ETH Zürich

DUALITY
Bachelor’s Thesis

written by
Franco de Bardeci

under the supervision of

Prof. Richard Pink

Abstract
This thesis has the purpose of illustrating the Principle of Dual-
ity. We present various examples in different fields of mathematics,
including Geometry, Algebra, Graph Theory and Probability Theory.

February 16, 2016

Contents

1 Introduction 3

2 Projective Geometry 3

3 Vector Spaces 6

4 Topological Groups 8

5 Category Theory 14

6 Plane Graphs 14

7 Percolation Theory 21

2
1 Introduction
Duality is a principle that appears throughout mathematics. It can manifest
itself in great elegance and be very useful at the same time. It often starts
out as a certain symmetry at the axiomatic level, extending across the theory.
In many fields of mathematics duality plays a fundamental role that might
also remain present in the applications of those fields.
An undergraduate student may encounter this principle many times, for
example when considering dual vector spaces or dual norms. Also perhaps,
in projective geometry. Sometimes without explicit mention of the fact that
a duality is involved, as for the case of the Fourier transform.
There is no precise definition of what a duality is, but the following is the
prototype for many of them. Say we have a mathematical structure that in
particular involves the sets A and B, and is defined by a set of axioms. Each
axiom is a statement that may refer to the sets A and B. For each axiom we
consider the dual statement, which is obtained by interchanging the roles of
A and B. This means that whenever A is mentioned, it is replaced by B, and
vice versa. For example, if at some point one of the axioms reads: “For each
element of A there is an element of B such that...”, then the dual statement
reads: “For each element of B there is an element of A such that...”. One
has a duality, if the dual statement of each of the axioms is derivable from
the axioms. When this is the case, any statement implies its dual. Thus, by
proving a theorem, unless the dual statement of the theorem is identical to
the theorem itself, we get a second theorem for free! In the following sections
we consider various examples of dualities. The sets A and B will for example
consist of the points and lines in the projective plane, or the vectors and
linear functionals of a finite vector space.

2 Projective Geometry
We begin by considering the Principle of Duality on the projective plane. A
projective plane consists of a set of points and a set of lines, together with
a relation between the two, such that a certain set of axioms is satisfied.
This relation is called the incidence relation, and two elements satisfying
this relation are called incident. Intuitively, a point and a line are incident
if the point lies on the line. The following set of axioms for this geometry
can be found in Coxeter [1], Chapter 3. We introduce them along with the
relevant terms.

3
Axiom 2.1. Any two lines are incident with at least one point.
Axiom 2.2. Any two distinct points are incident with exactly one line.
From these first two axioms, it follows that also two distinct lines are
incident with exactly one point. For any two distinct lines l and m, we can
therefore define their intersection as being the unique point incident with
both lines, which we denote by l · m. Analogously, for any two distinct points
P and Q, we define the line P Q, as the unique line incident with P and Q.
A set of points is called collinear if there is a line such that every point
in the set is incident with this line. A set of lines is called concurrent if there
is a point such that every line in the set is incident with this point.
Axiom 2.3. There exist four points of which no three are collinear.
Four points together with six distinct lines joining each pair of points are
called a complete quadrangle. Pairs of these lines that do not meet at one of
those four points are called opposite. We can form three different opposite
pairs, the points where these pairs intersect are called the diagonal points.
Axiom 2.4. The three diagonal points of a complete quadrangle are never
collinear.
The set of points incident with a given line is called the range of that
line. The set of lines incident with a given point is called the pencil of that
point. Given a line and a point which are not incident, there is a natural
one-to-one correspondence between the elements of the range and the ele-
ments of the pencil, corresponding elements being incident. Consider a finite
sequence whose elements alternate between lines and points, such that any
two consecutive elements are never incident, and any two elements that dif-
fer by two positions are distinct. Considering the one-to-one correspondence
between range and pencil of consecutive elements, we can define a transfor-
mation from the range or pencil of the first element to the range or pencil of
the last element. Such transformation is called a projectivity.
Axiom 2.5. If a projectivity leaves invariant three distinct points on a line,
it leaves invariant every point on the line.
Three non-collinear points A, B, C together with the lines AB, BC, and
CA are called a triangle denoted by ABC.
Axiom 2.6. Let ABC and A0 B 0 C 0 be triangles such that all six points, as well
as the six lines are different. If the lines AA0 , BB 0 , and CC 0 are concurrent,
then the points AB · A0 B 0 , BC · B 0 C 0 , and CA · C 0 A0 are collinear.

4
 The last axiom is also known as Desargues’s Theorem. Although we state
it as an axiom here, it might be a theorem if another set of axioms was chosen.
The reader may have noticed that a certain symmetry emerges. As noted
before, Axiom 2.2 still holds if we interchange the words points and lines.
Namely, if two distinct lines are given, Axiom 2.1 asserts that there is at
least one point of intersection. If there was more than one point, then by
Axiom 2.2 the lines would be equal, and therefore the point of intersection is
unique. Also, we observe that the definition of collinear and concurrent are
related in the same way. This relation is precisely the duality, and pairs of
statements, definitions or objects that are related in this way are said to be
dual. If we were able to prove the dual statement of each axiom, we would
know that the dual statement of any theorem must also hold.
This is indeed the case. The dual of Axiom 2.1 follows from Axiom 2.2
and the fact that more than one point exists, which follows from Axiom 2.3.
For the dual of Axiom 2.3, consider the four points given by this axiom.
Choose an ordering of the points and join consecutive points with a line as
well as the first point with the last point. We obtain four lines. First note
that the lines are all distinct. For if two pairs were to be joined by the same
line, there would be at least three collinear points. Suppose now that three of
the lines are concurrent. For each line consider the two points that gave rise
to them. The only way not to count more than four points is that the three
lines share one of these points. But by the way these lines were constructed,
such a point can be shared by two lines at most.
Instead of dealing with a complete quadrangle, the dual of Axiom 2.4 deals
with a set of four lines, such that every pair of lines intersects at one of six
distinct points. This is called a complete quadrilateral. The diagonal points
correspond to the three other lines that join pairs of this six points, they are
called the diagonal lines. We have to show that these are not concurrent.
Consider a complete quadrilateral formed by the lines a, b, c, d. Assume the
diagonal lines all meet at a point P . Say the quadrilateral has the vertices A,
B, C, D, E, F . The vertices A, B, C, D define a complete quadrangle. Its
edges are precisely the lines a, b, c, d, and two of the three diagonal lines of
the quadrilateral. Its diagonal points are E, F and P . The diagonal line of
the quadrilateral that is not an edge of the quadrangle is EF . The point P
must also lie on this line, and therefore the diagonal points of the quadrangle
are collinear, which is impossible by Axiom 2.4.
To prove the dual of Axiom 2.5, assume there is a projectivity on a pencil
that leaves invariant every of three lines but not the entire pencil. Such a
projectivity is given by a sequence a1 , a2 , . . . , an , a1 . By shifting the sequence

5
to a2 , . . . , an , a1 , a2 , we get a counterexample to Axiom 2.5.
The dual statement of Axiom 2.6 is precisely its converse. We prove it.
Say we have the triangles ABC and A0 B 0 C 0 as in the axiom. Set

P := AB · A0 B 0 , Q := BC · B 0 C 0 , R := CA · C 0 A0 .

By assumption these points are collinear. Consider the triangles AA0 P and
CC 0 Q. The lines AC, A0 C 0 , and P Q meet at the point R. By Axiom 2.6
this means that B, B 0 and S := AA0 · CC 0 are collinear. Therefore the lines
AA0 , BB 0 , and CC 0 meet at the point S.
This completes the proof that all dual statements of the axioms are true.
Any theorem will therefore imply its dual, and hence every time we prove a
theorem, unless it is self dual, we also get a second theorem. We illustrate
this fact. Six distinct points and six distinct lines, where the points can be
arranged in a sequence such that pairs of consecutive points, as well as the
first and the last, are joined by one of the lines, and each line joins one of
those pairs, are called a hexagon. The points of the hexagon are called its
vertices, and the lines of the hexagon are called its sides. Two of the vertices
that differ by three places in the sequence are called opposite. Likewise, two
of the sides that join pairs that differer by three places are called opposite.

Theorem 2.7. If the vertices of a hexagon lie alternately on two lines, the
points of intersection of the three pairs of opposite sides of the hexagon are
collinear.

This is Pappus’s Theorem. For a proof see Coxeter [1], Section 4.4. By
duality one obtains:

Theorem 2.8. If the sides of a hexagon alternately intersect two points,

the lines that join the three pairs of opposite vertices of the hexagon are
concurrent.

3 Vector Spaces
Throughout let V be a vector space over a field K. A linear functional is a
homomorphism from V to K. The set of all linear functionals, with pointwise
addition and pointwise scalar multiplication, also becomes a vector space
over K. It is called the dual space of V , and is denoted by V ∗ . How is here
the term dual related to the previous discussion? Just as before points and

6
Figure 1: Two hexagons. The dark lines and dark points are their sides and
vertices. The one the left satisfies the conditions in Pappus’s Theorem. The
one on the right, the conditions in the dual theorem.

lines were interchangeable, we will see that the same kind of relation holds
between vectors and linear functionals. Consider the map
e : V × V ∗ → K, (v, f ) 7→ f (v).
By definition, for each f ∈ V ∗ the map V → K, v 7→ f (v) is a linear func-
tional, and every linear functional can be obtained this way. Observe that
for each v ∈ V the map λv : V ∗ → K, f 7→ f (v) is a linear functional as well.
This already shows some symmetry. Define
α : V → V ∗∗ , v 7→ λv .
This is an injective isomorphism. We will show that for finite dimensional
vector spaces, it is an isomorphism. Therefore, in this case, the symmetry
is perfect: V ∗ is the space of linear functionals on V by definition, and V is
naturally isomorphic to the space of linear functionals of V ∗ .
Henceforth, let v1 , ..., vn form a basis for V . For a fixed index i ∈ {1, .., n},
consider the map fi : V → K that takes a vector to its i-th coordinate. It is
a linear functional and for any i, j ∈ {1, .., n} we have: fi (vj ) = δij .
Lemma 3.1. The maps f1 , ..., fn form a basis for V ∗ .
Proof. Let f be any element of V ∗ . By the homomorphism property for any
v ∈ V we have:
f (v) = f (f1 (v)v1 ) + ... + f (fn (v)vn ) = f (v1 )f1 (v) + ... + f (vn )fn (v).

7
Therefore f can be written as a linear combination in f1 , ..., fn . Further, if
b1 f1 + ... + bn fn = 0, then in particular we have:

b1 f1 (vi ) + ... + bn fn (vi ) = 0, ∀i ∈ {1, ..., n}.

This implies bi = 0 for every i ∈ {1, ..., n}, and hence f1 , ..., fn are lineary
independent.

Proposition 3.2. The above described map α is an isomorphism from V to

V ∗∗ .

Proof. It is directly verifiable that α is a homomorphism. Moreover, by

Lemma 3.1, f1 , ..., fn form a basis for V ∗ . By the same process we imple-
mented to obtain this basis from v1 , ..., vn , a basis for V ∗∗ can be obtained
from f1 , ..., fn , say µ1 , ..., µn . We have µi (b1 f1 + ... + bn fn ) = bi , for every
i ∈ {1, ..., n}. Also

λvi (b1 f1 + ... + bn fn ) = b1 f1 (vi ) + ... + bn fn (vi ) = bi , ∀i ∈ {1, ..., n}.

Therefore µi and λvi are identical and thus α sends basis elements to basis
elements.

Note that even though when proving that α is an isomorphism we made

use of bases, from its definition it is clear that α does not depend on any
choice of basis. Note also that by Lemma 3.1 there is an isomorphism from
V to V ∗ . However, this isomorphism may depend on the basis chosen.

4 Topological Groups
A topological group is a group with a topology such that the law of com-
position, as well as the inversion operation, are continuous functions. A
locally compact abelian group, LCA group for short, is a topological group
whose group structure is abelian and whose topology is locally compact and
Hausdorff. Note that the latter is not represented in the name.
In a topological group, translation by an element of the group defines a
homeomorphism from the topological group to itself. The open neighbor-
hoods of the identity therefore describe the whole topology, other open sets
being translations of them. To show that a homomorphism from a topolog-
ical group to another is continuous, it suffices to show that it is continuous
at the identity.

8
 Throughout let G be an LCA group and T the topological group of com-
plex numbers of absolute value 1 under multiplication. A character is a
continuous homomorphism from G to T. The set of all characters becomes
a group by defining the composition of two characters to be the character
obtained by pointwise multiplication. This group, together with a topology
we describe next, is called the dual group of G and is denoted by Ĝ.
Let X and Y be topological spaces, and F a set of functions from X
to Y . For a compact set K of X, and an open set U of Y , we define S(K, U )
to be the set of all elements f of F such that f (K) ⊂ U . The collection of
all sets of the form S(K, U ) is subbasis for a topology on F . It is called
the compact-open topology. The dual group Ĝ is given the compact-open
topology.

Lemma 4.1. Let X be a topological space and Y a metric space. Let K be

a compact set of X, let U be an open set of Y , and f : X → Y a continuous
map that sends K to U . There is an  > 0 such that for any point of f (K)
the open ball of radius  around that point is contained in U .

Proof. For y ∈ Y define r(y) to be the supremum of all radii r0 , such that
Br0 (y) ⊂ U . Suppose no such  exists. Then there is a sequence (yn )n∈N in
f (K) such that r(yn ) → 0 for n → ∞. Since f (K) is compact (yn )n∈N has a
subsequence that converges to some point of f (K), say y0 . But r(y0 ) > 0,
and hence we have a contradiction.

Proposition 4.2. The dual group is a topological group.

Proof. Let φ : Ĝ × Ĝ → Ĝ be the map that takes (χ1 , χ2 ) to χ1 χ−12 . It suffices

to show that φ is continuous. Consider an element of the subbasis, S(K, U )
say. Let (χ1 , χ2 ) be any element of the preimage of S(K, U ) under φ. We
will show that there is an open set in Ĝ × Ĝ that contains (χ1 , χ2 ) and is
mapped into S(K, U ) by φ.
As χ1 χ−1
2 is continuous, the set χ1 χ−1 2 (K) is compact. By Lemma 4.1
there is an  > 0 such that a ball of radius  around any point of χ1 χ−1 2 (K) is
contained in U . Let i ∈ {1, 2}. As χi is continuous, for any point x in G there
is a neighborhood such that the image under χi of its closure is contained in
a ball of radius /4. Let (Vi,j )j∈Ji be a finite collection of such neighborhoods
covering K. Let Ki,j be the intersection of the closure of Vi,j with K, and
Ui,j a ball of radius /4 that contains the image of Ki,j under χi . Consider
the set: \ \
S(K1,j , U1,j ) × S(K2,j , U2,j ).
j∈J1 j∈J2

9
It is open in Ĝ × Ĝ and contains (χ1 , χ2 ). Further, it is mapped into S(K, U )
by φ, since for every (ξ1 , ξ2 ) in Ĝ × Ĝ and every x ∈ G we have:

|χ1 χ−1 −1
2 (x) − ξ1 ξ2 (x)| ≤ |χ1 (x) − ξ1 (x)| + |χ2 (x) − ξ2 (x)|.

Note that we have not used the fact that G is locally compact in the proof.
Using this fact, we can prove that Ĝ is an LCA group. It follows directly
from the definition that Ĝ is abelian. To see that it is Hausdorff, consider
two different characters χ1 and χ2 . They must differ at some element of G,
say x. As T is Hausdorff, we can separate χ1 (x) from χ2 (x) with two open
sets, U1 and U2 . The open sets S({x}, U1 ) and S({x}, U2 ) then separate
χ1 from χ2 . To show local compactness, we use a generalization of Ascoli’s
Theorem which characterizes compact subspaces in certain function spaces.
Before we state the theorem, we introduce a required definition.
For a topological space X and a metric space Y , a set of functions
from X to Y is called equicontinuous at x ∈ X if for every  > 0 there is
a neighborhood V of x such that for every function f in the set, we have
f (V ) ⊂ B (f (x)). If the set of functions is equicontinuous at every x ∈ X,
then its called equicontinuous. We define C(X, Y ) to be the set of all con-
tinuous functions from X to Y with the compact-open topology, and T+ to
be the of elements of T that have positive real part.
Theorem 4.3. Let X be a locally compact space and Y a compact metric
space. A subset in C(X, Y ) has compact closure if and only if it is equicon-
tinuous.
A proof of the theorem can be found in Bourbaki [5]. It is obtained by
combining Theorem 2 in Section 2 of Chapter 10, and Theorem 2 in Section 3
of the same chapter. Alternatively, it can be found in von Querenburg [6],
Theorem 14.22.
Proposition 4.4. The dual group is an LCA group.
Proof. It sufficient to show that the identity element 1 of Ĝ has a compact
neighborhood. Let C be a compact neighborhood of the identity element e
of G. Consider the set S(C, T+ ), it is an open neighborhood of 1. We show
that it is equicontinuous, then the claim follows by Theorem 4.3.
Showing that S(C, T+ ) is equicontinuous amounts to showing that it is
equicontinuous at the identity e. Let ϕ be the branch of the argument func-
tion on C that takes values in (−π, π]. For n ∈ N let Un be the set of all

10
z ∈ T such that |ϕ(z)| < π/n. Let φ : G × G → G be the composition map,
the map that takes (χ1 , χ2 ) to χ1 χ2 . Let V be an open neighborhood of e
in G that is contained in C. As φ is continuous, φ−1 (V ) is open, also it
is a neighborhood of (e, e). Therefore, there is a standard basis element of
the product topology of G × G, say W1 × W2 , that is contained in φ−1 (V )
and is a neighborhood of (e, e). The set W1 ∩ W2 has the property, that the
composition of any two of its elements is in C. By iterating this process,
one can obtain for any n ∈ N, a neighborhood Vn of e, such that the com-
position of any n elements of Vn is in C. Any open neighborhood of 1 in
T contains a set of the form Un . We show that Vn is mapped into Un by
any character in S(C, T+ ), as any character maps e to 1, this shows that
S(C, T+ ) is equicontinuous. Assume the opposite, then there is a character
χ in S(C, T+ ) and x ∈ Vn such that |ϕ(χ(x))| ≥ π/n. There is k < n such
that χ(x)k ∈/ T+ . But χ(x)k = χ(xk ) and therefore, because xk is an element
of C, we have χ(x)k ∈ T+ .

In analogy to what we did for vector spaces we consider the map:

e : G × Ĝ → T, (x, χ) 7→ χ(x).

Lemma 4.5. The map e is continuous.

Proof. Consider the preimage of an arbitrary open set U under this map.
Choose a point on the preimage, say (x0 , χ0 ). Consider the open set χ−1
0 (U ).
Because G is locally compact we can choose a compact set K, and an open
set V such that {x0 } ⊂ V ⊂ K ⊂ χ−1
0 (U ). Then V × S(K, U ) is an open set
in G × Ĝ and is mapped into U by e.

By Proposition 4.4 the dual group is also an LCA group, we can therefore
ˆ . Consider the map e. For x ∈ G, the map
consider the dual of the dual: Ĝ
λx : Ĝ → T, χ 7→ χ(x) is continuous. It is also a homomorphism and thus
ˆ . We therefore have a well defined map:
an element of Ĝ
ˆ , x 7→ λ .
α : G → Ĝ x

It is called the Pontryagin map.

Theorem 4.6. The Pontryagin map is a homeomorphism and an isomor-

phism.

This is the Pontryagin Duality Theorem. In contrast to its analog for

finite vector spaces, it is hard to prove and we will not do so here. It was

11
first proven by Soviet mathematician Lev Pontryagin, although in less gener-
ality. His results were reported at the International Mathematical Congress
in Zurich in 1932, see Pontryagin [11]. There are many proofs available. For
a proof involving abstract Fourier analysis see Rudin [8], Chapter 1, or Deit-
mar and Echterhoff [9], Chapter 3. However, the following partial result is
not difficult.

Proposition 4.7. The Pontryagin map is a continuous homomorphism.

Proof. For x1 , x2 ∈ G we have λ(x1 +x2 ) = λx1 + λx2 so that α is a homomor-

phism. We prove that it is also continuous. By definition, the sets of the
form S(K, U ), where K is a compact set of Ĝ and U is an open set of T,
form a subbasis for the topology on Ĝ ˆ . For same K and U , let Ŝ(K, U ) be
the set of all x ∈ G that are taken into U by any character in K. Then the
preimage of S(K, U ) under α is precisely Ŝ(K, U ). We show that every set
of the form Ŝ(K, U ), for K a compact set of Ĝ and U an open set of T, is
open. Take any point in Ŝ(K, U ), say x0 . Consider the preimage of U under
the map e. It is open. For any character χ in K we can therefore choose a
standard basis set of the product topology around the point (x0 , χ) that is
contained in e−1 (U ). Such a basis set consists of a product of two open sets,
say Uχ × Vχ . The set of all sets of the form Vχ , where χ ∈ K, is an open
cover of K. Choose a finite subcover indexed by I ⊂ K. Then ∩χ∈I Uχ is an
open neighborhood of x0 contained in Ŝ(K, U ).

Consider the group of integers Z with the discrete topology. It is an LCA

group. We calculate its dual. A homomorphism from Z to T is already
defined by specifying the target of the element 1. As Z has the discrete
topology, any such homomorphism is continuous. We therefore have a one-
to-one correspondence between Ẑ and T, defined by the map that takes a
character to the target of the element 1. It is an isomorphism. We show that
it is also a homeomorphism. For an arbitrary open set U of T, the open set
S({1}, U ) corresponds to U , hence the map is continuous. As the topology
of Z is discrete, its compact sets are exactly its finite sets. The sets of the
form S({a}, U ) therefore form a subbasis for the topology of Ẑ. A character
χ in S({a}, U ) must take 1 to an |a|-th root of χ(a), or to the inverse of such
a root. It follows that we can find an open set U 0 of T such that S({1}, U 0 )
contains χ and is contained in S({a}, U ). As S({1}, U 0 ) corresponds to U 0 ,
our correspondence also has continuous inverse, and thus T is the dual of Z.
By the Pontryagin Duality Theorem we just have also shown that Z is the
dual of T.

12
 As another example, R turns out to be self dual. In abstract Fourier
analysis, LCA groups are given a certain measure, and one considers complex
valued Lebesgue integrable functions on them. The Fourier transform of such
a function is then a function defined on the dual group. When taking the
group to be T, one obtains the special case of Fourier series. When taking
R, one obtains the classical Fourier transform. For more on abstract Fourier
analysis see for example Rudin [8] or Deitmar and Echterhoff [9].
As the examples of Z, T and R hint at, discrete and compact LCA groups
are in duality. More precisely:

Proposition 4.8. If G is compact then Ĝ is discrete.

Proof. It suffices to show that the set containing only the trivial character is
open in Ĝ. Consider the open set S(G, T + ). Let χ be a character in that set.
Then χ(G) is a subgroup of T contained in T+ . Let ϕ be the branch of the
argument function on C that takes values in (−π, π]. Let a be any element
of T+ different from 1. We have 0 < |ϕ(a)| < π/2. There is an integer n such
that |ϕ(an )| = n|ϕ(a)| ∈ (π/2, π), and therefore an ∈ / T+ . Hence, the only
subgroup of T contained in T+ is the trivial subgroup. Therefore S(G, T + )
contains the trivial character alone.

Proposition 4.9. If G is discrete then Ĝ is compact.

Proof. By Tychonoff’s Theorem, the set of all functions from G to T endowed

with the product topology is compact. The fact that G is discrete implies that
the subbasis in the definition of the compact-open topology is the same as the
standard subbasis of the product topology. Also because G is discrete, every
function from G to T is continuous and thus Ĝ is the set of homomorphisms
from G to T. This is a closed subset of the set of all functions from G
to T endowed with compact-open topology. For if a character χ is not a
homomorphism, there are a and b in G such that χ(ab) 6= χ(a)χ(b). For a
sufficiently small r we then have that




S {ab}, Br (χ(ab)) ∩ S {a}, Br (χ(a)) ∩ S {b}, Br (χ(b))

contains no homomorphism.

By the Pontryagin Duality Theorem the converse of each of the last two
propositions holds as well.

13
5 Category Theory
A category is a collection of objects, together with a collection of arrows,
and three operations that satisfy certain axioms. The first operation takes
an arrow, say f , and returns an object called the domain of the arrow,
denoted by d(f ). The second operation takes the arrow, and returns an object
called the codomain of the arrow, denoted by cod(f ). The third operation, a
composition of arrows, takes an ordered pair of arrows, say (f, g), such that
cod(f ) = d(g), and returns an arrow, denoted by g ◦ f . The axioms are the
following.

Axiom 5.1. d(g ◦ f ) = d(f ).

Axiom 5.2. cod(g ◦ f ) = cod(g).

Axiom 5.3. The composition of arrows is associative.

Axiom 5.4. For every object, there is an arrow that has the object as do-
main and codomain, such that the arrow is a left and right identity for the
composition of arrows.

Here the duality emerges when interchanging the operations for the do-
main and codomain, as well as reversing the order of the composition op-
eration. This means, we write cod(f ) for d(f ), d(f ) for cod(f ), and f ◦ g
for g ◦ f . We can proceed as we did for the projective plane, and prove the
dual of every axiom, although this time there is almost nothing to prove.
Axiom 5.1 is the dual of Axiom 5.2, and vice versa. Axiom 5.3 can be rewrit-
ten as: For any three objects: f , g and h, that satisfy cod(f ) = d(g) and
cod(f ) = d(h), we have: (h ◦ g) ◦ f = h ◦ (g ◦ f ). Which is the same as: For
any three objects: f , g and h, that satisfy c(f ) = cod(g) and d(f ) = cod(h),
we have: f ◦ (g ◦ h) = (f ◦ g) ◦ h. Likewise for Axiom 5.4.

6 Plane Graphs
A simple polygonal curve is a subset of the real plane that is the union of
finitely many straight line segments which can be arranged in a sequence
such that any pair of these straight line segments, with exception of the pair
consisting of the first and the last segment, intersects if and only if the two
are consecutive, and intersections only occur at endpoints of the straight line
segments.

14
 A plane multigraph consists of a finite set of points in the real plane,
called vertices, and a finite set of simple polygonal curves, called edges, such
that the endpoints of any edge are vertices, and no two edges intersect at
a point that is not an endpoint of both, nor an edge intersects a vertex
at a point that is not an endpoint. Consider the complement of the union
of all edges and all vertices of a plane multigraph. It divides the plane in
finitely many connected components. Those are called the faces of a plane
multigraph. The sets of vertices and edges form an underling multigraph.
A plane multigraph is said to be connected if the underlying multigraph is
connected. Likewise, we adopt other terms like path or cycle. We call the set
of all points of an edge that are not its endpoints, the interior of the edge.
Throughout let G be a connected plane multigraph, and V , E, and F its
sets of vertices, edges, and faces respectively. A plane dual for G, is a plane
multigraph G∗ with vertices V ∗ , edges E ∗ , and faces F ∗ , such that there
is a one-to-one correspondence between F and V ∗ , as well as a one-to-one
correspondence between E and E ∗ , such that the following conditions are
satisfied:

• Any vertex in V ∗ is contained in the face in F to which it corresponds.

• Any edge in E intersects its corresponding edge in E ∗ at exactly one

point, this point is not an endpoint of a straight line segment of either
edge, and the edge does not intersect any other edge of E ∗ .

Figure 2: Example of a connected plane multigraph and a dual.

15
 Any connected plane multigraph has a plane dual. To construct one, on
each face f choose a point f ∗ . From the fact that each face is an open,
connected subset of the plane, it follows that for each edge of G one can
choose a corresponding edge, such that the second condition is satisfied. The
endpoints are already determined.
A vertex and an edge are said to be incident, if the vertex is one of the
endpoints of the edge. Two different edges are said to be incident if they
share an endpoint. Choose a point on the interior of an edge. We can put
an open ball of sufficiently small radius around that point, such that the
intersection of the ball with the complement of the graph in the plane has
two connected components, each of which is contained in one of the faces,
possibly both in the same. These faces do not depend on the particular point
chosen on the interior of the edge, and are said to be incident with the edge.
The number of faces incident with a given edge is called its degree. An edge
can have degree 1 or 2. The set of edges that are incident with a given face,
is called the frontier of that face.
The plane with the vertices of G removed is connected, therefore we can
join any two vertices of V ∗ with a curve that does not intersect any vertex
of G. We can modify any such curve to intersect any edge of G at most once,
and also, to enter a face at most once. Then, the edges that correspond to
edges that the curve intersects is a path in G∗ that joins the two vertices.
Therefore G∗ is connected.
Lemma 6.1. Let v0 be a vertex of G, let e0 be an edge incident with v0 ,
and e∗0 the edge of G∗ corresponding to e0 . If e0 has two different endpoints,
then there is a cycle in G∗ that contains e∗0 such that every edge of the cycle
corresponds to an edge incident with v0 .
Proof. If e∗0 has only one endpoint, then it forms such a cycle. Assume
therefore e∗0 has two different endpoints. Put a closed ball of sufficiently
small radius around v0 such that the ball only intersects the edges of G that
are incident with v0 , these intersections are straight line segments, and the
ball does not intersect any other vertex of G, nor any edge of G∗ . Consider the
straight line segments obtained by the intersection of the ball and the edges.
Arrange them in a sequence s0 , s1 , ..., sn such that consecutive segments lie
beside each other, in the sense that they can be joined by a path in the plane
that stays inside the ball and does not intersect any other segment, and such
that the sequence starts with a segment that belongs to e0 . Then s0 and sn
also lie beside each other. As e0 has two different endpoints by assumption,
s0 is the only segment in the sequence that belongs to e0 . The sequence has

16
at least two elements: if s0 was the only such segment, then the vertices of e∗0
could be joined by a path in the plane traveling alongside the edges e∗0 and e0
that does not intersect G, which by the definition of a dual plane multigraph
contradicts the assumption that the vertices are different. For i ∈ {0, ..., n}
let ϑ(si ) be the edge of G to which the straight line segment si belongs, and
ϑ(si )∗ the corresponding edge in G∗ .
For two different i, j ∈ {0, ..., n}, if si and sj lie beside each other, then
there is an endpoint v1 of ϑ(si )∗ and an endpoint v2 of ϑ(sj )∗ , such that both
endpoints can be joined by a path in the plane that travels alongside the
edges ϑ(si )∗ , ϑ(si ), ϑ(sj ), and ϑ(sj )∗ , which does not intersect any edge of G.
Therefore v1 and v2 are vertices of G∗ that lie on the same face of G, and
thus we have v1 = v2 . So ϑ(si ) and ϑ(sj ) are incident.
In particular e∗0 is incident with ϑ(s1 )∗ and ϑ(sn )∗ . Further if e∗0 and ϑ(s1 )∗
share a given vertex of e∗0 , then e∗0 and ϑ(sn )∗ share the other vertex of e∗0 .
Consider the edges ϑ(s1 )∗ ,...,ϑ(sn )∗ . By the previous paragraph consecutive
edges are incident, therefore the union of those edges forms a connected
subgraph of G∗ . It contains both endpoints of e∗0 , but it does not contain e∗0
itself. Take a path in this subgraph that connects both endpoints, adding e∗0
to the path yields a cycle with the required properties.

Proposition 6.2. The plane multigraph G is a dual for G∗ .

Proof. The second condition in the definition of a dual plane multigraph is

self dual. It is therefore left to show that every face of G∗ contains exactly
one vertex of G.
Assume there are two vertices v1 and v2 of G that lie on same face of G∗ .
As G is connected, there is a path from v1 to v2 in G. Consider the first
edge of the path starting from v1 . Consider its corresponding edge in G∗ . By
Lemma 6.1, there is a cycle in G∗ that contains this edge, such that every
edge in the cycle corresponds to an edge that is incident with v1 . This cycle
divides the plane into two connected components. The vertices v1 and v2
must be in the same of those connected components. With the first edge
the path exits the component in which v1 is contained. In order to reach
v2 it must enter again. But any edge that crosses the cycle to enter this
component, arrives at v1 . We therefore have a contradiction and thus a face
of G∗ contains at most one vertex of G.
Also, every face of G∗ that is not the whole plane, must have an edge of
G∗ on its boundary. The corresponding edge of G to that edge has a vertex
on the face.

17
 Let S be a plane multigraph that has the same vertices as G and its edges
are a subset of the edges of G. Let S 0 be the plane multigraph that has the
same vertices as G∗ and its edges are those edges of G∗ that correspond to
edges of G that are not edges of S.
Lemma 6.3. The plane multigraph S is connected if and only if S 0 has no
cycles.
Proof. By the Jordan Curve Theorem, a cycle in S 0 divides the plane into
two connected components. An edge of G that corresponds to an edge of
the cycle, has one endpoint in each connected component. Therefore they
cannot be joined by a path that does not meet the given cycle. Thus if S 0
has a cycle, then S is not connected.
Conversely, assuming that S is not connected, we construct a cycle in S 0 .
For the purpose of this proof we call an edge of G∗ that corresponds to an edge
of G which connects two vertices that are in different connected components
of S, a separating edge. As S is not connected, there is at least one separating
edge, say e∗0 . If the endpoints of e∗0 are equal, this edge is already a cycle
in S 0 . Assume therefore the endpoints of e∗0 are distinct. We show that for
each endpoint of e∗0 there is another separating edge that is incident with
the endpoint. Given this, a cycle can be found by iteration. Let v0 be any
endpoint of e∗0 . Let e0 be the edge of G that corresponds to e∗0 . By the dual
statement of Lemma 6.1, there is a cycle in G that contains e0 such that every
edge in the cycle corresponds to an edge that is incident with v0 . Remove
e0 from the cycle. We obtain a path in G that connects the endpoints of e0 .
Consider the sequence of vertices of this path. The first and the last vertices
are the endpoints of e0 , and these belong to different connected components
of S. Hence, there must be a pair of consecutive vertices in the sequence that
are in different connected components of S. The edge in the path joining this
pair of vertices corresponds to a separating edge that is incident with v0 .
If S is connected and has no cycles it is called a spanning tree of G. By
Lemma 6.3 and by duality, S is a spanning tree for G if and only if S 0 is a
spanning tree of G∗ .
A consequence of this, is that a connected plane multigraph satisfies Eu-
ler’s Formula. Let T be a spanning tree for G. As the number of edges in a
tree is always one less than its number of vertices we have:

[ Number of edges of T ] + 1 = |V |,

[ Number of edges of T 0 ] + 1 = |F |.

18
The number of edges in T and T 0 together, sum to |E|. Thus, putting both
equations together we obtain:

|V | − |E| + |F | = 2.

Under the definition for plane multigraphs, two different plane multi-
graphs might still be such that we would consider them to be essentially the
same. For example, when one is a translation of the other, or obtainable after
certain deformations. One can introduce many different equivalence relations
that capture this idea. But is there an equivalence relation that is compati-
ble with duality? More precisely, we want an equivalence relation such that
assigning a dual to an equivalence class by considering any representative,
any dual of the representative, and finally the equivalence class of that dual,
yields a well defined map. A first try could be to consider only the underlying
multigraph structure, whithout the embedding into the plane, and therefore
define two plane multigraphs to be equivalent if there is a one-to-one corre-
spondence between their vertices and a one-to-one correspondence between
their edges such that the incidence relations are conserved. However, Fig-
ure 3 shows that two plane multigraphs can be equivalent under this relation
but have duals that are not. In a second attempt, we could ask for these
correspondences to be induced by a homeomorphism of the plane to itself, so
that the homeomorphism would map every vertex to its corresponding ver-
tex, and every edge to its corresponding edge. This is not adequate either.
Figure 4 shows that a graph can have two duals that are not equivalent under
this relation. It turns out, that this would be actually a precise equivalence
relation for our purpose if instead of drawing the graphs on the plane, we
would do so on a sphere! We only need the topological properties of a sphere
though. So we can keep our definitions for plane multigraphs on the plane,
only adding a point at infinity.
We give R2 ∪ {∞} the topology of the Riemann sphere, this means that
a subset that does not contain infinity is open if it is open in R, and a subset
that contains infinity is open if its complement is compact in R. We define
two plane multigraphs to be equivalent if there is a homeomorphism from
R2 ∪ {∞} to itself that maps vertices to vertices and edges to edges.

Proposition 6.4. Let G1 and G2 be connected plane multigraphs. Then G∗1

is equivalent to G∗2 if and only if G1 is equivalent to G2 .

We sketch a proof for the proposition. By duality it suffices to prove

only one direction. Say φ is a homeomorphism from R2 ∪ {∞} to itself that

19
Figure 3: Counting the degree of the vertices of the dashed graphs, reveals
that their elements cannot paired such that the incidence relations are con-
served.

Figure 4: A homeomorphism from the plane to itself that maps the vertices
of one of the dashed graph to the vertices to the other dashed graph, as well
as their edges, would map a set that it compact to a set that is not.

induces an equivalence between G1 and G2 . We construct a homeomorphism

φ∗ from R2 ∪ {∞} to itself that induces an equivalence between G∗1 and G∗2 .
We do so by defining it first on the vertices of G∗1 , then extending it on the
edges, and finally to the faces. Because φ is a homeomorphism it maps faces
of G1 to faces of G2 . Consider the following diagram:

20
 V1 E1 F1 V1∗ E1∗ F1∗

V2 E2 F2 V2∗ E2∗ F2∗

The dashed lines represent one-to-one correspondences given by the dualities,

while the non-dashed lines represent the one-to-one correspondences given
by the equivalences. The correspondences between the elements of the duals
are not defined yet, since they are determined by φ∗ . First we define φ∗
on V1∗ such that the diagram commutes. Again by the fact that φ is a
homeomorphism, it follows that if we pair the edges of G∗1 with the edges of G∗2
such that the digram commutes, corresponding edges will have corresponding
endpoints. We therefore can extend φ∗ to the edges on E1∗ such that it
maps edges to edges and induces the correspondence for which the diagram
commutes. We proceed in the same fashion to extend φ∗ to the faces. The key
is that by pairing the faces such that the digram commutes, corresponding
faces have corresponding frontiers. Also, one uses the fact that every face is
homeomorphic to the open unit disk.

7 Percolation Theory
In Percolation Theory, a similar duality to that presented in the previous sec-
tion plays a prominent role. Consider Z2 ⊂ R2 . Consider the infinite graph,
where the vertices are the points in Z2 and there is an edge between any two
points that are separated by an Euclidean distance of 1. A function from
the set of edges to {0, 1} is called a configuration. We define a probability
space on the set of all configurations. In a given configuration, an edge is
called open if it is mapped to 1, and closed otherwise. For an edge e, define
Ae to be subset of all configurations for which e is open. Our σ-algebra will
be the one generated by all such subsets. Take an arbitrary p ∈ [0, 1]. We
call it the percolation parameter. There is a unique probability measure Pp
on our σ-algebra, such that given any edge e, the probability Pp (Ae ) equals

21
p, and the set of all events of the form Ae is independent. This will be our
probability measure.

(a) p = 1/4 (b) p = 3/4

(c) p = 1/2

Figure 5: Simulations of for distinct values of p on an excerpt of the infinite

graph. Only the open edges are shown.

A path is a finite sequence of distinct points in Z2 , such that consecutive

points have Euclidean distance 1. Consider a path of strictly more than two
points, for which the first point has Euclidean distance of 1 to the last point.
For such a path, consider the union of the set of edges that join consecutive
points with the edge that joins the first point with the last. Such a set of
edges is called a cycle. Given a configuration, a path for which consecutive
points are joined by open edges is said to be open, or closed if those edges
are closed. A cycle that consists of open edges is likewise said to be open, or

22
closed if it consists of closed edges. An open cluster is a subset of Z2 , such
that any two points can be joined by an open path within the subset, and is
maximal in the sense that there is no subset strictly containing it that also
satisfies this condition.
The name Percolation Theory has the following explanation. Our model
is a two-dimensional analog of what could be seen as a model for some porous
material, the open edges being conduits where a liquid can pass through. An
interesting question is how connected these conduits are, in dependance of p.
Suppose this is a rock. If the conduits were only sparely connected, water
would only penetrate superficially. On the contrary, if they were largely
connected, forming large tunnels that go trough the rock, water would strain
through.
A formal question we can ask in this regard is: Given a point in Z2 ,
what is the probability that the open cluster containing this point is infinite?
First, that probability is independent of the given point, so we can assume
it to be the origin. Let C be the event: “The open cluster containing the
origin is infinite”. Our probability is then Pp (C). If p = 0, this probability
is 0. If p = 1, this probability is 1. But what exactly happens in the middle?
Consider yet another analogous model, for which we take Z instead of Z2 .
This case is easier. Here, if p < 1, the probability for having only open edges
on at least one of both sides of the origin is 0, as is any probability for an
event that fixes the state of an infinite number of edges. Therefore, unless p
is 1, the probability for having the open cluster containing the origin to be
infinite, is also 0. Is the same true for our model in Z2 ? The answer is no,
as we will prove using duality.

Claim 7.1. There is p < 1 such that Pp (C) > 0.

Recall how we defined the dual graph in the previous section. We proceed
similarly here. Consider our infinite graph with vertices Z2 and the edges as
previously defined. Let it be denoted by L. Consider the infinite graph L∗
that has vertices the points of the form (1/2 + a, 1/2 + b), where a, b ∈ Z, and
edges the straight line segments that connect points of distance 1. Each edge
of L intersects with exactly one edge of L∗ and vice versa. We therefore have a
correspondence between the edges of L and those of L∗ . For a configuration
of L, the dual configuration is the configuration of L∗ where open edges
correspond to open edges. A cycle in L∗ divides the plane into two connected
components, and only one is bounded. We say that a cycle in L∗ contains the
origin, if the origin is contained in the bounded connected component. Using

23
a similar argument as in the proof of Lemma 6.3, one obtains the following
lemma. Compare with Figure 6.
Lemma 7.2. Given a configuration of L, the open cluster containing the
origin is finite if and only if there is a closed cycle in the dual configuration
that contains the origin.

Figure 6: A depiction of the excerpt of a configuration around the origin.

Only the open edges are shown. The cluster containing the origin is finite.
The dotted edges form a closed cycle in the dual configuration, which contains
the origin.

We are now ready to prove Claim 7.1. Let C 0 be the event: “There
is a closed cycle in the dual configuration that contains the origin”. By
Lemma 7.2 we have:
Pp (C) = 1 − Pp (C 0 ).
We look for an estimate for Pp (C 0 ). Consider the positive part of the real
line. Any cycle in L∗ that contains the origin, must intersect this set at
least once. This can only happen at the points of the form (0, −1/2 + a), for
a ∈ N. A cycle of length n that contains the origin must intersect the positive
part of the real line somewhere before n. Also, for a given point of the form
(0, −1/2 + a), with a ∈ N, there are not more than 3n cycles of length n that
intersect that point. To see this, imagine constructing a cycle edge by edge,
starting with the edge that intersects the given point on the positive real

24
line, then placing the second edge incident on the upper endpoint of the first
edge, and continuing to place each new edge such that it is incident with
the previous one. Each time we add a new edge there are not more three
possibilities for placing it. Therefore n3n is an upper bound for the numbers
of cycles of length n that contain the origin. The probability for a given cycle
of length n to be closed is (1 − p)n . We therefore have:
∞
X
0
Pp (C ) ≤ (1 − p)n n3n .
n=1

The right hand side converges for p > 2/3, and approaches 0 when p ap-
proaches 1. From this the claim follows. This is just a glimpse of how
duality is used in Percolation Theory.
Much more can be proved about the behavior of Pp (C) in dependance of p.
As a function in p, we have that Pp (C) is monotonically non-decreasing, and
it turns out that there is pc ∈ [0, 1], called the critical value, such that:

= 0, for p < p .
c
Pp (C)
> 0, for p > pc .

Claim 7.1 can be used as a key ingredient to prove this. By the use
of duality one can show that in fact pc = 1/2. This requires substantially
more work than the proof of Claim 7.1. See Grimmett [14], Chapter 1 and
Chapter 11, for a treatment of these results.

25
References
[1] H. S. M. Coxeter, Projective Geometry, Second Edition, Springer, 2003.

[2] A. N. Whitehead, The Axioms of Projective Geometry, Cambridge Uni-

versity Press, 1913.

[3] S. Lang, Algebra, Springer, 2002.

[4] G. Fischer, Lineare Algebra, Springer, 2008.

[5] N. Bourbaki, General Topology, Volume 2, Hermann, Paris, 1966.

[6] B. von Querenburg, Mengentheoretische Topologie, Springer, 2001.

[7] J. Munkres, Topology, Second Edition, Pearson, 2000.

[8] W. Rudin, Fourier Analysis on Groups, Intersience Publishers, New York,

1962.

[9] A. Deitmar and S. Echterhoff, Principles of Harmonic Analysis, Springer,

2014.

[10] S. A. Morris, Pontryagin Duality and the Structure of Locally Compact

Abelian Groups, Cambridge University Press, 1977.

[11] L. Pontrjagin, The Theory of Topological Commutative Groups, The

Annals of Mathematics, Volume 35, Number 2, 1934.

[12] S. Mac Lane, Categories for the Working Mathematician, Springer, 1998.

[13] R. Diestel, Graph Theory, Springer, 2005.

[14] G. Grimmett, Perlocation, Springer, 1999.

26