Vous êtes sur la page 1sur 56

CHAPTER 6

LINEAR TRANSFORMATIONS
6.1 WHAT IS A LINEAR TRANSFORMA-
TION?
You know what a function is - its a RULE which
turns NUMBERS INTO OTHER NUMBERS: f(x) =
x
2
means please turn 3 into 9, 12 into 144 and so
on.
Similarly a TRANSFORMATION is a rule which
turns VECTORS into other VECTORS. For exam-
ple, please rotate all 3-dimensional vectors through
an angle of 90

clockwise around the z-axis. A LIN-


EAR TRANSFORMATION T is one that ALSO
satises these rules: if c is any scalar, and

u and

v
are vectors, then
T(c

u) = cT(

u) and T(

u +

v ) = T(

u) +T(

v ).
1
EXAMPLE: Let I be the rule I

u =

u for all

u.
You can check that I is linear! Called IDENTITY
Linear Transformation.
EXAMPLE : Let D be the rule D

u = 2

u for all

u.
D(c

u) = 2(c

u) = c(2

u) = cD

u
D(

u +

v ) = 2(

u +

v ) = 2

u + 2

v = D

u + D

v
LINEAR!
Note: Usually we write D(

u) as just D

u.
6.2. THE BASIC BOX, AND THE MATRIX
OF A LINEAR TRANSFORMATION
The usual vectors

i and

j dene a square:
Lets call this the BASIC BOX in two dimensions.
2
Similarly,

i,

j, and

k dene the BASIC BOX in 3
dimensions.
Now let T be any linear transformation. You know
that any 2-dimensional vector can be written as a

i +
b

j, for some numbers a and b. So for any vector, we


have
T(a

i +b

j) = aT

i +bT

j.
This formula tells us something very important: IF
I KNOW WHAT T DOES TO

i and

j, THEN I
KNOW EVERYTHING ABOUT T - because now I
can tell you what T does to ANY vector.
EXAMPLE: Suppose I know that T(

i) =

i +
1
4

j
and T(

j) =
1
4

i +

j. Then what is T(2

i + 3

j)?
Answer: T(2

i + 3

j) = 2T

i + 3T

j = 2
_

i +
1
4

j
_
+
3
_
1
4

i +

j
_
= 2

i +
1
2

j +
3
4

i + 3

j =
11
4

i +
7
2

j.
Since T

i and T

j tell me everything I need to know,


3
this means that I can tell you everything about T by
telling you WHAT IT DOES TO THE BASIC BOX.
EXAMPLE: Let T be the same transformation as
above, T(

i) =

i +
1
4

j and T(

j) =
1
4

i +

j.
The basic box has been squashed a bit! Pictures
of WHAT T DOES TO THE BASIC BOX tell us
everything about T!
EXAMPLE: If D is the transformation D

u = 2

u,
then the Basic Box just gets stretched:
So every LT can be pictured by
seeing what it does to the Basic Box.
There is another way!
Let T

i =
_
a
c
_
and T

j =
_
b
d
_
. Then we DEFINE
4
THE MATRIX OF T RELATIVE TO

i,

j as
_
a b
c d
_
,
that is, the rst COLUMN tells us what happened
to

i, and the second column tells us what happened
to

j.
EXAMPLE: Let I be the identity transformation.
Then I

i =

i =
_
1
0
_
, I

j =

j =
_
0
1
_
, so the matrix of
the identity transformation relative to

i

j is
_
1 0
0 1
_
.
EXAMPLE: If D

u = 2

u, then D

i =
_
2
0
_
and
D

j =
_
0
2
_
so the matrix of D relative to

i,

j is
_
2 0
0 2
_
.
EXAMPLE: If T

i =

i +
1
4

j and T

j =
1
4

i + j, then
the matrix is
_
1
1
4
1
4
1
_
.
EXAMPLE: If T

i =

j and T

j =

i, the matrix is
_
0 1
1 0
_
. Basic box is REFLECTED
5
EXAMPLE: Suppose in 3 dimensions T

i =

i+4

j +
7

k, T

j = 2

i + 5

j + 8

k, T

k = 3

i + 6

j + 9

k, then the
matrix is
_
_
1 2 3
4 5 6
7 8 9
_
_
, relative to

i

k.
EXAMPLE: Suppose T

i =

i +

j + 2

k and T

j =

i 3

k. This is an example of a LT that eats 2-


dimensional vectors but PRODUCES 3-dimensional
vectors. But it still has a matrix,
_
_
1 1
1 0
2 3
_
_
. Its
just that this matrix is not a SQUARE MATRIX,
that is, it is not 2 by 2 or 3 by 3. Instead it is 3 by
2.
We shall say that a linear transformation is a 2-
dimensional L.T. if it eats 2-dimensional vectors AND
produces 2-dimensional vectors. A 2-dimensional
L.T. has a square, 2 by 2 matrix relative to

i,

j. Sim-
ilarly a 3-dimensional linear transformation has a 3
by 3 matrix. In Engineering applications, most lin-
6
ear transformations are 2-dimensional or 3-dimensional,
so we are mainly interested in these two cases.
EXAMPLE: Suppose T is a linear transformation
that eats 3-dimensional vectors and produces
2-dimensional vectors according to the rule T

i = 2

i,
T

j =

i +

j, T

k =

j. What is its matrix?


Answer:
_
2 1 1
0 1 1
_
, a 2 by 3 matrix.
EXAMPLE: Suppose, in solid mechanics, you take
a at square of rubber and SHEAR it, as shown.
In other words, you dont
change its volume,
you just push it like a pack
of cards. The base stays xed but the top moves
a distance tan(). (The height remains the same, 1
unit.) Clearly the shearing transformation S satises
S

i =

i, S

j =

i tan +

j, so the matrix of S relative
to

i,

j is
_
1 tan
0 1
_
.
7
EXAMPLE: Suppose Ti =

i +

j and T

j =

i +

j. Matrix is
_
1 1
1 1
_
and basic box is SQUASHED
FLAT!
EXAMPLE: Rotations in the plane. Suppose you
ROTATE the whole plane through an angle (anti-
clockwise). Then simple trigonometry shows you
that
R

i = cos

i + sin

j
R

j = sin i + cos

j
So the rotation matrix is
R() =
_
cos sin
sin cos
_
.
Application: Suppose an object is moving on a cir-
cle at constant angular speed . What is its accel-
8
eration?
Answer: Let its position vector at t = 0 be

r
0
.
Because the object is moving on a circle, its position
at a later time t is given by rotating

r
0
by an angle
(t). So

r (t) =
_
cos sin
sin cos
_

r
0
Dierentiate
d

r
dt
=

_
sin cos
cos sin
_

r
0
by the chain rule. Here

is actually , so
d

r
dt
=
_
sin cos
cos sin
_

r
0
. Dierentiate again,
d
2

r
dt
2
=
_
cos sin
sin cos
_

r
0
=
2
_
cos sin
sin cos
_

r
0
.
9
Substitute the equation for

r (t),
d
2

r
dt
2
=
2

r ,
which is formula you know from physics.
6.3. COMPOSITE TRANSFORMATIONS
AND MATRIX MULTIPLICATION.
You know what it means to take the COMPOSITE
of two functions: if f(u) = sin(u), and u(x) = x
2
,
then f u means: please do u FIRST, THEN f, so
f u(x) = sin(x
2
). NOTE THE ORDER!!
u f(x) = sin
2
(x), NOT the same!
Similarly if A and B are linear transformations, then
AB means do B FIRST, then A.
NOTE: BE CAREFUL! According to our denition,
A and B both eat vectors and both produce vectors.
But then you have to take care that A can eat what
10
B produces!
EXAMPLE: Suppose Aeats and produces 2-dimensional
vectors, and B eats and produces 3-dimensional vec-
tors. Then AB would not make sense!
EXAMPLE: Suppose B eats 2-d vectors and pro-
duces 3-d vectors (so its matrix relative to

i

k looks
like this:
_
_
b
11
b
12
b
21
b
22
b
31
b
32
_
_
, a 3 by 2 matrix) and suppose
A eats 3-d vectors and produces 2-d vectors. Then
AB DOES make sense, because A can eat what B
produces. (In this case, BA also makes sense.).
IMPORTANT FACT: Suppose a
ij
is the matrix
of a linear transformation A relative to

k, and sup-
pose b
ij
is the matrix of the Linear Transformation
B relative to

i

k. Suppose that AB makes sense.


Then the matrix of AB relative to

i

j or

i

k is just
11
the matrix product of a
ij
and b
ij
.
EXAMPLE: What happens to the vector
_
1
2
_
if
we shear 45

parallel to the x axis and then rotate


90

anticlockwise? What if we do the same in the


reverse order?
Answer: Shear

_
1 tan
0 1
_
so in this case it is
_
1 1
0 1
_
. A rotation through
has matrix
_
cos sin
sin cos
_
, so here it is
_
0 1
1 0
_
Hence
SHEAR, THEN ROTATE

_
0 1
1 0
_ _
1 1
0 1
_
=
_
0 1
1 1
_
ROTATE, THEN SHEAR

_
1 1
0 1
_ _
0 1
1 0
_
=
_
1 1
1 0
_
.
12
So shear, then rotate
_
1
2
_

_
0 1
1 1
_ _
1
2
_
=
_
2
3
_
.
Rotate, then shear
_
1
2
_

_
1 1
1 0
_ _
1
2
_
=
_
1
1
_
Very dierent!
EXAMPLE: Suppose B is a LT with matrix
_
_
1 0
0 1
1 1
_
_
and A is a LT with matrix
_
0 1 1
1 1 0
_
.
What is the matrix of AB? Of BA?
Answer:
_
0 1 1
1 1 0
_
_
_
1 0
0 1
1 1
_
_
=
_
1 0
1 1
_
= AB
2 by 3 3 by 2 2 by 2
_
_
1 0
0 1
1 1
_
_
_
0 1 1
1 1 0
_
=
_
_
0 1 1
1 1 0
1 0 1
_
_
3 by 2 2 by 3 3 by 3
13
EXAMPLE: Suppose you take a piece of rubber in
2 dimensions and shear it parallel to the x axis by
degrees, and then shear it again by degrees. What
happens?
_
1 tan
0 1
_ _
1 tan
0 1
_
=
_
1 tan + tan
0 1
_
which is also a shear, but NOT through +!
The shear angles dont add up, since tan +tan =
tan( +).
EXAMPLE: Rotate 90

around z-axis, then rotate


90

around x-axis in 3 dimensions. [Always anti-


clockwise unless otherwise stated.] Is it the same
if we reverse the order? Rotate about z axis
i becomes j, j becomes i, k stays the same, so
_
_
0 1 0
1 0 0
0 0 1
_
_
. Rotate about x axis, i stays the same,
14
j becomes k, k becomes j, so
_
_
1 0 0
0 0 1
0 1 0
_
_
, and
_
_
1 0 0
0 0 1
0 1 0
_
_
_
_
0 1 0
1 0 0
0 0 1
_
_
=
_
_
0 1 0
1 0 0
0 0 1
_
_
_
_
1 0 0
0 0 1
0 1 0
_
_
,
so the answer is NO!
6.4 DETERMINANTS
You probably know that the AREA of the box de-
ned by two vectors is
|

v |, magnitude
of the vector product.
If you dont know it, you can easily check it, since
the area of any parallelogram is given by
15
AREA = HEIGHT Base
=
_
|

v | sin
_
|

u|
= |

u| |

v | sin
= |

v |.
Similarly, the VOLUME of a three-dimensional par-
allelogram [called a PARALLELOPIPED!] is given
by
VOLUME = (AREA OF BASE) HEIGHT.
If you take any 3 vectors in 3 dimensions, say

u,

v ,

w,
then they dene a 3-dimensional parallelogram. The
area of the base is |

v |,
height is |

w| | sin
_

2

_
|
where is the angle between

v and

w, so VOLUME
dened by

u,

v ,

w is just
16
|

v | |

w| | sin
_

2

_
|
=|

v | |

w| | cos |
=|

w|.
[Check: Volume of Basic Box dened by

i

k is
|

k| = |

k| = 1, correct!
Now let T be any linear transformation in two di-
mensions. [This means that it acts on vectors in the
xy plane and turns them into other vectors in the xy
plane.]
We let T act on the Basic Box, as usual.
Now T

i and T

j still lie in the same plane as



i and

j, so (T

i) (T

j) must be perpendicular to that


plane. Hence it must be some multiple of

k. We
17
dene the DETERMINANT of T to be that multi-
ple, that is, by denition, det(T) is the number given
[STRICTLY IN 2 DIMENSIONS] by
(T

i) (T

j) = det(T)

k.
EXAMPLE: If I = identity, then
I

i I

j =

j =

k = 1

k
so det(I) = 1.
EXAMPLE: D

u = 2

u
D

i D

j = 4

j = 4

k det(D) = 4
EXAMPLE: T

i =

i +
1
4

j, T

j =
1
4

i +

j,
T

i T

j =
_

i +
1
4

j
_

_
1
4

i +

j
_
=

j +
1
16

i
=
15
16

j =
15
16

k det T =
15
16
.
EXAMPLE: T

i =

j, T

j =

i,
T

i T

j =

j

i =

k det T = 1
18
EXAMPLE: Shear, S

i =

i, S

j =

i tan +

j,
S

i S

j =

k det S = 1.
EXAMPLE: T

i =

i +

j = T

j,
T

i T

j =

0
det T = 0.
EXAMPLE: Rotation
R

i R

j = (cos

i + sin

j) (sin

i + cos

j)
= (cos
2
sin
2
)

k =

k det(R) = 1.
The area of the Basic Box is initially |

j| = 1.
After we let T act on it, the area becomes
|T

i T

j| = |det T| |

k| = | det T|.
So
Final Area of Basic Box
Initial Area of Basic Box
=
| det T|
1
= | det T|
19
so | det T| TELLS YOU THE AMOUNT BY WHICH
AREAS ARE CHANGED BY T. So det T = 1
means that the area is UNCHANGED (Shears, ro-
tations, reections) while det T = 0 means that the
Basic Box is squashed FLAT, zero area.
Take a general 2 by 2 matrix M =
_
a b
c d
_
. We
know that this means M

i = a

i + c

j, M

j = b

i + d

j.
Hence M

i M

j =
_
a

i +c

j
_

_
b

i +d

j
_
= (ad
bc)

k, so
det
_
a b
c d
_
= ad bc.
Check: det
_
2 0
0 2
_
= 4, det
_
1 tan
0 1
_
= 1,
det
_
cos sin
sin cos
_
= 1, det
_
1 1
1 1
_
= 0.
IN THREE dimensions there is a similar gadget.
20
The Basic Box is dened by

i

k, and we can let any


3-dimensional L.T. act on it, to get a new box de-
ned by T

i, T

j, T

k. We dene
det T =
_
T

i
_

_
T

j
_

_
T

k
_
where the dot is the scalar product, as usual. Since
|T

i T

j T

k| is the volume of the 3-dimensional


parallelogram dened by T

i, T

j, T

k, we see that
| det T| =
Final Volume of Basic Box
Initial Volume of Basic Box
,
that is, | det T| tells you how much T changes vol-
umes. If T squashes the Basic Box at, then
det T = 0.
Just as det
_
a b
c d
_
= ad bc, there is a formula
for the determinant of a 3 by 3 matrix. The usual
notation is this. We DEFINE

a b
c d

= det
_
a b
c d
_
= ad bc.
21
Similarly

a
11
a
12
a
13
a
21
a
22
a
23
a
31
a
32
a
33

is the determinant of
_
_
a
11
a
12
a
13
a
21
a
22
a
23
a
31
a
32
a
33
_
_
and there is a formula for it, as
follows:

a
11
a
12
a
13
a
21
a
22
a
23
a
31
a
32
a
33

= a
11

a
22
a
23
a
32
a
33

a
12

a
21
a
23
a
31
a
33

+a
13

a
21
a
22
a
31
a
32

.
In other words, we can compute a three-dimensional
determinant if we know how to work out 2-dimensional
determinants.
COMMENTS:
[a] We worked along the top row. Actually, a THE-
OREM says that you can use ANY ROW OR ANY
COLUMN!
[b] How did I know that a
12
had to multiply the par-
ticular 2-dimensional determinant

a
21
a
23
a
31
a
33

? Easy:
22
I just struck out EVERYTHING IN THE SAME
ROW AND COLUMN as a
12
:


a
21
a
23
a
31
a
33

and
just kept the survivors!
This is the pattern, for example if you expand along
the second row you will get
a
21

a
12
a
13

a
32
a
33

+a
22

a
11
a
13

a
31
a
33

a
23

a
11
a
12


a
31
a
32

[c] What is the pattern of the + and signs? It is


an (Ang Moh ) CHESSBOARD, starting with a +
in the top left corner:
+ +
+
+ +
[d] You can do exactly the same thing in FOUR di-
23
mensions, following this pattern, using
+ +
+ +
+ +
+ +
because now you know how to work out 3-dimensional
determinants. And so on!
Example:

1 1 0
1 1 1
2 0 0

= 1

1 1
0 0

1 1
2 0

+ 0

1 1
2 0

= 0 + 2 + 0 = 2
(expanding along the top row) or, if you use the sec-
ond row,

1 1 0
1 1 1
2 0 0

= 1

1 0
0 0

+ 1

1 0
2 0

1 1
2 0

= 0 + 0 + 2 = 2
24
or

1 1 0
1 1 1
2 0 0

= 1

1 1
0 0

1 0
0 0

+ 2

1 0
1 1

= 0 + 0 + 2 = 2
(expanding down the rst column).
Important Properties of Determinants
[a] Let S and T be two linear transformations such
that det S and det T are dened. Then
det ST = det TS = (det S) (det T).
Therefore, det[STU] = det[UST] = det[TUS] and
so on: det doesnt care about the order. Remember
however that this DOES NOT mean that STU =
UST etc etc.
25
[b] If M is a square matrix, then
det M
T
= det M.
[c] If c is a number and M is an n by n matrix, then
det(cM) = c
n
det M.
EXAMPLE: Remember from Section 2[g] of Chap-
ter 5 that an ORTHOGONAL matrix satises MM
T
=
I. So det(MM
T
) = det I = 1. But det(MM
T
) =
det(M) det(M
T
) = det(M) det(M) = (det M)
2
,
thus
det M = 1
for any orthogonal matrix.
6.5. INVERSES.
If I give you a 3-dimensional vector

u and a 3-
dimensional linear transformation T, then T sends
26

u to a particular vector, it never sends



u to two
DIFFERENT VECTORS! So this picture is impos-
sible:
T

u
T

u
But what about this picture:

u
T

u = T

v
Can T send TWO DIFFERENT VECTORS TO
ONE? Yes!
_
_
1 0 0
0 0 0
0 0 0
_
_
_
_
0
1
0
_
_
=
_
_
0
0
0
_
_
=
_
_
1 0 0
0 0 0
0 0 0
_
_
_
_
0
0
1
_
_
So it can happen! Notice that this transformation
destroys

j (and also

k). In fact if

u =

v and
27
T

u = T

v , then T(

v ) = 0, that is, T

w =

0
where

w IS NOT THE ZERO VECTOR. So if this
happens, T destroys everything in the

w direction.
That is, T SQUASHES 3-dimensional space down
to two or even less dimensions. This means that
T LOSES INFORMATION it throws away all of
the information stored in the

w direction. Clearly T
squashes the basic box down to zero volume, so
det T = 0
and we say T is SINGULAR.
SUMMARY: A SINGULAR LINEAR TRANSFOR-
MATION
[a] Maps two dierent vectors to one vector
[b] Destroys all of the vectors in at least one direction
[c] Loses all information associated with those direc-
tions
[d] Satises det T = 0.
28
Conversely, a NON-SINGULAR transformation never
maps 2 vectors to one,

u T

v T

v
_
_
_
Dierent
Therefore if I give you T

u, THERE IS EXACTLY
ONE

u. The transformation that takes you fromT

u
back to

u is called the INVERSE OF T. The idea is
that since a NON-SINGULAR linear transformation
does NOT destroy information, we can re-construct

u if we are given T

u. Clearly T HAS AN INVERSE,


CALLED T
1
, if and only if det T = 0. All this
works in other dimensions too.
EXAMPLE:
_
1 1
1 1
_ _
3
4
_
=
_
7
7
_
,
_
1 1
1 1
_ _
4
3
_
=
_
7
7
_
,
29
two dierent vectors to one!
_
1 1
1 1
_ _
1
1
_
=
_
1 1
1 1
_ _
2
2
_
=
_
1 1
1 1
_ _
13.59
13.59
_
=
_
0
0
_
It destroys everything
in that direction!
Finally det
_
1 1
1 1
_
= 0.
So it is SINGULAR and
has NO inverse.
_
1
1
_
EXAMPLE: Take
_
0 1
1 0
_
and suppose it acts on
_

_
and
_
a
b
_
and sends them to the same vector,
so
_
0 1
1 0
_ _

_
=
_
0 1
1 0
_ _
a
b
_
.
Then
_

_
=
_
b
a
_

= a
= b

_

_
=
_
a
b
_
30
so
_

_
and
_
a
b
_
are the same this transforma-
tion never maps dierent vectors to the same vector.
No vector is destroyed, no information is lost, noth-
ing gets squashed! And det
_
0 1
1 0
_
= 1, NON-
SINGULAR.
How to FIND THE INVERSE.
By denition, T
1
sends T

u to

u, i.e.
T
1
(T(

u)) =

u = T(T
1
(

u)).
But

u = I

u (identity) so T
1
satises
T
1
T = TT
1
= I.
So to nd the inverse of
_
0 1
1 0
_
we just have to
nd a matrix
_
a b
c d
_
such that
_
a b
c d
_ _
0 1
1 0
_
=
_
1 0
0 1
_
, b = 1, a = 0, d = 0, c = 1 so answer is
31
_
0 1
1 0
_
. In fact its easy to show that
_
a b
c d
_
1
=
1
ad bc
_
d b
c a
_
.
For example, when we needed to nd the matrix S
in Section 4 of Chapter 5, we needed to nd a way
of solving
S
_
0.7 0.4
0.5 0.7
_
= I.
This just means that we need to inverse of
_
0.7 0.4
0.5 0.7
_
,
and the above formula does the job for us.
For bigger square matrices there are many tricks
for nding inverses. A general [BUT NOT VERY
PRACTICAL] method is as follows:
[a] Work out the matrix of COFACTORS. [A cofac-
tor is what you get when you work out the smaller
determinant obtained by striking out a row and a
column, for example the cofactor of 6 in

1 2 3
4 5 6
7 8 9

32
is

1 2
7 8

= 6. You can do this for each element


in a given matrix, to obtain a new matrix of the
same size. For example, the matrix of cofactors of
_
_
1 0 1
0 1 0
0 0 1
_
_
is
_
_
1 0 0
0 1 0
1 0 1
_
_
.
[b] Keep or reverse the signs of every element accord-
ing to
+ +
+
+ +
(you get
_
_
1 0 0
0 1 0
1 0 1
_
_
above.)
[c] Take the TRANSPOSE,
_
_
1 0 1
0 1 0
0 0 1
_
_
.
[d] Divide by the determinant of the original ma-
trix. THE RESULT IS THE DESIRED INVERSE.
_
_
1 0 1
0 1 0
0 0 1
_
_
in this example. Check:
_
_
1 0 1
0 1 0
0 0 1
_
_
_
_
1 0 1
0 1 0
0 0 1
_
_
=
_
_
1 0 0
0 1 0
0 0 1
_
_
.
33
INVERSE OF A PRODUCT:
(AB)
1
= B
1
A
1
Note the order! Easily checked:
(AB)
1
AB = B
1
(A
1
A)B = B
1
IB = I.
APPLICATION: SOLVING LINEAR SYSTEMS.
Suppose you want to solve
x + 2y + 3z = 1
4x + 5y + 6z = 2
7x + 8y + 9z = 4.
One way is to write it as
_
_
1 2 3
4 5 6
7 8 9
_
_
_
_
x
y
z
_
_
=
_
_
1
2
4
_
_
.
34
Then all you have to do is nd
_
_
1 2 3
4 5 6
7 8 9
_
_
1
and
multiply it on both sides, so you get
_
_
x
y
z
_
_
=
_
_
1 2 3
4 5 6
7 8 9
_
_
1
_
_
1
2
4
_
_
= answer.
So this is a systematic way of solving such problems!
Now actually det
_
_
1 2 3
4 5 6
7 8 9
_
_
= 0, and you can see
why:
_
_
1 2 3
4 5 6
7 8 9
_
_
_
_
1
2
1
_
_
=
_
_
0
0
0
_
_
.
So this transformation destroys everything in the di-
rection of
_
_
1
2
1
_
_
. In fact it squashes 3-dimensional
space down to a certain 2-dimensional space. [We
say that the matrix has RANK 2. If it had squashed
everything down to a 1-dimensional space, we would
say that it had RANK 1.] Now actually
_
_
1
2
4
_
_
. DOES
35
NOT lie in that two-dimensional space. Since
_
_
1 2 3
4 5 6
7 8 9
_
_
squashes EVERYTHING into that two-dimensional
space, it is IMPOSSIBLE for
_
_
1 2 3
4 5 6
7 8 9
_
_
_
_
x
y
z
_
_
to be
equal to
_
_
1
2
4
_
_
. Hence the system has NO solutions.
If we change
_
_
1
2
4
_
_
to
_
_
1
2
3
_
_
, this vector DOES lie
in the special 2-dimensional space, and the system
_
_
1 2 3
4 5 6
7 8 9
_
_
_
_
x
y
z
_
_
=
_
_
1
2
3
_
_
DOES have a solution
in fact it has innitely many!
SUMMARY:
Any system of linear equations can be written as
M

r =

a
where M is a matrix,

r = the vector of variables,
36
and

a is a given vector. Suppose M is square.
[a] If det M = 0, there is exactly one solution,

r = M
1

a.
[b] If det M = 0, there is probably no solution. But
if there is one, then there will be many.
PRACTICAL ENGINEERING PERSPECTIVE:
In the REAL world, NOTHING IS EVER EXACTLY
EQUAL TO ZERO! So if det M = 0, either [a] you
have made a mistake, OR [b] you are pretending that
your data are more accurate than they really are!
_
_
1 2 3
4 5 6
7 8 9
_
_
REALLY means
_
_
1.01 2.08 3.03
3.99 4.97 6.02
7.01 7.96 8.98
_
_
and of course the determinant of THIS is non-zero!
Actually, det = 0.597835!
37
6.6 EIGENVECTORS AND EIGENVALUES.
Remember we said that a linear transformation
USUALLY changes the direction of a vector. But
there may be some special vectors which DONT
have their direction changed!
EXAMPLE:
_
1 2
2 2
_
clearly DOES change the
direction of

i and

j, since
_
1
2
_
is not parallel to

i
and
_
2
2
_
is not parallel to

j. BUT
_
1 2
2 2
_ _
2
1
_
=
_
4
2
_
= 2
_
2
1
_
which IS parallel to
_
2
1
_
.
In general if a transformation T does not change the
direction of a vector

u, that is
T

u =

u
38
for some (SCALAR), then

u is called an EIGEN-
VECTOR of T. The scalar is called the EIGEN-
VALUE of

u.
6.7 FINDING EIGENVALUES AND EIGEN-
VECTORS.
There is a systematic way of doing this. Take the
equation
T

u =

u
and write

u = I

u, I = identity. Then
(T I)

u =

0
Lets suppose

u =

0
[of course,

0
is always an eigen-
vector, that is boring]. So the equation says that
T I SQUASHES everything in the

u direction.
Hence
det(T I) = 0.
39
This is an equation which can be SOLVED to nd
.
EXAMPLE: Find the eigenvalues of
_
1 2
2 2
_
:
det
__
1 2
2 2
_

_
1 0
0 1
__
= 0
det
_
1 2
2 2
_
= 0
(1 )(2 +) 4 = 0
= 2 OR 3
So there are TWO answers for a 2 by 2 matrix. Sim-
ilarly, in general there are three answers for 3 by 3
matrices, etc.
What are the eigenvectors for = 2, = 3?
IMPORTANT POINT: Let

u be an eigenvector
of T. Then 2

u is also an eigenvector with the same


eigenvalue!
T(2

u) = 2T

u = 2

u = (2

u).
40
Similarly 3

u, 13.59

u etc are all eigenvectors! SO


YOU MUST NOT EXPECT A UNIQUE ANSWER!
OK, with that in mind, lets nd an eigenvector for
= 2. Lets call an eigenvector
_

_
. Then
(T I)

u =

0

_
1 2
2 2
_ _

_
= 0

_
1 2
2 4
_ _

_
= 0
+ 2 = 0
2 4 = 0
But these equations are actually the SAME, so we
really only have ONE equation for 2 unknowns. We
arent surprised, because we did not expect a unique
answer anyway! We can just CHOOSE = 1 (or
13.59 or whatever) and then solve for . Clearly
=
1
2
, so an eigenvector corresponding to = 2
is
_
1
1
2
_
. But if you said
_
2
1
_
or
_
100
50
_
that is also
41
correct!
What about = 3?
_
4 2
2 1
_ _

_
= 0
4 + 2 = 0
2 + = 0
Again we can set = 1, then = 2, so an eigen-
vector corresponding to = 3 is
_
1
2
_
or
_
2
4
_
or
_
10
20
_
etc.
EXAMPLE: Find the eigenvalues, and correspond-
ing eigenvectors, of
_
0 1
1 0
_
.
Answer: We have det
_
1
1
_
= 0
2
+ 1 =
0 = i, i =

1.
Eigenvector for i: we set
_
i 1
1 i
_ _
1

_
= 0
42
i = 0 = i so an eigenvector for i is
_
1
i
_
. For = i we have
_
i 1
1 i
_ _
1

_
= 0
i = 0 = i so an eigenvector for i
is
_
1
i
_
. Note that a REAL matrix can have COM-
PLEX eigenvalues and eigenvectors! This is hap-
pening simply because
_
0 1
1 0
_
is a ROTATION
through 90

, and of course such a transformation


leaves NO [real] vectors direction unchanged (apart
from the zero vector).
6.8. DIAGONAL FORM OF A LINEAR TRANS-
FORMATION.
Remember that we dened the matrix of a linear
transformation T WITH RESPECT TO

i,

j by let-
ting T act on

i and

j and then putting the results in
43
the columns. So to say that T has matrix
_
a b
c d
_
with respect to

i,

j means that
T

i = a

i +c

j
T

j = b

i +d

j.
Whats so special about the two vectors

i and

j?
Nothing, except that EVERY vector in two dimen-
sions can be written as

i +

j for some , .
Now actually we only really use

i and

j for CONVE-
NIENCE. In fact, we can do this with ANY pair of
vectors

u,

v in two dimensions,
PROVIDED that they are not parallel.
That is, any vector

w can be
expressed as

w =

u +

v
for some scalars , . You can see this from the
diagram by stretching

u to

u and

v to

v , we
can make their sum equal to

w.
44
We call

u,

v a BASIS for 2-dimensional vectors. Let

u = P
11

i +P
21

j =
_
P
11
P
21
_

v = P
12

i +P
22

j =
_
P
12
P
22
_
Then the transformation that takes
_

i,

j
_
to (

u,

v )
has matrix
_
P
11
P
12
P
21
P
22
_
= P. In order for

u,

v to be
a basis, P must not squash the volume of the Basic
Box down to zero, since otherwise

u and

v will be
parallel. So we must have
det P = 0.
The same idea works in 3 dimensions: ANY set of
3 vectors forms a basis PROVIDED that the matrix
of components satises det P = 0.
EXAMPLE: The pair of vectors

u =
_
1
0
_
,

v =
_
1
1
_
forms a basis, because det
_
1 1
0 1
_
= 1 = 0.
45
Now of course the COMPONENTS of a vector will
change if you choose a dierent basis. For example,
_
1
2
_
= 1

i + 2

j BUT
_
1
2
_
= 1

u + 2

v .
Instead,
_
1
2
_
=

u +2

v , so the components of this


vector relative to

u,

v are
_
1
2
_
(

u,

v )
. Where did I
get these numbers?
As usual, set

u = P

i,

v = P

j where P =
_
1 1
0 1
_
.
We want to nd , such that
_
1
2
_
=

u +

v .
We have, in this particular case,

u =

i,

v =

i +

j, so
_
1
2
_
=

i +[

i +

j] =
_
+

_
= P
_

_
We know P is not singular, so we can take P over to
the left side by multiplying both sides of this equa-
46
tion by the inverse of P. So we get
_

_
= P
1
_
1
2
_
and this is our answer: this is how we nd and !
So to get and we just have to work out
P
1
_
1
2
_
=
_
1 1
0 1
_ _
1
2
_
=
_
1
2
_
,
that is, the components of this vector relative to

u,

v
are found as
_
1
2
_
(

u,

v )
= P
1
_
1
2
_
(

i,

j)
THE COMPONENTS RELATIVE TO

u,

v ARE
OBTAINED BY MULTIPLYING P
1
INTO THE
COMPONENTS RELATIVE TO

i,

j. Similarly for
linear transformations if a certain linear transfor-
mation T has matrix
_
1 2
0 1
_

j
relative to

i,

j it
47
will have a DIFFERENT matrix relative to

u,

v .
We have
_
1 2
0 1
_
(

i,

j)
_
1
2
_
(

i,

j)
=
_
5
2
_
(

i,

j)
That is, the matrix of T relative to

i,

j sends
_
1
2
_
(

j)
to
_
5
2
_
(

j)
. In the same way, the matrix of T rela-
tive to (

u,

v ), which we dont know and want to nd,


sends
_
1
2
_
(

u,

v )
to
_
7
2
_
(

u,

v )
, because these are
the components of these two vectors relative to

u,

v ,
as you can show by multiplying P
1
into
_
1
2
_
(

j)
and
_
5
2
_
(

j)
respectively.
So the unknown matrix we want satises
_
? ?
? ?
_
(

u,

v )
_
1
2
_
(

u,

v )
=
_
7
2
_
(

u,

v )
48
But we know
_
1
2
_
(

u,

v )
= P
1
_
1
2
_
(

i,

j)
and
_
7
2
_
(

u,

v )
= P
1
_
5
2
_
(

i,

j)
so
_
? ?
? ?
_
(

u,

v )
P
1
_
1
2
_
(

j)
= P
1
_
5
2
_
(

j)
.
Multiply both sides by P and get
P
_
? ?
? ?
_
(

u,

v )
P
1
_
1
2
_
(

j)
=
_
5
2
_
(

j)
Compare this with
_
1 2
0 1
_
(

i,

j)
_
1
2
_
(

i,

j)
=
_
5
2
_
(

i,

j)

_
1 2
0 1
_
(

i,

j)
= P
_
? ?
? ?
_
(

u,

v )
P
1

_
? ?
? ?
_
(

u,

v )
= P
1
_
1 2
0 1
_
(

i,

j)
P.
49
[In the last step, we multiplied both sides on the
LEFT by P
1
, and on the RIGHT by P.]
We conclude that THE MATRIX OF T REL-
ATIVE TO

u,

v , IS OBTAINED BY MULTIPLY-
ING P
1
ON THE LEFT AND P ON THE RIGHT
INTO THE MATRIX OF T RELATIVE TO

i,

j.
In this example,
_
? ?
? ?
_
(

u,

v )
=
_
1 1
0 1
_ _
1 2
0 1
_ _
1 1
0 1
_
=
_
1 1
0 1
_ _
1 3
0 1
_
=
_
1 4
0 1
_
.
So now we know how to work out the matrix of any
linear transformation relative to ANY basis.
Now let T be a linear transformation in 2 dimensions,
with eigenvectors

e
1
,

e
2
, eigenvalues
1
,
2
. Now

e
1
and

e
2
may or may not give a basis for 2-dimensional
space. But suppose they do.
50
QUESTION: What is the matrix of T relative to

e
1
,

e
2
?
ANSWER: As always, we see what T does to

e
1
and

e
2
, and put the results into the columns!
By denition of eigenvectors and eigenvalues,
T

e
1
=
1

e
1
=
1

e
1
+ 0

e
2
T

e
2
=
2

e
2
= 0

e
1
+
2

e
2
So the matrix is
_

1
0
0
2
_
(

e
1
,

e
2
)
.
We say that a matrix of the form
_
a 0
0 d
_
or
_
_
0 0
0 0
0 0
_
_
is DIAGONAL. So we see that THE MATRIX OF
A TRANSFORMATION RELATIVE TO ITS OWN
EIGENVECTORS (assuming that these form a ba-
sis) is DIAGONAL.
EXAMPLE: We know that the eigenvectors of
_
1 2
2 2
_
are
_
1
1
2
_
and
_
1
2
_
. So here P =
_
1 1
1
2
2
_
,
51
P
1
=
2
5
_
2 1

1
2
1
_
,
P
1
_
1 2
2 2
_
P =
2
5
_
2 1

1
2
1
_ _
1 2
2 2
_ _
1 1
1
2
2
_
=
2
5
_
2 1

1
2
1
_ _
2 3
1 6
_
=
2
5
_
5 0
0
15
2
_
=
_
2 0
0 3
_
as expected since the
eigenvalues are 2 and -3.
EXAMPLE: The shear matrix
_
1 tan
0 1
_
.
Eigenvalues: det
_
1 tan
0 1
_
= 0 (1)
2
=
0 = 1. Only one eigenvector, namely
_
1
0
_
, so
the eigenvectors DO NOT give us a basis in this case
NOT possible to diagonalize this matrix!
52
6.9 APPLICATION MARKOV CHAINS.
We saw back in Section 3 of Chapter 5 that to predict
the weather 4 days from now, we needed the 4th
power of the matrix
_
0.6 0.3
0.4 0.7
_
.
But suppose I want the weather 30 days from now
I need M
30
! There is an easy way to work this
out using eigenvalues.
Suppose I can diagonalize M, that is, I can write
P
1
MP = D =
_

1
0
0
2
_
for some matrix P. Then
M = PDP
1
M
2
= (PDP
1
)(PDP
1
)
= PDP
1
PDP
1
= PD
2
P
1
M
3
= MM
2
= PDP
1
PD
2
P
1
= PD
3
P
1
etc
M
30
= PD
30
P
1
.
53
But D
30
is very easy to work out it is just
_

30
1
0
0
30
2
_
.
Lets see how this works!
Eigenvectors and eigenvalues of
_
0.6 0.3
0.4 0.7
_
are
_
1
1
_
(eigenvalue 0.3) and
_
1
4
3
_
(eigenvalue 1) so
P =
_
1 1
1
4
3
_
, D =
_
0.3 0
0 1
_
, P
1
=
_
4
7

3
7
3
7
3
7
_
D
30
=
_
(0.3)
30
0
0 1
_

_
2 10
16
0
0 1
_
so
M
30
=
_
1 1
1
4
3
_ _
2 10
16
0
0 1
_ _
4
7

3
7
3
7
3
7
_
=
1
7
_
3 + 8 10
16
3 6 10
16
4 8 10
16
4 + 6 10
16
_

_
3
7
3
7
4
7
4
7
_
So if it is rainy today, the probability of rain tomor-
row is 60%, but the probability of rain 30 days from
54
now is only
3
7
43%. As we go forward in time, the
fact that it rained today becomes less and less im-
portant! The probability of rain in 31 days is almost
the same as the probability of rain in 30 days!
6.10 THE TRACE OF A MATRIX.
Let M be any square matrix. Then the TRACE
of M, denoted TrM, is dened as the sum of the
diagonal entries: Tr
_
1 0
0 1
_
= 2, Tr
_
_
1 2 3
4 5 6
7 8 9
_
_
=
15, Tr
_
_
1 5 16
7 2 15
11 9 8
_
_
= 11, etc.
In general it is NOT true that Tr(MN) = TrM TrN
BUT it is true that TrMN = TrNM.
Proof: TrM =

i
M
ii
so
TrMN =

j
M
ij
N
ji
=

i
N
ji
M
ij
= TrNM.
55
Hence Tr(P
1
AP) = Tr(APP
1
) = TrA so if A is
diagonalizable, TrA = Tr
_

1
0
0
2
_
=
1
+
2
.
56