Math Methods

L ECTURE N OTES
D EPARTMENT OF C OMPUTING
I MPERIAL C OLLEGE OF S CIENCE , T ECHNOLOGY AND M EDICINE
Mathematical Methods (CO-145)
Instructors:
Marc Deisenroth, Mahdi Cheraghchi
May 10, 2016

Contents
1 Sequences 1
1.1 The Convergence Definition . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Illustration of Convergence . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Common Converging Sequences . . . . . . . . . . . . . . . . . . . . . 3
1.4 Combinations of Sequences . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Sandwich Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.6 Ratio Tests for Sequences . . . . . . . . . . . . . . . . . . . . . . . . 6
1.7 Proof of Ratio Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.8 Useful Techniques for Manipulating Absolute Values . . . . . . . . . 8
1.9 Properties of Real Numbers . . . . . . . . . . . . . . . . . . . . . . . 9
2 Series 10
2.1 Geometric Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Harmonic Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Series of Inverse Squares . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Common Series and Convergence . . . . . . . . . . . . . . . . . . . . 13
2.5 Convergence Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6 Absolute Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.7 Power Series and the Radius of Convergence . . . . . . . . . . . . . . 18
2.8 Proofs of Ratio Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3 Power Series 22
3.1 Basics of Power Series . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2 Maclaurin Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Taylor Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.4 Taylor Series Error Term . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.5 Deriving the Cauchy Error Term . . . . . . . . . . . . . . . . . . . . . 29
3.6 Power Series Solution of ODEs . . . . . . . . . . . . . . . . . . . . . 29
4 Differential Equations and Calculus 31

4.1 Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.1.1 Differential Equations: Background . . . . . . . . . . . . . . . 31
4.1.2 Ordinary Differential Equations . . . . . . . . . . . . . . . . . 31
4.1.3 Coupled ODEs . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Partial Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2.1 Definition of Derivative . . . . . . . . . . . . . . . . . . . . . 36
I
CONTENTS Table of Contents
4.2.2 Partial Differentiation . . . . . . . . . . . . . . . . . . . . . . 36

4.2.3 Extended Chain Rule . . . . . . . . . . . . . . . . . . . . . . . 38
4.2.4 Jacobian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5 Complex Numbers 40
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.1.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.1.2 Imaginary Number . . . . . . . . . . . . . . . . . . . . . . . . 41
5.1.3 Complex Numbers as Elements of R2 . . . . . . . . . . . . . . 42
5.1.4 Closure under Arithmetic Operators . . . . . . . . . . . . . . 42
5.2 Representations of Complex Numbers . . . . . . . . . . . . . . . . . . 43
5.2.1 Cartesian Coordinates . . . . . . . . . . . . . . . . . . . . . . 43
5.2.2 Polar Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.2.3 Euler Representation . . . . . . . . . . . . . . . . . . . . . . . 44
5.2.4 Transformation between Polar and Cartesian Coordinates . . 45
5.2.5 Geometric Interpretation of the Product of Complex Numbers 47
5.2.6 Powers of Complex Numbers . . . . . . . . . . . . . . . . . . 47
5.3 Complex Conjugate . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.3.1 Absolute Value of a Complex Number . . . . . . . . . . . . . . 49
5.3.2 Inverse and Division . . . . . . . . . . . . . . . . . . . . . . . 49
5.4 De Moivre’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.4.1 Integer Extension to De Moivre’s Theorem . . . . . . . . . . . 51
5.4.2 Rational Extension to De Moivre’s Theorem . . . . . . . . . . 51
5.5 Triangle Inequality for Complex Numbers . . . . . . . . . . . . . . . 52
5.6 Fundamental Theorem of Algebra . . . . . . . . . . . . . . . . . . . . 53
5.6.1 nth Roots of Unity . . . . . . . . . . . . . . . . . . . . . . . . 53
5.6.2 Solution of zn = a + ib . . . . . . . . . . . . . . . . . . . . . . 54
5.7 Complex Sequences and Series* . . . . . . . . . . . . . . . . . . . . . 55
5.7.1 Limits of a Complex Sequence . . . . . . . . . . . . . . . . . . 55
5.8 Complex Power Series . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.8.1 A Generalized Euler Formula . . . . . . . . . . . . . . . . . . 57
5.9 Applications of Complex Numbers* . . . . . . . . . . . . . . . . . . . 57
5.9.1 Trigonometric Multiple Angle Formulae . . . . . . . . . . . . 57
5.9.2 Summation of Series . . . . . . . . . . . . . . . . . . . . . . . 58
5.9.3 Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
6 Linear Algebra 60
6.1 Linear Equation Systems . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.1.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.2 Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.2.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.2.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.3 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.3.1 Matrix Multiplication . . . . . . . . . . . . . . . . . . . . . . . 64
6.3.2 Inverse and Transpose . . . . . . . . . . . . . . . . . . . . . . 66
6.3.3 Multiplication by a Scalar . . . . . . . . . . . . . . . . . . . . 67
II
Table of Contents CONTENTS
6.3.4 Compact Representations of Linear Equation Systems . . . . . 67

6.4 Gaussian Elimination . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.4.1 Example: Solving a Simple Linear Equation System . . . . . . 68
6.4.2 Elementary Transformations . . . . . . . . . . . . . . . . . . . 69
6.4.3 The Minus-1 Trick for Solving Homogeneous Equation Systems 73
6.4.4 Applications of Gaussian Elimination in Linear Algebra . . . . 75
6.5 Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.5.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.5.2 Generating Set and Vector Subspaces . . . . . . . . . . . . . . 78
6.6 Linear (In)Dependence . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.6.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.7 Basis and Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.7.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.7.2 Example: Determining a Basis . . . . . . . . . . . . . . . . . . 84
6.7.3 Rank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.8 Intersection of Subspaces . . . . . . . . . . . . . . . . . . . . . . . . 87
6.8.1 Approach 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.8.2 Approach 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.9 Linear Mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.9.1 Image and Kernel (Null Space) . . . . . . . . . . . . . . . . . 94
6.9.2 Matrix Representation of Linear Mappings . . . . . . . . . . . 96
6.9.3 Basis Change . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.10 Determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.11 Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.11.1 Geometric Interpretation . . . . . . . . . . . . . . . . . . . . . 105
6.11.2 Characteristic Polynomial . . . . . . . . . . . . . . . . . . . . 105
6.11.3 Example: Eigenspace Computation . . . . . . . . . . . . . . . 105
6.11.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.12 Diagonalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.12.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6.12.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.12.3 Cayley-Hamilton Theorem . . . . . . . . . . . . . . . . . . . . 112
6.13 Scalar Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.13.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.13.2 Lengths, Distances, Orthogonality . . . . . . . . . . . . . . . . 114
6.13.3 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.13.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.14 Affine Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.14.1 Intersection of Affine Spaces . . . . . . . . . . . . . . . . . . . 117
6.15 Affine Mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.16 Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.16.1 Projection onto a Line . . . . . . . . . . . . . . . . . . . . . . 121
6.16.2 Projection onto General Subspaces . . . . . . . . . . . . . . . 123
6.16.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.17 Rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
III
CONTENTS Table of Contents
6.18 Vector Product (Cross Product) . . . . . . . . . . . . . . . . . . . . . 128
IV
Chapter 1
Sequences
1.1 The Convergence Definition

Given a sequence of real numbers a1 , a2 , . . ., we would like to have a precise definition
of what it means for this sequence to converge to a limit l. You may already have a
good intuitive understanding of what it means for an → l as n → ∞ or limn→∞ an = l.
Below is the mathematical definition that formally represents the concept.
Definition A sequence an for n ≥ 1 converges to a limit l ∈, written an → l as

n → ∞, if and only if the following statement can be shown to be true:
For all > 0, we can always find a positive integer N , such that, for all
n > N:
|an − l| <
This statement is best understood by taking the last part first and working back-
wards.
The limit inequality |an − l| < is called the limit inequality and says that the dis-
tance from the nth term in the sequence to the limit should be less than .
For all n > N . . . There should be some point N after which all the terms an in the
sequence are within of the limit.
For all > 0 . . . No matter what value of we pick, however tiny, we will still be able
to find some value of N such that all the terms to the right of aN are within
of the limit.
How do we prove convergence So that is the definition, but it also tells us how
to prove that a sequence converges. We need to find the N for any given value of .
If we can find this quantity then we have a rigorous proof of convergence for a given
sequence. We have an example below for an = 1/n.
1
1.2. Illustration of Convergence Chapter 1. Sequences
1.2 Illustration of Convergence

1.2
an
1
0.8
0.6
0.4

0.2
|an − l|
0 N
0 2 4 6 8 10 12 14
nth term
Figure 1.1: The convergence of the sequence an = 1/n for n ≥ 1 tending to 0
A demonstration of the convergence of an = 1/n for n ≥ 1 is shown in Figure 1.1.

The game is to find a value of N for whatever value of is chosen. N represents the
point after which all an will be within of the limit. Clearly the smaller the value
of chosen, the further to the right in Figure 1.1 we will have to go to achieve this,
thus the larger the value of N we will have to choose.
Applying the limit inequality In the example in Figure 1.1 we have taken a par-
ticular value of to demonstrate the concepts involved. Taking = 0.28 (picked
arbitrarily), we apply the |an − l| < inequality, to get:
|1/n| < 0.28
1/n < 0.28
n > 1/0.28
n > 3.57
This states that for n > 3.57 the continuous function 1/n will be less than 0.28 from
the limit. However, we are dealing with a sequence an which is only defined at
integer points on this function. Hence we need to find the N th term, aN , which is
the first point in the sequence to also be less than this value of . In this case we can
see that the next term above 3.57, that is N = 4, satisfies this requirement.
For all n > N We are further required to ensure that all other points to the right of
aN in the sequence an for n > N are also within 0.28 of the limit. We know this to be
2
Chapter 1. Sequences 1.3. Common Converging Sequences
true because the condition on n we originally obtained was n > 3.57, so we can be
sure that aN , aN +1 , aN +2 , . . . are all closer to the limit than 0.28.
For all > 0 The final point to note is that it is not sufficient to find N for just a
single value of . We need to find N for every value of > 0. Since N will vary with
as mentioned above, we are clearly required to find a function N ().
In the case of an = 1/n, this is straightforward. We apply the limit inequality in
general and derive a condition for n.
1 1
< ⇒ n>
n
We are looking for a greater-than inequality. So long as we get one, we can select
the next largest integer value; we use the ceiling function to do this, giving:
1

N () =

Now whatever value of we are given, we can find a value of N using this function
and we are done.
1.3 Common Converging Sequences

It is very useful to have an intuition about certain common sequences and whether
they converge. Below is a list of useful results, which can be proved directly. I would
recommend you get practice at direct proof on some of the sequences below.
1 1 1
1. an = → 0, also an = 2 → 0, an = √ → 0
n n n
1
2. In general, an = → 0 for some positive real constant c > 0
nc
1 1 1
3. an = n
→ 0, also an = n → 0, an = n → 0
2 3 e
1
4. In general, an = → 0 for some real constant |c| > 1;
cn n
or equivalently an = c → 0 for some real constant |c| < 1
1
5. an = → 0 (Hard to prove directly, easier to use a ratio test in Section 1.6).
n!
1
6. an = → 0 for n > 1
log n
3
1.4. Combinations of Sequences Chapter 1. Sequences
1.4 Combinations of Sequences

Given sequences an and bn which converge to limits a and b respectively, then
1. lim (λan ) = λa for some real constant λ

n→∞
2. lim (an + bn ) = a + b
n→∞
3. lim (an − bn ) = a − b
n→∞
4. lim (an bn ) = ab
n→∞
an a
5. lim = provided that b , 0
n→∞ bn b
We proved rule (2) in lectures, you can find proofs of the other results in “Analysis:
an introduction to proof” by Stephen Lay.
Example These sequence combination results are useful for being able to under-
stand convergence properties of combined sequences. As an example, if we take the
sequence:
3n2 + 2n
an = 2
6n + 7n + 1
We can transform this sequence into a combination of sequences that we know how
to deal with, by dividing numerator and denominator by n2 .
3 + n2
an =
6 + n7 + n12
b
= n
cn
where bn = 3 + n2 and cn = 6 + n7 + n12 . We know from rule (5) above that if bn → b and
cn → c, then bcnn → bc . This means that we can investigate the convergence of bn and
cn separately.
Similarly we can rewrite bn as 3 + dn , for dn = n2 . We know that dn = n2 → 0 as it is
one of the common sequence convergence results, thus bn → 3 by rule (2) above. By
a similar argument, we can see that cn → 6 using composition rule (2) and common
results.
Finally, we get the result that an → 21 .
1.5 Sandwich Theorem

The Sandwich theorem is an alternative way, and often a simpler way, of proving that
a sequence converges to a limit. In order to use the Sandwich theorem, you need
to have two sequences (one of which can be constant) which bound the sequence
4
Chapter 1. Sequences 1.5. Sandwich Theorem
you wish to reason about. The two bounding sequences should be given as known
results, or proved to converge prior to use. The bounding sequences should be easier
to reason about than the sequence to be reasoned about.
1
un
ln
an
0.5
-0.5
-1
0 2 4 6 8 10 12 14
nth term
Figure 1.2: Demonstration of the Sandwich theorem. The sequence an is sandwiched

between two bounding sequences un and ln which converge to the same limit.
Sandwich Theorem Given sequences un and ln for n ≥ 1, where both un → l and

ln → l as n → ∞.
If we can show that a third sequence an satisfies:
ln ≤ an ≤ un for all n ≥ N for some N ∈ N
Then an → l as n → ∞.
Demonstration We provide an illustration of the sandwich theorem in Figure 1.2.

Here, we have a sequence an = sinn n which oscillates around its limit of 0. We con-
struct two sandwiching sequences un = 1/n and ln = −1/n. Since we know (or can
easily prove) that un and ln both tend to 0, we only need to show that ln ≤ an and
an ≤ un from some point n > N . In this case we can show it for all n > 0.
In this instance, we can prove this in one line and refer to the property of the sin
function.
1 sin n 1
− ≤ ≤ ⇒ −1 ≤ sin n ≤ 1
n n n
which is true for all n by the range of sin n.
5
1.6. Ratio Tests for Sequences Chapter 1. Sequences
Proof of sandwich theorem We know that ln and un tend to the same limit l. We
use the definition of convergence for both these sequences.
Given some > 0, we can find an N1 and N2 such that for all n > N1 , |ln − l| < and
for all n > N2 , |un − l| < . If we want both sequences to be simultaneously less than
from the limit l for a single value of N 0 , then we have to pick N 0 = max(N1 , N2 ).
So for N 0 = max(N1 , N2 ), for all n > N 0 , we know that both:
|ln − l| < and |un − l| <
Therefore, removing the modulus sign and using the fact that ln < un :
l − < l n < un < l +
We also have proved in the course of applying the sandwich theorem that ln ≤ an ≤ un
for all n > N . Hence, for all n > max(N , N 0 ):
l − < ln ≤ an ≤ un < l +
l − < an < l +
− < an − l <
⇒ |an − l| <
This allows us to say that an converges to l as a result.
1.6 Ratio Tests for Sequences

A third technique (and a very simple one) for proving sequence convergence is called
the ratio test. So-called as it compares the ratio of consecutive terms in the sequence.
The test verifies whether a sequence converges to 0 or diverges. If a sequence bn has
a non-zero limit, l, then the ratio test can still be applied to a modified sequence
an = bn − l, which will converge to 0 (if it passes the test!).

an+1
Ratio convergence test If an ≤ c < 1 for some c ∈ and for all sufficiently large n

(i.e., for all n ≥ N for some integer N ), then an → 0 as n → ∞.

an+1
Ratio divergence test If an ≥ c > 1 for some c ∈ and for all sufficiently large n

(i.e., for all n ≥ N for some integer N ), then an diverges as n → ∞.
Example of convergence ratio test As a simple

example of the ratio test in use
a −(n+1)
we can look at an = 3−n , then an+1 = 3 −n = 1 < 1. So we can say that an = 3−n
n 3 3
converges to 0 as n → ∞.
6
Chapter 1. Sequences 1.7. Proof of Ratio Tests
Limit ratio test The limit ratio test is a direct corollary of the standard ratio tests.
Instead of looking at the ratio of consecutive terms in the sequence, we look at the
limit of that ratio. This is useful when the standard test is not conclusive and leaves
terms involving n in the ratio expression. To use this test, we compute:

an+1
r = lim
n→∞ an
If r < 1 the an sequence converges to 0, and if r > 1, it diverges
Example of limit ratio test

An
example of the limit ratio test, we consider the
1 an+1 1/(n+1)! 1
sequence an = n! . Taking an = 1/n! = n+1 . This is not a conclusive result by the

normal ratio test (although clearly less than 1, there is no constant that sits between
1
it and 1). So we consider the limit r = limn→∞ n+1 = 0. Hence an converges to 0.
1.7 Proof of Ratio Tests

For completeness, I will include proofs of the ratio tests above. If you are finding the
material fairly straightforward then you may enjoy looking at these. I do not expect
you to be able to reproduce these arguments in an exam although I do expect you to
be able to apply the tests in question.

Proof of convergence ratio test In the case where aan+1 n
≤ c < 1 for n ≥ N , we can
take the left hand inequality and multiply by |an | to get |an+1 | ≤ c|an | for n ≥ N . Also
|an | ≤ c|an−1 | and so on. In fact we can create a chain of inequalities down to the
point n = N , where the property is first true:
|an | ≤ c|an−1 | ≤ c2 |an−2 | · · · ≤ cn−N |aN |
Thus:
|an | ≤ cn−N |aN |

|aN |
≤ kcn where k = is a constant
cN
As c < 1, we know that a new sequence bn = kcn converges to 0 (standard result (4)
from Section 1.3) and we have bounded the sequence an between bn and −bn . Thus
by the sandwich theorem, an must also converge to 0.
an+1
A very similar proof can be constructed to show the divergence of an should an ≥

c > 1 for n ≥ N .
Proof of limit ratio test Justifying that this works demonstrates an interesting

an+1
(and new) use of the limit definition. We define a new sequence, bn = an , and we

can consider the limit of this new sequence, r.
7
1.8. Useful Techniques for Manipulating Absolute Values Chapter 1. Sequences
As with all sequences, we can say that for all > 0, we can find an N such that for
all n > N , |bn − r| < .
As before we can expand the modulus to give:
r − < bn < r + (∗)
Now we take the definition of the limit literally and pick a carefully chosen value for
> 0 in both cases.
In the case where r < 1 (remember that this corresponds to a sequence that we wish
to show converges), we choose = 1−r2 > 0. Equation (∗) becomes:
1−r 1−r
r− < bn < r +
2 2
3r − 1 r +1
< bn <
2 2
Taking the right hand side, we see that bn < r+1

2 and since r < 1, we can show that
r+1 r+1
2 < 1 also. If we take c = 2 < 1, then by the original ratio test, we can say that
the original sequence an must converge.
In the case where r > 1 (remember that this corresponds to a sequence that we wish
to show diverges), we choose = r−1
2 > 0. Equation (∗) becomes:
r −1 r −1
r− < bn < r +
2 2
r +1 3r − 1
< bn <
2 2
In this case, we take the left hand side of the inequality, bn > r+1
2 and since r > 1, we
r+1 r+1
can show that 2 > 1 also. If we now take c = 2 > 1, then by the original ratio test,
we can say that the original sequence an must diverge.
1.8 Useful Techniques for Manipulating Absolute Val-

ues
Many of the reasoning techniques behind limits require the ability to manipulate the
modulus or absolute values of expressions. There are some properties it is useful
to know. These will come in handy for other topics, including series and radius of
convergence reasoning.
A commonly used decomposition of the absolute value operator is:
|x| < a ⇔ −a < x < a
Where we are combining absolute values of expressions, the following are useful,
where x, y as real numbers:
1. |xy| = |x| × |y|
8
Chapter 1. Sequences 1.9. Properties of Real Numbers

x |x|
2. =
y |y|
3. |x + y| ≤ |x| + |y|, the triangle inequality
4. |x − y| ≥ ||x| − |y||
1.9 Properties of Real Numbers

Now we have looked at limits of sequences of real numbers, we will see how this
relates to some fundamental properties of sets of real numbers, which will lead us
to describe the fundamental axiom of analysis.
We will cover some definitions of basic concepts for sets of real numbers: If S is a
set of real numbers.
1. u is an upper bound on S if u ≥ s for any s ∈ S
2. l is an lower bound on S if l ≤ s for any s ∈ S
3. we say that S is bounded above if S has an upper bound
4. we say that S is bounded below if S has a lower bound
5. we say that S is bounded if S has an upper and lower bound
A set S can have many upper or lower bounds (or none), so:
1. we call sup(S) the least upper bound of S
2. we call inf(S) the greatest lower bound of S
By convention, if a set S has no upper bound, e.g., S = {x | x > 0}, then sup(S) = ∞.
Similarly if S has no lower bound then inf(S) = −∞.
The fundamental axiom of analysis states that every increasing sequence of real
numbers that is bounded above, must converge.
9
Chapter 2
Series
Series are sums of sequences. Whether they converge or not is an important factor
in determining whether iterative numerical algorithms terminate.
An infinite series is a summation of the form:
∞
X
S= an
n=1
for some real sequence an for n ≥ 1.
Partial sums One easy way of determining the convergence of series is to construct
the partial sum – the sum up to the nth term of the series:
n
X
Sn = ai
i=1
Sn itself is a sequence and can be treated as such. A series converges or diverges if

and only if its sequence of partial sums converges or diverges.
Bounded above When ai > 0 for all i, then the partial sum Sn is an increasing
sequence, that is:
S1 < S2 < S3 < · · ·
If you can show that the sequence is bounded above, then as with other increasing
sequences, it must converge to some limit.
Series terms tend to 0 For a series S = ∞

P
n=1 an to have any chance of converging,
the terms an have to converge to 0 as n → ∞. If an diverges or tends to a non-zero
limit then the series S will diverge. Note that an → 0 is not sufficient to ensure that
S itself converges (otherwise we would not need the rest of this section).
10
Chapter 2. Series 2.1. Geometric Series
The tail of a series It is important to understand that, as with sequences, the

convergence behaviour of a series lies in the tail andPwe do not really care about
transient values
P∞ of early P
terms such as a1 , a2 , a3 . If S= ∞ n=1 an converges or diverges
then so will n=10 an or ∞ ∞
P
a
n=1000 n or in general n=N n for any natural number N .
a
This explains why, in comparison tests, we need only worry about later terms in the
series from a certain point N , rather than the whole series.
2.1 Geometric Series

This is an example of the use of partial sums to determine the convergence of a
series. A common and well-known example is the geometric series
∞
X
G= xn
n=1
P∞ n P∞ n
If G does exist, G = x + n=2 x = x+x n=1 x = x + xG so:
x
G=
1−x
How can we determine when G does exist? When the series and thus the sequence of
partial sums converges. In this case, we can find an expression
Pnof the nth partial
Pn−1 sum
of the geometric series. Similarly, nth partial sum Gn = x + i=2 x = x + x i=1 xi =
i
x + x(Gn − xn ), so:
x − xn+1
Gn =
1−x
Gn is a sequence, so we can determine if Gn converges or diverges in the limit, taking
care to take into account behaviour that may depend on x. The xn+1 term is the only
varying term in the sequence, so from our knowledge of sequences we can say that
this and thus Gn will converge if |x| < 1.
x
Gn → as n → ∞ for |x| < 1
1−x
by rules of convergence for sequences.
Similarly, for |x| > 1, Gn diverges as n → ∞. For x = 1, Gn = n which also diverges.
2.2 Harmonic Series

This is a very important result as it is used to show that other series diverge by using
the comparison test, described later.
The harmonic series is given by
∞
X 1 1 1 1 1
S= = 1+ + + + + ···
n 2 3 4 5
n=1
11
2.3. Series of Inverse Squares Chapter 2. Series
It is an important result to show that it diverges to ∞.

By grouping terms in the series, we can see this intuitively:
1 1 1 1 1 1 1 1 1 1

S = 1 + + + + + + + +··· > 1 + + + + ···
2 3 4 5 6 7 8 2 2 2
| {z } | {z }
> 14 + 14 > 18 + 18 + 18 + 18
So considering partial sums, we get that

2n
X 1 n
Sn = > 1+
i 2
i=1
Clearly 1 + n2 diverges and since Sn is greater than this sequence, it must also diverge
as n → ∞ and so does S.
2.3 Series of Inverse Squares

Similarly to the harmonic series, this result is important since it is used to show that
other series converge by comparison.
In order to understand the convergence of S = ∞ 1
P
n=1 n2 , the series of inverse squares,
we must first consider the limit of the series:
∞
X 1
T =
i(i + 1)
i=1
1 1 1
Using partial fractions, we can rewrite i(i+1)
as i − i+1 . The the nth partial sum of T
can therefore be written as:
n
1 1 1
X
Tn = − = 1− (2.1)
i i +1 n+1
i=1
(To see why the right hand side is true, try writing out the first and last few terms
of the sum and observing which cancel). From this, we see that Tn converges giving
T = 1.
We will use the partial sum result to gives us a handle on being able to understand
the convergence or divergence of:
∞
X 1
S=
n2
n=1
Pn 1
We start by considering terms of the partial sum Sn = i=1 i 2 , and we notice that:
1 1 1
< 2< for i ≥ 2
i(i + 1) i (i − 1)i
12
Chapter 2. Series 2.4. Common Series and Convergence
We sum from i = 2 (to avoid a 0 denominator on the right hand side) to n to get:
n
1 1 1 1 X1 1 1 1 1
+ + + ··· + < 2
< + + + ··· +
2·3 3·4 4·5 n(n + 1) i 2·1 3·2 4·3 n(n − 1)
i=2
We note that the left hand and right hand series above differ only in the first and last
terms, so can be rewritten as:
 n  n n−1
X 1  1 X 1 X 1
  − < <
 i(i + 1)  2 i2 i(i + 1)
i=1 i=2 i=1
1
Now we add 1 across the whole inequality to allow us to sum i2
from 1 and get the
required partial sum Sn as our middle expression:
n n−1
1 X 1 X 1
+ < Sn < 1 +
2 i(i + 1) i(i + 1)
i=1 i=1
This effectively bounds the sequence of partial sums, Sn . We use our result of (2.1)
to get:
3 1 1
− < Sn < 2 −
2 n+1 n
We see that the upper bound on Sn (and thus also S) is 2 and the lower bound is 32 .
Since the partial sum sequence is increasing, the existence of an upper bound proves
that the series converges. The value of the bound itself must lie between 32 and 2.
2.4 Common Series and Convergence

As with sequences, it is extremely useful to have a good knowledge of whether some
common series converge or not. These will be very useful in comparison tests for
both convergence and divergence arguments.
Diverging series It should be noted that if S = ∞

P
n=1 an and the sequence an does
not itself tend to 0 as n → ∞, then S will certainly diverge.
∞
X 1
1. Harmonic series: S = diverges.
n
n=1
X 1
2. Harmonic primes: S = diverges.
p
p : prime
∞
X
3. Geometric series: S = xn diverges for |x| ≥ 1.
n=1
13
2.5. Convergence Tests Chapter 2. Series
Converging series The limit for the inverse squares series is given for interest only,
the essential thing is to know that the series converges.
∞
X x
1. Geometric series: S = xn converges to so long as |x| < 1.
1−x
n=1
∞
X 1 π2
2. Inverse squares series: S = converges to .
n2 6
n=1
∞
1
X 1
3. nc series: S = converges for c > 1.
nc
n=1
2.5 Convergence Tests

In this section we focus on series with non-negative an terms. In these series the
partial sums are increasing and hence either they converge, if the partial sums are
bounded, or they diverge, if they are not.
In the following tests, we take:
∞
X
1. ai is a non-negative term in the series ai that we wish to reason about
i=1
∞
X
2. ci is a series that we have already established converges with sum c
i=1
∞
X
3. di is a series that we have already established diverges
i=1
Comparison test Let λ > 0 and N ∈ N. Then

∞
X
1. if ai ≤ λci for all i > N , then ai converges
i=1
∞
X
2. if ai ≥ λdi for all i > N , then ai diverges
i=1
Limit comparison test Sometimes the following form of comparison test is easier,
as it does not involve having to find a constant λ to make the first form of the test
work:
∞
ai X
1. if lim exists, then ai converges
i→∞ ci
i=1
∞
di X
2. if lim exists, then ai diverges
i→∞ ai
i=1
14
Chapter 2. Series 2.5. Convergence Tests
D’Alembert’s ratio test The D’Alembert ratio test is a very useful test for quickly
discerning the convergence or divergence of a series. It exists in many subtly differ-
ent forms.
From some point N ∈ N in the series, for all i ≥ N :
∞
a X
1. If i+1 ≥ 1, then ai diverges
ai
i=1
∞
a X
2. If there exists a k such that i+1 ≤ k < 1, then ai converges
ai
i=1
D’Alembert’s limit ratio test From this definition we can obtain (Section 2.8) the
easier-to-use and more often quoted version of the test which looks at the limit of
the ratio of successive terms. This version states:
∞
a X
1. if lim i+1 > 1 then ai diverges
i→∞ ai
i=1
∞
a X
2. if lim i+1 = 1 then ai may converge or diverge
i→∞ ai
i=1
∞
a X
3. if lim i+1 < 1 then ai converges
i→∞ ai
i=1
There are some simple examples of series, ∞

P
n=1 an , which converge or diverge but
give a limit value of 1 when applying the limit ratio test. an = 1 for all n diverges but
gives a limit of 1 in the above test. By contrast, an = n12 for n ≥ 1 converges but also
gives a limit ratio of 1.
P∞
Integral test This test is based on the idea1 that for a sequence S = n=1 an where
an = f (n) is a decreasing function, e.g., n1 as n ≥ 1:
∞
X Z ∞ ∞
X
an+1 < f (x) dx < an
n=1 1 n=1
In practice, you evaluate the integral and if it diverges you know that the series
diverges by the right hand inequality. If the integral converges, you know the series
does also; this time by the left hand inequality.
Formally stated: suppose that f (x) is a continuous, positive and decreasing function
on the interval [N , ∞) and that an = f (n) then:
Z∞ ∞
X
1. If f (x) dx converges, so does an
N n=N
you have come across Riemann Sums before, we Rare essentially using ∞
1 If P P∞
i=1 an and i=1 an+1 as
∞
left and right Riemann sums which bound the integral 1 f (x) dx where an = f (n).
15
2.5. Convergence Tests Chapter 2. Series
1.2
f (x) = 1x
1 an = n1
0.8
0.6
0.4
0.2
0
0 1 2 3 4 5 6 7 8 9 10
R∞
Figure 2.1: A divergence test: the series ∞ 1 1
P
n=1 n as compared to 1 x dx, showing that
the series is bounded below by the integral.
Z ∞ ∞
X
2. If f (x) dx diverges, so does an
N n=N
As usual we do not have to concern ourselves with the first N − 1 terms as long as
they are finite. If we can show that the tail converges or diverges, i.e., an for n ≥ N ,
then the whole series will follow suit.
Example:
P∞ 1 Integral test divergenceP∞ We can apply the integral test to show that
1
n=1 n diverges. Figure 2.1 shows a
n=1 n where an = n as a sum of 1×an rectangular
areas for n ≥ 1. When displayed like Pthis, and only because f (x) = 1x is a decreasing
∞
function,R we can see that the series n=1 an is strictly greater than the corresponding
∞
integral 1 f (x) dx. If we evaluate this integral:
Z ∞ Z b
1
f (x) dx = lim dx
1 b→∞ 1 x
= lim [ln x]b1
b→∞
= lim (ln b − ln 1)
b→∞
P∞
we get a diverging result. Hence the original series n=1 an must also diverge.
Example:
P∞ 1 Integral test convergence P We can apply the integral test to show that
∞ 1
n=1 n2 converges. Figure 2.2 shows n=1 an+1 where an = n2 as a sum of 1 × an+1
16
Chapter 2. Series 2.6. Absolute Convergence
0.4
f (x) = x12
0.35
an = n12
0.3
0.25
0.2
0.15
0.1
0.05
0
0 1 2 3 4 5 6 7 8 9 10
R∞
Figure 2.2: A convergence test: the series ∞ 1 1
P
n=2 n2 as compared to 1 x2 dx, showing
that the series is bounded above by the integral.
rectangular areas for n ≥ 1. When displayed like this, and

P∞again only because f (x) =
1
x2
is a decreasing function, we can see that the series n=1 an+1 is strictly less than
R∞
the corresponding integral 1 f (x) dx. If we evaluate this integral:
Z ∞ Z b
1
f (x) dx = lim dx
1 b→∞ 1 x2
b
1

= lim −
b→∞ x 1
1

= lim 1 −
b→∞ b
=1
P∞
Since we get a converging result
P∞ in1 the integral, we know that the series n=1 an+1
must also converge. Hence n=1 n2 also converges as it is only one (finite) term
different.
2.6 Absolute Convergence

Thus far we have only dealt with series involving positive an terms. The difficulty
with series that involve a mixture of positive and negative terms, is that the conver-
gence behaviour can depend on the order in which you add the terms up.
Consider S = 1 − 1 + 1 − 1 + 1 − 1 . . ., the partial sums of this series oscillate between
1 and 0. If we take two positive terms (from later in the same series) for every
17
2.7. Power Series and the Radius of Convergence Chapter 2. Series
negative term, i.e., S = 1 + 1 − 1 + 1 + 1 − 1 + 1 + 1 − 1, we get a series that diverges to

+∞. However both series have the same terms and the same number of −1 and +1
terms.
Absolute convergence In cases of series with a mixture of positive and negative

terms, S = ∞
P
a
n=1 n , it makes sense to look at the series of absolute values of the
terms of a series, that is:
X∞
0
S = |an |
n=1
If this series,S0, converges then we can be sure that the original series S will also
converge. S is said to converge absolutely.
A series that converges absolutely also converges, however the same is not true the
other
P∞ wayn−1 round. A simple example of this is the alternating harmonic series, S =
1
n=1 (−1) n . This series converges to ln 2 (when summed in this order). However
the series involving
P∞ 1 the absolute values of the same terms is the normal harmonic
0
series, S = n=1 n , which we know diverges.
Tests for absolute convergence Where we have a series with a mixture of positive
and negative terms, an , we can test for absolute convergence by applying any of the
above tests to the absolute value of the terms |an |. For example the limit ratio test
becomes: for N ∈ N,
∞
an+1 X
1. if lim > 1 then an diverges
n→∞ an
n=1
∞
an+1 X
2. if lim = 1 then an may converge or diverge
n→∞ an
n=1
∞
an+1 X
3. if lim < 1 then an converges absolutely (and thus also converges)
n→∞ an
n=1
2.7 Power Series and the Radius of Convergence

As a glimpse forward to the next topic, we will look at how series convergence affects
power series. As you may know, it is possible to represent high-order functions as
power series, that is:
∞
X
f (x) = an xn
n=0
However, such series may only converge for certain ranges of input parameter x. We
have already seen this with the geometric series, which represents a power series
x
expansion of the function, f (x) = 1−x (from Section 2.1) as:
∞
X
f (x) = xn
n=1
18
Chapter 2. Series 2.8. Proofs of Ratio Tests
x
However this limit, 1−x , and therefore also the power series expansion are only
equivalent if |x| < 1 as we have seen.
The radius of convergence represents the size of x set for which a power series
converges. If a power series converges for |x| < R, then the radius of convergence
is said to be R. Sometimes, we consider series which only converge within some
distance R of a point a, in this case, we might get the following result |x − a| < R. In
this case also the radius of convergence is R. If the series converges for all values of
x, then the radius is given as infinite.
Example: calculating the radius of convergence Let us take a power series:

∞
X
S= n2 x n
n=1
We can apply the D’Alembert ratio test to this as with any other series, however
we have to leave the x variable free, and we have to cope with possible negative x
values.
We apply the version of the D’Alembert ratio test which tests for absolute conver-
gence (and thus also convergence) since, depending on the value of x, we may have
negative terms in the series.

a

(n + 1)2 xn+1
lim n+1 = lim

n→∞ an n→∞ n2 xn

1 2

= lim x 1 +
n→∞ n

= |x|
If we look at the ratio test, it says that if the result is < 1 it will converge. So our
general convergence condition on this series, is that |x| < 1. Therefore the radius of
convergence is 1.
2.8 Proofs of Ratio Tests

As with the ratio test proofs for converging sequences, I do not expect you to come
up with these arguments in an exam. However some of you will be interested in
how these tests work.
Proof of D’Alembert ratio test We take the two cases separately:

Case aai+1
i
≥ 1 for all i ≥ N : Taking aN = C and the non-increasing case of ai+1
ai = 1,
we get a partial sum from the N th term of:
n+N
X−1
ai = Cn
i=N
19
2.8. Proofs of Ratio Tests Chapter 2. Series
This diverges as it is proportional to n. A comparison test with this constant term

series and the case where aai+1
i
> 1 shows that the case of the ratio being greater than
1 also diverges.
Case aai+1
i
≤ k < 1: We can see that this series converges by performing a term-by-term
comparison with a geometric series with ratio x = k. From i ≥ N , we get:
aN +1 ≤ kaN
aN +2 ≤ kaN +1 ≤ k 2 aN
aN +3 ≤ kaN +2 ≤ k 2 aN +1 ≤ k 3 aN
..
.
aN +m ≤ k m aN
We can rewrite this last as aN +m ≤ λk N +m where λ = kaNN is a constant. Letting n =

N + m, allows us to write that an ≤ λk n for n ≥ N . Since we know that k < 1 we are
sure that the geometric series bn = k n converges. This is precisely
P∞ what we require
under the conditions of the comparison test, to show that n=1 an converges.
20
Chapter 2. Series 2.8. Proofs of Ratio Tests
Proof of D’Alembert limit ratio test Taking the two cases P∞ in turn:
Case limi→∞ aai+1
i
= l > 1: We are looking to show that i=1 ai diverges. Using the
definition of the limit, we know that from some point N for all i > N :

ai+1 − l <
ai
a
l − < i+1 < l +
ai
Since we are trying to show that this series diverges, we would like to use the left
hand inequality to show that the ratio aai+1 i
> 1 from some point onwards, for example
i > N . To do this we can pick = l − 1, recalling that l > 1, so this is allowed as > 0.
We now get that for all i > N :
a
1 < i+1
ai
P∞
and so by the non-limit D’Alembert ratio test, the series P∞ i=1 ai diverges.
Case limi→∞ aai+1
i
= l < 1: We are looking to show that i=1 ai converges in this case.
Again using the limit definition we get for all i > N :

ai+1 − l <
ai
a
l − < i+1 < l +
ai
We are aiming to show that aai+1

i
≤< k < 1 for some k, so that we can avail ourselves
of the second D’Alembert ratio test case. So we are looking to use the right hand
inequality on this occasion. This means that we will need to set k = l + but we still
need to guarantee that k < 1 (and obviously l < k also as we need to set k = l +
where > 0). We can do this easily by situating k between l and 1 with k = l+1
2 . This
1−l
sets = 2 and importantly > 0 since l < 1. Going forward from setting (and
because we have fixed it to work), we get for all i > N :
ai+1
<k<1
ai
P∞
So, by the non-limit D’Alembert ratio test, the series i=1 ai converges.
21
Chapter 3
Power Series
3.1 Basics of Power Series

We can represent a function f (x) by a series expansion in x
∞
X
f (x) = ai xi
i=0
for coefficients ai ∈ and i = 1, 2, . . .. This is called a power series because it is a series

of powers of the argument x.
For example:
f (x) = (1 + x)2 = 1 + 2x + x2
has a0 = 1, a1 = 2, a2 = 1 and ai = 0 for i > 2. However in general the series may be
infinite provided it converges. We will use ratio-of-convergence arguments to check
this.
3.2 Maclaurin Series

For a function f (x):
f (x) = a0 + a1 x + a2 x2 + a3 x3 + · · ·
Suppose the value of the function f (x) = is known at x = 0. Then we have an easy
way of finding the first coefficient of our power series, a0 . We can set x = 0, to get:
f (0) = a0 .
Suppose the derivatives of the function f (x) are known at x = 0, that is:
f 0 (0), f 00 (0), . . . , f (n) (0)
where f (n) (a) represents the function f (x) differentiated n times and then taking
the derivative expression with x = a. Now if we differentiate f (x) to get rid of the
constant term:
f 0 (x) = a1 + 2.a2 x + 3.a3 x2 + 4.a4 x3 + · · ·
and again set x = 0, we get a1 = f 0 (0).
22
Chapter 3. Power Series 3.2. Maclaurin Series
Differentiate again to get rid of the constant term:
f 00 (x) ≡ f (2) (x) = 2.1.a2 + 3.2.a3 x + 4.3.a4 x2 + · · ·
Set x = 0 and repeat the process, we get:
f (2) (0)
a2 =
2!
In general, for the nth term, we get:
f (n) (0)
an = for n ≥ 0
n!
Maclaurin series Suppose f (x) is differentiable infinitely many times and that it
has a power series representation (series expansion) of the form
∞
X
f (x) = ai xi
i=0
as above.
Differentiating n times gives
∞
X
(n)
f (x) = ai i(i − 1) . . . (i − n + 1)xi−n
i=n
Setting x = 0, we have f (n) (0) = n!an because all terms but the first have x as a factor.
Hence we obtain Maclaurin’s series:
∞
X xn
f (x) = f (n) (0)
n!
n=0
As we have seen in the Series notes, it is important to check the radius of convergence
(set of valid values for x) that ensures that this series converges.
The addition of each successive term in the Maclaurin series creates a closer approx-
imation to the original function around the point x = 0, as we will demonstrate.
Example Taking f (x) = (1+x)3 . We repeat the process of differentiating repeatedly

and setting x to zero in each derivative expression:
1. f (0) = 1 so a0 = 1
3
2. f 0 (x) = 3(1 + x)2 so f 0 (0) = 3 and a1 = 1! =3
6
3. f 00 (x) = 3.2(1 + x) so f 00 (0) = 6 and a2 = 2! =3
6
4. f 000 (x) = 3.2.1 so f 000 (0) = 6 and a3 = 3! =1
23
3.2. Maclaurin Series Chapter 3. Power Series
Higher derivatives are all 0 and so (as we know), so we end up with:
(1 + x)3 = 1 + 3x + 3x2 + x3
In this case, since the series expansion is finite we know that the Maclaurin series will
be accurate (and certainly converge) for all x, therefore the radius of convergence is
infinite.
P f (i) (x)
We now consider the partial sums of the Maclaurin expansion fn (x) = ni=0 i! xi .
For f (x) = (1 + x)3 these would be:
f0 (x) =1
f1 (x) =1 + 3x
f2 (x) =1 + 3x + 3x2
f3 (x) =1 + 3x + 3x2 + x3
We can plot these along with f (x) in Figure 3.1, and we see that each partial sum is
a successively better approximation to f (x) around the point x = 0.
10
f (x)
f0 (x)
8 f1 (x)
f2 (x)
-2
-4
-6
-2 -1.5 -1 -0.5 0 0.5 1
Figure 3.1: Successive approximations from the Maclaurin series expansion of f (x) =
(1 + x)3 where f0 (x) = 1, f1 (x) = 1 + 3x and f2 (x) = 1 + 3x + 3x2 .
Example 2: f (x) = (1 − x)−1 We probably know that the power series is for this
function is the geometric series in x, in which all ai = 1.
1. f (0) = 1, so far so good!
2. f 0 (x) = −(1 − x)−2 (−1) = (1 − x)−2 so f 0 (0) = 1
3. f 00 (x) = −2!(1 − x)−3 (−1) = 2!(1 − x)−3 so f 00 (0) = 2!
24
Chapter 3. Power Series 3.2. Maclaurin Series
Differentiating repeatedly,
f (n) (x) = (−1)(−2) . . . (−n)(1 − x)−(n+1) (−1)n

= n!(1 − x)−(n+1)
so
f (n) (0) n!(1)−(n+1)
an = = =1
n! n!
Thus (1 − x)−1 = ∞ i
P
i=0 x unsurprisingly since this is the sum of an infinite geometric
series in x.
We can check convergence by using the absolute convergence version of D’Alembert’s
ratio test, where bn = xn are the series terms:

xn+1

bn+1
lim = lim
n→∞ bn n→∞ xn

= |x|
Thus convergence is given by |x| < 1 and a radius of convergence of 1 for this power
series.
Example 3: f (x) = ln(1 + x)
1. f (0) = 0 because ln 1 = 0 so no constant term
2. f 0 (x) = (1 + x)−1 so f 0 (0) = 1
3. f 00 (x) = (−1)(1 + x)−2 so f 00 (0) = −1
Giving an nth derivative of:
(−1)n−1 (n − 1)! f (n) (0) (−1)n−1

f (n) (x) = and thus an = =
(1 + x)n n! n
Therefore series for f (x) = ln(1 + x) is:
x x2 x3 x4
ln(1 + x) = − + − + ···
1 2 3 4
(−1)n−1 xn
Taking series terms of bn = n , we get a ratio test of limn→∞ | bbn+1 | = |x| and a
n
convergence condition of |x| < 1.
Indeed we see that if we were to set x = −1, we would get ln 0 = −∞ and a corre-
sponding power series of:
∞
X 1
ln 0 = −
n
n=1
which we already know diverges.
25
3.3. Taylor Series Chapter 3. Power Series
On the other hand setting x = 1, gives the alternating harmonic series and a nice
convergence result:
∞
X (−1)n−1
ln 2 =
n
n=1
which goes to show that you cannot be sure whether a ratio test of 1 will give a
converging or diverging series.
3.3 Taylor Series

A more general result is:
f (1) (a) f (2) (a)

f (x) = f (a) + (x − a) + (x − a)2 + · · ·
1! 2!
∞ (n)
X f (a)
= (x − a)n
n!
n=0
allows us to create power series of a function around a particular point x = a that

we may be interested in. This allows us to create successively better approximations
to the function f (x) about x = a directly rather than having to use huge numbers of
terms from the Maclaurin series to achieve the same result.
Setting a = 0 gives us an expansion around zero and we recover the Maclaurin series
as a special case.
Example For a series like f (x) = ln x, we would not be able to create a Maclaurin
series since it ln 0 is not defined (is singular). So this is an example where an expan-
sion around another point is required. We show a Taylor expansion about x = 2:
1. f (x) = ln x
1
2. f 0 (x) = x
3. f 00 (x) = − x12
4. f 000 (x) = 2!(−1)2 x13
(−1)n−1 (n−1)!
5. f (n) = xn for n > 0
This gives us a Taylor series of:

∞
X (−1)n−1
f (x) = ln 2 + (x − 2)n
n2n
n=1
We can now show how the Taylor series approximates the function f (x) around the
point x = a in the same way as the Maclaurin series does around the point x = 0.
26
Chapter 3. Power Series 3.3. Taylor Series
3
f (x)
f0 (x)
f1 (x)
f2 (x)
f3 (x)
2
-1
-2
0 0.5 1 1.5 2 2.5 3 3.5 4
Figure 3.2: Successive approximations from the Taylor series expansion of f (x) = ln x
around the point x = 2.
P f (i) (a)
Figure 3.2 shows the first four partial sums fn (x) = ni=0 i! (x − a)i from the Taylor
series for f (x) = ln x around the point x = 2, where:
f0 (x) = ln 2
x−2
f1 (x) = ln 2 +
2
x − 2 (x − 2)2
f2 (x) = ln 2 + −
2 8
x − 2 (x − 2)2 (x − 2)3
f3 (x) = ln 2 + − +
2 8 24
As for Maclaurin series we still need to be aware of the radius of convergence for
this series. The techniques for calculating it are identical for Taylor series.
(−1)n−1
Taking the absolute ratio test for terms bn = n2n (x − 2)n , where we can ignore the
ln 2 term since we are looking in the limit of n → ∞:

(x − 2)n+1 /(n + 1)2n+1

bn+1
lim = lim
n→∞ bn n→∞ (x − 2)n /n2n

|x − 2| n
= lim
n→∞ 2 n+1
|x − 2|
=
2
This gives us a convergence condition of |x − 2| < 2 and a radius of convergence of 2
for this power series. In general the radius of convergence is limited by the nearest
singularity (such as x = 0 in this case) for a real series or pole if we extend complex
series to the complex plane (we may or may not have time to investigate poles!).
27
3.4. Taylor Series Error Term Chapter 3. Power Series
3.4 Taylor Series Error Term

In computing, numerical and scientific applications it is usual that we truncate the
series at the kth term so that the computation terminates. In these circumstances it
is important to know how close we are to the actual value of f (x) (in the worst case)
and are able to approximate the error that we have introduced by not summing the
entire series.
The infinite series gets translated into a finite one:
∞
X f (n) (a)
f (x) = (x − a)n
n!
n=0
k
X f (n) (a) f (k+1) (c)
= (x − a)n + (x − a)k+1 (3.1)
n! (k + 1)!
n=0
f (k+1) (c)
where c is a constant that lies between x and a. The term (k+1)! (x−a)k+1 is known as
the Lagrange error term and it replaces the tail of the infinite series from the (k +1)th
term onwards. Although it is precise there is no easy way to find c. So in practice
the bound a < c < x or x < c < a is used to generate a worst-case error for the term.
f (x)
Figure 3.3: Mean Value Theorem
Figure 3.3 shows a continuous and differentiable function f (x). The mean value
theorem states that between two points x and a, there is a point c such that f 0 (c) is
equal to the gradient between the points (a, f (a)) and (x, f (x)).
28
Chapter 3. Power Series 3.5. Deriving the Cauchy Error Term
Given that we can compute the gradient between those two points, we get:
f (x) − f (a)
f 0 (c) =
x−a
f (x) = f (a) + (x − a)f 0 (c) (3.2)
for some c between x and a. Taylor’s theorem can be thought of as a general version
of Mean Value Theorem but expressed in terms of the (k + 1)th derivative instead of
the 1st.
In fact if you set k = 0 in (3.1), you get the mean value theorem of (3.2).
3.5 Deriving the Cauchy Error Term

Let us start by defining the set of partial sums to the Taylor series. Here we are going
to fix x and vary the offset variable.
k
X f (n) (t)
Fk (t) = (x − t)n
n!
n=0
Now Fk (a) is the standard for of the nth partial sum and Fk (x) = f (x) for all k. Thus:
k
X f (n) (a)
Fk (x) − Fk (a) = f (x) − (x − a)n
n!
n=0
= Rk (x)
3.6 Power Series Solution of ODEs

Consider the differential equation
dy
= ky
dx
for constant k, given that y = 1 when x = 0.
Try the series solution
∞
X
y= ai xi
i=0
Find the coefficients ai by differentiating term by term, to obtain the identity, for
i ≥ 0:
Matching coefficients
∞
X ∞
X ∞
X
ai ixi−1 ≡ kai xi ≡ kai−1 xi−1
i=1 i=0 i=1
29
3.6. Power Series Solution of ODEs Chapter 3. Power Series
Comparing coefficients of xi−1 for i ≥ 1
iai = kai−1 hence
k k k ki
ai = ai−1 = · ai−2 = . . . = a0
i i i −1 i!
When x = 0, y = a0 so a0 = 1 by the boundary condition. Thus
∞
X (kx)i
y= = ekx
i!
i=0
30
Chapter 4
Differential Equations and Calculus
4.1 Differential Equations

4.1.1 Differential Equations: Background
• Used to model how systems evolve over time:
– e.g., computer systems, biological systems, chemical systems
• Terminology:
dy
– Ordinary differential equations (ODEs) are first order if they contain a
dx
term but no higher derivatives
d2 y
– ODEs are second order if they contain a term but no higher derivatives
dx2
4.1.2 Ordinary Differential Equations

• First order, constant coefficients:
dy
– For example, 2 +y = 0 (∗)
dx
– Try: y = emx
⇒ 2memx + emx = 0
⇒ emx (2m + 1) = 0
⇒ emx = 0 or m = − 21
– emx , 0 for any x, m. Therefore m = − 12
– General solution to (∗):
1
y = Ae− 2 x
31
4.1. Differential Equations Chapter 4. Differential Equations and Calculus
Ordinary Differential Equations

• First order, variable coefficients of type:
dy
+ f (x)y = g(x)
dx
R
f (x) dx
• Use integrating factor (IF): e
dy
– For example: + 2xy = x (∗)
dx R
2x dx 2
– Multiply throughout by IF: e = ex
2 dy x2 x2
⇒ ex dx + 2xe y = xe
d x2 x2
⇒ dx (e y) = xe
2 2 2
⇒ ex y = 12 ex + C So, y = Ce−x + 12

• Second order, constant coefficients:
d2 y dy
– For example, 2
+5 + 6y = 0 (∗)
dx dx
– Try: y = emx
⇒ m2 emx + 5memx + 6emx = 0
⇒ emx (m2 + 5m + 6) = 0
⇒ emx (m + 3)(m + 2) = 0
– m = −3, −2
– i.e., two possible solutions
– General solution to (∗):
y = Ae−2x + Be−3x

• Second order, constant coefficients:
– If y = f (x) and y = g(x) are distinct solutions to (∗)
– Then y = Af (x) + Bg(x) is also a solution of (∗) by following argument:
d2 d
* dx2
(Af(x) + Bg(x)) + 5 dx (Af (x) + Bg(x))
+ 6(Af (x) + Bg(x)) = 0
d2
!
d
* A f (x) + 5 f (x) + 6f (x)
dx2 dx
| {z }
=0
d2
!
d
+B g(x) + 5 g(x) + 6g(x) = 0
dx2 dx
| {z }
=0
32
Chapter 4. Differential Equations and Calculus 4.1. Differential Equations
• Second order, constant coefficients (repeated root):
d2 y dy
– For example, 2
−6 + 9y = 0 (∗)
dx dx
– Try: y = emx
⇒ m2 emx − 6memx + 9emx = 0
⇒ emx (m2 − 6m + 9) = 0
⇒ emx (m − 3)2 = 0
– m = 3 (twice)
– General solution to (∗) for repeated roots:
y = (Ax + B)e3x
Applications: Coupled ODEs
• Coupled ODEs are used to model massive state-space physical and computer
systems
• Coupled Ordinary Differential Equations are used to model:
– chemical reactions and concentrations

– biological systems
– epidemics and viral infection spread
– large state-space computer systems (e.g., distributed publish-subscribe
systems
4.1.3 Coupled ODEs

• Coupled ODEs are of the form:

dy1


 dx = ay1 + by2
 dy2


dx = cy1 + dy2
!
y1
• If we let y = , we can rewrite this as:
y2
dy1
  " # ! " #

dx  = a b
 y1 dy a b

dy2 or = y
  c d y2 dx c d
dx
33
4.2. Partial Derivatives Chapter 4. Differential Equations and Calculus
Coupled ODE Solutions

dy
• For coupled ODE of type: = Ay (∗)
dx
dy
• Try y = veλx so, = λveλx
dx
dy
• But also = Ay, so Aveλx = λveλx
dx
• Now solution of (∗) can be derived from an eigenvector solution of Av = λv
• For n eigenvectors v 1 , . . . , v n and corresp. eigenvalues λ1 , . . . , λn : general solu-
tion of (∗) is y = B1 v 1 eλ1 x + · · · + Bn v n eλn x
Coupled ODEs: Example

• Example coupled ODEs:

dy1


 dx = 2y1 + 8y2
 dy2


dx = 5y1 + 5y2
" #
dy 2 8
• So dx = y
5 5
• Require to find eigenvectors/values of
" #
2 8
A=
5 5
Coupled ODEs: Example

" #
2−λ 8
• Eigenvalues of A: det( ) = λ2 − 7λ − 30 = (λ − 10)(λ + 3) = 0
5 5−λ
• Thus eigenvalues λ = 10, −3
! !
1 8
• Giving: λ1 = 10, v 1 = ; λ2 = −3, v 2 =
1 −5
! !
1 10x 8 −3x
• Solution of ODEs: y = B1 e + B2 e
1 −5
4.2 Partial Derivatives

• Used in (amongst others):
– Computational Techniques (2nd Year)
– Optimisation (3rd Year)
– Computational Finance (4th Year)
34
Chapter 4. Differential Equations and Calculus 4.2. Partial Derivatives
Differentiation Contents
• What is a (partial) differentiation used for?
• Useful (partial) differentiation tools:
– Differentiation from first principles

– Partial derivative chain rule
– Derivatives of a parametric function
– Multiple partial derivatives
Optimisation
• Example: look to find best predicted gain in portfolio given different possible
share holdings in portfolio
Differentiation
δy
δx
• Gradient on a curve f (x) is approximately:
δy f (x + δx) − f (x)
=
δx δx
35
4.2.1 Definition of Derivative

• The derivative at a point x is defined by:
df f (x + δx) − f (x)
= lim
dx δx→0 δx
• Take f (x) = xn
– We want to show that:

df
= nxn−1
dx
Derivative of xn
df f (x+δx)−f (x)
• dx = limδx→0 δx
(x+δx)n −xn
= limδx→0 δx
Pn n n−i i n
( )x δx −x
= limδx→0 i=0 i δx
Pn n n−i i
( )x δx
= limδx→0 i=1 iδx
Pn n n−i
= limδx→0 i=1 i x δxi−1
n !
n
X n
= limδx→0 1 xn−1 + xn−i δxi−1
i
i=2
| {z }
→0 as δx→0
n!
= 1!(n−1)!
xn−1 = nxn−1
4.2.2 Partial Differentiation

df
• Ordinary differentiation dx applies to functions of one variable, i.e., f ≡ f (x)
• What if function f depends on one or more variables, e.g., f ≡ f (x1 , x2 )
• Finding the derivative involves finding the gradient of the function by varying
one variable and keeping the others constant
• For example for f (x, y) = x2 y + xy 3 ; partial derivatives are written:
∂f ∂f
– ∂x
= 2xy + y 3 and ∂y
= x2 + 3xy 2
36
Partial Derivative: Example
• f (x, y) = x2 + y 2
Partial Derivative: Example
• f (x, y) = x2 + y 2
– Fix y = k ⇒ g(x) = f (x, k) = x2 + k 2

dg ∂f
– Now dx = ∂x
= 2x
Further Examples
∂f ∂
• f (x, y) = (x + 2y 3 )2 ⇒ ∂x
= 2(x + 2y 3 ) ∂x (x + 2y 3 ) = 2(x + 2y 3 )
• If x and y are themselves functions of t then
df ∂f dx ∂f dy
= +
dt ∂x dt ∂y dt
• So if f (x, y) = x2 + 2y where x = sin t and y = cos t then:

df
– dt = 2x cos t − 2 sin t = 2 sin t(cos t − 1)
37
4.2.3 Extended Chain Rule

• If f is a function of x and y where x and y are themselves functions of s and t
then:
∂f ∂f ∂x ∂f ∂y
– ∂s
= ∂x ∂s
+ ∂y ∂s
∂f ∂f ∂x ∂f ∂y
– ∂t
= ∂x ∂t
+ ∂y ∂t
• which can be expressed as a matrix equation:

∂f
 ∂f
  ∂y
 
   ∂x   
∂s ∂s ∂s ∂x
 = 
     
∂f
   
 ∂f   ∂x ∂y  
∂t ∂t ∂t ∂y
• Useful for changes of variable, e.g., to polar coordinates
4.2.4 Jacobian
• The modulus of this matrix is called the Jacobian:
∂y

∂x
J = ∂s ∂s

∂x ∂y
∂t ∂t
• Just as when performing a substitution on the integral:

Z
f (x) dx
df (x)
we would use: du ≡ dx dx
• So if converting between multiple variables in an integration, we would use

du ≡ Jdx.
Formal Definition
• Similar to ordinary derivative. For a two variable function f (x, y) :
∂f f (x + δx, y) − f (x, y)
= lim
∂x δx→0 δx
• and in the y-direction:
∂f f (x, y + δy) − f (x, y)

= lim
∂y δy→0 δy
38
Further Notation
• Multiple partial derivatives (as for ordinary derivatives) are expressed:

∂2 f
– ∂x2
is the second partial derivative of f
∂n f
– ∂xn
is the nth partial derivative of f
∂2 f
– ∂x∂y
is the partial derivative obtained by first partial differentiating by y
and then x
∂2 f
– ∂y∂x
is the partial derivative obtained by first partial differentiating by x
and then y
∂2 f ∂2 f
• If f (x, y) is a nice function then: ∂x∂y
= ∂y∂x
39
Chapter 5
Complex Numbers
5.1 Introduction
We can see need for complex numbers by looking at the shortcomings of all the
simpler (more obvious) number systems that preceded them. In each case the next
number system in some sense fixes a perceived problem or omission with the previ-
ous one:
N Natural numbers, for counting, not closed under subtraction
Z Integers, the natural numbers with 0 and negative numbers, not closed under
division
Q Rational numbers, closed under arithmetic operations but cannot represent the
solution of all non-linear equations, e.g., x2 = 2
R Real numbers, solutions to some quadratic equations with real roots and some
higher-order equations, but not all, e.g., x2 + 1 = 0
C Complex numbers, we require these to represent all the roots of all polynomial
equations.1
Another important use of complex numbers is that often a real problem can be solved
by mapping it into complex space, deriving a solution, and mapping back again: a
direct solution may not be possible or would be much harder to derive in real space,
e.g., finding solutions to integration or summation problems, such as
Z x n
X
aθ
I= e cos bθ dθ or S= ak cos kθ . (5.1)
0 k=0
5.1.1 Applications
Complex numbers are important in many areas. Here are some:
1 Complex numbers form an algebraically closed field, where any polynomial equation has a root.
40
Chapter 5. Complex Numbers 5.1. Introduction
• Signal analysis (e.g., Fourier transformation to analyze varying voltages and

currents)
• Control theory (e.g., Laplace transformation from time to frequency domain)
• Quantum mechanics is founded on complex numbers (see Schrödinger equa-

tion and Heisenberg’s matrix mechanics)
• Cryptography (e.g., finding prime numbers).
• Machine learning: Using a pair of uniformly distributed random numbers (x, y),
we can generate random numbers in polar form (r cos(θ), r sin(θ)). This can
lead to efficient sampling methods like the Box-Muller transform (Box and
Muller, 1958).2 The variant of the Box-Muller transform using complex num-
bers was proposed by Knop (1969).
• (Tele)communication: digital coding modulations
5.1.2 Imaginary Number

An entity we cannot describe using real numbers are the roots to the equation
x2 + 1 = 0, (5.2)
which we will call i and define as

√
i := −1. (5.3)
There is no way√of squeezing this into R, it cannot be compared with a real number
(in contrast to 2 or π, which we can compare with rationals and get arbitrarily
accurate approximations in the rationals). We call i the imaginary number/unit,
orthogonal to the reals.
Properties From the definition of i in (5.3) we get a number of properties for i.
1. i 2 = −1, i 3 = i 2 i = −i, i 4 = (i 2 )2 = (−1)2 = 1 and so on
2. In general i 2n = (i 2 )n = (−1)n , i 2n+1 = i 2n i = (−1)n i for all n ∈ N

1 i
3. i −1 = i = i2
= −i
1 1
4. In general i −2n = i 2n
= (−1)n
= (−1)n , i −(2n+1) = i −2n i −1 = (−1)n+1 i for all n ∈ N
5. i 0 = 1
2 This is a pseudo-random number sampling method, e.g., for generating pairs of independent,
standard, normally distributed (zero mean, unit variance) random numbers, given a source of uni-
formly distributed random numbers.
41
5.1. Introduction Chapter 5. Complex Numbers
Im
z = (x, y) = x + iy
iy
1 x Re
Figure 5.1: Complex plane (Argand diagram). A complex number can be represented
in a two-dimensional Cartesian coordinate system with coordinates x and y. x is the real
part and y is the imaginary part of a complex number z = x + iy.
5.1.3 Complex Numbers as Elements of R2

It is convenient (and correct3 ) to consider complex numbers
C := {a + ib : a, b ∈ R, i 2 = −1} (5.4)
as the set of tuples (a, b) ∈ R2 with the following definition of addition and multipli-
cation:
(a, b) + (c, d) = (a + c, b + d) , (5.5)

(a, b) · (c, d) = (ac − bd, ad + bc) . (5.6)
In this context, the element i := (0, 1) is the imaginary number/unit. With the
complex multiplication defined in (5.6), we immediately obtain
i 2 = (0, 1)2 = (0, 1)(0, 1) = −1, (5.7)
which allows us to factorize the polynomial z2 + 1 fully into (z − i)(z + i).

Since elements of R2 can be drawn in a plane, we can do the same with complex
numbers z ∈ C. The plane is called complex plane or Argand diagram, see Fig-
ure 5.1.
The Argand diagram allows us to visualize addition and multiplication, which are
defined in (5.5)–(5.6).
5.1.4 Closure under Arithmetic Operators

Closing R ∪ {i} under the arithmetic operators +, · as defined in (5.5)–(5.6) gives the
complex numbers, C. To be more specific, if z1 , z2 ∈ C, then z1 + z2 ∈ C, z1 − z2 ∈ C,
z1 · z2 ∈ C and z1 /z2 ∈ C.
3 There exists a bijective linear mapping (isomorphism) between C and R2 . We will briefly discuss
this in the Linear Algebra part of the course.
42
Chapter 5. Complex Numbers 5.2. Representations of Complex Numbers
Im
z2
Re
z1 + z2
z1
Figure 5.2: Visualization of complex addition. As known from geometry, we simply add
the two vectors representing complex numbers.
5.2 Representations of Complex Numbers

In the following, we will discuss three important representations of complex num-
bers.
5.2.1 Cartesian Coordinates

Every element z ∈ C can be decomposed into
(x, y) = (x, 0) + (0, y) = (x, 0) + (0, 1)(y, 0) = (x, 0) +i (y, 0) = x + iy. (5.8)
|{z} |{z}
∈R ∈R
Therefore, every z = x + iy ∈ C has a coordinate representation (x, y), where x

is called the real part and y is called the imaginary part of z, and we write x =
<(z), y = =(z), respectively. z = x + iy is the point (x, y) in the xy-plane (com-
plex plane), which is uniquely determined by its Cartesian coordinates (x, y). An
illustration is given in Figure 5.1.
5.2.2 Polar Coordinates

Equivalently, (x, y) can be represented by polar coordinates, r, φ, where r is the
distance of z from the origin 0, and φ is the angle between the (positive) x-axis and
the direction 0z~ . Then,
z = r(cos φ + i sin φ), r ≥ 0, 0 ≤ φ < 2π (5.9)

uniquely determines z ∈ C. The polar coordinates of z are then
q
r = |z| = x2 + y 2 , (5.10)
φ = Arg z , (5.11)
~ (the distance of z from the origin) and φ is the argument
where r is the length of 0z
of z.
43
5.2. Representations of Complex Numbers Chapter 5. Complex Numbers
z = (x, y) = r(cos φ + i sin φ)

iy
i r
φ
Figure 5.3: Polar coordinates.
Im
r sin φ
φ r
r cos φ Re
Figure 5.4: Euler representation. In the Euler representation, a complex number z =

r exp(iφ) “lives” on a circle with radius r around the origin. Therefore, r exp(iφ) =
r(cos φ + i sin φ).
5.2.3 Euler Representation

The third representation of complex numbers is the Euler representation
z = r exp(iφ) (5.12)
where r and φ are the polar coordinates. We already know that z = r(cos φ + i sin φ),
i.e., it must also hold that r exp(iφ) = r(cos φ+i sin φ). This can be proved by looking
at the power series expansions of exp, sin, and cos:
∞
X (iφ)k (iφ)2 (iφ)3 (iφ)4 (iφ)5
exp(iφ) = = 1 + iφ + + + + + ··· (5.13)
k! 2! 3! 4! 5!
k=0
φ2 iφ3 φ4 iφ5
= 1 + iφ − − + + ∓ ··· (5.14)
2! 3! 4! 5!
44
φ2 φ4 φ3 φ5
! !
= 1− + ∓ ··· + i φ − + ∓ ··· (5.15)
2! 4! 3! 5!
∞ ∞
X (−1)k φ2k X (−1)k φ2k+1
= +i = cos φ + i sin φ . (5.16)
(2k)! (2k + 1)!
k=0 k=0
Therefore, z = exp(iφ) is a complex number, which lives on the unit circle (|z| = 1)
and traces out the unit circle in the complex plane as φ ranges through the real
numbers.
5.2.4 Transformation between Polar and Cartesian Coordinates
Cartesian coordinates Polar coordinates

x, y r, φ
p
x = r cos φ r r = x2 + y 2
y
y = r sin φ tan φ = xy + quadrant
z = x + iy = r(cos φ + i sin φ)
Figure 5.5: Transformation between Cartesian and polar coordinate representations of
complex numbers.
Figure 5.5 summarizes the transformation between Cartesian and polar coordinate
representations of complex numbers z. We have to pay some attention when com-
puting Arg (z) when transforming Cartesian coordinates into polar coordinates.
Example: Transformation from Polar to Cartesian Coordinates

Transform the polar representation z = (r, φ) = (2, 2π 3 ) into Cartesian coordinates
(x, y).
It is always useful to draw the complex number. Figure 5.6(a) shows the setting. We
are interested in the blue dots. With x = r cos φ and y = r sin φ, we obtain
x = r cos( 23 π) = −1 (5.17)
√
y = r sin( 32 π) = 3 . (5.18)
√
Therefore, z = −1 + i 3.
Example: Transformation from Cartesian to Polar Coordinates

Getting the Cartesian coordinates from polar coordinates is straightforward. The
transformation from Cartesian to polar coordinates is somewhat more difficult be-
cause of the argument φ. The reason is that tan has a period of π, which means
45
5.2. Representations of Complex Numbers Chapter 5. Complex Numbers
y
z r=2 Im
φ= 2π
3
Re
x
z = 2 − 2i
2π
(a) (r, φ) = (2, 3 ) (b) (x, y) = (2, −2)
z = −1 + i Im Im
1 1
Re Re
z= − 32 i
(c) (x, y) = (−1, 1) (d) (x, y) = (0, − 32 )
Figure 5.6: Coordinate transformations
that y/x has two possible angles, which differ by π, see Figure 5.7. By looking at the
quadrant in which the complex number z lives we can resolve this ambiguity. Let us
have a look at some examples:
√ √
1. z = 2 − 2i. We immediately obtain r = 22 + 22 = 2 2. For the argument,
we obtain tan φ = − 22 = −1. Therefore, φ ∈ { 34 π, 74 π}. We identify the correct
argument by plotting the complex number and identifying the quadrant. Fig-
ure 5.6(b) shows that z lies in the fourth quadrant. Therefore, φ = 74 π.
2. z = −1 + i.
√ √
r = 1+1 = 2 (5.19)
−1
tan φ = = −1 ⇒ φ ∈ { 34 π, 74 π} . (5.20)
1
Figure 5.6(c) shows that z lies in the second quadrant. Therefore, φ = 43 π.
3. z = − 23 i.
3
r= 2 (5.21)
− 32 π 3
tan φ = ⇒ φ ∈ { , π} (5.22)
0 2 2
Figure 5.6(d) shows that z is between the third and fourth quadrant (and not
between the first and second). Therefore, φ = 23 π
46
φ1 φ2 = φ1 + π
Figure 5.7: Tangens. Since the tangens possesses a period of π, there are two solutions
for the argument 0 ≤ φ < 2π of a complex number, which differ by π.
5.2.5 Geometric Interpretation of the Product of Complex Num-

bers
Let us now use the polar coordinate representation of complex numbers to ge-
ometrically interpret the product z = z1 z2 of two complex numbers z1 , z2 . For
z1 = r1 (cos θ1 + i sin θ1 ) and z2 = r2 (cos θ2 + i sin θ2 ) we obtain
z1 z2 = r1 r2 (cos θ1 cos θ2 − sin θ1 sin θ2 + i(sin θ1 cos θ2 + cos θ1 sin θ2 ))
= r1 r2 (cos(θ1 + θ2 ) + i sin(θ1 + θ2 )) . (5.23)
1. The length r = |z| = |z1 | |z2 | is the product of the lengths of z1 and z2 .
2. The argument of z is the sum of the arguments of z1 and z2 .
This means that when we multiply two complex numbers z1 , z2 , the corresponding
distances r1 and r2 are multiplied while the corresponding arguments φ1 , φ2 are
summed up. This means, we are now ready to visualize complex multiplication, see
Figure 5.8. Overall, multiplying z1 with z2 performs two (linear) transformations on
z1 : a scaling by r2 and a rotation by φ2 . Similarly, the transformations acting on z2
are a scaling by r1 and a rotation by φ1 .
5.2.6 Powers of Complex Numbers

We will encounter situations where we need to compute powers of complex numbers
of the form zn . For this, we can use some advantages of some representations of
complex numbers. For instance, if we consider the representation using Cartesian
coordinates computing zn = (x + iy)n for large n will be rather laborious. However,
the Euler representation makes our lives a bit easier since
zn = (r exp(iφ))n = r n exp(inφ) (5.24)
47
5.3. Complex Conjugate Chapter 5. Complex Numbers
Im
z1
Re
z2
z1 z2
Figure 5.8: Complex multiplication. When we multiply two complex numbers z1 , z2 ,
the corresponding distances r1 and r2 are multiplied while the corresponding arguments
φ1 , φ2 are summed up.
Im
z = x + iy
Re
z = x − iy
Figure 5.9: The complex conjugate z is a reflection of z about the real axis.
can be computed efficiently: The distance r to the origin is simply raised to the
power of n and the argument is scaled/multiplied by n. This also immediately gives
us the result
(r(cos φ + i sin φ))n = r n (cos(nφ) + i sin(nφ)) (5.25)
which will later (Section 5.4) know as de Moivre’s theorem.
5.3 Complex Conjugate

The complex conjugate of a complex number z = x+iy is z = x−iy. Some properties
of complex conjugates include:
1. <(z) = <(z)
2. =(z) = −=(z)
3. z + z = 2x = 2<(z) ∈ R
48
Chapter 5. Complex Numbers 5.3. Complex Conjugate
4. z − z = 2iy = 2i=(z) is purely imaginary
5. z1 + z2 = z1 + z2
6. z1 z2 = z1 z2 . This can be seen either by noting that the conjugate operation

simply changes every occurrence of i to −i or since
(x1 + iy1 )(x2 + iy2 ) = (x1 x2 − y1 y2 ) + i(x1 y2 + y1 x2 ) , (5.26)

(x1 − iy1 )(x2 − iy2 ) = (x1 x2 − y1 y2 ) − i(x1 y2 + y1 x2 ) , (5.27)
which are conjugates. Geometrically, the complex conjugate z is a reflection of

z where the real axis serves as the axis of reflection. Figure 5.9 illustrates this
relationship.
5.3.1 Absolute Value of a Complex Number

√
The absolute value (length/modulus) of z ∈ C is |z| = zz, where
zz = (x + iy)(x − iy) = x2 + y 2 ∈ R. (5.28)
Notice that the term ‘absolute value’ is the same as defined for real numbers when
=(z) = 0. In this case, |z| = |x|.
The absolute value of the product has the following nice property that matches the
product result for real numbers:
|z1 z2 | = |z1 | |z2 |. (5.29)
This holds since
|z1 z2 |2 = z1 z2 z1 z2 = z1 z2 z1 z2 = z1 z1 z2 z2 = |z1 |2 |z2 |2 . (5.30)
5.3.2 Inverse and Division

If z = x + iy, its inverse (reciprocal) is
1 z z x − iy
= = 2= 2 . (5.31)
z zz |z| x + y2
This can be written z−1 = |z|−2 z, using only the complex operators multiply and add,
see (5.5) and (5.6), but also real division, which we already know. Complex division
is now defined by z1 /z2 = z1 z2−1 . In practice, we compute the division z1 /z2 by ex-
panding the fraction by the complex conjugate of the denominator. This ensures that
the denominator’s imaginary part is 0 (only the real part remains), and the overall
fraction can be written as
z1 z1 z2 z1 z2
= = (5.32)
z2 z2 z2 |z2 |2
49
5.4. De Moivre’s Theorem Chapter 5. Complex Numbers
Geometric Interpretation of Division
When we use the Euler representations of two complex numbers z1 , z2 ∈ C, we can

write the division as
z1 r
= z1 z2−1 = r1 exp(iφ1 ) r2 exp(iφ2 ) = 1 exp(i(φ1 − φ2 )) . (5.33)
z2 r2
Geometrically, we divide r1 by r2 (equivalently: scale r1 by r12 ) and rotate z1 by −φ2 .

This is not overly surprising since the division by z2 does exactly the opposite of
a multiplication by r2 . Therefore, looking again at Figure 5.8, if we take z1 z2 and
divide by z2 , we obtain z1 .
Example: Complex Division
Bring the following fraction into the form x + iy:
3 + 2i
z = x + iy = (5.34)
7 − 3i
Solution:
3 + 2i (3 + 2i)(7 + 3i) 15 + 23i 15 23
= = = +i (5.35)
7 − 3i (7 − 3i)(7 + 3i) 49 + 9 58 58
15 23
Now, the fraction can be written as z = x + iy with x = 58 and y = 58 .
5.4 De Moivre’s Theorem

De Moivre’s theorem (or formula) is a central result because it connects complex
numbers and trigonometry.
Theorem 1 (De Moivre’s Theorem)

For any n ∈ N
(cos φ + i sin φ)n = cos nφ + i sin nφ (5.36)
The proof is done by induction (which you will see in detail in the course Reasoning
about Programs). A proof by induction allows you to prove that a property is true for
all values of a natural number n. To construct an induction proof, you have to prove
that the property, P (n), is true for some base value (say, n = 1). A further proof is
required to show that if it is true for the parameter n = k, then that implies it is also
true for the parameter n = k + 1: that is P (k) ⇒ P (k + 1) for all k ≥ 1. The two proofs
combined allow us to build an arbitrary chain of implication up to some value n = m:
P (1) and (P (1) ⇒ P (2) ⇒ · · · ⇒ P (m − 1) ⇒ P (m)) |= P (m)
50
Chapter 5. Complex Numbers 5.4. De Moivre’s Theorem
Proof 1
We start the induction proof by checking whether de Moivre’s theorem holds for n = 1:
(cos φ + i sin φ)1 = cos φ + i sin φ (5.37)
is trivially true, and we can now make the induction step: We assume that (5.36) is
true for k and show that it also holds for k + 1.
Assuming
(cos φ + i sin φ)k = cos kφ + i sin kφ (5.38)
we can write
(cos φ + i sin φ)k+1 = (cos φ + i sin φ)(cos φ + i sin φ)k

= (cos φ + i sin φ)(cos kφ + i sin kφ) using assumption (5.38)
= (cos(k + 1)φ + i sin(k + 1)φ) using complex product (5.23)
which concludes the proof.
5.4.1 Integer Extension to De Moivre’s Theorem

We can extend de Moivre to include negative numbers, n ∈ Z
(cos φ + i sin φ)n = cos nφ + i sin nφ
We have tackled the case for n > 0 already, n = 0 can be shown individually. So we
take the case n < 0. We let n = −m for m > 0.
1
(cos φ + i sin φ)n =
(cos φ + i sin φ)m
1
= by de Moivre’s theorem
cos mφ + i sin mφ
cos mφ − i sin mφ
=
cos2 mφ + sin2 mφ
= cos(−mφ) + i sin(−mφ) Trig. identity: cos2 mφ + sin2 mφ = 1
= cos nφ + i sin nφ
5.4.2 Rational Extension to De Moivre’s Theorem

Finally, for our purposes, we will show that if n ∈ Q, one value of (cos φ + i sin φ)n is
cos nφ + i sin nφ. Take n = p/q for p, q ∈ Z and q , 0. We will use both de Moivre’s
theorems in the following:
!q
p p
cos φ + i sin φ = cos pφ + i sin pφ (5.39)
q q
= (cos φ + i sin φ)p (5.40)
51
5.5. Triangle Inequality for Complex Numbers Chapter 5. Complex Numbers
p p
Hence cos q φ + i sin q φ is one of the qth roots of (cosφ + i sin φ)p .
The qth roots of cos φ + i sin φ are easily obtained. We need to use the fact that
(repeatedly) adding 2π to the argument of a complex number does not change the
complex number.
1 1
(cos φ + i sin φ) q = (cos(φ + 2nπ) + i sin(φ + 2nπ)) q (5.41)
φ + 2nπ φ + 2nπ
= cos + i sin for 0 ≤ n < q (5.42)
q q
We will use this later to calculate roots of complex numbers.
Finally, the full set of values for (cos +i sin φ)n for n = p/q ∈ Q is:
pφ + 2nπ pφ + 2nπ
cos + i sin for 0 ≤ n < q (5.43)
q q
Example: Multiplication using Complex Products

We require the result of:
(3 + 3i)(1 + i)3
We could expand (1+i)3 and multiply by 3+3i using real and imaginary components.
Alternatively, we could tackle this in polar form (cos φ + i sin φ) using the complex
product of (5.23) and de Moivre’s theorem.
(1 + i)3 = [21/2 (cos π/4 + i sin π/4)]3
= 23/2 (cos 3π/4 + i sin 3π/4)
by de Moivre’s theorem. 3 + 3i = 181/2 (cos π/4 + i sin π/4) and so the result is
181/2 23/2 (cos π + i sin π) = −12
Geometrically, we just observe that the Arg of the second number is 3 times that of
1 + i, i.e., 3π/4 (or 3 · 45◦ in degrees). The first number has the same Arg, so the Arg
of the result is π. √ √
Similarly, the absolute values (lengths) of the numbers multiplied are 18 and 23 ,
so the product has absolute value 12. The result is therefore −12.
5.5 Triangle Inequality for Complex Numbers

The triangle inequality for complex numbers is as follows:
∀z1 , z2 ∈ C : |z1 + z2 | ≤ |z1 | + |z2 | (5.44)
An alternative form, with w1 = z1 and w2 = z1 + z2 is |w2 | − |w1 | ≤ |w2 − w1 | and,
switching w1 , w2 , |w1 | − |w2 | ≤ |w2 − w1 |. Thus, relabelling back to z1 , z2 :

∀z , z ∈ C : |z | − |z | ≤ |z − z |
1 2 1 2 2 1 (5.45)
In the Argand diagram, this just says that “In the triangle with vertices at 0, z1 , z2 , the
length of side z1 z2 is not less than the difference between the lengths of the other
two sides”.
52
Chapter 5. Complex Numbers 5.6. Fundamental Theorem of Algebra
Proof 2
Let z1 = x1 + iy1 and z2 = x2 + iy2 . Squaring the left-hand side of (5.45) yields
(x1 + x2 )2 + (y1 + y2 )2 = |z1 |2 + |z2 |2 + 2(x1 x2 + y1 y2 ), (5.46)
and the square of the right-hand side is
|z1 |2 + |z2 |2 + 2|z1 ||z2 | (5.47)
It is required to prove x1 x2 + y1 y2 ≤ |z1 ||z2 |. We continue by squaring this inequality
x1 x2 + y1 y2 ≤ |z1 ||z2 | (5.48)

⇔ (x1 x2 + y1 y2 )2 ≤ |z1 |2 |z2 |2 (5.49)
⇔ x12 x22 + y12 y22 + 2x1 x2 y1 y2 ≤ x12 x22 + y12 y22 + x12 y22 + y12 x22 (5.50)
2
⇔ 0 ≤ (x1 y2 − y1 x2 ) , (5.51)
which concludes the proof.
The geometrical argument via the Argand diagram is a good way to understand the
triangle inequality.
5.6 Fundamental Theorem of Algebra

Theorem 2 (Fundamental Theorem of Algebra)
Any polynomial of degree n of the form
n
X
p(z) = ak zk , ak ∈ C, an , 0 (5.52)
k=0
possesses, counted with multiplicity, exactly n roots in C.
A root z∗ of p(z) satisfies p(z∗ ) = 0. Bear in mind that complex roots include all real
roots as the real numbers are a subset of the complex numbers. Also some of the
roots might be coincident, e.g., for z2 = 0. Finally, we also know that if ω is a root
and ω ∈ C\R, then ω is also a root. So all truly complex roots occur in complex
conjugate pairs.
5.6.1 nth Roots of Unity

In the following, we consider the equation
zn = 1 , n ∈ N, (5.53)
for which we want to determine the roots. The fundamental theorem of algebra tells
us that there exist exactly n roots, one of which is z = 1.
53
5.6. Fundamental Theorem of Algebra Chapter 5. Complex Numbers
Im
1 Re
Figure 5.10: Then nth roots of zn = 1 lie on the unit circle and form a regular polygon.
Here, we show this for n = 8.
To find the other solutions, we write (5.53) in a slightly different form using the
Euler representation:
zn = 1 = eik2π , ∀k ∈ Z . (5.54)
Then the solutions are z = ei2kπ/n for k = 0, 1, 2, . . . , n − 1.4

Geometrically, all n roots lie on the unit circle, and they form a regular polygon
with n corners where the roots are 360◦ /n apart, see an example in Figure 5.10.
Therefore, if we know a single root and the total number of roots, we could even
geometrically find all other roots.
Example: Cube Roots of Unity
The 3rd roots of 1 are z = e2kπi/3 for k = 0, 1, 2, i.e., 1, e2πi/3 , e4πi/3 . These are often
referred to as ω1 ω1 and ω3 , and simplify to
ω1 = 1
√
ω2 = cos 2π/3 + i sin 2π/3 = (−1 + i 3)/2 ,
√
ω3 = cos 4π/3 + i sin 4π/3 = (−1 − i 3)/2 .
Try cubing each solution directly to validate that they are indeed cubic roots.
5.6.2 Solution of zn = a + ib
Finding the n roots of zn = a + ib is similar to the approach discussed above: Let
a + ib = reiφ in polar form. Then, for k = 0, 1, . . . , n − 1,
zn = (a + ib)e2πki = re(φ+2πk)i (5.55)

1 (φ+2πk)
i
⇒ zk = r n e n , k = 0, . . . , n − 1 . (5.56)
4 Note that the solutions repeat when k = n, n + 1, . . .
54
Chapter 5. Complex Numbers 5.7. Complex Sequences and Series*
Example
Determine the cube roots of 1 − i.
√
1. The polar coordinates of 1 − i are r = 2, φ = 74 π, and the corresponding Euler
representation is
√
z = 2 exp(i 7π
4 ). (5.57)
2. Using (5.56), the cube roots of z are

1 1
z1 = 2 6 (cos 7π 7π 7π
12 + i sin 12 ) = 2 exp(i 12 )
6 (5.58)
1 1 1
z2 = 2 (cos 15π
6 15π 5π 5π 5π
12 + i sin 12 ) = 2 (cos 4 + i sin 4 ) = 2 exp(i 4 )
6 6 (5.59)
1 1
z3 = 2 6 (cos 23π 23π 23π
12 + i sin 12 ) = 2 exp(i 12 ) .
6 (5.60)
5.7 Complex Sequences and Series*

A substantial part of the theory that we have developed for convergence of sequences
and series of real numbers also applies to complex numbers. We will not reproduce
all the results here, there is no need; we will highlight a couple of key concepts
instead.
5.7.1 Limits of a Complex Sequence

For a sequence of complex numbers z1 , z2 , z3 , . . ., we can define limits of convergence,
zn → l as n → ∞ where zn , l ∈ C. This means that for all > 0 we can find a natural
number N , such that
∀n > N : |zn − l| < . (5.61)
The only distinction here is the meaning of |zn − l|, which refers to the complex
absolute value and not the absolute real value.
Example of complex sequence convergence Prove that the complex sequence

1
zn = n+i converges to 0 as n → ∞. Straight to the limit inequality:

1 < (5.62)
n + i
|n − i|
⇔ 2 < (5.63)
p +1
n
(n − i)(n + i)
⇔ < (5.64)
n2 + 1
1
⇔√ < (5.65)
n2 + 1
55
5.7. Complex Sequences and Series* Chapter 5. Complex Numbers
r
1
⇒n> −1 for ≤ 1 (5.66)
2
Thus, we can set

 q
1
−1 ≤ 1


2

N () =  (5.67)


 1 otherwise

We have to be a tiny bit careful as N () needs to be defined for all > 0 and the
penultimate line of the limit inequality is true for all n > 0 if > 1. In essence this
was no different in structure from the normal sequence convergence proof. The only
difference was how we treated the absolute value.
Absolute Convergence
Similarly, a complex series ∞
P P∞
n=1 zn is absolutely convergent if n=1 |zn | converges.
Again the |zn | refers to the complex absolute value.
Complex Ratio Test

A complex series ∞
P
n=1 zn converges if

zn+1
lim <1 (5.68)
n→∞ zn
and diverges if

zn+1
lim > 1. (5.69)
n→∞ zn
Example of Complex Series Convergence

Let us take a general variant of the geometric series:
∞
X
S= azn−1 (5.70)
n=1
We can prove that this will converge for some values of z ∈ C in the same way we
n
could for the real-valued series. Applying the complex ratio test, we get limn→∞ | azazn−1 | =
|z|. We apply the standard condition and get that |z| < 1 for this series to converge.
The radius of convergence is still 1 (and is an actual radius of a circle in the complex
plane). What is different here is that now any z-point taken from within the circle
centred on the origin with radius 1 will make the series converge, not just on the
real interval (−1, 1).
a
For your information, the limit of this series is 1−z , which you can show using
Maclaurin as usual, discussed below.
56
Chapter 5. Complex Numbers 5.8. Complex Power Series
5.8 Complex Power Series

We can expand functions as power series in a complex variable, usually z, in the
same way as we could with real-valued functions. The same expansions hold in C
because the functions below (at any rate) are differentiable in the complex domain.
Therefore, Maclaurin’s series applies and yields
∞ n
X z z2 z3
exp(z) = = 1 + z + + + ... (5.71)
n! 2! 3!
n=0
∞
X z2n+1 z3 z5
sin(z) = (−1)n = z − + − ... (5.72)
(2n + 1)! 3! 5!
n=0
∞
X z2n z2 z4
cos(z) = (−1)n = 1 − + − ... (5.73)
(2n)! 2! 4!
n=0
5.8.1 A Generalized Euler Formula

A more general form of Euler’s formula (5.12) is
∀z ∈ C, n ∈ Z : z = rei(φ+2nπ) (5.74)
since ei2nπ = cos 2nπ + i sin 2nπ = 1. This is the same general form we used in the
rational extension to De Moivres theorem to access the many roots of a complex
number.
In terms of the Argand diagram, the points ei(φ+2nπ) for i ≥ 1 lie on top of each other,
each corresponding to one more revolution (through 2π).
The complex conjugate of eiφ is e−iφ = cos φ − i sin φ. This allows us to get useful
expressions for sin φ and cos φ:
cos φ = (eiφ + e−iφ )/2 (5.75)

sin φ = (eiφ − e−iφ )/2i. (5.76)
We will be able to use these relationships to create trigonometric identities.
5.9 Applications of Complex Numbers*

5.9.1 Trigonometric Multiple Angle Formulae
How can we calculate cos nφ in terms of cos φ and sin φ? We can use de Moivre’s
theorem to expand einφ and equate real and imaginary parts: e.g., for n = 5, by the
Binomial theorem,
(cos φ + i sin φ)5 = cos5 φ + i5 cos4 φ sin φ − 10 cos3 φ sin2 φ (5.77)

− i10 cos2 φ sin3 φ + 5 cos φ sin4 φ + i sin5 φ
57
5.9. Applications of Complex Numbers* Chapter 5. Complex Numbers
Comparing real and imaginary parts now gives
cos 5φ = cos5 φ − 10 cos3 φ sin2 φ + 5 cos φ sin4 φ (5.78)
and
sin 5φ = 5 cos4 φ sin φ − 10 cos2 φ sin3 φ + sin5 φ (5.79)
Trigonometric Power Formulae

We can also calculate cosn φ in terms of cos mφ and sin mφ for m ∈ N: Let z = eiφ so
that z + z−1 = z + z = 2 cos φ. Similarly, zm + z−m = 2 cos mφ by de Moivre’s theorem.
Hence by the Binomial theorem, e.g., for n = 5,
(z + z−1 )5 = (z5 + z−5 ) + 5(z3 + z−3 ) + 10(z + z−1 ) (5.80)

5 5
2 cos φ = 2(cos 5φ + 5 cos 3φ + 10 cos φ) (5.81)
Similarly, z − z−1 = 2i sin φ gives sinn φ

When n is even, we get an extra term in the binomial expansion, which is constant.
For example, for n = 6, we obtain
(z + z−1 )6 = (z6 + z−6 ) + 6(z4 + z−4 ) + 15(z2 + z−2 ) + 20 (5.82)

26 cos6 φ = 2(cos 6φ + 6 cos 4φ + 15 cos 2φ + 10) (5.83)
and, therefore,
1
cos6 φ = (cos 6φ + 6 cos 4φ + 15 cos 2φ + 10) . (5.84)
32
5.9.2 Summation of Series

Some series with sines and cosines can be summed similarly, e.g.,
n
X
C= ak cos kφ (5.85)
k=0
n
ak sin kφ. Then,
P
Let S =
k=1
n
X 1 − (aeiφ )n+1
C + iS = ak eikφ = . (5.86)
k=0
1 − aeiφ
Hence,
(1 − (aeiφ )n+1 )(1 − ae−iφ )

C + iS = (5.87)
(1 − aeiφ )(1 − ae−iφ )
1 − ae−iφ − an+1 ei(n+1)φ + an+2 einφ
= . (5.88)
1 − 2a cos φ + a2
58
Chapter 5. Complex Numbers 5.9. Applications of Complex Numbers*
Equating real and imaginary parts, the cosine series is
1 − a cos φ − an+1 cos(n + 1)φ + an+2 cos nφ

C= , (5.89)
1 − 2a cos φ + a2
and the sine series is

a sin φ − an+1 sin(n + 1)φ + an+2 sin nφ
S= (5.90)
1 − 2a cos φ + a2
5.9.3 Integrals
We can determine integrals
Z x
C= eaφ cos bφdφ , (5.91)
Z0x
S= eaφ sin bφdφ (5.92)
0
by looking at the sum5

Z x
C + iS = e(a+ib)φ dφ (5.93)
0
e (a+ib)x −1(eax eibx − 1)(a − ib)
= = (5.94)
a + ib a2 + b2
ax ax
(e cos bx − 1 + ie sin bx)(a − ib)
= (5.95)
a2 + b2
The result is therefore
eax (a cos bx + b sin bx) − a + i(eax (a sin bx − b cos bx) + b)
C + iS = (5.96)
a2 + b2
and so we get
eax (a cos bx + b sin bx − a)

C= , (5.97)
a2 + b2
eax (a sin bx − b cos bx) + b
S= (5.98)
a2 + b2
as the solutions to the integrals we were seeking.
5 The reduction formula would require a and b to be integers.
59
Chapter 6
Linear Algebra
This chapter is largely based on the lecture notes and books by Drumm and Weil
(2001); Strang (2003); Hogben (2013) as well as Pavel Grinfeld’s Linear Algebra
series1 . Another excellent source is Gilbert Strang’s Linear Algebra lecture at MIT2 .
Linear algebra is the study of vectors. Generally, vectors are special objects that can
be added together and multiplied by scalars to produce another object of the same
kind. Any object that satisfies these two properties can be considered a vector. Here
are three examples of such vectors:
1. Geometric vectors. This example of a vector may be familiar from High School.
Geometric vectors are directed segments, which can be drawn, see Fig. 6.1.
Two vectors x~, y~ can be added, such that x~ + y~ = ~z is another geometric vector.
Furthermore, λx~, λ ∈ R is also a geometric vector. In fact, it is the original
vector scaled by λ. Therefore, geometric vectors are instances of the vector
concepts introduced above.
2. Polynomials are also vectors: Two polynomials can be added together, which
results in another polynomial; and they can be multiplied by a scalar λ ∈ R, and
the result is a polynomial as well. Therefore, polynomial are (rather unusual)
instances of vectors. Note that polynomials are very different from geometric
vectors. While geometric vectors are concrete “drawings”, polynomials are
abstract concepts. However, they are both vectors.
3. Rn is a set of numbers, and its elements are n-tuples. Rn is even more abstract
than polynomials, and the most general concept we consider in this course. For
example,
1
 
a = 2 ∈ R3 (6.1)
 
3
 
is an example of a triplet of numbers. Adding two vectors a, b ∈ Rn component-

wise results in another vector: a + b = c ∈ Rn . Moreover, multiplying a ∈ Rn by
λ ∈ R results in a scaled vector λa ∈ Rn .
1 http://tinyurl.com/nahclwm
2 http://tinyurl.com/29p5q8j
60
Chapter 6. Linear Algebra
~y
~x
Figure 6.1: Example of two geometric vectors in two dimensions.
Linear algebra focuses on the similarities between these vector concepts: We can
add them together and multiply them by scalars. We will largely focus on the third
kind of vectors since most algorithms in linear algebra are formulated in Rn . There
is a 1:1 correspondence between any kind of vector and Rn . By studying Rn , we
implicitly study all other vectors. Although Rn is rather abstract, it is most useful.
Practical Applications of Linear Algebra

Linear algebra centers around solving linear equation systems and is at the core of
many computer science applications. Here is a selection:3
• Ranking of web pages (web search)
• Linear programming (optimziation)
• Error correcting codes (e.g., in DVDs)
• Decomposition of sounds into different sources
• Projections, rotations, scaling (computer graphics)
• Data visualization
• En/Decryption algorithms (cryptography)
• State estimation and optimal control (e.g., in robotics and dynamical systems)
• Numerics: determine whether a computation is numerically stable (e.g., in

large-scale data-analytics systems, optimization, machine learning)
• Find out whether a function is convex (e.g., in Optimization)
Relevance of Linear Algebra in other Courses

• Matrices, linear transformations, eigenvectors are required for Quantum Com-
puting in CO-438 (Complexity).
• Solving linear equation systems using Gaussian elimination is an important

prerequisite for CO-343 (Operations Research).
3 More details can be found on Jeremy Kun’s blog: http://tinyurl.com/olkbkct
61
6.1. Linear Equation Systems Chapter 6. Linear Algebra
6.1 Linear Equation Systems

6.1.1 Example
A company produces products N1 , . . . , Nn for which resources R1 , . . . , Rm are required.
To produce a unit of product Nj , aij units of resource Ri are needed, where i =
1, . . . , m and j = 1, . . . , n.
The objective is to find an optimal production plan, i.e., a plan how many units xj of
product Nj should be produced if a total of bi units of resource Ri are available and
(ideally) no resources are left over.
If we produce x1 , . . . , xn units of the corresponding products, we need a total of
ai1 x1 + · · · + ain xn (6.2)
many units of resource Ri . The desired optimal production plan (x1 , . . . , xn ) ∈ Rn ,

therefore, has to satisfy the following system of equations:
a11 x1 + · · · + a1n xn b1
.. = ... , (6.3)
.
am1 x1 + · · · + amn xn bm
where aij ∈ R and bi ∈ R. Equation (6.3) is the general form of a linear equation
system, and x1 , . . . , xn are the unknowns of this linear equation system. Every n-
tuple (x1 , . . . , xn ) ∈ Rn that satisfies (6.3) is a solution of the linear equation system.
The linear equation system
x1 + x2 + x3 = 3 (1)
x1 − x2 + 2x3 = 2 (2) (6.4)
2x1 + 3x3 = 1 (3)
has no solution: Adding the first two equations yields (1)+(2) = 2x1 +3x3 = 5, which
contradicts the third equation (3).
Let us have a look at the linear equation system
x1 + x2 + x3 = 3 (1)
x1 − x2 + 2x3 = 2 (2) . (6.5)
x2 + x3 = 2 (3)
From the first and third equation it follows that x1 = 1. From (1)+(2) we get 2+3x3 =
5, i.e., x3 = 1. From (3), we then get that x2 = 1. Therefore, (1, 1, 1) is the only
possible and unique solution (verify by plugging in).
As a third example, we consider
x1 + x2 + x3 = 3 (1)
x1 − x2 + 2x3 = 2 (2) . (6.6)
2x1 + 3x3 = 5 (3)
62
Chapter 6. Linear Algebra 6.2. Groups
Since (1)+(2)=(3), we can omit the third equation (redundancy). From (1) and
(2), we get 2x1 = 5 − 3x3 and 2x2 = 1 + x3 . We define x3 = a ∈ R as a free variable,
such that any triplet
5 3 1 1

− a, + a, a , a ∈ R (6.7)
2 2 2 2
is a solution to the linear equation system, i.e., we obtain a solution set that contains
infinitely many solutions.
In general, for a real-valued linear equation system we obtain either no, exactly one
or infinitely many solutions.
For a systematic approach to solving linear equation systems, we will introduce a
useful compact notation. We will write the linear equation system from (6.3) in the
following form:
            
 a11   a12   a1n   b1   a11 · · · a1n  x1   b1 
x1  ...  + x2  ...  + · · · + xn  ...  =  ...  ⇔  ... ..   ..  =  ..  .
        
.   .   .  (6.8)
        
am1 am2 amn bm am1 · · · amn xn bm
   
In order to work with these matrices, we need to have a close look at the underlying
algebraic structures and define computation rules.
6.2 Groups
Groups play an important role in computer science. Besides providing a fundamental
framework for operations on sets, they are heavily used in cryptography, coding
theory and graphics.
6.2.1 Definitions
Consider a set G and an operation ⊗ : G → G defined on G. For example, ⊗ could be
+, · defined on R, N, Z or ∪, ∩, \ defined on P (B), the power set of B.
Then (G, ⊗) is called a group if
• Closure of G under ⊗: ∀x, y ∈ G : x ⊗ y ∈ G
• Associativity: ∀x, y, z ∈ G : (x ⊗ y) ⊗ z = x ⊗ (y ⊗ z)
• Neutral element: ∃e ∈ G∀x ∈ G : x ⊗ e = x and e ⊗ x = x
• Inverse element: ∀x ∈ G∃y ∈ G : x ⊗ y = e and y ⊗ x = e. We often write x−1 to

denote the inverse element of x.4
If additionally ∀x, y ∈ G : x ⊗ y = y ⊗ x then (G, ⊗) is Abelian group (commutative).

4 The inverse element is defined with respect to the operation ⊗ and does not necessarily mean 1x .
63
6.3. Matrices Chapter 6. Linear Algebra
6.2.2 Examples
(Z, +) is a group, whereas (N0 , +)5 is not: Although (N0 , +) possesses a neutral ele-
ment (0), the inverse elements are missing.
(Z, ·) is not a group: Although (Z, ·) contains a neutral element (1), the inverse
elements for any z ∈ Z, z , ±1, are missing.
(R, ·) is not a group since 0 does not possess an inverse element. However, (R\{0}) is
Abelian.
(Rn , +), (Zn , +), n ∈ N are Abelian if + is defined componentwise, i.e.,
(x1 , · · · , xn ) + (y1 , · · · , yn ) = (x1 + y1 , · · · , xn + yn ). (6.9)
Then, e = (0, · · · , 0) is the neutral element and (x1 , · · · , xn )−1 := (−x1 , · · · , −xn ) is the
inverse element.
6.3 Matrices
Definition 1 (Matrix)
With m, n ∈ N a real-valued (m, n) matrix A is an m·n-tuple of elements aij , i = 1, . . . , m,
j = 1, . . . , n, which is ordered according to a rectangular scheme consisting of m rows
and n columns:
 a11 a12 · · · a1n 
 
 a
 21 a22 · · · a2n 

A =  ..
 .. ..  , aij ∈ R . (6.10)
 . . . 

am1 am2 · · · amn

(1, n)-matrices are called rows, (m, 1)-matrices are called columns. These special ma-
trices are also called row/column vectors.
Rm×n is the set of all real-valued (m, n)-matrices. A ∈ Rm×n can be equivalently
represented as A ∈ Rmn . Therefore, (Rm×n , +) is Abelian group (with componentwise
addition as defined in (6.9)).
6.3.1 Matrix Multiplication

For A ∈ Rm×n , B ∈ Rn×k (note the size of the matrices!) the elements cij of the
product C = AB ∈ Rm×k are defined as
n
X
cij = ail blj , i = 1, . . . , m, j = 1, . . . , k. (6.11)
l=1
This means, to compute element cij we multiply the elements of the ith row of A
with the jth column of B 6 and sum them up.7
5N
0 = N ∪ {0}
6 They are both of length k, such that we can compute ail blj for l = 1, . . . , n.
7 Later, we will call this the scalar product or dot product of the corresponding row and column.
64
Chapter 6. Linear Algebra 6.3. Matrices
Remark 1
Matrices can only be multiplied if their “neighboring” dimensions match. For instance,
an n × k-matrix A can be multiplied with a k × m-matrix B, but only from the left side:
A B = C (6.12)
|{z} |{z} |{z}
n×k k×m n×m
The product BA is not defined if m , n since the neighboring dimensions do not match.
Remark 2
Note that matrix multiplication is not defined as an element-wise operation on matrix
elements, i.e., cij , aij bij (even if the size of A, B was chosen appropriately).8
Example
0 2
" #  
1 2 3  
∈ R2×3 , B = 1 −1 ∈ R3×2 , we obtain

For A =

3 2 1
0 1
 
" # 0 2  " #

1 2 3  2 3
∈ R2×2 ,

AB = 1 −1 = (6.13)
3 2 1  2 5
0 1

0 2  #  6 4 2
 "
1 2 3
= −2 0 2 ∈ R3×3 .
 
BA = 1 −1 (6.14)

 3 2 1
0 1 3 2 1
  
From this example, we can already see that matrix multiplication is not commuta-
tive, i.e., AB , BA.
Definition 2 (Identity Matrix)

In Rn×n , we define the identity matrix as
1 0 · · · · · · 0
 
0 1 0 · · · 0
I n =  ... . . . . . . . . . ...  ∈ Rn×n .
 
(6.15)
 
 
0 · · · 0 1 0
 
0 ··· ··· 0 1
With this, A·I n = A = I n A for all A ∈ Rn×n . Therefore, the identity matrix is the neutral
element with respect to matrix multiplication “·” in (Rn×n , ·).9
8 This kind of element-wise multiplication appears often in computer science where we multiply
(multi-dimensional) arrays with each other.
9 If A ∈ Rm×n then I is only a right neutral element, succh that AI = A. The corresponding
n n
left-neutral element would be I m since I m A = A.
65
6.3. Matrices Chapter 6. Linear Algebra
Properties
• Associativity: ∀A ∈ Rm×n , B ∈ Rn×p , C ∈ Rp×q : (AB)C = A(BC)
• Distributivity: ∀A1 , A2 ∈ Rm×n , B ∈ Rn×p : (A1 + A2 )B = A1 B + A2 B
A(B + C) = AB + AC
• ∀A ∈ Rm×n : I m A = AI n = A. Note that I m , I n for m , n.
6.3.2 Inverse and Transpose

Definition 3 (Inverse)
For a square matrix10 A ∈ Rn×n a matrix B ∈ Rn×n with AB = I n = BA is called inverse
and denoted by A−1 .
Not every matrix A possesses an inverse A−1 . If this inverse does exist, A is called
regular/invertible, otherwise singular. We will discuss these properties much more
later on in the course.
Remark 3
The set of regular (invertible) matrices A ∈ Rn×n is a group with respect to matrix
multiplication as defined in (6.11) and is called general linear group GL(n, R).
Definition 4 (Transpose)
For A ∈ Rm×n the matrix B ∈ Rn×m with bij = aji is called the transpose of A. We write
B = A> .
For a square matrix A> is the matrix we obtain when we “mirror” A on its main
diagonal.11 In general, A> can be obtained by writing the columns of A as the rows
of A> .
Remark 4
• AA−1 = I = A−1 A
• (AB)−1 = B −1 A−1
• (A> )> = A
• (A + B)> = A> + B >
• (AB)> = B > A>
• If A is invertible, (A−1 )> = (A> )−1
1 1
• Note: (A + B)−1 , A−1 + B −1 . Example: in the scalar case 2+4 = 6 , 12 + 14 .
A is symmetric if A = A> . Note that this can only hold for (n, n)-matrices (quadratic
matrices). The sum of symmetric matrices is symmetric, but this does not hold for
the product in general (although it is always defined). A counterexample is
" #" # " #
1 0 1 1 1 1
= . (6.16)
0 0 1 1 0 0
10 Thenumber columns equals the number of rows.
11 Themain diagonal (sometimes principal diagonal, primary diagonal, leading diagonal, or major
diagonal) of a matrix A is the collection of entries Aij where i = j.
66
Chapter 6. Linear Algebra 6.4. Gaussian Elimination
6.3.3 Multiplication by a Scalar

Let A ∈ Rm×n and λ ∈ R. Then λA = K, Kij = λ aij . Practically, λ scales each element
of A. For λ, ψ ∈ R it holds:
• Distributivity:
(λ + ψ)C = λC + ψC, C ∈ Rm×n
λ(B + C) = λB + λC, B, C ∈ Rm×n
• Associativity:
(λψ)C = λ(ψC), C ∈ Rm×n
λ(BC) = (λB)C = B(λC), B ∈ Rm×n , C ∈ Rn×k .
Note that this allows us to move scalar values around.
• (λC)> = C > λ> = C > λ = λC > since λ = λ> for all λ ∈ R.
6.3.4 Compact Representations of Linear Equation Systems

If we consider a linear equation system
2x1 + 3x2 + 5x3 = 1

4x1 − 2x2 − 7x3 = 8
9x1 + 5x2 − 3x3 = 2
and use the rules for matrix multiplication, we can write this equation system in a
more compact form as
2 3 5  x1  1

    
(6.17)
4 −2 −7 x  8
   2  =   .
9 5 −3 x3 2
    
Note that x1 scales the first column, x2 the second one, and x3 the third one.
Generally, linear equation systems can be compactly represented in their matrix form
as Ax = b, see (6.3), and the product Ax is a (linear) combination of the columns of
A.12
6.4 Solving Linear Equation Systems via Gaussian Elim-

ination
In (6.3), we have introduced the general form of an equation system, i.e.,
a11 x1 + · · · + a1n xn = b1
.. .. , (6.18)
. .
am1 x1 + · · · + amn xn = bm
12 We will discuss linear combinations in Section 6.5.
67
6.4. Gaussian Elimination Chapter 6. Linear Algebra
where aij ∈ R and bi ∈ R are known constants and xj are unknowns, i = 1, . . . , m,

j = 1, . . . , n. Thus far, we have introduced matrices as a compact way of formulating
linear equation systems, i.e., such that we can write Ax = b, see (6.8). Moreover, we
defined basic matrix operations, such as addition and multiplication of matrices. In
the following, we will introduce a constructive and systematic way of solving h linear
i
equation systems. Before doing this, we introduce the augmented matrix A | b of
the linear equation system Ax = b. This augmented matrix will turn out to be useful
when solving linear equation systems.
6.4.1 Example: Solving a Simple Linear Equation System

Now we are turning towards solving linear equation systems. Before doing this
in a systematic way using Gaussian elimination, let us have a look at an example.
Consider the following linear equation system:
# x1  " #

 
"
1 0 8 −4 x2  42
  = . (6.19)
0 1 2 12 x3  8
 
x4
This equation system is in a particularly easy form, where the first two columns
consist of a 1 and a 0.13 Remember that we want to find scalars x1 , . . . , x4 , such that
P 4
i=1 xi ci = b, where we define ci to be the ith column of the matrix and b the right-
hand-side of (6.19). A solution to the problem in (6.19) can be found immediately
by taking 42 times the first column and 8 times the second column, i.e.,
" # " # " #
42 1 0
b= = 42 +8 . (6.20)
8 0 1
Therefore, one solution vector is [42, 8, 0, 0]> . This solution is called a particular
solution or special solution. However, this is not the only solution of this linear
equation system. To capture all the other solutions, we need to be creative of gen-
erating 0 in a non-trivial way using the columns of the matrix: Adding a couple of
0s to our special solution does not change the special solution. To do so, we express
the third column using the first two columns (which are of this very simple form):
" # " # " #
8 1 0
=8 +2 , (6.21)
2 0 1
such that 0 = 8c1 + 2c2 − 1c3 + 0c4 . In fact, any scaling of this solution produces the
0 vector:
 8 
 
 2 
λ1   = 0 , λ1 ∈ R. (6.22)
−1
 
0
13 Later, we will say that this matrix is in reduced row echelon form.
68
Following the same line of reasoning, we express the fourth column of the matrix
in (6.19) using the first two columns and generate another set of non-trivial versions
of 0 as
−4
 
 12 
λ2   = 0 , λ2 ∈ R. (6.23)
 
 0 
 
−1
Putting everything together, we obtain all solutions of the linear equation system
in (6.19), which is called the general solution, as
42  8  −4
     
 8   2   12 
  + λ1   + λ2   . (6.24)
     
 0  −1  0 
     
0 0 −1
Remark 5
• The general approach we followed consisted of the following three steps:
1. Find a particular solution to Ax = b
2. Find all solutions to Ax = 0
3. Combine the solutions from 1. and 2. to the general solution.
• Neither the general nor the particular solution is unique.
The linear equation system in the example above was easy to solve because the
matrix in (6.19) has this particularly convenient form, which allowed us to find
the particular and the general solution by inspection. However, general equation
systems are not of this simple form. Fortunately, there exists a constructive way of
transforming any linear equation system into this particularly simple form: Gaussian
elimination.
The rest of this section will introduce Gaussian elimination, which will allow us to
solve all kinds of linear equation systems by first bringing them into a simple form
and then applying the three steps to the simple form that we just discussed in the
context of the example in (6.19), see Remark 5.
6.4.2 Elementary Transformations

Key to solving linear equation systems are elementary transformations that keep
the solution set the same14 , but that transform the equation system into a simpler
form:
• Exchange of two equations
• Multiplication of an equation with a constant λ ∈ R\{0}
• Addition of an equation to another equation
14 Therefore, the original and the modified equation system are equivalent.
69
Example
−2x1 + 4x2 − 2x3 − x4 + 4x5 = −3

4x1 − 8x2 + 3x3 − 3x4 + x5 = 2
, a∈R (6.25)
x1 − 2x2 + x3 − x4 + x5 = 0
x1 − 2x2 − 3x4 + 4x5 = a
Swapping rows 1 and 3 leads to
x1 − 2x2 + x3 − x4 + x5 = 0
4x1 − 8x2 + 3x3 − 3x4 + x5 = 2 | − 4R1
(6.26)
−2x1 + 4x2 − 2x3 − x4 + 4x5 = −3 | + 2R1
x1 − 2x2 − 3x4 + 4x5 = a | − R1
When we now apply the indicated transformations (e.g., subtract Row 1 4 times
from Row 2), we obtain
x1 − 2x2 + x3 − x4 + x5 = 0
− x3 + x4 − 3x5 = 2
(6.27)
− 3x4 + 6x5 = −3
− x3 − 2x4 + 3x5 = a | − R2
then
x1 − 2x2 + x3 − x4 + x5 = 0
− x3 + x4 − 3x5 = 2
(6.28)
− 3x4 + 6x5 = −3
− 3x4 + 6x5 = a−2 | − R3
and finally
x1 − 2x2 + x3 − x4 + x5 = 0
− x3 + x4 − 3x5 = 2 | · (−1)
(6.29)
− 3x4 + 6x5 = −3 | · (− 13 )
0 = a+1
If we now multiply the second equation with (−1) and the third equation with − 31 ,
we obtain the row echelon form
x1 − 2x2 + x3 − x4 + x5 = 0
x3 − x4 + 3x5 = −2
(6.30)
x4 − 2x5 = 1
0 = a+1
Only for a = −1, this equation system can be solved. A particular solution is given
by
   
x1   2 
x2   0 
   
x3  = −1 (6.31)
   
   
x4   1 
   
x5 0
70
and the general solution, which captures the set of all possible solutions, is given
as
       
x1   2  2  2 
x2   0  1  0 
       
x3  = −1 + λ1 0 + λ2 −1 , λ1 , λ2 ∈ R (6.32)
       
       
x4   1  0  2 
       
x5 0 0 1
Remark 6 (Pivots and Staircase Structure)

The leading coefficient (pivot) of a row (the first nonzero number from the left) is
always strictly to the right of the leading coefficient of the row above it. This ensures
that an equation system in row echelon form always has a “staircase” structure.
Remark 7 (Obtaining a Particular Solution)

The row echelon form makes our lives easier when we need to determine a particular
solution. To do this, we express
P the right-hand side of the equation system using the
pivot columns, such that b = Pi=1 λi pi , where pi , i = 1, . . . , P are the pivot columns.
The λi are determined easiest if we start with the most-right pivot column and work
our way to the left.
In the above example, we would try to find λ1 , λ2 , λ3 such that
1 1 −1  0 

       
0 1 −1 −2
λ1   + λ2   + λ3   =   (6.33)
       
0 0  1   1 
       
0 0 0 0
From here, we find relatively directly that λ3 = 1, λ2 = −1, λ1 = 2. When we put every-
thing together, we must not forget the non-pivot columns for which we set the coefficients
implicitly to 0. Therefore, we get the particular solution
 
 2 
 0 
 
x = −1 . (6.34)
 
 1 
 
0
 
Example 2
In the following, we will go through solving a linear equation system in matrix form.
Consider the problem of finding x = [x1 , x2 , x3 ]> , such that Ax = b, where
1 2 3 4

   
A = 4 5 6 , b = 6 (6.35)
   
7 8 9 8
   
First, we write down the augmented matrix [A | b], which is given by
71
1 2 3 4
 
 

 4 5 6 6  ,

7 8 9 8
 
which we now transform into row echelon form using the elementary row opera-
tions:
 1 2 3 4   1 2 3 4 
   
 4
 5 6 6  −4R1 ;  0 −3 −6 −10  ·(− 13 )
7 8 9 8 −7R1 0 −6 −12 −20 −2R2
   
 1 2 3 4 
 
;  0 1 2 10 
 
3  
0 0 0 0

From the row echelon form, we see that x3 is a free variable. To find a particular
solution, we can set x3 to any real number. For convenience, we choose x3 = 0, but
any other number would have worked. With x3 = 0, we obtain a particular solution
   8
x1  − 3 
x   10 
 2  =  3  . (6.36)
x3 0
   
To find the general solution, we combine the particular solution with the solution of
the homogeneous equation system Ax = 0. There are two ways of getting there: the
matrix view and the equation system view. Looking at it from the matrix perspective,
we need to express the third column of the row-echelon form in terms of the first
two columns. This can be done by seeing that
3 1 2 1 2 3
           
(6.37)
2
  = −   + 2   ⇔ −   + 2 1 − 2 = 0 .
0 1 0    
0 0 0 0 0 0
           
We now take the coefficients [−1, 2, −1]> of these colums that are a non-trivial repre-
sentation of 0 as the solution (and any multiple of it) to the homogeneous equation
system Ax = 0.
An alternative and equivalent way is to remember that we wanted to solve a linear
equation system, we find the solution to the homogeneous equation system by ex-
pressing x3 in terms of x1 , x2 . From the row echelon form, we see that x2 + 2x3 =
0 ⇒ x3 = − 21 x2 . With this, we now look at the first set of equations and obtain
x1 + 2x2 + 3x3 = 0 ⇒ x1 − x3 = 0 ⇒ x3 = x1 .
Independent of whether we use the matrix or equation system view, we arrive at the
general solution
 8
− 3   1 
 
 10 
 3  + λ −2 , λ ∈ R. (6.38)
 
0 1
   
Remark 8 (Reduced Row Echelon Form)

An equation system is in reduced row echelon form15 if
15 also: row reduced echelon form or row canonical form
72
• It is in row echelon form
• Every pivot must be 1 and is the only non-zero entry in its column.
The reduced row echelon form will play an important role in later sections because it
allows us to determine the general solution of a linear equation system in a straight-
forward way.
Example: Reduced Row Echelon Form Verify that the following matrix is in re-
duced row echelon form:
1 3 0 0 3 
 
A = 0 0 1 0 9  (6.39)
 
0 0 0 1 −4
 
The pivots are colored red.

To read out the solutions of Ax = 0, we are mainly interested in the non-pivot
columns, which we will need to express as a sum of the pivot columns. The re-
duced row echelon form makes this relatively straightforward, and we express the
non-pivot columns in terms of sums and multiples of the pivot columns that are on
their left: The second column is three times the first column (we can ignore the
pivot columns on the right of the second column). Therefore, to obtain 0, we need
to subtract the second column from three times the first column. Now, we look at
the fifth column, which is our second non-pivot column. The fifth column is given
by 3 times the first pivot column, 9 times the second pivot column, and -4 times the
third pivot column. We need to keep track of the indices of the pivot columns and
translate this into 3 times the first column, 9 times the third pivot column (which is
our second pivot column) and -4 times the fourth column (which is the third pivot
column). Then we need to subtract the fifth column to obtain 0—in the end, we are
still solving a homogeneous equation system.
To summarize, all solutions of Ax = 0, x ∈ R5 are given by
   
 3   3 
−1  0 
   
λ1  0  + λ2  9  , λ1 , λ2 ∈ R. (6.40)
   
 0  −4
   
0 −1
   
6.4.3 The Minus-1 Trick for Solving Homogeneous Equation Sys-

tems
In the following, we introduce a practical trick for reading out the solutions x of a
homogeneous linear equation system Ax = 0, where A ∈ Rk×n , x ∈ Rn .
73
To start, we assume that A is in reduced row echelon form without any rows that
just contain zeros (e.g., after applying Gaussian elimination), i.e.,
 
 0 · · · 0 1 ∗ · · · ∗ 0 ∗ · · · ∗ 0 ∗ · · · ∗ 
 .. .. .. .. .. 
 . . 0 0 · · · 0 1 ∗ · · · ∗ . . . 
 . . . . . . .. 
A =  .. .. .. .. .. 0 ∗ . . . ∗ 0 .. .  (6.41)
 . . . . . . . .
.. .. ∗ . . . ∗ 0 ..

 .. .. .. .. .. 

 
0 ··· 0 0 0 ··· 0 0 ∗ ··· ∗ 1 ∗ ··· ∗
Note that the columns j1 , . . . , jk with the pivots (marked red) are the standard unit
vectors e1 , . . . , ek ∈ Rk .
We now extend this matrix to an n × n-matrix Ã by adding n − k rows of the form
h i
0 · · · 0 −1 0 · · · 0 , (6.42)
such that the diagonal of the augmented matrix Ã contains only 1 or −1. Then,
the columns of Ã, which contain the −1 as pivots are solutions of the homogeneous
equation system Ax = 0.16 To be more precise, these columns form a basis (Sec-
tion 6.7) of the solution space of Ax = 0, which we will later call the kernel or null
space (Section 6.9.1).
Example
Let us revisit the matrix in (6.39), which is already in reduced row echelon form:
1 3 0 0 3 
 
A = 0 0 1 0 9  . (6.43)
 
0 0 0 1 −4
 
We now augment this matrix to a 5 × 5 matrix by adding rows of the form (6.42) at
the places where the pivots on the diagonal are missing and obtain
 
1 3 0 0 3 
0 −1 0 0 0 
 
Ã = 0 0 1 0 9  (6.44)
0 0 0 1 −4
 
0 0 0 0 −1
From this form, we can immediately read out the solutions of Ax = 0 by taking the
columns of Ã, which contain −1 on the diagonal:
   
 3   3 
−1  0 
   
λ1  0  + λ2  9  , λ1 , λ2 ∈ R, (6.45)
   
 0  −4
   
0 −1
   
which is identical to the solution in (6.40) that we obtained by “insight”.

16 The proof of this trick is out of the scope of this course.
74
6.4.4 Applications of Gaussian Elimination in Linear Algebra

Gaussian elimination can also be used to find the rank of a matrix (Chapter 6.7), to
calculate the determinant of a matrix (Chapter 6.10), the null space, and the inverse
of an invertible square matrix. Because of its relevance to central concepts in Linear
Algebra, Gaussian elimination is the most important algorithm we will cover.
Gaussian elimination is an important prerequisite for CO343 Operations Research.
The initial part of the course discusses in depth systems of linear equations with
more columns than rows, and the related notions of basic solutions and non-basic
solutions.
Calculating the Inverse

To compute the inverse A−1 of A ∈ Rn×n , we need to satisfy AA−1 = I n . We can write
this down as a set of simultaneous linear equations AX = I n , where we solve for
X = [x1 | · · · |xn ]. We use the augmented matrix notation for a compact representation
of this set of linear equation systems and obtain
h i h i
A I ; I A−1 .
This means that if we bring the augmented equation system into reduced row eche-
lon form, we can read off the inverse on the right-hand side of the equation system.
"
#
1 2
Example 1 For A = , we determine its inverse by solving the following linear
3 4
equation system:
" #
1 2 1 0
3 4 0 1
We bring this system now into reduced row echelon form

" # " #
1 2 1 0 1 2 1 0 +R2
;
3 4 0 1 −3R1 0 −2 −3 1 ·(− 12 )
" #
1 0 −2 1
; 3 .
0 1 2 − 12
The right-hand side of this augmented equation system contains the inverse
" #
−1 −2 1
A = 3 . (6.46)
2 − 21
Example 2 To determine the inverse of
1 0 2 0
 
1 1 0 0
A =  (6.47)

1 2 0 1

 
1 1 1 1
75
6.5. Vector Spaces Chapter 6. Linear Algebra
we write down the augmented matrix
 1 0 2 0 1 0 0 0
 

 1 1 0 0 0 1 0 0 
 
 1 2 0 1 0 0 1 0
 

 
1 1 1 1 0 0 0 1
and transform it into reduced row echelon form
 1 0 0 0 −1 2 −2 2 
 
 0 1 0 0 1 −1 2 −2 
,
 
 0 0 1 0 1 −1 1 −1 

 
0 0 0 1 −1 0 −1 2
such that the desired inverse is given as its right-hand side:
−1 2 −2 2 
 
 1 −1 2 −2
A−1 =  . (6.48)
 
 1 −1 1 −1
 
−1 0 −1 2
Remark 9
You may have encountered a way of computing the inverse of a matrix using co-factors
and/or cross-products. This approach only works in three dimensions and is not used
in practice.
6.5 Vector Spaces

When we discussed group theory, we were looking at sets G and inner operations on
G, i.e., mappings G × G → G. In the following, we will consider sets that in addition
to an inner operation + also contain an outer operation ·, the multiplication by a
scalar λ ∈ R.
Definition 5 (Vector space)
A real-valued vector space (also called an R-vector space) is a set V with two opera-
tions
+ : V ×V → V (6.49)
· : R×V → V (6.50)
where
1. (V , +) is an Abelian group
2. Distributivity:
(a) λ · (x + y) = λ · x + λ · y ∀λ ∈ R, x, y ∈ V
(b) (λ + ψ) · x = λ · x + ψ · x ∀λ, ψ ∈ R, x ∈ V
76
Chapter 6. Linear Algebra 6.5. Vector Spaces
3. Associativity (outer operation): λ · (ψ · x) = (λψ) · x ∀λ, ψ ∈ R, x ∈ V
4. Neutral element with respect to the outer operation: 1 · x = x, ∀x ∈ V

The elements x ∈ V are called vectors. The neutral element of (V , +) is the zero vector
0 = [0, . . . , 0]> , and the inner operation + is called vector addition. The elements λ ∈ R
are called scalars and the outer operation · is a multiplication by scalars.17
Remark 10
When we started the course, we defined vectors as special objects that can be added
together and multiplied by scalars to yield another element of the same kind (see p. 60).
Examples were geometric vectors, polynomials and Rn . Definition 5 gives now the
corresponding formal definition and applies to all kinds of vectors. We will continue
focusing on vectors as elements of Rn because it is the most general formulation, and
most algorithms are formulated in Rn .
Remark 11
Note that a “vector multiplication” ab, a, b ∈ Rn , is not defined. Theoretically, we
could define it in two ways: (a) We could define an element-wise multiplication, such
that c = a · b with cj = aj bj . This “array multiplication” is common to many program-
ming languages but makes mathematically only limited sense; (b) By treating vectors as
n × 1 matrices (which we usually do), we can use the matrix multiplication as defined
in (6.11). However, then the dimensions of the vectors do not match. Only the fol-
lowing multiplications for vectors are defined: ab> (outer product), a> b (inner/scalar
product).
6.5.1 Examples
• V = Rn , n ∈ N is a vector space with operations defined as follows:
– Addition: x+y = (x1 , . . . , xn )+(y1 , . . . , yn ) = (x1 +y1 , . . . , xn +yn ) for all x, y ∈ Rn

– Multiplication by scalars: λx = λ(x1 , . . . , xn ) = (λx1 , . . . , λxn ) for all λ ∈
R, x ∈ Rn
• V = Rm×n , m, n ∈ N is a vector space with

 
 a11 + b11 · · · a1n + b1n 
– Addition: A + B = 
 .. .. 
 is defined elementwise for
. . 

am1 + bm1 · · · amn + bmn

all A, B ∈ V
 
 λa11 · · · λa1n 
– Multiplication by scalars: λA =  ... ..  as defined in Section 6.3.

 . 
λam1 · · · λamn

Remember that Rm×n is equivalent to Rmn .
• V = C, where the addition is defined in (5.5).

17 Note: A scalar product is something different, and we will get to this in Section 6.13.
77
6.5. Vector Spaces Chapter 6. Linear Algebra
Remark 12 (Notation)
The three vector spaces Rn , Rn×1 , R1×n are only different with respect to the way of
writing. In the following, we will not make a distinction between Rn and Rn×1 , which
allows us to write n-tuples as column vectors
 
x1 
x =  ...  .
 
(6.51)
 
xn
This will simplify the notation regarding vector space operations. However, we will
distinguish between Rn×1 and R1×n (the row vectors) to avoid confusion with matrix
multiplication. By default we write x to denote a column vector, and a row vector is
denoted by x> , the transpose of x.
6.5.2 Generating Set and Vector Subspaces

Definition 6 (Linear Combination)
Consider a vector space V and a finite number of vectors x1 , . . . , xk ∈ V . Then, every
vector v ∈ V of the form
k
X
v = λ1 x 1 + · · · + λk x k = λi x i ∈ V (6.52)
i=1
with λ1 , . . . , λk ∈ R is a linear combination of the vectors x1 , . . . , xk .
Definition 7 (Generating Set/Span)

Consider an R-vector space V and A = {x1 , . . . , xk } ⊂ V . If every vector v ∈ V can be
expressed as a linear combination of x1 , . . . , xk , A is called a generating set or span,
which spans the vector space V . In this case, we write V = [A] or V = [x1 , . . . , xk ].
Definition 8 (Vector Subspace)

Let V be an R-vector space and U ⊂ V , U , ∅. U is called vector subspace of V
(or linear subspace) if U is a vector space with the vector space operations + and ·
restricted to U × U and R × U .
Remark 13
If U ⊂ V and V is a vector space, then U naturally inherits many properties directly
from V because they are true for all x ∈ V , and in particular for all x ∈ U ⊂ V .
This includes the Abelian group properties, the distributivity, the associativity and the
neutral element. What we still do need to show is
1. U , ∅, in particular: 0 ∈ U
2. Closure of U :
(a) With respect to the outer operation: ∀λ ∈ R∀x ∈ U : λx ∈ U

(b) With respect to the inner operation: ∀x, y ∈ U : x + y ∈ U .
78
Chapter 6. Linear Algebra 6.6. Linear (In)Dependence
3 miles North 5 miles Northeast
4 miles East
Figure 6.2: Linear dependence of three vectors in a two-dimensional space (plane).
Examples
• For every vector space V the trivial subspaces are V itself and {0}.
• The solution set of a homogeneous linear equation system Ax = 0 with n un-

knowns x = [x1 , . . . , xn ]> is a subspace of Rn .
• However, the solution of an inhomogeneous equation system Ax = b, b , 0 is

not a subspace of Rn .
• The intersection of arbitrarily many subspaces is a subspace itself.
• The intersection of all subspaces Ui ⊂ V is called linear hull of V .
Remark 14
Every subspace U ⊂ Rn is the solution space of a homogeneous linear equation system
Ax = 0.
6.6 Linear (In)Dependence

In Section 6.5, we learned about linear combinations of vectors, see (6.52). The 0
vector
P can always be written as the linear combination of k vectors x1 , . . . , xk because
0 = ki=1 0xi is always true. In the following, we are interested in non-trivial linear
combinations of a set of vectors to represent 0.
Definition 9 (Linear (In)dependence)

Let us consider a vector space V with k ∈ N and x1 , . . . , xk ∈ V . If there is a non-
trivial linear combination, such that 0 = ki=1 λi xi with at least one λi , 0, the vectors
P
x1 , . . . , xk are linearly dependent. If only the trivial solution exists, i.e., λ1 = . . . = λk =
0 the vectors x1 , . . . , xk are linearly independent.
Intuitively, a set of linearly dependent vectors contains some redundancy, whereas

linearly independent vectors are all essential. Throughout this chapter, we will for-
malize this intuition more.
79
6.6. Linear (In)Dependence Chapter 6. Linear Algebra
Remark 15 (From Wikipedia (2015))

A geographic example may help to clarify the concept of linear independence. A person
describing the location of a certain place might say, “It is 3 miles North and 4 miles East
of here.” This is sufficient information to describe the location, because the geographic
coordinate system may be considered as a 2-dimensional vector space (ignoring altitude
and the curvature of the Earth’s surface). The person might add, “The place is 5 miles
Northeast of here.” Although this last statement is true, it is not necessary to find this
place (see Fig. 6.2 for an illustration).
In this example, the “3 miles North” vector and the “4 miles East” vector are linearly
independent. That is to say, the north vector cannot be described in terms of the east
vector, and vice versa. The third “5 miles Northeast” vector is a linear combination of
the other two vectors, and it makes the set of vectors linearly dependent, that is, one of
the three vectors is unnecessary.
Remark 16
The following properties are useful to find out whether vectors are linearly independent.
• k vectors are either linearly dependent or linearly independent. There is no third
option.
• If at least one of the vectors x1 , . . . , xk is 0 then they are linearly dependent. The
same holds if two vectors are identical.
• The vectors x1 , . . . , xk , k ≥ 2, are linearly dependent if and only if (at least) one of
them is a linear combination of the others.
• In a vector space V m linear combinations of k vectors x1 , . . . , xk are linearly
dependent if m > k.
• Consider an R-vector space V with k vectors b1 , . . . , bk and m linear combinations
k
X
x1 = λi1 bi , (6.53)
i=1
..
. (6.54)
k
X
xm = λim bi . (6.55)
i=1
We want to test whether x1 , . . . , xm are linearly

Pmindependent. For this purpose, we
follow the general approach of testing when i=1 ψi xi = 0 and obtain
 k   
X m m
X X  Xm X
k k X
X m 

0= ψj x j = ψj  λij bi  = ψj λij bi = ψ λ  bi . (6.56)
 
 j ij  
j=1 j=1 i=1 j=1 i=1 i=1 j=1
Therefore, x1 , . . . , xm are linearly independent if and only if the column vectors

   
λ11  λ1m 
x̂1 =  ...  ∈ Rk , . . . , x̂m =  ...  ∈ Rk
   
(6.57)
   
λk1 λkm
80
Chapter 6. Linear Algebra 6.6. Linear (In)Dependence
are linearly independent.

Proof 3
Since b1 , . . . , bk are linearly independent it follows that for all j = 1, . . . , m we
get ki=1 λij bi = 0 with λij = 0, i = 1, . . . , k. Therefore, m
P P
j=1 ψj λij = 0 for i =
Pk Pm Pm
1, . . . , k. This implies i=1 ( j=1 ψj λij )bi = 0. Hence, j=1 ψj xj = 0 is equivalent
to m
P
j=1 ψj x̂j .
• A practical way of checking whether the column vectors are linearly independent is
to use Gaussian elimination: Write all vectors as columns of a matrix A. Gaussian
elimination yields a matrix in (reduced) row echelon form. The pivot columns
indicate the vectors, which are linearly independent of the previous18 vectors (note
that there is an ordering of vectors when the matrix is built). If all columns are
pivot columns, the column vectors are linearly independent.
• The non-pivot columns can be expressed as linear combinations of the columns

that were before (left of) them. If the matrix is in reduced row echelon form, we
can immediately see how the columns relate to each other. For instance, in
" #
1 3 0
(6.58)
0 0 1
first and third column are pivot columns. The second column is a non-pivot col-
umn because it is 3 times the first column. If there is at least one non-pivot
column, the columns are linearly dependent.
6.6.1 Examples
• Consider R4 with
 2  1 −2
     
−3 0  1 
x1 =   , x2 =   , x3 =   . (6.59)
     
 1  1 −1
     
4 2 1
To check whether they are linearly dependent, we follow the general approach
and solve
 2  1 −2
     
−3 0  1 
λ1 x1 + λ2 x2 + λ3 x3 = λ1   + λ2   + λ3   = 0 (6.60)
     
 1  1 −1
     
4 2 1
for λ1 , . . . , λ3 . We write the vectors xi , i = 1, 2, 3, as the columns of a matrix

and apply Gaussian elimination.
18 the vectors on the left
81
6.6. Linear (In)Dependence Chapter 6. Linear Algebra
   
 1 1 −1 0   1 1 −1 0 
   
 2 1 −2 0  −2R1 0 −1 0 0
 +3R ;
 
   
 −3
 0 1 0  1

 0 3 −2 0 
 +3R2
4 2 1 0 −4R1 0 −2 5 0 −2R2
   
   
 1 1 −1 0   1 1 −1 0 
   
 0 −1 0 0  ·(−1) 0 1 0 0
; ;
 
   
 0 0 −2 0   0 0 1 0 
   
 5
0 0 5 0 + 2 R3 0 0 0 0
  
Here, every column of the matrix is a pivot column19 , i.e., every column is
linearly independent of the columns on its left. Therefore, there is no non-
trivial solution, and we require λ1 = 0, λ2 = 0, λ3 = 0. Hence, the vectors
x1 , x2 , x3 are linearly independent.
• Consider a set of linearly independent vectors b1 , b2 , b3 , b4 ∈ Rn and
x1 = b1 − 2b2 + b3 − b4
x2 = −4b1 − 2b2 + 4b4
(6.61)
x3 = 2b1 + 3b2 − b3 − 3b4
x4 = 17b1 − 10b2 + 11b3 + b4
Are the vectors x1 , . . . , x4 ∈ Rn linearly independent? To answer this question,
we investigate whether the column vectors
       
 1  −4  2   17 
 
         
−2 −2  3  −10

 
 
 
 
 
 
 
 
 
, , , (6.62)

1 0 −1 11

        

 
 
 
 
 
 
 
 
 

−1  4  −3  1 
 
are linearly independent. The reduced row echelon form of the corresponding
linear equation system with coefficient matrix
 1 −4 2 17 
 
−2 −2 3 −10
A =  (6.63)
 
 1 0 −1 11 

 
−1 4 −3 1
is given as
1 0 0 −7 
 
0 1 0 −15
. (6.64)

0 0 1 −18

 
0 0 0 0
From the reduced row echelon form, we see that the corresponding linear
equation system is non-trivially solvable: The last column is not a pivot column,
and x4 = −7x1 − 15x2 − 18x3 . Therefore, x1 , . . . , x4 are linearly dependent as x4
lies in the span of x1 , . . . , x3 .
19 Note that the matrix is not in reduced row echelon form; it also does not need to be.
82
Chapter 6. Linear Algebra 6.7. Basis and Dimension
6.7 Basis and Dimension

In a vector space V , we are particularly interested in the set of linearly independent
vectors A that possesses the property that any vector v ∈ V can be obtained by a
linear combination of vectors in A.
Definition 10 (Basis)
Consider a real vector space V and A ⊂ V
• A generating set A of V is called minimal if there exists no smaller set Ã ⊂ A ⊂ V ,

which spans V .
• Every linearly independent generating set of V is minimal and is called basis of

V.
Let V be a real vector space and B ⊂ V , B , ∅. Then, the following statements are
equivalent:
• B is basis of V
• B is a minimal generating set
• B is a maximal linearly independent subset of V .
• Every vector x ∈ V is a linear combination of vectors from B, and every linear

combination is unique, i.e., with
k
X k
X
x= λi bi = ψi bi (6.65)
i=1 i=1
and λi , ψi ∈ R, bi ∈ B it follows that λi = ψi , i = 1, . . . , k.
6.7.1 Examples
• In R3 , the canonical/standard basis is
     

 1 0 0  
     

B= 0 , 1 , 0 . (6.66)

 
 
 
 
 
 
 
0 0 1
 
• Different bases in R3 are

       
1 1 1  0.53  1.83 −2.25
     
   
     
  
 
B1 =  0 , 1 , 1 , B2 =  0.86 , 0.31 , −1.30 (6.67)
  
 
 
 
 
 
  
 
 
 
 
 
 

 0 0 1 
         −0.43 0.34
     
3.57 
 
83
6.7. Basis and Dimension Chapter 6. Linear Algebra
• The set
     


 1  2   1  

2 −1  1 

 
A=   ,   ,   (6.68)



 3  0   0  

4  2  −4
 
is linearly independent, but not a generating set (and no basis): For instance,
the vector [1, 0, 0, 0]> cannot be obtained by a linear combination of elements
in A.
Remark 17
• Every vector space V possesses a basis B.
• The examples above show that there can be many bases of a vector space V , i.e.,
there is no unique basis. However, all bases possess the same number of elements,
the basis vectors.
• We only consider finite-dimensional vector spaces V . In this case, the dimension

of V is the number of basis vectors, and we write dim(V ).
• If U ⊂ V is a subspace of V then dim(U ) ≤ dim(V ) and dim(U ) = dim(V ) if and

only if U = V .
6.7.2 Example: Determining a Basis

• For a vector subspace U ⊂ R5 , spanned by the vectors
       
 1   2   3  −1
 2  −1 −4  8 
       
x1 = −1 , x2 =  1  , x3 =  3  , x4 = −5 ∈ R5 , (6.69)
       
     
−1  2   5  −6
       
−1 −2 −3 1
       
we are interested in finding out which vectors x1 , . . . , x4 are a basis for U . For
this, we need to check whether x1 , . . . , x4 are linearly independent. Therefore,
we need to solve
4
X
λi x i = 0 , (6.70)
i=1
which leads to a homogeneous equation system with the corresponding matrix

 
 1 2 3 −1
i  2 −1 −4 8 
 
h
x1 |x2 |x3 |x4 = −1 1 3 −5 . (6.71)
−1 2 5 −6
 
−1 −2 −3 1
 
84
Chapter 6. Linear Algebra 6.7. Basis and Dimension
With the basic transformation of linear equation systems, we obtain

   
 1 2 3 −1   1 2 3 −1 
·(− 15 )
   
 2 −1 −4 8  −2R1  0 −5 −10 10 
   
 −1
 1 3 −5  +R
 1 ;  0 3 6 −6 
 · 13 | − R2
 −1 2 5 −6  +R1 0 4 8 −7 −4R2
   
 
   
−1 −2 −3 1 +R1 0 0 0 0
   
 1 2 3 −1   1 2 3 0  −2R2
   
 0 1 2 −2  +R4  0 1 2 0 
   
;  0
 0 0 0  +2R
 4 ; 
 0 0 0 1 

 0

0 0 1  swap with R 
0 0 0 0

  3 



0 0 0 0 0 0 0 0

 
 1 0 −1 0 
 
 0 1 2 0 
 
;  0
 0 0 1 .

 0 0 0 0
 

 
0 0 0 0
From this reduced-row echelon form we see that x1 , x2 , x4 are linearly inde-
pendent (because the linear equation system λ1 x1 + λ2 x2 + λ4 x4 = 0 can only
be solved with λ1 = λ2 = λ4 = 0). Therefore, {x1 , x2 , x4 } is a basis of U .
• Let us now consider a slightly different problem: Instead of finding out which
vectors x1 , . . . , x4 of the span of U form a basis, we are interested in finding
a “simple” basis for U . Here, “simple” means that we are interested in basis
vectors with many coordinates equal to 0.
To solve this problem we replace the vectors x1 , . . . , x4 with suitable linear com-
binations. In practice, we write x1 , . . . , x4 as row vectors in a matrix and per-
form Gaussian elimination:
   
 1 2 −1 −1 −1   1 2 −1 −1 −1 
   
 2 −1 1 2 −2   0 −5 3 4 0 

 3 −4  ;  
3 5 −3   0 −10 6 8 0 

 
−1 8 −6 −6 1 0 10 −7 −7 0
   
   
 1 2 −1 −1 −1   1 2 −1 −1 −1 
0 − 75
   
 0 0 0 0 0   0 1 0 
;    ;  
 0 1 − 53 − 45 0   0 0 1 −1 0 

 
0 0 −1 1 0 0 0 0 0 0
  
0 − 54
   
 1 2 0 −2 −1   1 0 −1 
0 − 75 0 − 75
   
 0 1 0   0 1 0 
;    ;  
 0 0 1 −1 0   0 0 1 −1 0 

 
0 0 0 0 0 0 0 0 0 0
  
85
6.7. Basis and Dimension Chapter 6. Linear Algebra
From the reduced row echelon form, the simple basis vectors are the rows with
the leading 1s (the “steps”).
 1   0   0 
     
 0   1   
     0 
U = [  0  ,  0  ,  1  ] (6.72)
     
 4  − 7  −1
− 5   5   
−1 0 0
     
|{z} |{z} |{z}
b1 b2 b3
and B = {b1 , b2 , b3 } is a (simple) basis of U (check that they are linearly inde-
pendent!).
6.7.3 Rank
• The number of linearly independent columns of a matrix A ∈ Rm×n equals the
number of linearly independent rows and is called rank of A and is denoted
by rk(A).
• rk(A) = rk(A> ), i.e., the column rank equals the row rank.
• The columns of A ∈ Rm×n span a subspace U ⊂ Rm with dim(U ) = rk(A)
• A basis of a subspace U = [x1 , . . . , xm ] ⊂ Rn can be found by executing the

following steps:
1. Write the spanning vectors as columns of a matrix A

2. Apply Gaussian elimination algorithm to A.
3. The spanning vectors associated with the pivot columns form a basis of
U.
• The rows of A ∈ Rm×n span a subspace W ⊂ Rn with dim(W ) = rk(A). A basis

of W can be found by applying the Gaussian elimination algorithm to the rows
of A (or the columns of A> ).
• For all A ∈ Rn×n holds: A is regular (invertible) if and only if rk(A) = n.
• For all A ∈ Rm×n and all b ∈ Rm : The linear equation system Ax = b can be
solved if and only if rk(A) = rk(A|b), where A|b denotes the “extended” system.
• For A ∈ Rm×n the space of solutions for Ax = 0 possesses dimension n − rk(A).
• A matrix A ∈ Rm×n has full rank if its rank equals the largest possible for a
matrix of the same dimensions, which is the lesser of the number of rows and
columns, i.e., rk(A) = min(m, n). A matrix is said to be rank deficient if it does
not have full rank.
86
Chapter 6. Linear Algebra 6.8. Intersection of Subspaces
Examples
1 0 1
 
• A = 0 1 1. A possesses two linearly independent rows (and columns).
 
0 0 0
 
Therefore, rk(A) = 2.
" #
1 2 3
• A= . We see that the second row is a multiple of the first row, such
4 8 12
" #
1 2 3
that the row-echelon form of A is , and rk(A) = 1.
0 0 0
 1 2 1
 
• A = −2 −3 1 We use Gaussian elimination to determine the rank:
 
3 5 0
 
   
 1 2 1   1 2 1 
1 ; 1 +2R1
   
 −2 −3   −2 −3 
   
3 5 0 +R1 − R2 0 0 0
 
 1 2 1 
;  0 3 
 
 −1
0 0 0

Here, we see that the number of linearly independent rows and columns is 2,
such that rk(A) = 2.
6.8 Intersection of Subspaces

In the following, we consider two approaches to determining a basis of the intersec-
tion U1 ∩ U2 of two subspaces U1 , U2 ⊂ V . This means, we are interested in finding
all x ∈ V , such that x ∈ U1 and x ∈ U2 .
6.8.1 Approach 1
Consider U1 = [b1 , . . . , bk ] ⊂ V and U2 = [cP1 , . . . , cl ] ⊂ V . We know that and x ∈ U1 can
be represented as a linear combination ki=1 λi bi of the basis vectors (or spanning
vectors) b1 , . . . , bk . Equivalently x = lj=1 ψj cj . Therefore, the approach is to find
P
λ1 , . . . , λk and/or ψ1 , . . . , ψl , such that

k
X l
X
λi b i = x = ψj cj (6.73)
i=1 j=1
k
X l
X
⇔ λi b i − ψj cj = 0 . (6.74)
i=1 j=1
87
6.8. Intersection of Subspaces Chapter 6. Linear Algebra
For this, we write the basis vectors into a matrix

h i
A = b1 · · · bk −c1 · · · −cl (6.75)
and solve the linear equation system
 λ1 
 
 . 
 .. 
 
 λ 
A  k  = 0 (6.76)
 
ψ1 
 . 
 . 
 . 
 
ψl
to find either λ1 , . . . , λk or ψ1 , . . . , ψl , which we can then use to determine U1 ∩ U2 .
Example
We consider
1 0 0 −1 0

         
1 1 0  1  1
U1 = [  ,   ,  ] ⊂ R4 , U2 = [  ,  ] ⊂ R4 . (6.77)
         
0 1 1  2  0
         
0 0 1 0 0
To find a basis of U1 ∩ U2 , we need to find all x ∈ V that can be represented as linear

combinations of the basis vectors of U1 and U2 , i.e.,
3
X 2
X
λi b i = x = ψj cj , (6.78)
i=1 j=1
where bi and cj are the basis vectors of U1 and U2 , respectively. The matrix A =
[b1 |b2 |b3 | − c1 | − c2 ] from (6.75) is given as
1 0 0 1 0 
 
1 1 0 −1 −1
A =  . (6.79)
0 1 1 −2 0 

0 0 1 0 0
By using Gaussian elimination, we determine the corresponding reduced row eche-

lon form
1 0 0 1 0
 
0 1 0 −2 0
. (6.80)
 
0 0 1 0 0

 
0 0 0 0 1
88
We keep in mind that we are interested in finding λ1 , λ2 , λ3 ∈ R and/or ψ1 , ψ2 ∈ R

with
 
  λ1 
1 0 0 1 0  λ2 

0 1 0 −2 0  
0 0 1 0 0  λ3  = 0 .
  (6.81)
  ψ 
0 0 0 0 1  1 
ψ2
From here, we can immediately see that ψ2 = 0 and ψ1 ∈ R is a free variable since it
corresponds to a non-pivot column, and our solution is
−1
 
 1 
U1 ∩ U2 = ψ1 c1 = [ ] , ψ1 ∈ R . (6.82)
 
 2 
 
0
Remark 18
Alternatively, we could have used λ1 = −ψ1 , λ2 = 2ψ1 , λ3 = 0 and determined the (same)
solution via the basis vectors of U1 as
    
 1 0 −1 −1
   
 
  

 1  1  1   1 
    
ψ1 −   + 2   = ψ1   = [ ] , ψ1 ∈ R . (6.83)
 0 1
 
 2   2 
   
0 0 0 0

6.8.2 Approach 2
In the second approach, we exploit Remark 14, which says that any subspace is the
solution of a homogeneous linear equation system, to determine the intersection
U1 ∩ U2 of two subspaces U1 , U2 ⊂ Rn .
First, we show how to determine the linear equation system that generates a sub-
space; second, we exploit these insights to find U1 ∩ U2 .
Lemma 1
Consider U = [x1 , . . . , xm ] ⊂ Rn and dim(U ) = r. We write the vectors x1 , . . . , xm as rows
of a matrix
 >
x1 
A =  ...  ∈ Rm×n
 
(6.84)
 > 
xm
and investigate the homogeneous linear equation system Ay = 0. First, the solution
space Ũ possesses dimension k = n − rk(A) = n − dim(U ) = n − r. Second, we choose a
basis (b1 , . . . , bk ) in Ũ and again write these basis vectors as the rows of a matrix
 >
b1 
B =  ...  ∈ Rk×n
 
(6.85)
 > 
bk
89
with rk(B) = k. Then U is the solution space of By = 0.20
Proof 4
Define Sh as the solution space of By = 0. It holds that dim(Sh ) = n − rk(B) = n − k = r.
Therefore, dim(Sh ) = dim(U ). From Abj = 0, j = 1, . . . , k it follows that x> i bj = 0 for
i = 1, . . . , m and j = 1, . . . , k (remember how matrix-vector multiplication works), and at
the same time b> j xi = 0. Therefore, Bxi = 0, i = 1, . . . , m and, hence, U ⊂ Sh . However,
since dim(Sh ) = dim(U ) it follows that Sh = U .
Practical Algorithm
Let us summarize the main steps to determine U1 ∩ U2 :
1. Write U1 , U2 as solution spaces of two linear equation systems B 1 x = 0 and
B 2 x = 0:
(a) Write spanning vectors of U1 , U2 as the rows of two matrices A1 , A2 ,
respectively.
(b) Determine S1 as the solution of A1 x = 0 and S2 as the solution of A2 x = 0
(c) Write spanning vectors of S1 and S2 as the rows of the matrices B 1 and
B 2 , respectively.
" #
B
2. U1 ∩ U2 is the solution space of Cx = 0, where C = 1 , which we find by
B2
means of Gaussian elimination.
Example 1
To determine the intersection of two subspaces U1 , U2 ⊂ Rn , we use the above
method. We consider again the subspaces U1 , U2 ⊂ R4 from the example above
(and hopefully, we end up with the same solution):
1 0 0 −1 0

         
1 1 0  1  1
U1 = [  ,   ,  ] ⊂ R4 , U2 = [  ,  ] ⊂ R4 . (6.86)
         
0 1 1  2  0
         
0 0 1 0 0
1. To determine the intersection U1 ∩ U2 , we first write U1 , U2 as solution spaces

of linear equation systems.
(a) We write the spanning vectors of U1 , U2 as the rows of the matrices
1 1 0 0
  " #
−1 1 2 0
(6.87)
0 1 1 0
A1 =   , A2 = ,
0 1 0 0
0 0 1 1
 
respectively.
20 It also holds that Rn = U ∪ Ũ = U ⊕ Ũ , where ⊕ is the direct sum.
90
(b) We use Gaussian elimination to determine the corrsponding reduced row

echelon forms
1 0 0 1 
  " #
1 0 −2 0
(6.88)
0 1 0 −1
Ã1 =   , Ã2 = .
0 1 0 0
0 0 1 1
 
Third, we determine the solution spaces of A1 x = 0 and A2 x = 0, e.g.,

using the Minus-1 Trick from Section 6.4.3, as
 1  2 0
     
−1 0 0
S1 = [ ] , S2 = [  ,  ] . (6.89)
     
 1  1 0
−1 0 1
(c) U1 is now the solution space of the linear equation system B 1 x = 0 with
h i
B 1 = 1 −1 1 −1 , (6.90)
and U2 is the solution space of the linear equation system B 2 x = 0 with

" #
2 0 1 0
B2 = . (6.91)
0 0 0 1
2. U1 ∩ U2 is the solution space of the linear equation system Cx = 0 with
1 −1 1 −1
" #  
B1
C= = 2 0 1 0  . (6.92)
 
B2
0 0 0 1
 
To determine this solution space, we follow the standard procedure of (a) com-
puting the reduced row echeolon form
1
1 0 2 0
 
0 1 − 1 0 (6.93)
 2 
0 0 0 1
using Gaussian elimination and (b) finding the (general) solution using the
Minus-1 Trick from Section 6.4.3 as
−1
 
 1 
U1 ∩ U2 = [ ] , (6.94)
 
 2 
 
0
which is identical to the solution in (6.83) found by using Approach 1.
91
Example 2
We apply again Approach 2 and consider the two subspaces U1 , U2 ⊂ R5 , where

             
 1  0  1  −1 −5  1  3
−1 3 −3  0  −1  2  1
             
U1 = [−1 , 3 ,  1 ] , U2 = [−4 ,  2  , −1 , 0] . (6.95)
             
−2 3 −2 −5  2   3  3
             
1 0 4 1 −6 2 3
             
1. To determine the intersection U1 ∩ U2 , we first write U1 , U2 as solution spaces

of linear equation systems.
(a) We write the spanning vectors of U1 , U2 as the rows of the matrices
−1 0 −4 −5 1 
 
1 −1 −1 −2 1
 
−5 −1 2 2 −6
A1 = 0 3 3 3 0 , A2 =  , (6.96)
   
 1 2 −1 3 2 
1 −3 1 −2 4
 
 
3 1 0 3 3
respectively.
(b) We use Gaussian elimination to determine the corrsponding reduced row
echelon forms
1 0 0 0 12
 
13
1 0 0 −1 1 
   
6 

0 1 0 0
 
Ã1 = 0 1 0 21 − 34  , Ã2 =  13   . (6.97)
 
5
0 0 1 0 − 13 

1 3 
0 0 1 2

4 1 
0 0 0 1 − 13

(c) We determine the solution spaces of A1 x = 0 and A2 x = 0, e.g., using the

Minus-1 Trick from Section 6.4.3, as
     
−2  4   12 
 1  −3  6 
     
S1 = [ 1  ,  3 ] , S2 = [ −5 ] . (6.98)
     
−2  0   −1 
     
0 −4 −13
     
U1 is the solution space of the linear equation system B 1 x = 0 with

" #
−2 1 1 −2 0
B1 = (6.99)
4 −3 3 0 −4
and U2 is the solution space of the linear equation system B 2 x = 0 with

h i
B 2 = 12 6 −5 −1 −13 . (6.100)
92
Chapter 6. Linear Algebra 6.9. Linear Mappings
2. U1 ∩ U2 is the solution space of the linear equation system Cx = 0 with

" # −2 1 1 −2 0 
B
C = 1 =  4 −3 3 0 −4  .
 
(6.101)
B2
12 6 −5 −1 −13
 
To determine this solution space, we follow the standard procedure of (a) com-
puting the reduced row echeolon form
1 0 0 0 −1
 
(6.102)
0 1 0 −1 −1
 
0 0 1 −1 −1
 
using Gaussian elimination and (b) finding the (general) solution using the
Minus-1 Trick from Section 6.4.3 as
   
0 1
1 1
   
U1 ∩ U2 = [1 , 1] . (6.103)
   
1 0
   
0 1
   
6.9 Linear Mappings

In the following, we will study mappings on vector spaces that preserve their struc-
ture: Consider two real vector spaces V , W . A mapping Φ : V → W preserves the
structure of the vector space if
Φ(x + y) = Φ(x) + Φ(y) (6.104)

Φ(λx) = λΦ(x) (6.105)
for all x, y ∈ V and λ ∈ R. We can summarize this in the following definition:

Definition 11 (Linear Mapping)
For real vector spaces V , W , a mapping Φ : V → W is called linear (or vector space
homomorphism) if
Φ(λx + ψy) = λΦ(x) + ψΦ(y) (6.106)
for all x, y ∈ V and λ, ψ ∈ R.
Important special cases:

• Isomorphism: Φ : V → W linear and bijective
• Endomorphism: Φ : V → V linear
• Automorphism: Φ : V → V linear and bijective
• We define idV : V → V , x 7→ x as the identity mapping in V .
93
6.9. Linear Mappings Chapter 6. Linear Algebra
Φ:V →W
V W
Ker(Φ) Im(Φ)
0 0
Figure 6.3: Kernel and Image of a linear mapping Φ : V → W .
Example: Homomorphism
The mapping Φ : R2 → C, Φ(x) = x1 + ix2 , is a homomorphism:
" # " #! " #! " #!
x1 y1 x1 y1
Φ + = (x1 + y1 ) + i(x2 + y2 ) = x1 + ix2 + y1 + iy2 = Φ +Φ
x2 x2 x2 y2
" #! " #!
x x1
Φ λ 1 = λx1 + λix2 = λ(x1 + ix2 ) = λΦ
x2 x2
(6.107)
We have already discussed the representation of complex numbers as tuples in R2 ,
but now we know why we can do this: There is a bijective linear mapping (we only
showed linearity, but not the bijection) that converts the elementwise addition of
tuples in R2 the set of complex numbers with the corresponding addition.
Theorem 3
Finite-dimensional R-vector spaces V and W are isomorph if and only if dim(V ) =
dim(W ).
6.9.1 Image and Kernel (Null Space)

Definition 12 (Image and Kernel)
For Φ : V → W , we define the kernel/null space
ker(Φ) := Φ −1 ({0W }) = {v ∈ V : Φ(v) = 0W } (6.108)
and the image
Im(Φ) := Φ(V ) = {w ∈ W |∃v ∈ V : Φ(v) = w} . (6.109)
An illustration is given in Figure 6.3.

Remark 19
Consider a linear mapping Φ : V → W , where V , W are vector spaces.
94
• It always holds that Φ({0V }) = 0W and, therefore, 0v ∈ ker(Φ). In particular, the

null space is never empty.
• Im(Φ) ⊂ W is a subspace of W , and ker(Φ) ⊂ V is a subspace of V .
• Φ is injective (one-to-one) if and only if ker(Φ) = {0}
Remark 20 (Null Space and Column Space)

• For A ∈ Rm×n the mapping Φ : Rn → Rm , x 7→ Ax is linear. For A = (a1 |...|an ) we
obtain
Im(Φ) = {Ax : x ∈ Rn } = {λ1 a1 + ... + λn an : λ1 , . . . , λn ∈ R} = [a1 , . . . , an ] ⊂ Rm ,

(6.110)
i.e., the image is the span of the columns of A, also called the column space.
Therefore, the column space (image) is a subspace of Rm , i.e., where m is the
“height” of the matrix.
• The kernel/null space ker(Φ) is the general solution to the linear homogeneous
equation system Ax = 0 and captures all possible linear combinations of the ele-
ments in Rn that produce 0 ∈ Rm .
• The kernel (null space) is a subspace of Rn , where n is the “width” of the matrix.
• The null space focuses on the relationship among the columns, and we can use it
to determine whether/how we can express a column as a linear combination of
other columns.
• The purpose of the null space is to determine whether a solution of the linear
equation system is unique and, if not, to capture all possible solutions.
Example: Image and Kernel of a Linear Mapping

The mapping
x1  " x1  "

   
# #
x  1 2 −1 0 x2  x1 + 2x2 − x3
Φ : R4 → R2 ,
 2 
  7→  = (6.111)
x3  1 0 0 1 x3  x1 + x4
   
x4 x4
" # " # " # " #
1 2 −1 0
= x1 + x2 + x3 + x4 (6.112)
1 0 0 1
is linear. To determine Im(Φ) we can simply take the span of the columns of the
transformation matrix and obtain
" # " # " # " #
1 2 −1 0
Im(Φ) = [ , , , ]. (6.113)
1 0 0 1
95
To compute the kernel (null space) of Φ, we need to solve Ax = 0, i.e., we need to

solve a homogeneous equation system. To do this, we use Gaussian elimination to
transform A into reduced row echelon form:
" # " # " #
1 2 −1 0 1 0 0 1 1 0 0 1
; ;
1 0 0 1 1 2 −1 0 −R1 | · ( 12 ) 0 1 − 12 − 12
This matrix is now in reduced row echelon form, and we can now use the Minus-1
Trick to compute a basis of the kernel (see Section 6.4.3).21 This gives us now the
kernel (null space) as
   
 0  −1
 1   1 
ker(Φ) = [ 2  ,  2 ]. (6.114)
   
 1   0 
0 1
   
Theorem 4 (Rank-Nullity Theorem)

For vector spaces V , W and a linear mapping Φ : V → W it holds that
dim(ker(Φ)) + dim(Im(Φ)) = dim(V ) (6.115)
Remark 21
Consider R-vector spaces V , W , X. Then:
• For linear mappings Φ : V → W and Ψ : W → X the mapping Ψ ◦ Φ : V → X is
also linear.
• If Φ : V → W is an isomorphism then Φ −1 : W → V is an isomorphism as well.
• If Φ : V → W , Ψ : V → W are linear then Φ + Ψ and λΦ, λ ∈ R are linear, too.
6.9.2 Matrix Representation of Linear Mappings

Any n-dimensional R-vector space is isomorph to Rn (Theorem 3). If we define a
basis {b1 , . . . , bn } of V we can construct an isomorphism concretely. In the following,
the order of the basis vectors will be important. Therefore, we write
B = (b1 , . . . , bn ) (6.116)
and call this n-tuple an ordered basis of V .

Definition 13 (Coordinates)
Consider an R-vector space V and an ordered basis B = (b1 , . . . , bn ). For x ∈ V we obtain
a unique representation (linear combination)
x = α1 b1 + . . . + αn bn (6.117)
21 Alternatively, we can express the non-pivot columns (columns 3 an 4) as linear combinations of
the pivot-columns (columns 1 and 2). The third column a3 is equivalent to − 12 times the second
column a2 . Therefore, 0 = a3 + 21 a2 . In the same way, we see that a4 = a1 − 12 a2 and, therefore,
0 = a1 − 21 a2 − a4 .
96
of x with respect to B. Then α1 , . . . , αn are the coordinates of x with respect to B and

the vector
 
α1 
 . 
 ..  ∈ Rn (6.118)
 
αn
 
is the coordinate vector/coordinate representation of x with respect to B.
Remark 22
For an n-dimensional R-vector space V and a basis B of V , the mapping Φ : Rn → V ,
Φ(ei ) = bi , i = 1, . . . , n is linear (and because of Theorem 3 an isomorphism), where
(e1 , . . . , en ) is the standard basis of Rn .
Now we are ready to make a connection between linear mappings between finite-
dimensional vector spaces and matrices.
Definition 14 (Transformation matrix)

Consider R-vector spaces V , W with corresponding (ordered) bases = (b1 , . . . , bn ) and
C = (c1 , . . . , cm ). Moreover, we consider a linear mapping Φ : V → W . For j ∈ {1, . . . n}
Φ(bj ) = a1j c1 + · · · + amj cm (6.119)
is the unique representation of Φ(bj ) with respect to C. Then, we call the m × n-matrix
AΦ := ((aij )) (6.120)
the transformation matrix of Φ (with respect to the ordered bases B of V and C of

W ).
Remark 23
• The coordinates of Φ(bj ) are the j-th column of AΦ .
• rk(AΦ ) = dim(Im(Φ))
• Consider (finite-dimensional) R-vector spaces V , W with ordered bases B, C, Φ :

V → W linear with transformation matrix AΦ . If x̂ is the coordinate vector of
x ∈ V and ŷ the coordinate vector of y = Φ(x) ∈ W , then
ŷ = AΦ x̂ . (6.121)
This means that the transformation matrix can be used to map coordinates with
respect to an ordered basis in V to coordinates with respect to an ordered basis in
W.
97
Example: Transformation Matrix

Consider a homomorphism Φ : V → W and ordered bases B = (b1 , . . . , b3 ) of V and
C = (c1 , . . . , c4 ) of W . With
Φ(b1 ) = c1 − c2 + 3c3 − c4
Φ(b2 ) = 2c1 + c2 + 7c3 + 2c4 (6.122)
Φ(b3 ) = 3c2 + c3 + 4c4
P4
the transformation matrix AΦ with respect to B and C satisfies Φ(bk ) = i=1 αik ci for
k = 1, . . . , 3 and is given as
 1 2 0
 
−1 1 3
AΦ = (α1 |α2 |α3 ) =  , (6.123)

 3 7 1
 
−1 2 4
where the αj , j = 1, . . . , 3, are the coordinates/coordinate vectors of Φ(bj ) with re-

spect to C.
6.9.3 Basis Change

In the following, we will have a closer look at how transformation matrices of a
linear mapping Φ : V → W change if we change the bases in V and W . Consider
B = (b1 , . . . , bn ), B̃ = (b̃1 , . . . , b̃n ) (6.124)
ordered bases of V and
C = (c1 , . . . , cm ), C̃ = (c̃1 , . . . , c̃n ) (6.125)
ordered bases of W . Moreover, AΦ is the transformation matrix of the linear map-

ping Φ : V → W with respect to the bases B and C, and ÃΦ is the corresponding
transformation mapping with respect to B̃ and C̃. We will now investigate how A
and Ã are related, i.e., how/whether we can transform AΦ into ÃΦ if we choose to
perform a basis change from B, C to B̃, C̃.
We can write the vectors of the new basis B̃ of V as a linear combination of the basis
vectors of B, such that
b̃j = s1j b1 + · · · + snj bn , j = 1, . . . , n. (6.126)
Similarly, we write the new basis vectors C̃ of W as a linear combination of the basis
vectors of C, which yields
c̃k = t1k c1 + · · · tmk cm . (6.127)
We define S = ((sij )) ∈ Rn×n and T = ((tij )) ∈ Rm×m . In particular, the jth column of S
are the coordinate representations of b̃j with respect to B and the jth columns of T
98
is the coordinate representation of c̃j with respect to C. Note that both S and T are
regular.
For all j = 1, . . . , n, we get
m m m
 m
m X

X X X X 
Φ(b̃j ) = ãkj c̃k = ãkj tik ci = t ã (6.128)
 c

 ik kj  i
k=1 |{z} k=1 i=1 i=1 k=1
∈W
where we expressed the new basis vectors c̃k ∈ W as linear combinations of the basis
vectors ci ∈ W . When we express the b̃k ∈ V as linear combinations of bi ∈ V , we
arrive at
 n  n n m
 n
m X

X  X X X X 
Φ(b̃j ) = Φ  skj bk  = skj Φ(bk ) = skj aik ci = aik skj  ci (6.129)



k=1 k=1 k=1 i=1 i=1 k=1
Comparing (6.128) and (6.129), it follows for all j = 1, . . . , n and i = 1, . . . , m that

m
X n
X
tik ãkj = aik skj (6.130)
k=1 k=1
and, therefore,
T Ã = AS, (6.131)
such that
Ã = T −1 AS. (6.132)
Hence, with a basis change in V (B is replaced with B̃) and W (C is replaced with
C̃) the transformation matrix AΦ of a linear mapping Φ : V → W is replaced by an
equivalent matrix ÃΦ with
ÃΦ = T −1 AΦ S. (6.133)
Definition 15 (Equivalence)
Two matrices A, Ã ∈ Rm×n are equivalent if there exist regular matrices S ∈ Rn×n and
T ∈ Rm×m , such that Ã = T −1 AS.
Definition 16 (Similarity)
Two matrices A, Ã ∈ Rn×n are similar if there exists a regular matrix S ∈ Rn×n with
Ã = S −1 AS
Remark 24
Similar matrices are always equivalent. However, equivalent matrices are not necessar-
ily similar.
Remark 25
Consider R-vector spaces V , W , X. From Remark 21 we already know that for linear
mappings Φ : V → W and Ψ : W → X the mapping Ψ ◦ Φ : V → X is also linear.
With transformation matrices AΦ and AΨ of the corresponding mappings, the overall
transformation matrix AΨ ◦Φ is given by AΨ ◦Φ = AΨ AΦ .
99
In light of this remark, we can look at basis changes from the perspective of concate-
nating linear mappings:
• AΦ is the transformation matrix of a linear mapping ΦBC : V → W with respect

to the bases B, C.
• ÃΦ is the transformation matrix of a linear mapping ΦB̃C̃ : V → W with respect

to the bases B̃, C̃.
• S is the transformation matrix of a linear mapping ΦB̃B : V → V (automor-

phism) that represents B̃ in terms of B.
• T is the transformation matrix of a linear mapping ΦC̃C : W → W (automor-

phism) that represents C̃ in terms of C.
If we (informally) write down the transformations just in terms of bases then
• AΦ : B → C
• ÃΦ : B̃ → C̃
• S : B̃ → B
• T : C̃ → C and T −1 : C → C̃
and
B̃ → C̃ = B̃ → B→ C → C̃ (6.134)
ÃΦ = T −1 AΦ S . (6.135)
Note that the execution order in (6.135) is from right to left because vectors are
multiplied at the right-hand side.
Example
Consider a linear mapping Φ : R3 → R4 whose transformation matrix is
 1 2 0
 
−1 1 3
AΦ =  (6.136)

 3 7 1

 
−1 2 4
with respect to the standard bases
1 0 0 0

       
1 0 0
     
0 1 0 0
B = (0 , 1 , 0) , C = (  ,   ,   ,  ). (6.137)
             
0 0 1 0
0 0 1
     
       
0 0 0 1
100
Chapter 6. Linear Algebra 6.10. Determinants
We now want to perform a basis change toward the new bases
1 1 0 1

       
1 0 1
     
      1 0 1 0
B̃ = (1 , 1 , 0) ∈ R3 , C̃ = (  ,   ,   ,  ) . (6.138)
             
0 1 1 0
0 1 1
     
       
0 0 0 1
Then,
1 1 0 1
 
1 0 1
 
1 0 1 0
S = 1 1 0 , T =  , (6.139)
  
0 1 1 0
0 1 1
 
 
0 0 0 1
where the ith column of S is the coordinate representation of b̃i in terms of the basis
vectors of B.22 Similarly, the jth column of T is the coordinate representation of c̃j
in terms of the basis vectors of C.
Therefore, we obtain
 1/2 1/2 −1/2 −1/2 3 2 1 −4 −4 −2

    
 1/2 −1/2 1/2 −1/2 0 4 2  6 0 0 
−1
ÃΦ = T AΦ S =  =  . (6.140)
    
−1/2 1/2 1/2 1/2 10 8 4  4 8 4 

    
0 0 0 1 1 6 3 1 6 3
Soon, we will be able to exploit the concept of a basis change to find a basis with
respect to which the transformation matrix of an endomorphism has a particularly
simple (diagonal) form.
6.10 Determinants
Determinants are important concepts in linear algebra. For instance, they indicate
whether a matrix can be inverted or we can use them to check for linear indepen-
dence. A geometric intuition is that the absolute value of the determinant of real
vectors is equal to the volume of the parallelepiped spanned by those vectors. Deter-
minants will play a very important role for determining eigenvalues and eigenvectors
(Section 6.11).
Determinants are only defined for square matrices A ∈ Rn×n , and we write det(A) or
|A|.
Remark 26
• For n = 1, det(A) = det(a11 ) = a11
• For n = 2,

a11 a12
det(A) = = a11 a22 − a12 a21 (6.141)
a21 a22
22 Since
B is the standard basis, this representation is straightforward toPfind. For a general basis B
we would need to solve a linear equation system to find the λi such that λi=1 3 bi = b̃j , j = 1, . . . , 3.
101
6.10. Determinants Chapter 6. Linear Algebra
• For n = 3 (Sarrus rule):

a11 a12 a13
a21 a22 a23 = a11 a22 a33 + a21 a32 a13 + a31 a12 a23 (6.142)
a31 a32 a33

− a31 a22 a13 − a11 a32 a23 − a21 a12 a33
• For an upper/lower triangular

Q matrix A, the determinant is the product of the
diagonal elements: det(A) = ni=1 aii
Remark 27 (Properties of Determinants)

• det(AB) = det(A)det(B)
• det(A) = 0 ⇔ A is singular (not invertible)
• Alternatively: A is regular ⇔ det(A) , 0.
• det(A) = det(A> )
• If A is regular then det(A−1 ) = 1/det(A)
• Similar matrices possess the same determinant. Therefore, for a linear mapping
Φ : V → V all transformation matrices AΦ of Φ have the same determinant.
Theorem 5
For A ∈ Rn×n :
1. Adding a multiple of a column/row to another one does not change det(A).
2. Multiplication of a column/row with λ ∈ R scales det(A) by λ. In particular,

det(λA) = λn det(A).
3. Swapping two rows/columns changes the sign of det(A).
Because of this theorem, we can use Gaussian elimination to compute det(A). How-
ever, we need to pay attention to swapping the sign when swapping rows.
Example

2 0 1 2 0 2 0 1 2 0 2 0 1 2 0
2 −1 0 1 1 0 −1 −1 −1 1 0 −1 −1 −1 1
0 1 2 1 2 = 0 1 2 1 2 = 0 0 1 0 3
(6.143)
−2 0 2 −1 2 0 0 3 1 2 0 0 3 1 2
2 0 0 1 1 0 0 −1 −1 1 0 0 −1 −1 1

2 0 1 2 0 2 0 1 2 0
0 −1 −1 −1 1 0 −1 −1 −1 1
= 0 0 1 0 3 = 0 0 1 0 3 = 6 (6.144)
0 0 0 1 −7 0 0 0 1 −7
0 0 0 −1 4 0 0 0 0 −3

102
Chapter 6. Linear Algebra 6.10. Determinants
We first used Gaussian elimination to bring A into triangular form, and then ex-
ploited the fact that the determinant of a triangular matrix is the product of its
diagonal elements.
Theorem 6 (Laplace Expansion)
Consider a matrix A = ((aij )) ∈ Rn×n . We define Ai,j to be the matrix that remains if we
delete the ith row and the jth column from A. Then, for all j = 1, . . . , n:
1. det(A) = nk=1 (−1)k+j akj det(Ak,j )
P
“Development about column j”
2. det(A) = nk=1 (−1)k+j ajk det(Aj,k )

P
“Development about row j”
The scalar (−1)k+1 det(Ak,j ) is also called a cofactor.
Remark 28 (Checkerboard Pattern)
The signs +/− resulting from (−1)k+j follow a checkerboard pattern, e.g., for a 3 × 3-
matrix it looks as follows:
+ − +
 
(6.145)
− + − .

+ − +
 
Example 1: Row Development

In the following, we will go through an example of how we can simplify a determi-
nant by development about a row. In particular, we develop about the first row in
the following example (the Sarrus rule can be recovered from this):

1 −2 −3
1 0 1 = 1(−1)1+1 0 1 + (−2)(−1)1+2 1 1 + (−3)(−1)1+3 1 0 (6.146)
4 4 2 4 2 4
2 4 4

= −4+2(4 − 2)−3 · 4 = −4+4−12 = −12 (6.147)

It is usually advisable to develop about rows/columns with many 0 entries.
Example 2
Let us re-compute the example in (6.143)

2 0 1 2 0 2 0 1 2 0
1

2 −1 0 1 1 0 −1 −1 −1 1 −1 −1 −1
0 1 2 1 2 = 0 1 2 1 2 1st=col. (−1)1+1 2 · 1 2 1 2

0 3 1 (6.148)
2
−2 0 2 −1 2 0 0 3 1 2
0 −1 −1 1

2 0 0 1 1 0 0 −1 −1 1
If we now subtract the fourth row from the first row and multiply (−2) times the
third column to the fourth column we obtain

−1 0 0 0
2 1 0

1 2 1 0 1st row
3rd col. 3+3 2 1
2 = −2 3 1 0 = (−2) · 3(−1) ·
= 6 (6.149)
0 3 1 0 3 1
−1 −1 3

0 −1 −1 3
103
6.11. Eigenvalues Chapter 6. Linear Algebra
6.11 Eigenvalues
Definition 17 (Eigenvalue, Eigenvector)
For an R-vector space V and a linear map Φ : V → V , the scalar λ ∈ R is called
eigenvalue if there exists an x ∈ V , x , 0, with
Φ(x) = λx. (6.150)
The corresponding vector x is called eigenvector of Φ associated with eigenvalue λ.
Definition 18 (Eigenspace and Spectrum)

• The set of all eigenvectors of Φ associated with an eigenvalue λ spans a subspace
of V , which is called eigenspace of Φ with respect to λ and denoted by Eλ .
• The set of all eigenvalues of Φ is called spectrum of Φ.
Remark 29
• Apparently, Eλ = ker(Φ − λidV ) since
Φ(x) = λx ⇔ Φ(x) − λx = 0 ⇔ (Φ − λidV )x = 0 ⇔ x ∈ ker(Φ − λidV ). (6.151)
• A matrix A ∈ Rn×n uniquely determines the linear mapping Φ : Rn → Rn , x 7→

Ax. Therefore, we can also talk of eigenvalues, eigenvectors and eigenspaces of
square matrices.
• Similar matrices possess the same eigenvalues
• Eigenvectors are not unique: If x is an eigenvector of Φ with eigenvalue λ, then

αx, α ∈ R, is an eigenvector with the same eigenvalue. Therefore, there exists an
infinite number of eigenvectors for every eigenvalue λ.
Theorem 7
Consider an R-vector space V and a linear map Φ : V → V with pairwise differ-
ent eigenvalues λ1 , . . . , λk and corresponding eigenvectors x1 , . . . , xk . Then the vectors
x1 , . . . , xk are linearly independent.
An endomorphism Φ : V → V , V ⊂ Rn (and equivalently the corresponding trans-

formation matrix A ∈ Rn×n ) possesses at most n different eigenvalues.
The following statements are equivalent:
• λ is eigenvalue of A ∈ Rn×n
• There exists an x ∈ Rn , x , 0 with Ax = λx or, equivalently, (A − λI n )x = 0
• (A − λI n )x = 0 can be solved non-trivially, i.e., x , 0.
• rk(A − λI n ) < n
• det(A − λI n ) = 0
Note that A and A> possess the same eigenvalues, but not the same eigenvectors.
104
Chapter 6. Linear Algebra 6.11. Eigenvalues
6.11.1 Geometric Interpretation

Geometrically, an eigenvector corresponding to a real, nonzero eigenvalue points in
a direction that is stretched, and the eigenvalue is the factor by which it is stretched.
If the eigenvalue is negative, the direction is reversed. In particular, the eigenvector
does not change its direction under Φ.
6.11.2 Characteristic Polynomial

In the following, we will discuss how to determine the eigenspaces of an endomor-
phism Φ.23 For this, we need to introduce the characteristic polynomial first.
Definition 19 (Characteristic Polynomial)

For λ ∈ R and an endomorphism Φ on Rn
p(λ) = det(A − λI ) = a0 + a1 λ + a2 λ2 + · · · + an−1 λn−1 + (−1)n λn , a0 , . . . , an−1 ∈ R,

(6.152)
is the characteristic polynomial of A. In particular,
a0 = det(A), (6.153)
n−1
an−1 = (−1) tr(A), (6.154)
Pn
where tr(A) = i=1 aii is the trace of A and defined as the sum of the diagonal elements
of A.
Theorem 8
λ ∈ R is eigenvalue of A ∈ Rn×n if and only if λ is a root of the characteristic polynomial
p(λ) of A.
Remark 30
1. If λ is an eigenvalue of A ∈ Rn×n then the corresponding eigenspace Eλ is the
solution space of the homogeneous linear equation system (A − λI n )x = 0.
2. Similar matrices possess the same characteristic polynomial.
6.11.3 Example: Eigenspace Computation

" #
1 0
• A=
1 1

1−λ 0
1. Characteristic polynomial: p(λ) = |A − λI 2 | = = (1−λ)2 . There-
1 1 − λ
fore λ = 1 is the only root of p and, therefore, the only eigenvalue of A
23 It
turns out that it is sufficient to work directly with the corresponding transformation mappings
AΦ ∈ Rn×n .
105
6.11. Eigenvalues Chapter 6. Linear Algebra
2. To compute the eigenspace for the eigenvalue λ = 1, we need to compute

the null space of A − I:
" #
0 0
(A − 1 · I )x = 0 ⇔ x=0 (6.155)
1 0
" #
0
⇒ E1 = [ ] (6.156)
1
" #
0 1
• A=
−1 0
1. Characteristic polynomial: p(λ) = det(A − λI) = λ2 + 1. For λ ∈ R there

exist no eigenvalue of A. However, for λ ∈ C we find λ1 = i, λ2 = −i.
2. The corresponding eigenspaces (for λi ∈ C) are
" # " #
1 1
Ei = [ ], E−i = [ ]. (6.157)
i −i
 0 −1 1 1
 
−1 1 −2 3
• A = 
 
 2 −1 0 0

 
1 −1 1 0
1. Characteristic polynomial:

−λ −1 1 1 −λ −1 1 1
−1 1 − λ −2 3 0 −λ −1 3 − λ

p(λ) = = (6.158)
2 −1 −λ 0 0 1 −2 − λ 2λ
1 −1 1 −λ 1 −1 1 −λ

−λ −1 − λ 0 1
0 −λ −1 − λ 3 − λ

= (6.159)
0 1 −1 − λ 2λ
1 0 0 −λ

−λ −1 − λ 3 − λ −1 − λ 0 1
= (−λ) 1 −1 − λ 2λ − −λ −1 − λ 3 − λ (6.160)
0 0 −λ 1 −1 − λ 2λ

−1 − λ 0 1
2 −λ −1 − λ
= (−λ) − −λ −1 − λ 3 − λ (6.161)
1 −1 − λ
1 −1 − λ 2λ

= (1 + λ)2 (λ2 − 3λ + 2) = (1 + λ)2 (1 − λ)(2 − λ) (6.162)

Therefore, the eigenvalues of A are λ1 = −1, λ2 = 1, λ3 = 2.
2. The corresponding eigenspaces are the solutions of (A − λi I )x = 0, i =
1, 2, 3, and given by
0 1 1
     
1 1 0
E−1 = [ ], E1 = [ ], E2 = [ ]. (6.163)
     
1 1 1
     
0 1 1
106
Chapter 6. Linear Algebra 6.11. Eigenvalues
6.11.4 Applications
4 2 0 2 4 6
(a) Data set with eigenvectors

4 4
3 3
2 2
1 1
0 0
1 1
2 2
3 3
4 2 0 2 4 6 4 2 0 2 4 6
(b) Data set with projection onto the first princi- (c) Data set with projection onto the second prin-
pal component cipal component
Figure 6.4: (a) Two-dimensional data set (red) with corresponding eigenvectors (blue),
scaled by the magnitude of the corresponding eigenvectors. The longer the eigenvector
the higher the variability (spread) of the data along this axis. (b) For optimal (linear)
dimensionality reduction, we would project the data onto the subspace spanned by the
eigenvector associated with the largest eigenvalue. (c) Projection of the data set onto
the subspac spanned by the eigenvector associated with the smaller eigenvalue leads to
a larger projection error.
• Eigenvectors and eigenvalues are fundamental to principal components anal-

ysis (PCA24 , Hotelling (1933)), which is commonly used for dimensionality
reduction in face recognition, data visualization and other machine learning
applications. Eigenvalues of a “data matrix” tell us, which dimensions of a
high-dimensional data vector are important, i.e., which dimensions contain
24 also known as Kosambi-KarhunenLoève Transform
107
6.12. Diagonalization Chapter 6. Linear Algebra
a significant signal in the data.25 These dimensions are important to keep,

whereas dimensions associated with small eigenvalues can be discarded with-
out much loss. The eigenvectors associated with the largest eigenvalues are
called principal components. Figure 6.4 illustrates this for two-dimensional
data.
• Eigenvalues were used by Claude Shannon to determine the theoretical limit to

how much information can be transmitted through a communication medium
like your telephone line or through the air. This is done by calculating the
eigenvectors and eigenvalues of the communication channel (expressed a ma-
trix), and then waterfilling on the eigenvalues. The eigenvalues are then, in
essence, the gains of the fundamental modes of the channel, which themselves
are captured by the eigenvectors.
• Google uses the eigenvector corresponding to the maximal eigenvalue of the

Google matrix to determine the rank of a page for search. The idea that the
PageRank algorithm26 brought up was that the importance of any web page
can be judged by looking at the pages that link to it. For this, we write down
all websites as a huge (weighted) directed graph that shows which page links to
which. Then, the navigation behavior of a user can be described by a transition
matrix A of this graph that tells us with what (click) probability somebody
will end up on a different website. The matrix A has the property that for
any initial rank/importance vector x of a website the sequence x, Ax, A2 x, . . .
converges to a vector x∗ . This vector is called the PageRank and satisfies Ax∗ =
x∗ , i.e., it is an eigenvector (with corresponding eigenvalue 1).27 More details
and different perspectives on PageRank can be found at http://tinyurl.com/
83tehpk.
• Eigenvalues are frequently used to determine numerical stability, e.g., when

inverting matrices. Since a computer can only represent numbers with a finite
precision, we often look at condition numbers of matrices, i.e., the ratio |λ max |
|λmin |
of the biggest to the smallest eigenvalue. If this ratio exceeds a threshold
(e.g., 108 ), matrix inversion may become numerically unstable and lead to
inaccurate results.
6.12 Diagonalization
Diagonal matrices are possess a very simple structure and they allow for a very fast
computation of determinants and inverses, for instance. In this section, we will have
a closer look at how to transform matrices into diagonal form. More specifically, we
25 Tobe more precise, PCA can be considered a method for (1) performing a basis change from the
standard basis toward the eigenbasis in Rd , d < n, (2) projecting the data in Rn onto the subspace
spanned by the d eigenvectors corresponding to the d largest eigenvalues (which are called the
principal components), (3) moving back to the standard basis.
26 Developed at Stanford University by Larry Page and Sergey Brin in 1996.
27 When normalizing x∗ , such that kx∗ k = 1 we can interpret the entries as probabilities.
108
Chapter 6. Linear Algebra 6.12. Diagonalization
will look at endomorphisms of finite-dimensional vector spaces, which are similar

to a diagonal matrix, i.e., endomorphisms whose transformation matrix attains di-
agonal structure for a suitable basis. Here, we finally have a practically important
application of the basis change we discussed in Section 6.9.3 and eigenvalues form
Section 6.11.
Definition 20 (Diagonal Form)
A matrix A ∈ Rn×n is diagonalizable if it is similar28 to a diagonal matrix
c1 0 · · · · · · 0 
 
 0 c2 0 · · · 0 

 . .. .. ..  .
 .. . . .  (6.164)

 0 · · · 0 cn−1 0 
 
 
0 · · · · · · 0 cn
Theorem 9
For an endomorphism Φ of an n-dimensional R-vector space V the following statements
are equivalent:
1. Φ is diagonalizable.
2. The transformation matrix AΦ is diagonalizable.
3. There exists a basis in V consisting of the eigenvectors of Φ.29
4. The sum of the dimensions of the eigenspaces of Φ is n.30
Theorem 10
For an n-dimensional R-vector space V and a linear mapping Φ : V → V the following
holds: Φ is diagonalizable if and only if
1. Its characteristic polynomial p(λ) is given in the form
p(λ) = (−1)n (λ − c1 )r1 · · · (λ − ck )rk (6.165)
with ri ∈ N and pairwise different roots ci ∈ R and
2. For i = 1, . . . , k
dim(Im(Φ − ci idV )) = n − ri (6.166)

Rank-Nullity
⇐⇒ dim(ker(Φ − ci idV )) = ri . (6.167)
28 Remember: Two matrices A, D are similar if and only if there exists an invertible matrix S, such
that D = S −1 AS.
29 Therefore, we need n eigenvectors.
30 In particular, an endomorphism Φ of an n-dimensional R-vector space with n different eigenval-
ues is diagonalizable.
109
In (6.165) we say that the characteristic polynomial decomposes into linear factors.
The second requirement in (6.166) says that the dimension of the eigenspace Eci
must correspond to the (algebraic) multiplicity ri of the eigenvalues in the character-
istic polynomial, i = 1, . . . , k.31 The dimension of the eigenspace Eci is the dimension
of the kernel/null space of Φ − ci idV .
Theorem 10 holds equivalently if we replace Φ with A ∈ Rn×n and idV with I n .
If Φ is diagonalizable it possesses a transformation matrix of the form
 
c1 0 · · · · · · · · · · · · 0 
 0 . . . . . .
 .. 
 . 
 . . ... .. 
 .. . . c . 
 1
 .. . .
.. .. .. . .. 
AΦ =  . .  (6.168)
 .. .

. .. c . . . .. 

 . k
 . .. ..

 .. . . 0 


0 · · · · · · · · · · · · 0 ck
 
where each eigenvalue ci appears ri times (its multiplicity in the characteristic poly-
nomial and the dimension of the corresponding eigenspace) on the diagonal.
6.12.1 Examples
"
#
1 0
• A= .
1 1
1. Characteristic polynomial: p(λ) = (1 − λ)2

2. The dimension of eigenspace is 2 − rk(A − I ) = 1 , 2. Because of Theo-
rem 10 A is not diagonalizable.
" #
0 1
• A= .
−1 0
• Characteristic polynomial: p(λ) = 1 + λ2 .

– For λ ∈ R there exist no roots of p(λ), and the characteristic polynomial
does not decompose into linear factors. Therefore, A is not diagonaliz-
able.
– If we consider a C-vector space and λ ∈ C, then p(λ) = (i −λ)(−i −λ) and A
has two eigenvalues, the characteristic polynomial decomposes into linear
factors, and A is diagonalizable (verify the dimension of the eigenspaces).
 0 −1 1 1
 
−1 1 −2 3
• A =  .
 
 2 −1 0 0
 
1 −1 1 0
31 More explicitly, the algebraic multiplicity of ci is the number of times ri it is repeated as a root
of the characteristic polynomial.
110
Chapter 6. Linear Algebra 6.12. Diagonalization
1. Characteristic polynomial: p(λ) = (1 + λ)2 (1 − λ)(2 − λ). The eigenvalues

are c1 = −1, c2 = 1, c3 = 2 with algebraic multiplicities r1 = 2, r2 = 1, r3 = 1,
respectively.
2. Dimension of eigenspaces: dim(Ec1 ) = 1 , r1 .
Therefore, A cannot be diagonalized.
3 2 −1
 
• A = 2 6 −2.
 
0 0 2
 
1. Characteristic polynomial: p(λ) = (2 − λ)2 (7 − λ). Therefore, c1 = 2, c2 =

7, r1 = 2, r2 = 1
2. Dimension of eigenspaces: rk(A−c1 I 3 ) = 1 = n−r1 , rk(A−c2 I 3 ) = 2 = n−r2
Therefore, A is diagonalizable.
Let us now discuss a practical way of constructing diagonal matrices.
Remark 31
If A ∈ Rn×n is diagonalizable and (b1 , . . . , bn ) is an ordered basis of eigenvectors of A
with Abi = ci bi , i = 1, . . . , n, then it holds that for the regular matrix S = (b1 | . . . |bn )
c1 0 ··· ··· 0 

 
 0 c2 0 ··· 0 
 . .. 

... ... ...
S AS =  ..
−1
.  . (6.169)
 . 
 .
 . · · · 0 cn−1 0 

0 ··· ··· 0 cn
 
The diagonal matrix in (6.169) is the transformation matrix of x 7→ Ax with respect to

the eigenbasis (b1 , . . . , bn ).
Coming backto the above example, where we wanted to determine the diagonal
3 2 −1

 
form of A = 2 6 −2. We already know that A is diagonalizable. We now de-

0 0 2
 
termine the eigenbasis of R3 that allows us to transform A into a similar matrix in
diagonal form via S −1 AS:
1. The eigenspaces are
1 −2 1

     
E2 = [ 0 ,  1  ] , E7 = [ 2 ] (6.170)
     
1 0 0
     
|{z} |{z} |{z}
=:b1 =:b2 =:b3
111
2. We now collect the eigenvectors in a matrix and obtain
1 −2 1
 
S = (b1 |b2 |b3 ) = 0 1 2 (6.171)
 
1 0 0
 
such that
2 0 0
 
S −1 AS = 0 2 0 . (6.172)
 
0 0 7
 
Remark 32
The dimension of the eigenspace Eλ cannot be greater than the algebraic multiplicity of
the corresponding eigenvalue λ.
Remark 33
So far, we computed diagonal matrices as D = S −1 AS. However, we can equally write
A = SDS −1 . Here, we can interpret the transformation matrix A as follows: S −1 per-
forms a basis change from the standard basis into the eigenbasis. Then, D then scales
the vector along the axes of the eigenbasis, and S transforms the scaled vectors back
into the standard/canonical coordinates.
6.12.2 Applications
Diagonal matrices D = S −1 AS exhibit the nice properties that they can be easily
raised to a power:
Ak = (S −1 DS)k = S −1 D k S . (6.173)
Computing D k is easy because we apply this operation individually to any diagonal

element. As an example, this allows to compute inverses of D in O(n) instead of
O(n3 ).
A different property of diagonal matrices is that they decouple variables. This is
important in probability theory to interpret random variables, e.g., for Gaussian
distributions.
With diagonal matrices, it is easier to analyze properties of differential equations,
which play an important role in any kind of (linear) dynamical system.
6.12.3 Cayley-Hamilton Theorem

Theorem 11 (Cayley-Hamilton)
Let V be an n-dimensional R-vector space and Φ : V → V an endomorphism with
transformation matrix AΦ and characteristic polynomial p. Then,
p(Φ) = 0 (6.174)
(and equivalently, p(AΦ ) = 0).
112
Chapter 6. Linear Algebra 6.13. Scalar Products
Remark 34
• Note that the right hand side of (6.174) is the zero mapping (or the 0-matrix
when we use the transformation matrix AΦ ).
• The importance of the Cayley-Hamilton theorem is not the existence of a (non-

trivial) polynomial q, such that q(Φ) = 0, but that the characteristic polynomial
has this property.
Applications
# "
−1 2 n−1 1 −1
• Find an expression for A in terms of I , A, A , . . . , A . Example: A =
2 1
2
has the characteristic polynomial p(λ) = λ − 2λ + 3. Then Theorem 11 states
that A2 − 2A + 3I = 0 and, therefore, −A2 + 2A = 3I ⇔ A−1 = 13 (2I − A)
• Find an expression of Am , m ≥ n, in terms of I , A, A2 , . . . , An−1
6.13 Scalar Products

Definition 21
Let β : V × V → R be a bilinear mapping (i.e., linear in both arguments).
• β is called symmetric if β(x, y) = β(y, x) for all x, y ∈ V .
• β is called positive definite if for all x , 0: β(x, x) > 0. β(0, 0) = 0.
• A positive definite, symmetric bilinear mapping β : V × V → R is called scalar

product/dot product/inner product on V . We typically write hx, yi instead of
β(x, y).
• The pair (V , h·, ·i) is called Euclidean vector space or (real) vector space with
scalar product.
6.13.1 Examples
Pn
• For V = Rn we define the standard scalar product hx, yi := x> y = i=1 xi yi .
• V = R2 . If we define β(x, y) = hx, yi := x1 y1 − (x1 y2 + x2 y1 ) + 2x2 y2 then β is a

scalar product but different from the standard scalar product.
In a Euclidean vector space, the scalar product allows us to introduce concepts, such
as lengths, distances and orthogonality.
113
6.13. Scalar Products Chapter 6. Linear Algebra
6.13.2 Lengths, Distances, Orthogonality

Definition 22 (Norm) √
Consider a Euclidean vector space (V , h·, ·i). Then kxk := hx, xi is the length or norm
of x ∈ V . The mapping
k·k : V → R (6.175)
x 7→ kxk (6.176)
is called norm.
6.13.3 Example
In geometry, we are often interested in lengths of vectors. We can now use the scalar
product to compute them. For instance, in a Euclidean vector√ space√with standard
>
scalar product, if x = [1, 2] then its norm/length is kxk = 1 + 22 = 5
2
Remark 35
The norm k · k possesses the following properties:
1. kxk ≥ 0 for all x ∈ V and kxk = 0 ⇔ x = 0
2. kλxk = |λ| · kxk for all x ∈ V and λ ∈ R
3. Minkowski inequality: kx + yk ≤ kxk + kyk for all x, y ∈ V
Definition 23 (Distance and Metric)

Consider a Euclidean vector space (V , h·, ·i). Then d(x, y) := kx − yk is called distance of
x, y ∈ V . The mapping
d : V ×V → R (6.177)
(x, y) 7→ d(x, y) (6.178)
is called metric.
A metric d satisfies:
1. d is positive definite, i.e., d(x, y) ≥ 0 for all x, y ∈ V and d(x, y) = 0 ⇔ x = y
2. d is symmetric, i.e., d(x, y) = d(y, x) for all x, y ∈ V .
3. Triangular inequality: d(x, z) ≤ d(x, y) + d(y, z).

Definition 24 (Orthogonality)
Vectors x and y are orthogonal if hx, yi = 0, and we write x ⊥ y
Theorem 12
Let V be a Euclidean vector space (V , h·, ·i) and x, y, z ∈ V . Then:
1. Cauchy-Schwarz inequality: |hx, yi| ≤ kxk kyk
114
Chapter 6. Linear Algebra 6.13. Scalar Products
2. Minkowski inequality: kx + yk ≤ kxk + kyk

3. Triangular inequality: d(x, z) ≤ d(x, y) + d(y, z)
4. Parallelogram law: kx + yk + kx − yk = 2kxk2 + 2kyk2
5. 4hx, yi = kx + yk2 − kx − yk2
6. x ⊥ y ⇔ kx + yk2 = kxk2 + kyk2
The Cauchy-Schwarz inequality allows us to define angles ω in Euclidean vector
spaces between two vectors x, y. Assume that x , 0, y , 0. Then
hx, yi
−1 ≤ ≤1 (6.179)
kxk kyk
Therefore, there exists a unique ω ∈ [0, π) with
hx, yi
cos ω = (6.180)
kxk kyk
The number ω is the angle between x and y.
6.13.4 Applications
Scalar products allow us to compute angles between vectors or distances. A major
purpose of scalar products is to determine whether vectors are orthogonal to each
other; in this case hx, yi = 0. This will play an important role when we discuss projec-
tions in Section 6.16. The scalar product also allows us to determine specific bases of
vector (sub)spaces, where each vector is orthogonal to all others (orthogonal bases)
using the Gram-Schmidt method32 . These bases are important optimization and
numerical algorithms for solving linear equation systems. For instance, Krylov sub-
space methods33 , such as Conjugate Gradients or GMRES, minimize residual errors
that are orthogonal to each other (Stoer and Burlirsch, 2002).
In machine learning, scalar products are important in the context of kernel meth-
ods (Schölkopf and Smola, 2002). Kernel methods exploit the fact that many linear
algorithms can be expressed purely by scalar product computations.34 Then, the
“kernel trick” allows us to compute these scalar products implicitly in a (potentially
infinite-dimensional) without even knowing this feature space explicitly. This al-
lowed the “non-linearization” of many algorithms used in machine learning, such
as kernel-PCA (Schölkopf et al., 1998) for dimensionality reduction. Gaussian pro-
cesses (Rasmussen and Williams, 2006) also fall into the category of kernel methods
and are the current state-of-the-art in probabilistic regression (fitting curves to data
points).
32 not discussed in this course
33 The basis for the Krylov subspace is derived from the CayleyHamilton theorem, see Sec-
tion 6.12.3, which allows us to compute the inverse of a matrix in terms of a linear combination
of its powers.
34 Matrix-vector multiplication Ax = b falls into this category since b is a scalar product of the ith
i
row of A with x.
115
6.14. Affine Subspaces Chapter 6. Linear Algebra
L y
x0
Figure 6.5: Points y on a line lie in an affine subspace L with support point x0 and
direction u.
6.14 Affine Subspaces

In the following, we will have a closer look at geometric properties of vector spaces.
For this purpose, we define an affine subspace.
Definition 25 (Affine Subspace)
Let V be a vector space, x0 ∈ V and U ⊂ V a subspace. Then the subset
L = x0 + U := {x0 + u : u ∈ U } = {v ∈ V |∃u ∈ U : v = x0 + u} ⊂ V (6.181)
is called affine subspace of V .35 U is called direction or direction space, and x0 is

called support point.
Note that the definition of an affine subspace excludes 0 if the support point x0 < U .
Therefore, an affine subspace is not a (linear) subspace of V for x0 < U .
Examples of affine subspaces are points, lines and planes in R2 , which do not (nec-
essarily) go through the origin.
Remark 36
• Consider two affine subspaces L = x0 + U and L̃ = x̃0 + Ũ of a vector space V .
L ⊂ L̃ if and only if U ⊂ Ũ and x0 − x̃0 ∈ Ũ .
• Affine subspaces are often described by parameters: Consider a k-dimensional

affine subspace L = x0 + U of V . If (b1 , . . . , bk ) is an (ordered) basis of U , then
every element x ∈ L can be (uniquely) described as
x = x 0 + λ1 b 1 + . . . + λk b k , (6.182)
where λ1 , . . . , λk ∈ R.
This representation is called parametric equation of L with directional vectors
b1 , . . . , bk and parameters λ1 , . . . , λk .
Examples
• One-dimensional affine subspaces are called lines and can be written as y =
x0 + λx1 , where λ ∈ R, where U = [x1 ] ⊂ Rn is a one-dimensional subspace of
Rn .
35 L is also called linear manifold.
116
Chapter 6. Linear Algebra 6.14. Affine Subspaces
• Two-dimensional affine subspaces of Rn are called planes. The parametric

equation for planes is y = x0 + λ1 x1 + λ2 x2 , where λ1 , λ2 ∈ R and [x1 , x2 ] is a
basis of U ⊂ Rn .
• In an n-dimensional vector space V , the (n − 1)-dimensional affine subspaces
are called hyperplanes.
• For A ∈ Rm×n and b ∈ Rm the solution of the linear equation system Ax = b
is either the empty set or an affine subspace of Rn of dimension n − rk(A).
In particular, the solution of the linear equation λ1 x1 + . . . + λn xn = b, where
(λ1 , . . . , λn ) , (0, . . . , 0), is a hyperplane in Rn .
6.14.1 Intersection of Affine Spaces

• In Rn every k-dimensional affine subspace is the solution of a linear equation
system Ax = b, where A ∈ Rm×n , b ∈ Rm and rk(A) = n − k.36
• Geometric interpretation: The intersection of a finite set of hyperplanes in Rn
is either empty or an affine subspace. Moreover, every k-dimensional affine
subspace in Rn is the intersection of finitely many hyperplanes.
Consider two linear subspaces L1 , L2 ⊂ R4 :
2 1 0 0 3 −1 0
             
0 1 1 0 1  1  1
L1 =   + [  ,   ,  ] , L2 =   + [  ,  ] (6.183)
             
0 0 1 1 0  2  0
             
1 0 0 1 0 0 0
|{z} | {z } |{z} | {z }
x1 U1 x2 U2
We want to determine the intersection L1 ∩ L2 . This can be done by exploiting above

remark that every affine subspace is the solution of an inhomogeneous equation
system. Alternatively (and perhaps more intuitively), we also know that it holds
that x ∈ L1 ∩ L2 ⇔ x ∈ L1 ∧ x ∈ L2 ⇔ ∃λ1 , λ2 , λ3 , ψ1 , ψ2 ∈ R, such that
2 1 0 0 3 −1 0
             
0 1 1 0 1  1  1
  + λ1   + λ2   + λ3   = x =   + ψ1   + ψ2   . (6.184)
             
0 0 1 1 0  2  0
             
1 0 0 1 0 0 0
| {z } | {z }
∈L1 ∈L2
From this, we obtain the following inhomogeneous equation system with unknowns
λ1 , λ 2 , λ 3 , ψ 1 , ψ 2 :
 1 0 0 1 0 1 
 
 1 1 0 −1 −1 1 
 
 0 1 1 −2 0 0 
 
 
0 0 1 0 0 −1
36 Recall that for homogeneous equation systems Ax = 0 the solution was a vector subspace (not
affine).
117
6.14. Affine Subspaces Chapter 6. Linear Algebra
L2 1
0 1
L1
Figure 6.6: Parallel lines. The affine subspaces L1 and L2 are parallel with L1 ∩ L2 = ∅.
Using Gaussian elimination, we quickly obtain the reduced row echelon form
 1 0 0 1 0 1 
 
 0 1 0 −2 −1 0 
,

 0 0 1 0 1 0 

 
0 0 0 0 1 1
which yields ψ2 = 1, ψ1 ∈ R. Plugging this into (6.184), we obtain that x ∈ L1 ∩ L2

satisfies
3 −1 0 3 −1 3 −1

             
1  1  1 2  1  2  1 
x =   + ψ1   + 1 ·   =   + ψ1   =   + [ ] , (6.185)
             
0  2  0 0  2  0  2 
             
0 0 0 0 0 0 0
and L1 ∩ L2 is a line (1D affine subspace) in R4 .
Parallelism
Two affine subspaces L1 and L2 are parallel (L1 ||L2 ) if the following holds for the
corresponding direction spaces U1 , U2 : U1 ⊂ U2 or U2 ⊂ U1 .37
Parallel affine subspaces, which do not contain each other (such that L1 ⊂ L2 or
L2 ⊂ L1 ), have no points in common, i.e., L1 ∩ L2 = ∅. For instance,
" # " # " # " #
1 1 2 1
L1 = + [ ] , L2 = +[ ] (6.186)
−1 0 1 0
are parallel (U1 = U2 in this case) and L1 ∩ L2 = ∅ because the lines are offset as
illustrated in Figure 6.6.
We talk about skew lines if they are neither parallel nor have an intersection point.
Imagine non-parallel lines in 3D that “miss” each other.
37 Note that this definition of parallel allows for L1 ⊂ L2 or L2 ⊂ L1 .
118
Chapter 6. Linear Algebra 6.14. Affine Subspaces
Examples
1. Consider two lines g = x1 + U1 , h = x2 + U2 in R2 , where U1 , U2 ⊂ R are one-
dimensional subspaces.
(a) If g ∩ h , ∅:
• For dim(U1 ∩ U2 ) = 0, we get a single point as the intersection.
• For dim(U1 ∩ U2 ) = 1 it follows that U1 = U2 and g = h.
(b) If g ∩ h = ∅:
• For dim(U1 ∩ U2 ) = 1 it follows again that U1 = U2 , and g||h, g , h
(because they do not intersect).
• The case dim(U1 ∩ U2 ) = 0 cannot happen in R2 .
2. Consider two lines g = x1 + U1 , h = x2 + U2 in Rn , n ≥ 3
(a) For g ∩ h , ∅ we obtain (as in R2 ) that either g and h intersect in a single
point or g = h.
(b) If g ∩ h = ∅:
• For dim(U1 ∩ U2 ) = 0, we obtain that g and h are skew lines (this
cannot happen in R2 ). This means, there exists no plane that contains
both g and h.
• If dim(U1 ∩ U2 ) = 1 it follows that g||h.
3. Consider two hyper-planes L1 = x1 + U1 and L2 = x2 + U2 in Rn , n = 3, where
U1 , U2 ⊂ R2 are two-dimensional subspaces.
(a) If L1 ∩ L2 , ∅ the intersection is an affine subspace. The kind of subspace
depends on the dimension dim(U1 ∩ U2 ) of the intersection of the corre-
sponding direction spaces.
• dim(U1 ∩ U2 ) = 2: Then U1 = U2 and, therefore, L1 = L2 .
• dim(U1 ∩ U2 ) = 1: The intersection is a line.
• dim(U1 ∩ U2 ) = 0: Cannot happen in R3 .
(b) If L1 ∩ L2 = ∅, then dim(U1 ∩ U2 ) = 2 (no other option possible in R3 ) and
L1 ||L2 .
4. Consider two planes L1 = x1 + U1 and L2 = x2 + U2 in Rn , n = 4, where U1 , U2 ⊂
R2 are two-dimensional subspaces.
(a) For L1 ∩ L2 , ∅ the additional case is possible that the planes intersect in a
point.
(b) For L1 ∩ L2 = ∅ the additional case is possible that dim(U1 ∩ U2 ) = 1. This
means that the planes are not parallel, they have no point in common, but
there is a line g such that g||L1 and g||L2 .
5. For two planes L1 = x1 + U1 and L2 = x2 + U2 in Rn , n > 4, where U1 , U2 ⊂ R2
are two-dimensional subspaces, all kinds of intersections are possible.
119
6.15. Affine Mappings Chapter 6. Linear Algebra
6.15 Affine Mappings

Definition 26 (Affine mapping)
For two R-vector spaces V , W and a linear mapping Φ : V → W and a ∈ W the mapping
φ :V → W (6.187)
x 7→ a + Φ(x) (6.188)
is an affine mapping from V to W . The vector a is called translation vector of φ.
• Every affine mapping φ : V → W is also the composition of a linear mapping

Φ : V → W and a translation τ in W , such that φ = τ ◦ Φ. The mappings Φ and
τ are uniquely determined.
• The composition φ0 ◦ φ of affine mappings φ : V → W , φ0 : W → X is affine.
• Affine mappings keep the geometric structure invariant, and preserve the di-
mension and parallelism.
Theorem 13
Let V , W be finite-dimensional R-vector spaces and φ : V → W an affine mapping.
Then:
• If L is an affine subspace of V then φ(L) is an affine subspace of W and dim(φ(L)) ≤

dim(L).
• φ preserves parallelism, i.e., for all affine subspaces L1 , L2 ⊂ V it follows that if

L1 ||L2 in V then φ(L1 )||φ(L2 ) in W .
6.16 Projections
Projections are an important class of linear transformations (besides rotations and
reflections38 ). Projections play an important role in graphics (see e.g., Figure 6.7),
coding theory, statistics and machine learning. We often deal with data that is very
high-dimensional. However, often only a few dimensions are important. In this case,
we can project the original very high-dimensional data into a lower-dimensional
feature space39 and work in this lower-dimensional space to learn more about the
data set and extract patterns. For example, machine learning tools such as Principal
Components Analysis (PCA) by Hotelling (1933) and Deep Neural Networks (e.g.,
deep auto-encoders, first applied by Deng et al. (2010)) exploit this idea heavily.
In the following, we will focus on linear orthogonal projections.
Definition 27 (Projection)
Let V be an R-vector space and W ⊂ V a subspace of V . A linear mapping π : V → W
is called a projection if π2 = π ◦ π = π.
38 Note that translations are not linear. Why?
39 “Feature” is just a commonly word for “data representation”.
120
Chapter 6. Linear Algebra 6.16. Projections
Figure 6.7: The shade is the projection of the ball onto a plane (table) with the center
being the light source. Adapted from http://tinyurl.com/ka4t28t
Remark 37 (Projection matrix)

Since linear mappings can be expressed by transformation matrices, the definition above
applies equally to a special kind of transformation matrices, the projection matrices
P π , which exhibit the property that P 2π = P π .
Since P 2π = P π it follows that all eigenvalues of P π are either 0 or 1. The corresponding
eigenspaces are the kernel and image of the projection, respectively.40
In the following, we will derive projections from points in the Euclidean vector space
(Rn , h·, ·i) onto subspaces. We will start with one-dimensional subspaces, the lines. If
not mentioned otherwise, we assume the standard scalar product hx, yi = x> y.
6.16.1 Projection onto a Line

Assuming we are given a line through the origin with direction b ∈ Rm . This is a
one-dimensional subspace U ⊂ Rm spanned by b. When we project x ∈ Rm onto U ,
we want to find the point πU (x) = p in U that is closest to x. Let us characterize
some properties of the projection p = πU (x):
• p is closest to x, where “closest” implies that kx − pk is minimal. This means,

the segment p − x from p to x is orthogonal to U and, therefore, the basis b
of U . The orthogonality condition yields hp − x, bi = (p − x)> b = 0 (when we
assume the standard scalar product).
• The projection p = πU (x) of x onto U must be a multiple of the basis/direction

b (remember that in this example the support point is 0, and the line is a
subspace). Thus, p = λb, for some λ ∈ R.
40 A good illustration is given here: http://tinyurl.com/p5jn5ws.
121
6.16. Projections Chapter 6. Linear Algebra
b
x
p = πU (x)
sin ω
ω ω
cos ω
b
(a) General case (b) Special case: Unit circle
with kxk = 1
Figure 6.8: Projection πU (x) of x onto a subspace U with basis b.
In the following three steps, we determine λ, the projection point πU (x) = p ∈ U and
the projection matrix P π that maps arbitrary x ∈ Rn onto U .
1. Finding λ. We know that p = λb. Therefore,
x − p = x − λb ⊥ b (6.189)
⇔b> (x − λb) = b> x − λb> b = 0 (6.190)
b> x b> x
⇔λ = b = (6.191)
b> b kbk2
2. Finding the projection point p = πU (x) =∈ U on U . With p = λb, with (6.191)

we immediately obtain
b> x
p = λb = b. (6.192)
kbk2
We can also compute the length of p (i.e., the distance from 0) we can deter-
mine by means of Definition 22:
|b> x| kbk
kpk = kλbk = |λ| kbk = 2
kbk = | cos ω| kxk kbk = cos ω kxk . (6.193)
kbk kbk2
Here, ω is the angle between x and b. This equation should be familiar from
trigonometry: If kxk = 1 it lies on the unit circle. Then the projection onto the
horizontal axis b is exactly cos ω. An illustration is given in Figure 6.8(b)
3. Finding the projection matrix P π . We know that a projection is a linear map-

ping (see Definition 27). Therefore, there exists a projection matrix P π , such
that πU (x) = p = P π x. With
b> x bb>
p = λb = b = x (6.194)
kbk2 kbk2
122
we immediately see that
bb>
Pπ= . (6.195)
kbk2
Note that bb> is a matrix (with rank 1) and kbk2 = hb, bi is a scalar.
The projection matrix P π projects any vector x onto the line through the origin with
direction b (equivalently, the subspace U spanned by b).
Example
h i>
Find the projection matrix P π onto the line through the origin and b = 1 2 2 .
b is a direction and a basis of the one-dimensional subspace (line through origin).
With (6.195), we obtain
1 i 1 1 2 2
   
bb> 1   h
P π = > = 2 1 2 2 = 2 4 4 . (6.196)
b b 9 2  9
2 4 4

Let us nowh choose ia particular x and see whether it lies in the subspace spanned by
>
b. For x = 1 1 1 , the projected point is
1 2 2 1  5  1

      
1  1
p = P π x = 2 4 4 1 = 10 ∈ [2] . (6.197)

 
 
    
9   9    
2 4 4 1 10 2
Note that the application of P π to p does not change anything, i.e., P π p = p.41 This
is expected because according to Definition 27 we know that a projection matrix P π
satisfies P 2π x = P π x.
6.16.2 Projection onto General Subspaces

In the following, we generalize projections to the case where we project vectors
x ∈ Rn onto general subspaces U ⊂ Rn . Assume that (b1 , . . . , bm ) is an ordered basis
of U .42
Projections πU (x) onto U exhibit the property that they can be represented as linear
combinations of the basis vectors, i.e., a projected point is mapped onto the subspace
spanned by the columns of B ∈ Rn×m , where B = (b1 |...|bm ). Projections πU (x) ∈ U
of x onto the subspace U spanned by (b1 , . . . , bmP) can be expressed as linear combi-
nations of the basis vectors of U , such that p = m i=1 λi bi .
As before, we follow a three-step procedure to find p and the projection matrix P π :
41 pis therefore an eigenvector of P π , and the corresponding eigenvalue is 1.
42 IfU is given by a set of spanning vectors, make sure you determine a basis b1 , . . . , bm before
proceeding.
123
6.16. Projections Chapter 6. Linear Algebra
1. Find λ1 , . . . , λm , such that the linear combination p = m

P
i=1 λi bi = Bλ is closest
to x ∈ Rn . As in the 1D case, “closest” means “minimum distance”, which
implies that the line connecting p ∈ U and x ∈ Rn must be orthogonal to all
basis vectors of U . Therefore, we obtain the conditions (assuming the standard
scalar product)
hb1 , x − pi = b>
1 (x − p) = 0 (6.198)
..
. (6.199)
>
hbm , x − pi = bm (x − p) = 0 (6.200)
which, with p = Bλ, can be written as
b>
1 (x − Bλ) = 0 (6.201)
..
. (6.202)
>
bm (x − Bλ) = 0 (6.203)
such that we obtain a homogeneous linear equation system

 > 
b1  

 ..  x − Bλ

 .    = 0 ⇔ B > (x − Bλ) = 0 . (6.204)
 >   
bm
Factorizing yields
B > (x − Bλ) = 0 ⇔ B > Bλ = B > x , (6.205)
and the expression on the right-hand side is called normal equation.43 Since
the vectors b1 , . . . , bm are a basis and, therefore, linearly independent, B > B is
regular and can be inverted. This allows us to solve for the optimal coefficients
λ = (B > B)−1 B > x . (6.206)
The matrix (B > B)−1 B > is often called the pseudo-inverse of B, which can be
computed for non-square matrices B. It only requires that B > B is positive
definite. In practical applications (e.g., linear regression), we often add I to
B > B to guarantee positive definiteness or increase numerical stability. This
“ridge” can be rigorously derived using Bayesian inference.44
2. Find the projection of πU (x) = p ∈ U . We already established that p = Bλ.

Therefore, with (6.206)
p = B(B > B)−1 B > x . (6.207)

43 You may see the normal equation again when you take courses on machine learning, state esti-
mation or robotics.
44 For more details, please see CO-493 or other machine learning courses.
124
3. Find the projection matrix P π . From (6.207) we can immediately see that the
projection matrix that solves P π x = p must be
P π = B(B > B)−1 B > . (6.208)
Remark 38
Comparing the solutions for projecting onto a line (1D subspace) and the general case,
we see that the general case includes the line as a special case: If dim(U ) = 1 then B > B
is just a scalar and we can rewrite the projection matrix in (6.208) P π = B(B > B)−1 B >
>
as P π = BB
B>B
, which is exactly the projection matrix in (6.195).
Example
1 0 6

     
For a subspace U = [1 , 1] ⊂ R3 and x = 0 ∈ R3 find λ, the projection point p
     
1 2 0
     
and the projection matrix P π .
First, we see that the generating set of U is a basis (linear independence) and write
1 0
the basis vectors of U into a matrix B = 1 1.
 
1 2
 
Second, we compute the matrix B > B and the vector B > x as
" # 1 0 " # " # 6 " #
1 1 1  3 3 1 1 1   6
B>B = B>x =

1 1 = 0 = (6.209)
0 1 2  3 5 0 1 2   0
1 2 0

Third, we solve the normal equation B > Bλ = B > x to find λ:

" #" # " # " #
3 3 λ1 6 5
= ⇒λ= (6.210)
3 5 λ2 0 −3
Fourth, the projection πU (x) = p of x onto U , i.e., into the column space of B, can
be directly computed via
 5 
 
p = Bλ =  2  (6.211)
 
−1
 
h i> √
The corresponding projection/reconstruction error is kx−pk = 1 −2 1 = 6

Fifth, the general projection matrix (for and x ∈ R3 ) is given by
5 2 1
 
1 
 
P π = B(B > B)−1 B > =  2 2 2 (6.212)

6
−1 2 5

To verify the results, we can (a) check whether the error vector p − x is orthogonal
to all basis vectors of U , (b) verify that P π = P 2π (see Definition 27).
Remark 39
In vector spaces with non-standard scalar products, we have to pay attention when
computing angles and distances, which are defined by means of the scalar product.
125
6.17. Rotations Chapter 6. Linear Algebra
Figure 6.9: The robotic arm needs to rotate its joints in order to pick up objects or to
place them correctly. Figure taken from (Deisenroth et al., 2015).
6.16.3 Applications
Projections are often used in computer graphics, e.g., to generate shadows, see Fig-
ure 6.7. In optimization, orthogonal projections are often used to (iteratively) min-
imize residual errors. This also has applications in machine learning, e.g., in linear
regression where we want to find a (linear) function that minimizes the residual er-
rors, i.e., the lengths of the orthogonal projections of the data onto the line (Bishop,
2006). PCA (Hotelling, 1933) also uses projections to reduce the dimensionality
of high-dimensional data: First, PCA determines an orthogonal basis of the data
space. It turns out that this basis is the eigenbasis of the data matrix. The impor-
tance of each individual dimension of the data is proportional to the correspond-
ing eigenvalue. Finally, we can select the eigenvectors corresponding to the largest
eigenvalues to reduce the dimensionality of the data, and this selection results in
the minimal residual error of the data points projected onto the subspace spanned
by these principal components. Figure 6.4 illustrates the projection onto the first
principal component for a two-dimensional data set.
6.17 Rotations
An important category of linear mappings are the rotations. A rotation is acting
to rotate an object (counterclockwise) by an angle θ about the origin. Important
application areas of rotations include computer graphics and robotics. For example,
it is often important to know how to rotate the joints of a robotic arm in order to
pick up or place an object, see Figure 6.9.
( " # " #)
2 1 0
Consider the standard basis in R given by e1 = , e2 = . Assume we want
0 1
to rotate this coordinate system by an angle θ as illustrated in Figure 6.10. Since
rotations Φ are linear mappings, we can express them by a transformation matrix
RΦ (θ). Trigonometry allows us to determine the coordinates of the rotated axes with
126
Chapter 6. Linear Algebra 6.17. Rotations
Φ(e2 ) = [− sin θ, cos θ]T

cos θ
e2
Φ(e1 ) = [cos θ, sin θ]T

sin θ
θ
− sin θ e1 cos θ
Figure 6.10: Rotation of the standard basis in R2 by an angle θ
e2
θ e1
e3
Figure 6.11: Rotation of a general vector (black) in R3 by an angle θ about the e2 -axis.
The rotated vector is shown in blue.
respect to the standard basis in R2 . We obtain

" # " #
cos θ − sin θ
Φ(e1 ) = , Φ(e2 ) = . (6.213)
sin θ cos θ
Therefore, the transformation matrix RΦ (θ) is given as
" #
cos θ − sin θ
RΦ (θ) = (6.214)
sin θ cos θ
Generally, in R2 rotations of any vector x ∈ R2 by an angle θ are given by
" #" # " #
0 cos θ − sin θ x1 x1 cos θ − x2 sin θ
x = Φ(x) = RΦ (x) = = . (6.215)
sin θ cos θ x2 x1 cos θ + x2 sin θ
In R3 there are three (planar) rotations about three standard basis vectors (see Fig-
ure 6.11):
• Anti-clockwise rotation about the e1 -axis
cos θ − sin θ 0
 
R1 =  sin θ cos θ 0 (6.216)
 
0 0 1
 
127
6.18. Vector Product (Cross Product) Chapter 6. Linear Algebra

 cos θ 0 sin θ 
 
R2 =  0 1 0  (6.217)

− sin θ 0 cos θ
 

1 0 0 
 
R3 = 0 cos θ − sin θ  (6.218)
 
0 sin θ cos θ
 
Remark 40
• Composition of rotations sums their angles:
Rφ Rθ = Rφ+θ (6.219)
• kxk = kRΦ xk, i.e., the original vector and the rotated vector have the same length.
• This implies that det(RΦ ) = 1
• x and x0 are separated by an angle θ.
• Only in two dimensions vector rotations are commutative (R1 R2 = R2 R1 ) and
form an Abelian group (with multiplication) if they rotate about the same point
(e.g., the origin).
• Rotations in three dimensions are generally not commutative. Therefore, the order
in which rotations are applied is important, even if they rotate about the same
point.
• Rotations have no real eigenvalues.
6.18 Vector Product (Cross Product)

The vector product is usually defined in three-dimensional vector spaces. Therefore,
we will focus on R3 in the following.
Definition 28 (Vector product)
Let V ⊂ R3 be a vector space and (e1 , e2 , e3 ) be a fixed orthonormal45 basis of V , e.g.,
the standard basis. For x = λ1 e1 + λ2 e2 + λ3 e3 and y = ψ1 e1 + ψ2 e2 + ψ3 e3 , where
λi , ψi ∈ R, then the vector/cross product of x and y is defined as
x × y := (λ2 ψ3 − λ3 ψ2 )e1 + (λ3 ψ1 − λ1 ψ3 )e2 + (λ1 ψ1 − λ2 ψ2 )e3 ∈ R3 . (6.220)
This can equivalently be written as

e1 e2 e3
λ2 λ3 λ1 λ3 λ1 λ2
x × y = λ1 λ2 λ3 = e − e + e (6.221)
ψ2 ψ3 1 ψ1 ψ3 2 ψ1 ψ2 3

ψ1 ψ2 ψ3

using the rules for computing determinants.
128
Chapter 6. Linear Algebra 6.18. Vector Product (Cross Product)
x×y
y kx × yk
0 x
Figure 6.12: Vector product. The vector product x×y is orthogonal to the plane spanned
by x, y, and its length corresponds to the area of the parallelogram having x and y as
sides.
Remark 41
• The vector product x × y of two linearly independent vectors x and y is a vector
that is orthogonal to both and, therefore, normal to the plane spanned by them
(see Fig. 6.12).
• The length of the vector product x × y corresponds to the (positive) area of the
parallelogram having x and y as sides (see Fig. 6.12), which is given as kx × yk =
kxk kyk sin θ, where θ is the angle between x, y.
• For a parallelepiped (3-dimensional parallelogram) with sides x, y, z, the vector

product can be used to compute the volume
V = |hx, y × zi| = |hy, z × xi| = |hz, x × yi| (6.222)
• Since the magnitude of the cross product goes by the sine of the angle between its
arguments, the cross product can be thought of as a measure of orthogonality in
the same way that the scalar product is a measure of parallelism (which goes by
the cosine of the angle between its arguments).
Remark 42
The vector/cross product possesses the following properties:
• Anti-commutative: x × y = −y × x
• Distributive over addition: x × (y + z) = x × y + x × z
• (λx) × y = x × (λy) = λ(x × y) for λ ∈ R
• Grassmann identity: x × (y × z) = hx, ziy − hx, yiz
• Jacobi identity: x × (y × z) + y × (z × x) + z × (x × y) = 0
45 All vectors are orthogonal and have unit length k · k = 1.
129
6.18. Vector Product (Cross Product) Chapter 6. Linear Algebra
• The cross product is not associative, i.e., (x × y) × z , x × (y × z)
• hx × y, xi = hx × y, yi = 0
• x, y are linearly dependent if and only if x × y = 0
130
Bibliography
Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Information Science

and Statistics. Springer-Verlag. pages 126
Box, G. E. P. and Muller, M. E. (1958). A Note on the Generation of Random Normal

Deviates. Annals of Mathematical Statistics, 29(2):610–611. pages 41
Deisenroth, M. P., Fox, D., and Rasmussen, C. E. (2015). Gaussian Processes for
Data-Efficient Learning in Robotics and Control. IEEE Transactions on Pattern Anal-
ysis and Machine Intelligence, 37(2):408–423. pages 126
Deng, L., Seltzer, M. L., Yu, D., Acero, A., Mohamed, A.-r., and Hinton, G. E. (2010).
Binary Coding of Speech Spectrograms using a Deep Auto-Encoder. In Interspeech,
pages 1692–1695. pages 120
Drumm, V. and Weil, W. (2001). Lineare Algebra und Analytische Geometrie. Lec-
ture Notes, Universität Karlsruhe. pages 60
Hogben, L., editor (2013). Handbook of Linear Algebra. Discrete Mathematics and
Its Applications. Chapman and Hall, 2nd edition. pages 60
Hotelling, H. (1933). Analysis of a Complex of Statistical Variables into Principal

Components. Journal of Educational Psychology, 24:417–441. pages 107, 120,
126
Knop, R. (1969). Remark on Algorithm 334 [G5]: Normal Random Deviates. Com-
munications of the ACM, 12(5):281. pages 41
Rasmussen, C. E. and Williams, C. K. I. (2006). Gaussian Processes for Machine Learn-

ing. Adaptive Computation and Machine Learning. The MIT Press, Cambridge,
MA, USA. pages 115
Schölkopf, B. and Smola, A. J. (2002). Learning with Kernels—Support Vector Ma-

chines, Regularization, Optimization, and Beyond. Adaptive Computation and Ma-
chine Learning. The MIT Press, Cambridge, MA, USA. pages 115
Schölkopf, B., Smola, A. J., and Müller, K.-R. (1998). Nonlinear Component Analysis
as a Kernel Eigenvalue Problem. Neural Computation, 10(5):1299–1319. pages
115
131
Bibliography BIBLIOGRAPHY
Stoer, J. and Burlirsch, R. (2002). Introduction to Numerical Analysis. Springer. pages

115
Strang, G. (2003). Introduction to Linear Algebra. Wellesley-Cambridge Press, 3rd

edition. pages 60
Wikipedia (2015). Linear Dependence. https://en.wikipedia.org/wiki/Linear independence.

pages 80
132

Math Methods

Uploaded by

Copyright:

Available Formats

Math Methods

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Math Methods

Uploaded by

Copyright:

Available Formats

L ECTURE N OTES

Mathematical Methods (CO-145)

May 10, 2016

4 Differential Equations and Calculus 31

4.2.2 Partial Differentiation . . . . . . . . . . . . . . . . . . . . . . 36

6.3.4 Compact Representations of Linear Equation Systems . . . . . 67

6.18 Vector Product (Cross Product) . . . . . . . . . . . . . . . . . . . . . 128

1.1 The Convergence Definition

Definition A sequence an for n ≥ 1 converges to a limit l ∈, written an → l as

1.2 Illustration of Convergence

Figure 1.1: The convergence of the sequence an = 1/n for n ≥ 1 tending to 0

A demonstration of the convergence of an = 1/n for n ≥ 1 is shown in Figure 1.1.

1.3 Common Converging Sequences

1.4 Combinations of Sequences

1. lim (λan ) = λa for some real constant λ

1.5 Sandwich Theorem

Figure 1.2: Demonstration of the Sandwich theorem. The sequence an is sandwiched

Sandwich Theorem Given sequences un and ln for n ≥ 1, where both un → l and

ln ≤ an ≤ un for all n ≥ N for some N ∈ N

Demonstration We provide an illustration of the sandwich theorem in Figure 1.2.

|ln − l| <  and |un − l| < 

l −  < l n < un < l + 

This allows us to say that an converges to l as a result.

1.6 Ratio Tests for Sequences

Example of convergence ratio test As a simple

If r < 1 the an sequence converges to 0, and if r > 1, it diverges

Example of limit ratio test

1.7 Proof of Ratio Tests

|an | ≤ c|an−1 | ≤ c2 |an−2 | · · · ≤ cn−N |aN |

|an | ≤ cn−N |aN |

r −  < bn < r +  (∗)

Taking the right hand side, we see that bn < r+1

1.8 Useful Techniques for Manipulating Absolute Val-

|x| < a ⇔ −a < x < a

1. |xy| = |x| × |y|

1.9 Properties of Real Numbers

1. u is an upper bound on S if u ≥ s for any s ∈ S

2. l is an lower bound on S if l ≤ s for any s ∈ S

3. we say that S is bounded above if S has an upper bound

4. we say that S is bounded below if S has a lower bound

5. we say that S is bounded if S has an upper and lower bound

1. we call sup(S) the least upper bound of S

2. we call inf(S) the greatest lower bound of S

for some real sequence an for n ≥ 1.

Sn itself is a sequence and can be treated as such. A series converges or diverges if

Series terms tend to 0 For a series S = ∞

The tail of a series It is important to understand that, as with sequences, the

2.1 Geometric Series

2.2 Harmonic Series

It is an important result to show that it diverges to ∞.

So considering partial sums, we get that

2.3 Series of Inverse Squares

2.4 Common Series and Convergence

Diverging series It should be noted that if S = ∞

2.5 Convergence Tests

Comparison test Let λ > 0 and N ∈ N. Then

There are some simple examples of series, ∞

rectangular areas for n ≥ 1. When displayed like this, and

2.6 Absolute Convergence

|ln − l| < and |un − l| <

l − < l n < un < l +

r − < bn < r + (∗)