Introduction to Applied Algebraic

Topology

Tom Needham

Last Updated: December 10, 2017

 These are lecture notes for the course MATH 4570 at the Ohio State University.
They are a work in progress and certainly contain mistakes/typos. Please contact need-
[email protected] to report any errors or to make comments.

Work on these notes was supported by the NSF RTG grant Algebraic Topology and
Its Applications, # 1547357.

2
Contents
1 Preliminaries 6
1.1 Basic Set Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.1 Set Theoretic Notation . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.2 Combining Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.3 Sets of Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.4 Functions on Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Infinite Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.1 Countable and Uncountable Sets . . . . . . . . . . . . . . . . . . 9
1.2.2 Arbitrary Unions and Intersections . . . . . . . . . . . . . . . . . 10
1.3 Equivalence Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2 Review of Linear Algebra 13

2.1 Abstract Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.1 Vector Spaces over R . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.2 Vector Spaces over Arbitrary Fields . . . . . . . . . . . . . . . . . 15
2.2 Basis and Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.1 Basis of a Vector Space . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.2 Dimension of a Vector Space . . . . . . . . . . . . . . . . . . . . . 18
2.3 Linear Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.1 Abstract Linear Transformations . . . . . . . . . . . . . . . . . . 20
2.3.2 Linear Transformations of Finite-Dimensional Vector Spaces . . . 23
2.3.3 Determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4 Vector Space Constructions . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.1 Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.2 Special Subspaces Associated to a Linear Transformation . . . . . 27
2.4.3 Rank and Nullity . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4.4 Direct Sums . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.5 Quotient Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4.6 Row and Column Operations . . . . . . . . . . . . . . . . . . . . 30
2.4.7 Vector Space Associated to a Linear Map . . . . . . . . . . . . . . 30
2.5 Structures on Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5.1 Inner Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5.2 Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

3
3 Metric Space Topology 35
3.1 Metric Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1.1 Definition of a Metric Space . . . . . . . . . . . . . . . . . . . . . 35
3.1.2 Examples of Metric Spaces . . . . . . . . . . . . . . . . . . . . . . 36
3.1.3 Open and Closed Sets . . . . . . . . . . . . . . . . . . . . . . . . 39
3.1.4 Topological Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.1.5 Limit Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2 Continuous Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3 Topological Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3.1 Compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3.2 Connectedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.4 Equivalence Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.4.1 Isometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.4.2 Homeomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.4.3 Homotopy Equivalence . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

4 Homology of Simplicial Complexes 53

4.1 Motivation: Distinguishing Topological Spaces . . . . . . . . . . . . . . . 53
4.2 Simplicial Complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.2.1 Geometric Simplicial Complexes . . . . . . . . . . . . . . . . . . . 55
4.2.2 Abstract Simplicial Complexes . . . . . . . . . . . . . . . . . . . . 58
4.3 Topological Invariants of Simplicial Complexes . . . . . . . . . . . . . . . 59
4.3.1 Connected Components . . . . . . . . . . . . . . . . . . . . . . . 59
4.3.2 Back to Linear Algebra: Free Vector Spaces . . . . . . . . . . . . 60
4.3.3 First Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.3.4 Second Example . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.4 Homology of Simplicial Complexes over F2 . . . . . . . . . . . . . . . . . 66
4.4.1 Chain Complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.5 More Advanced Topics in Homology . . . . . . . . . . . . . . . . . . . . . 69
4.5.1 Variants of Homology . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.5.2 Functoriality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

5 Persistent Homology 75
5.1 Vietoris-Rips Complex . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.1.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.1.2 A Simple Example . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.1.3 Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.1.4 Other Complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.2 Linear Algebra with Persistence Vector Spaces . . . . . . . . . . . . . . . 78
5.2.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.2.2 Matrix Representations of Linear Maps of Persistence Vector Spaces 80
5.3 Persistence Homology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

4
 5.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.4.1 Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.4.2 Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

6 Representions of Persistent Homology 83

6.1 Structure Theorem for Persistence Vector Spaces . . . . . . . . . . . . . . 83
6.1.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.2 Main Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.3 Barcodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.3.1 Examples of Barcodes . . . . . . . . . . . . . . . . . . . . . . . . 89
6.4 Persistence Diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.4.1 Examples of Persistence Diagrams . . . . . . . . . . . . . . . . . . 89
6.5 Computing Barcodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.6 Other Representations of Persistent Homology . . . . . . . . . . . . . . . 89
6.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

7 Metrics on the Space of Barcodes (Under Construction) 90

8 Applications (Under Construction) 91

9 Appendix 92
9.1 Every Vector Space Has a Basis . . . . . . . . . . . . . . . . . . . . . . . 92

Bibliography 94

5
1 Preliminaries
In this chapter we review some basic notions of set theory and equivalence relations.
The reader is presumably familiar with these concepts, so this chapter should be treated
mainly as a refresher and to fix notation.

1.1 Basic Set Theory

1.1.1 Set Theoretic Notation
A set is a collection of elements. We are taking the naive view of set theory in assuming
that such a definition is intuitively clear and proceeding from there.
We generally use capital letters A, B, X, Y, etc. to denote sets and lower case letters
a, b, x, y, etc. to denote their elements. We use a P A to denote that the element a belongs
to the set A. The expression a R A means that a is not an element of A. Two sets A
and B are called equal if they contain exactly the same elements, in which case we write
A “ B. The contents of a set are specified by listing them or using set builder notation,
as in the following examples.
Example 1.1.1. 1. A “ ta, b, cu denotes the set with three elements a, b and c.
2. Z “ t. . . , ´3, ´2, ´1, 0, 1, 2, 3, . . .u denotes the set of integers.
3. B “ tb P Z | b is evenu “ tb P Z | b “ 2k for some k P Zu denotes the set of even
integers.
4. H denotes the empty set containing no elements.
We note that set notation doesn’t account for multiplicity; that is, a set should not
include more than one copy of any particular element, e.g., ta, a, b, cu. The notion of a set
which respects multiplicity is called a multiset. Such objects will appear naturally later
in the text, so we will treat them when they arise. Also note that sets are not ordered;
for example, the set ta, b, cu is equal to the set tb, c, au.

1.1.2 Combining Sets

Let A and B be sets. The union of A and B is the set

A Y B “ tx | x P A or x P Bu.

The use of “or” here is non-exclusive. This means that the defining condition of A Y B
can be read as “x P A or x P B or x is in both A and B”.

6
 The intersection of A and B is the set

A X B “ tx | x P A and x P Bu.

The difference of A and B is the set

AzB “ tx | x P A and x R Bu.

The product of A and B is the set of ordered pairs px, yq such that x P A and y P B,
denoted
A ˆ B “ tpx, yq | x P A and y P Bu.
Example 1.1.2. Let A “ ta, b, c, du and B “ tb, d, x, yu. Then
1. A Y B “ ta, b, c, d, x, yu (note that b and d are not included twice!),
2. A X B “ tb, du,
3. AzB “ ta, cu,
4. AˆB “ tpa, bq, pa, dq, pa, xq, pa, yq, pb, bq, pb, dq, pb, xq, pb, yq, . . .u (there are 16 “ 4ˆ4
elements total in the set).
Example 1.1.3. For any set A,
1. A Y H “ A,
2. AzH “ A,
3. A X H “ H,
4. A ˆ H “ H.
The set A is called a subset of B if a P A implies a P B. This is denoted A Ă B.
Example 1.1.4. Let A “ ta, bu, B “ ta, b, c, du and C “ tbu. Then A Ă B but A Ć B
(this should be read as “A is not a subset of B”) because a P A but a R C, so the defining
implication fails.

1.1.3 Sets of Sets

The elements of a set can themselves be sets!
Example 1.1.5. Let B “ ttau, tbu, xu. Then the elements of B are tau, tbu and x. Let
A “ ta, bu. Then
A Y B “ ta, b, tau, tbu, xu.
Note that this doesn’t contradict our convention that a set can’t contain multiple copies
of the same element, since a and tau represent different objects. We also have

A X B “ H,

since, for example, a and tau are different elements.

7
 The power set of A, denoted PpAq is the set of all subsets of A.

Example 1.1.6. Let A “ ta, b, cu. Then

PpAq “ tH, A, tau, tbu, tcu, ta, bu, ta, cu, tb, cuu.

The cardinality of a set A is the number of elements in A. If A contains infinitely many

elements, we say that its cardinality is infinity and that A is an infinite set. Otherwise
we say that A is a finite set. We denote the cardinality of A by |A|.

Proposition 1.1.1. If A is a finite set, then so is PpAq and

|PpAq| “ 2|A| .

Proof. Let A “ ta1 , . . . , an u, so that |A| “ n. To form a subset B of A, we have to make

the binary choice of whether or not to include each aj in B. There are n such choices to
make and they are independent of one another, so

|PpAq| “ 2n “ 2|A| .

1.1.4 Functions on Sets

A function from the set A to the set B is a subset f Ă AˆB such that each a P A appears
in exactly one ordered pair pa, bq P f . The more typical notation used for a function is
f : A Ñ B, with f paq “ b denoting pa, bq P f . This is a precise way to say that f maps
each element of A to exactly one element of B. The set A is called the domain of f and
B is called the range B of f .

Example 1.1.7. 1. All of the usual functions from Calculus are functions in this
sense. For example the function f pxq “ x2 should be thought of as the function
f : R Ñ R with px, x2 q P f Ă R ˆ R.

2. Let A “ ta, b, cu and B “ tx, yu. Then the set f “ tpa, xq, pb, xq, pc, yqu Ă A ˆ B
defines a function f : A Ñ B. In this case, we would write f paq “ x, f pbq “ x and
f pcq “ y.

A function f : A Ñ B is called injective (or one-to-one) if f paq “ f pa1 q if and only

if a “ a1 . It is called surjective (or onto) if for all b P B there exists a P A such that
f paq “ b. The function is called bijective if it is both injective and surjective. An inverse
to f is a function f ´1 : B Ñ A such that f ´1 pf paqq “ a for all a P A and f pf ´1 pbqq “ b
for all b P B.

Proposition 1.1.2. A function f : A Ñ B has an inverse if and only if f is a bijection.

If an inverse for f exists, then it is unique.

8
Proof. If f is a bijection, define f ´1 : B Ñ A by f ´1 pbq “ a, where a P A satisfies
f paq “ b. Such an a must exist by surjectivity and it is unique by injectivity. In fact, if
an inverse for f exists, then it must be exactly of this form and this shows that inverses
are unique. Conversely, suppose that f ´1 exists. Let a, a1 P A satisfy f paq “ f pa1 q. Then
f ´1 pf paqq “ f ´1 pf pa1 qq and this implies a “ a1 , thus f is injective. To show surjectivity,
let b P B. Then a “ f ´1 pbq satisfies f paq “ f pf ´1 pbqq “ b.

1.2 Infinite Sets

1.2.1 Countable and Uncountable Sets
Let R denote the set of real numbers. The sets Z and R both have cardinality infinity,
but they feel different in the sense that Z is “discrete” while R is “continuous”. We will
make this difference precise in this section.
Let N “ t1, 2, 3, . . .u denote the natural numbers . A set A is called countable if there
exists an injective function f : A Ñ N and countably infinite if it is countable and has
infinite cardinality. If a set is not countable, then we call it uncountable .
Example 1.2.1. 1. Z is countably infinite. To see this, define f : Z Ñ N by
"
2k if k ą 0
f pkq “
´2k ` 1 if k ď 0.

Then it is easy to check that f is injective.

2. The set Q of rational numbers is countably infinite. We leave this as an exercise.

Lemma 1.2.1. If a set A is countable, then there exists a surjective map g : N Ñ A.
Proof. Assume that A is countably infinite (if |A| is finite, the existence of such a map
g is obvious). Let f : A Ñ N be an injection and let B “ tf paq | a P Au denote the
image of f . For each b P B, there exists a unique (by injectivity) element a P A such that
f paq “ b; we denote this element by f ´1 pbq. We fix any element b0 P B and define a map
g : N Ñ A by " ´1
f pkq if k P B
gpkq “
f ´1 pb0 q if k R B.
Then g is a well-defined, surjective function.
Remark 1.2.2. The converse of this lemma is also true, but requires the Axiom of Choice.
We wish to avoid treating the Axiom of Choice for now, but the interested reader is invited
to read the Appendix for a short discussion of one of its important consequences.
Let S denote the set of ordered sequences of 1’s and 0’s. That is, elements of S are of
the form p1, 1, 1, 1, . . .q or p0, 0, 0, 0, . . .q or p1, 0, 1, 1, 0, 0, 0, 1, 0, 1, . . .q, etc.
Theorem 1.2.3. The set S is uncountable.
The proof of the theorem uses a technique called Cantor’s Diagonal Argument.

9
Proof. Let g : N Ñ S be any function. We wish to show that g is not surjective. Since g
was arbitrary, it follows from Lemma 1.2.1 that S is uncountable. We list the values of
g as

gp1q “ pa11 , a12 , a13 , . . .q,

gp2q “ pa21 , a22 , a23 , . . .q,
gp3q “ pa31 , a32 , a33 , . . .q,
..
.

with each akj P t0, 1u. We define an element s P S by

s “ pa11 ` 1, a22 ` 1, a33 ` 1, . . .q,

where we add mod 2, i.e., 0 ` 1 “ 1 and 1 ` 1 “ 0. Then for all k P N,

gpkq “ pak1 , ak2 , . . . , akk´1 , akk , akk`1 , . . .q ‰ pa11 `1, a22 `1, . . . , ak´1 k k`1
k´1 `1, ak `1, ak`1 `1, . . .q “ s,

because gpkq and s differ in their k-th entry. It follows that g is not surjective.
An essentially straightforward corollary is left to the reader:
Corollary 1.2.4. The set R of real numbers is uncountable.

1.2.2 Arbitrary Unions and Intersections

We will frequently need to consider infinite collections of sets. We use the notation

U “ tUα uαPA .

Each Uα is a set, α is an index for the set, and A is an indexing set. We can consider
unions and intersections of sets in this collection, which are denoted, respectively, by
ď č
Uα and Uα .
αPA αPA

Example 1.2.2. Let Un denote the interval p1{n, 1s Ă R, where n is a natural number.
We can consider the collection
U “ tUn unPN .
The union of the elements of this collection is
ď
Un “ p0, 1s.
nPN

To see this, note that any element r of the union must be an element of some p1{n, 1s Ă
p0, 1s, so the union is a subset of p0, 1s. On the other hand, for every r P p0, 1s, there
exists some n P N such that 1{n ă r and it follows that r P Un , so that r is an element
of the union.

10
 The intersection of the elements of this collection is
č
Un “ t1u.
nPN

1.3 Equivalence Relations

Let S be a set. A binary relation on S is a subset R Ă S ˆ S. We typically use the
notation x „ x1 or x „R x1 to indicate that px, x1 q P R. A binary relation R is an
equivalence relation if the following conditions hold:

1. (Reflexivity) s „ s for all s P S

2. (Symmetry) s „ s1 if and only if s1 „ s

3. (Transitivity) if s „ s1 and s1 „ s2 , then s „ s2 .

Example 1.3.1. Consider the set R of real numbers. The most obvious equivalence
relation on R is equality; that is x „ y if and only if x “ y. Another equivalence relation
„2 is defined by x „2 y if and only if x ´ y is an integer multiple of 2. We can define
a similar equivalence relation „r for any fixed r P R. You will examine some other
equivalence relations on R in the exercises.

Let „ be a fixed equivalence relation on a set S. The equivalence class of s P S, denoted

rss, is the set
ts1 P S | s1 „ su.
We denote the set of all equivalence classes of S by S{ „. That is,

S{ „“ trss | s P Su.

Example 1.3.2. Consider the equivalence relation „2 restricted to the set of integers Z;
that is, integers a and b satisfy a „2 b if and only if a ´ b is an integer multiple of 2. Then
the set of equivalence classes Z{ „2 contains exactly two elements r0s and r1s. Indeed,
for any even integer 2k, 2k „2 0 so that r2ks “ r0s. Likewise, for any odd integer 2k ` 1,
r2k ` 1s “ r1s.
For the equivalence relation „2 on all of R, the set of equivalence classes R{ „2 is in
bijective correspondence with the interval r0, 2q. This is the case because for any real
number x, there is a unique y P r0, 2q such that x ´ y is an integer multiple of 2. To see
this, note that the set YkPZ r0 ` k, 2 ` kq is a partition of R, so there exists a unique k P Z
such that x P r0 ` k, 2 ` kq, and we define y “ x ´ 2k.

1.4 Exercises
1. Show that the set of rational numbers Q is uncountable by finding an injective map
Q Ñ N.

11
2. Show that a subset of a countable set must be countable.

3. Show that if there is a bijection between sets A and B, then A is countable if and
only if B is countable.

4. Prove that R is uncountable. One suggested strategy is to show that there is a

bijection from the set S (from Theorem 1.2.3) to the interval p0, 1q and to then
apply Theorem 1.2.3 and the results of the previous two exercises.

5. Let r P R be a fixed real number. Define a binary relation „r on R by declaring

x „r y if and only if x´y is an integer multiple of r. Show that „r is an equivalence
relation.

6. Show that ď is not and equivalence relation on R.

7. Let Mn denote the set of n ˆ n matrices with real entries. Define a binary relation
„ on Mn by declaring A „ B if and only if A “ B T , where the superscript denotes
matrix transpose. Show that „ is an equivalence relation.

12
2 Review of Linear Algebra
In this chapter we review the basic concepts of linear algebra. A strong grasp of abstract
linear algebra will be essential for the latter material in these notes. For a more in-depth
treatment of linear algebra, see, for example, [3, 4].

2.1 Abstract Vector Spaces

2.1.1 Vector Spaces over R
Definition of a Vector Space
A vector space over R is a set V together with an operation

`:V ˆV ÑV
pv1 , v2 q ÞÑ v1 ` v2

called vector addition and an operation

¨:RˆV ÑV
pλ, vq ÞÑ λ ¨ v

called scalar multiplication such that the following axioms are satisfied:

1. (Additive Associativity) For any elements v1 , v2 and v3 in V ,

pv1 ` v2 q ` v3 “ v1 ` pv2 ` v3 q.

2. (Scalar Multiple Associativity) For any λ1 and λ2 in R and any v in V ,

λ1 ¨ pλ2 ¨ vq “ pλ1 λ2 q ¨ v.

3. (Additive Commutativity) For any v1 and v2 in V ,

v1 ` v2 “ v2 ` v1 .

4. (Additive Identity) There exists an element 0V P V called the additive identity such
that for any v P V ,
v ` 0V “ v.

13
 5. (Additive Inverse) For any v P V , there exists an element ´v P V called the additive
inverse of v such that
v ` p´vq “ 0V .

6. (Distributive Law I) For any λ P R and v1 and v2 in V ,

λ ¨ pv1 ` v2 q “ λ ¨ v1 ` λ ¨ v2 .

7. (Distributive Law II) For any λ1 and λ2 in R and any v P V ,

pλ1 ` λ2 q ¨ v “ λ1 ¨ v ` λ2 ¨ v.

8. (Scalar Multiple Identity) For any v in V ,

1 ¨ v “ v.

Examples of Vector Spaces over R

Example 2.1.1. The example of a vector space over R that you are probably most
familiar working with is Rn (for some positive integer n). That is, Rn is the set of
n-tuples of real numbers px1 , x2 , . . . , xn q, vector addition is given by

px1 , x2 , . . . , xn q ` py1 , y2 , . . . , yn q “ px1 ` y1 , x2 ` y2 , . . . , xn ` yn q

and scalar multiplication is given by

λ ¨ px1 , x2 , . . . , xn q “ pλx1 , λx2 , . . . , λxn q.

It is an easy but instructive exercise to check that these operations satisfy the axioms of
a vector space over R.

Familiarity with Rn gives good intuition for working with abstract vector spaces, but
the material for this course will require us to work with more exotic vector spaces.

Example 2.1.2. Let Pn pRq denote the set of degree-n polynomials with real coefficients.
Any element of Pn pRq can be written in the form an xn ` an1 xn´1 ` ¨ ¨ ¨ ` a1 x ` a0 for some
ai P R. The set Pn pRq forms a vector space with vector addition

pan xn ` an´1 xn´1 ` ¨ ¨ ¨ ` a0 q ` pbn xn ` bn´1 xn´1 ` ¨ ¨ ¨ ` b0 q

“ pan ` bn qxn ` pan´1 ` bn´1 qxn´1 ` ¨ ¨ ¨ ` pa0 ` b0 q

and scalar multiplication

λ ¨ pan xn ` an´1 xn´1 ` ¨ ¨ ¨ ` a0 q “ pλan qxn ` pλan´1 qxn´1 ` ¨ ¨ ¨ ` pλa0 q.

In one of the exercises you will show that these operations on Pn pRq satisfy vector space
axioms. You may notice that the operations of Pn pRq share some similarity with those of

14
Rn`1 . Indeed, we will see later that Pn pRq and Rn`1 are actually “equivalent” as vector
spaces in a precise sense (to be defined in Section 2.3.1).

Example 2.1.3. As one last even more exotic example, consider the set of differentiable
functions

C 8 pr0, 1s, Rq “ tf : r0, 1s Ñ R | f is infinitely differentiableu.

We claim that this set forms a vector space with pointwise addition and scalar multipli-
cation. That is, for functions (vectors) f and g in C 8 pr0, 1s, Rq and a scalar λ, we define
the functions (vectors) f ` g and λf by

pf ` gqptq “ f ptq ` gptq

and
pλ ¨ f qptq “ λf ptq,
respectively. While Rn and Pn´1 (from the previous example) appear to be equivalent in
some way, C 8 pr0, 1s, Rq should feel “different”. In fact, Rn and Pn´1 are n-dimensional
and C 8 pr0, 1s, Rq is infinite-dimensional (dimension will be defined precisely in a couple
of sections), so the vector spaces are quite different. In this course we will primarily be
concerned with finite-dimensional vector spaces, but it is good to keep infinite-dimensional
spaces in mind (at least as motivation for the necessity of an abstract definition of vector
space!).

2.1.2 Vector Spaces over Arbitrary Fields

We will have need to consider a slightly more general notion of a vector space, where
scalars are not required to be elements of R, but of some field F.

Fields
A field is a set F endowed with operations ‚ and ` called multiplication and addition,
respectively, satisfying the following axioms for all a, b, c P F:

1. (Identities) There exists an additive identity denoted 0F such that a`0F “ a. There
also exists a multiplicative identity denoted 1F such that 1F ‚ a “ a.

3. (Associativity) Addition and multiplication are associative:

pa ` bq ` c “ a ` pb ` cq
pa ‚ bq ‚ c “ a ‚ pb ‚ cq.

4. (Commutativity) Addition and multiplication are commutative:

a`b“b`a
a ‚ b “ b ‚ a.

15
 5. (Inverses) Each a P F has additive inverse denoted ´a such that a ` p´aq “ 0F .
Each a P F besides 0F also has a multiplicative inverse denoted a´1 such that
a ‚ a´1 “ 1F .

6. (Distributivity) Multiplication distributes over addition:

a ‚ pb ` cq “ pa ‚ bq ` pa ‚ cq.

We have the following standard examples of fields.

Example 2.1.4. The real numbers R form a field with the obvious multiplication and
addition operations.

Example 2.1.5. The complex numbers C form a field with complex multiplication and
addition.

Our last basic example of a field will play a very important role in later chapters.

Example 2.1.6. Let F2 denote the field with two elements. As a set, F2 “ t0, 1u. The
addition and multiplication operations are described by the following tables.

+ 0 1 ‚ 0 1
0 0 1 0 0 0
1 1 0 1 0 1

You will verify that F2 is indeed a field in the exercises.

Vector Spaces over F

We then modify the definition of a vector space over R to obtain the more general notion
of a vector space over a field F. To be precise, a vector field over F is a set V together
with a vector addition operation and a scalar multiplication operation

¨:FˆV ÑV
pλ, vq ÞÑ λ ¨ v

satisfying the axioms:

1. (Additive Associativity) For any elements v1 , v2 and v3 in V ,

pv1 ` v2 q ` v3 “ v1 ` pv2 ` v3 q.

2. (Scalar Multiple Associativity) For any λ1 and λ2 in K and any v in V ,

λ1 ¨ pλ2 ¨ vq “ pλ1 λ2 q ¨ v.

16
 3. (Additive Commutativity) For any v1 and v2 in V ,

v1 ` v2 “ v2 ` v1 .

4. (Additive Identity) There exists an element 0V P V called the additive identity such
that for any v P V ,
v ` 0V “ v.

5. (Additive Inverse) For any v P V , there exists an element ´v P V called the additive
inverse of v such that
v ` p´vq “ 0V .

6. (Distributive Law I) For any λ P F and v1 and v2 in V ,

λ ¨ pv1 ` v2 q “ λ ¨ v1 ` λ ¨ v2 .

7. (Distributive Law II) For any λ1 and λ2 in F and any v P V ,

pλ1 ` λ2 q ¨ v “ λ1 ¨ v ` λ2 ¨ v.

8. (Scalar Multiple Identity) For any v in V ,

1F ¨ v “ v.

Example 2.1.7. We can modify our examples of vector spaces over R to get examples of
vector spaces over F. For example, Fn is the collection of n-tuples of elements of F. We
can similarly define Pn pFq to be the collection of degree-n polynomials with coefficients in
F. For F “ C we can likewise define an “infinite-dimensional” vector space C 8 pr0, 1s, Cq
over C.

2.2 Basis and Dimension

The vector space Rn of n-tuples of real numbers comes with a way to decompose its
elements in a canonical way. Let ej P Rn denote the n-tuple with a 1 in the jth entry
and zeros elsewhere; e.g., e1 “ p1, 0, 0, . . . , 0q. Then any element px1 , . . . , xn q P Rn can be
decomposed as
px1 , . . . , xn q “ x1 ¨ e1 ` x2 ¨ e2 ` ¨ ¨ ¨ ` xn ¨ en .
The set te1 , . . . , en u is called a basis for Rn . In this section we develop the notion of a
basis for an abstract vector space over a field F.

17
2.2.1 Basis of a Vector Space
Let S Ă V be a subset of a vector space V over the field F. A linear combination of
elements of S is an expression of the form
ÿ
λs s,
sPS

where each λs P F and only finitely many of them are nonzero. The span of S, denoted
spanF pSq (or simply spanpSq when the field F is clear), is the set of all linear combinations
of elements of S. The set S is called a spanning set for V if spanpSq “ V . The set is
called linearly dependent if there exist scalars λs P F, not all zero, such that
ÿ
λs s “ 0V .
sPS

If no such collection of scalars exist, the set is called linearly independent . A basis for
V is a linearly independent spanning set.
We have the following fundamental theorem.
Theorem 2.2.1. Every vector space V admits a basis. Moreover, any linearly indepen-
dent set S Ă V can be extended to a basis for V .
The proof of the theorem requires the Axiom of Choice. We will skip it for now, but
the interested reader is invited to read a proof sketch in Section 9.1.

2.2.2 Dimension of a Vector Space

Let B Ă V be a basis for a vector space V . We define the dimension of V to be the
number of elements in B. If B contains a finite number of elements n, we say that V is
n-dimensional and otherwise we say that V is infinite-dimensional —most of the vector
spaces that we will see in this course are finite-dimensional. We will use the notation
dimpV q for the dimension of the vector space V . In Proposition 2.2.4 below we will show
that our definition of dimension actually makes sense; that is, if we choose two different
bases for V then they will always have the same number of elements. To prove it, we will
need some lemmas. The first lemma tells us that if B is a basis for V , then any element
of V can be written as a linear combination of elements of B in a unique way.
Lemma 2.2.2. Let S be a linearly independent set. For any v P V , there exists at most
one way to write v as a linear combination of elements of S.
Proof. Assume that for some v P V there exist sets of scalars λs and µs such that
ÿ ÿ
v“ λs s “ µs s.
s s

This implies that

ÿ ÿ ÿ
0V “ v ´ v “ λs s ´ µs s “ pλs ´ µs qs.
s s s

18
Since S is linearly independent, this implies λs ´ µs “ 0F for all s, which implies in turn
that λs “ µs . We conclude that the two representations of v as a linear combination of
elements of S were actually the same to begin with.

Lemma 2.2.3. Let B Ă V be a basis for V . Then for any nonzero v P V zB, the set
B Y tvu is linearly dependent.

Proof. We can express v uniquely as a linear combination

ÿ
v“ λb b
bPB

for some scalars λb , not all of which are zero. Then we have a linear combination
ÿ
p´1qv ` λb b “ 0V ,
bPB

with not all coefficients equal to zero. It follows that B Y tvu is linearly dependent.

Proposition 2.2.4. The dimension of V is independent of choice of basis.

Proof. Let B and B 1 be bases for V . Our goal is to show that |B| “ |B 1 |. If |B| “ |B 1 | “ 8
then we are done, so let’s assume by way of obtaining a contradiction (and without loss
of generality) that |B| “ m and |B| ă |B 1 |. Write B “ tb1 , . . . , bm u and let b10 denote
some distinguished element of B 1 which is not an element of B (using our assumption on
the cardinalities of the sets). By Lemma 2.2.3, the set B Y tb10 u is linearly dependent, so
there exist scalars µj , j “ 1, . . . , m, and µb10 such that not all of them are zero and
ÿ
µb10 b10 ` µj bj “ 0V . (2.1)
j

Moreover, the linear independence of B implies that µb10 ‰ 0F .

Since B 1 is a basis, there exist scalars λjb1 such that
ÿ j
bj “ λb1 b1
b1 PB 1

for all j. Plugging this into (2.1), we have

˜ ¸
ÿ ÿ
0V “ µb10 b10 ` µj λjb1 b1 .
j b1 PB 1

Collecting terms, this can be rewritten as

˜ ¸ ˜ ¸ ˜ ¸
ÿ ÿ j
ÿ j
ÿ ÿ
0V “ µb10 b10 ` µj λb1 b1 “ µb10 ` µj λb1 b10 ` µj λjb1 b1 .
0
b1 PB 1 j j b1 PB 1 ztb10 u j

19
This gives a linear combination of the elements of B 1 which is equal to 0V such that not
all coefficients in the combination are zero. This is a contradiction to the assumption
that B 1 is a basis for V .
Example 2.2.1. 1. The set te1 , e2 , . . . , en u Ă Rn defined at the beginning of this
section is a basis for Rn , and is frequently referred to as the canonical basis for Rn .
The dimension of Rn is therefore equal to n.
2. In general dimpFn q “ n, as we would hope.
3. The vector space C 8 pr0, 1s, Rq is infinite-dimensional (see the exercises).
4. The vector space of polynomials Pn pRq defined in Example 2.1.2 has a basis given
by monomials t1 “ x0 , x “ x1 , x2 , x3 , . . . , xn u. It follows that the dimension of
Pn pRq is n ` 1.

2.3 Linear Transformations

2.3.1 Abstract Linear Transformations
Linear Transformations
Let V and W be vector spaces over the same field F. A linear transformation (also called
a linear map) from V to W is a function L : V Ñ W with the properties
1. Lpv1 ` v2 q “ Lpv1 q ` Lpv2 q for all v1 , v2 P V ,
2. Lpλvq “ λLpvq for all v P V and λ P F.
Put more simply, a linear map is just a map between vector spaces which preserves vector
space structure; that is, it takes addition to addition and scalar multiplication to scalar
multiplication.
Example 2.3.1. Let e1 , e2 , e3 denote the standard basis for R3 . Consider the linear map
L defined by
Lpe1 q “ 2e2 ` e3 , Lpe2 q “ e1 , Lpe3 q “ 0.
Note that we have only defined L explicitly on 3 elements of the vector space R3 . The
linear structure of L and the fact that the ej determine a basis for R3 allow us to extend
the map to all of R3 . Indeed, an arbitrary element v P R3 can be expressed as a sum

v “ λ1 e1 ` λ2 e2 ` λ3 e3

for some scalars λj . The linear structure of L allows us to evaluate Lpvq as

Lpvq “ λ1 Lpe1 q ` λ2 Lpe2 q ` λ3 Lpe3 q “ 2λ1 e2 ` λ1 e3 ` λ2 e1 .

As a concrete example, the vector v “ p1, 2, 3q “ e1 ` 2e2 ` 3e2 takes the value

Lpvq “ e2 ` e3 ` 2e1 “ p2, 1, 1q.

20
Linear Extensions
Let us expand on the observation in Example 2.3.1 that linear maps can be defined by
defining their values on basis elements.
Proposition 2.3.1. Let e1 , . . . , en be a fixed basis for the vector space V and let W be an
arbitrary vector space. Then for any choices of images Lpei q P W , there exists a unique
linear map L : V Ñ W which takes these values on the basis.
Proof. For any v P V , there is a unique representation of v as a linear combination

v “ λ1 e1 ` λ2 e2 ` ¨ ¨ ¨ ` λn en

for some scalars λj . We then define Lpvq as

Lpvq “ λ1 Lpe1 q ` λ2 Lpe2 q ` ¨ ¨ ¨ ` λn Lpen q.

On the other hand, if L is linear then it follows from the defining properties of a linear
map that L must take this value on v, and we see that this is the unique linear map
taking the prescribed values on the basis.
The process of defining a linear map from its values on basis vectors is called extending
linearly.

Linear Isomorphisms
A linear transformation which is a bijection is called a linear isomorphism . A pair of
vector spaces are called isomorphic if there is a linear isomorphism between them. In
this case we write V « W .
Proposition 2.3.2. Let L : V Ñ W be a linear isomorphism. The inverse function
L´1 : W Ñ V is a linear map.
Proof. Let w, w1 P W and λ P F. Since L is a bijection, there exist unique v, v 1 P V such
that Lpvq “ w and Lpv 1 q “ w1 . We can see that L´1 satisfies the conditions making it
linear map by direct calculation:

L´1 pw ` w1 q “ L´1 pLpvq ` Lpv 1 qq “ L´1 pLpv ` v 1 qq “ v ` v 1 “ L´1 pwq ` L´1 pw1 q

and
L´1 pλwq “ L´1 pλLpvqq “ L´1 pLpλvqq “ λv “ λL´1 pwq.

We can now see that finite-dimensional vector spaces have a simple classification up
to isomorphism. It requires the following simple but useful lemmas, which hold even for
infinite-dimensional vector spaces.
Lemma 2.3.3. A linear map L : V Ñ W is injective if and only if Lpvq “ 0W implies
v “ 0V .

21
Proof. Any linear map L : V Ñ W satisfies Lp0V q “ 0W . If L is injective, it follows that
Lpvq “ 0W implies v “ 0V . On the other hand, assume that the only element of V which
maps to 0W is 0V . Then for any v, v 1 P V with v ‰ v 1 ,

v ‰ v 1 ñ v ´ v 1 ‰ 0v ñ Lpv ´ v 1 q ‰ 0W ñ Lpvq ´ Lpv 1 q ‰ 0W ñ Lpvq ‰ Lpv 1 q

and it follows that L is injective.

Lemma 2.3.4. An injective linear map takes linearly independent sets to linearly inde-
pendent sets. A surjective linear map takes spanning sets to spanning sets.
Proof. Let L : V Ñ W be a linear map. Assuming that L is injective, let S Ă V be a
linearly independent set. Because L is injective, any element w in the image of S under
L can be written uniquely as Lpsq for some s P S. A linear combination of elements of
the image of S then satisfies
ÿ ÿ ÿ
0W “ λs Lpsq “ Lpλs sq “ Lp λs sq
sPS sPS sPS
ř
only if λs s “ 0V by Lemma 2.3.3. The independence of S then implies that λs “ 0 for
all s and it follows that the image of S under L is linearly independent.
Now assume that L is surjective and let S be a spanning set. We wish to show that
the image of S is spanning. For any w P W , the surjectivity of L implies that there exists
v P V with Lpvq “ w (although such a vř is not necessarily unique). Since
ř S is spanning,
there exist coefficients λs such that v “ λs s and it follows that w “ λs Lpsq.
We have the following immediate corollary.
Corollary 2.3.5. Let L : V Ñ W be a linear transformation of finite-dimensional vector
spaces of the same dimension. If L is injective then it is an isomorphism. Likewise, if L
is surjective then it is an isomorphism.
Proof. If L is injective, choose a basis B for V . The image LpBq of this basis is linearly
independent in W , and since the dimension of W is the same as the dimension of V ,
it follows that L is surjective as well. A similar argument works in the case that L is
surjective.
Finally, we have the following classification result for finite-dimensional vector spaces.
Theorem 2.3.6. Let V and W be finite-dimensional vector spaces over F. Then V « W
if and only if dimpV q “ dimpW q.
Proof. Let tv1 , . . . , vn u and tw1 , . . . , wm u be bases for V and W , respectively. If V « W ,
there exists a linear isomorphism L : V Ñ W . By Lemma 2.3.4, the injectivity of
L implies that the image of the basis for V is linearly independent in W , while the
surjectivity of L implies that the image of the basis for V is spanning. Therefore n “ m
and V and W have the same dimension.
Conversely, suppose that n “ m. We define a linear map L : V Ñ W by defining it on
the basis by Lpvj q “ wj and extending. This is clearly an isomorphism.
Example 2.3.2. It follows from Example 2.2.1 and Proposition 2.3.6 that the spaces
Pn pRq and Rn`1 are isomorphic.

22
2.3.2 Linear Transformations of Finite-Dimensional Vector Spaces
Matrix Representations: An Example
You are probably used to writing linear maps between finite-dimensional vector spaces
in terms of matrices, as in the following example.
Example 2.3.3. Consider the linear map L : R3 Ñ R3 from Example 2.3.1. Using the
standard column vector notation
¨ ˛
λ1
λ1 e1 ` λ2 e2 ` λ3 e3 Ø ˝ λ2 ‚,
λ3

the map can be written as matrix multiplication:

¨ ˛¨ ˛ ¨ ˛
0 1 0 λ1 λ2
Lpλ1 e1 ` λ2 e2 ` λ3 e3 q “ ˝ 2 0 0 ‚˝ λ2 ‚ “ ˝ 2λ1 ‚ Ø λ2 e1 ` 2λ1 e2 ` λ1 e3 .
1 0 0 λ3 λ1

We will see in a moment that linear maps in finite dimensions can always be expressed
as matrices, but that this representation depends on choices of bases. This choice is
sometimes unnatural, so it is important to understand the abstract definition of a linear
map. To further convince you, the next example gives a linear map between infinite-
dimensional vector spaces, where there is no hope to represent it using a matrix.
Example 2.3.4. Consider the map D : C 8 pr0, 1s, Rq Ñ C 8 pr0, 1s, Rq, where Dpf q is
defined at each x P r0, 1s by
Dpf qpxq “ f 1 pxq.
You will show that this is a linear map of vector spaces in the exercises.

Matrix Representations: Formal Theory

We now turn to the matrix representation of a linear map for finite-dimensional vector
spaces. Let L : V Ñ W be a linear map. Abstractly this just means that it satisfies
certain properties which mean that L respects the vector space structures of V and W .
However, if we fix ordered bases e1 , . . . , en for V and f1 , . . . , fm for W , we can represent
L by a size m ˆ n matrix as follows. For each ej , we can write

Lpej q “ λ1j f1 ` λ2j f2 ` ¨ ¨ ¨ ` λmj fm

for some scalars λij . Cycling through the n basis vectors of V , we obtain m ¨ n such
scalars, and our matrix representation of L (with respect to these ordered bases basis) is
given by ¨ ˛
λ11 λ12 ¨ ¨ ¨ λ1n
˚ λ21 λ22 ¨ ¨ ¨ λ2n ‹
˚ ‹
˚ .. .. .. .. ‹ “: pλij qij
˝ . . . . ‚
λm1 λm2 ¨ ¨ ¨ λmn

23
This is called the matrix representation of L with respect to the chosen bases.
To do calculations in matrix representations, note that we can also express an arbitrary
v P V as a matrix with respect to this basis. Namely, as the n ˆ 1 matrix pλ1 , . . . , λn qT
(superscript T denotes matrix transpose), where v “ λ1 e1 ` ¨ ¨ ¨ ` λn en . To evaluate the
linear map on the vector, we just perform the matrix multiplication
¨ ˛¨ ˛
λ11 λ12 ¨ ¨ ¨ λ1n λ1
˚ λ21 λ22 ¨ ¨ ¨ λ2n ‹ ˚ λ2 ‹
.. ‹ ˚ .. ‹ .
˚ ‹˚ ‹
˚ .. .. ..
˝ . . . . ‚˝ . ‚
λm1 λm2 ¨ ¨ ¨ λmn λn

The resulting m ˆ 1 matrix is interpreted as a list of coefficients for Lpvq in the fixed
basis for W .

2.3.3 Determinants
Let L : V Ñ W be a linear transformation of finite-dimensional vector spaces of the same
dimension. There is an algorithmic way to tell when L is an isomorphism from any of its
matrix representation.
From the previous section, we see that once we have chosen bases for V and W , the
linear map L can be represented as matrix multiplication by an n ˆ n matrix A, where
n “ dimpV q. Let A “ p~a1 ~a2 ¨ ¨ ¨ ~an q, where the ~aj are column vectors representing the
columns of A. The determinant is a function det from the set of n ˆ n matrices over F
to F which satisfies three properties:

1. For any scalars λ1 , λ2 P F, any j “ 1, . . . , n and any column vector ~v ,

detp~a1 ¨ ¨ ¨ λ1~aj ` λ2~v ¨ ¨ ¨ ~an q “ λ1 ¨ detp~a1 ¨ ¨ ¨ ~aj ¨ ¨ ¨ ~an q ` λ2 ¨ detp~a1 ¨ ¨ ¨ ~v ¨ ¨ ¨ ~an q.

2. For every j “ 1, . . . , n,

detp~a1 ¨ ¨ ¨ ~aj ~aj`1 ¨ ¨ ¨ ~an q “ ´detp~a1 ¨ ¨ ¨ ~aj`1 ~aj ¨ ¨ ¨ ~an q.

3. detpIn q “ 1.

Theorem 2.3.7. There is a unique map det satisfying the properties of a determinant.
If A is a square matrix representing a linear map L : V Ñ W , then detpAq ‰ 0 if and
only if L is an isomorphism.

To prove the theorem, let us first examine the n “ 2 case. Let

ˆ ˙ ˆˆ ˙ ˆ ˙˙
a b a b
A“ “ ,
c d c d

24
be an arbitrary 2 ˆ 2 matrix with a, b, c, d P F. Using the Property 1 of a determinant,
we have
ˆ ˆ ˙ ˆ ˙ ˆ ˙˙ ˆ ˙ ˆ ˙
1 0 b 1 b 0 b
detpAq “ det a `c , “ a ¨ det ` c ¨ det .
0 1 d 0 d 1 d

Continuing with a similar calculation shows

ˆ ˆ ˙ ˆ ˙˙
1 1 1 0
detpAq “ a b ¨ det ` d ¨ det
0 0 0 1
ˆ ˆ ˙ ˆ ˙˙
0 1 0 0
` c b ¨ det ` d ¨ det . (2.2)
1 0 1 1

Property 3 of a determinant tells us that detpI2 q “ 1. Moreover, Property 2 says that

ˆ ˙ ˆ ˙
1 1 1 1
det “ ´det ,
0 0 0 0

since the matrix is unchanged by switching the order of the columns. This implies that
ˆ ˙
1 1
det “0
0 0

and the same reasoning yields ˆ ˙

0 0
det “ 0.
1 1
Finally, Property 2 implies
ˆ ˙
0 1
det “ ´detpI2 q “ ´1.
1 0

Putting all of this into (2.2), we have

detpAq “ ad ´ bc.

This shows that in the case of n “ 2, a determinant map exists. Moreover, the map is
unique since our formula was forced by the properties of a determinant.
It is also easy to see that A is an isomorphism if and only if detpAq ‰ 0. Indeed,
detpAq “ ad ´ bc “ 0 if and only if the columns of A are linearly dependent. To see this,
assume some entry of A is nonzero (otherwise we are done). Without loss of generality,
say c ‰ 0. Then the columns of A satisfy
ˆ ˙ ˆ ˙
b d a
“ ¨ .
d c c

25
Having linearly dependent columns is equivalent to the failure of A to be the matrix
representation of an isomorphism. This completes the proof of the theorem in the n “ 2
case.
For the general case, we build the determinant by induction. For a 3 ˆ 3 matrix, we
define
¨ ˛
a b c ˆ ˙ ˆ ˙ ˆ ˙
e f d f d e
det ˝ d e f ‚ “ a ¨ det ` b ¨ det ` c ¨ det
h i g i g h
g h i

Inductively, for an n ˆ n matrix A we define the j-th pn ´ 1q ˆ pn ´ 1q minor Aj by

deleting the first row and the j-th column of A. Then

detpAq “ a11 ¨ detpA1 q ` a12 ¨ detpA2 q ` ¨ ¨ ¨ ` a1n ¨ detpAn q.

It is easy to check that this map satisfied the desired properties. To show that this map
is unique is straightforward, but somewhat time consuming. A full proof can be found in
any linear algebra textbook, e.g. Chapter 5 of [3]. A proof that detpAq ‰ 0 if and only
if A is an isomorphism can be found there as well.

2.4 Vector Space Constructions

2.4.1 Subspaces
Let V be a vector space over F. A subset U Ă V is a vector subspace (also called a linear
subspace or simply subspace) of V if it is itself a vector space with respect to operations
obtained by restricting the vector space operations of V . More concretely, U is a vector
subspace if and only if:

1. (Closure Under Addition) for all u, v P U , u ` v P U

2. (Closure Under Scalar Multiplication) for all u P U and λ P F, λu P U .

The dimension of a vector subspace is just its dimension as a vector space, using the
usual definition.

Example 2.4.1. For an n-dimensional vector space V , the vector subspaces of V take
one of the following forms:

1. the subset containing only the zero vector t0V u

2. spans of collections of linearly independent vectors; for v1 , . . . , vm P V linearly

independent, the set
spanptv1 , . . . , vm uq
is an m-dimensional vector subspace

3. the full space V .

26
Example 2.4.2. As you might expect, subspaces of infinite-dimensional vector spaces
can be much more exotic. For example, the set

tf P C 8 pr0, 1s, Rq | f p0q “ 0u

is a vector subspace of C 8 pr0, 1s, Rq.

2.4.2 Special Subspaces Associated to a Linear Transformation

Kernel
Let L : V Ñ W be a linear transformation. We define the kernel of V to be the set

kerpLq “ tv P V | Lpvq “ 0W u.

Proposition 2.4.1. The kernel of a linear transformation L : V Ñ W is a vector

subspace of V .

Proof. Let u, v P kerpLq and λ a scalar. We need to show that u ` v and λu are elements
of kerpLq. Indeed,
Lpu ` vq “ Lpuq ` Lpvq “ 0W ` 0W “ 0W
and
Lpλuq “ λLpuq “ λ0W “ 0W .

Image
The image of the linear map L : V Ñ W is the set

imagepLq “ tw P W | w “ Lpvq for some v P V u.

Proposition 2.4.2. The image of a linear transformation L : V Ñ W is a vector

subspace of W .

We leave the proof of this proposition as an exercise.

2.4.3 Rank and Nullity

Let L : V Ñ W be a linear map of finite-dimensional vector spaces. We define the rank
of L to be the dimension of imagepLq. We define the nullity of L to be the dimension
of kerpLq. These quantities are denoted rankpLq and nullpLq, respectively. We have the
following fundamental theorem.

Theorem 2.4.3 (Rank-Nullity Theorem). For a linear map of L : V Ñ W of finite-

dimensional vector spaces,

rankpLq ` nullpLq “ dimpV q.

27
Proof. Let dimpV q “ n. Let tv1 , . . . , vk u be a basis for kerpLq Ă V , so that nullpLq “ k.
By Theorem 2.2.1, the set tv1 , . . . , vk u can be extended to a basis tv1 , . . . , vk , w1 , . . . , wn´k u
for V . We claim that the set B “ tLpw1 q, . . . , Lpwn´k qu forms a basis for imagepLq, and
this will complete the proof of the theorem.
To see that B is a spanning set for imagepLq, let w P imagepLq. Then there exists
v Př V with Lpvq
ř “ w. There exist unique scalars λ1 , . . . , λk , ν1 , . . . , νn´k such that
v “ λj vj ` ν` w` . Because the vj lie in the kernel of L, it follows that
˜ ¸
k
ÿ n´k
ÿ
w “ Lpvq “ L λj vj ` ν` w `
j“1 `“1
ÿ ÿ
“ λj Lpvj q ` ν` Lpw` q
ÿ
“ ν` Lpw` q,

and this shows that B is spanning.

To see that B is linearly independent, suppose that
n´k
ÿ ´ÿ ¯
0W “ ν` Lpw` q “ L ν` w ` .
`“1
ř
Since L is injective on spantw1 , . . . , w` u, it follows from Lemma 2.3.3 that ν` w` “ 0V .
Since the w` are linearly independent, this implies that all ν` “ 0.

2.4.4 Direct Sums

Given two vector spaces V and W over F, we define the direct sum to be the vector space
V ‘ W with V ‘ W “ V ˆ W as a set, addition defined by

pv1 , w1 q ` pv2 , w2 q “ pv1 ` v2 , w1 ` w2 q

and scalar multiplication defined by

λ ¨ pv, wq “ pλ ¨ v, λ ¨ wq.

Proposition 2.4.4. The dimension of V ‘ W is dimpV q ` dimpW q.

Proof. If either V or W is infinite-dimensional, then so is V ‘W . Indeed, assume without

loss of generality that dimpV q “ 8. Let S Ă V be a linearly independent set containing
infinitely many elements. Then for any w P W , the set tps, wq | s P Su is an infinite
linearly independent subset of V ‘ W .
On the other hand, if V and W are both finite-dimensional, let tv1 , . . . , vn u be a basis
for V and tw1 , . . . , wm u a basis for W . Then it is easy to check that

tpvj , 0qu Y tp0, wk qu

gives a basis for V ‘ W . It follows that dimpV ‘ W q “ n ` m “ dimpV q ` dimpW q.

28
2.4.5 Quotient Spaces
Let U be a vector subspace of V . We define an equivalence relation „U on V by v „U w
if and only if v ´ w P U . As usual, we denote by rvs the equivalence class of v P V ,

rvs “ tw P V | w „U vu “ tw P V | w ´ v P U u.

The collection of equivalence classes is called the quotient of V by U and is denoted V {U

.

Proposition 2.4.5. The quotient space V {U has a natural vector space structure.

Proof. We define the zero vector to be

0V {U “ r0V s “ U,

we define vector addition by the formula

rus ` rvs “ ru ` vs

and we define scalar multiplication by the formula

λrus “ rλus.

We leave it as an exercise to show that the vector space axioms are satisfied with respect
to these operations.
Let V be finite-dimensional. Then the dimension of V {U is readily computable.

Proposition 2.4.6. The dimension of V {U is dimpV q ´ dimpU q.

Proof. Let B 1 be a basis for U and let B denote its completion to a basis for V (which
exists by Theorem 2.2.1). We claim that

trbs | b P BzB 1 u (2.3)

is a basis for V {W . Indeed, this set is spanning, since any rvs P V {U can be written as
« ff
ÿ ÿ ÿ
rvs “ λb b “ λb rbs “ λb rbs.
bPB bPB bPBzB 1

The existence of the coefficients λb comes from the fact that B is a basis for V , the
first equality follows by the definition of the vector space structure of V {U , and the last
equality follows because rbs “ r0s for any b P B 1 . Moreover, the set (2.3) is linearly
independent, as » fi
ÿ ÿ
r0s “ λb rbs “ – λb bfl
bPBzB 1 bPBzB 1

29
 ř
implies that bPBzB 1 λb b P B 1 and this can only be the case if all λb “ 0 by the linear
independence of BzB 1 Ă B.
Finally, we claim that the set (2.3) contains |B| ´ |B 1 | distinct elements. Its cardinality
is certainly bounded above by this number, so we need to check that if b1 , b2 P BzB 1
satisfy b1 ‰ b2 , then rb1 s ‰ rb2 s. This holds because rb1 s “ rb2 s if and only if b1 ´ b2 P B 1 ,
which is impossible by linear independence.

2.4.6 Row and Column Operations

Let L : V Ñ W be a linear map between finite-dimensional vector spaces and let A
denote the matrix representation of L with respect to some fixed choices of ordered bases
for V and W . We have the following row and column operations on A:
1. Multiply all entries in a row/column by the same nonzero element of F ,

2. Permute two rows/columns,

3. Add a nonzero multiple of a row/column to another row/column.

Each row/column operation corresponds to matrix multiplication:

1. To multiply the jth row of A by nonzero a P F, we multiply on the left by the

square matrix of size dimpW q which is diagonal with entries aii “ 1 for all i ‰ j
and aj j “ a. To multiply a column of A a constant, we multiply on the right by a
similar matrix with size dimpV q.

2. To permute the i and j row of A, we multiply on the left by the matrix obtained
by form the identity (of dimension dimpW q) by switching its ith and jth row. A
similar trick, using right multiplication, can be used to permute columns of A.

3. To add a times the ith row of A to its jth row, we left multiply by the matrix
which is equal to the identity except for the entry aij “ a. Once again, a similar
procedure using right multiplication gives the column operation.

Proposition 2.4.7. Let L : V Ñ W be a linear map between finite-dimensional vector

spaces of the same dimension and let A be a matrix representation of L with respect to
some choices of bases for V and W . Let M : V Ñ V and N : W Ñ W be matrices
corresponding to row and column operations on A. Then M and N are isomorphisms.
Proof. The matrices corresponding to row and column operations have nonzero determi-
nant, so the claim follows by Proposition 2.3.7.

2.4.7 Vector Space Associated to a Linear Map

Let L : V Ñ W be a linear map of finite-dimensional vector spaces. There is a vector
space associated to L, denoted θpLq and defined as

θpLq “ W {imagepV q.

30
We refer to the map taking L to θpLq as the θ-correspondence .
Let A be its matrix representation with respect to some choice of ordered bases for V
and W . We have the following characterization of the circumstances where vector spaces
obtained via the θ-correspondence are isomorphic.
Proposition 2.4.8. Let M : V Ñ V and N : W Ñ W be linear isomorphisms. Then
θpN LM q « θpLq. It follows that if a matrix A1 is obtained from A by row and column
operations, then the linear transformation L1 associated to A1 satisfies θpL1 q « θpLq.
Proof. Since M and N are isomorphisms, imagepN LM q has the same dimension as
imagepLq and it follows from Proposition 2.3.6 and Proposition 2.4.6 that θpLq « θpN LM q.
The second part of the proposition follows from Proposition 2.4.7.

2.5 Structures on Vector Spaces

In this section we introduce some extra structures on vector spaces. Let V denote a
vector space over R throughout this section.

2.5.1 Inner Products

An inner product on V is a map

h¨, ¨i : V ˆ V Ñ R
pv1 , v2 q ÞÑ hv1 , v2 i

with the following properties for all u, v, w P V and scalars λ P R:

1. (Positive-Definititeness) hv, wi ě 0 and equality holds if and only if v or w is equal
to 0V ,

2. (Symmetry) hv, wi “ hw, vi,

3. (Bilinearity) hu ` v, wi “ hu, wi ` hv, wi and hλv, wi “ λ hv, wi . It follows from the

symmetry property that hv, u ` wi “ hv, ui ` hv, wi and hv, λwi “ λ hv, wi.
Example 2.5.1. It is a useful exercise to verify that the standard dot product on Rn

pa1 , a2 , . . . , an q ‚ pb1 , b2 , . . . , bn q “ a1 ¨ b1 ` a2 ¨ b2 ` . . . ` an ¨ bn

defines an inner product.

Example 2.5.2. Notice that the dot product can be expressed in matrix form as vwT
for any v, w P Rn , where the superscript T denotes matrix transpose. More generally, let
M be an n ˆ n matrix with real entries. Then the map

pv, wq ÞÑ v ¨ M ¨ wT

defines an inner product on Rn provided:

31
 1. M is symmetric; i.e., M T “ M ,
2. M is positive-definite; i.e. v ¨ M ¨ v T ą 0 for all v ‰ ~0.
This gives a large collection of examples of inner products on Rn which can be easily
generalized to any finite-dimensional vector space over R.
Example 2.5.3. As a more exotic example, consider the vector space C 8 pr0, 1s, Rq with
the map h¨, ¨iL2 defined for functions f and g by
ż1
hf, giL2 “ f ptq ¨ gptq dt.
0

You will show in the exercises that this map defines an inner product.
A pair pV, h¨, ¨iq consisting of a vector space together with a choice of inner product is
called an inner product space.
Remark 2.5.1. One can similarly define an inner product on a vector space over C with
a slight change to the axioms. In this case, the definition is meant to be a generalization
of the map on Cn ˆ Cn defined by

ppz1 , z2 , . . . , zn q, pw1 , w2 , . . . , wn qq ÞÑ z1 ¨ w1 ` z2 ¨ w2 ` . . . zn ¨ wn ,

where w denotes the complex conjugate of w. Can you guess what needs to be changed in
the definition of an inner product in this case?

2.5.2 Norms
A norm on V is a map

}¨}:V ÑR
v ÞÑ }v}

with the following properties for all u, v P V and scalars λ P R:

1. (Positive-Definiteness) }v} ě 0 and equality holds if and only if v “ 0V ,
2. (Linearity Over Scalar Multiplication) }λv} “ |λ| ¨ }v},
3. (Triangle Inequality) }u ` v} ď }u} ` }v}.
There is one immediate source of norms on V .
Proposition 2.5.2. Any inner product h¨, ¨i on V determines a norm on V .
To prove the proposition, we need to make use of a famous lemma.
Lemma 2.5.3 (Cauchy-Schwarz Inequality). For any inner product h¨, ¨i and any u, v P
V,
| hu, vi | ď hu, ui hv, vi .

32
Proof. If v “ 0V we are done, so suppose not. Define λ “ hu, vi { hv, vi2 . Then positive-
definiteness and bilinearity of the inner product implies

0 ď hu ´ λv, u ´ λvi
“ hu, ui2 ` λ2 hv, vi2 ´ 2λ hu, vi
hu, ui2 hv, vi2 hu, vi2 hu, vi2
“ ` ´ 2 .
hv, vi2 hv, vi2 hv, vi2

Rearranging the terms and taking a square root proves the claim.
We can now prove the proposition.
Proof. We define a candidate for a norm } ¨ } on V by the formula
a
}v} “ hv, vi.

We need to check that this definition satisfies the definition of a norm. Positive-definiteness
and linearity over scalar multiplication follow immediately from the corresponding prop-
erties of h¨, ¨i. It remains to check the triangle inequality.
Let u, v P V . Then the bilinearity of the inner product and the Cauchy-Schwarz
Inequality imply

}u ` v}2 “ hu ` v, u ` vi
“ hu, ui2 ` hv, vi2 ` 2 hu, vi
ď }u}2 ` }v}2 ` 2}u}}v}
“ p}u} ` }v}q2 ,

and taking square roots proves the result.

Example 2.5.4. An important family of examples of norms on Rn are the `p -norms ,

defined as follows. For each 1 ď p ă 8, define the norm } ¨ }p on v “ pa1 , . . . , an q P Rn
by the formula
}v}p “ p|v1 |p ` ¨ ¨ ¨ ` |vn |p q1{p .
For p “ 8, define
}v}8 “ max |vi |.
i

Clearly, }v}2 is the standard norm on Rn , which can be written (as in the proposition) in
the form
}v}2 “ hv, vi .
Perhaps surprisingly, is a fact that none of the other `p norms are induced by inner
products!

A pair pV, } ¨ }q consisting of a vector space together with a choice of norm is called a
normed vector space.

33
2.6 Exercises
1. Show that in any vector space V over F, 0F v “ 0V for any v P V .

2. Show that for the vector space Rn , the additive inverse of v P Rn is given by

´v “ ´1 ¨ v.

3. Show that Pn pRq is a vector space over R. (See Example 2.1.2.)

4. Show that C 8 pr0, 1s, Rq is a vector space over R (See Example 2.1.3.)

5. Show that F2 is a field. (See Example 2.1.6.)

6. Show that C 8 pr0, 1s, Cq is a vector space over C. (See Example 2.1.7.)

7. Show that the derivative map defined in Exercise 2.3.4 is a linear transformation
between vector spaces.

8. Let C 0 pr0, 1s, Rq denote the set of continuous functions f : r0, 1s Ñ R. Show that
C 0 pr0, 1s, Rq is a vector space over R and then show that C 8 pr0, 1s, Rq is a vector
subspace of C 0 pr0, 1s, Rq.

9. Prove that C 8 pr0, 1s, Rq is infinite-dimensional. Hint: Find a countably infinite

collection of functions which you can prove are linearly independent.

10. Prove Proposition 2.4.2.

11. Complete the proof of Proposition 2.4.5.

12. Show that the h¨, ¨iL2 defines an inner product on C 8 pr0, 1s, Rq (see Example 2.5.3).

34
3 Metric Space Topology
In applications, we often think of data as some collection of datapoints; i.e., the data
forms a set. Realistically, the set of datapoints usually comes with the extra structure
of a notion of distance between the points. For example, if each datapoint is a vector
of (real) numbers, then we can think of the data set as a collection of points in a vector
space. There is a natural notion of distance between v, w P Rn given by }v ´ w}, where
} ¨ } is any choice of norm on Rn .
It is easy to imagine that the situation of the previous example can generalized to
more exotic structures on the dataset. Perhaps the datapoints actually all lie on or
near a sphere (or more complicated surface) inside of Rn . Perhaps the points are more
naturally represented as the nodes of some graph.
The correct abstract version of this idea is to represent the dataset as a metric space.
A metric space is simply a set X together with a choice of distance function d on X. The
distance function is an abstract function d : X ˆ X Ñ R which satisfies some natural
axioms (see the following section).
We will see in this chapter that the simple idea of treating sets with distance functions
abstractly produces a very rich theory. The study of metric spaces is a subfield of topology.
We will also introduce some basic ideas from topology in this section. For more in depth
coverage of metric spaces and more general topological spaces, a standard reference is
Munkres’ textbook [6].

3.1 Metric Spaces

3.1.1 Definition of a Metric Space
Let X be a set. A metric (or distance function) on X is a map

d:X ˆX ÑR

satisfying the following properties for all elements x, y and z of X:

1. (Positive Definite) dpx, yq ě 0 and dpx, yq “ 0 if and only if x “ y,

2. (Symmetry) dpx, yq “ dpy, xq,

3. (Triangle Inequality) for any elements x, y and z of the set X,

dpx, zq ď dpx, yq ` dpy, zq.

A set together with a choice of metric pX, dq is called a metric space.

35
 A metric subspace of a metric space pX, dq is a metric space pY, dY q, where Y Ă X and
dY “ d|Y ˆY ; that is dY is obtained by restricting the function d to Y ˆ Y Ă X ˆ X. We
will frequently abuse notation and continue to denote the restricted metric by d.

3.1.2 Examples of Metric Spaces

Basic Examples
Example 3.1.1. As a very basic example, consider the set of real numbers R together
with the metric
dpx, yq “ |x ´ y|.
This is called the standard metric on R.
Example 3.1.2. More generally, for any normed vector space pV, } ¨ }q we define a metric
by the formula
dpv, wq “ }v ´ w}.
It follows immediately from the definition of a norm that this function satisfies the prop-
erties required for it to be a metric.
Example 3.1.3. Let X be any nonempty set. Define a metric d on X by setting
"
1 if x ‰ y
dpx, yq “
0 if x “ y.

Let’s check that this formula really defines a metric. It is clear that d is positive-definite
and symmetric, so we just need to show that it satisfies the triangle inequality. Let
x, y, z P X. There are five cases to consider:
1. x “ y “ z: Then dpx, zq “ 0 ď 0 ` 0 “ dpx, yq ` dpy, zq.
2. x “ y, y ‰ z: Then dpx, zq “ 1 ď 0 ` 1 “ dpx, yq ` dpy, zq.
3. x ‰ y, y “ z: Then dpx, zq “ 1 ď 1 ` 0 “ dpx, yq ` dpy, zq.
4. x ‰ y, y ‰ z, z “ x: Then dpx, zq “ 0 ď 1 ` 1 “ dpx, yq ` dpy, zq.
5. x ‰ y, y ‰ z, z ‰ x: Then dpx, zq “ 1 ď 1 ` 1 “ dpx, yq ` dpy, zq.
This shows that the triangle inequality is always satisfied. The reader should check that
there are no other possibilities to consider! This metric is called the discrete metric on
X and it is useful to keep in mind when you are trying to think of counterexamples.
Example 3.1.4. Consider the standard unit sphere S 2 Ă R3 . We define a metric on S 2
as follows. Let u, v P S 2 (i.e., u and v are unit vectors in R3 ). The intersection of the
plane spanned by u and v with the sphere S 2 is called the great circle associated to u
and v. There are two segments along the great circle joining u to v. We define the metric
dS 2 by taking dS 2 pu, vq to be the length of the shorter of these two segments. A similar
construction works for spheres of all dimensions S n´1 Ă Rn . You will show that dS 2 is
really a metric in the exercises.

36
 length “ dS 2 pu, vq
v
u

form
great circle

Example 3.1.5. Consider the square T :“ r0, 1q ˆ r0, 1q Ă R2 . We define the distance
dT between points p1 “ px1 , y1 q and p2 “ px2 , y2 q in T by the formula

dT pp1 , p2 q “ min }px1 , y1 q ´ px2 ` k, y2 ` `q},

k,`PZ

where the norm is the standard one on R2 . This is called the torus metric on T and a
similar construction works for cubes of all dimensions r0, 1sn Ă Rn .
A torus is the geometric shape formed by the surface of a donut. An explanation of
this name for dT is given by the figure below. In the figure we form a donut shape by
identifying edges of the square which have “distance zero”.

You will show that dT is really a metric in the exercises.

Common Examples Arising in Data Analysis

Example 3.1.6. A point cloud in a metric space pX, dq is a metric subspace pY, d|Y ˆY q,
where Y is some finite set. The figure below shows some examples of point clouds. The
figure on the left shows a simple point cloud in R2 . The figure on the fight shows a more
complicated point cloud which appears to lie along the surface of a sphere. The point
clouds that we are interested in—those coming from real-world data—typically have a
large number of points and exhibit some underlying structure. The tools that we develop
will help us to discern this structure!

37
 1

-1 1

-1

Real-world data is often naturally represented as a point cloud in some metric space.
For example, consider customer records for a movie streaming service. Say the service
offers streaming for n titles (n some large integer). Then, as a vastly simplified model,
the record of a single customer could consist of a sequence of 0’s (for movies that have
not been watched) and 1’s (for movies that have been watched). This record can then be
represented a vector in Rn . For two customer records v and w in Rn , the number }v ´ w}
(i.e. the distance between v and w in Rn ) represents the similarity in viewing patterns
between the two customers. The collection of all customer records therefore forms a point
cloud in Rn .
Of course, the customer records of any streaming service are much more detailed and
include information such as when titles were viewed and what ratings the customer as-
signed. Thus the vectors of information can live in a space with much higher dimension
and can contain numbers besides 0’s and 1’s. When comparing the viewing patterns
between two customers, different types of information should potentially be weighted dif-
ferently. This can be interpreted as the statement that the vector space containing the
pointcloud should be endowed with a more complicated metric!
Example 3.1.7. Data frequently has the structure of a graph. A graph G “ pV, Eq
consists of a set of points V called vertices and a set E of edges e “ tv, wu, where v, w P V .
Graphs are realized geometrically by drawing the vertex set and joining vertices v and w
by a line segment when tv, wu P E. The figure below shows a realization of the graph

G “ pV “ tu, v, w, x, y, zu, E “ ttu, vu, tv, wu, tv, zu, tw, xu, tw, yu, tx, yu, ty, zuuq.

Graphs are a convenient representation of data which describes relationships between

points; for example, one could take as a vertex set the members of a social media platform
with a connection between vertices when the corresponding members are “friends”.

38
 A graph G defines a metric space as follows. We define the graph distance between
vertices v, w, dG pv, wq, to be the length of the shortest path joining v to w in G. Here
length means number of edges along the path. In the above example, dG pv, yq “ 2,
because we could take either path tv, wu, tw, yu or tv, zu, tz, yu to join the vertices and
there is no shorter path.
A graph in which there is a unique path joining any two points is called a tree . Consider
the tree T pictured below. For the particular v and w marked in the figure, dT pv, wq “ 3.

3.1.3 Open and Closed Sets

An open metric ball in a metric space pX, dq is defined for a center point x P X and a
radius r ą 0 to be the set

Bpx, rq “ ty P X | dpx, yq ă ru.

The notation will sometimes be decorated. For example, we may use Bd px, rq to empha-
size the choice of metric or BX px, rq to emphasize the set when the choice of metric is
clear from context.

Example 3.1.8. For R with its standard metric, the open metric balls are open intervals.

Example 3.1.9. For a set X with the discrete metric, the open metric balls are of the
form "
x if r ă 1
Bpx, rq “
X if r ě 1.
for all x P X.

Example 3.1.10. In the figure below we show metric balls in R2 of radius 1 with the
metrics induced by the `1 , `2 and `8 norms respectively.

1 1 1

-1 1 -1 1 -1 1

-1 -1 -1

B`1 p~0, 1q B`2 p~0, 1q B`8 p~0, 1q

39
Example 3.1.11. The next figure shows a point cloud of 11 points in R2 . The 5 red
points comprise the radius-1 open metric ball centered at the point ~0 with respect to the
subspace metric induced by the `2 norm.

-1 1

-1

Example 3.1.12. The next example returns to the tree metric space T of Example 3.1.7.
The figure shows the open metric ball Bpv, 3q highlighted in red. Notice that the far
endpoints are not included in the ball, since the definition Bpv, 3q “ tv P T | dpv, wq ă 3u
uses a strict inequality.

The following proposition characterizes the open metric balls of a metric subspace. The
proof is left as an exercise.
Proposition 3.1.1. Let pX, dq be a metric space and let Y Ă X be endowed with the
subspace metric. Then the metric open balls BY py, rq are of the form

BY py, rq “ BX py, rq X Y,

where BX py, rq is the metric open ball in X.

A subset U Ă X of a metric space is called open if for all x P U there exists r ą 0
such that Bpx, rq Ă U . A subset C Ă X is called closed if it can be expressed as the
complement of an open set; that is,

C “ XzU “ tx P X | x R U u

for some open subset of X.

Proposition 3.1.2. Open sets have the following properties:

40
 1. X and H are open;

2. for any collection U of open sets, the set

ď
U
U PU

is also open;

3. for any finite collection U “ tU1 , . . . , Un u of open sets, the set

n
č
Ui
i“1

is open.
Proof. For the first point, note that for any x P X, any r ą 0 satisfies Bpx, rq Ă X. Thus
X is open. Moreover, the statement that H isŤopen is vacuously true. To show that
arbitrary unions of open sets are open, let x P U PU U . Then Ť x P U for some element
of the collection, so there
Şn is r ą 0 such that Bpx, rq Ă U Ă U PU U . Finally, for finite
intersections, let x P i“1 Ui . Then for each i “ 1, . . . , n thereŞexists ri ą 0 such that
Bpx, ri q Ă Ui . Let r be the minimum of the ri ’s. Then Bpx, rq Ă Ui , and this completes
the proof.
Example 3.1.13. Note that an arbitrary union of open sets is open, while the corre-
sponding statement for intersections only concerns finite collections. Indeed, it is easy
to find infinite collections of open sets whose intersection is not open. For example, let
U “ tUn un“1,2,3,... be a collection of open subsets of R, where

Un “ p´1{n, 1{nq

is an open interval. Each set is open, but the intersection

8
č
Un “ t0u
n“1

is not.
We have a similar proposition for closed sets, whose proof we leave as an exercise.
Proposition 3.1.3. Closed sets have the following properties:
1. X and H are closed;

2. for any collection C of closed sets, the set

č
C
CPC

is also open;

41
 3. for any finite collection C “ tC1 , . . . , Cn u of closed sets, the set
n
ď
Ci
i“1

is closed.

We remark that a subset Y of a metric space pX, dq, can be open, closed, both open
and closed, or neither open nor closed.

Example 3.1.14. Consider R with its standard metric induced by | ¨ |. The following
are examples of open sets:

• R

• pa, bq for any a ă b

• p0, 1q Y p2, 3q
Ť8
• p0, 1q Y p2, 3q Y p4, 5q Y ¨ ¨ ¨ “ i“0 p2i, 2i ` 1q.

The following are examples of closed sets:

• R

• ra, bs for any a ă b

• t1u

• r0, 1s Y r2, 3s Y r4, 5s Y ¨ ¨ ¨ “ 8

Ť
i“0 r2i, 2i ` 1s. Note that the union of infinitely many
closed sets is not necessarily closed; this example is just a special case.

The whole real line R and the empty set H are examples of sets which are both open and
closed. We will see in Section 3.3.2 below that these are the only subsets of R with this
property. The following are examples of sets which are neither open nor closed:

• r0, 1s Y p2, 3q

• r0, 1q.

It will be useful to characterize the open sets of a metric subspace.

Proposition 3.1.4. Let pX, dq be a metric space and let Y Ă X be endowed with the
subspace metric. The open subsets of Y are of the form U X Y , where U is an open subset
of X.

Proof. Let U be an open set in X and let y P U X Y . Then there exists an open
metric ball BX py, rq which is contained in U , and it follows that the metric open ball
BY py, rq “ BX py, rq is contained in U X Y . This shows that U X Y is an open subset of
Y.

42
 Now we wish to show that every open subset of Y is of the form U X Y . Let V be an
open subset of Y . For each y P V there exists rpyq ą 0 such that BY py, rpyqq Ă V . Now
consider the set ď
U“ BX py, rpyqq.
yPV

This set is open in X (since it is the union of open sets) and has the property that
V “ Y X U.

3.1.4 Topological Spaces

While we are primarily concerned with metric spaces, it will occasionally be useful to
use more general terminology. For some of the ideas about a metric space pX, dq that we
will introduce, the metric d is auxillary, and we are really interested in the open sets of
pX, dq (as defined in the last section). Based on the properties of open sets that we just
derived, we make the following definition: a topological space is a set X together with a
collection T of subsets of X satisfying the following axioms:

1. X and H are in T ,

2. for any collection U Ă T of elements of T , the set

ď
U
U PU

is also in T ;

3. for any finite collection tU1 , . . . , Un u of elements of T , the set

n
č
Ui
i“1

is in T .

The collection T is called a topology on X and elements of T are called open sets.

Example 3.1.15. A metric space pX, dq is an example of a topological space. The

topology T consists of the open subsets with respect to the metric, as we defined in the
previous section. This topology is called the metric topology on pX, dq.

All of the topological spaces that we will study will be metric spaces. However, many
of the concepts that we will cover can be applied to arbitrary topological spaces; that is,
they are defined in terms of topologies and the metric is of secondary importance. We
refer the reader interested in studying general topological spaces to the excellent textbook
[6].
The notion of a topological space is strictly more general than that of a metric space;
that is, there exist topological spaces whose topologies are not induced by a metric. A
simple example of such a space is given below.

43
Example 3.1.16. Let X “ ta, b, cu. Define a topology on X to be the collection of sets
T “ tH, X, ta, bu, tb, cu, tbuu. Note that there are no open sets U and V such that a P U ,
c P V and U X V “ H; in standard terminology, X is not a Hausdorff space. On the other
hand, any metric space pY, dq has the Hausdorff property: for any x, y P Y with x ‰ y,
the open sets U “ Bd px, {2q and V “ Bd py, {2q, where  “ dpx, yq, have the properties
that x P U , y P V and U X V “ H (i.e., any metric space is Hausdorff).

3.1.5 Limit Points

Let Y Ă X be a subset of a metric space. The interior of Y is the set of points y P Y
such that there exists r ą 0 with Bpy, rq Ă Y . The interior of Y is denoted intpY q.
A limit point of Y is a point x P X such that any open metric ball Bpx, rq intersects
Y in some point besides x. In set notation, this condition is written

pBpx, rq X Y qztxu ‰ H.

Proposition 3.1.5. A subset Y Ă X is closed if and only if it contains all of its limit
points.

Proof. First assume that Y is closed. Then Y “ XzU for some open set U Ă X. For any
x P U there exists r ą 0 such that Bpx, rq X Y “ H and this implies that x is not a limit
point of Y . Therefore Y must contain all of its limit points.
Now assume that Y contains all of its limit points. We claim that XzY is open, whence
it follows that Y is closed. If XzY “ H we are done, so assume not and let x P XzY .
Then there exists r ą 0 such that Bpx, rq X Y “ H; i.e., Bpx, rq Ă XzY . Thus XzY is
open.
The closure of a subset Y is the set Y together with all limit points of Y and is denoted
Y . By the previous proposition, Y is a closed set. Moreover, Y is the “smallest” closed
set containing Y in sense which is made precise by the following proposition.

Proposition 3.1.6. The closure of a subset Y Ă X can be characterized as

č
Y “ tC Ă X | Y Ă C and C is closedu.

Proof. To save space with notation, let

č
Z “ tC Ă X | Y Ă C and C is closedu.

First note that Y is a closed set which contains Y , so it must be that Z Ă Y . It remains
to show that Y Ă Z. Let y P Y . If y P Y , then y is an element of each set in the
intersection defining Z, so it is an element of Z and we are done. Assume that y is a
limit point of Y such that y R Y and let C be a closed set with Y Ă C. Since y is a limit
point of Y it must also be a limit point of C and it follows from Proposition 3.1.5 that
y P C. Since C was arbitrary, it must be that y P Z.

44
 The boundary of a set Y Ă X is the set

BY :“ Y X pXzY q.

In the next example we give an example to demonstrate the intuitive meaning of the
interior, boundary and closure of a given set. You will work out a similar example in the
exercises.
Example 3.1.17. The figure below shows a set Y Ă R2 consisting of a disk with part
of its boundary circle included, a closed disk removed and a point removed. The other
figures show its interior intpY q, boundary BY , closure Y and the boundary of its closure
BY .

Y intpY q BY Y BY

3.2 Continuous Maps

The notion of a continuous map between metric spaces is of fundamental importance.
Accordingly, it has several equivalent definitions which are useful in different contexts.
The next proposition gives two of them. For a function f : X Ñ Y between sets and a
subset Z Ă Y , we use
f ´1 pZq “ tx P X | f pxq P Zu
to denote the preimage set of Z.
Proposition 3.2.1. Let pX, dX q and pY, dY q be metric spaces and let f : X Ñ Y be a
function. The following are equivalent:
1. for any open set U Ă Y , the preimage set f ´1 pU q is open in X;

2. for any  ą 0 and any x P X, there exists δ ą 0 such that dY pf pxq, f px1 qq ă 
whenever dX px, x1 q ă δ.
Proof. Assume that the first property holds and let  ą 0 and x P X. Consider the open
metric ball BY pf pxq, q. By our assumption, the preimage set U “ f ´1 pBY pf pxq, qq is
open. It certainly contains x, and by definition this means that there exists δ ą 0 such
that BX px, δq Ă U . Then whenever dX px, x1 q ă δ, we have x1 P BX px, δq which implies
x1 P U and this in turn implies that dY pf pxq, f px1 qq ă .
We now turn to the reverse implication. Assume that the second property holds and
let U Ă Y be an open set. We wish to show that f ´1 pU q is open. Assuming that the
preimage set is nonempty (otherwise we are done), let x P f ´1 pU q. Then f pxq P U and
since U is open this implies that there is an  ą 0 such that BY pf pxq, q Ă U . By our
assumption, we can choose δ ą 0 such that BX px, δq Ă f ´1 pBY pf pxq, q Ă f ´1 pU q, and
this implies that f ´1 pU q is open.

45
 If a function f satisfies the properties of the previous proposition, we say that it is
continuous.
Example 3.2.1. Consider the metric space pR, | ¨ |q; that is, use the metric on R induced
by absolute value. Let f : R Ñ R be a function. Then the second definition of conti-
nuity reads: f is continuous if for all  ą 0 and all x P R, there exists δ ą 0 such that
|f pxq ´ f px1 q| ă  whenever |x ´ x1 | ă δ. This is the usual definition of continuous that
you have used since Calculus I! This means that all of the elementary functions (polyno-
mials, exponentials, trigonometric functions with appropriately restricted domains) are
continuous in the metric space sense.
The following lemma will be useful and we leave its proof as an exercise.
Lemma 3.2.2. Let f : X Ñ Y and g : Y Ñ Z be continuous maps of metric spaces.
Then g ˝ f : X Ñ Z is continuous as well.

3.3 Topological Properties

Continuous maps are extremely important in the study of metric spaces, as they preserve
the “large scale” metric structures of metric spaces. More generally, they preserve the
open set structure of the metric spaces; that is, they preserve the metric topologies.
Because of this, properties which are preserved by continuous maps are called topological
properties. In the next two subsections we will introduce the two most basic topological
properties.

3.3.1 Compactness
An open cover of a metric space pX, dq is a collection U of open sets such that YU PU U “ X.
A subcover is a subset U 1 Ă U which is still an open cover of X. The space is said to be
compact if every open cover admits a finite subcover. We call a subset Y Ă X compact
if it is compact as a metric space with its subspace metric.
Example 3.3.1. A basic example of a compact space is a finite set of points Y “
ty1 , . . . , yn u in a metric space pX, dq. For any open cover U of Y , there exists an open
set Uj such that xj P Uj for all j (this must be the case, since U covers Y !). Then
tU1 , . . . , Un u is a finite subcover of U.
Example 3.3.2. The space R with its standard topology is not compact. To see this,
consider the open cover U “ YkPZ p´k, kq. Any finite subcollection of elements of U
is of the form tp´k1 , k1 q, p´k2 , k2 q, . . . , p´kn , kn qu for some positive integer n. Let kM
denote the maximum kj . Then Yj p´kj , kj q Ă p´kM , kM q, and the point kM ` 1 P R
is not contained in the subcollection. Therefore the open cover does not admit a finite
subcover.
Example 3.3.3. The subspace p0, 1q Ă R is not compact. To see this, consider the open
cover
U “ tp1{k, 1q | k P Z and k ą 0u.

46
This is an open cover, since for any x P p0, 1q, there exists a positive integer k such
that 1{k ă x. By an argument similar to the last example, any finite subcollection of
elements of U will have its union contained in an interval p1{kM , 1q. Then 1{pkM ` 1q is
not contained in the union of the subcollection, so that U contains no finite subcover.

We see from this example that it is fairly easy to show that the open interval p0, 1q is
not compact. As you might guess, the closed interval r0, 1s is compact, but this takes
much more work to prove. We omit the proof here, but include it for the interested reader
in the appendix.

Theorem 3.3.1. The closed interval r0, 1s is a compact subset of R with its standard
metric.

Proof. Let A be an open cover of r0, 1s. Let

C “ tx P r0, 1s | ra, xs is covered by finitely many sets of Au.

Then 0 P C, since A is an open cover. Then C is nonempty and bounded above (by 1),
so it has a suprememum c. We wish to show that c P C and that c “ 1, hence A admits
a finite subcover of r0, 1s.
We first note that c ą 0. Indeed, since there is some open set A P A with 0 P A,
it must be that the whole half-interval r0, q Ă A for sufficiently small  ą 0. Then
r0, {2s Ă A P A, and it follows that c ě {2 ą 0.
Now we can show that c P C. Certainly c P r0, 1s, so it must be contained in some
open set A P A. Then pc ´ , cs Ă A for sufficiently small  ą 0. Writing r0, cs “
r0, c ´ s Y pc ´ , cs, we have that r0, cs is contained in a funite subcover of A, so c P C.
Finally, we show that c “ 1. If not, c ă 1. Since c P A P A for some open set A, it
must be that rc, c ` s Ă A for some small  ą 0. Then r0, c ` s “ r0, cs Y rc, c ` s is
contained in a finite subcover of A, contradicting the definition of c. Therefore c “ 1,
and this completes the proof of the theorem.

Theorems About Compactness

This subsection includes some fundamental theorems about compact spaces. We leave
most of the proofs as guided exercises.

Proposition 3.3.2. Let f : X Ñ Y be a continuous map of metric spaces. If X is

compact then the image of f is also compact.

Proof. Let U be an open cover of f pXq. We form an open cover of X by pulling back
each open set U P U to the open set f ´1 pU q. The collection of these preimages forms an
open cover of X and since X is compact there is a finite subcover f ´1 pU1 q, . . . , f ´1 pUn q.
Then U1 , . . . , Un forms an open subcover of f pXq, and since U was arbitrary it follows
that f pXq is compact.
An easy way to get examples of compact spaces is to take products of compact spaces.

Proposition 3.3.3. A finite product of compact metric spaces is also compact.

47
 The Extreme Value Theorem is a theorem you learned in calculus which is of funda-
mental importance in optimization problems. The next theorem shows that it holds more
generally, and is really a statement about topology.

Theorem 3.3.4 (Extreme Value Theorem). Let pX, dq be a compact metric space. Any
continuous function X Ñ R achieves its maximum and minimum values.

The definition of compactness for a general metric space is somewhat abstract. The
next two theorems show that in special circumstances, we can replace it with a simpler
definition.

Theorem 3.3.5 (Heine-Borel Theorem). A subset A Ă Rd is compact if and ony if it is

closed and bounded with respect to the standard metric.

Theorem 3.3.6 (Sequential Compactness Theorem). Let pX, dq be a metric space. Then
X is compact if and only if every sequence in X has a convergent subsequence.

3.3.2 Connectedness
A separation of a metric space pX, dq is a pair of nonempty open sets U and V such that
U Y V “ X and U X V “ H. The metric space is called connected if it does not admit a
separation. We call a subset Y Ă X connected if it is connected as a metric space with
its subspace metric.

Proposition 3.3.7. Let f : X Ñ Y be a continuous map between metric spaces. If X is

connected then the image of f is connected as well.

Proof. Let U and V be open subsets of Y such that U Y V “ f pXq and U X V “ H.

Consider the preimages f ´1 pU q and f ´1 pV q. Since f is continuous, the preimages are
open. Moreover, it must be that f ´1 pU q Y f ´1 pV q “ X and f ´1 pU q X f ´1 pV q “ H. Since
X is connected, this implies that one of the preimage sets is empty and it follows that
one of the sets U or V is empty as well. Since U and V were arbitrary, it must be that
no separation of f pXq exists.

Theorem 3.3.8. The subspace r0, 1s Ă R is connected.

Proof. By way of obtaining a contradiction, assume that U YV is a disconnection of r0, 1s.

The sets U and V are closed and bounded subsets of r0, 1s, so they must be compact as
well by the Heine-Borel theorem. Proposition 3.3.3 then implies that U ˆ V is compact.
The distance function U ˆ V Ñ R taking px, yq P U ˆ V to |x ´ y| is continuous, so
it achieves its minimum value, by Proposition 3.3.4. Let u P U and v P V be points
achieving this minimum, with |u ´ v| “ t and assume without loss of generality that
u ď v. It must be that t ą 0, because otherwise U and V intersect. Let x P pu, vq. If
x P U , then |x´v| ă t gives a contradiction to the assumption that pu, vq is a minimum of
the distance function on U ˆ V . Likewise, if x P V , then |u ´ x| ă t gives a contradiction.
Therefore no such disconnection of r0, 1s exists.

Corollary 3.3.9. The metric space pR, | ¨ |q is connected.

48
Proof. The proof of the previous theorem can be easily adapted to show that any closed
interval ra, bs is connected. To obtain a contradiction, suppose that U Y V is a discon-
nection of R. Choose a closed interval ra, bs such that U X ra, bs ‰ H and V X ra, bs ‰ H.
Then pU Xra, bsqYpV Xra, bsq forms a disconnection of ra, bs, giving us a contradiction.
The following is a useful alternate characterization of connectedness.
Proposition 3.3.10. Let pX, dq be a connected metric space. Then the only subsets of
X which are both open and closed are X and H.
Proof. Let Y Ă X be an arbitrary subset. If Y is both open and closed, then the pair
of open sets Y and XzY satisfies Y Y pXzY q “ X. Then the connectedness assumption
implies that one of Y or XzY is empty, hence Y “ H or Y “ X.

Path Connectedness
A path in a metric space pX, dq is a continuous map from the interval I “ r0, 1s into
X. The metric space is said to be path-connected if for any x, y P X there exists a path
γ : I Ñ X with γp0q “ x and γp1q “ y.
Proposition 3.3.11. If pX, dq is path-connected then it is connected.
Proof. We prove the statement by contrapositive. Assume that X is not connected and
let U Y V form a separation of X. Let x P U and y P V . We claim that there is no
path γ : I Ñ X joining x and y. Indeed, since I is connected, its image under γ must be
connected as well. But pU X γpIqq Y pV X γpIqq would form a separation of its image.

3.4 Equivalence Relations

3.4.1 Isometry
Let pX, dX q and pY, dY q be metric spaces. We say that a function f : X Ñ Y is an
isometry if it is onto and dY pf pxq, f px1 qq “ dX px, x1 q for all x, x1 P X. We say that the
metric spaces are isometric and write pX, dX q „iso pY, dY q if there exists an isometry
between them. We will show that „iso is an equivalence relation momentarily, but first
we need a lemma.
Lemma 3.4.1. If f : X Ñ Y is an isometry then it is invertible and its inverse is also
an isometry.
Proof. That f is onto is part of the definition, so we need to show that it is one-to-one.
Indeed, if f pxq “ f px1 q, then 0 “ dY pf pxq, f px1 qq “ dX px, x1 q implies that x “ x1 . It
follows that f is invertible and it remains to show that f ´1 : Y Ñ X is an isometry. Let
y, y 1 P Y . Then

dX pf ´1 pyq, f ´1 py 1 qq “ dY pf pf ´1 pyqq, f pf ´1 py 1 qqq “ dY py, y 1 q,

where the first equality follows from the assumption that f is an isometry.

49
Proposition 3.4.2. The relation „iso defines an equivalence relation on the set of metric
spaces.
Proof. For any metric space pX, dq, the identity map defines an isometry of the space
with itself and it follows that „iso is reflexive. The previous lemma shows that „iso is
symmetric. To show that „iso is transitive, let pX, dX q „iso pY, dY q and pY, dY q „iso
pZ, dZ q. Let f : X Ñ Y and g : Y Ñ Z denote isometries. We claim that g ˝ f : X Ñ Z
is also an isometry. Indeed, for any x, x1 P X,

dZ pg ˝ f pxq, g ˝ f px1 qq “ dY pf pxq, f px1 qq “ dX px, x1 q.

Example 3.4.1. Consider the unit disks Bpp0, 0q, 1q Ă R2 and Bpp1, 0q, 1q Ă R2 , each
endowed with subspace metrics for the standard metric on R2 . These metric spaces are
isometric, with the isometry f : Bpp0, 0q, 1q Ñ Bpp1, 0q, 1q given by the translation map

px, yq ÞÑ px ` 1, yq.

3.4.2 Homeomorphism
The equivalence relation „iso is very restrictive. For our purposes, we will typically want
an equivalence relation which isn’t required to completely preserve the metric structure,
but instead preserves topological structure. Let pX, dX q and pY, dY q be metric spaces.
A homeomorphism between them is a map f : X Ñ Y which is a continuous bijec-
tion with continuous inverse. Metric spaces are called homeomorphic if there exists a
homeomorphism between them. If X and Y are homemorphic metric spaces, we write
X «Y.
Proposition 3.4.3. The relation « defines an equivalence relation on the set of metric
spaces. If metric spaces X and Y are isometric, then they are homeomorphic.
Proof. The proof follows easily from the definition of homeomorphism. The most inter-
esting part of the first part of hte proposition is the transitivity of «, but this follows
easily from Lemma 3.2.2. To see that X „iso Y implies X « Y , it suffices to show that
an isomorphism f : X Ñ Y is continuous and this follows immediately from the  ´ δ
definition of continuity.
Example 3.4.2. The converse of the second part of the proposition does not hold. To
prove this, we need to find spaces X and Y which are homeomorphic but not isometric.
Consider X “ r0, 1s and Y “ r0, 2s, each endowed with the subspace metric from R.
Then the function f : X Ñ Y given by f pxq “ 2x is a homeomorphism, but it is not an
isometry because dp0, 1q ‰ dp0, 2q.
Homeomorphism is a much weaker notion of equivalence than isometry. That two
spaces are homeomorphic only depends on their underlying topological structure and
does not reference distance at all, whereas isometry is defined exactly in terms of dis-
tance preservation. In fact, the definition of homeomorphism extends without change to
topological spaces (which do not necessarily have a metric).

50
3.4.3 Homotopy Equivalence
We now arrive at our weakest form of equivalence for metric spaces, which also has the
most involved definition. This notion of equivalence is called homotopy equivalence. We
will not use it in practice too frequently, but it is a fundamental idea of topology and will
be useful for describing invariance properties of homology.
Let pX, dX q and pY, dY q be metric spaces. Two continuous maps f0 : X Ñ Y and
f1 : X Ñ Y are said to be homotopic if there exists a continuous map F : r0, 1s ˆ X Ñ Y
such that F p0, xq “ f0 pxq and F p1, xq “ f1 pxq for all x P X. Spaces pX, dx q and pY, dY q are
said to be homotopy equivalent if there exist continuous maps f : X Ñ Y and g : Y Ñ X
such that f ˝ g : Y Ñ Y is homotopic to the identity map on Y and g ˝ f : X Ñ X is
homotopic to the identity map on X. In this case, we write X „h.e. Y . We leave the
proof of the following proposition to the reader.

Proposition 3.4.4. Homotopy equivalence is an equivalence relation on the set of metric

spaces. If spaces X and Y are homeomorphic, then they are homotopy equivalent.

Example 3.4.3. Let X “ R and Y “ t0u Ă R. We claim that X and Y are homotopy
equivalent. Let f : X Ñ Y be the constant map x ÞÑ 0 and let g : Y Ñ X be the inclusion
map 0 ÞÑ 0 P R. Then g ˝ f is equal to the identity map on Y , so there is nothing to
prove here. On the other hand f ˝ g : X Ñ X is the constant map x ÞÑ 0, and we need
to show that this is homotopic to the identity on X. To do so, define F : r0, 1s ˆ X Ñ X
by
F pt, xq “ t ¨ x.
Then F is continuous, F p0, xq “ 0 ¨ x “ 0 is the constant-zero map and F p1, xq “ 1 ¨ x “ x
is the identity map on X. This proves our claim.
On the other hand, X and Y are clearly not homeomorphic, since the sets have different
cardinalities (i.e., Y is finite and X is uncountably infinite). This shows that the converse
of the second part of the previous proposition does not hold in general.
Homotopy equivalence is a much weaker notion of equivalence than homeomorphic or
isometric. Note that, like homeomorphism, homotopy equivalence is perfectly well-defined
for general topological spaces.

3.5 Exercises
1. A pair of metrics d and d1 on the same set X are said to be equivalent if for any
x P X and r ą 0, there exist positive numbers r1 and r2 such that Bd1 px, r1 q Ă
Bd px, rq Ă Bd1 px, r2 q. Show that metric equivalence is an equivalence relation on
the set of metrics on a fixed set.

2. Prove Proposition 3.1.3.

3. Work out a more explicit representation of the function dS 2 defined in Example

3.1.4. Hint: try to write the distance dS 2 pu, vq using the angle between the vectors
u, v P R3 , then relate this to the formula for standard dot product.

51
 4. Prove that the function dS 2 defined in Example 3.1.4 is a metric.

5. Prove that the function dT defined in Example 3.1.5 is a metric.

6. For the set Y shown below, draw the interior intpY q, boundary BY and closure Y .

7. Consider R2 with its standard metric. Classify the following sets as open, closed,
open and closed, or none of the above:
a) r0, 1q ˆ r0, 1s
b) R2 ztp0, 0qu
c) tpa, aq | a P Ru
d) pr0, 1s ˆ r0, 1sqzpp1{4, 3{4q ˆ p1{4, 3{4qq

8. Prove Lemma 3.2.2.

9. Prove Proposition 3.3.3.

10. Prove the Heine-Borel Theorem Add hints.

11. Prove the Sequential Compactness Theorem Add hints.

12. Prove Proposition 3.4.4.

13. Prove that the sets R2 and D2 “ tx P R2 | }x} ď 1u.

14. Prove that the sets R2 zt0u and S 1 “ tx P R2 | }x} “ 1u are homotopy equivalent.

52
4 Homology of Simplicial Complexes
In this chapter we begin to study a particular type of topological space called a simplicial
complex. Our next major goal is to understand how to encode the topological features
of a simplicial complex by a list of vector spaces called homology groups. The next few
chapters follow quite closely the notation and exposition of the excellent survey paper
[1].

4.1 Motivation: Distinguishing Topological Spaces

The fundamental question of topology is as follows: given two topological space (metric
spaces, if you like) X and Y , are X and Y homeomorphic? If you suspect that the
answer is “yes”, then you need only to produce such a homeomorphism. However, if you
think that the answer is “no”, then you need to prove that no such homeomorphism can
possibly exist! In this section we examine some simple examples which will convince us
that some sophisticated tools might be necessary to answer this question.

Example 4.1.1. Are the spaces X “ p0, 1q and Y “ R homeomorphic?

While the spaces X and Y are quite different from a metric space perspective (one
has diameter 1, the other is unbounded), they are the same topologically. To see this we
construct a homeomorphism. First note that p0, 1q and p´π{2, π{2q are homeomorphic
by a simple map f : p0, 1q Ñ p´π{2, π{2q defined by

f pxq “ π ¨ x ´ π{2.

It therefore suffices to find a homeomorphism g : p´π{2, π{2q Ñ R, and this is given by

gpxq “ tanpxq.

Indeed, g is continuous, bijective, and its inverse g ´1 pxq “ arctanpxq is also continuous.

Example 4.1.2. Are the spaces X “ r0, 1s and Y “ R homeomorphic?

This is similar to the last example, but X does feel topologically distinct from R in that
X contains some boundary points. Thus we claim that X and Y are not homeomorphic,
and our goal is to show that no homeomorphism f : X Ñ Y can possibly exist. A
standard trick is to look for a specific topological property that one space has and the
other doesn’t. In this case, we know that X is compact and that R is not. Then the
image of any continuous map f : X Ñ Y must also be compact and it follows that any
such continuous map cannot be surjective! Therefore X and Y are not homeomorphic.

Example 4.1.3. Are the spaces X “ r0, 1q and Y “ p0, 1q homeomorphic?

53
Figure 4.1: The spaces X “ r0, 1q ˆ p0, 1q and Y “ p0, 1q ˆ p0, 1q are shown in the top
row. The second row shows each space with a point removed. For X, the
point is chosen so that the resulting space has no “holes”. For Y , any choice
of point to remove results in hole in the space’s interior.

We once again suspect that the answer is “no”, but in this case neither space is compact,
so we will need a different strategy. The following lemma (and its obvious generalizations)
will be useful.
Lemma 4.1.1. Let f : X Ñ Y be a homeomorphism. Then for any x P X, the restriction
of f to Xztxu is a homeomorphism onto Y ztf pxqu.
Proof. The restricted map is clearly still a bijection. To see that it is continuous, let
U Ă Y ztf pxqu be an open set. Then U “ U 1 ztf pxqu for some open set U 1 Ă Y , and it
follows that
f ´1 pU q “ f ´1 pU 1 ztf pxquq “ f ´1 pU qztxu
is open in Xztxu. Therefore f is continuous. Continuity of f ´1 follows similarly.
Now we note that, for our particular example, Xzt0u “ p0, 1q is a connected set, but
Y ztf p0qu is not connected (removing any point from Y results in a set which is not
connected). Since connectedness is preserved by continuous maps, it follows that there
is no homeomorphism f : X Ñ Y , by contrapositive to the lemma.
Example 4.1.4. Are the spaces X “ r0, 1q ˆ r0, 1s and Y “ pr0, 1s ˆ r0, 1sqzpp1{4, 3{4q ˆ
p1{4, 3{4qq homeomorphic (see Figure 4.1)?
Your intuition should be that the answer is “no”. However, none of our previous tricks
will work here: both X and Y are connected and compact, and removing a finite number
of points from X or Y will not result in a disconnected space. However, Y is “obviously”
different from X because it has a “hole”. How do we detect the presence of this hole
using topology?
The goal of this chapter is to develop a tool called homology which is an algorithm
for counting “holes” of various dimensions in a topological space. To do so for a general
topological or metric space is quite technical (this is discussed briefly in Section 4.5), so
we will restrict to a special class of spaces called simplicial complexes. Roughly, these

54
are spaces which are pieced together in a controlled way from a collection of triangles
and higher-dimensional analogues of triangles. Since these objects have an intuitively
“linear” structure, one might hope that the process of counting holes in the spaces can
be reduced a linear algebra operation!
Once we have the tools of homology in hand, you will be able to distinguish the spaces
from Example 4.1.4 with ease. You will do so in the exercises.

4.2 Simplicial Complexes

4.2.1 Geometric Simplicial Complexes
Convex Sets
A subset S of Rk is said to be convex if for any points x, y P S, each point p1 ´ tqx ` ty,
t P r0, 1s, along the interpolation between x and y is also contained in S. Otherwise S is
said to be nonconvex.

Remark 4.2.1. In the above, we are using x, y P S to denote points in Rk , but the
expression p1 ´ tqx ` ty treats x and y as vectors. We will frequently conflate the notion
of a point x P Rk with the vector with basepoint at ~0 and endpoint at x.

Convex Nonconvex
The convex hull of S is smallest convex subset of Rk which contains S and is denoted
cvxpSq. More precisely,
č
cvxpSq “ tC | S Ă C Ă Rk and C is convexu.

The figure below shows a set overlaid with its convex hull.

55
Simplices
Let S “ tx0 , x1 , . . . , xn u be a finite subset of Rk . The set S is said to be in general
position if its points are not contained in any affine subspace of Rk of dimension less than
n (thus n ď k). Recall that an affine subspace of Rk is a set of the form

x ` V “ tx ` v | v P V u,

where V Ă Rk is a vector subspace (see the figure below).

Figure 4.2: The figure on the left shows a set tx0 , x1 , x2 u of 3 points in R2 which are
in general position—any line can only contain 2 of the points. The figure
on the right shows a set of points which are not in general position. The
1-dimensional affine subspace containing all the points is indicated in red.

For a set S in general position, the simplex associated to S is the set σpSq “ cvxpSq.
The points xi are called the vertices (the singular form is vertex ) of σpSq. Any pair of
distinct points xi , xj P S determing their own simplex, called an edge of σpSq. In general,
for subset T Ă S, σpT q is called a face of σpSq. The number n is called the dimension of
σpSq.
Frequently, we will only be interested in the simplex σpSq and not in the particular set
S which defines it. We will refer to a set σ Ă Rn as a simplex if σ “ σpSq for some set S
in general position.

Example 4.2.1. It will be convenient to have a standard picture to refer to. The standard
n-dimensional simplex is the simplex associated to the set

S “ tp1, 0, 0, . . . , 0q, p0, 1, 0, . . . , 0q, . . . , p0, 0, 0, . . . , 1qu Ă Rn`1

consisting of all points on the coordinate axes at Euclidean distance 1 from the origin.
The figure on the left shows the standard 2-dimensional simplex with the 1-dimensional
face (i.e., an edge) σpp1, 0, 0q, p0, 0, 1qq highlighted. The figure on the right shows a
(nonstandard) 3-dimensional simplex embedded in R3 with one of its 2-dimensional faces
highlighted.

56
Simplicial Complexes
A (geometric) simplicial complex is a collection of simplices X in some Rn satisfying:

1. for any simplex σ P X , all faces of σ are also contained in X ,

2. for any two simplices σ, τ P X , the intersection σ X τ is also a simplex and which is
a face of both σ and τ .

Intuitively a simplicial complex is a shape obtained by gluing together a collection of

simplices, where gluing is only allowed to take place along faces.

Example 4.2.2. The figure on the left shows a complicated simplicial complex in R3 .
The figure on the right is a collection of simplices in R3 which is not a simplicial complex
since the simplices do not interesect along faces.

57
4.2.2 Abstract Simplicial Complexes
As a subspace of Rn (for some n), any simplicial complex inherits the structure of a
metric space. We can therefore study simplicial complexes up to the equivalence rela-
tion of isometry. This notion of equivalence is too rigid in many applications, and we
are actually primarily interested in topological properties of simplicial complexes. This
means that we really wish to study simplicial complexes up to the equivalence relation
of homeomorphism.
Fortunately, simplicial complexes are a class of geometric objects whose topological
structure can be encoded very efficiently—this is exactly the reason that we wish to use
simplicial complexes as a way to encode the topology of data! The topological informa-
tion from a simplicial complex that we are after can be deduced from the combinatorial
structure of the complex, which is encoded in the number of its simplices of various di-
mensions and in the way that the various simplices intersect. The topological information
does not depend on the particular geometric embedding of the complex in a Euclidean
space. With this motivation in mind, we will give a more abstract definition of a simplex
in terms of the combinatorial (topological) information.
An abstract simplicial complex is a pair X “ pV pXq, ΣpXqq (we will also use the
notation X “ pV, Σq), where V pXq is a finite set and ΣpXq is a collection of subsets of
V pXq such that for any σ P ΣpXq and any nonempty τ Ă σ, τ P ΣpXq. The elements
of V pXq are called the vertices of X and the elements of ΣpXq are called the simplices
or faces of X. Faces containing exactly two vertices are called edges. Faces containing
exactly pk ` 1q-vertices are called k-dimensional faces, or just k-faces. If σ is a k-face,
then a pk ´ 1q-face of σ is a pk ´ 1q-face τ with τ Ă σ.

Example 4.2.3. Let X be a simplicial complex. We can construct an abstract simplicial

complex X associated to X by first taking V pXq to be the union of all vertices of all
simplices contained in X . We include a subset of V pXq in ΣpXq if and only if the subset
consists of the vertices of some simplex in X . We leave it as an exercise to show that X
is really an abstract simplicial complex.

Example 4.2.4. The standard n-dimensional abstract simplex is the abstract simplicial
complex ∆n with vertex set t0, 1, 2, . . . , nu and edge set consisting of every non-empty
subset of the vertex set. How does this compare to the standard simplex defined in
Example 4.2.1 (see the exercises)?

A map of abstract simplicial complexes X and Y is a map f : V pXq Ñ V pY q such

that for all σ P ΣpXq, f pσq P ΣpY q. This means that for any collection tv1 , . . . , vk`1 u of
vertices of X which define a k-simplex, the set tf pv1 q, . . . , f pvk`1 qu defines some simplex
in Y . Note that we do not require f to be injective, so it is possible that the image
set contains redundant entries, whence it defines a lower-dimensional simplex. A map of
abstract simplicial complexes f from X to Y is called an simplicial isomorphism if it is a
bijection and if for all τ P ΣpY q, f ´1 pτ q P ΣpXq. Two abstract simplicial complexes are
simplicially isomorphic if there is a simplicial isomorphism between them.
Example 4.2.3 shows that for any geometric simplicial complex X , we can construct an
associated abstract simplicial complex. In the other direction, for an abstract simplicial

58
complex X, a (geometric) simplicial complex X is called a geometric realization of X if
the abstract simpicial complex associated to X is simplicially isomorphic to X. We use
the notation |X| for a geometric realization of X. Note that |X| is highly nonunique!
Also note that we have yet to show that any such |X| exists.
Proposition 4.2.2. A geometric realization exists for any abstract simplicial complex
X “ pV pXq, ΣpXqq.
Proof. Assume V pXq contains n`1 points. We embed the vertex set in Rn`1 by mapping
the elements of V pXq to the vertices of the standard n-dimensional simplex. For each
subset ΣpXq, we include the simplex in the geometric realization which is formed by
taking the convex hull of the corresponding vertices.
The construction given in the proof always works, but it is somewhat inefficient in the
sense that an abstract simplicial complex with n`1 vertices can be geometrically realized
in Rk for k far smaller than n ` 1.
Example 4.2.5. Consider the abstract simplicial complex X “ pV pXq, ΣpXqq with
V pXq “ t0, 1, 2, 3, 4, 5u and

ΣpXq “ tt0, 1, 2u, t0, 1u, t0, 2u, t1, 2u

t3, 4, 5u, t3, 4u, t3, 5u, t4, 5u
t2, 3u, t1, 4uu

A geometric realization of X is shown in the figure below.

5
2

3 4

1
0

The realization is a subset of R2 , rather than the R6 required by the construction in the
proof of Proposition 4.2.2.
Remark 4.2.3. Given a geometric simplicial complex X , we can form its abstract sim-
plicial complex X as set with finitely many elements. The abstract simplicial complex is
a very compact representation of the topology and combinatorics of X , but it completely
loses the metric space structure of of X .

4.3 Topological Invariants of Simplicial Complexes

4.3.1 Connected Components
Let X be a simplicial complex and let X denote its associated abstract simplicial complex.
Let V pXq denote the vertex set of X. We can define an equivalence relation „ on V pXq

59
by
v „ v 1 ô tv, v 1 u is an edge of X.
We then have the following proposition which relates the connected components of the
metric space X to the set V pXq{ „.

Proposition 4.3.1. The connected components of X are in bijective correspondence with

V pXq{ „.

Proof. Let C “ tX0 , . . . , XN u denote the set of connected components of X (remember, we

are only considering finite simplicial complexes, so it is okay to assume that X has finitely
many components). We define a map C Ñ V pXq{ „ by taking Xk to the equivalence class
in V pXq{ „ containing any vertex of Xk . We need to check that this map is well-defined
and that it is a bijection.
To see that the map is well-defined, let v, v 1 be vertices of Xk . Since Xk is path-
connected, there exists a sequence of edges v0 , v1 , . . . , vm such that v “ v0 , v 1 “ vm and
tvi , vi`1 u is an edge of Xk for all i. Then the transitivity of the equivalence relation „
implies that v and v 1 lie in the same equivalence class in V pXq{ „.
To see that the map is a bijection, we can explicitly define its inverse. We define a map
V pXq{ „Ñ C by taking the equivalence class of v to the component of X which contains
v. By a similar argument to the last paragraph, this map is well-defined. Moreover, it is
easy to see that this is the inverse to the map defined in the first paragraph.

4.3.2 Back to Linear Algebra: Free Vector Spaces

Before moving on to studying more interesting topological invariants of simplicial com-
plexes, we pause to introduce the very important notion of a free vector space on a set.
Let F be a field and let S be a finite set. The free vector space over F on the set S is
the vector space VF pSq with underlying set consisting of functions φ : S Ñ F. We will
sometimes shorten notation to V pSq “ VF pSq when the field is understood to be fixed.
The vector space operations are defined pointwise: for φ, φ1 P VF pSq, λ P F and s P S,

pφ ` φ1 qpsq “ φpsq ` φ1 psq,

pλ ¨ φqpsq “ λ ¨ φpsq.

The zero vector in VF pSq is the zero function; i.e., the function which takes every element
of S to zero.
For each s P S, let φs denote the characteristic function for s defined by
"
1 1K s1 “ s
φs ps q “
0K s1 ‰ s.

Proposition 4.3.2. The set of characteristic functions of the elements of S forms a basis
for VF pSq. It follows that dimpVF pSqq “ |S|.

Proof. We need to show that tφs usPS is spanning and linearly independent. Let φ be an
arbitrary element of VF pSq. For each s P S, let λs “ φpsq. Then we can write φ as the

60
linear combination ÿ
φ“ λs φs ,
sPS

and this shows that tφs usPS is a spanning set.

ř To show that it is linearly independent, consider an arbitrary linear combination1 φ “
sPS αs φs . If the linear combination is equal to the zero function, then for each s P S,
we have ÿ
0F “ φps1 q “ αs φs ps1 q “ αs1 .
sPS
1
Since s was arbitrary, it must be that each coefficient in the linear combination is zero.
We will refer to the basis consisting of characteristic functions as the standard basis for
VF pSq.
Any map f : S Ñ T of sets induces a linear map VF pf q : VF pSq Ñ VF pT q by extending
linearly the function defined on basis functions by

VF pf qpφs q “ φf psq .

Proposition 4.3.3. Let f : S Ñ T be a map of sets. The induced linear map is given by
the following general formula for φ P VF pSq:
ÿ
pVF pf qpφqq ptq “ φpsq.
sPS|f psq“t

The proof of the proposition is left as an exercise.

Let S be a finite set and let R Ă S ˆ S be a binary relation. We define a subspace
VF pRq Ă VF pSq by
VF pRq “ spantφs ´ φs1 | ps, s1 q P Ru.

Proposition 4.3.4. There is an isomorphism of vector spaces

VF pSq{VF pRq « VF pS{Rq.

Proof. We define a map

L : VF pS{Rq Ñ VF pSq{VF pRq
by linearly extending the map defined on basis vectors by

Lpφrss q “ rφs s.

First we need to check that this map is actually well-defined. This means that we
need to show that s „ s1 implies rφs s “ rφs1 s. The latter equality holds if and only if
φs ´ φs1 P VF pRq, which holds by definition.

61
 ř
Next we need to show that the map is injective. Let rss λrss φrss denote an arbitrary
element of VF pS{Rq and assume that it maps by L to the zero vector. Then
¨ ˛
ÿ ÿ
0 “ L ˝ λrss φrss ‚ “ λrss rφs s
rss rss

implies that all λrss “ 0 by linear independence of the vectors rφs s. The kernel of L is
the zero vector, and L is therefore injective. Finally, the fact that the vector spaces are
of the same dimension shows that L is surjective as well.

4.3.3 First Example

We first consider the example shown below, which we have already identified as a metric
space with the metric inherited from R2 . Denote this metric space by X .

B C

A D

Now that we have the proper definitions, we easily see that this shape can be thought of
as a (geometric) simplicial complex in R2 . Let X denote the associated abstract simplicial
complex for X . Its vertex set is

VertpXq “ tA, B, C, Du

and its edge set is

EdgepXq “ ttA, Bu, tA, Cu, tA, Du, tB, Cu, tC, Duu.

To simpify notation, we will denote 1-dimensional simplices (edges) by, e.g., AB rather
than tA, Bu.
We can visually see that this metric space consists of one path component (it is con-
nected), and that it contains a pair of “loops” which apparently cannot be shrunk to a
point. Our goal is to develop a computational approach which will allow us to discern
the apparent topological features of the shape from the combinatorial information given
by its simplicial decomposition. This will be accomplished by using tools from linear
algebra.
Consider the vector space C0 pXq :“ VF2 pVertpXqq, the free vector space over F2 (the
field with 2 elements) generated by VertpXq. (This is a standard notation that will be

62
explained in the following chapter.) Similarly, let C1 pXq :“ VF2 pEdgepXqq denote the free
vector space over F2 generated by EdgepXq. There is a natural, geometrically-motivated
linear map
B1 : C1 pXq Ñ C0 pXq
called the boundary map. Each basis element of C1 pXq corresponds to an edge of X,
and the boundary map takes a basis element to the linear combination of its boundary
vertices. For example,
B1 pφAB q “ φA ` φB ,
where we are using the notation of Section 4.3.2 and denoting the basis element of C1 pXq
assocated to the edge AB by φAB . With respect to this basis, the matrix form of this
linear map is given by

AB AC AD BC CD
» fi
A 1 1 1 0 0
B — 1 0 0 1 0 ffi
B1 “ — ffi.
C – 0 1 0 1 1 fl
D 0 0 1 0 1

By construction, the column space of B1 is exactly the span of vectors of the form
φv ` φ1v , where v, v 1 P VertpXq and tv, v 1 u P EdgepXq. Said differently, imagepB1 q is the
free vector space associated to the set tpv, v 1 q | tv, v 1 u P EdgepXqu. Applying Proposition
4.3.4, we see that the quotient space

C0 pXq{imagepB1 q

is isomorphic to the free vector space associated to the set VertpXq{ „, where v „ v 1 ô
tv, v 1 u P EpXq. Proposition 4.3.1 says that V pXq{ „ is in bijective correspondence with
the set of connected componenets of X . We have just proved the following Proposition,
which works for any simplicial complex X (defining Cj pX, F2 q and B1 in the appropriately
generalized ways—see Section 4.4).
Proposition 4.3.5. The number of connected components of X is given by the dimension
of the vector space C0 pX, F2 q{imagepB1 q.
For our specific example, performing row-reduction on the matrix (keeping in mind
that we are working over F2 , where 1 ` 1 “ 0), we obtain
» fi
1 1 1 0 0
— 0 1 1 1 0 ffi
– 0 0 1 0 1 fl .
— ffi

0 0 0 0 0

By inspection, we see that the first three columns of the reduced matrix form a basis
for its column space. This means that the image of B1 is 3-dimensional. We conclude
that C0 pX, F2 q{imagepB1 q is 1-dimensional, corresponding to the fact that the space X is
connected!

63
 We have seen that the image of B1 gives us important geometric about X . The next
natural step would be to examine the kernel of B1 . By the rank-nullity theorem (Theorem
2.4.3) the kernel of B1 must be 2-dimensional. We leave it to the reader to check that the
vectors ¨ ˛ ¨ ˛
1 0
˚ 1 ‹ ˚ 1 ‹
˚ ‹ ˚ ‹
˚ 0 ‹ and ˚ 1 ‹
˚ ‹ ˚ ‹
˝ 1 ‚ ˝ 0 ‚
0 1
form a basis for kerpB1 q. In terms of our basis, this means that the kernel is spanned by
the vectors
φAB ` φAC ` φBC and φAC ` φAD ` φCD .
Looking back at the picture of X , we see that these vectors exactly describe the apparent
loops in the shape! Indeed, one can intuitively think of the sums in the vector space
C1 pX, F2 q as unions, so that the vectors listed above correspond to the unions of edges

tA, Bu Y tA, Cu Y tB, Cu and tA, Cu Y tA, Du Y tC, Du, (4.1)

respectively. It is visually obvious that these are loops in X .

Apparently (at least for this simple example) the dimension of the kernel of B1 counts
the number of loops in the simplicial complex. At this point, one might object: there are
other loops in X which are not contained in the list (4.1). The most obvious is the loop

tA, Bu Y tB, Cu Y tC, Du Y tA, Du.

This is where we see that the extra vector space structure is important in our construction.
This loop corresponds (under our informal association of unions with sums) to the vector

φAB ` φBC ` φCD ` φAD .

In matrix notation, we have

φAB ` φBC ` φCD ` φAD “ p1, 0, 1, 1, 1q,

which can be expressed as the linear combination

p1, 0, 1, 1, 1q “ p1, 1, 0, 1, 0q ` p0, 1, 1, 0, 1q “ pφAB ` φAC ` φBC q ` pφAC ` φAD ` φCD q.

(Remember that we are working over the field F2 !) Thus the loop tA, Bu Y tB, Cu Y
tC, Du Y tA, Du can be viewed as a linear combination of the loops in the list (4.1). We
finally conclude that the dimension of the kernel of B1 counts the loops in X which are
independent in this sense.

64
4.3.4 Second Example
Now consider the metric space shown below. We denote this simplicial complex by Y
and the associated abstract simplicial complex by Y .

B C

A D

The space Y is clearly a slight modification of the space X from the previous section.
In particular, the vertex and edge sets of Y the same as those of X:

V pY q “ tA, B, C, Du and EpY q “ ttA, Bu, tA, Cu, tA, Du, tB, Cu, tC, Duu.

To obtain Y from X , we add in a single 2-simplex. The face set for Y is thus

F pY q “ ttA, B, Cuu.

Similar to the previous section, we use C0 pY, F2 q and C1 pY, F2 q to denote the free
vector spaces over F2 generated by V pY q and EpY q, respectively. Due to the presence
of a 2-dimensional simplex, we also introduce the notation C2 pY, F2 q for the free vector
space over F2 generated by F pY q. As before, we are interested in the boundary map
B1 : C1 pY, F2 q Ñ C0 pY, F2 q, which has exactly the same matrix representation as the map
B1 in the previous section.
From the previous section, we see that B1 still has 3-dimensional image. Applying
Proposition 4.3.5, we easily deduce the (visually obvious) fact that Y has a single con-
nected component. Likewise, the kernel of B1 is 2-dimensional, and is spanned by the
vectors
φAB ` φAC ` φBC and φAC ` φAD ` φCD .
Now we have run into a problem: the loop corresponding to tA, Bu, tA, Cu and tB, Cu
has been “filled in” by a 2-simplex in order to construct Y. Due to the presence of a
higher-dimensional simplex in Y, there is another natural map of interest. This is the
map B2 : C2 pY, F2 q Ñ C1 pY, F2 q defined on the basis for (the 1-dimensional vector space)
C2 pY, F2 q by
φABC ÞÑ φAB ` φAC ` φBC .
That is, B2 takes the vector corresponding to the 2-simplex tA, B, Cu to the linear com-
bination of vectors corresponding to edges along its boundary. For this reason, B2 is also

65
called a boundary map. In matrix form, we have

ABC
» fi
AB 1
AC —
— 1 ffi
ffi
B2 “ AD —
— 0 ffi.
ffi
BC – 1 fl
CD 0

The image of B2 is clearly the span of the vector φAB `φAC `φBC , which corresponds to
the boundary of the single 2-simplex in Y. We therefore see that the number of linearly
independent loops in Y (which are not filled in by a face) is given by the dimension of
the vector space

kerpB1 q{imagepB2 q “ spanF2 tφAB ` φAC ` φBC , φAC ` φAD ` φCD u{spanF2 tφAB ` φAC ` φBC u
« spanF2 tφAC ` φAD ` φCD u.

This construction works for general simplicial complexes, so we state our conclusion in
the following (slightly ill-defined) proposition.
Proposition 4.3.6. The number of linearly independent loops in a simplicial complex X
is given by the dimension of the vector space kerpB1 q{imagepB2 q.
There is an ambiguity in the proposition: how do we know that imagepB2 q Ă kerpB1 q?
After all, if this is not the case then the proposition doesn’t even make sense. We will
delay the proof until the next section, where it will be done in far greater generality.
For now, we claim that this proposition makes intuitive sense by the construction of the
boundary maps. The kernel of B1 contains linear combinations of edges which “have no
boundary”—that is, the union of the edges gives a closed loop. The image of B2 contains
linear combinations of edges which are the boundary of a 2-simplex—that is, unions of
edges which bound a face. Thus the vector space kerpB1 q{imagepB2 q contains all loops,
modulo those loops which are filled in by a face.

4.4 Homology of Simplicial Complexes over F2

4.4.1 Chain Complexes
Chain Groups
Let X be a (finite) simplicial complex and let X denote its associated abstract simplicial
complex. We define the k-th chain group of X over F2 to be the free vector space over F2
generated by the set of k-dimensional simplices of X. The k-th chain group is denoted
Ck pXq, or sometimes simply by Ck when the simplicial complex X is understood to be
fixed.
Remark 4.4.1. As defined, these Ck are really just vector spaces. We call these vector
spaces chain groups in order to match with the terminology used in most literature. A

66
group is a more general algebraic structure than a vector space—in particular, the additive
structure of any vector space turns it into a group. By using this more general structure,
we can define chain groups with other coefficients; e.g. one frequently considers chain
groups over the integers Ck pX; Zq. For our purposes, it will be sufficient to consider the
chain groups as vector spaces. See Section 4.5 for a brief discussion of these more general
chain groups.

Boundary Maps
For each k, we define the boundary map

Bk : Ck Ñ Ck´1

by defining it on the basis

tφσ | σ is a k-simplex of Xu (4.2)

via the formula

ÿ
Bk φσ “ tφτ | τ is a pk ´ 1q-dimensional face of σu.

The reason for calling Bk the “boundary map” should be clear. Indeed, Bk takes a
k-simplex to the sum of pk ´ 1q-simplices which lie along its boundary!

Alternate Forms of Bk
We will give two alternative forms of the map Bk which will be convenient in different
contexts. First let σ “ tv0 , . . . , vk u be a k-simplex of X with vertices vi . Then Bk applied
to this simplex takes the explicit form
k
ÿ
Bk φσ “ φtv0 ,...,x
vj ,...,vk u .
j“0

Here we use the hat to denote ommision of the vertex; that is,

tv0 , . . . , vpj , . . . , vk u “ tv0 , v1 , . . . , vj´1 , vj`1 , . . . , vk´1 , vk u.

The linear map Bk also has a convenient matrix form. Let us pick an ordering of the
standard basis (4.2) for Ck ; that is, we write the basis as

pφσ1 , φσ2 , . . . , φσnk q,

where the σj form a complete list of the k-dimensional simplices of X. Likewise, we pick
an ordering for the standard basis for Ck´1 :

pφτ1 , φτ2 , . . . , φτnk´1 q,

67
where the τi form a complete list of the pk ´ 1q-simplices of X. Then the matrix form of
Bk (with respect to these bases) has a 1 for its pi, jq entry if the ith pk ´ 1q-simplex is a
face of the jth k-simplex. Otherwise its pi, jq entry is 0.

Main Property of Bk
The boundary maps Bk have a very important property:

Theorem 4.4.2. For every k,

Bk ˝ Bk`1 : Ck`1 Ñ Ck´1

is the zero map.

For brevity, this theorem is frequently stated as simply B 2 “ 0. Moreover, the theorem
can be stated more geometrically as “the boundary of a boundary is empty”. This is
intuitively clear, but requires a formal proof to check our intuition.
Proof. Let σ “ tv0 , v1 , . . . , vk`1 u be a pk ` 1q-simplex of X. Then
k`1
ÿ
Bk`1 φσ “ φtv0 ,...,x
vj ,...,vk`1 u .
j“0

By linearity,
k`1
ÿ
Bk ˝ Bk`1 φσ “ Bk φtv0 ,...,x
vj ,...,vk`1 u
j“0
k`1
ÿÿ
“ φtv0 ,...,vpi ,...,x
vj ,...,vk`1 u .
j“0 i‰j

Expanding this sum, we see that each vector φtv0 ,...,vpi ,...,x
vj ,...,vk`1 u appears exactly twice.
Since we are working over the field F2 , this means that Bk ˝ Bk`1 φσ is zero. This shows
that the claim holds on arbitrary basis vectors and it follows that it holds in general.

Cycles and Boundaries

There are a pair of interesting subspaces associated to each linear map Bk . Let Zk pXq
denote the kernel of Bk . As in the case of chain groups, we will shorten the notation to
Zk when the simplicial complex X is understood. We refer to elements of Zk as k-cycles.
Let Bk pXq “ Bk denote the image of Bk`1 (note the shift in index!). Elements of Bk are
called k-boundaries. We have the following immediate corollary of Theorem 4.4.2.

Corollary 4.4.3. For all k, Bk Ă Zk .

Proof. Let φ P Bk . Then, by definition, φ “ Bk`1 ψ for some ψ P Ck`1 . Theorem 4.4.2
implies that Bk ˝ Bk`1 ψ is zero; i.e. Bk φ “ 0 and φ P Zk .

68
Homology Groups
We finally define the main objects of interest for this section. The k-th homology group
of X is the quotient space
Hk pXq “ Zk pXq{Bk pXq.
Note that this is well-defined by Corollary 4.4.3.

Remark 4.4.4. As we remarked about the chain groups Ck pXq, each Hk pXq is actually
a vector space over F2 . We use the terminology “homology group” in order to agree with
the literature. In more general homology theories, the homology groups have a richer
algebraic structure—see Section 4.5.

Betti Numbers
The k-th Betti number of X is the integer

βk pX q :“ dimHk pXq.

Intuitively, the k-th Betti number of X counts the number of k-dimensional holes in the
topological space X .

Example 4.4.1. Consider the simplicial complex shown below, where the tetrahedron
is not filled in by a 3-simplex. The Betti numbers for the figure below are: β0 “ 3,
indicating that it has 3 connected components; β1 “ 2, represented by the two empty
triangles; β2 “ 1, represented by the empty tetrahedron. These Betti numbers were
computed “by inspection”, but we could also compute the homology groups explicitly to
arrive at the same conclusion.

4.5 More Advanced Topics in Homology

Homology theory is a fundamental part of the subfield of topology called algebraic topol-
ogy. As such, it can be developed in many incredibly sophisticated ways. In this section

69
we mention some variants of the homology theory discussed in these notes and some
more advanced properties of homology. A rigorous treatment of these topics is beyond
the scope of these notes, but an awareness of them will be useful for the reader who
wishes to study the topic further. We will therefore informally discuss some directions
one can take in further research into homology theory. A standard reference for learning
more advanced algebraic topology is [2].

4.5.1 Variants of Homology

Homology with Other Coefficients
A very natural way to generalize our construction of homology groups is to start by
taking our chain groups to be free vector spaces over different fields. An argument could
be made that it would be more natural to start by defining homology using chain groups
over R, since the majority of readers are likely to be most comfortable with real vector
spaces. Indeed, given an abstract simplicial complex X, we can define chain groups

Ck pX; Rq “ VR ptk-dimensional simplices of Xuq.

We can then adjust our definitions to get homology with real coefficients. The main
technical drawback is that we would then need to introduce a more complicated boundary
map which takes into account an “orientation” on each simplex of X. Our motivation
for working over F2 was precisely to avoid these technicalities!
More generally, we could define chain groups to be free vector spaces over more general
fields, such as the field with three elements F3 (for any prime number p, there is a
corresponding field with p elements Fp ). Homology over other finite fields can be useful
for distinguishing spaces, and is actually built into several of the available persistence
homology programs. Even more generally, one can define homology by starting with
chain groups defined as modules over some other group or ring, which are algebraic
structures that generalize vector spaces and fields.

Homology of Topological Spaces

In this chapter we defined homology of simplicial complexes with F2 -coefficients. The
previous subsection indicates that we could define homology theories of simplicial com-
plexes with more general coefficients by some mild adjustments to our definitions. There
is another natural question: can we extend homology to treat more general topological
spaces? The answer to the question is “yes”, but it turns out to be not so straightforward
to do so.
Let X be a topological space (or a metric space, if you prefer). A triangulation of X is a
geometric simplicial complex X such that X and X are homeomorphic (X Ă Rn endowed
with the subspace topology). One could then calculate the homology of the simplicial
complex X using our definition. Unfortunately, some obvious issue arise immediately:

• Triangulations are highly non-unique. Is it possible that two triangulations X and

X 1 of the space X have different homology groups? (See Figure 4.3.)

70
 Figure 4.3: A 2-sphere S 2 Ă R3 and a pair of distinct triangulations.

• Does every space X admit a triangulation? If not, then this strategy will not allow
us to calculate the homology of a general space.

The second question has been a subject of intense study in pure mathematics for
decades. One obvious obstruction is that our definition of a simplicial complex requires
that every simplicial complex is compact, hence any noncompact space cannot be tri-
angulated. It was recently proved [5] that there even exist relatively simple compact
topological spaces (compact 5-dimensional manifolds, which are 5-dimensional analogues
of 2-dimensional surfaces such as spheres and donuts) that have no triangulation. To
treat general topological spaces, we therefore need a more flexible version of homology.
One such theory is called singular homology . The idea is to form the k-th chain group
of a topological space X as

C sing pX; Fq “ VF ptcontinuous maps of the standard k-simplex into Xuq.

Boundary maps and a homology theory can be defined from there. Singular homology
has the important property that if two spaces are homeomorphic, then their singular ho-
mology groups must agree (this is called functoriality, and is treated in the next section).
It is known that for a simplicial complex, the singular homology groups and simplicial
homology groups (as we have defined in this chapter) are the same, and it follows that if
X and X 1 are different triangulations of the same space X, then all spaces will have the
same homology groups.
The problem with using this singular homology approach for applications is that the
singular chain groups are infinite-dimensional and therefore impossible to work with di-
rectly. There are many sophisticated tools used to treat them abstractly, but in practice
one must always convert a space into a finite simplicial complex in order to do direct
calculations. The discussion in this section shows that, if any such triangulation exists,

71
then it doesn’t matter which triangulation we choose for a space—the resulting homology
will always be the same!

4.5.2 Functoriality
Consider simplicial complexes X and Y with abstract simplicial complexes X and Y .
Let f : V pXq Ñ V pY q be a map of abstract simplicial complexes. This map induces a
well-defined map Ck pf q : Ck pXq Ñ Ck pY q of chain groups (over F2 ) by defining
"
φf pσq if φf pσq P Ck pY q
Ck pf qpφσ q “
0 otherwise

for each k-simplex σ P Ck pXq and extending linearly. Recall that a map of simplicial
complexes is not required to be injective, so it is possible that φf pσq is a lower-dimensional
simplex—this is the reason for the conditional definition of Ck pf q.
Theorem 4.5.1. The maps Ck pf q : Ck pXq Ñ Ck pY q induce well-defined linear maps
Hk pf q : Hk pXq Ñ Hk pY q on homology vector spaces.
Proof. We first claim that the following diagram commutes.
Bk
Ck pXq Ck´1 pXq
Ck pf q Ck´1 pf q

Bk
Ck pY q Ck´1 pY q

This means that if we start in the upper left corner and proceed to the lower right
through either of the two possible paths, the resulting map is the same. To check this,
let φσ P Ck pXq be a basis element. We will assume that f takes the k-simplex σ to a
k-simplex f pσq—the case in which f pσq is a lower-dimensional simplex follows similarly.
Then ÿ
Bk pCk pf qpφσ qq “ Bk φf pσq “ tφξ | ξ is a pk ´ 1q-face of f pσqu . (4.3)
On the other hand,
´ÿ ¯
Ck´1 pf qpBk pσqq “ Ck´1 pf q tφτ | τ is a pk ´ 1q-face of σu
ÿ
“ tφf pτ q | τ is a pk ´ 1q-face of σu. (4.4)

Next note that the assumption that f pσq is a k-simplex implies that f is injective on the
vertices of σ. It follows immediately that the expressions (4.3) and (4.4) are equal.
The fact that the diagram commutes implies that Ck pf q takes Zk pXq into Zk pY q and
Bk pXq into Bk pY q. Therefore Ck pf q induces a well-defined linear map Hk pf q from the
quotient space Hk pXq “ Zk pXq{Bk pXq into the quotient space Hk pY q “ Zk pY q{Bk pY q.

The property described by the theorem—that maps between spaces induce maps be-
tween homology vector spaces—is an example of a general mathematical principle called

72
functoriality. A much more general property is enjoyed by singular homology groups. We
state it here without proof because the result is very useful, although a bit too technical
to prove for our purposes.

Theorem 4.5.2. If topological spaces X and Y are homotopy equivalent, then their sin-
gular homology vector spaces Hksing pX; F2 q and Hksing pY ; F2 q are isomorphic.

Although we haven’t proved the theorem, we will use the following corollary.

Corollary 4.5.3. Let X and Y be topological spaces with triangulations X and Y. If the
simplicial homology groups of X and Y are not all isomorphic, then X and Y are not
homotopy equivalent.

4.6 Exercises
1. Let S Ă Rk be a set. Show that cvxpSq is equal to the set
ď
tp1 ´ tqx ` ty | t P r0, 1s, x, y P Su.

2. Write down the abstract simplicial complex associated to the geometric simplicial
complex shown below.

3. Prove Proposition 4.3.3.

4. For each of the simplicial complexes shown in Figure 4.4, write down matrix expres-
sions for B1 and B2 with respect to natural choices of bases. Compute the images
and kernels of each map (feel free to use a computer algebra system to do so).
Calculate the dimension of kerpB1 q{imagepB2 q in each case. Does your answer make
sense intuitively?

5. Let X be a simplicial complex and let X “ pV pXq, ΣpXqq be as defined in Example

4.2.3. Show that X defines an abstract simplicial complex.

6. Let X be the standard n-simplex defined in Example 4.2.1 and let X be the abstract
simplex associated to X (see Example 4.2.3). Show that X is simplicially isomorphic
to the standard abstract n-simplex defined in Example 4.2.4.

7. Prove that the spaces from Example 4.1.4 are not homeomorphic. Here is a sug-
gested strategy: find a triangulation of each space. Compute the homology vector
spaces for each triangulation and show that they are not all the same. Using func-
toriality (in particular, Corollary 4.5.3), conclude that the spaces are not homotopy
equivalent, hence not homeomorphic.

73
 D
B

B
A
D F

A C
C

Figure 4.4: The space on the left is a simplicial complex in R2 . The space on the right is
a simplicial complex in R3 . It has the 1-skeleton of a 3-simplex, with a single
face tA, B, Cu filled in.

74
5 Persistent Homology
As we stated in Chapter 3, a typical real-world dataset comes in the form of a point
cloud —that is, a finite subset of some ambient metric space. Every point cloud determines
a finite metric space pX, dq by taking d to be the restriction of the metric from the
ambient space. Our goal is therefore to study the topology of finite metric spaces. But
here we see a problem: finite metric spaces are classified up to homeomorphism by the
number of points in the space, and the number of datapoints in a large dataset is not
a very interesting invariant. On the other hand, we can intuitively distinguish between
topological types of finite metric spaces, as in the example shown in Figure 5.1. The
point clouds contain the same number of points, so they are homeomorphic. But the
point cloud on the left appears intuitively to be unstructured, while the point cloud on
the right appears to have a topological feature (a “hole”, or a 1-dimensional homology
cycle!)
So the question becomes, how do we algorithmically encode the apparent topological
differences between these finite metric spaces? The approach that we will take is through
persistent homology of their associated Vietoris-Rips complexes. The rough idea is as
follows. Let pX, dq be a finite metric space. For each distance parameter r, we associate
to X a simplicial complex VRpX, rq defined in terms of d. We can then calculate the
simplicial homology of this complex. The homology vector spaces change with the pa-
rameter r, and those homology classes which survive for a long interval of r values (i.e.,
those which “persist”) are deemed topologically relevant, while homology classes that
appear and quickly disappear are treated as noise. The goal of this chapter is to fill in
the details of this process.

Figure 5.1: A pair of point clouds with the same number of points.

75
5.1 Vietoris-Rips Complex
5.1.1 Definition
Let pX, dq be a finite metric space. For each real number r ě 0, we can associate
a simplicial complex to X called the Vietoris-Rips complex and denoted VRpX, rq as
follows. The vertex set of VRpX, rq is simply the set X. A subset tx0 , x1 , . . . , xn u of X
is declared to be a simplex of VRpX, rq if and only if

dpxi , xj q ď r @ i, j P t0, 1, . . . , nu.

5.1.2 A Simple Example

Clearly, the definition of VRpX, rq depends heavily on the choice of r. Let’s look at a
simple example for various choices of r.

Example 5.1.1. Consider the metric space X consisting of 6 points forming the vertices
of a regular hexagon of side length 1 in the Euclidean plane. We label the points of X
as A ´ F . These points form the vertex set of VRpX, rq for any choice of r ě 0. The
Vietoris-Rips complexes of X can be grouped into four categories.

B C

A D

F E

1. For 0 ď r ă 1, VRpX, rq is just the set of discrete vertices.

?
2. For 1 ď r ă 3, the Vietoris-Rips complex can be visualized as the simplicial
complex pictured in the figure second from the left. For each pair of consective
vertices, there is an edge in VRpX, rq, because consective vertices are
? at distance
1 from each other. Non-consecutive vertices are at distance at least 3 from one-
another, so there are no other simplices in VRpX, rq.
?
3. For 3 ď r ă 2, each triple of consecutive vertices forms a 2-simplex in VRpX, rq.
For example, the elements of the set tA, B, Cu satisfy
? ? ?
dpA, Bq “ 1 ď 3, dpB, Cq “ 1 ď 3, dpA, Cq “ 3.

Any triple of vertices which are not consecutive will contain a pair of vertices which
are 2 units apart, so there are no other 2-simplices in VRpX, rq. Moreover, any set
containing 4 points will contain a pair of vertices which are 2 units apart, so there
are no higher-dimensional simplices in VRpX, rq. The figure second from the right

76
 shows the 2-dimensional simplices in VRpX, rq. Note that the figure is not actually
a simplicial complex, since the simplices aren’t attached to each other in the correct
way! It will typically be best to think of VRpX, rq as an abstract simplicial complex.
Of course, VRpX, rq has a geometric realization, but it will typically be difficult or
impossible to visualize.
4. For r ě 2, every set of vertices is included in VRpX, rq—this is simply because the
greatest distance between any two points in X is 2. This means that VRpX, rq
contains 3, 4 and 5-dimensional simplices. The figure on the right shows the set of
edges in VRpX, rq (i.e., there is an edge joining any two vertices).

5.1.3 Observations
From Example 5.1.1, we can immediately make some observations about VRpX, rq for
any finite metric space pX, dq. The following propositions are clear.
Proposition 5.1.1. For r ď mintdpx, x1 q | x, x1 P Xu, VRpX, rq is homeomorphic to X.
Let S be a finite set. The complete simplicial complex on S is the simplicial complex
containing a simplex for every subset of S.
Proposition 5.1.2. For r ě maxtdpx, x1 q | x, x1 P Xu, VRpX, rq is the complete simpli-
cial complex on X.
Another important phenomenon illustrated by Example 5.1.1 is that the topology of
VRpX, rq changes with r according to the geometry of X. In the example, we see that
a loop is formed when r “ 1, that the loop “persists” as 2-dimensional simplices are
attached when 1 ě r ă 2, and the loop finally dissappears when r “ 2 as it is filled in by
higher-dimensional simplices. This is the key idea of persistent homology, which we will
define below.

5.1.4 Other Complexes

There are a variety of other ways to associate a simplicial complex to a finite metric
space, for example, the α-complex is defined for a finite metric space pX, dq which is
isometrically embedded in some larger metric space pY, dq—for simplicity, assume that
pY, dq is RN with Euclidean distance. For each x P X, the Voronoi cell of x is the set

Vorpxq “ ty P Y | dpx, yq ď dpx1 , yq for all x1 P Xu.

For each r ą 0, we define the α-cell

Apx, rq “ Bd px, rq X Vorpxq.

The α-complex of X with scale parameter r is the simplicial complex with vertex set X
and k-faces consisting of sets tx0 , . . . , xk u such that
k
č
Apxj , rq ‰ H.
j“1

77
Figure 5.2: A point cloud, its Voronoi decomposition and its α-complexes at scale param-
eters r “ 3{4 and r “ 5{4.

An example is shown in Figure 5.2. The α-complex is typically smaller than the Vietoris-
Rips complex in that it contains fewer simplices. However, the algorithms to compute it
require one to compute the Voronoi cells of the ambient space, which are computationally
expensive when the ambient space is high-dimensional.
Some other parameterized complexes that can be associated to a finite metric space
are the Cĕch complex and the witness complex. See [1] for more details.

5.2 Linear Algebra with Persistence Vector Spaces

Inspired by the previous section, and the Vietoris-Rips complex construction, we wish
to introduce a version of simplicial homology which is parameterized over r. All of the
concepts used in linear algebra generalize to give parameterized versions.

5.2.1 Definitions
Persistence Vector Space
Let F be a field (we will mainly be using F “ F2 , so it is okay to just keep this choice in
mind). A persistence vector space over F is a family tVr urPR` of vector spaces Vr over F
together with a family of linear maps Lr,r1 : Vr Ñ Vr1 for r ď r1 . Moreover, we require
the linear maps to satisfy the following compatibility condition: if r ď r1 ď r2 , then
Lr,r2 “ Lr1 ,r2 ˝ Lr,r1 .
We will denote a persistence vector space by tVr u, with the understanding that per-
sistence vector spaces are always parameterized over R` . When talking about multiply
persistence vector spaces, we will need to distinguish their families of linear maps. In this
case, we will use the notation LVr,r1 for the maps associated to tVr u.

Linear Transformation Between Persistence Vector Spaces

Let tVr u and tWr u be persistence vector spaces over F. A linear transformation of per-
sistence vector spaces is a family tfr u of linear maps fr : Vr Ñ Wr which preserves the

78
structure of the maps Lv,v1 . That is, for all r ď r1 the following diagram commutes:

LV
r,r 1
Vr Vr1
fr fr 1
LW
r,r 1
Wr Wr1

To say that the diagram commutes means that either one of the possible paths from Vr
to Wr1 yields the same result. More precisely,

fr1 ˝ LVr,r1 “ LW
r,r1 ˝ fr .

A linear transformation of persistence vector spaces is called an isomorphism if it

admits a two-sided inverse.

Sub-Persistence Vector Space

A sub-persistence vector space is a collection tUr u of linear subspaces Ur Ă Vr such that
LVr,r1 pUr q Ă Ur1 holds for each r ď r1 .
Let f “ tfr : Vr Ñ Wr u be a linear map of persistence vector spaces. The kernel of f
is the sub-persistence vector space

kerpf q “ tkerpfr qu

and the image of f is the sub-persistence vector space

impf q “ timpfr qu.

Quotient Persistence Vector Space

Let tUr u be a sub-persistence vector space of tVr u. The quotient persistence vector space
V {U
is the persistence vector space tVr {Ur u where the linear maps Lr,r1 are given for v P Vr
by the formula
V {U “ ‰
Lr,r1 prvsq “ LVr,r1 pvq .

Direct Sum of Persistence Vector Spaces

For persistence vector spaces tVr u and tWr u over F, we define the direct sum tVr u ‘ tWr u
to be the persistence vector space with

ptVr u ‘ tWr uqr “ Vr ‘ Wr

and linear maps

LVr,r‘W
1 : Vr ‘ Wr Ñ Vr1 ‘ Wr1
given by ` ˘
LVr,r‘W
1 pv, wq “ LVr,r1 pvq, LW
r,r1 pwq .

79
Free Persistence Vector Space
A filtered set is a set X together with a map ρ : X Ñ R` . For each r P R` , we define
the sublevel set
Xrrs “ tx P X | ρpxq ď ru.
The free vector space generated by pX, ρq is the persistence vector space denoted tVF pX, ρqr u
and defined by
VF pX, ρqr “ VF pXrrsq.
V pXq
Note that for each pair r ď r1 , Xrrs Ă Xrr1 s so that we can define the linear maps Lr,rF 1
can be defined by the formula
V pX,ρq
Lr,rF 1 pφx q “ φx
for each basis vector φx P VF pX, ρqr .
A general persistence vector space tVr u is called free if it can be expressed as tVF pX, ρqr u
for some filtered set pX, ρq. It is called finitely generated if it is free and the filtered set
pX, ρq can be chosen so that X is finite.

Example 5.2.1. Our main examples of free persistence vector spaces are persistence
chain complexes. Given a finite metric space pX, dq, we construct its Vietoris-Rips com-
plex VRpX, rq for each scale parameter r. We then obtain a filtered set pX, ρq, where
finish this

5.2.2 Matrix Representations of Linear Maps of Persistence Vector

Spaces
Recall that choosing bases for finite-dimensional vector spaces V and W allows us to
represent an abstract linear transformation L : V Ñ W as matrix multiplication. Our
goal is to define a similar representation for linear maps between persistence vector spaces.
As one would expect, the representation is a bit more involved. For this section, we will
restrict our attention to finitely generated persistence vector spaces.

pX, Y q-Matrices
Let pX, Y q be a pair of finite sets. An pX, Y q-matrix over F is a size |X| ˆ |Y | matrix over
F whose entries are indexed by the sets X and Y rather than integers. That is, it is a
matrix paxy q of entries axy P F. For each x P X, we use rpxq to denote the row associated
to x. Similarly, cpyq denotes the column associated to y P Y .
Note that if X and Y are ordered, then the ordering turns an pX, Y q-matrix into a
matrix in the usual sense.

The Matrix Associated to a Linear Map

Let pX, ρq and pY, σq be finite filtered sets and let

f : tVF pY, σqr u Ñ tVF pX, ρqr u

80
be a linear transformation of the associated persistence vector spaces. Since X and Y
are finite, it must be the case that there exists some large real number R such that
r ě R implies VF pX, ρqr “ VF pXq and VF pY, σqr “ VF pY q. Moreover, if both r, r1 ě R
then the linear maps fr and fr1 are the same map on their common domain VF pY, σqr “
VF pY, σqr1 “ VF pY q. When r ě R, we denote this common map by

f8 : VF pY q Ñ VF pXq.

Using the standard bases tφx uxPX and tφy uyPY for VF pXq and VF pY q, respectively, there
is an pX, Y q-matrix associated to f8 , which we denote Apf q “ paxy q. We once again
remark that, in order to actually think of paxy q as a matrix in the usual sense, we need
to pick an ordering of the bases. This will not be necessary in what follows.
We have a representation result which says that Apf q contains all of the information
carried by f . To prove it, we will need the following simple lemma.
ř
Lemma 5.2.1. A linear combination xPX ax φx P VF pXq lies in VF pX, ρqr if and only if
ax “ 0 whenever ρpxq ą r.

Proof. By definition, the vector space VF pX, ρqr consists exactly

ř of linear combinations
of φx with x P Xrrs; i.e. φpxq ď r. A linear combination ax φx is of this form if and
only if ax “ 0 when x R Xrrs; i.e. when φpxq ą r.
An pX, Y q-matrix paxy q is called pρ, σq-adapted if it has the property that axy “ 0
whenever ρpxq ą σpyq.

Proposition 5.2.2. The pX, Y q-matrix Apf q is pρ, σq-adapted and any pρ, σq-adapted
matrix A uniquely determines a linear transformation of perisistence vector spaces

fA : tVF pY, σqr u Ñ tVF pX, ρqr u

such that the correspondences f ÞÑ Apf q and A ÞÑ fA are inverses of each other.
ř
Proof. The image of the basis vector φy P VF pY q is the linear combination xPX axy φx ,
by definition of Apf q. On the other hand, the structure of f implies that the image of φy
lies in VF pX, ρqσpyq , so Apf q is pρ, σq-adapted by Lemma 5.2.1.
Now let A “ paxy q be an arbitrary pρ, σq-adapted matrix. We define a linear map of
persistence
ř vector spaces by restriction. That is, a basis element φy P VF pY, σqr is mapped
to xPX axy φx . The assumption that A is pρ, σq-adapted guarantees that the image of
this map lies in VF pX, ρq, by Lemma 5.2.1.

5.3 Persistence Homology

A filtered simplicial complex is a family tXr u of simplicial complexes indexed by R, such
that for each r ď r1 , Xr is a subcomplex of Xr1 . Let tXr u denote the associated family of
abstract simplicial complexes.

81
Example 5.3.1. Our main examples are the simplicial complexes constructed for point
clouds in Section ??. For example, the collection of Vietoris-Rips complexes tVRpX, rqu
defines a filtered simplicial complex.

The k-th peristence homology vector space of a filtered simplicial complex is the per-
sistence vector space tP Hk pXqr u with P Hk pXqr the k-th simplicial homology of the
simplicial complex tXr u. This does determine a persistence vector space: for r ď r1 , the
P H pXq
map Lr,r1 k is the map induced on homology by the inclusion map Xr ãÑ Xr1 .

5.4 Examples
5.4.1 Example 1
5.4.2 Example 2

5.5 Exercises
1. Consider the?point cloud formed by the 14 points are at the vertices of the cube of
side length 2 (this value is chosen to make calculations a bit simpler),
? ? ? ? ? ? ? ? ? ? ? ?
p0, 0, 0q, p0, 0, 2q, p 2, 0, 0q, p 2, 0, 2q, p0, 2, 0q, p0, 2, 2q, p 2, 2, 0q, p 2, 2, 2q

and the midpoints of the faces of the cube

? ? ? ? ? ?
p 2{2, 2{2, 0q, p 2{2, 0, 2{2q, p0, 2{2, 2{2q,
? ? ? ? ? ? ? ? ?
p 2{2, 2{2, 2q, p 2{2, 2, 2{2q, p 2, 2{2, 2{2q.

Work out the persistent homology of this point cloud.

82
6 Representions of Persistent
Homology
6.1 Structure Theorem for Persistence Vector Spaces
We have so far developed a way to assign a topological invariant to a point cloud: first turn
the point cloud into a filtered simplicial complex (via, say, the Vietoris-Rips construction),
then compute its persistent homology. The resulting topological invariant initially seems
quite complex; it is, after all, a collection of infinitely many vector spaces. The goal
of this section is to show that the persistence vector spaces arising from the persistent
homology construction can actually be described in a simple way.

6.1.1 Preliminaries
Before stating the main result, we need to introduce some notation.

Birth-Death Persistence Vector Spaces

Let b P r0, 8q and d P r0, 8s be real numbers with b ă d. Let PF pb, dq denote the
persistence vector space with
"
F r P rb, dq
PF pb, dqr “
0 r R rb, dq,

where 0 denotes the vector space containing only 0. The linear maps are defined by
$
& 0 răb
PF pb,dq
Lr,r1 “ 1 b ď r ď r1 ă d
0 r1 ě d,
%

where 0 is shorthand for the zero map and 1 is shorthand for the identity map F Ñ F.
We refer to these special vector spaces as birth-death persistence vector spaces . The
numbers pb, dq are called a birth-death pair .
The following example suggests that birth-death persistence vector spaces play a special
role when dealing with free persistence vector spaces on finite filtered sets.
Example 6.1.1. Let pX, ρq be a finite filtered set with X “ tx1 , . . . , xn u and consider
the persistence vector space tVF pX, ρqr u. We define a map
à
f : tVF pX, ρqr u Ñ PF pρpxj q, `8q
j

83
as follows. For fixed r, note that VF pX, ρqr “ VF pXrrsq and
˜ ¸
à à
PF pρpxj q, `8q « F.
j r xPXrrs

We can think of elements of the target space as pt1u, Xrrsq-matrices (that is, as row
vectors indexed by Xrrs). Then the map fr is given on the canonical basis for VF pXrrsq
by taking φx (x P Xrrs) to the pt1u, Xrrsq-matrix with a 1 in the position cpxq and zeros
elsewhere.
V pX,rhoq P pb,dq
It is easy to see that for any r ď r1 , fr1 ˝ Lr,rF 1 “ Lr,rF 1 ˝ fr and that each fr is
invertible. It follows that the map f is an isomorphism of persistence vector spaces.

Birth-death persistence vector spaces are useful as “building blocks” to construct other
persistence vector spaces. They have the following important uniqueness property.

Proposition 6.1.1. Let

n
à m
à
tVr u “ P pbi , di q and tV r u “ P pbj , dj q.
i“1 j“1

If tVr u « tV r u, then n “ m and the set of pairs pbi , di q with multiplicities is equal to the
set of pbj , dj q with multiplicities.

We state the proposition more succinctly as: direct sums of decompositions into birth-
death pairs are unique up to reordering.
Proof. Let bmin “ mintbi u and bmin “ mintbj u; i.e., the “first birth times”. These values
are given by
bmin “ mintr | Vr ‰ 0u and bmin “ mintr | V r ‰ 0u,
and the isomorphism tVr u « tV r u therefore implies bmin “ bmin (since Vr “ 0 if and only
if V r “ 0). We denote this common value by b.
Next we consider the “first death times”

dmin “ mintdi | bi “ bu and dmin “ mintdj | bj “ bu.

There is also an intrinsic characterization of these values. Indeed,

dmin “ mintr1 | kerpLr,r1 q ‰ 0u and dmin “ mintr1 | kerpLr,r1 q ‰ 0u.

Our isomorphism implies that Lr,r1 is injective if and only if Lr,r1 is as well, so we have
dmin “ dmin and we denote this common value by d.
The work above shows that P pb, dq appears in both decompositions at least once, say
k-times in the first decomposition and k in the second. Our goal is to show that it
appears the same number of times in each decomposition; that is k “ k. Without loss of
generality, we can assume that P pb, dq appears as the first k (respectively k) summands

84
of tVr u (respectively tV r u). Let

k
à k
à
tWr u “ P pb, dq Ă tVr u and tW r u “ P pb, dq Ă tV r u,
i“1 j“1

where each space is a sub-persistence vector space. We once again have intrinsic char-
acterizations; we claim that Wr and W r are each isomorphic to the kernel of the linear
transformation
Lr,d |
imagepLb,r q
image pLb,r q ÝÝÝÝÝÝÝÝÝÑ Vd .
This shows that tWr u « tW r u.
We therefore have
tVr u{tWr u « tV r u{tW r u.
Moreover,
n
à n
à
tVr u{tWr u « P pbi , di q and tV r u{tW r u « P pbi , di q,
i“k`1 i“k`1

and we the proof follows by induction on the number of summands in the direct sum
decompositions.

Finitely Presented Persistence Vector Spaces

We will also need the following definition. A persistence vector space is called finitely
presented if it is isomorphic to a persistence vector space of the form tWr u{imagepf q for
some linear transformation of persistence vector spaces

f : tVr u Ñ tWr u,

where tVr u and tWr u are both finitely generated free persistence vector spaces.

Example 6.1.2. The main example of a finitely presented persistence vector space is a
persistence homology vector space.
Let tXr u denote the set of abstract simplicial complexes associated to filtered simplicial
complex tXr u. For each k, the collection of chain groups tCk pXr qu forms a persistence
C pX q
vector space, with maps Lr,rk1 r induced by inclusions. Consider the map Bk : tCk pXr qu Ñ
tCk´1 pXr qu with pBk qr given by the usual boundary map on Ck pXr q. It is straightforward
C pX q C pXr q
to see that for all r ď r1 , we have pBk qr1 ˝ Lr,rk1 r “ Lr,rk´1
1 ˝ pBk qr , so that Bk defines
a linear map of persistence vector spaces. It follows that the collection tZk pXr qu of
k-dimensional cycles also forms a persistence vector space.
Therefore tP Hk pXr qu “ tCk pXr qu{imagepBk q is a finitely presented persistence vector
space.

85
Persistence Vector Spaces Associated to Adapted Matrices
Let pX, ρq and pY, σq be finite filtered sets. There is a map from the set of pρ, σq-adapted
matrices to the set persistence vector spaces called the θ-correspondence , defined for a
pρ, σq-adapted matrix A by

A ÞÑ θpAq “ tVF pX, ρqr u{imagepfA q.

The is entirely analogous to the θ-correspondence for vector spaces defined in Section
2.4.7.
We have the following facts, which are more-or-less obvious, but will be useful to refer
to.

Proposition 6.1.2. The persistence vector space θpAq is finitely presented. Moreover,
any finitely-presented persistence vector space is of the form θpAq for some adapted matrix
A.

Proposition 6.1.3. Let A be a pρ, σq-adapted matrix, B a pρ, ρq-adapted matrix and C a
pσ, σq-adapted matrix. Then BAC is pρ, σq-adapted and θpAq is isomorphic to θpBACq.

We also have the following example, which will prove very useful in proving the main
theorem of this section.

Example 6.1.3. Let pX, ρq and pY, σq be finite-filtered sets and let A be a pρ, σq-adapted
matrix with the special property that it has at most one nonzero entry in each row and
in each column and that the nonzero entries are all equal to 1. Consider the persistence
vector space θpAq.
Label the elements of Y as ty1 , . . . , yn u. Then for each yj P Y ,
ÿ
fA pφyj q “ axyj φx “ φxj
xPX

for some xj P X. Indeed, the xj is the unique element of X so that axj yj “ 1. We claim
that θpAq can be decomposed as
à à
θpAq « P pρpxj q, σpyj qq ‘ P pρpxq, `8q.
j xRtx1 ,...,xn u

To see this, note that

˜ ¸
à à à à
P pρpxj q, σpyj qq ‘ P pρpxq, `8q “ F‘ F. (6.1)
j xRtx1 ,...,xn u ρpxj qďrăσpyj q xPXrrs,x‰xj
r

On the other hand,

θpAqr “ tVF pX, ρqr u{imageppfA qr q “ spanF trφx s | x P Xrrsu, (6.2)

86
with rφx s “ r0s if and only if x “ xj with σpyj q ă r. Decomposing (6.2) further, we have
à à
spanF trφx s | x P Xrrsu « spantφxj u ‘ spanF tφx u,
ρpxj qďrăσpyj q xPXrrs,x‰xj

which is clearly isomorphic to (6.1). This idea can be used to create an isomorphism of
persistence vector spaces, as in Example 6.1.1.

Row and Column Operations for Adapted Matrices

Let pX, ρq and pY, σq be finite filtered sets and let A be a pρ, σq-adapted matrix. Just as
in the case of unadapted matrices (see Section 2.4.6), there are certain useful row and
column operations for pρ, σq-adapted matrices :
1. Multiply all entries in a row/column by the same nonzero element of F ,

2. Add a nonzero multiple of rpxq to rpx1 q when ρpxq ě ρpx1 q,

3. Add a nonzero multiple of cpyq to cpy 1 q when σpyq ď σpy 1 q

Proposition 6.1.4. The row and column operations for pρ, σq-adapted matrices listed
above preserve the property of being pρ, σq-adapted.
Proof. Let A “ paxy q be a pρ, σq-adapted matrix. The first operation doesn’t change
locations of zeros in the matrix, so it obviously preserves the property. Consider the
second operation of adding λrpxq to rpx1 q, where ρpxq ě ρpx1 q. Assume that ρpx1 q ą σpyq;
we then need to show that the matrix A1 “ pa1xy q obtained after the row operation has
a1x1 y “ 0. This entry is given by a1x1 y “ ax1 y ` λaxy “ 0 ` λ ¨ 0, since ρpxq ě ρpx1 q ą σpyq
implies that ax1 y “ axy “ 0. The proof in the case of the third operation is similar.
We have the following proposition in analogy with Proposition 2.4.7, whose proof is
left as an exercise.
Proposition 6.1.5. Each adapted row and column operation corresponds to an isomor-
phism of a persistence vector space.
Proposition 6.1.6. For any pρ, σq-adapted matrix A, there exists a pρ, σq-adapted matrix
A1 with at most one nonzero entry in each row and each column, the nonzero entry equal
to 1, such that θpAq « θpA1 q.
Proof. Let A be a pρ, σq-adapted matrix for some filtered sets pX, ρq and pY, σq. We will
prove the proposition by constructing A1 inductively. Let y P Y be such that σpyq ď σpyq
for all y such that cpyq does not contain all zeros (assuming that this is a nonempty set,
because if this were not the case then we would already be done). Now let x P X be such
that ρpxq ě ρpxq for all x satisfying axy ‰ 0 (once again, this is a nonempty set; this
time by our assumption on cpyq). We wish to “zero out” entries in cpyq except for the
px, yq-entry using row operations. This is possible because we can add a nonzero multiple
of rpxq to any other row rpxq with ρpxq ě ρpxq and the rows satisfying this condition
are exactly those which potentially contain non-zero entries in the column cpyq! We can

87
similarly “zero out” entries in rpxq besides axy and the result is a pρ, σq-adapted matrix
such that rpxq and cpyq each contain exactly one nonzero entry. Multiplying this row or
column by 1{axy ensures that the nonzero entry is equal to 1.
Now we continue inductively; define ỹ to be the element of Y satisfying σpỹq ď σpyq
for all y such that y ą y and cpyq does not contain all zeros and x̃ to be the element of
X satisfying ρpx̃q ě ρpxq for all x ą x with axỹ ‰ 0. Running the same procedure will
“zero out” the row rpx̃q and column cpỹq as desired without effecting rpxq or cpyq. This
process can be repeated until we obtain a matrix A1 with the desired properties.

6.2 Main Theorem

We are now ready to state our main result. All vector spaces are assumed to be over the
same field F, so the field is supressed from the notation.

Theorem 6.2.1 (Fundamental Theorem of Persistent Homology). For any finitely pre-
sented persistence vector space tVr u, there exists a finite collection of birth-death pairs
pb1 , d1 q, pb2 , d2 q, . . . , pbn , dn q with bi P r0, 8q and di P r0, 8s such that

tVr u « P pb1 , d1 q ‘ P pb2 , d2 q ‘ ¨ ¨ ¨ ‘ P pbn , dn q.

Moreover, the decomposition is unique up to reordering the factors.

Proof. By Proposition 6.1.2, there exist finite filtered sets pX, ρq and pY, σq such that we
can represent tVr u as θpAq for some pρ, σq-adapted matrix A. Proposition 6.1.6 implies
that there is a matrix A1 with at most one entry equal to 1 in each row and in each
column, with all other entries equal to 0, such that θpA1 q « θpAq. We already showed
in Example 6.1.3 that θpA1 q has a decomposition of the desired form. This proves the
existence part of the theorem. The uniqueness part of the theorem follows immediately
from Proposition 6.1.1.

6.3 Barcodes
By applying Theorem 6.2.1, we can now associate to any filtered simplicial complex tXr u
a topological signature called a barcode . To do so, we calculate the persistence homology
of tXr u to obtain a collection of finitely presented persistence vector spaces tHk pXr qu for
k “ 0, 1, 2 . . .. Theorem 6.2.1 tells us that for each k there is a (unique up to reordering)
decomposition
tHk pXr qu « P pb1 , d1 q ‘ P pb2 , d2 q ‘ ¨ ¨ ¨ ‘ P pbn , dn q,
which gives a multiset of birth-death pairs. To form the barcode, we draw a line segment
for each birth-death pair. The barcode is then a collection of line segments, typically
represented in the form
To illustrate this idea, we provide examples below.

88
6.3.1 Examples of Barcodes

6.4 Persistence Diagrams

There is another convenient representation of the persistence homology of a filtered sim-
plicial complex called a persistence diagram . ...

6.4.1 Examples of Persistence Diagrams

6.5 Computing Barcodes

6.6 Other Representations of Persistent Homology
6.7 Exercises
1. Complete the proof of Proposition 6.1.4 by showing that the remaining operation
preserves the property of being pρ, σq-adapted.

2. Prove Proposition 6.1.5.

89
7 Metrics on the Space of Barcodes
(Under Construction)

90
8 Applications (Under Construction)

91
9 Appendix
9.1 Every Vector Space Has a Basis
In this section we provide a proof of Theorem 2.2.1, that every vector space has a basis.
The proof relies on the Axiom of Choice :
Let A be a collection of nonempty, disjoint sets. There exists a set A containing exactly
one element from each set in A.
The Axiom of Choice can’t be proved from any of the other usual axioms of set theory,
and can therefore only be accepted or not accepted. Most modern mathematicians choose
to accept the Axiom of Choice as a basic axiom of set theory. In fact, a major motivation
for accepting it is that it is equivalent to the statement that every vector space has a
basis!
It is well known that the Axiom of Choice is equivalent to Zorn’s Lemma , which is a
statement about ordered sets. To state it, we need to introduce some definitions.
Let A be a set. A partial order on A is a relation ă (i.e. we denote that a, b P A
are related by the notation a ă b) such that a ă a holds for any a P A and for every
a, b, c P A, a ă b and b ă c implies a ă c. An element a P A is a maximum of A if a ă b
implies that a “ b. Let B Ă A. An element a P A is an upper bound on B if b ă a for
every b P B. We say that B is totally ordered if for every a, b P B, either a ă b or b ă a.

Example 9.1.1. The set R has partial order ă“ď. In this case, every subset is totally
ordered.
On the other hand, we could take our set A to be the power set PpRq of all subsets
of R. This set admits a partial order given by inclusion, ă“Ď. Many subsets are not
totally ordered. The whole set PpRq is not totally ordered: r0, 1s and r1, 2s that neither
is a subset of the other. One example of a totally ordered subset of PpRq is the set

tr0, ns | n P Zą0 u.

We can now state Zorn’s Lemma.

Theorem 9.1.1 (Zorn’s Lemma). Let A be a set with partial order ă. If every totally
ordered subset B admits an upper bound, then A has a maximum.

We are now prepared to prove Theorem 2.2.1.

Proof. Let X denote the collection of all linearly independent subsets of V . This set is
partially ordered by inclusion. We wish to apply Zorn’s Lemma to show that X contains
a maximum element B. If such a maximum exists, then it must be a basis. Indeed,
it is linearly independent by definition. It must also be spanning, since for any v P V ,

92
B Ă B Y tvu, there are two possibilities: either v P B in which case it is clear that
v P spanpBq or v R B, so it must be that B Y tvu is not linearly independent and we once
again conclude that v P spanpBq.
It remains to show that we can apply Zorn’s Lemma. Let Y Ă X be a collection of
linearly independent subsets which is totally ordered by inclusion. We wish to show that
it is upper bounded by an element of X . Let Y0 “ YY PY Y . Then for any Y P Y, we
certainly have Y Ă Y0 . It therefore remains to show that Y0 P X ; that is, Y0 is linearly
independent. Any finite linear combination of elements of Y0 can be written as a linear
combination of elements of some set Y P Y, by virtue of the total ordering of Y. Therefore
Y0 is linearly independent.
To show that any linearly independent subset S of V can be extended to a basis, we
can mimic the above argument by replacing X with the set of all linearly independent
subsets of V which contain S.

93
Bibliography
[1] Carlsson, Gunnar. “Topological pattern recognition for point cloud data.” Acta Nu-
merica 23 (2014): 289-368.

[2] Hatcher, Allen. “Algebraic topology. 2002.” Cambridge UP, Cambridge 606, no. 9
(2002).

[3] Hoffman, K. and Kunze, R., Linear algebra. 1971. Englewood Cliffs, New Jersey.

[4] Johnson, Lee W., Ronald Dean Riess, and Jimmy Thomas Arnold. Introduction to
linear algebra. Addison-Wesley Longman, 1993.

[5] Manolescu, Ciprian. “Pin (2)-equivariant Seiberg-Witten Floer homology and the
triangulation conjecture.” Journal of the American Mathematical Society 29, no. 1
(2016): 147-176.

[6] Munkres, James R. Topology. Prentice Hall, 2000.

94
Index

pX, Y q-matrix, 80 bijective, 8

pρ, σq-adapted, 81 domain, 8
α-complex, 77 injective, 8
θ-correspondence inverse, 8
persistence vector spaces, 86 range, 8
vector spaces, 31 surjective, 8
functoriality, 73
affine subspace, 56 Fundamental Theorem of Persistent
Axiom of Choice, 92 Homology, 88
barcode, 88 general position, 56
basis, 18 graph, 38
binary relation, 11 distance, 39
birth-death
pair, 83 homeomorphism, 50
persistence vector space, 83 homology
boundary, 45 singular, 71
with real coefficients, 70
Cauchy-Schwarz Inequality, 32 homotopic maps, 51
closure, 44 homotopy equivalent, 51
compact, 46
connected, 48 image (of a linear map), 27
continuous function, 46 inner product, 31
convex interior, 44
hull, 55 isometry, 49
set, 55
kernel, 27
determinant, 24
direct sum limit point, 44
of vector spaces, 28 linear combination, 18
linear isomorphism, 21
empty set, 6 linear transformation, 20
equivalence class, 11 matrix representation, 24
equivalence relation, 11 nullity of, 27
rank of, 27
field, 15 linearly independent, 18
filtered set, 80
filtered simplicial complex, 81 metric ball, 39
function, 8 metric space, 35

95
 subspace of, 36 power set, 8
metric topology, 43 product, 7
subset, 7
natural numbers, 9 uncountable, 9
norm, 32 union, 6
`p , 33 simplex, 56
open cover, 46 associated to a set, 56
dimension, 56
partial order, 92 edge, 56
path connected, 49 face, 56
persistence diagram, 89 standard, 56
persistence homology vector space, 82 vertex, 56
persistence vector space, 78 simplicial complex
direct sum, 79 abstract, 58
finitely generated, 80 complete, 77
finitely presented, 85 geometric, 57
free, 80 span, 18
linear transformation of, 78 spanning set, 18
quotient, 79 sublevel set, 80
subspace, 79
point cloud, 37 topological properties, 46
topological space, 43
quotient total order, 92
vector space, 29 tree, 39
triangulation, 70
row and column operations, 30
for adapted matrices, 87 vector space, 13, 14
dimension of, 18
separation, 48 free, 60
set, 6 over R, 13
cardinality, 8 over a field F, 16
closed, 40 structure of Rn , 14
countable, 9 vector subspace, 26
countably infinite, 9 Vietoris-Rips complex, 76
difference, 7 Voronoi cell, 77
intersection, 7
open, 40, 43 Zorn’s Lemma, 92

96