Linear Algebra I

Course No. 100 221

Fall 2006
Michael Stoll
With corrections, version of December 10, 2007
Contents
1. Preliminaries 2
2. What Is Linear Algebra? 2
3. Vector Spaces 4
4. Fields 7
5. Linear Subspaces and Linear Hulls 10
6. Linear Independence and Dimension 14
7. Digression: Innite-Dimensional Vector Spaces
and Zorns Lemma 23
8. Linear Maps 25
9. Quotient Spaces 32
10. Digression: Finite Fields 35
11. Matrices 37
12. Computations with Matrices: Row and Column Operations 41
13. Linear Equations 47
14. Matrices and Linear Maps 50
15. Determinants 54
16. The Determinant and the Symmetric Group 60
17. Eigenvalues and Eigenvectors 62
References 68
2
1. Preliminaries
In the following, we will assume that you are familiar with the basic notions and
notations of (naive) set theory. This includes, for example, the following.
Notation for sets: 1, 2, 3, x S : some property of x, the empty set .
Symbols for the set of natural numbers:
N = 1, 2, 3, . . . , N
0
= 0, 1, 2, 3, . . . ,
integers: Z, rational numbers: Q, real numbers: R, and complex numbers:
C (which we will introduce shortly).
Notations x S (x is an element of S), A B (intersection of A and B),
A B (union of A and B) A B = x A : x / B (set dierence of A
and B), A B (A is a subset of B; this includes the case A = B my
convention), A B (A is a proper subset of B).
Pairs (a, b), triples (a, b, c) and general n-tuples (a
1
, a
2
, . . . , a
n
), and carte-
sian products A B = (a, b) : a A, b B etc.
The notion of a map f : A B and properties like injectivity, surjectivity,
bijectivity (we will recall these when we need them).
The notion of an equivalence relation on a set S: formally, a relation on S
is a subset R of SS. For simplicity, we write a b for (a, b) R. Then R
is an equivalence relation if it is reexive (a a for all a S), symmetric
(if a b, then b a), and transitive (if a b and b c, then a c).
Given an equivalence relation on S, the set S has a natural decomposition
into equivalence classes, and we can consider the quotient set S/: the set
of all equivalence classes. We will recall this in more detail later.
We will also use notation like the following.
n
_
i=1
A
i
= A
1
A
2
A
n
with
0
_
i=1
A
i
=
n

i=1
A
i
= A
1
A
2
A
n
;
0

i=1
A
i
usually does not make sense
_
AS
A =
_
o is the union of all the sets A that are elements of o
n

i=1
a
i
= a
1
+ a
2
+ + a
n
and is zero for n = 0
n

i=1
a
i
= a
1
a
2
a
n
and is one for n = 0
2. What Is Linear Algebra?
Linear Algebra is the theory of linear structures. So what is a linear structure?
Well, the notion of linearity involves addition you want to be able to form sums,
and this addition should behave in the way you expect (it is commutative (x+y =
y + x) and associative ((x + y) + z = x + (y + z)), has a zero element, and every
element has a negative) and multiplication, not between the elements of your
structure, but by scalars, for example, real numbers. This scalar multiplication
should also behave reasonably (you want to have 1 x = x and ( x) = () x)
3
and be compatible with the addition (in the sense that both distributive laws hold:
( + )x = x + x, (x + y) = x + y). We will make this precise in the next
section.
A set with a linear structure in the sense of our discussion is called a linear space or
vector space. So Linear Algebra studies these linear spaces and the maps between
them that are compatible with the linear structure: linear maps. This may sound
somewhat abstract, and indeed, it is. However, it is exactly this level of abstraction
that makes Linear Algebra an extremely useful tool. The reason for this is that
linear structures abound in mathematics, and so Linear Algebra has applications
everywhere (see below). It is this method of abstraction that extracts the common
features of various situations to create a general theory, which forms the basis of
all essential progress in mathematics. It took the mathematicians quite a long
time before they came up with our modern clean formulation of the theory.
As already hinted at above, many other structures in mathematics are built on
linear spaces, usually by putting some additional structure on them. Here are
some examples. Even though the various notions probably do not have much of
a meaning to you now, this list should show you that you can expect to use what
you will learn in this course quite heavily in your further studies.
Geometry. First of all, there is of course the elementary analytic geometry
of the plane and space, which is based on the fact that plane and space can
be turned into linear spaces by choosing a point as the origin. Usually one
adds some structure that allows to talk about distances and angles. This
is based on the dot product, which is a special case of an inner product,
which turns a linear space into a Euclidean vector space. We will discuss
this in detail later in the course.
A more advanced branch of geometry is Dierential Geometry that stud-
ies smooth curves, surfaces and higher-dimensional manifolds. Every
point on such a manifold has a tangent space, which is a linear space,
and there are many other linear spaces associated in a natural way to a
manifold (for example, spaces of dierential forms).
Analysis. The notion of derivative is based on linearity the derivative
describes the best linear approximation to a function in a given point.
In the case of functions of one variable, this comes down to the slope of
the tangent line to the graph. However, the concept can be generalized
to functions in several variables (which are functions dened on a vector
space!) and even to functions between manifolds. Also the operations
of dierentiating and integrating functions are linear. Finally, spaces of
functions usually carry a structure as a linear space (for example, the
space of dierentiable functions on the real line). Functional Analysis is
concerned with these function spaces; usually one adds some structure
by introducing a suitable norm, leading to Banach Spaces and Hilbert
Spaces.
Algebra. Algebra studies more general algebraic structures (like groups,
rings, elds), many of which are based on linear spaces, like for example
Lie Algebras. Then there is a whole bunch of so-called homology and
cohomology theories involving (co-)homology groups, which are in may
cases linear spaces. Such groups occur, for example, in Algebraic Topology,
where they can be used to show that two spaces are essentially distinct from
each other, in that one cannot continuously deform one into the other.
They also play an important role in Algebraic Geometry.
4
Another, somewhat dierent, application (which may be interesting to
the EECS, CS and ECE students) is to Coding Theory, where one tries
to construct good error-correcting codes. Essentially all the codes that
are considered are linear codes, which means that the codewords form a
vector space (where the scalar multiplication is not by real numbers, but
by elements from a nite eld like the eld F
2
= 0, 1).
To illustrate the versatility of Linear Algebra, let me write down a few linear
equations from various areas of mathematics. Even though they involve objects
of quite dierent nature, the general theory applies to all of them to provide (for
example) general information on the structure of the solution set.
A system of linear equations in real numbers:
x + y = 3 , x + z = 5 , y + z = 6 .
A linear recurrence relation for a sequence (F
n
)
n0
of real numbers (the
Fibonacci numbers provide one solution):
F
n+2
= F
n+1
+ F
n
for all n 0.
A linear ordinary dierential equation for a (twice dierentiable) function
y : R R (the sine and cosine functions solve it):
y

+ y = 0 .
A linear partial dierential equation for a (twice dierentiable) function f
of time t and space x, y, z (the Heat Equation):
f
t
=

2
f
x
2


2
f
y
2


2
f
z
2
.
The fact that Linear Algebra applies to all of them accounts for the general feeling
that linear equations are easy, whereas nonlinear equations are dicult.
3. Vector Spaces
In this section, we will give the complete formal denition of what a (real) vector
space or linear space is. Real here refers to the fact that the scalars are real
numbers.
3.1. Denition. A real vector space or linear space over R is a set V , together
with two maps + : V V V (addition) and : R V V (scalar multi-
plication) satisfying the following axioms. Note that for x, we usually write
x.
(1) For all x, y V , x + y = y + x (addition is commutative).
(2) For all x, y, z V , (x + y) + z = x + (y + z) (addition is associative).
(3) There is a 0 V such that for all x V , x + 0 = x (existence of a zero
element).
(4) For every x V , there is x V such that x + (x) = 0 (existence of
negatives).
(5) For all , R and x V , ( x) = () x (scalar multiplication is
associative).
5
(6) For all x V , 1 x = x (multiplication by 1 is the identity).
(7) For all R and x, y V , (x + y) = x + y (distributivity I).
(8) For all , R and x V , ( + )x = x + x (distributivity II).
The elements of a vector space are usually called vectors.
3.2. Remarks.
(1) The rst four axioms above exactly state that (V, +) is an (additive) abelian
group. (If you didnt know what an abelian group is, then this is the
denition.)
(2) Instead of writing (V, +, ) (which is the complete data for a vector space),
we usually just write V , with the addition and scalar multiplication being
understood.
Before we continue with the general theory, we should look at some examples.
3.3. Example. The simplest (and perhaps least interesting) example of a vector
space is V = 0, with addition given by 0 + 0 = 0 and scalar multiplication by
0 = 0 (these are the only possible choices). Trivial as it may seem, this vector
space, called the zero space, is important. It plays a role in Linear Algebra similar
to the role played by the empty set in Set Theory.
3.4. Example. The next (still not very interesting) example is V = R, with
addition and multiplication the usual ones on real numbers. The axioms above in
this case just reduce to well-known rules for computing with real numbers.
3.5. Example. Now we come to a very important example, which is the model
of a vector space. We consider V = R
n
, the set of n-tuples of real numbers. We
dene addition and scalar multiplication component-wise:
(x
1
, x
2
, . . . , x
n
) + (y
1
, y
2
, . . . , y
n
) = (x
1
+ y
1
, x
2
+ y
2
, . . . , x
n
+ y
n
)
(x
1
, x
2
, . . . , x
n
) = (x
1
, x
2
, . . . , x
n
)
Of course, we now have to prove that our eight axioms are satised by our choice
of (V, +, ). In this case, this is very easy, since everything reduces to known facts
about real numbers. As an example, let us show that the rst distributive law is
satised.

_
(x
1
, x
2
, . . . , x
n
) + (y
1
, y
2
, . . . , y
n
)
_
= (x
1
+ y
1
, x
2
+ y
2
, . . . , x
n
+ y
n
)
=
_
(x
1
+ y
1
), (x
2
+ y
2
), . . . , (x
n
+ y
n
)
_
= (x
1
+ y
1
, x
2
+ y
2
, . . . , x
n
+ y
n
)
= (x
1
, x
2
, . . . , x
n
) + (y
1
, y
2
, . . . , y
n
)
= (x
1
, x
2
, . . . , x
n
) + (y
1
, y
2
, . . . , y
n
)
Of course, for n = 2 and n = 3, this is more or less what you know as vectors
from high school. For n = 1, this example reduces to the previous one (if one
identies 1-tuples (x) with elements x); for n = 0, it reduces to the zero space.
(Why? Well, like an empty product of numbers should have the value 1, an empty
product of sets like R
0
has exactly one element, the empty tuple (), which we can
call 0 here.)
6
3.6. Exercise. Write down proofs for the other seven axioms.
3.7. Example. The preceding example can be generalized. Note that we can
identify R
n
with the set of maps f : 1, 2, . . . , n R, by letting such a map f
correspond to the n-tuple (f(1), f(2), . . . , f(n)). But there is no need to limit
ourselves to these specic sets. So let us consider any set X and look at the set
of all maps (or functions) from X to R:
V = R
X
= f : X R .
In order to get a vector space, we have to dene addition and scalar multiplication.
To dene addition, for every pair of functions f, g : X R, we have to dene a
new function f + g : X R. The only reasonable way to do this is as follows
(point-wise):
f + g : X R, x f(x) + g(x) ,
or, in a more condensed form, by writing (f +g)(x) = f(x)+g(x). (Make sure that
you understand these notations!) In a similar way, we dene scalar multiplication:
f : X R, x f(x) .
We then have to check the axioms in order to convince ourselves that we really
get a vector space. Let us do again the rst distributive law as an example. We
have to check that (f + g) = f + g, which means that for all x X, we have
_
(f + g)
_
(x) = (f + g)(x) .
So let R, f, g : X R, and x X.
_
(f + g)
_
(x) =
_
(f + g)(x)
_
=
_
f(x) + g(x)
_
= f(x) + g(x)
= (f)(x) + (g)(x)
= (f + g)(x) .
Note the parallelism of this proof with the one in the previous example! What do
we get when X is the empty set?
3.8. Exercise. Write down proofs for the other seven axioms.
Before we can continue, we have to deal with a few little things.
3.9. Remark. In a vector space V , there is only one zero element.
Proof. The way to show that there is only one element with a given property is to
assume there are two and then to show they are equal. So assume 0 and 0

were
two zero elements of V . Then we must have
0 = 0 + 0

= 0

+ 0 = 0

(using axioms (3) for 0, (1), (3) for 0

, respectively), so 0 = 0

.
3.10. Remark. In a vector space V , there is a unique negative for each element.
Proof. Let x V and assume that a, b V are both negatives of x, i.e., x+a = 0,
x + b = 0. Then
a = a + 0 = a + (x + b) = (a + x) + b = (x + a) + b = 0 + b = b + 0 = b ,
so a = b.
7
3.11. Notation. As usual, we write x y for x + (y).
Here are some more harmless facts.
3.12. Remarks. Let (V, +, ) be a real vector space.
(1) For all x V , we have 0 x = 0.
(2) For all x V , we have (1) x = x.
(3) For R and x V such that x = 0, we must have = 0 or x = 0.
Proof. Exercise.
4. Fields
So far, we have required our scalars to be real numbers. It turns out, however,
that we can be quite a bit more general without adding any complications, by
replacing the real numbers with other structures with similar properties. These
structures are called elds.
In the formulation of the axioms below, we will use the shorthands x X : . . .
and x X : . . . for the phrases for all x X, we have . . . and there exists
some x X such that . . . .
4.1. Denition. A eld is a set F, together with two maps + : F F F
(addition) and : F F F (multiplication), written, as usual, (, ) +
and (, ) or , respectively, satisfying the following axioms.
(1) , F : + = + .
(2) , , F : ( + ) + = + ( + ).
(3) 0 F F : + 0 = .
(4) F F : + () = 0.
(5) , F : = .
(6) , , F : () = ().
(7) 1 F : 1 ,= 0 and F : 1 = .
(8) F 0
1
F :
1
= 1.
(9) , , F : ( + ) = + .
Well-known examples of elds are the eld Q of rational numbers and the eld R
of real numbers. We will introduce the eld C of complex numbers shortly. But
there are other examples of elds. For example, the smallest possible eld just
has the required elements 0 and 1, with the only interesting denition being that
1 + 1 = 0.
4.2. Exercise. Check the axioms for this eld F
2
= 0, 1.
As before for vector spaces, we have some simple statements that easily follow
from the axioms.
8
4.3. Remarks. Let F be a eld.
(1) The zero 0 and unit 1 of F are uniquely determined. Likewise, for every
F, is uniquely determined, and for every F 0,
1
is
uniquely determined.
(2) We have 0 = 0 and (1) = for all F. In particular (taking
= 1), (1)(1) = 1.
(3) If = 0 for , F, then = 0 or = 0 (or both).
Proof. Exercise.
4.4. Remark. It is perhaps easier to remember the eld axioms in the following
way.
(1) (F, +) is an (additive) abelian group with zero element 0, called the additive
group of F.
(2) (F 0, ) is a (multiplicative) abelian group, called the multiplicative
group of F. (But note that the multiplication map is dened on all of F,
not just on F 0.)
(3) The distributive law holds: , , F : ( + ) = + .
Given this, we can now dene the notion of an F-vector space for an arbitrary
eld F.
4.5. Denition. Let F be a eld. An F-vector space or linear space over F is a
set V , together with two maps + : V V V (addition) and : F V V
(scalar multiplication), satisfying the vector space axioms given in Def. 3.1 with
R replaced by F.
Note that in order to prove the statements in Remarks 3.9, 3.10 and 3.12, we only
had to use that R is a eld. Hence these statements are valid for F-vector spaces
in general.
The same observation applies to our examples of real vector spaces.
4.6. Examples.
(1) V = 0, with the unique addition and scalar multiplication maps, is an
F-vector space, again called the zero space (over F).
(2) F, with the addition and multiplication from its structure as a eld, is an
F-vector space.
(3) F
n
, with addition and scalar multiplication dened component-wise, is an
F-vector space.
(4) For any set X, the set F
X
of all maps X F, with addition and scalar
multiplication dened point-wise, is an F-vector space.
4.7. Example. There are other examples that may appear more strange. Let X
be any set, and let V be the set of all subsets of X. (For example, if X = a, b,
then V has the four elements , a, b, a, b.) We dene addition on V as the
symmetric dierence: A+B = (A B) (B A) (this is the set of elements of X
that are in exactly one of A and B). We dene scalar multiplication by elements
of F
2
in the only possible way: 0 A = , 1 A = A. These operations turn V into
an F
2
-vector space.
9
To prove this assertion, we can check the vector space axioms (this is an instructive
exercise). An alternative (and perhaps more elegant) way is to note that subsets
of X correspond to maps X F
2
(a map f corresponds to the subset x X :
f(x) = 1) there is a bijection between V and F
X
2
and this correspondence
translates the addition and scalar multiplication we have dened on V into that
we had dened earlier on F
X
2
.
4.8. The Field of Complex Numbers. Besides real vector spaces, complex
vector spaces are very important in many applications. These are linear spaces
over the eld of complex numbers, which we now introduce.
The rst motivation for the introduction of complex numbers is a shortcoming of
the real numbers: while positive real numbers have real square roots, negative real
numbers do not. Since it is frequently desirable to be able to work with solutions
to equations like x
2
+ 1 = 0, we introduce a new number, called i, that has the
property i
2
= 1. The set C of complex numbers then consists of all expressions
a + bi, where a and b are real numbers. (More formally, one considers pairs of
real numbers (a, b) and so identies C with R
2
as sets.) In order to turn C into a
eld, we have to dene addition and multiplication. If we want the multiplication
to be compatible with the scalar multiplication on the real vector space R
2
, then
(bearing in mind the eld axioms) there is no choice: we have to set
(a + bi) + (c + di) = (a + c) + (b + d)i
and
(a + bi)(c + di) = ac + adi + bci + bdi
2
= (ac bd) + (ad + bc)i
(remember i
2
= 1). It is then an easy, but tedious, matter to show that the
axioms hold. (Later, in the Introductory Algebra course, you will learn that
there is a rather elegant way of doing this.)
If z = a + bi as above, then we call Re z = a the real part and Imz = b the
imaginary part of z.
The least straightforward statement is probably the existence of multiplicative
inverses. In this context, it is advantageous to introduce the notion of conjugate
complex number.
4.9. Denition. If z = a +bi C, then the complex conjugate of z is z = a bi.
Note that z z = a
2
+ b
2
0. We set [z[ =

z z; this is called the absolute value

or modulus of z. It is clear that [z[ = 0 only for z = 0; otherwise [z[ > 0. We
obviously have

z = z and [ z[ = [z[.
4.10. Remark.
(1) For all w, z C, we have w + z = w + z and wz = w z.
(2) For all z C 0, we have z
1
= [z[
2
z.
(3) For all w, z C, we have [wz[ = [w[ [z[.
Proof.
(1) Exercise.
(2) First of all, [z[ , = 0, so the expression makes sense. Now note that
[z[
2
z z = [z[
2
z z = [z[
2
[z[
2
= 1 .
(3) Exercise.
10

For example:
1
1 + 2i
=
1 2i
(1 + 2i)(1 2i)
=
1 2i
1
2
+ 2
2
=
1 2i
5
=
1
5

2
5
i .
4.11. Remark. Historically, the necessity of introducing complex numbers was
realized through the study of cubic (and not quadratic) equations. The reason
for this is that there is a solution formula for cubic equations that in some cases
requires complex numbers in order to express a real solution. See Section 2.7 in
Janichs book [J].
The importance of the eld of complex numbers lies in the fact that they pro-
vide solutions to all polynomial equations. This is the Fundamental Theorem of
Algebra:
Every non-constant polynomial with complex coecients has a root in C.
We will have occasion to use it later on. A proof, however, is beyond the scope of
this course.
4.12. Denition. A complex vector space is a linear space over C.
5. Linear Subspaces and Linear Hulls
In many applications, we do not want to consider all elements of a given vector
space V , rather we only consider elements of a certain subset. Usually, it is
desirable that this subset is again a vector space (with the addition and scalar
multiplication it inherits from V ). In order for this to be possible, a minimal
requirement certainly is that addition and scalar multiplication make sense on the
subset.
5.1. Denition. Let V be an F-vector space. A subset U V is called a vector
subspace or linear subspace of V if it has the following properties.
(1) 0 U.
(2) If u
1
, u
2
U, then u
1
+ u
2
U.
(3) If F and u U, then u U.
Here the addition and scalar multiplication are those of V .
Note that, given the third property, the rst is equivalent to saying that U is
non-empty. Indeed, let u U, then by (3), 0 = 0 u U.
We should justify the name subspace.
5.2. Lemma. Let (V, +, ) be an F-vector space. If U V is a linear subspace
of V , then (U, +[
UU
, [
FU
) is again an F-vector space.
The notation +[
UU
means that we take the addition map + : V V , but restrict it
to U U. (Strictly speaking, we also restrict its target set from V to U. However,
this is usually suppressed in the notation.)
11
Proof. By denition of what a linear subspace is, we really have well-dened ad-
dition and scalar multiplication maps on U. It remains to check the axioms. For
the axioms that state for all . . . , we have . . . and do not involve any existence
statements, this is clear, since they hold (by assumption) even for all elements
of V . This covers axioms (1), (2), (5), (6), (7) and (8). Axiom (3) is OK, since we
require 0 U. Finally, for axiom (4), we need that for all u U, u U as well.
But u = (1)u, so this is OK by the third property of linear subspaces.
It is time for some examples.
5.3. Examples. Let V be a vector space. Then 0 V and V itself are linear
subspaces of V .
5.4. Example. Consider V = R
2
and, for a R, U
a
= (x, y) R
2
: x + y = a.
When is U
a
a linear subspace?
We check the rst condition: 0 = (0, 0) U
a
0 +0 = a, so U
a
can only be a
linear subspace when a = 0. Let us check the other properties for U
0
:
(x
1
, y
1
), (x
2
, y
2
) U
0
= x
1
+ y
1
= 0, x
2
+ y
2
= 0
= (x
1
+ x
2
) + (y
1
+ y
2
) = 0
= (x
1
, y
1
) + (x
2
, y
2
) = (x
1
+ x
2
, y
1
+ y
2
) U
0
and
R, (x, y) U
0
= x + y = 0
= x + y = (x + y) = 0
= (x, y) = (x, y) U
0
.
5.5. Examples. Consider V = R
R
= f : R R, the set of real-valued func-
tions on R. You will learn in Analysis that if f and g are continuous functions,
then f + g is again continuous, and f is continuous for any R. Of course,
the zero function x 0 is continuous as well. Hence, the set of all continuous
functions
((R) = f : R R [ f is continuous
is a linear subspace of V .
Similarly, you will learn that sums and scalar multiples of dierentiable functions
are again dierentiable. Also, derivatives respect sums and scalar multiplication:
(f + g)

= f

+ g

, (f)

= f

. From this, we conclude that

(
n
(R) = f : R R [ f is n times dierentiable and f
(n)
is continuous
is again a linear subspace of V .
In a dierent direction, consider the set of all periodic functions (with period 1):
U = f : R R [ f(x + 1) = f(x) for all x R .
The zero function is certainly periodic. If f and g are periodic, then
(f + g)(x + 1) = f(x + 1) + g(x + 1) = f(x) + g(x) = (f + g)(x) ,
so f + g is again periodic. Similarly, f is periodic (for R). So U is a linear
subspace of V .
The following result now tells us that U ((R), the set of all continuous periodic
functions, is again a linear subspace.
12
5.6. Lemma. Let V be an F-vector space, U
1
, U
2
V linear subspaces of V .
Then U
1
U
2
is again a linear subspace of V .
More generally, if (U
i
)
iI
(with I ,= ) is any family of linear subspaces of V , then
their intersection U =

iI
U
i
is again a linear subspace of V .
Proof. It is sucient to prove the second statement (take I = 1, 2 to obtain the
rst). We check the conditions.
(1) By assumption 0 U
i
for all i I. So 0 U.
(2) Let u
1
, u
2
U. Then u
1
, u
2
U
i
for all i I, hence (by assumption)
u
1
+ u
2
U
i
for all i I. But this means u
1
+ u
2
U.
(3) Let F, u U. Then u U
i
for all i I, hence (by assumption)
u U
i
for all i I. This means that u U.

Note that in general, if U

1
and U
2
are linear subspaces, then U
1
U
2
is not (it is
if and only if U
1
U
2
or U
2
U
1
Exercise!).
The property we just proved is very important, since it tells us that there is always
a smallest linear subspace of V that contains a given subset S of V . This means
that there is a linear subspace U of V such that S U and such that U is
contained in every other linear subspace of V that contains S.
5.7. Denition. Let V be a vector space, S V a subset. The linear hull or
linear span of S, or the linear subspace generated by S is
L(S) =

U V : U linear subspace of V , S U .
(This notation means the intersection of all elements of the specied set: we
intersect all linear subspaces containing S. Note that V itself is such a subspace,
so this set of subspaces is non-empty, so by the preceding result, L(S) really is a
linear subspace.)
If we want to indicate the eld F of scalars, we write L
F
(S). If v
1
, v
2
, . . . , v
n
V ,
we also write L(v
1
, v
2
, . . . , v
n
) for L(v
1
, v
2
, . . . , v
n
).
If L(S) = V , we say that S generates V , or that S is a generating set for V .
Be aware that there are various dierent notations for linear hulls in the literature,
for example S) (which in L
A
T
E
X is written $\langle S \rangle$ and not $<S>$!).
5.8. Example. What do we get in the extreme case that S = ? Well, then we
have to intersect all linear subspaces of V , so we get L() = 0.
Denition 5.7 above has some advantages and disadvantages. Its main advantage
is that it is very elegant. Its main disadvantage is that it is rather abstract and
non-constructive. To remedy this, we show that we can build the linear hull in a
constructive way from below instead of abstractly from above.
13
5.9. Example. Let us look at a specic case rst. Given v
1
, v
2
V , how can we
describe L(v
1
, v
2
)?
According to the denition of linear subspaces, we must be able to add and multi-
ply by scalars in L(v
1
, v
2
); also v
1
, v
2
L(v
1
, v
2
). This implies that every element
of the form
1
v
1
+
2
v
2
must be in L(v
1
, v
2
). So set
U =
1
v
1
+
2
v
2
:
1
,
2
F
(where F is the eld of scalars); then U L(v
1
, v
2
). On the other hand, U is itself
a linear subspace:
0 = 0 v
1
+ 0 v
2
U
(
1
+
1
)v
1
+ (
2
+
2
)v
2
= (
1
v
1
+
2
v
2
) + (
1
v
1
+
2
v
2
) U
(
1
)v
1
+ (
2
)v
2
= (
1
v
1
+
2
v
2
) U
(Exercise: which of the vector space axioms have we used where?)
Therefore, U is a linear subspace containing v
1
and v
2
, and hence L(v
1
, v
2
) U
by our denition. We conclude that
L(v
1
, v
2
) = U =
1
v
1
+
2
v
2
:
1
,
2
F .
This observation generalizes.
5.10. Denition. Let V be an F-vector space, v
1
, v
2
, . . . , v
n
V. A linear combi-
nation (or, more precisely, an F-linear combination) of v
1
, v
2
, . . . , v
n
is an element
of V of the form
v =
1
v
1
+
2
v
2
+ +
n
v
n
with
1
,
2
, . . . ,
n
F. If n = 0, then the only linear combination of no vectors
is (by denition) 0 V .
If S V is any subset, then an (F-)linear combination on S is a linear combination
of nitely many elements of S.
5.11. Proposition. Let V be a vector space, v
1
, v
2
, . . . , v
n
V . Then the set
of all linear combinations of v
1
, v
2
, . . . , v
n
is a linear subspace of V ; it equals the
linear hull L(v
1
, v
2
, . . . , v
n
).
More generally, let S V be a subset. Then the set of all linear combinations
on S is a linear subspace of V, equal to L(S).
Proof. Let U be the set of all linear combinations of v
1
, v
2
, . . . , v
n
. We have to check
that U is a linear subspace of V . First of all, 0 U, since 0 = 0v
1
+ 0v
2
+ + 0v
n
(this even works for n = 0). To check that U is closed under addition, let
v =
1
v
1
+
2
v
2
+ +
n
v
n
and w =
1
v
1
+
2
v
2
+ +
n
v
n
be two elements
of U. Then
v + w = (
1
v
1
+
2
v
2
+ +
n
v
n
) + (
1
v
1
+
2
v
2
+ +
n
v
n
)
= (
1
+
1
)v
1
+ (
2
+
2
)v
2
+ + (
n
+
n
)v
n
is again a linear combination of v
1
, v
2
, . . . , v
n
. Also, for F,
v = (
1
v
1
+
2
v
2
+ +
n
v
n
)
= (
1
)v
1
+ (
2
)v
2
+ + (
n
)v
n
14
is a linear combination of v
1
, v
2
, . . . , v
n
. So U is indeed a linear subspace of V . We
have v
1
, v
2
, . . . , v
n
U, since
v
j
= 0 v
1
+ + 0 v
j1
+ 1 v
j
+ 0 v
j+1
+ + 0 v
n
,
so, by denition of the linear hull, L(v
1
, v
2
, . . . , v
n
) U. On the other hand, it
is clear that any linear subspace containing v
1
, v
2
, . . . , v
n
has to contain all linear
combinations of these vectors, hence U L(v
1
, v
2
, . . . , v
n
). Therefore
L(v
1
, v
2
, . . . , v
n
) = U .
For the general case, the only possible problem is with checking that the set of
linear combinations on S is closed under addition. For this, we observe that if v is a
linear combination on the nite subset I of S and w is a linear combination on the
nite subset J of S, then v and w can both be considered as linear combinations
on the nite subset I J of S; now our argument above applies.
5.12. Example. Let us consider again the vector space ((R) of continuous func-
tions on R. The power functions f
n
: x x
n
(n = 0, 1, 2, . . . ) are certainly
continuous and dened on R, so they are elements of ((R). We nd that their
linear hull L(f
n
: n N
0
) is the linear subspace of polynomial functions, i.e,
functions that are of the form
x a
n
x
n
+ a
n1
x
n1
+ + a
1
x + a
0
with n N
0
and a
0
, a
1
, . . . , a
n
R.
6. Linear Independence and Dimension
This section essentially follows Chapter 3 in Janichs book [J].
In the context of looking at linear hulls, it is a natural question whether we really
need all the given vectors in order to generate their linear hull. Also (maybe in
order to reduce waste. . . ), it is interesting to consider minimal generating sets.
These questions lead to the notions of linear independence and basis.
6.1. Denition. Let V be an F-vector space, v
1
, v
2
, . . . , v
n
V. We say that
v
1
, v
2
, . . . , v
n
are linearly independent, if for all
1
,
2
, . . . ,
n
F,

1
v
1
+
2
v
2
+ +
n
v
n
= 0
implies
1
=
2
= =
n
= 0. (The zero vector cannot be written as a
nontrivial linear combination of v
1
, . . . , v
n
.)
In a similar way, we can dene linear independence for arbitrary subsets of V.
This is equivalent to the following. S V is linearly independent if every choice
of nitely many (distinct) vectors from S is linearly independent.
As a special case, the empty sequence or empty set of vectors is considered to be
linearly independent.
If we want to refer to the eld of scalars F, we say that the given vectors are
F-linearly independent or linearly independent over F.
If v
1
, v
2
, . . . , v
n
(resp., S) are not linearly independent, we say that they are linearly
dependent.
15
6.2. Example. For an easy example that the eld of scalars matters in the context
of linear independence, consider 1, i C, where C can be considered as a real or
as a complex vector space. We then have that 1 and i are R-linearly independent
(essentially by denition of C 0 = 0 1+0 i, and this representation is unique),
whereas they are C-linearly dependent i 1 + (1) i = 0.
6.3. Remark. Let V be a vector space, v
1
, v
2
, . . . , v
n
V. Then v
1
, v
2
, . . . , v
n
are
linearly dependent if and only if one of the v
j
is a linear combination of the others,
i.e., if and only if
L(v
1
, v
2
, . . . , v
n
) = L(v
1
, . . . , v
j1
, v
j+1
, . . . , v
n
)
for some j 1, 2, . . . , n. A similar statement holds for subsets S V.
Proof. Let us rst assume that v
1
, v
2
, . . . , v
n
are linearly dependent. Then there
are scalars
1
,
2
, . . . ,
n
, not all zero, such that

1
v
1
+
2
v
2
+ +
n
v
n
= 0 .
Let j be such that
j
,= 0. Then
v
j
=
1
j
(
1
v
1
+ +
j1
v
j1
+
j+1
v
j+1
+ +
n
v
n
) .
Conversely, assume that v
j
is a linear combination of the other vectors:
v
j
=
1
v
1
+ +
j1
v
j1
+
j+1
v
j+1
+ +
n
v
n
.
Then

1
v
1
+ +
j1
v
j1
v
j
+
j+1
v
j+1
+ +
n
v
n
= 0 ,
so the given vectors are linearly dependent. Bearing in mind that S is linearly
dependent if and only if some nite subset of S is linearly dependent, the last
statement also follows.
6.4. Example. In ((R), the functions
x 1 , x sin x , x cos x , x sin
2
x , x cos
2
x
are linearly dependent, since 1 sin
2
x cos
2
x = 0 for all x R.
On the other hand,
x 1 , x sin x , x cos x
are linearly independent. To see this, assume that + sin x + cos x = 0 for
all x R. Plugging in x = 0, we obtain + = 0. For x = , we get = 0,
which together imply = = 0. Then taking x = /2 shows that = 0 as well.
6.5. Denition. Let V be a vector space. A sequence v
1
, v
2
, . . . , v
n
V (or a
subset S V ) is called a basis of V if v
1
, v
2
, . . . , v
n
are (resp., S is) linearly
independent, and V = L(v
1
, v
2
, . . . , v
n
) (resp., V = L(S)).
From Remark 6.3 above, we see that a basis of V is a minimal generating set of V,
in the sense that we cannot leave out some element and still have a generating set.
What is special about a basis among generating sets?
16
6.6. Lemma. If v
1
, v
2
, . . . , v
n
is a basis of the F-vector space V, then for every
v V , there are unique scalars
1
,
2
, . . . ,
n
F such that
v =
1
v
1
+
2
+ +
n
v
n
.
Proof. The existence of (
1
,
2
, . . . ,
n
) F
n
such that
v =
1
v
1
+
2
+ +
n
v
n
follows from the fact that v
1
, v
2
, . . . , v
n
generate V.
To show uniqueness, assume that (
1
,
2
, . . . ,
n
) F
n
also satisfy
v =
1
v
1
+
2
v
2
+ +
n
v
n
.
Taking the dierence, we obtain
0 = (
1

1
)v
1
+ (
2

2
)v
2
+ + (
n

n
)v
n
.
Since v
1
, v
2
, . . . , v
n
are linearly independent, it follows that

1

1
=
2

2
= =
n

n
= 0 ,
i.e., (
1
,
2
, . . . ,
n
) = (
1
,
2
, . . . ,
n
).
6.7. Exercise. Formulate and prove the corresponding statement for subsets.
6.8. Example. The most basic example of a basis is the canonical basis of F
n
.
This is e
1
, e
2
, . . . , e
n
, where
e
1
= (1, 0, 0, . . . , 0, 0)
e
2
= (0, 1, 0, . . . , 0, 0)
.
.
.
.
.
.
e
n
= (0, 0, 0, . . . , 0, 1) .
In a precise sense (to be discussed in detail later), knowing a basis v
1
, v
2
, . . . , v
n
of a vector space V allows us to express everything in V in terms of the standard
vector space F
n
.
Since we seem to know everything about a vector space as soon as we know a
basis, it makes sense to use bases to measure the size of vector spaces. In order
for this to make sense, we need to know that any two bases of a given vector space
have the same size. The key to this (and many other important results) is the
following.
6.9. Basis Extension Theorem. Let V be an F-vector space, and let v
1
, . . . , v
r
,
w
1
, . . . , w
s
V. If v
1
, . . . , v
r
are linearly independent and V is generated by
v
1
, . . . , v
r
, w
1
, . . . , w
s
, then by adding suitably chosen vectors from w
1
, . . . , w
s
, we
can extend v
1
, . . . , v
r
to a basis of V.
Note that this is an existence theorem what it says is that if we have a bunch of
vectors that is too small (linearly independent, but not necessarily generating)
and a larger bunch of vectors that is too large (generating but not necessarily
linearly independent), then there is a basis in between. However, it does not tell
us how to actually nd such a basis (i.e., how to select the w
j
that we have to
add).
The Basis Extension Theorem implies another important statement.
17
6.10. Exchange Lemma. If v
1
, . . . , v
n
and w
1
, . . . , w
m
are two bases of a vector
space V, then for each i 1, 2, . . . , n there is some j 1, 2, . . . , m such that
v
1
, . . . , v
i1
, w
j
, v
i+1
, . . . , v
n
is again a basis of V.
This says that we can trade any vector of our choice in the rst basis for a vector
in the second basis in such a way as to still have a basis.
We will postpone the proofs and rst look at some consequences.
6.11. Theorem. If v
1
, v
2
, . . . , v
n
and w
1
, w
2
, . . . , w
m
are two bases of a vector
space V, then n = m.
Proof. Assume that, for example, n > m. By repeatedly applying the Exchange
Lemma, we can successively replace v
1
, v
2
, . . . , v
n
by some w
j
and still have a basis.
Since there are more vs than ws, the resulting sequence must have repetitions
and therefore cannot be linearly independent, contradiction.
This implies that the following denition makes sense.
6.12. Denition. If the vector space V has a basis v
1
, v
2
, . . . , v
n
, then n 0 is
called the dimension of V , written n = dimV = dim
F
V.
6.13. Examples. The empty sequence is a basis of the zero space, so dim0 = 0.
The canonical basis of F
n
has length n, so dimF
n
= n.
6.14. Theorem. Let V be a vector space, dimV = n. Then any sequence of
vectors v
1
, . . . , v
m
V such that m > n is linearly dependent.
Proof. Let w
1
, . . . , w
n
be a basis of V. Assume that v
1
, . . . , v
m
were linearly
independent. Then by the Basis Extension Theorem (note that
V = L(w
1
, . . . , w
n
) = L(v
1
, . . . , v
m
, w
1
, . . . , w
n
) ),
we could extend v
1
, . . . , v
m
to a basis of V by adding some vectors from w
1
, . . . , w
n
.
Since m > n, the resulting basis would be too long, contradiction.
Note that this theorem is a quite strong existence statement: it guarantees the
existence of a nontrivial linear relation among the given vectors without the need
to do any computation. This is very useful in many applications. On the other
hand, it is quite a dierent matter to actually nd such a relation: the proof
is non-constructive (as is usually the case with proofs by contradiction), and we
usually need some computational method to exhibit a relation.
The preceding theorem tells us that in a vector space of (nite) dimension n, there
is an upper bound (namely, n) for the length of a linearly independent sequence
of vectors. We can use this to show that there are vector spaces that do not have
dimension n for any n 0.
18
6.15. Example. Let us consider again the linear subspace of polynomial functions
in ((R) (the vector space of continuous functions on R), compare Example 5.12.
Let us call this space P:
P = f ((R) : n N
0
a
0
, . . . , a
n
R x R : f(x) = a
n
x
n
+ + a
1
x + a
0

Denote as before by f
n
the nth power function: f
n
(x) = x
n
. I claim that
f
0
, f
1
, f
2
, . . . is linearly independent. Recall that this means that the only way
of writing zero (i.e., the zero function) as a nite linear combination of the f
j
is
with all coecients equal to zero. If we let n be the largest number such that
f
n
occurs in the linear combination, then it is clear that we can write the linear
combination as

0
f
0
+
1
f
1
+ +
n
f
n
= 0 .
We have to show that this is only possible when
0
=
1
= =
n
= 0.
Note that our assumption means that

n
x
n
+ +
1
x +
0
= 0 for all x R.
There are various ways to proceed from here. For example, we can make use of
the fact that a polynomial of degree n 0 can have at most n zeros in R. Since
there are innitely many real numbers, the polynomial above has innitely many
zeros, hence it must be the zero polynomial (which does not have a well-dened
degree).
Another possibility is to use induction on n (which, by the way, is implicit in the
proof above: it is used in proving the statement on zeros of polynomials). Let us
do this in detail. The claim we want to prove is
n N
0

0
, . . . ,
n
R :
_
x R :
n
x
n
+ +
0
= 0 =
0
= =
n
= 0
_
.
We now have to establish the induction base: the claim holds for n = 0. This is
easy let
0
R and assume that for all x R,
0
= 0 (the function is constant
here: it does not depend on x). Since there are real numbers, this implies
0
= 0.
Next, and this is usually the hard part, we have to do the induction step. We
assume that the claim holds for a given n (this is the induction hypothesis) and
deduce that it then also holds for n+1. To prove the statement for n+1, we have
to consider coecients
0
, . . . ,
n+1
R such that for all x R,
f(x) =
n+1
x
n+1
+
n
x
n
+ +
1
x +
0
= 0 .
Now we want to use the induction hypothesis, so we have to reduce this to a
statement involving a polynomial of degree at most n. One way of doing that is to
borrow some knowledge from Analysis about dierentiation. This tells us that the
derivative of f is zero again, and that it is a polynomial function of degree n:
0 = f

(x) = (n + 1)
n+1
x
n
+ n
n
x
n1
+ +
1
.
Now we can apply the induction hypothesis to this polynomial function; it tells
us that (n + 1)
n+1
= n
n
= =
1
= 0, hence
1
= =
n
=
n+1
= 0.
So f(x) =
0
is in fact constant, which nally implies
0
= 0 as well (by our
reasoning for the induction base).
This completes the induction step and therefore the whole proof.
Note that since P = L(f
n
: n N
0
), we have shown that f
n
: n N
0
is a
basis of P.
So we see that P cannot have a nite basis, since we can nd arbitrarily many
linearly independent elements. This motivates the following denition.
19
6.16. Denition. If a vector space V does not have a nite basis, then V is said
to be innite-dimensional, and we write dimV = .
In particular, we see that dimP = .
The following result shows that our intuition that dimension is a measure for the
size of a vector space is not too far o: larger spaces have larger dimension.
6.17. Lemma. Let U be a linear subspace of the vector space V . Then we have
dimU dimV . If dimV is nite, then we have equality if and only if U = V .
Here we use the usual convention that n < for n N
0
.
Note that in the case that dimV is nite, the statement also asserts the existence
of a (nite) basis of U.
Proof. There is nothing to show if dimV = . So let us assume that V has
a basis v
1
, . . . , v
n
. If u
1
, . . . , u
m
U are linearly independent, then m n by
Thm. 6.14. Hence there is a sequence u
1
, . . . , u
m
of linearly independent vectors
in U of maximal length m (and m n). We claim that u
1
, . . . , u
m
is in fact a
basis of U. The rst claim follows, since then dimU = m n = dimV .
We have to show that u
1
, . . . , u
m
generate U. So assume that there is u U that
is not a linear combination of the u
j
. Then u
1
, . . . , u
m
, u are linearly independent,
which contradicts our choice of u
1
, . . . , u
m
as a maximal linearly independent se-
quence in U. So there is no such u, hence U = L(u
1
, . . . , u
m
).
To prove the second part, rst note that dimU < dimV implies U V (if U = V ,
a basis of U would also be a basis of V , but it is too short for that by Thm. 6.11).
On the other hand, assume U V , and consider a basis of U. It can be extended
to a basis of V by the Basis Extension Theorem 6.9. Since it does not generate V ,
at least one element has to be added, which implies dimU < dimV .
6.18. Examples. Since
P (

(R) =

n=0
(
n
(R) (
2
(R) (
1
(R) ((R) R
R
,
all these spaces are innite-dimensional.
6.19. Exercise. Let F be a nite eld, and consider the F-vector space V of
functions from F to F (so V = F
F
in our earlier notation). Consider again the
linear subspace of polynomial functions:
P
F
= L
F
(f
0
, f
1
, f
2
, . . . )
where f
n
: x x
n
(for x F). Show that dim
F
P
F
is nite.
We have seen that the intersection of linear subspaces is again a linear subspace,
but the union usually is not. However, it is very useful to have a replacement
for the union that has similar properties, but is a linear subspace. Note that the
union of two (or more) sets is the smallest set that contains both (or all) of them.
From this point of view, the following denition is natural.
20
6.20. Denition. Let V be a vector space, U
1
, U
2
V two linear subspaces. The
sum of U
1
and U
2
is the linear subspace generated by U
1
U
2
:
U
1
+ U
2
= L(U
1
U
2
) .
More generally, if (U
i
)
iI
is a family of subspaces of V (I = is allowed here),
then their sum is again

iI
U
i
= L
_
_
iI
U
i
_
.
As before in our discussion of linear hulls, we want a more explicit description of
these sums.
6.21. Lemma. If U
1
and U
2
are linear subspaces of the vector space V , then
U
1
+ U
2
= u
1
+ u
2
: u
1
U
1
, u
2
U
2
.
If (U
i
)
iI
is a family of linear subspaces of V , then

iI
U
i
=
_

jJ
u
j
: J I nite and u
j
U
j
for all j J
_
.
Proof. It is clear that the sets on the right hand side contain U
1
U
2
, resp.

iI
U
i
,
and that they are contained in the left hand sides (which are closed under addi-
tion). So it suces to show that they are linear subspaces (this implies that the
left hand sides are contained in them).
We have 0 = 0 + 0 (resp., 0 =

j
u
j
), so 0 is an element of the right hand side
sets. Closure under scalar multiplication is easy to see:
(u
1
+ u
2
) = u
1
+ u
2
,
and u
1
U
1
, u
2
U
2
, since U
1
, U
2
are linear subspaces. Similarly,

jJ
u
j
=

jJ
u
j
,
and u
j
U
j
, since U
j
is a linear subspace. Finally, for u
1
, u

1
U
1
and u
2
, u

2
U
2
,
we have
(u
1
+ u
2
) + (u

1
+ u

2
) = (u
1
+ u

1
) + (u
2
+ u

2
)
with u
1
+ u

1
U
1
, u
2
+ u

2
U
2
. And for J
1
, J
2
nite subsets of I, u
j
U
j
for
j J
1
, u

j
U
j
for j J
2
, we nd
_

jJ
1
u
j
_
+
_

jJ
2
u

j
_
=

jJ
1
J
2
v
j
,
where v
j
= u
j
U
j
if j J
1
J
2
, v
j
= u

j
U
j
if j J
2
J
1
, and v
j
= u
j
+u

j
U
j
if j J
1
J
2
.
Now we have the following nice formula relating the dimensions of U
1
, U
2
, U
1
U
2
and U
1
+ U
2
. In the following, we use the convention that + n = n + =
+ = for n N
0
.
21
6.22. Theorem. Let U
1
and U
2
be linear subspaces of a vector space V. Then
dim(U
1
+ U
2
) + dim(U
1
U
2
) = dimU
1
+ dimU
2
.
Proof. First note that the statement is trivially true when U
1
or U
2
is innite-
dimensional, since then both sides are . So we can assume that U
1
and U
2
are
both nite-dimensional.
For the proof, we use the Basis Extension Theorem 6.9 again. Since U
1
U
2
U
1
and U
1
is nite-dimensional, we know by Lemma 6.17 that U
1
U
2
is also nite-
dimensional. Let v
1
, . . . , v
r
be a basis of U
1
U
2
. Using the Basis Extension
Theorem, we can extend it on the one hand to a basis v
1
, . . . , v
r
, w
1
, . . . , w
s
of U
1
and on the other hand to a basis v
1
, . . . , v
r
, z
1
, . . . , z
t
of U
2
. I claim that then
v
1
, . . . , v
r
, w
1
, . . . , w
s
, z
1
, . . . , z
t
is a basis of U
1
+ U
2
. It is clear that these vectors
generate U
1
+U
2
(since they are obtained by putting generating sets of U
1
and of U
2
together). So it remains to show that they are linearly independent. Consider a
general linear combination

1
v
1
+ +
r
v
r
+
1
w
1
+ +
s
w
s
+
1
z
1
+ . . .
t
z
t
= 0 .
Then z =
1
z
1
+ . . .
t
z
t
U
2
, but also
z =
1
v
1

r
v
r

1
w
1

s
w
s
U
1
,
so z U
1
U
2
, which implies that
z =
1
z
1
+ . . .
t
z
t
=
1
v
1
+ +
r
v
r
for suitable
j
, since v
1
, . . . , v
r
is a basis of U
1
U
2
. But v
1
, . . . , v
r
, z
1
, . . . , z
t
are
linearly independent (being a basis of U
2
), so this is only possible if
1
= =

r
=
1
= =
t
= 0. This then implies that

1
v
1
+ +
r
v
r
+
1
w
1
+ +
s
w
s
= 0 ,
and since v
1
, . . . , v
r
, w
1
, . . . , w
s
are linearly independent (being a basis of U
1
),
1
=
=
r
=
1
= =
t
= 0 as well. So we have dim(U
1
+ U
2
) = r + s + t,
dim(U
1
U
2
) = r, dimU
1
= r + s and dimU
2
= r + t, from which the claim
follows.
6.23. Remark and Exercise. Note the analogy with the formula
#(X Y ) + #(X Y ) = #X + #Y
for the number of elements in a set. However, there is no analogue of the corre-
sponding formula for three sets:
#(XY Z) = #X+#Y +#Z#(XY )#(XZ)#(Y Z)+#(XY Z) .
Find a vector space V and linear subspaces U
1
, U
2
, U
3
V such that
dim(U
1
+ U
2
+ U
3
) + dim(U
1
U
2
) + dim(U
1
U
3
) + dim(U
2
U
3
)
,= dimU
1
+ dimU
2
+ dimU
3
+ dim(U
1
U
2
U
3
) .
Note that if U
1
U
2
= 0, then we simply have dim(U
1
+U
2
) = dimU
1
+dimU
2
(and conversely). So this is an especially nice case; it motivates the following
denition.
22
6.24. Denition. Let V be a vector space. Two linear subspaces U
1
, U
2
V are
said to be complementary if U
1
U
2
= 0 and U
1
+ U
2
= V .
Note that by the above, we then have dimU
1
+ dimU
2
= dimV .
6.25. Lemma. Let V be a vector space.
(1) If V is nite-dimensional and U V is a linear subspace, then there is a
linear subspace U

V that is complementary to U.
(2) If U and U

are complementary linear subspaces of V , then for every v V

there are unique u U, u

such that v = u + u

.
Proof.
(1) In this case, U is nite-dimensional, with basis u
1
, . . . , u
m
(say). By the Ba-
sis Extension Theorem 6.9, we can extend this to a basis u
1
, . . . , u
m
, v
1
, . . . , v
n
of V . Let U

= L(v
1
, . . . , v
n
). Then we clearly have V = U + U

. But we
also have U U

= 0: if v U U

, then
v =
1
u
1
+ +
m
u
m
=
1
v
1
+ +
n
v
n
,
but u
1
, . . . , u
m
v
1
, . . . , v
n
are linearly independent, so all the s and s must
be zero, hence v = 0.
(2) Let v V. Since V = U + U

, there certainly are u U and u

such that v = u + u

. Now assume that also v = w + w

with w U and
w

. Then u + u

= w + w

, so u w = w

U U

, hence
u w = w

= 0, and u = w, u

= w

6.26. Example. Given U V , there usually are many complementary subspaces.

For example, consider V = R
2
and U = (x, 0) : x R. What are its comple-
mentary subspaces U

? We have dimV = 2 and dimU = 1, so we must have

dimU

= 1 as well. Let u

= (x

, y

) be a basis of U

. Then y

,= 0 (otherwise
0 ,= u

U U

). Then we can scale u

by 1/y

to obtain a basis of the form

u

= (x

, 1), and U

= L(u

) then is a complementary subspace for every x

R
note that (x, y) = (x yx

, 0) +y(x

, 1) U + U

.
6.27. Proof of Basis Extension Theorem. It remains to prove the Basis Ex-
tension Theorem 6.9 and the Exchange Lemma 6.10. Here we prove the Basis
Extension Theorem. A precise version of the statement is as follows.
Let V be a vector space, and let v
1
, . . . , v
r
, w
1
, . . . , w
s
V such that v
1
, . . . , v
r
are
linearly independent and V = L(v
1
, . . . , v
r
, w
1
, . . . , w
s
). Then there is t N
0
and
indices i
1
, . . . , i
t
1, . . . , s such that v
1
, . . . , v
r
, w
i
1
, . . . , w
i
t
is a basis of V.
Make sure you understand how we have formalized the notion of suitably chosen
vectors from w
1
, . . . , w
s
!
The idea of the proof is simply to add vectors from the w
j
s as long as this
is possible while keeping the sequence linearly independent. When no further
lengthening is possible, we should have a basis. So we are looking for a maximal
linearly independent sequence v
1
, . . . , v
r
, w
i
1
, . . . , w
i
t
. Note that there cannot be
repetitions among the w
i
1
, . . . , w
i
t
if this sequence is to be linearly independent.
Therefore t s, and there must be such a sequence of maximal length. We have to
show that it generates V. It suces to show that w
j
L(v
1
, . . . , v
r
, w
i
1
, . . . , w
i
t
)
23
for all j 1, . . . , s. This is clear if j = i
k
for some k 1, . . . , t. Other-
wise, assume that w
j
is not a linear combination of v
1
, . . . , v
r
, w
i
1
, . . . , w
i
t
. Then
v
1
, . . . , v
r
, w
i
1
, . . . , w
i
t
, w
j
would be linearly independent, which would contradict
our choice of a linearly independent sequence of maximal length. So w
j
must be
a linear combination of our vectors, and the theorem is proved.
Here is an alternative proof, using induction on the number s of vectors w
j
.
The base case is s = 0. In this case, the assumptions tell us that v
1
, . . . , v
r
are
linearly independent and generate V, so we have a basis.
For the induction step, we assume the statement of the theorem is true for
w
1
, . . . , w
s
(and any choice of linearly independent vectors v
1
, . . . , v
r
), and we have
to prove it for w
1
, . . . , w
s
, w
s+1
. First assume that L(v
1
, . . . , v
r
, w
1
, . . . , w
s
) = V.
Then the induction hypothesis immediately gives the result. So we assume now
that L(v
1
, . . . , v
r
, w
1
, . . . , w
s
) V. Then w
s+1
is not a linear combination of
v
1
, . . . , v
r
, hence v
1
, . . . , v
r
, w
s+1
are linearly independent. Now we can apply the
induction hypothesis again (to v
1
, . . . , v
r
, w
s+1
and w
1
, . . . , w
s
); it tells us that we
can extend v
1
, . . . , v
r
, w
s+1
to a basis by adding suitable vectors from w
1
, . . . , w
s
.
This gives us what we want.
6.28. Proof of Exchange Lemma. Now we prove the Exchange Lemma 6.10.
Recall the statement.
If v
1
, v
2
, . . . , v
n
and w
1
, w
2
, . . . , w
m
are two bases of a vector space V, then for each
i 1, . . . , n there is some j 1, . . . , m such that v
1
, . . . , v
i1
, w
j
, v
i+1
, . . . , v
n
is again a basis of V.
So x i 1, . . . , n. Since v
1
, . . . , v
n
are linearly independent, v
i
cannot be a
linear combination of the remaining vs. So U = L(v
1
, . . . , v
i1
, v
i+1
, . . . , v
n
) V .
This implies that there is some j 1, . . . , m such that w
j
/ U (if all w
j
U,
then V U). This in turn implies that v
1
, . . . , v
i1
, w
j
, v
i+1
, . . . , v
n
is linearly
independent. If it is not a basis of V , then by the Basis Extension Theorem,
v
1
, . . . , v
i1
, w
j
, v
i+1
, . . . , v
n
, v
i
must be a basis (we apply the Basis Extension The-
orem to the linearly independent vectors v
1
, . . . , v
i1
, w
j
, v
i+1
, . . . , v
n
and the addi-
tional vector v
i
; together they generate V ). However, the vectors in this latter se-
quence are not linearly independent, since w
j
is a linear combination of v
1
, . . . , v
n
.
So v
1
, . . . , v
i1
, w
j
, v
i+1
, . . . , v
n
must already be a basis of V.
7. Digression: Infinite-Dimensional Vector Spaces
and Zorns Lemma
We have seen that a vector space V such that the number of linearly independent
vectors in V is bounded has a (nite) basis and so has nite dimension. What can
we say about the existence of a basis in an innite-dimensional vector space?
We have seen an example of an innite-dimensional vector space that has a basis:
this was the space of polynomial functions on R, with basis given by the monomials
x x
n
, n N
0
.
On the other hand, you would be very hard put to write down a basis for (say)
((R), or also a basis for R as a Q-vector space.
In order to prove the existence of a basis and other related results, we would need
an innite version of the Basis Extension Theorem.
24
7.1. General Basis Extension Theorem. Let V be a vector space, X, Y V
two subsets such that X is linearly independent and V = L(X Y ). Then there
is a subset Z Y such that X Z is a basis of V.
Now, how can we prove such a statement? Recall that in the proof of the nite
version (if we formulate it for sets instead of sequences), we took a maximal subset
Z of Y such that X Z is linearly independent and showed that X Z is already
a basis. This last step will work here in the same way: assume that Z is maximal
as above, then for every y Y Z, X Z y is linearly dependent, and so
y L(XZ). This implies that V = L(XY ) L(XZ), so XZ generates V
and is therefore a basis.
However, the key point is the existence of a maximal set Z with the required
property. Note that if o is an arbitrary set of subsets of some set, o need not
necessarily have maximal elements. For example, o could be empty. Or consider
the set of all nite subsets of N. So we need some extra condition to ensure the
existence of maximal elements. (Of course, when o is nite (and nonempty), then
there is no problem we can just take a set of maximal size.)
This condition is formulated in terms of chains.
7.2. Denition. Let X be a set, and let o be a set of subsets of X. A subset
( o is called a chain if all elements of ( are comparable, i.e., if for all U, V (,
we have U V or V U. (Note that this is trivially true when ( is empty.)
7.3. Remark. The notion of chain (as well as Zorns Lemma below) applies
more generally to (partially) ordered sets: a chain then is a subset that is totally
ordered.
Now a statement of the kind we need is the following.
7.4. Zorns Lemma. Let X be a set, and let o be a collection of subsets of X.
If for every chain ( o, there is a set U o such that Z U for all Z (,
then o has a maximal element.
Note that the condition, when applied to the empty chain, ensures that o , = .
Also note that there can be more than one maximal element in o.
Let us see how we can apply this result to our situation. The set o we want to
consider is the set of all subsets Z Y such that X Z is linearly independent.
We have to verify the assumption on chains. So let ( o be a chain. We have to
exhibit a set U o containing all the elements of (. In such a situation, our rst
guess is to try U =

( (the union of all sets in (); usually it works. In our case,

we have to show that this U has the property that X U is linearly independent.
Assume it is not. Then there is a nite non-trivial linear combination of elements
of X U that gives the zero vector. This linear combination will only involve
nitely many elements of U, which come from nitely many sets Z (. Since
( is a chain, there is a maximal set Z
max
among these, and our nontrivial linear
combination only involves elements from X Z
max
. But Z
max
is in o, and so
XZ
max
is linearly independent, a contradiction. Therefore our assumption must
be false, and X U must be linearly independent.
7.5. Exercise. Use Zorns Lemma to prove that for every subset X of a vector
space V such that X contains the zero vector, there is a maximal linear subspace
of V contained in X.
25
7.6. Discussion. Based on Zorns Lemma, we can prove the general Basis Exten-
sion Theorem. In particular, this shows that every vector space must have a basis
(take X = and Y = V ). However, Zorns Lemma is an extremely inconstructive
result it does not give us any information on how to nd a maximal element.
And in fact, nobody has ever been able to write down (or explicitly construct) a
Q-basis of R, say. Still, such bases must exist.
The next question then is, how does one prove Zorns Lemma? It turns out that
it is equivalent (given the more harmless axioms of set theory) to the Axiom of
Choice, which states the following.
Let I be a set, and let (X
i
)
iI
be a family of nonempty sets indexed by I. Then
there is a choice function f : I

iI
X
i
such that f(i) X
i
for all i I.
In other words, if all the X
i
are nonempty, then the product

iI
X
i
of these sets
is also nonempty. This looks like a natural property, however it has consequences
like the existence of Q-bases of R which are not so intuitive any more. Also, as
it turned out, the Axiom of Choice is independent from the other axioms of set
theory: it is not implied by them.
For some time, there was some discussion among mathematicians as to whether
the use of the Axiom of Choice (and therefore, of Zorns Lemma) should be al-
lowed or forbidden (because of its inconstructive character). By now, a pragmatic
viewpoint has been adapted by almost everybody: use it when you need it. For
example, interesting parts of analysis and algebra need the Axiom of Choice, and
mathematics would be quite a bit poorer without it.
Finally, a historical remark: Zorns Lemma was rst discovered by Kazimierz
Kuratowski in 1922 (and rediscovered by Max Zorn about a dozen years later),
so it is not really appropriately named. In fact, when I was a student, one of my
professors told us that he talked to Zorn at some occasion, who said that he was
not at all happy that the statement was carrying his name. . .
8. Linear Maps
So far, we have dened the objects of our theory: vector spaces and their elements.
Now we want to look at relations between vector spaces. These are provided by
linear maps maps between two vector spaces that preserve the linear structure.
But before we give a denition, we have to review what a map or function is and
their basic properties.
8.1. Review of maps. A map or function f : X Y is a black box that for
any given x X gives us back some f(x) Y that only depends on x. More
formally, we can dene functions by identifying f with its graph

f
= (x, f(x)) : x X X Y .
In these terms, a function or map from X to Y is a subset f XY such that for
every x X there is a unique y Y such that (x, y) f; we then write f(x) = y.
It is important to keep in mind that the data of a function include the domain X
and target (or codomain) Y .
If f : X Y is a map, then we call f(x) : x X Y the image of f, im(f).
The map f is called injective or one-to-one (11) if no two elements of X are
mapped to the same element of Y . More formally, if x, x

X and f(x) = f(x

),
then x = x

. The map f is called surjective or onto if its image is all of Y .

Equivalently, for all y Y there is some x X such that f(x) = y. The map f
26
is called bijective if it is both injective and surjective. In this case, there is an
inverse map f
1
such that f
1
(y) = x f(x) = y.
A map f : X Y induces maps from subsets of X to subsets of Y and conversely,
which are denoted by f and f
1
again (so you have to be careful to check the
datatype of the argument). Namely, if A X, we set f(A) = f(x) : x A
(for example, the image of f is then f(X)), and for a subset B Y , we set
f
1
(B) = x X : f(x) B; this is called the preimage of V under f. Note
that when f is bijective, there are two meanings of f
1
(B) one as just dened,
and one as g(B) where g is the inverse map f
1
. Fortunately, both meanings agree
(Exercise), and there is no danger of confusion.
Maps can be composed: if f : X Y and g : Y Z, then we can dene a map
X Z that sends x X to g(f(x)) Z. This map is denoted by g f (g after
f) keep in mind that it is f that is applied rst!
Composition of maps is associative: if f : X Y , g : Y Z and h : Z W,
then (h g) f = h (g f). Every set X has a special map, the identity
map id
X
: X X, x x. It acts as a neutral element under composition: for
f : X Y , we have f id
X
= f = id
Y
f. If f : X Y is bijective, then its
inverse satises f f
1
= id
Y
and f
1
f = id
X
.
When talking about several sets and maps between them, we often picture them
in a diagram like the following.
X
f
/
g

Y
g

U
f

/
V
X
f

@
@
@
@
@
@
@
Y
g
/
Z
We call such a diagram commutative if all possible ways of going from one set to
another lead to the same result. For the left diagram, this means that g

f = f

g,
for the right diagram, this means that h = g f.
Now we want to single out among all maps between two vector spaces V and W
those that are compatible with the linear structure.
8.2. Denition. Let V and W be two F-vector spaces. A map f : V W is
called an (F-)linear map or a homomorphism if
(1) for all v
1
, v
2
V , we have f(v
1
+ v
2
) = f(v
1
) + f(v
2
),
(2) for all F and all v V , we have f(v) = f(v).
(Note: the rst property states that f is a group homomorphism between the
additive groups of V and W.)
An injective homomorphism is called a monomorphism, a surjective homomor-
phism is called an epimorphism, and a bijective homomorphism is called an iso-
morphism. Two vector spaces V and W are said to be isomorphic, written V

= W,
if there exists an isomorphism between them.
A linear map f : V V is called an endomorphism of V ; if f is in addition
bijective, then it is called an automorphism of V.
27
8.3. Lemma. Here are some simple properties of linear maps.
(1) If f : V W is linear, then f(0) = 0.
(2) If f : V W is an isomorphism, then the inverse map f
1
is also an
isomorphism.
(3) If f : U V and g : V W are linear maps, then g f : U W is also
linear.
Proof.
(1) This follows from either one of the two properties of linear maps:
f(0) = f(0 + 0) = f(0) + f(0) = f(0) = 0
or
f(0) = f(0 0) = 0 f(0) = 0 .
(Which of the zeros are scalars, which are vectors in V , in W?)
(2) The inverse map is certainly bijective; we have to show that it is linear.
So let w
1
, w
2
W and set v
1
= f
1
(w
1
), v
2
= f
1
(w
2
). Then f(v
1
) = w
1
,
f(v
2
) = w
2
, hence f(v
1
+ v
2
) = w
1
+ w
2
. This means that
f
1
(w
1
+ w
2
) = v
1
+ v
2
= f
1
(w
1
) + f
1
(w
2
) .
The second property is checked in a similar way.
(3) Easy.

Associated to a linear map there are two important linear subspaces: its kernel
and its image.
8.4. Denition. Let f : V W be a linear map. Then the kernel of f is dened
to be
ker(f) = v V : f(v) = 0 .
8.5. Lemma. Let f : V W be a linear map.
(1) ker(f) V is a linear subspace. More generally, if U W is a linear
subspace, then f
1
(U) V is again a linear subspace; it contains ker(f).
(2) im(f) W is a linear subspace. More generally, if U V is a linear sub-
space, then f(U) W is again a linear subspace; it is contained in im(f).
(3) f is injective if and only if ker(f) = 0.
Proof.
(1) We have to check the three properties of subspaces for ker(f). By the
previous remark, f(0) = 0, so 0 ker(f). Now let v
1
, v
2
ker(f). Then
f(v
1
) = f(v
2
) = 0, so f(v
1
+ v
2
) = f(v
1
) + f(v
2
) = 0 + 0 = 0, and
v
1
+v
2
ker(f). Finally, let be a scalar and v ker(f). Then f(v) = 0,
so f(v) = f(v) = 0 = 0, and v ker(f).
The more general statement is left as an exercise.
28
(2) We check again the subspace properties. We have f(0) = 0 im(f). If
w
1
, w
2
im(f), then there are v
1
, v
2
V such that f(v
1
) = w
1
, f(v
2
) = w
2
,
hence w
1
+ w
2
= f(v
1
+ v
2
) im(f). If is a scalar and w im(f), then
there is v V such that f(v) = w, hence w = f(v) im(f).
The more general statement is proved in the same way.
(3) If f is injective, then there can be only one element of V that is mapped
to 0 W, and since we know that f(0) = 0, it follows that ker(f) = 0.
Now assume that ker(f) = 0, and let v
1
, v
2
V such that f(v
1
) = f(v
2
).
Then f(v
1
v
2
) = f(v
1
)f(v
2
) = 0, so v
1
v
2
ker(f). By our assumption,
this means that v
1
v
2
= 0, hence v
1
= v
2
.

8.6. Remark. If you want to show that a subset U in a vector space V is a linear
subspace, it may be easier to nd a linear map f : V W such that U = ker(f)
than to check the properties directly.
It is time for some examples.
8.7. Examples.
(1) Let V be any vector space. Then the unique map f : V 0 into the
zero space is linear. More generally, if W is another vector space, then
f : V W, v 0, is linear. It is called the zero homomorphism; often it
is denoted by 0. Its kernel is all of V, its image is 0 W.
(2) For any vector space, the identity map id
V
is linear; it is even an automor-
phism of V. Its kernel is trivial (= 0); its image is all of V.
(3) If V = F
n
, then all the projection maps pr
j
: F
n
F, (x
1
, . . . , x
n
) x
j
are linear.
(In fact, one can argue that the vector space structure on F
n
is dened in
exactly such a way as to make these maps linear.)
(4) Let P be the vector space of polynomial functions on R. Then the following
maps are linear.
(a) Evaluation: given a R, the map ev
a
: P R, p p(a) is linear.
The kernel of ev
a
consists of all polynomials having a zero at a; the
image is all of R.
(b) Dierentiation: D : P P, p p

is linear.
The kernel of D consists of the constant polynomials; the image of D
is P (since D I
a
= id
P
).
(c) Denite integration: given a < b, the map
I
a,b
: P R, p
b
_
a
p(x) dx
is linear.
(d) Indenite integration: given a R, the map
I
a
: P P , p
_
x
x
_
a
p(t) dt
_
29
is linear. This map is injective; its image is the kernel of ev
a
(see
below).
(e) Translation: given a R, the map
T
a
: P P , p
_
x p(x + a)
_
is linear. This map is an isomorphism: T
1
a
= T
a
.
The Fundamental Theorem of Calculus says that D I
a
= id
P
and that
I
a,b
D = ev
b
ev
a
and I
a
D = id
P
ev
a
. This implies that ev
a
I
a
=
0, hence im(I
a
) ker(ev
a
). On the other hand, if p ker(ev
a
), then
I
a
(p

) = p p(a) = p, so p im(I
a
). Therefore we have shown that
im(I
a
) = ker(ev
a
).
The relation DI
a
= id
P
implies that I
a
is injective and that D is surjective.
Let C P be the subspace of constant polynomials, and let Z
a
P be
the subspace of polynomials vanishing at a R. Then C = ker(D) and
Z
a
= ker(ev
a
) = im(I
a
), and C and Z
a
are complementary subspaces. D
restricts to an isomorphism Z
a

P, and I
a
restricts (on the target side)
to an isomorphism P

Z
a
(Exercise!).
One nice property of linear maps is that they are themselves elements of vector
spaces.
8.8. Lemma. Let V and W be two F-vector spaces. Then the set of all linear
maps V W, with addition and scalar multiplication dened point-wise, forms
an F-vector space. It is denoted by Hom(V, W).
Proof. It is easy to check the vector space axioms for the set of all maps V W
(using the point-wise denition of the operations and the fact that W is a vector
space). Hence it suces to show that the linear maps form a linear subspace:
The zero map is a homomorphism. If f, g : V W are two linear maps, we have
to check that f + g is again linear. So let v
1
, v
2
V ; then
(f + g)(v
1
+ v
2
) = f(v
1
+ v
2
) + g(v
1
+ v
2
) = f(v
1
) + f(v
2
) + g(v
1
) + g(v
2
)
= f(v
1
) + g(v
1
) + f(v
2
) + g(v
2
) = (f + g)(v
1
) + (f + g)(v
2
) .
Similarly, if F and v V , we have
(f + g)(v) = f(v) + g(v) = f(v) + g(v) =
_
f(v) + g(v)
_
= (f + g)(v) .
Now let F, and let f : V W be linear. We have to check that f is again
linear. So let v
1
, v
2
V ; then
(f)(v
1
+ v
2
) = f(v
1
+ v
2
) =
_
f(v
1
) + f(v
2
)
_
= f(v
1
) + f(v
2
) = (f)(v
1
) + (f)(v
2
) .
Finally, let F and v V . Then
(f)(v) = f(v) =
_
f(v)
_
= ()f(v) =
_
f(v)
_
= (f)(v) .

Now the next question is, how do we specify a general linear map? It turns out
that it suces to specify the images of the elements of a basis. If our vector
spaces are nite-dimensional, this means that only a nite amount of information
is necessary (if we consider elements of the eld of scalars as units of information).
30
8.9. Theorem. Let V and W be two F-vector spaces. Let v
1
, . . . , v
n
be a basis
of V , and let w
1
, . . . , w
n
W. Then there is a unique linear map f : V W
such that f(v
j
) = w
j
for all j 1, . . . , n.
More generally, let B V be a basis, and let : B W be a map. Then there
is a unique linear map f : V W such that f[
B
= (i.e., f(b) = (b) for all
b B).
Proof. The statement has two parts: existence and uniqueness. In many cases, it
is a good idea to prove uniqueness rst, since this usually tells us how to construct
the object we are looking for, thus helping with the existence proof. So let us look
at uniqueness now.
We show that there is only one way to dene a linear map f such that f(v
j
) = w
j
for all j. Let v V be arbitrary. Then v is a linear combination on the basis:
v =
1
v
1
+
2
v
2
+ +
n
v
n
.
If f is to be linear, we must then have
f(v) =
1
f(v
1
) +
2
f(v
2
) + +
n
f(v
n
) =
1
w
1
+
2
w
2
+ +
n
w
n
,
which xes f(v). So there is only one possible choice for f.
To show existence, it suces to prove that f as dened above is indeed linear. Note
that f is well-dened, since every v V is given by a unique linear combination
of the v
j
, see Lemma 6.6. Let v, v

V with
v =
1
v
1
+
2
v
2
+ +
n
v
n
v

1
v
1
+

2
v
2
+ +

n
v
n
, so
v + v

= (
1
+

1
)v
1
+ (
2
+

2
)v
2
+ + (
n
+

n
)v
n
then
f(v + v

) = (
1
+

1
)w
1
+ (
2
+

2
)w
2
+ + (
n
+

n
)w
n
= (
1
w
1
+
2
w
2
+ +
n
w
n
) + (

1
w
1
+

2
w
2
+ +

n
w
n
)
= f(v) + f(v

)
and similarly for f(v) = f(v).
The version with basis sets is proved in the same way.
We can use the images of basis vectors to characterize injective and surjective
linear maps.
8.10. Proposition. Let V and W be vector spaces, f : V W a linear map, and
let v
1
, . . . , v
n
be a basis of V. Then
(1) f is injective if and only if f(v
1
), . . . , f(v
n
) are linearly independent;
(2) f is surjective if and only if L(f(v
1
), . . . , f(v
n
)) = W;
(3) f is an isomorphism if and only if f(v
1
), . . . , f(v
n
) is a basis of W.
Proof. The proof of the rst two statements is an exercise; the third follows from
the rst two.
This leads to an important fact: essentially (up to isomorphism), there is only
one F-vector space of any given nite dimension n.
31
8.11. Corollary. If V and W are two F-vector spaces of the same nite dimen-
sion n, then V and W are isomorphic. In particular, if V is an F-vector space of
dimension n < , then V is isomorphic to F
n
: V

= F
n
.
Proof. Let v
1
, . . . , v
n
be a basis of V , and let w
1
, . . . , w
n
be a basis of W. By
Thm. 8.9, there exists a linear map f : V W such that f(v
j
) = w
j
for all
j 1, . . . , n. By Prop, 8.10, f is an isomorphism. For the second statement,
take W = F
n
.
Note, however, that in general there is no natural (or canonical) isomorphism
V

F
n
. The choice of isomorphism is equivalent to the choice of a basis, and
there are many bases of V. In particular, we may want to choose dierent bases
of V for dierent purposes, so it does not make sense to identify V with F
n
in a
specic way.
There is an important result that relates the dimensions of the kernel, image and
domain of a linear map.
8.12. Denition. Let f : V W be a linear map. Then we call the dimension
of the image of f the rank of f: rk(f) = dimim(f).
8.13. Theorem (Dimension Formula for Linear Maps). Let f : V W be
a linear map. Then
dimker(f) + rk(f) = dimV .
Proof. By Lemma 6.25, there is a complementary subspace U of ker(f) in V .
(If dimV = , this is still true by the General Basis Extension Theorem 7.1,
based on Zorns Lemma.) We show that f restricts to an isomorphism between
U and im(f). This implies that dimU = dimim(f). On the other hand, dimV =
dimker(f) + dimU, so the dimension formula follows.
Let f

: U im(f) be the linear map given by restricting f. We note that

ker(f

) = ker(f) U = 0, so f

is injective. To show that f

is also surjective,
take w im(f). Then there is v V such that f(v) = w. We can write v = u

+u
with u

ker(f) and u U. Now

f

(u) = f(u) = 0 + f(u) = f(u

) + f(u) = f(u

+ u) = f(v) = w,
so w im(f

) as well.
For a proof working directly with bases, see Chapter 4 in Janichs book [J].
As a corollary, we have the following criterion for when an endomorphism is an
automorphism.
8.14. Corollary. Let V be a nite-dimensional vector space, and let f : V V
be a linear map. Then the following statements are equivalent.
(1) f is an isomorphism.
(2) f is injective.
(3) f is surjective.
Proof. Note that f is injective if and only if dimker(f) = 0 and f is surjective if
and only if rk(f) = dimV . By Thm.8.13, these two statements are equivalent.
32
9. Quotient Spaces
We have seen in the last section that the kernel of a linear map is a linear subspace.
One motivation for introducing quotient spaces is the question, is any given linear
subspace the kernel of a linear map?
But before we go into this, we need to review the notion of an equivalence relation.
9.1. Review of Equivalence Relations. Recall the following denition. An
equivalence relation on a set X is a relation on X (formally, we can consider the
relation to be a subset R X X, and we write x y when (x, y) R) that is
(1) reexive: x x for all x X;
(2) symmetric: for all x, y X, if x y, then y x;
(3) transitive: for all x, y, z X, if x y and y z, then x z.
One extreme example is the relation of equality on X. The other extreme example
is when all elements of X are equivalent to each other.
The most important feature of an equivalence relation is that it leads to a partition
of X into equivalence classes: for every x X, we consider its equivalence class
C
x
= y X : x y. Note that x C
x
(by reexivity), so C
x
,= . Then x y
is equivalent to C
x
= C
y
, and x , y is equivalent to C
x
C
y
= . To see this,
let z C
x
, so x z, so z x, and (since x y) therefore z y, so y z, and
z C
y
. The other inclusion follows in a similar way. Conversely, assume that
z C
x
C
y
. Then x z and y z, so (using symmetry and transitivity) x y.
So two equivalence classes are either equal or disjoint. We can then consider the
quotient set X/, which is the set of all equivalence classes,
X/ = C
x
: x X .
Note that we have a natural surjective map : X X/, x C
x
.
We use this construction when we want to identify objects with one another that
are possibly distinct, but have a common property.
9.2. Example. Let X = Z. I claim that
n m n m is even
denes an equivalence relation. This is easy to check (Exercise). What are the
equivalence classes? Well, C
0
= n Z : n is even is the set of all even integers,
and C
1
= n Z : n 1 is even is the set of all odd integers. Together, they
partition Z, and Z/ = C
0
, C
1
. The natural map Z Z/ maps all even
numbers to C
0
and all odd numbers to C
1
.
But now on to quotient spaces.
9.3. Denition and Lemma. Let V be an F-vector space, U V a linear
subspace. For v, v

V , we set
v v

mod U v v

U .
This denes an equivalence relation on V , and the equivalence classes have the
form
C
v
= v + U = v + u : u U ;
33
these sets v + U are called the cosets of U in V . We write
V/U =
V
U
= v + U : v V
for the quotient set. We dene an addition and scalar multiplication on V/U by
(v + U) + (v

+ U) = (v + v

) + U , (v + U) = v + U .
These operations are well-dened and turn V/U into an F-vector space, the quo-
tient vector space of V mod U. The natural map : V V/U is linear; it is
called the canonical epimorphism. We have ker() = U.
Proof. There is a number of statements that need proof. First we need to show
that we have indeed dened an equivalence relation:
(1) v v = 0 U, so v v mod U;
(2) if v v

mod U, then v v

U, so v

v = (v v

) U, hence
v

v mod U;
(3) if v v

mod U and v

mod U, then v

v U and v

U, so
v

v = (v

v) + (v

) U, hence v v

mod U.
Next we have to show that the equivalence classes have the form v + U. So let
v V ; then v v

mod U if and only if u = v v

U, if and only if v

= v + u
for some u U, if and only if v

v + U.
Next we have to check that the addition and scalar multiplication on V/U are
well-dened. Note that a given coset can (usually) be written as v + U for many
dierent v V , so we have to check that our denition does not depend on the
specic representatives chosen. So let v, v

, w, w

V such that v + U = w + U
and v

+ U = w

+ U. We have to show that (v + v

) + U = (w + w

) + U,
which is equivalent to (w + w

) (v + v

) U. But this follows easily from

w v, w

U. So addition is OK. For scalar multiplication, we have to show

that v + U = w + U. But w v U implies w v U, so this is ne, too.
Then we have to show that V/U with the addition and scalar multiplication we
have dened is an F-vector space. This is clear from the denitions and the
validity of the vector space axioms for V, if we take U = 0+U as the zero element
and (v) + U as the additive inverse of v + U.
It remains to show that the canonical map V V/U is linear and has kernel U.
But linearity is again clear from the denitions:
(v + v

) = (v + v

) + U = (v + U) + (v

+ U) = (v) + (v

) etc.
In fact, the main reason for dening the vector space structure on V/U in the way
we have done it is to make linear! Finally, ker() = v V : v + U = U =
U.
So we see that indeed every linear subspace of V is the kernel of some linear map
f : V W.
However, the following property of quotient spaces is even more important.
34
9.4. Proposition. Let f : V W be a linear map and U V a linear subspace.
If U ker(f), then there is a unique linear map : V/U W such that f = ,
where : V V/U is the canonical epimorphism. In other words, there is a
unique linear map that makes the following diagram commutative.
V
f
/

W
V/U

<
z
z
z
z
If ker(f) = U, then is injective.
Note that this property allows us to construct linear maps with domain V/U.
Proof. It is clear that we must have (v + U) = f(v) (this is what f =
means). This already shows uniqueness. For existence, we need to show that ,
dened in this way, is well-dened. So let v, w V such that v + U = w + U.
We have to check that f(v) = f(w). But we have w v U ker(f), so
f(w) f(v) = f(w v) = 0. It is clear from the denitions that is linear.
Finally,
ker() = v + U : f(v) = 0 = U
if ker(f) = U; this is the zero subspace of V/U, hence is injective.
9.5. Corollary. If f : V W is a linear map, then V/ ker(f)

= im(f).
Proof. Let U = ker(f). By Prop. 9.4, there is a linear map : V/U W such
that f = , and is injective. So gives rise to an isomorphism V/U
im() = im(f).
9.6. Corollary. Let V be a vector space and U V a linear subspace. Then
dimU + dimV/U = dimV .
Proof. Let : V V/U be the canonical epimorphism. By Thm. 8.13, we have
dimker() +dimim() = dimV . But ker() = U and im() = V/U, so the claim
follows.
9.7. Denition. Let V be a vector space and U V a linear subspace. Then we
call dimV/U the codimension of U in V, written codim
V
U.
9.8. Examples. Note that codim
V
U can be nite, even though U and V both
are innite-dimensional. A trivial example is codim
V
V = 0.
For some less trivial examples, consider again the vector space P of polynomial
functions on R. For the subspace C of constant functions, we have dimC = 1
and codim
P
C = . For the subspace Z = p P : p(0) = 0 of polynomials
vanishing at 0, we have dimZ = and codim
P
Z = 1. (Indeed, Z = ker(ev
0
), so
P/Z

= R = im(ev
0
).) Finally, for the subspace
E = p P : x R : p(x) = p(x)
of even polynomials, we have dimE = and codim
P
E = .
35
9.9. Exercise. Let U
1
, U
2
V be linear subspaces of a (not necessarily nite-
dimensional) vector space V. Show that
codim
V
(U
1
+ U
2
) + codim
V
(U
1
U
2
) = codim
V
U
1
+ codim
V
U
2
.
For this exercise, the following results may be helpful.
9.10. Proposition. Let U
1
, U
2
V be two linear subspaces of the vector space V.
Then there is a natural isomorphism
U
1
U
1
U
2

U
1
+ U
2
U
2
.
Proof. Exercise.
9.11. Proposition. Let U V W be vector spaces. Then V/U W/U is a
linear subspace, and there is a natural isomorphism
W
V

W/U
V/U
.
Proof. It is clear that V/U = v + U : v V is a linear subspace of W/U =
w + U : w W. Consider the composite linear map
f : W W/U
W/U
V/U
where both maps involved in the composition are canonical epimorphisms. What
is the kernel of f? The kernel of the second map is V/U, so the kernel of f consists
of all w W such that w + U V/U. This is equivalent to w v U for some
v V , or w U + V . Since U V , we have U + V = V , hence ker(f) = V .
The map : W/V (W/U)/(V/U) given to us by Prop. 9.4 then is injective and
surjective (since f is surjective), hence an isomorphism.
10. Digression: Finite Fields
Before we embark on studying matrices, I would like to discuss nite elds. First
a general notion relating to elds.
10.1. Lemma and Denition. Let F be a eld with unit element 1
F
. As in any
abelian group, integral multiples of elements of F are dened:
n =
_

_
+ + + (n summands) if n > 0,
0 if n = 0,
(n) () if n < 0.
Consider the set S = n N : n 1
F
= 0. If S = , we say that F has charac-
teristic zero. In this case, all integral multiples of 1
F
are distinct. Otherwise, the
smallest element p of S is a prime number, and we say that F has characteristic p.
We write char F = 0 or char F = p.
36
Proof. First assume that S is empty. Then we have to show that m 1
F
and n 1
F
are distinct when m and n are distinct integers. So assume m 1
F
= n 1
F
and
(without loss of generality) n > m. Then we have (nm) 1
F
= 0, so nm S,
a contradiction.
We also have to show that the smallest element of S is a prime number when S
is nonempty. So assume the smallest element n of S is not prime. Then n = km
with integers 2 k, m < n. But then we have
0 = n 1
F
= km 1
F
= (k 1
F
)(m 1
F
) .
Since F is a eld, one of the two factors must be zero, so k S or m S. But
this contradicts the assumption that n is the smallest element of S.
10.2. Corollary. If F is a nite eld, then char F = p for some prime number p.
Proof. If char F = 0, then the integral multiples of 1
F
are all distinct, so F must
be innite.
In order to see that nite elds of characteristic p exist, we will construct the
smallest of them.
10.3. Denition and Lemma. Let p be a prime number. The following denes
an equivalence relation on Z:
a b mod p p divides a b.
Its equivalence classes have the form
a + pZ = a + kp : k Z .
Let F
p
= a + pZ : a Z be the quotient set. Then
F
p
= 0 + pZ, 1 + pZ, 2 + pZ, . . . , (p 1) +pZ .
The following addition and multiplication on F
p
are well-dened:
(a + pZ) + (b + pZ) = (a + b) + pZ, (a + bZ)(b + pZ) = ab + pZ.
F
p
with this addition and multiplication is a eld, and #F
p
= p and char F
p
= p.
Proof. That we have dened an equivalence relation is seen in much the same way
as in the case of quotient vector spaces. The same statement applies to the form
of the equivalence classes and the proof that addition is well-dened. To see that
multiplication is well-dened, assume that
a a

mod p and b b

mod p .
Then there are k, l Z such that a

= a + kp, b

= b + lp. Then
a

= (a + kp)(b + lp) = ab + (kb + al + klp)p ,

so a

+ pZ = ab + pZ.
It is then clear that all the eld axioms are satised (with zero 0 + pZ and one
1+pZ), except possibly the existence of multiplicative inverses. For this, we show
rst that (a+pZ)(b+pZ) = 0+pZ implies that a+pZ = 0+pZ or b+pZ = 0+pZ.
Indeed, the vanishing of the product means that p divides ab, so p (as a prime
number) must divide a or b.
Now consider a + pZ ,= 0 + pZ. Then I claim that the map
F
p
F
p
, x + pZ (a + pZ)(x + pZ)
37
is injective. Indeed, if (a + pZ)(x + pZ) = (a + pZ)(y + pZ), then
(a + pZ)
_
(x + pZ) (y + pZ)
_
= 0 + pZ,
hence x+pZ = y +pZ. Since F
p
is nite, the map must then be surjective as well,
so there is x + pZ F
p
such that (a + pZ)(x + pZ) = 1 + pZ.
That #F
p
= p is clear from the description of the equivalence classes. Finally,
char F
p
= p follows from n (1 +pZ) = n + pZ.
10.4. Theorem. Let F be a eld of characteristic p. Then F is a vector space
over F
p
. In particular, if F is nite, then #F = p
n
for some n N.
Proof. We dene scalar multiplication F
p
F F by (a +pZ) x = a x (where
on the right, we use integral multiples). We have to check that this is well-dened.
So let a

= a + kp. Then
(a

+ pZ) x = a

x = a x + kp x = a x + (p 1
F
)(k x) = a x
(since p 1
F
= 0). The relevant axioms are then clearly satised (they are for
integral multiples, as one can prove by induction).
If F is nite, then its dimension over F
p
must be nite, say n. Then every element
of F is a unique linear combination with coecients from F
p
of n basis elements,
hence F must have p
n
elements.
So we see that there can be no eld with exactly six elements, for example.
We know that for every prime number p, there is a eld with p elements. What
about existence of elds with p
n
elements for every n?
10.5. Theorem. If p is a prime number and n N, then there exists a eld with
p
n
elements, and all such elds are isomorphic (in a suitable sense).
Proof. The proof of this result is beyond this course. You should see it in the
Introductory Algebra course.
10.6. Example. Let us show that there is a eld F
4
with four elements. It will
be of characteristic 2. Its elements will be 0, 1, and +1 (since +1 has to be
something and cannot be one of 0, 1 or , it has to be the fourth element). What
is
2
? It cannot be 0 or 1 or , so we must have

2
= + 1 .
This implies ( + 1) = 1, which shows that our new two elements have multi-
plicative inverses. Here are the addition and multiplication tables.
+ 0 1 + 1
0 0 1 + 1
1 1 0 + 1
+ 1 0 1
+ 1 + 1 1 0
0 1 + 1
0 0 0 0 0
1 0 1 + 1
0 + 1 1
+ 1 0 + 1 1
11. Matrices
Matrices are a convenient way of specifying linear maps from F
n
to F
m
. Since
every nite-dimensional F-vector space is isomorphic to some F
n
, they can also be
used to describe linear maps between nite-dimensional vector spaces in general.
Last, but not least, matrices are a very convenient tool for performing explicit
computations with linear maps.
38
11.1. Denition. Recall that by Thm. 8.9, a linear map f : F
n
F
m
is uniquely
determined by the images of a basis. Now, F
n
has a canonical basis e
1
, . . . , e
n
(where e
j
= (
1j
, . . . ,
nj
) and
ij
= 1 if i = j and 0 otherwise;
ij
is called the
Kronecker symbol), and so f is uniquely specied by
f(e
1
) = (a
11
, a
21
, . . . , a
m1
) F
m
f(e
2
) = (a
12
, a
22
, . . . , a
m2
) F
m
.
.
.
.
.
.
.
.
.
f(e
n
) = (a
1n
, a
2n
, . . . , a
mn
) F
m
.
We arrange the various coecients a
ij
F in a rectangular array
A =
_
_
_
_
a
11
a
12
a
1n
a
21
a
22
a
2n
.
.
.
.
.
.
.
.
.
a
m1
a
m2
a
mn
_
_
_
_
= (a
ij
)
1im,1jn
and call this an m n matrix (with entries in F). The a
ij
are called the entries
or coecients of A. For i 1, . . . , m, (a
i1
, a
i2
, . . . , a
in
) is a row of A, and for
j 1, . . . , n,
(a
1j
, a
2j
, . . . , a
mj
)

:=
_
_
_
_
a
1j
a
2j
.
.
.
a
mj
_
_
_
_
is called a column of A. Note that the coecients of f(e
j
) appear in the jth column
of A. The set of all mn matrices with entries in F is denoted by Mat(mn, F).
Note that as a boundary case, m = 0 or n = 0 (or both) is allowed; in this case
Mat(mn, F) has only one element, which is an empty matrix and corresponds
to the zero homomorphism.
If m = n, we sometimes write Mat(n, F) for Mat(n n, F). The matrix I = I
n

Mat(n, F) that corresponds to the identity map id
F
n is called the identity matrix;
we have
I
n
=
_
_
_
_
1 0 0
0 1 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 1
_
_
_
_
= (
ij
)
1i,jn
.
11.2. Remark. By Thm. 8.9, the matrices in Mat(m n, F) correspond bijec-
tively to linear maps in Hom(F
n
, F
m
). Therefore, we will usually not distinguish
between a matrix A and the linear map F
n
F
m
it describes.
In this context, the elements of F
n
(and similarly for F
m
) are considered as column
vectors, and we write the linear map given by the matrix A = (a
ij
), as applied to
x = (x
j
) F
n
in the form
Ax =
_
_
_
_
a
11
a
12
a
1n
a
21
a
22
a
2n
.
.
.
.
.
.
.
.
.
a
m1
a
m2
a
mn
_
_
_
_
_
_
_
_
x
1
x
2
.
.
.
x
n
_
_
_
_
=
_
_
_
_
a
11
x
1
+ a
12
x
2
+ + a
1n
x
n
a
21
x
1
+ a
22
x
2
+ + a
2n
x
n
.
.
.
a
m1
x
1
+ a
m2
x
2
+ + a
mn
x
n
_
_
_
_
.
Note that the result is again a column vector, this time of length m (the length
of the columns of A). Note also that Ae
j
is the jth column of A, hence A really
39
corresponds (in the sense introduced above) to the linear map we have dened
here.
11.3. Denition. We know that Hom(F
n
, F
m
) has the structure of an F-vector
space (see Lemma 8.8). We can transport this structure to Mat(mn, F) using
the identication of matrices and linear maps. So for A, B Mat(m n, F), we
dene A + B to be the matrix corresponding to the linear map x Ax + Bx.
It is then a trivial verication to see that (a
ij
) + (b
ij
) = (a
ij
+ b
ij
), i.e., that
addition of matrices is done coecient-wise. Similarly, for F and A = (a
ij
)
Mat(m n, F), we dene A to be the matrix corresponding to the linear map
x Ax; we then see easily that (a
ij
) = (a
ij
). With this addition and scalar
multiplication, Mat(mn, F) becomes an F-vector space, and it is clear that it is
the same as (i.e., isomorphic to) F
mn
the only dierence is the arrangement
of the coecients in a rectangular fashion instead of in a row or column.
11.4. Denition. By Lemma 8.3, the composition of two linear maps is again
linear. How is this reected in terms of matrices?
Let A Mat(l m, F) and B Mat(m n, F). Then B gives a linear map
F
n
F
m
, and A gives a linear map F
m
F
l
. We dene the product AB to be
the matrix corresponding to the composite linear map F
n
B
F
m
A
F
l
. So AB
will be a matrix in Mat(l n, F).
To nd out what this means in terms of matrix entries, recall that the kth column
of AB gives the image of the basis vector e
k
. So the kth column of AB is given
by ABe
k
= AB
k
, where B
k
= Be
k
denotes the kth column of B. The ith entry of
this column is then
a
i1
b
1k
+ a
i2
b
2k
+ + a
im
b
mk
=
m

j=1
a
ij
b
jk
.
As a mnemonic, to compute the entry in row i and column k of AB, we take row i
of A: (a
i1
, a
i2
, . . . , a
im
) and column k of B: (b
1k
, b
2k
, . . . , b
mk
)

, and compute their

dot product

m
j=1
a
ij
b
jk
.
If the linear map corresponding to A Mat(m n, F) is an isomorphism, then
A is called invertible. This implies that m = n. The matrix corresponding to the
inverse linear map is (obviously) denoted A
1
; we then have AA
1
= A
1
A = I
n
,
and A
1
is uniquely determined by this property.
11.5. Remark. From the corresponding statements on linear maps, we obtain
immediately that matrix multiplication is associative:
A(BC) = (AB)C
for A Mat(k l, F), B Mat(l m, F), C Mat(mn, F), and is distributive
with respect to addition:
A(B + C) = AB + AC for A Mat(l m, F), B, C Mat(mn, F);
(A + B)C = AC + BC for A, B Mat(l m, F), C Mat(mn, F).
However, matrix multiplication is not commutative in general BA need not
even be dened even though AB is and AB = 0 (where 0 denotes a zero matrix
40
of suitable size) does not imply that A = 0 or B = 0. For a counterexample (to
both properties), consider (over a eld of characteristic ,= 2)
A =
_
1 1
0 0
_
and B =
_
0 1
0 1
_
.
Then
AB =
_
0 2
0 0
_
,=
_
0 0
0 0
_
= BA.
The identity matrix acts as a multiplicative identity:
I
m
A = A = AI
n
for A Mat(mn, F).
If A, B Mat(n, F) are both invertible, then AB is also invertible, and (AB)
1
=
B
1
A
1
(note the reversal of the factors!).
11.6. Denition. Let A Mat(mn, F). Then the rank of A, rk(A), is the rank
of A, considered as a linear map F
n
F
m
. Note that we have rk(A) minm, n,
since it is the dimension of a subspace of F
m
, generated by n vectors.
By this denition, the rank of A is the same as the column rank of A, i.e., the
dimension of the linear hull of the columns of A (as a subspace of F
m
). We can
as well dene the row rank of A to be the dimension of the linear hull of the rows
of A (as a subspace of F
n
). The following result tells us that these additional
denitions are not really necessary.
11.7. Proposition. Let A Mat(m n, F) be a matrix. Then the row and
column ranks of A are equal.
Proof. We rst note that the dimension of the linear hull of a sequence of vectors
equals the length of a maximal linearly independent subsequence (which is then a
basis of the linear hull). If we call a row (column) of A redundant if it is a linear
combination of the remaining rows (columns), then the row (column) rank of the
matrix is therefore unchanged if we remove a redundant row (column). We want
to show that removing a redundant column also does not change the row rank
and conversely. So suppose that the jth column is redundant. Now assume that
a sequence of rows is linearly dependent after removing the jth column. Since the
jth column is a linear combination of the other columns, this dependence relation
extends to the jth column, hence the rows are also linearly dependent before
removing the jth column. This shows that the row rank does not drop (since
linearly independent rows stay linearly independent), and as it clearly cannot
increase, it must be unchanged. Similarly, we see that removing a redundant row
leaves the column rank unaected.
We now successively remove redundant rows and columns until this is no longer
possible. Let the resulting matrix A

have r rows and s columns. Without loss of

generality, we can assume r s. The column rank is s (since there are s linearly
independent columns), but can at most be r (since the columns have r entries),
so r = s, and row and column rank of A

are equal. But A has the same row and

column ranks as A

, hence the row and column ranks of A must also be equal.

There is another way of expressing this result. To do this, we need to introduce
another notion.
41
11.8. Denition. Let A = (a
ij
) Mat(m n, F) be a matrix. The transpose
of A is the matrix
A

= (a
ji
)
1in,1jm
Mat(n m, F) .
(So we get A

from A by a reection on the main diagonal.)

11.9. Remark. The result of Prop. 11.7 can be stated as rk(A) = rk(A

).
As simple properties of transposition, we have that
(A + B)

= A

+ B

, (A)

= A

, (AB)

= B

(note the reversal of factors!) this is an exercise. If A Mat(n, F) is invertible,

this implies that A

is also invertible, and (A

)
1
= (A
1
)

.
12. Computations with Matrices: Row and Column Operations
Matrices are not only a convenient means to specify linear maps, they are also
very suitable for doing computations. The main tool for that are the so-called
elementary row and column operations.
12.1. Denition. Let A be a matrix with entries in a eld F. We say that we
perform an elementary row operation on A, if we
(1) multiply a row of A by some F 0, or
(2) add a scalar multiple of a row of A to another (not the same) row of A, or
(3) interchange two rows of A.
Note that the third type of operation is redundant, since it can be achieved by a
sequence of operations of the rst two types (Exercise).
We dene elementary column operations on A in a similar way, replacing the word
row by column each time it appears.
12.2. Remark. If A

is obtained from A by a sequence of elementary row opera-

tions, then there is an invertible matrix B such that A

= BA. Similarly, if A

is
obtained from A by a sequence of elementary column operations, then there is an
invertible matrix C such that A

= AC. In both cases, we have rk(A

) = rk(A).
Proof. Let A Mat(mn, F). We denote by E
ij
Mat(m, F) the matrix whose
only non-zero entry is at position (i, j) and has value 1. (So E
ij
= (
ik

jl
)
1k,lm
.)
Also, we set M
i
() = I
m
+ ( 1)E
ii
; this is a matrix whose non-zero entries are
all on the diagonal, and have the value 1 except the entry at position (i, i), which
has value .
Then it is easily checked that multiplying the ith row of A by amounts to
replacing A by M
i
()A, and that adding times the jth row of A to the ith row
of A amounts to replacing A by (I
m
+ E
ij
)A.
Now we have that M
i
() and I
m
+ E
ij
(for i ,= j) are invertible, with inverses
M
i
(
1
) and I
m
E
ij
, respectively. (We can undo the row operations by row
operations of the same kind.) Let B
1
, B
2
, . . . , B
r
be the matrices corresponding
to the row operations we have performed on A to obtain A

, then
A

= B
r
_
B
r1

_
B
2
(B
1
A)
_

_
= (B
r
B
r1
B
2
B
1
)A,
and B = B
r
B
r1
B
2
B
1
is invertible as a product of invertible matrices.
42
The statement on column operations is proved in the same way, or by applying
the result on row operations to A

.
Finally, the statement on the ranks follows from the fact that invertible matrices
represent isomorphisms, or also from the simple observation that elementary row
(column) operations preserve the row (column) rank, together with Prop. 11.7.
The following algorithm is the key to most computations with matrices.
12.3. The Row Echelon Form Algorithm. Let A Mat(mn, F) be a matrix.
The following procedure applies successive elementary row operations to A in order
to transform it into a matrix A

in row echelon form. This means that A

has the
following shape.
A

=
_
_
_
_
_
_
_
_
_
_
_
0 0 1
0 0 0 0 0 1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 0 0 0 0 0 1
0 0 0 0 0 0 0 0 0 0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 0 0 0 0 0 0 0 0
_
_
_
_
_
_
_
_
_
_
_
So there are 0 r m and 1 j
1
< j
2
< < j
r
n such that if A

= (a

ij
),
then a

ij
= 0 if i > r or if i r and j < j
i
, and a

ij
i
= 1 for 1 i r.
1. Set A

= A, r = 0 and j
0
= 0.
2. (At this point, a

ij
= 0 if i > r and j j
r
or if 1 i r and 1 j < j
i
. Also,
a

ij
i
= 1 for 1 i r.)
If the (r + 1)st up to the mth rows of A

are zero, then stop.

3. Find the smallest j such that there is some a

ij
,= 0 with r < i m. Replace
r by r + 1, set j
r
= j, and interchange the rth and the ith row of A

if r ,= i.
Note that j
r
> j
r1
.
4. Multiply the rth row of A

by (a

rj
r
)
1
.
5. For each i = r + 1, . . . , m, add a

ij
r
times the rth row of A

to the ith row

of A

.
6. Go to Step 2.
Proof. The only changes that are done to A

are elementary row operations of the

third, rst and second kinds in steps 3, 4 and 5, respectively. Since in each pass
through the loop, r increases, and we have to stop when r = m, the procedure
certainly terminates. We have to show that when it stops, A

is in row echelon
form.
We check that the claim made at the beginning of step 2 is correct. It is trivially
satised when we reach step 2 for the rst time. We now assume it is OK when
we are in step 2 and show that it is again true when we come back to step 2. Since
the rst r rows are not changed in the loop, the part of the statement referring
to them is not aected. In step 3, we increase r and nd j
r
(for the new r) such
that a

ij
= 0 if i r and j < j
r
. By our assumption, we must have j
r
> j
r1
.
The following actions in steps 3 and 4 have the eect of producing an entry with
value 1 at position (r, j
r
). In step 5, we achieve that a

ij
r
= 0 for i > r. So a

ij
= 0
for i > r and j j
r
and for i = r and j < j
r
. This shows that the condition in
step 2 is again satised.
43
So at the end of the algorithm, the statement in step 2 is true. Also, we have seen
that 0 < j
1
< j
2
< < j
r
, hence A

has row echelon form when the procedure is

nished.
12.4. Proposition. The value of the number r at the end of the Row Echelon
Form Algorithm is the rank of A. More precisely, the r nonzero rows form a basis
of the row space of A.
Proof. It is clear that the rst r rows of A

are linearly independent. Since all

remaining rows are zero, the (row) rank of A

is r. But elementary row operations

do not change the rank, so rk(A) = rk(A

) = r. Since elementary row operations

do not even change the row space (exercise), the second claim also follows.
12.5. Example. Consider the following matrix.
A =
_
_
1 2 3
4 5 6
7 8 9
_
_
Let us bring it into row echelon form and nd its rank!
Since the upper left entry is nonzero, we have j
1
= 1. We subtract 4 times the
rst row from the second and 7 times the rst row from the third. This leads to
A

=
_
_
1 2 3
0 3 6
0 6 12
_
_
.
Now we have to distinguish two cases. If char(F) = 3, then
A

=
_
_
1 2 0
0 0 0
0 0 0
_
_
is already in row echelon form, and rk(A) = 1. Otherwise, 3 ,= 0, so we divide
the second row by 3 and then add 6 times the new second row to the third. This
gives
A

=
_
_
1 2 3
0 1 2
0 0 0
_
_
.
This is in row echelon form, and we nd rk(A) = 2.
12.6. Proposition. If A Mat(n, F) is invertible, then we can transform it into
the identity matrix I
n
by elementary row operations. The same operations, applied
to I
n
in the same order, produce the inverse A
1
.
Proof. If A is invertible, then its rank is n. So the row echelon form our algorithm
produces looks like this:
A

=
_
_
_
_
_
_
1
0 1
0 0 1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 1
_
_
_
_
_
_
(A

is an upper triangular matrix.)

44
By adding suitable multiples of the ith row to the rows above it, we can clear all
the entries away from the diagonal and get I
n
.
By Remark 12.2, I
n
= BA, where the left multiplication by B corresponds to the
row operations performed on A. This implies that A
1
= B = BI
n
, hence we
obtain A
1
by performing the same row operations on I
n
.
12.7. Example. Let us see how to invert the following matrix (where we assume
char(F) ,= 2).
A =
_
_
1 1 1
1 2 4
1 3 9
_
_
It is convenient to perform the row operations on A and on I in parallel:
_
_
1 1 1 1 0 0
1 2 4 0 1 0
1 3 9 0 0 1
_
_

_
_
1 1 1 1 0 0
0 1 3 1 1 0
0 2 8 1 0 1
_
_

_
_
1 0 2 2 1 0
0 1 3 1 1 0
0 0 2 1 2 1
_
_

_
_
1 0 0 3 3 1
0 1 0
5
2
4
3
2
0 0 1
1
2
1
1
2
_
_
So
A
1
=
_
_
3 3 1

5
2
4
3
2
1
2
1
1
2
_
_
.
12.8. Remark. This inversion procedure will also tell us whether the matrix is
invertible or not. Namely, if at some point in the computation of the row echelon
form, the lower part of the next column has no non-zero entries, then the matrix is
not invertible. This corresponds to a gap (j
i+1
j
i
+2) in the sequence j
1
, . . . , j
r
,
which implies that r < n.
12.9. Corollary. If A Mat(n, F) is invertible, then A can be written as a prod-
uct of matrices M
i
() ( ,= 0) and I
n
+E
ij
(i ,= j). (Notation as in the proof of
Remark 12.2.)
Proof. By Prop. 12.6, A
1
can be transformed into I
n
by a sequence of elementary
row operations, and A = (A
1
)
1
then equals the matrix B such that left mul-
tiplication by B eects the row operations. A look at the proof of Remark 12.2
shows that B is a product of matrices of the required form.
The next application of the Row Echelon Form Algorithm is to compute a basis
for the kernel of a matrix (considered as a linear map F
n
F
m
).
45
12.10. Denition. A matrix A = (a
ij
) Mat(mn, F) is in reduced row echelon
form, if it is in row echelon form and in addition a
ij
k
= 0 for all i ,= k. This means
that the entries above the leading 1s in the nonzero rows are zero as well:
A =
_
_
_
_
_
_
_
_
_
_
_
0 0 1 0 0
0 0 0 0 0 1 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 0 0 0 0 0 1
0 0 0 0 0 0 0 0 0 0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 0 0 0 0 0 0 0 0
_
_
_
_
_
_
_
_
_
_
_
It is clear that every matrix can be transformed into reduced row echelon form by
a sequence of elementary row operations we only have to change Step 5 of the
algorithm to
5. For each i = 1, . . . , r 1, r +1, . . . , m, add a

ij
r
times the rth row of A

to the
ith row of A

.
12.11. Remark. The reduced row echelon form is unique in the sense that if
A, A

Mat(m n, F) are both in reduced row echelon form, and A

= BA
with B Mat(m, F) invertible, then A = A

.
In other words, if we declare two m n matrices to be equivalent if one can
be obtained from the other by row operations, then the matrices in reduced row
echelon form give a complete system of representatives of the equivalence classes.
Proof. Exercise.
12.12. Lemma. If A = (a
ij
) Mat(m n, F) is in reduced row echelon form,
then dimker(A) = n r, and a basis of ker(A) is given by
v
k
= e
k

1ir
j
i
<k
a
ik
e
j
i
, k 1, . . . , n j
1
, . . . , j
r
,
where e
1
, . . . , e
n
is the canonical basis of F
n
.
Proof. It is clear that the given vectors are linearly independent, since v
k
is the
only vector in the list whose kth entry is nonzero. We know that dimker(A) =
n rk(A) = n r. So we only have to check that Av
k
= 0 for all k. For this,
note that Ae
k
= (a
1k
, a
2k
, . . . , a
mk
)

is the kth column of A, that Ae

j
i
= e

i
, where
e

1
, . . . , e

m
is the canonical basis of F
m
, and that a
ik
= 0 if i > r or j
i
> k. So
Av
k
= Ae
k

1ir
j
i
<k
a
ik
e

i
=
m

i=1
a
ik
e

i

m

i=1
a
ik
e

i
= 0 .

12.13. Lemma. If A Mat(mn, F) is a matrix and A

is obtained from A by
a sequence of elementary row operations, then ker(A) = ker(A

).
Proof. By Remark 12.2, A

= BA with an invertible matrix B. We then have

x ker(A) Ax = 0 BAx = 0 A

x = 0 x ker(A

) .

46
We can therefore compute (a basis of) the kernel of A by rst bringing it into
reduced row echelon form; then we read o the basis as described in Lemma 12.12.
12.14. Example. Let us compute the kernel of the telephone matrix
A =
_
_
1 2 3
4 5 6
7 8 9
_
_
.
We have seen earlier that we can transform it into the row echelon form (for
char(F) ,= 3)
_
_
1 2 3
0 1 2
0 0 0
_
_
.
From this, we obtain the reduced row echelon form
_
_
1 0 1
0 1 2
0 0 0
_
_
.
Hence the kernel has dimension 1 and is generated by (1, 2, 1)

.
If char(F) = 3, the reduced row echelon form is
_
_
1 2 0
0 0 0
0 0 0
_
_
.
Here, the kernel has dimension 2; a basis is given by (1, 1, 0)

, (0, 0, 1)

.
As a nal application, we show how we can write a given linear subspace of F
n
(given as the linear hull of some vectors) as the kernel of a suitable mn matrix.
12.15. Proposition. Let V = L
_
v
1
, . . . , v
k
_
F
n
be a linear subspace. Write
v
i
= (a
ij
)
1jn
, and let A = (a
ij
) be the k n matrix whose rows are given by
the coecients of the v
i
. If U = L
_
w
1
, . . . , w
m
_
is the kernel of A, then V is the
kernel of the mn matrix B whose rows are given by the coecients of the w
j
.
Proof. Let l = dimV ; then dimU = n l. In terms of matrices, we have the
relation AB

= 0, which implies BA

= 0, hence the v
i
, which are the columns
of A

, are in the kernel of B. We also know that rk(A) = l, rk(B) = nl, therefore
dimker(B) = n rk(B) = l = dimV . Since, as we have seen, V ker(B), this
implies V = ker B.
Of course, we use the Row Echelon Form Algorithm in order to compute (a basis
of) the kernel of the matrix A.
12.16. Example. Let us use Prop. 12.15 to nd the intersection of two linear
subspaces. Consider
U = L
_
(1, 1, 1), (1, 2, 3)
_
and V = L
_
(1, 0, 0), (1, 1, 1)
_
as linear subspaces of R
3
. We want to nd (a basis of) U V .
To do this, we rst write U and V as kernels of suitable matrices. To get that
for U, we apply the Row Echelon Form Algorithm to the following matrix.
_
1 1 1
1 2 3
_

_
1 1 1
0 1 2
_

_
1 0 1
0 1 2
_
47
The kernel of this matrix is therefore generated by (1, 2, 1)

, and U is the kernel

of
_
1 2 1
_
.
For V , we proceed as follows.
_
1 0 0
1 1 1
_

_
1 0 0
0 1 1
_

_
1 0 0
0 1 1
_
So V is the kernel of the matrix
_
0 1 1
_
.
Then U V will be the kernel of the following matrix, which we compute using
the Row Echelon Form Algorithm again.
_
1 2 1
0 1 1
_

_
1 0 3
0 1 1
_
We see that U V = L
_
(3, 1, 1)
_
.
12.17. Remark. Let A Mat(kn, F) and B Mat(mn, F). Performing row
operations, we can nd a basis for the row space of a given matrix, and we can nd
a basis for the kernel of a given matrix. If we stack the matrix A on top of B to
form a matrix C Mat((k +m) n, F), then the row space of C will be the sum
of the row spaces of A and of B, and the kernel of C will be the intersection of the
kernels of A and of B. Since we can switch between representing a linear subspace
of F
n
as a row space (i.e., as the linear hull of given vectors) or as the kernel of a
matrix (i.e., of a linear map F
n
F
m
for some m), we can use our Row Echelon
Form Algorithm in order to compute (bases of) sums and intersections of linear
subspaces.
13. Linear Equations
One of the very useful applications of Linear Algebra is to linear equations.
13.1. Denition. Let f : V W be a linear map between two F-vector spaces.
The equation
f(x) = 0 ,
to be solved for x V, is called a homogeneous linear equation. If V = F
n
and W = F
m
(with m > 1), we also speak of a homogeneous system of linear
equations. (Since the equation consists of m separate equations in F, coming from
the coordinates of F
m
.)
If b W 0, the equation
f(x) = b
(again to be solved for x V ) is called an inhomogeneous linear equation, or in
the case V = F
n
, W = F
m
, an inhomogeneous system of linear equations. The
equation or system of equations is called consistent, if it has a solution, i.e., if
b im(f).
With the theory we have built so far, the following result is essentially trivial.
48
13.2. Theorem. Let f : V W be a linear map between two F-vector spaces.
(1) The solution set of the homogeneous linear equation f(x) = 0 forms a
linear subspace U of V.
(2) Let b W 0. If the inhomogeneous linear equation f(x) = b is con-
sistent, and a V is a solution, then the set of all solutions is the coset
a + U.
Proof.
(1) The solution set U is exactly the kernel of f, which is a linear subspace
of V by Lemma 8.5.
(2) Let x be any solution. Then f(x a) = f(x) f(a) = b b = 0, so
x a U, and x a +U. Conversely, if x a +U, then f(x) = f(a) = b.

13.3. Example. Consider the wave equation

2
f
t
2
= c
2

2
f
x
2
for f (
2
(R [0, ]), with boundary conditions f(t, 0) = f(t, ) = 0 and initial
conditions f(0, x) = f
0
(x) and
f
t
(0, x) = 0. If we ignore the rst initial condition
for a moment, we can consider this as a homogeneous linear equation, where we
let
V = f (
2
(R[0, ]) : t R : f(t, 0) = f(t, ) = 0, x ]0, [ :
f
t
(0, x) = 0
and W = ((R [0, ]), and the linear map V W is the wave operator
w : f

2
f
t
2
c
2

2
f
x
2
.
We can nd fairly easily a bunch of solutions using the trick of separating the
variables we look for solutions of the form f(t, x) = g(t)h(x). This leads to an
equation
1
c
2
g

(t)
g(t)
=
h

(x)
h(x)
,
and the common value of both sides must be constant. The boundary conditions
then force h(x) = sin kx (up to scaling) for some k 1, and then g(t) = cos kct
(again up to scaling). Since we know that the solution set is a linear subspace, we
see that all linear combinations
f(t, x) =
n

k=1
a
k
cos kct sin kx
are solutions. Such a solution has
f(0, x) =
n

k=1
a
k
sin kx ,
so if f
0
is of this form, we have found a (or the) solution to the original prob-
lem. Otherwise, we have to use some input from Analysis, which tells us that we
can approximate f
0
by linear combinations as above and that the corresponding
solutions will approximate the solution we are looking for.
Let us now look at the more familiar case where V = F
n
and W = F
m
, so that
we have a system of m linear equations in n variables. This is most conveniently
49
written in matrix notation as Ax = 0 in the homogeneous case and Ax = b in
the inhomogeneous case, where x F
n
and 0 F
m
or b F
m
are considered as
column vectors.
13.4. Algorithm. To solve a homogeneous system of linear equations Ax = 0,
use elementary row operations to bring A into reduced row echelon form; then read
o a basis of the kernel (which is the solution space) according to Lemma 12.12.
13.5. Algorithm. To solve an inhomogeneous system of linear equations Ax = b,
let A

= (A[b) denote the extended matrix of the system (the matrix A with b
attached as an (n + 1)st column). Use elementary row operations to bring A

into reduced row echelon form. The system is consistent if and only if n + 1 is
not one of the j
k
, i.e., the last column does not contain the leading 1 of a row.
In this case, the rst n coordinates of v
n+1
(in the notation of Lemma 12.12)
give a solution of the system. A basis of the solution space of the corresponding
homogeneous system can be read o from the rst n columns of the reduced row
echelon form of A

.
Note that the last column does not contain a leading 1 of a row if and only if the
rank of the rst n columns equals the rank of all n + 1 columns, i.e., if and only
if rk(A) = rk(A

). The latter is equivalent to saying that b is in the linear hull of

the columns of A, which is the image of A as a linear map. The statement on how
to nd a solution is then easily veried.
13.6. Example. Consider the following system of linear equations:
x + y + z + w = 0
x + 2y + 3z + 4w = 2
x + 3y + 5z + 7w = 4
We will solve it according to the procedure outlined above. The extended matrix is
A

=
_
_
1 1 1 1 0
1 2 3 4 2
1 3 5 7 4
_
_
.
We transform it into reduced row echelon form:
_
_
1 1 1 1 0
1 2 3 4 2
1 3 5 7 4
_
_

_
_
1 1 1 1 0
0 1 2 3 2
0 2 4 6 4
_
_

_
_
1 0 1 2 2
0 1 2 3 2
0 0 0 0 0
_
_
Since the last column does not contain the leading 1 of a row, the system is
consistent, and a solution is given by (x, y, z, w) = (2, 2, 0, 0). The kernel of the
non-extended matrix has basis (1, 2, 1, 0), (2, 3, 0, 1). So all solutions are given
by
(x, y, z, w) = (2 + r + 2s, 2 2r 3s, r, s) ,
where r and s are arbitrary.
50
14. Matrices and Linear Maps
So far, we have considered matrices as representing linear maps between F
n
and F
m
. But on the other hand, we have seen earlier (see Cor. 8.11) that any
n-dimensional F-vector space is isomorphic to F
n
, the isomorphism coming from
the choice of a basis. This implies that we can use matrices to represent linear
maps between arbitrary nite-dimensional vector spaces. One important thing to
keep in mind here is that this representation will depend on the bases chosen for
the two vector spaces it does not make sense to say that A is the matrix of f,
one has to say that A is the matrix of f with respect to the chosen bases.
14.1. Denition. If V is an F-vector space with basis v
1
, . . . , v
n
, then the iso-
morphism

(v
1
,...,v
n
)
: F
n
V , (
1
, . . . ,
n
)
1
v
1
+ . . .
n
v
n
is called the canonical basis isomorphism with respect to the basis v
1
, . . . , v
n
.
14.2. Denition. Let V and W be F-vector spaces with bases v
1
, . . . , v
n
and
w
1
, . . . , w
m
, respectively. Let f : V W be a linear map. Then the matrix
A Mat(mn, F) that is dened by the commutative diagram
V
f
/
W
F
n

(v
1
,...,v
n
)

=
O
A
/
F
m

(w
1
,...,w
m
)

=
O
is called the matrix associated to f relative to the chosen bases. In terms of maps,
we then have

(w
1
,...,w
m
)
A
1
(v
1
,...,v
n
)
= f and
1
(w
1
,...,w
m
)
f
(v
1
,...,v
n
)
= A.
Now what happens when we change to a dierent basis? Let v
1
, . . . , v
n
and
v

1
, . . . , v

n
be two bases of the F-vector space V. For simplicity, write =
(v
1
,...,v
n
)
and

=
(v

1
,...,v

n
)
. Then we have a diagram as follows.
V
F
n

=
>
|
|
|
|
|
|
|
|
P
/
F
n

=
`B
B
B
B
B
B
B
B
14.3. Denition. The matrix P dened by the diagram above is called the basis
change matrix associated to changing the basis from v
1
, . . . , v
n
to v

1
, . . . , v

n
. Since

is an isomorphism, P Mat(n, F) is invertible.

Conversely, given an invertible matrix P Mat(n, F) and the basis v
1
, . . . , v
n
,
we can dene

= P and hence the new basis v

1
=

(e
1
), . . . , v

n
=

(e
n
)
(where, as usual, e
1
, . . . , e
n
is the canonical basis of F
n
).
51
14.4. Lemma. Let V and W be F-vector spaces, let v
1
, . . . , v
n
and v

1
, . . . , v

n
be
bases of V , and let w
1
, . . . , w
m
and w

1
, . . . , w

m
be bases of W. Let f : V W
be a linear map, and let A be the matrix associated to f relative to the bases
v
1
, . . . , v
n
and w
1
, . . . , w
m
, and let A

be the matrix associated to f relative to the

bases v

1
, . . . , v

n
and w

1
, . . . , w

m
. Then
A

= Q
1
AP
where P is the basis change matrix associated to changing the basis of V from
v
1
, . . . , v
n
to v

1
, . . . , v

n
, and Q is the basis change matrix associated to changing
the basis of W from w
1
, . . . , w
m
to w

1
, . . . , w

m
.
Proof. Write =
(v
1
,...,v
n
)
,

=
(v

1
,...,v

n
)
, =
(w
1
,...,w
m
)
and

=
(w

1
,...,w

m
)
.
We have a commutative diagram
V
f
/
W
F
n

=
z
z
z
z
z
z
z
z
P
/
A

4 F
n

O
A
/
F
m

O
F
m

bE
E
E
E
E
E
E
E
Q
o
from which the statement can be read o.
14.5. Corollary. If f : V W is a linear map between nite-dimensional F-
vector spaces and A Mat(mn, F) is the matrix associated to f relative to some
choice of bases of V and W, then the set of all matrices associated to f relative
to any choice of bases is
QAP : P Mat(n, F), Q Mat(m, F), P and Q invertible .
Proof. By Lemma 14.4, every matrix associated to f is in the given set. Conversely,
given invertible matrices P and Q, we can change the bases of V and W in such
a way that P and Q
1
are the corresponding basis change matrices. Then (by
Lemma 14.4 again) QAP is the matrix associated to f relative to the new bases.

If we choose bases that are well-adapted to the linear map, then we will obtain a
very nice matrix. This is used in the following result.
14.6. Corollary. Let A Mat(m n, F). Then there are invertible matrices
P Mat(n, F) and Q Mat(m, F) such that
QAP =
_
_
_
_
_
_
_
_
_
_
_
1 0 0 0 0
0 1 0 0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 1 0 0
0 0 0 0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 0 0
_
_
_
_
_
_
_
_
_
_
_
=
_
I
r
0
r(nr)
0
(mr)r
0
(mr)(nr)
_
,
where r = rk(f).
52
Proof. Let V = F
n
, W = F
m
, and let f : V W be the linear map given by A.
Let v
1
, . . . , v
n
be a basis of V such that v
r+1
, . . . , v
n
is a basis of ker(f). Then
w
1
= f(v
1
), . . . , w
r
= f(v
r
) are linearly independent in W, and we can extend to
a basis w
1
, . . . , w
m
. We then have
f(v
i
) =
_
w
i
if 1 i r
0 if r + 1 i n.
So the matrix A

associated to f relative to these bases has the required form.

Let P and Q
1
be the basis change matrices associated to changing the bases of
V and W from the canonical bases to the new ones. By Lemma 14.4, we then
have A

= QAP (since A is the matrix associated to f relative to the canonical

bases).
14.7. Remarks.
(1) If we say that two matrices A, A

Mat(m n, F) are equivalent if there

are invertible matrices P Mat(n, F) and Q Mat(m, F) such that
A

= QAP (exercise: this really denes an equivalence relation), then

Cor. 14.6 tells us that A and A

are equivalent if and only if rk(A) = rk(A

).
To see this, rst note that if A and A

are equivalent, they must have

the same rank (since the rank does not change under multiplication by
invertible matrices). Then Cor. 14.6 tells us that if A has rank r, it is
equivalent to the matrix given there, so any two matrices of rank r are
equivalent to the same matrix.
(2) Recall that by Remark 12.2, row operations on a matrix A correspond to
multiplication on the left by an invertible matrix, and column operations
on A correspond to multiplication on the right by an invertible matrix.
Interpreting A as the matrix associated to a linear map relative to some
bases, we see that row operations correspond to changing the basis of the
target space (containing the columns) of A, whereas column operations
correspond to changing the basis of the domain space (containing the rows)
of A. The result of Cor. 14.6 then also means that any matrix A can be
transformed into the given simple form by elementary row and column
operations. The advantage of this approach is that by keeping track of the
operations, we can also determine the matrices P and Q explicitly, much
in the same way as when inverting a matrix.
14.8. Endomorphisms. If we consider endomorphims f : V V , then there is
only one basis to choose. If A is the matrix associated to f relative to one basis,
and P is the basis change matrix associated to changing that basis to another one,
then the matrix associated to f relative to the new basis will be A

= P
1
AP, see
Lemma 14.4. Matrices A, A

Mat(n, F) such that there is an invertible matrix

P Mat(n, F) with A

= P
1
AP are said to be similar. This denes again an
equivalence relation (exercise).
We have seen that it is easy to classify matrices with respect to equivalence: the
equivalence class is determined by the rank. In contrast to this, the classication
of matrices with respect to similarity is much more complicated. For example, the
multiplication by endomorphism (for F) has matrix I
n
regardless of the
basis, and so I
n
and I
n
are not similar if ,= .
53
14.9. Example. As another example, consider the matrices
M
,t
=
_
t
0
_
.
The corresponding endomorphism f
,t
has ker(f
,t
id) = 0 if ,= , and has
nontrivial kernel otherwise. This shows that M
,t
and M
,u
can be similar only
when = . Since dimker(f
,t
id) is 1 if t ,= 0 and 2 if t = 0, M
,0
and M
,1
are not similar. On the other hand, M
,t
is similar to M
,1
if t ,= 0, since
_
t
0
_
=
_
1 0
0 t
1
__
1
0
__
1 0
0 t
_
.
This example gives you a rst glimpse of the classication theorem, the Jordan
Normal Form Theorem, which will be a topic later.
14.10. The Trace. For purposes of classication, it is useful to have invariants,
i.e., functions that are constant on the equivalence classes. In the case of equiva-
lence of matrices, the rank is an invariant, and in this case, it gives the complete
classication. The rank is (of course) still an invariant with respect to similar-
ity, but as the example above shows, it is by no means sucient to separate the
classes. Here is another invariant.
14.11. Denition. For A = (a
ij
) Mat(n, F), we dene the trace of A to be
Tr(A) = a
11
+ a
22
+ + a
nn
.
14.12. Lemma. If A Mat(m n, F) and B Mat(n m, F), so that both
products AB and BA are dened, then
Tr(AB) = Tr(BA) .
Proof. The (i, i)-entry of AB is

n
j=1
a
ij
b
ji
. The (j, j)-entry of BA is

m
i=1
b
ji
a
ij
.
So we get
Tr(AB) =
m

i=1
n

j=1
a
ij
b
ji
=
n

j=1
m

i=1
b
ji
a
ij
= Tr(BA) .

14.13. Corollary. Let A, A

Mat(n, F) be similar. Then Tr(A) = Tr(A

).
Proof. There is an invertible matrix P Mat(n, F) such that A

= P
1
AP. It
follows that
Tr(A

) = Tr(P
1
AP) = Tr(AP P
1
) = Tr(A) .

This allows us to make the following denition.

54
14.14. Denition. Let V be a nite-dimensional F-vector space and f : V V
an endomorphism of V. We dene the trace of f, Tr(f), to be the trace of any
matrix associated to f relative to some basis of V.
Note that Tr(f) is well-dened, since all matrices associated to f have the same
trace according to Cor. 14.13.
In the next section, we will introduce another invariant, which is even more im-
portant than the trace: the determinant.
14.15. Remark. To nish o this section, let us remark that, having chosen bases
of the F-vector spaces V and W of dimensions n and m, respectively, we obtain
an isomorphism
Hom(V, W)

=
Mat(mn, F) , f A,
where A is the matrix associated to f relative to the chosen bases. In particular,
we see that dimHom(V, W) = mn.
15. Determinants
Let V be a real vector space of dimension n. The determinant will be a number
associated to an endomorphism f of V that tells us how f scales oriented volume
in V. So we have to think a little bit about functions that dene oriented volume.
We will only consider parallelotopes; these are the bodies spanned by n vectors
in V :
P(v
1
, . . . , v
n
) =
1
v
1
+ +
n
v
n
:
1
, . . . ,
n
[0, 1]
If V = R
n
, then P(v
1
, . . . , v
n
) is the image of the unit cube P(e
1
, . . . , e
n
) under
the linear map that sends the canonical basis vectors e
1
, . . . , e
n
to v
1
, . . . , v
n
.
Now let D : V
n
R be a function that is supposed to measure oriented volume of
parallelotopes D(v
1
, . . . , v
n
) gives the volume of P(v
1
, . . . , v
n
). What properties
should such a function D satisfy?
One property should certainly be that the volume vanishes when the parallelotope
is of lower dimension, i.e., when its spanning vectors are linearly dependent. It
will be sucient to only consider the special case when two of the vectors are
equal:
D(v
1
, . . . , v
n
) = 0 if v
i
= v
j
for some 1 i < j n.
Also, volume should scale corresponding to scaling of the vectors:
D(v
1
, . . . , v
i1
, v
i
, v
i+1
, . . . , v
n
) = D(v
1
, . . . , v
n
) .
Finally, volumes are additive in the following sense:
D(v
1
, . . . , v
i1
, v
i
+ v

i
, v
i+1
, . . . , v
n
)
= D(v
1
, . . . , v
i1
, v
i
, v
i+1
, . . . , v
n
) + D(v
1
, . . . , v
i1
, v

i
, v
i+1
, . . . , v
n
) .
The last two properties can be stated simply by saying that D is linear in each
argument separately. Such a function is said to be multilinear. A multilinear
function satisfying the rst property is said to be alternating. So the functions we
are looking for are alternating multilinear functions from V
n
to R.
Note that it makes sense to talk about alternating multilinear functions over any
eld F, not just over R (even though we cannot talk about volumes any more).
So we will from now on allow arbitrary elds again.
55
15.1. Denition. Let V be an n-dimensional F-vector space. An alternating
multilinear function D : V
n
F is called a determinantal function on V.
How many determinantal functions are there? First, it is pretty clear that the set
of all determinantal functions on V forms an F-vector space. So the question we
should ask is, what is the dimension of this vector space?
Before we state the relevant theorem, let us rst prove a few simple properties of
determinantal functions.
15.2. Lemma. Let V be an n-dimensional F-vector space, and let D : V
n
F
be a determinantal function on V.
(1) If v
1
, . . . , v
n
V are linearly dependent, then D(v
1
, . . . , v
n
) = 0.
(2) If we add a scalar multiple of v
i
to v
j
, where i ,= j, then D(v
1
, . . . , v
n
) is
unchanged.
(3) If we interchange two of the vectors v
1
, . . . , v
n
V , then D(v
1
, . . . , v
n
)
changes sign.
Proof.
(1) If v
1
, . . . , v
n
are linearly dependent, then one of them, say v
i
, will be a
linear combination of the others, say
v
i
=

j=i

j
v
j
.
This implies
D(v
1
, . . . , v
i
, . . . , v
n
) = D(v
1
, . . . ,

j=i

j
v
j
, . . . , v
n
)
=

j=i

j
D(v
1
, . . . , v
j
, . . . , v
j
, . . . , v
n
)
=

j=i

j
0 = 0 .
(2) Say, we replace v
j
by v
j
+ v
i
. Assuming that i < j, we have
D(v
1
, . . . ,v
i
, . . . , v
j
+ v
i
, . . . , v
n
)
= D(v
1
, . . . , v
i
, . . . , v
j
, . . . , v
n
) + D(v
1
, . . . , v
i
, . . . , v
i
, . . . , v
n
)
= D(v
1
, . . . , v
n
) + 0 = D(v
1
, . . . , v
n
) .
(3) To interchange v
i
and v
j
(with i < j), we proceed as follows.
D(v
1
, . . . , v
i
, . . . , v
j
, . . . , v
n
) = D(v
1
, . . . , v
i
, . . . , v
j
+ v
i
, . . . , v
n
)
= D(v
1
, . . . , v
i
(v
j
+ v
i
), . . . , v
j
+ v
i
, . . . , v
n
)
= D(v
1
, . . . , v
j
, . . . , (v
j
+ v
i
) + (v
j
), . . . , v
n
)
= D(v
1
, . . . , v
j
, . . . , v
i
, . . . , v
n
) .
Alternatively, we can use that (omitting all except the ith and jth argu-
ments)
0 = D(v
i
+ v
j
, v
i
+ v
j
)
= D(v
i
, v
i
) + D(v
i
, v
j
) + D(v
j
, v
i
) + D(v
j
, v
j
)
= D(v
i
, v
j
) + D(v
j
, v
i
) .

56
15.3. Theorem. Let V be an n-dimensional F-vector space, with basis v
1
, . . . , v
n
,
and let F. Then there is a unique determinantal function D : V
n
F such
that D(v
1
, . . . , v
n
) = . In particular, the determinantal functions on V form a
one-dimensional F-vector space.
Proof. As usual, we have to prove existence and uniqueness, and we will start
with uniqueness. Let w
1
, . . . , w
n
be vectors in V , and let A be the matrix whose
columns are given by the coecients of the w
j
when written as linear combinations
of the v
i
. If the w
j
are linearly dependent, then D(w
1
, . . . , w
n
) = 0. Otherwise,
the matrix A is invertible, and we can transform it into the identity matrix by
elementary column operations. The multilinearity of D and Lemma 15.2 tell us
how the value of D changes in the process: we see that
D(w
1
, . . . , w
n
) = (1)
k

1
D(v
1
, . . . , v
n
) = (1)
k

1
,
where k is the number of times we have swapped two columns and is the product
of all the scaling factors we have used when scaling a column. Note that the
identity matrix corresponds to v
1
, . . . , v
n
. This shows that there is at most one
choice for D(w
1
, . . . , w
n
).
We cannot use the observation made in the uniqueness proof easily to show exis-
tence (we would have to show that (1)
k

1
does not depend on the sequence of
elementary column operations we have performed in order to obtain I
n
). Instead,
we use induction on the dimension n of V.
As the base case, we consider n = 0. Then V = 0, the basis is empty, and
V
0
has just one element (which coincides with the empty basis). So the function
that sends this element to is trivially a determinantal function with the required
property. (If you suer from horror vacui, i.e. you are afraid of the empty set,
you can consider n = 1. Then V = L(v
1
), and the required function is given by
sending v
1
V
1
= V to .)
For the induction step, we assume n 1 and let W = L(v
2
, . . . , v
n
). By the induc-
tion hypothesis, there is a determinantal function D

on W that takes the value

on (v
2
, . . . , v
n
). Any element w
j
V can be written uniquely as w
j
=
j
v
1
+ w

j
with w

j
W. We now set
D(w
1
, . . . , w
n
) =
n

j=1
(1)
j1

j
D

(w

1
, . . . , w

j1
, w

j+1
, . . . , w

n
)
and have to check that D is a determinantal function on V . We rst verify that
D is linear in w
k
. This follows from the observation that each term in the sum is
linear in w
k
=
k
v
1
+w

k
the term with j = k only depends on w
k
through
k
,
and the other terms only depend on w
k
through w

k
, which is linear in w
k
; also
D

is linear in each of its arguments. Next assume that w

k
= w
l
for k < l. Then
w

k
= w

l
, and so in all terms that have j / k, l, the value of D

is zero. The
remaining terms are, writing w
k
= w
l
= v
1
+ w

,
(1)
k1
D

(w

1
, . . . , w

k1
, w

k+1
, . . . , w

l1
, w

, w

l+1
, . . . , w

n
)
+ (1)
l1
D

(w

1
, . . . , w

k1
, w

, w

k+1
, . . . , w

l1
, w

l+1
, . . . , w

n
) .
These terms cancel since we have to swap adjacent arguments (l k 1) times to
go from one value of D

to the other, which results in a sign of (1)

lk1
.
Finally, we have D(v
1
, v
2
, . . . , v
n
) = 1 D

(v
2
, . . . , v
n
) = .
57
We can now make use of this fact in order to dene determinants of matrices and
of endomorphisms.
15.4. Denition. Let n 0. The determinant on Mat(n, F) is the unique deter-
minantal function on the columns of the n n matrices that takes the value 1 on
the identity matrix I
n
. If A Mat(n, F), then then its value det(A) on A is called
the determinant of A.
If A = (a
ij
) is written as an n n array of entries, we also write
det(A) =

a
11
a
12
a
1n
a
21
a
22
a
2n
.
.
.
.
.
.
.
.
.
.
.
.
a
n1
a
n2
a
nn

15.5. Remarks.
(1) Note that det(A) ,= 0 is equivalent to A invertible.
(2) The uniqueness proof gives us a procedure to compute determinants: we
perform elementary column operations on A, keeping track of the scalings
and swappings, until we get a zero column (then det(A) = 0), or we reach
the identity matrix.
15.6. Example. We compute a determinant by elementary column operations.
Note that we can avoid divisions (and hence fractions) by choosing the operations
cleverly.

1 2 3 4
2 1 4 3
3 4 2 1
4 3 1 2

1 0 0 0
2 3 2 5
3 2 7 11
4 5 11 14

1 0 0 0
2 1 2 5
3 12 7 11
4 17 11 14

1 0 0 0
0 1 0 0
21 12 17 49
30 17 23 71

1 0 0 0
0 1 0 0
21 12 17 2
30 17 23 2

= 2

1 0 0 0
0 1 0 0
21 12 1 17
30 17 1 23

= 2

1 0 0 0
0 1 0 0
0 0 1 0
51 29 1 40

= 2 40

1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1

= 80
15.7. Lemma (Expansion by Rows). Let A Mat(n, F) with n 1. For
1 i, j n, denote by A
ij
the matrix obtained from A by deleting the ith row and
jth column. Then for all 1 i n, we have
det(A) =
n

j=1
(1)
ji
a
ij
det(A
ij
) .
This is called the expansion of the determinant by the ith row.
Proof. As in the proof of Thm. 15.3 above, we check that the right hand side is a
determinantal function that takes the value 1 on I
n
.
58
15.8. Examples. For n = 0, we nd that the empty determinant is 1. For n = 1
and A = (a), we have det(A) = a. For n = 2 and n = 3, we obtain by an
application of the lemma the following formulas.

a b
c d

= ad bc

a b c
d e f
g h i

= aei + bfg + cdh afh bdi ceg

We will see later how this generalizes to larger n.
15.9. Denition. Let V be an F-vector space of dimension n, and let f : V V
be an endomorphism. We x a nontrivial determinantal function D : V
n
F.
Then
D

: V
n
F , (v
1
, . . . , v
n
) D
_
f(v
1
), . . . , f(v
n
)
_
is again a determinantal function, hence there is F such that D

= D (since
D is a basis of the space of determinantal functions on V ). We set det(f) = and
call det(f) the determinant of f.
Note that this is well-dened, since any other choice of D diers from the one we
have made by a non-zero scalar, by which D

is also multiplied, and so it drops

out when computing .
15.10. Theorem. Let V be an F-vector space of dimension n, and let f, g : V
V be two endomorphisms. Then det(f g) = det(f) det(g).
Proof. Let D be a xed nontrivial determinantal function on V , and let D

be D
after f and let D

be D after g as in the denition above. Then

D
_
f g(v
1
), . . . , f g(v
2
)
_
= D

_
g(v
1
), . . . , g(v
n
)
_
= det(f)D
_
g(v
1
), . . . , g(v
n
)
_
= det(f)D

(v
1
, . . . , v
n
)
= det(f) det(g)D(v
1
, . . . , v
n
) .
The denition then implies that det(f g) = det(f) det(g).
15.11. Lemma. Let V be an F-vector space of dimension n, let f : V V be an
endomorphism. We x a basis v
1
, . . . , v
n
of V and let A be the matrix associated
to f relative to this basis. Then det(A) = det(f).
Proof. Let D : V
n
F be the determinantal function that takes the value 1 on
the xed basis. The columns of A contain the coecients of f(v
1
), . . . , f(v
n
) with
respect to this basis, hence
det(A) = D
_
f(v
1
), . . . , f(v
n
)
_
= det(f)D(v
1
, . . . , v
n
) = det(f) .

59
15.12. Theorem (Multiplicativity of the Determinant).
Let A, B Mat(n, F). Then det(AB) = det(A) det(B).
Proof. Consider V = F
n
and let (for sake of clarity) f
A
, f
B
: V V be the
endomorphisms given by A and B. By Lemma 15.11 and Thm. 15.10, we have
det(AB) = det(f
A
f
B
) = det(f
A
) det(f
B
) = det(A) det(B) .

15.13. Theorem. Let A Mat(n, F). Then det(A

) = det(A).
Proof. We show that A det(A

) is a determinantal function of the columns

of A. First, we have
det(A) = 0 rk(A) < n rk(A

) < n det(A

) = 0 ,
so our function is alternating. Second, we have to show that det(A

) is linear in
each of the columns of A. This is obviously equivalent to saying that det(A) is
linear in each of the rows of A. To check that this is the case for the ith row, we
expand det(A) by the ith row according to Lemma 15.7. For A = (a
ij
),
det(A) =
n

j=1
(1)
ji
a
ij
det(A
ij
) .
Now in A
ij
the ith row of A has been removed, so det(A
ij
) does not depend
on the ith row of A; linearity is then clear from the formula. Finally, we have
det(I

n
) = det(I
n
) = 1, so det(A

) must coincide with det(A) because of the

uniqueness of determinantal functions.
15.14. Corollary (Expansion by Columns). We can also expand determinants
by columns. Let n 1 and A = (a
ij
) Mat(n, F); we use the notation A
ij
as
before. Then for 1 j n,
det(A) =
n

i=1
(1)
ji
a
ij
det(A
ij
) .
Proof. We have
det(A) = det(A

) =
n

j=1
(1)
ji
a
ji
det
_
(A

)
ij
_
=
n

j=1
(1)
ji
a
ji
det
_
(A
ji
)

_
=
n

j=1
(1)
ji
a
ji
det(A
ji
)
=
n

i=1
(1)
ji
a
ij
det(A
ij
) .

15.15. Example. A matrix A Mat(n, F) is said to be orthogonal if AA

= I
n
.
What can we deduce about det(A)? Well,
1 = det(I
n
) = det(AA

) = det(A) det(A

) = det(A)
2
,
so det(A) = 1.
60
15.16. Denition. Let A Mat(n, F) with n 1. Then the adjugate matrix
of A (sometimes called the adjoint matrix, but this has also other meanings) is
the matrix

A Mat(n, F) whose (i, j)-entry is (1)
ji
det(A
ji
). Here A
ij
is, as
before, the matrix obtained from A by removing the ith row and jth column. Note
the reversal of indices

A
ij
= (1)
ji
det(A
ji
) and not det(A
ij
)!
15.17. Proposition. Let A Mat(n, F) with n 1. Then
A

A =

AA = det(A)I
n
.
In particular, if A is invertible, then det(A) ,= 0, and
A
1
= det(A)
1

A.
Proof. The (i, k)-entry of A

A is
n

j=1
a
ij
(1)
kj
det(A
kj
) .
If i = k, then this is just the formula that expands det(A) by the ith row, so A

A
has diagonal entries equal to det(A). If i ,= k, then the result is unchanged if we
modify the kth row of A (since A
kj
does not involve the kth row of A). So we get
the same result as for the matrix A

= (a

ij
) which we obtain from A by replacing
the kth row by the ith row. We nd that
0 = det(A

) =
n

j=1
(1)
kj
a

kj
det(A

kj
) =
n

j=1
(1)
kj
a
ij
det(A
kj
) .
This shows that the o-diagonal entries of A

A vanish. The assertion on

AA is
proved in the same way (or by applying what we have just proved to A

).
16. The Determinant and the Symmetric Group
If A = (a
ij
) Mat(n, F) and we recursively expand det(A) along the rst row
(say), then we end up with an expression that is a sum of products of n matrix
entries with a plus or minus sign. For example:

a
11
a
12
a
21
a
22

= a
11
a
22
a
12
a
21

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

= a
11
a
22
a
33
a
11
a
23
a
32
a
12
a
21
a
33
+ a
12
a
23
a
31
+ a
13
a
21
a
32
a
13
a
22
a
31
Since in the process of recursively expanding, the row and column of the entry we
multiply with are removed, each product in the nal expression has the form
a
1,(1)
a
2,(2)
a
n,(n)
where : 1, 2, . . . , n 1, 2, . . . , n is a bijective map, a so-called permutation
of 1, 2, . . . , n. All these permutations together form a group, the symmetric
group.
61
16.1. Denition. A group is a set G, together with a map m : G G G,
usually written m(x, y) = x y or xy, such that the following axioms are satised.
(1) For all x, y, z G, we have (xy)z = x(yz) (associativity).
(2) There is e G such that for all x G, we have ex = xe = x (identity).
(3) For all x G, there is x
1
G such that xx
1
= x
1
x = e (inverse).
As usual, the identity element and the inverse of x are uniquely determined.
16.2. Examples.
(1) If F is a eld, then we have the additive group of F (F, +) and the mul-
tiplicative group of F (F 0, ). If V is a vector space, we have the
additive group (V, +). All these groups are commutative or abelian: they
satisfy xy = yx (or x + y = y + x in additive notation).
(2) If X is a set, then the set of all bijective maps f : X X forms a group,
where the multiplication is given by composition of maps. The identity
element is given by id
X
, the inverse by the inverse map. This group is
called the symmetric group of X and denoted o(X). If X = 1, 2, . . . , n,
we also write o
n
. If X has more than two elements, this group is not
abelian (Exercise!). Note that #o
n
= n!.
(3) Let F be a eld, n 0. The set of all invertible matrices in Mat(n, F)
forms a group under matrix multiplication. This group is called the general
linear group and denoted by GL
n
(F).
16.3. The Leibniz Formula. From the considerations above, we know that
there is a formula, for A = (a
ij
) Mat(n, F):
det(A) =

S
n
() a
1,(1)
a
2,(2)
a
n,(n)
with a certain map : o
n
1. We call () the sign of the permutation .
It remains to determine the map . Let us introduce the permutation matrix
P() = (
(i),j
); this is the matrix whose entries at positions (i, (i)) are 1, and
the other entries are 0. By specializing the formula above, we nd that
() = det
_
P()
_
.
We also have that
() = det
_
P()
_
= det
_
P()P()
_
= det
_
P()
_
det
_
P()
_
= ()()
so is a group homomorphism, and () = 1 when is a transposition, i.e., a
permutation that exchanges two elements and leaves the others xed. Since every
permutation can be written as a product of such transpositions (even transposi-
tions of neighboring elements), this determines uniquely: write as a product of,
say, k transpositions, then () = (1)
k
. We can also give an explicit formula.
16.4. Proposition. Let n 0, o
n
. Then
() =

1i<jn
(j) (i)
j i
.
Note that for each pair i < j, we will either have j i or i j as a factor in the
numerator (since permutes the 2-element subsets of 1, 2, . . . , n). Therefore
the right hand side is 1. Since the factor ((j) (i))/(j i) is negative if and
62
only if (j) < (i), the right hand side is (1)
m
, where m counts the number of
pairs i < j such that (i) > (j).
Proof. Denote the right hand side by

(). We rst show that

is a group
homomorphism.

() =

i<j
((j)) ((i))
j i
=

i<j
((j)) ((i))
(j) (i)
(j) (i)
j i
=

i<j
(j) (i)
j i

i<j
(j) (i)
j i
=

()

()
(Note that (i), (j) runs exactly through the two-element subsets of 1, 2, . . . , n,
and the ordering within the subset does not inuence the value of the fraction.)
Next we show that

() = 1 when is a transposition. So assume that

interchanges k and l, where k < l. Then

() = (1)
m
, where m is the number of
pairs i < j such that (j) > (i). This is the case when (i) i = k and k < j < l,
or (ii) k < i < l and j = l, or (iii) i = k and j = l. Therefore m = 2(l k 1) +1
is odd, and

() = 1.
As discussed above, these two properties determine

uniquely, hence =

.
16.5. Remark. As a concluding remark, let us recall the following equivalences,
which apply to an endomorphism f of a nite-dimensional vector space V.
f is an automorphism ker(f) = 0 det(f) ,= 0 .
17. Eigenvalues and Eigenvectors
We continue with the study of endomorphisms f : V V . Such an endomorphism
is particularly easy to understand if it is just multiplication by a scalar: f = id
V
for some F. Of course, these are very special (and also somewhat boring)
endomorphisms. So we look for linear subspaces of V on which f behaves in this
way.
17.1. Denition. Let V be an F-vector space, and let f : V V be an endo-
morphism. Let F. If there exists a vector 0 ,= v V such that f(v) = v,
then is called an eigenvalue of f, and v is called an eigenvector of f for the
eigenvalue . In any case, the linear subspace E

(f) = v V : f(v) = v
is called the -eigenspace of f. (So that is an eigenvalue of f if and only if
E

(f) ,= 0.)
17.2. Examples.
(1) Let V = R
2
and consider f(x, y) = (y, x). Then 1 and 1 are eigenvalues
of f, and E
1
(f) = (x, x) : x R, E
1
(f) = (x, x) : x R. The
eigenvectors (1, 1) and (1, 1) form a basis of V , and the matrix of f
relative to that basis is
_
1 0
0 1
_
.
63
(2) Let V = (

(R) be the space of innitely dierentiable functions on R.

Consider the endomorphism D : f f

. Then every R is an eigen-

value, and all eigenspaces are of dimension two:
E

(D) =
_

_
L(x 1, x x) if = 0
L(x e
x
, x e
x
) if =
2
> 0
L(x sin x, x cos x) if =
2
< 0
Since matrices can be identied with linear maps, it makes sense to speak about
eigenvalues and eigenvectors of a square matrix A Mat(n, F).
17.3. The Characteristic Polynomial. How can we nd the eigenvalues (and
eigenvectors) of a given endomorphism f, when V is nite-dimensional?
Well, we have
is an eigenvalue of f there is 0 ,= v V with f(v) = v
there is 0 ,= v V with (f id
V
)(v) = 0
ker(f id
V
) ,= 0
det(f id
V
) = 0
It is slightly more convenient to consider det(id
V
f) (which of course vanishes
if and only if det(f id
V
) = 0). If A = (a
ij
) Mat(n, F) is a matrix associated
to f relative to some basis of V, then
det(id
V
f) = det(I
n
A) =

a
11
a
12
a
1n
a
21
a
22
a
2n
.
.
.
.
.
.
.
.
.
.
.
.
a
n1
a
n2
a
nn

.
Expanding the determinant, we nd that
det(I
n
A) =
n
Tr(A)
n1
+ + (1)
n
det(A) .
This polynomial of degree n in is called the characteristic polynomial of A (or
of f). We will denote it by P
A
() (or P
f
()). By the discussion above, the
eigenvalues of A (or of f) are exactly the roots (in F) of this polynomial.
17.4. Examples.
(1) Let us come back to the earlier example f : (x, y) (y, x) on R
2
. With
respect to the canonical basis, the matrix is
_
0 1
1 0
_
, so the characteristic polynomial is

1
1

=
2
1
and has the two roots 1 and 1.
(2) Let us consider the matrix
A =
_
_
5 2 6
1 0 1
3 1 4
_
_
.
64
What are its eigenvalues and eigenspaces? We compute the characteristic
polynomial:

5 2 6
1 1
3 1 + 4

= ( 5)
_
( + 4) 1
_
+ 2
_
( + 4) 3
_
+ 6
_
1 + 3
_
=
3

2
+ 1 = ( 1)
2
( + 1) .
The roots are 1 and 1; these are therefore the eigenvalues. To nd (bases
of) the eigenspaces, note that E

(A) = ker(A I
3
). For = 1, we have
A I
3
=
_
_
4 2 6
1 1 1
3 1 5
_
_

_
_
1 0 2
0 1 1
0 0 0
_
_
(by elementary row operations), so E
1
(A) is generated by (2, 1, 1)

. For
= 1, we obtain
A + I
3
=
_
_
6 2 6
1 1 1
3 1 3
_
_

_
_
1 0 1
0 1 0
0 0 0
_
_
and E
1
(A) is generated by (1, 0, 1)

.
17.5. Diagonalizable Matrices and Endomorphisms. If the canonical basis
of F
n
consists of eigenvectors of A, then A is diagonal:
A =
_
_
_
_

1
0 0
0
2
0
.
.
.
.
.
.
.
.
.
.
.
.
0 0
n
_
_
_
_
where
1
, . . . ,
n
are the eigenvalues corresponding to the basis vectors. More
generally, if F
n
has a basis consisting of eigenvectors of A, then changing from
the canonical basis to that basis will make A diagonal: P
1
AP is diagonal, where
P is invertible. In this case, A is called diagonalizable. Similarly, if f is an
endomorphism of a nite-dimensional F-vector space V, then f is diagonalizable if
V has a basis consisting of eigenvectors of f. The matrix associated to f relative
to this basis is then diagonal.
The big question is now: when is a matrix or endomorphism diagonalizable?
This is certainly not always true. For example, in our second example above, we
only found two linearly independent eigenvectors in F
3
, and so there cannot be a
basis of eigenvectors. Another kind of example is f : (x, y) (y, x) on R
2
. The
characteristic polynomial comes out as
2
+ 1 and does not have roots in R, so
there are no eigenvalues and therefore no eigenvectors. (If we take C instead as
the eld of scalars, then we do have two roots i, and f becomes diagonalizable.)
17.6. Denition. Let V be a nite-dimensional F-vector space, f : V V an
endomorphism and F. Then dimE

(f) is called the geometric multiplicity of

the eigenvalue of f. (So the geometric multiplicity is positive if and only if is
indeed an eigenvalue.)
65
17.7. Lemma. Let V be an F-vector space and f : V V an endomorphism.
Let
1
, . . . ,
m
F be distinct, and for i = 1, . . . , m, let v
i
E

i
(f). If
v
1
+ v
2
+ + v
m
= 0 ,
then v
i
= 0 for all i.
Proof. By induction on m. The case m = 0 (or m = 1) is trivial. So assume the
claim is true for m, and consider the case with m + 1 eigenvalues. We apply the
endomorphism f
m+1
id
V
to the equation
v
1
+ v
2
+ + v
m
+ v
m+1
= 0
and obtain (note (f
m+1
id
V
)(v
m+1
) = 0)
(
1

m+1
)v
1
+ (
2

m+1
)v
2
+ + (
m

m+1
)v
m
= 0 .
By induction, we nd that (
i

m+1
)v
i
= 0 for all 1 i m. Since
i
,=
m+1
,
this implies v
i
= 0 for 1 i m. But then we must also have v
m+1
= 0.
17.8. Corollary. In the situation above, the union of bases of distinct eigenspaces
of f is linearly independent.
Proof. Consider a linear combination on the union of such bases that gives the zero
vector. By the preceding lemma, each part of this linear combination that comes
from one of the eigenspaces is already zero. Since the vectors involved there form
a basis of this eigenspace, they are linearly independent, hence all the coecients
vanish.
17.9. Example. We can use this to show once again that the power functions
f
n
: x x
n
for n N
0
are linearly independent as elements of the space P of
polynomial functions on R (say). Namely, consider the endomorphism D : P P,
f (x xf

(x)). Then D(f

n
) = nf
n
, so the f
n
are eigenvectors of D for
eigenvalues that are pairwise distinct, hence they must be linearly independent.
17.10. Corollary. Let V be a nite-dimensional F-vector space and f : V V
an endomorphism. Then f is diagonalizable if and only if

F
dimE

(f) = dimV .
Proof. By Cor. 17.8, we always have . If f is diagonalizable, then there is
a basis consisting of eigenvectors, and so we must have equality. Conversely, if
we have equality, then the union of bases of the eigenspaces will be a basis of V ,
which consists of eigenvectors of f.
17.11. Proposition. Let V be an n-dimensional F-vector space and f : V V
an endomorphism. If P
f
() has n distinct roots in F, then f is diagonalizable.
Proof. In this case, there are n distinct eigenvalues
1
, . . . ,
n
. Therefore, E

i
(f)
is nontrivial for 1 i n, which means that dimE

i
(f) 1. So
dimV = n
n

i=1
dimE

i
(f) dimV ,
and we must have equality. The result then follows by the previous corollary.
66
The converse of this statement is false in general, as the identity endomorphism
id
V
shows (for dimV 2).
However, some statement in the converse direction is true. In order to state it, we
need some preparations.
17.12. Denition. Let F be a eld. The polynomial ring in x over F, F[x], is an
F-vector space with basis 1 = x
0
, x = x
1
, x
2
, . . . , x
n
, . . . , on which a multiplication
F[x] F[x] F[x] is dened by the following two properties: (i) it is F-bilinear,
and (ii) x
m
x
n
= x
m+n
.
Written out, this means that
(a
n
x
n
+ + a
1
x + a
0
)(b
m
x
m
+ + b
1
x + b
0
)
= a
n
b
m
x
n+m
+ (a
n
b
m1
+ a
n1
b
m
)x
n+m1
+ . . .
+
_

i+j=k
a
i
b
j
_
x
k
+ + (a
1
b
0
+ a
0
b
1
)x + a
0
b
0
.
It can then be checked that F[x] is a commutative ring with unit, i.e., it satises
the axioms of a eld with the exception of the existence of multiplicative inverses.
If p(x) = a
n
x
n
+ +a
1
x +a
0
F[x] and a
n
,= 0, then p is said to have degree n;
we write deg(p) = n. In this case a
n
is called the leading coecient of p(x); if
a
n
= 1, p(x) is said to be monic.
For example, if V is an n-dimensional vector space and f : V V is an endomor-
phism, then the characteristic polynomial P
f
(x) of f is monic of degree n.
17.13. Theorem. Let p(x) = x
n
+ a
n1
x
n1
+ + a
1
x + a
0
F[x] be a monic
polynomial, and let F. If p() = 0, then there is a polynomial q(x) =
x
n1
+ b
n2
x
n2
+ + b
0
such that p(x) = (x )q(x).
Proof. If = 0, this is certainly true, since then 0 = p(0) = a
0
, and visibly p(x) =
xq(x). In general, we replace x by x + . Then the polynomial p(x) = p(x + )
is again monic of degree n, and p(0) = p() = 0, so p(x) = x q(x) with a monic
polynomial q of degree n 1. Then
p(x) = p(x ) = (x ) q(x ) = (x )q(x) ,
where q(x) = q(x ) is monic of degree n 1.
17.14. Corollary and Denition. Let p(x) = x
n
+a
n1
x
n1
+ +a
1
x +a
0

F[x] and F. Then there is a largest m N
0
such that p(x) = (x )
m
q(x)
with a polynomial q(x); we then have q() ,= 0.
This number m is called the multiplicity of the root of p; we have m > 0 if and
only if p() = 0.
Proof. Write p(x) = (x)
m
q(x) with m as large as possible. (Note that deg(p) =
m+deg(q), so m n.) The we must have q() ,= 0, since otherwise we could write
q(x) = (x )r(x), so p(x) = (x )
m+1
r(x), contradicting our choice of m.
Now we can make another denition.
67
17.15. Denition. Let V be a nite-dimensional F-vector space, f : V V an
endomorphism. Then the multiplicity of F as a root of the characteristic
polynomial P
f
(x) is called the algebraic multiplicity of the eigenvalue of f.
Note that the following statements are then equivalent.
(1) is an eigenvalue of f;
(2) the geometric multiplicity of is 1;
(3) the algebraic multiplicity of is 1.
We also know that the sum of the geometric multiplicities of all eigenvalues as well
as the sum of the algebraic multiplicities of all eigenvalues are bounded by dimV .
There is one further important relation between the multiplicities.
17.16. Theorem. Let V be a nite-dimensional F-vector space, f : V V an
endomorphism, and F. Then the geometric multiplicity of as an eigenvalue
of f is not larger than its algebraic multiplicity.
Proof. We can choose a basis v
1
, . . . , v
k
, v
k+1
, . . . , v
n
of V such that v
1
, . . . , v
k
form
a basis of the eigenspace E

(f); then k is the geometric multiplicity. The matrix

associated to f relative to this basis then has the form
A =
_
_
_
_
_
_
_
_
_
_
_
0 . . . 0 . . .
0 . . . 0 . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 . . . . . .
0 0 . . . 0 . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 . . . 0 . . .
_
_
_
_
_
_
_
_
_
_
_
=
_
I
k
B
0 C
_
.
We then have
P
f
(x) = det(xI
n
A) = det
_
(x )I
k
B
0 xI
nk
C
_
= det
_
(x )I
k
_
det(xI
nk
C) = (x )
k
q(x)
where q(x) is some monic polynomial of degree nk. We see that has multiplicity
at least k as a root of P
f
(x).
17.17. Lemma. Let f : V V be an endomorphism of an n-dimensional F-
vector space V, and let P
f
be its characteristic polynomial. Then the sum of the
algebraic multiplicities of the eigenvalues of f is at most n; it is equal to n if and
only if P
f
(x) is a product of linear factors x (with F).
Proof. By Thm. 17.13, if is a root of P
f
, we can write P
f
(x) = (x )q(x) with
a monic polynomial q of degree n 1. Continuing in this way, we can write
P
f
(x) = (x
1
)
m
1
(x
k
)
m
k
q(x)
with a monic polynomial q that does not have roots in F and distinct elements

1
, . . . ,
k
F. If F, then
P
f
() = (
1
)
m
1
(
k
)
m
k
q() ,
so if P
f
() = 0, then
1
, . . . ,
k
(since q() ,= 0). Therefore the eigenvalues
are exactly
1
, . . . ,
k
, with multiplicities m
1
, . . . , m
k
, and
m
1
+ m
2
+ + m
k
m
1
+ m
2
+ + m
k
+ deg(q) = n.
68
We have equality if and only if deg(q) = 0, i.e., if and only if q(x) = 1; then
P
f
() = (
1
)
m
1
(
k
)
m
k
is a product of linear factors.
17.18. Corollary. Let V be a nite-dimensional F-vector space and f : V V
an endomorphism. Then f is diagonalizable if and only if
(1) P
f
(x) is a product of linear factors, and
(2) for each F, its geometric and algebraic multiplicities as an eigenvalue
of f agree.
Proof. By Cor. 17.10, f is diagonalizable if and only if the sum of the geometric
multiplicities of all eigenvalues equals n = dimV. By Thm. 17.16, this implies that
the sum of the algebraic multiplicities is at least n; however it cannot be larger
than n, so it equals n as well. This already shows that geometric and algebraic
multiplicities agree. By Lemma 17.17, we also see that P
f
(x) is a product of linear
factors.
Conversely, if we can write P
f
(x) as a product of linear factors, this means that
the sum of the algebraic multiplicities is n. If the geometric multiplicities equal
the algebraic ones, their sum must also be n, hence f is diagonalizable.
17.19. Remark. If F is an algebraically closed eld, for example F = C, then
condition (1) in the corollary is automatically satised (by denition!). However,
condition (2) can still fail. It is then an interesting question to see how close we
can get to a diagonal matrix in this case. This is what the Jordan Normal Form
Theorem is about, which will be a topic for the second semester.
References
[BR1] T.S. Blyth and E.F. Robertson: Basic Linear Algebra. Springer Undergraduate
Mathematics Series, 2002.
[BR2] T.S. Blyth and E.F. Robertson: Further Linear Algebra. Springer Undergraduate
Mathematics Series, 2002.
[J] K. J anich: Linear Algebra. Springer Undergraduate Texts in Mathematics, 1994
[KM] A. Kostrykin and Y. Manin: Linear Algebra and Geometry. Gordan and Breach,
1988.