Exposrnonas Expo. Math.

4 (1986), 193—215
MW © Bibliographisches Institut 1986

A general theory of Vandermonde matrices.

T.Y. Lam 1

§1 Introduction
In history books of mathematics, Alexandre Théophile Vandermonde (1735—1796)
is usually credited for his study of resolvents and the group of permutations of the
roots of an equation, and for his role in isolating the theory of determinants as a
subject for independent study. Nowadays he is perhaps best remembered through
the beautiful matrix
1 l 1 (I)
“{1 a; an
V,,(a1,...,an)=
ail—1 a; 1... a:—1

which is known to all students of linear algebra as the Vandermonde matrix.2

Here, (11,...,a,l are, say, elements from a field F. A simple inductive argument
yields the following classic formula for the Vandermonde determinant:
det V,,(a1, ...,an)=l—I(ai—-aj). (2)
i>j

Therefore, the Vandermonde matrix (1) is nonsingular (=invertib1e 3) over F iff

the elements a,, ...,a,l are distinct. If a,, ...,a,l are not necessarily distinct, then the
rank of Vn(a1, ...,a,,) is given by the number r of distinct elements among
{a1,...,a,,}. This follows by applying the determinant formula to an r x r Vander-
monde submatrix of V,,(a1, ...,a,.).
Needless to say, the Vandermonde matrix enjoys its great fame for a good reason.
It is perhaps not an exaggeration to say that every working mathematician, pure
or applied, would encounter the Vandermonde matrix at least once or twice some
time in his/her career. The Vandermonde matrix is used, for instance, in dealing

1) Supported by NSF.
2) Actually, it is not altogether clear that the matrix (1) had occurred explicitly in
Vandermonde’s work. According to Lebesgue, “Vandermonde matrix” might have been a
misnomer because someone had misread subscripts for exponents (cf. [Edwards: 1984,
p. 18]). However, V3(a, b, c) did appear in Vandermonde’s work and the generalization from
3 x 3 to n x n seems to be a natural step.
‘ 3) In this paper, we use the terms “nonsingular” and “invertible” interchangeably.

0723-0869/86/030193-23 $6.10
 17-7 A . l . hill“

with finite Fourier series, the polynomial interpolation problem, in solving differ-
'ential and difference equations, in curve fitting, in coding theory and in the study
of recursively defined sequences. For a survey of some of these applications, see
[Kalmanz 1984]. Of course, the Vandermonde matrix is also important for many
theoretical considerations in abstract algebra, for instance, in ring theory, algebraic
number theory, group representation theory, invariant theory and the theory of
symmetric functions [MacDonald: 1979].
Like so many other good notions in mathematics, the Vandermonde matrix is
capable of interesting generalizations. For instance, one may form the more
general n x n matrix whose i"l row is (a;‘, ...,a;‘) where 0=r1 <r2 < <r,, are given
integers. The resulting determinant, called a “higher alternant” ([Muir: 1933,
Ch. 11]), is important for combinatorics [Stanley: 1971] and for the study of the
irreducible characters of the symmetric group [Mumagham 1938], [Littlewoodz
1950]. Another way to generalize V,,(a,,...,a,,) is to form certain “differentiated”
Vandermonde matrices [Aitken'r 1959, p.119], [Norton: 1984]; their deter-
minants, called “confluent alternants”, are useful for solving “osculatory” in-
terpolation problems [Kalmanz 1984] and for studying shift register sequences
[Grundofer-Liineburg: 1980]. Yet other generalizations appeared in [Hirsch:
1949], [Carlitzz 1955] and [Famelk 1959].
In this paper, we shall give an exposition on a new generalization of the Vander-
monde matrix. Our generalization differs from the aforementioned ones in that it
contains the following two new ingredients: (1) we allow the elements a1, ...,a,l in
the Vandermonde matrix to be elements from a division ring D, so that they may
no longer commute with one another, and (2) our new Vandermonde matrix takes
into account an arbitrarily given endomorphism a of the division ring D. More
precisely, for given elements a1, ...,a,,eD, and a field endomorphism aHa‘ of D,
we define the a-Vandermonde matrix to be
1 1 1
N192.) naz) NM
V,,"(a1,...,a,,)= * (3)
Ml-l(al) N_1(a2) “Ml-1(an)

where, for any aeD, we define N,(a) to be a“"...a‘2a"a. In the special case when
a is the identity automorphism of D (we write a=1 for short), we have N,(a)=a",
so in this case the a-Vandermonde matrix boils down to the ordinary Vander-
monde matrix V,,(a1 , . . . , a").
Note that our a—Vandermonde matrix represents a nontrivial generalization of the
ordinary Vandermonde matrix even in the case when D is a field. We shall
mention two examples among many interesting ones.
(1) If D is a field of characteristic p>0, we can take a to be the endomorphism of
D defined by a"=a’ (aeD). A 4 x 4 a-Vandermonde matrix, say, looks like
 A general theory of Vandermonde matrices 195

l 1 1 l
a b c d
ap+1 bp+l cp+1 dp+l ' (4)

ap1+p+1 bp2+p+1 cp1+p+1 dp1+p+1

Its determinant is one of the higher altemants.

(2) If D is the field (E of complex numbers and a is its automorphism given by
complex conjugation, a typical 4 x 4 a-Vandermonde matrix looks like
1 l l 1
a b c d
lal2 lbl2 ICI2 ldl2
(5)
alal2 blbl2 CICI2 dldl2
Here, |z = 22 1’2, where 2 denotes the comPlex conjugate of 2643.
It turns out that, to treat determinants of type (4), (5), even though it appears to
be a “commutative” problem, the best way of attack is to use certain noncom-
mutative techniques. In fact, it is this somewhat surprising use of techniques from
noncommutative algebra that is the raison d’étre for this article. As long as
noncommutative techniques are to be employed, we may as well go one step
further, and work more generally with division rings rather than just with fields.
Our methods yield uniform results which are valid over any division ring. On the
other hand, most of our results seem to be new already over fields.
The basic problem concerning the a-Vandermonde matrix V:(a1, ...,a,,) is the
following: When is this matrix nonsingular, and, if it happens to be singular, in what
way does its rank depend an al, ...,a,,? It turns out that.both questions have fairly
natural answers, as we shall see in this exposition.
In this introductory section, we shall first try to whet the reader’s appetite by
giving the answers to the above questions in a couple of special cases. The
solutions for these special cases will serve two useful purposes. First, through
working with them, we shall become better acquainted with, our generalized
Vandermonde matrix, and, secondly, the form of the solutions obtained in the
special cases will give us a good illustration of, and useful empirical evidence for,
the general results to come.
In both of the special cases we treat below, we shall work with 3x3 Vander-
monde matrices, the 2 x 2 case being completely trivial. For the first special case,
we let D be the complex field C, and a be “bar” as before. Consider the 3 x 3 a-
Vandermonde matrix
‘ 1 l l
V3"(a, b, c)= a b c (6)
Id!2 Ibl2 ICI2
where a, b,c are assumed distinct. Obeyling our first impulse, we write down its
determinant :
det Va'(a,b, C)= -a(|6|2 -|b|2)+b(|6|2 -|a|2) -C(lb|2 —|a|2)- (7)
First, let us assume that lal, lbl, |c| are distinct, say |a|<|b|<|c| (this may be
achieved by -a permutation of the columns). Applying twice the inequality
Ix-y|>IXI-lyl we get
IdetV$’(a,b,'0)l2|b(ICI2—Ia|2)l-|a(16|2—|b|2)|—IC(|b|2-la|2)l
= —|al -(|6|2 -|b|2)+|b|(|c|2 -|al2)-|6|(lblz —|a|’)
=detV(lal,|bl,Icl) ' ‘8)
=(Icl-Ibl)(l6|—|a|)(|b|-|a|)>0.‘)
Thus, V;(a, b, c) is invertible in case |a|, lbl, lcl are distinct. On the other hand, if
|a|=|b| =|=|c|, then, going back to (7), we have
det V,°(a, b, c)=(a —b)(|b|2 —|cI’)=I=0,
so V;(a, b, c) is again invertible. Finally, if |a|=|b|=|c|, then det V;(a, b, c)=0 so
this is the only case when Vg’(a, b, c) fails to be invertible.
At this point, even with only scanty evidence, one might begin to make some
conjectures. It would appear that, for D=C and a=“bar”, V:(a1,...,a,,) is in-
vertible if |a1|,...,|a,,| are distinct (real) numbers; or, even better, .V:(a1,...,a,,) is
invertible if no three of the ai’s have the same modulus. As we shall see, these
conjectures do turn out to be true. In fact, we shall be able to prove that the rank
of V‘(a1,.,a,,) is determined 1n a straightforward way by the distribution of the
distinct moduli among |a1|,.. .,Ia |. -
For the second motivating example, we shall treat the special case V3(a,b, 0) over
a division ring D, taking 0 here to be the identity. We assume that a, b, c are
distinct. The following left multiplication by an elementary matrix corresponds to
two successive row operations on V3(a,b, c):
' 1 0 o 1 1 1
(—a l ’ 0) V3(a,b,c)=(0 b—a c—a ). (9)
_0 —a 1 0 (b—a)b (c—a)c
Note that, in carrying out elementary row (resp. column), operations on matrices
over a division ring, we can only multiply rows (resp. columns) by scalars from
the left (resp. from the right). In view of the equation (9), we are reduced to
b—
—a c— a
lookin at the 2x2 matrix M =(
g , (b —a)b (c—a)c ). Here, though, we are faced

4) Based on this inequality, it seems tempting to conjecture that, for any n,

Idet V"(al, .. .,a")IzldetVndaIL...., Ia,|)|. I do not have a proof for n24. (See Note added
in Proof at the end of the paper)
 A general theory of Vandermonde matrices 197

with a dilemma. In the matrix M, it would be nice to “extract” the factors b—a
and c— a. If these were right factors instead of left factors, we would be able to
scale the two columns simultaneously by right multiplication of M by
((b (:1)_ (c (2)4). Since we are stuck with b-a and c-a as left factors, the
— _ —1
next best thing is to left multiply M by ((6 (:1) (c _:)_1) to get
((c—a)'1(b —a) I)
. Another row operation on the latter leads to
(c—a)‘1(b —a)b c
( (c—a)‘1(b—a) . l)
(c—a)‘1(b—a)b—c(c—a)‘l(b—a) 0 '
—a)“(b—a—-)b
Thus, for distinct a, b, ceD, V3(,a b, c) is singular if and only if (c—
=c(c—a)'1(b—a), or equivalently
(b—a)b(b -a)‘ =(c-a)c(c-—a)'l. , (10a)
At this point, to give the reader the proper perspective, we should point out that
the computation above is in fact what is needed to determine the so-called
Dieudonné determinant of V3(a, b, c) over the division ring D (see [Artinz 1957],
p. 151). For an invertible matrix V over a division ring D, the Dieudonné de-
terminant, det V, of V takes value in the commutator quotient group D*/[D*,D*],
where [D*,D*] is the commutator subgroup of the multiplicative group D*
=D\{0}. (For singular M, it is customary to take detV=0-[D*,D*].) Since
elementary matrices must have determinant 1, and det(diag(u1, ...,u,,))
=u1...u,,[D*,D*], the computation above shows that det V5(a, b, c) is given by the '
coset of
—(c —a)2[(c —a)' 1(b —a)b —c(c —a)‘ 1(b —a)]
=(c —a) [(c -a)c(c —a)‘ 1(b —a) —(b —a)b].
To those readers who may be appalled by this seemingly ugly expression, it is
perhaps comforting to note that, in case a commutes with both b and c, this
simplifies to the usual Vandermonde determinant (c —a)(b —a)(c —b)!
The above computation leads us to make several interesting observations. First,
by symmetry, the necessary and sufficient condition in (10a) for the singularity of
V3(a, b, c) has two other parallel forms:
(a—b)a(a—b)‘1 =(c—b)c(c—-b)_1, (10b)
(a—c)a(q—c)'1=(b—c)b(b—c)‘1, g (10c)
and therefore, (10a), (10b), (10c) are in fact equivalent, which does not seem’to be
obvious at first sight. Secondly, a necessary condition for (10a), (10b), (10c) is, that
a, b, c be pairwise conjugate. In particular, we see that, if a, b, c are not in a single
conjugacy class of D, then the Vandermonde matrix is necessarily nonsingular. If
a, b, c are indeed in one conjugacy class, then V3(a, b, c) may be singular. For
.-- 1.1.uulll

instance, over the division algebra of real quatemions, a=i, b=j, c=k are con-
1 1 1
jugate to one another; here, V3(i,j, k)= i j k is clearly singular, so we
—1 —1 —1
can predict that (10a) holds. There also exist examples (though not over the real
quaternions) where a, b, c are in one conjugacy class, but (10a) does not hold. For
such triples, V3(a, b, c) will be nonsingular. (The existence of these examples will be
clear later on.)

If we recall the conjecture which we made for the “bar”-Vandermonde matrix

over 03, we might be tempted to make a corresponding conjecture for the Vander-
monde matrix V,,(al, ...,a,,) over a division ring D: If no three of a1, ...,an lie in
one conjugacy class of D, then Vn(a1,...,an) is nonsingular. As 'we shall see later,
this conjecture is indeed true!

At this point, we could not have failed to observe a certain similarity between the
two motivating examples treated above. In the first example, we have to dis-
tinguish complex numbers by their moduli, and in the second example, we have
to distinguish elements in a division ring by the conjugacy classes they determine.
It is not difficult to see how these two cases can be unified. For any division ring
D equipped with an endomorphism a, let us say that two elements a,b are a-
conjugate if b=d"ad‘l for some deD“. (We also say that b is a o-conjugate of a.)
As is easily verified, being a-conjugate is an equivalence relation on D, so D is
partitioned into its a-conjugacy classes. If 0:], we get back the usual conjugacy
classe_s. If D=C and a=“bar”, then the a-conjugacy equation amounts to b
=a-d/d for some (1603*. A simple argument shows that this means exactly that a
and b have the same modulus.

§2 The Method of Skew Polynomials

Clearly, the method we used to treat the two special examples in §1 relies
basically on determinant theory. However, in both examples, we have definitely
exploited the fact that the matrices are only 3 x 3. It would be nice to extend this
method to treat the general n x n case, but I have had no luck with it so far. In
any case, such an extension will not lead to a unified treatment since the two
proofs are already quite different in the 3 x3 case. To find a way to solve the
problem, we must now abandon determinant theory and try to find a new tool.
The right tool to use turns out to be the theory of skew polynomials in one
variable over a division ring. With this brief warning, we now enter the arena of
noncommutative algebra.

Let a be a fixed endomorphism of the division ring D. For an indeterminate I

over D, we write D[t,a] for the ring of skew left polynomials over D. By this, we
 A general theory of Vandermonde matrices 199

mean that D[t,a] is the set of all left polynomials 5) Zeit‘ which are added in the
usual way, and multiplied by using the distributive law together with the rule that
t- c =c" - t for any ceD. Thus, the coefficients need not commute with the variable t.
The fact that (c1c2)‘=c‘{ c; guarantees the associative law for polynomial multipli-
cations, so D[t,a] is a ring. This so-called skew polynomial ring is a basic object
of study in noncommutative ring theory [Jacobson: 1943, Ch. 3]; the idea of the
twist equation tc=c"t goes as far back as [Hilbertz 1899]. Note that any right
polynomial Zt‘bi gives an element in D[t,a], but not every element in D[t,a] can
be expressed as a right polynomial, unless a is an automorphism of D.
As is easily seen, the usual division algorithm prevails in R=D[t,a]: For f (t)eR
and g(t)eR\{0}, we can write uniquely f (t)=q(t)g(t)+r(t) where r(t)=0 or
deg r(t)<degg(t). In particular, R is a left PID (principal ideal domain), i.e. any
(nonzero) left ideal 19R has the form R~g; here, g is any polynomial in I of the
smallest degree. It can further be shown that, in a suitable sense, R has the
Unique Factorization property, so it is a (not necessarily commutative) UFD (see
[Cohnz 1973]). However, the notion of unique factorization in the noncom-
mutative setting is a rather sophisticated one. In keeping with the elementary
nature of this article, we shall therefore try to avoid using the unique factorization
property in D[t,a]. (The expert reader will notice that in one or two places, our
proofs could be somewhat streamlined by invoking such a property.)
We shall now introduce the notion of evaluating a skew polynomial f (t)eD[t, a]
at a point deD. We write f as a left polynomial, say Zcit‘, and then define
f(d)=ZciNi(d), (11)
where N,(d)=d""'...d‘d for i 2 1, and No(d)El. Note that, in case 0= 1, then N,(d).
=di and so f (d)=Zc,d‘ is the usual evaluation off at d. In general, we must take
a into account in the evaluation process; this is why we introduced the functions
Ni. When we evaluate at d, we have to exercise caution. For instance, if f happens
to be given in the form of a right polynomial f =Ztibi, we must first express f as
a left polynomial, viz. f =b't‘, before “plugging in” (1. Thus, f (d) is given by
Zbg’iNM), not ZN,(d)b,). If f (d)=0, we say that d is a (right) root (or zero) of f.
Unless otherwise specified, the word “root” shall mean right root in this paper. If
A is a subset of D, we say that f vanishes on A if f(A)=0, i.e. if f(d)=0 for every
deA. Note that a nonzero f may vanish on an infinite set in the noncommutative
setting. For instance, in the division algebra of real quaternions, for 0:1, the
quadratic polynomial t2+l vanishes on all conjugates of i, and it is easy to see
that i has infinitely many conjugates. (This is a special case of a theorem of
[Hersteinz 1956]; for a quick proof, see MAA Monthly, Vol. 90 (1983), p. 292.)
Roots of polynomial equations over division rings have been studied by various
authors. Some of the key papers on this topic are [Richardson: 1927], [Nivenz
5) The terminology refers to the fact that the coefficients are written on the left.of the
variable.
200 T.Y. Lam

1941, 1942], [Brand: 1942], [Eilenberg-Niven: 1944], [Pollakz 1960], [Gordon-

Motzkin: 1965], [Fein-Schacher: 1971], [R6hrlz 1978], [Beck: 1979], [Cerlienco-
Mureddu: 1981] and [Bray-Whaples: 1983]. For expositions, see [Smith: 1974],
[Cohnz 1977, Ch. 8] and [Dauns: 1982, Appendix A4].
The basic relation between skew polynomials and a-Vandermonde matrices is
provided by the equation
(CO’cl’"'9cn—1)Vn‘(dl!""dn)=(f(d1)a-'-:f(dn—-1»s (12)
n— 1
where f (t)- ‘20 c t‘eD[t, 0]. Therefore, to study the Vandermonde matrix, a useful

tool is the theory of roots of skew left polynomials. We shall now establish the
fundamental facts about such roots for later use. We start with two easy formulas
for the functions M.
Lemma 1. For deD, aeD“, and integers i,j 20, we have
(1) M+j(d)=1vj(d)dilvi(d);
(2) M(a°da‘1)=a”i1\l}(d)a‘i.
Proof. The right side of (1) equals
(d‘j"...d"d)‘i(d“"...d"d)=d"”’.‘ ..=d"d N,+J(d).
The left side of (2) equals
(aada‘ 1)"""(a"da‘ l)""“...(a"da- l)”(a"da" 1)=a*"‘(d“" therma-1
=a”i1\l,-(d)a‘ 1.
This completes the proof.
Ingeneral, evaluating skew polynomials at a point deD may not be a ring
homomorphism from D[t,a] to D. In other words, if f (t)=g(t)h(t) in D[t,a], f (d)
may not be equal to g(d)h(d), due to the lack of commutativity. However, we have
the following remarkable result which, to a large extent, recovers the homomor-
phism property.

Theorem 2. Let f (t)= g(t)h(t)eD[t, a] and deD. If h(d)=0 then f(d)=0. If a==h(d)
=|=O, then f (d): g(a"da'1)h(d). In particular, if d is a root off but not a root of h,
then a a-conjugate of d is a root of g.

Proof. Let g(t)= 2b. z' h(t)= 2“,- ti sofa)... Zbic,”it1+: and

f(d)= ZJNTN.+101)
=zbici‘Nj(d)"‘Ni(d) (by Lemma 1(1))
i-i .

=Zbia°‘Ni(d)
 A general theory of Vandermonde matrices 201

Where a=h(d)- If a=0, clearly f(d)=0. If a=l=0, then by (2) of Lemma 1,

f(d)=2biN.-(a‘da‘ ‘)'a=g(a°da- ‘)h(d). Q.E.D.
Using the first part of the theorem above, we obtain an analogue of the Re-
mainder Theorem in a noncommutative setting.
Theorem 3 (Remainder Theorem). Let f(t)eD[t, a] and deD. Then f (d)=0 ifl' f (t)
= g(t)(t —d) for some polynomial g(t)eD[t, a].
Proof. Since t—d vanishes on d, the “if” part is clear from the first statement in
Theorem 2. For the “only if” part, assume f(d)=0. By the (left) division algo-
rithm, we can write f (t)=g(t)(t —d)+r for some reD. Evaluating at d shows that r
= f(d)=0. Q.E.D.
The familiar result that a polynomial of degree n over a field has at most n roots
can now be generalized. The following theorem has appeared in [Smith: 1974]
and [Daunsz 1982]. In the special case when 0: 1, it is due to [Gordon-Motzkin:
1965].
Theorem 4. Let f (t)eD[t, a] be of degree n. Then the (right) roots of f (t) lie in at
most n a-conjugacy classes of D. If f (t) has the form (t—al)...(t—an) where
a1, ...,aneD, then any root off is a-conjugate to one of the ai’s.
Proof. Both statements follow quickly from Theorem 2 and Theorem 3, by in-
duction on n.
Let us say that a subset A ED is a—algebraic if there exists a nonzero polynomial
f (t)eD[t, a] which is zero on A. The set 1(A) of polynomials vanishing on A forms
a left ideal in D[t,a], by Theorem 3. If f is a monic polynomial of the smallest
degree in I(A), then f is uniquely determined and I(A)=D[t, a] -f. We shall write
fA for f and call it the minimal polynomial of A. The degree of fA will be called the
rank of A.
Lemma 5. Let A be a o-algebraic set as above, and n=rankA=den. Then fA(t)
=(t —a,,)...(t —-a 1) where each ai is a-conjugate to some element of A.
Proof. For any aeA, fA(t) has a right factor t—a. Choose m to be the largest
integer such that f4(t) has a right factor of the form h(t)=(t—am)...(t—a1) where
each ai is a-conjugate to some element of A. We finish by showing that m=n.
Indeed, if m<n, then h(t) cannot vanish identically on A. Pick deA such that h(d)
4:0. Writing fA(t) g(t)h(t). Theorem 2 implies that a a-conjugate d’ of d is a zero
of g(t). But then g(t) has a right factor t—d—’ and so fA(t) has a right factor
(t— d’)(t—aa..,,,) .(—t a1), a contradiction. Q...ED
Proposition 6. A finite set A ED is always a-algebraic, with rank A s|Al.
Proof. It suffices to find a monic polynomial f of degreesr==|A| which vanishes
on A. We proceed by induction on r, the case r=1 being clear. If r22, let aeA
and let f (t)=g(t)(t—a) where g is to be determined. In order that f vanish on A0
202 1.1. Lam

==A\{a}, g must vanish on A’={(b —a)‘b(b —a)‘1: be}. Since IA’ISr—l, we

can inductively choose such a g with deggSr— 1. Q.E.D.
Incidentally, the argument above provides an effective inductive procedure to
calculate the rank of A. In particular, the minimal polynomial of a doubleton set
A = {a, b} is

fun“) = (t “(17 —a)°b(b -a)_ 1)(t —a)- (13)

Consequently, a 2-element set always has rank 2.

§3 The Rank of the a-Vandermonde Matrix

In this section, we shall explain the relevance of the methods in §2 by relating
them to the computation of the rank of the a—Vandermonde matrix. As before, the
division ring D and the endomorphism a of D are both fixed.
We start with a quick review of the notion of the rank of a matrix over a division
ring D. Let A=(aij) be an m x n matrix over D. By the row space of A, we mean
the subspace .92 of the left n-tuple vector space D(D") spanned by the row vectors
of A; the dimension of .9? as a left D-vector space is called the row rank of A. The
column space and column rank of A are similarly defined, by working with the
right m-tuple vector space (D"')D. If D is a field, it is well-known in linear algebra
that, for any matrix A, the row rank of A is equal to its column rank. This fact
actually does not depend on the commutativity of D, i.e. it holds for any division
ring D. Since we shall need this result in a crucial way, we include a proof here
for completeness. I like the following proof best because it shows very clearly that
cummutativity is not required, and also shows why the left row space should be
related to the right column space.

Theorem 7. row rank (A)=column rank (A).

Proof. Let r=row rank (A) and c=column rank (A). Let B be an rxn matrix
whose r rows form a basis of the (left) row space. Expressing the rows of A as left
linear combinations of those of B, we get A=B'B for a suitable m xr matrix B’.
But the equation A=B’B may also be construed as expressing the columns of A
as right linear combinations of the r columns of B', hence cSr. A similar
argument shows that rsc. Q.E.D.

The common value of rowrank (A) and column rank (A) is, of course, called the
rank of A. If A is a square nxn matrix, then A is invertible iff rankA=n. In
general, rankA is the largest integer s such that A contains an invertible
sxs submatrix. (The easy proof is left to the reader.) As a word of caution, we
should point out that the rank of A need not be equal to the dimension of the
subspace spanned by the rows in the right n-tuple vector space (D")D. In other
words, rankA may not be equal to rankAT, where AT denotes the transpose of A. In
 A general theory of Vandermonde matrices 203

fact, for a,beD, A= (‘11 abb) has rank 1, but AT= (1 a) has rank2 unless ab
b ab
=ba. Thus, rankA =rank AT holds for all matrices A iff D is a field.
We are now ready to compute the rank of the a-Vandermonde matrix V
= V,,"(a1, ...,an). Let A ED be the subset {a1,...,a,,}. We have then
Theorem 8. rank V = rank A.
Proof. Let r=row rank V, c=column rank V and 6=rank A. We shall show that
r<5<c. Then Theorem 7 gives the desired result.
11— 1
The basic observation needed here is that a polynomial g(t)= go b t' (of degree
Sn- 1) vanishes on A iff (b0, bn_1) V: 0 (cf. (12)). To show thato6<c, it suffices
to find a nonzero polynomial g(t)eD[t, a] of degree Sc such that g(A)= 0. Among the
columns C1, ..., C,I of V, c of them form a basis of the column space. Without loss
of generality, we may assume that these are the first c columns. Let g(t)=2biti be
the minimal polynomial of the set {an ...,ac}. Then by Proposition 6, deggScSn,
and (b0, ...,bu_ il'C1=0 for lSiSC. Since each C]. is a right linear combination of
C1,...,Cc, we also have (bo,...,b,,_1)-Cj=0 and so (bo,...,bn_1)-V=0, i.e. g(A)
=0. Next, to show that r 36, let f = fA(t) be the minimal polynomial of A. It
suffices to show that each row Ri of V is a left linear combination of the first 6
rows R1, ...,R5. By the (left) division algorithm, we can write
=q(t)f(t)+(do+d1t+...+d,_1t""), (14)
since 6=rankA is the degree of f. Evaluating at a} (and using Theorem 2), we
have
M(aj)=do+d1N1(aj)+ u. +d‘5_lN;,_1(aJ-).

Thus, Ri=d0Rl +d1R2+ ..o +d6—1R6’ as deSil‘ed. Q.E.D.

Next we shall extend the above result to the case of a rectangular a-Vandermonde
matrix. To this end, let us write V;n(a1, ...,a,) for the m x n a-Vandermonde
matrix whose (i+1)"t row is (Ni(al), ...,M(a,,)).
Lemma 9. Suppose man. Then
rank V" (a1, ...,a,l)=rank V‘(al, ,a ).
Proof. Let f(t) and 6 be as in the proof of the theorem above. The last part of the
argument in this proof 1s applicable to any row ofV,” n.Hence
row rank V3”, 5 6 = row rank V:.
The reverse inequality is clear. Q.E.D.
A curious consequence of this lemma is that, if some nxn submatrix of V‘
(man) is invertible, then V: itself must must be invertible. In the case when D is
commutative, this means that if det V": 0, then any n x n submatrix ofV,1, (man)
204 T.Y. Lam

also has determinant zero. This is obvious when a is the identity, but seems to be
far from obvious otherwise (say when D= C and a=complex conjugation).
Combining Theorem '8 and Lemma 9, we have the following result, valid for any
pair (m, n).
Theorem 10. rank V;,,,(a1, . . . , a”) = min {m, rank A}
Proof. Let 6=rank A sn. We may assume that the first 6 columns of V,“ are right
independent. Hence a5(a1, ...,a,,) has rank 6. By Lemma 9, Vggal, ...,a,,) also has
rank 6. This implies that the 6 rows of 1453011, ...,a,,) are left independent, and so
the first 6 rows of K,°(a1,...,a,,) are also left independent. If msé, then clearly
rank V;,,,(a1,...,a,.)=m=min{m,6}. Next assume m_>_6. If in fact man, we are
done by Lemma 9. If msn, then rank a'nsrank V,,"=6. But the first 5 rows of
V5,, are left independent so rank V,,‘,’,,,=6=min{m,6}. Q.E.D. '
Of course, it also makes sense to talk about the infinite a-Vandermonde matrix
V;,n(a1, ...,a,,) whose i'h column is the infinite tuple (1,N1(ai),N2(a,.), ...)T. The row
rank and the column rank of V3,," are defined as before, and the proof of
Theorem 7 carries over verbatim to show that they are equal. By Theorem 10 (or
_ its proof), we see easily that rank V;,,,=rank A.
Finally, we note the following remarkable result:
Proposition 11. Let C, (resp. C?) be the i'h column of V,f(a1, ...,a,,) (resp.
V;'n(a1, ...,a,,)), and let b1, ...,bneD. IfZCibi=0, then Ecrb,=o.
Proof. This follows since, as we saw before, any row of V1,... is a left linear
combination of the n rows of V,,". The details will be left to the reader. Q.E.D.

§4 Polynomial Independence in Division Rings

The results on a-Vandermonde matrices in the last section suggests naturally a
notion of polynomial independence (or P-independence) in a division ring D
equipped with an endomorphism 0. Many of the results in this section were
motivated by the consideration of a-Vandermonde matrices. However, it is of
interest to first try to develop the theory of P—independence without reference to
Vandermonde matrices.
Let A be a a-algebraic set in D. We say that an element deD is P-dependent on A
if every polynomial in D[t,a] vanishing on A also vanishes on (1. Further, we say
that a a-algebraic set A is P-independent if no element beA is P-dependent on
A\{b}. (Since a is regarded as fixed, we suppress 'a in the terminology defined
above.) Clearly, any subset of a P-independent set is also P-independent.
Lemma 12. An n-element set A ED is P-independent 11f rankA =n.
Proof. If some beA is P-dependent on A\{b}, then rankA=rank(A\{b})Sn—1.
Conversely, suppose A is P-independent, and fix an element beA. Inductively, we
 A general theory of Vandermonde matrices 205

may assume that rank(A\{b})=n—l. Let f (resp. g) be the minimal polynomial of

A (resp. A\{b}). Then f is a left multiple of g but f 4:3. Hence deg f 21+deg g
=n. On the other hand, deg f SIAI =n, hence rankA =deg f =n. Q.E.D.
Corollary 13. If a a-algebraic set A is P-independent, then A is a finite set (of
cardinality rank A).
Proof. Let A() be any. finite subset of A. Then |Ao|=rankAoErank A. Thus
|A| srank A < 00. In fact, by the Lemma, |A|=rank A. Q.E.D.
The next lemma is a parody of a familiar fact about linear dependence in a vector
space.
Exchange Lemma 14. Let A be a a-algebraic set in D, and let b, deD. If b is P-
dependent on Au {d} but not on A, then d is P-dependent on A u {b}.
Proof. Let h be the minimal polynomial of A. If h(d)=0, then d is P-dependent on
A and hence on Au{b}. We may therefore assume h(d)=|=0. By Theorem 2, the
minimal polynomial of Au{d} is given by (t-h(d)"dh(d)‘1)h(t). By hypothesis,
this is satisfied by b. But since h(b)=1=0, Theorem 2 implies that
h(b)"b h(b)‘ 1 = h(d)"dh(d)‘.
1 (15)
Now the minimal polynomial for Au{b} is (t— h(b)"bh(b) l)h(t), so (15) implies
that this is satisfied by d, i.e. d 1s P-dependent on Au {b}. Q.E.D.
For any a-algebraic set A ED, let us define a subset BEA to be a P-basis of A if B
is P-independent, and every deA is P-dependent on B. As the reader would
suspect, we have the following characterization of a P-basis:
Proposition 15. B is a P-basis of the a-algebraic set A ED ifi“ B is a maximal P-
independent subset of A.
Proof. The “only if” part is clear. For the “if” part, assume BEA is maximal P-
independent. Consider any deA\B; we want to show that d is P-dependent on B.
By assumption, Bu{d} is not P-independent, so some element beBu{d} is P-
dependent on the others. If b=d, we are done so assume beB. Then b is not P-
dependent on Ao==B\{b}, but P-dependent on A0U{d}. By the Exchange Lem-
ma, we conclude that d is P-dependent on Aou {b} =B, as desired. Q.E.D.
Let A be any a—algebraic set with r=rank A. For any P-independent subset A0
EA, we have |A0|=rankAosrankA=r, so P-bases of A do exist. In fact, a P-
independent subset BEA is a P-basis for A ijf |B| =r. And, if B is any P—basis of A,
then the minimal polynomial for B coincides with that for A. On the other hand,
for any set A1 EA on which A is P-dependent (we say A1 is a P-spanning set for
A), we have |A1| 2r. We leave it to the reader to show that a P-spanning set BEA
is a P-basis for A iff it is a minimal P-spanning set, ijf |B| =r.
It is also of interest to note that we have the following complete analogue of the
Steinitz Replacement Theorem in linear algebra:
206 T.Y. Lam

Theorem 16. Let A ED be a a—algebraic set. Let {b,, ...,b,,,} be a P-independent set
in A, and {(11, ...,d,,} be a P-spanning set for A. Then, after a reindexing of the di’s,
{b1, ...,b,,,, dm+1,...,d,,} is a P-spanning set for A.
Proof. Assume, inductively, that {b1, ...,b,,d,+1, ...,d,,} is a P-spanning set for A,
where r<m. Suppose br H is P-dependent on {b1,...,b,, dil,...,di'}, where s is
chosen minimal. (Note that s21 by the P-independence of the bi’s.) By the
Exchange Lemma, d,l is P-dependent on {b1,...,b,, b,+1, di2,...,d,..}. Thus, the
latter set is a P-spanning set for A, and the proof proceeds by induction. Q.E.D.
Of course, the remarkable parallel of the theory of P—independence and the usual
theory of linear independence in vector spaces is not an accident. The exact
connection between these two theories can be seen by using a-Vandermonde
matrices, which make possible the following interpretation of P-dependence in
terms of linear dependence in linear algebra.
Proposition 17. For deD and any a-algebraic set A ED, the following statements are
equivalent:
(1) d is P-dependent on A;
(2) (1,N1(d),N2(d), ...)T is right linearly dependent on {(1,N1(a),Nz(a), ...)T: aeA};
(3) There exist a1,...,a,,eA such that (1,Nl(d),...,N _1(d))T is a right linear com-
bination of {(1,N1(ai), ...,N,,,_ l(ai))T: lsisn} for some m2n+ 1.
Proof. Note that d is P-dependent on A iff d is P-dependent on some finite subset
of A, iff rank({d}uA)=rank A. Using the Vandermonde matrix interpretation of
these ranks (Theorem 10), we get the equivalence of (1) and (3). Finally, the
equivalence of (2) and (3) follows from Proposition 11. Q.E.D.
The Proposition above unifies the theories of P-independence and linear inde-
pendence, and puts our earlier results on P-independence in perspective. Of
course, it would have been possible to deduce all these results from the corre-
sponding facts in linear algebra, by using Proposition 17. Nevertheless, the ad hoc
treatment of the theory of P-independence and P-bases given earlier in this
section seems to be of interest in its own right.
With the theory developed so far, we are able to write down effectively the
condition for a a-Vandermonde matrix V:(a1,...,a_) to be singular. In fact, we
can write down effectively the condition for V:(a1,...,a,,) to have rank equal to
any prescribed number r (1 srsn). By Theorem 8, this amounts to A ={a1, ...,a,,}
having rank r, which means that there exist r distinct elements ah,...,aireA
forming a P-basis of A, i.e. they are P-independent, and each a, is a root of the
minimal polynomial f of {a,1,...,air}. Since f can be computed effectively, we
know how to express the condition that V:(a1,...,a,,) has rank r by a finite
number of equations in D. For instance, if n=3, the condition for V3"(a, b, c) to be
singular is that either two or more of a, b, c are equal, or all three are distinct,
and c satisfies the minimal polynomial (t—(b—a)"b(b—a)'1) (t—a) of {a,b}, i.e.
 A general theory of Vandermonde matrices 207

(c —a)"c(c —a)‘ 1 =(b —a)"b(b -a)' 1

(cf. Theorem 2). In case a= 1, this agrees with the condition we derived earlier in
(1021) by using elementary row transformations.
At this point, it is appropriate to pause for some examples of the notion of
minimal polynomials and P-independence.
Let F be the center of D. It is well-known that F is a field. Let 0:1 and consider
a a-conjugacy class A of D. (This is just an ordinary conjugacy class.) When is A
a-algebraic? Suppose A is o-algebraic and let f(t)=z ai t‘eD[t, a] be the minimal
polynomial of A in the sense defined in §3. Let xeD*; for any d eD, write
d’=xdx’l. If deA, we can conjugate Zaidi=0 to get Zagd"=0. On the other
hand, Zaid"=0, so 2(a; —a,.)d"'=0. This means that g(t)==2(a§—a,.)t" vanishes on
A. Since f is monic, we have deg g<degf. This implies that a;=a,. for all i. By
varying xeD*, we see now that each aieF. Thus, to say that A is a-algebraic means
exactly that one, and hence all, of the elements of A are algebraic over F in the
usual sense. Moreover, the minimal polynomial for A in the sense of §3 is exactly
the common minimal polynomial of the elements of A over F in the usual sense.
Applying Lemma 5 and Theorem 4 to A, we recover the following classical
theorem in the theory of division rings (of. [Dicksom 1923, p. 230]):
Theorem'18 (Wedderburn, Dickson). Let D be a division ring with center F. Let
deD be algebraic over F with minimal polynomial f(t)eD[t]. Then f(t)=(t—d,,)...
(t—dl) for suitable conjugates di of d. Moreover, an element d’eD is conjugate to
d iff d and d’ have the same minimal polynomial f(t) over F.
For a more concrete example, let F be a field of characteristic not 2, and let
a, beF*. Let D be the F-algebra generated by two elements i,j, with relations
i2=a, j2=b, ij=—ji. (16)
In the literature, D is called a generalized quaternion algebra over F. It is not hard
to show that it has F-dimension 4, and that it is a division ring iff the equation
ax2+b y2=1 has no solution in F (see, e.g. [Lamz 1973, p. 58]). Assume we are in
this case, and consider the conjugacy class A of any noncentral element deD. By
dimension consideration, the field F(d) obtained by adjoining d to F must be a
quadratic extension of F, and so the minimal polynomial f (t) of A (say in the
sense of §3) is given by

f (t)=(t -d’)(t -d)=(t -d)(t -d’) (17)

where d’ is the image of d under the nontrivial F-automorphism of F(d). This
shows that rankA=2 and consequently any 2-element set {a,b}§A forms a P-
basis for A.
For the next illustration (and application) of our general theory of minimal
polynomials, consider a field D with an automorphism a of order n, and let F be
the fixed field of 0. Then D/F is a cyclic extension of degree n. Let A be the a-
conjugacy class of a nonzero element d 60*. Let c=d‘""...d"d=N,,,p(d)eF*, and
h(t)=t"—ceF[t]§D[t,a]. Fer any d’eA, we have d’=x"dx‘l for some xeD*, so
by Lemma 1(2), '
Nn(d’) = Nn(x”dx‘ 1) =x‘”N,,(d)x' 1 = a(d)x‘ 1 = c.
This shows that h vanishes on A; in particular, A is a—algebraic. Let f (t)
=Zait‘eD[t,a] be the (monic) minimal polynomial of A. Then m==deg f Sn, and
the constant term ao4=0. Let xeD*. For any eeA, we have 0=ZaiN,(x"ex‘1)
=Zaix"Ni(e)x‘l. Thus, A also satisfies the polynomial Zaixflt‘. By the unique-
ness of the minimal polynomial, we must have x""'a,.=a,.x“ for every i. In
particular, since ao=|=0, this implies that a”: 1. Therefore, we have m=n and f (t)
=h(t)=t"—c. Applying Lemma 5 and Theorem 4 to A, we get the following
theorem, the last part of which recovers the so-called “Hilbert 90” for cyclic
extensions: ' ‘
Theorem 19. In the notation above, we have t"—c=(t—d,,)...(t—dl) for suitable a-
eonjugates d, of d, with (11:11. The a-conjugacy class A is always of rank n,
Moreover, an element d’eD is o—conjugate to d if ND,F(d')=N,,/,(d).
Note that, in the theorem above, the decomposition of t”—c in D[t,a] is obtained
without assuming the existence of any roots of unity in D. On the other hand, for
any ceD*, if t"—c splits completely in D[t,a], then in particular it has a root deD
and so c=N,,(d)=ND,,-(d).
For a more concrete example, let D=C and a=“bar”' (the complex conjugation).
A typical a-conjugacy class is of the form A,= {aeCt |a|=r}. If r=0, AP has rank 1
and minimal polynomial t. If r=|=0, then A, has rank2 and minimal polynomial
f4'(t)=t2 —r2 =(t+a)(t —a)=(t+a)(t —a). (18)
(Recall that ta=at!) In the latter case, as in the case of the quaternions, any 2-
element set {a, b} EA, forms a P-basis for A,.
The remarkable thing about Theorem 19 is that we have used noncommutative
techniques to prove a result in a commutative situation. This is perhaps a good
instance of Emmy Noether’s philosophy that noncommutative algebra may be
important for the investigation of commutative arithmetic. In the treatment we
gave above, the apparently unrelated classical theorems of Wedderburn, Dickson
and Hilbert are unified in the light of Theorem 4 and Lemma 5. In fact, by
working more carefully, it is possible to obtain a single theorem which encom-
passes both Theorem 18 and Theorem 19. Such a common generalization will
give a Hilbert 90 Theorem valid in the noncommutative setting which is consider-
ably stronger than the presently known version in [Cohnz 1977, p. 68]. Its proof,
however, involves some technical computations, and is not entirely in keeping
with the elementary nature of this article. Therefore, we shall postpone its pre-
sentation to a sequel of this paper, [Lamz 1985]. In the latter, more results on P-
dependence, P-basis and polynomial equations over division rings can be found.
 A general theory of Vandermonde matrices 209

§5 The Excision and Union Theorems

In this section, we continue to fix a division ring D and an endomorphism a of D.
We shall prove some interesting results which say essentially that in studying P-
independence in D, we may work with one a-conjugacy class at a time. Stated in
terms of a-Vandermonde matrices, these results suggest that the consideration of
these matrices may be reduced to the case V,”(a1, ...,a,,) where all the ai’s lie in a
single o-conjugacy class. The fact crucial for making this reduction is the follow-
mg.

Excision Theorem 20. Let C be the a-conjugacy class of D determined by an

element ceD. If c is P-dependent on a a-algebraic set 1", then c is P-dependent on
1" n C.
Proof. We need to show that any h(t)eD[t,a] vanishing on In C rnust vanish on
c. For this, we may as well assume that h is the minimal polynomial of I'm C. For
any deI‘\C, we have then h(d)=l=0 (cf. Lemma 5 and Theorem 4). Let g(t) be the
minimal polynomial of the set
{h(d)"dh(d)‘ 1: deF\C} (19)
and let f (t)=g(t)h(t). By Theorem 2, f vanishes on 1"\C as well as on F n C. Thus
f vanishes on F, and hence on c. If h(c)=|=0, then by Theorem 2 again, g(t) must
vanish on h(c)‘ch(c)‘1eC. However, since 3 is the minimal polynomial of the set
(19), each root of g is a—conjugate to some deI‘\C. This is a contradiction.
Therefore, h(c)=0, as desired. Q.E.D.
Corollary 21. Let B, B’ be o—algebraic subsets of D each of which is P-independent,
such that no element of B is a-conjugate to an element of B’. Then BUB’ is also P-
independent.
Proof. Deny the conclusion. Then (say) some ceB is P-dependent on F==BouB’,
where BO=B\{c}. Letting C be the a-conjugacy class of c, the theorem above
implies that c is P-dependent on 1" n C. By hypothesis, To C is disjoint from B’ so
FnCEBo. Thus c is P-dependent on Bo, contradicting the P-independence of
B. Q.E.D.
Remark. The corollary above can also be deduced from the Unique Factorization
Theorem on D[t,a]. However, the proof we just gave is more elementary, and
Theorem 20 yields further useful information.
Union Theorem 22. Let A, A’ be a—algebraic sets in D such that no element of A is
a-conjugate to an element of A’. Let B, B’ be, respectively, P-bases for A and A’.
Then A UA’ is a a-algebraic set, and BuB’ i a P-basis for it. In particular,
rank(A u A’) = rank A + rank A’.
Proof. If a polynomial vanishes on BuB’, then it also vanishes on A and A’.
Therefore AuA’ is a-algebraic, and every element of it is P-dependent on BuB’.
Since BUB’ is P-independent by Corollary 21, it follows that BUB’ is a P-basis
for AuA’. Q.E.D.
Theorem 23. Let A be any a-algebraic set in D. Then
(1) rankA=Zrank(AnC), where C ranges over the (finitely many) a-conjugacy
classes which intersect C.
(2) Let fA and fMC be, respectively, the minimal polynomials of A and Am C. Then
deD is a zero offA ifi' it is a zero offdnc for some C as in (1).
(3) If IA n CI $2 for each C as above, then |A|< 00 and A is P-independent.
Proof. (1) follows from Theorem 22 and induction. (3) follows from (1) and the
fact that any doubleton set is P-independent. For (2), we need only prove the
“only if” part. Let ceD be a root of f4- Then c is a—conjugate to some element c’
of A. Let C be the a-conjugacy class of c’ so An C=|=¢. Since c is P-dependent on
A, it is already P-dependent on An C by the Excision Theorem. Therefore, c is a
root of fMC. Q.E.D.
Specializing to the case when a=1 and |AnC|sl for each conjugacy class C,
part (2) and part (3) of Theorem 23 recover, in a somewhat stronger form,
Proposition 5 of [Bray-Whaples: 1983].
In terms of a-Vandermonde matrices, Theorem 23 means the following: If A is
any finite set and A=A1u...uAm is the partition of A into a-conjugacy classes,
then the square a-Vandermonde matrix V built on A has rank Zr,- where ri is the
rank of the square a-Vandermonde matrix Vi built on A... Keeping this notation,
we may, in particular, draw the following conclusions:

Corollary 24. rank V2 E1 min{2, IA |} The matrix V is invertible iff each square
submatrix V is invertible If no three of the elements of A are o-conjugate, then V'is
invertible. If V is invertible, and A: {d1,... ,dn}, then for any given elements
b1,...,b,,eD, there exists a unique polynomial f(t)eD[t, a] of degreeSn—l such
that f (d,)=b, for all i.
(For the last statement, recall the interpolation formula (12).)
We shall mention two interesting cases of Corollary 24 in which we get a direct
and complete computation of the rank of the a—Vandermonde matrix. The general
notation above will be fixed.
Application (A). For any generalized quaternion division algebra D over a field F of
characteristic =|=2 (cf. (16)), let 0= 1. Then rank V= 2 min{2,|A,|}.
l-l

Application (B). For D=C and a=“bar”, let 61,...,6,,, be the different values
among |d|: deA. Then rank V= Z r, where ri=1 if exactly one element in A has
modulus 6,- and ri=2 otherwise. "1
 A general theory of Vandermonde matrices 211

While Corollary 24 gives a lower bound on the rank of a a-Vandermonde matrix

V= V,,"(d1, ...,d,,), it is also possible to give an upper bound in case 0:1 and the
division ring D is centrally finite, i.e. D is finite dimensional as a vector space over
its center F. In this case, it is well-known that dimFF =k2 for some natural
number k ([Herstein: 1968, p. 91]), which is called the degree (or Schur index) of
D. In the same reference, it is also shown that k can be characterized as
max{dim,. K} where K ranges over all subfields of D containing F. For any deD,
it follows that the minimal polynomial of d over F has degree dimF F(d)sk. By
what we said in the paragraph preceding Theorem 18, this implies that any
conjugacy class C of D has ranksk. Thus, if A ={dl, ...,d,,} intersects m conjugacy
classes of D, then, by Theorem 23(1), rank V,,(d1, ...,d,,)smk.
In view of the results in this section, the study of o-Vandermonde matrices is
reduced to the case V=V,,"(a,,...,a,,) where a1,...,a,I lie in a single a-conjugacy
class of D. More results concerning the rank of V in this case will be given in
[Lamz 1985].

§6 The Transpose of a Vandermonde Matrix

Throughout this section, we take 0:11.)
We have remarked in an earlier section that, over a division ring D, the transpose
MT of a matrix M may not have the same rank as that of M. In dealing with
Vandermonde matrices, one may perhaps wonder: If V=Vm_,,(a,,...,a,,) with all
aieD, is rank V=rank VT? Since V seems to have a certain degree of internal
symmetry, one might hope for an affirmative answer. To treat this question, we
can first dispose of the case when m or n is $2, or when D is a field, for the
answer is indeed affirmative in these cases. Of course, our results are applicable to
the transpose of V by using right polynomials over D instead of left polynomials.
So the problem of deciding whether rank V=rank VT amounts to a problem of
studying the relation between roots of left and right polynomials. If D is a
generalized quaternion division algebra over a field F (charF 4:2), we have al-
ready shown that rank V is determined completely by the way the ai’s distribute
into conjugacy classes. Since this distribution is left-right symmetric, we have the
following pleasant result:
Theorem 25. If V is any rectangular Vandermonde matrix over a generalized quater-
nion division algebra, then rank V =rank VT.
Since generalized quaternion division algebras have degree 2 (over their centers),
the next natural case to consider is the case of (centrally finite) division algebras
of degree 3. The situation in this case turns out to be completely different, for we
have the following result:
Theorem 26. There exist division algebras D of degree 3 over Q such that some
3 x 3 Vandermonde matrix V= V3(a, b, c) over D has rank 2 but the transpose VT of V
has rank 3.
41‘ LY. Lam

To produce such an example, we use a construction due to LB. Dickson. Follow-

ing [Dickson: 1923, p.71], we take the field K=Q(a) where a has minimal
polynomial m(t)=t3+t2—2t—1 over Q. The field K is easily seen to be Galois
over Q, the other roots of m(t) being a2—2 and l—a—az. We then construct D
=K®K - x@K - x2, where x is a symbol satisfying the relations x3 =2 and xa=
(a2 —2)x. These relations (and their consequences) define a multiplication on D, and
make D into a 9-dimensional Q-division algebra with center Q. (The 'details were
very readably written out in Dickson’s book; D is an explicit example of
Dickson’s famous cyclic algebras, ca. 1906.)
Let b = xax‘ 1 = a2 —2, and consider the quadratic polynomial

f(t)=(t—a)(t —b)=t2 —(a+b)t+abeD[t]. (20)

Since a,b commute, they are both right roots of f (t). Let us look for conjugates
dad‘1 of a which may be right roots off. We need to have dazd'1 —(a+b)dad'1
+ ab =0, or equivalently,
da2 —(a+b)da+abd=0. (21)
We do know that a, and b(=xax") are right roots of f, so (21) is satisfied by d
=1 and by d =x. Since (21) is linear in (1, this assures that (21) is also satisfied by
d=1+x. Thus, a third right root off is

c==(1+x)a(1+x)'l
=(1+x)a(l —x+x2)/3 (since x3=2)
=(a+bx)(l —x+x2)/3 (22)
= ((a + 2b) + (b —a)x + (a —b)x2)/3.
Noting that f (t) is the minimal polynomial f4 for A={a,b}, the fact that f (c)=0
implies that c is P-dependent on A. Thus, V= V3(a, b, c) has rank 2.
Of course, the word “P-dependent” above should have been written as “left P-
dependent”, since it is checked by using left polynomials. We have a parallel
notion of right P-dependence by using, instead, right polynomials. We claim that
the 3-element set {a, b, c} is right P-independent. If this is the case, then applying
our results to right polynomials, it follows that VT is invertible, i.e. rank VT=3.
Assume, instead, that {a,b,c} is not right P-independent. Then there exists a
monic quadratic right polynomial g(t) vanishing on {a,b,c} as left roots. Since a,b
commute, g(t) must be just
(t —a)(t —b)=t2 -t(a+b)+ab.
Thus, we have c2 —c(a + b) +ab=0 as well as c2 —(a+b)c +ab =0. This implies
that a+b commutes with c. Since K - x‘=x‘- K and D=K$K~x$K-x2, we see
from (22) that a +b commutes with all three “components” of c in (22). In
particular (b —a)x(a + b) = (a + b)(b —a)x, or (b —a)(b + b2 —2)x = (a + a2 —2)(b —a)x.
 A general theory ‘of Vandermonde matrices 213

But then b + b2 -2 =a +a2 —2, which simplifies to (b —a)(b +a + 1)=0. This is

absurd since a is not a quadratic irrationality.
Although the computations above may seem rather special, the method used is, in
fact, general. Using this method, it can be shown that, for any field F and any F- -
central division algebra D of degree d, we have rank V= rank VT for all Vander-
monde matrices V over 'D iff dsZ. Moreover, as observed by K.H. Leung, it is
possible to construct examples of division algebras of degree d in which there are
sets A whose right rank is d, but whOse left rank is 2. Further results in this
direction may be found in [Lamz 1985]. -

Epilogue
Just as this article was being typed up for publication, I received, with perfect
timing, the doctoral disseration of André Leroy at Université de l’Etat a Mons,
entitled “Derivations Algébriques.” In Chapter 3 of this dissertation, Leroy also
considered Vandermonde matrices over division rings D through the connection
with Wronskian matrices, and observed the fact that V,,"(a1, ...,a,,) is invertible iff
“the left LCM of t—a1,...,t—a,, in D[t,a] is of degree n.” In a letter to me dated
March 18, 1985, Leroy further made the following important observations.
For a division ring D, let a be a fixed endomorphism and let 6: D—vD be a a-
derivation, i.e. an additive homomorphism such that 6(ab)=6(a)b+a"6(b) for all
a,beD. One can form the Ore skew polynomial ring D[t, 0,6] whose elements are
left polynomials Zait‘, multiplied according to the rule ta=a"t+6(a), for any
aeD. (For 6:0, one gets back the D[t,a] used in the text.) For aeD, redefine
Nn(a) by taking No(a)=1 and inductively M+1(a)=l\ln(a)‘a+6(Nn(a)). Then define
the (a,6)-Vandermonde matrix K,""(al, ...,a,,) as in (3). For f (t)=2ait‘eD[t, a, 6],
define f (a)=2aiN,-(a) as before, and, for c,deD, define c to be a (a,6)-conjugate of
d if c=a"da‘1 +6(a)a'l for some aeD*. Using the techniques in his dissertation,
Leroy has proved the following generalization of our Theorem 2:
Theorem (Leroy). Let f (t)=g(t)h(t)eD[t, a, 6] and deD. If h(d)=0 then f (d)=0. If
a==h(d)=|=0, then f(d)=g(a"da'l +6(a)a'1)h(d).
In view of this new form of Theorem 2, Leroy observed (and I happily agree) that
all of the main results in this paper carry over, with the same proofs, to the more
general case of (a, 5)-Vandermonde matrices and (a,6)-polynomial dependence.
One nice consequence of this observation is the following. Given (D, a, 6), and
x1, ...,x,,eD*, let W(x1, ...,x,,) be the following Wronskian matrix
x1 x2 x,l
6(x,) 6(x2) 6(xn)
............................................. (23)
.............................................

6n— 1(x1) 5::- 1(x2) au— 1 (xn)

41-! LY. Lam

Let a,= 6—(x,)x'l. According to Lemma 4 in Chapter 3 of Leroy’s dissertation,

6(x)= a x. implies that 6j(x)= Nj(a-,..)x Therefore, one has the following matrix
equation:
W(x1, ...,xm)= K,""(a1, ...,a") diag(x1, ...,x,), (24)
and so W(x1, ...,x,,) has the same rank as Vu“(al, ...,an). In particular, the results
of this paper, in the more general (0,6) form, will yield information on the rank of
the Wronskian matrices.

Note added in Proof. In a footnote to (51, the author conjectured that, over the complex
numbers with a =complex conjugation, one has the determinantal inequality
Idet V,,"(a1, ..., an)|2|det V,,(|a1|, ..., |a,,|)|. This has recently been proved to be true by David
Richman et al. at the University of South Carolina.

References
A. Aitken, Determinants and Matrices, Interscience Publishers, New York. Reprinted 1959.
E. Artin, Geometric Algebra, Interscience Publishers, New York, 1957.
B. Beck, Sur les equations polynomiales dans les quatemions, L'Enseignement Math. 25
(1979), 195—201.
L. Brand, The roots of a quaternion. MAA Monthly 49 (1942), 519—520.
U. Bray and G. Whaples, Polynomials with coefficients from a division ring, Canad. J.
Math. 35 (1983), 509—515.
L. Carlitz, A special determinant, MAA Monthly 62 (1955), 242—243.
C. Cerlienco and M. Mureddu, A note on polynomial equations in quatemions, Rend. Sem.
Fae. Sci. Univ. Cagliari 51 (1981), 95—99.
P.M. Cohn, Unique factorization domains, MAA Monthly 80 (1973), l—18. .
P.M. Cohn, Skew Field Constructions, London Math. Society Lecture Notes Series, Vol. 27,
Cambridge University Press, 1977.
J. Dauns, A Concrete Approach to Division Rings, Heldermann Verlag, Berlin, 1982.
LE. Dickson, Algebras and their Arithmetics, 1923. (Reprinted by Dover Publications,
1960).
H. Edwards, Galois Theory, Graduate Text in Mathematics, Vol. 101, Springer-Verlag,
Berlin, Heidelberg, New York, 1984.
S. Eilenberg and I. Niven, The “Fundamental Theorem of Algebra” for quatemions, Bull.
Amer. Math. Soc. 50 (1944), 246—248.
A. Farnell, A special Vandermondian determinant, MAA Monthly 66 (1959), 564—569.
B. Fein and M. Schacher, Solutions of pure equations in rational division algebras. I., 'J.
Algebra 17 (1971), 83~93.
B. Gordon and T. Motzkin, On the zeros of polynomials over division rings, Trans. Amer.
Math. Soc. 116 (1965), 218-226. (Correction ibid. 122 (1966), 547.)
T. Grundhofer and H. Liineburg, Uber die Summation von Reihen aus linearen Schiebe-
registerfolgen, Archiv der Mathematik 34 (1980), 428—439.
I. Herstein, Conjugates in division rings, Proc. Amer. Math. Soc. 7 (1956), 1021—1022.
I. Herstein, Non-commutative Rings, Carus Monograph, Vol.15, Math. Assoc. America,
1968. '
 A general theory of Vandermonde matrices 215

D. Hilbert, Grundlagen der Geometric, 1899. Leipzig and Berlin, Teubner, (9th ed., 1962,
Stuttgart)
K.A. Hirsch, A note on the Vandermonde determinant, J. London Math. Soc. 24 (1949),
144—145.
N. Jacobson, The Theory of Rings, Amer. Math. 800., Providence, R.I., 1943.
D. Kalman, The generalized Vandermonde matrix, Math. Magazine 57 (1984), 15—21.
T.Y. Lam, The Algebraic Theory of Quadratic Forms, Benjamin, Addison-Wesley, 1973.
(Second Printing with revisions, 1980.)
T.Y. Lam, Polynomial dependence and polynomial equations in division rings, to appear,
1985. .
D.E. Littlewood, The Theory of Group Characters, 2nd edition, Oxford University Press,
1950.
LG. MacDonald, Symmetric Functions and Hall Polynomials, Clarendon Press, Oxford,
1979.
T. Muir, A Theatise on the Theory of Determinants, 1933. (Reprinted by Dover Publications,
1960.)
F. Murnaghan, The Theory of Group Representations, 1938. (Reprinted by Dover Publi-
cations, 1963.)
I. Niven, Equations in quarternions, MAA Monthly 48 (1941), 654—661.
I. Niven, The roots of a quaternion, MAA Monthly 49 (1942), 386—388.
V. Norton, A generalized Vandermonde determinant (Problem E2942), solution in MAA
Monthly (1984), 586—587.
B. Pollak, The equation tat=b in a quaternion algebra, Duke Math. J. 27 (1960), 261—271.
A.R. Richardson, Equations over a division algebra, Messenger of Math. 57 (1927), 1-6.
H. Rohrl, On the zeros of polynomials over arbitrary finite dimensional algebras, Manu-
scripta Math. 25 (1978), 359—390.
K.C. Smith, Theory and Applications of Noncommutative Polynomials, University of Okla-
homa Lecture Notes, 1974.
R. Stanley, Theory and application of plane partitions. 1., Studies in Applied Math. 50
(1971), 167—188.

Received 21.5.85

Department of Mathematics
University of California
Berkeley CA 94720