February 2, 2003

45

Chapter 3

Linear transformations and

matrices
3.1

Linear transformations

3.1.1

Definitions and examples

In this section, we will define the notion of a linear transformation between

two vector spaces V and W which are defined over the same field. First
of all, a transformation, or vector valued function, T : V W is simply a
function or mapping which has domain V and takes its values in W . Linear
transformations a special class of transformations which are closely related
to matrices. Hence we could have chosen to study matrices first. However,
linear transformations are conceptually more natural than matrices, and
introducing them first will enable us to see how certain definitions involving
matrices and matrix multiplication came about.
Suppose F is a transformation between two vector spaces V and W . The
notation for such a transformation is F : V W . By definition, F assigns
to each vector x V a unique vector F (x) = y W . Note that we will
usually call W the target of F . A transformation F is sometimes called
a vector valued function. A transformation F : Fn Fm is completely
determined by its component functions. That is, fi : Fn F. Since the
target of F is Fm , there are m of these. Thus we write
F (x1 , x2 , . . . , xn ) = (f1 (x1 , x2 , . . . , xn ), f2 (x1 , x2 , . . . , xn ),
. . . , fm (x1 , x2 , . . . , xn )).

46
The transformations which are considered in linear algebra are those
which preserve linear combinations. These are called linear transformations.
Definition 3.1. Suppose V and W are vector spaces over a field F. Then
a transformation T : V W is called linear if
(1) for all x, y V , T (x + y) = T (x) + T (y), and
(2) for all r F and all x V , T (rx) = rT (x).
The statement that a linear transformation T preserves linear combinations means that for all r, s F and all x, y V
T (rx + sy) = rT (x) + sT (y).
An obvious property is that every linear transformation maps the zero vector
in the domain to the zero vector in the target. This follows from the fact
that
T (x) = T (x + 0) = T (x) + T (0)
for any vector x. This can only happen if T (0) = 0.
Example 3.1. Here are some examples of transformations, not necessarily
linear.
(1) (Polar coordinates) Let (r, ) denote rectangular coordinates on the
domain and let (x, y) denote coordinates on the target. Then the transformation F : R2 R2 given by
F (r, ) = (r cos , r sin )
relates rectangular coordinates and polar coordinates. It should be clear
that F is not a linear transformation.
(2) Let a R3 be nonzero. Recall that the transformation
Pa (x) =

ax
a
aa

is called the projection on the line spanned by a. One of the exercises in 2

asked you to show that Pa is linear. This is proved as follows.
Pa (x + y) =

a (x + y)
a x + a y

a=
a = Pa (x) + Pa (y).
aa
aa

In addition, for any scalar r,

Pa (rx) =

a (rx)
a x

a=r
a = rPa (x).
aa
aa

47
This verifies the linearity of any projection. Using the above formula, we
get an explicit expression for Pa . Let a = (a1 , a2 ) and x = (x1 , x2 ). Then
Pa (x1 , x2 ) = (

a1 x1 + a2 x2
a1 x1 + a2 x2
.
)a
,
(
)a
1
2
a21 + a22
a21 + a22

The thing to notice is that each component is linear in x1 and x2 .

(3) Let V be any vector space. Then the identity transformation is the
transformation F : V V defined by F (x) = x. When V = Fn , we denote
the identity transformation by In .
(4) If we choose a vector a Rn , the dot product with a defines a linear
transformation Ta : Rn R by Ta (x) = a x. The linearity follows from
the properties of the dot product. We will consider Ta further below.
(5) We can add and subtract transformations to get new transformations.
For example, let Pa : R3 R3 be the projection of part (1). Then Q =
I3 Pa : R3 R3 turns out to be the projection of R3 onto the plane P in
R3 through the origin orthogonal to a. We will say more about this later in
the section.
(6) There are many natural examples of linear transformations involving
abstract vector spaces. To take one, suppose the domain is C(a, b), the
space of continuous real valued functions on [a, b]. Then the definite integral
defines a linear transformation
Z

: C(a, b) R

Rb
by the rule f 7 a f (t)dt. The assertion that this is a linear transformation
is just the fact that for all r, s R and f, g C(a, b),
Z

(rf + sg)(t)dt = r
a

f (t)dt + s
a

g(t)dt.
a

The fact that the definite integral defines a linear transformation is usually
very useful for computing integrals. You may have already realized that this
example is nothing but the analogue for C(a, b) of the linear transformation
Ta on Rn defined just above, where a = 1. In fact, by definition,
Z

f (t)dt = (f, 1).

a

48

3.1.2

The geometry of linear transformations from R2 to R2

Linear transformations with domain and target R2 have a rich geometry.

Suppose T : R2 R2 is linear, and let e1 = (1, 0) and e2 = (0, 1). We
will call e1 , e2 the standard basis of R2 . The term basis will be explained in
a later section. Lets begin by finding an expression for the value of T on
(x, y). Strictly speaking, we should express this value as T ((x, y)), but for
simplicity, we will just express it as T (x, y). Since (x, y) = xe1 + ye2 and T
preserves linear combinations, then we see that
T (x, y) = T (x(1, 0) + y(0, 1))
= xT (1, 0) + yT (0, 1)
= xT (e1 ) + yT (e2 ).
The important point this calculation demonstrates is that all values of T
are uniquely determined once we know the particular values T (e1 ), T (e2 ).
This is one of the remarkable properties of linear transformations. The
property of being determined by only two values will never hold for arbitrary
non-linear transformations. And because of this, we can see how a linear
transformation acts. For example, condition (2) says that T sends each one
of the coordinate axes Rei to the line RT (ei ), unless T (ei ) = 0, in which case
the whole line is mapped to the origin instead of a proper line. In general,
the line spanned by x is either mapped to the line spanned by T (x) or to
0. In addition, if T (e1 ) and T (e2 ) are non collinear, then T transforms the
square S determined by e1 and e2 into a parallelogram, namely the unique
parallelogram with edges T (e1 ) and T (e2 ). Indeed, this parallelogram is
P = {rT (e1 ) + sT (e2 ) | 0 r, s 1},
and since T (re1 + se2 ) = rT (e1 ) + sT (e2 ), T clearly sends S to P. More
generally, T sends the parallelogram with sides x and y to the parallelogram with sides T (x) and T (y), provided T (x) and T (y) are not collinear
(Exercise 8).
On the other hand, projections are linear transformations which dont
send parallelograms to parallelograms due to the fact that any two values
Pb (x) and Pb (y) are multiples of b, hence always collinear. Nevertheless,
projections are still very useful and have an interesting geometry in their
own right. Namely, each vector on the line spanned by b is preserved by
Pb , and every vector orthogonal to b is mapped by Pb to 0.

49

3.1.3

Orthogonal transformations: reflections and rotations

When you look in a mirror, what you see is your reflection through the plane
of the mirror. This is an example of a transformation called a reflection. Let
us analyze reflections, starting in the case of the plane. Suppose we consider
a line ` in R2 which passes through the origin. The the reflection of R2
through ` acts as follows: every point on ` is left fixed, and the points on
the line through the origin orthogonal to ` are sent to their negatives.
Perhaps somewhat surprisingly, every reflection is linear. We can see
this fact by deriving a formula. Let b be a vector orthogonal to `, and let
us denote the reflection through ` by Hb . Choose an arbitrary v R2 , and
consider its orthogonal decomposition (2)
v = Pb (v) + c
with c on `. Using the parallelogram law, one sees that
Hb (v) = c Pb (v).

FIGURE (diagram for Hb (v))

Replacing c by v Pb (v), we get the formula
Hb (v) = v 2Pb (v)
v b

= v2
b.
bb
b determined by b
b gives us the
Expressing this in terms of the unit vector b
simpler expression
b b.
b
Hb (v) = v 2(v b)

(3.1)

Clearly Hb has the right properties: Hb (v) = v if b v = 0, and Hb (b) =

b.
Notice also that our reflection Hb can be expressed as I2 2Pb . Indeed,
Hb (v) = I2 (v) 2Pb (v).
Thus every reflection is a linear combination of two linear transformations,
the identity and a projection.

50
The above discussion of reflections goes through not just for R2 but for
R3 as well. More generally, let b be any nonzero vector in Rn , and let W
be the hyperplane in Rn consisting of all vectors orthogonal to b. Then the
transformation H : Rn Rn defined by (3.1) is the reflection of Rn through
W.
Now lets compute some reflections of R2 .
Example 3.2. Let b = (1, 1). Then Hb is the reflection through the line
x = y. We have
H(1,1) (a, b) = (a, b) 2((a, b) (1, 1)/2)(1, 1)
= (a (a + b), b (a + b))
= (b, a).
Similarly (or by inspection), the reflection through the x and y axes are
H(0,1) (a, b) = (a, b)
and
H(1,0) (a, b) = (a, b).
There are several consequences of formula (3.1) that will be left as an
exercise. First of all, every reflection is linear. Secondly, reflecting any
v R2 twice returns v to itself, i.e. Hb Hb = I2 .
Another property of reflections is that they preserve inner products.
That is, for all v, w R2 ,
Hb (v) Hb (w) = v w.
We will also leave this as an exercise. A consequence of this property is that
since we are able to express lengths, distances and angles between vectors
in terms of the dot product, reflections preserve all these quantites. In
other words, a vector and its reflection have the same length, and the angle
(measured with respect to the origin) between a vector and the reflecting
line is the same as the angle between the reflection and the reflecting line.
In general, we say that a linear transformation T : Rn Rn is orthogonal if
it preserves the dot product. So T is orthogonal iff
T (v) T (w) = v w
for all v, w Rn .

51
Proposition 3.1. If a linear transformation T : Rn Rn is orthogonal,
then for any v Rn , |T (v)| = |v|. In particular, if v 6= 0, then T (v) 6= 0.
Moreover, the angle between any two nonzero vectors v, w Rn is the same
as the angle between the vectors T (v) and T (w), which are both nonzero
by the last assertion.
We leave the proof as an easy exercise.
Example 3.3. Besides reflections, another class of examples of orthogonal
linear transformations of R2 are the rotations. Let R : R2 R2 stand for
the counter-clockwise rotation of R2 through . It is clear that R should
have the properties that
R (e1 ) = cos e1 + sin e2 ,
and
R (e2 ) = sin e1 + cos e2 .
Granting that a rotation is linear (which will be proven shortly), then
R (x, y) = xR (e1 ) + yR (e2 ),
so
R (x, y) = (x cos y sin , x sin + y cos ).
DIAGRAM
(ROTATIONS)
The following proposition asserts a surprising property of orthogonal
linear transformations of R2 .
Proposition 3.2. Any orthogonal linear transformation of R2 is either a
reflection or a rotation.
This will be proved in 3.4. The motivated reader may wish to go ahead
and prove it on their own. The structure of orthogonal transformations in
higher dimensions is more complicated. For example, the rotations and reflections of R3 do not give all the possible orthogonal linear transformations
of R3 .

3.1.4

Some further examples

Let us next consider a class of transformations which we will encounter again

when we study eigentheory.

52
Example 3.4. A linear transformation T : R2 R2 of the form T (x, y) =
(x, y), where and are scalars, will be called a diagonal transformation.
Since T stretches e1 by a factor of and stretches e2 by a factor of , it
maps a rectangle with sides parallel to e1 and e2 another such rectangle
whose sides have been dilated by and and whose area has been changed
by the factor ||. Diagonal transformations also map circles to ellipses. For
example, let C denote the unit circle x2 + y 2 = 1. Put w = x and z = y.
Assuming both and are nonzero, then if x2 + y 2 = 1, then
w
z
( )2 + ( )2 = 1.

Thus, for every point (x, y) on C, T (x, y) lies on the ellipse. To see that
T (C) is actually the whole ellipse, observe that any point (w, z) on the
ellipse comes from a point (x, y) on the circle. Indeed, just set x = w and
y = z . More generally, we can call a linear transformation diagonalizable if
there exist a pair of noncollinear vectors v, w R2 for which T (v) = v and
T (w) = w for some , R. It will turn out a big problem we will take
up later is how to determine when a linear transformation is diagonalizable.
We wont be able to solve this until we take up eigentheory.
FIGURE
(DIAGONAL TRANSFORMATION)
Example 3.5. A pretty example of a linear transformation on R3 is given
by the cross product. Let a R3 and define Ca : R3 R3 by
Ca (v) = a v.
Notice that Ca (a) = 0, and that Ca (x) is orthogonal to a for any x. In
particular, Ca cannot possibly be orthogonal. The transformation Ca is
used in mechanics to express angular momentum.

3.1.5

The algebra of linear transformations

As indicated in our first example, one can form linear combinations of linear transformations, as long as they have the same domain and target. In
fact, one can form linear combinations of arbitrary transformations with
the same domain and target. The sum of two such transformations is just
the usual pointwise addition of real valued functions discussed above. The
only difference is that now the target is Rm instead of R. For example, if
S, T : Rn Rm are linear and a, b R, then
(aS + bT ) : Rn Rm

53
is defined at x Rn by
(aS + bT )(x) := aS(x) + bT (x).
By the next result, aS + bT is also linear.
Proposition 3.3. Any linear combination of linear transformations from
Rn to Rm is linear. In fact, the set of all linear transformations T : Rn Rm
is a vector space over R.
We leave the proof as an exercise. This gives another proof of the fact
that reflections are linear.
Definition 3.2. We will denote the space of all linear transformations T :
Rn Rm by L(Rn , Rm ).
A linear transformation T : Rn R is called a linear function. If a is
a scalar, then the function Ta (x) := ax is a linear function T : R R. In
fact, every linear function of one variable has this form. The n-dimensional
generalization of this linear function uses the dot product. Recall that for a
given a Rn , we put
Ta (x) = a x =

n
X

ai xi .

i=1

Then Ta : Rn R is a linear function, and, in fact, it isnt hard to see that

every that every linear function T : Rn R is a Ta for some a Rn . That
is, there exist a1 , a2 , . . . , an R such that
T (x1 , x2 , . . . , xn ) =

n
X

ai xi .

i=1

Example 3.6. Let a = (1, 2, 0, 1). Then the linear function Ta : R4 R

has the explicit form
Ta (x1 , x2 , x3 , x4 ) = x1 + 2x2 + x4 .
Notice that the right hand side contains only terms of degree 1 in the variables x1 , x2 , x3 , x4 .
To test whether a transformation Ta : Rn Rm is linear, the idea is to
apply the criterion in the previous example to every component of T as in
the following:

54
Proposition 3.4. Suppose T : Rn Rm is an arbitrary transformation
and write
T (v) = (f1 (v), f2 (v), . . . , fm (v)).
Then T is linear if and only if each component fi is a linear function. In
particular, T is linear if and only if there exist a1 , a2 , . . . , am in Rn such that
for all v Rn ,
T (v) = (a1 v, a2 v, . . . , am v).
Note that each ai is chosen so that fi (v) = ai v if 1 i m. We will
leave the proof as an exercise.
Example 3.7. Consider the transformation T : R2 R2 given by
T (x, y) = (2x + y, x y).
Then each component of T is linear, so T is linear.

3.1.6

Gradients and differentials

Since arbitrary transformations can be very complicated, we should view

linear transformations as one of the simplest are types of transformations.
In fact, we can make a much more precise statement about this. One of the
most useful principals about smooth transformations is that no matter how
complicated such transformations are, they admit linear approximations,
which means that one certain information may be obtained by constructing
a taking partial derivatives. Suppose we consider a transformation F : Rn
Rm such that each component function fi of F has continuous first partial
derivatives throughout Rn , that is F is smooth. Then it turns out that in
a sense which can be made precise, the differentials of the components fi
are the best linear approximations to the fi . Recall that if f : Rn R
is a smooth function, the differential df (x) of f at x is the linear function
df (x) : Rn R whose value at v Rn is
df (x)v = f (x) v.
Here,
f
f
(x), . . . ,
(x)) Rn
x1
xn
is the called the gradient of f at x. In other words,
f (x) = (

df (x)v =

n
X
f
(x)vi ,
xi
i=1

55
so the differential is the linear transformation induced by the gradient and
the dot product. Note that in the above formula, x is not a variable. It
represents the point at which the differential of f is being computed.
The differential of the transformation F at x is the linear function
DF (x) : Rn Rn defined by DF (x) = (df1 (x), df2 (x), . . . , dfm (x)). The
components of DF at x are the differentials of the components of F at x.
We will have to leave further discussion of the differential for a course in
vector analysis.

56
Exercises
Exercise 3.1. Show that every linear function T : R R has the form
T (x) = ax for some a R.
Exercise 3.2. Determine whether the following are linear or not:
(i) f (x1 , x2 ) = x2 x2 .
(ii) g(x1 , x2 ) = x1 x2 .
(iii) f (x) = ex .
Exercise 3.3. Verify from the formula that the projection Pb fixes every
vector on the line spanned by b and sends every vector orthogonal to b to
0.
2
Exercise 3.4. Let Hb : R2 R2 be the reflection

of R through the line
vb
orthogonal to b. Recall that Hb (v) = v 2 bb b.

(i) Use this formula to show that every reflection is linear.

(ii) Show also that Hb (Hb (x)) = x.
(iii) Find formulas for Hb ((1, 0)) and Hb ((0, 1)).
Exercise 3.5. Consider the transformation Ca : R3 R3 defined by
Ca (v) = a v.
(i) Show that Ca is linear.
(ii) Descibe the set of vectors x such that Ca (x) = 0.
Exercise 3.6. Let u and v be two orthogonal unit length vectors in R2 .
Show that
(a) Pu + Pv = I2 , and
(b) Pu Pv = Pv Pu = 0.
Conclude from (a) that x = (x u)u + (x v)v, for all x R2 .
Exercise 3.7. Suppose T : R2 R2 is a linear transformation which sends
any two non collinear vectors to non collinear vectors. Suppose x and y
in R2 are non collinear. Show that T sends any parallelogram with sides
parallel to x and y to another parallelogram with sides parallel to T (x) and
T (y).

57
Exercise 3.8. Show that reflections are orthogonal linear transformations.
In other words, show that for all
x and y in R2 ,
Hb (x) Hb (y) = x y.
Exercise 3.9. Show that rotations R of R2 also give orthogonal linear
transformations.
Exercise 3.10. Show that every orthogonal linear transformation not only
preserves dot products, but also lengths of vectors and angles and distances
between two distinct vectors. Do reflections and rotations preserve lengths
and angles and distances?
Exercise 3.11. Suppose F : Rn Rn is a transformation with the property
that for all x, y Rn ,
F (x) F (y) = x y.
(a) Show that for all x, y Rn , ||F (x + y) F (x) F (y)||2 = 0.
(b) Show similarly that for all x Rn and r R, ||F (rx) rF (x)||2 = 0
Conclude that F is in fact linear. Hence F is an orthogonal linear transformation.
Exercise 3.12. How would you define the reflection H of R3 through the
plane ax + by + cz = 0. (Hint: imitate the construction for R2 .)
Exercise 3.13. Find the reflection of R3 through the plane P if:
(a) P is the plane x + y + z = 0; and
(b) P is the plane 3x + 4y = 0.
Exercise 3.14. Find a formula for the composition of two rotations. That
is, compute R R in terms of sines and cosines. Give an interpretation of
the result.
Exercise 3.15. Prove Proposition 3.4.
Exercise 3.16. * Let f (x1 , x2 ) = x21 + 2x22 .
(a) Find both the gradient and differential of f at (1, 2).
(b) If u R2 is a unit vector, then df (1, 2)u is called the directional
derivative of f at (1, 2) in the direction u. Find the direction u R2 which
maximizes the value of df (1, 2)u.
(c) What has your answer in part (b) got to do with the length of the
gradient of f at (1, 2)?

58
Exercise 3.17. Consider the transformation
T : C C
z 7 z .
(i) Is T linear as a transformation of C-vector spaces?
(ii) Is T linear as a transformation of R-vector spaces R2 R2 ?
Justify your answers.
Exercise 3.18. Let V = C as in the previous exercise and consider the
transformation R : V V defined by R(z) = ei z. Interpret R as a
transformation from R2 to R2 .
Exercise 3.19. *. Find the differential at any (x1 , x2 ) of the polar coordinate map of Example 3.8.

59

3.2

Matrices

In this section, we will introduce matrices and study their algebra. Linear
transformations and matrices turn out to be two sides of the same coin.
However, matrices also have a myriad of other uses which dont seem to
involve linear transformations at all. A matrix is like a chessboard, or, more
accurately, a table. It turns out to be an extremely useful device for storing
certain types of information, such as the data in a digit photo. Whereas we
usually only considered linear transformations in the setting of Rn , in order
to emphasize their geometry, we will now broaden the scope and consider
matrices over an arbitrary field F.

3.2.1

The notion of a matrix

Definition 3.3. An m n matrix over a field F is an array of elements of

F arranged in a grid with m rows and n columns.
For example, if one puts a real number on each square of a chessboard,
one gets an 8 8 matrix over R. An m n matrix will obviously have mn
entries. We will denote the entry of A in the ith row and jth column by aij
and will write A = (aij ). For example, a 2 3 matrix A will be written as


a11 a12 a13
.
a21 a22 a23
Matrices with only one row are called row matrices or row vectors, and
those with just one column are called column matrices or column vectors.
For convenience, if A is m n, we will say that A has dimension m by n.
There are three basic matrix operations: scalar multiplication, addition
and multiplication. There are conditions as to when two matrices can be
added and when they can be multiplied. In particular, one cannot add or
multiply any two matrices. First of all, if A = (aij ) is a matrix over F and
r F, then the scalar multiple rA of A is the matrix rA = (raij ) in which
every element of A has been multiplied by r. For example, if A is the 2 3
matrix above, then


3a11 3a12 3a13
3A =
.
3a21 3a22 3a23
Matrix addition can only be carried out on matrices of the same dimension. (This is for the same reason we dont attempt to add a row vector
to a column vector.) However, when A and B are two matrices of the
same dimension over the same field, then we take their sum in the obvious

60
manner. Let A = (aij ) and B = (bij ). Then A + B is the m n matrix
A + B := (aij + bij ). In other words, the (i, j) entry of A + B is the sum of
the corresponding entries of A and B. For example,

 
 

a11 a12
b11 b12
a11 + b11 a12 + b12
+
=
.
a21 a22
b21 b22
a21 + b21 a22 + b22
Addition and scalar multiplication can be combined in the usual way to
give linear combinations of matrices (of the same dimension). Here is an
example.
Example 3.8. Let


1 1 0 2
A=
2 4 0 1
Then


3A =


3
3
0 6
6 12 0 3


and

B=


1 1 1 0
.
0 2 1 2


and

A+B =


2 2 1 2
.
2 2 1 3

A matrix such that every entry is 0 is called a zero matrix. Clearly, the
m n zero matrix is an additive identity for addition of m n matrices.
Any matrix A has an additive inverse, namely A, since A + (A) = O, the
zero matrix. We can summarize all these comments by noting the following
Proposition 3.5. The set of all m n matrices over a field F is a vector
space over F.
In fact, a moments reflection suggests that the vector space of all m n
matrices over F is indistinguishable from Fmn .

3.2.2

Matrices Over F2 : Lorenz Codes and Scanners

Let us consider a couple of applications of matrices over the field F2 . Since F2

has only two elements, there are exactly 2mn mn such matrices. Obviously,
scalar multiplication is not very interesting. On the other hand, addition
can be.
Example 3.9. Lorenz Codes. For example,

 
 

1 0 1
1 1 1
0 1 0
+
=
,
0 1 1
1 1 1
1 0 0
and

 
 

1 0 1
1 0 1
0 0 0
+
=
.
0 1 1
0 1 1
0 0 0

61
The first sum changed the parity of every element in the left hand matrix.
The second illustrates that over F2 , every matrix is its own additive inverse.
This property leads to a way of encrypting a matrix over F2 called a Lorenz
code. Suppose a fixed m n matrix E over F2 has been selected. Then if
A is any other m n matrix over F2 , put L(A) = A + E. Then one may
consider L(A) as an encryption of A and thus transmit it securely. Anyone
who knows beforehand the matrix E can recover the original matrix A from
L(A), since
A = A + O = A + 2E = (A + E) + E = L(A) + E.
Example 3.10. (Scanners) We can also interpret matrices over F2 in another natural way. Consider a black and white photograph as being a rectangular array consisting of many black and white dots. By giving the white
dots the value 0 and the black dots the value 1, our black and white photo
is therefore transformed into a matrix over F2 . Now suppose we want to
compare two black and white photographs whose matrices A and B have
the same dimensions, that is, both are m n for some m, n. It turns out to
be difficult to make a computer scan the two matrices to see in how many
positions they agree. So the idea is to consider the sum A + B. If A + B
has a 1 in the (i, j)-component, then aij 6= bij . If it has a 0, then aij = bij .
Hence two identical photographs will sum to the zero matrix, and two complementary photographs will sum to the all ones matrix. A measure of how
similar the two photos represented by A and B are might be the number of
non zero entries of A + B. The number of non zero elements of a matrix A
over F2 is called the weight of A. We will encounter this notion of weight
again when we consider coding thory.

3.2.3

Matrix multiplication

The third algebraic operation is matrix multiplication. Perhaps unexpectedly, the product of two matrices of the same dimension is not always defined. (But square matrices of the same dimension can be multiplied.) To
be able to take the product AB of two matrices A and B, they have to be
defined over the same field, and the number of columns of A has to coincide
with the number of rows of B. To be precise, let A be m n and B n p.
Then the product AB of A and B is the m p matrix C whose entry in the
ith row and kth column is
cik =

n
X
j=1

aij bjk .

62
Thus
AB =

n
X

aij bjk .

j=1

Put another way, if the columns of A are a1 , . . . , an , then the rth column
of AB is
b1r a1 + b2r a2 + . . . bnr an .
Hence the columns of AB are linear combinations of the columns of A using
the entries of B as scalars. One can also express AB in terms of linear
cominations of the rows of B. This will turn out to be connected with row
operations, which we will take up in detail later.
Example 3.11. Here are two examples.


 
 

1 3
6 0
1 6 + 3 (2) 1 0 + 3 7
0 21
=
=
.
2 4
2 7
2 6 + 4 (2) 2 0 + 4 7
4 28
Note how the columns of the product are linear combinations. Computing
the product in the opposite order gives a different result:

 
 


6 0
1 3
61+02
63+04
6 18
=
=
.
2 4
2 1 + 7 2 2 3 + 7 4
12 22
2 7
From this example, we see that AB 6= BA. Therefore we have to distinguish between the products AB and BA. Of course, both products are
defined only when A and B are square of the same size. In fact, the only
time we dont need to distinguish is when m = n = 1 and we are multiplying
elements of F.

3.2.4

The transpose of a matrix

There is another operation on matrices called transposition. If A is m n,

the transpose AT of A is the n m matrix AT := (crs ), where crs = asr .
This is easy to remember: the ith row of AT is just the ith column of A.
Note two obvious facts. First,
(AT )T = A.
Second, a matrix and its transpose have the same diagonal. A matrix A
which is equal to its transpose (that is, A = AT ) is called symmetric. Clearly,
every symmetric matrix is square. The symmetric matrices over R turn out
to be especially important, as we will see later in the course.

63
Example 3.12. If

A=

1 2
3 4



1 3
A =
.
2 4
T

An example of a 2 2 symmetric matrix is



1 3
.
3 5
Notice that the dot product v w of any two vectors v, w Rn can be
expressed as a matrix product, provided we use the transpose. In fact,

w1
w2

n
X

vw =
vi wi = v1 v2 vn
.

i=1

wn
Thus,
v w = vwT .
Finally, the transpose of a product has an amusing property:

T
AB = B T AT .
This transpose identity can be seen as follows. The (i, j) entry of B T AT is
the dot product of the ith row of B T and the jth column of AT , which is
the same thing as the dot product of the jth row of A and the ith column
of B. Since this is the (j, i) entry of AB, the (i, j) entry of (AB)T is the
same as the (i, j) entry of B T AT . Therefore (AB)T = B T AT . Suggestion:
try this out on an example.

3.2.5

The algebra of matrices

Except for commutativity of multiplication, the expected algebraic properties of addition and multiplication all hold for matrices. Assuming all the
sums and products below are defined, matrix algebra obeys following laws:
(1) Associative Law: Matrix addition and multiplication are associative:








A+B +C =A+ B+C
and
AB C = A BC .

64
(2) Distributive Law: Matrix addition and multiplication are distributive:




A B + C = AB + AC
and
A + B C = AC + BC.
(3) Scalar Multiplication Law: For any scalar r,






rA B = A rB = r AB .
(4) Commutative Law for Addition: Matrix addition is commutative: A + B = B + A.
Verifying these properties is a routine exercise. I suggest working a
couple of examples to convince yourself, if necessary. Though seemingly
uninteresting, the associative law for multiplication will often turn out to be
the key property.

3.2.6

The n n identity matrix

The n n identity matrix In is the n n matrix which has a one in each

diagonal entry and a zero at each entry off the diagonal. For example,



1 0 0
1 0
I2 =
and I3 = 0 1 0 .
0 1
0 0 1
Note that it makes sense to refer to the identity matrix for any field F
(why?). The most important property of the identity matrix In is that it is
a multiplicative identity. To make this precise, we state the next proposition.
Proposition 3.6. If A is an m n matrix over any field F, then AIn = A
and Im A = A.
Proof. This is an exercise.

65
Exercises
Exercise 3.20. Make up three matrices A, B, C so that AB and BC are
defined. Then compute AB and (AB)C. Next compute BC and A(BC).
Compare your results.
Exercise 3.21. Suppose A and B are symmetric n n matrices. (You can
even assume n = 2.)
(a) Decide whether or not AB is always symmetric.
(AB)T = AB for all symmetric A and B?

That is, whether

(b) If the answer to (a) is no, what condition ensures AB is symmetric?

Exercise 3.22. Suppose B has a column of zeros. How does this affect any
product of the form AB? What if A has a row or a column of zeros?
Exercise 3.23. Let A be the 2 2 matrix over F2 such that aij = 1 for
each i, j. Compute A2 . Generalize this to matrices over Fp .
Exercise 3.24. Let A be the n n matrix over R such that aij = 2 for
all i, j. Find a formula for Aj for any positive integer j. (Note Aj is the
product AA A of A with itself j times.)
Exercise 3.25. Verify Proposition 3.6 for all m n matrices A over an
arbitrary field.
Exercise 3.26. Give an example of a 2 2 matrix A such that every entry
of A is either 0 or 1 and A2 = I2 as a matrix over F2 , but A2 6= I2 as a
matrix over the reals.
Exercise 3.27. Use matrix addition and matrix multiplication to define
two binary operations on the set of all real 2 2-matrices of the form


x y
.
y x
Prove that this turns the set


{ ac db M (2 2, R) | a = d and b + c = 0}
into a field. (You can use the results of 3.2.5 and 3.2.6 to prove most
of the field axioms. But commutativity of multiplication and existence of
multiplicative inverses will require special attention.) Do you see any relationship with the field of complex numbers?

66
Exercise 3.28. Find all matrices A M (2 2, F2 ) satisfying the matrix
equations
(i) A2 = 0,
(ii) A2 = I2 .

Exercise 3.29. Show that the real 2 2-matrix 10 54
has a multiplicative
1
inverse. Find it.


Prove that the real 2 2-matrix 00 10 admits no multiplicative inverse.

67

3.3

Matrices and linear transformations

As we mentioned at the beginning of this Chapter, matrices and linear

transformations are intimately related. We will now make that connection
precise.

3.3.1

The linear transformation of a matrix

Our plan is next to identify the set of all m n matrices A over F with
the set of all linear transformations T : Fn Fm . (This is why we usually
considered a T : Fn Fm instead of a T : Fm Fn .) First of all, an
arbitrary m n matrix A over F defines a linear transformation
TA : Fn Fm
by the rule
TA (x1 , . . . , xn ) = TA (x) := (AxT )T .
Since up to now we have been viewing the elements of Fn as row vectors,
the tranposes are necessary in order to turn column vectors back into row
vectors at the crucial times. Actually, it will simplify matters to adopt
the convention that from now on, all vectors in Fn will be written as
column vectors. Thus the above linear transformation can now be written

x1
x2

TA (x) = Ax = A ... .

xn1
xn
To spell everything out completely, express the matrix A in terms of its
columns, say A = (a1 a2 an ). Then
TA (x) = Ax =

n
X

xi a i .

i=1

In other words, the value of TA at x is the linear combination of the columns

of A using the ith component xi of x as the coeffficient of the ith column ai
of A.
By the distributive and scalar multiplication laws for matrix multiplication, TA is a linear transformation. In the next subsection, we will see the
nice fact we can now forget abstract linear transformations from Fn to Fm
and concentrate only on matrix linear transformations.

68

3.3.2

The matrix of a linear transformation

We now see that every linear transformation comes from a matrix.

Proposition 3.7. Every linear transformation T : Fn Fm is of the form
TA for a unique m n matrix A. The ith column of A is T (ei ), where ei is
the standard basis vector introduced below.
Here, ei = (0, . . . , 0, 1, 0, . . . , 0)T , where the 1 is in the ith component
and the other n 1 components are 0. The vectors e1 , . . . , en make up what
we call the standard basis of Fn . (The general notion of a basis will be spelled
out in Chapter 5.) The main point is that any x = (x1 , . . . , xn )T Fn has
the unique expansion
n
X
x=
xi e i
i=1

as a linear combination of the standard basis vectors. Therefore,

T (x) = T

n
X
i=1

xi e i =

n
X

xi T (ei ) = Ax,

i=1

where A is the m n matrix T (e1 ) . . . T (en ) . If A and B are m n
and A 6= B, then Aei 6= Bei for some i, so TA (ei ) 6= TB (ei ). This means
different matrices define different linear transformations, which establishes
the uniqueness statement of the Proposition.
Notice that this proof gives an explicit description of how to write the
matrix A of T . The ith column of the matrix is just T (ei ), and so


A = T (e1 ) T (en ) .
We have now established that linear transformations and matrices are
the same thing. All that is left is to do some examples.
Example 3.13. Lets find the matrix of the reflection of R2 through the
line x = y. The linear transformation in question is Hb where b = (1, 1)T .
We need to find Hb (ei ) for i = 1, 2. Clearly, e1 is reflected to e2 and e2 is
reflected to e1 . Hence the matrix of Hb is


0 1
.
1 0
In 3.1, we dealt only with linear transformations we could express by
writing down a formula. It is sometimes much easier to express a linear
transformation by giving its matrix. An example of this is given by considering the rotations of R2 .

69
Example 3.14. For this example, you will need to refer back Example 3.3.
Let R denote the matrix of the counter clockwise rotation R of R2 through
. Then


cos sin
R =
.
sin cos
Example 3.15. Suppose T (x1 , x2 )T
(3, 1)T and T (e2 ) = (2, 1)T . Hence

3
1

3.3.3

= (3x1 + 2x2 , x1 x2 )T . Then T (e1 ) =

the matrix of T is

2
.
1

A remark on the algebra of linear transformations

Recall that we have also stated that any linear combination of linear transformations is a linear transformation. We can also consider linear combinations
of m n matrices. The next theorem states that these two operations are
the same.
Proposition 3.8. If S, T : Fn Fm are linear transformations with matrices A and B respectively, then, for any scalars a, b F, the matrix of
aS + bT is aA + bB.

70
Exercises
Exercise 3.30. Find a linear transformation T : R3 R3 with the property
that T (e1 ) = (1, 2, 2)T , T (e2 ) = (0, 1, 2)T and T (e3 ) = (3, 1, 1)T .
(a) What is the value of T on (1, 2, 0)T ?
(b) Find all values of x R3 for which T (x) = (2, 3, 1)T .
Exercise 3.31. What is the matrix of the linear transformation T : Rn
Rn defined by T (x) = x for all x Rn ?
Exercise 3.32. Let b = (1, 2)T . Find the matrices A of Pb and B of Hb .
Compute A2 and B 2 ? (Try to predict the answer without computing.)
Exercise 3.33. For the matrices A and B of the previous problem, is AB =
BA?
Exercise 3.34. Find the matrices A + 2B and 3A B, where A and B are
the matrices in the previous two problems. Hence verify Theorem 3.8 in this
special case.
Exercise 3.35. Is there a linear transformation T : R2 R2 so that
T (1, 1)T = (3, 4)T and T (1, 1)T = (1, 2)T ? If so, find it.
Exercise 3.36. Let ` be the line in R2 passing through the origin which
meets the positive x-axis at the angle , and let b be a unit vector in R2
orthogonal to `.
(a) Find the matrix of the reflection Hb of R2 through ` in terms of .
(b) Show that the matrix of the reflection of part (a) is symmetric.
Exercise 3.37. An n n matrix Q over R is called orthogonal if QT Q = In .
(a) Show that the matrix of Hb is orthogonal.
(b) Prove that every rotation matrix R is also orthogonal.
Exercise 3.38. * Let T : Rn Rn be an orthogonal linear transformation.
Show that the matrix of T is an orthogonal matrix.

71

3.4

Multiplying matrices and composing linear transformations

The definition of matrix multiplication we gave in the previous section is

just an extension of the dot product which we studied at length in 2, it
should seem to be a natural operation (though perhaps a trifle unwieldy).
Yet the definition of matrix multiplication we have given turns out to give
us the property that product BA of two matrices A and B is the matrix of
the linear transformation TB TA obtained by composing TA and TB .

3.4.1

Composing linear transformations

Lets first review the meaning of the composition of two transformations.

Suppose S : Fp Fn and T : Fn Fm are two transformations. Since the
target of S is the domain of T , we may define the composite transformation
T S : Fp Fm by putting by


T S(x) = T S(x) .
If S and T are linear transformations, we have the following facts.
Proposition 3.9. Suppose S : Fp Fn and T : Fn Fm are two linear
transformations and assume A is the matrix of S and B is the matrix of
T . Then the composite transformation T S : Fp Fm is also linear.
Furthermore, the matrix of T S is BA. In other words, T S = TB TA =
TBA .
Proof. To prove T S is linear, note that

T S(rx + sy) = T S(rx + sy)

= T rS(x) + sS(y)




= rT S(x) + sT S(y) .

In other words, T S(rx + sy) = rT S(x) + sT S(y), so T S is linear

as claimed. To find the matrix of T S, we observe that
T S(x) = T (Ax) = B(Ax) = (BA)x.
This implies that the matrix of T S is the product BA as asserted.
We can put our result in another form. Let T : Fn Fm and S : Fp Fn
be linear, and let MT and MS denote their matrices. Then what we have
just shown is the identity
MT S = MT MS .

72
Let us emphasize that in the proof of Proposition 3.9, the key fact is that
matrix multiplication is associative. In fact, the main observation is that
T S(x) = T (Ax) = B(Ax). Given this, it is immediate that T S is linear,
so the first step in the proof was actually unnecessary. This Proposition is
one of the main applications of the associativity of matrix multiplication.

3.4.2

An example: rotations of R2

A nice way of illustrating the above discussion is by considering rotations of

the plane.
Example 3.16. We already briefly considered rotations of R2 in Example
3.3. We will now consider them in a little more detail. Recall the counter
clockwise rotation R is the transformation which rotates any vector in R2
counter clockwise through . One can argue that a rotation R is linear as
follows. Suppose x and y are any two non collinear vectors in R2 , and let P
be the parallelogram they span. Then R rotates the whole parallelogram
P about the origin to a new parallelogram, namely R (P ). Clearly R (P )
is spanned by R (x) and R (y). Hence, the diagonal x + y of P is rotated
to the diagonal of R (P ). Thus
R (x + y) = R (x) + R (y).
Similarly, for any scalar r,
R (rx) = rR (x).
Therefore R is in fact linear. We suggest another way to deduce R is
linear in the exercises. Now R sends e1 = (1, 0)T to (cos , sin )T and
e2 = (0, 1)T to ( sin , cos )T . Thus, as we saw above, the matrix R of
R is


cos sin
R =
.
sin cos
Clearly, if one first applies the rotation R and follows that by the rotation
R , the outcome is the rotation R+ through + . In other words, ]
R+ = R R .
Therefore, by Proposition 3.9, we see that

 


cos( + ) sin( + )
cos sin
cos sin
=
.
sin( + ) cos( + )
sin cos
sin cos

73
Expanding the product gives the angle sum formulas for cos( + ) and
sin( + ). Namely,
cos( + ) = cos cos sin sin ,
and
sin( + ) = sin cos + cos sin .
Thus the angle sum formulas for cosine and sine via matrix algebra come
from the fact that rotations are linear transformations (together with Proposition 3.9).

3.4.3

Orthogonal linear transformations

Recall (see 3.1) that a linear transformation T : Rn Rn is called orthogonal if

T (v) T (w) = v w
for all v, w Rn . As we mentioned earlier, since orthogonal linear transformations preserve dot products, they preserve also preserve all the other
associated quantities such as lengths and the angles between vectors. For example, reflections and rotations of R2 are orthogonal transformations. That
reflections are orthogonal follows by applying the formula for a reflection
and then simply computing. On the other hand, the reason rotations are
orthogonal is that they obviously preserve lengths and also the angles between two vectors. Given that, they also have to preserve the dot product
of any two vectors (why?).
In the exercises of 3.3, we introduced the notion of an orthogonal matrix.
An n n matrix Q over R is orthogonal if QT Q = In . As the terminology
suggests, there is a connection between orthogonal matrices and orthogonal
transformations.
Proposition 3.10. The matrix Q of an orthogonal linear transformation
T : Rn Rn is orthogonal. Conversely, if Q is an orthogonal matrix,
then its associated linear transformation TQ is orthogonal. Furthermore,
the composition of two orthogonal linear transformations is orthogonal, and
hence the product of two orthogonal matrices is orthogonal.
This Proposition now gives an easier proof all rotations R of R2 are
orthogonal. Indeed, it is only necessary to check that (R )T R = I2 .
Proof of Proposition 3.10. First, recall that the dot product v w of any
v, w Rn is expressed as the matrix product
v w = vT w.

74
Thus let Q be an orthogonal matrix. Then
Qv Qw = (Qv)T Qw
= (vT QT )Qw
= vT (QT Q)w
= vT w
= v w,
since QT Q = In . Therefore the linear transformation given by an orthogonal
matrix is orthogonal. On the other hand, let T be an orthogonal linear
transformation with matrix Q. We must show that QT Q = In . Since
T (v) = Qv, the law for computing the transpose of a product and the
associativity of matrix multiplication, tell us that
T (v) T (w) = Qv Qw
= (Qv)T Qw
= vT QT Qw

Since T is orthogonal, this implies

vT (QT Q)w = vT w
for all v, w Rn . But if A is any n n matrix,
eTi Aej = aij ,
so putting v = ei and w = ej gives the formula eTi (QT Q)ej = 0 if i 6= j and
eTi (QT Q)ej = 1 if i = j. In other words, QT Q = In . Therefore, the matrix
Q is orthogonal.
To verify that the composition of any two orthogonal linear transformations S, T : Rn Rn is orthogonal, we have to show that
S T (v) S T (w) = v w.
But


S T (v) S T (w) = S T (v) S(T (w)) = T (v) T (w) = v w,
so we get what is needed. Hence the product of two orthogonal matrices is
orthogonal, and the proof is complete.

75
We could also check directly that if Q and R are orthogonal, then QR
is. For
(QR)T QR = (RT QT )QR = RT (QT Q)R = RT R = In .
This is another way of showing the composition of two orthogonal transformations is orthogonal.
There is a simple way to determine whether a matix is orthogonal.
Proposition 3.11. An nn matrix Q is othogonal if and only if the columns
of Q are mutually orthogonal unit vectors.
Proof. Let Q = (q1 q2 . . . qn ) be orthogonal. Since QT Q = In , it follows
that qTi qj , which is, by definition, the entry of QT Q in the (i, j) component,
satisfies
(
0 if i 6= j
T
qi qj =
1 if i = j
Hence the columns of Q are indeed mutually orthogonal unit vectors.
We will now use this proposition give a complete description of the orthogonal linear transformations of R2 .
Proposition 3.12. Every orthogonal linear transformation of R2 is either
a reflection or a rotation.
Proof. We have to show that every 2 2 orthogonal matrix is either a
rotation matrix or a reflection matrix. Consider an arbitrary orthogonal
matrix


a b
Q=
.
c d
By the previous proposition, the columns of Q are of unit length and the
first and second columns are orthogonal. Thus
a2 + c2 = 1,

b2 + d2 = 1,

and
ac + bd = 0.
It is easy to deduce from this that Q is one of the two possibilties:




cos sin
cos sin
Q1 =
or Q2 =
.
sin cos
sin cos

76
But Q1 is the rotation through . It remains to show Q2 is a reflection. In
fact, it is the reflection of R2 through the line R(cos /2, sin /2)T . We leave
this as an exercise.
This proposition makes an interesting point about composing rotations
and reflections of the plane. The composition of two rotations is another
rotation, as we observed earlier in this section and elsewhere, but the composition of two reflections isnt a reflection. However, it is still an orthogonal
transformation, so it has to be a rotation. You can guess what is by trying
some special cases. We will also leave this as an exercise.

3.4.4

Powers of a matrix and the Fibonacci sequence

Now let us turn to something different. One of the problems that linear
algebra is frequently applied to is the problem of finding what happens
when a linear process is iterated over and over. This is a special case of a
dynamical system. If we want to compute what happens when we compose
a linear transformation T : Fn Fn with itself k times, and if T has matrix
A, then by Proposition 3.9, we need to compute Ak . It turns out that
linear algebra, in particular eigentheory, can sometimes solve this problem
(as well as some related problems such as whether the sequence of powers
Ak converges as k and to what it converges if it does).
Let us now consider a very pretty example to illustrate how matrix powers can occur.
Example 3.17. The Fibonacci sequence
a0 , a1 , a2 , . . . , am , . . .
is the infinite sequence defined by giveing the first two terms a0 and a1 two
arbitrary values and setting an = an1 + an2 if n 2. In the original
Fibonacci sequence, a0 = 0 and a1 = 1. In that case, the first eight terms
are
0, 1, 1, 2, 3, 5, 8, 13.
The defining relation an = an1 + an2 (n 2) can be expressed in matrix
form as follows:

 


an+1
1 1
an
=
.
1 0
an
an1

77
Putting F =

11
10

and iterating the above relation, we see that



an+1
an


an
= F
an1


2 an1
= F
an2
..
.
 
a
= Fn 1 .
a0

Thus


 
an+1
n a1
=F
.
an
a0

(3.2)

Note that to obtain the identity (3.2) for F n , we have once again used the
associativity of matrix multiplication. Hence, if F n is found, then we have
an and an+1 in terms of a0 and a1 . We will see later that each an can be
expressed in terms of a famous irrational number called the golden number.
The Fibonacci sequence turns out to appear in biology in connection with
a phenomenon called phyllotaxis.

78
Exercises
Exercise 3.39. Let L be the line in R2 spanned by (1, 1)T . Find the matrices of the following linear transformations.
(a) The transformation of R2 obtained by first reflecting R2 through the
line L and then reflecting it through the line spanned by (0, 1)T .
(b) The transformation of R2 obtained by first projecting R2 onto the line
L and then projecting it onto the line spanned by (0, 1)T .
Exercise 3.40. Let Pb : R2 R2 be the projection on the line spanned by
b. Find an general expression for the matrix of Pb and use it to verify that
Pb Pb = Pb in two ways.
Exercise 3.41. Let L and M be two orthogonal lines through the origin in
R2 , and let H and J denote respectively the matrices of the reflections of R2
through L and M . Show that HJ = JH = I2 . (Hint: Use Exercise 3.6.)
Exercise 3.42. Compute the matrix of the linear transformation that first
rotates R2 though /4 then reflects R2 through the y-axis. Is this transformation a reflection or a rotation? Explain.
Exercise 3.43. Let H the matrix of the reflection of R2 though the line
x = y and let R be the ccw rotation through /4.
(a) Describe the transformation given by HRH 1 . That is, identify it as
either a reflection or a rotation, and describe it.
(b) Repeat part (a) for RHR1 .
Exercise 3.44. Let H be the reflection of R2 through the line whose angle
with the x axis, measured counter clockwise, is .


cos 2 sin 2
(i) Show directly that the matrix of H is
.
sin 2 cos 2
(ii) Show that the linear transformation H H is a rotation R and find
. (Answer: = 2( ).)
Exercise 3.45. In general, what is the product (or composition) of a rotation of R2 and a reflection of R2 ?
Exercise 3.46. Why cant there be a linear transformation T : R2 R2
that is both a reflection and a rotation?

79
Exercise 3.47. Let A be the 3 3 matrix such that aij = 2 for all 1
i, j 3. Find a formula for An for all n > 0.


Exercise 3.48. Let B be the 2 2 matrix 10 11 . Find a formula for B n ,
for all n > 0.

80

3.5

The null space and range of a linear transformation

In the previous two sections, we have seen that m n matrices over a field F
are the same as linear transformations with domain Fn and target Fm . We
have also seen that matrix multiplication is the same as composition of the
corresponding linear transformations. In the next section, we will start the
study of linear systems. It turns out that linear systems are closely involve
two fundamental subspaces associated to any linear transformation, namely
the null space, or kernel, and the image of the transformation. The purpose
of this section is to define these subspaces and get some of their properties.

3.5.1

Basic definitions

Let V and W be two vector spaces over the field F, and let T : V W
be an arbitrary linear transformation. The null space of T , often also called
the kernel of T , is defined as the subset N (T ) of V consisting of all vectors
which T sends to the zero vector. That is,
N (T ) := {x V | T (x) = 0}.
When V = Fn , W = Fm and A is the matrix of T , then we will also consider
the null space of A, which is defined in exactly the same way, i.e.
N (A) = {x Fn | Ax = 0}.
The other basic subspace associated to T (and A) is the range or image
of T , which we will denote as Im(T ). Namely,
Im(T ) = {y W |y = T (x) x V }.
Linear transformations have the special property that both the null space
and range are subspaces (of the domain and target respectively).
Proposition 3.13. The null space and image of a linear transformation
T : V W are subspaces of the domain V and the target W respectively.
When V = Fn , W = Fm and A is the matrix of T , then the image of T is
precisely the subspace of Fm spanned by the columns of A.
Proof. Let us prove first that the null space is a subspace. Suppose a and
b are in N (T ). Then, by the fact that T is linear,
T (a + b) = T (a) + T (b) = 0 + 0 = 0.

81
Hence the sum of any two vectors in the null space is also in the null space.
Similarly, if r F is any scalar, then
T (ra) = rT (a) = r0 = 0.
Hence, N (T ) is a subspace. The proof that the image is a subspace is similar.
Another way to see that the image is a subspace is to note that if the matrix
A has the form A = (a1 . . . an ), then
Ax = x1 a1 + + xn an .
Thus, the vector Ax is a linear combination of the columns of A. Hence
Im(T ) is the subsace of Fm spanned by the columns of A as asserted. This
proves all the claims of the Proposition.
Example 3.18. Consider the linear transformation given by the 33 matrix

1 4
2
A= 3
2 7 .
2 3 2
The null space N (A) is the set of solutions of
x1 + 4x2 + 2x3 = 0
3x1 + 2x2 7x3 = 0
2x1 3x2 + 2x3 = 0
The set of all solutions to this linear system is a subspace of R3 . The crucial
property is that each equation is linear and every constant term on the right
side is zero. A reasonable question is how to find N (A). For example, that
0 N (A) is clear. But are there any non-zero vectors in the null space? At
this point, we dont have the tools for determining this, but we will develope
a method called Gaussian elimination in the next chapter. The image is the
span of the columns

1
4
2
3 , 2 , 7 .
2
3
2
Example 3.19. Suppose T : R2 R2 is a diagonal linear transformation.
That is, T (x, y)T = (x, y)T . Then if both , 6= 0, then N (T ) = 0. In
this case, the image of T is R2 . Otherwise, N (T ) is non zero. For example,
if = 0, then N (T ) = {(x, y)T | x = 0}.

82
The subspace of Fm spanned by the columns of an m n matrix A is
called the column space of A. In other words, the column space of A is the
image of TA . The subspace of Fn spanned by the rows of A is called the row
space of A. We will use both of these subspaces extensively when we treat
linear systems.
Here is a useful comment that relates N (A) to a general property of the
linear transformation TA . We call a transformation F : Fn Fm (linear or
not) one to one if F (x) = F (y) implies x = y.
FIGURE
(ONE TO ONE TRANSFORMATION)
Proposition 3.14. A linear transformation T : Fn Fm is one to one if
and only if N (T ) = {0}.
Proof. Suppose first that T is one to one. Then clearly N (T ) = {0} since
T (0) = 0. Conversely, suppose N (T ) = {0}. If x, y Rn are such that
T (x) = T (y), then
T (x) T (y) = T (x y).
Therefore,
T (x y) = 0.
Thus x y N (T ), which means x y = {0} hence x = y. Thus T is one
to one, and the proof is finished.

83
Exercises
Exercise 3.49. Find a description of both the column space and null space
1 1 0

of the matrix 2 3 1 .
121

Exercise 3.50. Using only the basic definition of a linear transformation,

show that the image of a linear transformation T is a subspace of the target
of T .
Exercise 3.51. Let A and B be n n matrices.
(a) Explain why the null space of A is contained in the null space of BA.
(b) Explain why the column space of A contains the column space of
AB.
(c) If AB = O, show that the column space of B is contained in N (A).
Exercise 3.52. Consider the subspace W of R4 spanned by (1, 1, 1, 2)T
and (1, 1, 0, 1)T . Find a system of homogeneous linear equations whose set
of solutions is W .
Exercise 3.53. What are the null space and image of:
(a) a projection Pb : R2 R2 ?
(b) a reflection Hb : R2 R2 ?
(c) a rotation R : R2 R2 ?
Exercise 3.54. Repeat Exercise 3.53 for the projection P : R3 R3 defined
by
P (x, y, z)T = (x, y)T .
Exercise 3.55. Let A be a real 3 3 matrix such that the first row of A is
a linear combination of As second and third rows.
(a) Show that N (A) is either a line through the origin or a plane containing
the origin.
(b) Show that if the second and third rows of A span a plane P , then
N (A) is the line through the origin orthogonal to P .
Exercise 3.56. Let T : Fn Fn be a linear transformation such that
N (T ) = 0 and Im(T ) = Fn . Prove the following statements.
(a) There exists a transformation S : Fn Fn with the property that
S(y) = x if and only if T (x) = y. Note: S is called the inverse of T .

84
(b) Show that in addition, S is also a linear transformation.
(c) If A is the matrix of T and B is the matrix of S, then BA = AB = In .
Exercise 3.57. Let S : Rn Rm and T : Rm Rp be two linear transformations both of which are one to one. Show that the composition T S
is also one to one. Conclude that if A is m n has N (A) = {0} and B is
n p has N (B) = {0}, then N (BA) = {0} too.
Exercise 3.58. If A is any n n matrix over a field F, then A is said
to be invertible with inverse B if B is an n n matrix over F such that
AB = BA = In . (In other words, A is invertible
 if and
 only if its associated
a b
linear transformation is.) Show that if A =
, then A is invertible
c d


d b
1
provided ad bc 6= 0 and that in this case, the matrix (ad bc)
c a
is an inverse of A.
Exercise
Let F : R2 R2 be the linear transformation given by
  3.59. 
x
y
F
=
.
y
xy
(a) Show that F has an inverse and find it.
(b) Verify that if A is the matrix of F , then AB = BA = I2 if B is a
matrix of the inverse of F .