eigenvalues with side conditions,
then employ Langrange multipliers. Au-Yeung [3] later shortened the proof
by showing that the upper bild coincides with the union of certain consecu-
tive triangles.
lust as for the case of complex matrices [23, p. 181, the convexity of the
upper bild of a quaternion matrix is easily reduced to that of 2 X 2 cases.

THEOREM 9.5. If the upper hild of any 2 X 2 quaternion matrix is

convex, then the upper hild (Ifay n x n quaternion matrix A is convex.

By splitting a general 2 X 2 matrix into Hermitian and skew-Hermitian

parts and using a pair of suitable unitary matrices on A [35], one may further
reduce the convexity for a 2 X 2 matrix to that for the matrix

k,i
G=
-a+pj

where CY, p, k,, and k, are nonnegative numbers.

Unlike the normal case in which an upper bild is generally a polygon, the
boundary of the upper bild of G is a curve, not necessarily formed by line
segments. It is readily seen from Kippenhahn’s theorem that the upper
bounding curve (envelope) of B+(G) IS concave downward. To prove the
convexity, it must be shown that the lower bounding curve is concave
upward. This is extremely difficult. With the help of Mathematics computa-
tions and in a great effort over years, So and Thompson have accomplished
the proof in a very recent manuscript [35] (65 pages long).

TIIEOHEM 9.6 (So and Thompson, 1995). Let A be an n X n quaternion

matrix. Then B + ( A) is a convex subset of the complex plane @.

Many questions regarding a quaternionic numerical range may be asked

even though the main problem for the finite-dimensional case has been
settled.

QUESTION 9.1. Is there a short and conceptual proof for Theorem 9.6?
QUATERNIONS AND MATRICES OF QUATERNIONS 55

QUESTION 9.2. How are B(A) and W,( xA) related? What is the area
ratio of these two sets?

QUESTION 9.3. Is there a simple proof (without using Theorem 9.6) for
the statement that the quatemionic numerical range of the diagonal matrix
formed by the diagonal entries (or eigenvalues) of A is contained in WC,<A)?

QUESTION 9.4. Investigate W,(A) and B+(A) when A is a linear

operator on a left quatemionic Hilbert space of infinite dimension.

The author wishes to express his deep thanks to Professor R. C. Thompson

and Professor W. So for their encouragement and helpful discussions, and
Professor L. C. Chen for sending him some preprints. Thanks also go to the
referee for many suggestions.

REFERENCES

1 S. L. Adler, Quaternionic Quantum Mechanics and Quantum Fields, Oxford

U.P., New York, 1994.
2 T. M. Apostol, Mathematical Analysis, 2nd ed., Addison-Wesley, 1977.
3 Y. H. Au-Yeung, A short proof of a theorem on the range of a normal quater-
nionic matrix, Linear and Multilinear Algebra, to appear.
4 Y. H. Au-Yeung, 1994. On the eigenvalues and numerical range of a quatemionic
matrix, Preprint.
5 Y. H. Au-Yeung, On the convexity of numerical range in quaternionic Hilbert
spaces, Linear and M&linear Algebra 1693-100 (1984).
6 W. C. Brown, Matrices over Commutatiae Rings. Marcel1 Dekker, 1992.
7 G. Birkhoff and S. MacLane, A Survey of Modern Algebra, 4th ed. Macmillan,
1977.
8 J. L. Brenner, Matrices of quaternions, Pa+ 1. Math. 1:329-335 (1951).
9 A. Bunse-Gerstner, R. Byers, and V. Mehrmann, A quaternion QR algorithm,
Numer. Math. 55:83-95 (1989).
10 C. G. Cao, The numerical radius of real quaternion matrix. J. Xinjiang Univ. 7,
No. 2 (1990); also see MR 89k:15031.
11 L. X. Chen, Definition of determinant and Cramer solutions over the quatemion
field, Acta Math. Sinica N. S. 7(2):171-180 (1991).
12 L. X. Chen, The extension of Cayley-Hamilton theorem over the quatemion
field, Chinese Sci. Bull. 17:1291-1293 (1991).
13 L. X. Chen, Inverse matrix and properties of double determinant over quaternion
field, Sci. China Ser. A 34(5):528-540 (1991).
14 P. M. Cohn, Skew Field Constructions, London Math. Sot. Lecture Note Ser. 27,
Cambridge U.P., 1977.
15 P. M. Cohn, The similarity reduction of matrices over a skew field, Math. 2.
132:151-163 (1973).
56 FUZHEN ZHANG

16 S. Eilenberg and I. Niven, The “fundamental theorem of algebra” for quater-

nions, Bull. Amer. Math. Sot. 50:246-248 (1944).
17 S. Eilenberg and N. Steenrod, Foundation.s of Algebraic Topology, Princeton
U.P., 1952.
18 D. Finkelstein, J. M. Jauch, S. Schiminovich, and D. Speiser, Foundations of
quatemion quantum mechanics, J. Math. Phys., 1962, pp. 3207-3220.
19 D. Finkelstein, J. M. Jauch, and D. Speiser, Notes on quatemion quantum
mechanics, CEBiV 59(9):59- 17 (1959).
20 K. Guerlebeck and W. Sproessig, Quaternionic Analysis, Akademie Verlag,
Berlin, 1989.
21 P. R. Halmos, A Hilbert Space Problem Book, Van Nostrand, 1967.
22 R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge U.P., 1985.
23 R. A. Horn and C. R. Johnson, Topics in Matrix Analysis, Cambridge U.P., 1994.
24 N. Jacobson, Basic Algebra I, W. H. Freeman, 1974.
25 J. E. Jamison, Numerical Range and Numerical Radius in Quatemionic Hilbert
Spaces, Doctoral Dissertation, Univ. of Missouri, 1972.
26 R. E. Johnson, On the equation xcr = 7~ + /3 over an algebraic division ring,
Bull. Amer. Math. Sot. 50:202-207 (1944).
27 R. Von Kippenhahn, Uber der Wertevorrat einer Matrix, Math. Nachr.
6(3/4):193-228 (1951).
28 A. Krieg, Modular Forms on Half-Spaces of Quaternions, Lecture Notes in Math.
1143, Springer-Verlag, 1985.
29 H. C. Lee, Eigenvalues of canonical forms of matrices with quatemion coeffi-
cients, Proc. R.I. A. 52, Sect. A (1949).
30 M. L. Mehta, Determinants of quaternion matrices, I. Math. Phys. Sci. 8, No. 6
(1974).
31 N. Mackey, Hamilton and Jacobi meet again: Quaternions and the eigenvalue
problem, SlAM J. Matrix Anal. Appl., to appear.
32 I. Niven, Equations in quaternions, Amer. Math. Monthly 48:654-661 (1995).
33 N. Steenrod, The Topology of Fibre BundZe.s, Princeton, U.P., 1951.
34 W. So, Left eigenvalues of quaternionic matrices, a Private communication, 1994.
35 W. So and R. C. Thompson, Convexity of the upper complex plane part of the
numerical range of a quaternionic matrix, Manuscript, 1995.
36 W. So, R. C. Thompson, and F. Zhang, Numerical ranges of matrices with
quatemion entries, Linear and Multilinear Algebra 37:175-195 (1994).
37 K. Viswanath, Normal operators on quaternionic Hilbert spaces, Trans. Amer.
Math. Sot. 162 (1971).
38 K. Viswanath, Contributions to Linear Quaternionic Hilbert Analysis, Thesis,
Indian Statistical Inst., Calcutta, 1968.
39 L. A. Wolf, Similarity of matrices in which the elements are real quaternions,
Bull. Amer. Math. Sot. 42~737-743 (1936).
40 N. A. Wiegmann, Some theorems on matrices with real quaternion elements,
Canad. J. Math. 7:191-201 (19551.
41 R. M. W. Wood, Quatemionic eigenvalues, Bull. London Math. Sot. 17:137-138
(1985).
QUATERNIONS AND MATRICES OF QUATERNIONS 57

42 B. J. Xie, An extension of Hadamard’s theorem over the skew field of quater-

nions, Sci. Sinica, Special Issue I on Math., pp. 88-93 (1970); also see MR
83g: 15014.
43 B. J. Xie, Applications of characteristic roots and standard forms of matrices over
a skew field, Acta Math. Sinica 23(4):522-533 (1980); also see MR 82m:15021.
44 B. J. Xie, Existence theorems for the canonical diagonal form and the weak
canonical form of characteristic matrices over a skew field, Acta Math. Sinica
23(3):398-410 (1980); also see MR 82k:15015.
45 B. J. Xie, An expansion theorem for determinants of self-adjoint quatemion
matrices and its applications, Acta Math. Sinica 23(5):668-683 (1980); also see
M R 83a: 15007.
46 B. J. Xie, Several theorems concerning self-conjugate matrices and determinants
of quatemions, Acta Sci. Natur. Univ. J&n, No. 1, pp. l-7 (1982); also see MR
84~115022.
47 F. Zhang, Permanent Inequalities and Quaternion Matrices, Ph.D. Dissertation,
Univ. of California at Santa Barbara, 1993.
48 F. Zhang, On Numerical Range of Normal Matrices of Quaternions, /. Math.
Physical Sciences, 29(6):235-251 (199.5).

Received 29 August 1994; f ma 1 mmmript accepted 13 June 199.5