MATH 233 - Linear Algebra I Lecture Notes: Cesar O. Aguilar

MATH 233 - Linear Algebra I
Lecture Notes
Cesar O. Aguilar
Department of Mathematics
SUNY Geneseo
Lecture 0
Contents
1 Systems of Linear Equations 1

1.1 What is a system of linear equations? . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Solving linear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Geometric interpretation of the solution set . . . . . . . . . . . . . . . . . . 8
2 Row Reduction and Echelon Forms 11

2.1 Row echelon form (REF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Reduced row echelon form (RREF) . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Existence and uniqueness of solutions . . . . . . . . . . . . . . . . . . . . . . 17
3 Vector Equations 19
3.1 Vectors in Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 The linear combination problem . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 The span of a set of vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4 The Matrix Equation Ax = b 31

4.1 Matrix-vector multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 Matrix-vector multiplication and linear combinations . . . . . . . . . . . . . 33
4.3 The matrix equation problem . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5 Homogeneous and Nonhomogeneous Systems 41

5.1 Homogeneous linear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2 Nonhomogeneous systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6 Linear Independence 49
6.1 Linear independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6.2 The maximum size of a linearly independent set . . . . . . . . . . . . . . . . 53
7 Introduction to Linear Mappings 57

7.1 Vector mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
7.2 Linear mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
7.3 Matrix mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
7.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3
CONTENTS
8 Onto, One-to-One, and Standard Matrix 67

8.1 Onto Mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
8.2 One-to-One Mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
8.3 Standard Matrix of a Linear Mapping . . . . . . . . . . . . . . . . . . . . . . 71
9 Matrix Algebra 75
9.1 Sums of Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
9.2 Matrix Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
9.3 Matrix Transpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
10 Invertible Matrices 83
10.1 Inverse of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
10.2 Computing the Inverse of a Matrix . . . . . . . . . . . . . . . . . . . . . . . 85
10.3 Invertible Linear Mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
11 Determinants 89
11.1 Determinants of 2 × 2 and 3 × 3 Matrices . . . . . . . . . . . . . . . . . . . . 89
11.2 Determinants of n × n Matrices . . . . . . . . . . . . . . . . . . . . . . . . . 93
11.3 Triangular Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
12 Properties of the Determinant 97

12.1 ERO and Determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
12.2 Determinants and Invertibility of Matrices . . . . . . . . . . . . . . . . . . . 100
12.3 Properties of the Determinant . . . . . . . . . . . . . . . . . . . . . . . . . . 100
13 Applications of the Determinant 103

13.1 The Cofactor Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
13.2 Cramer’s Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
13.3 Volumes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
14 Vector Spaces 109

14.1 Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
14.2 Subspaces of Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
15 Linear Maps 117

15.1 Linear Maps on Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 117
15.2 Null space and Column space . . . . . . . . . . . . . . . . . . . . . . . . . . 121
16 Linear Independence, Bases, and Dimension 125

16.1 Linear Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
16.2 Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
16.3 Dimension of a Vector Space . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
17 The Rank Theorem 133

17.1 The Rank of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
4
Lecture 0
18 Coordinate Systems 137

18.1 Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
18.2 Coordinate Mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
18.3 Matrix Representation of a Linear Map . . . . . . . . . . . . . . . . . . . . . 142
19 Change of Basis 147

19.1 Review of Coordinate Mappings on Rn . . . . . . . . . . . . . . . . . . . . . 147
19.2 Change of Basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
20 Inner Products and Orthogonality 153

20.1 Inner Product on Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
20.2 Orthogonality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
20.3 Coordinates in an Orthonormal Basis . . . . . . . . . . . . . . . . . . . . . . 158
21 Eigenvalues and Eigenvectors 163

21.1 Eigenvectors and Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . 163
21.2 When λ = 0 is an eigenvalue . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
22 The Characteristic Polynomial 169

22.1 The Characteristic Polynomial of a Matrix . . . . . . . . . . . . . . . . . . . 169
22.2 Eigenvalues and Similarity Transformations . . . . . . . . . . . . . . . . . . 176
23 Diagonalization 179
23.1 Eigenvalues of Triangular Matrices . . . . . . . . . . . . . . . . . . . . . . . 179
23.2 Diagonalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
23.3 Conditions for Diagonalization . . . . . . . . . . . . . . . . . . . . . . . . . . 182
24 Diagonalization of Symmetric Matrices 187

24.1 Symmetric Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
24.2 Eigenvectors of Symmetric Matrices . . . . . . . . . . . . . . . . . . . . . . . 188
24.3 Symmetric Matrices are Diagonalizable . . . . . . . . . . . . . . . . . . . . . 188
25 The PageRank Algortihm 191

25.1 Search Engine Retrieval Process . . . . . . . . . . . . . . . . . . . . . . . . . 191
25.2 A Description of the PageRank Algorithm . . . . . . . . . . . . . . . . . . . 192
25.3 Computation of the PageRank Vector . . . . . . . . . . . . . . . . . . . . . . 195
26 Discrete Dynamical Systems 197

26.1 Discrete Dynamical Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
26.2 Population Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
26.3 Stability of Discrete Dynamical Systems . . . . . . . . . . . . . . . . . . . . 199
5
Lecture 1
Lecture 1
Systems of Linear Equations
In this lecture, we will introduce linear systems and the method of row reduction to solve
them. We will introduce matrices as a convenient structure to represent and solve linear
systems. Lastly, we will discuss geometric interpretations of the solution set of a linear
system in 2- and 3-dimensions.
1.1 What is a system of linear equations?

Definition 1.1: A system of m linear equations in n unknown variables x1 , x2 , . . . , xn
is a collection of m equations of the form
a11 x1 + a12 x2 + a13 x3 + · · · + a1n xn = b1

a21 x1 + a22 x2 + a23 x3 + · · · + a2n xn = b2
a31 x1 + a32 x2 + a33 x3 + · · · + a3n xn = b3 (1.1)
.. .. .. .. ..
. . . . .
am1 x1 + am2 x2 + am3 x3 + · · · + amn xn = bm
The numbers aij are called the coefficients of the linear system; because there are m equa-
tions and n unknown variables there are thefore m × n coefficients. The main problem with
a linear system is of course to solve it:
Problem: Find a list of n numbers (s1 , s2 , . . . , sn ) that satisfy the system of linear equa-
tions (1.1).
In other words, if we substitute the list of numbers (s1 , s2 , . . . , sn ) for the unknown
variables (x1 , x2 , . . . , xn ) in equation (1.1) then the left-hand side of the ith equation will
equal bi . We call such a list (s1 , s2 , . . . , sn ) a solution to the system of equations. Notice
that we say “a solution” because there may be more than one. The set of all solutions to a
linear system is called its solution set. As an example of a linear system, below is a linear
1
system consisting of m = 2 equations and n = 3 unknowns:
x1 − 5x2 − 7x3 = 0
5x2 + 11x3 = 1
Here is a linear system consisting of m = 3 equations and n = 2 unknowns:
−5x1 + x2 = −1
πx1 − 5x2 = 0
√
63x1 − 2x2 = −7
And finally, below is a linear system consisting of m = 4 equations and n = 6 unknowns:
−5x1 + x3 − 44x4 − 55x6 = −1

√
πx1 − 5x2 − x3 + 4x4 − 5x5 + 5x6 =0
√ 1
63x1 − 2x2 − 15 x3 + ln(3)x4 + 4x5 − 33 x6 =0
√ 1
63x1 − 2x2 − x3 − 18 x4 − 5x6 =5
5
Example 1.2. Verify that (1, 2, −4) is a solution to the system of equations
2x1 + 2x2 + x3 = 2
x1 + 3x2 − x3 = 11.
Is (1, −1, 2) a solution to the system?

Solution. The number of equations is m = 2 and the number of unknowns is n = 3. There
are m × n = 6 coefficients: a11 = 2, a12 = 1, a13 = 1, a21 = 1, a22 = 3, and a23 = −1. And
b1 = 0 and b2 = 11. The list of numbers (1, 2, −4) is a solution because
2 · (1) + 2(2) + (−4) = 2
(1) + 3 · (2) − (−4) = 11
On the other hand, for (1, −1, 2) we have that
2(1) + 2(−1) + (2) = 2
but
1 + 3(−1) − 2 = −4 6= 11.
Thus, (1, −1, 2) is not a solution to the system.
A linear system may not have a solution at all. If this is the case, we say that the linear
system is inconsistent:
2
Lecture 1
INCONSISTENT ⇔ NO SOLUTION
A linear system is called consistent if it has at least one solution:
CONSISTENT ⇔ AT LEAST ONE SOLUTION
We will see shortly that a consistent linear system will have either just one solution or
infinitely many solutions. For example, a linear system cannot have just 4 or 5 solutions. If
it has multiple solutions, then it will have infinitely many solutions.
Example 1.3. Show that the linear system does not have a solution.
−x1 + x2 = 3
x1 − x2 = 1.
Solution. If we add the two equations we get
0=4
which is a contradiction. Therefore, there does not exist a list (s1 , s2 ) that satisfies the
system because this would lead to the contradiction 0 = 4.
Example 1.4. Let t be an arbitrary real number and let
s1 = − 32 − 2t
s2 = 23 + t
s3 = t.
Show that for any choice of the parameter t, the list (s1 , s2 , s3 ) is a solution to the linear
system
x1 + x2 + x3 = 0
x1 + 3x2 − x3 = 3.
Solution. Substitute the list (s1 , s2 , s3 ) into the left-hand-side of the first equation

− 32 − 2t + 32 + t + t = 0
and in the second equation

− 23 − 2t + 3( 23 + t) − t = − 23 + 9
2
=3
Both equations are satisfied for any value of t. Because we can vary t arbitrarily, we get an
infinite number of solutions parameterized by t. For example, compute the list (s1 , s2 , s3 )
for t = 3 and confirm that the resulting list is a solution to the linear system.
3
1.2 Matrices
We will use matrices to develop systematic methods to solve linear systems and to study
the properties of the solution set of a linear system. Informally speaking, a matrix is an
array or table consisting of rows and columns. For example,
 
1 −2 1 0
A= 0 2 −8 8 
−4 7 11 −5
is a matrix having m = 3 rows and n = 4 columns. In general, a matrix with m rows and
n columns is a m × n matrix and the set of all such matrices will be denoted by Mm×n .
Hence, A above is a 3 × 4 matrix. The entry of A in the ith row and jth column will be
denoted by aij . A matrix containing only one column is called a column vector and a
matrix containing only one row is called a row vector. For example, here is a row vector

u = 1 −3 4
and here is a column vector
3
v= .
−1
We can associate to a linear system three matrices: (1) the coefficient matrix, (2) the
output column vector, and (3) the augmented matrix. For example, for the linear system
5x1 − 3x2 + 8x3 = −1
x1 + 4x2 − 6x3 = 0
2x2 + 4x3 = 3
the coefficient matrix A, the output vector b, and the augmented matrix [A b] are:
     
5 −3 8 −1 5 −3 8 −1
A= 1 4 −6 , b =  0  , [A b] = 1 4 −6 0  .
0 2 4 3 0 2 4 3
If a linear system has m equations and n unknowns then the coefficient matrix A must be a
m × n matrix, that is, A has m rows and n columns. Using our previously defined notation,
we can write this as A ∈ Mm×n .
If we are given an augmented matrix, we can write down the associated linear system in
an obvious way. For example, the linear system associated to the augmented matrix
 
1 4 −2 8 12
0 1 −7 2 −4
0 0 5 −1 7
is
x1 + 4x2 − 2x3 + 8x4 = 12
x2 − 7x3 + 2x4 = −4
5x3 − x4 = 7.
4
Lecture 1
We can study matrices without interpreting them as coefficient matrices or augmented ma-
trices associated to a linear system. Matrix algebra is a fascinating subject with numerous
applications in every branch of engineering, medicine, statistics, mathematics, finance, biol-
ogy, chemistry, etc.
1.3 Solving linear systems

In algebra, you learned to solve equations by first “simplifying” them using operations that
do not alter the solution set. For example, to solve 2x = 8 − 2x we can add to both sides
2x and obtain 4x = 8 and then multiply both sides by 14 yielding x = 2. We can do
similar operations on a linear system. There are three basic operations, called elementary
operations, that can be performed:
1. Interchange two equations.
2. Multiply an equation by a nonzero constant.
3. Add a multiple of one equation to another.
These operations do not alter the solution set. The idea is to apply these operations itera-
tively to simplify the linear system to a point where one can easily write down the solution
set. It is convenient to apply elementary operations on the augmented matrix [A b] repre-
senting the linear system. In this case, we call the operations elementary row operations,
and the process of simplifying the linear system using these operations is called row reduc-
tion. The goal with row reducing is to transform the original linear system into one having
a triangular structure and then perform back substitution to solve the system. This is
best explained via an example.
Example 1.5. Use back substitution on the augmented matrix

 
1 0 −2 −4
0 1 −1 0 
0 0 1 1
to solve the associated linear system.
Solution. Notice that the augmented matrix has a triangular structure. The third row
corresponds to the equation x3 = 1. The second row corresponds to the equation
x2 − x3 = 0
and therefore x2 = x3 = 1. The first row corresponds to the equation
x1 − 2x3 = −4
and therefore
x1 = −4 + 2x3 = −4 + 2 = −2.
Therefore, the solution is (−2, 1, 1).
5
Example 1.6. Solve the linear system using elementary row operations.
−3x1 + 2x2 + 4x3 = 12

x1 − 2x3 = −4
2x1 − 3x2 + 4x3 = −3
Solution. Our goal is to perform elementary row operations to obtain a triangular structure
and then use back substitution to solve. The augmented matrix is
 
−3 2 4 12
1 0 −2 −4 .
2 −3 4 −3
Interchange Row 1 (R1 ) and Row 2 (R2 ):

   
−3 2 4 12 1 0 −2 −4
R ↔R2
1 0 −2 −4 −−1−−→ −3 2 4 12 
2 −3 4 −3 2 −3 4 −3
As you will see, this first operation will simplify the next step. Add 3R1 to R2 :
   
1 0 −2 −4 1 0 −2 −4
3R +R2
−3 2 4 12  −−1−−→ 0 2 −2 0 
2 −3 4 −3 2 −3 4 −3
Add −2R1 to R3 :
   
1 0 −2 −4 1 0 −2 −4
−2R1 +R3
0 2 −2 0  − −−−−→ 0 2 −2 0 
2 −3 4 −3 0 −3 8 5
Multiply R2 by 12 :
   
1 0 −2 −4 1
R2
1 0 −2 −4
0 2 −2 0  −2−→ 0 1 −1 0 
0 −3 8 5 0 −3 8 5
Add 3R2 to R3 :    
1 0 −2 −4 1 0 −2 −4
3R2 +R3
0 1 −1 0  − −−−→ 0 1 −1 0 
0 −3 8 5 0 0 5 5
Multiply R3 by 15 :
   
1 0 −2 −4 1
R3
1 0 −2 −4
0 1 −1 0  −5−→ 0 1 −1 0 
0 0 5 5 0 0 1 1
We can continue row reducing but the row reduced augmented matrix is in triangular form.
So now use back substitution to solve. The linear system associated to the row reduced
6
Lecture 1
augmented matrix is
x1 − 2x3 = −4
x2 − x3 = 0
x3 = 1
The last equation gives that x3 = 1. From the second equation we obtain that x2 − x3 = 0,
and thus x2 = 1. The first equation then gives that x1 = −4 + 2(1) = −2. Thus, the solution
to the original system is (−2, 1, 1). You should verify that (−2, 1, 1) is a solution to the
original system.
The original augmented matrix of the previous example is

 
−3 2 4 12 −3x1 + 2x2 + 4x3 = 12
M= 1 0 −2 −4 →
x1 − 2x3 = −4
2 −3 4 −3
2x1 − 3x2 + 4x3 = −3.
After row reducing we obtained the row reduced matrix
 
1 0 −2 −4 x1 − 2x3 = −4
N = 0 1 −1 0  →
x2 − x3 = 0
0 0 1 1
x3 = 1.
Although the two augmented matrices M and N are clearly distinct, it is a fact that they
have the same solution set.
Example 1.7. Using elementary row operations, show that the linear system is inconsistent.
x1 + 2x3 = 1
x2 + x3 = 0
2x1 + 4x3 = 1
Solution. The augmented matrix is
 
1 0 2 1
 0 1 1 0
2 0 4 1
Perform the operation −2R1 + R3 :
   
1 0 2 1 1 0 2 1
−2R1 +R3
 0 1 1 0 − −−−−→ 0 1 1 0 
2 0 4 1 0 0 0 −1
The last row of the simplified augmented matrix
 
1 0 2 1
0 1 1 0
0 0 0 −1
7
corresponds to the equation

0x1 + 0x2 + 0x3 = −1
Obviously, there are no numbers x1 , x2 , x3 that satisfy this equation, and therefore, the
linear system is inconsistent, i.e., it has no solution. In general, if we obtain a row in an
augmented matrix of the form

0 0 0 ··· 0 c
where c is a nonzero number, then the linear system is inconsistent. We will call this type
of row an inconsistent row. However, a row of the form

0 1 0 0 0
corresponds to the equation x2 = 0 which is perfectly valid.
1.4 Geometric interpretation of the solution set

The set of points (x1 , x2 ) that satisfy the linear system
x1 − 2x2 = −1
(1.2)
−x1 + 3x2 = 3
is the intersection of the two lines determined by the equations of the system. The solution
for this system is (3, 2). The two lines intersect at the point (x1 , x2 ) = (3, 2), see Figure 1.1.
Figure 1.1: The intersection point of the two lines is the solution of the linear system (1.2)
Similarly, the solution of the linear system

x1 − 2x2 + x3 = 0
2x2 − 8x3 = 8 (1.3)
−4x1 + 5x2 + 9x3 = −9
8
Lecture 1
is the intersection of the three planes determined by the equations of the system. In this case,
there is only one solution: (29, 16, 3). In the case of a consistent system of two equations,
the solution set is the line of intersection of the two planes determined by the equations of
the system, see Figure 1.2.
x1 − 2x2 + x3 = 0
−4x1 + 5x2 + 9x3 = −9
the solution set is this line
Figure 1.2: The intersection of the two planes is the solution set of the linear system (1.3)
After this lecture you should know the following:

• what a linear system is
• what it means for a linear system to be consistent and inconsistent
• what matrices are
• what are the matrices associated to a linear system
• what the elementary row operations are and how to apply them to simplify a linear
system
• what it means for two matrices to be row equivalent
• how to use the method of back substitution to solve a linear system
• what an inconsistent row is
• how to identify using elementary row operations when a linear system is inconsistent
• the geometric interpretation of the solution set of a linear system
9
10
Lecture 2
Lecture 2
Row Reduction and Echelon Forms
In this lecture, we will get more practice with row reduction and in the process introduce
two important types of matrix forms. We will also discuss when a linear system has a unique
solution, infinitely many solutions, or no solution. Lastly, we will introduce a convenient
parameter called the rank of a matrix.
2.1 Row echelon form (REF)

Consider the linear system
x1 + 5x2 − 2x4 − x5 + 7x6 = −4

2x2 − 2x3 + 3x6 = 0
−9x4 − x5 + x6 = −1
5x5 + x6 = 5
0=0
having augmented matrix

 
1 5 0 −2 −1 7 −4
0 2 −2 0 0 3 0
 
0 0 0 −9 −1 1 −1
 .
0 0 0 0 5 1 5
0 0 0 0 0 0 0
The above augmented matrix has the following properties:
P1. All nonzero rows are above any rows of all zeros.
P2. The leftmost nonzero entry of a row is to the right of the leftmost nonzero entry of
the row above it.
11
Any matrix satisfying properties P1 and P2 is said to be in row echelon form (REF). In
REF, the leftmost nonzero entry in a row is called a leading entry:
 
1 5 0 −2 −1 7 −4
 0 2 −2 0 0 3 0
 
 0 0 0 −9 −1 1 −1
 
0 0 0 0 5 1 5
0 0 0 0 0 0 0
A consequence of property P2 is that every entry below a leading entry is zero:
 
1 5 0 −2 −4 −1 −7
0 2 −2 0 0 3 0
 
0 0 0 −9 −1 1 −1
 
0 0 0 0 5 1 5
0 0 0 0 0 0 0
We can perform elementary row operations, or row reduction, to transform a matrix into
REF.
Example 2.1. Explain why the following matrices are not in REF. Use elementary row
operations to put them in REF.
   
3 −1 0 3 7 5 0 −3
M = 0 0 0 0 N = 0 3 −1 1 
0 1 3 0 0 6 −5 2
Solution. Matrix M fails property P1. To put M in REF we interchange R2 with R3 :
   
3 −1 0 3 3 −1 0 3
R ↔R3
M = 0 0 0 0 −−2−−→ 0 1 3 0
0 1 3 0 0 0 0 0
The matrix N fails property P2. To put N in REF we perform the operation −2R2 + R3 →
R3 :    
7 5 0 −3 7 5 0 −3
−2R2 +R3
0 3 −1 1  − −−−−→ 0 3 −1 1 
0 6 −5 2 0 0 −3 0
Why is REF useful? Certain properties of a matrix can be easily deduced if it is in REF.
For now, REF is useful to us for solving a linear system of equations. If an augmented matrix
is in REF, we can use back substitution to solve the system, just as we did in Lecture 1.
For example, consider the system
8x1 − 2x2 + x3 = 4
3x2 − x3 = 7
2x3 = 4
12
Lecture 2
whose augmented matrix is already in REF:

 
8 −2 1 4
0 3 −1 7
0 0 2 4
From the last equation we obtain that 2x3 = 4, and thus x3 = 2. Substituting x3 = 2 into
the second equation we obtain that x2 = 3. Substituting x3 = 2 and x2 = 3 into the first
equation we obtain that x1 = 1.
2.2 Reduced row echelon form (RREF)

Although REF simplifies the problem of solving a linear system, later on in the course we
will need to completely row reduce matrices into what is called reduced row echelon form
(RREF). A matrix is in RREF if it is in REF (so it satisfies properties P1 and P2) and in
addition satisfies the following properties:
P3. The leading entry in each nonzero row is a 1.
P4. All the entries above (and below) a leading 1 are all zero.
A leading 1 in the RREF of a matrix is called a pivot. For example, the following matrix
in RREF:  
1 6 0 3 0 0
0 0 1 −4 0 5
0 0 0 0 1 7
has three pivots:  
1 6 0 3 0 0
0 0 1 −4 0 5
0 0 0 0 1 7
Example 2.2. Use row reduction to transform the matrix into RREF.
 
0 3 −6 6 4 −5
 3 −7 8 −5 8 9 
3 −9 12 −9 6 15
Solution. The first step is to make the top leftmost entry nonzero:
   
0 3 −6 6 4 −5 3 −9 12 −9 6 15
R3 ↔R1
3 −7 8 −5 8 9  − −−−→ 3 −7 8 −5 8 9 
3 −9 12 −9 6 15 0 3 −6 6 4 −5
Now create a leading 1 in the first row:
   
3 −9 12 −9 6 15 1
R1
1 −3 4 −3 2 5
3 −7 8 −5 8 9  −3−→ 3 −7 8 −5 8 9 
0 3 −6 6 4 −5 0 3 −6 6 4 −5
13
Create zeros under the newly created leading 1:

   
1 −3 4 −3 2 5 1 −3 4 −3 2 5
−3R1 +R2
3 −7 8 −5 8 9  − −−−−→ 0 2 −4 4 2 −6
0 3 −6 6 4 −5 0 3 −6 6 4 −5
Create a leading 1 in the second row:

   
1 −3 4 −3 2 5 1
R2
1 −3 4 −3 2 5
0 2 −4 4 2 −6 −2−→ 0 1 −2 2 1 −3
0 3 −6 6 4 −5 0 3 −6 6 4 −5
Create zeros under the newly created leading 1:

   
1 −3 4 −3 2 5 1 −3 4 −3 2 5
−3R2 +R3
0 1 −2 2 1 −3 − −−−−→ 0 1 −2 2 1 −3
0 3 −6 6 4 −5 0 0 0 0 1 4
We have now completed the top-to-bottom phase of the row reduction algorithm. In the
next phase, we work bottom-to-top and create zeros above the leading 1’s. Create zeros
above the leading 1 in the third row:
   
1 −3 4 −3 2 5 1 −3 4 −3 2 5
−R3 +R2
0 1 −2 2 1 −3 − −−−−→ 0 1 −2 2 0 −7
0 0 0 0 1 4 0 0 0 0 1 4
   
1 −3 4 −3 2 5 1 −3 4 −3 0 −3
−2R3 +R1
0 1 −2 2 0 −7 − −−−−→ 0 1 −2 2 0 −7
0 0 0 0 1 4 0 0 0 0 1 4
Create zeros above the leading 1 in the second row:
   
1 −3 4 −3 0 −3 1 0 −2 3 0 −24
3R +R1
0 1 −2 2 0 −7 −−2−−→ 0 1 −2 2 0 −7 
0 0 0 0 1 4 0 0 0 0 1 4
This completes the row reduction algorithm and the matrix is in RREF.
Example 2.3. Use row reduction to solve the linear system.
2x1 + 4x2 + 6x3 = 8

x1 + 2x2 + 4x3 = 8
3x1 + 6x2 + 9x3 = 12

 
2 4 6 8
1 2 4 8 
3 6 9 12
14
Lecture 2
Create a leading 1 in the first row:

   
2 4 6 8 1
R1
1 2 3 4
1 2 4 8  −2−→ 1 2 4 8 
3 6 9 12 3 6 9 12
Create zeros under the first leading 1:
   
1 2 3 4 1 2 3 4
−R1 +R2
1 2 4 8 −−−−−→ 0
  0 1 4
3 6 9 12 3 6 9 12
   
1 2 3 4 1 2 3 4
−3R +R3
0 0 1 4  −−−1−−→ 0 0 1 4
3 6 9 12 0 0 0 0
The system is consistent, however, there are only 2 nonzero rows but 3 unknown variables.
This means that the solution set will contain 3 − 2 = 1 free parameter. The second row
in the augmented matrix is equivalent to the equation:
x3 = 4.
The first row is equivalent to the equation:
x1 + 2x2 + 3x3 = 4
and after substituting x3 = 4 we obtain
x1 + 2x2 = −8.
We now must choose one of the variables x1 or x2 to be a parameter, say t, and solve for the
remaining variable. If we set x2 = t then from x1 + 2x2 = −8 we obtain that
x1 = −8 − 2t.
We can therefore write the solution set for the linear system as
x1 = −8 − 2t
x2 = t (2.1)
x3 = 4
where t can be any real number. If we had chosen x1 to be the parameter, say x1 = t,
then the solution set can be written as
x1 = t
x2 = −4 − 12 t (2.2)
x3 = 4
Although (2.1) and (2.2) are two different parameterizations, they both give the same solution
set.
15
In general, if a linear system has n unknown variables and the row reduced augmented
matrix has r leading entries, then the number of free parameters d in the solution set is
d = n − r.
Thus, when performing back substitution, we will have to set d of the unknown variables
to arbitrary parameters. In the previous example, there are n = 3 unknown variables and
the row reduced augmented matrix contained r = 2 leading entries. The number of free
parameters was therefore
d = n − r = 3 − 2 = 1.
Because the number of leading entries r in the row reduced coefficient matrix determine the
number of free parameters, we will refer to r as the rank of the coefficient matrix:
r = rank(A).
Later in the course, we will give a more geometric interpretation to rank(A).
Example 2.4. Solve the linear system represented by the augmented matrix
 
1 −7 2 −5 8 10
0 1 −3 3 1 −5
0 0 0 1 −1 4
Solution. The number of unknowns is n = 5 and the augmented matrix has rank r = 3
(leading entries). Thus, the solution set is parameterized by d = 5 − 3 = 2 free variables,
call them t and s. The last equation of the augmented matrix is x4 − x5 = 4. We choose x5
to be the first parameter so we set x5 = t. Therefore, x4 = 4 + t. The second equation of
the augmented matrix is
x2 − 3x3 + 3x4 + x5 = −5
and the unassigned variables are x2 and x3 . We choose x3 to be the second parameter, say
x3 = s. Then
x2 = −5 + 3x3 − 3x4 − x5
= −5 + 3s − 3(4 + t) − t
= −17 − 4t + 3s.
We now use the first equation of the augmented matrix to write x1 in terms of the other
variables:
x1 = 10 + 7x2 − 2x3 + 5x4 − 8x5

= 10 + 7(−17 − 4t + 3s) − 2s + 5(4 + t) − 8t
= −89 − 31t + 19s
16
Lecture 2
Thus, the solution set is
x1 = −89 − 31t + 19s

x2 = −17 − 4t + 3s
x3 =s
x4 =4+t
x5 =t
where t and s are arbitrary real numbers.. Choose arbitrary numbers for t and s and
substitute the corresponding list (x1 , x2 , . . . , x5 ) into the system of equations to verify that
it is a solution.
2.3 Existence and uniqueness of solutions

The REF or RREF of an augmented matrix leads to three distinct possibilities for the
solution set of a linear system.
Theorem 2.5: Let [A b] be the augmented matrix of a linear system. One of the following
distinct possibilities will occur:
1. The augmented matrix will contain an inconsistent row.
2. All the rows of the augmented matrix are consistent and there are no free parameters.
3. All the rows of the augmented matrix are consistent and there are d ≥ 1 variables
that must be set to arbitrary parameters
In Case 1., the linear system is inconsistent and thus has no solution. In Case 2., the linear
system is consistent and has only one (and thus unique) solution. This case occurs when
r = rank(A) = n since then the number of free parameters is d = n − r = 0. In Case 3., the
linear system is consistent and has infinitely many solutions. This case occurs when r < n
and thus d = n − r > 0 is the number of free parameters.

• what the REF is and how to compute it
• what the RREF is and how to compute it
• how to solve linear systems using row reduction (Practice!!!)
• how to identify when a linear system is inconsistent
• how to identify when a linear system is consistent
• what is the rank of a matrix
• how to compute the number of free parameters in a solution set
• what are the three possible cases for the solution set of a linear system (Theorem 2.5)
17
18
Lecture 3
Lecture 3
Vector Equations
In this lecture, we introduce vectors and vector equations. Specifically, we introduce the
linear combination problem which simply asks whether it is possible to express one vector
in terms of other vectors; we will be more precise in what follows. As we will see, solving
the linear combination problem reduces to solving a linear system of equations.
3.1 Vectors in Rn
Recall that a column vector in Rn is a n × 1 matrix. From now on, we will drop the
“column” descriptor and simply use the word vectors. It is important to emphasize that a
vector in Rn is simply a list of n numbers; you are safe (and highly encouraged!) to forget
the idea that a vector is an object with an arrow. Here is a vector in R2 :

3
v= .
−1
Here is a vector in R3 :  
−3
v =  0 .
11
Here is a vector in R6 : 
9
0
 
−3
 6 .
v= 
 
0
3
To indicate that v is a vector in Rn , we will use the notation v ∈ Rn . The mathematical
symbol ∈ means “is an element of”. When we write vectors within a paragraph, we willwrite

−1
them using list notation instead of column notation, e.g., v = (−1, 4) instead of v = .
4
19
Vector Equations
We can add/subtract vectors, and multiply vectors by numbers or scalars. For example,
here is the addition of two vectors:
     
0 4 4
−5 −3 −8
  +   =  .
9 0 9
2 1 3
And the multiplication of a scalar with a vector:
   
1 3
3 −3 = −9 .
5 15
And here are both operations combined:
         
4 −2 −8 −6 −14
−2 −8 + 3  9  =  16  +  27  =  43  .
3 4 −6 12 6
These operations constitute “the algebra” of vectors. As the following example illustrates,
vectors can be used in a natural way to represent the solution of a linear system.
Example 3.1. Write the general solution in vector form of the linear system represented
by the augmented matrix
 
1 −7 2 −5 8 10
A b = 0 1 −3 3 1 −5
0 0 0 1 −1 4
Solution. The number of unknowns is n = 5 and the associated coefficient matrix A has
rank r = 3. Thus, the solution set is parametrized by d = n − r = 2 parameters. This
system was considered in Example 2.4 and the general solution was found to be
x1 = −89 − 31t1 + 19t2

x2 = −17 − 4t1 + 3t2
x3 = t2
x4 = 4 + t1
x5 = t1
where t1 and t2 are arbitrary real numbers. The solution in vector form therefore takes the
form          
x1 −89 − 31t1 + 19t2 −89 −31 19
x2   −17 − 4t1 + 3t2  −17  −4  3
         
x3  =  t2
x=    =  0  + t1  0  + t2  1 
      
x4   4 + t1   4   1  0
x5 t1 0 1 0
20
Lecture 3
A fundamental problem in linear algebra is solving vector equations for an unknown

vector. As an example, suppose that you are given the vectors
     
4 −2 −14
v1 = −8 , v2 =  9  , b =  43  ,
3 4 6
and asked to find numbers x1 and x2 such that x1 v1 + x2 v2 = b, that is,

     
4 −2 −14
x1 −8 + x2  9  =  43  .
3 4 6
Here the unknowns are the scalars x1 and x2 . After some guess and check, we find that
x1 = −2 and x2 = 3 is a solution to the problem since
     
4 −2 −14
−2 −8 + 3  9  =  43  .
3 4 6
In some sense, the vector b is a combination of the vectors v1 and v2 . This motivates the
following definition.
Definition 3.2: Let v1 , v2 , . . . , vp be vectors in Rn . A vector b is said to be a linear

combination of the vectors v1 , v2 , . . . , vp if there exists scalars x1 , x2 , . . . , xp such that
x1 v1 + x2 v2 + · · · + xp vp = b.
The scalars in a linear combination are called the coefficients of the linear combination.
As an example, given the vectors
       
1 −2 −1 −3
v1 = −2 , v2 =  4  , v3 =  5  , b =  0 
3 −6 6 −27
you can verify (and you should!) that
3v1 + 4v2 − 2v3 = b.
Therefore, we can say that b is a linear combination of v1 , v2 , v3 with coefficients x1 = 3,

x2 = 4, and x3 = −2.
3.2 The linear combination problem

The linear combination problem is the following:
21
Vector Equations
Problem: Given vectors v1 , . . . , vp and b, is b a linear combination of v1 , v2 , . . . , vp ?
For example, say you are given the vectors

     
1 1 2
v1 = 2 , v2 = 1 , v3 = 1
1 0 2
and also  
0
b =  1 .
−2
Does there exist scalars x1 , x2 , x3 such that
x1 v1 + x2 v2 + x3 v3 = b? (3.1)
For obvious reasons, equation (3.1) is called a vector equation and the unknowns are x1 ,
x2 , and x3 . To gain some intuition with the linear combination problem, let’s do an example
by inspection.
Example 3.3. Let v1 = (1, 0, 0), let v2 = (0, 0, 1), let b1 = (0, 2, 0), and let b2 = (−3, 0, 7).
Are b1 and b2 linear combinations of v1 , v2 ?
Solution. For any scalars x1 and x2
       
x1 0 x1 0
x1 v1 + x2 v2 = 0 + 0 = 0 6= 2
      
0 x2 x2 0
and thus no, b1 is not a linear combination of v1 , v2 , v3 . On the other hand, by inspection
we have that      
−3 0 −3
−3v1 + 7v2 = 0 + 0 = 0  = b2
    
0 7 7
and thus yes, b2 is a linear combination of v1 , v2 , v3 . These examples, of low dimension,
were more-or-less obvious. Going forward, we are going to need a systematic way to solve
the linear combination problem that does not rely on pure inspection.
We now describe how the linear combination problem is connected to the problem of
solving a system of linear equations. Consider again the vectors
       
1 1 2 0
v1 = 2 , v2 = 1 , v3 = 1 , b = 1  .
      
1 0 2 −2
Does there exist scalars x1 , x2 , x3 such that
x1 v1 + x2 v2 + x3 v3 = b? (3.2)
22
Lecture 3
First, let’s expand the left-hand side of equation (3.2):

       
x1 x2 2x3 x1 + x2 + 2x3
x1 v1 + x2 v2 + x3 v3 = 2x1  + x2  +  x3  = 2x1 + x2 + x3  .
x1 0 2x3 x1 + 2x3
We want equation (3.2) to hold so let’s equate the expansion x1 v1 + x2 v2 + x3 v3 with b. In

other words, set    
x1 + x2 + 2x3 0
2x1 + x2 + x3  =  1  .
x1 + 2x3 −2
Comparing component-by-component in the above relationship, we seek scalars x1 , x2 , x3
satisfying the equations
x1 + x2 + 2x3 = 0
2x1 + x2 + x3 = 1 (3.3)
x1 + 2x3 = −2.
This is just a linear system consisting of m = 3 equations and n = 3 unknowns! Thus, the
linear combination problem can be solved by solving a system of linear equations for the
unknown scalars x1 , x2 , x3 . We know how to do this. In this case, the augmented matrix of
the linear system (3.3) is  
1 1 2 0
[A b] = 2 1 1 1 
1 0 2 −2
Notice that the 1st column of A is just v1 , the second column is v2 , and the third column
is v3 , in other words, the augment matrix is

[A b] = v1 v2 v3 b
Applying the row reduction algorithm, the solution is
x1 = 0, x2 = 2, x3 = −1
and thus these coefficients solve the linear combination problem. In other words,
0v1 + 2v2 − v3 = b
In this case, there is only one solution to the linear system, so b can be written as a
linear combination of v1 , v2 , . . . , vp in only one (or unique) way. You should verify these
computations.
We summarize the previous discussion with the following:
The problem of determining if a given vector b is a linear combination of the vectors

v1 , v2 , . . . , vp is equivalent to solving the linear system of equations with augmented matrix

A b = v1 v2 · · · vp b .
23
Vector Equations
Applying the existence and uniqueness Theorem 2.5, the only three possibilities to the linear
combination problem are:
1. If the linear system is inconsistent then b is not a linear combination of v1 , v2 , . . . , vp ,

i.e., there does not exist scalars x1 , x2 , . . . , xp such that x1 v1 + x2 v2 + · · · + xp vp = b.
2. If the linear system is consistent and the solution is unique then b can be written as a
linear combination of v1 , v2 , . . . , vp in only one way.
3. If the the linear system is consistent and the solution set has free parameters, then b
can be written as a linear combination of v1 , v2 , . . . , vp in infinitely many ways.
Example 3.4. Is the vector b = (7, 4, −3) a linear combination of the vectors
   
1 2
v1 = −2 , v2 = 5?

−5 6
Solution. Form the augmented matrix:

 
1 2 7
v1 v2 b = −2 5 4 
−5 6 −3
The RREF of the augmented matrix is

 
1 0 3
 0 1 2
0 0 0
and therefore the solution is x1 = 3 and x2 = 2. Therefore, yes, b is a linear combination of

v1 , v2 :      
1 2 7
3v1 + 2v2 = 3 −2 + 2 5 = 4  = b
    
−5 6 −3
Notice that the solution set does not contain any free parameters because n = 2 (unknowns)
and r = 2 (rank) and so d = 0. Therefore, the above linear combination is the only way to
write b as a linear combination of v1 and v2 .
Example 3.5. Is the vector b = (1, 0, 1) a linear combination of the vectors

     
1 0 2
v1 = 0 ,
 v2 = 1 ,
 v3 = 1?

2 0 4
24
Lecture 3
Solution. The augmented matrix of the corresponding linear system is

 
1 0 2 1
0 1 1 0 .
2 0 4 1
After row reducing we obtain that
 
1 0 2 1
0 1 1 0  .
0 0 0 −1
The last row is inconsistent, and therefore the linear system does not have a solution. There-
fore, no, b is not a linear combination of v1 , v2 , v3 .
Example 3.6. Is the vector b = (8, 8, 12) a linear combination of the vectors
     
2 4 6
v1 = 1 , v2 = 2 , v3 = 4?
    
3 6 9
   
2 4 6 8 1 2 3 4
REF
1 2 4 8  −−→ 0 0 1 4 .
3 6 9 12 0 0 0 0
The system is consistent and therefore b is a linear combination of v1 , v2 , v3 . In this case,
the solution set contains d = 1 free parameters and therefore, it is possible to write b as a
linear combination of v1 , v2 , v3 in infinitely many ways. In terms of the parameter t, the
solution set is
x1 = −8 − 2t
x2 = t
x3 = 4
Choosing any t gives scalars that can be used to write b as a linear combination of v1 , v2 , v3 .
For example, choosing t = 1 we obtain x1 = −10, x2 = 1, and x3 = 4, and you can verify
that        
2 4 6 8
−10v1 + v2 + 4v3 = −10 1 + 2 + 4 4 =  8  = b
3 6 9 12
Or, choosing t = −2 we obtain x1 = −4, x2 = −2, and x3 = 4, and you can verify that
       
2 4 6 8
−4v1 − 2v2 + 4v3 = −4 1 − 2 2 + 4 4 = 8  = b
      
3 6 9 12
25
Vector Equations
We make a few important observations on linear combinations of vectors. Given vectors

v1 , v2 , . . . , vp , there are certain vectors b that can be written as a linear combination of
v1 , v2 , . . . , vp in an obvious way. The zero vector b = 0 can always be written as a linear
combination of v1 , v2 , . . . , vp :
0 = 0v1 + 0v2 + · · · + 0vp .
Each vi itself can be written as a linear combination of v1 , v2 , . . . , vp , for example,
v2 = 0v1 + (1)v2 + 0v3 + · · · + 0vp .
More generally, any scalar multiple of vi can be written as a linear combination of v1 , v2 , . . . , vp ,

for example,
xv2 = 0v1 + xv2 + 0v3 + · · · + 0vp .
By varying the coefficients x1 , x2 , . . . , xp , we see that there are infinitely many vectors b
that can be written as a linear combination of v1 , v2 , . . . , vp . The “space” of all the possible
linear combinations of v1 , v2 , . . . , vp has a name, which we introduce next.
3.3 The span of a set of vectors

Given a set of vectors {v1 , v2 , . . . , vp }, we have been considering the problem of whether
or not a given vector b is a linear combination of {v1 , v2 , . . . , vp }. We now take another
point of view and instead consider the idea of generating all vectors that are a linear
combination of {v1 , v2 , . . . , vp }. So how do we generate a vector that is guaranteed to be
a linear combination of {v1 , v2 , . . . , vp }? For example, if v1 = (2, 1, 3), v2 = (4, 2, 6) and
v3 = (6, 4, 9) then
       
2 4 6 8
−10v1 + v2 + 4v3 = −10 1 + 2 + 4 4 = 8  .
      
3 6 9 12
Thus, by construction, the vector b = (8, 8, 12) is a linear combination of {v1 , v2 , v3 }. This
discussion leads us to the following definition.
Definition 3.7: Let v1 , v2 , . . . , vp be vectors. The set of all vectors that are a linear
combination of v1 , v2 , . . . , vp is called the span of v1 , v2 , . . . , vp , and we denote it by
S = span{v1 , v2 , . . . , vp }.
By definition, the span of a set of vectors is a collection of vectors, or a set of vectors. If b is

a linear combination of v1 , v2 , . . . , vp then b is an element of the set span{v1 , v2 , . . . , vp },
and we write this as
b ∈ span{v1 , v2 , . . . , vp }.
26
Lecture 3
By definition, writing that b ∈ span{v1 , v2 , . . . , vp } implies that there exists scalars x1 , x2 , . . . , xp

such that
x1 v1 + x2 v2 + · · · + xp vp = b.
Even though span{v1 , v2 , . . . , vp } is an infinite set of vectors, it is not necessarily true that
it is the whole space Rn .
The set span{v1 , v2 , . . . , vp } is just a collection of infinitely many vectors but it has some
geometric structure. In R2 and R3 we can visualize span{v1 , v2 , . . . , vp }. In R2 , the span of
a single nonzero vector, say v ∈ R2 , is a line through the origin in the direction of v, see
Figure 3.1.
Figure 3.1: The span of a single non-zero vector in R2 .
In R2 , the span of two vectors v1 , v2 ∈ R2 that are not multiples of each other is all of
R2 . That is, span{v1 , v2 } = R2 . For example, with v1 = (1, 0) and v2 = (0, 1), it is true
that span{v1 , v2 } = R2 . In R3 , the span of two vectors v1 , v2 ∈ R3 that are not multiples
of each other is a plane through the origin containing v1 and v2 , see Figure 3.2. In R3 , the
3
span{v,w}
z 0
−1
−2
−3
−4
−4
−3
−2
−1 −4
−3
0 −2
y 1 −1
0
2 1 x
3 2
3
4
Figure 3.2: The span of two vectors, not multiples of each other, in R3 .
span of a single vector is a line through the origin, and the span of three vectors that do not
depend on each other (we will make this precise soon) is all of R3 .
Example 3.8. Is the vector b = (7, 4, −3) in the span of the vectors v1 = (1, −2, −5), v2 =
(2, 5, 6)? In other words, is b ∈ span{v1 , v2 }?
27
Vector Equations
Solution. By definition, b is in the span of v1 and v2 if there exists scalars x1 and x2 such
that
x1 v1 + x2 v2 = b,
that is, if b can be written as a linear combination of v1 and v2 . From our previous
discussion

on the linear combination problem, we must consider the augmented matrix v1 v2 b .
Using row reduction, the augmented matrix is consistent and there is only one solution (see
Example 3.4). Therefore, yes, b ∈ span{v1 , v2 } and the linear combination is unique.
Example 3.9. Is the vector b = (1, 0, 1) in the span of the vectors v1 = (1, 0, 2), v2 =
(0, 1, 0), v3 = (2, 1, 4)?
Solution. From Example 3.5, we have that

 
REF 1 0 2 1
v1 v2 v3 b −−→ 0 1 1 0 
0 0 0 −1
The last row is inconsistent and therefore b is not in span{v1 , v2 , v3 }.
Example 3.10. Is the vector b = (8, 8, 12) in the span of the vectors v1 = (2, 1, 3), v2 =
(4, 2, 6), v3 = (6, 4, 9)?
Solution. From Example 3.6, we have that

 
REF 1 2 3 4
v1 v2 v3 b −−→ 0 0 1 4 .
0 0 0 0
The system is consistent and therefore b ∈ span{v1 , v2 , v3 }. In this case, the solution set
contains d = 1 free parameters and therefore, it is possible to write b as a linear combination
of v1 , v2 , v3 in infinitely many ways.
Example 3.11. Answer the following with True or False, and explain your answer.
(a) The vector b = (1, 2, 3) is in the span of the set of vectors
     
 −1 2 4 
 3  , −7 , −5 .
 
0 0 0

(b) The solution set of the linear system whose augmented matrix is v1 v2 v3 b is the
same as the solution set of the vector equation
x1 v1 + x2 v2 + x3 v3 = b.
(c) Suppose that the augmented matrix v1 v2 v3 b has an inconsistent row. Then
either b can be written as a linear combination of v1 , v2 , v3 or b ∈ span{v1 , v2 , v3 }.
(d) The span of the vectors {v1 , v2 , v3 } (at least one of which is nonzero) contains only the
vectors v1 , v2 , v3 and the zero vector 0.
28
Lecture 3

• what a vector is
• what a linear combination of vectors is
• what the linear combination problem is
• the relationship between the linear combination problem and the problem of solving
linear systems of equations
• how to solve the linear combination problem
• what the span of a set of vectors is
• the relationship between what it means for a vector b to be in the span of v1 , v2 , . . . , vp
and the problem of writing b as a linear combination of v1 , v2 , . . . , vp
• the geometric interpretation of the span of a set of vectors
29
Vector Equations
30
Lecture 4
Lecture 4
The Matrix Equation Ax = b
In this lecture, we introduce the operation of matrix-vector multiplication and how it relates
to the linear combination problem.
4.1 Matrix-vector multiplication

We begin with the definition of matrix-vector multiplication.
Definition 4.1: Given a matrix A ∈ Mm×n and a vector x ∈ Rn ,

   
a11 a12 a13 · · · a1n x1
 a21 a22 a23 · · · a2n   x2 
   
A =  .. .. .. .. ..  , x =  ..  ,
 . . . . .  .
am1 am2 am3 · · · amn xn
we define the product of A and x as the vector Ax in Rm given by

    
a11 a12 a13 · · · a1n x1 a11 x1 + a12 x2 + · · · + a1n xn
 a21 a22 a23 · · · a2n 
  x2   a21 x1 + a22 x2 + · · · + a2n xn 
   

Ax =  .. .. .. .. =
..   ..   .. .
 . . . . .  .   . 
am1 am2 am3 · · · amn xn am1 x1 + am2 x2 + · · · + amn xn
| {z } | {z }
A x
For the product Ax to be well-defined, the number of columns of A must equal the number
of components of x. Another way of saying this is that the outer dimension of A must equal
the inner dimension of x:
(m × n) · (n × 1) → m × 1
Example 4.2. Compute Ax.
31
(a)
 
2
−4
A = 1 −1 3 0 , x=
−3

(b)
 
1
3 3 −2
A= , x= 0 
4 −4 −1
−1
(c)
 
−1 1 0  
4 −1
1 −2
 3 −3 3  ,
A=  x= 2 
−2
0 −2 −3
Solution. We compute:
(a)
 
2
−4
Ax = 1 −1 3 0  −3

8

= (1)(2) + (−1)(−4) + (3)(−3) + (0)(8) = −3
(b)
 
1
3 3 −2  
Ax = 0
4 −4 −1
−1

(3)(1) + (3)(0) + (−2)(−1)
=
(4)(1) + (−4)(0) + (−1)(−1)

5
=
5
32
Lecture 4
(c)
 
−1 1 0  
4 −1
1 −2
Ax = 
   2
3 −3 3 
−2
0 −2 −3
 
(−1)(−1) + (1)(2) + (0)(−2)
 (4)(−1) + (1)(2) + (−2)(−2) 
= 
 (3)(−1) + (−3)(2) + (3)(−2) 
(0)(−1) + (−2)(2) + (−3)(−2)
 
3
 2 
=−15

We now list two important properties of matrix-vector multiplication.
Theorem 4.3: Let A be an m × n a matrix.

(a) For any vectors u, v in Rn it holds that
A(u + v) = Au + Av.
(b) For any vector u and scalar c it holds that
A(cu) = c(Au).
Example 4.4. For the given data, verify that the properties of Theorem 4.3 hold:

3 −3 −1 2
A= , u= , v= , c = −2.
2 1 3 −1
4.2 Matrix-vector multiplication and linear combina-

tions
Recall the general definition of matrix-vector multiplication Ax is
    
a11 a12 a13 · · · a1n x1 a11 x1 + a12 x2 + · · · + a1n xn
 a21 a22 a23 · · · a2n   x2   a21 x1 + a22 x2 + · · · + a2n xn
    
 
 .. .. .. .. ..   ..  =  ..  (4.1)
 . . . . .  .   . 
am1 am2 am3 · · · amn xn am1 x1 + am2 x2 + · · · + amn xn
33
There is an important way to decompose matrix-vector multiplication involving a linear

combination. To see how, let v1 , v2 , . . . , vn denote the columns of A and consider the
following linear combination:
     
x1 a11 x2 a12 xn a1n
 x1 a21   x2 a22   xn a2n 
     
x1 v1 + x2 v2 + · · · + xn vn =  ..  +  ..  + · · · +  .. 
 .   .   . 
x1 am1 x2 am2 xn amn
 
x1 a11 + x2 a12 + · · · + xn a1n

 x1 a21 + x2 a22 + · · · + xn a2n


= .. . (4.2)
 . 
x1 am1 + x2 am2 + · · · + xn amn

We observe that expressions (4.1) and (4.2) are equal! Therefore, if A = v1 v2 · · · vn
and x = (x1 , x2 , . . . , xn ) then
Ax = x1 v1 + x2 v2 + · · · + xn vn .
In summary, the vector Ax is a linear combination of the columns of A where the scalar
in the linear combination are the components of x! This (important) observation gives an
alternative way to compute Ax.
Example 4.5. Given  

−1 1 0  
4 −1
1 −2
 3 −3 3  ,
A=  x =  2 ,
−2
0 −2 −3
compute Ax in two ways: (1) using the original Definition 4.1, and (2) as a linear combination
of the columns of A.
4.3 The matrix equation problem

As we have seen, with a matrix A and any vector x, we can produce a new output vector
via the multiplication Ax. If A is a m × n matrix then we must have x ∈ Rn and the output
vector Ax is in Rm . We now introduce the following problem:
Problem: Given a matrix A ∈ Mm×n and a vector b ∈ Rm , find, if possible, a vector

x ∈ Rn such that
Ax = b. (⋆)
Equation (⋆) is a matrix equation where the unknown variable is x. If u is a vector such
that Au = b, then we say that u is a solution to the equation Ax = b. For example,
34
Lecture 4
suppose that
1 0 −3
A= , b= .
1 0 7

x
Does the equation Ax = b have a solution? Well, for any x = 1 we have that
x2

1 0 x1 x
Ax = = 1
1 0 x2 x1
and thus any output vector Ax has equal entries. Since b does not have equal entries then
the equation Ax = b has no solution.
We now describe a systematic way to solve matrix equations. As we have seen, the vector
Ax is a linear combination of the columns of A with the coefficients given by the components
of x. Therefore, the matrix equation problem is equivalent to the linear combination problem.
In Lecture 2, we showed that the linear combination problem can be solved by solving
a
system of linear equations. Putting all this together then, if A = v1 v2 · · · vn and
b ∈ Rm then:
To find a vector x ∈ Rn that solves the matrix equation
Ax = b
we solve the linear system whose augmented matrix is

A b = v1 v2 · · · vn b .
From now on, a system of linear equations such as
a11 x1 + a12 x2 + a13 x3 + · · · + a1n xn = b1

a21 x1 + a22 x2 + a23 x3 + · · · + a2n xn = b2
a31 x1 + a32 x2 + a33 x3 + · · · + a3n xn = b3
.. .. .. .. ..
. . . . .
am1 x1 + am2 x2 + am3 x3 + · · · + amn xn = bm
will be written in the compact form

Ax = b
where A is the coefficient matrix of the linear system, b is the output vector, and x is the
unknown vector to be solved for. We summarize our findings with the following theorem.
Theorem 4.6: Let A ∈ Mm×n and b ∈ Rm . The following statements are equivalent:
(a) The equation Ax = b has a solution.
(b) The vector b is a linear combination of the columns of A.

(c) The linear system represented by the augmented matrix A b is consistent.
35
Example 4.7. Solve, if possible, the matrix equation Ax = b if

   
1 3 −4 −2
A= 1
 5 2 , b = 4 .
 
−3 −7 −6 12
Solution. First form the augmented matrix:

 
1 3 −4 −2
[A b] =  1 5 2 4
−3 −7 −6 12
Performing the row reduction algorithm we obtain that

   
1 3 −4 −2 1 3 −4 −2
1 5 2 4  ∼ 0 1 3 3 .
−3 −7 −6 12 0 0 −12 0
Here r = rank(A) = 3 and therefore d = 0, i.e., no free parameters. Peforming back

substitution we obtain that x1 = −11, x2 = 3, and x3 = 0. Thus, the solution to the matrix
equation is unique (no free parameters) and is given by

−11
x= 3 
0
Let’s verify that Ax = b:

      
1 3 −4 −11 −11 + 9 + 0 −2
Ax =  1 5 2   3  = −11 + 15 + 0 =  4  = b
−3 −7 −6 0 33 − 21 12
In other words, b is a linear combination of the columns of A:

       
1 3 −4 −2
−11 1 + 3 5 + 0 2 = 4 
      
−3 7 −6 12
36
Lecture 4

1 2 3
A= , b= .
2 4 −4

Solution. Row reducing the augmented matrix A b we get

1 2 3 −2R1 +R2 1 2 3
−−−−−→ .
2 4 −4 0 0 −10
The last row is inconsistent and therefore there is no solution to the matrix equation Ax = b.
In other words, b is not a linear combination of the columns of A.

1 −1 2 2
A= , b= .
0 3 6 −1
Solution. First note that the unknown vector x is in R3 because A has n = 3 columns. The
linear system Ax = b has m = 2 equations and n = 3 unknowns. The coefficient matrix A
has rank r = 2, and therefore
the solution set will contain d = n − r = 1 parameter. The
augmented matrix A b is

1 −1 2 2
A b = .
0 3 6 −1
Let x3 = t be the parameter and use the last row to solve for x2 :
x2 = − 13 − 2t
Now use the first row to solve for x1 :
x1 = 2 + x2 − 2x3 = 2 + (− 31 − 2t) − 2t = 5
3
− 4t.
Thus, the solution set to the linear system is
x1 = 53 − 4t
x2 = − 13 − 2t
x3 = t
where t is an arbitrary number. Therefore, the matrix equation Ax = b has an infinite

number of solutions and they can all be written as
 5 
3
− 4t
x = − 31 − 2t
t
37
where t is an arbitrary number. Equivalently, b can be written as a linear combination of

the columns of A in infinitely many ways. For example, choosing t = −1 gives the particular
solution  
17/3
x = −7/3
−1
and you can verify that  
17/3
A −7/3 = b.
−1
Recall from Definition 3.7 that the span of a set of vectors v1 , v2 , . . . , vp , which we denoted
by span{v1 , v2 , . . . , vp }, is the space of vectors that can be written as a linear combination
of the vectors v1 , v2 , . . . , vp .
Example 4.10. Is the vector b in the span of the vectors v1 , v2 ?
     
0 3 −5
b = 4 , v1 = −2 , v2 = 6 
    
4 1 1
Solution. The vector b is in span{v1 , v2 } if we can find scalars x1 , x2 such that
x1 v1 + x2 v2 = b.
If we let A ∈ R3×2 be the matrix

 
3 −5
A = [v1 v2 ] = −2 6 
1 1

x1
then we need to solve the matrix equation Ax = b. Note that here x = ∈ R2 .
x2
Performing row reduction on the augmented matrix [A b] we get that
   
3 −5 0 1 0 2.5
−2 6 4 ∼ 0 1 1.5
1 1 4 0 0 0
Therefore, the linear system is consistent and has solution

2.5
x=
1.5
Therefore, b is in span{v1 , v2 }, and b can be written in terms of v1 and v2 as
2.5v1 + 1.5v2 = b
38
Lecture 4
If v1 , v2 , . . . , vp are vectors in Rn and it happens to be true that span{v1 , v2 , . . . , vp } = Rn

then we would say that the set of vectors {v1 , v2 , . . . , vp } spans all of Rn . From Theorem 4.6,
we have the following.
Theorem 4.11: Let A ∈ Mm×n be a matrix with columns v1 , v2 , . . . , vn , that is, A =

v1 v2 · · · vn . The following are equivalent:
(a) span{v1 , v2 , . . . , vn } = Rm
(b) Every b ∈ Rm can be written as a linear combination of v1 , v2 , . . . , vn .
(c) The matrix equation Ax = b has a solution for any b ∈ Rm .
(d) The rank of A is m.
Example 4.12. Do the vectors v1 , v2 , v3 span R3 ?

     
1 2 −1
v1 = −3 , v2 = −4 , v3 = 2 
    
5 2 3

Solution. From Theorem 4.11, the vectors v1 , v2 , v3 span R3 if the matrix A = v1 v2 v3
has rank r = 3 (leading entries in its REF/RREF). The RREF of A is
   
1 2 −1 1 0 0
−3 −4 2  ∼ 0 1 0
5 2 3 0 0 1
which does indeed have r = 3 leading entries. Therefore, regardless of the choice of b ∈ R3 ,
the augmented matrix [A b] will be consistent. Therefore, the vectors v1 , v2 , v3 span R3 :
span{v1 , v2 , v3 } = R3 .
In other words, every vector b ∈ R3 can be written as a linear combination of v1 , v2 , v3 .

• how to multiply a matrix A with a vector x
• that the product Ax is a linear combination of the columns of A
• how to solve the matrix equation Ax = b if A and b are known
• how to determine if a set of vectors {v1 , v2 , . . . , vp } in Rm spans all of Rm
• the relationship between the equation Ax = b, when b can be written as a linear
combination of the columns of A, and when the augmented matrix A b is consistent
(Theorem 4.6)
• when the columns of a matrix A ∈ Mm×n span all of Rm (Theorem 4.11)
• the basic properties of matrix-vector multiplication Theorem 4.3
39
40
Lecture 5
Lecture 5
Homogeneous and Nonhomogeneous

Systems
5.1 Homogeneous linear systems

We begin with a definition.
Definition 5.1: A linear system of the form Ax = 0 is called a homogeneous linear

system.
A homogeneous system Ax = 0 always has at least one solution, namely, the zero solution
because A0 = 0. A homogeneous system is therefore always consistent. The zero solution
x = 0 is called the trivial solution and any non-zero solution is called a nontrivial
solution. From the existence and uniqueness theorem (Theorem 2.5), we know that a
consistent linear system will have either one solution or infinitely many solutions. Therefore,
a homogeneous linear system has nontrivial solutions if and only if its solution set has at
least one parameter.
Recall that the number of parameters in the solution set is d = n − r, where r is the rank
of the coefficient matrix A and n is the number of unknowns.
Example 5.2. Does the linear homogeneous system have any nontrivial solutions?
3x1 + x2 − 9x3 = 0
x1 + x2 − 5x3 = 0
2x1 + x2 − 7x3 = 0
Solution. The linear system will have a nontrivial solution if the solution set has at least one
free parameter. Form the augmented matrix:
 
3 1 −9 0
1 1 −5 0
2 1 −7 0
41
Homogeneous and Nonhomogeneous Systems
The RREF is:    

3 1 −9 0 1 0 −2 0
1 1 −5 0 ∼ 0 1 −3 0
2 1 −7 0 0 0 0 0
The system is consistent. The rank of the coefficient matrix is r = 2 and thus there will be
d = 3 − 2 = 1 free parameter in the solution set. If we let x3 be the free parameter, say
x3 = t, then from the row equivalent augmented matrix
 
1 0 −2 0
0 1 −3 0
0 0 0 0
we obtain that x2 = 3x3 = 3t and x1 = 2x3 = 2t. Therefore, the general solution of the
linear system is
x1 = 2t
x2 = 3t
x3 = t
The general solution can be written in vector notation as

 
2
x = 3 t

1
 
2
Or more compactly if we let v = 3 then x = vt. Hence, any solution x to the linear
1
 
2
system can be written as a linear combination of the vector v = 3. In other words, the
1
solution set of the linear system is the span of the vector v:
span{v}.
Notice that in the previous example, when solving a homogeneous

system Ax = 0 using
row reduction, the last column of the augmented matrix A 0 remains unchanged (always
0) after every elementary row operation. Hence, to solve a homogeneous system, we can row
reduce the coefficient matrix A only and then set all rows equal to zero when performing
back substitution.
Example 5.3. Find the general solution of the homogenous system Ax = 0 where
 
1 2 2 1 4
A = 3 7 7 3 13 .
2 5 5 2 9
42
Lecture 5
Solution. After row reducing we obtain

   
1 2 2 1 4 1 0 0 1 2
A = 3 7 7 3 13 ∼ 0 1 1 0 1
2 5 5 2 9 0 0 0 0 0
Here n = 5, and r = 2, and therefore the number of parameters in the solution set is
d = n − r = 3. The second row of rref(A) gives the equation
x2 + x3 + x5 = 0.
Setting x5 = t1 and x3 = t2 as free parameters we obtain that
x2 = −x3 − x5 = −t2 − t1 .
From the first row we obtain the equation
x1 + x4 + 2x5 = 0
The unknown x5 has already been assigned, so we must now choose either x1 or x4 to be a
parameter. Choosing x4 = t3 we obtain that
x1 = −x4 − 2x5 = −t3 − 2t1
In summary, the general solution can be written as
       
−t3 − 2t1 −2 0 −1
 −t2 − t1  −1 −1 0
       
x=  t2  = t1  0  +t2  1  +t3  0  = t1 v1 + t2 v2 + t3 v3
      
 t3  0 0 1
t1 1 0 0
| {z } | {z } | {z }
v1 v2 v3
where t1 , t2 , t3 are arbitrary parameters. In other words, any solution x is in the span of
v1 , v2 , v3 :
x ∈ span{v1 , v2 , v3 }.
The form of the general solution in Example 5.3 holds in general and is summarized in
the following theorem.
Theorem 5.4: Consider the homogenous linear system Ax = 0, where A ∈ Mm×n and
0 ∈ Rm . Let r be the rank of A.
1. If r = n then the only solution to the system is the trivial solution x = 0.
2. Otherwise, if r < n and we set d = n − r, then there exist vectors v1 , v2 , . . . , vd such

that any solution x of the linear system can be written as
x = t1 v1 + t2 v2 + · · · + tp vd .
43
In other words, any solution x is in the span of v1 , v2 , . . . , vd :
x ∈ span{v1 , v2 , . . . , vd }.
A solution x to a homogeneous system written in the form
x = t1 v1 + t2 v2 + · · · + tp vd
is said to be in parametric vector form.
5.2 Nonhomogeneous systems

As we have seen, a homogeneous system Ax = 0 is always consistent. However, if b is non-
zero, then the nonhomogeneous linear system Ax = b may or may not have a solution. A
natural question arises: What is the relationship between the solution set of the homogeneous
system Ax = 0 and that of the nonhomogeneous system Ax = b when it is consistent? To
answer this question, suppose that p is a solution to the nonhomogeneous system Ax = b,
that is, Ap = b. And suppose that v is a solution to the homogeneous system Ax = 0, that
is, Av = 0. Now let q = p + v. Then
Aq = A(p + v)
= Ap + Av
=b+0
= b.
Therefore, Aq = b. In other words, q = p + v is also a solution of Ax = b. We have

therefore proved the following theorem.
Theorem 5.5: Suppose that the linear system Ax = b is consistent and let p be a
solution. Then any other solution q of the system Ax = b can be written in the form
q = p + v, for some vector v that is a solution to the homogeneous system Ax = 0.
Another way of stating Theorem 5.5 is the following: If the linear system Ax = b is consistent
and has solutions p and q, then the vector v = q−p is a solution to the homogeneous system
Ax = 0. The proof is a simple computation:
Av = A(q − p) = Aq − Ap = b − b = 0.
More generally, any solution of Ax = b can be written in the form
q = p + t1 v1 + t2 v2 + · · · + tp vd
where p is one particular solution of Ax = b and the vectors v1 , v2 , . . . , vd span the solution
set of the homogeneous system Ax = 0.
44
Lecture 5
There is a useful geometric interpretation of the solution set of a general linear system.
We saw in Lecture 3 that we can interpret the span of a set of vectors as a plane containing
the zero vector 0. Now, the general solution of Ax = b can be written as
x = p + t1 v1 + t2 v2 + · · · + tp vd .
Therefore, the solution set of Ax = b is a shift of the span{v1 , v2 , . . . , vd } by the vector p.
This is illustrated in Figure 5.1.
p + span{v}
p + tv
b
p span{v}
b
b
tv
b
v
b
Figure 5.1: The solution sets of a homogeneous and nonhomogeneous system.
Example 5.6. Write the general solution, in parametric vector form, of the linear system
3x1 + x2 − 9x3 = 2
x1 + x2 − 5x3 = 0
2x1 + x2 − 7x3 = 1.
Solution. The RREF of the augmented matrix is:
   
3 1 −9 2 1 0 −2 1
1 1 −5 0 ∼ 0 1 −3 −1
2 1 −7 1 0 0 0 0
The system is consistent and the rank of the coefficient matrix is r = 2. Therefore, there
are d = 3 − 2 = 1 parameters in the solution set. Letting x3 = t be the parameter, from the
second row of the RREF we have
x2 = 3t − 1
And from the first row of the RREF we have
x1 = 2t + 1
Therefore, the general solution of the system in parametric vector form is
     
2t + 1 1 2
x = 3t − 1 = −1 +t 3
    
t 0 1
| {z } |{z}
p v
45
You should check that p = (1, −1, 0) solves the linear system Ax = b, and that v = (2, 3, 1)
solves the homogeneous system Ax = 0.
Example 5.7. Write the general solution, in parametric vector form, of the linear system
represented by the augmented matrix
 
3 −3 6 3
−1 1 −2 −1 .
2 −2 4 2
Solution. Write the general solution, in parametric vector form, of the linear system repre-
sented by the augmented matrix
 
3 −3 6 3
−1 1 −2 −1
2 −2 4 2
The RREF of the augmented matrix is

   
3 −3 6 3 1 −1 2 1
−1 1 −2 −1 ∼ 0 0 0 0
2 −2 4 2 0 0 0 0
Here n = 3, r = 1 and therefore the solution set will have d = 2 parameters. Let x3 = t1
and x2 = t2 . Then from the first row we obtain
x1 = 1 + x2 − 2x3 = 1 + t2 − 2t1
The general solution in parametric vector form is therefore

     
1 −2 1
x = 0 +t1  0  +t2 1
0 1 0
|{z} | {z } |{z}
p v1 v2
You should verify that p is a solution to the linear system Ax = b:
Ap = b
And that v1 and v2 are solutions to the homogeneous linear system Ax = 0:
Av1 = Av2 = 0
46
Lecture 5
5.3 Summary
The material in this lecture is so important that we will summarize the main results. The
solution set of a linear system Ax = b can be written in the form
x = p + t1 v1 + t2 v2 + · · · + td vd
where Ap = b and where each of the vectors v1 , v2 , . . . , vd satisfies Avi = 0. Loosely

speaking,
{Solution set of Ax = b} = p + {Solution set of Ax = 0}
or
{Solution set of Ax = b} = p + span{v1 , v2 , . . . , vd }
where p satisfies Ap = b and Avi = 0.

• what a homogeneous/nonhomogenous linear system is
• when a homogeneous linear system has nontrivial solutions
• how to write the general solution set of a homogeneous system in parametric vector
form Theorem 5.4)
• how to write the solution set of a nonhomogeneous system in parametric vector form
Theorem 5.5)
• the relationship between the solution sets of the nonhomogeneous equation Ax = b
and the homogeneous equation Ax = 0
47
48
Lecture 6
Lecture 6
Linear Independence
6.1 Linear independence

In Lecture 3, we defined the span of a set of vectors {v1 , v2 , . . . , vn } as the collection of all
possible linear combinations
t1 v1 + t2 v2 + · · · + tn vn
and we denoted this set as span{v1 , v2 , . . . , vn }. Thus, if x ∈ span{v1 , v2 , . . . , vn } then by
definition there exists scalars t1 , t2 , . . . , tn such that
x = t1 v1 + t2 v2 + · · · + tn vn .
A natural question that arises is whether or not there are multiple ways to express x as a
linear combination of the vectors v1 , v2 , . . . , vn . For example, if v1 = (1, 2), v2 = (0, 1),
v3 = (−1, −1), and x = (3, −1) then you can verify that x ∈ span{v1 , v2 , v3 } and x can be
written in infinitely many ways using v1 , v2 , v3 . Here are three ways:
x = 3v1 − 7v2 + 0v3

x = −4v1 + 0v2 − 7v3
x = 0v1 − 4v2 − 3v3 .
The fact that x can be written in more than one way in terms of v1 , v2 , v3 suggests that there
might be a redundancy in the set {v1 , v2 , v3 }. In fact, it is not hard to see that v3 = −v1 +v2 ,
and thus v3 ∈ span{v1 , v2 }. The preceding discussion motivates the following definition.
Definition 6.1: A set of vectors {v1 , v2 , . . . , vn } is said to be linearly dependent if

some vj can be written as a linear combination of the other vectors, that is, if
vj ∈ span{v1 , . . . , vj−1 , vj+1, . . . , vn }.
If {v1 , v2 , . . . , vn } is not linearly dependent then we say that {v1 , v2 , . . . , vn } is linearly

independent.
49
Linear Independence
Example 6.2. Consider the vectors

     
1 4 2
v1 = 2 , v2 = 5 , v3 = 1 .
    
3 6 0
Show that they are linearly dependent.

Solution. By inspection, we have
     
2 2 4
2v1 + v3 = 4 + 1 = 5 = v2
    
6 0 6
Thus, v2 ∈ span{v1 , v3 } and therefore {v1 , v2 , v3 } is linearly dependent.

Notice that in the previous example, the equation 2v1 + v3 = v2 is equivalent to
2v1 − v2 + v3 = 0.
Hence, because {v1 , v2 v3 } is a linearly dependent set, it is possible to write the zero vector
0 as a linear combination of {v1 , v2 v3 } where not all the coefficients in the linear
combination are zero. This leads to the following characterization of linear independence.
Theorem 6.3: The set of vectors {v1 , v2 , . . . , vn } is linearly independent if and only if 0
can be written in only one way as a linear combination of {v1 , v2 , . . . , vn }. In other words,
if
t1 v1 + t2 v2 + · · · + tn vn = 0
then necessarily the coefficients t1 , t2 , . . . , tn are all zero.
Proof. If {v1 , v2 , . . . , vn } is linearly independent then every vector x ∈ span{v1 , v2 , . . . , vn }

can be written uniquely as a linear combination of {v1 , v2 , . . . , vn }, and this applies to the
particular case of the zero vector x = 0.
Now assume that 0 can be written uniquely as a linear combination of {v1 , v2 , . . . , vn }.
In other words, assume that if
t1 v1 + t2 v2 + · · · + tn vn = 0
then t1 = t2 = · · · = tn = 0. Now take any x ∈ span{v1 , v2 , . . . , vn } and suppose that there

are two ways to write x in terms of {v1 , v2 , . . . , vn }:
r1 v1 + r2 v2 + · · · + rn vn = x
s1 v1 + s2 v2 + · · · + sn vn = x.
Subtracting the second equation from the first we obtain that
(r1 − s1 )v1 + (r2 − s2 )v2 + · · · + (rn − sn )vn = x − x = 0.
50
Lecture 6
The above equation is a linear combination of v1 , v2 , . . . , vn resulting in the zero vector 0.

But we are assuming that the only way to write 0 in terms of {v1 , v2 , . . . , vn } is if all the
coefficients are zero. Therefore, we must have r1 − s1 = 0, r2 − s2 = 0, . . . , rn − sn = 0, or
equivalently that r1 = s1 , r2 = s2 , . . . , rn = sn . Therefore, the linear combinations
r1 v1 + r2 v2 + · · · + rn vn = x
s1 v1 + s2 v2 + · · · + sn vn = x
are actually the same. Therefore, each x ∈ span{v1 , v2 , . . . , vn } can be written uniquely in
terms of {v1 , v2 , . . . , vn }, and thus {v1 , v2 , . . . , vn } is a linearly independent set.
Because of Theorem 6.3, an alternative definition of linear independence of a set of vectors
{v1 , v2 , . . . , vn } is that the vector equation
x1 v1 + x2 v2 + · · · + xn vn = 0
has only the trivial solution, i.e., the solution x1 = x2 = · · · = xn = 0. Thus, if {v1 , v2 , . . . , vn }
is linearly dependent, then there exist scalars x1 , x2 , . . . , xn not all zero such that
x1 v1 + x2 v2 + · · · + xn vn = 0.
Hence, if we suppose for instance that xn 6= 0 then we can write vn in terms of the vectors
v1 , . . . , vn−1 as follows:
xn−1
vn = − xxn1 v1 − x2
v
xn 2
−···− xn
vn−1 .
In other words, vn ∈ span{v1 , v2 , . . . , vn−1 }.
According to Theorem 6.3, the set of vectors {v1 , v2 , . . . , vn } is linearly independent if
the equation
x1 v1 + x2 v2 + · · · + xn vn = 0 (6.1)
has only the trivial solution. Now, the vector equation (6.1) is a homogeneous linear system
of equations with coefficient matrix

A = v1 v2 · · · vn .
Therefore, the set {v1 , v2 , . . . , vn } is linearly independent if and only if the the homogeneous
system Ax = 0 has only the trivial solution. But the homogeneous system Ax = 0 has only
the trivial solution if there are no free parameters in its solution set. We therefore have the
following.
Theorem 6.4: The set {v1 , v2 , . . . , vn } is linearly independent if and only if the the rank
of A is r = n, that is, if the number of leading entries r in the REF (or RREF) of A is
exactly n.
Example 6.5. Are the vectors below linearly independent?

     
0 1 4
v1 = 1 , v2 = 2 , v3 = −1
5 8 0
51
Linear Independence
Solution. Let A be the matrix

 
0 1 4
A = v1 v2 v3 = 1 2 −1
5 8 0
Performing elementary row operations we obtain
 
1 2 −1
A ∼ 0 1 4 
0 0 13
Clearly, r = rank(A) = 3, which is equal to the number of vectors n = 3. Therefore,
{v1 , v2 , v3 } is linearly independent.
     
1 4 2
v1 = 2 , v2 = 5 , v3 = 1
3 6 0
Solution. Let A be the matrix
 
1 4 2
A = v1 v2 v3 = 2 5 1
3 6 0
Performing elementary row operations we obtain
 
1 4 2
A ∼ 0 −3 −3
0 0 0
Clearly, r = rank(A) = 2, which is not equal to the number of vectors, n = 3. Therefore,
{v1 , v2 , v3 } is linearly dependent. We will find a nontrivial linear combination of the vectors
v1 , v2 , v3 that gives the zero vector 0. The REF of A = [v1 v2 v3 ] is
 
1 4 2
A ∼ 0 −3 −3
0 0 0
Since r = 2, the solution set of the linear system Ax = 0 has d = n − r = 1 free parameter.
Using back substitution on the REF above, we find that the general solution of Ax = 0
written in parametric form is  
2
x = t −1
1
The vector  
2
v = −1
1
52
Lecture 6
spans the solution set of the system Ax = 0. Choosing for instance t = 2 we obtain the
solution    
2 4
x = t −1 = −2 .
  
1 2
Therefore,
4v1 − 2v2 + 2v3 = 0
is a non-trivial linear combination of v1 , v2 , v3 that gives the zero vector 0. And, for instance,
v3 = −2v1 + v2
that is, v3 ∈ span{v1 , v2 }.
Below we record some simple observations on the linear independence of simple sets:
• A set consisting of a single non-zero vector {v1 } is linearly independent. Indeed, if v1

is non-zero then
tv1 = 0
is true if and only if t = 0.
• A set consisting of two non-zero vectors {v1 , v2 } is linearly independent if and only if
neither of the vectors is a multiple of the other. For example, if v2 = tv1 then
tv1 − v2 = 0
is a non-trivial linear combination of v1 , v2 giving the zero vector 0.
• Any set {v1 , v2 , . . . , vp } containing the zero vector, say that vp = 0, is linearly depen-
dent. For example, the linear combination
0v1 + 0v2 + · · · + 0vp−1 + 2vp = 0
is a non-trivial linear combination giving the zero vector 0.
6.2 The maximum size of a linearly independent set

The next theorem puts a constraint on the maximum size of a linearly independent set in
Rn .
Theorem 6.7: Let {v1 , v2 , . . . , vp } be a set of vectors in Rn . If p > n then v1 , v2 , . . . , vp

are linearly dependent. Equivalently, if the vectors v1 , v2 , . . . , vp in Rn are linearly inde-
pendent then p ≤ n.
53
Linear Independence

Proof. Let A = v1 v2 · · · vp . Thus, A is a n × p matrix. Since A has n rows, the
maximum rank of A is n, that is r ≤ n. Therefore, the number of free parameters d = p − r
is always positive because p > n ≥ r. Thus, the homogeneous system Ax = 0 has non-trivial
solutions. In other words, there is some non-zero vector x ∈ Rp such that
Ax = x1 v1 + x2 v2 + · · · + xp vp = 0
and therefore {v1 , v2 , . . . , vp } is linearly dependent.

Theorem 6.7 will be used when we discuss the notion of the dimension of a space.
Although we have not discussed the meaning of dimension, the above theorem says that in
n-dimensional space Rn , a set of vectors {v1 , v2 , . . . , vp } consisting of more than n vectors
is automatically linearly dependent.

         
8 4 2 3 0
3  11  0 −9 −2
 0  , v2 = −4 , v3 = 1 , v4 = −5 , v5 = −7 .
v1 =          
−2 6 1 3 7
Solution. The vectors v1 , v2 , v3 , v4 , v5 are in R4 . Therefore,

by Theorem 6.7,
the set {v1 , . . . , v5 }
is linearly dependent. To see this explicitly, let A = v1 v2 v3 v4 v5 . Then
 
1 0 0 0 −1
0 1 0 0 1 
A∼ 0 0 1 0 0 

0 0 0 1 −2
One solution to the linear system Ax = 0 is x = (−1, 1, 0, −2, −1) and therefore
(−1)v1 + (1)v2 + (0)v3 + (−2)v4 + (−1)v5 = 0
Example 6.9. Suppose that the set {v1 , v2 , v3 , v4 } is linearly independent. Show that the
set {v1 , v2 , v3 } is also linearly independent.
Solution. We must argue that if there exists scalars x1 , x2 , x3 such that
x1 v1 + x2 v2 + x3 v3 = 0
then necessarily x1 , x2 , x3 are all zero. Suppose then that there exists scalars x1 , x2 , x3 such
that
x1 v1 + x2 v2 + x3 v3 = 0.
Then clearly it holds that
x1 v1 + x2 v2 + x3 v3 + 0v4 = 0.
But the set {v1 , v2 , v3 , v4 } is linearly independent, and therefore, it is necessary that x1 , x2 , x3
are all zero. This proves that v1 , v2 , v3 are also linearly independent.
54
Lecture 6
The previous example can be generalized as follows: If {v1 , v2 , . . . , vd } is linearly inde-

pendent then any (non-empty) subset of the set {v1 , v2 , . . . , vd } is also linearly independent.

• the definition of linear independence and be able to explain it to a colleague
• how to test if a given set of vectors are linearly independent (Theorem 6.4)
• the relationship between the linear independence of {v1 , v2 , . . . , vp } and the solution
set of the homogeneous system Ax = 0, where A = v1 v2 · · · vp
• that in Rn , any set of vectors consisting of more than n vectors is automatically linearly
dependent (Theorem 6.7)
55
Linear Independence
56
Lecture 7
Lecture 7
Introduction to Linear Mappings
7.1 Vector mappings

By a vector mapping we mean simply a function
T : Rn → Rm .
The domain of T is Rn and the co-domain of T is Rm . The case n = m is allowed of

course. In engineering or physics, the domain is sometimes called the input space and the
co-domain is called the output space. Using this terminology, the points x in the domain
are called the inputs and the points T(x) produced by the mapping are called the outputs.
Definition 7.1: The vector b ∈ Rm is in the range of T, or in the image of T, if there

exists some x ∈ Rn such that T(x) = b.
In other words, b is in the range of T if there is an input x in the domain of T that outputs
b = T(x). In general, not every point in the co-domain of T is in the range of T. For
example, consider the vector mapping T : R2 → R2 defined as
" 2 #
x1 sin(x2 ) − cos(x21 − 1)
T(x) = .
x21 + x22 + 1
The vector b = (3, −1) is not in the range of T because the second component of T(x) is
positive. On the other hand, b = (−1, 2) is in the range of T because
2
1 1 sin(0) − cos(12 − 1) −1
T = = = b.
0 12 + 02 + 1 2
Hence, a corresponding input for this particular b is x = (1, 0). In Figure 7.1 we illustrate
the general setup of how the domain, co-domain, and range of a mapping are related. A
crucial idea is that the range of T may not equal the co-domain.
57
Rm , Co-domain
x b T(x)
b
R
an
ge
Rn , domain
Figure 7.1: The domain, co-domain, and range of a mapping.
7.2 Linear mappings
For our purposes, vector mappings T : Rn → Rm can be organized into two categories: (1)
linear mappings and (2) nonlinear mappings.
Definition 7.2: The vector mapping T : Rn → Rm is said to be linear if the following

conditions hold:
• For any u, v ∈ Rn , it holds that T(u + v) = T(u) + T(v).
• For any u ∈ Rn and any scalar c, it holds that T(cu) = cT(u).
If T is not linear then it is said to be nonlinear.
As an example, the mapping
" #
x21 sin(x2 ) − cos(x21 − 1)
T(x) =
x21 + x22 + 1
is nonlinear. To see this, previously we computed that

1 −1
T = .
0 2
58
Lecture 7
If T were linear then by property (2) of Definition 7.2 the following must hold:

3 1
T =T 3
0 0

1
= 3T
0

−1
=3
2

−3
= .
6
However,
2
3 3 sin(0) − cos(32 − 1)
T =
0 32 + 02 + 1

− cos(8)
=
10

−3
6= .
6
Example 7.3. Is the vector mapping T : R2 → R3 linear?

 
2x1 − x2
x1
T =  x1 + x2 
x2
−x1 − 3x2
Solution. We must verify that the two conditions in Definition 7.2 hold. For the first condi-
tion, take arbitrary vectors u = (u1 , u2 ) and v = (v1 , v2 ). We compute:

u1 + v1
T (u + v) = T
u2 + v2
 
2(u1 + v1 ) − (u2 + v2 )
=  (u1 + v1 ) + (u2 + v2 ) 
−(u1 + v1 ) − 3(u2 + v2 )
 
2u1 + 2v1 − u2 − v2
=  u1 + v1 + u2 + v2 
−u1 − v1 − 3u2 − 3v2
 
2u1 − u2 + 2v1 − v2
=  u1 + u2 + v1 + v2 
−u1 − 3u2 − v1 − 3v2
   
2u1 − u2 2v1 − v2
=  u1 + u2  +  v1 + v2 
−u1 − 3u2 −v1 − 3v2
= T(u) + T(v)
59
Therefore, for arbitrary u, v ∈ R2 , it holds that
T(u + v) = T(u) + T(v).
To prove the second condition, let c ∈ R be an arbitrary scalar. Then:

 
2(cu1 ) − (cu2 )
cu1
T(cu) = T =  (cu1) + (cu2) 
cu2
−(cu1 ) − 3(cu2)
 
c(2u1 − u2 )
=  c(u1 + u2 ) 
c(−u1 − 3u2 )
 
2u1 − u2
= c  u1 + u2 
−u1 − 3u2
= cT(u)
Therefore, both conditions of Definition 7.2 hold, and thus T is a linear map.
Example 7.4. Let α ≥ 0 and define the mapping T : Rn → Rn by the formula T(x) = αx.
If 0 ≤ α ≤ 1 then T is called a contraction and if α > 1 then T is called a dilation. In
either case, show that T is a linear mapping.
Solution. Let u and v be arbitrary. Then
T(u + v) = α(u + v) = αu + αv = T(u) + T(v).
This shows that condition (1) in Definition 7.2 holds. To show that the second condition
holds, let c is any number. Then
T(cx) = α(cx) = αcx = c(αx) = cT(x).
Therefore, both conditions of Definition 7.2 hold, and thus T is a linear mapping. To see a
particular example, consider the case α = 12 and n = 3. Then,
1 
x
2 1
 
 
T(x) = 21 x =  12 x2  .
 
1
x
2 3
60
Lecture 7
7.3 Matrix mappings

Given a matrix A ∈ Rm×n and a vector x ∈ Rn , in Lecture 4 we defined matrix-vector
multiplication between A and x as an operation that produces a new output vector Ax ∈ Rm .
We discussed that we could interpret A as a mapping that takes the input vector x ∈ Rn
and produces the output vector Ax ∈ Rm . We can therefore associate to each matrix A a
vector mapping T : Rn → Rm defined by
T(x) = Ax.
Such a mapping T will be called a matrix mapping corresponding to A and when con-
venient we will use the notation TA to indicate that TA is associated to A. We proved in
Lecture 4 (Theorem 4.3), that for any u, v ∈ Rn , and scalar c, matrix-vector multiplication
satisfies the properties:
1. A(u + v) = Au + Av
2. A(cu) = cAu.
The following theorem is therefore immediate.
Theorem 7.5: To a given matrix A ∈ Rm×n associate the mapping T : Rn → Rm defined

by the formula T(x) = Ax. Then T is a linear mapping.
Example 7.6. Is the vector mapping T : R2 → R3 linear?

 
2x1 − x2
x1
T =  x1 + x2 
x2
−x1 − 3x2
Solution. In Example 7.3 we showed that T is a linear mapping using Definition 7.2. Alter-
natively, we observe that T is a mapping defined using matrix-vector multiplication because
   
2x1 − x2 2 −1
x1 x
T =  x1 + x2  =  1 1 1
x2 x2
−x1 − 3x2 −1 −3
Therefore, T is a matrix mapping corresponding to the matrix

 
2 −1
A= 1 1
−1 −3
that is, T(x) = Ax. By Theorem 7.5, T is a linear mapping.
61
Let T : Rn → Rm be a vector mapping. Recall that b ∈ Rm is in the range of T if there

is some input vector x ∈ Rn such that T(x) = b. In this case, we say that b is the image
of x under T or that x is mapped to b under T. If T is a nonlinear mapping, finding a
specific vector x such that T(x) = b is generally a difficult problem. However, if T(x) = Ax
is a matrix mapping, then it is clear that finding such a vector x is equivalent to solving the
matrix equation Ax = b. In summary, we have the following theorem.
Theorem 7.7: Let T : Rn → Rm be a matrix mapping corresponding to A, that is,

T(x) = Ax. Then b ∈ Rm is in the range of T if and only if the matrix equation Ax = b
has a solution.
Let TA : Rn → Rm be a matrix mapping, that is, TA (x) = Ax. We proved that the
output vector Ax is a linear combination of the columns of A where the coefficients in the
linear combination are the components of x. Explicitly, if A = [v1 v2 · · · vn ] and the
components of x = (x1 , x2 , . . . , xn ) then
Ax = x1 v1 + x2 v2 + · · · + xn vn .
Therefore, the range of the matrix mapping TA(x) = Ax is
Range(TA ) = span{v1 , v2 , . . . , vn }.
In words, the range of a matrix mapping is the span of its columns. Therefore, if v1 , v2 , . . . , vn
span all of Rm then every vector b ∈ Rm is in the range of TA .
Example 7.8. Let    

1 3 −4 −2
A= 1 5 2 , b =  4 .
−3 −7 −6 12
Is the vector b in the range of the matrix mapping T(x) = Ax?
Solution. From Theorem 7.7, b is in the range of T if and only if the the matrix equation
Ax = b has a solution. To solve the system Ax = b, row reduce the augmented matrix
A b :    
1 3 −4 −2 1 3 −4 −2
1 5 2 4  ∼ 0 1 3 3
−3 −7 −6 12 0 0 −12 0
The system is consistent and the (unique) solution is x = (−11, 3, 0). Therefore, b is in the
range of T.
7.4 Examples
If T : Rn → Rm is a linear mapping, then for any vectors v1 , v2 , . . . , vp and scalars
c1 , c2 , . . . , cp , it holds that
T(c1 v1 + c2 v2 + · · · + cp vd ) = c1 T(v1 ) + c2 T(v2 ) + · · · + cd T(vp ). (⋆)
62
Lecture 7
Therefore, if all you know are the values T(v1 ), T(v2 ), . . . , T(vp ) and T is linear, then you
can compute T(v) for every
v ∈ span{v1 , v2 , . . . , vp }.
Example 7.9. Let T : R2 → R2 be a linear transformation that maps u to T(u) = (3, 4)

and maps v to T(v) = (−2, 5). Find T(2u + 3v).
Solution. Because T is a linear mapping we have that
T(2u + 3v) = T(2u) + T(3v) = 2T(u) + 3T(v).
We know that T(u) = (3, 4) and T(v) = (−2, 5). Therefore,

3 −2 0
T(2u + 3v) = 2T(u) + 3T(v) = 2 +3 = .
4 5 23
Example 7.10. (Rotations) Let Tθ : R2 → R2 be the mapping on the 2D plane that rotates
every v ∈ R2 by an angle θ. Write down a formula for Tθ and show that Tθ is a linear
mapping.
Tθ (v)
b
θ b
v
α
b
Solution. If v = (cos(α), sin(α)) then

" #
cos(α + θ)
Tθ (v) = .
sin(α + θ)
Then from the angle sum trigonometric identities:
" # " #
cos(α + θ) cos(α) cos(θ) − sin(α) sin(θ)
Tθ (v) = =
sin(α + θ) cos(α) sin(θ) + sin(α) cos(θ)
But " # " #" #
cos(α) cos(θ) − sin(α) sin(θ) cos(θ) − sin(θ) cos(α)
Tθ (v) = = .
cos(α) sin(θ) + sin(α) cos(θ) sin(θ) cos(θ) sin(α)
| {z }
v
63
If we scale v by any c > 0 then performing the same computation as above we obtain that
Tθ (cv) = cT(v). Therefore, Tθ is a matrix mapping with corresponding matrix
" #
cos(θ) − sin(θ)
A= .
sin(θ) cos(θ)
Thus, Tθ is a linear mapping.

Example 7.11. (Projections) Let T : R3 → R2 be the vector mapping
   
x1 x1
T   x2  = x2  .

x3 0
Show that T is a linear mapping and describe the range of T.

Solution. First notice that
      
x1 x1 1 0 0 x1
T x2  = x2  = 0 1 0 x2  .
x3 0 0 0 0 x3
Thus, T is a matrix mapping corresponding to the matrix

 
1 0 0
A =  0 1 0 .
0 0 0
Therefore, T is a linear mapping. Geometrically, T takes the vector x and projects it to the
(x1 , x2 ) plane, see Figure 7.2. What is the range of T? The range of T consists of all vectors
in R3 of the form  
t
b = s

0
where the numbers t and s are arbitrary. For each b in the range of T, there are infinitely
many x’s such that T(x) = b.
 
x1
b x = x2 

x3
b
 
x1
b T(x) = x2 

0
Figure 7.2: Projection onto the (x1 , x2 ) plane
64
Lecture 7

• what a vector mapping is
• what the range of a vector mapping is
• that the co-domain and range of a vector mapping are generally not the same
• what a linear mapping is and how to check when a given mapping is linear
• what a matrix mapping is and that they are linear mappings
• how to determine if a vector b is in the range of a matrix mapping
• the formula for a rotation in R2 by an angle θ
65
66
Lecture 8
Lecture 8
Onto and One-to-One Mappings,

and the Matrix of a Linear Mapping
8.1 Onto Mappings

We have seen through examples that the range of a vector mapping (linear or nonlinear) is
not always the entire co-domain. For example, if TA (x) = Ax is a matrix mapping and b
is such that the equation Ax = b has no solutions then the range of T does not contain b
and thus the range is not the whole co-domain.
Definition 8.1: A vector mapping T : Rn → Rm is said to be onto if for each b ∈ Rm

there is at least one x ∈ Rn such that T(x) = b.
For a matrix mapping TA (x) = Ax, the range of TA is the span of the columns of A.
Therefore:
Theorem 8.2: Let TA : Rn → Rm be the matrix mapping TA (x) = Ax, where A ∈

Mm×n . Then TA is onto if and only if the columns of A span all of Rm .
Combining Theorem 4.11 and Theorem 8.2 we have:
Theorem 8.3: Let TA : Rn → Rm be the matrix mapping TA (x) = Ax, where A ∈ Rm×n .
Then TA is onto if and only if r = rank(A) = m.
Example 8.4. Let T : R3 → R3 be the matrix mapping with corresponding matrix

 
1 2 −1
A = −3 −4 2 
5 2 3
Is TA onto?
67
Onto, One-to-One, and Standard Matrix
Solution. The rref(A) is    

1 2 −1 1 0 0
−3 −4 2  ∼ 0 1 0
5 2 3 0 0 1
Therefore, r = rank(A) = 3. The dimension of the co-domain is m = 3 and therefore TA is
onto. Therefore, the columns of A span all of R3 , that is, every b ∈ R3 can be written as a
linear combination of the columns of A:
     
 1 2 −1 
span −3 , −4 , 2  = R3
   
 
2 2 3
Example 8.5. Let TA : R4 → R3 be the matrix mapping with corresponding matrix

 
1 2 −1 4
A = −1 4
 1 8
2 0 −2 0
Is TA onto?
Solution. The rref(A) is
   
1 2 −1 4 1 0 −1 0
A = −1 4 1 8 ∼ 0 1 0 2
2 0 −2 0 0 0 0 0
Therefore, r = rank(A) = 2. The dimension of the co-domain is m = 3 and therefore TA is

not onto. Notice that v3 = −v1 and v4 = 2v2 . Thus, v3 and v4 are already in the span of
the columns v1 , v2 . Therefore,
6 R3 .
span{v1 , v2 , v3 , v4 } = span{v1 , v2 } =
Below is a theorem which places restrictions on the size of the domain of an onto mapping.
Theorem 8.6: Suppose that TA : Rn → Rm is a matrix mapping corresponding to

A ∈ Mm×n . If TA is onto then m ≤ n.
Proof. If TA is onto then the rref(A) has r = m leading 1’s. Therefore, A has at least m
columns. The number of columns of A is n. Therefore, m ≤ n.
An equivalent way of stating Theorem 8.6 is the following.
68
Lecture 8
Corollary 8.7: If TA : Rn → Rm is a matrix mapping corresponding to A ∈ Mm×n and

n < m then TA cannot be onto.
Intuitively, if the domain Rn is “smaller” than the co-domain Rm and TA : Rn → Rm is

linear then TA cannot be onto. For example, a matrix mapping TA : R → R2 cannot be
onto. Linearity plays a key role in this. In fact, there exists a continuous (nonlinear) function
f : R → R2 whose range is a square! In this case, the domain is 1-dimensional and the range
is 2-dimensional. This situation cannot happen when the mapping is linear.
Example 8.8. Let TA : R2 → R3 be the matrix mapping with corresponding matrix

 
1 4
A = −3 2
2 1
Is TA onto?
Solution. TA is onto because the domain is R2 and the co-domain is R3 . Intuitively, two
vectors are not enough to span R3 . Geometrically, two vectors in R3 span a 2D plane going
through the origin. The vectors not on the plane span{v1 , v2 } are not in the range of TA.
8.2 One-to-One Mappings

Given a linear mapping T : Rn → Rm , the question of whether b ∈ Rm is in the range of T
is an existence question. Indeed, if b ∈ Range(T) then there exists a x ∈ Rm such that
T(x) = b. We now want to look at the problem of whether x is unique. That is, does there
exist a distinct y such that T(y) = b.
Definition 8.9: A vector mapping T : Rn → Rm is said to be one-to-one if for each

b ∈ Range(T) there exists only one x ∈ Rn such that T(x) = b.
When T is a linear mapping, we have all the tools necessary to give a complete description
of when T is one-to-one. To do this, we use the fact that if T : Rn → Rm is linear then
T(0) = 0. Here is one proof: T(0) = T(x − x) = T(x) − T(x) = 0.
Theorem 8.10: Let T : Rn → Rm be linear. Then T is one-to-one if and only if T(x) = 0

implies that x = 0.
If TA : Rn → Rm is a matrix mapping then according to Theorem 8.10, TA is one-to-one

if and only if the only solution to Ax = 0 is x = 0. We gather these facts in the following
theorem.
69
Theorem 8.11: Let TA : Rn → Rm be a matrix mapping, where A = [v1 v2 · · · vn ] ∈

Mm×n . The following statements are equivalent:
1. TA is one-to-one.
2. The rank of A is r = rank(A) = n.
3. The columns v1 , v2 , . . . , vn are linearly independent.
Example 8.12. Let TA : R4 → R3 be the matrix mapping with matrix

 
3 −2 6 4
A = −1 0 −2 −1 .
2 −2 0 2
Is TA one-to-one?
Solution. By Theorem 8.11, TA is one-to-one if and only if the columns of A are linearly
independent. The columns of A lie in R3 and there are n = 4 columns. From Lecture 6, we
know then that the columns are not linearly independent. Therefore, TA is not one-to-one.
Alternatively, A will have rank at most r = 3 (why?). Therefore, the solution set to Ax = 0
will have at least one parameter, and thus there exists infinitely many solutions to Ax = 0.
Intuitively, because R4 is “larger” than R3 , the linear mapping TA will have to project R4
onto R3 and thus infinitely many vectors in R4 will be mapped to the same vector in R3 .
Example 8.13. Let TA : R2 → R3 be the matrix mapping with matrix

 
1 0
A = 3 −1
2 0
Is TA one-to-one?
Solution. By inspection, we see that the columns of A are linearly independent. Therefore,
TA is one-to-one. Alternatively, one can compute that
 
1 0
rref(A) = 0 1
0 0
Therefore, r = rank(A) = 2, which is equal to the number columns of A.
70
Lecture 8
8.3 Standard Matrix of a Linear Mapping

We have shown that all matrix mappings TA are linear mappings. We now want to answer
the reverse question: Are all linear mappings matrix mappings in disguise? If T : Rn → Rm
is a linear mapping, then to show that T is in fact a matrix mapping we must show that
there is some matrix A ∈ Mm×n such that T(x) = Ax. To that end, introduce the standard
unit vectors e1 , e2 , . . . , en in Rn :
       
1 0 0 0
0 1 0 0
       
e1 = 0 , e2 = 0 , e3 = 1 , · · · , en = 0 .
       
 ..   ..   ..   .. 
. . . .
0 0 0 1
Every x ∈ Rn is in span{e1 , e2 , . . . , en } because:

       
x1 1 0 0
 x2  0 1 0
       
x =  ..  = x1  ..  + x2  ..  + · · · + xn  ..  = x1 e1 + x2 e2 + · · · + xn en
. . . .
xn 0 0 1
With this notation we prove the following.
Theorem 8.14: Every linear mapping is a matrix mapping.
Proof. Let T : Rn → Rm be a linear mapping. Let
v1 = T(e1 ), v2 = T(e2 ), . . . , vn = T(en ).
The co-domain of T is Rm , and thus vi ∈ Rm . Now, for arbitrary x ∈ Rn we can write
x = x1 e1 + x2 e2 + · · · + xn en .
Then by linearity of T, we have
T(x) = T(x1 e1 + x2 e2 + · · · + xn en )
= x1 T(e1 ) + x2 T(e2 ) + · · · + xn T(en )
= x1 v1 + x2 v2 + · · · + xn vn

= v1 v2 · · · vn x.

Define the matrix A ∈ Mm×n by A = v1 v2 · · · vn . Then our computation above
shows that
T(x) = x1 v1 + x2 v2 + · · · + xn vn = Ax.
Therefore, T is a matrix mapping with the matrix A ∈ Mm×n .
71
If T : Rn → Rm is a linear mapping, the matrix

A = T(e1 ) T(e2 ) · · · T(en )
is called the standard matrix of T. In words, the columns of A are the images of the
standard unit vectors e1 , e2 , . . . , en under T. The punchline is that if T is a linear mapping,
then to derive properties of T we need only know the standard matrix A corresponding to
T.
Example 8.15. Let T : R2 → R2 be the linear mapping that

rotates every vector by an
1 0
angle θ. Use the standard unit vectors e1 = and e2 = in R2 to write down the
0 1
matrix A ∈ R2×2 corresponding to T.
Tθ (e2 ) b e2
b b
Tθ (e1 )
θ b
e1
Solution. We have " #

cos(θ) − sin(θ)
A = T(e1 ) T(e2 ) =
sin(θ) cos(θ)
Example 8.16. Let T : R3 → R3 be a dilation of factor k = 2. Find the standard matrix

A of T.
Solution. The mapping is T(x) = 2x. Then
           
1 2 0 0 0 0
T(e1 ) = 2 0 = 0 , T(e2 ) = 2 1 = 2 , T(e3 ) = 2 0 = 0
0 0 0 0 1 2
Therefore,  
2 0 0
A = T(e1 ) T(e2 ) T(e3 ) = 0 2 0
0 0 2
is the standard matrix of T.
72
Lecture 8
• the relationship between the range of a matrix mapping T(x) = Ax and the span of
the columns of A
• what it means for a mapping to be onto and one-to-one
• how to verify if a linear mapping is onto and one-to-one
• that all linear mappings are matrix mappings
• what the standard unit vectors are
• how to compute the standard matrix of a linear mapping
73
74
Lecture 9
Lecture 9
Matrix Algebra
9.1 Sums of Matrices

We begin with the definition of matrix addition.
Definition 9.1: Given matrices

   
a11 a12 · · · a1n b11 b12 · · · b1n
 a21 a22 · · · a2n   b21 b22 · · · b2n 
   
A =  .. .. .. ..  , B =  .. .. .. ..  ,
 . . . .   . . . . 
am1 am2 · · · amn bm1 bm2 · · · bmn
both of the same dimension m × n, the sum A + B is defined as

 
a11 + b11 a12 + b12 ··· a1n + b1n
 a21 + b21 a22 + b22 ··· a2n + b2n 
 
A+B= .. .. .. .. .
 . . . . 
am1 + bm1 am2 + bm2 · · · amn + bmn
Next is the definition of scalar-matrix multiplication.
Definition 9.2: For a scalar α we define αA by

   
a11 a12 · · · a1n αa11 αa12 · · · αa1n
 a21 a22 · · · a2n   αa21 αa22 · · · αa2n 
   
αA = α  .. .. .. ..  =  .. .. .. ..  .
 . . . .   . . . . 
am1 am2 · · · amn αam1 αam2 · · · αamn
75
Matrix Algebra
Example 9.3. Given A and B below, find 3A − 2B.

   
1 −2 5 5 0 −11
A = 0 −3
 9 , B =  3 −5 1 
4 −6 7 −1 −9 0
   
3 −6 15 10 0 −22
3A − 2B =  0 −9 27 −  6 −10 2 
12 −18 21 −2 −18 0
 
−7 −6 37
= −6 1 25
14 0 21
Below are some basic algebraic properties of matrix addition/scalar multiplication.
Theorem 9.4: Let A, B, C be matrices of the same size and let α, β be scalars. Then
(a) A + B = B + A (d) α(A + B) = αA + αB

(b) (A + B) + C = A + (B + C) (e) (α + β)A = αA + βA
(c) A + 0 = A (f) α(βA) = (αβ)A
9.2 Matrix Multiplication

Let TB : Rp → Rn and let TA : Rn → Rm be linear mappings. If x ∈ Rp then TB (x) ∈ Rn
and thus we can apply TA to TB (x). The resulting vector TA (TB (x)) is in Rm . Hence, each
x ∈ Rp can be mapped to a point in Rm , and because TB and TA are linear mappings the
resulting mapping is also linear. This resulting mapping is called the composition of TA
and TB , and is usually denoted by TA ◦ TB : Rp → Rm (see Figure 9.1). Hence,
(TA ◦ TB )(x) = TA (TB (x)).
Because (TA ◦ TB ) : Rp → Rm is a linear mapping it has an associated standard matrix,

which we denote for now by C. From Lecture 8, to compute the standard matrix of any
linear mapping, we must compute the images of the standard unit vectors e1 , e2 , . . . , ep under
the linear mapping. Now, for any x ∈ Rp ,
TA (TB (x)) = TA (Bx) = A(Bx).
Applying this to x = ei for all i = 1, 2, . . . , p, we obtain the standard matrix of TA ◦ TB :

C = A(Be1 ) A(Be2 ) · · · A(Bep ) .
76
Lecture 9
Rm
Rp TB Rn TA
b b
x b TB (x) TA (TB(x))
(TA ◦ TB )(x)
Figure 9.1: Illustration of the composition of two mappings.
Now Be1 is

Be1 = b1 b2 · · · bp e1 = b1 .
And similarly Bei = bi for all i = 1, 2, . . . , p. Therefore,

C = Ab1 Ab2 · · · Abp
is the standard matrix of TA ◦ TB . This computation motivates the following definition.

Definition 9.5: For A ∈ Rm×n and B ∈ Rn×p , with B = b1 b2 · · · bp , we define the
product AB by the formula

AB = Ab1 Ab2 · · · Abp .
The product AB is defined only when the number of columns of A equals the number of
rows of B. The following diagram is useful for remembering this:
(m × n) · (n × p) → m × p
From our definition of AB, the standard matrix of the composite mapping TA ◦ TB is
C = AB.
In other words, composition of linear mappings corresponds to matrix multiplication.
Example 9.6. For A and B below compute AB and BA.

 
−4 2 4 −4
1 2 −2
A= , B =  −1 −5 −3 3 
1 1 −3
−4 −4 −3 −1
77
Matrix Algebra
Solution. First AB = [Ab1 Ab2 Ab3 Ab4 ]:
 
−4 2 4 −4
1 2 −2
AB =  −1 −5 −3 3 
1 1 −3
−4 −4 −3 −1

2
=
7

2 0
=
7 9

2 0 4
=
7 9 10

2 0 4 4
=
7 9 10 2
On the other hand, BA is not defined! B has 4 columns and A has 2 rows.
Example 9.7. For A and B below compute AB and BA.
   
−4 4 3 −1 −1 0
A =  3 −3 −1  , B =  −3 0 −2 
−2 −1 1 −2 1 −2
Solution. First AB = [Ab1 Ab2 Ab3 ]:
  
−4 4 3 −1 −1 0
AB =  3 −3 −1   −3 0 −2 
−2 −1 1 −2 1 −2

−14
=  8
3

−14 7
=  8 −4
3 3
 
−14 7 −14
= 8 −4 8 
3 3 0
78
Lecture 9
Next BA = [Ba1 Ba2 Ba3 ]:

  
−1 −1 0 −4 4 3
BA =  −3 0 −2   3 −3 −1 
−2 1 −2 −2 −1 1

1
= 16

15

1 −1
=  16 −10
15 −9
 
1 −1 −2
=  16 −10 −11 
15 −9 −9
On the other hand:  

−14 7 −14
AB =  8 −4 8 
3 3 0
Therefore, in general AB 6= BA, i.e., matrix multiplication is not commutative.
An important matrix that arises frequently is the identity matrix In ∈ Rn×n of size
n:  
1 0 0 ··· 0
0 1 0 · · · 0
 
In =  .. .. .. .. 
. . . · · · .
0 0 0 ··· 1
You should verify that for any A ∈ Rn×n it holds that AIn = In A = A. Below are some
basic algebraic properties of matrix multiplication.
Theorem 9.8: Let A, B, C be matrices, of appropriate dimensions, and let α be a scalar.

Then
(1) A(BC) = (AB)C
(2) A(B + C) = AB + AC
(3) (B + C)A = BA + CA
(4) α(AB) = (αA)B = A(αB)
(5) In A = AIn = A
If A ∈ Rn×n is a square matrix, the kth power of A is
Ak = |AAA
{z· · · A}
k times
79
Matrix Algebra
Example 9.9. Compute A3 if

−2 3
A= .
1 0
Solution. Compute A2 :

2 −2 3 −2 3 7 −6
A = =
1 0 1 0 −2 3
And then A3 :

3 2 7 −6 −2 3
A =A A=
−2 3 1 0

−20 21
=
7 −6
We could also do:

3 2 −2 3 7 −6 −20 21
A = AA = = .
1 0 −2 3 7 −6
9.3 Matrix Transpose

We begin with the definition of the transpose of a matrix.
Definition 9.10: Given a matrix A ∈ Rm×n , the transpose of A is the matrix AT whose
ith column is the ith row of A.
If A is m × n then AT is n × m. For example, if

 
0 −1 8 −7 −4
 −4 6 −10 −9 6 
A=  9

5 −2 −3 5 
−8 8 4 7 7
then  
0 −4 9 −8

 −1 6 5 8 

AT = 
 8 −10 −2 4 
.
 −7 −9 −3 7 
−4 6 5 7
Example 9.11. Compute (AB)T and BT AT if

 
−2 1 2
−2 1 0
A= , B =  −1 −2 0 .
3 −1 −3
0 0 −1
80
Lecture 9
Solution. Compute AB:

 
−2 1 2
−2 1 0
AB =  −1 −2 0 
3 −1 −3
0 0 −1

3 −4 −4
=
−5 5 9
Next compute BT AT :
  
−2 −1 0 −2 3
BT AT =  1 −2 0   1 −1 
2 0 −1 0 −3
 
3 −5
= −4
 5  = (AB)T
−4 9
The following theorem summarizes properties of the transpose.
Theorem 9.12: Let A and B be matrices of appropriate sizes. The following hold:
(1) (AT )T = A
(2) (A + B)T = AT + BT
(3) (αA)T = αAT
(4) (AB)T = BT AT
A consequence of property (4) is that
(A1 A2 . . . Ak )T = ATk ATk−1 · · · AT2 AT1
and as a special case

(Ak )T = (AT )k .
Example 9.13. Let T : R2 → R2 be the linear mapping that first contracts vectors by a
factor of k = 3 and then rotates by an angle θ. What is the standard matrix A of T?
Solution. Let e1 = (1, 0) and e2 = (0, 1) denote the standard
unit vectors in R2 . From
Lecture 8, the standard matrix of T is A = T(e1 ) T(e2 ) . Recall that the standard matrix
of a rotation by θ is
cos(θ) − sin(θ)
sin(θ) cos(θ)
Contracting e1 by a factor of k = 3 results in ( 13 , 0) and then rotation by θ results in
1
cos(θ)
3
1 = T(e1 ).
3
sin(θ)
81
Matrix Algebra
Contracting e2 by a factor of k = 3 results in (0, 13 ) and then rotation by θ results in

1
− 3 sin(θ)
1 = T(e2 ).
3
cos(θ)
Therefore, "1 #
3
cos(θ) − 13 sin(θ)
A = T(e1 ) T(e2 ) =
1 1
3
sin(θ) 3
cos(θ)
1
On the other hand, the standard matrix corresponding to a contraction by a factor k = 3
is
"1 #
3
0
1
0 3
Therefore, " # "1 # "1 #

cos(θ) − sin(θ) 3
0 3
cos(θ) − 13 sin(θ)
= =A
sin(θ) cos(θ) 0 1 1
sin(θ) 1
cos(θ)
| {z } | {z 3 } 3 3
rotation contraction

• know how to add and multiply matrices
• that matrix multiplication corresponds to composition of linear mappings
• the algebraic properties of matrix multiplication (Theorem 9.8)
• how to compute the transpose of a matrix
• the properties of matrix transposition (Theorem 9.12)
82
Lecture 10
Lecture 10
Invertible Matrices
10.1 Inverse of a Matrix

The inverse of a square matrix A ∈ Rn×n generalizes the notion of the reciprocal of a non-
zero number a ∈ R. Formally speaking, the inverse of a non-zero number a ∈ R is the unique
number c ∈ R such that ac = ca = 1. The inverse of a 6= 0, usually denoted by a−1 = a1 , can
be used to solve the equation ax = b:
ax = b ⇒ a−1 ax = a−1 b ⇒ x = a−1 b.
This motivates the following definition.
Definition 10.1: A matrix A ∈ Rn×n is called invertible if there exists a matrix C ∈

Rn×n such that AC = In and CA = In .
If A is invertible then can it have more than one inverse? Suppose that there exists C1 , C2
such that ACi = Ci A = In . Then
C2 = C2 (AC1 ) = (C2 A)C1 = In C1 = C1 .
Thus, if A is invertible, it can have only one inverse. This motivates the following definition.
Definition 10.2: If A is invertible then we denote the inverse of A by A−1 . Thus,

AA−1 = A−1 A = In .
Example 10.3. Given A and C below, show that C is the inverse of A.

   
1 −3 0 −14 −3 −6
A =  −1 2 −2  , C =  −5 −1 −2 
−2 6 1 2 0 1
83
Invertible Matrices
Solution. Compute AC:

    
1 −3 0 −14 −3 −6 1 0 0
AC =  −1 2 −2   −5 −1 −2  = 0 1 0
−2 6 1 2 0 1 0 0 1
Compute CA:
    
−14 −3 −6 1 −3 0 1 0 0
CA =  −5 −1 −2   −1 2 −2  = 0 1 0
2 0 1 −2 6 1 0 0 1
Therefore, by definition C = A−1 .
Theorem 10.4: Let A ∈ Rn×n and suppose that A is invertible. Then for any b ∈ Rn
the matrix equation Ax = b has a unique solution given by A−1 b.
Proof: Let b ∈ Rn be arbitrary. Then multiplying the equation Ax = b by A−1 from the
left we obtain that
A−1 Ax = A−1 b
⇒ In x = A−1 b
⇒ x = A−1 b.
Therefore, with x = A−1 b we have that
Ax = A(A−1 b) = AA−1 b = In b = b
and thus x = A−1 b is a solution. If x̃ is another solution of the equation, that is, Ax̃ = b,
then multiplying both sides by A−1 we obtain that x̃ = A−1 b. Thus, x = x̃.
Example 10.5. Use the result of Example 10.3. to solve the linear system Ax = b if
   
1 −3 0 1
A = −1
 2 −2 , b = −3 .
 
−2 6 1 −1
Solution. We showed in Example 10.3 that

 
−14 −3 −6
A−1 =  −5 −1 −2  .
2 0 1
Therefore, the unique solution to the linear system Ax = b is

    
−14 −3 −6 1 1
A−1 b =  −5 −1 −2  −3 = 0
2 0 1 −1 1
84
Lecture 10
Verify:     
1 −3 0 1 1
 −1 2 −2  0 = −3
−2 6 1 1 −1
The following theorem summarizes the relationship between the matrix inverse and ma-
trix multiplication and matrix transpose.
Theorem 10.6: Let A and B be invertible matrices. Then:

(1) The matrix A−1 is invertible and its inverse is A:
(A−1 )−1 = A.
(2) The matrix AB is invertible and its inverse is B−1 A−1 :
(AB)−1 = B−1 A−1 .
(3) The matrix AT is invertible and its inverse is (A−1 )T :
(AT )−1 = (A−1)T .
Proof: To prove (2) we compute
(AB)(B−1A−1 ) = ABB−1A−1 = AIn A−1 = AA−1 = In .
To prove (3) we compute
AT (A−1)T = (A−1 A)T = ITn = In .
10.2 Computing the Inverse of a Matrix

If A ∈ Mn×n is invertible, how do we find A−1 ? Let A−1 = c1 c2 · · · cn and we will
find expressions for ci . First note that AA−1 = Ac1 Ac2 · · · Acn . On the other hand,
we also have AA−1 = In = e1 e2 · · · en . Therefore, we want to find c1 , c2 , . . . , cn such
that
Ac1 Ac2 · · · Acn = e1 e2 · · · en .
| {z } | {z }
AA−1 In
To find ci we therefore need to solve the linear system

Ax
= ei . Here the image vector “b”
is ei . To find c1 we form the augmented matrix A e1 and find its RREF:

A e1 ∼ In c1 .
85
Invertible Matrices
We will need to do this for each

c2 , . . . , cn so we might as well form the combined augmented
matrix A e1 e2 · · · en and find the RREF all at once:

A e1 e2 · · · en ∼ In c1 c2 · · · cn .
In summary, to determine if A−1 exists and to simultaneously compute it, we compute the
RREF of the augmented matrix
A In ,
that is, A augmented with the n × n identity matrix. If the RREF of A is In , that is

A I n ∼ I n c1 c2 · · · cn
then
A−1 = c1 c2 · · · cn .
If the RREF of A is not In then A is not invertible.

1 3
Example 10.7. Find the inverse of A = if it exists.
−1 −2

Solution. Form the augmented matrix A I2 and row reduce:

1 3 1 0
A I2 =
−1 −2 0 1
Add rows R1 and R2 :

1 3 1 0 R1 +R2 1 3 1 0
−−−−→
−1 −2 0 1 0 1 1 1
−3R +R
Perform the operation −−−2−−→
1
:

1 3 1 0 −3R2 +R1 1 0 −2 −3
−−−−−→
0 1 1 1 0 1 1 1
Thus, rref(A) = I2 , and therefore A is invertible. The inverse is

−1 −2 −3
A =
1 1
Verify:
−1 1 3 −2 −3 1 0
AA = = .
−1 −2 1 1 0 1
 
1 0 3
Example 10.8. Find the inverse of A =  1 1 0  if it exists.
−2 0 −7
86
Lecture 10

   
1 0 3 1 0 0 1 0 3 1 0 0
−R1 +R2 , 2R1 +R2
1 1 0 0 1 0 −−−−−−−−−−→ 0 1 −3 −1 1 0
−2 0 −7 0 0 1 0 0 −1 2 0 1
−R3 :    
1 0 3 1 0 0 1 0 3 1 0 0
−R3
0 1 −3 −1 1 0 − −→ 0 1 −3 −1 1 0 
0 0 −1 2 0 1 0 0 1 −2 0 −1
3R3 + R2 and −3R3 + R1 :
   
1 0 3 1 0 0 1 0 0 7 0 3
3R3 +R2 , −3R3 +R1
0 1 −3 −1 1 0  − −−−−−−−−−−→ 0 1 0 −7 1 −3
0 0 1 −2 0 −1 0 0 1 −2 0 −1
Therefore, rref(A) = I3 , and therefore A is invertible. The inverse is
 
7 0 3
A−1 = −7 1 −3
−2 0 −1
Verify:     
1 0 3 7 0 3 1 0 0
AA−1 =  1 1 0  −7 1 −3 = 0 1 0
−2 0 −7 −2 0 −1 0 0 1
 
1 0 1
Example 10.9. Find the inverse of A =  1 1 −2  if it exists.
−2 0 −2

   
1 0 1 1 0 0 1 0 1 1 0 0
−R1 +R2 , 2R1 +R2
 1 1 −2 0 1 0 − −−−−−−−−−→ 0 1 −3 −1 1 0
−2 0 −2 0 0 1 0 0 0 2 0 1
We need not go further since the rref(A) is not I3 (rank(A) = 2 ). Therefore, A is not
invertible.
10.3 Invertible Linear Mappings

Let TA : Rn → Rn be a matrix mapping with standard matrix A and suppose that A is
invertible. Let TA−1 : Rn → Rn be the matrix mapping with standard matrix A−1 . Then
the standard matrix of the composite mapping TA−1 ◦ TA : Rn → Rn is
A−1 A = In .
87
Invertible Matrices
Therefore, (TA−1 ◦ TA )(x) = In x = x. Let’s unravel (TA−1 ◦ TA )(x) to see this:
(TA−1 ◦ TA )(x) = TA−1 (TA (x)) = TA−1 (Ax) = A−1 Ax = x.
Similarly, the standard matrix of (TA ◦ TA−1 ) is also In . Intuitively, the linear mapping TA−1
undoes what TA does, and conversely. Moreover, since Ax = b always has a solution, TA is
onto. And, because the solution to Ax = b is unique, TA is one-to-one.
The following theorem summarizes equivalent conditions for matrix invertibility.
Theorem 10.10: Let A ∈ Rn×n . The following statements are equivalent:

(a) A is invertible.
(b) A is row equivalent to In , that is, rref(A) = In .
(c) The equation Ax = 0 has only the trivial solution.
(d) The linear transformation TA (x) = Ax is one-to-one.
(e) The linear transformation TA (x) = Ax is onto.
(f) The matrix equation Ax = b is always solvable.
(g) The columns of A span Rn .
(h) The columns of A are linearly independent.
(i) AT is invertible.
Proof: This is a summary of all the statements we have proved about matrices and matrix
mappings specialized to the case of square matrices A ∈ Rn×n . Note that for non-square
matrices, one-to-one does not imply ontoness, and conversely.
Example 10.11. Without doing any arithmetic, write down the inverse of the dilation
matrix " #
3 0
A= .
0 5
Example 10.12. Without doing any arithmetic, write down the inverse of the rotation
matrix " #
cos(θ) − sin(θ)
A= .
sin(θ) cos(θ)

• how to compute the inverse of a matrix
• properties of matrix inversion and matrix multiplication
• relate invertibility of a matrix with properties of the associated linear mapping (1-1,
onto)
• the characterizations of invertible matrices Theorem 10.10
88
Lecture 11
Lecture 11
Determinants
11.1 Determinants of 2 × 2 and 3 × 3 Matrices

Consider a general 2 × 2 linear system
a11 x1 + a12 x2 = b1
a21 x1 + a22 x2 = b2 .
Using elementary row operations, it can be shown that the solution is

b1 a22 − b2 a12 b2 a11 − b1 a21
x1 = , x2 = ,
a11 a22 − a12 a21 a11 a22 − a12 a21
provided that a11 a22 − a12 a21 6= 0. Notice the denominator is the same in both expressions.
The number a11 a22 − a12 a21 then completely characterizes when a 2 × 2 linear system has a
unique solution. This motivates the following definition.
Definition 11.1: Given a 2 × 2 matrix

a a
A = 11 12
a21 a22
we define the determinant of A as

a11 a12
det A = det = a11 a22 − a12 a21 .
a21 a22
An alternative notation for det A is using vertical bars:

a11 a12 a11 a12
det = .
a21 a22 a21 a22
89
Determinants
Example 11.2. Compute the determinant of A.

3 −1
(i) A =
8 2

3 1
(ii) A =
−6 −2

−110 0
(iii) A =
568 0
Solution. For (i):
3 −1
det(A) =
= (3)(2) − (8)(−1) = 14
8 2
For (ii):
3 1
det(A) =
= (3)(−2) − (−6)(1) = 0
−6 −2
For (iii):
−110 0
det(A) = = (−110)(0) − (568)(0) = 0
568 0
As in the 2 × 2 case, the solution of a 3 × 3 linear system Ax = b can be shown to be

Numerator1 Numerator2 Numerator3
x1 = , x2 = , x3 =
D D D
where
D = a11 (a22 a33 − a23 a32 ) − a12 (a21 a33 − a23 a31 ) + a13 (a21 a32 − a22 a31 ).
Notice that the terms of D in the parenthesis are determinants of 2 × 2 submatrices of A:
D = a11 (a22 a33 − a23 a32 ) − a12 (a21 a33 − a23 a31 ) + a13 (a21 a32 − a22 a31 ).
| {z } | {z } | {z }
a22 a23 a21 a23 a21 a22

a32 a33 a31 a33 a31 a32

Let
a22 a23 a21 a23 a21 a22
A11 = , A12 = , and A13 = .
a32 a33 a31 a33 a31 a32
Then we can write
D = a11 det(A11 ) − a12 det(A12 ) + a13 det(A13 ).

a22 a23
The matrix A11 = is obtained from A by deleting the 1st row and the 1st column:
a32 a33
 
a11 a12 a13
a 22 a23
A = a21 a22 a23  −→ A11 =
 .
a32 a33
a31 a32 a33
90
Lecture 11

a21 a23
Similarly, the matrix A12 = is obtained from A by deleting the 1st row and the
a31 a33
2nd column:  
a11 a12 a13
a21 a 23
A = a21 a22 a23  −→ A12 = .
a31 a33
a31 a32 a33

a21 a22
Finally, the matrix A13 = is obtained from A by deleting the 1st row and the 3rd
a31 a32
column:  
a11 a12 a13
a21 a22
A = a21 a22 a23  −→ .
a31 a32
a31 a32 a33
Notice also that the sign in front of the coefficients a11 , a12 , and a13 , alternate. This motivates
the following definition.
Definition 11.3: Let A be a 3 × 3 matrix. Let Ajk be the 2 × 2 matrix obtained from
A by deleting the jth row and kth column. Define the cofactor of ajk to be the number
Cjk = (−1)j+k det Ajk . Define the determinant of A to be
det A = a11 C11 + a12 C12 + a13 C13 .
This definition of the determinant is called the expansion of the determinant along the
first row. In the cofactor Cjk = (−1)j+k det Ajk , the expression (−1)j+k will evaluate to
either 1 or −1, depending on whether j + k is even or odd. For example, the cofactor of a12
is
C12 = (−1)1+2 det A12 = − det A12
and the cofactor of a13 is
C13 = (−1)1+3 det A13 = det A13 .
We can also compute the cofactor of the other entries of A in the obvious way. For example,
the cofactor of a23 is
C23 = (−1)2+3 det A23 = − det A23 .
A helpful way to remember the sign (−1)j+k of a cofactor is to use the matrix
 
+ − +
− + − .
+ − +
This works not just for 3 × 3 matrices but for any square n × n matrix.
Example 11.4. Compute the determinant of the matrix

 
4 −2 3
A = 2 3 5
1 0 6
91
Determinants
Solution. From the definition of the determinant
det A = a11 C11 + a12 C12 + a13 C13
= (4) det A11 − (−2) det A12 + (3) det A13

3 5 2 5 2 3
= 4 1 6 + 3 1 0
+2
0 6
= 4(3 · 6 − 5 · 0) + 2(2 · 6 − 1 · 5) + 3(2 · 0 − 1 · 3)
= 72 + 14 − 9
= 77
We can compute the determinant of a matrix A by expanding along any row or column.
For example, the expansion of the determinant for the matrix
 
a11 a12 a13
A = a21 a22 a23 
a31 a32 a33
along the 3rd row is

a12 a13 a11 a13 a11 a12
det A = a31 − a32
a21 a23 + a33 a21 a22 .

a22 a23
And along the 2nd column:

a21 a23 a11 a13 a11 a13
det A = −a12
+ a22
− a32
.
a31 a33 a31 a33 a21 a23
The punchline is that any way you choose to expand (row or column) you will get the same
answer. If a particular row or column contains zeros, say entry ajk , then the computation of
the determinant is simplified if you expand along either row j or column k because ajk Cjk = 0
and we need not compute Cjk .
Example 11.5. Compute the determinant of the matrix

 
4 −2 3
A= 2  3 5
1 0 6
Solution. In Example 11.4, we computed det(A) = 77 by expanding along the 1st row.
92
Lecture 11
Notice that a32 = 0. Expanding along the 3rd row:
det A = (1) det A31 − (0) det A32 + (6) det A33

−2 3 4 −2
= +6
3 5 2 3
= 1(−2 · 5 − 3 · 3) + 6(4 · 3 − (−2) · 2)
= −19 + 96
= 77
11.2 Determinants of n × n Matrices

Using the 3 × 3 case as a guide, we define the determinant of a general n × n matrix as
follows.
Definition 11.6: Let A be a n × n matrix. Let Ajk be the (n − 1) × (n − 1) matrix

obtained from A by deleting the jth row and kth column, and let Cjk = (−1)j+k det Ajk
be the (j, k)-cofactor of A. The determinant of A is defined to be
det A = a11 C11 + a12 C12 + · · · + a1n C1n .
The next theorem tells us that we can compute the determinant by expanding along any
row or column.
Theorem 11.7: Let A be a n × n matrix. Then det A may be obtained by a cofactor

expansion along any row or any column of A:
det A = aj1 Cj1 + aj2 Cj2 + · · · + ajn Cjn .
We obtain two immediate corollaries.
Corollary 11.8: If A has a row or column containing all zeros then det A = 0.
Proof. If the jth row contains all zeros then aj1 = aj2 = · · · = ajn = 0:
det A = aj1 Cj1 + aj2 Cj2 + · · · + ajn Cjn = 0.
93
Determinants
Corollary 11.9: For any square matrix A it holds that det A = det AT .
Sketch of the proof. Expanding along the jth row of A is equivalent to expanding along
the jth column of AT .
Example 11.10. Compute the determinant of
 
1 3 0 −2
 1 2 −2 −1 
A=
 0

0 2 1 
−1 −3 1 0
Solution. The third row contains two zeros, so expand along this row:
det A = 0 det A31 − 0 det A32 + 2 det A33 − det A34

1 3 −2 1 3 0

= 2 1 2 −1 − 1 2 −2
−1 −3 0 −1 −3 1

2 −1 1 −1 1 2
= 2 1 − 3 − 2
−3 0 −1 0 −1 −3

2 −2 1 −2
− 1 − 3
−3 1 −1 1
= 2((0 − 3) − 3(0 − 1) − 2(−3 + 2)) − ((2 − 6) − 3(1 − 2))

=5
Example 11.11. Compute the determinant of
 
1 3 0 −2
 1 2 −2 −1 
A=
 0

0 2 1 
−1 −3 1 0
94
Lecture 11
Solution. Expanding along the second row:

det A = − det A21 + 2 det A22 − (−2) det A23 − 1 det A24

3 0 −2 1 0 −2

= − 0 2 1 + 2 0 2 1
−3 1 0 −1 1 0

1 3 −2 1 3 0

+ 2 0 0 1 − 0 0 2
−1 −3 0 −1 −3 1
= −1(−3 − 12) + 2(−1 − 4) + 2(0) − (0)

=5
11.3 Triangular Matrices

Below we introduce a class of matrices for which the determinant computation is trivial.
Definition 11.12: A square matrix A ∈ Rn×n is called upper triangular if ajk = 0

whenever j > k. In other words, all the entries of A below the diagonal entries aii are
zero. It is called lower triangular if ajk = 0 whenever j < k.
For example, a 4 × 4 upper triangular matrix takes the form

 
a11 a12 a13 a14
 0 a22 a23 a24 
A= 0

0 a33 a34 
0 0 0 a44
Expanding along the first column, we compute

a22 a23 a24
a33 a34
det A = a11 0 a33 a34 = a11 a22 = a11 a22 a33 a44 .
0 0 a44

0 a44
The general n × n case is similar and is summarized in the following theorem.
Theorem 11.13: The determinant of a triangular matrix is the product of its diagonal
entries.

• how to compute the determinant of any sized matrix
• that the determinant of A is equal to the determinant of AT
• the determinant of a triangular matrix is the product of its diagonal entries
95
Determinants
96
Lecture 12
Lecture 12
Properties of the Determinant
12.1 ERO and Determinants

Recall that for a matrix A ∈ Rn×n we defined
det A = aj1 Cj1 + aj2 Cj2 + · · · + ajn Cjn
where the number Cjk = (−1)j+k det Ajk is called the (j, k)-cofactor of A and

aj = aj1 aj2 · · · ajn
denotes the jth row of A. Notice that
 
Cj1
 Cj2 


det A = aj1 aj2 · · · ajn  ..  .
 . 
Cjn

If we let cj = Cj1 Cj2 · · · Cjn then
det A = aj · cTj .
In this lecture, we will establish properties of the determinant under elementary row opera-
tions and some consequences. The following theorem describes how the determinant behaves
under elementary row operations of Type 1.
Theorem 12.1: Suppose that A ∈ Rn×n and let B be the matrix obtained by interchang-
ing two rows of A. Then det B = − det A.

a11 a12 a21 a22
Proof. Consider the 2 × 2 case. Let A = and let B = . Then
a21 a22 a11 a12
det B = a12 a21 − a11 a22 = −(a11 a22 − a12 a21 ) = − det A.
The general case is proved by induction.
This theorem leads to the following corollary.
97
Corollary 12.2: If A ∈ Rn×n has two rows (or two columns) that are equal then
det(A) = 0.
Proof. Suppose that A has rows j and k that are equal. Let B be the matrix obtained by
interchanging rows j and k. Then by the previous theorem det B = − det A. But clearly
B = A, and therefore det B = det A. Therefore, det(A) = − det(A) and thus det A = 0.
Now we consider how the determinant behaves under elementary row operations of Type
2.
Theorem 12.3: Let A ∈ Rn×n and let B be the matrix obtained by multiplying a row of
A by β. Then det B = β det A.
Proof. Suppose that B is obtained from A by multiplying the jth row by β. The rows of A
and B different from j are equal, and therefore
Bjk = Ajk , for k = 1, 2, . . . , n.
In particular, the (j, k) cofactors of A and B are equal. The jth row of B is βaj . Then,
expanding det B along the jth row:
det B = (βaj ) · cTj
= β(aj · cTj )
= β det A.
Lastly we consider Type 3 elementary row operations.
Theorem 12.4: Let A ∈ Rn×n and let B be the matrix obtained from A by adding β
times the kth row to the jth row. Then det B = det A.
Proof. For any matrix A and any row vector r = [r1 r2 · · · rn ] the expression
r · cTj = r1 Cj1 + r2 Cj2 + · · · + rn Cjn
is the determinant of the matrix obtained from A by replacing the jth row with the row r.
Therefore, if k 6= j then
ak · cTj = 0
98
Lecture 12
since then rows k and j are equal. The jth row of B is bj = aj + βak . Therefore, expanding
det B along the jth row:
det B = (aj + βak ) · cTj

= aj · cTj + β ak · cTj
= det A.
Example 12.5. Suppose that A is a 4 × 4 matrix and suppose that det A = 11. If B is
obtained from A by interchanging rows 2 and 4, what is det B?
Solution. Interchanging (or swapping) rows changes the sign of the determinant. Therefore,
det B = −11.
Example 12.6. Suppose that A is a 4 × 4 matrix and suppose that det A = 11. Let
a1 , a2 , a3 , a4 denote the rows of A. If B is obtained from A by replacing row a3 by 3a1 + a3 ,
what is det B?
Solution. This is a Type 3 elementary row operation, which preserves the value of the de-
terminant. Therefore,
det B = 11.
a1 , a2 , a3 , a4 denote the rows of A. If B is obtained from A by replacing row a3 by 3a1 + 7a3 ,
what is det B?
Solution. This is not quite a Type 3 elementary row operation because a3 is multiplied by
7. The third row of B is b3 = 3a1 + 7a3 . Therefore, expanding det B along the third row
det B = (3a1 + 7a3 ) · cT3
= 3a1 · cT3 + 7a3 · cT3
= 7(a3 · cT3 )
= 7 det A
= 77
99
a1 , a2 , a3 , a4 denote the rows of A. If B is obtained from A by replacing row a3 by 4a1 + 5a2 ,
what is det B?
Solution. Again, this is not a Type 3 elementary row operation. The third row of B is
b3 = 4a1 + 5a2 . Therefore, expanding det B along the third row
det B = (4a1 + 5a2 ) · cT3
= 4a1 · cT3 + 5a2 · cT3
=0+0
=0
12.2 Determinants and Invertibility of Matrices

The following theorem characterizes invertibility of matrices with the determinant.
Theorem 12.9: A square matrix A is invertible if and only if det A 6= 0.
Proof. Beginning with the matrix A, perform elementary row operations and generate a
sequence of matrices A1 , A2 , . . . , Ap such that Ap is in row echelon form and thus triangular:
A ∼ A1 ∼ A2 ∼ · · · ∼ Ap .
Thus, matrix Ai is obtained from Ai−1 by performing one of the elementary row operations.
From Theorems 12.1, 12.3, 12.4, if det Ai−1 6= 0 then det Ai 6= 0. In particular, det A = 0 if
and only if det Ap = 0. Now, Ap is triangular and therefore its determinant is the product
of its diagonal entries. If all the diagonal entries are non-zero then det A = det Ap 6= 0. In
this case, A is invertible because there are r = n leading entries in Ap . If a diagonal entry
of Ap is zero then det A = det Ap = 0. In this case, A is not invertible because there are
r < n leading entries in Ap . Therefore, A is invertible if and only if det A 6= 0.
12.3 Properties of the Determinant

The following theorem characterizes how the determinant behaves under scalar multiplication
of matrices.
Theorem 12.10: Let A ∈ Rn×n and let B = βA, that is, B is obtained by multiplying
every entry of A by β. Then det B = β n det A.
100
Lecture 12
Proof. Consider the 2 × 2 case:

βa11 βa12
det(βA) =
βa12 βa22
= βa11 · βa22 − βa12 · βa21
= β 2 (a11 a22 − a12 a21 )
= β 2 det A.
Thus, the statement holds for 2 × 2 matrices. Consider a 3 × 3 matrix A. Then
det(βA) = βa11 |βA11| − βa12 |βA12 | + βa13 |βA13 |
= βa11 β 2 |A11 | − βa12 β 2 |A12 | + βa13 β 2 |A13 |
= β 3 (a11 |A11 | − a12 |A12 | + a13 |A13 |)
= β 3 det A.
The general case can be treated using mathematical induction on n.

Example 12.11. Suppose that A is a 4 × 4 matrix and suppose that det A = 11. What is
det(3A)?
Solution. We have
det(3A) = 34 det A
= 81 · 11
= 891
The following theorem characterizes how the determinant behaves under matrix multi-
plication.
Theorem 12.12: Let A and B be n × n matrices. Then
det(AB) = det(A) det(B).
Corollary 12.13: For any square matrix det(Ak ) = (det A)k .
101
Corollary 12.14: If A is invertible then

1
det(A−1 ) = .
det A
Proof. From AA−1 = In we have that det(AA−1) = 1. But also
det(AA−1 ) = det(A) det(A−1 ).
Therefore
det(A) det(A−1 ) = 1
or equivalently
1
det A−1 = .
det A
Example 12.15. Let A, B, C be n × n matrices. Suppose that det A = 3, det B = 0, and

det C = 7.
(i) Is AC invertible?
(ii) Is AB invertible?
(iii) Is ACB invertible?
Solution. (i): We have det(AC) = det A det C = 3 · 7 = 21. Thus, AC is invertible.
(ii): We have det(AB) = det A det B = 3 · 0 = 0. Thus, AB is not invertible.
(iii): We have det(ACB) = det A det C det B = 3·7·0 = 0. Thus, ACB is not invertible.
• how the determinant behaves under elementary row operations
• that A is invertible if and only if det A 6= 0
• that det(AB) = det(A) det(B)
102
Lecture 13
Lecture 13
Applications of the Determinant
13.1 The Cofactor Method

Recall that for A ∈ Rn×n we defined
det A = aj1 Cj1 + aj2 Cj2 + · · · + ajn Cjn
where Cjk = (−1)j+k det Ajk is called the (j, k)-Cofactor of A and

aj = aj1 aj2 · · · ajn

is the jth row of A. If cj = Cj1 Cj2 · · · Cjn then
 
Cj1
 C
 j2 
det A = aj1 aj2 · · · ajn  ..  = aj · cTj .

 . 
Cjn
Suppose that B is the matrix obtained from A by replacing row aj with a distinct row ak .
To compute det B expand along its jth row bj = ak :
det B = ak · cTj = 0.
The Cofactor Method is an alternative method to find the inverse of an invertible matrix.
Recall that for any matrix A ∈ Rn×n , if we expand along the jth row then
det A = aj · cTj .
On the other hand, if j 6= k then

aj · cTk = 0.
In summary, (
det A, if j = k
aj · cTk =
0, if j 6= k.
103
Form the Cofactor matrix

   
C11 C12 · · · C1n c1
 C21 C22 · · · C2n   c2 
 
 
Cof(A) =  .. .. ..  =  ..  .
 . . ··· .   .
Cn1 Cn2 · · · Cnn cn
Then,

a1
 a2 
T  
A(Cof(A)) =  ..  cT1 cT2 · · · cTn
.
an
 
a1 cT1 a1 cT2 · · · a1 cTn
 
 a cT a cT · · · a cT 
 2 1 2 2 2 n
 
= 
 .. .. .. .. 
 . . . . 
 
T T T
an c1 an c2 · · · an cn
 
det A 0 ··· 0
 
 
 0 det A · · · 0 
 
= 
 .. .. .. .. 
 . . . . 
 
0 0 · · · det A
This can be written succinctly as
A(Cof(A))T = det(A)In .
Now if det A 6= 0 then we can divide by det A to obtain

1
A (Cof(A))T = In .
det A
This leads to the following formula for the inverse:
1
A−1 = (Cof(A))T
det A
Although this is an explicit and elegant formula for A−1, it is computationally intensive,
even for 3 × 3 matrices. However, for the 2 × 2 case it provides a useful formula to compute
104
Lecture 13

a b d −c
the matrix inverse. Indeed, if A = we have Cof(A) = and therefore
c d −b a

−1 1 d −b
A = .
ad − bc −c a
When does an integer matrix have an integer inverse? We can answer this question
using the Cofactor Method. Let us first be clear about what we mean by an integer matrix.
Definition 13.1: A matrix A ∈ Rm×n is called an integer matrix if every entry of A is

an integer.
Suppose that A ∈ Rn×n is an invertible integer matrix. Then det(A) is a non-zero integer
and (Cof(A))T is an integer matrix. If A−1 is also an integer matrix then det(A−1 ) is also
an integer. Now det(A) det(A−1) = 1 thus it must be the case that det(A) = ±1. Suppose
on the other hand that det(A) = ±1. Then by the Cofactor method
1
A−1 = (Cof(A))T = ±(Cof(A))T
det(A)
and therefore A−1 is also an integer matrix. We have proved the following.
Theorem 13.2: An invertible integer matrix A ∈ Rn×n has an integer inverse A−1 if and
only if det A = ±1.
We can use the previous theorem to generate integer matrices with an integer inverse
as follows. Begin with an upper triangular matrix M0 having integer entries and whose
diagonal entries are either 1 or −1. By construction, det(M0 ) = ±1. Perform any sequence
of elementary row operations of Type 1 and Type 3. This generates a sequence of matrices
M1 , . . . , Mp whose entries are integers. Moreover,
M0 ∼ M1 ∼ · · · ∼ Mp .
Therefore,
±1 = det(M) = det(M1 ) = · · · = det(Mp ).
105
13.2 Cramer’s Rule

The Cofactor method can be used to give an explicit formula for the solution of a linear
system where the coefficient matrix is invertible. The formula is known as Cramer’s Rule.
To derive this formula, recall that if A is invertible then the solution to Ax = b is x = A−1 b.
Using the Cofactor method, A−1 = det1 A (Cof(A))T , and therefore
  
C11 C21 · · · Cn1 b1
1  12 C22
 C · · · Cn2   b2 
 
x=  . .. .. ..   ..  .
det A  .. . . .  . 
C1n C2n · · · Cnn bn
Consider the first component x1 of x:
1
x1 = (b1 C11 + b2 C21 + · · · + bn Cn1 ).
det A
The expression b1 C11 + b2 C21 + · · · + bn Cn1 is the expansion of the determinant along the
first column of the matrix obtained from A by replacing the first column with b:
 
b1 a12 · · · a1n
 b2 a22 · · · a2n 
 
det  .. .. .. .  = b1 C11 + b2 C21 + · · · + bn Cn1
. . . .. 
bn an2 · · · ann
Similarly,
1
x2 = (b1 C12 + b2 C22 + · · · + bn Cn2 )
det A
and (b1 C12 + b2 C22 + · · · + bn Cn2 ) is the expansion of the determinant along the second
column of the matrix obtained from A by replacing the second column with b. In summary:
Theorem 13.3: (Cramer’s Rule) Let A ∈ Rn×n be an invertible matrix. Let b ∈ Rn

and let Ai be the matrix obtained from A by replacing the ith column with b. Then the
solution to Ax = b is  
det A1
1   det A2 

x=  ..  .
det A  . 
det An
Although this is an explicit and elegant formula for x, it is computationally intensive, and
used mainly for theoretical purposes.
106
Lecture 13
13.3 Volumes
The volume of the parallelepiped determined by the vectors v1 , v2 , v3 is

Vol(v1 , v2 , v3 ) = abs(v1T (v2 × v3 )) = abs(det v1 v2 v3 )
where abs(x) denotes the absolute value of the number x. Let A be an invertible matrix and
let w1 = Av1 , w2 = Av2 , w3 = Av3 . How are Vol(v1 , v2 , v2 ) and Vol(w1 , w2 , w2 ) related?
Compute:

Vol(w1 , w2 , w3 ) = abs(det w1 w2 w3 )

= abs det Av1 Av2 Av3

= abs det(A v1 v2 v3 )

= abs det A · det v1 v2 v3
= abs(det A) · Vol(v1 , v2 , v3 ).
Therefore, the number abs(det A) is the factor by which volume is changed under the linear
transformation with matrix A. In summary:
Theorem 13.4: Suppose that v1 , v2 , v3 are vectors in R3 that determine a parallelepiped

of non-zero volume. Let A be the matrix of a linear transformation and let w1 , w2 , w3 be
the images of v1 , v2 , v3 under A, respectively. Then
Vol(w1 , w2 , w3 ) = abs(det A) · Vol(v1 , v2 , v3 ).
Example 13.5. Consider the data

       
4 1 −1 1 0 −1
A= 2 4 1 , v1 = −1 , v2 = 1 , v3 = 5  .
     
1 1 4 0 2 1
and let w1 = Av1 , w2 = Av2 , and w3 = Av3 . Find the volume of the parallelepiped
spanned by the vectors {w1 , w2 , w3 }.

Vol(v1 , v2 , v3 ) = abs(det( v1 v2 v3 )) = abs(−7) = 7
We compute:
det(A) = 55.
Therefore, the volume of the parallelepiped spanned by the vectors {w1 , w2 , w3 } is
Vol(w1 , w2 , w3 ) = abs(55) × 7 = 385.
107

• what the Cofactor Method is
• what Cramer’s Rule is
• the geometric interpretation of the determinant (volume)
108
Lecture 14
Vector Spaces
14.1 Vector Spaces

When you read/hear the word vector you may immediately think of two points in R2 (or
R3 ) connected by an arrow. Mathematically speaking, a vector is just an element of a
vector space. This then begs the question: What is a vector space? Roughly speaking,
a vector space is a set of objects that can be added and multiplied by scalars. You
have already worked with several types of vector spaces. Examples of vector spaces that you
have already encountered are:
1. the set Rn ,
2. the set of all n × n matrices,
3. the set of all functions from [a, b] to R, and
4. the set of all sequences.
In all of these sets, there is an operation of “addition“ and “multiplication by scalars”. Let’s
formalize then exactly what we mean by a vector space.
Definition 14.1: A vector space is a set V of objects, called vectors, on which two
operations called addition and scalar multiplication have been defined satisfying the
following properties. If u, v, w are in V and if α, β ∈ R are scalars:
(1) The sum u + v is in V. (closure under addition)
(2) u + v = v + u (addition is commutative)
(3) (u + v) + w = u + (v + w) (addition is associativity)
(4) There is a vector in V called the zero vector, denoted by 0, satisfying v + 0 = v.
(5) For each v there is a vector −v in V such that v + (−v) = 0.

Vector Spaces
(6) The scalar multiple of v by α, denoted αv, is in V. (closure under scalar multiplica-
tion)
(7) α(u + v) = αu + αv
(8) (α + β)v = αv + βv
(9) α(βv) = (αβ)v
(10) 1v = v
It can be shown that 0 · v = 0 for any vector v in V. To better understand the definition of
a vector space, we first consider a few elementary examples.
Example 14.2. Let V be the unit disc in R2 :

V = {(x, y) ∈ R2 | x2 + y 2 ≤ 1}
Is V a vector space?
Solution. The circle is not closed under scalar multiplication. For example, take u = (1, 0) ∈
V and multiply by say α = 2. Then αu = (2, 0) is not in V. Therefore, property (6) of the
definition of a vector space fails, and consequently the unit disc is not a vector space.
Example 14.3. Let V be the graph of the quadratic function f (x) = x2 :
n o
V = (x, y) ∈ R2 | y = x2 .
Solution. The set V is not closed under scalar multiplication. For example, u = (1, 1) is a
point in V but 2u = (2, 2) is not. You may also notice that V is not closed under addition
either. For example, both u = (1, 1) and v = (2, 4) are in V but u + v = (3, 5) and (3, 5) is
not a point on the parabola V. Therefore, the graph of f (x) = x2 is not a vector space.
110
Lecture 14
Example 14.4. Let V be the graph of the function f (x) = 2x:
V = {(x, y) ∈ R2 | y = 2x}.
Solution. We will show that V is a vector space. First, we verify that V is closed under
addition. We first note that an arbitrary point in V can be written as u = (x, 2x). Let then
u = (a, 2a) and v = (b, 2b) be points in V. Then
u + v = (a + b, 2a + 2b) = (a + b, 2(a + b)).
Therefore V is closed under addition. Verify that V is closed under scalar multiplication:
αu = α(a, 2a) = (αa, α2a) = (αa, 2(αa)).
Therefore V is closed under scalar multiplication. There is a zero vector 0 = (0, 0) in V:
u + 0 = (a, 2a) + (0, 0) = (a, 2a).
All the other properties of a vector space can be verified to hold; for example, addition is
commutative and associative in V because addition in R2 is commutative/associative, etc.
Therefore, the graph of the function f (x) = 2x is a vector space.
The following example is important (it will appear frequently) and is our first example
of what we could say is an “abstract vector space”. To emphasize, a vector space is a set
that comes equipped with an operation of addition and scalar multiplication and these two
operations satisfy the list of properties above.
Example 14.5. Let V = Pn [t] be the set of all polynomials in the variable t and of degree
at most n: n o
Pn [t] = a0 + a1 t + a2 t2 + · · · + an tn | a0 , a1 , . . . , an ∈ R .
Solution. Let u(t) = u0 + u1 t + · · · + un tn and let v(t) = v0 + v1 t + · · · + vn tn be polynomials

in V. We define the addition of u and v as the new polynomial (u + v) as follows:
(u + v)(t) = u(t) + v(t) = (u0 + v0 ) + (u1 + v1 )t + · · · + (un + vn )tn .
111
Vector Spaces
Then u + v is a polynomial of degree at most n and thus (u + v) ∈ Pn [t], and therefore this
shows that Pn [t] is closed under addition. Now let α be a scalar, define a new polynomial
(αu) as follows:
(αu)(t) = (αu0 ) + (αu1 )t + · · · + (αun )tn
Then (αu) is a polynomial of degree at most n and thus (αu) ∈ Pn [t]; hence, Pn [t] is closed
under scalar multiplication. The 0 vector in Pn [t] is the zero polynomial 0(t) = 0. One can
verify that all other properties of the definition of a vector space also hold; for example,
addition is commutative and associative, etc. Thus Pn [t] is a vector space.
Example 14.6. Let V = Mm×n be the set of all m × n matrices. Under the usual operations
of addition of matrices and scalar multiplication, is Mn×m a vector space?
Solution. Given matrices A, B ∈ Mm×n and a scalar α, we defined the sum A + B by adding
entry-by-entry, and αA by multiplying each entry of A by α. It is clear that the space
Mm×n is closed under these two operations. The 0 vector in Mm×n is the matrix of size
m × n having all entries equal to zero. It can be verified that all other properties of the
definition of a vector space also hold. Thus, the set Mm×n is a vector space.
Example 14.7. The n-dimensional Euclidean space V = Rn under the usual operations of
addition and scalar multiplication is vector space.
Example 14.8. Let V = C[a, b] denote the set of functions with domain [a, b] and co-domain
R that are continuous. Is V a vector space?
14.2 Subspaces of Vector Spaces

Frequently, one encounters a vector space W that is a subset of a larger vector space V. In
this case, we would say that W is a subspace of V. Below is the formal definition.
Definition 14.9: Let V be a vector space. A subset W of V is called a subspace of V

if it satisfies the following properties:
(1) The zero vector of V is also in W.
(2) W is closed under addition, that is, if u and v are in W then u + v is in W.
(3) W is closed under scalar multiplication, that is, if u is in W and α is a scalar then
αu is in W.
Example 14.10. Let W be the graph of the function f (x) = 2x:
W = {(x, y) ∈ R2 | y = 2x}.
Is W a subspace of V = R2 ?
112
Lecture 14
Solution. If x = 0 then y = 2 · 0 = 0 and therefore 0 = (0, 0) is in W. Let u = (a, 2a) and

v = (b, 2b) be elements of W. Then
u + v = (a, 2a) + (b, 2b) = (a + b, 2a + 2b) = (a + }b, 2 (a + b)).
| {z | {z }
x x
Because the x and y components of u + v satisfy y = 2x then u + v is inside in W. Thus, W

is closed under addition. Let α be any scalar and let u = (a, 2a) be an element of W. Then
αa , 2 (αa))
αu = (αa, α2a) = (|{z}
|{z}
x x
Because the x and y components of αu satisfy y = 2x then αu is an element of W, and thus

W is closed under scalar multiplication. All three conditions of a subspace are satisfied for
W and therefore W is a subspace of V.
Example 14.11. Let W be the first quadrant in R2 :

W = {(x, y) ∈ R2 | x ≥ 0, y ≥ 0}.
Is W a subspace?
Solution. The set W contains the zero vector and the sum of two vectors in W is again in
W; you may want to verify this explicitly as follows: if u1 = (x1 , y1 ) is in W then x1 ≥ 0
and y1 ≥ 0, and similarly if u2 = (x2 , y2) is in W then x2 ≥ 0 and y2 ≥ 0. Then the sum
u1 + u2 = (x1 + x2 , y1 + y2 ) has components x1 + y1 ≥ 0 and x2 + y2 ≥ 0 and therefore u1 + u2
is in W. However, W is not closed under scalar multiplication. For example if u = (1, 1)
and α = −1 then αu = (−1, −1) is not in W because the components of αu are clearly not
non-negative.
Example 14.12. Let V = Mn×n be the vector space of all n × n matrices. We define the
trace of a matrix A ∈ Mn×n as the sum of its diagonal entries:
tr(A) = a11 + a22 + · · · + ann .
Let W be the set of all n × n matrices whose trace is zero:
W = {A ∈ Mn×n | tr(A) = 0}.
Is W a subspace of V?
Solution. If 0 is the n × n zero matrix then clearly tr(0) = 0, and thus 0 ∈ Mn×n . Suppose
that A and B are in W. Then necessarily tr(A) = 0 and tr(B) = 0. Consider the matrix
C = A + B. Then
tr(C) = tr(A + B) = (a11 + b11 ) + (a22 + b22 ) + · · · + (ann + bnn )
= (a11 + · · · + ann ) + (b11 + · · · + bnn )
= tr(A) + tr(B)
=0
113
Vector Spaces
Therefore, tr(C) = 0 and consequently C = A + B ∈ W, in other words, W is closed under

addition. Now let α be a scalar and let C = αA. Then
tr(C) = tr(αA) = (αa11 ) + (αa22 ) + · · · + (αann ) = α tr(A) = 0.
Thus, tr(C) = 0, that is, C = αA ∈ W, and consequently W is closed under scalar multipli-
cation. Therefore, the set W is a subspace of V.
Example 14.13. Let V = Pn [t] and consider the subset W of V:
W = {u ∈ Pn [t] | u′ (1) = 0}
In other words, W consists of polynomials of degree n in the variable t whose derivative at

t = 1 is zero. Is W a subspace of V?
Solution. The zero polynomial 0(t) = 0 clearly has derivative at t = 1 equal to zero, that is,
0′ (1) = 0, and thus the zero polynomial is in W. Now suppose that u(t) and v(t) are two
polynomials in W. Then, u′ (1) = 0 and also v′ (1) = 0. To verify whether or not W is closed
under addition, we must determine whether the sum polynomial (u + v)(t) has a derivative
at t = 1 equal to zero. From the rules of differentiation, we compute
(u + v)′ (1) = u′ (1) + v′ (1) = 0 + 0.
Therefore, the polynomial (u + v) is in W, and thus W is closed under addition. Now let α
be any scalar and let u(t) be a polynomial in W. Then u′ (1) = 0. To determine whether or
not the scalar multiple αu(t) is in W we must determine if αu(t) has a derivative of zero at
t = 1. Using the rules of differentiation, we compute that
(αu)′ (1) = αu′ (1) = α · 0 = 0.
Therefore, the polynomial (αu)(t) is in W and thus W is closed under scalar multiplication.
All three properties of a subspace hold for W and therefore W is a subspace of Pn [t].
Example 14.14. Let V = Pn [t] and consider the subset W of V:
W = {u ∈ Pn [t] | u(2) = −1}
In other words, W consists of polynomials of degree n in the variable t whose value t = 2 is

−1. Is W a subspace of V?
Solution. The zero polynomial 0(t) = 0 clearly does not equal −1 at t = 2. Therefore, W
does not contain the zero polynomial and, because all three conditions of a subspace must be
satisfied for W to be a subspace, then W is not a subspace of Pn [t]. As an exercise, you may
want to investigate whether or not W is closed under addition and scalar multiplication.
114
Lecture 14
Example 14.15. A square matrix A is said to be symmetric if AT = A. For example,

here is a 3 × 3 symmetric matrix:
 
1 2 −3
A= 2
 4 5
−3 5 7
Verify for yourself that we do indeed have that AT = A. Let W be the set of all symmetric
n × n matrices. Is W a subspace of V = Mn×n ?
Example 14.16. For any vector space V, there are two trivial subspaces in V, namely, V
itself is a subspace of V and the set consisting of the zero vector W = {0} is a subspace of
V.
There is one particular way to generate a subspace of any given vector space V using the
span of a set of vectors. Recall that we defined the span of a set of vectors in Rn but we can
define the same notion on a general vector space V.
Definition 14.17: Let V be a vector space and let v1 , v2 , . . . , vp be vectors in V. The

span of {v1 , . . . , vp } is the set of all linear combinations of v1 , . . . , vp :
n o
span{v1 , v2 , . . . , vp } = t1 v1 + t2 v2 + · · · + vp vp | t1 , t2 , . . . , tp ∈ R .
We now show that the span of a set of vectors in V is a subspace of V.
Theorem 14.18: If v1 , v2 , . . . , vp are vectors in V then span{v1 , . . . , vp } is a subspace of

V.
Solution. Let u = t1 v1 +· · ·+tp vp and w = s1 v1 +· · ·+sp vp be two vectors in span{v1 , v2 , . . . , vp }.

Then
u + w = (t1 v1 + · · · + tp vp ) + (s1 v1 + · · · + sp vp ) = (t1 + s1 )v1 + · · · + (tp + sp )vp .
Therefore u + w is also in the span of v1 , . . . , vp . Now consider αu:
αu = α(t1 v1 + · · · + tp vp ) = (αt1 )v1 + · · · + (αtp )vp .
Therefore, αu is in the span of v1 , . . . , vp . Lastly, since 0v1 + 0v2 + · · · + 0vp = 0 then the
zero vector 0 is in the span of v1 , v2 , . . . , vp . Therefore, span{v1 , v2 , . . . , vp } is a subspace
of V.
Given a general subspace W of V, if w1 , w2 , . . . , wp are vectors in W such that
span{w1 , w2 , . . . , wp } = W
then we say that {w1 , w2 , . . . , wp } is a spanning set of W. Hence, every vector in W can
be written as a linear combination of the vectors w1 , w2 , . . . , wp .
115
Vector Spaces
• what a vector space/subspace is
• be able to give some examples of vector spaces/subspaces
• that the span of a set of vectors in V is a subspace of V
116
Lecture 15
Lecture 15
Linear Maps
Before we begin this Lecture, we review subspaces. Recall that W is a subspace of a vector
space V if W is a subset of V and
1. the zero vector 0 in V is also in W,
2. for any vectors u, v in W the sum u + v is also in W, and
3. for any vector u in W and any scalar α the vector αu is also in W.
In the previous lecture we gave several examples of subspaces. For example, we showed that
a line through the origin in R2 is a subspace of R2 and we gave examples of subspaces of
Pn [t] and Mn×m . We also showed that if v1 , . . . , vp are vectors in a vector space V then
W = span{v1 , v2 , . . . , vp }
is a subspace of V.
15.1 Linear Maps on Vector Spaces

In Lecture 7, we defined what it meant for a vector mapping T : Rn → Rm to be a linear
mapping. We now want to introduce linear mappings on general vector spaces; you will
notice that the definition is essentially the same but the key point to remember is that the
underlying spaces are not Rn but a general vector space.
Definition 15.1: Let T : V → U be a mapping of vector spaces. Then T is called a linear

mapping if
• for any u, v in V it holds that T(u + v) = T(u) + T(v), and
• for any scalar α and u in V is holds that T(αv) = αT(v).
Example 15.2. Let V = Mn×n be the vector space of n × n matrices and let T : V → V be
the mapping
T(A) = A + AT .
117
Linear Maps
Is T is a linear mapping?
Solution. Let A and B be matrices in V. Then using the properties of the transpose and
regrouping we obtain:
T(A + B) = (A + B) + (A + B)T
= A + B + AT + BT
= (A + AT ) + (B + BT )
= T(A) + T(B).
Similarly, if α is any scalar then
T(αA) = (αA) + (αA)T

= αA + αAT
= α(A + AT )
= αT(A).
This proves that T satisfies both conditions of Definition 15.1 and thus T is a linear mapping.
Example 15.3. Let V = Mn×n be the vector space of n × n matrices, where n ≥ 2, and let
T : V → R be the mapping
T(A) = det(A)
Solution. If T is a linear mapping then according to Definition 15.1, we must have T(A +
B) = det(A + B) = det(A) + det(B) and also T(αA) = αT(A) for any scalar α. Do
these properties actually hold though? For example, we know from the properties of the
determinant that det(αA) = αn det(A) and therefore it does not hold that T(αA) = αT(A)
unless α = 1. Therefore, T is not a linear mapping. Also, it does not hold in general that
det(A + B) = det(A) + det(B); in fact it rarely holds. For example, if

2 0 −1 1
A= , B=
0 1 0 3
then det(A) = 2, det(B) = −3 and therefore det(A) + det(B) = −1. On the other hand,

1 1
A+B=
0 4
and thus det(A + B) = 4. Thus, det(A + B) 6= det(A) + det(B).
Example 15.4. Let V = Pn [t] be the vector space of polynomials in the variable t of degree
no more than n ≥ 1. Consider the mapping T : V → V define as
T(f (t)) = 2f (t) + f ′ (t).
118
Lecture 15
For example, if f (t) = 3t6 − t2 + 5 then
T(f (t)) = 2f (t) + f ′ (t)

= 2(3t5 − t2 + 5) + (18t5 − 2t)
= 6t5 + 18t5 − 2t2 − 2t + 10.
Solution. Let f (t) and g(t) be polynomials of degree no more than n ≥ 1. Then
T(f (t) + g(t)) = 2(f (t) + g(t)) + (f (t) + g(t))′

= 2f (t) + 2g(t) + f ′ (t) + g ′ (t)
= (2f (t) + f ′ (t)) + (2g(t) + g ′ (t))
= T(f (t)) + T(g(t)).
Therefore, T(f (t) + g(t)) = T(f (t)) + T(g(t)). Now let α be any scalar. Then
T(αf (t)) = 2(αf (t)) + (αf (t))′

= 2αf (t) + αf ′ (t)
= α(2f (t) + f ′ (t))
= αT(f (t)).
Therefore, T(αf (t)) = αT(f (t)). Therefore, T is a linear mapping.

We now introduce two important subsets associated to a linear mapping.
Definition 15.5: Let T : V → U be a linear mapping.
1. The kernel of T is the set of vectors v in the domain V that get mapped to the zero
vector, that is, T(v) = 0. We denote the kernel of T by ker(T):
ker(T) = {v ∈ V | T(v) = 0}.
2. The range of T is the set of vectors b in the codomain U for which there exists at
least one v in V such that T(v) = b. We denote the range of T by Range(T):
Range(T) = {b ∈ U | there exists some v ∈ U such that T(v) = b}.
You may have noticed that the definition of the range of a linear mapping on an abstract
vector space is the usual definition of the range of a function. Not surprisingly, the kernel
and range are subspaces of the domain and codomain, respectively.
119
Linear Maps
Theorem 15.6: Let T : V → U be a linear mapping. Then ker(T) is a subspace of V and

Range(T) is a subspace of U.
Proof. Suppose that v and u are in ker(T). Then T(v) = 0 and T(u) = 0. Then by linearity
of T it holds that
T(v + u) = T(v) + T(u) = 0 + 0 = 0.
Therefore, since T(u + v) = 0 then u + v is in ker(T). This shows that ker(T) is closed
under addition. Now suppose that α is any scalar and v is in ker(T). Then T(v) = 0 and
thus by linearity of T it holds that
T(αv) = αT(v) = α0 = 0.
Therefore, since T(αv) = 0 then αv is in ker(T) and this proves that ker(T) is closed under
scalar multiplication. Lastly, by linearity of T it holds that
T(0) = T(v − v) = T(v) − T(v) = 0
that is, T(0) = 0. Therefore, the zero vector 0 is in ker(T). This proves that ker(T) is a
subspace of V. The proof that Range(T) is a subspace of U is left as an exercise.
Example 15.7. Let V = Mn×n be the vector space of n × n matrices and let T : V → V be
the mapping
T(A) = A + AT .
Describe the kernel of T.
Solution. A matrix A is in the kernel of T if T(A) = A + AT = 0, that is, if AT = −A.
Hence,
ker(A) = {A ∈ Mn×n | AT = −A}.
What type of matrix A satisfies AT = −A? For example, consider the case that A is the
2 × 2 matrix
a11 a12
A=
a21 a22
and AT = −A. Then
a11 a21 −a11 −a12
= .
a12 a22 −a21 −a22
Therefore, it must hold that a11 = −a11 , a21 = −a12 and a22 = −a22 . Then necessarily
a11 = 0 and a22 = 0 and a12 can be arbitrary. For example, the matrix

0 7
A=
−7 0
satisfies AT = −A. Using a similar computation as above, a 3 × 3 matrix satisfies AT = −A
if A is of the form  
0 a b
A = −a 0 c 
−b −c 0
120
Lecture 15
where a, b, c are arbitrary constants. In general, a matrix A that satisfies AT = −A is called

skew-symmetric.
Example 15.8. Let V be the vector space of differentiable functions on the interval [a, b].
That is, f is an element of V if f : [a, b] → R is differentiable. Describe the kernel of the
linear mapping T : V → V defined as
T(f (x)) = f (x) + f ′ (x).
Solution. A function f is in the kernel of T if T(f (x)) = 0, that is, if f (x) + f ′ (x) = 0.
Equivalently, if f ′ (x) = −f (x). What functions f do you know of satisfy f ′ (x) = −f (x)?
How about f (x) = e−x ? It is clear that f ′ (x) = −e−x = −f (x) and thus f (x) = e−x is in
ker(T). How about g(x) = 2e−x ? We compute that g ′(x) = −2e−x = −g(x) and thus g is
also in ker(T). It turns out that the elements of ker(T) are of the form f (x) = Ce−x for a
constant C.
15.2 Null space and Column space

In the previous section, we introduced the kernel and range of a general linear mapping
T : V → U. In this section, we consider the particular case of matrix mappings TA : Rn → Rm
for some m×n matrix A. In this case, v is in the kernel of TA if and only if TA(v) = Av = 0.
In other words, v ∈ ker(TA) if and only if v is a solution to the homogeneous system Ax = 0.
Because the case when T is a matrix mapping arises so frequently, we give a name to the set
of vectors v such that Av = 0.
Definition 15.9: The null space of a matrix A ∈ Mm×n , denoted by Null(A), is the
subset of Rn consisting of vectors v such that Av = 0. In other words, v ∈ Null(A) if
and only if Av = 0. Using set notation:
Null(A) = {v ∈ Rn | Av = 0}.
Hence, the following holds
ker(TA) = Null(A).
Because the kernel of a linear mapping is a subspace we obtain the following.
Theorem 15.10: If A ∈ Mm×n then Null(A) is a subspace of Rn .
Hence, by Theorem 15.10, if u and v are two solutions to the linear system Ax = 0 then
αu + βv is also a solution:
A(αu + βv) = αAu + βAv = α · 0 + β · 0 = 0.
121
Linear Maps
Example 15.11. Let V = R4 and consider the following subset of V:
W = {(x1 , x2 , x3 , x4 ) ∈ R4 | 2x1 − 3x2 + x3 − 7x4 = 0}.
Is W a subspace of V?
Solution. The set W is the null space of the matrix 1 × 4 matrix A given by

A = 2 −3 1 −7 .
Hence, W = Null(A) and consequently W is a subspace.

From our previous remarks, the null space of a matrix A ∈ Mm×n is just the solution set
of the homogeneous system Ax = 0. Therefore, one way to explicitly describe the null space
of A is to solve the system Ax = 0 and write the general solution in parametric vector form.
From our previous work on solving linear systems, if the rref(A) has r leading 1’s then the
number of parameters in the solution set is d = n − r. Therefore, after performing back
substitution, we will obtain vectors v1 , . . . , vd such that the general solution in parametric
vector form can be written as
x = t1 v1 + t2 v2 + · · · + td vd
where t1 , t2 , . . . , td are arbitrary numbers. Therefore,
Null(A) = span{v1 , v2 , . . . , vd }.
Hence, the vectors v1 , v2 , . . . , vn form a spanning set for Null(A).
Example 15.12. Find a spanning set for the null space of the matrix
 
−3 6 −1 1 −7
A = 1 −2
 2 3 −1 .
2 −4 5 8 −4
Solution. The null space of A is the solution set of the homogeneous system Ax = 0.
Performing elementary row operations one obtains
 
1 −2 0 −1 3
A ∼ 0 0 1 2 −2 .
0 0 0 0 0
Clearly r = rank(A) and since n = 5 we will have d = 3 vectors in a spanning set for
Null(A). Letting x5 = t1 , and x4 = t2 , then from the 2nd row we obtain
x3 = −2t2 + 2t1 .
Letting x2 = t3 , then from the 1st row we obtain
x1 = 2t3 + t2 − 3t1 .
122
Lecture 15
Writing the general solution in parametric vector form we obtain

     
−3 1 2
0 0 1
     
x = t1  2
 
 + t2 −2 + t3 0
   
0 1 0
1 0 0
Therefore,  

 

       



 −3 1 2 

 
 0   0  1
       
Null(A) = span  2 , −2 0
     

  0   1  0

 

 


 1 0 0 


| {z } | {z } |{z}
 
v1 v2 v3
You can verify that Av1 = Av2 = Av3 = 0.
Now we consider the range of a matrix mapping TA : Rn → Rm . Recall that a vector

b in the co-domain Rm is in the range of TA if there exists some vector x in the domain
Rn such
that TA (x) = b. Since, TA (x) = Ax then Ax = b. Now, if A has columns
A = v1 v2 · · · vn and x = (x1 , x2 , . . . , xn ) then recall that
Ax = x1 v1 + x2 v2 + · · · + xn vn
and thus Ax = x1 v1 + x2 v2 + · · · + xn vn = b. Thus, a vector b is in the range of A if it can

be written as a linear combination of the columns v1 , v2 , . . . , vn of A. This motivates the
following definition.
Definition 15.13: Let A ∈ Mm×n be a matrix. The span of the columns of A is called
the column
space of A. The column space of A is denoted by Col(A). Explicitly, if
A = v1 v2 · · · vn then
Col(A) = span{v1 , v2 , . . . , vn }.
In summary, we can write that
Range(TA ) = Col(A).
and since Range(TA) is a subspace of Rm then so is Col(A).
Theorem 15.14: The column space of a m × n matrix is a subspace of Rm .
123
Linear Maps
Example 15.15. Let

   
2 4 −2 1 3
A = −2 −5 7 3 , b = −1 .
3 7 −8 6 3
Is b in the column space Col(A)?

Solution. The vector b is in the column space of A if there exists x ∈ R4 such that Ax = b.
Hence, we must determineif Ax= b has a solution. Performing elementary row operations
on the augmented matrix A b we obtain
 
2 4 −2 1 3
A b ∼ 0 1 −5 −4 −2
0 0 0 17 1
The system is consistent and therefore Ax = b will have a solution. Therefore, b is in

Col(A).
• what the null space of a matrix is and how to compute it
• what the column space of a matrix is and how to determine if a given vector is in the
column space
• what the range and kernel of a linear mapping is
124
Lecture 16
Linear Independence, Bases, and

Dimension
16.1 Linear Independence

Roughly speaking, the concept of linear independence evolves around the idea of working
with “efficient” spanning sets for a subspace. For instance, the set of directions
{EAST, NORTH, NORTH-EAST}
are redundant since a total displacement in the NORTH-EAST direction can be obtained
by combining individual NORTH and EAST displacements. With these vague statements
out of the way, we introduce the formal definition of what it means for a set of vectors to be
“efficient”.
Definition 16.1: Let V be a vector space and let {v1 , v2 , . . . , vp } be a set of vectors in
V. Then {v1 , v2 , . . . , vp } is linearly independent if the only scalars c1 , c2 , . . . , cp that
satisfy the equation
c1 v1 + c2 v2 + · · · + cp vp = 0
are the trivial scalars c1 = c2 = · · · = cp = 0. If the set {v1 , . . . , vp } is not linearly
independent then we say that it is linearly dependent.
We now describe the redundancy in a set of linear dependent vectors. If {v1 , . . . , vp } are
linearly dependent, it follows that there are scalars c1 , c2 , . . . , cp , at least one of which is
nonzero, such that
c1 v1 + c2 v2 + · · · + cp vp = 0. (⋆)
For example, suppose that {v1 , v2 , v3 , v4 } are linearly dependent. Then there are scalars
c1 , c2 , c3 , c4 , not all of them zero, such that equation (⋆) holds. Suppose, for the sake of
argument, that c3 6= 0. Then,
c1 c2 c4
v3 = − v1 − v2 − v4 .
c3 c3 c3
Linear Independence, Bases, and Dimension
Therefore, when a set of vectors is linearly dependent, it is possible to write one of the vec-
tors as a linear combination of the others. It is in this sense that a set of linearly dependent
vectors are redundant. In fact, if a set of vectors are linearly dependent we can say even
more as the following theorem states.
Theorem 16.2: A set of vectors {v1 , v2 , . . . , vp }, with v1 6= 0, is linearly dependent if

and only if some vj is a linear combination of the preceding vectors v1 , . . . , vj−1 .
Example 16.3. Show that the following set of 2 × 2 matrices is linearly dependent:

1 2 −1 3 5 0
A1 = , A2 = , A3 = .
0 −1 1 0 −2 −3
Solution. It is clear that A1 and A2 are linearly independent, i.e., A1 cannot be written as
a scalar multiple of A2 , and vice-versa. Since the (2, 1) entry of A1 is zero, the only way to
get the −2 in the (2, 1) entry of A3 is to multiply A2 by −2. Similary, since the (2, 2) entry
of A2 is zero, the only way to get the −3 in the (2, 2) entry of A3 is to multiply A1 by 3.
Hence, we suspect that 3A1 − 2A2 = A3 . Verify:

3 6 −2 6 5 0
3A1 − 2A2 = − = = A3
0 −3 2 0 −2 −3
Therefore, 3A1 − 2A2 − A3 = 0 and thus we have found scalars c1 , c2 , c3 not all zero such
that c1 A1 + c2 A2 + c3 A3 = 0.
16.2 Bases
We now introduce the important concept of a basis. Given a set of vectors {v1 , . . . , vp−1 , vp }
in V, we showed that W = span{v1 , v2 , . . . , vp } is a subspace of V. If say vp is linearly
dependent on v1 , v2 , . . . , vp−1 then we can remove vp and the smaller set {v1 , . . . , vp−1 } still
spans all of W:
W = span{v1 , v2 , . . . , vp−1 , vp } = span{v1 , . . . , vp−1 }.
Intuitively, vp does not provide an independent “direction” in generating W. If some other

vector vj is linearly dependent on v1 , . . . , vp−1 then we can remove vj and the resulting
smaller set of vectors still spans W. We can continue removing vectors until we obtain a
minimal set of vectors that are linearly independent and still span W. The following remarks
motivate the following important definition.
Definition 16.4: Let W be a subspace of a vector space V. A set of vectors B = {v1 , . . . , vk }

in W is said to be a basis for W if
(a) the set B spans all of W, that is, W = span{v1 , . . . , vk }, and
126
Lecture 16
(b) the set B is linearly independent.
A basis is therefore a minimal spanning set for a subspace. Indeed, if B = {v1 , . . . , vp }

is a basis for W and we remove say vp , then B̃ = {v1 , . . . , vp−1 } cannot be a basis for W.
Why? If B = {v1 , . . . , vp } is a basis then it is linearly independent and therefore vp cannot
be written as a linear combination of the others. In other words, vp ∈ W is not in the span of
B̃ = {v1 , . . . , vp−1 } and therefore B̃ is not a basis for W because a basis must be a spanning
set. If, on the other hand, we start with a basis B = {v1 , . . . , vp } for W and we add a new
vector u from W then B̃ = {v1 , . . . , vp , u} is not a basis for W. Why? We still have that
span B̃ = W but now B̃ is not linearly independent. Indeed, because B = {v1 , . . . , vp } is a
basis for W, the vector u can be written as a linear combination of {v1 , . . . , vp }, and thus B̃
is not linearly independent.
Example 16.5. Show that the standard unit vectors form a basis for V = R3 :
     
1 0 0
e1 = 0 , e2 = 1 , e3 = 0
0 0 1
Solution. Any vector x ∈ R3 can be written as a linear combination of e1 , e2 , e3 :

       
x1 1 0 0
x = x2  = x1 0 + x2 1 + x3 0 = x1 e1 + x2 e2 + x3 e3
x3 0 0 1
Therefore, span{e1 , e2 , e3 } = R3 . The set B = {e1 , e2 , e3 } is linearly independent. Indeed, if

there are scalars c1 , c2 , c3 such that
c1 e1 + c2 e2 + c3 e3 = 0
then clearly they must all be zero, c1 = c2 = c3 = 0. Therefore, by definition, B = {e1 , e2 , e3 }

is a basis for R3 . This basis is called the standard basis for R3 . Analogous arguments hold
for {e1 , e2 , . . . , en } in Rn .
Example 16.6. Is B = {v1 , v2 , v3 } a basis for R3 ?

     
2 −4 4
v1 =  0  , v2 = −2 , v3 = −6
−4 8 −6
Solution. Form the matrix A = [v1 v2 v3 ] and row reduce:

 
1 0 0
 
A∼  0 1 0 

0 0 1
127
Therefore, the only solution to Ax = 0 is the trivial solution.

Therefore,
B is linearly inde-
3
pendent. Moreover, for any b ∈ R , the augmented matrix A b is consistent. Therefore,
the columns of A span all of R3 :
Col(A) = span{v1 , v2 , v3 } = R3 .
Therefore, B is a basis for R3 .
Example 16.7. In V = R4 , consider the vectors

     
1 2 −1
 3 −1  4
v1 =  0 , v2 = −2 , v3 =  2 .
    
−2 1 −3
Let W = span{v1 , v2 , v3 }. Is B = {v1 , v2 , v3 } a basis for W?

Solution. By definition, B is a spanning set for W, so we need only determine if B is linearly
independent. Form the matrix, A = [v1 v2 v3 ] and row reduce to obtain
 
1 0 1
0 1 −1
A∼ 0 0 0 

0 0 0
Hence, rank(A) = 2 and thus B is linearly dependent. Notice v1 − v2 = v3 . Therefore, B is

not a basis of W.
Example 16.8. Find a basis for the vector space of 2 × 2 matrices.
Example 16.9. Recall that a n × n is skew-symmetric A if AT = −A. We proved that

the set of n × n matrices is a subspace. Find a basis for the set of 3 × 3 skew-symmetric
matrices.
16.3 Dimension of a Vector Space

The following theorem will lead to the definition of the dimension of a vector space.
Theorem 16.10: Let V be a vector space. Then all bases of V have the same number of
vectors.
Proof: We will prove the theorem for the case that V = Rn . We already know that the
standard unit vectors {e1 , e2 , . . . , en } is a basis of Rn . Let {u1 , u2 , . . . , up } be nonzero vec-
tors in Rn and suppose first that p > n. In Lecture 6, Theorem 6.7, we proved that any set
of vectors in Rn containing more than n vectors is automatically linearly dependent. The
reason is that the RREF of A = u1 u2 · · · up will contain at most r = n leading ones,
128
Lecture 16
and therefore d = p − n > 0. Therefore, the solution set of Ax = 0 contains non-trivial

solutions. On the other hand, suppose instead that p < n. In Lecture 4, Theorem 4.11, we
proved that a set of vectors {u1 , . . . , up } in Rn spans Rn if and only if the RREF of A has
exactly r = n leading ones. The largest possible value of r is r = p < n. Therefore, if p < n
then {u1 , u2 , . . . , up } cannot be a basis for Rn . Thus, in either case (p > n or p < n), the set
{u1 , u2 , . . . , up } cannot be a basis for Rn . Hence, any basis in Rn must contain n vectors.
The previous theorem does not say that every set {v1 , v2 , . . . , vn } of nonzero vectors in
R containing n vectors is automatically a basis for Rn . For example,
n
     
1 0 2
v1 = 0 , v2 = 1 , v3 = 3
0 0 0
do not form a basis for R3 because

 
0
x =  0
1
is not in the span of {v1 , v2 , v3 }. All that we can say is that a set of vectors in Rn containing
fewer or more than n vectors is automatically not a basis for Rn . From Theorem 16.10, any
basis in Rn must have exactly n vectors. In fact, on a general abstract vector space V, if
{v1 , v2 , . . . , vn } is a basis for V then any other basis for V must have exactly n vectors also.
Because of this result, we can make the following definition.
Definition 16.11: Let V be a vector space. The dimension of V, denoted dim V, is the
number of vectors in any basis of V. The dimension of the trivial vector space V = {0} is
defined to be zero.
There is one subtle issue we are sweeping under the rug: Does every vector space have a
basis? The answer is yes but we will not prove this result here.
Moving on, suppose that we have a set B = {v1 , v2 , . . . , vn } in Rn containing exactly n
vectors. For B = {v1 , v2 , . . . , vn } to be a basis of Rn , the set B must be linearly independent
and span B = Rn . In fact, it can be shown that if B is linearly independent then the spanning
condition span B = Rn is automatically satisfied, and vice-versa. For example, say the vec-
tors {v1 , v2 , . . . , vn } in Rn are linearly independent, and put A = [v1 v2 · · · vn ]. Then A−1
exists and therefore Ax = b is always solvable. Hence, Col(A) = span {v1 , v2 , . . . , vn } = Rn .
In summary, we have the following theorem.
Theorem 16.12: Let B = {v1 , . . . , vn } be vectors in Rn . If B is linearly independent

then B is a basis for Rn . Or if span{v1 , v2 , . . . , vn } = Rn then B is a basis for Rn .
129
Example 16.13. Do the columns of the matrix A form a basis for R4 ?

 
2 3 3 −2
 
 4 7 8 −6 
A=  0


 0 1 0 
−4 −6 −6 3
Solution. Let v1 , v2 , v3 , v4 denote the columns of A. Since we have n = 4 vectors in Rn , we
need only check that they are linearly independent. Compute
det A = −2 6= 0
Hence, rank(A) = 4 and thus the columns of A are linearly independent. Therefore, the
vectors v1 , v2 , v3 , v4 form a basis for R4 .
A subspace W of a vector space V is a vector space in its own right, and therefore also
has dimension. By definition, if B = {v1 , . . . , vk } is a linearly independent set in W and
span{v1 , . . . , vk } = W, then B is a basis for W and in this case the dimension of W is k.
Since an n-dimensional vector space V requires exactly n vectors in any basis, then if W is
a strict subspace of V then
dim W < dim V.
As an example, in V = R3 subspaces can be classified by dimension:
1. The zero dimensional subspace in R3 is W = {0}.
2. The one dimensional subspaces in R3 are lines through the origin. These are spanned
by a single non-zero vector.
3. The two dimensional subspaces in R3 are planes through the origin. These are spanned
by two linearly independent vectors.
4. The only three dimensional subspace in R3 is R3 itself. Any set {v1 , v2 , v3 } in R3 that
is linearly independent is a basis for R3 .
Example 16.14. Find a basis for Null(A) and the dim Null(A) if
 
−2 4 −2 −4
 
A=  2 −6 −3 1 .

−3 8 2 −3
Solution. By definition, the Null(A) is the solution set of the homogeneous system Ax = 0.
Row reducing we obtain  
1 0 6 5
 
A∼  0 1 5/2 3/2 

0 0 0 0
130
Lecture 16
The general solution to Ax = 0 in parametric form is

   
−5 −6
−3/2
 + s −5/2 = tv1 + sv2
 
x = t
 0   1 
1 0
By construction, the vectors

   
−5 −6
−3/2 −5/2
 0 ,
v1 =   v2 = 
 1 

1 0
span the null space (A) and they are linearly independent. Therefore, B = {v1 , v2 } is a
basis for Null(A) and therefore dim Null(A) = 2. In general, the dimension of the Null(A)
is the number of free parameters in the solution set of the system Ax = 0, that is,
dim Null(A) = d = n − rank(A)
Example 16.15. Find a basis for Col(A) and the dim Col(A) if
 
1 2 3 −4 8
 
 1 2 0 2 8 
A=
 2 4 −3
.
 10 9 

3 6 0 6 9
Solution. By definition, the column space of A is the span of the columns of A, which we
denote by A = [v1 v2 v3 v4 v5 ]. Thus, to find a basis for Col(A), by trial and error we could
determine the largest subset of the columns of A that are linearly independent. For example,
first we determine if {v1 , v2 } is linearly independent. If yes, then add v3 and determine if
{v1 , v2 , v3 } is linearly independent. If {v1 , v2 } is not linearly independent then discard v2
and determine if {v1 , v3 } is linearly independent. We continue this process until we have
determined the largest subset of the columns of A that is linearly independent, and this will
yield a basis for Col(A). Instead, we can use the fact that matrices that are row equivalent
induce the same solution set for the associated homogeneous system. Hence, let B be the
RREF of A:  
1 2 0 2 0
 
 0 0 1 −2 0 
B = rref(A) =   
 0 0 0 0 1 
0 0 0 0 0
131
By inspection, the columns b1 , b3 , b5 of B are linearly independent. It is easy to see that

b2 = 2b1 and b4 = 2b1 − 2b3 . These same linear relations hold for the columns of A:
 
1 2 3 −4 8
 
 1 2 0 2 8 
A=  

 2 4 −3 10 9 
3 6 0 6 9
By inspection, v2 = 2v1 and v4 = 2v1 − 2v3 . Thus, because b1 , b3 , b5 are linearly inde-
pendent columns of B =rref(A), then v1 , v3 , v5 are linearly independent columns of A.
Therefore, we have
     
 1 3 8 

 

 1   0   8 
      

Col(A) = span{v1 , v3 , v5 } = span   , 
  
 ,  9 
  


  2   −3    

 

3 0 9
and consequently dim Col(A) = 3. This procedure works in general: To find a basis
for the Col(A), row reduce A ∼ B until you can determine which columns of B are linearly
independent. The columns of A in the same position as the linearly independent columns
of B form a basis for the Col(A).
WARNING: Do not take the linearly independent columns of B as a basis for Col(A).
Always go back to the original matrix A to select the columns.
• what it means for a set to be linearly independent/dependents
• what a basis is (a spanning set that is linearly independent)
• what is the meaning of the dimension of a vector space
• how to determine if a given set in Rn is linearly independent
• how to find a basis for the null space and column space of a matrix A
132
Lecture 17
The Rank Theorem
17.1 The Rank of a Matrix

We now give the definition to the rank of a matrix.
Definition 17.1: The rank of a matrix A is the dimension of its column space. We will
use rank(A) to denote the rank of A.
Recall that Col(A) = Range(TA ), and thus the rank of A is the dimension of the range of
the linear mapping TA . The range of a mapping is sometimes called the image.
We now define the nullity of a matrix.
Definition 17.2: The nullity of a matrix A is the dimension of its nullspace Null(A).
We will use nullity(A) to denote the nullity of A.
Recall that (A) = ker(TA ), and thus the nullity of A is the dimension of the kernel of the
linear mapping TA .
The rank and nullity of a matrix are connected via the following fundamental theorem
known as the Rank Theorem.
Theorem 17.3: (Rank Theorem) Let A be a m × n matrix. The rank of A is the

number of leading 1’s in its RREF. Moreover, the following equation holds:
n = rank(A) + nullity(A).
Proof. A basis for the column space is obtained by computing rref(A) and identifying the
columns that contain a leading 1. Each column of A corresponding to a column of rref(A)
with a leading 1 is a basis vector for the column space of A. Therefore, if r is the number
of leading 1’s then r = rank(A). Now let d = n − r. The number of free parameters in the
The Rank Theorem
solution set of Ax = 0 is d and therefore a basis for Null(A) will contain d vectors, that is,
nullity(A) = d. Therefore,
nullity(A) = n − rank(A).
Example 17.4. Find the rank and nullity of the matrix

 
1 −2 2 3 −6
A =  0 −1 −3 1 1 .
−2 4 −3 −6 11
Solution. Row reduce far enough to identify where the leading entries are:
 
1 −2 2 3 −6
2R +R2
A −−1−−→  0 −1 −3 1 1 
0 0 1 0 −1
There are r = 3 leading entries and therefore rank(A) = 3. The nullity is therefore
nullity(A) = 5 − rank(A) = 2.
Example 17.5. Find the rank and nullity of the matrix

 
1 −3 −1
A =  −1 4 2 .
−1 3 0
Solution. Row reduce far enough to identify where the leading entries are:
 
1 −3 −1
R1 +R2 ,R1 +R3
A −−−−−−−−→  0 1 1 
0 0 −1
There are r = 3 leading entries and therefore rank(A) = 3. The nullity is therefore
nullity(A) = 3 − rank(A) = 0. Another way to see that nullity(A) = 0 is as follows. From
the above computation, A is invertible. Therefore, there is only one vector in Null(A) = {0}.
The subspace {0} has dimension zero.
Using the rank and nullity of a matrix, we now provide further characterizations of
invertible matrices.
Theorem 17.6: Let A be a n × n matrix. The following statements are equivalent:
(i) The columns of A form a basis for Rn .
(ii) Col(A) = Rn
(iii) rank(A) = n
(iv) Null(A) = {0}
134
Lecture 17
(v) nullity(A) = 0
(vi) A is an invertible matrix.
• what the rank of a matrix is and how to compute it
• what the nullity of a matrix is and how to compute it
• the Rank Theorem
135
The Rank Theorem
136
Lecture 18
Coordinate Systems
18.1 Coordinates
Recall that a basis of a vector space V is a set of vectors B = {v1 , v2 , . . . , vn } in V such that
1. the set B spans all of V, that is, V = span(B), and
2. the set B is linearly independent.

Hence, if B is a basis for V, each vector x∗ ∈ V can be written as a linear combination of B:
x∗ = c1 v1 + c2 v2 + · · · + cn vn .
Moreover, from the definition of linear independence given in Definition 6.1, any vector
x ∈ span(B) can be written in only one way as a linear combination of v1 , . . . , vn . In other
words, for the x∗ above, there does not exist other scalars t1 , . . . , tn such that also
x∗ = t1 v1 + t2 v2 + · · · + tn vn .
To see this, suppose that we can write x∗ in two different ways using B:
x∗ = c1 v1 + c2 v2 + · · · + cn vn
x∗ = t1 v1 + t2 v2 + · · · + tn vn .
Then
0 = x∗ − x∗ = (c1 − t1 )v1 + (c2 − t2 )v2 + · · · + (cn − tn )vn .
Since B = {v1 , . . . , vn } is linearly independent, the only linear combination of v1 , . . . , vn
that gives the zero vector 0 is the trivial linear combination. Therefore, it must be the case
that ci − ti = 0, or equivalently that ci = ti for all i = 1, 2 . . . , n. Thus, there is only one way
to write x∗ in terms of B = {v1 , . . . , vn }. Hence, relative to the basis B = {v1 , v2 , . . . , vn },
the scalars c1 , c2 , . . . , cn uniquely determine the vector x, and vice-versa.
Our preceding discussion on the unique representation property of vectors in a given basis
leads to the following definition.
Coordinate Systems
Definition 18.1: Let B = {v1 , . . . , vn } be a basis for V and let x ∈ V. The coordinates
of x relative to the basis B are the unique scalars c1 , c2 , . . . , cn such that
x = c1 v1 + c2 v2 + · · · + cn vn .
In vector notation, the B-coordinates of x will be denoted by

 
c1
 c2 
 
[x]B =  .. 
.
cn
and we will call [x]B the coordinate vector of x relative to B.
The notation [x]B indicates that these are coordinates of x with respect to the basis B.
If it is clear what basis we are working with, we will omit the subscript B and simply write
[x] for the coordinates of x relative to B.
Example 18.2. One can verify that

1 −1
B= ,
1 1

2 3
is a basis for R . Find the coordinates of v = relative to B.
1
Solution. Let v1 = (1, 1) and let v2 = (−1, 1). By definition, the coordinates of v with
respect to B are the scalars c1 , c2 such that

1 −1 c1
v = c1 v1 + c2 v2 =
1 1 c2
If we put P = [v1 v2 ], and let [v]B = (c1 , c2 ), then we need to solve the linear system
v = P[v]B
Solving the linear system, one finds that the solution is [v]B = (2, −1), and therefore this is
the B-coordinate vector of v, or the coordinates of v, relative to B.
It is clear how the procedure of the previous example can be generalized. Let B =
n n
{v1 , v2 , . . . , vn } be a basis for R and let v be any vector in R . Put P = v1 v2 · · · vn .
Then the B-coordinates of v is the unique column vector [v]B solving the linear system
Px = v
138
Lecture 18
that is, x = [v]B is the unique solution to Px = v. Because v1 , v2 , . . . , vn are linearly

independent, the solution to Px = v is
[v]B = P−1 v.
We
remark
that if an inconsistent row arises when you row reduce the augmented matrix
P v then you have made an error in your row reduction algorithm. In summary, to find
coordinates with respect to a basis B in Rn , we need to solve a square linear system.
Example 18.3. Let      

3 −1 3
v1 = 6 , v2 = 0 , x = 12
    
2 1 7
and let B = {v1 , v2 }. One can show that B is linearly independent and therefore a basis for
W = span{v1 , v2 }. Determine if x is in W, and if so, find the coordinate vector of x relative
to B.
Solution. By definition, x is in W = span{v1 , v2 } if we can write x as a linear combination

of v1 , v2 :
x = c1 v1 + c2 v2
Form the associated augmented matrix and row reduce:
   
3 −1 3 1 0 2
6 0 12 ∼ 0 1 3
2 1 7 0 0 0
The system is consistent with solution c1 = 2 and c2 = 3. Therefore, x is in W, and the

B-coordinates of x are
2
[x]B =
3
Example 18.4. What are the coordinates of

 
3
v =  11 
−7
in the standard basis E = {e1 , e2 , e3 }?
Solution. Clearly,        
3 1 0 0
v =  11  = 3 0 + 11 1 − 7 0
−7 0 0 1
139
Coordinate Systems
Therefore, the coordinate vector of v relative to {e1 , e2 , e3 } is

 
3
[v]E = 11 

−7
Example 18.5. Let P3 [t] be the vector space of polynomials of degree at most 3.
(i) Show that B = {1, t, t2 , t3 } is a basis for P3 [t].
(ii) Find the coordinates of v(t) = 3 − t2 − 7t3 relative to B.

Solution. The set B = {1, t, t2, t3 } is a spanning set for P3 [t]. Indeed, any polynomial
u(t) = c0 + c1 t + c2 t2 + c3 t3 is clearly a linear combination of 1, t, t2 , t3 . Is B linearly
independent? Suppose that there exists scalars c0 , c1 , c2 , c3 such that
c0 + c1 t + c2 t2 + c3 t3 = 0.
Since the above equality must hold for all values of t, we conclude that c0 = c1 = c2 = c3 = 0.
Therefore, B is linearly independent, and consequently a basis for P3 [t]. In the basis B, the
coordinates of v(t) = 3 − t2 − 7t3 are
 
3
0
[v(t)]B = 
−1

−7
The basis B = {1, t, t2 , t3 } is called the standard basis in P3 [t].
Example 18.6. Show that

1 0 0 1 0 0 0 0
B= , , ,
0 0 0 0 1 0 0 1

3 0
is a basis for M2×2 . Find the coordinates of A = relative to B.
−4 −1

m11 m12
Solution. Any matrix M = can be written as a linear combination of the ma-
m21 m22
trices in B:

m11 m12 1 0 0 1 0 0 0 0
= m11 + m12 + m21 + m22
m21 m22 0 0 0 0 1 0 0 1
If
1 0 0 1 0 0 0 0 c1 c2 0 0
c1 + c2 + c3 + c4 = =
0 0 0 0 1 0 0 1 c3 c4 0 0
140
Lecture 18
then clearly c1 = c2 = c3 = c4 = 0. Therefore,

B is linearly independent, and consequently
3 0
a basis for M2×2 . The coordinates of A = in the basis
−4 −1

1 0 0 1 0 0 0 0
B= , , ,
0 0 0 0 1 0 0 1
are  
3
0
[A]B = 
−4

−1
The basis B above is the standard basis of M2×2 .
18.2 Coordinate Mappings

Let B = {v1 , v2 , . . . , vn } be a basis of Rn and let P = [v1 v2 · · · vn ] ∈ Mn×n . If x ∈ Rn and
[x]B are the B-coordinates of x relative to B then
x = P[x]B . (⋆)
Hence, thinking of P : Rn → Rn as a linear mapping, P maps B-coordinate vectors to

coordinate vectors relative to the standard basis of Rn . For this reason, we call P the
change-of-coordinates matrix from the basis B to the standard basis in Rn . If we need
to emphasize that P is constructed from the basis B we will write PB instead of just P.
Multiplying equation (⋆) by P−1 we obtain
P−1 x = [x]B .
Therefore, P−1 maps coordinate vectors in the standard basis to coordinates relative to B.
Example 18.7. The columns of the matrix P form a basis B for R3 :

 
1 3 3
P =  −1 −4 −2  .
0 0 −1
(a) What vector x ∈ R3 has B-coordinates [x]B = (1, 0, −1).
(b) Find the B-coordinates of v = (2, −1, 0).
Solution. The matrix P maps B-coordinates to standard coordinates in R3 . Therefore,

 
−2
x = P[x]B =  1 
1
141
Coordinate Systems
On the other hand, the inverse matrix P−1 maps standard coordinates in R3 to B-coordinates.
One can verify that  
4 3 6
P−1 =  −1 −1 −1 
0 0 −1
Therefore, the B coordinates of v are
    
4 3 6 2 5
−1
[v]B = P v = −1 −1 −1
   −1 = −1
 
0 0 −1 0 0
When V is an abstract vector space, e.g. Pn [t] or Mn×n , the notion of a coordinate
mapping is similar as the case when V = Rn . If V is an n-dimensional vector space and
B = {v1 , v2 , . . . , vn } is a basis for V, we define the coordinate mapping P : V → Rn relative
to B as the mapping
P(v) = [v]B .
Example 18.8. Let V = M2×2 and let B = {A1 , A2 , A3 , A4} be the standard basis for
M2×2 . What is P : M2×2 → R4 ?
Solution. Recall,

1 0 0 1 0 0 0 0
B = {A1 , A2 , A3 , A4} = , , ,
0 0 0 0 1 0 0 1

a11 a12
Then for any A = we have
a21 a22
 
a11
a11 a12 a12 
P a21  .
= 
a21 a22
a22
18.3 Matrix Representation of a Linear Map

Let V and W be vector spaces and let T : V → W be a linear mapping. Then by definition
of a linear mapping, T(v + u) = T(v) + T(u) and T(αv) = αT(v) for every v, u ∈ V and
α ∈ R. Let B = {v1 , v2 , . . . , vn } be a basis of V and let γ = {w1 , w2 , . . . , wm } be a basis of
W. Then for any v ∈ V there exists scalars c1 , c2 , . . . , cn such that
v = c1 v1 + c2 v2 + · · · + cn vn
142
Lecture 18
and thus [v]B = (c1 , c2 , . . . , cn ) are the coordinates of v in the basis B By linearity of the
mapping T we have
T(v) = T(c1 v1 + c2 v2 + · · · + cn vn )
= c1 T(v1 ) + c2 T(v2 ) + · · · + cn T(vn )
Now each vector T(vj ) is in W and therefore because γ is a basis of W there are scalars
a1,j , a2,j , . . . , am,j such that
T(vj ) = a1,j w1 + a2,j w2 + · · · + am,j wm
In other words,
[T(vj )]γ = (a1,j , a2,j , . . . , am,j )
Substituting T(vj ) = a1,j w1 + a2,j w2 + · · · + am,j wm for each j = 1, 2, . . . , n into
T(v) = c1 T(v1 ) + c2 T(v2 ) + · · · + cn T(vn )
and then simplifying we get !

m
X n
X
T(v) = cj ai,j wi
i=1 j=1
Therefore,
[T(v)]γ = A[v]B
where A is the m × n matrix given by

A = [T(v1 )]γ [T(v2 )]γ · · · [T(vn )]γ
The matrix A is the matrix representation of the linear mapping T in the bases B and γ.
Example 18.9. Consider the vector space V = P2 [t] of polynomial of degree no more than
two and let T : V → V be defined by
T(v(t)) = 4v′(t) − 2v(t)
It is straightforward to verify that T is a linear mapping. Let
B = {v1 , v2 , v3 } = {t − 1, 3 + 2t, t2 + 1}.
(a) Verify that B is a basis of V.
(b) Find the coordinates of v(t) = −t2 + 3t + 1 in the basis B.
(c) Find the matrix representation of T in the basis B.
143
Coordinate Systems
Solution. (a) Suppose that there are scalars c1 , c2 , c3 such that
c1 v1 + c2 v2 + c3 v3 = 0
Then expanding and then collecting like terms we obtain
c3 t2 + (c1 + 2c2 )t + (−c1 + 3c2 + c3 ) = 0
Since the above holds for all t ∈ R we must have
c3 = 0, c1 + 2c2 = 0, −c1 + 3c2 + c3 = 0
Solving for c1 , c2 , c3 we obtain c1 = 0, c2 = 0, c3 = 0. Hence, the only linear combination of

the vectors in B that produces the zero vector is the trivial linear combination. This proves
by definition that B is linearly independent. Since we already know that dim(P2 ) = 3 and
B contains 3 vectors, then B is a basis for P2
(b) The coordinates of v(t) = −t2 + 3t + 1 are the unique scalars (c1 , c2 , c3 ) such that
c1 v1 + c2 v2 + c3 v3 = v
In this case the linear system is
c3 = −1, c1 + 2c2 = 3, −c1 + 3c2 + c3 = 1
and solving yields c1 = 1, c2 = 1, and c3 = −1. Hence,
[v]B = (1, 1, −1)
(c) The matrix representation A of T is

A = [T(v1 )]B [T(v2 )]B [T(v3 )]B
Now we compute directly that
T(v1 ) = −2t + 6, T(v2 ) = −4t + 2, T(v3 ) = −2t2 + 8t − 2
And then one computes that

     
−18/5 −6/5 24/5
[T(v1 )]B =  4/5  , [T(v2 )]B = −2/5 , [T(v3 )]B =  8/5 
0 0 −2
And therefore  
−18/5 −6/5 24/5
 
 
A =  4/5 −2/5 8/5 
 
0 0 −2
144
Lecture 18
• what coordinates are (you need a basis)
• how to find coordinates relative to a basis
• the interpretation of the change-of-coordinates matrix as a mapping that transforms

one set of coordinates to another
145
Coordinate Systems
146
Lecture 19
Change of Basis
19.1 Review of Coordinate Mappings on Rn

Let B = {v1 , . . . , vn } be a basis for Rn and let
PB = [v1 v2 · · · vn ].
If x ∈ Rn and [x]B is the coordinate vector of x in the basis B then
x = PB [x]B .
The components of the vector x are the coordinates of x in the standard basis E = {e1 , . . . , en }.
In other words,
[x]E = x.
Therefore,
[x]E = PB [x]B .
We can therefore interpret PB as the matrix mapping that maps the B-coordinates of x to
the E-coordinates of x. To make this more explicit, we sometimes use the notation
E PB
to indicate that E PB maps B-coordinates to E-coordinates:
[x]E = (E PB )[x]B .
If we multiply the equation

[x]E = (E PB )[x]B
on the left by the inverse of E PB we obtain
(E PB )−1 [x]E = [x]B
Hence, the matrix (E PB )−1 maps standard coordinates to B-coordinates, see Figure 19.1. It
is natural then to introduce the notation
B PE = (E PB )−1
Change of Basis
V = Rn
b
x
B PE = (E PB )−1
[x]B
Figure 19.1: The matrix B PE maps E coordinates to B coordinates.
Example 19.1. Let
       
1 −3 3 −8
v1 = 0 , v2 = 4 , v2 = −6 , x = 2  .
      
0 0 3 3
(a) Show that the set of vectors B = {v1 , v2 , v3 } forms a basis for Rn .
(b) Find the change-of-coordinates matrix from B to standard coordinates.
(c) Find the coordinate vector [x]B for the given x.
Solution. Let
 
1 −3 3
PB = 0 4 −6
0 0 3
It is clear that det(PB ) = 12, and therefore v1 , v2 , v3 are linearly independent. Therefore,
B is a basis for Rn . The matrix PB takes B-coordinates to standard coordinates. The
B-coordinate vector [x]B = (c1 , c2 , c3 ) is the unique solution to the linear system
x = PB [x]B
Solving the linear system with augmented matrix [PB x] we obtain
[x]B = (−5, 2, 1)
We verify that [x]B = (−5, 2, 1) are indeed the coordinates of x = (−8, 2, 3) in the basis
148
Lecture 19
B = {v1 , v2 , v3 }:
     
1 −3 3
(−5)v1 + (2)v2 + (1)v3 = −5 0 + 2 4 + −6
    
0 0 3
     
−5 −6 3
=  0  +  8  + −6
0 0 3
 
−8
= 2 
3
| {z }
x
19.2 Change of Basis

We saw in the previous section that the matrix
E PB
takes as input the B-coordinates [x]B of a vector x and returns the coordinates of x in the
standard basis. We now consider the situation of dealing with two basis B and C where
neither is assumed to be the standard basis E. Hence let B = {v1 , v2 , . . . , vn } and let
C = {w1 , . . . , wn } be two basis of Rn and let
E PB = [v1 v2 · · · vn ]
E PC = [w1 w2 · · · wn ].
Then if [x]C is the coordinate vector of x in the basis C then
x = (E PC )[x]C .
How do we transform B-coordinates of x to C-coordinates of x, and vice-versa? To answer

this question, start from the relations
x = (E PB )[x]B
x = (E PC )[x]C .
Then
(E PC )[x]C = (E PB )[x]B
and because E PC is invertible we have that
[x]C = (E PC )−1 (E PB )[x]B .
149
Change of Basis
Hence, the matrix (E PC )−1 (E PB ) maps the B-coordinates of x to the C-coordinates of x. For
this reason, it is natural to use the notation (see Figure 19.2)
C PB = (E PC )−1 (E PB ).
V = Rn
b
x
E PC E PB
b b
[x]C C PB [x]B
Figure 19.2: The matrix C PB maps B-coordinates to C-coordinates.
If we expand (E PC )−1 (E PB ) we obtain that

(E PC )−1 (E PB ) = (E PC )−1 v1 (E PC )−1 v2 · · · (E PC )−1 vn .
Therefore, the ith column of (E PC )−1 (E PB ), namely
(E PC )−1 vi ,
is the coordinate vector of vi in the basis C = {w1 , w2 , . . . , wn }. To compute C PB we

augment E PC and E PB and row reduce fully:

E P C E P B ∼ In C P B .
Example 19.2. Let

1 −2 −7 −5
B= , , C= ,
−3 4 9 7
It can be verified that B = {v1 , v2 } and C = {w1 , w2 } are bases for R2 .

(a) Find the matrix the takes B-coordinates to C-coordinates.
(b) Find the matrix that takes C-coordinates to B-coordinates.
(c) Let x = (0, −2). Find [x]B and [x]C .
Solution. The matrix E PB = [v1 v2 ] maps B-coordinates to standard E-coordinates. The

matrix E PC = [w1 w2 ] maps C-coordinates to standard E-coordinates. As we just showed,
the matrix that maps B-coordinates to C-coordinates is
C PB = (E PC )−1 (E PB )
150
Lecture 19
It is straightforward to compute that

" #
−7/4 −5/4
(E PC )−1 =
9/4 7/4
Therefore,
" #" # " #
−7/4 −5/4 1 −2 2 −3/2
C PB = (E PC )−1 (E PB ) = =
9/4 7/4 −3 4 −3 5/2
To compute B PC , we can simply invert C PB . One finds that

−1 5 3
(C P B ) =
6 4
and therefore
5 3
B PC =
6 4
Given that x = (0, −2), to find [x]B we must solve the linear system
E PB [x]B =x
Row reducing the augmented matrix [E PB x] we obtain

2
[x]B =
1
Next, to find [x]C we can solve the linear system
E PC [x]C =x
Alternatively, since we now know [x]B and C PB has been computed, to find [x]C we simply
multiply C PB by [x]B :
" #" # " #
2 −3/2 2 5/2
[x]C = C PB [x]B = =
−3 5/2 1 −7/2

5/2 0
Let’s verify that [x]C = are indeed the C-coordinates of x = :
−7/2 −2
" #" # " #
−7 −5 5/2 0
P
E C [x] C = = .
9 7 −7/2 −2

• how to compute a change of basis matrix
• and how to use the change of basis matrix to map one set of coordinates into another
151
Change of Basis
152
Lecture 20
Lecture 20
Inner Products and Orthogonality
20.1 Inner Product on Rn

The inner product on Rn generalizes the notion of the dot product of vectors in R2 and R3
that you may are already familiar with.
Definition 20.1: Let u = (u1 , u2, . . . , un ) and let v = (v1 , v2 , . . . , vn ) be vectors in Rn .

The inner product of u and v is
u • v = u1 v1 + u2 v2 + · · · + un vn .
Notice that the inner product u • v can be computed as a matrix multiplication as follows:
 
v1
 v2 
u • v = uT v = u1 u2 · · · un  ..  .
 
.
vn
The following theorem summarizes the basic algebraic properties of the inner product.
Theorem 20.2: Let u, v, w be vectors in Rn and let α be a scalar. Then
(a) u • v = v • u
(b) (u + v) • w = u • w + v • w
(c) (αu) • v = α(u • v) = u • (αv)
(d) u • u ≥ 0, and u • u = 0 if and only if u = 0
153
Example 20.3. Let u = (2, −5, −1) and let v = (3, 2, −3). Compute u • v, v • u, u • u, and
v • v.
Solution. By definition:
u • v = (2)(3) + (−5)(2) + (1)(−3) = −1

v • u = (3)(2) + (2)(−5) + (−3)(1) = −1
u • u = (2)(2) + (−5)(−5) + (−1)(−1) = 30
v • v = (3)(3) + (2)(2) + (−3)(−3) = 22.
We now define the length or norm of a vector in Rn .
Definition 20.4: The length or norm of a vector u ∈ Rn is defined as

√ q
kuk = u • u = u21 + u22 + · · · + u2n .
A vector u ∈ Rn with norm 1 will be called a unit vector:
kuk = 1.
Below is an important property of the inner product.
Theorem 20.5: Let u ∈ Rn and let α be a scalar. Then
kαuk = |α|kuk.
Proof. We have
p
kαuk = (αu) • (αu)
p
= α2 (u • u)
√
= |α| u • u
= |α|kuk.
By Theorem 20.5, any non-zero vector u ∈ Rn can be scaled to obtain a new unit vector
in the same direction as u. Indeed, suppose that u is non-zero so that kuk =
6 0. Define the
new vector
1
v= u
kuk
154
Lecture 20
1
Notice that α = kuk
is just a scalar and thus v is a scalar multiple of u. Then by Theorem 20.5
we have that
1
kvk = kαuk = |α| · kuk = · kuk = 1
kuk
and therefore v is a unit vector, see Figure 20.1. The process of taking a non-zero vector u
1
and creating the new vector v = kuk u is sometimes called normalization of u.
u
1
v= kuk
u
Figure 20.1: Normalizing a non-zero vector.
Example 20.6. Let u = (2, 3, 6). Compute kuk and find the unit vector v in the same
direction as u.
Solution. By definition,
√ √ √
kuk = u•u= 22 + 32 + 62 = 49 = 7.
Then the unit vector that is in the same direction as u is
   
2 2/7
1 1
v= u = 3 = 3/7
kuk 7
6 6/7
Verify that kvk = 1:
p p p √
kvk = (2/7)2 + (3/7)2 + (6/7)2 = 4/49 + 9/49 + 36/49 = 49/49 = 1 = 1.
Now that we have the definition of the length of a vector, we can define the notion of
distance between two vectors.
Definition 20.7: Let u and v be vectors in Rn . The distance between u and v is the
length of the vector u − v. We will denote the distance between u and v by d(u, v). In
other words,
d(u, v) = ku − vk.

3 7
Example 20.8. Find the distance between u = and v = .
−2 −9
p √
d(u, v) = ku − vk = (3 − 7)2 + (−2 + 9)2 = 65.
155
20.2 Orthogonality
In the context of vectors in R2 and R3 , orthogonality is synonymous with perpendicularity.
Below is the general definition.
Definition 20.9: Two vectors u and v in Rn are said to be orthogonal if u • v = 0.
In R2 and R3 , the notion of orthogonality should be familiar to you. In fact, using the
Law of Cosines in R2 or R3 , one can prove that
u • v = kuk · kvk cos(θ) (20.1)
where θ is the angle between u and v. If θ = π2 then clearly u • v = 0. In higher dimensions,

i.e., n ≥ 4, we can use equation (20.1) to define the angle between vectors u and v. In other
words, the angle between any two vectors u and v in Rn is define to be

u•v
θ = arccos .
kuk · kvk
The general notion of orthogonality in Rn leads to the following theorem from grade
school.
Theorem 20.10: (Pythagorean Theorem) Two vectors u and v are orthogonal if and
only if ku + vk2 = kuk2 + kvk2 .
p
Solution. First recall that ku + vk = (u + v) • (u + v) and therefore
ku + vk2 = (u + v) • (u + v)
=u•u+u•v+v•u+v•v
= kuk2 + 2(u • v) + kvk2.
Therefore, ku + vk2 = kuk2 + kvk2 if and only if u • v = 0.
We now introduce orthogonal sets.
Definition 20.11: A set of vectors {u1 , u2 , . . . , up } is said to be an orthogonal set if

any pair of distinct vectors ui , uj are orthogonal, that is, ui • uj = 0 whenever i 6= j.
In the following theorem we prove that orthogonal sets are linearly independent.
156
Lecture 20
Theorem 20.12: Let {u1 , u2 , . . . , up } be an orthogonal set of non-zero vectors in Rn .

Then the set {u1 , u2 , . . . , up } is linearly independent. In particular, if p = n then the set
{u1 , u2 , . . . , un } is basis for Rn .
Solution. Suppose that there are scalars c1 , c2 , . . . , cp such that
c1 u1 + c2 u2 + · · · + cp up = 0.
Take the inner product of u1 with both sides of the above equation:
c1 (u1 • u1 ) + c2 (u2 • u1 ) + · · · + cp (up • u1 ) = 0 • u1 .
Since the set is orthogonal, the left-hand side of the last equation simplifies to c1 (u1 • u1 ).
The right-hand side simplifies to 0. Hence,
c1 (u1 • u1 ) = 0.
But u1 • u1 = ku1 k2 is not zero and therefore the only way that c1 (u1 • u2 ) = 0 is if c1 = 0.
Repeat the above steps using u2 , u3 , . . . , up and conclude that c2 = 0, c3 = 0, . . . , cp =
0. Therefore, {u1 , . . . , up } is linearly independent. If p = n, then the set {u1 , . . . , up } is
automatically a basis for Rn .
Example 20.13. Is the set {u1 , u2 , u3 } an orthogonal set?
     
1 0 −5
u1 = −2 , u2 = 1 , u3 = −2
    
1 2 1
Solution. Compute
u1 • u2 = (1)(0) + (−2)(1) + (1)(2) = 0

u1 • u3 = (1)(−5) + (−2)(−2) + (1)(1) = 0
u2 • u3 = (0)(−5) + (1)(−2) + (2)(1) = 0
Therefore, {u1 , u2 , u3 } is an orthogonal set. By Theorem 20.12, the set {u

1 , u2 , u3 } is linearly
independent. To verify linear independence, we computed that det( u1 u2 u3 ) = 30,
which is non-zero.
157
We now introduce orthonormal sets.
Definition 20.14: A set of vectors {u1 , u2 , . . . , up } is said to be an orthonormal set if

it is an orthogonal set and if each vector ui in the set is a unit vector.
Consider the previous orthogonal set in R3 :

     
 1 0 −5 
{u1 , u2 , u3 } = −2  , 1  , −2 .

 
1 2 1
It
√ is not an orthonormal
√ set because
√ none of u1 , u2 , u3 are unit vectors. Explicitly, ku1 k =
6, ku2 k = 5, and ku3 k = 30. However, from an orthogonal set we can create an
orthonormal set by normalizing each vector. Hence, the set
 √     √ 
 1/ √6 0√ −5/√30 
{v1 , v2 , v3 } = −2/√ 6 , 1/√5 , −2/√ 30
 
1/ 6 2/ 5 1/ 30
is an orthonormal set.
20.3 Coordinates in an Orthonormal Basis

As we will see in this section, a basis B = {u1 , u2 , . . . , un } of Rn that is also an orthonormal
set is highly desirable when performing computations with coordinates. To see why, let x
be any vector in Rn and suppose we want to find the coordinates of x in the basis B, that is
we seek to find [x]B = (c1 , c2 , . . . , cn ). By definition, the coordinates c1 , c2 , . . . , cn satisfy the
equation
x = c1 u1 + c2 u2 + · · · + cn un .
Taking the inner product of u1 with both sides of the above equation and using the fact that
u1 • u2 = 0, u1 • u3 = 0, and u1 • un = 0, we obtain
u1 • x = c1 (u1 • u1 ) = c1 (1) = c1
where we also used the fact that ui is a unit vector. Thus, c1 = u1 • x! Repeating this
procedure with u2 , u3 , . . . , un we obtain the remaining coefficients c2 , . . . , cn :
c2 = u2 • x
c3 = u3 • x
.. .
. = ..
cn = un • x.
Our previous computation proves the following theorem.
158
Lecture 20
Theorem 20.15: Let B = {u1 , u2 , . . . , un } be an orthonormal basis for Rn . The coordi-

nate vector of x in the basis B is
 
u1 • x
 u2 • x 
 
[x]B =  ..  .
 . 
un • x
Hence, computing coordinates with respect to an orthonormal basis can be done without
performing any row operations and all we need to do is compute inner products! We make
the important observation that an alternate expression for [x]B is
   
u1 • x uT1
 u2 • x  uT 
  2
[x]B =  ..  =  ..  x = UT x

 .   . 
un • x uTn
where U = [u1 u2 · · · un ]. On the other hand, recall that by definition [x]B satisfies
U[x]B = x, and therefore [x]B = U−1 x. If we compare the two identities
[x]B = U−1 x and [x]B = UT x
we suspect then that U−1 = UT . This is indeed the case. To see this, let B = {u1 , u2 , . . . , un }
be an orthonormal basis for Rn and put
U = [u1 u2 · · · un ].
Consider the matrix product UT U, and recalling that ui • uj = uTi uj , we obtain

 
uT1
uT 
 2
UT U =  ..  u1 u2 · · · un
 . 
uTn
 T 
u1 u1 uT1 u2 · · · uT1 un
 
 T
u2 u1 uT2 u2 · · · uT2 un 

 
= 
 .. .. .. .. 
 . . . . 
 
uTn u1 uTn u2 T
· · · un un
= In .
159
Therefore,
U−1 = UT .
A matrix U ∈ Rn×n such that
UT U = UUT = In
is called a orthogonal matrix. Hence, if B = {u1 , u2 , . . . , un } is an orthonormal set then
the matrix
U = u1 u2 · · · un
is an orthogonal matrix.
Example 20.16. Consider the vectors

       
1 −1 2 1
v1 = 0 , v2 = 4 , v3 = 1 , x = 2  .
      
1 1 −2 −1
(a) Show that {v1 , v2 , v3 } is an orthogonal basis for R3 .

(b) Then, if necessary, normalize the basis vectors vi to obtain an orthonormal basis B =
{u1 , u2 , u3 } for R3 .
(c) For the given x find [x]B .
Solution. (a) We compute that v1 • v2 = 0, v1 • v3 = 0, and v2 • v3 = 0, and thus {v1 , v2 , v3 }

is an orthogonal set. Since orthogonal sets are linearly independent and {v1 , v2 , v3 }
3
consists of three vectors then
√ {v1 , v2 , v√3 } is basis for R .
(b) We compute that kv1 k = 2, kv2 k = 18, and kv3 k = 3. Then let
 √   √   
1/ 2 −1/√ 18 2/3
u1 =  0√  , u2 =  4/√18  , u3 =  1/3 
1/ 2 1/ 18 −2/3
Then B = {u1 , u2 , u3 } is now an orthonormal set and thus since B consists of three
vectors then B is an orthonormal basis of R3 .
(c) Finally, computing coordinates in an orthonormal basis is easy:
   
u1 • x 0
√
[x]B = u2 • x = 2/ 18
u3 • x 5/3
Example 20.17. The standard unit basis

     
 1 0 0 
E = {e1 , e2 , e3 } = 0 , 1 , 0
   
 
0 0 1
160
Lecture 20
in R3 is an orthonormal basis. Given any x = (x1 , x2 , x3 ), we have [x]E = x. On the other

hand, clearly
x1 = x • e1
x2 = x • e2
x3 = x • e3
Example 20.18. (Orthogonal Complements) Let W be a subspace of Rn . The orthogonal

complement of W, which we denote by W⊥ , consists of the vectors in Rn that are orthogonal
to every vector in W. Using set notation:
W⊥ = {u ∈ Rn : u • w = 0 for every w ∈ W}.
(a) Show that W⊥ is a subspace.

(b) Let w1 = (0, 1, 1, 0), let w2 = (1, 0, −1, 0), and let W = span{w1 , w2 }. Find a basis for
W⊥ .
Solution. (a) The vector 0 is orthogonal to every vector in Rn and therefore it is certainly
orthogonal to every vector in W. Thus, 0 ∈ W⊥ . Now suppose that u1 , u2 are two
vectors in W⊥ . Then for any vector w ∈ W it holds that
(u1 + u2 ) • w = u1 • w + u2 • w = 0 + 0 = 0.
Therefore, u1 + u2 is also orthogonal to w and since w is an arbitrary vector in W then

(u1 + u2 ) ∈ W⊥ . Lastly, let α be any scalar and let u ∈ W⊥ . Then for any vector w in
W we have that
(αu) • w = α(u • w) = α · 0 = 0.
Therefore, αu is orthogonal to w and since w is an arbitrary vector in W then (αu) ∈ W⊥ .
This proves that W⊥ is a subspace of Rn .
(b) A vector u = (u1 , u2, u3 , u3) is in W⊥ if u • w1 = 0 and u • w2 = 0. In other words, if
u2 + u3 = 0
u1 − u3 = 0
This is a linear system for the unknowns u1 , u2, u3 , u4. The general solution to the linear
system is    
1 0
0 1
1 + s −1 .
u = t   
0 0
Therefore, a basis for W⊥ is {(1, 0, 1, 0), (0, 1, −1, 0)}.
161
• how to compute inner products, norms, and distances

• how to normalize vectors to unit length
• what orthogonality is and how to check for it
• what an orthogonal and orthonormal basis is
• the advantages of working with orthonormal basis when computing coordinate vectors
162
Lecture 21
Lecture 21
Eigenvalues and Eigenvectors
21.1 Eigenvectors and Eigenvalues

An n × n matrix A can be thought of as the linear mapping that takes any arbitrary vector
x ∈ Rn and outputs a new vector Ax. In some cases, the new output vector Ax is simply
a scalar multiple of the input vector x, that is, there exists a scalar λ such that Ax = λx.
This case is so important that we make the following definition.
Definition 21.1: Let A be a n × n matrix and let v be a non-zero vector. If Av = λv

for some scalar λ then we call the vector v an eigenvector of A and we call the scalar λ
an eigenvalue of A corresponding to v.
Hence, an eigenvector v of A is simply scaled by a scalar λ under multiplication by A.

Eigenvectors are by definition nonzero vectors because A0 is clearly a scalar multiple of 0
and then it is not clear what that the corresponding eigenvalue should be.
Example 21.2. Determine if the given vectors v and u are eigenvectors of A? If yes, find
the eigenvalue of A associated to the eigenvector.
     
4 −1 6 −3 −1
A = 2 1 6 , v =  0  , u =  2  .
2 −1 8 1 1
Solution. Compute
    
4 −1 6 −3 −6
Av = 2 1 6  0  =  0 
2 −1 8 1 2
 
−3
=2 0 

1
= 2v
163
Hence, Av = 2v and thus v is an eigenvector of A with corresponding eigenvalue λ = 2. On

the other hand,     
4 −1 6 −1 0
Au = 2 1 6  2  = 6 .
2 −1 8 1 4
There is no scalar λ such that    
0 −1
 6 = λ  2  .
4 1
Therefore, u is not an eigenvector of A.
Example 21.3. Is v an eigenvector of A? If yes, find the eigenvalue of A associated to v:

   
2 −1 −1 1
A = −1
 2 −1 , v = 1 .
 
−4 2 2 1
Solution. We compute  
0
Av = 0 = 0.
0
Hence, if λ = 0 then λv = 0 and thus Av = λv. Therefore, v is an eigenvector of A with
corresponding eigenvalue λ = 0.
How does one find the eigenvectors/eigenvalues of a matrix A? The general procedure
is to first find the eigenvalues of A and then for each eigenvalue find the corresponding
eigenvectors. In this section, however, we will instead suppose that we have already found
the eigenvalues of A and concern ourselves with finding the associated eigenvectors. Suppose
then that λ is known to be an eigenvalue of A. How do we find an eigenvector v corresponding
to the eigenvalue λ? To answer this question, we note that if v is to be an eigenvector of A
with eigenvalue λ then v must satisfy the equation
Av = λv.
We can rewrite this equation as

Av − λv = 0
which, after using the distributive property of matrix multiplication, is equivalent to
(A − λI)v = 0.
The last equation says that if v is to be an eigenvector of A with eigenvalue λ then v must
be in the null space of A − λI:
v ∈ Null(A − λI).
164
Lecture 21
In summary, if λ is known to be an eigenvalue of A, then to find the eigenvectors corre-

sponding to λ we must solve the homogeneous system
(A − λI)x = 0.
Recall that the null space of any matrix is a subspace and for this reason we call the subspace
Null(A − λI) the eigenspace of A corresponding to λ.
Example 21.4. It is known that λ = 4 is an eigenvalue of

 
−4 6 3
A= 1 7 9 .
8 −6 1
Find a basis for the eigenspace of A corresponding to λ = 4.
Solution. First compute
     
−4 6 3 4 0 0 −8 6 3
A − 4I =  1 7 9 − 0 4 0 =  1 3 9 
8 −6 1 0 0 4 8 −6 −3
Find a basis for the null space of A − 4I:
   
−8 6 3 1 3 9
R1 lR2
 1 3 9  −−−→  −8 6 3 
8 −6 −3 8 −6 −3
   
1 3 9 8R1 +R2 1 3 9
−8R1 +R3
 −8 6 3  −−−− −−→  0 30 75 
8 −6 −3 0 −30 −75
Finally,    
1 3 9 1 3 9
R +R3
 0 30 75  −−2−−→  0 30 75 
0 −30 −75 0 0 0
Hence, the general solution to the homogenous system (A − 4I)x = 0 is
 
−3/2
x = t −5/2
1
where t is an arbitrary scalar. Therefore, the eigenspace of A corresponding to λ = 4 is
   
 −3/2   −3 
span  −5/2  = span −5 = span{v}
   
1 2
and {v} is a basis for the eigenspace. The vector v is of course an eigenvector of A with
eigenvalue λ = 4 and also (of course) any multiple of v is also eigenvector of A with λ = 4.
165
Example 21.5. It is known that λ = 3 is an eigenvalue of

 
11 −4 −8
A= 4 1 −4  .
8 −4 −5
Find the eigenspace of A corresponding to λ = 3.

Solution. First compute
     
11 −4 −8 3 0 0 8 −4 −8
A − 3I =  4 1 −4  −  0 3 0  =  4 −2 −4 
8 −4 −5 0 0 3 8 −4 −8
Now find the null space of A − 3I:

   
8 −4 −8 4 −2 −4
R1 lR2
 4 −2 −4  − −−→  8 −4 −8 
8 −4 −8 8 −4 −8
   
4 −2 −4 −2R1 +R2 4 −2 −4
 8 −4 −8  −−2R
−−− 1 +R3
−−→  0 0 0 
8 −4 −8 0 0 0
Hence, any vector in the null space of
 
4 −2 −4
A − 3I =  0 0 0 
0 0 0
can be written as    
1 1
x = t1 0 + t2 2
  
1 0
Therefore, the eigenspace of A corresponding to λ = 3 is
   
 1 1 
Null(A − 3I) = span{v1 , v2 } = span 0 , 2 .
 
1 0
The vectors v1 and v2 are two linearly independent eigenvectors of A with eigenvalue λ = 3.
Therefore {v1 , v2 } is a basis for the eigenspace of A with eigenvalue λ = 3. You can verify
that Av1 = 3v1 and Av2 = 3v2 .
As shown in the last example, there may exist more than one linearly independent eigen-
vector of A corresponding to the same eigenvalue, in other words, it is possible that the
dimension of the eigenspace Null(A − λI) is greater than one. What can be said about the
eigenvectors of A corresponding to different eigenvalues?
166
Lecture 21
Theorem 21.6: Let v1 , . . . , vk be eigenvectors of A corresponding to distinct eigenvalues

λ1 , . . . , λk of A. Then {v1 , . . . , vk } is a linearly independent set.
Solution. Suppose by contradiction that {v1 , . . . , vk } is linearly dependent and {λ1 , . . . , λk }

are distinct. Then, one of the eigenvectors vp+1 that is a linear combination of v1 , . . . , vp ,
and {v1 , . . . , vp } is linearly independent:
vp+1 = c1 v1 + c2 v2 + · · · + cp vp . (21.1)
Applying A to both sides we obtain
Avp+1 = c1 Av1 + c2 Av2 + · · · + cp Avp
and since Avi = λi vi we can simplify this to
λp+1vp+1 = c1 λ1 v1 + c2 λ2 v2 + · · · + cp λp vp . (21.2)
On the other hand, multiply (21.1) by λp+1 :
λp+1 vp+1 = c1 λp+1v1 + c2 λp+1v2 + · · · + cp vp λp+1 . (21.3)
Now subtract equations (21.2) and (21.3):
0 = c1 (λ1 − λp+1)v1 + c2 (λ2 − λp+1)v2 + · · · + cp (λp − λp+1 )vp .
Now {v1 , . . . , vp } is linearly independent and thus ci (λi − λp+1 ) = 0. But the eigenvalues
{λ1 , . . . , λk } are all distinct and so we must have c1 = c2 = · · · = cp = 0. But from (21.1)
this implies that vp+1 = 0, which is a contradiction because eigenvectors are by definition
non-zero. This proves that {v1 , v2 , . . . , vk } is a linearly independent set.
Example 21.7. It is known that λ1 = 1 and λ2 = −1 are eigenvalues of

 
−4 6 3
A= 1 7 9 .
8 −6 1
Find bases for the eigenspaces corresponding to λ1 and λ2 and show that any two vectors
from these distinct eigenspaces are linearly independent.
Solution. Compute  
−5 6 3
A − λ1 I =  1 6 9 
8 −6 0
and one finds that  
 −3 
(A − λ1 I) = span −4
 
3
167
Hence, v1 = (−3, −4, 3) is an eigenvector of A with eigenvalue λ1 = 1, and {v1 } forms a

basis for the corresponding eigenspace. Next, compute
     
−4 6 3 1 0 0 −3 6 3
A − λ2 I =  1 7 9 + 0 1 0 =  1 8 9 
8 −6 1 0 0 1 8 −6 2
and one finds that  

 −1 
A − λ2 I = span −1
 
1
Hence, v2 = (−1, −1, 1) is an eigenvector of A with eigenvalue λ2 = −1, and {v2 } forms a
basis for the corresponding eigenspace. Now verify that v1 and v2 are linearly independent:
   
−3 −1 −3 −1
R +R3
v1 v2 = −4 −1 −−1−−→ −4 −1
3 1 0 0
The last matrix has rank r = 2, and thus v1 , v2 are indeed linearly independent.
21.2 When λ = 0 is an eigenvalue

What can we say about A if λ = 0 is an eigenvalue of A? Suppose then that A has eigenvalue
λ = 0. Then by definition, there exists a non-zero vector v such that
Av = 0 · v = 0.
In other words, v is in the null space of A. Thus, A is not invertible (Why?).
Theorem 21.8: The matrix A ∈ Rn×n is invertible if and only if λ = 0 is not an eigenvalue
of A.
In fact, later we will see that det(A) is the product of its eigenvalues.

• what eigenvalues are
• what eigenvectors are and how to find them when eigenvalues are known
• the behavior of a discrete dynamical system when the initial condition is set to an
eigenvector of the system matrix
168
Lecture 22
Lecture 22
The Characteristic Polynomial
22.1 The Characteristic Polynomial of a Matrix

Recall that a number λ is an eigenvalue of A ∈ Rn×n if there exists a non-zero vector v such
that
Av = λv
or equivalently if v ∈ Null(A − λI). In other words, λ is an eigenvalue of A if and only

if the subspace Null(A − λI) contains a vector other than the zero vector. We know that
any matrix M has a non-trivial null space if and only if M is non-invertible if and only if
det(M) = 0. Hence, λ is an eigenvalue of A if and only if λ satisfies det(A − λI) = 0. Let’s
compute the expression det(A − λI) for a generic 2 × 2 matrix:

a11 − λ a 12

det(A − λI) =
a21 a22 − λ
= (a11 − λ)(a22 − λ) − a12 a22
= λ2 − (a11 + a22 )λ + a11 a22 − a12 a22 .
Thus, if A is 2 × 2 then
det(A − λI) = λ2 − (a11 + a22 )λ + a11 a22 − a12 a22
is a polynomial in the variable λ of degree n = 2. This motivates the following definition.
Definition 22.1: Let A be a n × n matrix. The polynomial
p(λ) = det(A − λI)
is called the characteristic polynomial of A.
169
In summary, to find the eigenvalues of A we must find the roots of the characteristic poly-
nomial:
p(λ) = det(A − λI).
The following theorem asserts that what we observed for the case n = 2 is indeed true for
all n.
Theorem 22.2: The characteristic polynomial p(λ) = det(A − λI) of a n × n matrix A

is an nth degree polynomial.
Solution. Recall that for the case n = 2 we computed that
det(A − λI) = λ2 − (a11 + a22 )λ + a11 a22 − a12 a22 .
Therefore, the claim holds for n = 2. By induction, suppose that the claims hold for n ≥ 2.
If A is a (n + 1) × (n + 1) matrix then expanding det(A − λI) along the first row:
n
X
det(A − λI) = (a11 − λ) det(A11 − λI) + (−1)1+k a1k det(A1k − λI).
k=2
By induction, each of det(A1k −λI) is a nth degree polynomial. Hence, (a11 −λ) det(A11 −λI)
is a (n + 1)th degree polynomial. This ends the proof.
Example 22.3. Find the characteristic polynomial of

−2 4
A= .
−6 8
What are the eigenvalues of A?
Solution. Compute

−2 4 λ 0 −2 − λ 4
A − λI = − = .
−6 8 0 λ −6 8−λ
Therefore,
p(λ) = det(A − λI)

−2 − λ 4
=
−6 8 − λ
= (−2 − λ)(8 − λ) + 24
= λ2 − 6λ + 8
= (λ − 4)(λ − 2)
The roots of p(λ) are clearly λ1 = 4 and λ2 = 2. Therefore, the eigenvalues of A are λ1 = 4
and λ2 = 2.
170
Lecture 22
Example 22.4. Find the eigenvalues of

 
−4 −6 −7
A= 3 5 3 .
0 0 3
Solution. Compute
   
−4 −6 −7 λ 0 0
A − λI =  3 5 3  − 0 λ 0
0 0 3 0 0 λ
 
−4 − λ −6 −7
= 3 5−λ 3 
0 0 3−λ
Then

5 − λ 3 −6 −7
det(A − λI) = (−4 − λ) − 3
−λ 3 − λ −λ 3 − λ
= (−4 − λ)[(3 − λ)(5 − λ) + 3λ] − 3[−6(3 − λ) − 7λ]
= λ3 − 4λ2 + λ + 6
Factor the characteristic polynomial:
p(λ) = λ3 − 4λ2 + λ + 6 = (λ − 2)(λ − 3)(λ + 1)
Therefore, the eigenvalues of A are
λ1 = 2, λ2 = 3, λ3 = −1.
Now that we know how to find eigenvalues, we can combine our work from the previous
lecture to find both the eigenvalues and eigenvectors of a given matrix A.
Example 22.5. For each eigenvalue of A from Example 22.4, find a basis for the corre-
sponding eigenspace.
Solution. Start with λ1 = 2:
 
−6 −6 −7
A − 2I =  3 3 3 
0 0 1
After basic row reduction and back substitution, one finds that the null space of A − 2I is
spanned by  
−1
v1 =  1  .
0
171
Therefore, v1 is an eigenvector of A with eigenvalue λ1 . For λ2 = 3:

 
−7 −6 −7
A − 3I =  3 2 3 
0 0 0
The null space of A − 3I is spanned by

 
−1
v2 =  0 
1
and therefore v2 is an eigenvector of A with eigenvalue λ2 . Finally, for λ3 = −1 we compute

 
−3 −6 −7
A − λ3 I =  3 6 3 
0 0 4
and the null space of A − λ3 I is spanned by

 
−2
v3 =  1 
0
and therefore v3 is an eigenvector of A with eigenvalue λ3 . Notice that in this case, the 3 × 3
matrix A has three distinct eigenvalues and the eigenvectors
     
 −1 −1 −2 
{v1 , v2 , v3 } =  1 , 0 , 1
   
 
0 1 0
correspond to the distinct eigenvalues λ1 , λ2 , λ3 , respectively. Therefore, the set β = {v1 , v2 , v3 }

is linearly independent (by Theorem 21.6), and therefore β is a basis for R3 . You can verify,
for instance, that det([v1 v2 v3 ]) 6= 0.
By Theorem 21.6, the previous example has the following generalization.
Theorem 22.6: Suppose that A is a n × n matrix and has n distinct eigenvalues

λ1 , λ2 , . . . , λn . Let vi be an eigenvector of A corresponding to λi . Then {v1 , v2 , . . . , vn }
is a basis for Rn .
Hence, if A has distinct eigenvalues, we are guaranteed the existence of a basis of Rn

consisting of eigenvectors of A. In forthcoming lectures, we will see that it is very convenient
to work with matrices A that have a set of eigenvectors that form a basis of Rn ; this is one of
the main motivations for studying eigenvalues and eigenvectors in the first place. However,
we will see that not every matrix has a set of eigenvectors that form a basis of Rn . For
example, what if A does not have n distinct eigenvalues? In this case, does there exist a
172
Lecture 22
basis for Rn of eigenvectors of A? In some cases, the answer is yes as the next example
demonstrates.
Example 22.7. Find the eigenvalues of A and a basis for each eigenspace.
 
2 0 0
A=  4 2 2 
−2 0 1
Does R3 have a basis of eigenvectors of A?

Solution. The characteristic polynomial of A is
p(λ) = det(A − λI) = λ3 − 5λ2 + 8λ − 4 = (λ − 1)(λ − 2)2
and therefore the eigenvalues are λ1 = 1 and λ2 = 2. Notice that although p(λ) is a
polynomial of degree n = 3, it has only two distinct roots and hence A has only two
distinct eigenvalues. The eigenvalue λ2 = 2 is said to be repeated and λ1 = 1 is said to be
a simple eigenvalue. For λ1 = 1 one finds that the eigenspace Null(A − λ1 I) is spanned by
 
0
v1 = −2
1
and thus v1 is an eigenvector of A with eigenvalue λ1 = 1. Now consider λ2 = 2:
 
0 0 0
A − 2I =  4 0 2 
−2 0 −1
Row reducing A − 2I one obtains
   
0 0 0 −2 0 −1
A − 2I =  4 0 2  ∼  0 0 0 .
−2 0 −1 0 0 0
Therefore, rank(A − 2I) = 1 and thus by the Rank Theorem it follows that Null(A − 2I) is
a 2-dimensional eigenspace. Performing back substitution, one finds the following basis for
the λ2 -eigenspace:    
 −1 0 
{v2 , v3 } =  0  , 1
 
2 0
Therefore, the eigenvectors
     
 0 −1 0 
{v1 , v2 , v3 } = −2 , 0 , 1
   
 
1 2 0
form a basis for R3 . Hence, for the repeated eigenvalue λ2 = 2 we were able to find two
linearly independent eigenvectors.
173
Before moving further with more examples, we need to introduce some notation regard-
ing the factorization of the characteristic polynomial. In the previous Example 22.7, the
characteristic polynomial was factored as p(λ) = (λ − 1)(λ − 2)2 and we found a basis for
R3 of eigenvectors despite the presence of a repeated eigenvalue. In general, if p(λ) is an
nth degree polynomial that can be completely factored into linear terms, then p(λ) can be
written in the form
p(λ) = (λ − λ1 )k1 (λ − λ2 )k2 · · · (λ − λp )kp
where k1 , k2 , . . . , kp are positive integers and the roots of p(λ) are then λ1 , λ2 , . . . , λk . Because
p(λ) is of degree n, we must have that k1 + k2 + · · ·+ kp = n. Motivated by this, we introduce
the following definition.
Definition 22.8: Suppose that A ∈ Mn×n has characteristic polynomial p(λ) that can be
factored as
p(λ) = (λ − λ1 )k1 (λ − λ2 )k2 · · · (λ − λp )kp .
The exponent ki is called the algebraic multiplicity of the eigenvalue λi . The dimension
Null(A − λi I) of the eigenspace associated to λi is called the geometric multiplicity of
λi .
For simplicity and whenever it is convenient, we will denote the geometric multiplicity of the
eigenvalue λi as
gi = dim(Null(A − λi I)).
Example 22.9. A 6 × 6 matrix A has characteristic polynomial
p(λ) = λ6 − 4λ5 − 12λ4 .
Find the eigenvalues of A and their algebraic multiplicities.

Solution. Factoring p(λ) we obtain
p(λ) = λ4 (λ2 − 4λ − 12) = λ4 (λ − 6)(λ + 2)
Therefore, the eigenvalues of A are λ1 = 0, λ2 = 6, and λ3 = −2. Their algebraic multiplic-

ities are k1 = 4, k2 = 1, and k3 = 1, respectively. The eigenvalue λ1 = 0 is repeated, while
λ2 = 6 and λ3 = −2 are simple eigenvalues.
In Example 22.7, we had p(λ) = (λ−1)(λ−2)2 and thus λ1 = 1 has algebraic multiplicity
k1 = 1 and λ2 = 2 has algebraic multiplicity k2 = 2. For λ1 = 1, we found one linearly
independent eigenvector, and therefore λ1 has geometric multiplicity g1 = 1. For λ1 = 2,
we found two linearly independent eigenvectors, and therefore λ2 has geometric multiplicity
g2 = 2. However, as we will see in the next example, the geometric multiplicity gi is in
general less than the algebraic multiplicity ki :
g i ≤ ki
174
Lecture 22
Example 22.10. Find the eigenvalues of A and a basis for each eigenspace:
 
2 4 3
A = −4 −6
 −3
3 3 1
For each eigenvalue of A, find its algebraic and geometric multiplicity. Does R3 have a basis
of eigenvectors of A?
Solution. One computes
p(λ) = −λ3 − 3λ2 + 4 = −(λ − 1)(λ + 2)2
and therefore the eigenvalues of A are λ1 = 1 and λ2 = −2. The algebraic multiplicity of λ1
is k1 = 1 and that of λ2 is k2 = 2. For λ1 = 1 we compute
 
1 4 3
A − I = −4 −7 −3
3 3 0
and then one finds that  
1
v1 = −1
1
is a basis for the λ1 -eigenspace. Therefore, the geometric multiplicity of λ1 is g1 =. For
λ2 = −2 we compute
     
4 4 3 4 4 3 1 1 1
A − λ2 I = −4 −4 −3 ∼ 1 1 1 ∼ 0 0 1
3 3 3 0 0 0 0 0 0
Therefore, since rank(A − λ2 I) = 2, the geometric multiplicity of λ2 = −2 is g2 = 1, which
is less than the algebraic multiplicity k2 = 2. An eigenvector corresponding to λ2 = −2 is
 
−1
v2 = 1 

0
Therefore, for the repeated eigenvalue λ2 = −2, we are able to find only one linearly inde-
pendent eigenvector. Therefore, it is not possible to construct a basis for R3 consisting of
eigenvectors of A.
Hence, in the previous example, there does not exist a basis of R3 of eigenvectors of A
because for one of the eigenvalues (namely λ2 ) the geometric multiplicity was less than the
algebraic multiplicity:
g2 < d2 .
In the next lecture, we will elaborate on this situation further.
Example 22.11. Find the algebraic and geometric multiplicities of each eigenvalue of the
matrix  
−7 1 0
A =  0 −7 1  .
0 0 −7
175
22.2 Eigenvalues and Similarity Transformations

To end this lecture, we will define a notion of similarity between matrices that plays an
important role in linear algebra and that will be used in the next lecture when we dis-
cuss diagonalization of matrices. In mathematics, there are many cases where one is inter-
ested in classifying objects into categories or classes. Classifying mathematical objects into
classes/categories is similar to how some physical objects are classified. For example, all
fruits are classified into categories: apples, pears, bananas, oranges, avocados, etc. Given a
piece of fruit A, how do you decide what category it is in? What are the properties that
uniquely classify the piece of fruit A? In linear algebra, there are many objects of interest.
We have spent a lot of time working with matrices and we have now reached a point in our
study where we would like to begin classifying matrices. How should we decide if matrices
A and B are of the same type or, in other words, are similar? Below is how we will decide.
Definition 22.12: Let A and B be n × n matrices. We will say that A is similar to B

if there exists an invertible matrix P such that
A = PBP−1.
If A is similar to B then B is similar to A because from the equation A = PBP−1 we

can multiply on the left by P−1 and on the right by P to obtain that
P−1 AP = B.
Hence, with Q = P−1 , we have that B = QAQ−1 and thus B is similar to A. Hence, if A is
similar to B then B is similar to A and therefore we simply say that A and B are similar.
Matrices that are similar are clearly not necessarily equal. However, there is a reason why
the word similar is used. Here are a few reasons why.
Theorem 22.13: If A and B are similar matrices then the following are true:
(a) rank(A) = rank(B)
(b) det(A) = det(B)
(c) A and B have the same eigenvalues
Proof. We will prove part (c). If A and B are similar then A = PAP−1 for some matrix P.
Then
det(A − λI) = det(A − λPP−1 )
= det(PBP−1 − λPP−1)
= det(P(B − λI)P−1)
= det(P) det(B − λI) det(P−1 )
= det(B − λI)
176
Lecture 22
Thus, A and B have the same characteristic polynomial, and hence the same eigenvalues.
In the next lecture, we will see that if Rn has a basis of eigenvectors of A then A is similar
to a diagonal matrix.

• what the characteristic polynomial is and how to compute it
• how to compute the eigenvalues of a matrix
• that when a matrix A has distinct eigenvalues, we are guaranteed a basis of Rn con-
sisting of the eigenvectors of A
• that when a matrix A has repeated eigenvalues, it is still possible that there exists a
basis of Rn consisting of the eigenvectors of A
• what is the algebraic multiplicity and geometric multiplicity of an eigenvalue
• that eigenvalues of a matrix do not change under similarity transformations
177
178
Lecture 23
Diagonalization
23.1 Eigenvalues of Triangular Matrices

Before discussing diagonalization, we first consider the eigenvalues of triangular matrices.
Theorem 23.1: Let A be a triangular matrix (either upper or lower). Then the eigen-
values of A are its diagonal entries.
Proof. We will prove the theorem for the case n = 3 and A is upper triangular; the general
case is similar. Suppose then that A is a 3 × 3 upper triangular matrix:
 
a11 a12 a13
A =  0 a22 a23 
0 0 a33
Then  
a11 − λ a12 a13
A − λI =  0 a22 − λ a23  .
0 0 a33 − λ
and thus the characteristic polynomial of A is
p(λ) = det(A − λI) = (a11 − λ)(a22 − λ)(a33 − λ)
and the roots of p(λ) are
λ1 = a11 , λ2 = a22 , λ3 = a33 .
In other words, the eigenvalues of A are simply the diagonal entries of A.
Example 23.2. Consider the following matrix

 
6 0 0 0 0
−1 0 0 0 0
 
A= 0 0 7 0 0.
−1 0 0 −4 0
8 −2 3 0 7
Diagonalization
(a) Find the characteristic polynomial and the eigenvalues of A.

(b) Find the geometric and algebraic multiplicity of each eigenvalue of A.
We now introduce a very special type of a triangular matrix, namely, a diagonal matrix.
Definition 23.3: A matrix D whose off-diagonal entries are all zero is called a diagonal
matrix.
For example, here is 3 × 3 diagonal matrix

 
3 0 0
D = 0 −5 0  .
0 0 −8
and here is a 5 × 5 diagonal matrix

 
6 0 0 0 0
0 0 0 0 0 
 
0
D= 0 − 72 0 0  .
0 0 0 2 0 
1
0 0 0 0 − 11
A diagonal matrix is clearly also a triangular matrix and therefore the eigenvalues of a
diagonal matrix D are simply the diagonal entries of D. Moreover,
the powers of a diagonal
λ 0
matrix are easy to compute. For example, if D = 1 then
0 λ2
2
2 λ1 0 λ1 0 λ1 0
D = =
0 λ2 0 λ2 0 λ22
and similarly for any integer k = 1, 2, 3, . . ., we have that

k
k λ1 0
D = .
0 λk2
23.2 Diagonalization
Recall that two matrices A and B are said to be similar if there exists an invertible matrix
P such that
A = PBP−1.
A very simple type of matrix is a diagonal matrix since many computations with diagonal
matrices are trivial. The problem of diagonalization is thus concerned with answering the
question of whether a given matrix is similar to a diagonal matrix. Below is the formal
definition.
180
Lecture 23
Definition 23.4: A matrix A is called diagonalizable if it is similar to a diagonal matrix

D. In other words, if there exists an invertible P such that
A = PDP−1 .
How do we determine when a given matrix A is diagonalizable? Let us first determine what
conditions need to be met for a matrix A to be diagonalizable. Suppose then
that A is diag-
onalizable. Then by Definition 23.4, there exists an invertible matrix P = v1 v2 · · · vn
and a diagonal matrix  
λ1 0 . . . 0
 0 λ2 . . . 0 
 
D =  .. .. . . .. 
. . . .
0 0 . . . λn
such that A = PDP−1. Multiplying on the right both sides of the equation A = PDP−1
by the matrix P we obtain that
AP = PD.
Now
AP = Av1 Av2 · · · Avn
while on the other hand
PD = λ1 v1 λ2 v2 · · · λn vn .
Therefore, since it holds that AP = PD then

Av1 Av2 · · · Avn = λ1 v1 λ2 v2 · · · λn vn .
or if we compare columns we must have that
Avi = λi vi .
Thus, the columns v1 , v2 , . . . , vn of P are eigenvectors of A and form a basis for Rn because
P is invertible. In conclusion, if A is diagonalizable then Rn has a basis consisting of
eigenvectors of A.
Suppose instead that {v1 , v2 , . . . , vn } is a basis of Rn consisting of eigenvectors of A. Let
λ1 , λ2 , . . . , λn be the eigenvalues of A associated to v1 , v2 , . . . , vn , respectively, and set

P = v1 v2 · · · vn .
Then P is invertible because {v1 , v2 , . . . , vn } are linearly independent. Let

 
λ1 0 . . . 0
 0 λ2 . . . 0 
 
D =  .. .. . . ..  .
. . . .
0 0 . . . λn
181
Diagonalization
Now, since Avi = λi vi we have that

AP = A v1 v2 · · · vn

= Av1 Av2 · · · Avn

= λ1 v1 λ2 v2 · · · λn vn .

Therefore, AP = λ1 v1 λ2 v2 · · · λn vn . On the other hand,
 
λ1 0 . . . 0
 0 λ2 . . . 0 

PD = v1 v2 · · · vn  .. .. . . .
. . . .. 
0 0 . . . λn

= λ1 v1 λ2 v2 · · · λn vn .
Therefore, AP = PD, and since P is invertible we have that
A = PDP−1.
Thus, if Rn has a basis of consisting of eigenvectors of A then A is diagonalizable. We have
therefore proved the following theorem.
Theorem 23.5: A matrix A is diagonalizable if and only if there is a basis {v1 , v2 , . . . , vn }

of Rn consisting of eigenvectors of A.
The punchline with Theorem 23.5 is that the problem of diagonalization of a matrix A
is equivalent to finding a basis of Rn consisting of eigenvectors of A. We will see in some of
the examples below that it is not always possible to diagonalize a matrix.
23.3 Conditions for Diagonalization

We first consider the simplest case when we conclude that a given matrix is diagonalizable,
namely, the case when all eigenvalues are distinct.
Theorem 23.6: Suppose that A ∈ Rn×n has n distinct eigenvalues λ1 , λ2 , . . . , λn . Then

A is diagonalizable.
Proof. Each eigenvalue λi produces an eigenvector vi . The set of eigenvectors {v1 , v2 , . . . , vn }

are linearly independent because they correspond to distinct eigenvalues (Theorem 21.6).
Therefore, {v1 , v2 , . . . , vn } is a basis of Rn consisting of eigenvectors of A and then by
Theorem 23.5 we conclude that A is diagonalizable.
What if A does not have distinct eigenvalues? Can A still be diagonalizable? The
following theorem completely answers this question.
182
Lecture 23
Theorem 23.7: A matrix A is diagonalizable if and only if the algebraic and geometric
multiplicities of each eigenvalue are equal.
Proof. Let A be a n × n matrix and let λ1 , λ2 , . . . , λp denote the distinct eigenvalues of A.

Suppose that k1 , k2 , . . . , kp are the algebraic multiplicities and g1 , g2, . . . , gp are the geometric
multiplicities of the eigenvalues, respectively. Suppose that the algebraic and geometric
multiplicities of each eigenvalue are equal, that is, suppose that gi = ki for each i = 1, 2 . . . , p.
Since k1 +k2 +· · ·+kp = n, then because gi = ki we must also have that g1 +g2 +· · ·+gp = n.
Therefore, there exists n linearly eigenvectors of A and consequently A is diagonalizable.
On the other hand, suppose that A is diagonalizable. Since the geometric multiplicity is at
most the algebraic multiplicity, the only way that g1 + g2 + · · · + gp = n is if gi = ki , i.e.,
that the geometric and algebraic multiplicities are equal.
Example 23.8. Determine if A is diagonalizable. If yes, find a matrix P that diagonalizes

A.  
−4 −6 −7
A= 3 5 3 
0 0 3
p(λ) = det(A − λI) = (λ − 2)(λ − 3)(λ + 1)
and therefore λ1 = 2, λ2 = 3, and λ3 = −1 are the eigenvalues of A. Since A has n =

3 distinct eigenvalues, then by Theorem 23.6 A is diagonalizable. Eigenvectors v1 , v2 , v3
corresponding to λ1 , λ2 , λ3 are found to be
     
−1 −1 −2
v1 =  1  , v2 =  0  , v3 =  1 
0 1 0
Therefore, a matrix that diagonalizes A is

 
−1 −2 −2
P= 1 0 1
0 1 0
You can verify that  

λ1 0 0
P  0 λ2 0  P−1 = A
0 0 λ3
The following example demonstrates that it is possible for a matrix to be diagonalizable

even though the matrix does not have distinct eigenvalues.
183
Diagonalization

A.  
2 0 0
A= 4 2 2 
−2 0 1
p(λ) = det(A − λI) = (λ − 1)(λ − 2)2
and therefore λ1 = 1, λ2 = 2. An eigenvector corresponding to λ1 = 1 is
 
0
v1 = −2
1
One finds that g2 = dim(Null(A − λ2 I)) = 2, and two linearly independent eigenvectors for
λ2 are    
 −1 0 
{v2 , v3 } =  0 , 1
 
 
2 0
Therefore, A is diagonalizable, and a matrix that diagonalizes A is
 
0 −1 0
P = v1 v2 v3 = −2 0 1
1 2 0
You can verify that  
λ1 0 0
P  0 λ2 0  P−1 = A
0 0 λ3

A.  
2 4 3
A = −4 −6 −3
3 3 1
p(λ) = det(A − λI) = −λ3 − 3λ2 + 4 = −(λ − 1)(λ + 2)2
and therefore the eigenvalues of A are λ1 = 1 and λ2 = −2. For λ2 = −2 one computes
 
1 1 1
A − λ2 I ∼ 0 0 1
0 0 0
We see that the dimension of the eigenspace of λ2 = −2 is g2 = 1, which is less than the
algebraic multiplicity k2 = 2. Therefore, from Theorem 23.7 we can conclude that it is not
possible to construct a basis of eigenvectors of A, and therefore A is not diagonalizable.
184
Lecture 23
Example 23.11. Suppose that A has eigenvector v with corresponding eigenvalue λ. Show
that if A is invertible then v is an eigenvector of A−1 with corresponding eigenvalue λ1 .
Example 23.12. Suppose that A and B are n × n matrices such that AB = BA. Show
that if v is an eigenvector of A with corresponding eigenvalue then v is also an eigenvector
of B with corresponding eigenvalue λ.
• Determine if a matrix is diagonalizable or not
• Find the algebraic and geometric multiplicities of an eigenvalue
• Apply the theorems introduced in this lecture
185
Diagonalization
186
Lecture 24
Lecture 24
Diagonalization of Symmetric
Matrices
24.1 Symmetric Matrices

Recall that a square matrix A is said to be symmetric if AT = A. As an example, here is
a 3 × 3 symmetric matrix:  
1 −3 7
A = −3 2 8 .
7 8 4
Symmetric matrices are ubiquitous in mathematics. For example, let f (x1 , x2 , . . . , xn ) be a
function having continuous second order partial derivatives. Then Clairaut’s Theorem from
multivariable calculus says that
∂f ∂f
= .
∂xi ∂xj ∂xj ∂xi
Therefore, the Hessian matrix of f is symmetric:
 
∂f ∂f ∂f
∂x1 ∂x1 ∂x1 ∂x2
··· ∂x1 ∂xn
 
 ∂f ∂f ∂f 

Hess(f ) =  ∂x2 ∂x1 ∂x2 ∂x2
··· 
∂x2 ∂xn  .
 .. .. .. .. 
 . . . . 
∂f ∂f ∂f
∂xn ∂x1 ∂xn ∂x2
··· ∂xn ∂xn
The Second Derivative Test of multivariable calculus then says that if P = (a1 , a2 , . . . , an )
is a critical point of f , that is
∂f ∂f ∂f
(P ) = (P ) = · · · = (P ) = 0
∂x1 ∂x2 ∂xn
then
(i) P is a local minimum point of f if the matrix Hess(f ) has all positive eigenvalues,
(ii) P is a local maximum point of f if the matrix Hess(f ) has all negative eigenvalues,
and
187
Diagonalization of Symmetric Matrices
(iii) P is a saddle point of f if the matrix Hess(f ) has negative and positive eigenvalues.
In general, the eigenvalues of a matrix with real entries can be complex numbers. For
example, the matrix
0 −1
A=
1 0
has characteristic polynomial
p(λ) = λ2 + 1
√ √
the roots of which are clearly λ1 = −1 = i and λ2 = − −1 = −i. Thus, in general,
a matrix whose entries are all real numbers may have complex eigenvalues. However, for
symmetric matrices we have the following.
Theorem 24.1: If A is a symmetric matrix then all of its eigenvalues are real numbers.
The proof is easy but we will omit it.
24.2 Eigenvectors of Symmetric Matrices

We proved earlier that if {v1 , v2 , . . . , vk } are eigenvectors of a matrix A corresponding to
distinct eigenvalues λ1 , λ2 , . . . , λk then the set {v1 , v2 , . . . , vk } is linearly independent (The-
orem 21.6). For symmetric matrices we can say even more as the next theorem states.
Theorem 24.2: Let A be a symmetric matrix. If v1 and v2 are eigenvectors of A

corresponding to distinct eigenvalues then v1 and v2 are orthogonal, that is, v1 • v2 = 0.
Proof. Recall that v1 • v2 = v1T v2 . Let λ1 6= λ2 be the eigenvalues associated to v1 and v2 .

Then
λ1 v1T v2 = (λ1 v1 )T v2
= (Av1 )T v2
= v1T AT v2
= v1T Av2
= v1T (λ2 v2 )
= λ2 v1T v2 .
Therefore, λ1 v1T v2 = λ2 v1T v2 which implies that (λ1 − λ2 )v1T v2 = 0. But since (λ1 − λ2 ) 6= 0
then we must have v1T v2 = 0, that is, v1 and v2 are orthogonal.
24.3 Symmetric Matrices are Diagonalizable

As we have seen, the main criteria for diagonalization is that for each eigenvalue the geometric
and algebraic multiplicities are equal; not all matrices satisfy this condition and thus not
188
Lecture 24
all matrices are diagonalizable. As it turns out, any symmetric A is diagonalizable and
moreover (and perhaps more importantly) there exists an orthogonal eigenvector matrix P
that diagonalizes A. The full statement is below.
Theorem 24.3: If A is a symmetric matrix then A is diagonalizable. In fact, there is an

orthonormal basis of Rn of eigenvectors {v1 , v2 , . . . , vn } of A. In other words, the matrix
P = [v1 v2 · · · vn ] is orthogonal, PT P = I, and A = PDPT .
The proof of the theorem is not hard but we will omit it. The punchline of Theorem 24.3
is that, for the case of a symmetric matrix, we will never encounter the situation where
the geometric multiplicity is strictly less than the algebraic multiplicity. Moreover, we are
guaranteed to find an orthogonal matrix that diagonalizes a given symmetric matrix.
Example 24.4. Find an orthogonal matrix P that diagonalizes the symmetric matrix
 
1 0 −1
A= 0 1 1 .
−1 1 2
p(λ) = det(A − λI) = λ3 − 4λ2 + 3λ = λ(λ − 1)(λ − 3)
The eigenvalues of A are λ1 = 0, λ2 = 1 and λ3 = 3. Eigenvectors of A associated to
λ1 , λ2 , λ3 are      
1 1 −1
u1 = −1 , u2 = 1 , u3 = 1  .
    
1 0 2
As expected by Theorem 24.2, the eigenvectors u1 , u2 , u3 form an orthogonal set:
uT1 u2 = 0, uT1 u3 = 0, uT2 u3 = 0.
To find an orthogonal matrix P that diagonalizes A we must normalize the eigenvectors
u1 , u2 , u3 to obtain an orthonormal basis {v1 , v2 , v3 }. To that end, first compute uT1 u1 = 3,
uT2 u2 = 2, and uT3 u3 = 6. Then let v1 = √13 u1 , let v2 = √12 u2 , and let v3 = √16 u3 . Therefore,
an orthogonal matrix that diagonalizes A is
 1 
√ √1 − √16
3 2
 − √1 √1 1 

P = v1 v2 v3 =  3 2
√
6 
 
√1 0 √2
3 6
You can easily verify that PT P = I, and that

 
0 0 0
A = P 0 1 0 PT
0 0 3
189
Diagonalization of Symmetric Matrices
Example 24.5. Let A and B be n × n matrices. Show that if A is symmetric then the
matrix C = BABT is also a symmetric matrix.
• a symmetric matrix is diagonalizable with an orthonormal set of eigenvectors
190
Lecture 25
The PageRank Algortihm
In this lecture, we will see how linear algebra is used in Google’s webpage ranking algorithm
used in everyday Google searches.
25.1 Search Engine Retrieval Process

Search engines perform a two-stage process to retrieve search results1 . In Stage 1, traditional
text processing is used to find all relevant pages (e.g. keywords in title, body) and produces
a content score. After Stage 1, there is a large amount of relevant pages. For example,
the query “symmetric matrix ” results in about 3,830,000 pages (03/31/15). Or “homework
help” results in 49,400,000 pages (03/31/15). How should the relevant pages be displayed?
In Stage 2, the pages are sorted and displayed based on a pre-computed ranking that is
query-independent, this is the popularity score. The ranking is based on the hyperlinked
or networked structure of the web, and the ranking is based on a popularity contest; if many
pages link to page Pi then Pi must be an important page and should therefore have a high
popularity score.
In January 1998, John Kleinberg from IBM (now a CS professor at Cornell) presented
the HITS algorithm2 (e.g., www.teoma.com). At Stanford, doctoral students Sergey Brin
and Larry Page were busy working on a similar project which they had begun in 1995. Below
is the abstract of their paper3 :
“In this paper, we present Google, a prototype of a large-scale search engine which makes
heavy use of the structure present in hypertext. Google is designed to crawl and index the
Web efficiently and produce much more satisfying search results than existing systems. The
prototype with a full text and hyperlink database of at least 24 million pages is available at
http://google.stanford.edu/ .”
1
A.N. Langville and C.D. Meyer, Google’s PageRank and Beyond, Princeton University Press, 2006
2
J. Kleinberg, Authoritative sources in a hyperlinked environment, Journal of ACM, 46, 1999, 9th ACM-
SIAM Symposium on Discrete Algorithms
3
S. Brin and L. Page, The anatomy of a large-scale hypertextual Web search engine, Computer Networks
and ISDN Systems, 33:107-117, 1998
In both models, the web is defined as a directed graph, where the nodes represent
webpages and the directed arcs represent hyperlinks, see Figure 25.1.
2 3
Figure 25.1: A tiny web represented as a directed graph.
25.2 A Description of the PageRank Algorithm

In the PageRank algorithm, each inlink is viewed as a recommendation (or vote). In general,
pages with many inlinks are more important than pages with few inlinks. However, the
quality of the inlink (vote) is important. The vote of each page should be divided by the
total number of recommendations made by the page. The PageRank of page i, denoted xi ,
is the sum of all the weighted PageRanks of all the pages pointing to i:
X xj
xi =
j→i
|Nj |
where
(1) Nj is the number of outlinks from page j
(2) j → i means page j links to page i

Example 25.1. Find the PageRank of each page for the network in Figure 25.1.
From the previous example, we see that the PageRank of each page can be found by
solving an eigenvalue/eigenvector problem. However, when dealing with large networks such
as the internet, the size of the problem is in the billions (8.1 billion in 2006) and directly
solving the equations is not possible. Instead, an iterative method called the power method
is used. One starts with an initial guess, say x0 = ( 14 , 41 , 41 , 41 ). Then one updates the guess
by computing
x1 = Hx0 .
In other words, we have a discrete dynamical system
xk+1 = Hxk .
A natural question is under what conditions will the the limiting value of the sequence
lim xk = lim (Hk x0 ) = q

k→∞ k→∞
192
Lecture 25
converge to an equilibrium of H? Also, if lim xk exists, will it be a positive vector? And

k→∞
lastly, can x0 6= 0 be chosen arbitrarily? To see what situations may occur, consider the
network displayed in Figure 25.2. Starting with x0 = ( 51 , . . . , 51 ) we obtain that for k ≥ 39,
the vectors xk = Hk x0 cycle between (0, 0, 0, 0.28, 0.40) and (0, 0, 0, 0.40, 0.28). Therefore,
the sequence x0 , x1 , x2 , . . . does not converge. The reason for this is that nodes 4 and 5 form
a cycle.
 1

0
0 0 0
3
4 5  
 1 
 0 0 2
0 0
 
 1 
2 3 H = 0 3 0 0 0


 
 1 1 
0 3 2 0 1
 
1 0 0 0 1 0
Figure 25.2: Cycles present in the network
Now consider the network displayed in Figure 25.3. If we remove the cycle we are still
left with a dangling node, namely node 1 (e.g. pdf file, image file). Starting with x0 =
( 15 , . . . , 15 ) results in
lim xk = 0.
k→∞
Therefore, in this case the sequence x0 , x1 , x2 , . . . converges to a non-positive vector, which

for the purposes of ranking pages would be an undesirable situation.
 1 
0 3 0 0 0
4 5  
0 0 21 12 0
 
 
 1 
2 3 H=0
 3 0 0 0 

 
 1 1 
0 3 2 0 1
 
1 0 0 0 12 0
Figure 25.3: Dangling node present in the network
To avoid the presence of dangling nodes and cycles, Brin and Page used the notion of
a random surfer to adjust H. To deal with a dangling node, Brin and Page replaced
the associated zero-column with the vector n1 1 = ( n1 , n1 , . . . , n1 ). The justification for this
adjustment is that if a random surfer reaches a dangling node, the surfer will “teleport” to
any page in the web with equal probability. The new updated hyperlink matrix H∗ may still
not have the desired properties. To deal with cycles, a surfer may abandon the hyperlink
structure of the web by ocassionally moving to a random page by typing its address in the
193
browser. With these adjustments, a random surfer now spends only a proportion of his
time using the hyperlink structure of the web to visit pages. Hence, let 0 < α < 1 be
the proportion of time the random surfer uses the hyperlink structure. Then the transition
matrix is
G = αH∗ + (1 − α) n1 J.
The matrix G goes by the name of the Google matrix, and it is reported that Google uses
α = 0.85 (here J is the all ones matrix). The Google matrix G is now a primitive and
stochastic matrix. Stochastic means that all its columns are probability vectors, i.e., non-
negative vectors whose components sum to 1. Primitive means that there exists k ≥ 1 such
that Gk has all positive entries (k = 1 in our case). With these definitions, we now have the
following theorem.
Theorem 25.2: If G is a primitive stochastic matrix then:
(i) There is a stochastic G∗ such that limk→∞ Gk = G∗ .

(ii) G∗ = q q · · · q where q is a probability vector.
(iii) For any probability vector q0 we have limk→∞ Gk q0 = q.
(iv) The vector q is the unique probability vector which is an eigenvector of G with
eigenvalue λ1 = 1.
(v) All other eigenvalues λ2 , . . . , λn have |λj | < 1.
Proof. We will prove a special case4 . Assume for simplicity that G is positive (this is the
case of the Google Matrix). If x = Gx, and x has mixed signs, then

Xn X n

|xi | = G x
ij j < Gij |xj |.

j=1 j=1
Then
n
X n X
X n X
|xi | < Gij |xj | = |xj |
i=1 i=1 j=1 j=1
which is a contradiction. Therefore, all the eigenvectors in the λ1 = 1 eigenspace are either
negative or positive. One then shows that the eigenspace corresponding to λ1 = 1 is 1-
dimensional. This proves that there is a unique probability vector q such that
q = Gq.
4
K. Bryan, T. Leise, The $25,000,000,000 Eigenvector: The Linear Algebra Behind Google, SIAM Review,
48(3), 569-581
194
Lecture 25
Let λ1 , λ2 , . . . , λn be the eigenvalues of G. We know that λ1 = 1 is a dominant eigenvalue:
|λ1 | > |λj |, j = 2, 3, . . . , n.
Let q0 be a probability vector and let q be as above, and let v2 , . . . , vn be the remaining
eigenvectors of G. Then q0 = q + c2 v2 + · · · + cn vn and therefore
Gk q0 = Gk (q + c2 v2 + · · · + cn vn )
= Gk q + c2 Gk v2 + · · · + cn Gk vn
= q + c2 λk2 v2 + · · · + cn λkn vn .
From this we see that

lim Gk q0 = q.
k→∞
25.3 Computation of the PageRank Vector

The Google matrix G is completely dense, which is computationally undesirable. Fortu-
nately,
G = αH∗ + (1 − α) n1 eeT
= α(H + n1 11T ) + (1 − α) n1 11T
= αH + (αa + (1 − α)1) n1 1T
and H is very sparse and requires minimal storage. A vector-matrix multiplication generally
requires O(n2) computation (n ≈ 8, 000, 000, 000 in 2006). Estimates show that the average
webpage has about 10 outlinks, so H has about 10n non-zero entries. This means that
multiplication with H reduces to O(n) computation. Aside from being very simple, the
power method is a matrix-free method, i.e., no manipulation of the matrix H is done. Brin
and Page, and others, have confirmed that only 50-100 iterations are needed for a satisfactory
approximation of the PageRank vector q for the web.
• Setup a Google matrix and compute PageRank vector
195
196
Lecture 26
Lecture 26
Discrete Dynamical Systems
26.1 Discrete Dynamical Systems

Many interesting problems in engineering, science, and mathematics can be studied within
the framework of discrete dynamical systems. Dynamical systems are used to model systems
that change over time. The state of the system (economic, ecologic, engineering, etc.) is
measured at discrete time intervals producing a sequence of vectors x0 , x1 , x2 , . . .. The
relationship between the vector xk and the next vector xk+1 is what constitutes a model.
Definition 26.1: A linear discrete dynamical system on Rn is an infinite sequence

{x0 , x1 , x2 , . . .} of vectors in Rn and a matrix A such that
xk+1 = Axk .
The vectors xk are called the state of the dynamical system and x0 is the initial condition
of the system. Once the initial condition x0 is fixed, the remaining state vectors x1 , x2 , . . . ,
can be found by iterating the equation xk+1 = Axk .
26.2 Population Model

Consider the dynamic system consisting of the population movement between a city and its
suburbs. Let x ∈ R2 be the state population vector whose first component is the population
of the city and the second component is the population of the suburbs:

c
x= .
s
For simplicity, we assume that c + s = 1, i.e., c and s are population percentages of the total
population. Suppose that in the year 1900, the city population was c0 and the suburban
population was s0 . Suppose it is known that after each year 5% of the city’s population
197
moves to the suburbs and that 3% of the suburban population moves to the city. Hence, the
population in the city in year 1901 is
c1 = 0.95c0 + 0.03s0 ,
while the population in the suburbs in year 1901 is
s1 = 0.05c0 + 0.97s0 .
The equations
c1 = 0.95c0 + 0.03s0
s1 = 0.05c0 + 0.97s0
can be written in matrix form as
" # " #" #
c1 0.95 0.03 c0
= .
s1 0.05 0.97 s0
Performing the same analysis for the next year, the population in 1902 is
" # " #" #
c2 0.95 0.03 c1
= .
s2 0.05 0.97 s1
Hence, the population movement is a linear dynamical system with matrix and state vector
" # " #
0.95 0.03 ck
A= , xk = .
0.05 0.97 sk
Suppose that the initial population state vector is
" #
0.70
x0 = .
0.30
Then, " #" # " #
0.95 0.03 0.70 0.674
x1 = Ax0 = = .
0.05 0.97 0.30 0.326
Then, " #" # " #
0.95 0.03 0.674 0.650
x2 = Ax1 = = .
0.05 0.97 0.326 0.349
In a similar fashion, one can compute that up to 3 decimal places:
" # " #
0.375 0.375
x500 = , x1000 = .
0.625 0.625
It seems as though the population distribution converges to a steady state or equilibrium.
We predict that in the year 2400, 38% of the total population will live in the city and 62%
in the suburbs.
Our computations in the population model indicate that the population distribution is
reaching a sort of steady state or equilibrium, which we now define.
198
Lecture 26
Definition 26.2: Let xk+1 = Axk be a discrete dynamical system. An equilibrium

state for A is a vector q such that Aq = q.
Hence, if q is an equilibrium for A and the initial condition is x0 = q then x1 = Ax0 = x0 ,

and x2 = Ax1 = x0 , and iteratively we have that xk = x0 = q for all k. Thus, if the system
starts at the equilibrium q then it remains at q for all time.
How do we find equilibrium states? If q is an equilibrium for A then from Aq = q we
have that
Aq − q = 0
and therefore
(A − I)q = 0.
Therefore, q is an equilibrium for A if and only if q is in the nullspace of the matrix A − I:
q ∈ Null(A − I).
Example 26.3. Find the equilibrium states of the matrix from the population model
" #
0.95 0.03
A= .
0.05 0.97
Does the initial condition of the population x0 change the long term behavior of the
discrete dynamical system? We will know the answer once we perform an eigenvalue analysis
on A (Lecture 22). As a preview, we will use the fact that
xk = Ak x0
and then write x0 in an appropriate basis that reveals how A acts on x0 . To see how the
last equation was obtained, notice that
x1 = Ax0
and therefore
x2 = Ax1 = A(Ax0 ) = A2 x0
and therefore
x3 = Ax2 = A(A2x0 ) = A3 x0
etc.
26.3 Stability of Discrete Dynamical Systems

We first formally define the notion of stability of a discrete dynamical system.
199
Definition 26.4: Consider the discrete dynamical system xk+1 = Axk where A ∈ Rn×n .
The origin 0 ∈ Rn is said to be asymptotically stable if for any initial condition x0 ∈ Rn
of the dynamical system we have
lim xk = lim Ak x0 = 0.
k→∞ k→∞
The following theorem characterizes when a discrete linear dynamical system is asymptoti-
cally stable.
Theorem 26.5: Let λ1 , . . . , λn be the eigenvalues of A. If |λj | < 1 for all j = 1, 2, . . . , n

then the origin 0 is asymptotically stable for xk+1 = Axk .
Solution. For simplicity, we suppose that A is diagonalizable. Let {v1 , . . . , vn } be a basis

of eigenvectors of A with eigenvalues λ1 , . . . , λn respectively. Then, for any vector x0 ∈ Rn ,
there exists constants c1 , . . . , cn such that
x0 = c1 v1 + · · · + cn vn .
Now, for any integer k ≥ 1 we have that.
Ak vi = λki vi
Then
xk = Ak x0
= Ak (c1 v1 + · · · + cn vn )
= c1 Ak v1 + · · · + cn Ak vn
= c1 λk1 v1 + · · · + cn λkn vn .
Since |λi | < 1 we have that lim λki = 0. Therefore,

k→∞
lim xk = lim (c1 λk1 v1 + · · · + cn λkn vn )

k→∞ k→∞

= c1 lim λk1 v1 + · · · + cn lim λkn vn
k→∞ k→∞
= 0v1 + · · · + 0vn
= 0.
This completes the proof.
200
Lecture 26
As an example of an asymptotically stable dynamical system, consider the 2D system

1.1 −0.4
xk+1 = x.
0.15 0.6

1.1 −0.4
The eigenvalues of A = are λ1 = 0.8 and λ2 = 0.9. Hence, by Theorem 26.5,
0.15 0.6
for any initial condition x0 , the sequence {x0 , x1 , x2 , . . . , } converges to the origin in R2 . In
Figure 26.1, we plot four different
state
sequences
{x0 , x1 , x2, . . . , } corresponding
to the four
3 3 −3 −3
distinct initial conditions x0 = , x0 = , x0 = , and x0 = . As expected,
7 −7 7 −7
all trajectories converge to the origin.
Figure 26.1: A 2D asymptotically stable linear system
• what a dynamical system is
• and how to find its equilibrium states
• how to determine if a discrete dynamical system has the origin as an asymptotically

stable equilibrium
201

MATH 233 - Linear Algebra I Lecture Notes: Cesar O. Aguilar

Uploaded by

MATH 233 - Linear Algebra I Lecture Notes: Cesar O. Aguilar

Uploaded by

MATH 233 - Linear Algebra I

1 Systems of Linear Equations 1

2 Row Reduction and Echelon Forms 11

4 The Matrix Equation Ax = b 31

5 Homogeneous and Nonhomogeneous Systems 41

7 Introduction to Linear Mappings 57

8 Onto, One-to-One, and Standard Matrix 67

12 Properties of the Determinant 97

13 Applications of the Determinant 103

14 Vector Spaces 109

15 Linear Maps 117

16 Linear Independence, Bases, and Dimension 125

17 The Rank Theorem 133

18 Coordinate Systems 137

19 Change of Basis 147

20 Inner Products and Orthogonality 153

21 Eigenvalues and Eigenvectors 163

22 The Characteristic Polynomial 169

24 Diagonalization of Symmetric Matrices 187

25 The PageRank Algortihm 191

26 Discrete Dynamical Systems 197

Systems of Linear Equations

1.1 What is a system of linear equations?

a11 x1 + a12 x2 + a13 x3 + · · · + a1n xn = b1

system consisting of m = 2 equations and n = 3 unknowns:

Here is a linear system consisting of m = 3 equations and n = 2 unknowns:

And finally, below is a linear system consisting of m = 4 equations and n = 6 unknowns:

−5x1 + x3 − 44x4 − 55x6 = −1

Is (1, −1, 2) a solution to the system?

2 · (1) + 2(2) + (−4) = 2

(1) + 3 · (2) − (−4) = 11

On the other hand, for (1, −1, 2) we have that

2(1) + 2(−1) + (2) = 2

A linear system is called consistent if it has at least one solution:

CONSISTENT ⇔ AT LEAST ONE SOLUTION

Solution. If we add the two equations we get

Example 1.4. Let t be an arbitrary real number and let

and in the second equation

1.3 Solving linear systems

1. Interchange two equations.

2. Multiply an equation by a nonzero constant.

3. Add a multiple of one equation to another.

Example 1.5. Use back substitution on the augmented matrix

−3x1 + 2x2 + 4x3 = 12

Interchange Row 1 (R1 ) and Row 2 (R2 ):

The original augmented matrix of the previous example is

corresponds to the equation

corresponds to the equation x2 = 0 which is perfectly valid.

1.4 Geometric interpretation of the solution set

Similarly, the solution of the linear system

the solution set is this line

After this lecture you should know the following:

Row Reduction and Echelon Forms

2.1 Row echelon form (REF)

x1 + 5x2 − 2x4 − x5 + 7x6 = −4

having augmented matrix

The above augmented matrix has the following properties:

whose augmented matrix is already in REF:

2.2 Reduced row echelon form (RREF)

P3. The leading entry in each nonzero row is a 1.

Create zeros under the newly created leading 1:

Create a leading 1 in the second row:

Create zeros under the newly created leading 1:

Example 2.3. Use row reduction to solve the linear system.

2x1 + 4x2 + 6x3 = 8

Solution. The augmented matrix is

Create a leading 1 in the first row:

The first row is equivalent to the equation:

and after substituting x3 = 4 we obtain

Later in the course, we will give a more geometric interpretation to rank(A).

x1 = 10 + 7x2 − 2x3 + 5x4 − 8x5

Thus, the solution set is

x1 = −89 − 31t + 19s

2.3 Existence and uniqueness of solutions

1. The augmented matrix will contain an inconsistent row.

After this lecture you should know the following: