BULL. AUSTRAL. MATH. SOC.

I 5AI 5

VOL.

( I 5A09 )

21 ( 1 9 8 0 ) ,

137-146.

A DETERMINANT FOR RECTANGULAR MATRICES

V.N. JOSHI

The familiar notion of the determinant is generalised to include

rectangular matrices. An expression for a normalised generalised
inverse of a matrix is given in terms of its determinant and a
possible generalisation of the Schur complement is discussed as a
simple application.

1.

Introduction

The well known concept of the determinant has been defined to cover
square matrices alone.

This concept is very intimately related to the

concept of inversion of square nonsingular matrices in the sense that

square matrices which have nonvanishing determinant can be inverted.

In

extending the idea of inversion to cover rectangular matrices [6], the need
to generalise the concept of the determinant was not felt as a number of
methods devised to compute the inverse did not use any determinants
whatsoever.

It is felt by the author that many of the identities from

Linear Algebra can be generalised and be used in, say, the theory of
permanants and its evaluation [4],

Certain results in respect of this will

be communicated for publication elsewhere.

It is shown [6] that for an m * n
unique

n x m

matrix

X = A

real matrix

A , there exists a

which satisfies the following conditions:

(1.1)

AXA = A ,

(1.2)

MX = X ,

(1.3)

UX)T = AX ,
Received 16 August 1979137

138

V.N. Joshi

(1.10

(XA)T = XA .

The matrix

is known as the Moore-Penrose generalised inverse of A .

For some purposes, however, matrices which satisfy fewer than all of the
above conditions are of interest. A matrix which satisfies the condition
(l.l) is called a generalised inverse A *'

of A . A matrix satisfying

conditions (l.l)-(l.2) is called a reflexive generalised inverse A

of
1

A . A matrix which satisfies the conditions (l.l)-(l.2) and (l.3)/(l. 0 is

known as a right/left normalised generalised inverse A
class of inverses mentioned above is nonempty;

A+ {AM}d

of A . Each

in fact

{AM}<z

The Moore-Penrose generalised inverse A

of A

is primarily important

because of its uniqueness [6]. The other types of inverses are, however,
required in many applications [?, 7, S, 9].

2.
Let
m, K

be the set of integers

, ...,

m5 n ,

(ii)

K . J
"pi*
K

If

for a l l

{l, 2, . . . , n} .

i i J

Yl

pX<KP2<

For an integer

\A \
Let the integers

be such t h a t

(i)

( i i i )

A determinant

and p = 1, 2, . . . ,

TTl

<K

pm-

d , 1 2 d 5 (n-m+l) , define a set S, such that

N, = n~C
, , then the cardinal number of 5 , is tf, . The sets
a
m-1
a
a

S , , 1 <i 5 -m+l , will be ordered as follows. A set S

u < v . Moreover, the elements
e < e
p
q

whenever

K
< K
ps
qs

and

for all

< S

whenever

will be placed in the order

s = 2, 3, ..., m . All the

m-tuples, therefore, admit of the following order; namely,

determinant

Consider a r e a l matrix

for

A = [a. .)

be a submatrix of order

m * m of

ordering of integers in

DEFINITION
elements, l e t

rectangular

of order

matrices

mx

|39

m n .

/4 whose columns conform to the

1 d (n-m+l) ,

2 . 1 . For an m * n ,

1 p tf, .

m n , matrix

A with r e a l

be defined as above, then the number

(2.1)

n-m+l

N
d

d=l

p=i

det/
p

will be defined as the determinant of A , and will be denoted by

DEFINITION

Let

\A\ .

2.2. For an m x n , m > n , matrix A with real

1 T\

11

elements, \A\ will be defined as |A | , where A denotes the

transposed matrix of A .
LEMMA. (1) jf m = x , then \A\ = a + a + ... + a .
(2) If m = n , then

by

\A\ = det i4 = det A^ .

(3) If any rou of A is multiplied by c , then

0.
(k) If any two rows of A are identical, then

\A\ is multiplied

\A\ = 0 .

(5) If any row of A is a linear combination of the remaining vows

of A , -then \A\ = 0 .

by

(6) If any two rows of A are interchanged, then

-1 .

\A\ is multiplied

Proof. For m = 1 , N, = 1 and 1 d n . In view of Definition

2.1, (1) now follows. If m = n , then d = 1 = N, and A

= A ; from

this it follows that

is multiplied

\A\ = det A . If the ith row of A

by c , then the tth row of the submatrix A

p

for every value of d and

will be multiplied by O , in which case det A

will be multiplied by

140

V.N. Josh i

c . From this (3) follows. The proof of (k), (5), (6) follows on the same
lines by noting the corresponding properties of A
and

for every value of p

d .
The

det A

contains

permutation of

elements of

m\

terms each of which corresponds to a

. More precisely, each term of

det A

is a product of

taken from its each row and column once only with

positive or negative sign attached to it according as the permutation of

,
e

is even or odd.

Since

UI

contains

m\ x nc

= m\ x

n-m+1
T
N, = np m

J?2_

terms in all with each one containing the entries from different rows and
different columns of
DEFINITION

A , we can define \A\ equivalently as follows:

2.3.

Let

be a real matrix of order

m xn , m 5 n ;

then
I* I
where
m

l }

a m t m

ranges over all the permutations of integers

at a time from

J
n

and

required to bring

t,

..., t

EXAMPLE

2.1.

taken

Find the determinant of A , where

1 2 - 1 1

Since

= 1 . The sets

,t

to its natural order.

1 1 0

1 3
Solution.

t.,

is defined as the number of inversions

A=

it

m = 3
and

and
S

n = k , d = 1

or

2 and

= 3 and

contain the following elements, namely,

S = {(1, 2, 3 ) , (1, 2, k), (1, 3, 1*)} , S2 = {(2, 3, k)} . In view of

Definition 2.1, therefore,

A determinant

for rectangular

1 2 - 1
\A\

1 1

1 3

= -12 + 2 + 6 + 2 =

141

matrices

-1

-1

-2

3. Cofactors and expansion of \A\

In view of Definition 2.3 of

j ^4 j , it is clear that

and homogeneous function of the entries in the

ith

row of

\A\ is a linear
A . If

C..
z-J

denotes the coefficient of

a. . , j = 1, ..., n , then we get the

expression
(3.1)

\A\ = a. C.

+ ... + a. C.

The coefficient
Let

E, F, G

C.. of a., in (3.1) is called the cofactor of a.. .

**3
1*3
1*3
and E be the submatrices of A of the order

(i-l) x (j-i), (i-l) x (n-j), {m-i) x (n-j) and

(m-i) x (j-i)

respectively such that

E
A =

a
E

ith
I

row;

jth

column

then the determinant of the submatrix

(3.2)

M. . =
1-3

of the order

(w-l) x (n-l)

-F

-E

corresponds to the cofactor of

a.. .

Alternately we have
Ci;j = \Mid\ .

(3.3)

Using (3.l)-(3.3) we can evaluate

lower order.
EXAMPLE

\A\ in terms of the determinants of

This is illustrated in the following example.

3.1.

For the matrix

cofactors of the second row.

of Example 2.1, find

\A\ using the

142

V.N.

Solution.
j = 1, . . . , h

We have

Joshi

1
\A\ = 2C
^ , where
2 +
C2
+ 1C^
1C^ +
+ 1C
1C2233+ OC
^

C^

a r e t h e following determinants:
-2

-1

21

= -2|(2 1)| + l | ( - 3 1 ) | - l | ( - 3 -2)|

= -2(2+1) + l ( - 3 + l ) - l ( - 3 - 2 ) = -3 ,

"22

1 1 - 1
- 1 2
1

= l|(2 1)| + l | ( l 1)| - l | ( l -2)|

= 1(2+1) + 1(1+1) - 1(1-2) = 6 .

Similarly

= -2 and C ^ = -11 , which yields

\A\ = -2 . Note that

we have used the cofactors of the first row of M. . to evaluate C. . in

13
13
terms of the determinant of row vectors according to (l) of the lemma.

4. Normalised generalised inverse of A

We recall that for an m x n matrix A , the class of m x n
matrices

satisfying the conditions

ABA = A , BAB = B and AB or BA symmetric,

is not empty.

In spite of the fact that such a matrix

is not unique,

it is required in statistics for many estimation purposes [/, 7, 8, 9]. In

what follows, we shall present a realisation of such a B in terms of the
cofactors of elements of A whenever

\A | ?t 0 . In view of (3.1) and (h)

of the lemma, it is easy to see that

n

Y, a. .C-, . = & .

\A\

3=1

Let

C = (C.
c .) , be the
13

cofactors of

a.
a.13.. .
13

If

mx n

matrix whose elements,

C = R = [r.. .)
.) , then
13
13

n
13'

13'

1 ->> V 1 . 1 '

k=l
= \A\I

From this i t follows that for

(U.2)

mxm

\A\ # 0 , the matrix

B = R/\A\

13

r . . = C.
C. . .
13
13
13
13

AR = [a. v.) [r. .) =

*

C. . , are the
Now,

determinant

for

rectangular

matrices

4 3

is a normalised generalised inverse of A . For the matrix A of

Example 2.1, we find that the B , as defined above, is given by
(l
-1*

-3

l\
-2

2
7

-2
-11

0
3.

The expression (U.2) breaks down when

\A| = 0 . In the next theorem

we give a necessary and sufficient condition for matrix

deficient in which case
THEOREM

\A\ = 0 .

to be rank deficient is that for all

d = 1, ..., (n-m+l) and p = 1, ..., N-, ,

Proof.

Let r < m n

rows.

rows.

= 0 .

be the rank of A . This implies that r

(m-r) rows can be expressed as linear combinations of

For every value of

therefore, A
r

det A

are linearly independent and form a basis for the row-space of

A . The remaining
r

to be rank

4.1. A necessary and sufficient condition for an m x n ,

m S n , matrix A

rows of A

Since

d = 1, ..., (n-m+l) and p = 1, ..., N, ,

contains at least

(m-r) rows as linear combinations of

r < m , it follows that

det A

= 0 . Conversely, if

det A = 0 for every value of d and Fp , the column rank of A

will be
P
P
smaller than m . This implies that the column space of il is spanned by
a basis containing less than
column rank deficient matrix.

columns of A . In other words, A

is

Since the column rank of a matrix is the

same as its row rank, the result now follows.

COROLLARY
m xn ,

4.1.1.

A necessary and sufficient condition for an

m 5 n , matrix A

to be of full row rank is that

zero for at least one value of p

COROLLARY
matrix A

4.1.2.

and d .

to be of full rank is that det A t 0 .

is not

A necessary and sufficient condition for a square

5.
Let

det A

Schur complement

be an m x n , m S n , matrix partitioned as

I 44

V .N. Josh i

11
A =
22.
where

is an

v x v

Schur complement of

nonsingular principal submatrix of A . The

in A , denoted by

[A/A ) , is defined as the

matrix
A

Define

=
v

B = [b

22

~ C
and f o r
mv

) , where

, construct

1 x t

matrix
v

\B\ = det A

The matrix A

A A

^21 11 12 '

I S p <i

= det A

(5.D

- A

We s h a l l show t h a t
11

can be written as
12

11
A =
A

say, where the matrices

respectively.
1 - p - t

Let

5 c S

A~

21 11

m-r

and

A
A

are of order

to

are the integers

entries of such

1, ..., r, e\\

1, 2, ..., r

are denoted by

P>

, 1 2 p < tr

= det A1 = det

= det Q
Hence

m xm

and m x n
e

for

. In view of the order relation introduced amongst the elements

in Section 2, we now note that the first

which contains the elements

e~ =

A1

= det

entries in

in that order.
e

Now we have
= det P det Q1

pe

belonging

If the last {m-r)

, then we can write

Pi

A determinant

for

rectangular

matrices

= det A

&et[A/A

\B\ = iet A,. &et[A/A.,)l

11

<' I V p c

= det A.,\[A/A.A\

11

11 L*
p=l

145

)X

* ' 1 1 >pc
F

11

The l a s t step has followed from Definition 2 . 3 .

If

m = n , then

= 1 and, in view of (2) of the lemma, equation

(5.1) reduces t o det A = det A

det[A/A

) .

This r e s u l t has been proved

in [ 2 , 3 ] .

References
[J]

A.J. Goldman and M. Zelan, "Weak generalized inverses and minimum

variance linear unbiased estimation", J. Res. Nat. Bur.
Sect.

[2]

Standards

B 68 (196U), 151-172.

Emilie V. Haynsworth, "Applications of an inequality for the Schur

complement", Proa. Amer. Math. Soo. 24 (1970), 512-516.

[3]

Alston S. Householder, The theory of matrices in numerical

analysis
(Blaisdell [Ginn and Company], New York, Toronto, London, 196U).

[4]

W.B. Jurkat and H.J. Ryser, "Matrix factorizations of determinants and

permanents", J. Algebra 3 (1966), 1-27.

[5]

Gerald L. Morris and Patrick L. OdelI, "A characterization for

generalized inverses of matrices", SIAMRev. 10 (1968), 208-211.

[6]

R. Penrose, "A generalized inverse for matrices", Proc. Cambridge

Philos. Soc. 51(1955), U06-U13.

[7]

S.R. Searle, "Additional r e s u l t s concerning estimable functions and

generalized inverse matrices", J. Roy. Statist.
Soc. Ser. B 27
(1965), U86-U9O.

[S]

N.S. Urquhart, "Computation of generalized inverse matrices which

satisfy specified conditions", SIAMRev. 10 (1968), 216-218.-

146

[9]

V.N. Josh i

N.S. Urguhart and J.S. Williams, "A multivariate test of significance

for responses from any population I:

Unrestricted randomization"

(Biometrics Unit Mimeo, Ser. No. BU-156, 1967. Cornell

University, Ithaca, New York, 1967).

Victoria Jubilee Technical Institute,

H.R. Mahajani Marg,
Matunga,
Bombay 400 019,
Ind ia.