Supplemental Materials for EE203001 Students

II. Determinant Functions

Chung-Chin Lu
Department of Electrical Engineering
National Tsing Hua University
May 22, 2003

1 Three Axioms for a Determinant Function

Let d be a scalar-valued function on the space Mn×n of all n × n matrices. Let A1 , A2 , . . . , An
be the n rows of an n × n matrix A. We will denote the value d(A) of A under the function d
as d(A) = d(A1 , A2 , . . . , An ), indicating that d is a scalar-valued function of n n-dimensional
row vectors A1 , A2 , . . . , An . The scalar-valued function d is called a determinant function of
order n if it satisfies the following three axioms:
Axiom 1. Homogeneity in each row . If matrix B is obtained from matrix A by multiplying
one row, says the ith row, of A by a scalar α, then d(B) = α d(A), i.e.,
d(A1 , . . . , αAi , . . . , An ) = α d(A1 , . . . , Ai , . . . , An ).

Axiom 2. Invariance under row addition. If matrix B is obtained from matrix A by adding
one row, says the kth row, of A to another row, says the ith row, of A, then d(B) =
d(A), i.e.,
d(A1 , . . . , Ai + Ak , . . . , Ak , . . . , An ) = d(A1 , . . . , Ai , . . . , Ak , . . . , An ).

Axiom 3. The determinant of the identity matrix is one.

d(In×n ) = d(e1 , e2 , . . . , en ) = 1.

The primary purpose of this supplement is to show that there exists such a determinant
function d of order n and it is unique. Before being able to do so, we will derive important
properties of such a determinant function d if exists.

1
Professor Chung-Chin Lu for EE203001 Linear Algebra 2

Theorem 1 If some row of matrix A is the zero vector, then d(A) = 0.

Proof. Without loss of generality, assume the ith row Ai of A is the zero vector. Then
d(A) = d(. . . , Ai−1 , 0, Ai+1 , . . .) = d(. . . , Ai−1 , (−1)0, Ai+1 , . . .)
= (−1)d(. . . , Ai−1 , 0, Ai+1 , . . .) = (−1)d(A),
where Axiom 1 is applied to the third equality. Thus we must have d(A) = 0. 2
Theorem 2 If matrix B is obtained from matrix A by adding a scalar multiple of one row,
says the kth row, of A to another row, says the ith row, of A, then d(B) = d(A), i.e.,
d(A1 , . . . , Ai + αAk , . . . , Ak , . . . , An ) = d(A1 , . . . , Ai , . . . , Ak , . . . , An )
for any scalar α.
Proof. It is trivial if α = 0. Assume that α 6= 0. Then we have
1
d(A1 , . . . , Ai , . . . , Ak , . . . , An ) = d(A1 , . . . , Ai , . . . , αAk , . . . , An )
α
1
= d(A1 , . . . , Ai + αAk , . . . , αAk , . . . , An )
α 
α

= d(A1 , . . . , Ai + αAk , . . . , Ak , . . . , An ),
α
where Axiom 1 is applied to the 1st and the 3rd equalities and Axiom 2 is applied to the
2nd equality. This completes the proof. 2
Theorem 3 If matrix B is obtained from matrix A by interchanging two rows of A, says
the ith and the kth rows with i 6= k, then d(B) = (−1)d(A), i.e.,
d(A1 , . . . , Ak , . . . , Ai , . . . , An ) = (−1)d(A1 , . . . , Ai , . . . , Ak , . . . , An ).
Proof. We compute
d(A1 , . . . , Ai , . . . , Ak , . . . , An ) = d(A1 , . . . , Ai + Ak , . . . , Ak , . . . , An )
= d(A1 , . . . , Ai + Ak , . . . , Ak + (−1)(Ai + Ak ), . . . , An )
= d(A1 , . . . , Ai + Ak , . . . , (−1)Ai , . . . , An )
= d(A1 , . . . , Ai + Ak + (−1)Ai , . . . , (−1)Ai , . . . , An )
= d(A1 , . . . , Ak , . . . , (−1)Ai , . . . , An )
= (−1)d(A1 , . . . , Ak , . . . , Ai , . . . , An ),
where Axiom 1 is applied to the last equality, Axiom 2 is applied to the 1st and 4th
equality, and Theorem 2 is applied to the 2nd equality. 2
Theorem 4 If two rows of matrix A are equal, then d(A) = 0.
Proof. Since if we switch the two equal rows of A, the resulted matrix remains the same
as A but the determinant value must change sign by the previous theorem, we must have
d(A) = 0. 2
Note that Theorems 1–4 are consequences of Axioms 1 and 2 and are independent of Axiom
3.
Professor Chung-Chin Lu for EE203001 Linear Algebra 3

2 Gauss-Jordan Process and the Uniqueness of Deter-

minant Functions
In this section, we will show that a determinant function d of order n is unique if exists by
considering the effect of performing elementary row operations on an n×n matrix. Recall that
we perform three types of elementary row operations in Gauss-Jordan elimination process:
1. Interchanging two rows.
2. Multiplying a row by a nonzero scalar.
3. Adding to one row a scalar multiple of another.
By Theorem 3, a type 1 row operation will produce a sign change in the determinant of a
square matrix. By Axiom 1, a type 2 row operation will leave the determinant of a square
matrix unchanged and so does a type 3 row operation by Theorem 2.
Consider an upper triangular matrix
· · · u1n
 
u11 u12

 0 u22 · · · u2n 

U = .. .. .. . .

 . . . .. 

0 0 · · · unn
If unn is zero, then the last row of U is the zero vector and by Theorem 1, d(U ) = 0 =
u11 u22 · · · unn . If unn 6= 0, then by applying at most (n − 1) type 3 row operations, we have
···
 
u11 u12 0

 0 u22 ··· 0 

d(U ) = d( .. .. .. .. ).

 . . . .


0 0 · · · unn
If un−1,n−1 = 0, then again we have d(U ) = 0 = u11 u22 · · · unn . If not, then again by applying
at most (n − 2) type 3 row operations, we have
 
u11 u12 ··· 0 0
0 u22 ··· 0 0
 
 
.. .. .. .. ..
 
d(U ) = d( . ).
 
 . . . . 

 0 0 · · · un−1,n−1 0 

0 0 ··· 0 unn
Continuing this process, we either have some uii = 0 such that d(U ) = 0 = u11 u22 · · · unn or
have all uii 6= 0 such that
···
 
u11 0 0

 0 u22 ··· 0 

d(U ) = d( .. .. .. ) = (u11 u22 · · · unn ) d(In×n ). (1)
...
. . .
 
 
0 0 · · · unn
Professor Chung-Chin Lu for EE203001 Linear Algebra 4

In general, an n × n matrix A can be transformed to an upper triangular matrix U by

applying a sequence of elementary row operations. If there are p row exchanges and q scalar
multiplications by nonzero scalars c1 , c2 , . . . , cq included in the sequence, then we have
d(U ) = (−1)p (c1 c2 · · · cq )d(A),
i.e.,
d(A) = (−1)p (c1 c2 · · · cq )−1 d(U ) = (−1)p (c1 c2 · · · cq )−1 (u11 u22 · · · unn ) d(In×n ), (2)
by (1). Since Theorems 2 and 3 depend only on Axioms 1 and 2 and are independent of
Axiom 3, (2) is obtained from Axioms 1 and 2 and is independent of Axiom 3. Thus we
conclude that if f is a scalar-valued function of n × n matrices satisfying Axioms 1 and 2,
then we must have
f (A) = αf (In×n ), (3)
where the scalar α depends on the matrix A (and probably depends on the Gauss-Jordan
process proceeded as you might conceive). The following lemma gives a further characteri-
zation of (3) if there exists a determinant function d of order n (which in turn shows that
the scalar α in (3) depends only on the matrix A and is independent of the Gauss-Jordan
process proceeded).
Lemma 5 If f is another scalar-valued function on the space Mn×n of all n × n matrices
satisfying Axioms 1 and 2, then we have
f (A) = d(A)f (In×n )
for all A ∈ Mn×n .
Proof. Define a scalar-valued function on the space Mn×n as
g(A) = f (A) − d(A)f (In×n ).
It is easy to see that g satisfies both Axioms 1 and 2, but not Axiom 3. In fact, we have
g(In×n ) = f (In×n ) − d(In×n )f (In×n ) = f (In×n ) − f (In×n ) = 0,
since d(In×n ) = 1 by Axiom 3. But from (3), we have
g(A) = αg(In×n ) = 0
for all A. This completes the proof. 2
Now by (1) and Axiom 3, if d is a determinant function of order n, then
d(U ) = u11 u22 · · · unn
for any upper triangular matrix U with diagonal entries u11 , u22 , · · · , unn . Note that the
value d(U ) is invariant for any determinant function d of order n. In fact, this is true for
any n × n matrix A, as stated in the following uniqueness theorem.
Professor Chung-Chin Lu for EE203001 Linear Algebra 5

Theorem 6 [Uniqueness Theorem] If f is another scalar-valued function on the space Mn×n

of all n × n matrices satisfying all the three axioms, then we have

f (A) = d(A) ∀ A ∈ Mn×n .

Proof. By the above lemma, we have

f (A) = d(A)f (In×n ).

By Axiom 3, we have f (In×n ) = 1 and thus f (A) = d(A). 2

Although the above theorem states that a determinant function d is unique if exists, we
do not settle down the existence question yet. But for a 1 × 1 matrix A = [a11 ], the mapping

d(A) = a11

satisfies the three axioms and is the unique determinant function of order 1. It"is clear that
#
a11 a12
d is a linear function of the single row vector of A. And for a 2 × 2 matrix A = ,
a21 a22
it is easy to check that the following definition

d(A) = a11 a22 − a12 a21

satisfies the three axioms and thus gives the unique determinant function of order 2. When
we regard the determinant function d of order 2 as a scalar function of two 2-dimensional
row vectors, we have the following additive property in each row:

d(A1 + A01 , A2 ) = (a11 + a011 )a22 − (a12 + a012 )a21 = (a11 a22 − a12 a21 ) + (a011 a22 − a012 a21 )
= d(A1 , A2 ) + d(A01 , A2 )
0
d(A1 , A2 + A2 ) = a11 (a22 + a022 ) − a12 (a21 + a021 ) = (a11 a22 − a12 a21 ) + (a11 a022 − a12 a021 )
= d(A1 , A2 ) + d(A1 , A02 ).

Together with the homogeneity axiom in each row, we have

d(αA1 + βA01 , A2 ) = αd(A1 , A2 ) + βd(A01 , A2 )

d(A1 , αA2 + βA02 ) = αd(A1 , A2 ) + βd(A1 , A02 ),

for any scalars α, β, which says that d is a linear function of one row when the other row is
held fixed. We call d to be 2-linear. In the next section, we will show that if a determinant
function d of order (n − 1) exists and is (n − 1)-linear, then a determinant function d1 of
order n also exists and is n-linear. Thus by induction, there is a unique determinant function
d of order n for every n and d is n-linear.
1
While we have a little bit abused the notation, we are able to distinguish different determinant function
d for different order n in the context.
Professor Chung-Chin Lu for EE203001 Linear Algebra 6

3 Multilinearity and the Existence of a Determinant

Function of Order n
Assume that d is a determinant function of order n − 1 and is (n − 1)-linear. Let Aij be the
(n − 1) × (n − 1) matrix obtained from an n × n matrix A by deleting the ith row and the jth
column of A. This submatrix Aij of A is called the i, j minor of A. For each j, 1 ≤ j ≤ n,
consider the following scalar-valued function Ej as
n
(−1)i+j aij d(Aij ), ∀ A ∈ Mn×n .
X
Ej (A) =
i=1

Let B be the matrix obtained from matrix A by multiplying the kth row Ak of A by a
scalar α. Then we have
n
(−1)i+j bij d(Bij )
X
Ej (B) = Ej (A1 , . . . , αAk , . . . , An ) =
i=1
n
= (−1)k+j (αakj )d(Akj ) + (−1)i+j aij (αd(Aij ))
X

i=1
i6=k
n
(−1)i+j aij d(Aij ) = αEj (A1 , . . . , Ak , . . . , An ) = αEj (A),
X
= α
i=1

where (1) the k, j minor Bkj of B is the same as the k, j minor Akj of A and (2) for all
1 ≤ i ≤ n, i 6= k, the i, j minor Bij of B is obtained by multiplying the kth row or (k − 1)th
row (i > k or i < k) of the i, j minor Aij of A by scalar α. Thus the function Ej satisfies
Axiom 1.
Now let B be the matrix obtained from matrix A by adding the lth row to the kth row
of A. Then we have
n
(−1)i+j bij d(Bij )
X
Ej (B) = Ej (A1 , . . . , Ak + Al , . . . , Al , . . . , An ) =
i=1
n
= (−1)k+j (akj + alj )d(Akj ) + (−1)l+j alj (d(Alj ) + d(Âlj )) + (−1)i+j aij d(Aij )
X

i=1
i6=k,l
n
= (−1)k+j alj d(Akj ) + (−1)l+j+|k−l|−1 alj d(Akj ) + (−1)i+j aij d(Aij )
X

i=1
= Ej (A1 , . . . , Ak , . . . , Al , . . . , An ) = Ej (A),

where (1) the k, j minor Bkj of B is the same as the k, j minor Akj of A, (2) the l, j minor
(j)
Blj of B is obtained by adding the shortened (n − 1)-dimensional row vector Al (obtained
from the lth row vector Al of A by removing its jth component) to the shortened (n − 1)-
(j)
dimensional row vector Ak of the l, j minor Alj of A and then from the (n − 1)-linearity
Professor Chung-Chin Lu for EE203001 Linear Algebra 7

of d, d(Blj ) = d(Alj ) + d(Âlj ) with Âlj being the matrix obtained from Alj by replacing the
(j) (j)
shortened row vector Ak with the shortened row vector Al , and (3) for all 1 ≤ i ≤ n,
(j)
i 6= k, l, the i, j minor Bij of B is obtained by adding the shortened row vector Al of the
(j)
i, j minor Aij of A to the shortened row Ak of Aij . Note that matrix Âlj can be obtained
(j)
from the k, j minor Akj of A by exchanging the shortened row vector Al of Akj with the
adjacent row vector |k − l| − 1 times and then d(Âlj ) = (−1)|k−l|−1 d(Akj ). Thus the function
Ej satisfies Axiom 2.
For A = In×n , we have

Ej (In×n ) = (−1)j+j d((In×n )jj ) = d(I(n−1)×(n−1) ) = 1,

which says that Ej satisfies Axiom 3. Then by Theorem 6, Ej is the unique determinant
function of order n, also denoted as d, and
n
(−1)i+j aij d(Aij ), ∀ A ∈ Mn×n ,
X
d(A) = (4)
i=1

for any j, 1 ≤ j ≤ n. Now, for each k, 1 ≤ k ≤ n, we have from in (4)

d(A1 , . . . , Ak + A0k , . . . , An )
n
(j) (j) (j) 0(j)
= (−1)k+j (akj + a0kj )d(. . . , Ak−1 , Ak+1 , . . .) + (−1)i+j aij d(. . . , Ak + Ak , . . .)
X

i=1
i6=k
 
n
k+j (j) (j) (j)
(−1)i+j aij d(. . . , Ak , . . .)
 X 
= 
(−1) akj d(. . . , Ak−1 , Ak+1 , . . .) + 
i=1
i6=k
 
n
k+j 0 (j) (j) 0(j)
(−1)i+j aij d(. . . , Ak , . . .)
 X 
+
(−1) akj d(. . . , Ak−1 , Ak+1 , . . .) + 
i=1
i6=k
= d(A1 , . . . , Ak , . . . , An ) + d(A1 , . . . , A0k , . . . , An ),

which shows that the determinant function d of order n is n-linear.

Now we have proved that if a determinant function d of order (n−1) exists and is (n−1)-
linear, then a determinant function d of order n exists and is n-linear. Since a determinant
function d of order 2 exists and is 2-linear, we conclude by induction that for every n, there
exists a unique determinant function of order n and it is n-linear, as stated in the following
theorem.

Theorem 7 [Existence Theorem] For every positive integer n, there is a unique determinant
function of order n. Furthermore, this function is n-linear.
Professor Chung-Chin Lu for EE203001 Linear Algebra 8

This unique function is frequently denoted as det. The expansion formula in (4) is called
the expansion of the determinant det(A) of A by the jth column minors, rewritten here as
n
(−1)i+j aij det(Aij ).
X
det(A) = (5)
i=1

We next show that the determinant det(A) of matrix A can be expanded by minors in a row.
For each i, 1 ≤ i ≤ n, consider the ith row Ai of an n × n matrix A,

Ai = ai1 I1 + ai2 I2 + · · · + ain In ,

where I1 = (1, 0, . . . , 0), . . . , In = (0, . . . , 0, 1) are standard row vectors. Then we have
n
X
det(A) = d(A1 , . . . , Ai−1 , aij Ij , Ai+1 , . . . , An )
j=1
n
X
= aij d(A1 , . . . , Ai−1 , Ij , Ai+1 , . . . , An )
j=1
n
aij det(A0ij )
X
= (6)
j=1

by the n-linearity of d, where

··· · · · a1n
   
A1 a11 a1j
 .  .. ... .. .. ..
 .. 
 
  
 . . . . 

 Ai−1  ai−1,1 · · · ai−1,j · · · ai−1,n
   
 
A0ij
   
 Ij
= =
  0 ··· 1 ··· 1 .

 Ai+1
 
  ai+1,1

· · · ai+1,j · · · ai+1,n 

 . .. .. ..
 .
 
.. .. 
 .



 . . . . .


An an1 · · · anj · · · ann

By adding a multiple −akj Ij of the ith row Ij of A0ij to the kth row Ak of A0ij for each k,
1 ≤ k ≤ n and k 6= i, we obtain the matrix
··· 0 · · · a1n
 
a11
 .. .. .. . . .. 

 . . . . . 

 ai−1,1 ··· 0 · · · ai−1,n
 

A00ij
 
=
 0 ··· 1 ··· 1 

 ai+1,1

··· 0 · · · ai+1,n 

 .. ... .. . . .. 

 . . . .


an1 · · · 0 · · · ann

and by Theorem 2,
det(A00ij ) = det(A0ij ). (7)
Professor Chung-Chin Lu for EE203001 Linear Algebra 9

With the expansion of the determinant d(A00ij ) by the jth column minors in (5), we have
det(A00ij ) = (−1)i+j det(Aij ) (8)
since the i, j minor of A00ij is equal to the the i, j minor Aij of A. By (6)–(8), we have
n
(−1)i+j aij det(Aij ),
X
det(A) = (9)
j=1

which is called the expansion of the determinant det(A) of A by the ith row minors.
The next theorem is an application of the multilinearity of the determinant function det.
Theorem 8 If the rows of an n × n matrix A are linearly dependent, then det(A) = 0.
Proof. Suppose there exist scalars c1 , c2 , . . . , cn , not all zeros and says ck 6= 0, such that
Pn Pn
i=1 ci Ai = 0. Then we have Ak = i=1,i6=k ti Ai , where ti = −ci /ck . Thus
n
X
det(A) = d(. . . , Ak−1 , ti Ai , Ak+1 , . . .)
i=1
i6=k
n
X
= ti d(. . . , Ak−1 , Ai , Ak+1 , . . .).
i=1
i6=k

But for each i 6= k, row Ai is equal to at least one of the rows A1 , . . . , Ak−1 , Ak+1 , . . . , An
and hence
d(. . . , Ak−1 , Ai , Ak+1 , . . .) = 0
by Theorem 4. Thus det(A) = 0. 2
The next theorem is an application of the expansion formulas in (5) and (9).
Theorem 9 An n × n matrix A and its transpose At have the same determinants, i.e.,
det(A) = det(At ).
Proof. The proof is by induction on n, the size of the matrix A. For n = 1, we have
A = At = [a11 ] and det(A) = det(At ) = a11 trivially. Assume that the theorem is true for
(n − 1). Let A = [aij ] be an n × n matrix and B = At = [bij ]. Then bij = aji . By expanding
det(B) by the ith row minors, we have
n
(−1)i+j bij det(Bij ).
X
det(B) =
j=1

The i, j minor Bij of B is equal to the transpose of the j, i minor Aji of A, i.e., Bij = Atji .
And by the induction step, det(Aji ) = det(Atji ). Hence we have
n
(−1)j+i aji det(Aji ) = det(A)
X
det(B) =
j=1

from the expansion of det(A) by the ith column minors. We conclude that
det(A) = det(At ). 2
Professor Chung-Chin Lu for EE203001 Linear Algebra 10

Since the row vectors (At )i of At is just the transpose (ai )t of the column vectors ai of
A, we have

det(A) = det(At ) = d((At )1 , . . . , (At )i , . . . , (At )n ) = d((a1 )t , . . . , (ai )t , . . . , (an )t ) = d(a1 , . . . , ai , . . . , an )

which implies that we can regard the determinant function det as a scalar function of n n-
dimensional column vectors a1 , a2 , . . . an . Thus by Theorem 9, all the row properties of the
determinant function in Theorems 1–4 and 7–8 have the corresponding versions for columns
as stated in the following corollary.
Corollary 10 The determinant function det(A) = d(a1 , . . . , ai , . . . , an ) of an n × n matrix
A = [a1 a2 . . . an ], where a1 , a2 , . . . an are n column vectors of A, has the following column
properties:
1. It is n-linear with respect to column vectors, i.e., for each i, 1 ≤ i ≤ n,

d(a1 , . . . , αai + βa0i , . . . , an ) = αd(a1 , . . . , ai , . . . , an ) + βd(a1 , . . . , a0i , . . . , an ),

for any scalars α, β.

2. It is invariant under the addition of a multiple of a column of A to another column of
A, i.e., for any scalar α,

d(a1 , . . . , ai + αak , . . . , ak , . . . , an ) = d(a1 , . . . , ai , . . . , ak , . . . , an ).

3. It changes its sign under the exchange of two columns of A, i.e.,

d(a1 , . . . , ak , . . . , ai , . . . , an ) = (−1)d(a1 , . . . , ai , . . . , ak , . . . , an ).

And det(A) = 0 if A has linearly dependent columns. 2

4 The Cofactor Matrix and Cramer’s Rule

The scalar (−1)i+j det(Aij ) corresponding to the i, j minor Aij of an n × n matrix A is called
the cofactor of the i, j entry aij of A, denoted as cof aij , i.e.,

cof aij = (−1)i+j det(Aij ).

The n × n matrix
cof A = [cof aij ]ni,j=1
is called the cofactor matrix of A.
Consider the product A(cof A)t of A and the transpose of the cofactor matrix of A. The
i, j entry of the product is
n n
(A(cof A)t )ij = (−1)j+k aik det(Ajk ) = d(. . . , Aj−1 , Ai , Aj+1 , . . .)
X X
aik cof ajk =
k=1 k=1
Professor Chung-Chin Lu for EE203001 Linear Algebra 11

by the expansion formula along the jth row and is

(
det(A), if i = j,
(A(cof A)t )ij =
0, if i 6= j,
by Theorem 4. Thus we have
A(cof A)t = det(A)In×n .
Next Consider the product (cof A)t A. The i, j entry of this product is
n n
((cof A)t A)ij = (−1)k+i akj det(Aki ) = d(. . . , ai−1 , aj , ai+1 , . . .)
X X
cof aki akj =
k=1 k=1

by the expansion formula along the ith column and is

(
det(A), if i = j,
((cof A)t A)ij =
0, if i 6= j,
by Corollary 10. Thus we have
(cof A)t A = det(A)In×n .
We conclude in the following theorem.
Theorem 11 For any n × n matrix A with n ≥ 2, we have
A(cof A)t = (cof A)t A = det(A)In×n .
In particular, if det(A) 6= 0, then A is invertible and its inverse is
1
A−1 = (cof A)t .
det(A)
2
As an application of the above theorem, we consider a system of n linear equations in n
unknowns x1 , x2 , . . . , xn ,
· · · a1n
    
a11 a12 x1 b1

 a21 a22 · · · a2n 


x2 


 b2 

Ax =  .. .. . . . ..  
 .. = .. =b
. . .  . .
    
   
an1 an2 · · · ann xn bn
If det(A) 6= 0, then A is invertible and the above system of linear equations has a unique
solution
1
x = A−1 b = (cof A)t b,
det(A)
where
n n
1 X 1 X d(a1 , . . . , ai−1 , b, ai+1 , . . . , an )
xi = cof aji bj = (−1)j+i bj det(Aji ) =
det(A) j=1 det(A) j=1 d(a1 , . . . , ai−1 , ai , ai+1 , . . . , an )
(10)
for each i, 1 ≤ i ≤ n. Equation (10) is called Cramer’s rule.
Professor Chung-Chin Lu for EE203001 Linear Algebra 12

5 The Product Formula for Determinants

Consider the product AB of two n × n matrices A and B. It is easy to see that the ith row
(AB)i of AB is just the ith row Ai of A times B, i.e.,
(AB)i = Ai B, ∀i.
Thus we have
det(AB) = d(A1 B, A2 B, . . . , An B)
And by keeping B fixed, we can define a scalar-valued function f (A) = det(AB) of matrix
A and
f (A1 , A2 , . . . , An ) = d(A1 B, A2 B, . . . , An B).
Since for each i,
f (A1 , . . . , αAi , . . . , An ) = d(A1 B, . . . , αAi B, . . . , An B)
= αd(A1 B, . . . , Ai B, . . . , An B)
= αf (A1 , . . . , αAi , . . . , An ),
and for i 6= k,
f (A1 , . . . , Ai + Ak , . . . , Ak , . . . , An ) = d(A1 B, . . . , (Ai + Ak )B, . . . , Ak B, . . . , An B)
= d(A1 B, . . . , Ai B, . . . , Ak B, . . . , An B)
= f (A1 , . . . , Ai , . . . , Ak , . . . , An ),
function f satisfies Axioms 1 and 2 for a determinant function and by Lemma 5,
f (A) = det(A)f (In×n ).
Since
f (In×n ) = det(In×n B) = det(B),
we have the following theorem.
Theorem 12 [Product Formula] For any two square matrices A and B of the same size, we
have
det(AB) = det(A) det(B).
2
As an application of the above theorem, consider an invertible n × n matrix A, i.e.,
AA−1 = In×n .
By the product formula, we have
det(A) det(A−1 ) = det(In×n ) = 1.
Together with Theorem 11, we have the following theorem.
Professor Chung-Chin Lu for EE203001 Linear Algebra 13

Theorem 13 A square matrix A is invertible if and only if det(A) 6= 0. In this case, we

have
det(A−1 ) = det(A)−1 .
2
Another application of the product formula is to prove the converse of Theorem 8 and
the last sentence of Corollary 10.
Theorem 14 If an n×n matrix A has n linearly independent rows (columns), then det(A) 6=
0.

Proof. Firstly assume that the n columns a1 , a2 , . . . , an of A are linearly independent in

the linear space V of all n-dimensional column vectors. Since V is n-dimensional,
{a1 , a2 , . . . , an } is a basis of V and every n-dimensional column vector in V can be
expressed uniquely as a linear combination of a1 , a2 , . . . , an . In particular, for standard unit
column vectors ei , we have

e1 = β11 a1 + β21 a2 + · · · + βn1 an

e2 = β12 a1 + β22 a2 + · · · + βn2 an
.. .. ..
. . .
en = β1n a1 + β2n a2 + · · · + βnn an ,

i.e.,
· · · β1n
 
β11 β12

 β21 β22 · · · β2n 

[e1 e2 · · · en ] = [a1 a2 · · · an ]  .. .. .. . ,

 . . . .. 

βn1 βn2 . . . βnn
i.e.,
In×n = AB.
By the product formula, we have det(A) det(B) = det(In×n ) = 1 and then det(A) 6= 0.
Secondly assume that A has n linearly independent row vectors. Then At has n linearly
independent column vectors. As proved in above, det(At ) 6= 0. Thus we have det(A) 6= 0
since det(A) = det(At ). 2

Corollary 15 An n × n matrix A has n linearly independent rows (columns) if and only if

det(A) 6= 0. 2

Combined with Theorem 13 and Corollary 15, we have the following corollary.
Corollary 16 An n × n matrix A is invertible if and only if A has n linearly independent
rows (columns). 2
Note that Corollary 16 can be proved directly without the concept of determinants.