MTH6140 Linear Algebra II

Notes 3 28th October 2010

3 Linear maps between vector spaces

We return to the setting of vector spaces in order to define linear maps between them.
We will see that these maps can be represented by matrices, decide when two matrices
represent the same linear map, and give another proof of the canonical form for equiv-
alence.

3.1 Definition and basic properties

Definition 3.1 Let V and W be vector spaces over a field K. A function T from V to
W is a linear map if it preserves addition and scalar multiplication, that is, if

• T (v + v0 ) = T (v) + T (v0 ) for all v, v0 ∈ V ;

• T (cv) = cT (v) for all v ∈ V and c ∈ K.

Remark We can combine the two conditions into one as follows:

T (cv + c0 v0 ) = cT (v) + c0 T (v0 ).

Definition 3.2 Let T : V → W be a linear map. The image of T is the set

Im(T ) = {w ∈ W : w = T (v) for some v ∈ V },

and the kernel of T is

Ker(T ) = {v ∈ V : T (v) = 0}.

Proposition 3.1 Let T : V → W be a linear map. Then the image of T is a subspace

of W and the kernel is a subspace of V .

1
Proof We have to show that both Im(T ) and Ker(T ) are closed under addition and
scalar multiplication. For the image, if w, w0 ∈ Im(T ) then we can find v, v0 ∈ V such
that w = T (v) and w0 = T (v0 ). Then
w + w0 = T (v) + T (v0 ) = T (v + v0 ) ∈ Im(T ) ,
and if c ∈ K then
cw = cT (v) = T (cv) ,
so Im(T ) is indeed a subspace of W .
For the kernel, if v, v0 ∈ Ker(T ) then T (v) = T (v0 ) = 0. Therefore
T (v + v0 ) = T (v) + T (v0 ) = 0 + 0 = 0,
so v + v0 ∈ Ker(T ). Also, if c ∈ K then
T (cv) = cT (v) = c0 = 0 ,
so cv ∈ Ker(T ).
Definition 3.3 We define the rank of T to be rank(T ) = dim(Im(T )) and the nullity
of T to be nul(T ) = dim(Ker(T )). (Recall that in Chapter 2 we defined the rank of a
matrix. This will turn out to be closely related to the rank of a linear map!)
Theorem 3.2 (Rank–Nullity Theorem) Let T : V → W be a linear map. Then rank(T )+
nul(T ) = dim(V ).
Proof Choose a basis u1 , u2 , . . . , uq for Ker(T ), where q = dim(Ker(T )) = nul(T ).
The vectors u1 , . . . , uq are linearly independent vectors of V , so by Corollary 1.3(b)
we can add further vectors to get a basis for V , say u1 , . . . , uq , v1 , . . . , vs , where q + s =
dim(V ).
We claim that the vectors T (v1 ), . . . , T (vs ) form a basis for Im(T ). We have to
show that they are linearly independent and spanning.
Linearly independent: Suppose that c1 T (v1 ) + · · · + cs T (vs ) = 0. Then T (c1 v1 +
· · · + cs vs ) = 0, so that c1 v1 + · · · + cs vs ∈ Ker(T ). But then this vector can be
expressed in terms of the basis for Ker(T ):
c1 v1 + · · · + cs vs = a1 u1 + · · · + aq uq ,
whence
−a1 u1 − · · · − aq uq + c1 v1 + · · · + cs vs = 0.
But the us and vs form a basis for V , so they are linearly independent. So this
equation implies that all the as and cs are zero. The fact that c1 = · · · = cs = 0
shows that the vectors T (v1 ), . . . , T (vs ) are linearly independent.

2
 Spanning: Take any vector in Im(T ), say w. Then w = T (v) for some v ∈ V . Write v
in terms of the basis for V :
v = a1 u1 + · · · + aq uq + c1 v1 + · · · + cs vs
for some a1 , . . . , aq , c1 , . . . , cs . Applying T , we get
w = T (v)
= a1 T (u1 ) + · · · + aq T (uq ) + c1 T (v1 ) + · · · + cs T (vs )
= c1 w1 + · · · + cs ws ,
since T (ui ) = 0 (as ui ∈ Ker(T )) and T (vi ) = wi . So the vectors w1 , . . . , ws span
Im(T ).
Thus, rank(T ) = dim(Im(T )) = s. Since nul(T ) = q and q + s = dim(V ), the
theorem is proved.

3.2 Representation by matrices

We come now to the second role of matrices in linear algebra: they represent linear
maps between vector spaces.
Let T : V → W be a linear map, where dim(V ) = m and dim(W ) = n. As we saw
in Chapter 1, we can take V and W in their coordinate representation: V = K m and
W = K n (the elements of these vector spaces being represented as column vectors).
Let v1 , . . . , vm be the standard basis for V (so that vi is the vector with ith coordinate 1
and all other coordinates zero), and w1 , . . . , wn the standard basis for W . Then for
i = 1, . . . , m, the vector T (vi ) belongs to W , so we can write it as a linear combination
of w1 , . . . , wn .
Definition 3.4 The matrix representing the linear map T : V → W relative to the bases
B = (v1 , . . . , vm ) for V and C = (w1 , . . . , wn ) for W is defined to be the n × m matrix A
whose (i, j) entry is Ai j , where
n
T (vi ) = ∑ A jiw j
j=1

for i = 1, . . . , n.
In practice this means the following. Take T (vi ) and write it as a column vector
 
A1i
 A2i 
 . .
 .. 
Ani

3
This vector is the ith column of the matrix representing T . So, for example, if m = 3,
n = 2, and

T (v1 ) = w1 + w2 , T (v2 ) = 2w1 + 5w2 , T (v3 ) = 3w1 − w2 ,

then the vectors T (vi ) as column vectors are

     
1 2 3
T (v1 ) = , T (v2 ) = , T (v3 ) = ,
1 5 −1
and so the matrix representing T is
 
1 2 3
.
1 5 −1
The most important thing about this representation is that the action of T is now
easily described:

Proposition 3.3 Let T : V → W be a linear map. Choose bases for V and W and let
A be the matrix representing T . Then, if we represent vectors of V and W as column
vectors relative to these bases, we have

T (v) = Av.

Proof Let v1 , . . . , vm be the basis for V , and w1 , . . . , wn for W . Take v = ∑m

i=1 ci vi ∈ V ,
so that in coordinates  
c1
.
v =  ..  .
cm
Then
!
m m m n n m
T (v) = T ∑ civi = ∑ ci T (vi ) = ∑ ci ∑ A ji w j = ∑ ∑ ciA jiw j ,
i=1 i=1 i=1 j=1 j=1 i=1

so the jth coordinate of T (v) is ∑m

i=1 A ji ci , which is precisely the jth coordinate in the
matrix product Av.
 
2
In our example, if v = 2v1 + 3v2 + 4v3 = 3 , then

4
 
  2  
1 2 3   20
T (v) = Av = 3 = .
1 5 −1 13
4

4
 One of the important things about this representation is that addition and compo-
sition of linear maps correspond to addition and multiplication of the matrices repre-
senting them.

Definition 3.5 Let S and T be linear maps from V to W . Define their sum S + T by
the rule
(S + T )(v) = S(v) + T (v)
for all v ∈ V . It is easy to check that S + T is a linear map.

Proposition 3.4 If S and T are linear maps represented by matrices A and B respec-
tively, then S + T is represented by the matrix A + B.

The proof of this is not difficult: just use the definitions.

Definition 3.6 Let U,V,W be vector spaces over K, and let S : U → V and T : V → W
be linear maps. The composition T ◦ S (also sometimes denoted by T S) is the function
U → W defined by the rule
(T ◦ S)(u) = T (S(u))
for all u ∈ U. Again it is easily checked that T ◦ S is a linear map. Note that the order
is important: we take a vector u ∈ U, apply S to it to get a vector in V , and then apply
T to get a vector in W . So T ◦ S means “apply S, then T ”.

Proposition 3.5 If S : U → V and T : V → W are linear maps represented by matrices

A and B respectively, then T ◦ S is represented by the matrix BA.

Again the proof is tedious but not difficult.

Remark Let l = dim(U), m = dim(V ) and n = dim(W ), then A is m × l, and B is

n × m; so the product BA is defined, and is n × l, which is the right size for a matrix
representing a map from an l-dimensional to an n-dimensional space.
The significance of all this is that the strange rule for multiplying matrices is cho-
sen so as to make Proposition 3.5 hold. We could say: what definition of matrix
multiplication should we choose to make the Proposition valid? We would find that
the usual definition was forced upon us.

5
3.3 Change of basis
The matrix representing a linear map depends on the choice of bases we used to rep-
resent it. Now we have to discuss what happens if we change the basis.
Remember the notion of transition matrix from Chapter 1. If B = (v1 , . . . , vm ) and
B0 = (v01 , . . . , v0m ) are two bases for a vector space V , the transition matrix PB,B0 is the
matrix whose jth column is the coordinate representation of v0j in the basis B. Then
(see Proposition 1.5) we have
[v]B = P[v]B0 ,
where [v]B is the coordinate representation of an arbitrary vector in the basis B, and
similarly for B0 . Recall that the inverse of PB,B0 is PB0 ,B (see Corollary 1.6 (b)). Let pi j
denote the (i, j) entry of P = PB,B0 .
Now let C = (w1 , . . . , wn ) and C0 = (w01 , . . . , w0n ) be two different bases for the same
vector space W , with transition matrix QC,C0 and inverse QC0 ,C . Let Q = QC,C0 and let
R = QC0 ,C be its inverse, with (i, j) entry denoted by ri j .
Let T be a linear map from V to W . Then T is represented by a matrix A using
the bases B and C, and by a matrix A0 using the bases B0 and C0 . What is the relation
between A and A0 ?
We just do it and see. To get A0 , we have to represent the vectors T (v0i ) in the basis
C0 . We have
m
v0j = ∑ pi j vi ,
i=1
so
m
T (v0j ) = ∑ pi j T (vi )
i=1
m m
= ∑∑ pi j Aki wk
i=1 k=1
m n n
= ∑ ∑ ∑ pi j Akirlk w0l .
i=1 k=1 l=1

This means, on turning things around, that

n m
0
(A )l j = ∑ ∑ rlk Aki pi j ,
k=1 i=1
so, according to the rules of matrix multiplication,
A0 = RAP = Q−1 AP.
So we have proved the following result:

6
Proposition 3.6 Let T : V → W be a linear map represented by matrix A relative to
the bases B for V and C for W , and by the matrix A0 relative to the bases B0 for V and
C0 for W . If P = PB,B0 and Q = PC,C0 are the transition matrices from the unprimed to
the primed bases, then
A0 = Q−1 AP.

This is rather technical; you need it for explicit calculations, but for theoretical
purposes the importance is the following corollary. Recall (see Definition 2.5) that
two matrices A and B are equivalent if B is obtained from A by multiplying on the left
and right by invertible matrices.

Proposition 3.7 Two matrices represent the same linear map with respect to different
bases if and only if they are equivalent.

This holds because transition matrices are always invertible, and any invertible
matrix can be regarded as a transition matrix.

3.4 Canonical form revisited

Now we can give a simpler proof of Theorem 2.3 about canonical form for equiva-
lence. First, we make the following observation.

Theorem 3.8 Let T : V → W be any linear map. Then there are bases for V and W
such that the matrix representing T is, in block form,
 
Ir O
,
O O

where r = rank(T ) is the rank of T .

Proof As in the proof of Theorem 3.2, choose a basis u1 , . . . , us for Ker(T ), and extend
to a basis u1 , . . . , us , v1 , . . . , vr for V . Then T (v1 ), . . . , T (vr ) is a basis for Im(T ), and
so can be extended to a basis T (v1 ), . . . , T (vr ), x1 , . . . , xt for W . Now we will use the
bases

v1 , . . . , vr , vr+1 = u1 , . . . , vr+s = us for V,

w1 = T (v1 ), . . . , wr = T (vr ), wr+1 = x1 , . . . , wr+s = xs for W.

We have 
wi if 1 ≤ i ≤ r,
T (vi ) =
0 otherwise;

7
so the matrix of T relative to these bases is
 
Ir O
O O

as claimed.

We recognise the matrix in the theorem as the canonical form for equivalence.
Combining Theorem 3.8 with Proposition 3.7, we see:

Theorem 3.9 A matrix of rank r is equivalent to the matrix

 
Ir O
.
O O

We also see, by the way, that the rank of a linear map (that is, the dimension of its
image) is equal to the rank of any matrix which represents it. So all our definitions of
rank agree!
The conclusion is that

two matrices are equivalent if and only if they have the same rank.

So how many equivalence classes of m × n matrices are there, for given m and n? The
rank of such a matrix can take any value from 0 up to the minimum of m and n; so the
number of equivalence classes is min{m, n} + 1.